Philip E. Pace-Detecting and Classifying Low Probability of Intercept Radar (2009) PDF

Detecting and Classifying
Low Probability of Intercept Radar

Second Edition
DISCLAIMER OF WARRANTY
The technical descriptions, procedures, and computer programs in this book
have been developed with the greatest of care and they have been useful to the
author in a broad range of applications; however, they are provided as is, with-
out warranty of any kind. Artech House, Inc. and the author and editors of the
book titled Detecting and Classifying Low Probability of Intercept Radar, Second
Edition make no warranties, expressed or implied, that the equations, programs,
and procedures in this book or its associated software are free of error, or are
consistent with any particular standard of merchantability, or will meet your
requirements for any particular application. They should not be relied upon for
solving a problem whose incorrect solution could result in injury to a person or
loss of property. Any use of the programs or procedures in such a manner is at
the users own risk. The editors, author, and publisher disclaim all liability for
direct, incidental, or consequent damages resulting from use of the programs or
procedures in this book or the associated software.
For a listing of recent titles in the

Artech House Radar Library,
turn to the back of this book.
Detecting and Classifying
Low Probability of Intercept Radar
Second Edition
Phillip E. Pace
Library of Congress Cataloging-in-Publication Data
A catalog record for this book is available from the Library of Congress.
British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library.
ISBN-13 978-1-59693-234-0
Cover design by Igor Valdman
2009 ARTECH HOUSE

685 Canton Street
Norwood, MA 02062
All rights reserved. Printed and bound in the United States of America. No part of this book may
be reproduced or utilized in any form or by any means, electronic or mechanical, including pho-
tocopying, recording, or by any information storage and retrieval system, without permission in
writing from the publisher. All terms mentioned in this book that are known to be trademarks or
service marks have been appropriately capitalized. Artech House cannot attest to the accuracy of
this information. Use of a term in this book should not be regarded as affecting the validity of
any trademark or service mark.
10 9 8 7 6 5 4 3 2 1
Disclaimer:
This eBook does not include the ancillary media that was
packaged with the original printed version of the book.
To my wife,
Ann Marie Pace,
and
to our children,
Amanda, Zachary, and Molly
Contents
Foreword xix
Preface xxi
Acknowledgments xxix
PART I: FUNDAMENTALS OF LPI RADAR DESIGN 1
1 To See and Not Be Seen 3

1.1 The Requirement for LPI . . . . . . . . . . . . . . . . . . . . 3
1.2 Characteristics of LPI Radar . . . . . . . . . . . . . . . . . . 5
1.2.1 Antenna Considerations . . . . . . . . . . . . . . . . . 5
1.2.2 Achieving Ultra-Low Side Lobes . . . . . . . . . . . . 7
1.2.3 Antenna Scan Patterns for Search Processing . . . . . 10
1.2.4 Advanced Multifunction RF Concept . . . . . . . . . . 13
1.2.5 Transmitter Considerations . . . . . . . . . . . . . . . 14
1.2.6 Power Management . . . . . . . . . . . . . . . . . . . 16
1.2.7 Carrier Frequency Considerations . . . . . . . . . . . . 17
1.3 Pulse CompressionThe Key to LPI Radar . . . . . . . . . . 18
1.4 Radar Detection Range . . . . . . . . . . . . . . . . . . . . . 24
1.5 Interception Range . . . . . . . . . . . . . . . . . . . . . . . . 27
1.6 Comparing Radar Range and Interception Range . . . . . . . 29
1.7 The Pilot LPI Radar . . . . . . . . . . . . . . . . . . . . . . . 31
1.8 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . 36
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2 LPI Technology and Applications 41

2.1 Altimeters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 41
2.1.2 Fielded LPI Altimeters . . . . . . . . . . . . . . . . . 42
2.2 Landing Systems . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 45
vii
viii Detecting and Classifying LPI Radar
2.2.2 Fielded LPI Landing Systems . . . . . . . . . . . . . . 46

2.3 Surveillance and Fire Control Radar . . . . . . . . . . . . . . 48
2.3.1 Battlefield Awareness . . . . . . . . . . . . . . . . . . 48
2.3.2 LPI Ground-Based Systems . . . . . . . . . . . . . . . 48
2.3.3 LPI Airborne Systems . . . . . . . . . . . . . . . . . . 56
2.4 Antiship Capable Missile and Torpedo Seekers . . . . . . . . 58
2.4.1 A Significant Threat to Surface Navies . . . . . . . . . 58
2.4.2 Fielded LPI Seeker Systems . . . . . . . . . . . . . . . 58
2.5 Summary of LPI Radar Systems . . . . . . . . . . . . . . . . 62
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3 Ambiguity Analysis of LPI Waveforms 67

3.1 The Ambiguity Function . . . . . . . . . . . . . . . . . . . . . 68
3.2 Periodic Autocorrelation Function . . . . . . . . . . . . . . . 68
3.3 Periodic Ambiguity Function . . . . . . . . . . . . . . . . . . 69
3.3.1 Periodicity of the PAF . . . . . . . . . . . . . . . . . . 70
3.3.2 Peak and Integrated Side Lobe Levels . . . . . . . . . 70
3.4 Frank Phase Modulation Example . . . . . . . . . . . . . . . 71
3.4.1 Transmitted Waveform . . . . . . . . . . . . . . . . . . 71
3.4.2 Simulation Results . . . . . . . . . . . . . . . . . . . . 72
3.5 Reducing the Doppler Side Lobes . . . . . . . . . . . . . . . . 75
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4 FMCW Radar 81
4.1 Advantages of FMCW . . . . . . . . . . . . . . . . . . . . . . 81
4.2 Single Antenna LPI Radar for Target Detection . . . . . . . . 83
4.3 Transmitted Waveform Design . . . . . . . . . . . . . . . . . 86
4.3.1 Triangular Waveform . . . . . . . . . . . . . . . . . . . 86
4.3.2 Waveform Spectrum . . . . . . . . . . . . . . . . . . . 89
4.3.3 Generating Linear FM Waveforms . . . . . . . . . . . 91
4.4 Receiver-Transmitter Isolation . . . . . . . . . . . . . . . . . 94
4.4.1 Transmission Line Basics . . . . . . . . . . . . . . . . 95
4.4.2 Single Antenna Isolation Using a Circulator . . . . . . 96
4.4.3 Single Antenna Isolation Using a Reflected Power
Canceler . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.5 The Received Signal . . . . . . . . . . . . . . . . . . . . . . . 100
4.6 LPI Search Mode Processing . . . . . . . . . . . . . . . . . . 101
4.7 Track Mode Processing Techniques . . . . . . . . . . . . . . . 104
4.8 Eect of Sweep Nonlinearities . . . . . . . . . . . . . . . . . . 105
4.9 Moving Target Indication Filtering . . . . . . . . . . . . . . . 107
4.10 Matched Receiver Response . . . . . . . . . . . . . . . . . . . 108
4.11 Mismatched Receiver Response . . . . . . . . . . . . . . . . . 110
Table of Contents ix
4.12 PANDORA FMCW Radar . . . . . . . . . . . . . . . . . . . 113

4.13 Electronic Attack Considerations . . . . . . . . . . . . . . . . 115
4.14 Technology Trends for FMCW Emitters . . . . . . . . . . . . 115
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5 Phase Shift Keying Techniques 125

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.2 The Transmitted Signal . . . . . . . . . . . . . . . . . . . . . 126
5.3 Binary Phase Codes . . . . . . . . . . . . . . . . . . . . . . . 128
5.4 Polyphase Codes . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.5 Polyphase Barker Codes . . . . . . . . . . . . . . . . . . . . . 134
5.6 Frank Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.7 P1 Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
5.8 P2 Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5.9 P3 Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5.10 P4 Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
5.11 Polytime Codes . . . . . . . . . . . . . . . . . . . . . . . . . . 163
5.11.1 T1(n) Code . . . . . . . . . . . . . . . . . . . . . . . . 163
5.11.2 T2(n) Code . . . . . . . . . . . . . . . . . . . . . . . . 165
5.11.3 T3(n) Code . . . . . . . . . . . . . . . . . . . . . . . . 169
5.11.4 T4(n) Code . . . . . . . . . . . . . . . . . . . . . . . . 169
5.12 Omnidirectional LPI Radar . . . . . . . . . . . . . . . . . . . 179
5.13 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
6 Frequency Shift Keying Techniques 187

6.1 Advantages of the FSK Radar . . . . . . . . . . . . . . . . . . 187
6.2 Description of the FSK CW Signal . . . . . . . . . . . . . . . 189
6.3 Range Computation in FSK Radar . . . . . . . . . . . . . . . 189
6.4 Costas Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
6.4.1 Characteristics of a Costas Array or Sequence . . . . . 191
6.4.2 Computing the Dierence Triangle . . . . . . . . . . . 192
6.4.3 Deriving the Costas Sequence PAF . . . . . . . . . . . 192
6.4.4 Welch Construction of Costas Arrays . . . . . . . . . . 193
6.5 Hybrid FSK/PSK Technique . . . . . . . . . . . . . . . . . . 195
6.5.1 Description of the FSK/PSK Signal . . . . . . . . . . 195
6.6 Matched FSK/PSK Signaling . . . . . . . . . . . . . . . . . . 199
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
x Detecting and Classifying LPI Radar
7 Noise Techniques 207

7.1 Historical Perspective . . . . . . . . . . . . . . . . . . . . . . 207
7.2 Ultrawideband Considerations . . . . . . . . . . . . . . . . . . 210
7.3 Principles of Random Noise Radars . . . . . . . . . . . . . . . 212
7.4 Narayanan Random Noise Radar Design . . . . . . . . . . . . 215
7.4.1 Operating Characteristics . . . . . . . . . . . . . . . . 216
7.4.2 Model of RNR Transmitter . . . . . . . . . . . . . . . 219
7.4.3 Periodic Ambiguity Results . . . . . . . . . . . . . . . 219
7.5 Random Noise Plus FMCW Radar . . . . . . . . . . . . . . . 222
7.5.1 RNFR Spectrum . . . . . . . . . . . . . . . . . . . . . 223
7.5.2 Model of RNFR Transmitter . . . . . . . . . . . . . . 225
7.6 Random Noise FMCW Plus Sine . . . . . . . . . . . . . . . . 227
7.6.1 Model of RNFSR Transmitter . . . . . . . . . . . . . . 229
7.7 Random Binary Phase Modulation . . . . . . . . . . . . . . . 234
7.7.1 Model of RBPC Transmitter . . . . . . . . . . . . . . 236
7.8 Millimeter Wave Noise Radar . . . . . . . . . . . . . . . . . . 238
7.9 Correlation Receiver Techniques . . . . . . . . . . . . . . . . 238
7.9.1 Ideal Correlation . . . . . . . . . . . . . . . . . . . . . 239
7.9.2 Digital-Analog Correlation . . . . . . . . . . . . . . . 239
7.9.3 Fully Digital Correlation . . . . . . . . . . . . . . . . . 241
7.9.4 Acousto-Optic Correlation . . . . . . . . . . . . . . . . 242
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
8 Over-the-Horizon Radar 249

8.1 Two Types of OTHR . . . . . . . . . . . . . . . . . . . . . . . 249
8.2 Sky Wave OTHR . . . . . . . . . . . . . . . . . . . . . . . . . 252
8.2.1 Characteristics of the Ionosphere . . . . . . . . . . . . 253
8.2.2 Example of F2-Layer Propagation . . . . . . . . . . . 259
8.2.3 Doppler Clutter Spectrum . . . . . . . . . . . . . . . . 259
8.2.4 Example Sky Wave OTHR System . . . . . . . . . . . 261
8.2.5 Sky Wave Processing . . . . . . . . . . . . . . . . . . . 261
8.3 Sky Wave LPI Waveform Considerations . . . . . . . . . . . . 265
8.3.1 Phase Modulation Techniques . . . . . . . . . . . . . . 265
8.3.2 Costas Frequency Hopping . . . . . . . . . . . . . . . 266
8.3.3 Reducing the CIT . . . . . . . . . . . . . . . . . . . . 266
8.3.4 Multiple Waveform Repetition Frequencies . . . . . . 266
8.3.5 Out-of-Band Emission Suppression . . . . . . . . . . . 270
8.4 Sky Wave Maximum Detection Range . . . . . . . . . . . . . 271
8.5 Sky Wave Footprint Prediction . . . . . . . . . . . . . . . . . 274
Table of Contents xi
8.6 Surface Wave OTHR . . . . . . . . . . . . . . . . . . . . . . . 276

8.6.1 Example Surface Wave OTHR System . . . . . . . . . 281
8.7 Surface Wave LPI Waveform Considerations . . . . . . . . . . 282
8.7.1 FMICW Characteristics . . . . . . . . . . . . . . . . . 282
8.7.2 FMICW Ambiguity Space . . . . . . . . . . . . . . . . 287
8.8 Surface Wave Maximum Detection Range . . . . . . . . . . . 288
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
9 Case Study: Antiship LPI Missile Seeker 301

9.1 History of ASCM Seeker Technology . . . . . . . . . . . . . . 301
9.2 The Future for ASCM Technology . . . . . . . . . . . . . . . 302
9.3 Detecting the Threat . . . . . . . . . . . . . . . . . . . . . . . 305
9.4 ASCM Target Scenario . . . . . . . . . . . . . . . . . . . . . . 306
9.4.1 Low RCS Targets . . . . . . . . . . . . . . . . . . . . . 306
9.4.2 Sea Clutter Model . . . . . . . . . . . . . . . . . . . . 308
9.4.3 Linear FMCW Emitter Power Management . . . . . 310
9.4.4 Target-to-Clutter Ratio . . . . . . . . . . . . . . . . . 312
9.5 ASCM Ship Target Model . . . . . . . . . . . . . . . . . . . . 315
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
10 Network-Centric Warfare and Netted LPI Radar Systems 319

10.1 Network-Centric Warfare . . . . . . . . . . . . . . . . . . . . 319
10.1.1 NCW Requirements . . . . . . . . . . . . . . . . . . . 322
10.1.2 Situational Awareness . . . . . . . . . . . . . . . . . . 323
10.1.3 Maneuverability . . . . . . . . . . . . . . . . . . . . . 323
10.1.4 Decision Speed and Operational Tempo . . . . . . . . 324
10.1.5 Agility . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
10.1.6 Lethality . . . . . . . . . . . . . . . . . . . . . . . . . 326
10.2 Metrics for Information Grid Analysis . . . . . . . . . . . . . 326
10.2.1 Generalized Connectivity Measure . . . . . . . . . . . 326
10.2.2 Reference Connectivity Measure . . . . . . . . . . . . 328
10.2.3 Network Reach . . . . . . . . . . . . . . . . . . . . . . 329
10.2.4 Suppression Example . . . . . . . . . . . . . . . . . . 331
10.2.5 Extended Generalized Connectivity Measure . . . . . 333
10.2.6 Entropy and Network Richness . . . . . . . . . . . . . 333
10.2.7 Maximum Operation Tempo . . . . . . . . . . . . . . 336
10.3 Electronic Attack . . . . . . . . . . . . . . . . . . . . . . . . . 337
10.4 Information Network Analysis Using LPIsimNet . . . . . . . . 338
10.5 Netted LPI Radar Systems . . . . . . . . . . . . . . . . . . . 342
10.5.1 Advantages of the Netted LPI Radar Systems . . . . . 346
10.5.2 Netted LPI Radar Sensitivity . . . . . . . . . . . . . . 348
xii Detecting and Classifying LPI Radar
10.5.3 Signal Model . . . . . . . . . . . . . . . . . . . . . . . 349

10.5.4 Netted Radar Electronic Attack . . . . . . . . . . . . 352
10.6 Netted Radar Analysis Using LPIsimNet . . . . . . . . . . . . 353
10.6.1 Monostatic LPI Emitter and the SNR Contour Chart 353
10.6.2 Three Netted LPI Emitters . . . . . . . . . . . . . . . 354
10.6.3 Two Netted LPI Emitters with Jammer . . . . . . . . 358
10.7 Orthogonal Waveforms for Netted Radar . . . . . . . . . . . . 358
10.7.1 Orthogonal Polyphase Codes . . . . . . . . . . . . . . 362
10.7.2 Addressing Doppler Shift Degradation . . . . . . . . . 365
10.7.3 Orthogonal Frequency Hopping Sequences . . . . . . . 370
10.7.4 Noise Waveforms . . . . . . . . . . . . . . . . . . . . 374
10.8 Netted Over-the-Horizon Radar Systems . . . . . . . . . . . . 377
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
PART II: INTERCEPT RECEIVER STRATEGIES

AND SIGNAL PROCESSING 385
11 Strategies for Intercepting LPI Radar Signals 387

11.1 EW Intercept Receiver Techniques . . . . . . . . . . . . . . . 387
11.1.1 Traditional Approach . . . . . . . . . . . . . . . . . . 387
11.1.2 The Look-Through Problem . . . . . . . . . . . . . . . 388
11.1.3 Modern Network-Centric Concepts Arriving . . . . . . 389
11.2 Detecting the LPI Radar with UAVs . . . . . . . . . . . . . . 391
11.3 Noncooperative Intercept Receivers . . . . . . . . . . . . . . 392
11.3.1 Comparison of Classic Receiver Architectures
for Detecting LPI Waveforms . . . . . . . . . . . . . . 392
11.3.2 Digital EW Receivers . . . . . . . . . . . . . . . . . . 396
11.3.3 Direct RF Sampling . . . . . . . . . . . . . . . . . . . 398
11.4 Demodulation of the LPI Waveform . . . . . . . . . . . . . . 400
11.5 EW Receiver Challenges . . . . . . . . . . . . . . . . . . . . . 400
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
12 Wigner-Ville Distribution Analysis of LPI Radar

Waveforms 405
12.1 Wigner-Ville Distribution . . . . . . . . . . . . . . . . . . . . 406
12.1.1 Continuous WVD . . . . . . . . . . . . . . . . . . . . 406
12.1.2 Example Calculation: Real Input Signal . . . . . . . . 409
12.1.3 Example Calculation: Complex Input Signal . . . . . 411
12.1.4 Two-Tone Input Signal Results . . . . . . . . . . . . . 414
12.2 FMCW Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 419
12.3 BPSK Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 421
12.4 Polyphase Code Analysis . . . . . . . . . . . . . . . . . . . . 426
Table of Contents xiii
12.5 Polytime Code Analysis . . . . . . . . . . . . . . . . . . . . . 429

12.6 Distinguishing Between Phase Codes . . . . . . . . . . . . . . 431
12.7 FSK and FSK/PSK Analysis . . . . . . . . . . . . . . . . . . 438
12.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
13 Choi-Williams Distribution Analysis of LPI Radar

Waveforms 445
13.1 Mathematical Development . . . . . . . . . . . . . . . . . . . 446
13.2 LPI Signal Analysis . . . . . . . . . . . . . . . . . . . . . . . 448
13.2.1 FMCW Analysis . . . . . . . . . . . . . . . . . . . . . 449
13.2.2 BPSK Analysis . . . . . . . . . . . . . . . . . . . . . . 449
13.2.3 Polyphase Code Analysis . . . . . . . . . . . . . . . . 455
13.2.4 Polytime Code Analysis . . . . . . . . . . . . . . . . . 455
13.2.5 FSK and FSK/PSK Analysis . . . . . . . . . . . . . . 458
13.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
14 LPI Radar Analysis Using Quadrature Mirror Filtering 467

14.1 Time-Frequency Wavelet Decomposition . . . . . . . . . . . . 468
14.1.1 Basis Functions . . . . . . . . . . . . . . . . . . . . . . 468
14.1.2 Short-Time Fourier Transform Decomposition . . . . . 469
14.1.3 Wavelets and the Wavelet Transform . . . . . . . . . . 469
14.1.4 Wavelet Filters . . . . . . . . . . . . . . . . . . . . . . 472
14.2 Discrete Two-Channel Quadrature Mirror Filter Bank . . . . 474
14.3 Tree Structure to Filter the Lowpass Component . . . . . . . 476
14.4 Tree Structure to Filter the Highpass Component . . . . . . . 477
14.5 QMFB Tree Receiver . . . . . . . . . . . . . . . . . . . . . . . 478
14.6 Example Calculations . . . . . . . . . . . . . . . . . . . . . . 482
14.6.1 Complex Single-Tone Signal . . . . . . . . . . . . . . . 482
14.6.2 Complex Two-Tone Signal . . . . . . . . . . . . . . . . 485
14.7 FMCW Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 487
14.8 BPSK Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 489
14.11 Costas Frequency Hopping Analysis . . . . . . . . . . . . . . 499
14.12 FSK/PSK Signal Analysis . . . . . . . . . . . . . . . . . . . 499
14.13 Noise Waveform Analysis . . . . . . . . . . . . . . . . . . . . 499
14.14 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510
xiv Detecting and Classifying LPI Radar
15 Cyclostationary Spectral Analysis for Detection of LPI

Radar Parameters 513
15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
15.1.1 Cyclic Autocorrelation . . . . . . . . . . . . . . . . . . 514
15.1.2 Spectral Correlation Density . . . . . . . . . . . . . . 515
15.2 Spectral Correlation Density Estimation . . . . . . . . . . . . 516
15.3 Discrete Time Cyclostationary Algorithms . . . . . . . . . . . 520
15.3.1 The Time-Smoothing FFT Accumulation Method . . 520
15.3.2 Direct Frequency-Smoothing Method . . . . . . . . . . 522
15.4 Test Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
15.5 BPSK Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 528
15.6 FMCW Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 531
15.9 Costas Frequency Hopping Results . . . . . . . . . . . . . . . 540
15.10 Random Noise Analysis . . . . . . . . . . . . . . . . . . . . . 543
15.11 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 545
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 547
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 548
16 Antiradiation Missiles 551

16.1 Suppression of Enemy Air Defense . . . . . . . . . . . . . . . 551
16.1.1 The Beginning of SEAD . . . . . . . . . . . . . . . . . 553
16.1.2 Early ARM Developments . . . . . . . . . . . . . . . . 554
16.1.3 Vietnam . . . . . . . . . . . . . . . . . . . . . . . . . . 555
16.1.4 Post Vietnam . . . . . . . . . . . . . . . . . . . . . . . 556
16.1.5 Miniature Air-Launched Decoys . . . . . . . . . . . . . 558
16.2 Antiradiation Missile Seeker Design . . . . . . . . . . . . . . . 559
16.2.1 Antenna Design . . . . . . . . . . . . . . . . . . . . . . 559
16.2.2 Receiver and Signal Processing . . . . . . . . . . . . . 566
16.2.3 Dual-Mode Design . . . . . . . . . . . . . . . . . . . . 567
16.2.4 Signal Processing . . . . . . . . . . . . . . . . . . . . . 571
16.2.5 Future ARMsAddressing the LPI Emitter . . . . . . 572
16.3 ARM Performance Metrics . . . . . . . . . . . . . . . . . . . 577
16.4 Former Soviet Union and Warsaw Pact Allies . . . . . . . . . 578
16.4.1 AA-10 Alamo . . . . . . . . . . . . . . . . . . . . . . . 578
16.4.2 AS-4 Kitchen . . . . . . . . . . . . . . . . . . . . . . . 579
16.4.3 AS-5 Kelt . . . . . . . . . . . . . . . . . . . . . . . . . 580
16.4.4 AS-6 Kingfish . . . . . . . . . . . . . . . . . . . . . . . 581
16.4.5 AS-9 Kyle . . . . . . . . . . . . . . . . . . . . . . . . . 582
16.4.6 AS-11 Kilter . . . . . . . . . . . . . . . . . . . . . . . 584
16.4.7 Kh-27 . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
16.4.8 AS-12 Kegler . . . . . . . . . . . . . . . . . . . . . . . 585
16.4.9 AS-16 Kickback . . . . . . . . . . . . . . . . . . . . . . 587
Table of Contents xv
16.4.10 AS-17 Krypton . . . . . . . . . . . . . . . . . . . . . . 587

16.5 United States . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
16.5.1 Shrike . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
16.5.2 Standard ARM . . . . . . . . . . . . . . . . . . . . . . 591
16.5.3 HARM . . . . . . . . . . . . . . . . . . . . . . . . . . 591
16.5.4 AARGM . . . . . . . . . . . . . . . . . . . . . . . . . 592
16.5.5 Aordable Reactive Strike Missile . . . . . . . . . . . 593
16.5.6 Sidearm . . . . . . . . . . . . . . . . . . . . . . . . . . 593
16.5.7 Rolling Airframe Missile . . . . . . . . . . . . . . . . . 594
16.5.8 Army UAVs . . . . . . . . . . . . . . . . . . . . . . . . 595
16.6 France . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 596
16.7 United Kingdom . . . . . . . . . . . . . . . . . . . . . . . . . 597
16.8 Taiwan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 598
16.9 Germany . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600
16.10 Israel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
16.10.1 Harpy . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
16.10.2 STAR-1 . . . . . . . . . . . . . . . . . . . . . . . . . . 603
16.11 China . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 604
16.12 Anti-ARM Techniques . . . . . . . . . . . . . . . . . . . . . . 606
16.12.1 Decoys . . . . . . . . . . . . . . . . . . . . . . . . . . 607
16.12.2 Gazetchik . . . . . . . . . . . . . . . . . . . . . . . . 610
16.12.3 AN/TLQ-32 ARM-D Decoy . . . . . . . . . . . . . . 611
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 616
17 Autonomous Classification of LPI Radar Modulations 619

17.1 Classification Using Time-Frequency Imaging . . . . . . . . . 620
17.2 Classification Authority and Automation . . . . . . . . . . . . 621
17.2.1 Human-Computer Interface Considerations . . . . . . 621
17.2.2 Automation and the Human Operator . . . . . . . . . 622
17.2.3 Autonomous Modulation Classification . . . . . . . . . 623
17.3 Nonlinear Classification Networks . . . . . . . . . . . . . . . . 624
17.3.1 Single Perceptron Networks . . . . . . . . . . . . . . . 625
17.3.2 Multilayer Perceptron Networks . . . . . . . . . . . . 629
17.3.3 Radial Basis Function . . . . . . . . . . . . . . . . . . 632
17.4 Feature Extraction Signal Processing . . . . . . . . . . . . . . 634
17.4.1 Marginal Frequency Adaptive Binarization . . . . . . 634
17.4.2 Classification Results with Multilayer Perceptron . . . 638
17.4.3 Classification Results with Radial Basis Function
Network . . . . . . . . . . . . . . . . . . . . . . . . . . 642
17.4.4 Discussion of Classification Results . . . . . . . . . . . 647
17.5 Modified Feature Extraction Signal Processing . . . . . . . . 648
17.5.1 Lowpass Filtering for Cropping Consistency . . . . . . 648
17.5.2 Calculating the Marginal Frequency Distribution . . . 651
xvi Detecting and Classifying LPI Radar
17.5.3 Principal Components Analysis . . . . . . . . . . . . . 656

17.5.4 Classification Using Modified Feature Extraction . . . 660
17.5.5 Classification Results with the Multilayer Perceptron . 667
17.5.6 Classification Results with the Radial Basis Function . 674
17.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 682
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 682
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 685
18 Autonomous Extraction of Modulation Parameters 687

18.1 Emitter Clustering . . . . . . . . . . . . . . . . . . . . . . . . 687
18.2 Polyphase Parameters Using Wigner-Ville DistributionRadon
Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 688
18.2.1 Time-Frequency Algorithm Description . . . . . . . . 689
18.2.2 Testing the Algorithm . . . . . . . . . . . . . . . . . . 694
18.3 Polyphase Parameters from Quadrature
Mirror Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . 695
18.3.1 Wavelet Decomposition Algorithm Description . . . . 695
18.4 FMCW Parameters from Cyclostationary Bifrequency Plane . 699
18.4.1 Cyclostationary Algorithm Description . . . . . . . . . 700
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705
APPENDIXES
A Low Probability of Intercept Toolbox 709

A.1 Introduction to the LPIT . . . . . . . . . . . . . . . . . . . . 709
A.2 Naming Convention and Example . . . . . . . . . . . . . . . . 710
B Generating PAF Plots Using the LPIT Files 713
C Primitive Roots and Costas Sequences 715

C.1 Primes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715
C.2 Complete and Reduced Residue Systems . . . . . . . . . . . . 716
C.3 Primitive Roots . . . . . . . . . . . . . . . . . . . . . . . . . . 717
D LPIsimNet 721
D.1 Overview of LPIsimNet Architecture . . . . . . . . . . . . . . 721
D.1.1 Loading the Default Sensor Network . . . . . . . . . . 722
D.1.2 Building a Scenario File and Running the Simulation . 722
D.2 Setting the Node Properties . . . . . . . . . . . . . . . . . . . 726
D.3 Viewing the Simulation Results . . . . . . . . . . . . . . . . . 728
Table of Contents xvii
D.4 Adding a Moving Jammer to the Scenario . . . . . . . . . . . 731

D.5 Netted Radar with a Jammer . . . . . . . . . . . . . . . . . . 733
E PWVD for FMCW with F = 500 Hz 741
F PWVD for Frank Code with T = 64 ms 745
G PWVD Results for P1, P2, P3, and P4 Codes 749

G.1 P1 Code Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 749
G.2 P2 Code Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 749
G.3 P3 Code Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 752
G.4 P4 Code Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 752
H PWVD Results for Polytime Codes T2, T3, and T4 759

H.1 T2(2) Polytime Code . . . . . . . . . . . . . . . . . . . . . . . 759
H.2 T3(2) Polytime Code . . . . . . . . . . . . . . . . . . . . . . . 763
H.3 T4(2) Polytime Code . . . . . . . . . . . . . . . . . . . . . . . 763
I QMFB Results for FMCW with F = 500 Hz 771
J QMFB Results for 11-Bit BPSK 773
K QMFB Results for Frank Signal with Nc = 16 777
L QMFB Results for P1, P2, P3, and P4 781

L.1 P1 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 781
L.2 P2 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 782
L.3 P3 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 782
L.4 P4 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 788
M QMFB Results for T2, T3, and T4 797
N Cyclostationary Processing Results with FMCW,

F = 500 Hz 805
O Cyclostationary Processing Results with Frank Signal,

Nc = 16 809
P Cyclostationary Processing Results for P1, P2, P3,

and P4 813
P.1 P1 Code Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 813
P.2 P2 Code Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 816
P.3 P3 Code Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 816
P.4 P4 Code Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 816
xviii Detecting and Classifying LPI Radar
Q Cyclostationary Processing Results for T2, T3, and T4

Polytime Codes 821
Q.1 Polytime T2(2) Code Analysis . . . . . . . . . . . . . . . . . 821
List of Symbols 829
Glossary 841
About the Author 847
Index 849
Foreword
In the foreword of Detecting and Classifying Low Probability of Intercept

Radar, 1st Edition, I noted that there is considerable interest in radars that
can see and not be seen, commonly called low probability of intercept or
LPI radars. If anything, interest has grown in the intervening years and
this new book on the subject is both timely and welcome. The problem of LPI
radar design is dicult to solve for long range radars because the signal avail-
able to the listener is reduced by the square of the distance from transmitter
to listening receiver, whereas signal available to the radar receiver decreases
in proportion to the fourth power of the distance between the radar and its
target. Phillip E. Pace has included the many facets of LPI radar from his
earlier work and has added valuable insights in nearly every area. He has
also added much that is entirely new to this volume, including topics of noise
radar and network centric warfare and radar netting. He also considers the
interception problem and has added material on use of the Choi-Williams
distribution, as well as chapters on autonomous extraction and recognition
architectures. This coverage of both the radar and interception problems in
one volume provides a valuable reference work for this important technical
field.
As radar interception techniques evolved over the past half-century, the
generally high signal strength available to the intercept receiver led to inter-
cept receivers which detect each radar pulse using threshold detection and
then estimate parameters such as carrier frequency, angle of arrival, pulse du-
ration, time of arrival, polarization, and other single pulse parameters. These
form pulse descriptor words and are further sorted, deinterleaved and an-
alyzed to discern PRI patterns. This approach to signal interception and
threat recognition requires a high probability of detection for each individual
xix
xx Detecting and Classifying LPI Radar
pulse. Antiradiation missiles and other approaches to suppression of enemy

air defenses makes reduction of peak power a matter of radar survivability.
This in turn forces a reexamination of the single pulse detection approach for
signal interception as well as a reexamination of the use of high peak power
transmissions for performing radar functions.
Whether you are interested in techniques used in the design of LPI radar
or in techniques which may be useful for countering such LPI designs, this
book provides a good starting point for rethinking both the radar problem
and the interception problem.
Richard G. Wiley, Ph.D.

Vice President-Chief Scientist of Research Associates of Syracuse, Inc.
East Syracuse, New York
December, 2008
Preface
Introduction
The second edition of Detecting and Classifying Low Probability of Intercept

Radar is designed to meet the needs of electrical engineering, physics, and
systems engineering students at the senior undergraduate and beginning
graduate levels and especially those of practicing engineers. A low proba-
bility of intercept (LPI) radar course must present, as they say, both sides
of the story. Whereas radar proper has little appeal and seems even less
pointed to most of these students, the subject becomes highly significant to
them when it is presented along with the digital intercept receiver and signal
processing techniques for counter-LPI. My experiences as a student, engineer,
and teacher have led to the thought that a successful text for this study must
present both the radar design characteristics as well as the noncooperative
detection strategies and algorithms. In doing so, the course provides an inter-
esting opportunity to study the various trade-os that are involved not only
in intercept receiver architectures but also in the design of LPI waveform
modulations.
This book has grown out of research and teaching in the field of network-
centric radar electronic warfare, signal processing, and wideband digital re-
ceiver technologies at the Naval Postgraduate School. Even though the first
edition of this book was published barely four years ago, based on the helpful
reviews published in the IEEE Aerospace and Electronic Systems Magazine
and the feedback from the many students in industry and universities, it be-
came evident that a new edition was needed to incorporate the suggested
topics and changes to the contents.
LPI radar systems are seeing unprecedented levels of growth. In many
countries, new milestones are being established for streamlined acquisition of
these emitters for all types of applications. On the other hand, the recent
xxi
xxii Detecting and Classifying LPI Radar
advances in LPI radar technology have pushed the door open for the de-
sign of extremely sensitive intercept receivers and high-speed signal proces-
sors for autonomous LPI emitter detection, classification, and counter-LPI
operations.
Whats New
LPI radar techniques added to this second edition include; random noise radar
waveforms, their periodic ambiguity characteristics, and the dierent types
of correlation receivers used (Chapter 7); sky wave and surface wave over-
the-horizon radar systems and their move away from the traditional wave-
forms to the incorporation of new LPI modulations (Chapter 8); netted LPI
radar sensors and orthogonal polyphase modulations, network-centric warfare
principles, frequency hopping waveforms, and information network analysis
(Chapter 10).
New intercept receiver strategies and signal processing algorithms supplied
in the second edition include; the Choi-Williams time-frequency analysis of
LPI waveforms (Chapter 13); antiradiation missiles and the new seeker de-
signs for detecting LPI emitters (Chapter 16); autonomous feature
extraction and classification algorithms for identifying the intercepted modu-
lation (Chapter 17); and autonomous modulation parameter extraction signal
processing (Chapter 18).
A distinguishing feature of this book is investigating the LPI techniques
that go beyond the use of a single emitter and use a network to integrate sev-
eral distributed sensors to provide additional aspects of the target. Employing
a sensor network can unfold new capabilities in many important applications.
Secondly, this book examines extending the detection and classification algo-
rithms to execute autonomously, independent of any human interpretation to
the extent desired. Executing these modulation decisions autonomously can
draw these techniques closer to providing the intercept receiver the real-time
response capability needed for fast, reactive counter-LPI.
Course Structure
The book is written to serve not only as a textbook, but also as a reference for
the practicing radar and digital intercept receiver design engineer. The layout
was intended to be applicable to many dierent course structures including,
a one-semester (two quarters) course of study in low probability of intercept
radar systems design (Part I) and the noncooperative detection and classifi-
cation of these types of emitters (Part II). The book is especially appropriate
for 2-, 3-, and 4-day short courses. For the prerequisites, it is assumed that
the student has at least senior-level academic experience in engineering and
Preface xxiii
mathematics, and has the ability to write and run computer programs. A
course in radar and a course in signal processing would provide a very useful
background.
Overview of the Book

As with the first edition, this book is divided into two parts with the main
objective in Part I being the unified presentation of the fundamental design
principles of LPI radar. This includes a thorough treatment of the numerous
types of wideband modulations that can be used to reduce the probabil-
ity of a noncooperative intercept receivers ability to extract the waveform
parameters (which may easily lead to an eective jammer response). The
main objective in Part II is to present the intercept receiver time-frequency
and bifrequency signal processing techniques that can extract the wideband
waveform parameters. Autonomous classification and parameter extraction
algorithms are also an objective such that a real-time jammer response can be
developedjust what we did not want to happen in Part I! In summary, a bal-
anced coverage is provided of both LPI radar and waveform design concepts
(Part I) and the signal processing techniques for LPI waveform detection and
characterization for counter-LPI (Part II).
Each chapter ends with exercises that are an essential part of any course
using the text. A key feature of this book is the extensive use of MATLAB.
The CD accompanying this book contains many programs that should be
used for the problem exercises in order to further the understanding of the
concepts, and also to generate new and useful results that are of special
interest to the reader. The exercises are often used to complete the readers
understanding of a concept or to present dierent applications of ideas in the
text.
A distinguishing feature of this book is that it includes many graphical
illustrations of the results, especially in Part II. It is hoped that this will lead
to a better understanding of the underlying principles of waveform design and
will provide a clearer visualization of how the waveform parameters can be
extracted. As they say, a picture is worth a thousand words. Identification
of the waveform parameters is the first step to the development of autonomous
classification and parameter extraction algorithms.
The text contains sucient mathematical detail to enable the average
undergraduate electrical engineering and physics student to follow, without
too much diculty, the flow of analysis and design. A certain amount of
analytical detail, rigor, and thoroughness allows many of the topics to be
investigated further with the aid of many references. A brief overview of each
chapter is given below.
xxiv Detecting and Classifying LPI Radar
PART I:
Fundamentals of LPI Radar Design
In Chapter 1, an introduction to LPI radar is presented which provides the

reasons for the LPI requirement that include advanced intercept receivers
and the threat of antiradiation missiles. The characteristics of LPI radar
that distinguish them from conventional radar are also presented, as well as
the LPI radar architectures emphasizing continuous waveform (CW) radar.
The detection range of the LPI radar is examined and the advantage of the
LPI radar is quantified in terms of the intercept range and processing gain.
To illustrate the analysis, several examples using the Pilot LPI radar are
presented.
In Chapter 2, an updated and comprehensive review of the applications
that utilize LPI radar technology is presented. Applications include altime-
ters, surveillance, navigation, and landing radar for unmanned aerial vehicles
(UAVs). Also discussed are the tactical multimode airborne radar, antiship
capable missile (ASCM) seekers, and torpedo seekers.
In Chapter 3, the ambiguity analysis of LPI waveforms is introduced in
order to quantify their delay-Doppler properties. The concepts are used
throughout Part I to examine the various waveforms being studied. The
mathematical tools include the autocorrelation function (ACF), the periodic
autocorrelation function (PACF), and the periodic ambiguity function (PAF).
The eect of weighting functions on the PAF is also discussed. The low prob-
ability of intercept toolbox (LPIT) is a collection of MATLAB routines that
enable the student to quickly design all of the LPI waveforms. The LPIT is
introduced in Appendix A. Appendix B discusses the download of MATLAB
code from N. Levanons Web site in order to compute the ACF, PACF, and
PAF.
Chapter 4 investigates the characteristics of frequency modulation CW
(FMCW) LPI radar. A detailed architecture is analyzed. Mathematical for-
mulations of the transmitted waveform and the received signal are developed,
and there is an analysis of the receiver-transmitter isolation problems being
overcome (single antenna systems). The search mode signal processing is
described, including the details of the system components (e.g., filter band-
widths, analog-to-digital converter speeds, and so forth). Track mode process-
ing techniques are also presented. Nonlinearities in the frequency sweep wave-
form are addressed, and the PAF of the FMCW is analyzed. As an example
of an FMCW LPI radar, details of the Parallel Array for Numerous Dier-
ent Operational Research Activities (PANDORA) are presented. Finally, the
technology trends and latest developments in FMCW emitters are presented.
Preface xxv
In Chapter 5, phase shift keying (PSK) LPI radar is discussed. Details

on polyphase Barker sequences, Frank code, P1, P2, P3, and P4 codes are
presented, and their spectral and ambiguity properties investigated. Also
presented are polytime codes T1(n), T2(n), T3(n), and T4(n). As an example
of a phase coding LPI radar, the details of the Omnidirectional LPI radar are
presented.
Chapter 6 discusses frequency shift keying (FSK) radar waveform design.
The design of Costas codes is presented. By combining Costas coding with
PSK, an additional advantage is obtained for the LPI radar. Tailoring the
FSK/PSK waveform to the power spectral density of a particular target of
interest can improve detection probabilities by transmitting (randomly) at
those frequencies where the target resonates the most. This concept is also
presented and examples of the waveform are given.
In Chapter 7, random noise radar concepts are introduced. Four types are
presented including random noise, random noise plus FMCW, random noise
FMCW plus sine, and random binary phase code modulation. The ambiguity
analysis of the waveforms is discussed and the correlation receiver techniques
used in the radar receiver are examined including an acousto-optic approach.
In Chapter 8, over-the-horizon radar concepts are discussed emphasizing
the new movement away from the traditional FMCW waveforms to the more
LPI type waveforms. Ionospheric eects are presented and both surface wave
and sky wave emitter concepts are investigated. The maximum detection
range is also quantified for both types of emitters.
In Chapter 9, the design of LPI seekers for antiship capable missiles is
discussed. The design of a modern 9.3-GHz homodyne triangular-FMCW
emitter for detection of low radar cross section (RCS) ships is described. To
predict target detection capability, clutter and target models are developed
as the emitter is flown at 300 m/s in a scenario that starts at a range of 15
nmi from the target. To evaluate the feasibility of detecting low RCS ships
at the horizon, a low RCS ship design is examined. Each sea state (0-4) is
characterized by using a second-order polynomial that describes the normal-
ized mean sea backscatter coecient as a function of the grazing angle. The
emitter transmit power is adapted in time to measure the target character-
istics (power management). The emitter transmit power level is consistent
with the RCS and range to the target, while keeping a target-to-clutter power
ratio at 20 dB. For detection analysis, 50, 100, and 500 m2 RCS values are
considered.
In Chapter 10, the concept of network-centric warfare is introduced and
the use of a sensor network is analyzed. Performance of the information
grid is quantified. Netted radar concepts are introduced and minimum input
minimum output techniques are reviewed. Sensor network and netted radar
performance are examined including their capabilities under electronic attack
using the MATLAB program LPIsimNet. The use of orthogonal polyphase
modulations and orthogonal frequency hopping waveforms is also discussed.
xxvi Detecting and Classifying LPI Radar
PART II:
Intercept Receiver Strategies and Signal
Processing
To begin Part II, Chapter 11 takes a look at (noncooperative) digital intercept

receiver strategies. The trend today is toward the all-digital receiver with
the analog-to-digital conversion taking place directly at the antenna (direct
conversion). Network-centric and swarm intercept strategies are discussed.
The trade-os of various receiver architectures is presented and a new digital
analog-to-information receiver is discussed. Problems that intercept receivers
must deal with are presented as well as future trends in intercept receiver
architectures. For the remaining chapters, it is assumed that the sampled
signal is available within bulk memory of the receiver, and used as input to
the signal processor.
Chapter 12 examines the Wigner-Ville distribution (WD) time-frequency
analysis technique, including an ecient kernel transformation that helps
speed up the computation time. Two small examples are carried through
(real input signal and complex input signal) to demonstrate the WD time-
frequency calculation. A two-tone input signal is analyzed to further the un-
derstanding of the WD output and to demonstrate the presence of the cross
term. Although not an LPI waveform, the binary PSK (BPSK) signal is an-
alyzed first for various signal-to-noise ratios (SNRs), so that the WD results
can be verified and compared to other phase coding techniques. Extraction of
the signal parameters such as code period, subcode period, number of phase
codes, carrier frequency, and signal bandwidth is developed. The LPI wave-
forms developed in Part I are analyzed. These include the FMCW technique
and the phase coding techniques: Frank, P1, P2, P3, and P4. The advanced
phase coding techniques where the subcode width is not uniform throughout
the code period are examined next. These include the T1(n) through T4(n).
Using the WD, it is shown that the numerous LPI signals can be distinguished
and the signal parameters can be extracted, even for moderately low SNR.
The frequency coding techniques are examined last and include Costas se-
quences (FSK), Costas sequences with phase modulation (FSK/PSK), and
the target matched FSK/PSK signals.
In Chapter 13, the Choi-Williams distribution is presented. Using an ex-
ponential kernel and the same transformation as outlined for the Wigner-Ville
distribution, the amplitude of the cross terms is significantly reduced making
the identification of the modulation parameters easier. The LPI modulations
are calculated using the Choi-Williams to quantify the amplitude reduction
of the cross terms and to compare the results with those shown in Chapter
12. LPI modulations examined include FMCW, BPSK and polyphase mod-
ulations. Also examined are polytime, FSK, and FSK/PSK modulations.
Preface xxvii
Chapter 14 investigates the use of quadrature mirror filter banks (QMFBs)

for the extraction of LPI radar waveform parameters. The introduction of
time-frequency wavelets and the wavelet transform are presented first, fol-
lowed by the development of the discrete two-channel quadrature mirror filter
bank. This leads to a discussion on filtering the lowpass component and the
highpass component, and the arrangement of the filters into a tree structure.
The QMFB tree is then considered, and the results for a complex single-tone
signal are shown as an example of the time-frequency output. A complex
two-tone signal is then considered, followed by the QMFB analysis of the LPI
signals. This investigation then examines the LPI waveforms and parallels
the analysis carried out in Chapters 12 and 13, so a direct comparison of the
methods can be made.
The fundamentals of cyclostationary signal processing are presented in
Chapter 15. Discrete time algorithms are presented to generate the spectral
correlation density and include the time-smoothing fast Fourier transform
(FFT) accumulation method and the direct frequency smoothing method. A
single-tone test frequency is used to illustrate the cyclostationary results on
the bifrequency plane for both methods. The extraction of the waveform
parameters on the bifrequency plane provides some significant advantages
when compared to the time-frequency methods discussed in Chapters 1214.
Chapter 16 introduces the concept of suppression of integrated defense
systems using antiradiation missiles (ARMs). The ARM seeker and signal
processing are detailed and the algorithms used to address the LPI threat are
introduced. Performance metrics are examined and the important ARMs of
the world are presented. Anti-ARM techniques are also reviewed.
Chapter 17 examines the task of autonomously classifying the types of
signal modulation using time-frequency imaging and detection. Classification
authority and the human computer interface considerations are emphasized.
Feature extraction algorithms are presented and nonlinear neural network
classification architectures are introduced. Classification results using the
LPI emitter modulations discussed in Part I are presented.
Chapter 18 introduces the algorithms that can be used to autonomously
extract the modulation parameters from the time-frequency and bifrequency
results. The concept of emitter clustering is presented and the extraction
of polyphase modulation parameters from a WD-Radon transform algorithm
is discussed. Extraction of the polyphase modulation parameters from the
QMFB are also discussed. An algorithm for extracting FMCW parameters
from the cyclostationary bifrequency plane is presented. Results are shown
to illustrate the performance of the techniques.
xxviii Detecting and Classifying LPI Radar
Final Message
Every attempt has been made to ensure the accuracy of all materials in this
book, including the many MATLAB programs contained on the CD. I would,
however, appreciate readers bringing to my attention any errors that may
appear.
I have been extremely gratified by the tremendous success of this text.
The many improvements and additions in the second edition have been made
possible by the feedback and suggestions of a large number of instructors and
students at many companies and universities.
Finally, on a personal note, it continues to be very encouraging to learn
that many people working with or having to learn about detecting and clas-
sifying LPI radar systems have found the first edition useful. It is still my
hope that this second edition, with its new chapters and additional software,
will be of value not only to new readers, but will also be worthwhile to those
who have already read the first edition.
Acknowledgments
This book would not have been possible without the help, encouragement,
and support received during its preparation. First, I thank God for giving me
the strength and endurance to complete this work. I would also like to thank
my family Ann, Amanda, Zachary, and Molly. I could not have completed this
enormous task without their support, patience, sacrifice, and understanding
for the many hours of neglect during the completion of the first and second
editions of this book and it is to them to whom this book is dedicated.
I would also like to thank the following people who were invaluable in
reviewing the first edition of this work. Foremost, I would like to thank
Dr. David K. Barton, ANRO Engineering Inc., and Dr. Richard G. Wiley,
Research Associates of Syracuse, Inc., for taking the time to oer numerous
helpful suggestions that improved the quality of the manuscript. Many thanks
also go to Professor Nadav Levanon, Tel Aviv University, for working with
me tirelessly on the ambiguity analysis, and to Professor Herschel H. Loomis
Jr., Naval Postgraduate School, for helpful discussions in cyclostationary sig-
nal processing. I am also grateful to Professor David Styer, University of
Cincinnati, for sharing his insights into the world of number theory.
Reviewers for various portions of this second edition include Dr. Carlo
Kopp, defense analyst and consulting engineer, Air Power Australia for his
insights into antiradiation weapons, Dr. Ram Narayanan, Penn State Uni-
versity for his help with noise radar concepts, Dr. Jerey B. Knorr, Naval
Postgraduate School, for his many years of experience in the HF world, and
again Dr. David Barton, and Dr. Richard Wiley. I would also like to thank
graduate students Fernando Taboada, Antonio Lima, Jen Gau, Pedro Jarpa,
xxix
xxx Detecting and Classifying LPI Radar
Siew-Yam Yeo, and Christer Persson, Taylan Gulum, You-Chen, Bin-Yi Liu,
You-Quan Chen, Teresa and Gary Upperman, Patrick Kistner, Eugene R.
Heuschel III, Micael Grahn, Jason Phillips, Pick Guan Hui, and Sharon Ai
Lin Tan for their eort in helping develop the software tools, and the many
graduate students who have contributed their valuable time to understanding
the results in the text.
I am also very grateful to the sta of Artech House, especially Mark
Walsh, senior acquisitions editor, for his interest, support, and cooperation
of this second edition; Barbara Lovenvirth, developmental editor, for helping
me along; Erin Donahue, production editor, for the production of the book;
and Igor Valdman, for managing the production of the cover. It has been a
satisfying but sometimes overwhelming task.
Phillip E. Pace
Naval Postgraduate School
Monterey, CA
pepace@nps.edu
PART I:
FUNDAMENTALS OF LPI RADAR DESIGN

Chapter 1
To See and Not Be Seen

This chapter addresses the questions: What is a low probability of inter-
cept radar, and why is this capability needed? After answering these basic
questions, the radar design characteristics that make these type of sensors
dierent are presented. The radar range equation is used to quantify the de-
tection performance of an LPI radar design. The range at which an intercept
receiver can detect the LPI radar emission is also addressed. The Pilot radar
is used to illustrate a complete design, and its performance is also examined.
1.1 The Requirement for LPI

Many users of radar today are specifying a low probability of intercept (LPI)
and low probability of identification (LPID) as an important tactical require-
ment. As of 2008, the ANSI/IEEE Standard 686: Radar Terms and Defini-
tions, does not address this type of radar. The term LPI is that property of a
radar that, because of its low power, wide bandwidth, frequency variability,
or other design attributes, makes it dicult for it to be detected by means of
a passive intercept receiver. An LPID radar is an LPI radar with a waveform
that makes it dicult for an intercept receiver to correctly identify the para-
meters and radar type. More formal definitions for LPI and LPID are oered
below:
Definition 1.1
A low probability of intercept (LPI) radar is defined as a radar

that uses a special emitted waveform intended to prevent a non-
cooperative intercept receiver from intercepting and detecting its
emission.
3
4 Detecting and Classifying LPI Radar
Definition 1.2
Low probability of identification (LPID) radar is defined as a radar

that uses a special emitted waveform intended to prevent a non-
cooperative intercept receiver from intercepting and detecting its
emission but if intercepted, makes identification of the emitted
waveform modulation and its parameters dicult.
According to the definitions 1.1 and 1.2 above, an LPID radar is an LPI
radar but and LPI radar is not necessarily an LPID radar. It follows that the
LPI and LPID radar attempts detection of targets at longer ranges than the
intercept receiver can accomplish detection/jamming of the radar [13]. It is
important to note that defining a radar to be LPI and/or LPID necessarily
involves the definition of the corresponding intercept receiver. That is, the
success of an LPI radar is measured by how hard it is for the intercept receiver
to detect/intercept the radar emissions.
The LPI requirement is in response to the increase in capability of modern
intercept receivers to detect and locate a radar emitter [4]. One thing is for
certain. For every improvement in LPI radar, improvements in intercept re-
ceiver design can be expected (which is why this book addresses both areas).
In applications such as altimeters, tactical airborne targeting, surveillance,
and navigation, the interception of the radar transmission can quickly lead
to electronic attack (or jamming) if the parameters of the emitter can be
determined. Due to the wideband nature of these pulse compression wave-
forms, however, this is typically a dicult task. The LPI requirement is also
in response to the ever-present threat of being destroyed by precision guided
munitions and antiradiation missiles (ARMs). ARMs are designed to home
in on active, ground-based, airborne or shipboard radars, and disable them
by destroying their antenna systems and/or killing or wounding their opera-
tor crews [4]. ARMs are typically used for suppression of enemy air defense
(SEAD). The intercept receiver on board the aircraft (or the ARM system
itself) locates the victim radar. The victim radar is then designated to the
ARM if the parameters of the intercepted signal are correct. In Chapter
16, a thorough treatment of the ARM threat and the new signal processing
techniques to counter the LPI emitter are presented.
The denial of signal intercept protects the emitters from most of these
types of threats and is the objective of using a low probability of intercept
waveform. Since LPI radar tries to use signals that are dicult to inter-
cept and/or identify, they have dierent design characteristics compared to
conventional radar systems. These characteristics are discussed below.
To See and Not Be Seen 5
1.2 Characteristics of LPI Radar

Many combined features help the LPI radar prevent its detection by modern
intercept receivers. These features are centered on the antenna (antenna
pattern and scan patterns) and the transmitter (radiated waveform).
1.2.1 Antenna Considerations

The antenna is the interface, or connecting link, between some guiding system
and (usually) free space. Its function is to either radiate electromagnetic
energy (the transmitter feeds the guiding system) or receive electromagnetic
energy (the guiding system feeds a receiving system). The antenna pattern is
the electric field radiated as a function of the angle measured from boresight
(center of the beam). The various parts of a radiation pattern are referred
to as lobes that may be subclassified into main, side, and back lobes [5].
The main lobe is defined as the lobe containing the direction of maximum
radiation. The side lobe is a radiation lobe in any direction other than the
intended lobe. A back lobe refers to a lobe that occupies the hemisphere in
a direction opposite to that of the main lobe. The side lobe level is usually
expressed as a ratio of the power density in the lobe in question to that of the
main lobe. That is, the side lobe level is amplitude of the side lobe normalized
to the main beam peak. The highest side lobe is usually that lobe closest to
the main beam. It is also convenient to use the side lobe ratio (SLR) which
is the inverse of the side lobe level.
The radiation intensity of an antenna is the power per unit solid angle.
The power gain of an antennas main lobe is defined as 4 times the ratio of
the radiation intensity in the maximum direction to the net power accepted
by the antenna from the transmitter. The power gain can be estimated closely
using Krauss approximation [5]
4
G= (1.1)
a e
where a is the half-power beamwidth in the azimuth plane, e is the half-
power beamwidth in the elevation plane (in radians), and is the antenna
aperture eciency
Prad
= (1.2)
Pin
or the ratio of the radiated power of the antenna to the total input power.
The half-power beamwidth is the angle between two directions in which the
radiation intensity is one-half the maximum value of the beam. The gain of
the antenna can also be approximated using the physical aperture area A as
4A
G
2
For any antenna aperture, the antenna radiation pattern is obtained by

taking the Fourier transform of the field distribution across the aperture; for
example, in a rectangular aperture
0.88
a , e = (1.3)
da , de
where da is the aperture dimension in the azimuth plane and de is the aperture
dimension in the elevation plane (same dimensions as ).
There are two types of antenna beams that can be used. These are the pen-
cil beam and the fan beam. The pencil beam antenna pattern has a beamwidth
in the horizontal plane that is approximately equal to the beamwidth in the
vertical plane (e a ). The beamwidth for a radar pencil beam is generally
only a few degrees, since a small angular resolution is usually desired. From
(1.3), the resolution depends on the aperture size as well as the wavelength of
operation. For the fan beam pattern, one angular dimension is smaller than
the other (usually a < e to maintain good angular resolution in azimuth).
The bandwidth of the antenna is defined as the range of frequencies for
which the performance of the antenna conforms to a specific standard. It is
usually specified as a range of frequencies about the center frequency of radi-
ation. The polarization of a radiated waveform is that property of the wave
that describes the time-varying direction and relative magnitude of the elec-
tric field vector (the curve traced by the instantaneous electric field vector).
Polarization of the radiation can be linear, circular, or elliptical. Polarization
modulation can also reduce the probability of intercept.
A phased array is an array antenna whose beam direction or radiation
pattern is controlled primarily by the relative phases of the excitation coef-
ficients of the radiating elements. A single multifunction phased array radar
system can perform surveillance, fire control, communications, and electronic
warfare without requiring separate radars and antennas for these functions.
Phased arrays generally have bandwidths less than 10% and are steered by
using passive phase shifters that are controlled over electrical paths (usually
by digital signals).
More advanced phased arrays are being developed where the transmit and
receive modules employ photonic switching (at optical frequencies), allowing
high accuracy pointing and instantaneous beam positioning. They also al-
low multiple pulse compression modulation signals to be scanned over large
angles. An example of a recent pioneering development is shown in Figure
1.1. This figure shows the phased array used in the F-22 multimode fire
control radar [6]. The F-22s AN/APG-77 electronically scanned array an-
tenna is composed of several thousand transmit/receive modules, circulators,
radiators, and manifolds assembled into subarrays and then integrated into
a complete array. The baseline design used thousands of hand-soldered flex
circuit interconnects to make the numerous radio frequency, digital, and di-
rect current connections between the components and manifolds that make
Figure 1.1: Phased array antenna for the F-22 multimode radar [6].
up the subarray. The phased array aids the APG-77 with the capability to
transmit an LPI waveform. More of these types of systems are discussed in
Chapter 2.
1.2.2 Achieving Ultra-Low Side Lobes

The fields radiated from a linear array are a superposition (sum) of the fields
radiated by each element in the presence of the other elements. Each element
has an excitation parameter (current for a dipole, voltage for a slot, and mode
voltage for a multiple-mode element) [7]. The excitation of each element in
the aperture has a dierent amplitude and phase and is known as the aperture
distribution. The far-field radiation pattern is the discrete Fourier transform
of the array excitation. The array pattern can be written as
Ne
3
F (u) = An ej2(n1)u (1.4)
n=1
where An are the excitation coecients of the array which has Ne elements
and
d
u = (sin sin 0 ) (1.5)

and represents the angle from broadside, d the element spacing, and u the
array variable. The main lobe peak is at 0 . Using
w = ej2u (1.6)
(1.4) can be written as

Ne
3
F (u) = An wn1 (1.7)
n=1
If the aperture excitation is uniform (An = 1), it can be shown that [7]
sin Ne u j(Ne 1)u

F (u) = e (1.8)
Ne sin u
In this case the radiation intensity has a (sin x/x)2 pattern. The field strength
voltage pattern has a sin x/x pattern with a highest side lobe level of
13 dB.
The LPI antenna must have a transmit radiation pattern with very low
side lobes. The low side lobes in the transmit pattern reduce the possibility
of an intercept receiver detecting the radio frequency (RF) emissions from
the side lobe structures of the antenna pattern. The important general rules
for developing low side lobe antennas are [7]:
Symmetric amplitude distributions give lower side lobes.
F (u) should be an entire function of u.
A distribution with a pedestal produces a far-out side lobe envelope of
1/u.
A distribution going linearly to zero at the ends produces a far-out side
lobe envelope of 1/u2 .
A distribution that is nonzero at the ends (pedestal) is more ecient.
Zeros should be real (located on the unit circle).
Far-out zeros should be separated by unity (in u).
By applying a tapered (apodized) excitation from the center to the ends
of the antenna, the side lobe level can be lowered below 13 dB. A level of
20 dB is normally acceptable, but with LPI radar, ultra-low side lobes are
required (45 dB). Table 1.1 shows three excitation tapers (cosine, triangular,
and parabolic) for a rectangular array of length d, and the resulting antenna
performance [8]. A circular array has similar numbers. Note that as the side
lobe level goes down (SLR gets larger), the beamwidth gets larger and the
antenna gain decreases.
Another significant aperture excitation is the Taylor distribution devel-
oped by T. T. Taylor in 1960 [9, 10]. Taylor realized that to produce a linear
aperture distribution with a side lobe envelope approximating a 1/u fallo,
the uniform amplitude sin x/x pattern could be used as a starting point by
realizing that the height of each side lobe is controlled by the spacing between
Table 1.1: Aperture Taper Functions and Resulting Characteristics

3-dB Beamwidth Side Lobe Ratio Relative Full Null
Excitation (rad) (dB) Gain Position
Cosine
G(x) = cosN (x/2); |x| < 1
N =0 0.88/d 13.2 1.000 1.0/d

N =1 1.20/d 23.0 0.810 1.5/d
N =2 1.45/d 32.0 0.667 2.0/d
N =3 1.66/d 40.0 0.575 2.5/d
N =4 1.94/d 48.0 0.515 3.0/d
Triangular
G(x) = 1 |x|; |x| 1
1.28/d 26.4 0.75 2.0/d
Parabolic
G(x) = 1 (1 )x2 ; |x| < 1
= 1.0 0.88/d 13.2 1.00 1.00/d

= 0.8 0.92/d 15.8 0.99 1.06/d
= 0.5 0.97/d 17.1 0.97 1.14/d
=0 1.15/d 20.6 0.83 1.43/d
the aperture pattern factor zeros on each side of the side lobe. That is, since
the sinc pattern has a 1/u side lobe envelope it is only necessary to modify the
close-in zeros to reduce the close-in side lobes. The shifting is accomplished
by setting zeros equal to 0
u = n2 + B 2 (1.9)
where B is a positive real parameter. The resulting pattern with the zeros
shifted can be written as

sinh B 2 u2
F (u) = (1.10)
B 2 u2
for u B and
sin B 2 u2
F (u) = (1.11)
B 2 u2
for u B and is a modified sinc pattern where the one parameter B controls
all of the characteristics (side lobe level, beamwidth, directivity and so forth).
Often known as the one-parameter Taylor scheme, the SLR (in decibels) can
be expressed as
sinh B
SLR = 20 log + 13.2614 (1.12)
B
Table 1.2: Taylor Weighting Characteristics
Side Lobe Ratio (dB) B 3 (rad)

13.26 0 0.8858/d
15 0.3558 0.9230/d
20 0.7386 1.0238/d
25 1.0229 1.1160/d
30 1.2762 1.2004/d
35 1.5136 1.2782/d
40 1.7415 1.3504/d
45 1.9628 1.4182/d
50 2.1793 1.4822/d
The SLR for the Taylor weighting as a function of the B parameter, and the
3-dB beamwidth is shown in Table 1.2 as a function of the array length d and
the wavelength . Tables of circular aperture distributions and the design
process for the Taylor scheme are given in [11].
1.2.3 Antenna Scan Patterns for Search Processing

LPI radar systems are often identified by the type of scanning the emitter
uses. Scanning is the systematic movement of a radars antenna beam in a
particular pattern to search or track a target. The two methods of scanning
an antenna beam are mechanically and electronically. The antenna can be
mechanically scanned by using gimbals to move the entire antenna aperture in
any direction. Most often used are the two-dimensional arrays and parabolic
reflectors (where instead of moving the reflector, the reflector feed can be
nutated to provide the scan coverage needed). The antenna can also be elec-
tronically scanned by varying the phase between antenna elements (phased
array).
The simplest case of a search radar scan is the use of a stationary pencil
beam that is fixed in elevation and rotated mechanically at a scan rate of
r r s1 to obtain an a = 2 r coverage in azimuth and an e = e cov-
erage in elevation. If range information is obtained for each beam position
in space, this is an example of a one-dimensional (1D) scan pattern. In this
case the antenna searches or scans a solid angle field of view or scan volume
s = a e = 2e sr (steradian). With a total solid angle coverage of the
sensor s sr, the number of resolution elements is this value divided by the
instantaneous field of view of the antenna or
s
nr = (1.13)
a e
or 2/a . The elevation of the scan can also be changed after each complete
Figure 1.2: LPI scan patterns: (a) conventional transmit-receive raster,

(b) multibeam sector scan, and (c) omnitransmit multibeam receive.
rotation. This results in range, azimuth, and elevation information being

obtained from the field of view (e.g., a 2D scan pattern). To increase the
eciency of this 2D scan pattern, a set of n contiguous fixed pencil beams
can be stacked in elevation [12]. The azimuth scan can also be limited as
shown in Figure 1.2(a) (azimuth dimension only) and, at the end of the limit,
the elevation can be changed for the next azimuth scan (in the opposite
direction). This type of transmit-receive pattern is called a raster scan and is
used frequently since it provides good coverage both in azimuth and elevation.
An example of an LPI radar using this scan pattern is the Signaals SMART-L
(mechanical scanning implementation) [13, 14]. The SMART-L is discussed
further in Chapter 2.
The time Tf s required to cover the solid angle of coverage s is called
the frame time. For frame time Tf , the dwell time (sometimes known as
time-on-target) is
Tf
d = s (1.14)
nr
That is, for d s, energy is received from any point target at a range RT in
space that is illuminated by the transmitted radiation. The value of the dwell
time given by (1.14) is for one pencil beam scanning in azimuth and elevation.
If the radar system uses a stacked beam configuration to scan in azimuth and
elevation, the on target time will be increased by this factor
nTf nTf a e
d = = s (1.15)
nr s
where n is the number of contiguous beams in the stack (usually six to 16
[12]). Note that (1.15) does not include any scanning loss factor such as the
time necessary for the mechanically scanning antenna to move from the end-
of-frame position to the beginning-of-frame position. Accounting for this loss
Ls > 1, the dwell time is expressed as
nTf a e
d = s (1.16)
s Ls
Phased arrays provide the ability to form multiple beams at dierent frequen-
cies to selectively search dierent portions of the scan volume. In most cases,
the transmit scan pattern of the LPI radar is controlled precisely to limit the
illumination time to short and infrequent intervals (aperiodic scan cycle). An
example of an LPI radar with this capability is described in the next section.
By subdividing the scan volume into Nd sections, with every sector simul-
taneously searched by a dierent stack of beams using a dierent frequency,
the dwell time in each beam direction can also be increased by a factor of Nd
(see p. 530 in [15]) as
nTf a e
d = Nd s (1.17)
s Ls
Figure 1.2(b) shows the multibeam sector scan where the same scan volume
s is divided into Nd sectors. For this technique (electronic scanning), each
sector has its own transmit-receive beam. Matching the coherent integration
time to the dwell time, the power emitted in any one beam direction can be
reduced by the factor 1/Nd .
To electronically scan a phased array antenna a progressive phase shift is
typically used [5]. To maintain a low side lobe level a method based on a
pattern search algorithm (PSA) has been recently proposed [16]. Instead of
an amplitude taper (such as a Taylor excitation), low side lobe scanning can
be achieved using a phase taper. The PSA is a direct search algorithm. By
defining a fitness function, the pattern search finds the best group of phase
variations to scan the main beam to the desired position while also reducing
the peak side lobe level [17].
Figure 1.2(c) shows a nonscanning single-beam transmit, multibeam re-
ceive array where enough receive beams are formed to completely fill the scan
volume s . This technique requires increased signal processing throughput
and uses a single beamwidth nonscanning transmit antenna with many si-
multaneous receive beams. Since no scanning is involved, the dwell time is
equal to the frame time
d = Tf s (1.18)
One important flexibility with this pattern is that the transmitter does not
need to be colocated with the receive array. An example of an LPI radar that
uses this technique is the omnidirectional LPI radar (OLPI) [18, 19]. The
OLPI is discussed further in Chapter 5.
As discussed in Part II (Chapter 11 through Chapter 18) of the text, inter-
cept receivers use a variety of strategies to identify the LPI radar, including
angle of arrival, carrier frequency, scan rate, modulation period, bandwidth,
and polarization. Randomly altering any of these parameters can therefore
provide confusion to the intercept receiver. Scan methodologies can be used
to help confuse identification if intercept occurs. For example, a scan tech-
nique that attempts to confuse identification might include amplitude modu-
lation of a monopulse array at conical scan frequencies that are not considered
threatening. These types of scan methodologies require significant additional
processing requirements (and power) that limit the platforms that can carry
this type of capability.
1.2.4 Advanced Multifunction RF Concept

A radar that has the capability of forming multiple beams is the advanced
multifunction RF concept (AMRFC). The AMRFC is a United States Navy
program to investigate the capability to integrate radar (including an LPI
navigation radar in the high band 4.518.0 GHz), electronic warfare, and
communication functions into a common set of wideband, low power level
RF apertures, where the functionality is mostly defined by the software [20].
Consequently, the AMRFC reduces the number of topside RF system antenna
apertures while increasing the eective functionality (through software), as
well as increasing the capability for ship signature control/reduction. The
AMRFC divides the frequency band into an optimal number of segments,
based on cost and functionality, and then utilizes separate transmit and re-
ceive apertures. The separate transmit and receive apertures allow full uti-
lization of the entire time line for the transmitter as well as for the receiver.
The transmit array is composed of dynamically allocable subarrays that are
sectioned to form multiple simultaneous transmit beams. Since having more
than one signal present in a power amplifier is not currently feasible, each
transmit subarray is used by one function at a time. However, for the receive
array, more than one signal can be present simultaneously in a subarray.
The wideband arrays are electronically scanned and use solid state trans-
mit and receive apertures. The use of the contiguous subarray architecture
using phase shifters at the element level and true time delays at the subar-
Figure 1.3: High band multifunction transmit array [22].
ray level results in high side lobes, due to subarray dispersion and grating
lobe generation [21]. The AMRFC phased arrays achieve side lobe reduc-
tion by using an overlapping subarray architecture. Digital beamforming at
the subarray level makes it possible to generate multiple cluster beams and
achieve pattern control or interference cancelation simultaneously. The over-
lapping subarray allows the grating lobes to be pushed away from the main
lobe and shape the subarray patterns in such a way that all grating lobes are
suppressed in the subarrays low side lobe region. The overlap architecture
is further discussed in [21]. Figure 1.3 shows the high-band multifunction
transmit array that supports up to four simultaneous transmit beam con-
figurations. Transmit array quadrants may also be combined to form larger
apertures [22].
1.2.5 Transmitter Considerations

A conventional radar that uses a coherent pulse train has independent con-
trol of both range and Doppler resolution. This type of radar waveform
also exhibits a range window that can be inherently free of side lobes. The
main drawback of a coherent pulse train waveform is the high peak-to-average
power ratio put out by the transmitter. The average power is what deter-
mines the detection characteristics of the radar. For high average power, a
short pulse (high range resolution) transmitter must have a high peak power,
necessitating vacuum tubes and high voltages. The high peak power trans-
missions can also easily be detected by noncooperative intercept receivers.
The duty cycle dc for a pulsed emitter relates the average transmitted power
Figure 1.4: Comparison of a pulsed radar and a CW radar.
Pavg to the peak power Pt as

Pavg
dc = (1.19)
Pt
The duty cycle can also be calculated as
R
dc = (1.20)
TR
where TR is the pulse repetition interval (time between pulses) and R is
the emitters pulse width or duration (in seconds). Typical duty cycles are
dc = 0.001 (the average power 0.001 times the peak power) for navigation
radar.
In modulated CW signals, however, the average-to-peak power ratio is
one or 100% duty cycle. This allows a considerably lower transmit power to
maintain the same detection performance as the coherent pulse train radar.
Also, solid state transmitters can be used that are lighter in weight. A com-
parison of a coherent pulse train radar and the CW radar is shown in Figure
1.4. The CW radar has a low continuous power compared to the high peak
power of the pulse radar but, as will be demonstrated, both can give the same
detection performance. On the other hand, the final peak power for a pulsed
system may be only a few decibels (dB) higher than that of CW systems
having equivalent performance.
Consequently, most LPI emitters use continuous wave (CW) signals. A
CW (tone) signal is easily detected with a narrowband receiver and cannot
resolve targets in range. LPI radars use periodically modulated CW signals
resulting in large bandwidths and small resolution cells, and are ideally suited
for pulse compression.1
1 The pulse compression concept is being extended here to unpulsed CW waveforms
since the techniques are similar and the objectives are the same.
There are many pulse compression modulation techniques available that

provide a wideband LPI CW transmit waveform. Any change in the radars
signature can help confuse an intercept receiver and make intercept dicult.
The wide bandwidth makes the interception of the signal more dicult. For
the intercept receiver to demodulate the waveform, the particular modula-
tion technique used must be known (which is typically not the case). Pulse
compression (wideband) CW modulation techniques include:
Linear, nonlinear frequency modulation (Chapter 4);
Phase modulation (phase shift keying PSK) (Chapter 5);
Frequency hopping (frequency shift keying FSK), Costas arrays
(Chapter 6);
Combined phase modulation and frequency hopping (PSK/FSK)
(Chapter 6);
Noise modulation (Chapter 7).
With the above modulation techniques, the radiated energy is spread over
a wide frequency range in a manner that is initially unknown to a hostile
receiver. The phase and frequency modulation are not inherently wideband
or narrowband. The LPI radar designer chooses the necessary bandwidth
in order to get the range resolution properties needed. He then chooses the
modulation code necessary to get the ambiguity properties needed. This is
where the implementation issues must be addressed. The major goal for the
LPI radar designer is to get a 100% duty cycle and still retain the range
and velocity performance required. In single antenna systems where leakage
from transmitter to receiver can desensitize the target detection capability,
an interrupted CW waveform is often used.
1.2.6 Power Management

Another feature of the LPI transmitter is power management (one of the ben-
efits to using a solid-state radar/phased array combination). Of course, the
best LPI strategy is to not radiate at all, but the next best strategy is to man-
age the power that is radiated. Power management is the ability to control the
power level emitted by the antenna, and limit the power to the appropriate
range/radar cross section detection requirement. The emissions are also lim-
ited in time (short dwell time). With the use of wideband pulse compression
CW emissions, it is only necessary to transmit a few watts (instead of tens of
kilowatts of peak power required by low duty cycle pulsed radars with similar
detection performance). The LPI radar operates under low SNR conditions
and it is important to recall that the radars ability to detect targets depends
Figure 1.5: Regions of maximum atmospheric absorption in the millimeter

wave spectrum from measured data [23] ( c 1999 Artech House, reprinted with
permission).
not on the waveform characteristics, but on the transmitted energy returned

from the target.
Many intercept receivers depend on seeing an increase in intercepted power
as a closing missile approaches. With power managed seekers, the radar emits
only the power required for detection. As the range-to-target is reduced,
the intercepted power level decreases and varies directly as a function of
R2 . This LPI strategy can then force the intercept receiver into incorrectly
placing its priorities for electronic attack. That is, since the intercepted power
decreases, the receiver identifies the threat as nonapproaching; therefore no
attack should be necessary (unfortunately, a deadly decision). The concept
and usefulness of power management is quantified further in Chapter 9.
1.2.7 Carrier Frequency Considerations

Another LPI radar technique is choosing the emitter frequency strategically.
The use of a high operating frequency band that is within atmospheric ab-
sorption lines makes interception dicult, but also makes the target detection
by the radar even more dicult in most cases. The absorption spectrum is
shown in Figure 1.5 [23]. Peak absorption occurs at frequencies of 22, 60,
118, 183, and 320 GHz. The RF frequency can be chosen at these frequencies
to maximize the attenuation in order to mask the transmit signal and limit
reception by a hostile receiver (atmospheric attenuation shielding). Since the
physics of radar detection, however, depends only on the energy placed on
the target, LPI radars must still radiate sucient eective radiated power
(ERP) to accomplish detection. The loss for the radar due to atmospheric
absorption is over its total two-way path (out to the target and back), while
the interceptors loss is over the one-way path (from the radar to the inter-
cept receiver). Because of the high absorption of the emitters energy, this
technique is always limited to short range systems. In the case of an inter-
cept receiver on a radar target platform (such as a radar warning receiver),
the advantage lies with the interceptor, since there is only one-half the path
loss. Another approach to achieving a lower probability of interception is to
interleave the LPI radar mode with an infrared sensor (dual mode approach),
reducing the amount of time that the RF transmitter is radiating.
In summary, the important characteristics of LPI radar include wideband
CW emission, low antenna side lobes with infrequent scan modulation, or the
use of a broad nonscanning transmitting beam combined with a stationary set
of receive beams. Polarization modulation can also be used. The transmitter
uses a wideband modulation technique (for the range resolution desired) in
combination with power management and a strategic selection of frequency
to achieve the desired amount of atmospheric attenuation. That is, the wide-
band signals are diused in time, appearing in pseudorandom directions at
pseudorandom times.
1.3 Pulse CompressionThe Key to LPI Radar

The three general types of CW LPI radar architectures are the: (a) fre-
quency modulating radar that includes FMCW and frequency shift keying
(hopping), (b) the phase modulating radar that includes polyphase modula-
tion (polyphase shift keying) and polytime modulation, and (c) the radar that
is a combination of both (a) and (b). The FMCW radar architecture is now in
widespread use. A block diagram of an FMCW radar is shown in Figure 1.6.
The radar uses two antennas (one transmit and one receive). The transmitted
waveform for the FMCW radar is a linear (or nonlinear) frequency modulated
waveform, and can be generated by using a direct digital synthesizer. The
received waveform is amplified by a low noise amplifier (LNA) and correlated
(or mixed) with the transmit waveform in order to derive the target beat fre-
quencies (homodyne detection). After the analog demodulation is used to
generate the intermediate frequency (IF) beat signals, they are digitized with
an analog-to-digital converter. The digital signal with input signal-to-noise
ratio SNRRi is processed by one or more fast Fourier transform signal proces-
sors to derive the range (and possibly Doppler) profile. As shown in Figure
1.6, a certain amount of integration is also commonly used to increase the
Figure 1.6: Block diagram of an FMCW radar.
output signal-to-noise ratio SNRRo . Integration improves SNRRo since the

noise energy that accumulates in each range bin varies from one integration
period to the next, whereas the target return increases in direct proportion
to the integration time. Increasing the integration time can improve SNRRo
significantly. After integration, the target detection and tracking function is
performed.
The radar input SNRRi and output SNRRo are related by the processing
gain of the signal processor as
SNRRo
P GR = (1.21)
SNRRi
and depends on the time-bandwidth characteristics of the transmit signal
modulation as well as any noncoherent integration. The processing gain is
also referred to as the pulse compression ratio. A simple empirical formula
can also be used for the relationship between SNRRo and the probability of
detection Pd , and the probability of false alarm Pf a due to Albersheim is
given as [12]
SNRRo = A + 0.12AB + 1.7B (1.22)
where w W
0.62
A = ln (1.23)
Pf a
and w W
Pd
B = ln (1.24)
1 Pd
Here SNRRo is in linear units and not decibels.
If the radar uses an FMCW waveform, the processing gain (excluding any
noncoherent integration) is the sweep or modulation period, tm , multiplied
by the sweep (input) bandwidth, F . That is,
P GR = tm F (1.25)
The modulation period tm in an FMCW radar plays the same role as TR in

a pulsed radar and, in either case, both systems normally perform nonco-
herent integration over as many such intervals as occupy the dwell time of
the beam. When noncoherent integration
is performed for NI such intervals,
the processing gain is increased by NI . Also note that although Figure 1.6
shows an analog processing approach, the cross correlation (or homodyne de-
tection) could also be done digitally. Additional details on FMCW LPI radar
design are discussed in Chapters 4, 7, 8, and 9.
A block diagram of a phase coded radar is shown in Figure 1.7. The
phase coded radar can also use a direct digital synthesizer to generate the
transmitted waveform. The phase coded radar transmit waveform is gener-
ated using various phase modulations and/or frequency modulations. The
target return signal is amplified and downconverted using a local oscillator
(LO), and digitized with an ADC. The digitized samples are then processed
by a digital compressor, which cross correlates the transmitted code with the
received signal. For phase modulation of a CW waveform using Nc num-
ber of subcodes, the processing gain is the code period, T , multiplied by the
transmitted bandwidth, 1/tb , where tb is the subcode period. That is,
P GR = T (1/tb ) = (Nc tb )/tb = Nc (1.26)
In the phase coded radar, the return signal is compressed using digital tech-
niques, and noncoherent integration will also add to this processing gain.
Additional details on phase shift keying radar are given in Chapter 5 and an
example of the processing gain distribution in a phase coded LPI radar is
given in Section 5.12.
Note that in the FMCW radar example shown in Figure 1.6, the return
signal from the target is compressed using an analog processor. Although
Figure 1.6 shows an analog processing approach, the processing could equally
well be digital. That is, the distinction between the two is the modulation, not
the method of processing or the location of the ADC in the receiver path. For
both the frequency modulation and phase modulation LPI radar, the trans-
mitted CW signal is coded with a reference signal to spread the transmitted
energy in frequency, to avoid detection and identification by the noncoop-
erative intercept receiver. The reference signal can take the form of either
Figure 1.7: Block diagram of a phase coding radar.
a linear (nonlinear) frequency modulation, a frequency hopping sequence, a

phase code sequence, or a combination of these techniques. The FMCW
technique has been the most popular implementation, but with the current
capabilities of digital processors, phase coding CW waveforms are becoming
the standard, since many codes and variations can be employed. Note also
that frequency modulated CW signals can be approximated by phase coded
signals; a concept also discussed more thoroughly in Chapters 5, 8 and 10.
Although the CW signal is continuous, this does not imply that the por-
tion processed by the receiver in order to make a measurement or detect a
target is infinitely long. There are physical constraints, such as the illumi-
nation time and the size of the receivers correlation processor. Fast Fourier
transform processors (for frequency modulated waveforms) and finite dura-
tion coherent correlation processors (for phase modulation waveforms), as
well as combinations of both, are among the most-often-used techniques to
derive the target information.
The LPI receiver must correlate (or compress) the received signal from the
target using the stored reference signal, in order to perform target detection.
The correlation receiver is a matched receiver if the reference signal is
exactly the same duration as the finite duration return signal. Figure 1.8
shows a transmitted waveform (represented as a phase coded signal) of length
P T , where T is the code period, tb is the subcode period, and P 1. Also
Figure 1.8: CW transmitted waveform and receiver reference signal [24, 25].
shown is the reference signal of length N T used in the receiver to compress

the received signal. Increasing the number of receiver reference waveforms
N improves the target detection capabilities by increasing the resolution of
the receiver response. The ambiguity analysis in Chapter 3 investigates this
concept in more detail.
The LPI radar receiver can be modeled as a coherent correlation processor
of finite duration N T as shown in Figure 1.9 [27]. The return signal is received
by the correlation receiver containing a reference signal which is the conjugate
of N periods of the transmitted signal with N < P . The correlation receiver
performs a cross correlation between the received signal and a reference signal,
whose envelope is the complex conjugate of N periods of the transmitted
signal envelope. To do this, the return signal (a binary phase coded signal in
this example) is first processed by a filter matched to a rectangular subcode
of length tb , followed by a detector that sends forward a one or a zero. The
detected output signal is then piped through a tapped delay line where each
delay D is tb s. The signal in the tapped delay line is first multiplied by the
reference signal. The output of each multiplication is then summed separately
for each of the N code periods. The output of the sum block can then be
weighted as C1 through CN . With uniform weights, the first stage represents
the response of the receiver for a zero-Doppler shift signal ( = 0), and
is identical to the ideal autocorrelation function [26]. The response of the
receiver to a Doppler shifted signal is obtained from the second stage
by first multiplying the output (before addition) from the first stage with q 0
through q MN1 where q = ej2tb .
In phase coded CW radar systems, return signals with Doppler do not cor-
Figure 1.9: Doppler matrix correlation receiver matched to N periods of a

phase coded signal of length M = 5 including weighting Ci for Doppler side
c
lobe reduction [26] (s1992 IEEE).
Figure 1.10: LPI radar and intercept receiver configuration.
relate perfectly because the Doppler shift changes the phase of the code across
its period. This causes imperfect compression. Since the received signal is
usually delayed and Doppler shifted, there is a special interest in the response
of a matched receiver, such as in Figure 1.9, to its own signal as a function of
the two parameters delay and Doppler. To reduce the side lobes, weighting
may be factored into the reference signal. If the reference signal is weighted
in order to reduce side lobes, the receiver is called a mismatched receiver. In
Chapter 3, the ambiguity response of these LPI receivers is discussed, as well
as the weighting functions.
1.4 Radar Detection Range

In this section, the maximum detection range for a CW radar is examined.
The CW radar has a low continuous power with a 100% duty cycle. The
LPI radar and intercept receiver configuration is shown in Figure 1.10. To
determine the detection range of a CW radar, we start with the power density
at a range R m from an isotropic antenna given as [27]
PCW 2
PD = W/m (1.27)
4R2
where PCW is the average power of the CW transmitter in watts. With

a directive antenna having a transmit gain Gt along the boresight, the
directed power density at a range R from the radar is
PCW Gt L1 2
P DD = W/m (1.28)
4R2
The term L1 (< 1) is the one-way atmospheric transmission factor
L1 = eRk (1.29)
where Rk is the range or path length in kilometers and is the one-way extinc-
tion coecient or power attenuation coecient in nepers per km (Np/km).
The one-way attenuation coecient as a function of frequency is shown in
Figure 1.5 in more useful engineering units (dB/km). To convert dB/km into
Np/km, multiply the attenuation coecient in Figure 1.5 by 0.23.
The reradiated power density reflected o a target with radar cross section
T (m2 ) at range RT and appearing back at the radar is
w W
PCW Gt L2 T
P DDR = W/m2 (1.30)
4RT2 4RT2
where RT is the range between the LPI radar and the target. The term
L2 (< 1) is the two-way atmospheric transmission factor
L2 = e2Rk (1.31)
The LPI radar captures the reflected energy with its receive antenna. The
received signal power at the radar receiver from the target is
w W
PCW Gt L2 T
PRT = Ae (1.32)
4RT2 LRT LRR 4RT2
where Ae is the eective area of the radar receive antenna and related to the
receive antenna gain Gr as
Gr 2
Ae = (1.33)
4
and LRT is the loss between the radars transmitter and antenna, and LRR
is the loss between the radars antenna and receiver. Substituting (1.33) into
(1.32) gives the reflected power at the radar receiver as
PCW Gt Gr 2 L2 T
PRT = (1.34)
(4)3 RT4 LRT LRR
It is often necessary to know the minimum input signal power at which a

receiver can detect and process an incoming target signal. This is called the
receivers sensitivity or R . Substituting the sensitivity for PRT in (1.34), the

maximum range at which the LPI radar can detect a target is
} ]1/4
PCW Gt Gr 2 T L2
RR max = (1.35)
(4)3 (R )LRT LRR
The sensitivity R is the product of the minimum signal-to-noise ratio required

at the input (SNRRi ) times the noise power in the input bandwidth of the
receiver. The sensitivity of the radar receiver can be expressed as
R = kT0 FR BRi (SNRRi ) (1.36)
where k = 1.38(1023 ) joule/K (Boltzmanns constant), T0 is the standard

noise temperature (T0 = 290K), FR is the receiver noise factor, and BRi is
the radar receivers input bandwidth in Hz, and is usually matched to the
particular waveform being transmitted. The maximum detection range can
be expressed as
} ]1/4
PCW Gt Gr 2 T L2
RR max = (1.37)
(4)3 kT0 FR BRi (SNRRi )LRT LRR
Also recall that the processing gain of the radar is

SNRRo
P GR = (1.38)
SNRRi
and depends on the particular waveform characteristics and integration tech-
niques being used by the LPI radar. Note also that the sensitivity R can
be expressed as a function of the output signal-to-noise SNRRo required for
detection and the output bandwidth BRo as
R = kT0 FR BRo (SNRRo ) (1.39)
For example, consider an LPI radar with PCW = 1W, Gt = Gr = 30

dB, fc = 9.375 GHz, FR = 5 dB, and BRi = 1 MHz. If all losses are 0
dB and L2 = 1, Figure 1.11 shows the LPI radar maximum detection range
as a function of the required input signal-to-noise ratio SNRRi for T = 1,
10, and 100 m2 . With an SNRRi = 10 dB, a 1 m2 -target can be detected
at a range of 1,450m while a T = 100 m2 can be detected at a range of
4,500m. This information, however, does not reveal the benefit of the LPI
radar. To quantify this, the LPI radars maximum target detection range
must be compared to the intercept receivers maximum interception range.
Figure 1.11: LPI radar maximum detection range for T =1, 10, 100 m2 .
1.5 Interception Range

From the configuration shown in Figure 1.10, the signal power available at
the intercept receiver from the LPI radar is
w W
PCW Gt L1 GI 2
PIR = (1.40)
4RI2 LRT LIR 4
where RI is the range from the LPI radar to the intercept receiver, and Gt is
the gain of the LPI radars transmit antenna in the direction of the intercept
receiver. Also, GI is the gain of the intercept receivers antenna, and LIR
represents the losses from the antenna to the receiver. If the intercept receiver
detects the radar main lobe, Gt = Gt . If the intercept receiver detects the
radar emission from the side lobes, Gt represents the gain of the antenna side
lobe in the intercept receiver direction. Replacing the signal power available,
PIR , by the intercept receivers sensitivity, I , the maximum interception
range of the receiver can be defined as

PCW Gt GI L1 2
RI max = (1.41)
(4)2 LRT LIR (I )
where the sensitivity in the intercept receiver is similarly defined as
I = kT0 FI BI (SNRIi ) (1.42)
Figure 1.12: Block diagram of an intercept receiver model showing both the
predetection stage and the postdetection stage.
where FI is the intercept receiver noise factor, BI is the bandwidth of the

intercept receiver, and SNRIi is the SNR at the intercept receiver signal
processor input. The maximum interception range can then be expressed as

PCW Gt GI L1 2
RI max = 2
(1.43)
(4) LRT LIR kT0 FI BI (SNRIi )
Also recall that the intercept receiver processing gain P GI is defined as

SNRIo
P GI = (1.44)
SNRIi
Contrary to communication or radar system receiver design where the band-
width is matched to the known transmitted signal, the intercept receiver does
not know the exact nature of the threat signals. Figure 1.12 shows a block
diagram of an intercept receiver model showing the predetection stage and
the postdetection stage. The three major components include the RF (prede-
tection) amplifier with bandwidth BIR , the detector (e.g., square law), and
the postdetection video amplifier with bandwidth BIV .
In the intercept receiver design it is most often necessary to match the
front-end RF bandwidth BIR to the largest coherent radar bandwidth ex-
pected, and to match the video bandwidth BIV to the inverse of the smallest
radar coherent integration time expected tI . Exact analysis of intercept re-
ceiver bandwidths and sensitivities is complicated. However, since the I in
(1.42) is dependent on the intercept receivers overall bandwidth, it is desir-
able to have an approximate expression for BI that includes the eects of
both the predetection and postdetection bandwidths. The approximate ex-
pressions have been derived by Klipper [28]. The bandwidth of the intercept
receiver for BIR >> BIV (which is typically the case) can be expressed as
0
BI = 2BIR BIV (1.45)
for a square law detector and

10
BI = 2BIR BIV (1.46)
2
for a linear detector. The time-bandwidth product or processing gain of the
intercept receiver often takes the form of [3]
P GI = (tI BIR ) (1.47)
and depends on the eciency . The noncoherent integration eciency is

on the order of 0.5 0.8 [3]. For example, if an FMCW radar has a
modulation bandwidth of F = 55 MHz, and a coherent integration time
of tm = 1 ms, the processing gain of the radar is P GR = tm F = 55,000,
whereas the processing gain of the intercept receiver is only P GI = tm F =
234 ( = 0.5). This is the real origin of the LPI radar.
Figure 1.13 shows the maximum interception range of the intercept re-
ceiver as a function of the required input SNRIi . This is the maximum range
at which the passive intercept receiver can intercept an LPI radar operating
at fc = 9.375 GHz with a transmitting antenna gain Gt = 1,000 and trans-
mitter power PCW = 1W. The intercept receiver has an isotropic antenna
with gain GI = 1, a noise figure FI = 5 dB, and LRT , LIR = 1. Both a
square law and linear detector configuration are considered with BIV = 1
kHz (radar coherent integration time tm = 1 ms). The performance for pre-
detection bandwidths of BIR = 60 MHz and 120 MHz is also compared. Note
that the smaller the required predetection RF bandwidth is, the larger the
maximum interception range. The use of a square law detector over a linear
detector also gives a larger interception range.
1.6 Comparing Radar Range and Interception

Range
The radar sensitivity R (1.36) and intercept receiver sensitivity I (1.42) can
be used to quantify the benefit of the LPI radar. The ratio of the intercept
receiver sensitivity to the radar sensitivity is
w W
I kT0 BI FI SNRIi
= = (1.48)
R kT0 BRi FR SNRRi
In terms of the processing gains and output signal-to-noise ratios, the ratio
of sensitivities can be expressed as
w Ww W
FI BI SNRIo P GR
= (1.49)
FR BRo SNRRo P GI
Figure 1.13: Intercept receiver with square law and linear detector BIV = 1
kHz, BIR = 60, and 120 MHz, showing maximum interception range for LPI
radar with PCW = 1W, Gt = 30 dB, and fc = 9.375 GHz.
Also, the sensitivity ratio can be written as a function of the radar and
intercept receiver antenna parameters as
w W w 2
W2
4 Gt GI L1 RR max
= (1.50)
T Gt Gr L2 RI max
and conveniently expresses the sensitivity ratio as a function of the maxi-

mum detection ranges. Note that this equation is independent of the radar
wavelength (directly) and the radars average transmit power.
To directly compare the radar detection range and the intercept receiver
detection range, we can solve (1.50) for the ratio of the two maximum detec-
tion ranges as
} w W ]1/2
RI max 1 4 Gt GI L1
= RR max (1.51)
RR max T Gt Gr L2
Here the ratio of the radar receiver sensitivity to the intercept receiver sen-
sitivity () is in the denominator. If RI max /RR max < 1, then the radar can
be considered a quiet radar. If the ratio RI max /RR max = 1, then the radar
cannot be intercepted beyond the range at which it can detect targets. This
is an important consequence. With RI max /RR max = 1, from (1.51)

} p Q ]1/2
T L2 Gt Gr
RR max = (1.52)
4 L1 Gt GI
Note that this is the maximum detection range of the LPI radar without
being intercepted by the intercept receiver. This is also the noncooperative
receivers maximum intercept range. An example is developed next in order
to examine this result.
To illustrate (1.52), the detection of an LPI radar is considered with both
main lobe detection (Gt = Gt = 1,000) and side lobe detection with side
lobes Gt = 0.1 (40 dB below main beam). The target RCS T = 1m2 and
we consider that L2 = L1 = 1. The intercept receiver antenna gain GI = 1.
Figure 1.14 shows the sensitivity ratio as a function of the maximum detection
range for both main lobe intercepts and side lobe intercepts. The figure shows
the large dierence in the sensitivity ratio due to the dierence of detecting
the radar in the side lobes versus the main lobes. The figure shows that a
sensitivity ratio of 60 dB is required for a maximum radar detection range,
noncooperative intercept range of 104 m (intercept receiver intercepting the
main lobe). If the intercept receiver is required to intercept the radar in the
side lobes at this range, the intercept receiver must decrease the sensitivity
ratio from 60 dB to 20 dB. From (1.49), one of the ways this can happen
is when the intercept receiver increases its processing gain P GI which is
typically dicult to do without sophisticated signal processing techniques
(discussed in Part II).
1.7 The Pilot LPI Radar

During 1988, the Philips Research Laboratory developed a quiet radar
known as Pilot, which was marketed by the then Philips subsidiaries PEAB
in Sweden and Signaal in the Netherlands. With the sell-o of Philips de-
fense assets, PEAB was taken over by Bofors (subsequently CelsiusTech and
now SaabTech), and maintained the name of Pilot for this radar. For its
part, Signaal was taken over by the then Thomson-CSF (now Thales), and
modified and improved the FMCW Pilot concept and changed the name of
the radar to Scout.
The Pilot is a well-published example of an FMCW tactical navigation
LPI radar [2933]. It can easily be added on to an existing navigation radar,
retaining the original X-band antenna, transceiver, and display system. In
a tactical situation, the Pilot can be switched out and the pulsed radar can
be switched in when higher signal-to-noise ratios are required. It also has
standard video output to simplify integration with standard pulsed navigation
radar.
Figure 1.14: LPI radar maximum detection range as a function of the

sensitivity ratio .
The Pilot uses an FMCW 1-kHz sweep repetition frequency with a low
noise figure (FR = 5 dB) and very low output power to ensure that it is un-
detectable by hostile intercept receivers. Other features include a 1,024-point
FFT (512 range cells) high range resolution (2.7m to 86m), high reliability,
small lightweight designs, and ease of installation. The technical parame-
ters are given in Table 1.3. Note that 1 nautical mile (nmi) = 1.852 km.
Figure 1.15 shows the equipment that makes up the Pilot Mk3 version that
was developed by Saab Bofors Dynamics AB (formerly CelsiusTech Electron-
ics). The MK3 has an improved LPI performance by combining an FMCW
waveform with frequency agility. In this section we use the formulations in
previous sections to quantify the Pilot performance.
The most important LPI characteristics of the Pilot are that it uses only
one low side lobe antenna, transmits a maximum CW power of only 1W, and
uses an FMCW waveform with a variable modulation bandwidth F to vary
the range resolution. We will return to a detailed discussion of this type of
LPI radar modulation in Chapter 4. Below, several examples are shown to
illustrate the performance of the radar and compare it with a conventional
low pulse repetition frequency (LPRF) navigation radar.
Table 1.3: Technical Characteristics of the Pilot Mk3
Antenna Type: Single or dual slotted-waveguide

Gain: 30 dB
Side lobes: < 25 dB
< 30 dB
Beamwidth (3 dB)
horizontal: 1.2 deg
vertical: 20 deg
Rotational speed: 24/48 RPM
Polarization: horizontal
Transmitter Output power: 1.0, 0.1, 0.01, or 0.001W (CW)

Frequency: 9.375 GHz (X-band)
Range selection: 24, 12, 6, 3, 1.5, 0.75 nmi
Frequency sweep: 1.7, 3.4, 6.8, 13.75, 27.5, 55 MHz
Sweep repetition
frequency: 1 kHz
Receiver IF bandwidth: 512 kHz

Noise figure: 5 dB
Processor Unit No. of range cells: 512 (1,024-point FFT)

Range resolution: < 75m at 6 nmi scale
Range accuracy: < 25m at 6 nmi scale
Azimuth accuracy: 2 degrees
Azimuth resolution: 1.4 degrees
Display System Type: Color

Minimum eective
PPI diameter: 250 mm
Resolution: 768 1,024V
Tracking capacity: 40
Range ring accuracy: 1.5% of selected scale or
50m, whichever is greater
Figure 1.15: Equipment that makes up the Pilot Mk3.
Example 1: Sensitivity
It is known that the Pilot radar using only PCW = 1W has a maximum
detection range of RR max = 28 km for a T = 100m2 target. Using the
system parameter values given in Table 1.3, determine the sensitivity R of
the Pilot receiver. Using (1.35),
PCW Gt Gr 2 T L2
R = 4 (1.53)
(4)3 (RR max )LRT LRR
with Gt = Gr = 1,000, = 0.032m, and substitution of the Pilot parameters

(assuming that L2 = LRT = LRR = 1)
1 (1,000)2 (0.032)2 100

R = = 8.4 1017 W (1.54)
(4)3 (28,000)4
or 130 dBm.
Example 2: Required Input SNR
Figure 1.6 shows SNRRi is located at the output of the ADC (input to the
signal processor). Since the Pilot radar uses down conversion processing to
translate the received signal frequency to IF, BRi = 512 kHz which corre-
sponds to the Pilots IF bandwidth in Table 1.3. To determine the input
signal-to-noise ratio SNRRi we know from (1.36)
Using FR = 100.5 , kT0 = 4 1021 , and BRi = 512 kHz,

R
SNRRi = = 0.013 (1.56)
kT0 FR BRi
or 19 dB.
Example 3: Processing Gain, Output SNR
For the maximum modulation bandwidth F = 55 MHz, we can calculate

the processing gain P GR and the output SNRRo . The processing gain for a
single sweep can be calculated from (1.25) as
P GR = tm F = 55,000 (1.57)
The corresponding output SNR is then
SNRRo = P GR (SNRRi ) = 715 (1.58)
or 28 dB. The addition of noncoherent integration of more than one modula-

tion period within the signal processor can increase the processing gain and
the SNRRo .
Example 4: Comparison with Conventional Pulsed Radar
If the emitter is a conventional low pulse repetition frequency (PRF) naviga-

tion radar with a peak power of Pt = 10 kW, pulse width of = 1 s, and
FR = 5 dB, neglecting losses, (a) determine the maximum detection range for
a T = 100m2 target if the minimum required receiver input signal-to-noise
ratio SNRRi = 1.7 dB and (b) for the intercept system above (I = 80
dBmi), determine the maximum intercept range (main lobe intercepts). For
(a), we can use (1.37) with PCW replaced by the peak power Pt = 10 kW and
Table 1.4: Pilot Detection and Intercept Range Calculations

Radar Detection Intercept Range
Range (km) (km)
Radar Intercept Intercept Intercept
Output 100 m2 1 m2 I I I
Power target target 40 dBmi 60 dBmi 80 dBmi
Pilot Mk2
1W 28 8.8 0.25 2.5 2.5
0.1 W 16 5.0 0 0.8 8.0
10 mW 9.0 2.8 0 0.25 2.5
1 mW 5.0 1.5 0 0 0.8
LPRF Radar
10 kW 49.6 15.7 25 254 2.546
BRi = 1(106 ) = 1/ .2
} ]1/4
Pt Gt Gr 2 T L2
Rconv = (1.59)
(4)3 kT0 FR BRi (SNRRi )
or } ]1/4
104 (1, 000)2 (0.032)2 100
Rconv = (1.60)
(4)3 4(1021 ) 100.5 1(106 ) 0.67
or Rconv = 49.6 km. For (b), the maximum intercept range can be determined
from (1.41) as
} ]1/2
Pt Gt GI L1 2
RI max = (1.61)
(4)2 I LRT LIR
or } ]1/2
10, 000 1(103 ) (1) (0.032)2
RI max = (1.62)
(4)2 1011.0
or RI max = 2,546 km. A summary of the above results and other additional
calculations are shown in Table 1.4. Note that we use dBmi to represent dB
in mW with reference to a system containing an isotropic antenna GI = 1.
1.8 Concluding Remarks

LPI modulation techniques include frequency modulation such as FMCW and
frequency shift keying. Also used are phase modulations such as the polyphase
2 A pulsed radar receiver usually has an input bandwidth that is matched to the transmit-
ted pulse width at either the null-to-null bandwidth (BRi = 2/ ) or the 3-dB bandwidth
(BRi = 1/ ).
codes Frank, P1, P2, P3, P4, and polytime codes T1, T2, T3, and T4. There
are several trade-os in the design of LPI emitters. The LPI modulations are
not inherently wideband (or narrowband). The radar designer chooses the
emitter bandwidth to achieve the range resolution properties needed. He also
chooses the particular code to get the ambiguity (delay Doppler frequency)
code properties needed. Implementation issues must also be addressed (such
as digital versus analog). The major question is how to get a 100% duty factor
and still get the desired range and velocity performance needed to perform
the mission.
A larger processing gain can be obtained by wideband coding of the trans-
mitted waveform with a modulation that is known only to itself. What is
important is if the coding degrades the sensitivity of the intercept receiver
relative to the radar receiver. The coding may or may not have an eect on
the sensitivity ratio . For example, if the intercept receiver is a simple crys-
tal video receiver, then the wideband coding has no eect on the intercept
receivers sensitivity. That is, the value of coding in LPI has to do with the
eect imposed on the interceptornot on the radar if it uses a matched filter.
The intercept receiver bandwidth BI is typically larger than the radars
coherent bandwidth in order to maximize the detection of the unknown signals
and perform well against large time-bandwidth signals. Also, the intercept
receivers noncoherent integration time should match the radars coherent
integration time. The design of the modern intercept receiver, however, is a
complicated issue due to the combined capability of an electronic support (ES)
receiver, radar warning receiver (RWR), and electronic intelligence (ELINT)
receiver in a single system, and many architectures are possible. These issues
are addressed in further detail in Part II.
References
[1] Wiley, R. G., Electronic Intelligence: The Interception of Radar Signals,
Artech House, Dedham, MA, 1985.
[2] Schleher, D. C., Low probability of intercept radar, Record of the IEEE
International Radar Conference, pp. 346349, 1985.
[3] Schrick, G., and Wiley, R. G., Interception of LPI radar signals, Record of
the IEEE International Radar Conference, Arlington, VA, pp. 108111, May
710, 1990.
[4] Rue, L. I., and Stott, G. F., LPI considerations for surveillance radars,
Proc. of the International Conference on Radar, Brighton, U.K., pp. 200
202, 1992.
[5] Balanis, C. A., Antenna Theory Analysis and Design, Harper and Row, Pub-
lishers, New York, 1982.
[6] http://www.f22fighter.com/radar.htm and
http://www.globalsecurity.org/military/systems/aircraft/f-22-avionics.htm
(APG-77).
[7] Rudge, A. W., Milne, K., Olver, A. D., and Knight, P., The Handbook of
Antenna Design, Vol. 2, IET, 1983.
[8] Forrest, J. R., Antenna design tradeos examined, Microwave Systems
News, Vol. 13, No. 12, pp. 237-243, Nov. 1983.
[9] Taylor, T. T., Design of circular apertures for narrow beamwidth and low
sidelobes, IRE Trans. on Antennas and Propagation, AP-8, pp. 1722,
1960.
[10] Hansen, R. C., Tables of Taylor distributions for circular aperture antennas,
IRE Trans. on Antennas and Progagation, pp. 2326, Jan. 1960.
[11] Hansen, R. C., A one parameter circular aperture distribution with narrow
beamwidth and low sidelobes, IEEE Trans. on Antennas and Propagation,
pp. 477480, July, 1976.
[12] Skolnik, M. I., Introduction to Radar Systems, 3rd Edition, McGraw-Hill,
Boston, MA, 2001.
[13] http://www.naval-technology.com/contractors/weapon control/thales5/
(SMART-L).
[14] Pietrasinski, J. F., Brenner, T. W., and Lesnik, C. J., Selected tendencies
of modern radars and radar systems development, 12th International Con-
ference on Microwaves and Radar, MIKON 98, Krakow, Poland, Vol. 1, pp.
133137, May 20-22, 1998.
[15] Stimson, G. W., Introduction to Airborne Radar, 2nd Edition, Scitech Pub-
lishing Inc., Mendham, NJ, 1998.
[16] Ebadi, S., Forouraghi, K., and Sattarzadef, S. A., Optimum low sidelobe
level phased array antenna design using pattern search algorithms, IEEE
International Symposium on Antennas and Propagation, pp. 770773, Vol.
1B, Washington DC, 2005.
[17] Ebadi, S., Forouraghi, K., Pattern scanning in low sidelobe phased array
antennas using pattern search algorithms, Proceedings of the 4th European
Radar Conference, pp. 347349, 2007.
[18] Wirth, W. D., Long term coherent integration for a floodlight radar, Record
of the IEEE International Radar Conference, pp. 698703, 1995.
[19] Wirth, W. D., Radar Techniques Using Array Antennas, IEE Radar, Sonar,
Navigation, and Avionics Series 10, 2001.
[20] Hughes, P. K., and Choe, J. Y., Overview of advanced multifunction RF
system (AMRFS), Proc. of the IEEE International Conference on Phased
Array Systems and Technology, pp. 2124, 2000.
[21] Ching-Tai Lin, and Ly, Hung, Sidelobe reduction through subarray overlap-
ping for wideband arrays, Proc. of IEEE Radar Conference, pp. 228233
2001.
[22] Tavik, G. C., Hilterbrick, C. L., Evins, J. B., Alter, J. J., Crnkovich, J. G.,
de Graaf, J. W., Habicht, W., Hrin, G. P., Lessin, S. A., Wu, D. C., and
Hagewood, S. M., The advanced multifunction RF concept, IEEE Trans.
on Microwave Theory and Techniques, Vol. 53, No. 3, pp. 10091020, March
2005.
[23] Klein, L. A., Millimeter-Wave and Infrared Multisensor Design and Signal
Processing, Artech House, Inc., Norwood, MA, 1997.
[24] Levanon, N., and Freedman, A. Periodic ambiguity function of CW signals
with perfect periodic autocorrelation, IEEE Trans. on Aerospace and Elec-
tronic Systems, Vol. 28, No. 2, pp. 387395, April 1992.
[25] Levanon, N., and Getz, B., Weight eects on the periodic ambiguity func-
tion, IEEE Trans. on Aerospace and Electronic Systems, Vol. 31, No. 1,
pp. 182193, July 1994.
[26] Levanon, N., CW alternatives to the coherent pulse trainsignals and proces-
sors, IEEE Trans. on Aerospace and Electronic Systems, Vol. 28, No. 2, pp.
387395, April 1992.
[27] Nathanson, F.E., Radar Design Principles, 2nd Ed., McGraw-Hill, New York,
1991.
[28] Klipper, H. Sensitivity of crystal video receivers with RF preamplification,
Microwave Journal pp. 8592, Aug. 1965.
[29] Beasley, P. D. L., and Stove, A. G., Pilot-an example of advanced FMCW
techniques, IEE Colloquium on High Time-Bandwidth Product Waveforms
in Radar and Sonar, pp. 10/110/5, May 1, 1991.
[30] Fuller, K. L., To see and not be seen, IEE Proc. F Radar, Sonar and,
Navigation Signal Processing, Vol. 137, Issue: 1, pp. 110, Feb. 1990.
[31] Pengelley, R. Philips Pilot, covert naval radar, International Defense Re-
view, pp. 11771178, Sept. 1988.
[32] Lok, J. J, Navigation radars, sensors working overtime, Janes Defence
Weekly, pp. 3940, Oct. 1992.
[33] Scott, R., Covert operations: navies seek discreet radars for surveillance,
Janes International Defence Review, 9 Aug. 2006.
Problems
1. (a) Estimate the beamwidth in azimuth and elevation of a rectangular
array 10 cm by 10 cm if the wavelength is 3 cm. (b) Estimate the gain
of the antenna if the eciency is 90%.
2. An LPI radar has an active phased array antenna with a = e = 3 de-
grees and a total solid angle of coverage, s = 2.4 sr. (a) If the antenna
takes only 0.5s to cover the solid angle of coverage and the transmitted
CW signal is a low power phase-coded signal with 11 subcodes and a
code period of T = 11 s (repeats every 11 s), how many code periods
would be integrated in a correlation receiver? (b) If noncoherent inte-
gration is performed over the entire dwell time, what is the processing
gain of the radar?
3. (a) Write a MATLAB program to determine the detection range for

both a Pilot LPI radar and a conventional 10kW pulsed emitter (as
a function of the transmitter power, antenna parameters, wavelength,
and target radar cross section), using the values for the Pilot radar.
(b) Compare your output results with those in Table 1.4. (c) To verify
the ES intercept range, include in your MATLAB program, a calcula-
tion of the intercept range for a I = 40, 60, 80 dBmi and a high
performance intercept receiver 110 dBmi.
4. A CW LPI radar has an average transmit power of PCW = 0.1W,
LRT = LRR = 1, and an antenna with Gt = Gr = 30 dB. The radar
illuminates a T = 50m2 target at a range of 2 km. (a) Calculate
the expected reradiated power back at the radar receiver if fc = 94
GHz. (b) Calculate the reradiated power back at the radar receiver if
fc = 9.3 GHz. (Assume the gain of the antenna remains the same).
5. An airborne platform carrying an LPI CW emitter is moving toward a
target (T = 100m2 ), at a speed of V = 150 m/s. The emitter is turned
on at a distance of R = 20 km from the target. Every 100 ms the
emitter adjusts its transmit power level to keep the required SNRRo
a constant, and equal to the minimum acceptable value. The other
characteristics of the emitter are Gt = Gr = 32 dB, fc = 10 GHz, and
R = 8 1017 W. The target also carries a noncooperative intercept
receiver with the following characteristics: I = 80 dBmi (GI = 0 dB).
(a) Plot the emitter transmitted power in dBW versus time, (b) Plot
the radars maximum detection range and the interceptors maximum
intercept range as a function of time for the entire engagement. (c)
What can you conclude about the radars quietness?
6. Using the Taylor GUI software in the Chapter 1 folder, examine the
antenna patterns for side lobe level of 10 to 50 dB (in increments of
5 dB). For each pattern, estimate the 3-dB beamwidth and plot these
values as a function of the side lobe level.
Chapter 2
LPI Technology and

Applications
In this chapter, we examine the applications of LPI radar technology. Al-
timeters are discussed first. An altimeter is an instrument that measures the
vertical distance (or altitude) of an object (such as a missile) with respect to
a reference level. The next application discussed is LPI landing systems. A
fully automatic LPI landing system can compensate for wind and platform
roll, and can perform ship-based landings under day/night, all-weather con-
ditions. Surveillance and fire control radar systems are also presented, and
depend on LPI technology to remain functional on the battlefield. Finally,
antiship capable missiles and torpedo seekers that use LPI technology are
reviewed.
2.1 Altimeters
2.1.1 Introduction
In 1928, German inventor Paul Kollsman changed the world of aviation with
the invention of the worlds first accurate barometric altimeter, also called
the Kollsman Window. Barometric altimeters are operated by air pressure
but have two limitations: (a) If the atmospheric pressure changes while the
platform is in flight the altimeter reading will change, and (b) the barometric
altimeter indicates height above sea level, or some other preset level, and does
not reveal the actual platform altitude above the surface.
In 1924, Lloyd Espenschied invented the first radio altimeter. The ra-
dio altimeter is a device, most often used in aircraft and cruise missiles,
that makes use of the reflection of radio waves from the land or water to
41
determine the height of the platform above the surface. In 1938, the fre-
quency modulation (FM) radio altimeter was first demonstrated in New York
by Bell Labs. In the first public display of the device, radio signals were
bounced o the ground, showing pilots the altitude of an aircraft.
Another choice for an altimeter is the use of a pulse-modulated radar. Al-
timeters that work on this principle give satisfactory results if the platform is
at a high altitude. At low altitudes, however, they have significant problems.
This is because pulse-modulated radar have a blind zone area surrounding
their installation where no targets can be detected. The blind zone area de-
pends upon the pulse width. For example, with a pulse width of 0.2 s, no
target within 100 feet of the radar can be detected. Consequently, altimeters
of this type are not useful for aerial vehicles such as cruise missiles flying near
the surface.
For vehicles that fly near the surface, it is necessary to detect and to
measure the distance from the surface to the radar, down to almost zero
feet. Frequency modulation continuous wave radar is the simplest of radar
ranging techniques, and the most resistant to false-lock to undesired targets
such as the missile structure. For example, in a typical FMCW altimeter, the
transmitters carrier frequency changes linearly over a 120-MHz modulation
bandwidth that ranges from 4.24 to 4.36 GHz. The transmitter works contin-
uously to produce the CW output, and changes frequency at a constant rate
in either a sawtooth pattern or a triangular pattern. A fixed, broad-beam
antenna system is used to illuminate a large area of the underlying terrain.
The broad beam allows for correct operation over the normal range of missile
pitch and roll. The FMCW ranging process occurs by mixing a sample of
the linearly varying frequency with the signal reflected from the surface. The
dierence produced after mixing is a low-frequency beat signal proportional
to the range of the surface being measured. A simple limiter then selects
the strongest signal from the surface directly below the vehicle. With proper
antenna installation, the FMCW processor can accurately select the surface
directly below the missile and ignore any atmospheric variations.
2.1.2 Fielded LPI Altimeters

The NavCom Defense Electronics Inc. Combined Altitude Radar Altimeter
(CARA) AN/APN-232 uses a solid-state, FMCW emitter centered on 4.2 to
4.36 GHz with 100-MHz modulation bandwidth, and features a wideband LPI
output and electronic protection (EP) features to prevent inoperability due
to jamming or electronic attack (EA). The device assemblies are shown in
Figure 2.1. LPI operation is achieved by using automatic power management
that depends on the aircraft attitude and altitude, and the terrain type. That
is, the transmitter output power is adjusted automatically so that the least
amount required for signal acquisition and tracking is transmitted [1]. The
AN/APN-232 measures the altitude from 0 to 50,000 ft. The system uses two
LPI Technology and Applications 43
Figure 2.1: Assemblies of the AN/APN-232 LPI radar altimeter [1] ( c 2003
Janes Information Group).
identical antennas mounted along the bottom of the aircraft (one for transmit
and one for receive) [2]. When the system is energized, it remains in search
until the reflected signal strength is sucient for the receiver portion to lock
on to the return signal. The AN/APN-232 is used on the Lockheed Martin
C-130 Hercules and its F-16 Fighting Falcon [3]. The display presented to
the pilot is similar to that shown in Figure 2.2.
The HG-9550 LPI radar altimeter system developed by Honeywell Sensor
and Guidance Products uses power management by controlling the emitter
power to produce an echo signal at a level 10 dB above the track threshold,
and transmits less than 1Wmaking it virtually undetectable. Other pro-
grammable LPI features include high sensitivity, frequency agility, jittered
code, and pulse repetition frequency. A microprocessor allows the track rate
and EP response to be varied as a function of real-time inputs, or to be pre-
programmed according to mission requirements [4]. The HG-9550 operates
at a frequency of 4.3 GHz, has a range of 0 to 50,000 ft, and a track rate of
2,000 ft/s. It also maintains an altitude accuracy of 4 ft. The HG-9550
is an o-the-shelf system currently in production for U.S. Air Force HC-130J
and C-17 Globemaster, U.K. C-130J, Argentine A-4 upgrade, the F-16 Block
60, the Boeing Joint Strike Fighter, and the Lockheed Martin Joint Strike
Fighter aircraft.
The cruise missile radar altimeter (CMRA) built by Honeywell Inc. Mili-
tary Avionics was developed specifically for cruise missile programs, including
the air launched cruise missile (ALCM) and Tomahawk missile. Honeywells
CMRA is a derivative product in which a variety of features from other Hon-
eywell altimeters are incorporated. The system has the capability to perform
terrain correlation and navigation functions [5]. Another Honeywell LPI al-
timeter is the AN/APN-209 LPI radar altimeter that is standard on all U.S.
Figure 2.2: AN/APN-232 LPI radar pilot display.
Army helicopters. Functions include transmitter power management, low-

and high-altitude warnings, analog and digital outputs, and integration of
the indicator, receiver, and transmitter [6].
The GRA-2000 LPI radar altimeter is being developed and tested by
the integrated product team of NAVAIRs Air Combat Electronics Program
Oce (PMA-2091), along with GEC-Marconi Hazeltine, General Microwave
Corporation, and Systems Maintenance and Technology, Inc. The GRA-2000
LPI altimeter has been selected by the U.S. Joint Services Program Oce
to replace the AN/APN-194, -171, -209, and -232 series altimeters on the
majority of tactical jet, helicopter, and transport aircraft employed by the
U.S. Department of Defense [7]. The design is based on using a high gain
receiver with a single-stage IF downconversion and specialized algorithms
to provide LPI and jam resistance. The altimeter has a high-speed digital
signal processor and achieves the LPI characteristics by combining a frequency
hopping, phase-coded waveform with a low power transmitter output signal
[8]. The assemblies for the GRA-2000 are shown in Figure 2.3. The small size
enables easy mounting to a variety of platforms. The GRA-2000 has a range
of 0 to 35,000 ft. It also maintains an altitude accuracy of 2 ft. (05,000
ft) and 50 ft. (5,00035,000 ft).
The PA-5429 pulsed airborne radar altimeter built by Tellumat, South
Africa, provides the height between the altimeter and the underlying ter-
rain/surface for heights from 0 to 5,000 ft. The altimeter operates in the
mid-J-band (15 GHz) and features a self-contained installation, eliminating
the need for separate RF feed cables and antennas. The altimeter has good EP
Figure 2.3: GRA-2000 LPI radar altimeter set [6] ( c 2003 Janes Information
Group).
performance, with a low probability of intercept and comprehensive EP, mak-

ing it suitable for a wide range of applications, including high-performance
and transport aircraft, helicopters, and missiles. The accuracy of the PA-5429
is 3 ft for heights 0-100 ft and 3% for heights between 100 and 5,000 ft
[9].
Other LPI altimeters include the Thales (originally Thompson CSF) AHV-
2100 digital radar altimeter [10] and the BAE AD1990, both operating at 4.3
GHz. The AHV-2100 uses power management of the RF output to reduce
the probability of interception at low altitude over water, and the combina-
tion of a narrow receiver bandwidth with digital signal processing to provide
EP from jamming. The AD1990 radar altimeter was designed for the U.K.
Royal Air Forces Tornado in the 1990s and has a maximum operating alti-
tude of 5,000 ft down to ground level. The altimeter was ahead of its time
with LPI being achieved by spreading the transmitted signal over a very wide
bandwidth through the application of pseudorandom phase modulation and
adaptive power management.
2.2 Landing Systems

2.2.1 Introduction
Landing an aircraft and especially an unmanned aerial vehicle (UAV) is dif-
ficult for several reasons. Landing involves the air vehicle switching between
dierent modes of operation (e.g., takeo, landing, and hovering). The air
vehicle must also coordinate with the landing site using voice or data links.
Automatic and precision landing systems transmit a beacon and can aid in
the landing operation, but must be LPI to remain active on the battlefield.
2.2.2 Fielded LPI Landing Systems

The AN/SPN-46 is an automatic precision approach and landing system
(PALS) for aircraft carriers and amphibious assault ships. Built by Tex-
tron, the AN/SPN-46 PALSs are installed on all U.S. Navy aircraft carriers,
and provide safe and reliable final approach and landing guidance for Marine
Corps helicopters and AV-8B Harrier vertical and/or short takeo and land-
ing (VSTOL) attack aircraft during day/night operations and adverse weather
conditions. The PALS employs LPI technology using both an fc = 9.3 and
33.2-GHz carrier frequency mainly to obtain adequate accuracy. As in many
other system examples where two bands are used, one band is used for search
while another higher frequency band is used for tracking. Sometimes one
lower band can be used for initial tracking through weather, and a higher
frequency for more precision at short range.
The PALS employs the 9.3-GHz coherent transmitter and receiver with
monopulse tracking and Doppler processing on received signals for clutter
rejection and rain attenuation at an operating range of 15 km [11]. The
PALS is also capable of controlling up to two aircraft simultaneously in a
leapfrog pattern, because of two dual-band radar antennas/transmitters.
As each approaching aircraft being assisted by the system lands, another
can be acquired. As of 2002, Janes sources were reporting the SPN-46(V)
as being in service aboard the U.S. Navy aircraft carrier Enterprise (two
radar installed), the Kitty Hawk, the John F. Kennedy (two radar installed),
and the Nimitz (two radar installed) class aircraft carriers. The SPN-46(V)
was also noted as being a retrofit option for the U.S. Navys Wasp class
amphibious assault ships [12].
The Sierra Nevada tactical automatic landing system (TALS) is an all-
weather, transponder tracking radar system designed for land-based environ-
ments and interoperability with any ground control station (GCS). It is an
upgrade of the AN/UPN-51(V) UAV Common Automatic Recovery System
(UCARS) and features a millimeter wave (MMW) K-band (35 GHz) radar. It
uses a narrow beamwidth antenna for close-range LPI acquisition in fog and
rain, and an omnidirectional antenna for rollout. Due to its LPI signature,
the TALS has minimal impact to the host aircraft [13]. At a range of 3.7 km,
the transmit power for the airborne transponder is 100 mW with a 60% duty
cycle, while the ground tracking radar transmits a maximum of 1W with a
0.04% duty cycle. A photo of the TALS system is shown in Figure 2.4. The
ground tracking subsystem, contained in a portable unit, locates and accu-
rately tracks the airborne transponder, using high-bandwidth tracking loops
to cover touchdown and roll-out. Recovery software, proven in UCARS, per-
forms air vehicle guidance and control calculations. The recovery of a UAV
using the TALS is similar to the UCARS system shown in Figure 2.5 [14].
The TALS has been ordered for the AAI RQ-7A Shadow 200 tactical UAV
system and recently underwent U.S. Army trials in 2000.
Figure 2.4: The tactical automatic landing system (TALS) showing the 35-
GHz antenna [13] ( c 2002 Janes Information Group).
Figure 2.5: Steps in the recovery of a UAV using a UCARS [14] ( c 2002
2.3 Surveillance and Fire Control Radar

2.3.1 Battlefield Awareness
On the battlefield, situational awareness and threat evaluation are achieved
using tactical surveillance radar to detect and track targets. For covert op-
erations, the detection and tracking of targets should be as quiet as possible.
These systems employ LPI technology to decrease the probability of passive
detection by hostile forces; that is, to see without being seen. The role of
multimode airborne fire control radar is to provide the eyes for tactical fighter
aircraft within an air dominance mission and also require LPI operation.
2.3.2 LPI Ground-Based Systems

Ericsson Microwave Systems has produced several LPI radar systems for
fire control and surveillance, including the Improved Helicopter and Air-
craft/Radar Detection (HARD)-3D, the Eagle, and the Pointer. The Im-
proved Hard-3D is a solid-state, 3D search and acquisition radar that has
been designed for use in short-range air defense systems. The Improved
HARD features an LPI capability that is due to a low output peak power
of 240W (30W average), broadband frequency agility, low side lobes, and a
narrow antenna beam [15]. The 3D capability is achieved by an electronically
scanned beam in elevation with intelligent beam control, providing a short
reaction time after the target is detected. The elevation coverage is obtained
by steering the antenna beam to a number of fixed elevations on a pulse-to-
pulse basis. Figure 2.6 shows the HARD-3D radar mounted on a Hagglunds
vehicle. The elevation search pattern covers up to 35 degrees in elevation
within two antenna revolutions. Upon target detection, the beam pattern is
controlled so that a secondary detection will always occur in the next revolu-
tion for immediate confirmation and track initiation [15]. In a special pop-up
mode, the track will start automatically after the first detection. Up to 20
targets and five jammers can be tracked automatically in range, azimuth, and
elevation. Instrumented ranges are 12 and 20 km.
The Ericsson Microwave Systems Eagle is a fire-control LPI radar intended
for mobile ground and naval-based air defense systems. The equipment op-
erates in the K-band (35 GHz) and is used to track low-flying targets and
perform air-to-surface missile alert and closed-loop fire control. The Ka-band
waveform provides a narrow antenna beam for low-altitude tracking at short
range, as is required for gun fire control. The Eagle system is shown in Fig-
ure 2.7 and achieves LPI operation by using a low-output peak power (20W),
pulse compression, high antenna gain with extremely low side lobes [16]. The
radiation pattern and a new transmission technique claim to make it impossi-
ble for escort or stand-o jammers to degrade the radar performance. It can
track two targets simultaneously with an angular error of less than 0.2 mrad
Figure 2.6: HARD-3D radar on Hagglunds vehicle [15] ( c 2002 Janes Infor-
mation Group).
Figure 2.7: Missile control and launch vehicle with the mast-mounted Eagle
radar [16] ( c 2003 Janes Information Group).
Figure 2.8: Pointer LPI radar system antenna [17] ( c 2003 Janes Information
Group).
at 10 km.
The Ericsson Microwave Systems Pointer radar system is a short-range
LPI air surveillance 3D solid state radar system that was designed to be
integrated into short-range air defense missile systems such as the Mistral,
Stinger, and Starburst [17]. Pointer is a fully autonomous system that in-
cludes an X-band radar and the antenna shown in Figure 2.8. The range
of Pointer is typically over 20 km, and 9 to 10 km in altitude. Pointer can
be brought into action in 1 minute by a single operator. Most functions of
Pointer, including track initiation, tracking, classification of fixed-wing air-
craft and helicopters, threat evaluation, and data distribution via radio or
wire are fully automatic. Target information can be sent to the firing unit 2
or 3 seconds after the target enters the line of sight. All the missile system
operator then has to do is to acquire, track, and engage the target [17]. The
Pointer operator can be positioned a long distance from the radar, using the
radar remote control laptop computer shown in Figure 2.9. The Pointer was
designed to increase the overall eectiveness of short-range air defense sys-
tems by reducing target acquisition time, as well as enabling more targets to
be engaged [17]. Pointer builds on Ericsson Microwave Systems experience
in the development of both the HARD-3D and Eagle LPI radar.
Hollandse Signaalapparaten is developing a LPI radar as an alternative
Figure 2.9: Pointer operators radar remote control unit that is deployed away
from the radar unit [17] ( c 2003 Janes Information Group).
to the use of infrared sensors for short-range missile and gun systems. The
companys PAGE (portable air-defense guard equipment) is a lightweight and
inexpensive 810 GHz (I-band) FMCW emitter with a transmit power of only
1020W, providing a detection range of 1015 km [18]. The PAGE LPI radar
system, shown in Figure 2.10, exploits Signaals experience in developing its
Scout family of naval and land-based FMCW ground surveillance LPI radars.1
It is also being developed as a private venture by Thales Nederland as a low-
level air surveillance radar which provides early warning and cueing data for
short-range networked air defense applications. Especially of interest are the
man-portable surface-to-air missiles (SAM) and light anti-aircraft guns. The
PAGE can survive the most extreme EA conditions and is nearly undetectable
by ES and radar warning receivers.
The PAGE system consists of an antenna unit including a solid-state trans-
ceiver, a radar processor unit, an operator unit, and a small generator. It also
has the capability to integrate an identification friend-or-foe (IFF). (An au-
tomated datalink and weapons terminal can be added to PAGE for real-time
data processing at remote fire units). Configurations include a man-portable
tripod version, a light vehicle or trailer mounting and installation on existing
self-propelled anti-aircraft gun or SAM systems [18].
1 The GB-Squire is a variant of the PAGE, and has detected artillery shells and Browning
0.50-calibre machine gun bullets in flight during testing.

Figure 2.10: PAGE LPI radar system [18] ( c 2006 Janes Information Group).
The Thales Nederland (formerly Signaal) Variant shown in Figure 2.11,

is a dual-band (46 GHz and 810 GHz) radar (one octave apart) that is rel-
atively low-cost low-power, and lightweight making it ideal for a broad range
of vessels, including fast patrol boats, amphibious vessels and support ships
[19]. It has an autonomous target detection and tracking capability and is
intended to fill three principal functions. These include surface target de-
tection and tracking, air target detection and tracking, and gunfire targeting
support.
The Variant uses an integrated solid-state FMCW emitter with a transmit
power Pavg = 10 mW. The antenna rotates at 14 rpm for long-range surveil-
lance and at 28 rpm for a higher update rate (for self-defense applications).
The system is fully coherent and provides pulse Doppler detection and track-
ing algorithms for optimal clutter suppression and air targeting. Spread spec-
trum techniques are used to enable detection and classification of helicopters.
Surface gunfire support is provided for by three fire-control/splash-spotting
windows, eliminating the need for a dedicated tracking radar for engagement
of surface targets [19]. The system is able to detect air and surface targets out
to instrumented ranges of 60 km and 70 km respectively. Important features
include the ability to detect hovering and slow-moving helicopters, and a high
resistance to jamming, weather clutter, and multipath propagation due to the
dual-band operation and LPI operation.
Figure 2.11: Variant LPI radar system [19] ( c 2006 Janes Information
Group).
The quiet naval radar CRM-100, built by Przemyslowy Instytut Teleko-

munikacji Telecommunications Research Institute in Poland, is a solid-state
FMCW LPI radar that uses 10 switched frequencies in the 9.3- to 9.5-GHz
subband. It has a modulation period of 1 ms and a modulation bandwidth of
54 MHz (max). The modulation bandwidth chosen depends on the selected
range scale of 1.4, 3, 5.6, 11.1, 22.2, or 44.5 km, resulting in range cell sizes
3, 6, 12, 24, 48, and 96m, respectively [20]. The range resolution is three
times the range cell size. Designed as a surface surveillance radar, this range
coverage is similar to standard navigational radar that uses a pulsed signal.
The transmit waveform is power managed, depending on the range to the tar-
get, and ranges from 1 mW to 1W. A line drawing of the CRM-100 antenna
is shown in Figure 2.12. The radar has a beamwidth of 1.8 degrees in the
horizontal dimension, and 25 degrees in the vertical dimension with side lobes
27 dB. The scan rate is 30 revolutions per minute (RPM). The receiver has
an IF bandwidth of 500 kHz and a noise figure of 3 dB. The CRM-100 is
designed to detect surface targets and determine their coordinates [20]. It
provides automatic tracking of targets and automatic transfer of data on the
tracked targets to command and control systems. The radar can be installed
on a ground vehicle as shown in Figure 2.13 (shore version) or on board a
ship (marine version).
The Chinese JY-17A, shown in Figure 2.14, is a fully coherent, medium-
range pulse Doppler battlefield surveillance radar that is designed to detect,
locate, and identify moving ground or low-altitude air targets. Built by the
Figure 2.12: The CRM-100 quiet naval radar antenna (measurements shown
in millimeters) [20] ( c 2003 Janes Information Group).
Figure 2.13: The CRM-100 quiet naval radar installation [20] ( c 2003 Janes
Information Group).
Figure 2.14: JY-17A medium-range ground surveillance radar [21] ( c 2003

East China Institute of Electronic Engineering, the radar can be vehicle

mounted or grounddeployed. The radar features a solid-state, LPI trans-
mitter in the 8- to 12-GHz range, and a high-stability frequency synthesizer
[21]. It also has a selective linear and circular polarization antenna with low
side lobes and digital phase coding, random frequency shift keying, with pulse
Doppler processing that has automatic target detection and tracking. It can
detect a single pedestrian at 10 km, a light vehicle at 15 km, a helicopter at
20 km, and a ship at 30 km [21].
The Raytheon multirole survivable radar (MRSR) is a tactical target ac-
quisition and tracking LPI radar for the U.S. Army Missile Command to meet
the tactical air defense requirements in the high-to-medium air defense and
forward-area air defense mission areas. The radar is a 3D track-while-scan,
phased array in elevation radar; designed to acquire and track multiple air-
borne targets over a 360-degree azimuth at extended ranges and at all tactical
altitudes. Targets include tactical aircraft, UAVs, and hovering and slowly
moving helicopters. The LPI radar incorporates a solid-state, low noise trans-
mitter, and operates over a wide bandwidth with frequency agility [22]. The
radar aperture is optimized to resist advanced EA and antiradiation missiles
by employing very low side lobes combined with the LPI waveform. Multiple
beams are moved electronically in elevation, with one continuously scanning
the horizon with its bottom edge touching the ground, producing hot spots
to confuse antiradiation missile seekers [22].
2.3.3 LPI Airborne Systems

The AN/APS-147 multimode radar is an inverse synthetic aperture radar
(ISAR) system designed to support the multimission capability of the light
airborne multipurpose (LAMPS) SH-60B helicopter system during maritime
surveillance and patrol missions. Power management and frequency agility
give the operators the ability to perform missions at output power levels lower
than traditional maritime surveillance radar. This enables the radar to detect
medium-to-long-range targets with an LPI capability against enemy intercept
receivers. Radar modes include target imaging, small target (periscope) de-
tection, long-range surveillance, weather detection and avoidance, all-weather
navigation, short-range search and rescue, and enhanced LPI search and tar-
get designation [23].
The AN/APQ-181 is the LPI radar designed specifically for the Northrop
Grumman B-2 Spirit stealth bomber. The B-2 is in use by the U.S. Air Force
and is shown in Figure 2.15. The radar operates in the J-band (12.518 GHz),
using 21 separate modes for terrain following and terrain avoidance, naviga-
tion system updates, target search, location, identification and acquisition,
and weapons delivery [24]. The radar employs two electronically scanned an-
tennas and advanced LPI techniques that match the aircrafts overall stealth
qualities. The antenna is electronically steered in two dimensions and features
a monopulse feed design to enable fractional beamwidth angular precision. It
is designed to have a low RCS with respect to both in- and out-of-band RF
illumination [25].
The AN/APG-77 is an advanced multimode tactical radar and is the pri-
mary sensor for the F-22 Raptor fighter aircraft built by Northrop Grumman
(with Raytheon). A photo of the F-22 Raptor is shown in Figure 2.16. The
LPI nature of the APG-77 radar provides a significant advantage for the F-
22. The F-22 is able to detect RWR/ES-equipped fighter aircraft without
them knowing they are being illuminated [26]. The APG-77 emits low energy
pulses over a wide frequency band. That is, the emitter changes frequency
and power levels after every pulse, in order that no two transmitted pulses are
alike. When multiple echoes are sent back to the radar, the signal processor
converts the signals together instead of individually. The radar antenna is a
fixed, elliptical, electronically scanned active array that contains 2,000 trans-
mit and receive (TR) modules [27]. The antenna also contains circulators,
radiators, and manifolds assembled into subarrays and then integrated into a
complete array. The active array requires significantly less volume and prime
power than a gimbaled slotted array. The antenna is integrated both phys-
ically and electromagnetically with the airframe and has a low radar cross
section. The active array provides frequency agility, low radar cross section,
agile beam steering, and a wide bandwidth capability typical of LPI radar.
Figure 2.15: The Northrop Grumman B-2 bomber carries the Raytheon Elec-
tronic Systems AN/APQ-181 radar [24] ( c 2002 Janes Information Group).
Figure 2.16: The F-22 Raptor employs the LPI AN/APG-77 radar [27]
( c 2003 Janes Information Group).
The low-altitude navigation and targeting infrared for night (LANTIRN) is

a system consisting of two pods that allow aircrew to fly their aircraft by day
or night and in adverse meteorological conditions. The LANTIRN consists of
a navigation pod and a targeting pod. The navigation pod contains a wide
field of view forward looking infrared (FLIR) and a Ku-band LPI terrain
following radar, the AN/APN-237A, that can be linked directly to the F-16s
autopilot to automatically maintain a preset altitude down to 100 feet while
flying over virtually any kind of terrain. It has five modes: normal, weather,
EP, LPI, and very low clearance [28]. The targeting pod contains another
FLIR and a laser designator/rangefinder. The LANTIRN is deployed on the
F-16C/D, F-15E/I/S, and F-14 platforms.
2.4 Antiship Capable Missile and Torpedo

Seekers
2.4.1 A Significant Threat to Surface Navies
Antiship capable missiles (ASCMs) are a significant threat to navy surface
ships. Active RF ASCM seekers that radiate substantial transmitter power,
however, allow themselves to be detected by relatively modest intercept re-
ceivers in both the main and side lobes. The intercept of seeker transmissions
ultimately leads to vulnerability through the use of antiradiation missiles,
missile interceptors, or EA. In the future, RF seekers will have LPI, power-
managed operation in the 8- to 20-GHz range as well as the 35- and 96-GHz
ranges, by incorporating a number of advanced electronic technologies. These
technologies will enable the missile to generate a broad collection of wide-
band reprogrammable waveforms with bandwidths reaching 500 MHz to 1
GHz. Using a variety of wideband techniques and coherent range-Doppler
processing, these seekers will eectively target low radar cross section ships,
while simultaneously allowing the seeker to escape detection and reject decoys
such as cha. Chapter 7 examines ASCM seeker technology and explores a
missile-ship engagement scenario where the missile uses a power-managed,
LPI seeker to detect a low RCS ship in several sea states.
2.4.2 Fielded LPI Seeker Systems

The Saab Bofors Dynamics ABs RBS-15 medium-range, radar-guided, air-
to-surface missile is one of a family of long-range ASCMs produced in Sweden
that can be launched from the air, land, or sea [29]. The missile makes use of
low RCS materials to reduce the likelihood of early detection by enemy radar
and also has a low infrared signature to reduce the probability of detection
by infrared search and track systems. A picture of the RBS-15 is shown in
Figure 2.17. In the 1990s, the company developed and tested an LPI radar
Figure 2.17: RBS-15 missile.
seeker for the RBS-15 (Mk 2). The seeker uses FMCW technology and has
output power in the milliwatt range that is progressively reduced as the missile
approaches the target. In 1994, Saab began work on the next generation
RBS-15 (Mk 3) that incorporates an updated version of the current turbojet
engine, providing a range in excess of 200 km [29]. Saab is developing a
future land-attack version of the RBS-15 Mk 3 and is working on several new
seeker technologies that may be applicable and that may also be retrofitted to
existing variants. These include synthetic aperture radar, which would boost
seeker resolution by more than 100% and substantially increase the seekers
target discrimination capability as well as the terminal aimpoint accuracy.
Another option is an LPI radar seeker that would use long, coded pulses that
are dicult to detect and dicult to jam. Prototypes for both the synthetic
aperture and LPI seekers are currently under test [29]. The improved Mk 3
version uses a global positioning system (GPS) data link, and the range has
been increased to 400 km.
Sweden is also developing automatic target recognition (ATR) systems
that would give the missile a better discrimination capability. One option is
for a dual-mode seeker version that combines the LPI radar with an imaging
infrared (IIR) seeker, using ATR for terminal guidance. Figure 2.18 shows
the RBS-15 missile being fired from a ground-based launch site.
Another type of LPI approach is the random noise emitter. DARPA
is investigating this type of seeker for the miniature air-launched intercep-
tor (MALI). The MALI is a supersonic armed version of the miniature air-
launched decoy (MALD) and is used to intercept cruise missiles in flight from
the rear. Figure 2.19 shows the MALI mounted on an aircraft ready for
launch. The noise seeker is a Ka-band (35 GHz) seeker with 1-GHz band-
width, and transmits randomly generated noise signals to detect and home
in on the cruise missile. The randomly generated noise signals are copied
and stored in seeker memory in order to correlate with the radar return. Not
only does the randomness of the noise seeker make it harder for an intercept
Figure 2.18: RBS-15 missile firing.
Figure 2.19: Miniature air-launched interceptor [30] ( c 2003 Aviation Week

and Space Technology).
Figure 2.20: Spearfish being loaded [31] ( c 2003 Janes Information Group).
receiver to detect the seekers transmission, the wide bandwidth provides an

imaging capability that makes it easier to distinguish low-flying cruise mis-
siles from the clutter. The large bandwidth provides excellent range resolution
and a large processing gain, while the random noise pulse eliminates range
ambiguities and is resistant to certain advanced countermeasures. This ap-
proach has been made possible by recent advances in high-speed, low-power
processing [30].
Torpedo-homing performance in littoral regions has traditionally suered
due to poor acoustics found in the shallow-water environment. For example,
shallow water has more pronounced temperature gradients (particularly in
equatorial regions) that distort the sound-ray path and can result in non-
detection or skip zones. Also, active sonar performance is degraded by the
proximity of the surface- and bottom-reflecting boundaries, while passive
sonar suers as a result of wave noise and marine life.
Conceived during the Cold War, the Spearfish torpedo was optimized to
defeat fast, deep-diving, Soviet nuclear-powered submarine threats [31]. With
the emphasis now on operating in littoral zones against small, ultra-quiet
diesel-electric submarines, BAE Systems and QinetiQ have been researching
torpedo sonar and signal-processing techniques that form the basis for an
upgrade to the British Royal Navys Spearfish heavyweight torpedo shown
in Figure 2.20. The Advanced Spearfish update program is intended to im-
prove substantially the weapons performance against quiet targets in shal-
low water environments, while at the same time solving obsolescence issues
aecting Spearfishs existing hybrid processing architecture. Digital signal-
processing (DSP) techniques have been considered, along with microprocessor
technology, to handle the high computational loads demanded. New technol-
ogy includes wide bandwidth processing, complex waveforms with additional
modulations, LPI active waveforms, adaptive beam forming, neural net clas-
sification, and advanced tracking [31]. Adaptive beam-forming is also used to

overcome the eects of EA by noise jammers.
2.5 Summary of LPI Radar Systems

This chapter presented several LPI radar system applications where their
design intentionally (and sometimes unintentionally) makes their transmis-
sion dicult to intercept. Table 2.1 summarizes the systems discussed, along
with their application or use. It is important to note that indentifying an
LPI radar as any radar system that uses higher than conventional duty cy-
cles (through pulse compression or CW operation), solid state transmitters,
low side-lobe antennas, or low transmitter power, can end up leading to a
misclassification. For example, under this definition, the new ballistic missile
early warning system (BMEWS), Pave phased array warning system (PAWS)
radar, airborne warning and control system (AWACS) radar, air route surveil-
lance radar model 4 (ARSR-4), and any police CW radar would be classified
as LPI, which is certainly not correct. In the next few chapters, details on
the LPI technology and the important pulse compression techniques used in
the above applications are presented.
Table 2.1: LPI Radar Systems
Developer System LPI Use
NavCom Defense Electronics AN/APN-232 Combined altitude radar altimeter

Honeywell HG-9550 Radar altimeter
NAVAIR GRA-2000 Tri-service radar altimeter
Tellumat, South Africa PA-5429 Radar altimeter
Honeywell CMRA Cruise missile radar altimeter
Thompson CSF AHV-2100 Radar altimeter
BAE AD1990 Radar altimeter
Textron Systems AN/SPN-46 (V) Precision approach, automatic landing
Sierra Nevada TALS Tactical automatic landing system
Saab Bofors Pilot Surveillance, navigation
Signaal Scout Surveillance, navigation
Signaal Smart-L Surveillance
Ericsson Microwave Systems HARD-3D Fire control and surveillance
Ericsson Microwave Systems Eagle Fire control
Ericsson Microwave Systems Pointer Air surveillance radar
Thales Nederland PAGE Air surveillance
Thales Nederland Variant Surface and air target, gun fire detection
PITT Research Institute, Poland CRM-100 Surface target detection
China Inst. of Elec. Engineering JY-17A Battlefield surveillance radar
Raytheon MRSR Target acquisition and tracking radar
TI AN/APS-147 Enhanced search and target designation
Raytheon AN/APQ-181 Tactical multimode fire control radar
Northrop Grumman AN/APG-77 Multimode tactical radar
Raytheon AN/APG-70 Multimode tactical radar
TI LANTIRN Terrain following radar
Saab Dynamics RBS-15MR Radar guided air-to-surface missile
BAE Spearfish Torpedo for littoral environments
References
[1] AN/APN-232 combined altitude radar altimeter, Janes Avionics - Military
CNS, FMS, Data and Threat Management, Feb. 10, 2003.
[2] http://www.osmpf.wpafb.af.mil/.
[3] Lockheed Martin Aeronautics Company, Janes All the Worlds Aircraft-
Fixed Wing-Military, Jan. 10, 2003.
[4] HG9550 LPI radar altimeter system, Janes Avionics - Military CNS, FMS,
Data and Threat Management, April 26, 2002.
[5] Cruise missile radar altimeter, Janes Radar, July 17, 1994.
[6] AN/APN-209 radar altimeter, Janes Avionics - Military CNS, FMS, Data
and Threat Management, Feb. 10, 2003.
[7] GRA-2000 low probability of intercept (LPI) altimeter, Janes Avionics -
Military CNS, FMS, Data and Threat Management, Feb. 5, 2003.
[8] http://www.cni.na.baesystems.com/html/low probability of intercept a.html
[9] PA-5429 radar altimeter, Janes Avionics - Military CNS, FMS, Data and
Threat Management, Feb. 5, 2003.
[10] AHV-2100 digital radar altimeter, Janes Avionics - Military CNS, FMS,
Data and Threat Management, Feb. 10, 2003.
[11] http://www.fas.org/man/dod-101/sys/ship/weaps/an-spn-46.htm.
[12] AN/SPN-46(V) approach radar, Janes Radar and Electronic Warfare Sys-
tems, Military Air Trac Control, Instrumentation and Ranging Radars, Feb.
7, 2003.
[13] Sierra Nevada TALS, Janes Unmanned Aerial Vehicles and Targets - Launch
and Recovery Systems, April 17, 2002.
[14] Sierra Nevada UCARS, Janes Unmanned Aerial Vehicles and Targets -
Launch and Recovery Systems, April 17, 2002.
[15] Ericsson Microwave Systems Improved HARD-3D radar system, Janes
Land-Based Air Defence-Anti-Aircraft Control Systems, Oct. 23, 2002.
[16] Eagle fire-control radar, Janes Radar and Electronic Warfare Systems,
Battlefield, Missile Control and Ground Surveillance Radar Systems, Jan.
30, 2003.
[17] Ericsson Microwave Systems Pointer radar system, Janes Land-Based Air
Defence-Anti-Aircraft Control Systems, 1999.
[18] Hewish, M., Low-level air defensenew sensors enhance eectiveness, Janes
Defence Equipment and Technology, Vol. 27, No. 6, pp. 43, June, 1994.
[19] Aordable performers: surveillance radars balance cost with capability,
International Defence Review, Mar., 2008.
[20] CRM-100 surveillance radar, Janes Radar and Electronic Warfare Systems-
Naval/Coastal Surveillance and Navigation Radar, Jan. 30, 2003.
[21] JY-17 battlefield reconnaissance radar, Janes C4I Systems-Land Based
Surveillance and Location, April, 22, 2002.
[22] Multi-role survivable radar - Tactical target acquisition and tracking, Janes
Air Defence Radar - Land and Sea, Jan. 1997.
[23] AN/APS-147 multimode airborne radar, Janes Avionics - Military CNS,
FMS, Data and Threat Management, Feb. 10, 2003.
[24] AN/APQ-181 radar for the B-2, Janes Avionics - Military CNS, FMS,
Data and Threat Management, Jan. 17, 2003.
[25] http://www.raytheon.com/products/apq181/.
[26] http://f22rap.virtualave.net/avionics.html.
[27] AN/APG-77 multimode airborne radar, Janes Radar and Electronic War-
fare Systems-Airborne Fire Control Radar, Nov. 11, 2002.
[28] http://www.f-16.net/reference/armament/lantirn.html.
[29] RBS 15F, Janes Air-Launched Weapons-Air to Surface Missiles, Sept. 12,
2002.
[30] R. Wall, USAF eyes decoy, jammer as MALI demonstration ends, Aviation
Week and Space Technology, Jan. 13, 2003.
[31] Spearfish, Janes Underwater Warfare Systems - Torpedos, May 3, 2002.
Problems
1. In an FMCW altimeter such as the CARA, the frequency is swept
over the modulation bandwidth F during the modulation (coherent
processing) period tm . Ranging (determining the altitude) occurs by
mixing a sample of the transmitted signal with the reflected signal from
the surface to derive a dierence frequency (or beat frequency) f . (a)
Write an expression for the time interval that corresponds to the mea-
sured beat frequency f as a function of the sweep rate F . (b) Deter-
mine the beat frequency (in Hz) for the CARA if the modulation period
is 1 ms and the altimeter is at a height of 30m. HINT: the wavefront
takes 6.7 s/km to travel a round-trip path.
2. Estimate the maximum intercept range of the TALS ground tracking
radar if its MMW antenna is 60% ecient and the intercept receiver
sensitivity is I = 100 dBmi.
3. Estimate the HARD-3D pulse width R if the maximum unambiguous
range is 20 km.
4. Determine the maximum detection range of the Eagle radar, considering
that the antenna is 90% ecient.
Chapter 3
Ambiguity Analysis of LPI

Waveforms
In this chapter, the ambiguity (delay-Doppler) analysis of LPI waveforms is
addressed. Ambiguity analysis is important to understand the properties of
the CW waveform and its eect on measurement accuracy, target resolution,
ambiguities in range, and radial velocity, and its response to clutter. The
periodic autocorrelation function (PACF) is introduced, and it is shown that
CW signals can have a perfect PACF with zero side lobes. The periodic
ambiguity function (PAF) is also introduced, to analyze the response of a
matched receiver that uses N copies of the reference (transmitted) function to
cross-correlate the return CW signal and perform target detection. The PAF
is similar to the ambiguity function often used to represent the magnitude of
the matched receiver output for a coherent pulse train. The cut of the PAF
at zero Doppler ( = 0) is the PACF, and cuts of the PAF along zero delay
( = 0) yield the response of the correlation receiver at a given Doppler shift.
Several important properties of the PAF are presented.
The MATLAB low probability of intercept toolbox (LPIT) is described
(Appendix A) and is used to generate a CW Frank signal in order to demon-
strate the PACF and PAF concepts. The MATLAB code used to calculate
the PACF and PAF is also described (Appendix B). Modifying the reference
waveform in the receiver with a weighting function (mismatched receiver) can
help reduce the Doppler side lobes that appear. This subject is addressed,
and three important weighting functions are presented.
67
3.1 The Ambiguity Function

A matched radar receiver performs a cross-correlation of the received signal
and a reference signal, whose envelope is the complex conjugate of the enve-
lope of the transmitted signal. The ambiguity function describes the response
of this matched receiver to a finite duration signal. In ambiguity analysis,
the receiver is considered matched to a target signal at a given delay and
transmitted frequency. The ambiguity is then a function of any added delay
and additional Doppler shift from what the receiver was matched to. If u(t)
is the complex envelope of both the transmitted signal and received signal,
the ambiguity function is given by [1]
e8 e
e e
|(, )| = ee u(t)u (t )ej2t dtee (3.1)

where is the time delay and is the Doppler frequency shift. The 3D plot,
as a function of and , is called the ambiguity diagram. The maximum of
the ambiguity function occurs at the origin ( = 0, = 0), and |(0, 0)| is
the output if the target appears at the delay and Doppler shift for which the
filter was matched. The delay-Doppler response of the matched filter output is
important for understanding the properties of the radar waveform [2]. Ideally,
the ambiguity diagram would consist of a diagonal ridge centered at the origin,
and zero elsewhere (no ambiguities). The ideal ambiguity function, however,
is impossible to obtain. For a coherent pulse train consisting of NR pulses with
pulse duration R and pulse repetition interval Tr , the ambiguity function
indicates that the Doppler resolution is the inverse of the total duration of
the signal NR Tr while the delay resolution is the pulse duration [3].
3.2 Periodic Autocorrelation Function

LPI signals are typically low-power CW waveforms that are modulated by a
periodic function, such as a phase code sequence or linear frequency ramp. A
major advantage of the periodically modulated CW waveforms is that they
can yield a perfect PACF. For example, consider a phase-coded CW signal
with Nc phase codes each with subcode duration tb s. The transmitted CW
signal has a code period T = Nc tb s and a periodic complex envelope u(t)
given as
u(t) = u(t + nT ) (3.2)
for n = 0, 1, 2, 3 . . . . The values of the PACF as a function of the delay
r (which are multiples of tb ) are given by
Nc
1 3
R(rtb ) = u(n)u (n + r) (3.3)
Nc n=1
Ambiguity Analysis of LPI Waveforms 69
and ideally we would like a perfect PACF or

F
1, r = 0(modNc )
R(rtb ) = (3.4)
0, r = 0(modNc )
Since the CW signal is continuous, the perfect PACF is possible. Note how-
ever, that finite duration signals, such as a pulse train, cannot achieve this
ideal autocorrelation since as the first sample (or last sample) enters (or
leaves) the correlator, there is no sample that can cancel the product to yield
a zero output.
3.3 Periodic Ambiguity Function

The periodic ambiguity function or PAF, introduced by Levanon and Freed-
man [4], describes the response of a correlation receiver to a CW signal mod-
ulated by a periodic waveform with period T , when the reference signal is
constructed from an integral number N of periods of the transmitted signal
(coherent processor length N T ). The target illumination time (dwell time)
P T must be longer than N T (see Figure 1.9). As long as the delay is
shorter than the dierence between the dwell time and the length of the ref-
erence signal 0 (P N )T , the illumination time can be considered
infinitely long and the receiver response can be described by the PAF given
as [5] e e
e 1 8 NT e
e e
|N T (, )| = e u (t ) u (t) ej2t dte (3.5)
eNT 0 e
where is assumed to be a constant, and the delay rate of change is rep-
resented by the Doppler shift . The PAF for N periods is related to the
single-period ambiguity function by a universal relationship
e e
e sin(N T ) e
e
|NT (, )| = |T (, )| e e (3.6)
N sin(T ) e
where e8 e
1 e T e
e j2t e
|T (, )| = e u(t )u (t)e dte (3.7)
T e 0 e
is the single period ambiguity function. The single period ambiguity function
is multiplied by a universal function of N and T that is independent of the
complex envelope of the signal and that does not change with . The PAF
shows the eect of using a reference receiver consisting of N code periods (see
Section 1.3). Examination of (3.6) reveals that for a large number of code
periods N , the PAF is increasingly attenuated for all values of except at
multiples of 1/T . It also has main lobes at T = 0, 1, 2, . . . . Equation
(3.6) also reveals that the PAF has relatively strong Doppler side lobes.
The PAF serves CW radar signals in a similar role to which the traditional
ambiguity function serves finite duration signals. Note that for a large N , the
PAF is compressed to zero for all , except near = n/T, n = 0, 1, 2, . . . .
For an infinitely large N , the function |N T (, )| becomes a train of impulses.
For large N , the PAF of a sequence exhibiting perfect periodic autocorrelation
will strongly resemble the ambiguity function of a coherent pulse train.
3.3.1 Periodicity of the PAF

The PAF formulation given in (3.5) is not unique, and alternate definitions
have also been adopted [6]. The form of the PAF in (3.5), however, represents
the straightforward implementation of a matched filter to the signal u(t)
delayed by and Doppler shifted by . It can easily be shown that the cut
along the PAFs delay axis |NT (, 0)| (zero Doppler) is the magnitude of the
PACF of the signal given by (3.3) [4, 6]. The cut along the Doppler axis (zero
delay) is
8 NT
1
N T (0, ) = |u(t)|2 ej2t dt (3.8)
NT 0
Assuming a constant amplitude signal, |u(t)| = 1 (e.g., phase-modulated CW
signals) e e
e sin(N T ) e
|N T (0, )| = ee e (3.9)
N T e
and
|N T (0, 0)| = 1 (3.10)
For any integer n, the periodicity on the delay axis is
|N T (nT, )| = |NT (0, )| (3.11)
For the axis, for m = 0, 1, 2, . . .
|N T (, m/T )| = |NT ( + nT, m/T )| (3.12)
The symmetry cuts are a function of the three parameters: the code period
T , the number of phase codes Nc , and the number of code periods used in
the correlation receiver N . Additional symmetry and periodicity properties
are discussed in [4, 6].
3.3.2 Peak and Integrated Side Lobe Levels

The time side lobe levels in the autocorrelation function (ACF) help quantify
the LPI waveform in its ability to detect targets without interfering side lobe
targets. That is, if the ACF has high side lobes, a second nearby target
might be able to hide in a side lobe and go undetected. To quantify the LPI
waveform characteristics, the peak side lobe level (PSL) of the ACF can be
defined as
} ] } ]
max side lobe power max R2 (k)
PSL = 10 log10 = 10 log (3.13)
(peak response)2 R2 (0)
where k is the index for the points in the ACF, R(k) is ACF for all of the
output range side lobes except that at k = 0, and R(0) is the peak of the
ACF at k = 0. The integrated side lobe level is
} ] M
3
total power in side lobes R2 (k)
ISL = 10 log10 2
= 10 log (3.14)
(peak response) R2 (0)
k=M
and is a measure of the total power in the side lobes as compared with the
compressed peak. The PSL is a useful measure when a single point target
response is of concern. Values for the PSL depend on the number of subcodes
in the code sequence Nc as well as the number of code periods N within the
receiver. The ISL is considered a more useful measure than the PSL when
distributed targets are of concern. Typical matched filter ISL values range
from 10 to 20 dB.
3.4 Frank Phase Modulation Example

To demonstrate the properties of the ACF, PACF, and PAF, we look briefly
at one important type of phase modulation called the Frank code [7]. The
Frank code is a polyphase code (more than two phase states). It has a variable
length and can be used to phase modulate a complex signal every subcode
period tb .
3.4.1 Transmitted Waveform

The transmitted signal can be written as
+
s(t) = Ae(j2fc t+k ) (3.15)
where fc is the carrier frequency and k is the phase modulation that is

used to shift the phase of the carrier in time every subcode period according
to the particular phase modulation used. Note that the carrier frequency
remains constant. The Frank phase modulation code is derived from a step
approximation to a linear frequency modulation waveform using M frequency
steps and M samples per frequency. If i is the number of the sample in a
given frequency and j is the number of the frequency, the phase of the ith
sample of the jth frequency for the Frank code is
2
i,j = (i 1) (j 1) (3.16)
M
Figure 3.1: Frank phase modulation for M = 8 (Nc = 64).
where i = 1, 2, . . . , M , and j = 1, 2, . . . , M . The Frank code has a length

of Nc = M 2 subcodes, which is also the corresponding pulse compression
ratio or processing gain P GR . For tb s (the subcode period), if cpp represents
the number of carrier cycles per subcode, then tb = cpp/fc s resulting in a
transmitted signal bandwidth B = 1/tb = fc /cpp. The code period can also
be expressed as
T = Nc tb = M 2 tb (3.17)
Below we examine the ACF, PACF, and PAF properties of this signal. Chap-
ter 5 discusses the Frank code in more detail.
3.4.2 Simulation Results

A signal containing the Frank phase modulation can easily be generated with
the MATLAB low probability of intercept toolbox distributed on the enclosed
CD and described in Appendix A. The LPIT can also generate a host of other
LPI signals discussed in Chapters 47.
Figure 3.1 shows the Frank phase modulation (3.16) with M = 8
(Nc = 64). The plot is generated within the LPIT. The carrier frequency
is fc = 1 kHz, fs = 7 kHz, and cpp = 1. Figure 3.2 shows the power spectral
Figure 3.2: Power spectral density for Frank phase modulation for M = 8
(Nc = 64) with fc = 1 kHz, fs = 7 kHz, and cpp = 1.
density of the Frank signal. This plot is also generated within the LPIT. Note
that since the cpp = 1, the 3-dB bandwidth B = 1 kHz, as illustrated.
The ACF and PACF are shown in Figure 3.3 for the number of code
periods N = 1. These results can be obtained by using the output waveforms
from the LPIT in conjuction with Levanons ambfn7.m code as described
in Appendix B with r = 1, F M tb = 10, T = 1, N = K = 100. The
PSL can be read from Figure 3.3(a). The largest side lobe level is 28 dB
down from the peak. This is in agreement with the theoretical result PSL =
20 log10 (1/M ) = 28 dB (voltage ratio). Also note from Figure 3.3(b) that
the CW Frank signal has a perfect PACF (zero side lobes). The PAF for
N = 1 is shown in Figure 3.4. The phase modulation signals generated
using the LPIT contain
cppfs
bsc = (3.18)
fc
number of samples per subcode. The total number of samples within a code
period is then Nc bsc . When ambfn7.m is used to examine the signals from
the LPIT, the delay axis is normalized by the subcode period tb and so the
PAF repeats at Nc bsc since the waveform is sampled. That is, dividing this
axis by the number of samples per subcode bsc gives the delay axis in terms
of the subcode number. For the LPIT default Frank signal (cpp = 1, fs = 7
kHz, fc = 1 kHz), bsc = 7 and, as illustrated in the plot, the code repeats
Figure 3.3: Frank (a) ACF (PSL = 28 dB down) and (b) PACF for M = 8
(Nc = 64), cpp = 1 with number of reference waveforms N = 1.
Figure 3.4: PAF for Frank phase modulation for M = 8 (Nc = 64), cpp = 1
with number of reference waveforms N = 1.
Figure 3.5: Frank (a) ACF (PSL = 40 dB down) and (b) PACF for M = 8
(Nc = 64), cpp = 1 with number of reference waveforms N = 4.
every = 448/bsc = 64 = Nc . The Doppler axis is normalized with respect to

the entire signal duration Nc tb . Therefore, depending on the number of code
periods N integrated into the PAF calculation, the Doppler lobes appear at
kN for k {0, 1, 2, . . .} as illustrated in Figure 3.4 for N = 1.
Increasing the number of code periods N used in the receiver can help to
decrease the Doppler side lobes as well as the time side lobes in the ACF.
Figure 3.5 shows the ACF and PACF for when N = 4 code periods are used
within the reference receiver (r = 1, F M tb = 40, T = 0.3, N = K = 100).
Including N in the estimation of the peak side lobe level
w W
1
PSL = 20 log10 dB (3.19)
NM
Using N = 4, PSL = 40 dB down from the peak as shown in Figure 3.5.

Figure 3.6 shows the PAF for the Frank code with N = 4 and demon-
strates that by using more copies of the reference signal within the correlation
receiver, the delay-Doppler side lobe performance improves.
3.5 Reducing the Doppler Side Lobes

To reduce the Doppler side lobes it is necessary to modify the reference signal
with a weighting function w(t) that converts the receiver from a matched
receiver to a mismatched receiver (with a corresponding degradation in SNR
Figure 3.6: PAF for Frank phase modulation for M = 8 (Nc = 64), cpp = 1
with number of reference waveforms N = 4.
and decrease in resolution). Following the development in [5], the reference

signal u (t) in (3.5) is divided into a product of two signals: r(t) which
is periodic with the same period as u(t), and w(t) an aperiodic weighting
function. That is, u (t) = r(t)w(t). The delay-Doppler response of the
mismatched receiver is
e8 e
e e
e
|(, )| = e u(t )r(t)p(t)w(t)e j2t e
dte (3.20)

where p(t) is an aperiodic rectangular window function

F
1 0 t < NT
p(t) = (3.21)
0 elsewhere
Since (3.20) is the Fourier transform of two products (except for the miss-
ing negative sign in the exponential) it can be described by the convolution
(denoted ) of two Fourier transforms as
e8
e
|(, )| = ee u(t )r(t)ej2t dt

8 e
e
p(t)w(t)ej2t dtee (3.22)

With the first transform, since both u(t) and r(t) are infinitely long and
periodic with period T , the Fourier transform of their product (for any )
can be shown to be a series of delta functions at = n/T , n = 0, 1, 2, . . .
or 8 3 p nQ
u(t )r(t)ej2t dt = gn ( ) (3.23)
n=
T
where 8 T
1
gn ( ) = u(t )r(t)ej2nt/T dt (3.24)
T 0
The second integral in (3.22) is the Fourier transform of the product of the
rectangular window and the weight function
8
W () = p(t)w(t)ej2t dt (3.25)

or 8 NT
W () = w(t)ej2t dt (3.26)
0
Finally, the delay-Doppler response of the weighted correlation receiver is
obtained from the convolution between (3.23) and (3.26) yielding [5]
e e
e 3 p n Qee
e
|(, )| = e g ( )W e (3.27)
en= n T e
The significance of this equation is that at any given coordinate (, ), the

delay-Doppler receiver response is determined by contributions from gn ( )
and the weight function. The set of functions gn ( ) is determined by (3.24)
and depend on the transmitted signal modulation that is used.
Three important amplitude weighting windows have been described in [5]
and can be defined by selecting the parameter c in the following expression
w W
1 1c 2t
p(t)w(t) = 1 cos (3.28)
NT c NT
where 0 t N T and zero elsewhere. For uniform, Hann, and Hamming

weight windows, c is selected as c = 1.0, 0.5, and 0.53836 respectively. Using
(3.25) to transform p(t)w(t) yields
w W
sin(N T ) (1 c)(N T )2
W () = 1+ ejN T (3.29)
N T c[1 (N T )2 ]
with the exponent indicating that the weight function is not centered at t = 0.
Still to be determined is the modulation function gn ( ). This is discussed
in detail in the following chapters since it depends on the waveform being
considered. Note that a smooth weight, covering N periods of the signal,

aects only the Doppler behavior. It has no influence on the PACF (the zero
Doppler cut of the PAF). In phase-coded signals, the delay response remains
a triangle with base 2tb , regardless of any amplitude taper along N periods
of the signal.
References
[1] Levanon, N., Radar Principles, John Wiley & Sons, New York, 1988.
[2] Skolnik, M., Introduction to Radar Systems, 3rd Edition, McGraw Hill, Boston,
p. 331, 2001.
[3] Levanon, N., CW alternatives to the coherent pulse train - signals and
processors, IEEE Trans. on Aerospace and Electronic Systems, Vol. 29,
No. 1, pp. 250254, Jan. 1993.
[4] Levanon, N. and Freedman, A. Periodic ambiguity function of CW signals
with perfect periodic autocorrelation, IEEE Trans. on Aerospace and Elec-
tronic Systems, Vol. 28, No. 2, pp. 387395, April 1992.
[5] Getz, B. and Levanon, N., Weight eects on the periodic ambiguity func-
pp. 182193, Jan. 1995.
[6] Freedman, A. and Levanon, N., Properties of the periodic ambiguity func-
pp. 938941, July 1994.
[7] Frank, R. L., Polyphase codes with good nonperiodic correlation properties,
IEEE Trans. IT-9, pp. 4345, 1963.
Problems
1. When the reference signal is of duration N T , the response of the corre-
lation receiver is the PAF for N periods. Another form of the PAF can
be defined by the relationship
8 p
1 NT
Q p Q j2t
N T (, ) = u t+ u t e dt
NT 0 2 2
Starting from here, show the universal result

e e
e sin(N T ) e
|N T (, )| = |T (, )| ee e
N sin(T ) e
Hint: Split the integral into N sections and use the transformation of
variables t = t + (n 1)T .
2. The complex envelope of a signal with periodic phase modulation can

be expressed as
Nc
3
u(t) = un (t (n 1)tb )
n=1
where 0 t Nc tb and
un (t) = ejn
for 0 t < tb . There are periodic two-valued phase sequences that can
also yield a perfect periodic autocorrelation. That is, n = 0 or n =
and un can either be 1 or where
= ej
For Nc = 7,
n = {0 0 0 0 }
and
= cos1 (3/4)
For Nc = 11,
n = {0 0 0 0 0 }
and
= cos1 (5/6)
For both the Nc = 7 and Nc = 11 sequences, (a) generate the complex
CW signal using beta.m. Save the phase shift plot showing the Nc phase
values being used within a code period. (b) For the signal only, plot
the PACF and PAF (delay versus Doppler) when N = 1 and N = 4 to
verify the periodicity (how often the ambiguity function repeats itself)
and the reduction of the PSL. (c) Add the beta.m signal to your LPIT
menu.
3. Plot the weighting function (3.29) for (a) a uniform window, (b) a Hann
window, and (c) a Hamming window, for N = 1 and N = 5 for a code
period T = 0.021s (consistent with a CW LPI waveform with fc = 1
kHz, M = 7, and cpp = 3).
4. Using the LPIT, generate the Frank signal with M = 8, fc = 1 kHz,
fs = 7 kHz, and cpp = 1. Plot the ACF, PACF, and PAF for N = 8,
and compare your results with Figures 3.5 and 3.6.
5. A phase code signal is generated with a carrier of fc = 5 kHz. The
processing gain of the signal is 24 dB and the bandwidth of the signal
is B = 1.25 kHz. Determine (a) the subcode period tb and (b) the code
period T in s. (c) If the signal is intercepted with a receiver that has
an ADC with fs = 25 kHz, how many samples are within a subcode
period (bsc )?
Chapter 4
FMCW Radar
This chapter examines the advantages of the frequency modulation CW LPI
technique, gives detailed expressions for the transmitted signal and the re-
ceived signal, and discusses the isolation required when using a single antenna.
LPI search and track mode processing are discussed, and several FMCW emit-
ter configurations are presented. Also investigated are the eects of frequency
modulation nonlinearities. Moving target indication filtering is discussed, as
well as the FMCW periodic ambiguity function. The experimental PAN-
DORA multifrequency FMCW radar is presented as an example of FMCW
technology. Electronic attack considerations are also addressed. Finally, the
technology trends for FMCW emitters are examined.
4.1 Advantages of FMCW

CW radars that use unmodulated waveforms cannot measure a targets range.
To measure the targets range and/or speed, the transmit frequency must
be varied in time, and the frequency of the return signal from the target
measured. Correlation of the return signal with the transmit signal can give
a measure of both the range and Doppler information of the target. Since the
modulation cannot be continually changed in one direction (e.g., up or down),
a periodic modulation is normally used. Frequency modulation can create a
wideband LPI waveform and take many forms, with sinusoidal and linear
modulation being used most frequently. The most popular linear modulation
utilized is the triangular FMCW emitter, since it can measure the targets
range and range rate.1
1 Sinusoidal frequency modulation is mathematically more tractable than linear fre-
quency modulation and is presented well in [1].
81
FMCW is an eective LPI technique for many reasons. In some appli-

cations, such as radio altimeters, a key advantage is the simple architecture,
which is capable of giving a very high range resolution. Due to the very
low energy transmitted (low radiation hazard), the noncooperative intercept
receivers interception range is significantly reduced. This means that an
FMCW radar may be used in otherwise restrictive emission-control (EM-
CON) conditions that would preclude the operation of pulsed emitters. The
frequency modulation spreads the transmitted energy over a large modulation
bandwidth F , providing good range resolution that is critical for discrim-
inating targets from clutter. The power spectrum of the FMCW signal is
nearly rectangular over the modulation bandwidth, so noncooperative inter-
ception is dicult. Since the transmit waveform is deterministic, the form
of the return signals can be predicted. This gives it the added advantage of
being resistant to interference (such as jamming), since any signal not match-
ing this form can be suppressed. Consequently, it is a dicult matter for
a noncooperative receiver to detect the FMCW waveform and measure the
parameters accurately enough to match the jammer waveform to the radar
waveform (a subject we take up in Part II).
FMCW modulation is also readily compatible with solid-state transmit-
ters, and represents the best use of output power available from these solid
state devices. The return signal is correlated with the transmitted signal, and
is often done using analog techniques. The correlation receiver can also be
implemented using digital techniques. The frequency processing performed
to obtain the range information from the digitized IF signals can be done very
quickly with FFTs. The ease with which the range resolution can be changed,
and the way in which very high range resolutions can be obtained without
requiring wide IF and video bandwidths is also a significant advantage. That
is, the IF and video bandwidths can be matched to the required data rate
rather than to the RF bandwidth required to give the range resolution [2].
Due to the fourth power relationship between a radars return signal power
and the targets range, an adequate amount of sensitivity time control (STC)
must be used in the receiver to selectively attenuate the returns from close-
in targets in order to control the dynamic range and prevent saturation [3].
Due to the frequency-range relationship in the FMCW radar, this technique
may be easily implemented in the frequency domain early on in the signal
processing. FMCW is also easier to implement than phase code modulation,
as long as there is no strict demand on linearity specifications over the mod-
ulation bandwidth. The ability to use weighting to control the range and
Doppler side lobes (mismatched correlation receiver) also allows for ecient
use of the spectrum. Finally, the advanced transceiver design allows FMCW
radars to be connected to, and operated in parallel with, any available pulsed
I-band navigation radar using a common antenna. This means that the very
presence of the LPI radar cannot be ascertained by external observation.
FMCW Radar 83
Figure 4.1: Block diagram of a homodyne triangular FMCW radar.
4.2 Single Antenna LPI Radar for Target

Detection
A block diagram of a homodyne triangular FMCW emitter is shown in Fig-
ure 4.1. In this search mode configuration, both the target range and Doppler
information can be measured unambiguously, while maintaining a low proba-
bility of intercept. The system uses a single antenna. A triangular waveform
generator is used to modulate the CW source for transmission. For low power
single antenna systems, a circulator can be used to allow simultaneous trans-
mission and reception [4]. With higher power systems, the transmitter noise
side bands can hide valid targets and desensitize the receiver. In this case,
separate transmit and receive antennas must be used.
To enable the FMCW emitter to operate more eciently using a single an-
tenna for both transmission and reception, a reflected power canceler (RPC)
is shown [3, 5]. The RPC adaptively cancels the transmit/receive feedthrough
that can limit the dynamic range of single antenna CW radar. In the case of
a linear ramp, a simple RPC can adapt during the sweep to handle a wide
modulation bandwidth, since the instantaneous bandwidth is small.
The target echo is received through the antenna and consists of a delayed
replica of the transmitted waveform. The instantaneous frequency dierence
between the received signal and the transmitted signal is a constant propor-
Figure 4.2: Envelope approximation detection GOCFAR processor.
tional to the round trip delay, so a measurement of this frequency dierence

yields the target range. The frequency dierence is obtained by a homodyne
mixing process, and the frequencies of the received echos (beat frequencies)
are recovered by a spectral analysis of the mixer-lowpass filter output. The
lowpass filter is used to pass only the beat frequencies of interest (maximum
expected beat frequency fb ), and also to reduce the possibility of strong inter-
fering signals reaching the low noise amplifier (LNA), where they can generate
inband spurious signals and distortion that could prevent the detection of the
desired target signal. The LNA amplifies the signal after the mixing/LPF
operation. An analog-to-digital converter samples and quantizes the complex
LNA output, and an FFT computes the frequency spectrum in order to derive
the range profile for each sweep.
The complex FFT output is detected using an envelope approximation
detector
x = a max{|I|, |Q|} + b min{|I|, |Q|} (4.1)
where a and b are simple multiplying coecients (e.g., a = 1, b = 1) [6, 7].
This provides a reasonable approximation to the envelope detector but avoids
the squares and square roots of the envelope detector which impose additional
hardware complexity.
A greatest of constant false alarm rate (GOCFAR) processor shown in
Figure 4.2 is used to detect the targets in the presence of possible clutter edges
within a single modulation period. The envelope approximation detector
output values are strobed into the n reference cells with the test cell located
in the center. Both reference cell neighborhoods have n cells that are used
to determine the noise power levels y1 and y2 on each side of the test cell.
FMCW Radar 85
Figure 4.3: State transition diagram of the Markov chain used for postdetec-
tion integration.
Note that the width of each reference cell or filter is f Hz. The threshold
voltage Vt is obtained by choosing the greatest of y1 and y2 , normalizing by
the number of reference cells n and multiplying by the threshold multiplier
T . Targets are declared in range for both up slope and down slope (beat
frequencies f1b , f2b ), when the amplitude of the test filter is greater than the
threshold voltage. Other CFAR architectures can be used, depending on the
operating environment.
For each modulation period, a single target can result in a number of
GOCFAR range detections, depending on the targets extent and the size
of the range resolution R. Each detection is tagged by its range RT and
its azimuth angle a . To reduce the chance of declaring a false target, post-
detection integration can be used within a single scan. A simple method of
performing postdetection integration for each range detection is through the
use of a discrete time Markov chain [8] with NM states followed by a single
scan angle threshold processor. A state transition diagram of a postdetection
integration Markov chain is shown in Figure 4.3. When the state reaches
NX , start = a , and this marks the beginning of the target position in az-
imuth. For each detection at RT , the state of the chain advances one level
(with probability of detection p). Upon receiving subsequent reports for this
range bin from the GOCFAR, the state either moves up or down. For each
subsequent miss at RT , the state drops one level (with probability q = 1 p).
When the state drops below NY , stop = a and this marks the end of the
targets extent in azimuth.

Each postdetection integration output has both the targets range RT , and
its extent in azimuth = |start stop |. The single scan angle threshold
processor then compares and declares a target at this range and scan
if TA TB . The thresholds TA (lower limit) and TB (upper limit)
depend on the signal-to-noise ratio and are a function of the targets range,
RCS, and any frequency domain STC that is applied. The targets declared on
each scan are normally entered into a track file after going through a gating
process followed by a scan-to-scan correlation.
4.3 Transmitted Waveform Design

There are two main challenges in designing a high dynamic range FMCW
radar for the detection of small targets against a high density clutter back-
ground. The first is generating a frequency sweep that is linear. The second
challenge is controlling the leakage of transmitter phase noise into the receiver.
We begin by examining a triangular FMCW waveform and the Doppler
shifted received signal as shown in Figure 4.4. The triangular modulation
consists of two linear frequency modulation sections with positive and nega-
tive slopes. With a triangular waveform, the range and Doppler frequency of
the detected target can be extracted unambiguously by taking, respectively,
the sum and the dierence of the two beat frequencies. In this section, the
triangular waveform is described, and ways of generating the LPI waveform
are discussed.
4.3.1 Triangular Waveform

The frequency of the transmitted waveform for the first section is [9, 10]
F F
f1 (t) = fc + t (4.2)
2 tm
for 0 < t < tm and zero elsewhere. Here fc is the RF carrier, F is the trans-
mit modulation bandwidth, and tm is the modulation period. The modulation
(sweep) bandwidth F is chosen to provide the required range resolution
c
R = m (4.3)
2F
Note that the larger the bandwidth, the smaller the resolution and the more
LPI the signal becomes. The rate of frequency change or chirp rate F is
F
F = (4.4)
tm
FMCW Radar 87
Figure 4.4: Linear frequency modulated triangular waveform and the Doppler
shifted received signal.
The phase of the transmitted RF signal is

8 t
1 (t) = 2 f1 (x)dx (4.5)
0
Assuming that 0 = 0 at t = 0,
}w W ]
F F 2
1 (t) = 2 fc t+ t (4.6)
2 2tm
for 0 < t < tm . The transmit signal is given by
}w W ]
F F 2
s1 (t) = a0 sin 2 fc t+ t (4.7)
2 2tm
The frequency of the transmitted waveform for the second section is similarly
F F
f2 (t) = fc + t (4.8)
2 tm
Figure 4.5: A block diagram of an FMCW emitter simulation.
for 0 < t < tm . Finally, the transmit baseband signal is given by

}w W ]
F F 2
s2 (t) = a0 sin 2 fc + t t (4.9)
2 2tm
Figure 4.5 shows a block diagram of an FMCW emitter simulation. The

sinusoidal carrier is 9.3 GHz and the modulation bandwidth F = 1 GHz.
Also shown is the triangular waveform with modulation period tm = 0.5
106 s. The power spectral density for the fc = 9.3-GHz FMCW signal with
F = 1.0 GHz is shown in Figure 4.6 with an SNR = 0 dB. The SNR is
defined in terms of the signal and noise power as
A2
SNR = (4.10)
2 2
where A is the amplitude of the signal and 2 is the white Gaussian noise
power. Note the simulation shows that the power spectrum of the linear
FMCW waveform is nearly rectangular over the band fc F/2 < f <
fc + F/2 adding to the LPI properties of the transmitted signal.
FMCW Radar 89
Figure 4.6: The FMCW signal with fc = 9.3 GHz, F = 1.0 GHz, and
SNR = 0 dB.
4.3.2 Waveform Spectrum

Without loss of generality, the instantaneous frequency for the first section
(4.2) in transmitted waveform can be rewritten as
F
f1 = t + fc (4.11)
tm
for |t| tm /2 where the carrier frequency fc lies at the beginning of the sweep
in frequency. The phase of the signal with instantaneous frequency (4.11) can
be calculated as
8 t
F 2
(t) = 2 f1 (t )dt = t + 2fc t (4.12)
0 tm
where the signal has constant amplitude. To compute the spectrum of the
waveform we use the complex form of the transmit signal as [11]
s(t) = ej(t) (4.13)
and obtain the Fourier transform as

8
s(f ) = s(t)ej2f t dt (4.14)

Substituting in (4.13) and letting = F/tm and = (fc f )

8
2
s(f ) = ej(t +2t) dt (4.15)

Using the integral relationships

8 5 F w 2W w 2W k
2
cos(t + 2t)dt = cos C(x) + sin S(x) (4.16)
2
and
8 5 F w 2W w 2W k

sin(t2 + 2t)dt = cos S(x) sin C(x) (4.17)
2
where 5
2
x= (t + ) (4.18)

and C(x) and S(x) are the Fresnel integrals
8 x w 2W
t
C(x) = cos dt (4.19)
0 2
8 x w 2W
t
S(x) = sin dt (4.20)
0 2
the spectrum of a single sweep is then
5
j 2 / x(tm /2)
s(f ) = e [C(x) + jS(x)]|x(t m /2)
(4.21)
2
and the magnitude squared spectrum is
tm
|s(f )|2 = {[C(x2 ) C(x1 )]2 + [S(x2 ) S(x1 )]2 } (4.22)
2F
Using (4.22), the spectrum sidelobe roll-o rate may be approximated in
the side lobe region using the Fresnel integral approximations for large x > 5
as [11] w 2W w 2W
1 1 x 1 x
C(x) + sin 2 3 cos (4.23)
2 x 2 x 2
and w 2W w 2W
1 1 x 1 x
S(x) cos 2 3 sin (4.24)
2 x 2 x 2
This shows that the dominant frequency relationship is an inverse x2 term
since x f through , and thus the roll-o rate
w W
f2
s(f ) = 10 log10 = 20dB per decade (4.25)
(10f )2
This spectral behavior is important when considering the out-of-band emis-
sions that can degrade the LPI nature of the waveform. Various techniques
can be employed for reducing out-of-band emissions of FMCW waveforms.
These include amplitude tapering on a sweep-by-sweep basis (sometimes re-
ferred to as curbing) and using a smooth and finite flyback at the end of each
sweep.
FMCW Radar 91
4.3.3 Generating Linear FM Waveforms

Linear FM waveforms may be generated by either analog or digital methods.
One method consists of using a voltage controlled oscillator (VCO) to produce
an approximately linear FM sweep, with the nonlinearities being compensated
in the control voltage ramp [4]. The problem with this approach is being able
to achieve adequate linearity over a wide bandwidth. The oscillators can
also drift with temperature. Another commonly used technique involves a
dispersive delay line using surface acoustic wave (SAW) technology, but this
has limitations with large time-bandwidth product waveforms tm F .
Another approach is to synthesize the sweep in frequency directly by dig-
ital means [12]. Advantages of the direct digital synthesizer (DDS) method
include: Only the waveform bandwidth (not the time-bandwidth product) is
limited by the technology (and circuit complexity), and digital circuits are
less likely to be susceptible to temperature drift. Modern DDSs are fully inte-
grated, low-cost, single chip solutions that only need an external clock source
for generating the sinusoidal output signals. The DDS benefits from the to-
tally digital generation of the output signal, which allows full control of the
signals frequency and phase, both with very high precision and resolution.
The sequence of waveform samples can be precomputed, stored, and clocked
out of memory. The waveform can be generated at IF or baseband. If gen-
erated at IF, a high clock rate is required (that depends on the bandwidth).
The preferred approach is generating the complex (I and Q) waveform at
baseband and using a single sideband modulator to put the waveform on a
carrier for transmission. The clock rate for this approach is equal to the chirp
bandwidth (rather than twice the chirp bandwidth as in the IF approach).
From Figure 4.4, note that the frequency of the waveform increases linearly
with time. The phase for the upsweep is given by (4.6). To synthesize the
waveform in discrete steps, t is replaced by the sample index i and F = F/tm
becomes the angular frequency increment per sample F [12]. The sampled
frequency for the first section for N samples is then
p Q F 3
N
f1 N, F = fc + F (4.26)
2 i=0
The corresponding phase is then

p Q 3N p Q 3N
F
1 N, F = 2 f1 N, F = 2 fc + F N (4.27)
i=0 i=0
2
or p Q
1 N, F = F N 2 + 1 (0) (4.28)
From these phases, the complex baseband amplitudes can be generated with
two accumulators as shown in Figure 4.7. The first (frequency) accumulator
Figure 4.7: Generation of linear FMCW waveform using two accumulators.
is preloaded with the starting frequency fc F/2. The second (phase)

accumulator is loaded with zero. In each clock cycle, the frequency increment
is added to the frequency accumulator and the new frequency is added to the
phase accumulator [see (4.26) and (4.27)]. The width of the data path in the
accumulators is given by [12]
} ]
tm fclk F
n = log2 (4.29)
F
where F is the bandwidth increment and fclk is the clock frequency.
The phase accumulator output (e.g., 2s complement) is then used to ad-
dress both a sine look-up table (LUT) and a cosine LUT. Only one cycle of
the sine and cosine waveform needs to be stored in the LUT, since the wave-
form repeats every 2. The output LUT resolution depends on the fidelity
of the signal required, and the bandwidth and resolution of the digital-to-
analog converter (DAC) that is available. The lowpass filter (LPF) is needed
to reject the repeated spectra around the clock frequency and multiples of
the clock frequency [13]. If not eliminated, the high frequency components
cause spurious signals out of the single-sideband modulator (SSBM) used for
upconversion on to the RF carrier fc . The SSBM uses the DDS as a reference
for a phase-locked-loop stabilized VCO, where the DDS is driven by a high
speed reference clock.2
2 Software-driven digital upconversions are also an area of development that is promising
for FMCW generation. They can currently provide an intermediate stage upconversion,
FMCW Radar 93
Figure 4.8: Staircase approximation of the phase (t).
The SSBM mixes the video modulating signal fm with the carrier fc and
suppresses either fc + fm or fc fm . How well the sideband is suppressed
can be quantified as a function of |I|, |Q|, and the phase error from perfect
quadrature as
} ]
|I|2 |I|2
s = 10 log10 1 + 2 cos (4.30)
|Q|2 |Q|2
A detailed noise analysis of the approach above was completed by [14].
The analysis investigates a staircase phase approximation of the FMCW
chirp, and examines the spectrum of the transmitted and received signal.
The noise caused by the quantization error is then analyzed. It shows that
the FMCW signal can be digitally generated using a staircase approximation
of its quadratic phase term, without requiring a filter to smooth the signal
on transmission or reception. The noise on the transmitted waveform is not
usually a problem, except that demodulation of the return signal uses this
transmitted waveform as a reference.
A digital approximation of the parabolic phase is shown in Figure 4.8, and
ensures that each step of the phase staircase is a rectangular function. Also
shown is the holding time or subpulse width. The amplitude and phase noise
produce a noise power spectral density on the transmitted signal that depends
on the subpulse width T and the number of bits of the uniform quantizer n.
The requirement for good spectral conditions is [14]
1
T < (4.31)
5Ba
where Ba is the bandwidth after demodulation (on the order of 100 Hz). If
the amplitude and phase noise errors are both uniformly distributed over the
relaxing the requirements on the DAC and LUT resolution.
bandwidth of 1/T , and the spectral contributions of individual subpulses each

contain a constant error, the power spectral density (PSD) is
w W
T 2 T V2
PSD = sinc (4.32)
12 22n 2 Hz
Example 1: Determine the number of bits required in the accumulator

data path if the modulation period tm = 5.0 s, F = 500 MHz, the clock
frequency fclk = 500 MHz, and the bandwidth increment F = 1 MHz.
Using (4.29) J o
n = log2 1.25(106 ) = 20 bits (4.33)
Example 2: Consider a subpulse period T = 8 s and a 12-bit quantizer.
(a) Determine the bandwidth over which the total noise power will be spread.
(b) Determine the noise level due to amplitude and phase noise errors.
For (a), 1/T = 1/8s = 125 kHz. Note that this is much larger than the
low frequency stages of the radar receiver (e.g., 100 Hz). For (b), from (4.32)
above, w W
8 s V2
PSD = 10 log = 134 dB (4.34)
12 224 Hz
below the received carrier, which is quite good.
Recently a DDS was reported using InP double heterojunction bipolar
transistor technology. With a single 12-bit phase accumulator and a read
only memory LUT phase converter, the DDS is capable of synthesizing output
frequencies up to 12 GHz in steps that are 1/4,096 of the 24 GHz clock rate
[15]. The measured spurious free dynamic range (SFDR) is 30.7 dB and the
average SFDR over all frequency control words is 40.4 dB. The significance
of this is that the radar signals can be generated directly in the desired RF
band.
In summary, the DDS is a more complex approach to generating the
FMCW waveform than using a VCO. It has the advantage however, that
it provides a perfectly linear sweep and has greater stability which is espe-
cially important in LPI emitters where sweep-to-sweep Doppler processing is
required.
4.4 Receiver-Transmitter Isolation

One of the greatest problems facing CW radar designers is detecting target
returns on the order of a picowatt or less in the presence of a few watts of
transmitted power. This is due to the problem of achieving sucient isolation
between transmitter and receiver, since the transmission and reception are
simultaneous. The main two problems are:
FMCW Radar 95
Transmitter noise sidebands can hide valid targets.

Power leakage desensitizes the receiver.
Although a dual antenna configuration that is well isolated is a possible
solution, numerous applications require only a single antenna (e.g., a missile
seeker). In this section, a review of transmission line basics is presented,
followed by a discussion of two single antenna systems; one using a circulator,
and one using an RPC.
4.4.1 Transmission Line Basics

In the analysis of transmission line performance [16], the voltage reflection
coecient, of a transmission line-antenna interface is defined as
reflected voltage
= (4.35)
incident voltage
and is generally a complex quantity
= ||ejl (4.36)
where || is the magnitude, and never greater than unity (|| 1). The phase
l is the angle between the incident and reflected voltages at the receiving
end, and is usually called the phase angle of the reflection coecient.
The general solutions of the transmission line equations consist of two
waves traveling in opposite directions with unequal amplitudes. These waves
are called standing waves. The ratio of the maximum voltage of the standing
wave pattern to the minimum voltage is defined as the voltage standing wave
ratio V
|VMax |
V = (4.37)
|VMin |
and is usually found using Smith charts. The standing wave ratio results from
the fact that the two traveling wave components add in phase at some points,
and subtract it at other points. The standing wave ratio V is related to the
reflection coecient by
1 + ||
V = (4.38)
1 ||
and solving for ||,
V 1
|| = (4.39)
V + 1
These results can be used to quantify the various antenna configurations for
FMCW emitters.
Figure 4.9: Computing power at receiver using a single circulator including

losses.
4.4.2 Single Antenna Isolation Using a Circulator

One solution to isolate the single antenna transmit and receive functions is to
use a circulator. To highlight the problems with using a single antenna and
circulator, consider the configuration shown in Figure 4.9. To receive a target
echo signal and derive the correct beat frequencies, a significant amount of
isolation must be present between the transmitted waveform and the received
waveform. The transmitter sends an average power Pt to the circulator. A
certain fraction of the incident power Ic is leaked at the circulator output,
due to the finite amount of isolation. Circulators provide the best isolation
when they are terminated correctly (impedances matched). The isolation
between any two ports is the return loss due to third port mismatch. Including
transmission line loss (LRT = LRR = Lx 1) and circulator loss (Lc 1),
the average power into the antenna is Pt /Lx Lc . From above, the amount
transmitted out of the antenna is Pt (1 ||2 )/Lx Lc and the amount reflected
back to the receiver is Pt ||2 /Lx Lc . The power received from the target is Pr
and that portion of received power entering the receiver is Pr (1 ||2 )/Lx Lc .
In summary, the total signal appearing at the receiver is the addition of the
target return, the leakage, and the antenna mismatch or
(1 ||2 ) ||2 Pt
Ptot = Pr + Ic Pt + (4.40)
Lx Lc (Lx Lc )2
Example 3: Calculate the total power at the receiver if the standing wave
ratio V =2:1, the transmission line loss Lx = 0.5 dB, the circulator isolation
is Ic = 60 dB, and the circulator loss Lc =1 dB. The CW transmitter pro-
vides Pt =10 dBW at fc = 9.375 GHz. The antenna has a transmit, receive
gain Gt , Gr =30 dB. The target is located at a range or 28,000m and has a
FMCW Radar 97
RCS T = 50m2 .
The first step is to calculate the magnitude of the reflection coecient. From
(4.39), =0.333. Next, the power transmitted out of the antenna is calculated
as
Pt (1 ||2 )
PCW = (4.41)
Lx Lc
or PCW =6.3W. Using (1.23) with PCW = 6.3W and recognizing that (4.40)
takes into account the transmission line losses, the return power from the
target is Pr = PRT = 2.6(1016 ) (or 156 dBW). Using this value in (4.40),
Ptot = 0.56W or -2.5 dBW. This example shows that the amount of power
from the target that reaches the receiver is minimal, and highlights one of the
main problems with using a single circulator FMCW emitter.
4.4.3 Single Antenna Isolation Using a Reflected Power

Canceler
The reflective power canceler was discussed briefly in the first section and
shown schematically in Figure 4.1. It was developed in the early 1960s as
a coherent device that could be used to cancel the transmitter feedthrough
in an FMCW emitter [17, 18]. Many of the recent improvements have been
made possible by the availability of new microwave and digital components.
The RPC is shown in Figure 4.10. The RPC takes a sample of the signal
being transmitted and vector modulates it, so that it is of equal amplitude
and opposite in phase to the transmitter leakage signal. By adding this signal
into the receiver, using a directional coupler, the leakage and noise sidebands
of the transmitted signal can be canceled out [3, 5]. The eectiveness of the
RPC depends on how accurately the amplitude and phase can be adjusted.
To perform adequately, the RPC must operate in a closed-loop fashion, with
sucient gain and bandwidth to track the leakage variations.
The principle of leakage cancellation is to generate a signal with equal-
amplitude and opposite-phase to the original leakage. This signal summing
up with the original leakage signal realizes the cancellation. A good can-
cellation requires accurate match of amplitudes and phases. If the signal to
be cancelled or leakage signal is the complex signal A, and the cancellation
signal or feedthrough signal under vector modulator control is B, assuming a
certain phase dierence and amplitude dierence A between them, the
cancellation signal is
B = (A + A)ej (4.42)
The cancellation ratio or cancellation depth is given by [19]
e e
eB Ae
Rcancel,dB = 20 log10 ee e (4.43)
A e
Figure 4.10: Schematic diagram of a reflective power canceler.
or
D i
Rcancel,dB = 10 log10 1 2(1 + A/A) cos + (1 + A/A)2 (4.44)
The cancellation depth is very sensitive to the phase error and the amplitude
error. For example, to achieve a 30 dB cancellation depth, an amplitude
dierence of less than 0.25 dB (3%) and a phase dierence of less than 1
is required. An accurate phase match within 1o of error is very dicult to
realize using wideband analog microwave and millimeter wave circuits which
can vary with temperature and environmental changes.
One recent RPC that uses PIN (p into n) diodes and is also used by the
Pilot radar, is shown in Figure 4.10. The amplitude and phase of the leakage
power are estimated by measuring the dc levels of the I and Q outputs of the
receiver mixer. These I and Q signals are then used to control the amplitude
and phase of the leakage signal, forming a closed loop controller. The RPC
is quite robust to phase errors, on the order of 45 degrees around the loop.
Consequently, the vector modulation and quadrature mixer requirements can
be relaxed. The Pilot radar uses modern microwave components to improve
the transmit/receive isolation from about 20 dB (without RPC) to over 50 dB,
and is comparable to the isolation achieved by a dual antenna configuration.
The block diagram of a FMCW radar system that uses DDS technology
coupled with an RPC and a single antenna for shipboard surveillance is shown
in Figure 4.11 [20]. The DDS uses a clock frequency of 300 MHz integrated
with a phase-locked loop at L band which is upconverted to X band using a
FMCW Radar 99
mixer. The radar uses a solid state transmitter module for power management
(maximum transmit power of 1 W) and the modulation bandwidth can be
varied from 50 MHz to 200 MHz. The RPC provides 30 dB of cancellation
and has a vector modulator fabricated with PIN diodes, a commerical I-Q
demodulator, and a low frequency control circuit (to filter and amplify the
signals from the demodulator and apply them to the vector modulator to
generate the cancellation signal).
Figure 4.11: Block diagram of an FMCW radar using DDS technology, an

RPC, and a single antenna for shipboard surveillance (adapted from [20]).
In [19], a heterodyne scheme based on real-time digital signal processing

(DSP) is presented for leakage cancellation. In this approach, heterodyne
processing is used to generate an error signal modulated at a pre-selected
reference frequency. In this manner, the DC oset of the mixer can be sep-
arated from the modulated error signal using a band pass filter. Since the
modulated error signal contains the amplitude and phase information of the
leakage signal, the generation of the controlling error vector is carried out in
DSP by comparing the reference signal and the modulated error signal. Then
the error vector is used to adjust the vector modulator. Over 30 dB cancel-
lation of the leakage was achieved over a modulation bandwidth F = 1.7
GHz and modulation period tm = 1.4 ms [19]. A quadrature FMCW radar
topology using a leakage cancellation circuit at 24 GHz is presented in [21].
The canceller is composed of four branch-line hybrid couplers, a 90o delay
line and a Wilkinson combiner. For this architecture, a 35 dB cancellation
was achieved.
4.5 The Received Signal

The received signal from a stationary target is the transmit signal delayed in
time by the round-trip propagation time (or transit time) to the target and
back (td ), with reduced amplitude b0 [3]
b0
sr (t) = s1 (t td ) (4.45)
a0
or }w W ]
F F 2
s1r (t) = b0 sin 2 fc (t td ) + (t td ) (4.46)
2 2tm
For the homodyne FMCW emitter, the receive signal is mixed with the trans-
mit signal. The beat frequencies are derived as the dierence between the
transmitted and received signals. The beat frequency is sometimes referred
to as an intermediate frequency, although the information is not modulated
onto a conventional carrier [9]. The mixer output beat frequency signal is
}w W ]
F F 2 F
s1b (t) = c0 cos 2 fc td t + td t (4.47)
2 2tm d tm
For the second segment of the triangular waveform, the mixer output beat
frequency signal is
}w W ]
F F 2 F
s2b (t) = c0 cos 2 fc + td + td td t (4.48)
2 2tm tm
Equations (4.47) and (4.48) contain a frequency term that is time varying
and phase terms that are not. The beat frequency is the third term in (4.47)
and (4.48)
F 2RF 2R
fb = td = = F (4.49)
tm ctm c
where the delay time td = 2R/c for a stationary target at a range of R. If the
target is moving with velocity V , the beat frequency for the first segment is
2RF 2V 2R 2V
f1b = = F (4.50)
ctm c
and the beat frequency for the second segment is
2RF 2V 2R 2V
f2b = + = F + (4.51)
ctm c
where the second term is due to the targets Doppler frequency. For multiple
targets, multiple beat frequencies would be present and would depend on each
targets range and velocity.
FMCW Radar 101
Figure 4.12: ASCM LPI emitter-ship scenario.
In summary, the advantages of the linear FMCW waveform include:

(a) the presence of LPI operation with an ecient utilization of the spec-
trum, (b) the waveform is easier to implement than phase-coded modulation
if there are no strict demands on linearity over a wide bandwidth, and (c)
the received signal can be processed using one multiplication with a matched
reference signal, with the range being resolved using spectral analysis.
4.6 LPI Search Mode Processing

To illustrate LPI search mode processing, consider an antiship cruise missile
with a seeker (fc = 9.375 GHz) flying inbound to a target ship with RCS =
T . The incremental backscattering coecient of the sea surface is o . The
LPI seeker comes on at a range of R = 28,000m (15 nmi) from the ship.
Figure 4.12 shows the missile-ship scenario being investigated. The emitter
is flown at 300 m/s (Mach 1) at an altitude of h = 70m toward the ship for
a period of 91s. Note that this scenario assumes initially that the ship is not
moving. The scenario is analogous to the LPI radar being stationary and
the target approaching at 300m/s. In the search mode, the emitter uses, for
example, the conventional scan shown in Figure 1.4(a) and makes a single
scan every 3s with a scan rate of 70 deg/s. At end game (700m from target),
the missile dives to the ship for impact.
The first step in designing the LPI seeker is to determine the modulation
bandwidth F (peak-to-peak frequency deviation) in order to give the re-
quired range resolution. For example, the ASCM might require a R = 0.3m
range resolution, in order to calculate the ship orientation and select a wa-
terline aimpoint with good accuracy. With this range resolution, the ship
returns could easily extend over a large number of range bins, depending on
the aspect angle. Recall that the ideal range resolution is
c
R = m (4.52)
2F
For R = 0.3m, a F = 500 MHz is chosen. To increase the signal to
clutter ratio, a R equivalent to the size of the ship may be selected so that
the entire ships return lies within a single range bin. This approach would
require a smaller F .
The modulation period is chosen next, and two factors must be considered.
The first consideration is that
R
tm < (4.53)
Vt
where Vt is the maximum closing velocity of the target. This relationship
requires that the target must remain in a range bin for at least an entire
modulation period tm . Otherwise, the target return will smear across several
range bins. The second consideration is that tm should be several times
the maximum round-trip delay td , of the targets return signal. This is in
order to minimize the loss in eective transmit bandwidth and power and
to also provide a high velocity resolution [9]. Since an acquisition range of
R = 28, 000m corresponds to a maximum round-trip delay td = 186.7 s, a
modulation period of tm = 1 ms is chosen ( 5.5td ).
The resulting coherent processing interval is
t0 = tm td (4.54)
The spectral width of the beat frequency is the inverse of the coherent process-
ing interval or
1 1 1
w = = (4.55)
t0 tm td tm
and is the Doppler shift that causes a range error of exactly one range bin.
The eect that a Doppler shift can change the apparent range of the target is
the well-known FMCW range-Doppler cross-coupling eect. That is, the un-
ambiguous Doppler frequency is fu = 1/2t0 Hz. The corresponding velocity
resolution or first blind speed is
w
v = m/s (4.56)
2
The first blind speed is the speed at which the Doppler goes through one
complete cycle from one sweep to the next (beat frequency increases by
one cycle per sweep) [3]. For our example, with a target acquisition range
FMCW Radar 103
R = 28 km and a td = 186.7 s, the coherent processing interval t0 = 0.81

ms and the spectral width w = 1.23 kHz. The resulting first blind speed is
v = 19.7m/s.
For the first section of the triangular waveform, the partial overlap results
in a reduced processed bandwidth. The eective bandwidth is
w W
td
F = F 1 Hz (4.57)
tm
with an eective time bandwidth product of t0 F . The range resolution is
also slightly degraded as
c c
R = = m (4.58)
2F 2F (1 td /tm )
Continuing the example above, with td = 186.7 s, F = 406.7 MHz, R =
0.37m, and the eective time-bandwidth product t0 F = 330.75 103 . The
large time-bandwidth product contributes to the LPI nature of the radar.
The resulting beat frequencies are of the form
2RF 2V
f1b = (4.59)
ct0
and
2RF 2V
f2b = + (4.60)
ct0
Using the numbers from the example above with Vmax = 300m/s, the corre-
sponding maximum beat frequency is fbmax = 93.35 MHz.
The analog-to-digital converter section of the LPI radar receiver must
sample at least twice the highest beat frequency or
w W
2RF 2V
fs = 2 + samples/s (4.61)
ct0
resulting in the number of samples within a coherent processing interval of
N F = fs t0 . With fbmax = 93.35 MHz, fs = 186.7 MS/s.
To resolve the multiple echoes from the clutter and targets, an FFT
process is used for frequency analysis. The signal after frequency analysis
is a coherent video signal. Since the FFT requires integer powers of 2, the
FFT size is chosen to be N = 2x N F . To capture this many samples, the
ADC sampling frequency must be
N
fs = samples/s (4.62)
t0
with the sampling frequency and an FFT size of N , the resulting filter width
is
f
f = s (4.63)
N
Table 4.1: Eight LPI Emitter Designs for Comparison

Examples 1 2 3 4 5 6 7 8
F (MHz) 15 20 25 30 35 50 500 1000
R (m) 10.0 7.5 6.0 5.0 4.3 3.0 0.3 0.15
F I (MHz) 12.2 16.2 20.3 24.4 28.5 40.7 406.7 813.3
RI (m) 12.3 9.2 7.4 6.1 5.3 3.7 0.37 0.18
fbmax (MHz) 2.8 3.8 5.0 5.6 6.6 9.4 93.3 186.7
FFT size 8,192 8,192 8,192 16,384 16,384 16,384 262,144 524,288
t0 F I 9,923 13,230 16,537 19,845 23,153 33,075 330,755 661,511
and agrees with (4.55). For the example, N F = fs t0 = 151,853 samples

and N = 218 = 262,144 so fs = 322 MS/s giving a filter width of f =
1.23 kHz or 19.7m/s. Since the signal processed by the FFT is complex,
the unambiguous range is Ru = N R =96,993m. Since this exceeds the
required detection range of the target, the number of range cells processed
by the FFT can be limited by filtering the input to the FFT processor. This
reduces the input bandwidth, as well as the complexity of the digital process-
ing. Weighting can also reduce the frequency and range side lobes, but will in-
crease the overall frequency and range resolution [10]. For a Hamming weight-
ing we include the loss by multiplying by a factor of 1.8 or R = 0.67m.
Table 4.1 shows, for comparison, eight emitter design examples. For all ex-
amples shown, fc = 9.375 GHz, t0 = 0.81 ms, and f = 1.23 kHz.
After the GOCFAR detects the beat frequencies f1b , f2b from both transit
waveform sections, the targets range is calculated as
ctm
R= (f1b + f2b ) m (4.64)
4F
and the targets range rate is calculated as

R = (f f1b ) (4.65)
4 2b
The Markov chain counts up the range hits during a scan, and sends the start
and stop angles of the detected target to a single-scan detection process that
thresholds the dierence between the start and stop angle (see Section 4.2).
After target declaration, the emitter fixes the antenna position at the angle
of the target, and the track mode processing takes over.
4.7 Track Mode Processing Techniques

The LPI emitter starts at a range of R = 28,000m (15 nmi) from the ship.
Figure 4.12 shows the emitter-ship scenario being examined. For the track
mode processing, consider the return from a target showing up in a particular
filter fb with bandwidth f . This frequency will correspond to a target at a
FMCW Radar 105
range R where
cfb t0
R= (4.66)
2F
Consequently if the target range is varying, it may be tracked by adjusting the
sweep bandwidth F in order to keep the beat frequency equal to fb . This
relaxes the LPF requirements at the receiver frontend. Also recall that the
range resolution R = c/2F . The ratio of these gives the range resolution
expressed as a fraction of the range
R 1
= (4.67)
R fb t0
and is a constant. That is, the emitter will measure the range to the target
with a resolution that is proportional to the range. As the range-to-target
gets smaller, the bandwidth F gets larger. For example, if the target at
R = 28,000 is acquired on the first transmit waveform, the target shows up
at filter number 75,675 or fb = 75, 675 1.23 kHz = 93.1 MHz. In order to
keep the target at this beat frequency, the sweep bandwidth is calculated as
cfb t0 1.13 1013

F = = Hz (4.68)
2R R
For an endgame range-to-target of R = 700m, F = 16.2 GHz. A
block diagram of the track processing is shown in Figure 4.13. Note that
the major advantage of this technique is that the (narrow) bandpass filter
is now centered on fb . Since the range to the moving target is changing
with time, a range tracker (in Doppler space) is required that is constantly
adjusted to keep the target locked in range. After the range is computed, the
required sweep bandwidth is recalculated and sent to the triangular waveform
generator.
Another approach to the track processing is to keep the sweep bandwidth
constant, and to allow the targets beat frequency to change. The targets
position can be followed in signal processing by monitoring the position of
the FFT peak detector output. The advantage of this approach is that the
receiver LPF used in the search mode can also be used for the track processing,
at the expense of integrating a larger noise component.
4.8 Eect of Sweep Nonlinearities

Frequency sweep nonlinearities (frequency instability in the transmitter) act
to broaden the spectral width of the targets beat frequency. The requirement
on frequency stability in the transmitter is investigated in [10, 22], and tech-
niques to correct for these nonlinearities are examined in [23]. To quantify
these eects, a nonlinear term can be added to (4.8) as [10]
Figure 4.13: Block diagram of the track processing, for the triangular homo-
dyne FMCW emitter.
F F
f2n (t) = fc + t + An sin 2fn t (4.69)
2 tm
where 0 < t < tm and An is the amplitude of the sinusoidal nonlinearity, and
fn is the frequency of the sinusoidal nonlinearity. The transmitted signal is
}w W ]
F F 2 An
s2n (t) = a0 sin 2 fc + t t + (1 cos 2fn t) (4.70)
2 2tm 2fn
The corresponding mixer output beat frequency signal can be shown to have
the form
}w W
F F 2 F
s2bn (t) = c0 cos 2 fc + td + t td t
2 2tm d tm
]
An
+ [cos 2fn (t td ) cos 2fn t]
2fn
for 0 < t < tm . From this expression, the eects of the nonlinearity can be
evaluated and their significance evaluated.
FMCW Radar 107
Figure 4.14: Performance of a three-sweep canceler on S-band FMCW radar

(i-canceled, ii-uncanceled) [3] ( c IEE 1992).
4.9 Moving Target Indication Filtering

A moving target indication (MTI) filter can also be added to the FMCW
signal processing chain, and operates in a manner similar to a pulsed radar
[3]. Figure 4.14 shows the operation of an MTI canceler, an experimental S-
band FMCW radar built at Philips Research Laboratories. The upper trace
shows the video A-scope picture from one sweep of the radar. The lower trace
shows the signal at the output of a digital three-sweep MTI canceler with more
than 40 dB of cancellation. Moving target Doppler (MTD) processing can be
implemented by measuring the rate of change of phase of the output of each
FFT range bin, from one sweep to the next, as shown in Figure 4.15. This
capability can be added quite easily to existing FMCW radars.
Figure 4.15: MTD processing applied to FMCW radar with output a set of
range-Doppler cells [3] ( c IEE 1992).
4.10 Matched Receiver Response

For the FMCW emitter design shown in Figure 4.1, the weighting is uni-
form. That is, no weighting is applied to the homodyne detector during the
correlation process. Recall that the unweighted PAF describes the range-
Doppler response of a matched receiver, when the receivers reference signal
is constructed from an integral number of periods N of the transmitted signal
(reference signal duration N tm ). Figure 4.16(a) shows the unweighted ACF
for an FMCW waveform with tm =20 ms and F =500 Hz, where the re-
ceiver uses N = 1 reference signals for correlation. Note the peak side lobe
level PSL 15 dB. The PACF is shown in Figure 4.16(b) and indicates
that the FMCW does not have a perfect PACF. The PAF is shown in Figure
4.17. The plot shows the two peaks left out from the two diagonal ridges,
bifurcating from the main lobe at a level of 1/2. The high side lobes are un-
wanted, since additional targets could possibly hide at these positions. Note
that the main and side lobes do not have deep nulls, a situation that can be
rectified with additional copies of the reference signal used in the receiver.
Figure 4.18(a, b) shows the ACF and PACF, respectively, for the triangu-
lar FMCW with N = 4. Note the well-defined main lobe repeating at every
2tm fs = 280 samples, resulting from the use of N = 4 copies of the refer-
ence signal in the correlation receiver. Figure 4.19 shows the PAF for N = 4
and demonstrates the more pronounced Doppler lobes appearing at kN for
FMCW Radar 109
Figure 4.16: Triangular FMCW (a) ACF and (b) PACF with F = 500 Hz,
tm = 20 ms, and N = 1.
Figure 4.17: Triangular FMCW PAF with F = 500 Hz, tm = 20 ms, and
N = 1.
Figure 4.18: Triangular FMCW (a) ACF and (b) PACF with F = 500 Hz,
tm = 20 ms, and N = 4.
k {0, 1, . . .}. The side lobes are reduced significantly.
4.11 Mismatched Receiver Response

Recall from Chapter 3 the weighted PAF
e w We
e 3 n ee
e
|(, )| = e g ( )W e (4.71)
en= n tm e
where 8 T
1
gn ( ) = u(t )r(t)e2nt/T dt (4.72)
T 0
describes the performance of a mismatched correlation receiver. That is,
for any (, ) the receiver response is determined by contributions from all
the gn functions. To study close up the mismatched response, we follow
the development by Levanon and present the function gn ( ) for a sawtooth
FMCW signal [24]. In the receiver, the envelope of the reference signal (before
adding weights) is the complex conjugate of the envelope of the transmitted
signal (r(t) = u (t)). The complex envelope of the transmitted signal is
3
u(t) = uT (t ntm ) (4.73)
n
FMCW Radar 111
Figure 4.19: Triangular FMCW PAF with N = 4.
for n = 0, 1, 2, , and
w W2
tm
jF t
2
uT (t) = exp (4.74)
tm
where 0 t tm and zero elsewhere. Using (4.74) and (4.73) and the
reference signal r(t) = u (t) in (4.72)
} w W ]
sin sin(n ) jn /tm
gn ( ) = + (1)n 1 e (4.75)
tm tm n
where

= [F (tm ) + n] (4.76)
tm
and is shown in Figure 4.20(a) for F = 500 Hz and tm = 20 ms (tm F =
10) to demonstrate the extensive range side lobes that appear in the delay-
Doppler response for the FMCW signal.
The weight function in (4.71) suppresses the Doppler side lobes. To reduce
the range (or time) side lobes, frequency weighting is often used (similar to
the STC processing). Since the frequency deviation is linearly swept within
Figure 4.20: gn ( ) for FMCW waveform with tm F = 10 (a) unweighted

and (b) weighted (c = 0.53836 Hamming).
FMCW Radar 113
one modulation period tm , frequency weighting can be implemented by a

repetitive time weight function with a period of tm
3
r(t) = utm (t ntm )wtm (t ntm ) (4.77)
n
where n = 0, 1, 2, . . . . Using the same weighting function discussed pre-

viously, the new function to reduce the time (range) side lobes can be shown
to be
1c
gnw ( ) = gn ( ) [gn+1 ( ) + gn1 ( )] (4.78)
2c
The result for the tm F = 10 waveform is shown in Figure 4.20(b) and
demonstrates the reduction in the range (time) side lobes using c = 0.53836
(Hamming window).
4.12 PANDORA FMCW Radar

The parallel array for numerous dierent operational research activities
(PANDORA) is an experimental LPI radar that is designed to generate eight
separate (but simultaneous) narrowband FMCW signals at X-band, that are
additively mixed and radiated. A block diagram of a four-channel PANDORA
radar is shown in Figure 4.21 [25]. The multichannel multifrequency emitter
consists of an FMCW waveform generator and a power combiner block. The
receiver contains a wideband LNA, a power resolver block, stretch process-
ing for each FMCW channel, a noncoherent processor, and a high resolution
FFT.
The radar operates on two well-isolated antennas; one for transmission
and one for reception. In order to eliminate near field clutter, pencil beams
are used for transmission and reception. The range is unambiguous, and the
ambiguity in Doppler is controlled by ensuring that the change in Doppler
across the modulation bandwidth for a particular target is less than the spec-
tral width f (confined to a single range bin).
The major contribution of the PANDORA LPI radar is the ultrawideband
processing capability without the need for an ultrawideband instantaneous
bandwidth. The center frequency of each channel diers by the modulation
bandwidth (f1 = 48 MHz) plus a guard channel (fg = 56 MHz). The guard
channel helps to ensure channel isolation, as well as enabling a higher res-
olution to be obtained than the individual sweeps by themselves. A total
modulation bandwidth of 776 MHz is achieved from eight channels (from
9.378 GHz to 10.154 GHz) resulting in a range resolution of R = 0.19m
instead of R = 3.1m characteristic of a F = 48 MHz. The narrowband
FMCW signals are generated in each channel as shown in Figure 4.22 [26]. A
single channel has been demonstrated covering the entire bandwidth.
Figure 4.21: Block diagram of the PANDORA radar [25] ( c IEEE 2000).
Figure 4.22: PANDORA narrowband FMCW channel configuration [26]

( c IEEE 2000).
FMCW Radar 115
4.13 Electronic Attack Considerations

FMCW radars are hard to detect due to their wideband waveforms and con-
sequently, potential jammers have a significant problem measuring the wave-
form parameters with sucient accuracy in order to match the jamming wave-
form to the radar waveform. In a realistic environment with a large number
of other radar systems operating in the same frequency band, an FMCW
radar is significantly more dicult to detect. These types of jammers and
their requirements are discussed in Part II. Also, since the FMCW transmit
waveform is deterministic, a good deal of robustness against electronic attack
is inherent. This stems from the fact that with this deterministic transmitted
signal, the return target signature has a general form that may be predicated.
This leads to a significant suppression of many interfering waveforms that are
uncorrelated, such as narrowband interference and pulsed radar emissions.
Of course, if the modulation period tm and bandwidth F can be deter-
mined, then coherent deception jamming is feasible and very eective, since
the jammer waveform looks like the radar waveform.
Antijam aspects of linear FM waveforms using simulations have also been
performed [27]. White Gaussian noise, continuous wideband jamming, and
jamming signals that were identical to the transmitted chirp signal were eval-
uated. They conclude that the FMCW signal can be recovered in moderate
noise conditions, but the radar has a hard time distinguishing a genuine chirp
signal from a hostile jammer signal when the jammer produces signals that
have a similar frequency spectrum to the chirp signal.
4.14 Technology Trends for FMCW Emitters

The FMCW limitations discussed above are quickly being overcome, with
such devices as solid-state transmitters and high-speed DSPs. This section
discusses some of the recent advances and their impact on the FMCW radar
performance.
As shown in Figure 4.1, the same antenna is used for both transmission
and reception, and the signals are separated with a circulator connected to the
antenna. In the FMCW emitter, transmission and reception are simultaneous,
and it is necessary to detect target returns on the order of a picowatt or
less in the presence of watts of transmitted power. The transmitter noise
can swamp out the valid targets, and the power leakage can desensitize the
receiver. Although two antennas (one transmit, one receive) can solve the
problem, many FMCW emitters (e.g., LPI missile seekers) must use a single
antenna. Consequently, the leakage must be canceled before it desensitizes
the receiver performance.
Conventionally, the vector modulator RPC adjusts a sample of the signal
being transmitted so that it is of equal amplitude to, and directly out of
Figure 4.23: Block diagram of C-band FMCW radar MMIC with electronic
circulator and photograph of GaAs chip (1.08 2.15 0.25 mm) [28] ( c IEEE
1989).
phase to the transmitter leakage signal [3, 4]. By adding the signal into the
receiver via a directional coupler, the leakage, including the noise sidebands
of the transmitted signal, can be canceled out. This reflected power-canceling
circuit must operate closed loop, with sucient gain and bandwidth to track
the leakage variation.
The demand for the low cost and small size makes modern front-end so-
lutions based on microwave monolithic integrated circuits (MMICs) the most
favored approach. In order to allow single antenna operation, the first low-
power C-band FMCW radar MMIC chip that incorporated an electronic cir-
culator in a single gallium arsenide (GaAs) integrated circuit was reported
in [28] and shown in Figure 4.23. The chip also included the VCO, buer
amplifier, and the mixer. In this circuit, the VCO drives a two-stage ampli-
fier to form the FM transmitter. The receiver consists of an active field eect
transistor (FET) mixer with a bandpass input filter. The electronic circulator
circuit provides the interface to transmitter and receiver. The reference or
local oscillator signal for the mixer is provided by the reflected power from
the antenna mismatch. The signal reflected from the antenna mismatch is
suciently greater than the inherent circulator leakage to capture the mixer
and serve as the local oscillator signal [28].
The circulator can also be a passive ferrite device, but this typically has
to be placed outside the MMIC circuitry. The use of a power divider has
also been suggested, but wastes one half of both the transmitted and received
power. The FET transceiver is one promising approach to separate two sig-
nals that are closely spaced in frequency [29]. A circuit diagram of the FET
transceiver is shown in Figure 4.24, and overcomes the disadvantages asso-
ciated with diode circuits as well as being well suited to MMIC technology.
The circuit eliminates the need for dual antennas, a circulator, or a coupler
for the separation of the transmit and receive signal paths. The FET is used
simultaneously as an amplifier for the transmitted signal, and as a resistive
mixer to downconvert the received signal. At optimum bias point, the cir-
FMCW Radar 117
Figure 4.24: Circuit diagram of the X-band FET transceiver [29] ( c IEEE
2000).
cuit has an output power of 7 dBm and a conversion loss of 9 dB. Although
the output power in this technology is limited, by 2011 the output power
capabilities are expected to be suitable for missile seeker applications.
An indium phosphide (InP)-based MMIC for use in millimeter wave
FMCW emitters was reported in [30] for a two-antenna system. For the tran-
sistor of this MMIC, an indium aluminum arsenide/indium gallium arsenide
(InAlAs/InGaAs) on InP pseudomorphic high electron mobility transistor
(HEMT) was used with a 0.5m gate length. Because of the high electron
mobility and the high sheet charge density, the HEMT performed with 8
dBm output power gain in the millimeter wave frequency range. The mil-
limeter wave circuitry consisted of a 30-GHz voltage-controlled oscillator, a
30/60-GHz frequency doubler, a 60-GHz amplifier, and a 60-GHz single-
balanced mixer. Other GaAs HEMT technology for W-band FMCW receivers
have been recently reported in [3133].
A 25-GHz nonlinear, single antenna FMCW front-end, that uses a high-
precision 2.45-GHz surface acoustic wave (SAW) reference and adaptively
compensates for phase errors (linearizing the target signal) by software, is
reported in [34]. The compensation of phase errors is based on measuring
the target signal against an exactly known distance standard. This standard
is implemented using the SAW delay line. By moving the VCO and the ref-
erence delay line to a 2.45-GHz IF, a complete planar design of the 24-GHz
front-end is realized as shown in Figure 4.25. A control voltage m(t) sweeps
the frequency of the 2.45-GHz VCO monotonically over the sweep band-
width. The VCO feeds the SAW delay line and a mixer yielding the reference
signal sr(t). The other part is upconverted with a 21.7-GHz LO signal, band-
pass filter, amplified and fed through a directional coupler to the antenna.
The transmit/receiver hybrid diverts the delayed echo signal from the an-
Figure 4.25: Block diagram of a 24-GHz FMCW sensor with 2.45-GHz SAW
reference [34] ( c IEEE 1997).
tenna to the downconverter that is pumped by the LO. The resulting delayed
2.45-GHz IF signal is then mixed with the IF transmit signal, providing the
sensor signal sm (t) for further digital signal processing. A 77-GHz version of
the architecture incorporating a flip-chip MMIC VCO was also reported in
[35].
The precision of FMCW emitters depends largely on the linearity of the
frequency ramp generator. Many frequency synthesizer concepts have been
explored. One method is to directly linearize the VCO. The linearity, however,
that is achievable with a direct linearization circuit at the VCO input is poor
[36]. The linearity can also be improved by controlling it within a phase-
locked loop (PLL) that uses a stable crystal-controlled oscillator. This can
result in linearity better than 10-4. The conventional PLL linearization circuit
consists of a programmable frequency divider with a unity division ratio N .
The analog ramp frequency results from the moving average of the reference
frequency fref multiplied by N . To increase the number of steps the divider
executes on the ramp, a fractional divider circuit can be used to make any
desired step size. This way, many more division ratio steps can be executed
on the ramp [36]. A block diagram of the FMCW system with a fractional
ramp generator is shown in Figure 4.26, and includes a modified setup for
measuring the ramp quality. Digital techniques to generate the sweep signals
have also gained much attention. Use of a direct digital synthesizer, for
example, has many advantages over analog methods, including good flexibility
in changing the sweep bandwidth and sweep rate. Also, these techniques are
not as susceptible to environmental factors.
High-temperature superconductor (HTS)-based systems have made the
FMCW Radar 119
Figure 4.26: Block diagram of an FMCW system with fractional ramp gen-
erator [36] ( c IEEE 1999).
transition from the laboratory to the field. The use of HTS systems has re-
cently gained significant attention, since they can solve a number of recurring
problems in digital receiver designs, such as noise figure, bandwidth, gain,
loss, size, and dynamic range. For example, the insertion loss in the preLNA
filter shown in Figure 4.1 can degrade the system noise figure and dynamic
range. Superconductor tunable filters are very attractive due to the low loss
that is achievable. A compact superconducting-ferrite filter operating at 77K
with insertion loss of 1 dB is reported in [37]. An HTS filter with noise fig-
ure on the order of 0.3 dB is reported in [38]. Note that these filters have
a flat passband, sharp filter edge skirts, and superior out-of-band rejection
characteristics. The filter and the LNA are often contained together in a
cryo-cooled Dewar. The use of HTS for an FMCW radar is discussed in [39].
Here, a self-contained FMCW radar incorporating a compact 2.2-ns broad-
band superconducting delay line operating at 80K is described and is the first
demonstration of an integral microwave system utilizing HTS circuitry that
incorporates a closed-cycle cooler, and a long-life permanently sealed Dewar.
References
[1] Mahafza, B. R., Radar Systems Analysis and Design Using MATLAB, Chap-
man & Hall/CRC, Boca Raton, Jan. 2000.
[2] Stove, A. G., Modern FMCW radar - techniques and applications, European
Radar Conference, Amsterdam, pp. 149152, 2004.
[3] Stove, A. G., Linear FMCW radar techniques, IEE Proc. F, Vol. 139, No.
5, pp. 343350, Oct. 1992.
[4] Griths, H. D., New ideas in FM radar, IEE Electronics and Communica-
tions Engineering Journal, pp. 185194, Oct. 1990.
[5] Beasley, P. D. L., et al., Solving the problems of a single antenna frequency
modulated CW radar, Record of the IEEE 1990 International Radar Con-
ference, pp. 391395, 1990.
[6] Nathanson, F. E., and Luke, P. J., Loss from approximations to square-
law detectors in quadrature systems with postdetection integration, IEEE
Trans. on Aerospace and Electronic Systems, AES-8 pp. 7577, Jan. 1972.
[7] Filip, A. E. A bakers dozen magnitude approximations and their detection
statistics, IEEE Trans. of Aerospace and Electronic Systems, AES-12, pp.
8689, Jan. 1976.
[8] Cassandras, C. G., Discrete Event Systems Modeling and Performance Analy-
sis, Aksen and Irwin Associates, Homewood, IL, 1993.
[9] Piper, S. O., Receiver frequency resolution for range resolution in homo-
dyne FMCW radar, Proc. National Telesystems Conference, Commercial
Applications and Dual-Use Technology, pp. 169173, 1993.
[10] Piper, S. O., Homodyne FMCW radar range resolution eects with sinu-
soidal nonlinearities in the frequency sweep, Record of the IEEE Interna-
tional Radar Conference, pp. 563567, 1995.
[11] Turley, M. D. E., FMCW radar waveforms in the HF band, ITU-R JRG
1A-1C-8B meeting, Nov. 2006.
[12] Griths, H. D., and Bradford, W. J., Digital generation of high time-
bandwidth product linear FM waveforms for radar altimeters, IEE Proc.
F, Vol. 139, No. 2, pp. 160169, April 1992.
[13] Pace, P. E., Advanced Techniques for Digital Receivers, Artech House, Inc.,
Norwood, MA, July 2000.
[14] Abousetta, M. M., and Cooper, D. C., Noise analysis of digitized FMCW
radar waveforms, IEE Proc. F, pp. 209215, Aug. 1998.
[15] Turner, S. E., Chan, R. T., and Feng, J. T., ROM-based direct digital syn-
thesizer at 24 GHz clock frequency in InP DHT technology, IEEE Microwave
and Wireless Components Letters, Vol. 18, No. 8, pp. 566568, Aug. 2008.
[16] Liao, S. Y., Microwave Devices and Circuits, 2nd Edition Prentice Hall, Upper
Saddle River, New Jersey, 1980.
[17] Harmer, J. D., and OHare, W. S., Some advances in CW radar techniques,
IRE 5th Mil-E-Con Record, pp. 311323, 1961.
[18] OHara, F. J., and Moore, G. M., A high performance CW receiver using
feedthrough nulling, Microwave Journal, Vol. 6, No. 9, pp. 6371, Sept.
1963.
[19] Lin, K., Wang, Y. E., Pao, C.-K., and Shih, Y.-C., A Ka-Band FMCW radar
front-end with adaptive leakage cancellation, IEEE Trans. on Microwave
Theory and Techniques, Vol. 54, No. 12, pp. 4041 4048, Dec. 2006.
[20] Grajal, J., Asensio, A. and Requejo, L., From a high-resolution LFM-CW
shipborne radar to an airport surface detection equipment, Proceedings of
the IEEE Radar Conference, pp. 157160, Madrid, Spain, 2004.
[21] Kim, C.-Y., Kim, J.-G., and Hong, S., A quadrature radar topology with Tx
leakage canceller for 24-GHz radar applications, IEEE Trans. on Microwave
Theory and Techniques, Vol. 55, No. 7, pp. 14381444, July 2007.
FMCW Radar 121
[22] Tujaka, S., On frequency stability of transmitter in LFMCW MTD radar,

12th International Conference on Microwaves and Radar, MIKON 98, Vol.
3, pp. 776780 May 1998.
[23] Fuchs, J., et al., Simple techniques to correct for VCO nonlinearities in short
range FMCW radars, IEEE MTT-S International Microwave Symposium
Digest, pp. 11751178, 1996.
[24] Levanon, N., and Getz, B., Comparison between linear FM and phase-coded
CW radars, IEE Proc. F, Vol. 141, No. 4, pp. 230240, Aug. 1994.
[25] Jankiraman, M., Wessels, B. J., and van Genderen, P., Pandora multifre-
quency FMCW/SFCW radar, Record of the IEEE International Radar Con-
ference, pp. 750757, 2000.
[26] Jankiraman, M., de Jong, E. W., and van Genderen, P., Ambiguity analy-
sis of Pandora multifrequency FMCW/SFCW radar, Record of the IEEE
[27] Fu, J. S., and Ke, Y., Anti-jamming aspects of linear FM and phase coded
pulse compressions by simulation, CIE International Conference of Radar
Proc., pp. 605608, Oct. 810, 1996.
[28] Reynolds, L., and Ayasli, Y., Single chip FMCW radar for target velocity
and range sensing applications, Technical Digest of 11th Annual GaAs IC
Symposium, pp. 243246, 1989.
[29] Yhland, K., and Fager, C., A FET transceiver suitable for FMCW radars,
IEEE Microwave and Guided Wave Letters, Vol. 10, No. 9, pp. 377379,
Sept. 2000.
[30] Sasaki, K., et al., InP MMICs for V-band FMCW radar, IEEE MTT-S
International Microwave Symposium Digest, pp. 937940, 1997.
[31] Tessmann, A., et al., A 77 GHz GaAs pHEMT transceive MMIC for auto-
motive sensor applications, Proc. of the GaAs IC Symposium, pp. 207210,
1999.
[32] Lamberg, J. R., et al., A compact high performance W-band FMCW radar
front-end based on MMIC technology, IEEE MTT-S International Microwave
Symposium Digest, pp. 17971800, 1999.
[33] Haydl, W. H., et al., Single-chip coplanar 95 GHz FMCW radar sensors,
IEEE Microwave and Guided Wave Letters, Vol. 9, pp. 7375, Feb. 1999.
[34] Nalezinski, M., Vossiek, M., and Heide, P., Novel 25 GHz FMCW front-
end with 2.45 GHZ SAW reference path for high precision distance mea-
surements, IEEE MTT-S International Microwave Symposium Digest, pp.
185188, 1997.
[35] Vossiek, M., Kerssenbrock, T. V., and Heide, P., Novel nonlinear FMCW
radar for precise distance and velocity measurements, IEEE MTT-S Inter-
national Microwave Symposium Digest, pp. 511514, 1998.
[36] Musch, T., Rolfes, N., and Schiek, B., A highly linear frequency ramp gen-
erator based on a fractional divider phase-locked loop, IEEE Trans. Instru-
mentation and Measurement, pp. 634637, April 1999.
[37] Oates, D. C., Dionne, G. F., and Anderson, A. C., Magnetically tunable
superconductor filters, GOMAC Session 15, Advanced Receiver Technology,
Monterey, CA, pp. 396399, March 1999.
[38] Terrell, J., High temperature superconducting filters for military applica-
tions, GOMAC Session 15, Advanced Receiver Technology, Monterey, CA,
pp. 400403, March 1999.
[39] Kapolnek, D. J., et al., Integral FMCW radar incorporating an HTSC delay
line with user-transparent cyrogenic cooling and packaging, IEEE Trans. on
Applied Superconductivity, Vol. 3 No. 1, pp. 28202823, Aug. 1992.
Problems
0
1. Calculate the error between the envelope detector output xe = I 2 + Q2
and the envelope approximation detector output (4.1) when I = 1.3 and
Q = 3.1 for (a) a = 1, b = 1/2, and (b) a = 1 and b = 1/4.
2. An LPI emitter scans at a rate of 65 deg/s using a triangular FMCW
waveform with modulation period tm = 1 ms. The GOCFAR range
processor sends the report sequence
{0, 0, 1, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0}
for a target at RT , to a Markov chain with NX = 6 and NY = 3.

Calculate the targets azimuth extent .
3. An FMCW LPI waveform is shown in Figure 4.27. If the modulation
period tm = 5 ms, the noise factor FR = 10, and the signal-to-noise
ratio required at the receiver output is 13 dB, determine the receivers
sensitivity in dBmW.
4. For the FMCW waveform shown in Figure 4.27, if the modulation band-
width F = 1 GHz, determine the range resolution.
Figure 4.27: FMCW LPI waveform.

FMCW Radar 123
5. A range resolution of R= 5m is required to contain a ship target

completely within a range bin. (a) Calculate the modulation bandwidth
F required. (b) If the modulation period tm = 1 ms, what is the chirp
rate required of the radar?
6. Using the results in Section 4.3.2, plot the FMCW magnitude spectrum
for F = 500 kHz and tm = 1 ms.
7. The program lpi fmcw design.m, is useful for understanding FMCW
radar systems and can be used to evaluate the trade-os in the emit-
ter design. (a) Use the MATLAB program lpi fmcw design.m to de-
sign an FMCW LPI emitter at 9.3 GHz, with an ideal range resolution
R = 5.5m for a V = 200 m/s target velocity and a maximum target
distance of 18 km. Choose your modulation period to be tm = 5.5td .
Make a table listing the parameters of your design, including: the tar-
gets Doppler frequency; the maximum delay td ; the coherent processing
interval; the spectral width; the eective transmitted modulation band-
width; the degraded resolution, and minimum sampling frequency of
the ADC; the FFT size; the adjusted sampling rate of the ADC; the
maximum beat frequency; and the time bandwidth product. (b) What
is the unambiguous range and unambiguous Doppler of your design?
8. The receiver design.m program computes the beat frequencies for a par-
ticular target configuration and FMCW radar design (from the program
lpi fmcw design.m). The program asks for the range of the target and
the velocity. Keeping the velocity at 200 m/s, run the program 10 times
where, with each iteration, the range is decreased by 1m. For each itera-
tion, save both the range-to-the-target and the target velocity computed
by the FFT. Plot both of these parameters separately, as a function of
range along with the error (dierence between the true values and the
measured values). Comment on the size of this error that is computed.
9. If the maximum target velocity is VT = 30 m/s, calculate the required
modulation period such that the target moves a maximum of two range
cells during the coherent processing interval. What is the resulting
velocity resolution and spectral width?
Chapter 5
Phase Shift Keying

Techniques
In Chapter 4, we discussed the LPI technique of frequency modulation of
the CW signal. This chapter presents the phase shift keying techniques that
are useful for LPI radar waveform design. Although not a LPI modulation
technique, Barker binary PSK is discussed first, since it is the first PSK
technique to be investigated and is still widely used today. This is followed
by a discussion of polyphase shift keying techniques. These include Barker
polyphase sequences and the Frank code. Also presented are the P1, P2,
P3, and P4 codes, and polytime codes T1, T2, T3, and T4. Each of these
codes is shown to be useful as an LPI CW PSK technique due to its wideband
characteristics and the fact that it forces the intercept receiver to initially have
a large processing gain. For each code, the phase characteristics are examined,
along with the power spectrum magnitude. To quantify the usefulness as a
CW LPI waveform, the ACF, PACF, and PAF are examined. As an example
of a PSK LPI radar, the omnidirectional LPI radar which uses the Frank
code, is discussed.
5.1 Introduction
While linear FMCW has established itself as one of the most popular LPI
waveforms, PSK CW waveforms have recently been a topic of active inves-
tigation, due to the their wide bandwidth and inherently low PAF side lobe
levels achievable. For the LPI radar (as with pulsed radar), it is important
to have a low side lobe level to avoid the side lobes of large targets from
masking the main peak of smaller targets. The choice of PSK code aects
the radar performance and the implementation. For the PSK waveforms, the
125
bandwidth (inverse of the subcode period) is selected first by the designer,

in order to achieve the range resolution desired. Encompassing a large target
(such as a ship) within a single resolution cell can aid in detection, but results
in a narrow bandwidth signal. On the other hand, a wideband transmitted
signal can be chosen to divide the target echo into many resolution cells, and
is a technique that is useful for target recognition. The trade-o here is that
the radar requires a larger transmitted power to detect a target that has a
small cross section, decreasing the ability of the radar to remain quiet.
Binary phase shifting codes (e.g., 0 or 180 degrees) are popular, but pro-
vide little in the way of low side lobes and Doppler tolerance. Most useful for
the LPI radar designer are the polyphase codes where the phase shift value
within the subcode can take on many values (not just two) and the code
period T can be made extremely long. These codes have better side lobe
performance and better Doppler tolerance than the binary phase codes.1
The PSK techniques can result in a high range resolution waveform, while
also providing a large SNR processing gain for the radar. The average power
of the CW transmission is responsible for extending the maximum detection
range while improving the probability of target detection (as compared to a
pulsed signal of equal peak power). PSK techniques are also compatible with
new digital signal processing hardware, and a variety of side lobe suppression
methods [14] can be applied. Compatibility with solid state transmitters
enables power management techniques to be used that lower the average
power requirements of the transmitted CW signal. Power management al-
lows the radar to keep a targets SNR constant within the receiver, as the
range to the target changes. An example of this technique is described in
Chapter 9. In addition, the LPI radar designer can choose from a large se-
lection of PSK codes that are available, which is the subject of this chapter.
5.2 The Transmitted Signal

In the PSK radar, the phase shifting operation is performed in the radars
transmitter, with the timing information generated from the receiver-exciter.
The transmitted complex signal can be written as
s(t) = Aej(2fc t+k ) (5.1)
where k is the phase modulation function that is shifted in time, according

to the type of PSK code being used, and fc is the angular frequency of the
carrier. The inphase (I) and quadrature (Q) representation of the complex
signal from the transmitter can be represented as
I = A cos(2fc t + k ) (5.2)
1 Doppler tolerance is measured by how well the code compresses in the matched receiver,
when the received signal is Doppler shifted with respect to the reference code.
Phase Shift Keying Techniques 127
and
Q = A sin(2fc t + k ) (5.3)
Within a single code period, the CW signal is phase shifted Nc times, with
phase k every tb seconds, according to a specific code sequence. Here tb is
the subcode period. The resulting code period is
T = Nc tb s (5.4)
and the code rate is

Rc = 1/Nc tb s1 (5.5)
The transmitted signal can be expressed as
Nc
uT = uk [t (k 1)tb ] (5.6)
k=1
for 0 t T and zero elsewhere. The complex envelope uk is
uk = ejk (5.7)
for 0 t tb and zero otherwise. The range resolution of the phase coding
CW radar is
ctb
R = (5.8)
2
and the unambiguous range is
cT cNc tb
Ru = = (5.9)
2 2
If cpp is the number of cycles of the carrier frequency per subcode, the band-
width of the transmitted signal is
B = fc /cpp = 1/tb Hz (5.10)
The received waveform from the target is digitized and correlated in the
receiver using a matched (unweighted) or mismatched (weighted) filter that
contains a cascade of N sets of Nc reference coecients. The results from
each correlation are combined to concentrate the targets energy and produce
a compressed pulse having a time resolution equal to the subcode duration tb
and a height of Nc . For this reason, the number of phase code elements Nc is
also called the compression ratio. Recall that the PAF describes the range-
Doppler performance of this type of receiver, and depends on the number of
reference sets used.
5.3 Binary Phase Codes

In 1953, R. H. Barker presented binary sequences for synchronization pur-
poses in telecommunications [5]. The binary Barker sequences are finite
length, discrete time sequences with constant magnitude, and a phase of
either k = 0 or k = . The formal definition of a Barker sequence is given
below [6].
Definition 5.1
A Barker sequence is a finite length sequence A = [a0 , a1 , . . . , an ]

of +1s and 1s of length n 2 such that the aperiodic autocor-
relation coecients (or side lobes)
nk
rk = aj aj+k (5.11)
j=1
satisfies |rk | 1 for k = 0 and similarly rk = rk .
Consequently, a binary Barker sequence has elements ai {1, +1}, which

are only known for lengths Nc = 2, 3, 4, 5, 7, 11, and 13. A list of the nine
known Barker sequences is shown in Table 5.1 along with their PSL (dB) and
ISL (dB). The longest code is of length Nc =13. The nine sequences are listed
where a +1 is represented by a + and a 1 is represented by a . It has
been shown that binary Barker sequences with lengths greater than 13, with
Nc odd, do not exist. Also, it has been proven that binary Barker sequences
with 4 < Nc < 1, 898, 884 with Nc even do not exist. It has been conjectured
that sequences with Nc 1, 898, 884 with Nc even also do not exist [7].
Compound Barker codes (Barker code within a Barker code) can also be
created to have a large compression gain. An example of a compound Barker
code made from a Barker sequence of length Nc = 4 is shown in Figure 5.1.
Although a larger compression gain is achieved, the peak side lobes are not
proportionally decreased. The Nc = 169 compound Barker code is frequently
used and consists of a 13 Barker code inside a 13 Barker code. This represents
the longest binary code sequence from a single concatenation.
The Barker codes are the most frequently used binary code since they
result in an ambiguity function with side lobe levels, at zero Doppler, not
higher than 1/Nc relative to a main lobe of level 1. In fact, due to this
property, Barker codes are often called perfect codes. Figure 5.2(a) shows the
ACF (rk ) of a CW signal phase coded with an Nc = 13bit Barker sequence,
and reveals the side lobe structure of the code. For this signal, fc = 1 kHz
and the sampling frequency fs = 7 kHz. Note the side lobe characteristics
reflecting the perfect nature of the Barker codes. For the Nc = 13bit code
Figure 5.1: A compound Nc = 4 Barker code.
Table 5.1: Nine Barker Codes with Corresponding PSL and ISL
Code PSL ISL

Length Code Elements (dB) (dB)
2 +, + 6.0 3.0
3 ++ 9.5 6.5
4 + + + 12.0 6.0
4 + + + 12.0 6.0
5 ++++ 14.0 8.0
7 ++++ 16.9 9.1
11 +++++ 20.8 10.8
13 +++++++++ 22.3 11.5
Figure 5.2: (a) ACF and (b) PACF for the Nc = 13-bit Barker binary PSK
signal (PSL = 22 dB).
shown, PSL = 20 log10 (1/Nc ) = 22.3 dB. The number of cycles per phase
cpp = 1. Figure 5.2(b) shows the PACF and reveals the fact that the Barker
codes do not have a perfect PACF side lobe characteristic (zero side lobes),
but have a lowest side lobe level that equals the PSL shown for the ACF
(22 dB). In Figure 5.3, a plot of the PAF is shown for N = 1. The delay axis
is normalized by the bit period tb so the PAF repeats every = Nc bsc = 91
samples. Note the presence of the large Doppler side lobes.
Upon reception of the targets return signal, the receiver uses a detector to
generate a + or for each subcode. Figure 5.4 demonstrates the binary phase
coding technique and receiver architecture using an Nc = 13-bit Barker code.
In this figure, the receiver output uses a single tapped delay line matched
filter to compress the transmitted waveform. When the return signal vector
is centered within the filter, the + filter coecients line up with the signal
+s and filter coecients line up with signal s, and a maximum output
results as shown.
In addition to having a limited code length, Barker codes are very sensitive
to Doppler shifts, as illustrated by the large PAF Doppler lobes shown in
Figure 5.3. The Doppler shift of the return waveform (due to a motion of the
target) can compress the waveform within the filter such that the matched
filter gives incorrect results. This characteristic restricts binary Barker code
applications. As a final note, Barker codes are not considered LPI since
they are easily detected by an intercept receiver that uses frequency doubling.
This simple technique involves multiplying the received signal by itself and
Figure 5.3: PAF for 13-bit Barker binary PSK signal showing the large
Doppler side lobes.
Figure 5.4: Binary phase coding techniques and receiver architecture using a
13 Barker code (Nc = 13).
processing the result with an envelope detector.

Barker BPSK signals can easily be generated using the LPIT contained
on the enclosed CD (see Appendix A). The BPSK code generates three types
of CW Barker sequences (Nc =7, 11, or 13). Also selectable are the car-
rier frequency fc , the sampling frequency fs , the number of code periods to
generate, the cycles per subcode cpp, and the SNR.
5.4 Polyphase Codes

Polyphase sequences are finite length, discrete time complex sequences with
constant magnitude but with a variable phase k . Polyphase coding refers to
phase modulation of the CW carrier, with a polyphase sequence consisting
of a number of discrete phases. That is, the sequence elements are taken
from an alphabet of size Nc > 2. Increasing the number of elements or
phase values in the sequence allows the construction of longer sequences,
resulting in a high range resolution waveform with greater processing gain in
the receiver or equivalently a larger compression ratio. The trade-o is that
a more complex matched filter is required compared to a Barker code filter.
Note that a greater sequence length Nc does not aect the signal bandwidth
at the antenna and/or change the transmitted signal bandwidth (B = 1/tb ).
Polyphase sequences that satisfy the Barker criteria (so-called polyphase
Barker codes) are currently under investigation in order to try and find longer
sequences. Polyphase compression codes have also been derived from step-
approximation-to-linear-frequency modulation waveforms (Frank, P1, P2)
and linear-frequency modulation waveforms (P3, P4). These codes are de-
rived by dividing the waveform into subcodes of equal duration, and using a
phase value for each subcode that best matches the overall phase trajectory
of the underlying waveform. An alternate approach to approximating these
waveforms is to quantize the underlying waveform into a user-selected number
of phase states, where the time spent at each phase state changes (in time)
throughout the duration of the waveform. These codes are referred to as
polytime codes. Other codes, such as the P (n, k) polyphase codes, have been
derived using a step approximation of the phase function from a nonlinear
frequency modulation waveform with a favorable energy density.
The importance of polyphase coding to the LPI community is that by in-
creasing the alphabet size Nc , the autocorrelation side lobes can be decreased
significantly while providing a larger processing gain. By narrowing the sub-
code width tb (so there are fewer cycles per phase), the transmitted signal
can also be spread over a large bandwidth, forcing the receiver to integrate
over a larger band of frequencies. Polyphase signals can easily be generated
using the LPIT contained on the enclosed CD (see Appendix A). The user
can select any sequence length, carrier frequency fc , sampling frequency fs ,
and number of code periods to generate. The cycles per subcode cpp and the
SNR are also selected.
5.5 Polyphase Barker Codes

Polyphase Barker codes allow the LPI emitter a large amount of flexibility
in generating the phase modulated waveforms. Since the number of dierent
phase terms (or alphabet) is not two-valued, there is considerable advantage
to their use since they are unknown to the noncooperative intercept receiver.
Over the years Nc 63 codes have been discovered and are presented below.
Consider the generalized Barker sequences {aj } of finite length n where
the terms aj are allowed to be complex numbers of absolute value 1 where the
correlation is now the Hermitian dot product2
nk
rk = aj aj+k (5.12)
j=1
where z represents the complex conjugate of z and the same restrictions in

Definition 5.1 apply (|r(k)| 1 for k = 0). A class of transformations can be
developed that leave the absolute value of the correlation function unaltered,
so that, in particular, generalized Barker sequences are changed into other
generalized Barker sequences [6]. For example, let {ur } be a complex sequence
of length k and let its autocorrelation function be
n
rru = ur ur+ (5.13)
j=1
Now define a new complex-valued sequence {vr } of length n as
vr = ur ej2r/m (5.14)
where m is any nonzero integer. We can then observe the fact that the
autocorrelation function rv satisfies
k
rv =
vr vr+ (5.15)
r=1
or
k
rv = ur ur+ ej2 /m = ru ej2 /m (5.16)
r=1
for all . Since |ej2 /m | = 1, |rv | = |ru | for all Also from (5.13) |vr | = |ur |.
A more general transformation between vr and ur is given in [6].
2 The Hermitian dot product of two vectors (x1 , x2 , . . . , xn ), (y1 , y2 , . . . , yn ) is
n
x y .
i=1 i i
Table 5.2: Seven Alternating Quarternary Barker Codes [6]
Code
Length Code Elements
2 +1, +j
3 +1, +j, +1
4 +1, +j, 1, +j
5 +1, +j, 1, +j, +1
7 +1, +j, 1, +j, 1, +j, +1
11 +1, +j, 1, +j, 1, j, 1, +j, 1, +j, 1
13 +1, +j, 1, j, +1, j, +1, j, +1, j, 1, +j, +1
Taking m = 4 in (5.13), the sequence (u1 , u2 , u3 , u4 . . . , uk ) is trans-

formed into (u1 , ju2 , u3 , ju4 , . . . , (1)k1 uk ) which sets up a one-to-one
correspondence between ordinary (binary) Barker sequences and four-symbol
(1, j) Barker sequences in which the real and imaginary terms alternate.
This set of sequences is shown in Table 5.2. Note that the alternating Barker
sequences of odd length are all palindromic (i.e., read the same forward and
backward) and show a symmetry that is obscured in the binary represen-
tation. For a list of all the generalized Barker sequences whose terms are
restricted to the complex sixth roots of unity with Nc 13 the reader is
referred to [6].
Until recently, construction methods for generating Nc phase Barker se-
quences with low autocorrelation side lobes were not known and exhaustive
search routines were used. These methods and results are discussed below.
In [8] an exhaustive search for all sixth-root Barker sequences was conducted
through Nc = 22, for all eighth-root sequences through Nc = 15, and for all
12th-root sequences through Nc = 15. Table 5.3 shows these results. In the
results, the sequence values ai are rth roots of unity. That is, they are the
roots of the polynomial z r 1 = 0. Their search extended the list of known
Barker sequences up to Nc = 19 where the terms of the sequence are sixtieth
roots of unity. In Table 5.3, the terms ai are expressed in terms of their phase
angles as multiples of 6 degrees. Their results illustrate the smallest r known
for each Nc where r divides 60.
An iterative algorithm based on constrained iteration techniques is applied
to generate polyphase Barker sequences in [9]. Uniform sequences meeting the
Barker condition with Nc = 3 up to Nc = 25 elements are reported (except
for Nc =20 elements). The sequence values ai are shown in Table 5.4.
The list of uniform sequences was extended up to Nc =31 using the great
Table 5.3: Polyphase Barker Sequences Nc = 1 to 19 [8] ( c IEEE 1989)

Nc r Sequence ai
1 1 0
2 1 0 0
3 2 0 0 30
4 2 0 0 30 0
5 2 0 0 0 30 0
6 6 0 0 10 30 0 40
7 2 0 0 0 30 30 0
30
8 6 0 0 0 0 30 20
50 10
9 3 0 0 0 0 20 40
0 40 20
10 6 0 0 0 0 20 40
0 30 10 50
11 2 0 0 0 30 30 30
0 30 30 0 30
12 6 0 0 10 0 50 20
20 40 0 40 10 30
13 2 0 0 0 0 0 30
30 0 0 30 0 30
0
14 6 0 0 0 10 10 0
40 40 10 20 40 10
50 20
15 4 0 0 0 15 15 0
45 45 15 30 45 15
0 30 0
16 12 0 0 0 15 15 0
45 45 15 30 45 15
0 30 0 35
17 15 0 0 0 4 20 16
0 48 40 4 28 32
0 20 4 32 56
18 60 0 0 9 8 53 47
4 7 35 32 15 37
54 57 35 3 30 8
19 60 0 0 9 4 16 15
41 3 58 20 43 57
31 56 32 28 11 2
42
Table 5.4: Polyphase Barker Sequences Nc = 3 to 25 [9] ( c IEE 1989)
Nc ai Nc ai
3 0.0121 1.1943 +0.1409 18 +2.1297 2.4221 1.6995 +0.3822

4 2.5016 2.7749 1.2319 +2.1417 2.5280 0.2354 0.1811 +1.3279
5 +1.3354 0.8992 1.8595 1.4368 +0.2331 +2.3578 0.4065 +0.0976
+0.2608 0.8630 +0.5150 1.6150 2.7582
6 +1.3630 +3.0246 0.5804 +3.1349 +2.4890 +3.7406
+1.6235 +1.1919 19 1.9461 +2.6015 0.0648 1.3004
7 +0.8688 +0.9106 0.9048 0.6238 +2.1067 +0.5609 +2.4323 1.6685
+1.7924 +0.0223 +2.6777 2.7650 0.4415 +1.6424 1.4620
8 0.3131 1.2660 +2.8067 +2.6116 0.4287 +1.5349 +2.3556 +1.0432
2.9788 1.4428 +2.1516 1.8044 +1.1451 +0.3446 +0.7259
9 0.6896 2.6573 +7.2661 1.0342 20 +0.4786 +0.6706 +1.5770 +1.8712
0.7098 1.7158 0.6370 +1.5811 +2.2135 2.1501 +0.8608 +1.1924
+2.8670 2.5136 +0.0993 +1.7860 0.3836
10 0.5887 1.8332 +2.1561 0.3414 2.8818 +0.0677 1.3544 2.4389
1.7728 0.8423 1.9530 0.7544 +2.4420 +2.5909 +1.1860 0.6338
+1.6298 +2.8294 21 +1.8087 +0.7662 0.5432 1.5944
11 +2.0760 0.6452 +2.3212 +1.9586 +1.3209 +0.0701 0.8463 1.4062
0.8944 +1.5958 +1.8337 +1.6372 0.1876 1.9735 +1.7062 1.4345
1.2395 1.1485 1.6531 0.0882 +0.2233 +1.3810 +2.8241
12 +2.2363 2.6854 +3.1305 2.8078 1.7205 +2.1529 2.4726 0.5451
1.5832 +3.0342 0.0850 +2.2462 +0.9588
+1.9194 0.6520 0.3623 3.0102 22 +2.7208 +2.7943 +2.4522 +2.0044
13 +2.7188 +1.3744 +2.0519 +0.6899 +1.5822 +1.7885 2.9352 1.3904
+1.6773 0.2793 +3.0833 +2.2955 2.1716 0.9320 +0.8861 +2.2276
+0.6941 2.8756 1.8306 0.4901 1.3843 +0.8799 2.1043 +2.4527
+1.2749 1.6386 +1.3802 0.2948 1.8923
14 0.7019 2.2825 +1.2537 1.4914 +2.6576 +0.6222
+2.0131 +2.2090 +1.7266 +2.9683 23 +0.0179 +1.1156 +2.3422 +1.8448
+2.5855 +2.4207 0.6031 +0.1768 +3.1247 2.6283 1.4822 +1.2732
+2.9666 2.6327 2.7484 1.6192 +2.3777 0.5946
15 +0.5655 +0.2147 0.4469 0.5837 +2.3777 1.6192 2.7484 +1.2732
0.9322 +2.6090 +1.0432 0.8091 1.4822 2.6283 +3.1247 +1.8448
+1.8799 +2.0924 2.9651 0.6062 +2.3422 +1.1156 +0.0179
2.8457 0.4439 +1.5204 24 +1.4496 0.3890 2.1404 +1.4554
16 0.3146 +0.1656 +0.1834 +0.4540 1.4167 +2.3177 +0.2333 +0.4296
0.0918 +1.2243 2.7713 1.9528 1.5464 2.5852 +0.8093 +2.0150
+0.0895 +2.8601 +0.2766 +2.5677 1.1339 +1.2976 3.0525 0.8208
0.1218 1.1910 2.9903 +1.6009 0.8155 0.1505 +2.1155 +2.6104
17 1.6448 +0.8837 2.7783 0.0189 +2.8658 +3.0213 +2.3375 +2.6871
+0.7623 2.4826 +2.1345 0.1942 25 +2.6251 +2.9578 +1.8614 +2.4892
2.9279 2.7891 2.5196 0.3109 1.5638 0.4785 1.0745 0.9767
0.3226 0.0768 1.8421 1.9773 2.0911 0.9913 +1.2308 +2.8284
1.9249 1.3895 1.5432 2.6046 +1.4961
0.1824 2.7280 2.8259 0.1689
2.8375 +0.6143 2.2693 +1.4167
2.0333

Nc 20 26 27 28 29 30 31
ai 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000

0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000
4.553049 0.895841 0.184756 5.463798 0.120401 0.577184 0.495357
4.086800 2.043260 0.381913 4.811496 5.552993 0.603685 2.054263
2.215325 2.412575 0.501476 3.381014 4.186235 0.587563 2.880816
1.894461 4.632634 5.667001 3.739787 4.619877 0.207471 4.126907
3.945793 4.659854 5.383231 2.796088 4.174523 5.237221 5.387150
3.770581 3.061830 4.897648 4.448612 2.799691 4.913640 5.322754
0.092210 2.055563 2.065845 4.248550 5.262097 0.461842 4.127543
0.967971 4.540587 1.731439 5.258279 5.715868 0.945976 3.776533
2.931040 3.491350 1.958458 4.368063 0.326038 2.715992 5.713595
1.003241 2.376147 4.965856 0.595450 5.580307 3.698029 4.877785
2.452194 2.690606 3.500713 5.859119 1.481594 4.041807 3.687515
3.460843 3.123204 5.477305 2.969228 1.894634 2.346154 4.313581
1.330362 1.322308 2.029632 5.908823 3.910927 1.326526 3.350924
3.745265 5.951858 5.701172 2.858973 0.109814 5.544522 1.665117
1.916620 3.270461 3.226034 5.257347 0.547566 4.813630 0.296297
4.431156 5.356688 0.931898 0.584991 3.212265 1.179540 4.764131
2.297439 3.393189 0.153375 4.028262 2.928542 5.217885 0.920844
4.416524 1.614152 3.384669 1.763293 1.569419 3.222122 5.779411
3.319561 1.694949 0.929504 5.675662 1.267677 3.908770
0.300302 4.205310 4.128530 3.971280 2.683552 5.301284
1.920568 5.841580 4.342097 2.537828 0.115401 2.568618
4.368801 1.796881 0.824984 5.757173 4.583854 0.379536
0.674586 3.989645 1.073402 1.598000 1.642393 4.286983
3.485233 5.794346 3.249433 4.602239 4.237597 0.511796
1.622635 4.575381 1.641441 6.267467 2.539604
0.561518 4.413265 2.612215 5.184811
1.039439 5.346838 1.089083
1.248448 3.330664
0.135490
deluge algorithm (GDA) in [10]. The GDA is a stochastic optimization rou-

tine that outperforms other routines such as simulated annealing or threshold
accepting. First, a quality factor Q is established based on the autocorrela-
tion properties. Searching for a global maximum value for the quality factor
Q, the GDA allows every phase step size that does not result in a Q value
lower than a certain threshold. During optimization, this threshold is contin-
uously increased. The algorithm terminates when the phase step size becomes
smaller than a predefined minimum phase step size. These results are shown
in Table 5.5 for 20 Nc 31.
In [11] uniform complex sequences of lengths 32 Nc 36 are presented.
These sequences are also derived using a stochastic optimization algorithm to
optimize a set of continuous phase values after properly selecting the start-
ing vectors. After optimization, the phase values are quantized into a finite
alphabet. The results are shown in Table 5.6. Exhaustive search routines are
not feasible for large alphabet sizes. In addition, a suitable initial value for
the quality factors must be chosen. In [12], a systematic method is presented
Table 5.6: Polyphase Barker Sequences Nc = 32 to 36 [11] ( c IEEE 1996)

Nc P ai = P i /2, i = 2 Nc 1
32 720 27 33 181 220 190 121 666 614 578 563 171 328
497 670 343 152 128 443 596 220 74 545 359 39 358
576 165 584 266 659
33 720 286 307 678 665 361 267 38 217 332 433
451 455 637 477 369 452 283 227 150 72 371
654 453 217 605 233 546 700 376 713 329
34 360 11 1 307 245 200 184 231 293 300 348 45 227
247 57 335 1 127 249 68 91 315 221 57
116 238 58 287 127 273 127 5 216
35 11,520 2,984 2,094 5,326 4,236 11,012 8,940 10,804 9,642

6,322 1,797 1,176 295 10,427 10,697 782 5,051 9,316
9,634 4,748 3,612 4,521 9,491 4,120 4,014 10,924 4,157
7,826 2,363 10,287 5,043 9,622 3,440 8,140
36 180 41 59 114 114 29 30 77 54 10 117 106 131 118

98 110 58 6 113 89 61 63 38 133 57
128 54 160 50 133 15 62 123 30 93
based on a modified stochastic optimization procedure (similar to that in [11]).

These eorts extended the Nc phase Barker sequences to 37 Nc 45 and
are shown in Table 5.7.
Recently, the number of known polyphase Barker sequences was extended
to a length of Nc = 63 [13]. This set was discovered using an algorithm
that applies stochastic methods and calculus to the problem of finding the
polyphase sequences that are a good local minima for the base energy (sum
of the side lobe energies) [13]. Barker sequences of phase values with lengths
46 Nc 54 are shown in Table 5.8. The Barker sequences of phase values
with lengths 55 Nc 63 using the smallest alphabet size K are shown in
Table 5.9.
In summary, the search for longer polyphase Barker sequences is contin-
uing rapidly. Because the complexity of the applied numerical procedures
increases significantly with longer sequence lengths, the search will become
more and more demanding in terms of algorithm eciency and computer
resources. The entire set of polyphase Barker codes are contained within
an EXCEL file (ppbc.xls) and are called out by the MATLAB LPIT code
(ppbc.m).
5.6 Frank Code

In 1963, R. L. Frank devised a polyphase code that is closely related to the
linear frequency modulation and Barker codes [14]. The Frank code is well

Nc P ai = P i /2, i = 2 Nc 1
37 60 0 0 11 15 20 39 39 54 2 46 37 55 0 46 33 55 8 54 42 48 27
23 13 59 20 3 51 20 59 39 7 29 51 17 30 59 23
38 90 0 0 10 24 28 57 64 71 85 59 53 70 84 57 46 74 15 56 11 73
3 43 86 6 63 74 44 34 9 9 59 13 78 49 89 54 30 89
39 90 0 0 10 23 21 47 62 58 76 55 68 44 59 49 78 26 50 41 83 89
46 49 1 79 56 52 21 47 2 12 67 72 34 71 35 9 65 29 85
40 90 0 0 7 6 11 14 23 18 83 76 63 53 10 85 47 38 51 19 8 55 2
85 46 39 58 76 21 20 74 33 59 79 15 45 10 61 13 54 1 41
41 120 0 0 18 21 10 1 1 8 32 38 74 79 111 101 68 47 28 26 0 86 52

102 101 40 49 98 29 26 78 115 68 22 103 54 104 29 71 6 61
117 52
42 120 0 0 3 13 12 45 58 63 75 27 26 36 44 17 109 21 51 119 58

117 26 74 12 116 96 67 56 6 14 88 80 5 52 80 22 3 65 118
68 27 97 47
43 120 0 0 16 25 31 21 18 24 32 65 84 118 8 28 112 86 57 33 90

119 4 75 63 10 62 75 30 32 6 103 30 91 110 60 101 41 102
57 28 91 23 80 21
44 120 0 0 10 3 14 19 104 102 117 94 16 2 63 39 40 55 104 84 44

118 15 93 82 110 45 22 52 27 99 105 63 31 77 82 16 19 73
117 30 94 45 85 12 54
45 90 0 0 7 1 76 71 76 63 56 73 87 9 9 14 25 53 62 5 32 35 85 69
40 76 57 26 9 83 56 57 21 5 52 89 48 11 68 26 62 6 37 73
19 58 12
Table 5.8: Barker Sequences 46 Nc 54 With Alphabet Sizes K ( c IEEE

2005)
Nc K Phase
56 190 0, 0, 13, 37, 43, 95, 83, 115, 109, 145, 111, 12, 117, 86, 127, 116, 184, 109,
65, 121, 125, 116, 36, 92, 79, 85, 12, 1, 72, 183, 156, 135, 62, 139, 95, 16,
67, 134, 17, 138, 59, 92, 161, 46, 79, 176, 10, 127, 114, 48, 23, 148, 162,
88, 117, 35
55 150 0, 0, 8, 18, 18, 19, 22, 105, 100, 127, 119, 128, 117, 118, 53, 33, 112,
147, 132, 46, 30, 1, 133, 48, 117, 83, 31, 35, 38, 64, 144, 129, 100, 56, 39,
92, 104, 32, 140, 49, 110, 88, 14, 91, 134, 38, 84, 3, 111, 33, 95, 140, 43,
101, 19
54 200 0, 0, 23, 43, 16, 9, 40, 51, 20, 7, 67, 126, 178, 180, 71, 120, 144, 151, 61, 25,
45, 100, 86, 9, 172, 161, 142, 22, 85, 8, 96, 128, 81, 1, 18, 137, 0, 95, 132, 59,
44, 155, 16, 129, 157, 98, 47, 174, 73, 18, 145, 65, 170, 100
53 100 0, 0, 5, 3, 4, 5, 9, 13, 23, 58, 79, 99, 42, 68, 66, 99, 2, 41, 68, 29, 41, 76, 22,
25, 94, 98, 74, 59, 16, 58, 35, 62, 22, 93, 85, 19, 54, 17, 56, 94, 64, 92, 43,
26, 13, 70, 47, 95, 57, 21, 13, 86, 51
52 185 0, 0, 20, 11, 30, 26, 15, 27, 57, 26, 133, 97, 177, 149, 123, 45, 11, 140, 76,
85, 105, 3, 133, 31, 28, 58, 150, 103, 149, 39, 32, 137, 170, 100, 122, 58,
42, 86, 2, 172, 50, 128, 163, 49, 136, 76, 122, 17, 20, 108, 171
51 50 0, 0, 4, 4, 18, 20, 27, 25, 25, 26, 24, 15, 15, 14, 9, 32, 36, 2, 21, 17, 9, 27,
46, 49, 19, 29, 9, 32, 7, 43, 21, 46, 22, 47, 18, 35, 0, 22, 9, 31, 44, 5, 29, 21,
4, 49, 33, 24, 9, 49, 29
50 150 0, 0, 16, 20, 44, 48, 72, 66, 103, 40, 142, 59, 4, 92, 129, 96, 112, 82, 58, 71,
94, 67, 1, 52, 58, 112, 92, 37, 14, 59, 107, 3, 68, 146, 71, 102, 40, 58, 0, 124,
62, 67, 129, 41, 51, 138, 136, 76, 66, 13
49 90 0, 0, 5, 12, 7, 1, 0, 88, 6, 25, 43, 68, 72, 51, 29, 13, 55, 62, 10, 21, 78, 79, 28,
23, 63, 50, 81, 57, 37, 5, 9, 23, 84, 61, 47, 54, 24, 75, 23, 88, 51, 7, 43, 78,
35, 65, 15, 51, 7
48 70 0, 0, 1, 5, 14, 23, 35, 36, 26, 22, 17, 5, 68, 16, 16, 51, 53, 0, 21, 13, 63, 50,
59, 43, 21, 1, 52, 27, 53, 62, 28, 28, 0, 55, 24, 51, 5, 22, 51, 15, 50, 8, 44, 21,
64, 24, 52, 12
47 80 0, 0, 10, 13, 15, 11, 9, 15, 31, 41, 66, 74, 5, 77, 46, 35, 65, 53, 32, 15, 77,
59, 37, 30, 42, 4, 8, 39, 74, 71, 25, 57, 60, 24, 54, 23, 41, 75, 19, 58, 13,
55, 11, 61, 33, 65, 28
46 90 0, 0, 3, 14, 21, 34, 50, 70, 75, 79, 57, 61, 47, 61, 79, 22, 55, 71, 71, 25, 44,
85, 9, 67, 5, 56, 81, 59, 26, 64, 11, 58, 25, 14, 83, 85, 62, 42, 4, 56, 23, 81,
50, 24, 11, 71
Table 5.9: Barker Sequences 55 Nc 63 with Alphabet Sizes K ( c IEEE

2005).
Nc K Phase
63 2000 0, 0, 88, 200, 250, 89, 1832, 1668, 1792, 145, 308, 290, 528, 819, 1357,
1558, 1407, 1165, 930, 869, 274, 97, 10, 1857, 731, 789, 1736, 150, 1332,
1229, 390, 944, 1522, 1913, 648, 239, 1114, 1708, 200, 666, 1870, 1124,
1464, 265, 845, 1751, 1039, 53, 737, 1760, 798, 1880, 851, 1838, 1103, 419,
1711, 1155, 546, 1985, 1325, 754, 44
62 3000 0, 0, 459, 324, 361, 2987, 152, 432, 2963, 2907, 112, 598, 1276, 1489, 2216,
1814, 1505, 2536, 2949, 197, 1039, 1241, 2809, 2780, 1388, 590, 2233, 1352,
2458, 2284, 962, 172, 1453, 2245, 799, 558, 2461, 1258, 34, 1666, 2834,
1364, 2755, 1369, 2284, 796, 724, 2118, 198, 1327, 2858, 2962, 2021, 1774,
1604, 698, 1059, 100, 2995, 1923, 2278, 884
61 1930 0, 0, 58, 1761, 1762, 1703, 1724, 193, 721, 241, 247, 1855, 187, 416,
1379, 1421, 1385, 922, 362, 784, 1401, 1383, 584, 1709, 284, 807, 285, 373,
1404, 1739, 1173, 179, 750, 1, 1239, 1215, 1691, 1092, 490, 17, 160, 1047,
704, 536, 1515, 820, 1892, 1138, 1630, 139, 288, 1065, 1780, 733, 613,
1309, 1452, 550, 1673, 1049, 143
60 210 0, 0, 16, 208, 180, 153, 126, 161, 135, 78, 83, 98, 143, 127, 162, 153, 183,
141, 72, 207, 149, 167, 15, 13, 146, 58, 23, 109, 169, 208, 74, 143, 173, 199,
51, 50, 31, 142,152, 84, 74, 6, 147, 205, 151, 66, 31, 151, 27, 101, 170, 75,
172, 91, 20, 131, 1, 78, 166, 68
59 340 0, 0, 5, 321, 293, 253, 251, 285, 268, 262, 286, 14, 96, 65, 33, 43, 152, 220,
235, 71, 142, 49, 262, 176, 285, 31, 181, 150, 305, 337, 108, 138, 13, 209,
274, 163, 24, 100, 320, 169, 221, 4, 48, 209, 339, 109, 192, 33, 222, 301, 128,
45, 228, 130, 299, 188, 45, 288, 134
58 500 0, 0, 1, 47, 209, 191, 154, 364, 437, 363, 420, 51, 437, 413, 277, 382, 78,
4, 428, 267, 308, 352, 238, 115, 205, 179, 474, 425, 234, 52, 443, 311, 482,
491, 400, 234, 297, 495, 492, 169, 397, 464, 75, 259, 476, 121, 437, 183, 34,
263, 0, 64, 242, 496, 292, 68, 318, 127
57 240 0, 0, 18, 51, 31, 37, 6, 39, 43, 64, 128, 167, 187, 19, 22, 226, 163, 103, 97,
238, 200, 172, 111, 201, 72, 95, 75, 172, 2, 91, 49, 220, 209, 57, 212, 168, 116,
206, 110, 102, 25, 131, 2, 30, 143, 182, 42, 107, 216, 89, 10, 161, 29, 170, 106,
205, 86
56 190 0, 0, 13, 37, 43, 95, 83, 115, 109, 145, 111, 12, 117, 86, 127, 116, 184, 109,
65, 121, 126, 116, 36, 92, 79, 85, 12, 1, 72, 183, 156, 135, 62, 139, 95, 16,
67, 134, 17, 138, 59, 92, 161, 46, 79, 176, 10, 127, 114, 48, 23, 148, 162, 88,
117, 35
55 150 0, 0, 8, 18, 18, 19, 22, 105, 100, 127, 119, 128, 117, 118, 53, 33, 112,
147, 132, 46, 30, 1, 133, 48, 117, 83, 31, 35, 38, 64, 144, 129, 100, 56,
39, 92, 104, 32, 140, 49, 110, 88, 14, 91, 134, 38, 84, 3, 111, 33, 95, 140, 43,
101, 19
Figure 5.5: Phase relationship between the quantized linear FM and Frank
coded signals with M = 4.
documented and has recently been used successfully in LPI radars (such as
the OLPI). The Frank code is derived from a step approximation to a linear
frequency modulation waveform using M frequency steps and M samples per
frequency. The Frank code has a length or processing gain of Nc = M 2 . In
the case of a single side band detection, the result is the Frank code [15]. As
an example, consider that a local oscillator is at the start of the sweep of the
step approximation to the linear frequency waveform. The first M samples of
the polyphase code are 0 phase. The second M samples start with 0 phase,
and increase with phase increments of (2 /M ) from sample to sample. The
third group of M samples start with 0 phase and increase with (3-1)(2/M )
increments from sample to sample and so on.
Figure 5.5 shows the phase relationship between the quantized linear FM
and Frank code signal for M =4. If i is the number of the sample in a given
frequency and j is the number of the frequency, the phase of the ith sample
of the jth frequency is
2
i,j = (i 1) (j 1) (5.17)
M
where i = 1, 2, . . . , M , and j = 1, 2, . . . , M . The Frank polyphase code can
also be written as an M M matrix

0 0 0 0
0 1 2 (M 1)

0 2 4 2(M 1)
(5.18)
.. .. .. ..
. . . .
0 (M 1) 2(M 1) (M 1)2
where the numbers represent multiplying coecients of the basic phase angle
2/M . For the Frank code, the PSL = 20 log10 (1/(M )) [16, 17]. For M =8,
Nc =64, the PSL = 28 dB.
Figure 5.6(a) shows the discrete phase values that result for the Frank
code for M = 8 (Nc = 64). Figure 5.6(b) shows the signal phase modulo 2,
and demonstrates that the Frank code has the largest phase increments from
sample to sample in the center of the code. Consequently, when the Frank
code is passed through a bandpass amplifier in a radar receiver, the code
is attenuated most heavily in the center of the waveform. This attenuation
tends to increase the side lobes of the Frank code ACF.
Figure 5.7 shows the power spectrum magnitude of a Frank signal with
fc =1 kHz, fs =7 kHz, and cpp = 1 with M = 8. Figure 5.7(a) shows the
power spectrum for the signal only, and shows the wideband characteristics
(B =1 kHz) resulting from the phase modulation. Figure 5.7(b) shows the
power spectrum for the SNR = 0 dB case. Figure 5.8(a) shows the ACF and
the PACF for the Nc = 64 Frank code with N = 1. The ACF reveals the
peak side lobe level PSL = 28 dB. Figure 5.8(b) shows the PACF, and the
fact that the Frank code has a perfect PACF. Figure 5.9 shows the PAF for
the Frank code for Nc = 64 and N = 1. Note the delay and the Doppler side
lobe levels are much lower than the BPSK code examined in Section 5.3.
Another formulation to generate the Frank code can be found by exam-
ining a linear transformation. In a linear transformation of the Frank code,
the kth phase element can be expressed as [18]
2 k k(mod M )
k = [k(mod M )] (5.19)
M M
where M is any positive integer that defines the code sequence length Nc =
M 2 . If the phase-coded signal given in (5.1) is converted into digital form
with a sample period t = tb , then the kth signal sample of the polyphase
Frank code sequence envelope is
2 k k(mod M )
sk = A exp [jk ] = exp j k(mod M ) (5.20)
M M
for k = 0, 1, , Nc 1. In a radar application, the sampling rate must be

higher. Assuming that the sampling period t = tb /s where s is the number
Figure 5.6: Frank code phase values for M = 8 (Nc = 64), showing
(a) discrete phase jumps and (b) signal phase modulo 2.
Figure 5.7: Frank code power spectrum magnitude for M = 8 (Nc = 64

subcodes) for (a) signal only and (b) SNR = 0 dB.
Figure 5.8: Frank code (a) ACF and (b) PACF for Nc = 64, N = 1.
Figure 5.9: Frank code PAF for Nc = 64, N = 1.

of the samples per subcode, the kth sample of sk is
2 k k(mod M s) k k(mod M s) (k mod M s) mod s

sk = A exp j
M Ms s
(5.21)
where k = 0, 1, . . . , Nc s 1. Note that the influence of the Doppler eect is
not taken into consideration with this model.
5.7 P1 Code
The P1 code is also generated using a step approximation to a linear frequency
modulation waveform. In this code, M frequency steps and M samples per
frequency are obtained from the waveform using a double sideband detection
with the local oscillator at band center [15, 16]. The length of the resulting
code or compression ratio is Nc = M 2 . If i is the number of the sample in
a given frequency and j is the number of the frequency, the phase of the ith
sample of the jth frequency is

i,j = [M (2j 1)][(j 1)M + (i 1)] (5.22)
M
where i = 1, 2, . . . , M , and j = 1, 2, . . . , M , and M = 1, 2, 3, . . .. For the P1
code the PSL = 20 log10 (1/(M )) (the same as the Frank code).
Figure 5.10(a) shows the phase values that result for the P1 code for
M = 8 (Nc = 64). Figure 5.10(b) shows the signal phase modulo 2 and
demonstrates that the P1 code has the largest phase increments from sample
to sample at the ends of the code. When the P1 code is passed through the
bandpass amplifiers in a radar receiver the attenuation is heaviest on the
ends of the code. This tends to reduce the side lobes of the P1 AACF in the
receiver [16].3
Figure 5.11 shows the power spectrum magnitude of an fc =1-kHz P1
signal (fs =7 kHz, cpp =1) with M = 8, and reveals the wideband nature
of this phase modulation. Figure 5.11(a) shows the power spectrum for just
the signal and Figure 5.11(b) shows the power spectrum for the SNR = 0 dB
case.
Figure 5.12(a) shows the ACF and the corresponding side lobe structure
for the Nc = 64 P1 code with N = 1. Here PSL = 28 dB down from the peak
as predicted. Figure 5.12(b) shows the PACF. Note that the P1 code has a
perfect PACF with zero side lobes. Figure 5.13 shows the corresponding PAF
for the P1 code. Note that the PAF repeats at Nc bsc = Nc (cppfs /fc ) = 448
samples.
3 Recall that the bandpass amplifier attenuation increased the side lobes of the Frank
code AACF.
Figure 5.10: P1 code phase values for M = 8 (Nc = 64), showing (a) discrete
phase jumps and (b) signal phase modulo 2.
Figure 5.11: P1 code power spectrum magnitude for fc = 1-kHz signal

(fs = 7 kHz, cpp = 1) with M = 8 (Nc = 64), for (a) signal only and
(b) SNR = 0 dB.
Figure 5.12: P1 code (a) ACF and (b) PACF for Nc = 64, N = 1.
Figure 5.13: P1 code PAF for Nc = 64, N = 1.

5.8 P2 Code
For the P2 code M even, the phase increment within each phase group is the
same as the P1 code, except that the starting phases are dierent [15]. The
P2 code also has a length or compression ratio of Nc = M 2 . The P2 code is
given by [17]

i,j = [2i 1 M ][2j 1 M ] (5.23)
2M
where i = 1, 2, 3 . . . , M , and j = 1, 2, 3 . . . , M , and where M = 2, 4, 6, . . ..
The requirement for M to be even in this code stems from the desire for low
autocorrelation side lobes [15]. For the P2 code, the PSL = 20 log10 (1/(M ))
and is the same as the Frank code and P1 code. Figure 5.14(a) shows the
discrete phase values that result for the P2 code for M = 8. Figure 5.14(b)
shows the signal phase values modulo 2, and demonstrates the fact that
the phase changes are largest toward the end of the code. Figure 5.15 shows
the power spectrum magnitude of the P2 code for M = 8 (Nc2 =64) with
fc = 1 kHz, fs = 7 kHz, and cpp = 1. Figure 5.15(a) shows the power
spectrum for the signal only, and Figure 5.15(b) shows the power spectrum
for the SNR = 0 dB.
Figure 5.16(a, b) shows the corresponding ACF and PACF, respectively.
Note that the P2 code does not have a perfect PACF. In fact, the PACF is
identical to the ACF. Figure 5.17 shows the PAF for the P2 code for Nc = 64
and N = 1. An interesting observation is that the P2 PAF has an opposite
slope compared to the other PSK sequences.
5.9 P3 Code
The P3 code is conceptually derived by converting a linear frequency modu-
lation waveform to baseband, by using a synchronous oscillator on one end of
the frequency sweep (single sideband detection), and sampling the I and Q
video at the Nyquist rate (first sample of I and Q taken at the leading edge
of the waveform) [15]. The phase of the ith sample of the P3 code is given by

i = (i 1)2 (5.24)
Nc
where i = 1, 2, . . . , Nc , and Nc is the compression ratio. Figure 5.18(a)
shows the quadratic discrete phase values that result for the P3 code for
Nc = 64. Figure 5.18(b) shows the signal phase modulo 2. In the P3 code,
the largest phase increments occur at the center of the code. The P3 shares
the intolerance to precompression band limiting associated with the Frank
code [15].
Figure 5.19 shows the power spectrum magnitude of the P3 code for
Nc = 64. Figure 5.19(a) shows the power spectrum magnitude for just the
Figure 5.14: P2 code phase values for M = 8 (Nc2 = 64), showing (a) discrete
phase values and (b) signal phase modulo 2.
Figure 5.15: P2 code power spectrum magnitude for M = 8, for (a) signal
only and (b) SNR = 0 dB.
Figure 5.16: P2 code (a) ACF and (b) PACF for Nc = 64, N = 1.

Figure 5.18: P3 code phase values for Nc = 64, showing (a) discrete phase
steps and (b) signal phase modulo 2.
signal, and Figure 5.19(b) shows the power spectrum magnitude for the SNR
= 0 dB. Figure 5.20(a) shows ACF and the side lobe structure of the P3 for
Nc = 64 and N = 1. The peak side lobe ratio for the P3 code is larger than
the Frank, P1, and P2 codes. Here PSL = 20 log10 2/(Nc 2 ) dB, down
from the peak. With Nc = 64, PSL = 25 dB. This is revealed in Figure
5.20(a) which shows the corresponding ACF. Figure 5.20(b) shows the PACF
for the P3 code and indicates that the P3 has a perfect PACF. The PAF for
the P3 code is shown in Figure 5.21. Here Nc = 64 and N = 1. Note that
the PAF slope for the P3 code is opposite to that of the P2 code.
5.10 P4 Code
The P4 code is conceptually derived from the same linear frequency modu-
lation waveform as the P3 code, except that the local oscillator frequency is
oset in the I and Q detectors, resulting in coherent double sideband detec-
tion. Sampling at the Nyquist rate yields the polyphase code named the P4
[15, 16]. The P4 code consists of the discrete phases of the linear chirp wave-
form taken at specific time intervals, and exhibits the same range Doppler
coupling associated with the chirp waveform. However, the peak side lobe lev-
els are lower than those of the unweighted chirp waveform. Various weighting
techniques can be applied to reduce the side lobe levels further. The phase
sequence of a P4 signal is described by
(i 1)2
i = (i 1) (5.25)
Nc
for i = 1 to Nc where Nc is the pulse compression ratio. Figure 5.22(a)
shows the discrete phase values that result for the P4 code for Nc = 64.
Figure 5.22(b) shows the signal phase values modulo 2. The P4 code has
its largest phase increments from sample to sample on the ends of the code,
similar to the P1 code. Figure 5.23(a) shows the power spectrum magnitude
for just the signal, and Figure 5.23(b) shows the power spectrum for the SNR
= 0 dB case, and for the P4 code PSL = 20 log10 2/(Nc 2 ) (same as the
P3 code). Figure 5.24(a) shows the ACF and its corresponding side lobe
structure for the Nc = 64 P4 code with N = 1. Figure 5.24(b) shows the
PACF. The P4 is a Doppler-tolerant perfect code in that it exhibits a perfect
PACFnamely zero PACF side lobes.
Figure 5.25 shows the PAF for the P4 code for Nc = 64 and N = 1. Note
that the side lobe levels are smaller compared to nonperfect PACF codes,
such as the BPSK and P2 code.
Figure 5.19: P3 code power spectrum magnitude for Nc = 64, for (a) signal
Figure 5.20: P3 code (a) ACF and (b) PACF, for Nc = 64, N = 1.

Figure 5.22: P4 code phase values for Nc = 64, showing (a) discrete phase
steps and (b) signal phase modulo 2.
Figure 5.23: P4 code power spectrum magnitude for Nc = 64, for (a) signal
Figure 5.24: P4 code (a) ACF and (b) PACF, for Nc = 64, N = 1.

5.11 Polytime Codes

The Frank, P1, P2, P3, and P4 codes discussed above were developed by ap-
proximating a stepped frequency or linear frequency modulation waveform,
where the phase steps vary as needed to approximate the underlying wave-
form, and the time spent at any given phase state is a constant. Another
approach to approximating a stepped frequency or linear frequency modu-
lation waveform is to quantize the underlying waveform into a user-selected
number of phase states. In this case, the time spent at each phase state
changes throughout the duration of the waveform. The code sequences that
use fixed phase states with varying time periods at each phase state are given
the name polytime coding [19].
Two types of polytime coded waveforms can be generated from the stepped
frequency model and are denoted as T1(n) and T2(n), where n is the number
of phase states used to approximate the underlying waveform. The T3(n) and
T4(n) polytime sequences are approximations to a linear frequency modula-
tion waveform. Increasing the number of phase states increases the quality
of the polytime approximation to the underlying waveform, but also reduces
the time spent at any given phase state, complicating the generation of the
waveform. The phase state (or bit) durations change as a function of time.
The minimum bit duration sets the waveform bandwidth.
5.11.1 T1(n) Code

The T1(n) sequence is generated using the stepped frequency waveform that
is zero beat at the leading segment. The expression for the wrapped phase
versus time for the T1(n) polytime sequence is [19]
2 jn
T 1 (t) = mod INT (kt jT ) , 2 (5.26)
n T
where j = 0, 1, 2, . . . , k 1 is the segment number in the stepped frequency
waveform, k is the number of segments in the T1 code sequence, t is time, T
is the overall code duration, and n is the number of phase states in the code
sequence.
An example of how a stepped frequency waveform is converted into a
T1(2) polytime waveform with k = 4 segments and n = 2 phase states is
shown in Figure 5.26 (one period with length T = 16 ms). Figure 5.26(a)
shows the unwrapped phase change in the time domain. Figure 5.26(b) shows
the wrapped phase quantized to phase states of 0 and 180 degrees. Figure
5.26(c) shows a resulting 1-kHz time domain signal incorporating the phase
changes generated. In this example, the underlying waveform has k = 4
segments each, with duration of 4 ms (overall code period T = 16 ms). The
frequency step between adjacent segments is 1/4 ms = 250 Hz (B = 1,000 Hz).
In the T1 code, the first segment is zero frequency and the phase is constant
Figure 5.26: Stepped frequency waveform generating a polytime code T1(2)

showing (a) the unwrapped stepped frequency phase shift, (b) wrapped phase
quantized to phase state 0 and 180 degrees, and (c) the resulting time domain
waveform with phase modulation.
at zero. The second segment accumulates one full cycle (360 degrees) over the
duration of 4 ms. The third segment accumulates an additional two full cycles
(720 degrees) over its duration of 4 ms. The fourth segment accumulates an
additional three full cycles (1,080 degrees) over its duration of 4 ms resulting
in a total accumulated phase change of 2,160 degrees. As the phase of the
stepped frequency waveform crosses increments of 180 degrees, the quantized
phase changes to the alternate state (between 0 degrees and 180 degrees) and
remains there until the phase reaches the next 180-degree boundary as shown.
The power spectrum magnitude of the T1(2) signal only is shown in Figure
5.27(a). Figure 5.27(b) shows the power spectrum magnitude for the SNR
= 0 dB case. These plots serve to demonstrate the wideband nature of this
type of phase modulation. Note that the bandwidth is not equal to B but is
the result of the phase state with the smallest duration.
Figure 5.28(a) shows the ACF and its side lobe structure for the T1(2)
code with N = 1. The PSL is high (PSL 10 dB). From the PACF shown
in Figure 5.28(b), it is clear that the T1(2) also does not have a perfect PACF.
Figure 5.29 shows the PAF for the T1(2) code for N = 1. The plot reveals
the high Doppler side lobes expected.
5.11.2 T2(n) Code

The T2(n) sequence is generated by approximating a stepped frequency wave-
form that is zero at its center frequency. For stepped frequency waveforms
with an odd number of segments, the zero frequency is the frequency of the
center segment. If an even number of segments is used, the zero frequency is
the frequency halfway between the two centermost segments. The expression
for the wrapped phase versus time for the T2(n) polytime sequence is
2 2j k + 1 n
T 2 (t) = mod INT (kt jT ) , 2 (5.27)
n T 2
where the variables are as defined above. An example of how a stepped

frequency waveform is converted into a T2(2) polytime waveform, resulting
from n = 2 phase states and k = 4 segments, is shown in Figure 5.30 for
one period T = 16 ms. Figure 5.30(a) shows the unwrapped phase change
in the time domain. Figure 5.30(b) shows the wrapped phase quantized to
phase state 0 and 180 degrees. Figure 5.30(c) shows the resulting 1-kHz time
domain signal, illustrating the imposed phase modulation.
The underlying waveform has k = 4 segments, each with duration of
4 ms (overall code period T = 16 ms). The frequency step between adjacent
segments is 1/4 ms= 250 kHz (B = 1,000 Hz). In the T2 code, the phase
shift for the code period is significantly dierent as reflected in the unwrapped
stepped frequency phase shift and the wrapped phase that is quantized to n
phase states.
Figure 5.27: Polytime code T1(2) power spectrum magnitude, for (a) signal
Figure 5.28: T1(2) code (a) ACF and (b) PACF for N = 1.
Figure 5.29: T1(2) code PAF for N = 1.

Figure 5.30: Stepped frequency waveform generating a polytime code T2(2),

showing (a) the unwrapped stepped frequency phase shift, (b) the wrapped
phase quantized to phase state 0 and 180 degrees, and (c) the resulting time
domain waveform with imposed phase modulation.
The power spectrum magnitude of the T2(2) signal for signal only is shown
in Figure 5.31(a), and Figure 5.31(b) shows the power spectrum magnitude
for SNR = 0 dB. Note again, that the bandwidth is not equal to B but is the
result of the phase state with the smallest duration.
Figure 5.32(a) shows the ACF for the T2(2) code with N = 1. The
peak side lobe level is approximately the same as the T1(2) examined above,
except that it occurs at a larger delay ( /tb = 38). Figure 5.32(b) shows the
PACF and reveals that the T2(n) code also does not have a perfect PACF.
Figure 5.33 shows the PAF and, as also expected, has fairly large side lobes.
5.11.3 T3(n) Code

A linear FM waveform that is zero beat at its leading edge generates the
T3(n). The equation for the wrapped phase versus time for a T3 polytime
sequence is
2 nF t2
T 3 (t) = mod INT , 2 (5.28)
n 2tm
where tm is the modulation period and F is the modulation bandwidth.
An example of a T3(2) waveform generated using fc =1 kHz, F =1 kHz,
and tm =16 ms is given in Figure 5.34. Figure 5.34(a) shows the unwrapped
phase change in the time domain. Figure 5.34(b) shows the wrapped phase
quantized to phase state 0 and 180 degrees. Figure 5.34(c) shows the resulting
1-kHz time domain signal, illustrating the imposed phase modulations. The
quadratic phase accumulates 2,880 degrees after 16 ms. The wrapped phase
shifts between 0 and 180 degrees, as the quadratic phase of the linear FM
waveform passes through increments of 180 degrees.
The power spectrum magnitude of the T3(2) signal only is shown in
Figure 5.35(a) and Figure 5.35(b) shows the power spectrum magnitude for
SNR = 0 dB. Figure 5.36(a) shows the ACF for the T3(2) code with N = 1,
and indicates that the side lobe performance is somewhat better than the
T1(2) or T2(2) code (PSL 18 dB). Figure 5.36(b) shows the correspond-
ing PACF, also indicating that the T3(2) does not have a perfect PACF.
Figure 5.37 shows the PAF for the T3(2) code. As expected, the side lobes
are relatively high.
5.11.4 T4(n) Code

If the linear frequency modulation waveform is zero beat at its center and
is quantized into n discrete phase states, the T4(n) polytime sequence is
generated. The equation for the wrapped phase versus time for a T4(n)
polytime sequence is
2 nF t2 nF t
T 4 (t) = mod INT , 2 (5.29)
n 2tm 2


showing (a) the unwrapped phase shift, (b) the wrapped phase quantized to
phase state 0 and 180 degrees, and (c) the resulting time domain waveform
with phase shifts.

An example of a T4(2) waveform with fc =1 kHz, F =1 kHz, and tm =16 ms

is given in Figure 5.38. Figure 5.38(a) shows the unwrapped phase change in
the time domain. Figure 5.38(b) shows the binary phase code corresponding
to each phase shift. Figure 5.38(c) shows the resulting 1-kHz time domain
signal illustrating the phase modulation.
The power spectrum magnitude of the T4(2) signal only is shown in
Figure 5.39(a). Figure 5.39(b) shows the power spectrum magnitude for
SNR = 0 dB.
Figure 5.40(a) and (b) shows the ACF and PACF, respectively. The T4(2)
has side lobes that are higher than the T3(2). The T3(2) also does not have
a perfect PACF. Figure 5.41 shows the PAF for the T4(2) code for N = 1,
and shows similar performance to the T1(2) through T3(2).
Other sequences can be formed by quantizing the phase into n > 2 phase
states. Increasing the number of phase states increases the quality of the
polytime approximation of the underlying waveform [19]. Since the poly-
time sequences approximate the underlying stepped frequency and linear fre-
quency modulation waveforms, it is surprising that their properties do not
follow more closely the properties of the underlying waveform. As more phase
states are added to the polytime sequence, the agreement in time side lobe
behavior improves. Polytime coding also has the advantage that arbitrary
time-bandwidth waveforms can be generated with only a few phase states.

showing (a) the unwrapped stepped frequency phase shift, (b) the wrapped
phase quantized to phase state 0 and 180 degrees, and (c) the resulting time
domain waveform with phase shifts.

Figure 5.42: Photo of the OLPI radar (a) transmit antenna and (b) receive
antenna.
5.12 Omnidirectional LPI Radar

The objective of the omnidirectional LPI radar concept is to provide medium-
range surveillance while avoiding antiradiation missile attacks [20, 21]. The
transmitting antenna beam illuminates the observation space continuously
using a Frank phase-coded CW waveform. The CW signal is transmitted from
an antenna that uses a nonscanning main beam, as illustrated in Figure 1.4(c).
The transmitter is separated from the receiving system by approximately
100m. By separating the transmit and receive antennas, no direct coupling
exists, providing good isolation. In a tactical situation, several transmitters
can be used (with dierent frequencies and phase codes) to provide backup, or
as decoys. Figure 5.42 shows a picture of both the OLPI transmitter antenna
and receive antenna.
In the transmitting antenna, eight vertical dipoles in a column are com-
bined by a microstrip-feeding network, resulting in a fan-beam pattern with
a width of 20 degrees in elevation and 120 degrees in azimuth. The gain of
the transmitting antenna is Gt = 8. The objective here is to distribute the
energy evenly within the observation space. The transmitter power is 10W
at 2.82 GHz [22].
The receive antenna uses a multiple beam array to provide continuous
coverage of the illuminated space. The multiple beam antenna also provides
Figure 5.43: Block diagram of the OLPI radar signal processing.
directional information from the detected targets. The echo from nearby
objects is strong, while the echo from distant flying targets is weak. Conse-
quently, Doppler processing is used to extract the targets from the background
clutter. The receive antenna has a planar array of 64 columns, with each col-
umn containing eight dipoles combined by a network that is the same as for
the transmit antenna. The total number of antenna elements is 512. The
multiple beams in azimuth are formed by a 64-port Butler matrix that was
manufactured using microstrip technology. The Butler matrix is an analog
beamforming network, made up of 3-dB directional couplers and fixed phase
shifts, to form N continuous beams with an N element linear array where
N = 2x .
The half-power beamwidth of each beam in the OLPI is 2 degrees, which is
also the azimuth resolution. The gain of the receiving antenna is Gr = 512.
Since everything is illuminated by the transmit antenna, the signals received
through the receive antenna side lobes can pose a problem, making side lobe
cancelation in the receive beam particularly important.
A block diagram of the OLPI radar signal processing is shown in Figure
5.43. The received signal from the Butler matrix is passed through a low-
noise amplifier and downconverted to an IF frequency of 30 MHz, filtered,
and then converted to baseband I and Q. The ADCs performing the base-
band sampling are 12-bit devices sampling at a rate of 250 kHz. The received
signals for each resolution are integrated during a time of Ts =2s. The range
resolution is matched to the azimuth resolution at 30 km, which is 600m.
Therefore, the range resolution is 600m, resulting in a subcode width of
tb = 4 s. With fs = 250 kHz, each subcode period contains one I and
Q pair. The Frank code length is Nc = 64, resulting in a code period of
T = 256 s and is equivalent to an Ru =38.4 km.
The first step in the signal processing is to suppress the (mostly stationary)
clutter echoes, using a recursive notch filter. After clutter suppression, the
code compression is accomplished by using an FFT that eciently compresses
the Frank code. The processing gain due to code compression is Nc = 64.
Because the phase is unknown, however, there is a loss of 1 dB compared
to fully coherent integration resulting in a net processing gain of P GR =
10 log10 (Nc ) 1 or 17 dB.
The phase code compression is followed by Doppler filtering to extract the
moving targets. To detect targets with velocity v =250m/s at 38 km within
an integration time of Ts = 2s, the resulting Doppler spectral width is
2v 2v 2 Ts
= = (5.30)
R
or 60 Hz. The filter width is matched to this value, resulting in a coherent
integration time of 16 ms and corresponds to 64 code periods, each 256 s
long. That is, 64 code periods are integrated for each range bin.
To simplify the hardware complexity of having to process 4,096 range bins,
the signal is digitized into a single bit (1) [22]. The processing gain due to
the Doppler filtering of 64 phase codes is P G = 10 log10 (64) = 18 dB. Due
to the digitization into a single bit, however, a loss of 2 dB is encountered.
Also, since the Doppler frequency and phase are unknown, an additional loss
of 2.5 dB is included, resulting in a processing gain due to Doppler filtering
of P GR = 18 2.5 2 = 13.5 dB.
The fourth step described in [22] is the noncoherent integration of the
outputs of the Doppler filter bank. The noncoherent integration is carried
out over the frame time (T = 2s). During the total integration time of
Ts = 2s, a further 128 signals are integrated in amplitude individually. With
64 beams, 64 range bins, and 64 Doppler filters per range bin, a total of
262,144 resolution cells are available. The processing gain due to noncoherent
integration is 12.7 dB. The total processing gain for all three stages is then
P GR = 17 + 13.5 + 12.7 =43.2 dB. With an output detection threshold of
SNRRo = 13 dB, the required input SNRRi = 30.2 dB. With BRi = 1/tb =
250 kHz and FR = 5.6 dB (including Butler matrix loss), the sensitivity of
the OLPI can be estimated as
R = kT0 FR BRi SNRRi = 174 dBW (5.31)
The OLPI radar has been used experimentally to detect hovering helicopters
(above terrain masking for only a short time), and is described more fully
in [22]. Although the OLPI is only one example of an LPI radar that uses
phase modulation, there have also been others built that use much longer
phase codes such as the Hughes Aircraft Company quiet radar built in the
early 1980s.
5.13 Summary
In this chapter we have examined several popular PSK LPI radar schemes,
as well as some new techniques recently developed. The phase structure was
examined for each type of code, along with an analysis of the magnitude of
phase change throughout the code. The power spectrum results were also
presented. The correlation and ambiguity functions from Chapter 3 were
used to examine each waveforms ACF, PACF, and PAF in order to determine
its suitability for use as an LPI waveform. In the next chapter we look at
frequency shift keying techniques for LPI radar applications, as well as hybrid
FSK/PSK techniques.
References
[1] Lee, W. K., and Griths, H. D., Pulse compression filter generating optimal
uniform range sidelobe level, IEE Electronics Letters, Vol. 35, No. 11, pp.
873875, May 1999.
[2] Lee, W. K., Griths, H. D., and Benjamin, R., Integrated sidelobe energy
reduction technique using optimal polyphase codes, IEE Electronics Letters,
Vol. 35, No. 24, pp. 20902091, Nov. 1999.
[3] Grishin, Y. P., and Zankiewicz, A., A neural network sidelobe suppression
filter for a pulsecompression radar with powers-of-two weights, IEEE 10th
Mediterranean Electrotechnical Conference, Vol. 2, pp. 713716, 2000.
[4] Lee, W-K., and Griths, H. D., A new pulse compression technique gener-
ating optimal uniform range sidelobe and reducing integrated sidelobe level,
Record of the IEEE International Radar Conference, pp. 441446, 2000.
[5] Barker, R. H., Group synchronizing of binary digital systems, in Commu-
nications Theory, Butterworth, London, pp. 273287, 1953.
[6] Golomb, S. W., and Scholtz, R. A., Generalized Barker sequences, IEEE
Trans. on Information Theory, Vol. IT-11, No. 4, pp. 533537, Oct. 1965.
[7] Eliahou, S., and Kervaire, A., Barker sequences and dierence sets, L
Enseignement Mathematique, Vol. 38, pp. 345382, 1992.
[8] Zhang, N., and Golomb, S. W., Sixty-phase generalized Barker sequences,
IEEE Trans. on Information Theory, Vol. 35, No. 4, pp. 911912, April
1989.
[9] Bomer, L., and Antweiler, M., Polyphase Barker sequences, IEE Electronics
Letters, Vol. 25, No. 23, pp. 15771579, 1989.
[10] Friese, M., and Zottmann, H., Polyphase Barker sequences up to length 31,
IEE Electronics Letters, Vol. 30, No. 23, pp. 19301931, Nov. 1994.
[11] Friese, M., Polyphase Barker sequences up to length 36, IEEE Trans. on
Information Theory, Vol. 42, No. 4, pp. 12481250, July 1996.
[12] Brenner, A. R., Polyphase Barker sequences up to length 45 with small
alphabets, IEE Electronics Letters, Vol. 34, No. 16, pp. 15761577, Aug.
1998.
[13] Borwein, P., and Ferguson, R., Polyphase sequences with low autocorrela-
tion, IEEE Trans. on Information Theory, Vol. 51, No. 4, pp. 15641567,
April 2005.
[14] Frank, R. L., Polyphase codes with good nonperiodic correlation properties,
IEEE Trans. IT-9, pp. 4345, 1963.
[15] Lewis, B. L., Kretschmer, F. F., and Shelton, W. W., Aspects of Radar Signal
Processing, Artech House, Norwood, MA, 1986.
[16] Lewis, B. L., Range-time-sidelobes reduction technique for FM-derived poly-
phase PC codes, IEEE Trans. on Aerospace and Electronic Systems, Vol.
29, No. 3, pp. 834840, July 1993.
[17] Painchaud, G. R., et al., An experimental adaptive digital pulse compres-
sion subsystem for multi-function radar applications, Record of the IEEE
[18] Lesnik, C. J., et al., Ecient matched filtering of signal with polyphase
Frank coded sequences, IEEE 12th International Conference on Microwaves
and Radar, MIKON 98, Krakow, Poland, Vol. 3, pp. 815819, May 2022,
1998.
[19] Fielding, J. E., Polytime coding as a means of pulse compression, IEEE
Trans. on Aerospace and Electronic Systems, Vol. 35, No. 2, pp. 716721,
Apr. 1999.
[20] Wirth, W. D., Long term coherent integration for a floodlight radar, Record
of the IEEE International Radar Conference, pp. 698703, 1995.
[21] Wirth, W. D., Polyphase coded CW radar, Proc. of the IEEE Fourth
International Symposium on Spread Spectrum Techniques and Applications,
Mainz, Germany, Vol. 1, pp. 186190, Sept. 2225, 1996.
[22] Wirth, W. D., Radar Techniques Using Array Antennas, IEE, London, United
Kingdom, 2001.
Problems
1. For an LPI CW radar with an fc =9 GHz, (a) what is the subcode
period tb of the transmitted waveform if the cycles per subcode cpp = 5?
(b) What is the transmitted bandwidth of the signal? (c) If the number
of phase codes used is Nc = 128, what is the code rate Rc and the range
resolution?
Figure 5.44: LPI phase-coded waveform.
2. The generation of an LPI phase-coded CW signal is shown in

Figure 5.44. If the carrier frequency is 0.5 MHz, (a) what is the code
period (in seconds)? (b) What is the unambiguous range of the wave-
form (in km)? (c) What is the bandwidth of the transmitted signal
(in MHz)? (d) What is the range resolution of the waveform (in m)?
(e) If the phase code used is a P3 or P4 code, what would you expect
the ACF peak side lobe level to be (in decibels down from the main
lobe response)?
3. (a) Using the bpsk.m file as a template, generate a compound Barker
code with fc = 1 kHz, cpp = 2, and Nc = 7. (b) Add this capability
to the LPIT menu. (c) For this waveform, how many subcodes are
contained within a code period? (d) What is the subcode period tb ?
(e) Plot the ACF, PACF, and the PAF. (f) What can you say about
the PSL and the Doppler side lobe levels?
4. (a) Generate three code periods of an fc = 1 kHz (fs = 7 kHz), cpp = 3
polyphase Barker code with length Nc = 16. (b) Plot the ACF, PACF,
and PAF. (c) Compare these results with an fc =1 kHz (fs = 7 kHz),
cpp = 3 polyphase P4 code with length Nc = 16.
5. Use the lpit.m file in the LPIT toolbox to generate the default BPSK,
Frank, P1P4 and T1T4 codes, and both test signals by selecting the
no change option for each signal to accept the defaults. Be sure to
write down the default parameters for each case. The signals will be
saved automatically to your LPIT directory. Move these signals to a
separate folder for use in Part II of this text.
6. Run the LPIT and select the BPSK signal. (a) Change only the num-
ber of cycles per Barker bit to 1 (cpp = 1), and generate the sig-
nal. That is, each bit now contains only 1 cycle of the IF frequency.
(b) Plot the power spectrum magnitude and record the approximate
3-dB bandwidth of the waveform, comparing this with a calculation of
what you would theoretically expect. (c) Repeat for cpp = 7.
7. Consider a generalized Barker sequence with the general transformation
vr = ur ej2(r+)/x
where and x are any real numbers, x = 0 with |vr | = |ur | and |rv |
|ru |. Show that the transformation vr = ukr+1 preserves the Barker
property and runs the sequence backwards.
8. (a) Using the LPIT toolbox, generate polytime codes T1(8) and T2(8),
with fc = 1 kHz, fs = 7 kHz, k = 4, and tm = 16 ms. (b) Plot
the phase distribution within a code period for each signal. (c) Plot
the ACF, PACF, and PAF for each signal. (d) Compare the phase
distribution diagrams with those shown in Figure 5.26 and Figure 5.30.
(e) Compare the correlation and ambiguity diagrams with those shown
in Figure 5.28(a, b), Figure 5.29, Figure 5.32(a, b), and Figure 5.33.
9. (a) Using the LPIT toolbox, generate polytime codes T3(8) and T4(8),
with fc = 1 kHz, fs = 7 kHz, k = 4, F = 1 kHz, and tm = 16 ms.
(b) Plot the phase distribution within a code period for each signal. (c)
Plot the ACF, PACF, and PAF for each signal. (d) Compare the phase
distribution diagrams with those shown in Figure 5.34 and Figure 5.38.
(e) Compare the correlation and ambiguity diagrams with those shown
in Figure 5.36(a, b), Figure 5.37, Figure 5.40(a, b), and Figure 5.41.
10. (a) Determine the maximum detection range of the OLPI radar for a
T = 5 m2 target if the losses total 14 dB. (b) How does this range
compare with Ru ?
Chapter 6
Frequency Shift Keying

Techniques
CW waveforms with pulse compression allow the LPI radar to have a signif-
icant processing gain over an unintended intercept receiver, due to the code
secrecy and the large bandwidth signals that are transmitted. In Chapter
4, it was shown that an FMCW waveform is the simplest technique to im-
plement, with the compression of the waveform done using either analog or
digital hardware. In Chapter 5, PSK techniques were shown to have signif-
icant promise, due to advances in digital hardware and the fact that many
codes and code lengths are available to choose from. Another coding tech-
nique that increases the library of LPI radar waveforms is the use of frequency
shift keying. In this chapter three important FSK or frequency hopping (FH)
techniques for coding CW waveforms are presented.
6.1 Advantages of the FSK Radar

Much of the LPI radar technology fielded today is linear FMCW, with the
simplicity of this technology being its main advantage.1 The FMCW ap-
proach spreads the transmitted energy out over the modulation bandwidth
to eectively reduce the power spectral density (PSD). The main disadvan-
tage, as illustrated in Chapter 4, is the high side lobe values that occur on
the order of 13 dB down from the peak response, and so this type of wave-
form requires that some type of weighting be applied to the matched filter
response. The PSK radar uses polyphase codes to reduce the side lobe lev-
els, and the waveforms are directly compatible with digital generation and
1A review of fielded LPI radar systems is given in Chapter 2.
187
compression, making their use more attractive. In addition, the codes must
be chosen carefully in order to maintain Doppler tolerance.
An LPI radar that uses FH techniques hops or changes the transmitting
frequency in time over a wide bandwidth in order to prevent an unintended
receiver from intercepting the waveform.2 The frequency slots used are cho-
sen from an FH sequence, and it is this unknown sequence that gives the
radar the advantage in terms of processing gain. That is, the frequency se-
quence appears random to the intercept receiver, and so the possibility of it
following the changes in frequency is remote. This prevents a jammer from
reactively jamming the transmitted frequency.3 In contrast to the FMCW
and PSK techniques, the FH technique of rapidly changing the transmitter
frequency does not lower the PSD of the emission, but instead moves the
PSD about according to the FH sequence. Consequently, the FH radar has a
higher probability of detection than a PSK or FMCW waveform, but retains
a significantly low probability of interception.
In a PSK radar, all the control circuitry, modulators, and demodulators
must have enough bandwidth in order to avoid transmitting second order
eects, thereby making the overall system expensive [1]. A major advan-
tage of the FH radar is the simplicity of the FSK architecture, especially for
track processing and generating large bandwidth signals. Large bandwidth
frequency hopping radar waveforms can be generated by using coherent di-
rect analog synthesizers that generate the output frequencies using standard
VCOs and very simple digital circuitry. Drawbacks of this approach include
spurious frequencies and high levels of phase noise, due to the complex analog
circuitry required. Direct digital methods can also be used, and involve using
a digital frequency synthesizer and a digital-to-analog converter, followed by
a lowpass filter. The major disadvantage here is that the output bandwidth
is limited by the speed of the digital devices. Single or multiple phaselocked
loops can also be used and have the advantage of large bandwidths and the
ability to filter spurious frequencies outside the loop bandwidth [2].
Another advantage of the FH radar is that the range resolution is in-
dependent of the hopping bandwidth (unlike that of the FMCW and PSK
techniques). Range resolution in an FH radar depends only on the hop rate.
A significant benefit also resides in the secrecy of the FH sequence that is
used. FH radar performance depends only slightly on the code used, given
that certain properties are met. This allows for a larger variety of codes,
making it more dicult to intercept. By comparison, a PSK radar must
choose from a group of well-known codes, due to the ambiguity properties
2 A frequency hopping radar is dierent than a frequency agile radar, in that the fre-
quency agile radar is usually regarded as a pulse radar that uses a dierent frequency on a
pulse-to-pulse basis. The LPI frequency hopping radar transmits a CW frequency hopped
signal.
3 A jammer can preemptively jam a FH radar if its bandwidth and power are large enough
to cover the FH band.

Frequency Shift Keying Techniques 189
required. Although the length of the PSK code may be unknown to the in-
tercept receiver, it may still cycle through, and attempt to correlate specific
signal patterns for detection and jammer waveform construction (more about
this in Part II).
The order in which the frequencies are transmitted significantly aects the
ambiguity performance of the signal. The PAF for FH signals can easily be
approximated, because the cross-correlation signals at dierent frequencies
approach zero when the frequency dierence is large relative to the inverse of
the signal duration (or multiples of that inverse) [1]. In a multiple LPI emitter
environment, an important requirement is to keep the mutual interference
between transmitters as low as possible. Mutual interference occurs when
two or more emitters transmit the same frequency slot at the same time. The
degree of mutual interference is related to the cross correlation properties of
the FH sequences. Another advantage is that the glint (target scintillation)
error spectrum is broadened significantly, since the glint error is eectively
decorrelated when the transmitter changes frequency.
6.2 Description of the FSK CW Signal

In an FSK radar, the transmitted frequency fj is chosen from the FH se-
quence {f1 , f2 , . . . , fNF } of available frequencies for transmission at a set of
consecutive time intervals {t1 , t2 , . . . , tNF }. The frequencies are placed in the
various time slots corresponding to a binary time-frequency matrix. Each fre-
quency is used once within the code period, with one frequency per time slot
and one time slot per frequency. The expression for the complex envelope of
the transmitted CW FSK signal is given by
s(t) = Aej2fj t (6.1)
The transmitted waveform has NF contiguous frequencies within a band B,

with each frequency lasting tp s in duration.
6.3 Range Computation in FSK Radar

CW FSK radars using multiple frequencies can compute very accurate range
measurements. To illustrate, consider a CW radar that transmits the wave-
form
s(t) = A sin(2fj t) (6.2)
where the received signal from a target at a range RT is
w W
4fj RT
s(t) = A sin(2fj t T ) = A sin 2fj t (6.3)
c
Solving for RT
c
RT = T (6.4)
4fj
The unambiguous range occurs when T is maximum or T = 2 and there-
fore with one frequency, RT is limited to extremely small values that are not
practical [3]. If two frequencies are used
s1 (t) = A1 sin(2f1 t) (6.5)
and
s2 (t) = A2 sin(2f2 t) (6.6)
the received signals are
w W
4f1 RT
s1 (t) = A1 sin(2f1 t T 1 ) = A sin 2f1 t (6.7)
c
and w W
4f2 RT
s2 (t) = A2 sin(2f2 t T 2 ) = A sin 2f2 t (6.8)
c
After mixing with the carrier frequency in the receiver, the phase dierence
between the two signals is
4RT 4RT
T = (f2 f1 ) = f (6.9)
c c
Since RT is maximum when = 2, the maximum unambiguous range is
c
Ru = (6.10)
2f
and is very large since f << c. From the measurement of the phase dier-
ence, T , the range of the target is then
cT
RT = (6.11)
4f
This can also be written as a function of just the phase dierence as
T
RT = Ru (6.12)
2
Since the range to the target depends on the frequency dierence, the
range resolution then depends on the duration of each frequency as
ctp
R = (6.13)
2
The transmitted power for each frequency must be such that the energy con-
tent within the target echo is sucient for detection, and enough to ensure
that accurate phase measurements can be made.
In summary, for the FSK CW radar, the frequency dierence f de-
termines the maximum unambiguous detection range. The targets range is
computed by measuring the return signal phase dierence from two consecu-
tive transmitted frequencies. The range resolution, R, depends only on the
FH period.
6.4 Costas Codes

In a study by J. P. Costas, techniques were presented for generating a sequence
of frequencies that produce unambiguous range and Doppler measurements
while minimizing the cross talk between frequencies [4]. In general, the Costas
sequence of frequencies provides an FH code that produces peak side lobes
in the PAF, that are down from the main lobe response by a factor of 1/NF
for all regions in the delay-Doppler frequency plane. That is, the order of
frequencies in a Costas sequence or array is chosen in a manner to preserve
an ambiguity response with a thumbtack nature (the narrow main lobe and
side lobes are as low as possible). The firing order of these frequencies is
based on primitive roots (elements) of finite fields.
6.4.1 Characteristics of a Costas Array or Sequence

A Costas array or (frequency) sequence f1 , , fNF is a sequence that is a
permutation of the integers 1, , NF satisfying the property
fk+i fk = fj+i fj (6.14)
for every i, j, and k such that 1 k < i < i + j NF . An array that

results from a Costas sequence in this way is called a Costas array [5]. The
nonequivalence condition in (6.14) can be checked easily when the frequency
sequence is expressed in an NF 1 NF dierence triangle. As an added
bonus, the dierence triangle can also be used to derive the PAF. We discuss
how to derive Costas sequences in the next section and in Appendix C.
Consider the frequency sequence fj = {2, 4, 8, 5, 10, 9, 7, 3, 6, 1} kHz. Fig-
ure 6.1(a) shows the binary time-frequency matrix for this sequence. The
frequencies are fired at each ti and are indicated by a 1 in the matrix. The
power spectral density of the signal is shown in Figure 6.1(b). Also indicated
is the firing order for each frequency. If the subperiod tp (frequency duration)
is constant, the cycle density or number of cycles per frequency varies as tp fj .
Figure 6.1: Costas sequence fj = {2, 4, 8, 5, 10, 9, 7, 3, 6, 1} kHz, showing

(a) the binary time-frequency matrix, and (b) the power spectrum magnitude
for several code periods.
6.4.2 Computing the Dierence Triangle

The first step to verifying (6.14) and deriving the PAF is to form a dierence
triangle. To form the dierence triangle, we start by writing the NF frequency
values in the sequence (fj where j = 1, . . . , NF ) as column headers across the
top as illustrated in Figure 6.2(a). The NF 1 rows in the dierence triangle
correspond to the delays, with each row number i representing the delay value.
To calculate each cell value in the dierence triangle i,j
i,j = fj+i fj (6.15)
where i = 1 . . . NF 1, j = 1 . . . NF 1, and i + j NF . For example,

the first row (delay i = 1) is formed by taking dierences between adjacent
frequencies. By (6.14), all results in this row must be unique [6, 7]. The
second row in the dierence matrix is formed by taking dierences between
next-adjacent terms (delay i = 2). Results in this row must also be unique.
The process is repeated until the i = NF 1 delay is computed. All values of
i,j must be unique within each row, and is the defining criterion of a Costas
sequence.
6.4.3 Deriving the Costas Sequence PAF

The PAF can be approximated by overlaying the binary time-frequency ma-
trix upon itself, and shifting one relative to the other according to a particular
delay (horizontal shifts) and particular Doppler (vertical shifts). At each com-
bination of shifts, the sum of coincidences between points of the fixed and
the shifted matrix, represents the relative height of the PAF. An easier way
to derive the PAF is using the dierence triangle, as shown in Figure 6.2(a).
In Figure 6.2(b) the PAF of the 10-frequency Costas signal is derived from
the dierence triangle. The PAF is constructed by considering each row (de-
lay) in the dierence triangle, and placing a 1 in the PAF delay-Doppler
cell corresponding to each i,j . The delay i = 1 is shaded as an exam-
ple. The PAF derived from this complex Costas signal with fs = 27.5 kHz
(power spectrum magnitude shown in Figure 6.1) is shown for comparison in
Figure 6.2(c).
6.4.4 Welch Construction of Costas Arrays

There are many analytical procedures for constructing Costas frequency hop-
ping arrays. Although Costas arrays may exist in principle for any positive
integer NF , these analytical construction methods are typically limited to
values of NF related to prime numbers [6, 8, 9]. Most construction methods
to produce a large number of Costas arrays of equal length are based on the
properties of primitive roots (see the tutorial in Appendix C).
For the Welch construction of a Costas array, an odd prime number p is
chosen first. The number of frequencies and the number of time slots in the
Costas sequence are then NF = (p) = p 1 where (p) is the Euler function.
Next, a primitive root g modulo p is chosen. As discussed in Appendix C, if g
is a primitive root modulo p, then g is an integer belonging to the Euler-(p)
function modulo p. Since g is a primitive root modulo p, g, g 2 , . . . , g (p) are
mutually incongruent and form a permuted sequence of the reduced residues
p. Welch showed that this reduced residue sequence is a Costas sequence.
Theorem 6.1 Let g be a primitive root of an odd prime number p. Then the
(p1) by (p1) permutation matrix A has elements ai,j =1 i j g i (mod p)
for 1 i, j p 1 and this is a Costas array [10].
Example: The first step is choosing a prime number p. We choose the prime
number p = 7. For p = 7, NF = 6, which is the number of frequencies
in the FH code. The complete residue system is {0, 1, 2, 3, 4, 5, 6}. With p
prime, we know that the number of elements in the reduced residue system
is (p) = p 1 = 6. The reduced residue system is {1, 2, 3, 4, 5, 6} (deleting
the one element that is divisible by p). From Appendix C, for p = 7 we know
there are exactly {(7)} = {6} = 2 mutually incongruent primitive roots
modulo p = 7. We start with the smallest value g = 2, but the order of g =2
is 3 (not 6). Next we pick g = 3 and get the desired result {3, 2, 6, 4, 5, 1},
which is the Costas array as shown below. The left-hand column shows i
from 1 to p 1, and the right-hand column shows the frequency j using
Figure 6.2: Costas sequence with (a) the dierence triangle, (b) the PAF
derived from the dierence triangle, and (c) the PAF derived from the complex
signal.
Theorem 6.1.
i j = g i (mod7)
1 31 = 3
2 32 = 2
3 33 = 6
4 34 = 4
5 35 = 5
p1
6 3 = 36 = 1
where i, j = 1, 2, . . . , 6. Consequently, the Costas frequency sequence is
f = {3, 2, 6, 4, 5, 1} (6.16)
Note from the symmetry that f = {5, 4, 6, 2, 3, 1} is also a Costas sequence.
The Welch construction is also singly periodic [5]. That is, the sequence
fi , fi+1 , . . . , fi+p2 is also a Costas sequence. Any circular shift of the se-
quence is also a Costas sequence. The costas.m program within the LPIT
allows the user to quickly add additional sequences that may be of interest.
6.5 Hybrid FSK/PSK Technique
The hybrid LPI radar technique discussed in this section combines the tech-
nique of FSK (FH using Costas sequences) with that of a PSK modulation
using sequences of varying length [11, 12]. This type of signaling can achieve a
high time-bandwidth product or processing gain, enhancing the LPI features
of the radar. Ambiguity properties of the signal are retained by preserving
the desirable properties of the separate FSK and PSK signaling schemes. The
FSK/PSK techniques can maintain a high Doppler tolerance, while yielding
an instantaneous spreading of the component frequencies along with an en-
hanced range resolution [11]. Below, a Costas-based FSK/PSK signal (Barker
5-bit PSK over each frequency) is investigated as an example. Other PSK
techniques from Chapter 5 can also be investigated using the LPIT.
6.5.1 Description of the FSK/PSK Signal

Recall that for the FH LPI radar, the CW waveform has NF contiguous fre-
quencies within a bandwidth B, with each frequency lasting tp s in duration.
The hybrid FSK/PSK signal further subdivides each subperiod into NB phase
slots, each of duration tb as shown in Figure 6.3. The total number of phase
slots in the FSK/PSK waveform is then
NT = NF NB (6.17)
Figure 6.3: General FSK/PSK signal containing NF frequency subcodes

(hops) each with duration tp s. Each frequency subcode is subdivided into
NB phase slots, each with duration tb .
with the total code period T = tb NB NF . The expression for the complex
envelope of the transmitted CW FSK/PSK signal is given by
s(t) = Aej2fj t+k (6.18)
where k is one of NB Barker codes for this example, and fj is one of NF
Costas frequencies.
During each hop, the signal frequency (one of NF frequencies) is modu-
lated by a binary phase sequence, according to a Barker sequence of length
NB = 5, 7, 11, or 13. As an example, the FSK/PSK signal generated by
using the NF = 6 Costas sequence (6.16), and phase modulating it with a
Barker binary phase modulation of length NB = 5 gives the signal: S = 3+ ,
3+ , 3+ , 3 , 3+ , 2+ , 2+ , 2+ , 2 , 2+ , 6+ , 6+ , 6+ , 6+ , 6 , 4+ , 4+ , 4+ , 4 ,
4+ , 5+ , 5+ , 5+ , 5 , 5+ , 1+ , 1+ , 1+ , 1 , 1+ . The final waveform is a binary
phase modulation within each frequency hop, resulting in five phase subcodes
equally distributed within each frequency, for a total of NP NF =30 subcodes.
Figure 6.4 shows the power spectrum magnitude that reveals the spread
spectrum characteristic of the phase-modulated Costas signal
f = {3, 2, 6, 4, 5, 1} kHz. For this signal, the sampling frequency fs = 15
kHz, the subperiod for each frequency is tp =6 ms (B = 167 Hz) and an
NB = 5-bit Barker code is used. Figure 6.4(a) shows the Costas sequence
power spectrum magnitude before phase modulation. Figure 6.4(b) shows
the power spectrum magnitude of the Costas sequence FSK/PSK after phase
modulation. Figure 6.5(a, b) shows the ACF and the PACF, respectively, of
the FSK/PSK sequence. Note the phase modulation spikes that are present
with regular periodicity. Figure 6.6 shows the PAF and the Doppler side
lobes present. The fsk psk costas.m program within the LPIT allows addi-
tional phase modulations to be included with the Costas sequences (which can
also be easily changed). In this manner, the side lobe structure for various
phase modulations can be easily compared.
Figure 6.4: Power spectrum magnitude plot for a Costas waveform, with
(a) no phase modulation and (b) 5-bit phase modulation.
Figure 6.5: (a) ACF and (b) PACF plot for the Costas sequence with a 5-bit
Barker phase modulation.
Figure 6.6: PAF plot for the Costas sequence with a 5-bit Barker phase
modulation.
6.6 Matched FSK/PSK Signaling

The matched FSK/PSK radar concept uses a pulse compression technique
that allows it to synthesize uniform envelope signals with an arbitrary spectral
density. This type of radar, proposed in [13], concentrates the signal energy
in spectral locations of most importance within the spectrum bandwidth, but
does so with a random sequence. The radar uses FSK/PSK signals as above
but, instead of selecting the frequency from a Costas sequence, the frequency
is chosen randomly with a probability distribution that is defined by the
spectral characteristics of the target of interest. That is, the frequencies
transmitted most often are those that correspond to the spectral peaks of the
target signature. Since the FH sequence still appears random, this type of
radar can achieve a relatively low probability of intercept.
One method to estimate the spectral characteristics of a target (magnitude
and phase) is by Fourier-transforming the range profile. The range and spec-
tral characteristics depend on the details of the target structure and aspect
angle at which the target is observed. If the target is known ahead of time,
several range and frequency profiles for the various aspect angles anticipated
can be precomputed and stored in the radars bulk memory.
A random binary phase shift is also added to reduce the ambiguity func-
tion side lobes. A correlation receiver with a phase-mismatched reference
signal is used in order to allow the radar to generate signals that can match
a targets spectral response in both magnitude and phase.
Figure 6.7 shows a block diagram of the signaling scheme proposed in
[13]. The implementation starts with the complex range response of a se-
lected target with NF range samples. The targets range response is Fourier-
transformed, to give a magnitude spectrum with NF unique frequencies. The
frequency components with their corresponding magnitude and phase are
collected, to represent the probability density function of the transmitter. A
random selection process then chooses each frequency, with a probability de-
fined by the spectral characteristics of the target [14]. Consequently, the
frequencies corresponding to the spectral peaks of the target (highest magni-
tudes) are transmitted more often. Each frequency from the NF sequence is
transmitted a certain number of times and is also modulated in phase. Each
frequency starts with its initial phase value (from the FFT), but is modified
by a pseudorandom binary code with values zero or equally likely. Note
that although the spectral density function of the target may contain only NF
points, the frequencies are chosen randomly NC NT times in a particular
code period T , in order to obtain the proper probability density function.
Note that in the LPIT, the complex range response of only one target
is available and is configured to closely represent that shown in [13]. Fig-
ure 6.8(a) shows NF = 32 complex points describing the targets range
response. Figure 6.8(b) shows the probability density function or, equiv-
alently, the power spectrum magnitude computed from the targets range
Figure 6.7: Block diagram of the generation of an FSK/PSK target matched

waveform.
response. The histogram of the transmitted frequency hopping signal is

shown in Figure 6.9(a). This figure shows how many times each one of NF =
32 frequencies is transmitted in the NC =512 frequency code. Note the similar
appearance to the probability density function shown in
Figure 6.9(b), and the power spectrum magnitude given in Figure 6.8(b).
Increasing the number of random frequency selections NC > 512, will result
in an even better match. Figure 6.10 shows the resulting power spectrum
magnitude for the transmitted signal, revealing the wideband nature of this
type of hybrid FSK/PSK signaling. Figure 6.11(a) shows the ACF of the
transmitted waveform. Note the near-uniform side lobes that are down close
to 30 dB. The PACF is shown in Figure 6.11(b), and shows the absence
of any periodic components. The PAF shown in Figure 6.12 shows a spike
at ( /tb = 0, Nc tb = 0) and very low side lobes, resembling the PAF of
random noise.

FSK signals provide a higher probability of detection compared to PSK and
FMCW signals, but oer many advantages for LPI signaling. Combined with
PSK, significant LPI results can be obtained. The FSK, FSK/PSK pulse
compression signals discussed in this chapter can help a radar achieve LPI
goals. The waveforms can be generated using the LPI toolbox main menu
program lpit.m contained on the MATLAB CD.
The hybrid modulations presented in this chapter tend to make the trans-
mitted signal appear as noise-enhancing its low probability of intercept na-
ture. These hybrid techniques are a subset of a larger group of radar architec-
tures known as random signal or noise radar. Random signal radar techniques
can derive target detections using correlation, spectrum analysis, or anticor-
relation. Random signal radar modulations include: noise FMCW, sine plus
noise FMCW, random binary PSK CW, and random pulse modulation [15].
Because of the random nature of the transmitted waveform, random signal
radar also provides a good deal of electronic protection and has a counter-
electronic support capability [16]. These techniques are examined in detail in
the next chapter.
Figure 6.8: Transmitted signal showing (a) 32 complex points describing the
simulated range response and (b) the normalized power spectrum magnitude
representing the probability density function. (After [13].)
Figure 6.9: (a) Synthetic or transmitted signal histogram and (b) the original
histogram defined by the spectral characteristics of the target.
Figure 6.10: Transmitted signal power spectrum magnitude.

Figure 6.11: Transmitted signal (a) ACF and (b) PACF.
Figure 6.12: Transmitted signal PAF.

References
[1] Sanmartin-Jara, J., Burgos-Garcia, M., and Retamosa-Sanchez, J., Radar
sensor using low probability of interception SS-FH signals, IEEE Aerospace
and Electronics Magazine, pp. 2328, April 2000.
[2] Benn, H. P. and Jones, W. J., A fast hopping frequency synthesizer, Second
International Conference on Frequency Control and Synthesis, pp. 6972,
April 1989.
[3] Mahafza, B. R., Radar Systems Analysis and Design Using MATLAB, Chap-
man & Hall/CRC, New York, 2000.
[4] Costas, J. P., A study of a class of detection waveforms having nearly ideal
range-Doppler ambiguity properties, Proc. of the IEEE, Vol. 72, No. 8, pp.
9961009, August 1984.
[5] Golomb, S. W., and Moreno, O., On periodicity properties of Costas arrays
and a conjecture on permutation polynomials, IEEE Trans. on Information
Theory, Vol. 42, No. 6, pp. 22522253, Nov. 1996.
[6] Golomb, S. W., and Taylor, H., Construction and properties of Costas ar-
rays, Proc. of the IEEE, Vol. 72, No. 9, pp. 11431163, Sept. 1984.
[7] Levanon, N., Radar Principles, John Wiley & Sons, New York, NY 1988.
[8] Maric, S. V., Seskar, I., and Titlebaum, E. L., On cross-ambiguity properties
of Welch-Costas arrays, IEEE Trans. on Aerospace and Electronic Systems,
Vol. 30, No. 4, pp. 10631071, Oct. 1994.
[9] OCarroll, L., et al., A study of auto- and cross-ambiguity surface per-
formance for discretely coded waveforms, IEE Proc. F Radar and Signal
Processing, Vol. 137, No. 5, pp. 362370, Oct. 1990.
[10] Lemieux, J. A., Analysis of an optimum hybrid radar waveform using fre-
quency hopping and locally optimum signals, Proc. of the IEEE National
Radar Conference, pp. 98102, March 1213, 1991.
[11] Donohoe, J. P., and Ingels, F. M., The ambiguity properties of FSK/PSK
signals, Record of the IEEE 1990 International Radar Conference, 1990, pp.
268273 May 710, 1990.
[12] Skinner, B. J., Donohoe, J. P., and Ingels, F. M., Simplified performance
estimation of FSK/PSK hybrid signaling radar systems, Proc. of the IEEE
1993 National Aerospace and Electronics Conference, NAECON, Vol. 1, pp.
255261, May 2428, 1993.
[13] Skinner, B. J., Donohoe, J. P., and Ingels, F. M., Matched FSK/PSK radar,
Record of the 1994 IEEE National Radar Conference, pp. 251255, March
2931, 1994.
[14] Marsaglia, G., Random variables and computers, Proc. of the Third Prague
Conference on Information Theory, Statistical Decision Functions, Random
Processes, Liblice, pp. 499512, June 513, 1962.
[15] Guosui, L., Hong, G., and Weimin, S., Development of random signal radars,
IEEE Trans. on Aerospace and Electronic Systems, Vol. 35, No. 3, pp. 770
777, July 1999.
[16] Garmatuk, D. S., and Narayanan, R. M., ECCM capabilities of an ultrawide-

band bandlimited random noise imaging radar, IEEE Trans. on Aerospace
and Electronic Systems, Vol. 38, No. 4, pp. 12431255, Oct. 2002.
Problems
1. An FSK CW radar is required to perform ranging up to a maximum of
15 nmi. What is the required frequency dierence f ?
2. Derive the Costas frequency sequence given in Figure 6.1. Hint: Start
by choosing the correct prime modulus p and writing down the two
primitive roots.
3. The frequency hopping sequence
{3, 9, 10, 13, 5, 15, 11, 16, 14, 8, 7, 4, 12, 2, 6, 1}
is being considered for a new LPI radar. (a) Show that this is a Costas
sequence. (b) If so, how many primitive roots are there? (c) Derive the
sequence by determining p and the primitive roots.
4. Consider the Welch construction of a Costas frequency hopping se-
quence with p = 13. (a) How many frequencies are contained in the
frequency hopping sequence? (b) Write the elements of the reduced
residue system. (c) How many primitive roots are there in the system?
Do not forget to show your work. (d) What are the primitive roots of
the system? (e) Write out the Costas sequence for each primitive root.
(f) For the sequence resulting from the largest primitive root, show that
the sequence is Costas by forming the dierence triangle. (g) Draw a
contour grid of the periodic ambiguity function for the sequence in (f)
making sure that you label the side lobe levels and main peak amplitude.
5. (a) Edit the costas.m file to include the Costas sequence given in (6.16).
(b) Compute the power spectrum magnitude and PAF of this sequence.
6. Using the fsk psk costas.m file, generate the power spectrum, ACF,
PACF, and PAF for the first Costas sequence with (a) NB = 5 and
(b) NB = 13. What is the dierence in the side lobe level you observe?
Chapter 7
Noise Techniques
In this chapter the principles of random noise radar are presented. A discus-
sion of each noise technology is then described in detail including a compar-
ative discussion of the advantages and the disadvantages of each. The major
focus is on the radar systems transmitted waveform. Mathematical models
of each transmitter have been developed in MATLAB and are included on
the CD within the Part I, LPIT Toolbox folder. The autocorrelation function
(ACF), the periodic autocorrelation (PACF) and the periodic ambiguity func-
tion (PAF) for each transmitted waveform are examined in order to compare
their Doppler side lobe and time side lobe characteristics. The four types of
noise technology radar systems discussed include: random noise radar, ran-
dom noise plus FMCW, random noise FMCW plus sine, and random binary
phase modulation.
7.1 Historical Perspective

The concept of random noise radar (RNR) is not new but was considered as
early as the 1950s as a way to eliminate all the range-Doppler ambiguities
in the radar (i.e., thumbtack ambiguity function). RNR systems transmit a
random or random-like low power microwave noise waveform that may (or
may not be) modulated by a lower frequency waveform. The peak value of a
cross-correlation process (delayed copy of the transmitted signal corresponds
with the echoes of the target) can be used to determine the distance to the
target. The earliest reported investigations of noise technology used as a range
measurement system are given by Horton [1] and Craig [2] in the Institute of
Radio Electronics. A short time later Grant et al. [3], Cooper et al. [4], and
McGillem [5] at Purdue University put forth a theoretical analysis and some
prototypes were built. Further experimental results of a complete noise radar
207
system were obtained in [6] showing that it was possible to detect a target in
very poor SNR conditions even with nonoptimum hardware.
The research however, was quickly dropped since the development of noise
waveform sources with the required bandwidth was dicult. Also since cross-
correlation processing of the transmitted and received signal was necessary
the use of variable microwave delay lines was required [7]. Before the 1960s,
the manufacturing of these devices was complicated. From the 1960s to the
1970s, the research into RNR ramped up quickly and several experimental
systems were built and tested. A good overview of the dierent techniques is
given in [8] and an extensive bibliography of the early development in RNR
is given in [9].
With the development of solid-state microwave techniques and high-speed
integrated VLSI circuits the technology began to support the RNR concepts
and implement the required processing. Today the RNR waveform can be gen-
erated digitally followed by a digital-to-analog converter and up-conversion
onto a carrier signal. They are also relatively inexpensive to build and many
dierent variants on the RNR are possible including the use of UWB wave-
forms.
Noise technology radar can be used to detect targets in both range and
Doppler. RNR emitters have good electronic protection properties by pos-
sessing a natural immunity to jamming and interference from other radar
systems operating in the same theater of operations [10, 11]. The use of a
RNR provides the advantage that it is uncorrelated with the intentional and
unintentional interference as well as other noise sources. That is, the correla-
tion process used in the receiver allows it to sort out the incoming signals even
within the same band making it attractive in multi radar environments [10].
These advantages are due to the properties of the RNR featureless waveform.
These include transmitting the lowest obtainable instantaneous power spec-
tral density possible by spreading its energy over a wide signal bandwidth and
the use of non-redundant waveforms that appear random and are concealed
in the ambient thermal noise and interference environment [12].
The exception is the use of deception. Deception is a repeater technique
(constant gain) in which false targets are created in the radar receiver that
are interpreted as valid targets. The jam-to-signal ratio is independent of the
range between the repeater jammer and the radar. Post detection integration
of target signal returns can normally provide a significant decrease in jam-to-
signal ratio however, for deception techniques, the integration gain is equal
for valid and false targets.
RNR systems that use random noise also have a significant processing
gain unavailable to the noncooperative intercept receiver since their low mean
power and noise-like characteristics result in a very low SNR. Even if the
signal is detected, it is unlikely to be identified making these types of emitters
important for many LPI and LPID applications.
RNR systems and waveforms are becoming useful in certain (limited) ap-
Noise Techniques 209
plications such as acoustic radarusually called sodar (sound detection and

ranging) [13]. Also of growing importance are the high-resolution measure-
ment of range profiles for foliage- and ground-penetrating detection of buried
objects such as mines [1417], inverse synthetic aperture radar (ISAR) and
synthetic aperture (SAR) imaging [18, 19]. Covert tracking of targets using
RNR monopulse techniques are discussed in [2022]. Car collision warning
and avoidance systems and UAV landing systems are also being investigated
[7]. The detection of small-RCS targets in surface clutter depends on the
signal-to-clutter ratio (SCR) in the target range resolution cell and the abil-
ity to resolve the target signal in Doppler. The SCR (in the absence of range
ambiguity) is the ratio of target RCS to clutter RCS, and for dierent wave-
forms the clutter RCS depends on the signal bandwidth and the possible
presence of range side lobes. Surface clutter (especially at low grazing an-
gles) is notably spiky, with peaks having dimensions on the order of 1m, so
a range resolution cell smaller than the target length is of limited value in
target detection.
There are many waveform options that can be matched to the radial di-
mension of a small-RCS target, or to the dimension of dominant scatterers
within the target and noise waveforms have no unique advantage in this re-
spect. For the noise waveforms, the range cells are formed by stepwise delay of
the transmitted reference signal however, the use of a nonperiodic waveform
suppresses the range ambiguity. The Doppler cells are created by step-wise
varying the time compression of the reference signal at the correlation oper-
ation. The Doppler resolution depends on the coherent processing interval
and is limited by the same factors for all waveform types and therefore noise
waveforms do not have any particular advantage here as well. Due to the
randomness of the waveform, a noise floor is present in the correlation in-
tegral that limits the possible side lobe suppression. In strong clutter, the
integrated side lobe contribution can be very large and clutter cancelation
methods are required.
Recently, there has been much interest in multiple-input multiple-output
(MIMO) radar systems. MIMO techniques have been well studied in commu-
nications oering advantages where multipath environments can cause fading.
Radar waveform rejection of multipath requires that the range resolution cell
be smaller than the range dierence between the direct and multipath echoes
(even if range side lobes are absent). The range dierence between the direct
path and the multipath echoes can be expressed as
2hr ht

R
where hr and ht are the radar and target altitudes and R is the range. For a
low-grazing-angle target not resolvable by the antenna beamwidth, the range
dierence can be quite small. Even if the waveform bandwidth were increased
to make the resolution smaller than this range dierence, the scatterers in
the initial range cell of the target would have to be large enough to provide
a signal for tracking, before the multipath catches up with the direct signal
and generates a tracking error. Whether a noise waveform or other type of
waveform is used, resolution of this sort has not provided significant tracking
advantages in any known system.
The majority of MIMO radar configurations have focused on multistatic
arrays that have sucient spatial separation to decorrelate the targets radar
cross section scintillation. These networks combine the received data non-
coherently to average out the scintillations. Another form of MIMO radar
uses multiple orthogonally coded waveforms from individual transmitter ele-
ments of a phased array which are then combined coherently upon receive to
form multiple beams [23]. Recently, the extension of noise radar to MIMO
configurations has been explored. Two transmission techniques are described
and include an element-space and beam-space approach [24]. In the element-
space approach, K channels of independent (noncoherent) noise are transmit-
ted separately by K omnidirectional antennas. In the beam-space approach,
each independent noise source is fed into each antenna but is either delayed or
phase shifted so as to form a beam illuminating a selected sector of the radar
systems field of view-eectively coding each sector according to a particular
noise source. The direction of each noise sector is determined by the phase
shifts and the sector width is determined by the beamwidth of the array.
7.2 Ultrawideband Considerations

The combination of RNR and ultrawideband (UWB) technology can give sig-
nificant benefits and overcome inherent drawbacks of narrowband radar. A
few definitions concerning UWB waveforms are given below. Consider for ex-
ample a wideband RNR signal with bandwidth spanning fmin to fmax . The
first definition is for the absolute bandwidth B,
Definition 7.1
The absolute bandwidth B defines the width of the frequency

interval occupied by the signals spectrum and is the dierence
between the maximum frequency and the minimum frequency or
B = fmax fmin (7.1)
The term ultrawideband (UWB) refers to waveforms that have an

instantaneous fractional bandwidth greater than 0.25 (25%) with
respect to the center frequency.
Definition 7.2
The instantaneous fractional bandwidth f is defined as the ratio

of the absolute bandwidth (fmax fmin ) to the mean frequency
(fmin + fmax )/2. Thus the fractional bandwidth is defined as [25]
} ]
2(fmax fmin ) 100%
f = 100% = (7.2)
fmin + fmax 0.5 + fmin /B
The maximum fractional bandwidth is 200% and is reached if fmin = 0. Note

that this value is not fmax dependent. A small fractional bandwidth f indi-
cates that the behavior of the radar system devices being used are not likely to
change much within the absolute bandwidth B. With a large f , the device
behavior may show changes throughout B because the device characteristics
may be dierent across the large number of transmit frequencies. That is,
it is more dicult to build devices which cover a large bandwidth with the
same eciency or properties.
Another important term used to evaluate UWB RNR signal bandwidth
when the waveform is produced by random binary phase shifts is the spread-
ing ratio (or processing gain) [26].
Definition 7.3
The spreading ratio or processing gain of a random binary phase

shift keying signal is defined as
T
P GR = = Nc (7.3)
tb
where T is the code period, tb is the subcode period and Nc is the
number of subcodes within a period.
The motivation to use UWB random noise emitters comes from the need
to have fine range resolution and range measurement accuracy. In contrast
to conventional narrowband systems, the UWB radar obtains much more
information about the material properties and the structure of scanned tar-
gets. In the case of high SNR and precise calibration, range accuracy may
even approach millimeter wave (mm) and submillimeter levels. It also has
an enhanced clutter suppression capability which could someday lead to a
solution for the dicult task of detecting a sea-skimming antiship capable
missile. Foliage-, wall- and ground- penetrating detection and imaging also
benefit. For through-the-wall UWB radar, 1.99 GHz10.6 GHz are allowed in
the United States (FCC) and 30 MHz18 GHz frequency range is envisioned
for wall and ground penetrating radar in Europe [25].
An important aspect of the UWB radar in addition to its LPI and LPID
characteristics are its immunity from electromagnetic radiation eects which
enable frequency spectrum sharing and a significant immunity from deceptive
jamming. Note that noise jamming has the same eect on radar systems that
use noise waveforms as it does on systems that use of any other type of
waveforms of similar bandwidth. On the other hand, there are worries in the
community about the influence of the UWB RNR on small signal receiver
devices such as GPS, cell phone and wireless LAN communications [27].
Another major problem in the radar application of UWB noise waveforms
is that they exhibit the thumbtack ambiguity function and require the use
of a correlator that covers many range-Doppler cells to detect targets whose
range and velocity are unknown. The techniques examined in this chap-
ter describe systems in which the correlator covers a single cell, requiring
two-dimensional sequential search to detect targets with unknown position.
Although useful for some geophysical applications, the technology is currently
inadequate for most military radar applications. Detection of buried mines
and tracking a target after it has been detected by other conventional wave-
forms are examples of current military applications. The processing through-
put for parallel coverage of large regions in range-velocity space is a major
problem that must still be addressed.
7.3 Principles of Random Noise Radars

The ability to simultaneously measure range and Doppler is important for
target detection and imaging and requires that the radar be phase coherent
[28]. This simultaneous measurement would be especially advantageous if it
were implemented in a real-time range-Doppler processor covering a useful
field in both coordinates. Today no noise radars have this capability, and this
absence currently constitutes a major disadvantage being addressed.
Although RNR is by definition, totally incoherent, it is possible to inject
phase coherence into the radar using the technique of heterodyne correlation.
Figure 7.1 shows the main components of a RNR system. A microwave CW
noise signal S(t) is transmitted, and the return signal from the targets re-
flectivity, delayed by td , is received and coherently detected. A copy of the
transmitted noise, delayed by Tr (RF delay line), is used as a reference sig-
nal that is cross-correlated with the received signal. The amount of delay of
the reference signal is a measure of the targets distance. The down range
performance of a random noise radar depends mainly on the cross-correlation
function in the radar receiver. When Tr is varied, a strong correlation peak
is obtained for Tr = td , which gives an estimate of the targets range as
R = ctd /2. Doppler processing can also be performed and the output of the
Figure 7.1: Main components of a noise radar using a delay line. (After [28].)
Doppler filters following the correlator can be used to calculate the velocity
of the target.
Consider the detection of a point target. Following the development by
Axelsson, the transmitted noise signal can be modeled as a stationary process
in complex form as [29, 30]
S(t) = [X(t) + jY (t)] exp(j2fc t) (7.4)
where 2fc is the frequency of the carrier and X(t) and Y (t) are stationary
Gaussian processes with zero means and bandwidth B. For moving targets,
the received signal is compressed or expanded in time as a result of the relative
velocity between the point target and the noise radar. For a point target
positioned at R with a relative velocity v, the received signal is S(t td ),
where as before td = 2R/c. Also = (c v)/(c + v) 1 2v/c when v c.
Cross-correlation of S(t td ) with the reference signal S(r t Tr ), which is
delayed by Tr and time compressed by r = 1 2vr /c, becomes proportional
to [29]
8 Tint
C(td , ; Tr , r ) = w(t)S(t td )S (r t Tr )dt (7.5)
0
where Tint is the measurement time and a window function w(t) is included in
the correlation integral to improve the Doppler side lobe suppression. After
insertion of (7.4) into (7.5)

8 Tint
C(td , ; Tr , r ) = w(t) [X(t td ) + jY (t td )] (7.6)
0
[X(r t Tr ) jY (r t Tr )]
exp [jc (td Tr ) + j( r )c t] dt
The correlation output is close to its mean value when the time-bandwidth
product BTint is large. The average of (7.6) is found with = r and
T = td Tr
8 Tint
Rc (T, ) = 2 exp(jc T ) w(t) [RX (t T ) (7.7)
0
jRXY (t T )] exp [jc t] dt
where RX () and RXY () are the autocorrelation and cross-correlation func-

tions of X(t) and Y (t). For a symmetric noise power spectrum about the
carrier frequency, the cross-correlation term can be neglected and
8 Tint
Rc (T, ) = 2 exp(jc T ) w(t)RX (t T ) exp [jc t] dt
0
(7.8)
The parameters Tr and r are varied until the maximum is found, represented
by r0 and Tr0 , from which range and velocity are estimated: R = cTr0 /2
and v = c(1 r0 )/2. For narrowband noise processes (B fc ), the phase
term of (7.7) and (7.8) generates the dominant decorrelation. Hence, c t
should be kept small over the integration time to avoid a degradation of the
correlation peak [30].
As in previous LPI waveforms, the range resolution depends upon the
bandwidth B. There is a limiting relationship between the range resolution
R = c/(2B), the relative velocity of the target, v, and the available cor-
relation time Tint . Note that this is similar to the FMCW range-Doppler
cross coupling eect discussed in Chapter 4. The time taken for the target to
pass through a range resolution cell Tp = R/v should be greater than the
measurement time Tint , giving the limitation R/v = c/(2Bv) > Tint . If the
number of statistically independent samples is represented by N = 2BTint ,
an upper limit can be derived as [29]
N = 2BTint < c/v (7.9)
As an example, for v < 300 m/s, N < 106 is required. Longer sequences
can be applied if the delay of the reference signal is made variable and is
adapted to the predicted target velocity. From (7.7) and (7.8), the correlation
peak degrades as a result of the t term in RX (t T ) if Tint =
2|v|Tint /c exceeds the correlation time (1/B) of the noise process. This
gives the requirement 2BTint < c/|v|, which is equivalent to (7.9).
The RNR system shown in Figure 7.1 uses a noise source working at a
microwave frequency. Equally eective is the use of a baseband noise source
followed by the upconversion to a carrier frequency. A digital implementa-
tion could also generate the noise signal and provides the flexibility to include
other noise waveforms. For example, the use of tailored or colored noise wave-
forms can be used and have been shown to enhance target detectability [31].
The receiver can also include a homodyne or heterodyne detection of the
in-phase and quadrature components of the received signal and an I/Q de-
modulator used to generate the correlation response. As in a typical CW
LPI radar, leakage or lack of isolation between transmit and receive antennas
can degrade the receiver sensitivity and can aect long range target detection
performance. One approach to eliminate the leakage is the reflected power
canceler (discussed earlier in Chapter 4). In a bistatic configuration, the noise
radar can also use external transmitters where the correlation is between the
direct wave from one antenna and the target reflection from another antenna.
Another approach that can eliminate the CW leakage between the transmit-
ter and receiver entirely is to use an interrupted CW waveform (long noise
pulses). This technique improves the isolation by using a transmit/receive
(T/R) switch to switch the antenna between transmitter and receiver several
times per transmitted noise waveform. Typically, T/R switches can receive
more than 60 dB of isolation between transmitter and receiver. The system
is no longer a true continuous wave noise radar but under certain conditions
the essential properties of noise radars are preserved.
7.4 Narayanan Random Noise Radar Design

A well published hardware example of a UWB random noise radar is the
system first introduced by Narayanan et al. [32]. Figure 7.2 shows the block
diagram of the system configuration. An UWB Gaussian noise waveform is
transmitted and target detection is accomplished by employing a heterodyne
correlation receiver which cross-correlates the received signal with a time-
delayed and frequency shifted replica of the transmit signal. The transmitter
uses a microwave noise diode OSC1 that is band-limited using a bandpass
filter BPF and amplified using a broadband power amplifier AMP1. The
transmitted signal has a Gaussian amplitude distribution and an average
power output of 0 dBm (1 mW) in the 12 GHz frequency range.
The power is divided by PD1 that splits the transmitted waveform into
two equal in-phase components (the transmitted waveform and the reference
signal). The reference signal is connected to a fiber optic fixed delay line DL1
that is used to set the minimum range to the target [33].1 For example, in a
ground-penetrating system this minimum delay ensures that the correlation
operation is performed only at depths below the air-soil interface [14]. A
programmable delay line DL2 is also used to step through the entire range of
available delays so that various probing depths can be obtained.
The delay line output is mixed with a 160 MHz phase locked oscillator
OSC2 in a lower sideband up converter MXR1. The upconverter output
(0.841.84 GHz) feeds the mixer MXR2 that receives the 12 GHz return sig-
nal. The mixer MXR2 output is the 160 MHz correlation output (correlation
coecient) and is filtered in a 160 MHz BPF of 5 MHz and then fed to the I
and Q detector which is also fed by the 160 MHz oscillator OSC2 [34]. The
output of the I and Q detector is then sampled, integrated and the envelope
of the signal is extracted.
Worthy of mention in Figure 7.2 are a few points on the Doppler resolution.
The Doppler return from the slow-moving target will show up at roughly
50 Hz about the carrier frequency, when using fc = 1.5 GHz. Seeking to
isolate and keep this Doppler information, the model uses a low-pass filter
at 100 MHz. This filtering also rejects the undesirable harmonics of the 160
MHz LO frequency. Although not shown the system also houses a second
receive chain that may be used for fully polarimetric measurements or spaced
receiver interferometry [32]. A polarimetric noise radar system measures the
complex scattering matrix of a target [S] given by
} ]
SV V SV H
[S] = (7.10)
SHV SHH
where Sij represents the target scattering coecient for transmit and receive
polarizations i and j respectively. Subscripts V and H stand for vertical and
horizontal respectively. Reciprocity implies that SV H = SHV . The measured
scattering coecients of obscured targets are multiplied by the product TA TB
where Tk is the one-way complex transmission coecient through the medium
for the polarization k. The transmit polarization alternately switches between
V and H while the receiver processes both polarizations simultaneously [17].
Simulation and field test results have been shown to demonstrate the potential
of combining a UWB waveform with coherent processing for high-resolution
subsurface imaging.
7.4.1 Operating Characteristics

The system shown in Figure 7.2 can achieve a significant probability of detec-
tion Pd with arbitrarily small probability of false alarm Pf a . Taking the real
1 Erbium-doped fiber amplifiers can also been used as an ultrawideband microwave noise
source. Based on their amplified spontaneous emission characteristics, the low coherence
output light can have an absolute bandwidth approaching B = 1,200 GHz [33].
Figure 7.2: Block diagram of a random noise radar (from [34]).
part of (7.4), the transmitted Gaussian noise signal centered at 2fc , with a
bandwidth B (bandpass process) can be expressed as
St (t) = Re{S(t)} = X(t) cos(2fc t) Y (t) sin(2fc t) (7.11)
where
B
2fc > (7.12)
2
The received signal back at the radar can be expressed as [35]
Sr (t) = X (t) cos{c [(1 + )t td ]} Y (t) sin{c [(1 + )t td ]} (7.13)
where X (t) and Y (t) are given by
X (t) = AX[(1 + )t td ] (7.14)
and
Y (t) = AY [(1 + )t td ] (7.15)
where A2 is the power reflection coecient that is related to the targets
range, radar antenna gain, and target geometry. The delayed reference signal
Sd (t) at the output of the up converter MXR1 can be written as

} ]
X(t Tr )
Sd (t) = cos{(c IF )t c Tr } (7.16)
2
} ]
Y (t Tr )
sin{(c IF )t c Tr }
2
where IF is the intermediate frequency (IF) and Tr is the delay provided by
the delay line.
The cross-correlator MXR2 extracts the normalized complex correlation
coecient ej between the input signals Sr (t) and Sd (t) as
$T
Sr (t)Sd (t)dt
ej = lim 0
=1/2 (7.17)
T $ T $T
2 2
|Sr (t)| dt 0 |Sd (t)| dt
0
or $T
Xr (t)Xd (t)dt
ej = lim 0
(7.18)
T 1 2
where 1 is the total received power (signal + noise) and 2 is the total
power in the delayed replica. In practice, T cannot go to infinity and one
has to use the short-time correlation function over a finite time Tint . The
maximum value of the correlation coecient occurs when the received and
delayed reference signal are completely correlated.
The data acquisition, storage and analysis performs an integration of the
I and Q channels and then a square law detector
D i1/2
Z = I 2 + Q2 (7.19)
is used to compute the response. Expressions for the characteristic function

and joint PDF for the integrator output and the PDF and CDF for the detec-
tor output are given in [35]. For the sum of a large number of uncorrelated and
independent samples N with no target present, the envelope of the detector
output is approximately Rayleigh-distributed
w W
8Z 4Z 2
PN (Z) exp (7.20)
N (1 2 )2 N (1 2 )2
The probability of false alarm Pf a for a threshold Th can then be found by
integration of (7.19) w W
4Th2
Pf a exp (7.21)
N (1 2 )2
For large N , the threshold Th is related to the Pf a by
1 2
Th [N ln Pf a ]1/2 (7.22)
2
No closed form expressions are available for the probability of detection Pd

and numerical results for the Pf a vs. Pd are shown in [35] for N = 1, 25, 50
and 100. This maximum value of the correlation coecient is related to the
input signal to noise ratio SNRi as
} ]1/2
SNRi
max (7.23)
1 + SNRi
2 2 2
and SNRi = sr /n1 where sr = A2 s2 is the power in the received signal;
s2 is the transmitted signal power and A2 is the power reflection coecient
that is related to the antenna gain, targets range and geometry. The term
2
n1 is the receiver noise power.
Depending on the interference that corrupts the return signal, the corre-
lation coecient can have a value anywhere from 0 to 1.
7.4.2 Model of RNR Transmitter

The RNR transmitter uses a random white Gaussian microwave noise source
that is band-limited and amplified. To evaluate the characteristics of the
transmitted waveform, Figure 7.3 shows a block diagram of the transmitter
configuration used in this chapter to evaluate the PACF and PAF character-
istics [36]. The transmitter is modeled in MATLAB and is part of the LPIT
(contained on the CD). With inputs consisting of the carrier frequency fc ,
bandwidth B, amplitude A, and noise power level, the model produces the
in-phase and quadrature components of the waveform as it would appear at
the output of a noncooperative intercept receiver with an ADC sampling at
3 GS/s.
Figure 7.4 illustrates the wideband white Gaussian noise magnitude that
is generated prior to bandlimiting. A key performance element is the trans-
mitted noise bandwidth. For good range resolution, clutter discrimination,
and LPI characteristics, the noise signal needs to spread the energy over a
large modulation bandwidth. Without loss of generality, our example mod-
els the transmitted noise to have an absolute bandwidth of B = 300 MHz
(200 MHz B 500 MHz) giving a range resolution of 0.5 m. Figure 7.5
shows the white Gaussian noise after bandlimiting. The carrier frequency is
fc = 350 MHz. Note from (7.2) that this model demonstrates a UWB wave-
form with a fractional bandwidth of f = 85%.
7.4.3 Periodic Ambiguity Results

To examine the periodic ambiguity side lobe characteristics, Figure 7.6 shows
the ACF and the PACF for the CW noise signal shown in Figure 7.5. The
number of transmitted code periods used in the correlation receiver is N = 1.
The delay oset axis is normalized by the subcode period tb which in this case
Figure 7.3: Random noise radar transmitter model (from [36]).
Figure 7.4: Wideband microwave noise signal.

Figure 7.5: Band-limited microwave CW noise signal.
Figure 7.6: Random noise CW radar autocorrelation function and periodic

autocorrelation function with N = 1.
Figure 7.7: Random noise CW radar periodic ambiguity function with N = 1.
is the sampling period (1/fs ). With an integration period of

Tint = 2 s, the PACF mainlobe repeats every 6,000 samples. With
N = 1 copies of the reference in the correlation receiver the peak side lobe
level revealed after close examination of the ACF and PACF is 3 dB at
3 normalized delay osets. The side lobe level however drops o quickly to
15 dB at 8 osets. Figure 7.7 shows the PAF for the transmitted waveform.
The zero Doppler cut of the PAF is the PACF and the mainlobe repeats at
every integer multiple of the code length (6,000 samples). Note the relatively
large Doppler spread and high Doppler side lobes in the PAF at each code
period. Dawood et al. [28] examine expressions for the generalized ambiguity
function and show that the UWB RNR waveforms are not suitable for unam-
biguous range rate estimation due to this extended Doppler-spread parameter
(product of transmit bandwidth and range rate) unless the cross-correlator is
matched in the delay rate.
7.5 Random Noise Plus FMCW Radar

The random noise plus FMCW radar (RNFR) was first introduced by Liu
et al. in 1984 [37]. In the RNFR, a white Gaussian noise source is linearly
frequency modulated by an FMCW waveform. Figure 7.8 shows the radar
Figure 7.8: Block diagram of random noise plus FMCW radar system (from
[38]).
system and illustrates both the transmit and the receive functions. A portion
of the transmitted signal is used as a local oscillator input to the receivers
front-end mixer where the correlation between target echo and transmitted
signal takes place [38]. The mixer output is a beat frequency that represents
the targets range. The output spectrum of the mixer is a single Doppler
frequency for a zero-range target and becomes gradually larger with the tar-
get range increasing. Following the mixer is an amplifier and two bandpass
filters [9]. Optimized filters pass either the targets Doppler signal with some
noncorrelation signal or strictly the noncorrelation signal. The power detec-
tors detect the signal envelope and a dierence amplifier selects the correct
channel to determine the targets range from the measured power dierence
output. The emitter design has good electronic protection capability and also
good resolution for precise, simultaneous distance and velocity measurements
[79]. Its low mean power and noise-like characteristics result in a very low
SNR in the intercept receiver that does not have access to the noise waveform
to compress the signal. The RNFR has good distance measurement capabil-
ity but the CW leakage makes it dicult to measure target speed and detect
long range targets [9].
7.5.1 RNFR Spectrum

The RNFR transmitter uses a white Gaussian noise source that is linearly
frequency modulated by a triangular FMCW waveform. The transmitted
noise plus FMCW signal is
e(t) = E cos[c t + (t)] (7.24)

where 8 t
(t) = Df (t1 )dt1 (7.25)
0
and E is the amplitude and Df is the angular frequency per volt of the
FMCW and (t1 ) is the noise voltage of a stationary process with zero mean.
Following the development given in [38], the power spectrum of e(t) is
} ]
2 1 ( c )2
We (f ) = E exp (7.26)
2F 2F 2
where F is the angular frequency deviation of the transmitted noise signal.

The echo of the moving target is
eR (t) = ER cos[(c + d )(t td ) + (t td )] (7.27)
where ER is the amplitude of the echo, td is the delay between the echo and
the transmitted signal and fd = d /2 is the Doppler frequency. At the mixer
output
V (t) = A cos[d (t td ) c td + (t td ) (t)] (7.28)
The correlation function at the mixer output is then given approximately
by [38] } 2 ]
A2 1 (td )t2
Rv (t) = exp cos(d t) (7.29)
2 2
and 12 (td ) = K 2 t2d and assumes that K is a normalizing constant, and
2 (td ) = 12 (td )/42 . The Fourier transform of Rv (t) is
} ]
1 f2
SR1 (f ) = exp 2 (7.30)
2(td ) 2 (td )
Assuming that (td ) >> fd
A2 A2 A2
Sv (f ) = SR1 (f fd ) + SR1 (f + fd ) = SR1 (f ) (7.31)
4 4 2
and the spectrum after correlation by the mixer is approximately
} ]
A2 1 f2
Sv (f ) = exp 2 (7.32)
2 2(td ) 2 (td )
for f > 0. Note the spectrum behavior as a function of td .

Figure 7.9: Model of random noise plus FMCW transmitter (from [36]).
7.5.2 Model of RNFR Transmitter

A block diagram of the RNFR transmitter is shown in Figure 7.9 [36]. A
MATLAB model of the RNFR is also contained in the LPIToolbox. Indi-
cated on the diagram in parentheses are the places where intermediate results
of the model are shown below. A wideband microwave noise generator first
produces a signal that is band-limited to 300 MHz (200 f 500 MHz)
centered at fc = 350 MHz. After bandlimiting, the noise signal modulates
a triangular FMCW signal with a modulation bandwidth F = 300 MHz
and modulation period tm = 1s. The magnitude of the FMCW signal in
Figure 7.9 (1) with F = 300 MHz, and fc = 350 MHz are shown in
Figure 7.10. The resultant signal is then high-pass-filtered to remove the
lower sideband modulation leaving the transmitted signal with B = 600 MHz.
The final noise modulated FMCW signal with F = 300 MHz, and fc = 350
MHz is shown in Figure 7.11. In Figure 7.12 the magnitude of the noise
FMCW high-pass filter output signal in Figure 7.9 (3) is shown (transmitted
waveform) with F = 300 MHz and fc = 350 MHz.

Using the same signal duration as in Section 7.4, the ACF and the PACF
of the RNFR waveform are shown in Figure 7.13. For the RNFR waveform
the peak side lobes occur at approximately 21 dB down from the main lobe
showing the advantages of the FMCW modulation over a strictly random
noise modulation. To examine the side lobe performance in the delay-Doppler
Figure 7.10: Magnitude of the FMCW signal shown in Figure 7.9 (1) with
F = 300 MHz, and fc = 350 MHz.
Figure 7.11: Magnitude of the noise modulated FMCW signal in Figure 7.9
(2) with F = 300 MHz, and fc = 350 MHz.
Figure 7.12: Magnitude of the noise FMCW high-pass filter output signal in
Figure 7.9 (3) with F = 300 MHz, and fc = 350 MHz.
oset domain the PAF is then calculated and is shown in Figure 7.14. The
PAF main lobe repeats every code period or 6,000 samples. Note that the
zero delay Doppler side lobes are also less compared to the RNR results. As a
final note for comparison, closed form expressions for the average ambiguity
function for the RNFR waveform are given in [38].
7.6 Random Noise FMCW Plus Sine

The random noise FMCW plus sine radar (RNFSR) uses an additional sine
signal at a frequency of fm that modulates the noise source [8]. The com-
posite signal is then modulated by the FMCW waveform. The additional
sine signal that is added helps minimize the leakage from the transmitter to
the receiver as explained below. A block diagram of the RNFSR is shown in
Figure 7.15. The receiver is similar to the RNFR receiver with the received
signal correlated to a time-delayed version of the emitted waveform using a
mixer. The mixer output contains the noise and all the sinusoidal harmonics.
The amplifier is a wideband automatic gain control circuit with a large dy-
namic range [7]. To account for the injected sine wave, the receiver uses two
bandpass filters to expand the receivers frequency range for the Doppler plus
noncorrelation signal and for the noncorrelation signal only. The transmitted
Figure 7.13: Random noise plus FMCW autocorrelation function and periodic
autocorrelation function.
Figure 7.14: Random noise plus FMCW periodic ambiguity function.

signal for the RNFSR is [38]
ec (t) = E cos[c t + 1 (t) + 2 (t)] (7.33)
where 8 t
1 (t) = F cos(m t1 )dt1 = D1 sin(m t) (7.34)
0
with
F
D1 = (7.35)
m
and 8 t
2 (t) = Df V (t2 )dt2 (7.36)
0
where the variable Df is the angular frequency per volt and V (t2 ) is the
modulated noise voltage of a normal stationary process with zero mean and
m is the additional tone frequency. As expected, the derivation of the mixer
output spectrum for the RNFSR is considerably more complicated and the
reader is referred to [7].
The major dierence between the RNFR and the RNFSR are the filters
shown in Figure 7.15. The filters have bandwidths expanded to include mul-
tiples of the added sine signal within the return signal. The bandwidth of the
first bandpass filter is selected to be B1 = nfm + fdmin and the bandwidth
of the second bandpass filter is selected to be B2 = nfm + fdmax where n
represents the nth harmonic of the added sine waveform that is picked up
by the radar. This technique takes advantage of the harmonic characteris-
tics of the sinusoidal signal to eliminate the CW leakage at and close to zero
range [79]. This zero range hole characteristic does not let the radar respond
to close-in targets and only produces an output detection when a target is
present between the two chosen ranges determined by the filter bandwidths.
Even with the addition of the sinusoidal modulation, the RNFSR cannot mea-
sure the speed of a moving target or detect a long-range target (similar to
the RNFR). The RNFSR is suitable for short-range LPI applications such as
harbor control, missile fuse systems and UAV landing systems.
7.6.1 Model of RNFSR Transmitter

The RNFSR noise technology employs an additional tone signal that is mod-
ulated by the white Gaussian noise, which further modulates the FMCW
waveform. Figure 7.16 shows a block diagram of the transmitter model
used in this chapter [36]. A microwave noise generator produces a white
Gaussian noise waveform. After this, the signal is band-limited to 300 MHz,
centered on fc = 350 MHz. After band-limiting, the noise signal is added
(added in frequency, multiplied in time) to a single tone with a frequency of
fm = 350 MHz. The upper band is centered at 700 MHz (from the modulation
Figure 7.15: Random noise FMCW plus sine radar block diagram (from [9]).
product) and this new signal modulates an FMCW signal with a modulation
bandwidth of F = 300 MHz. After noise modulation, the resultant signal is
low-pass-filtered to remove the upper sideband modulation products. Finally,
the 600-MHz bandwidth signal with center frequency of 350 MHz is amplified
before transmission.
To examine the model development, Figure 7.17 shows the noise source
with the added tone modulation fm = 350 MHz. Note the frequency shift
of the waveform to a center frequency of 700 MHz. Figure 7.18 shows
the magnitude spectrum of the FMCW signal that is used to modulate
the noise with the added tone. The signal has a modulation bandwidth of
F = 300 MHz. In Figure 7.19, the modulation of the sine plus noise by the
FMCW signal is shown. Shown are the resulting upper and lower sideband
products. The upper sideband is not needed and eliminated. Figure 7.20
shows the magnitude spectrum of the output waveform.

For the RNFSR transmitted waveform, a sine wave is added to the noise plus
FMCW modulation. To compare the side lobe performance of this waveform
with the other noise modulations, the autocorrelation function and periodic
autocorrelation function are shown in Figure 7.21 for a 2-s period of the
waveform. In this case the peak side lobe level is approximately 21 dB.
Note that the side lobe structure is nearly the same as the RNFR waveform
and the peak side lobe is not the first side lobe. The periodic ambiguity
function is shown in Figure 7.22. Of interest here is the lower Doppler side
lobe level and the smaller extent of the Doppler side lobes.
Figure 7.16: Random noise FMCW plus sine transmitter model (from [36]).
Figure 7.17: Sine plus random noise FMCW model in Figure 7.16 showing
(1) the magnitude spectrum of the tone modulation of noise.
(2) the magnitude spectrum of FMCW signal.
Figure 7.19: Sine plus random noise FMCW model in Figure 7.16 showing (3)
the magnitude spectrum of noise plus sine after modulation by the FMCW
signal.
(4) the magnitude spectrum of output waveform after eliminating the upper
sideband.
Figure 7.21: Sine plus random noise FMCW autocorrelation and periodic
autocorrelation function.
Figure 7.22: Sine plus random noise FMCW periodic ambiguity function.
7.7 Random Binary Phase Modulation

The random binary phase code (RBPC) emitter is also a correlation CW
noise radar that uses a random phase modulation of a carrier frequency to
achieve LPI noise characteristics. The range resolution of a RBPC CW emit-
ter depends on the width of the subcode as
ctb
R = (7.37)
2
and the maximum range performance is
Nc ctb
Rmax = (7.38)
2
The Doppler tolerance depends on the length of the pulse compressor as
1
fd max = (7.39)
2Nc tb
where Nc is the number of phase codes and tb is the subcode width. Note that
simultaneously extending the Doppler tolerance and the range performance
is not possible since they are inverse relationships.
To improve the performance of the RBPC emitter, several pulse compres-
sors can be used in series. A block diagram of this RBPC emitter is shown in
Figure 7.23: Random binary phase modulation radar (from [39]).
Figure 7.23 [39]. The transmitted microwave signal is phase-modulated by a

random binary phase code (0 or ), which can be generated for example by
Bernoulli trials. The received target echo signal is detected at zero IF, am-
plified and digitized by an ADC. To cross-correlate the echo signal and the
delayed reference, a series of M parallel pulse compressors (transverse filters)
PC1, PC2, ... PC(M) are used. The pulse compressor outputs are rearranged
and range side lobe suppression techniques are used to limit the peak side
lobes to PSL < 30 dB [5]. The CFAR and threshold detector then give the
target range and velocity information.
The maximum distance performance is dependent on the total length of
the pulse compressor group as
M Nc ctb
Rmax = (7.40)
2
The maximum Doppler frequency that is measurable is set by the length of
a single pulse compressor. Consequently, the RBPC radar can detect long
range targets and high-speed targets simultaneously. That is, a reduction of
the length Nc of the pulse compressor can easily extend the Doppler toler-
ance while an increase in the number of pulse compressors M in the receiver
can satisfy a needed performance distance. The RBPC emitter also has good
Doppler sensitivity and overcomes the limitations of target velocity measure-
ment and long-range detection. The problem of CW leakage however, is still
Figure 7.24: Block diagram of a random binary phase code radar (from [36]).
present.
7.7.1 Model of RBPC Transmitter

A block diagram of the RBPC transmitter is shown in Figure 7.24 [36]. For
the example shown, the CW tone fc = 900 MHz and the phase change for each
subcode is randomly selected as either 0 or . The number of carrier cycles
per subcode cpp = 3 (B = 300 MHz) with number of subcodes Nc = 600. This
allows comparisons to be made with the previous noise radar configurations.

To compare the peak side lobe performance of the RBPC waveform,
fc = 900 MHz, cpp = 3, fs = 3 GS/s and Nc = 600. This results in
bsc = 6,000 samples being processed within the code period. The ACF and
PACF are shown in Figure 7.25. The peak side lobe level is approximately
19 dB. The PAF is shown in Figure 7.26. Note that although the peak range
side lobe is a bit higher, the Doppler side lobe performance is considerably
superior in that its extent is not as great.
Figure 7.25: Random binary phase code autocorrelation and periodic auto-
correlation function.
Figure 7.26: Random binary phase code periodic ambiguity function.

7.8 Millimeter Wave Noise Radar

Millimeter wave (MMW) signals are most appropriate for applications such as
environment monitoring, remote sensing, short range target detection, vehicle
collision warning and automatic landing systems [40]. Several MMW solid-
state noise transmitters have been investigated. These systems rely on the
use of a chaotic waveform generator as a source of CW noise and a digital-
analog correlator with an electronically controllable delay line as the main
part of the correlation receiver. Three MMW noise sources were investigated.
Dynamical chaotization2 was undertaken for: (1) microwave oscillations in
a waveguide multiresonant system (25 resonant frequencies) containing one
or more Gunn-diodes that can couple two or more modes, (2) microwave
oscillations in a microstrip ring or linear resonator, and (3) angle modulation
of a VCO signal by an RF noise signal [41].
Ka- and W-band solid state noise generators have also been developed
on the principles of chaotization of nonlinear systems using both IMPATT-
and Gunn-diodes. Power outputs on the order of 300 mW with bandwidths
ranging from 30 MHz to 300 MHz were generated. Special designs have
also been tested oering bandwidths up to 2 GHz with an output power of
40 maw. To process the received signal from the target, fast digital real-time
correlators were developed with clock frequencies on the order of 500 MHz.
The use of ADCs were avoided by using a simple two-level quantization of
the reference signal.
7.9 Correlation Receiver Techniques

The received noise signal reflected from the target is the delayed version of
the transmitted signal. By measuring the delay, the receiver can determine
the location of the target. To compute the position of the target, the receiver
computes the cross-correlation between the time delayed, conjugated transmit
signal and the received signal. The analog cross-correlation function is defined
as in (7.5). The correlation coecient or normalized correlation function is
defined as in (7.17). Signals are said to be correlated or alike to the extent that
their correlation coecient approaches unity. Below we examine the dierent
approaches to implementing the cross-correlation function in the receiver.
It can be shown that the received target response is a convolution between
the target reflectivity profile (t) and the ACF of the radar transmit waveform
p( ) [42]
g( ) = (t) p( ) (7.41)
where 8 T /2
p( ) = lim S(t)S (t )dt (7.42)
T T /2
2 Chaotization is the theoretical and experimental study of chaos in dynamical systems.
If there is a point scatterer at a distance R, with amplitude a0 and initial

phase 0 , the target reflectivity function can be expressed as
(t) = a0 ej0 (t td ) (7.43)
where denotes the Dirac impulse function.
With the radar transmit waveform a band limited Gaussian noise signal
with carrier frequency fc , bandwidth B, and a square spectrum with power
2 , the ACF is given by
p( ) = 2 sinc(B )ej2fc (7.44)
That is, for a radar that uses a band-limited Gaussian noise waveform as the
transmit signal, its ACF is a sinc-pulse modulated by the carrier frequency
fc [14].
7.9.1 Ideal Correlation

Figure 7.27 shows the block diagram of an ideal analog correlation receiver.
The receiver consists of a noise source and an ideal time delay line that
produces a delayed copy of the transmit signal. A mixer followed by a lowpass
filter performs the correlation integration between the received target returns
and the delayed transmit signals. The response is generated using an I & Q
demodulator. The cross-correlation function for the ideal analog correlation
receiver has been shown to be [42]
CI ( ) = a0 2 sinc(B )e(j2fc 0 ) (7.45)
and preserves both the target amplitude and the initial phase.
7.9.2 Digital-Analog Correlation

A digital-analog correlation receiver is shown below in Figure 7.28 [42]. In
this architecture, the delay line is implemented with a digital radio frequency
memory (DRFM) device. These devices use high-speed sampling and fast dig-
ital memory for storing and replicating the transmitted RF noise signal. They
provide the ability to capture the transmitted signal and generate a precise,
coherent replica for use in the correlation process. In the DRFM, the signal is
digitized and then multiplexed into high-speed dual-ported memory for stor-
age and delay processing [43]. The delayed signal is then converted back to
an analog signal using a digital-to-analog converter before being sent to the
mixer and lowpass filter. Expressed as a function of CI , the cross-correlation
function of a digital-analog cross-correlation with a 2N level ADC is [42]
N
a0 3 i2 22
CDA ( ) = e 2 CI ( ) (7.46)
i=0
where is the quantization step size.
Figure 7.27: Ideal analog correlation receiver (from [42]).
Figure 7.28: Digital-analog correlation receiver (after [42]).

Figure 7.29: Fully digital correlation receiver. (After [42].)
7.9.3 Fully Digital Correlation

A fully digital cross-correlation receiver can also be used. In this configura-
tion, both the received target echo signal and the copy of the transmit signal
are digitized and the cross-correlation is done digitally. A block diagram of
this approach is shown in Figure 7.29.
To avoid the use of ADCs, the principle of two level quantization of the
reference signal has been proposed. Fast correlator hardware can be developed
since only the monobit (two-level) version of the noise waveform reference is
delayed in a fast controllable digital delay line made up of shift registers or fast
random access memory. For the case when the delayed transmit signal and
the received signal are both clipped to be either 1 or +1 before performing
correlation, the cross-correlation function has been shown to be [42]
2
CD ( ) = sin1 [sinc(B )] (7.47)

Work on fast digital signal processing algorithms for computation of the
cross-correlation are of high interest recently since this presents the major
computational burden in random noise radar. By increasing the sample
size N being processed, the requirements for the discrete Fourier transform
processors and fast convolution processors which calculate the cross-correlation
increase sharply. Recently a system of orthogonal real-imaginary basis func-
tions representing a new version of Walsh functions were compared to using a
set of Walsh Cooley functions for calculating the correlation [44]. Parametric
and nonparametric algorithms are compared in [45].
7.9.4 Acousto-Optic Correlation

Due to the physical delay lines that are used a limited number of range bins is
available. This either limits the range gate extent (maximum detectable range
scanned by the variable delay line) or the range resolution must be sacrificed.
In addition, the slow switching speed (several seconds) of the variable delay
lines can limit the data acquisition rate and constrain the ability to do real-
time signal processing. This ultimately limits the ability to do range-Doppler
processing to detect fast moving targets.
Acousto-optic (AO) devices are well known for their utility in correlation
processing [46]. The use of an AO processor oers another approach to corre-
lation processing in the random noise radar. The receiver still employs a fiber
optic fixed-delay line to choose the coarse range delay but an acousto-optic
(AO) time-integrating correlator is used to allow parallel range bin process-
ing [47]. Here the received signal is heterodyned using the AO as the time
integrating correlation receiver as shown in Figure 7.30. The noise signal is
transmitted through the transmit antenna and the replica of the transmitted
Figure 7.30: Acousto-optic correlation receiver for noise radar (from [47]).
noise signal is delayed in the fixed fiber delay line (sets minimum detectable
range). After addition of a fixed bias and level adjustments, the signal mod-
ulates the laser diode of the AO correlator. The laser diode light is then
collimated with a lens and focused on the AO device.
The received signal is used to drive a piezoelectric transducer, which
launches a traveling acoustic wave into the AO device. The correlation of the
delayed transmit and received signal is achieved by imaging the AO device
aperture onto a 1-D CCD. The time-integrated correlation signal is produced
by detecting the interference between the undiracted beam and the polar-
ization switched, diracted traveling wave received signal by projecting both
optical beams through a polarizer that can be rotated to optimize the beam
ratio for good interferometric modulation depth.
0 The correlator provides the
amplitude as a function of range as A = I 2 + Q2 . Further details and
experimental results can be found in [47].
The use of the AO crystal provides up to 1,000 range cells that work
simultaneously over the aperture of the crystal providing the capability for
real-time data acquisition. The sensitivity is also improved up to 60 dB.
Dynamic range and linearity are limited by the acoustic nonlinearities and
scattering and by the limited dynamic range of CCD but could be increased by
improved CCD detector arrays providing higher bit resolution (e.g., 16 bits).
There is also a noise figure penalty for the electrical to fiber to electrical
conversion in the fiber delay line and AO correlator. This was oset by the
processing gain achieved by the long integration time.

Today noise waveforms can be generated digitally followed by a digital-to-
analog converter and up-conversion onto a carrier signal. With the increasing
integration of solid-state microwave techniques and high-speed VLSI circuits,
the correlation signal processing required for noise radar is leading the way
to real-time range-Doppler implementation. They are also relatively inex-
pensive to build and there is considerable interest in relating the technology
to military applications such as covert surveillance and reconnaissance, tar-
get detection and tracking, through-the-wall imaging, ground-penetration,
foliage-penetration profiling, synthetic aperture radar and inverse synthetic
aperture imaging.
The use of wideband noise waveforms can result in high resolution and
reduced ambiguities in range and Doppler estimation. The periodic ambi-
guity analysis for the four noise technology emitters are compared in this
Chapter. Table 7.1 summarizes the peak range side lobe level (in dB) and
the peak Doppler side lobe level of the noise radar configurations simulated
in this chapter. These results do not include any side lobe suppression tech-
niques which can lower these values significantly. The use of noise waveforms
can result in a large mismatch in processing gain between the radar and the
noncooperative intercept receiver making their presence hard to detect. The
most significant consequence on the traditional intercept receiver is a slight
increase in the receivers noise floor. Finally, the use of several noise wave-
forms in a netted radar configuration can minimize the mutual interference
between emitters while providing an increase in surveillance volume and also
lowering the CW power required even further (see Chapter 9).
Table 7.1: Summary of Ambiguity Peak Side Lobe Performance

Noise Technique Range Side Lobe Doppler Side Lobe
RNR (Figures 7.6 and 7.7) 3 dB 6
RNFR (Figures 7.13 and 7.14) 21 dB 6
RNFSR (Figures 7.21 and 7.22) 21 dB 7
RBPC (Figures 7.25 and 7.26) 19 dB 2
References
[1] Horton, B. M., Noise-modulated distance measuring system, Proceedings
of the IRE, Vol. 49, No. 5, pp. 821828, May 1959.
[2] Craig, S. E., Fishbein, W. and Rittenbach, O. E., Continuous-wave radar
with high range resolution and unambiguous velocity determination, IRE
Transactions MIL-6, pp. 153, April 1962.
[3] Grant, M. P., Cooper G. R., and Kamal, A. K., A class of noise radar
systems, Proceedings of the IEEE Vol. 51, No. 7, pp. 10601061, July, 1963.
[4] Cooper, G. R., and McGillem, C. D., Random Signal Radar, Final Report
NASA Grant-NSG 543, Purdue University, Lafayette, IN, June 1967.
[5] McGillem, C. D., Cooper, G. R., and Waltaman, W. B., An experimental
random signal radar, Proceedings of the National Electronics Conference,
Dayton, Ohio, pp. 409411, Oct. 1967.
[6] Smit, J. A., and Kneefel, W. B. S. M., RUDARAn experimental noise
radar, De Ingenieur, Vol. 83, No. 32, ppm. 99-110, Aug. 1971.
[7] Guosui, L., Hong, G., and Weimin, S., Hongbo, S., Jianhui, Z., Random
Signal RadarA Winner in Both the Military and Civilian Operating Envi-
ronments, IEEE Trans. on Aerospace and Electronic Systems, Vol. 39, No.
2, pp. 489498, April 2003.
[8] Guosui, L., Hong, G., Xiaohua, Z., and Weimin, S., The Present and Future
of Random Signal Radars, IEEE Aerospace and Electronic Systems Maga-
zine, Vol. 12, No. 10, pp. 3540, October 1997.
[9] Guosui, L., Hong, G., and Weimin, S., Development of random signal radars,
IEEE Transactions on Aerospace and Electronic Systems, Vol. 35, No. 3, pp.
770777, July 1999.
[10] Garmatyuk, D. S., and Narayanan, R. M., ECCM Capabilities of an

Ultrawideband Bandlimited Random Noise Imaging Radar, IEEE Trans. on
Aerospace and Electronic Systems, Vol. 38, No. 4, pp. 12431255, October
2002.
[11] Xianyi, Z., Weimin, S., and Hong, G., Anti-jamming performance analysis
for random noise UWB imaging radar, Proceedings of the International Conf.
on Radar (CIE 06), Shanghai, China, 1619 Oct. 2006.
[12] Turner, L., The evolution of featureless waveforms for LPI communications,
Proceedings of the NAECON, Dayton, OH, pp. 13251331, May 1991.
[13] Axelsson, S. R. J., Random noise radar/sodar with ultrawideband wave-
forms, IEEE Trans. on Geoscience and Remote Sensing, Vol. 45, No. 5, pp.
10991114, May 2007.
[14] Narayanan, R. M., and Dawood, M., Radar Penetration Imaging Using
Ultra-wideband Random Noise Waveforms, IEEE Trans. on Antennas and
Propagation, Vol. 48, No. 6, pp. 868878, June 2000.
[15] Lai, C-P., Ruan, Q., and Narayanan, R. M., Hilbert-Huang transform (HHT)
processing of through-wall noise radar data for human activity characteriza-
tion, Proceedings of the IEEE Workshop on Signal Processing Applications
for Public Security and Forensics (SAFE), Washington DC, April 2007.
[16] Xu, X., and Narayanan, R. M., FOPEN SAR Imaging Using UWB Step-
Frequency and Random Noise Waveforms, IEEE Trans. on Aerospace and
Electronic Systems, Vol. 37, No. 4, pp. 12871300, October 2001.
[17] Xu, Y., and Narayanan, R. M., Polarimetric Processing of Coherent Random
Noise Radar Data for Buried Object Detection, IEEE Trans. on Geoscience
and Remote Sensing, Vol. 39, No. 3, pp. 467478, March 2001.
[18] Bell, D. C., and Narayanan, R. M., Inverse synthetic aperture radar imaging
using a coherent ultrawideband random noise radar system, Optical Engi-
neering, Vol. 40, No. 11, pp. 26122623, November 2001.
[19] Garmatyuk, D. S., and Narayanan, R. M., Ultra-Wideband Continuous-
Wave Random Arc-SAR, IEEE Trans. on Geoscience and Remote Sensing,
Vol. 40, No. 12, pp. 25432552, December 2002.
[20] Zhang, Y., and Narayanan, R. M., Design Considerations for a Real-Time
Random-Noise Tracking Radar, IEEE Trans. on Aerospace and Electronic
Systems, Vol. 40, No. 2, pp. 434445, April 2004.
[21] Zhang, Y., and Narayanan, R. M., Monopulse radar based on spatiotemporal
correlation of stochastic signals, IEEE Trans. on Aerospace and Electronic
Systems, Vol. 42, No. 1, pp. 160173, Jan. 2006.
[22] Zhang, Y., Narayanan, R. M., and Xu, X., Theoretical and simulation analy-
sis of random noise monopulse radar, Proceedings of the IEEE Antennas and
Propagation Symposium, San Antonio, TX, pp. 386389, July 2002.
[23] Donnet, B. J., and Longsta, I. D., MIMO radar, techniques and oppor-
tunities, Proceedings of the 3rd European Radar Conference, pp. 112115,
Manchester UK, Sept. 2006.
[24] Gray, D. A., and Fry, R., MIMO noise radar element and beam space com-
parisons, Proceedings of the International Waveform Diversity and Design
Conference, Pisa Italy, pp. 344347, June 2007.
[25] Zetik, R., Sachs, J. and Thoma, R. S., UWB short range radar sensing,
IEEE Instrumentation & Measurement Magazine, Vol. 10, No. 2, pp. 3945,
April 2007.
[26] Sun, H. Possible ultra-wideband radar terminology, IEEE Aerospace and
Electronic Systems Magazine, Vol. 19, No. 8, pp. 38, Aug. 2004.
[27] Sun, H. Lu, Y., and Guosui, L., Ultra-Wideband Technology and Random
Signal Radar: An Ideal Combination, IEEE Aerospace and Electronic Sys-
tems Magazine, Vol. 18, No. 11, pp. 37, November 2003.
[28] Dawood, M., and Narayanan, R. M., Generalized wideband ambiguity func-
tion of a coherent ultrawideband random noise radar, IEE Proceedings-F
Radar, Sonar, Navigation Vol. 150, No. 5, pp. 379386, Oct. 2003.
[29] Axelsson, S. R. J., Noise Radar for Range/Doppler Processing and Digital
Beamforming using Low-bit ADC, IEEE Trans. on Geoscience and Remote
Sensing, Vol. 41, No. 12, pp 27032720, December 2003.
[30] Axelsson, S. R. J., Noise Radar Using Random Phase and Frequency Mod-
ulation, IEEE Trans. on Geoscience and Remote Sensing Vol. 42, No. 11,
pp. 23702384, Nov. 2004.
[31] Narayanan, R. M., Henning, J-A, and Dawood, M., Enhanced detection of
objects obscured by dispersive media using tailored noise waveforms, Pro-
ceedings of the SPIE Conf. on Detection and Remediation Technologies for
Mines and Minelike Targets III, Orlando, FL, Vol. 3392, pp. 604614, April
1998.
[32] Narayanan, R. M., Xu, Y., Homeyer, P. D., and Curtis, J. O., Design per-
formance and applications of a coherent ultra-wideband random noise radar,
Optical Engineering, Vol. 37 No. 6, pp. 18551869, June 1998.
[33] Jiang, R., Wolfe, K. W., and Nguyen, L., Low coherence fiber optics for ran-
dom noise radar, Proceedings of the IEEE Military Communications Con-
ference (MILCOM), Los Angeles, CA, pp. 907911, Oct. 2000.
[34] Li, Z., and Narayanan, R. M., Doppler visibility of coherent ultrawideband
random noise radar systems, IEEE Trans. in Aerospace and Electronic Sys-
tems, Vol. 42, No. 3, pp. 904916, July 2006.
[35] Dawood, M., and Narayanan, R. M., Receiver operating characteristics for
the coherent UWB random noise radar, IEEE Trans. on Aerospace and
Electronic Systems, Vol. 37, No. 2, pp. 586594, April 2001.
[36] Heuschel, E. R., III, Time-Frequency, Bi-Frequency Detection Analysis of
Noise Technology Radar, U. S. Naval Postgraduate School Masters Thesis,
Sept. 2006.
[37] Liu, G., and Xiangquan, S., Average ambiguity function for random FM-
CW radar signal, International Conference on Radar, Paris, pp. 339346,
2124 May 1984.
[38] Guosui, L., Xiangquan, S., Jinhui, L., Guoyu, Y., and Yaoliang, S., Design
of noise FM-CW radar and its implementation, IEE Proceedings-F Radar,
Sonar, Navigation , Vol. 138, No. 5, pp. 420426, October 1991.
[39] Hong, G., Guosui, L., Xiaohua, Z., Weimin, S., and Xi, L., A new kind of
noise radarrandom binary phase coded CW radar, Proceedings of the IEEE
National Radar Conference, Syracuse, NY, pp. 202206, May 1997.
[40] Lukin, K. A., Millimeter wave noise radar technology, Proceedings of the
3rd International Symposium on Physics and Engineering of Microwave, Mil-
limeter Wave and Submillimeter Waves, Kharkov Ukraine, Sept. 1517, 1998.
[41] Lukin, K. A., Millimeter wave noise radar applications: Theory and Exper-
iment, Proceedings of the 4th International Symposium on Physics and Engi-
neering of Microwave, Millimeter Wave and Submillimeter Waves, Kharkov
Ukraine,, June 49, 2001.
[42] Xu, X., and Narayanan, R. M., Impact of dierent correlation receiving tech-
niques on the imaging performance of UWB random noise radar, Proceed-
ings of the IEEE International Geoscience and Remote Sensing Symposium,
IGARSS, Toulouse, France pp. 45254527, 2125 July, 2003.
[43] Pace, P. E., Advanced Techniques for Digital Receivers, Artech House, Nor-
wood, MA, 2000.
[44] Sinitsyn, R. B., and Beletsky, A. J., Fast signal processing algorithms for
noise radar, Proceedings of the 3rd European Radar Conference, Manchester,
U.K., pp. 245248, Sept. 2006.
[45] Yanovsky, F. J., and Sinitsyn, R. B., Ultrawideband signal processing algo-
rithms for radar and sodars, Proceedings of the Ultrawideband and Ultrashort
Impulse Signals, Sevastopol, Ukraine, pp. 6671, 1822 Sept. 2006.
[46] Hecht, D., Characteristics of acoustooptic devices for signal processing,
Proceedings of the IEEE Symposium on Ultrasonics, San Francisco, CA, pp.
369380, 1985.
[47] Narayanan, R. M., Zhou, W., Wagner, K. H., and Kim, S., Acoustooptic
Correlation Processing in Random Noise Radar, IEEE Geoscience and Re-
mote Sensing Letters, Vol. 1, No. 3, pp. 166170, July 2004.
Problems
1. Determine the instantaneous fractional bandwidth of a noise waveform
if the absolute bandwidth B = 3 GHz and the maximum bandwidth
fmax = 4 GHz.
2. Consider a moving target. The instantaneous Doppler frequency is not
a constant but varies due to the varying nature of the instantaneous
wavelength . Since varies between min and max , the Doppler fre-
quencies vary from fdl to fdh . If fdc = (fmin + fmax )/2 is the average
Doppler corresponding to transmit frequency of fc , (a) show that the

minimum and maximum Doppler frequency are
2fmin fdc
fdl = (7.48)
fmin + fmax
and
2fmax fdc
fdh = (7.49)
fmin + fmax
(b) Derive the expression for the targets velocity as a function of fdl
and also as a function of fdh . (c) Calculate fdl , fdh and the targets
velocity if min = 0.15m and max = 0.3m.
3. Show that a transmitted noise signal (7.11)
St (t) = Re{S(t)} = X(t) cos(2fc t) Y (t) sin(2fc t) (7.50)
can be expressed as
1
St (t) = [sc (t)e2fc t + sc (t)e2fc t ] (7.51)
2

where sc (t) = X(t) + jY (t) is the complex conjugate envelope and
denotes the complex conjugate.
4. Using the algorithms in the noise folder, generate the four types of noise
waveforms discussed in the chapter. Examine the ACF, PACF and PAF
of each waveform and compare the highest time side lobe and Doppler
side lobe level (in dB) for each waveform.
Chapter 8
Over-the-Horizon Radar
In this chapter, we examine the sky wave over-the-horizon radar (OTHR)
concept including the characteristics of the ionosphere on the propagation of
the radio waves and the clutter spectrum. Sky wave processing and modern
LPI waveform considerations are presented. The sky wave maximum detec-
tion range is also quantified for the Chinese OTH-B. Simulation results using
PROPLAB PRO, an ionospheric radio propagation tool published by Solar
Terrestrial Dispatch, are shown to demonstrate the coverage region of the
emitter. Surface wave OTHR are also presented including the LPI waveform
considerations. The surface wave radar equation is developed and simula-
tion results of the maximum detection range are shown as a function of the
required input SNR.
8.1 Two Types of OTHR

Today, many countries require an OTHR to provide a long-range, wide area
surveillance capability due to modern day terrorists, smugglers and the need
to monitor ones economic zone and o shore resources [1]. OTHR systems
operate in the high frequency (HF) band 330 MHz and use either surface
wave propagation or sky wave propagation. Sky wave OTHR systems are
installed inland and make use of the ionospheric refraction of the radio waves
several hundred kilometers above the Earths surface to overcome the line-
of-sight limitation caused by the Earths curvature. Surface wave systems
operate in the lower part of the HF spectrum and are installed on the coast-
lines to make use of electromagnetic coupling of the emitted radio waves to
the sea surface allowing propagation to extend over the horizon. OTHR sur-
face wave systems were considered first and operated in the early 1950s with
eective sky wave systems coming along later [2]. Although work on sky wave
249
OTH radars continues around the world, systems that exploit surface wave
propagation are attracting greater long-term interest due to their more con-
venient size and transportability. The significant dierence between sky wave
and surface wave radars is that sky wave radars have large detection ranges
beyond the horizon (starting at about 800 km and extending out to 1,000
4,000 km) while surface wave systems can see out to 400 km. The reason for
the extended detection range for the sky wave propagation is that the losses
caused by the ionization and absorption in the ionosphere are much less than
the surface wave diraction loss. Ionospheric eects such as multipath and
Doppler spreading are also significant. Targets such as cruise missiles, stealth
aircraft, ballistic missiles and aircraft carriers can be detected with OTHR
systems at distances well beyond the horizon.
Several types of emitter waveforms for OTHR have been used in the past.
Transmitted waveforms such as a simple pulse (e.g., cosine-squared), a chirped
pulse or a pulse Doppler waveform have been used [3]. Due to the very small
duty cycles, large peak powers were required to overcome the propagation
losses incurred. With the necessity of having to operate across bands in
which other authorized users were emitting many anti-interference measures
had to be included [2]. The high-power, pulsed waveforms eectively detected
the targets however, they allowed the long-range interception of the emitter
by noncooperative intercept receivers leading to direction finding, emitter
identification, electronic attack (jamming), and deception.
In order to provide a more covert military capability as well as a more e-
cient use of the HF spectrum, the modern OTHR is moving towards the use of
CW LPI waveform modulations such as CW phase modulation and frequency
hopping [4]. Low power FMCW using multiple waveform repetition frequen-
cies [5, 6] are being used that can relax the transmit power requirements to
provide a more covert sky wave system. In addition these types of waveforms
can resolve the range/Doppler ambiguity usually associated with HF FMCW
radar. Surface wave systems using random low power FM interrupted CW
(FMICW) are also being pursued [7]. The FMICW is a FMCW waveform
that is gated on and o either randomly or with a well-defined sequence.
The main problem caused by the spectra discontinuity is the high-range side
lobes. Optimal sparse waveform designs [8] are being explored. They find
the interference-free channels in the HF band by frequency monitoring, en-
abling the target detection to be accomplished using clear channels while also
lowering the transmit power and minimizing the range side lobes. Shorter
coherent integration times (CIT) [9], adaptive transmit frequency techniques
[10], antijamming through the use of a radar waveform with discontinuous
spectra using two carrier frequencies [11] and orthogonal MIMO waveforms
[12] are also resulting in a quieter and more eective OTHR.
The ionospheric propagation and movement of the layers, contaminates
the transmitted waveform resulting in a low, and fading SNR at the target.
Traditionally, anti-interference measures such as adaptive frequency tuning,
Over-the-Horizon Radar 251
adaptive filtering (including tunable band reject filters) and spatial filtering
have been used by the emitter [2]. Modern OTHR signal processing tech-
niques such as adaptive interference suppression [1315] and clutter cance-
lation algorithms [16] are now able to eliminate the ionospheric propagation
path contaminants and compensate for the smearing of sea echo very eec-
tively. Adaptive time-frequency analysis has also been used to parameterize
the radar signal so the interference can be identified and removed [17].
On the other hand, the ionospheric propagation makes the noncooperative
detection of the OTHR more dicult. The ionospheric contamination makes
the polarization at the receiver indeterminable and consequently, the detec-
tion and DF (azimuth and elevation estimates) of the OTHR waveforms is
tricky especially in the presence of the other interference within the HF band
[18]. The antenna aperture required is large in size and must also be useful
across the HF band with no grating lobes. Large shipboard multifunction ar-
rays are hampered by limitations on the physical size of the aperture required
and problems with electromagnetic interference and compatibility [19]. Sin-
gle sight location techniques using complex time delay estimation algorithms
have been used to DF the OTHR signals [20]. Furthermore, the received HF
signals are nonstationary which limits the noncoherent integration eciency
within the signal processing. Due to multipath presence, high-resolution spec-
tral estimation techniques such as multiple signal classification (MUSIC) and
the cepstrum must also be used [18].
Since the HF signals have a large wavelength (10 100m) the OTHR
also has an inherent resistance to the ARM threat. This is because HF
wavelengths are greater than 10m and any antenna mounted in a missile
seeker (diameter = 0.5m) would have significant diculty deriving any useful
guidance information from the emitted HF waveform. In addition to counter-
ARM capacity, the HF CW waveforms can also detect stealth aircraft and
low level penetrators providing strategic and long distance early warning1 [21].
The OTHR systems are able to survey large areas of land and sea for air and
maritime targets. The OTHRs operating wavelength is nearly the same size
as many of the targets being pursued which puts the targets in the resonant
scattering region (increase in RCS and target detection performance).
In this chapter, sky wave OTHR systems and the eect the ionosphere
has on the waveforms is presented. LPI waveforms are discussed and PRO-
PLAB PRO simulation results are shown to demonstrate the typical footprint
coverage as a function of the HF frequency. Example results for the Chinese
OTH-B system are shown. MATLAB simulations showing the maximum
detection range as a function of the minimum required SNR are also dis-
cussed. Surface wave systems are examined including the FM interrupted
CW approach. With the focus on new emitter waveforms, incorporation of
electronic protection, signal processing and spectrum management, the inter-
1 Stealth aircraft are not optimized against bistatic over-the-horizon radars.
Figure 8.1: Sky wave radar concept (adapted from [23]).
ception and direction finding (DF) of both the sky wave and surface wave
OTHR systems present a significant challenge. For a review of the worlds
OTHR systems, see [22].
8.2 Sky Wave OTHR

The emitted waveform from the sky wave OTHR system is bounced o the
ionosphere and then down to the targets. Reflections from the targets are
bounced back through the ionosphere to the receiver array as shown in Fig-
ure 8.1 [23]. Note the similarity to the MIMO (or spatially waveform diverse)
architecture as discussed in Chapter 10. Due to the amount of sea clutter
being returned the term backscatter is often used. Use of the term backscat-
ter is intended to identify the system geometry, in which the small separation
between transmitter and receiver results in an eective monostatic radar, as
opposed to a bistatic geometry in which the large angle between the transmit
and receive path modifies the target and clutter RCS. These backscatter sys-
tems consist of two modes for detection of targets. Detection of air targets,
and ballistic missiles during the launch phase constitute an air mode. Detec-
tion of surface targets is called a surface mode. Although they are typically
separate operating modes, eorts to combine the modes is also being pursued.
Both modes are aected significantly by the ionosphere and these eects are
presented below.
Figure 8.2: Temperature and plasma density of neutral gas and ionized gas
as a function of altitude (after [25]).
8.2.1 Characteristics of the Ionosphere

The ionosphere is defined as that part of the upper atmosphere where su-
cient ionization can exist to aect the propagation of radio waves [24]. Ex-
amining the LPI characteristics of OTHR systems requires an understanding
of the ionosphere and its eects. The structure of the neutral atmosphere
and the ionosphere containing ionized gas surrounding the Earth is shown in
Figure 8.2 [25]. The left side of the figure (neutral gas) shows the altitude (in
kilometers) as a function of temperature (in Kelvin). The layer right above
the Earths surface up to 10 km is the troposphere and all weather phenomena
occur here. The layer above the troposphere is the stratosphere and the air
flow is horizontal. The layer above the stratosphere is the mesosphere where
the temperature increases with altitude. The layer above the mesosphere is
the ionosphere. The right side of Figure 8.2 (ionized gas) shows the altitude
(km) as a function of plasma density (in units of cm3 ). It also shows the
layer structure in both daytime (solid line) and night time (dashed line). The
plasma is typically the ionized gas and the plasma density refers to electron
density.
The ionosphere is the region at heights of above 80 km and is also the most
outlying area from the center of the Earth [26, 27]. The ionosphere consists of
ionized atoms. It makes long-distance propagation possible by reflecting the
radio waves typically at a height between 90 and 350 km above the Earths
surface. The free electron density is an indicator of the degree of ionization
and is used to measure the structure of the ionosphere in layers. They are
D-, E-, F1- and F2-layers in the daytime. There is always an F-layer at night
and sometimes an E-layer is present as well. The radio waves that propagate
through the troposphere are called surface waves or ground waves. The radio
waves refracted from the ionosphere are called sky waves.
The D-layer below the ionosphere is between the height of 48 and 80
km above the Earths surface. This layer only exists in daytime and its
absorption causes the shorter propagation distance for the radio waves [28].
The distribution of layers in the ionosphere, except the D-layer, is shown
in Figure 8.3. The E-layer exists between 88 and 145 km above the Earths
surface. The maximum electron density in this layer is 1.5105 electrons/cm3
at the height of about 110 km. The E-layer can refract the HF radio wave
inducing propagation distances up to 2,000 km in the daytime. The F-layer
splits into the F1- and F2-layers in the daytime and remains only the F-layer
at night. The F-layer exists between 273 and 321 km. The F1-layer usually
exists between 160 and 240 km and sometimes the electron density in this
layer is not great enough to distinguish it as a separate layer. The F2-layer
exists between 257 and 402 km and most HF radar signals are refracted from
this layer to maximize the propagation range. The nominal height for each
layers peak is 90 km for the D-layer, 110 km for E-layer, 200 km for F1-layer,
and 300 km for F2-layer.
The International Reference Ionosphere (IRI) is a joint project of the
Committee of Space Research (COSPAR) and the Union of Radio Science
International (URSI) [29]. The ionospheric model, IRI-2001, uses input data
that includes the time (universal or local time), date and year, the latitude and
longitude of the desired location, the profile type (height, latitude, longitude,
year, month, day of month, day of year and hour profile), and the parameters
of the profile itself. The optional input includes the sunspot number (SSN)
and ionosphere index (IG) [30].
Figures 8.4 and 8.5, produced by the IRI model, illustrate diagrams of
the electron density profile versus altitude at Nanjing, China (32.0 N and
241.7 W) for daytime and nighttime in the winter and summer, respectively.
In Figure 8.4, the time was set to be 1000 and 2400 (local time) in January
(winter). The same times were used in Figure 8.5 for July (summer). Both
months are in the year of 2007. In these figures, the value along the abscissa
Figure 8.3: Distribution of layers in the ionosphere during the daytime and
nighttime.
(from left to right) represents the electron density (electrons/cm3 ) for each al-
titude. The electron density is generated by semilog calculation method. The
numbers along the ordinate of the figure correspond to the altitude above the
ground (in kilometers). The first three lines on the top of the graph represent
the information of the geographic coordinates, time, day, and month. The
fourth line shows the optional inputs, SSN and IG index, that are generated
by the model itself unless input by the user. In the daytime results shown
in Figure 8.4(a) and Figure 8.5(a), the D, E and F2-layers are easily defined,
but the F1-layer is not well defined. In the night time results shown in Figure
8.4(b) and Figure 8.5(b), both E and F-layer are well defined. These results
demonstrate that the successful noncooperative interception of the OTHR
waveforms depend heavily on the conditions of the ionosphere, time of day,
and sun spot number (SSN) as well as the emitter power and range.
Modeling the ionospheric electron density Ne and refractive index is
useful for HF propagation studies and OTHR system planning and perfor-
mance prediction. Exploitation of this predictability by the targets can also
be used to avoid detection by the OTHR (e.g., by flying when the propaga-
tion losses and ionospheric modulation are the worst). The electron density
Figure 8.4: Electron density in (a) winter day (January 2007) and (b) winter
night (January 2007).
(electrons/cm3 ) of the ionosphere at the desired height from the Earths center
can be calculated using a quasiparabolic ray path to represent the waveform
within the ionosphere as [2628]
l w W2 p Q M
r rm rb 2
Ne = Nm 1 (8.1)
ym r
for
rm rb
rb r
rb ym
where Ne is the electron density having a maximum value of Nm
(electrons/cm3 ) at a radial distance rm (geocentric height of the maximum).
The distance r is the radial distance from Earths center to the height of in-
terest within the layer (r = re + h where re is the Earths radius and h is the
height), rb is the value of r at the layer base (geocentric base height) and ym
is the layer semithickness (half-thickness). This technique is developed for fit-
ting quasiparabolic layers to measured vertical electron density profiles. Note
that rb = rm ym and the Earths radius re = 6,378.1 km. The parameters
are illustrated in Figure 8.6.
Figure 8.5: Electron density in (a) summer day (July 2007) and (b) summer
night (July 2007).
The critical frequency (kHz) is the maximum frequency of the emitted

waveform that is returned from a layer at normal incidence. That is, when the
emitted waveform is transmitted straight up towards the ionosphere (vertical
incidence), the waveform will be returned to earth at all frequencies below
the critical frequency fcr (kHz) which takes the form [2]
0
fcr = 81Nm (8.2)
Thus at normal incidence (i = 0o ) a wave will penetrate the ionosphere if

f > fcr . A negative refractive index occurs when f < fcr and this results
in a ray at normal incidence being reflected from the ionosphere to return to
the Earth. In case the frequency exceeds the critical frequency the influence
the ionosphere layer has on the path of propagation depends upon the angle
of incidence i at the ionosphere. The angle of incidence i is measured
from the normal to the ionospheric layer. The critical frequency fcr is not
the highest frequency that can be reflected from the layer. The maximum
frequency that can be reflected back for a given distance of transmission is
called the maximum usable frequency (MUF). The MUF is related to the
critical frequency and the angle of incidence by [31]
MUF = fcr sec i (8.3)

Figure 8.6: Ray path geometry.
It has been shown that the attenuation varies approximately as the inverse
square of the frequency. Therefore it is desirable to use as high a frequency
as possible without approaching too close to the MUF.
Due to the curvature of the Earth and the ionospheric layer, the largest
angle of incidence i that can be obtained in F-layer reflection is on the order
of 74 degrees. The refractive index of the ionosphere can now be expressed
as a function of the height parameters of the layer and the critical frequency
as [2]
w W1/2 w W2 w W2
81Ne 1 rm r p rb Q2
= 1 = 1 + (8.4)
f2 F F ym r
where f is the HF frequency (in kHz) and F = f /fcr .
The refractive index decreases as the wave penetrates into regions of
greater electron density and the angle of refraction increases correspondingly.
The minimum distance from the transmitter at which a sky wave of given fre-
quency is returned to earth by ionosphere is called the skip zone or distance.
If the OTHR increases frequency, the range of the footprint (and skip zone)
also increases as shown in Figure 8.7. This summary set of skip zones were
derived from the PROPLAB PRO modeling of the Chinese OTH-B discussed
in Section 8.4.
Figure 8.7: Coverage range for fc = 6, 8, 10, 12, 14, 16, 18, 20 and 22 MHz for
SSN = 200 and the AI = 5 on 0:00 UTC 2007/7/31 [22].
8.2.2 Example of F2-Layer Propagation

Consider the electron density profile for the summer daytime in Figure 8.5.
Here the maximum electron density Nm of the F2-layer is at the height of
rm = 6618.1 km (240+6,378.1), with a value of 3 105 electrons/cm3 . The
F2-layer semithickness (half-thickness) ym = 50 km or ym = 0.5(300 200),
and the base height of the F2-layer rb = rm ym or 6568.1 km, which means
the base height of the F2-layer is 190 km above the Earths surface (also shown
in Figure 8.5). At a desired height of 230 km above the Earths surface, the
electron density is calculated from (8.1) as Ne = 2.8814 105 electrons/cm3 .
The critical frequency depends only on the maximum electron density of the
F2-layer fcr = 4.93 103 kHz. Based on (8.4), the refractive index is
a function of electron density Ne and operating frequency f (kHz). If a
frequency of f = 5.1 103 kHz is considered, the refractive index = 0.32.
If the waveform is launched at an angle of incidence of i = 10 degrees,
MUF= 5 103 kHz.
8.2.3 Doppler Clutter Spectrum

How the targets appear in the Doppler space relative to the clutter is im-
portant in determining the HF frequency to be used. The excessive noise
caused by ionospheric propagation is due to the electron density variations.
This process decorrelates the radio signal as it propagates and broadens the
Doppler spectrum. The Doppler radar spectrum of HF radiation backscat-
tered from the ocean surface is shown in Figure 8.8 and characterized by two
strong peaks appearing above and below the carrier frequency [32]. The phys-
Figure 8.8: Bragg peaks along with air and surface targets within the clutter-
target Doppler profile (after [32]).
ical mechanism producing this phenomenon is single-bounce Bragg scattering

from wave trains. The Bragg peaks represent the radiation being reflected.
The Doppler frequencies of the Bragg peaks are 0.102 fc Hz where fc is the
operating frequency in megahertz and the sign indicates the resonant ocean
waves advancing towards or receding away from the radar. The sidebands sur-
rounding the Bragg peaks are due to wave-wave interactions and higher order
Bragg scattering [33]. The ocean properties that can be extracted from fea-
tures of the HF radar sea echo spectrum include (from easiest to measure)
radial surface currents, sea state, surface wind speed, dominant wave period
and direction [34].
Radial velocity variation of maneuvering targets (aircraft) with high speed
may cause significant spread on the radar echo in the Doppler spectrum.
These eects make the echo energy disperse and degrades the eciency of
the coherent integration operation (coherent integration loss). Aircraft target
speeds separate them well from surface targets and clutter for many geome-
tries and provide a good match to the radar capabilities. Many ways have
been proposed to deal with maneuvering targets in OTHR systems. For air-
craft, the modulation periods are typically several seconds. Ships which have
radial speeds between 5 and 25 ms1 require coherent modulation periods
longer than tens of seconds to achieve high resolution in the Doppler domain
so as to distinguish the targets from the clutter [32]. The positioning accuracy
can also be improved by the use of active transponders, clutter storage and
display technology (to calibrate coast lines) and the use of one transmitting
site and multiple receiving sites. Many targets of interest may be concealed
by the clutter and this obscuration is predictable and hence exploitable by
ships wishing to evade detection. For OTHR, reducing the severity of clutter
masking is actively being investigated.
8.2.4 Example Sky Wave OTHR System

As an example, consider the Australias Jindalee Operational Radar Network.
The Jindalee Over-the-horizon Radar Network (JORN) is a system that pro-
vides surveillance over 4,000 km of Australias northern coastline. Full imple-
mentation of JORN has involved the construction of two OTH backscatter
radars that operate in the HF band and are able to detect airborne and ship
targets at ranges of between 1,000 and 3,000 km, with a range resolution
of 2040 km. The JORN radars incorporate bistatic transmission and re-
ception subsystems geographically separated to prevent mutual interference.
The transmission subsystem comprises 28 transmitter chains, each of which
incorporates a 20-kW power amplifier [35].
The receiver subsystem near Longreach, Queensland utilizes 480 receiver
chains, while the one near Laverton, Western Australia incorporates 960 such
chains. Figure 8.9 shows the JORN transmission site at Longreach, Queens-
land. An integral frequency management system determines which frequency
within the operating band will yield the best SNR while spectrum and noise
monitors identify clear channels and background noise levels. A backscat-
ter sounder is used to monitor ionospheric propagation characteristics in the
target area and operating frequency selection is made on the basis of inde-
pendent data that is gathered. General ionospheric structure characteristics
and target ground truths are obtained via a network of vertical and oblique
sounding facilities and transponders located along Australias northern coast-
line [32]. The Longreach 3-km reception array is shown in Figure 8.10 and
is positioned some 100 km from its associated transmitter. A JORN Coor-
dination Center (JCC) is located at the Royal Australian Air Force (RAAF)
base Edinburgh near Adelaide, South Australia. Here, the received data is
processed into usable tracking data. The Longreach site is equipped with a
0.4-km transmission array.
8.2.5 Sky Wave Processing

For sky wave systems, the receivers and transmitters are almost always sepa-
rated by as much as 100 km or more. Sky wave systems have large immovable
antenna arrays that are spread out over a long distance and are positioned
inland such that they are relatively immune to most forms of enemy attack by
rockets and missiles. Sky wave OTHR transmitters use adaptive frequency
Figure 8.9: The JORN transmission site at Longreach, Queensland [35].
Figure 8.10: Jindalee receiving antenna [35].

selection, low side lobe adaptive digital beam forming, and require sophisti-
cated frequency management systems using ionosondes in order to operate
via the ever-changing ionosphere [32]. Ionosondes are devices that send a
spectrum of radio wave pulses straight up to measure: (a) the length of time
it takes for a reflection to be returned, (b) the strength of the reflection, and
(c) how high of a frequency can be reflected. From these three measurements
(time, strength, frequency), the device can determine ionization density, al-
titude of the ionization, and the MUF. The transmitting system is usually
implemented as a number of separate antenna arrays, each covering a certain
frequency subband. Due to the large area being illuminated, these systems
provide the users with a significant surveillance capability to detect targets
at any altitude from the ionosphere to the surface of the Earth.
The receive antenna is usually a long (> 1 km) array of monopoles some-
times with a backscreen to reduce the back lobe radiation. The receiver
array is connected to a beamformer, receiver and ADC. The receiver output
is digitized by an ADC and strobed into a bulk memory for target detection
processing. The samples within a range gate (all range bins of interest) are
added together coherently for a period of time that may vary anywhere from
several seconds to several minutes depending on the targets being detected.
Beyond the time where coherent integration is performed, the returns from
the sea may be added noncoherently. That is, since the samples from beyond
the horizon are stored digitally, a good deal of flexibility in the processing
now exists (mostly to correct for the ionospheric modulation of the Doppler).
Pulse OTHR systems use short pulses or pulse compression to obtain high-
range resolution and a high peak power is required to obtain the necessary
average power for target detection. This high transmitted power can lead to
antenna design constraints and gives rise to impulsive interference that can
easily be identified within the HF band. The engineering compromise to using
high peak power, low duty cycle waveforms is to use CW frequency sweeping
such as the FMCW. Modern sky wave emitters for example, take advantage
of low power (30W in the case of WERA [36]) FMCW modulation. Although
there is currently no military OTHRs that use tens or hundreds of W, the
low power emitted makes it easy for them to hide within other HF radio
services and interference. sky wave OTHRs for air vehicle detection must use
much higher powers than 30W. JORN and the US OTH-B use hundreds of
kW average power. The use of FMCW modulation to transmit and receive
continuously maximizes the average power out of the transmitters amplifiers
providing the range resolution inherent in a given transmission bandwidth.
The disadvantages in using an FMCW are the spectral purity required in
the waveform generator and the high dynamic range required in the receiver
which has to handle the strong direct path (transmit antenna to receive an-
tenna) and the weak signals from far ranges. Doppler information must also
be derived by repeating the FM sweep a number of times (e.g., 64 or 128) and
then performing the FFT in each range gate to examine the phase history
of any target in that gate. After these repeated FMCW sweeps in any given
dwell or surveillance time, the OTHR moves to survey other azimuths and
ranges that are part of the surveillance plan.
Due to the frequency range of OTHR systems and the low volume data
rates, new digital receivers can now directly digitize the returned CW signal
with high resolution without the need for down conversion to intermediate
frequencies (as is conventionally done for microwave emitters). The signal
processor uses a primary FFT to sort the echo returns into range bins and a
second FFT is used to calculate the range-Doppler map. For the OTHR, the
unambiguous range of operation is given by [37]
c
Ru = (8.5)
2WRF
where WRF is the waveform repetition frequency WRF = 1/tm Hz and tm
is the modulation period. The range resolution depends on the modulation
bandwidth F that is used. The return signals are accumulated over this
bandwidth which determines the range resolution. The range resolution (m)
is given by
c
R = (8.6)
2F
For a WRF, the blind speed which varies as a function of fc as [37]
cWRF
= (8.7)
2fc
For the surface-mode (detection of ships), a narrow modulation bandwidth

F is used with a high WRF. For the air-mode (detection of aircraft targets),
a large modulation bandwidth is used with a smaller WRF. In both propaga-
tion cases the amount of energy scattered from the target back to the radar
receiver is extremely small. In addition, the HF bandwidth used represents
a wide percentage of the tunable bandwidth and the signals received after
propagation are nonstationary.
Each region within the coverage area is illuminated with the FMCW and
coherent integration is performed over a number of modulation periods. The
coherent integration time (CIT) must be long enough to extract the target
echoes. The CIT is variable to accommodate the changing ionospheric, clutter
conditions, and target types. Illumination for a long CIT however, works
against the OTHR in terms of avoiding a noncooperative intercept. If the CIT
is too short however, the low Doppler resolution cannot separate the ship from
the large ocean clutter. For the aircraft target, a short CIT is not a problem
since the speed separates it well from the clutter. A significant problem is
the robust high-resolution Doppler processing of accelerating or decelerating
targets [38]. The waveform parameters that can be set include the WRF,
the operating frequency, the modulation bandwidth and the CIT. Multiple
WRFs are used while the target is being illuminated and this solves the range
Doppler ambiguity. Multiple carrier frequencies (either simultaneous or time
multiplexed) are also used to enhance the target detection capability.
8.3 Sky Wave LPI Waveform Considerations

The repetitive nature of the frequency sweeping is one of the main drawbacks
of the FMCW technique. The concentration of power into narrow HF band-
widths and the additional element of repetition (e.g., with the same WRF)
makes the OTHR vulnerable to detection, which can lead to electronic attack
and deception. Consequently, the move away from these traditional wave-
forms to the incorporation of new LPI modulations is being actively pursued.
Changing the modulation parameters makes it even more dicult to DF the
emitter and identify the OTHR system location.
8.3.1 Phase Modulation Techniques

Phase modulation CW techniques such as Barker binary phase coding (with
low time side lobes) have been considered [4]. For a coherent integration time
CIT, more than one modulation period is integrated. Including the CIT, the
peak side lobe level or PSL (see Chapter 3) is given by
w W
tb
PSL = 20 log10 (8.8)
WRF
For example, for a tb = 50 s subcode width and WRF = 5 s, a

PSL = 100 dB is available everywhere in the ambiguity space. Time domain
weighting is used to control the Doppler side lobes at the zero range cut (at
the expense of Doppler resolution). Without weighting or uniform weighting,
the zero-range cut follows the sinc pattern with 13 dB side lobes. Polyphase
codes are also being explored since they provide low side lobes without the
use of weighting. Using polyphase codes with Nc subcodes
w W
1
PSL = 20 log10 (8.9)
Nc
For example, for Nc = 64, PSL = 36 dB. The polyphase codes however,
have a significant Doppler tolerance and therefore are not useful as a Doppler
sensitive waveform. As discussed in Chapter 5, polyphase codes oer the
LPI CW emitter good flexibility in achieving a large processing gain or time-
bandwidth product and can be quite useful in OTHR systems.
8.3.2 Costas Frequency Hopping

Frequency hopping Costas codes are also being explored. The peak side lobe
level for a given Costas time-bandwidth product is
w W
1
PSL = 20 log10 (8.10)
B(WRF)
For example, for B = 20 kHz and a WRF = 5 s, a PSL = 100 dB can be

achieved in addition to a thumbtack ambiguity function.
Increasing the time-bandwidth product of these CW modulations using
larger bandwidths is also being pursued entailing operations over a discon-
tinuous signal spectrum (due to all of the other radio trac), an approach
similar to the surface wave techniques that are used.
8.3.3 Reducing the CIT

In addition to the goal of LPI, a surveillance plan to scan, detect and track
surface targets over large areas requires a short CIT in order that the tracker
can receive the periodic updates in a timely fashion. To enable a short
CIT, ecient clutter cancelation algorithms are actively being pursued to
improve signal-to-clutter ratio. Fourier-based clutter cancelation algorithms
have shown success for OTHR and are based on modeling the first order clut-
ter as a sinusoid and subtracting it from the data. The Fourier technique
estimates the clutter frequency, amplitude and phase [9] from the Doppler
spectrum. A high clutter-to-noise ratio for the Bragg peaks is required in
order to estimate the initial phase for the clutter subtraction.
8.3.4 Multiple Waveform Repetition Frequencies

Unlike a pulsed signal, with FMCW, an ambiguous Doppler frequency is not
folded within the same time delay resolution cell but is shifted to a nearby
time delay cell. Recall that this cross-coupling eect is the range-Doppler am-
biguity of this waveform. One technique to overcome this ambiguity problem,
which also leads to a more LPI waveform, is the use of multiple WRFs within
a single illumination period [5]. It is also wellknown that ionospheric insta-
bility noise (due to polar and equatorial regions) has the potential to appear
through a range ambiguity on top of the signals of interest if the WRF 7.5
Hz. This causes problems for aircraft which have Doppler frequencies of 560
Hz. Consequently, Doppler frequencies greater than this must be resolved
without any range ambiguities.
To understand the use of multiple WRF technique, we revisit the FMCW
range-Doppler cross coupling eect as presented in Chapter 4. Recall that
the beat frequency is proportional to the range of a stationary target (or one
that has a Doppler shift fd < 1/2t0 ) as
cfb t0
R= (8.11)
2F
and the corresponding range resolution
c ctmin
R = = (8.12)
2F 2
where tmin = 1/F s is the minimum time delay that can be detected.
Due to the range-Doppler cross coupling eect, if the target is moving at
a velocity such that fd > 1/2t0 , the beat frequency is fm = fb fd and
corresponding range of the target is given by
cfb t0 cfd t0
Rm = (8.13)
2F 2F
and shows that the measured range is a function of the true range and an
error due to the Doppler shift or Rm = R Rd . The range error due to the
Doppler shift can be re-written in terms of range bins as [5]
Rd fd
= (8.14)
R WRF
where WRF = 1/t0 . This shows that the measured range is increased (or
decreased) by one range bin as the Doppler shift is decreased (or increased)
by a frequency equal to the WRF.
In the multiple WRF technique, three WRFs are used during a single
target illumination time as shown in Figure 8.11. Since the waveform repeti-
Figure 8.11: Coprime waveform repetition frequencies.
tion frequencies of the waveforms are dierent, aliasing causes the estimated
Doppler shifts to be dierent during each WRF. The Chinese remainder the-
orem can then be utilized to calculate the true Doppler shift where the max-
imum unambiguous Doppler range is limited by the least common multiple
(LCM) of the selected WRFs [5, 6]. The duration and number of coprime
WRF sweeps are chosen to resolve the range ambiguities and to achieve the
required Doppler resolution respectively.
The Doppler shifts are unique for each WRF block and estimated sequen-
tially using two FFTs as in normal FMCW processing. For OTHR surface
mode targets, it is only necessary to extend the unambiguous Doppler to
70 Hz so WRF3 = 7, WRF2 = 6, and WRF1 = 5 are sucient giving a
maximum unambiguous Doppler coverage of
3
1
fd = WRFi = 105 Hz (8.15)
2 i=1
In selecting the WRF, the first limitation is the maximum unambiguous range
ctm3 ctm2 ctm1
Rmax < < (8.16)
2 2 2
where tmi are the corresponding modulation periods for the three WRFs
(tmi = 1/WRFi ) [6]. The dierence in range bins between the three WRFs
should be large enough to separate the clutter. If NRB is the number of bins
covered by the clutter then to ensure a clutter free range
F F
NRB = (8.17)
WRF1 WRF2
and also
F F
NRB = (8.18)
WRF2 WRF3
which is the smallest dierence. Rewriting (8.18),
NRB WRF3 WRF2
< (8.19)
F WRF3 WRF2
and choosing WRF3 = WRF2 + 1, (8.19) is then
F
> WRF22 + WRF2 (8.20)
NRB
A fixed waveform repetition frequency emitter is able to resolve Doppler fre-
quencies within WRFi /2. With the WRFi satisfying WRF3 > WRF2 >
WRF1 , the Doppler shift of the target can be written as [5]
fd = n1 WRF1 + x1 = n2 WRF2 + x2 = n3 WRF3 + x3 (8.21)
where n1 , n2 and n3 are either a positive or a negative integer depending on

the Doppler shifts and x1 , x2 and x3 are the corresponding fractions that are
the measured Doppler shifts of certain targets with the limits WRFi /2
xi WRFi /2.
Below we present the alternate solution using the Chinese remainder the-
orem (CRT) and repeat the example in [5]. The problem is first set up as a
set of simultaneous congruences
fd x1 (mod WRF1 )
fd x2 (mod WRF2 )
fd x3 (mod WRF3 ) (8.22)
When WRFi are N positive integers that are coprime, the set of congruences
have a unique solution modulo
N

M= WRFi (8.23)
i=1
Any integer congruent modulo M to a given solution is also a solution. (The

proof and straightforward computation of the CRT is given in [39].) The
solution of the simultaneous congruences is
N
3 M
fd = bj xj (8.24)
j=1
WRFj
or for N = 3
f0 = WRF2 WRF3 b1 x1 + WRF1 WRF3 b2 x2 + WRF1 WRF2 b3 x3 (mod M )
(8.25)
The bi values are found by a repeated application of the Euclidean algorithm.
So to complete the CRT solution we consider the example given in [5] where
WRF1 = 5 Hz, WRF2 = 6 Hz, and WRF3 = 7 Hz and the corresponding
ambiguous Doppler shifts are x1 = 2 Hz, x2 = 2 Hz, and x3 = 3 Hz. For
b1 , since the greatest common divisor of 42 (M/WRF1 = WRF2 WRF3 ) and
5 (WRF1 ) is 1, the Euclidean algorithm is used to solve for x0 , y0 such that
42x0 + 5y0 = 1. Then we have 42x0 1(mod 5), so b1 = x0 . Applying the
algorithm [39]
42(1) + 5(0) = 42
42(0) + 5(1) = 5
42(1) + 5(8) = 2
42(2) + 5(17) = 1
and consequently, b1 = 2. Repeating the application of the Euclidean al-
gorithm for b2 and b3 gives b2 = 1 and b3 = 3. The solution (8.25) is
then
fd = 4(42) + 70 270(mod 210) = 368 (8.26)
which is out of the proper range (0 to 209). However, M = 210 goes into 368
two times, so 368 + 2(210) = 52 solves the problem and is the least positive
solution. That is fd = 52 Hz as also solved in [5]. The technique presented
here determines the actual Doppler frequency and is a more straightforward
method using the CRT.
8.3.5 Out-of-Band Emission Suppression

In MIMO systems, the design objectives are transmitter element waveform
orthogonality (using either temporal or spectral diversity). The implemen-
tation is limited by the performance of the transmission system especially
the out-of-band emissions. New approaches to FMCW waveform design and
to waveform generation have been adopted that are theoretically capable of
achieving much lower out-of-band spectral levels [13]. Spectrum leakage that
is out of band can be a significant beacon for HF/DF systems trying to nonco-
operatively identify the emitter. For FMCW sky wave emitters, out-of-band
emission suppression is a major consideration and steps are usually taken
to insure compliance with for example, the International Telecommunication
Union (ITU). The ITU-R spectral management document SM.1541-1 provides
recommendations on out-of-band emission advice for a number of radar wave-
forms including the FMCW waveform. The reduction of side lobes through
time and frequency weighting can be done to improve the out-of-band emis-
sion spectrum while improving the in-band performance. One choice is to
smooth the frequency transition region at the ends of each chirp by smooth-
ing the signal in either the instantaneous frequency or amplitude domains
[14].
Recall that a single sawtooth FMCW signal frequency can be written in
the time domain as
F
fI (t) = t + fc (8.27)
tm
for |t| tm /2. The complex signal is then
v(t) = ej(t) (8.28)
where
F 2
(t) = t + 2fc t (8.29)
tm
for |t| tm /2.
The spectral taper method begins by designing the waveform in the spec-
tral domain and including a modulation with a taper function wST (f )
8
vf m = wST (f )v(f )ej2f t dt (8.30)

The function vf m can then be inverse Fourier transformed to derive the signal.
Taper functions such as the Hann taper function have been used. The taper
however, does change slightly, the signals phase, amplitude and instantaneous
frequency.
The instantaneous frequency of the FMCW waveform is also a discontin-
uous function. A simple method to reduce this discontinuity is to introduce
a counter sloping chirp at the waveform end as [14]

w W
1 F
fI (t) = t + fc (8.31)
1 tm
for tm /2 t (1 2)tm /2 and fI (t) = F t/(tm ) + fc otherwise.

Here 0 1/2 is the flyback factor. The flyback factor accounts for the
amount of time required for the sawtooth FMCW to return to the beginning
frequency. To reduce the roll-o rate the application of amplitude tapering
can be applied. For example, a cosine-Tukey amplitude taper wCT (t) in the
time domain on the signal takes the form
vam (t) = wCT (t)v(t) (8.32)
This technique provides excellent out-of-band emission control even with low
percentage tapering (e.g., 10%) causing only a small amount of loss in coher-
ent gain.
The penalty when using this approach is that the Fresnel ripples, which
are usually a characteristic of the spectrum of the weighted chirp signal, now
appear in the time domain waveform. The waveform to be transmitted has
small amplitude ripples that must be preserved if the desired spectral perfor-
mance is to be maintained [13]. That is, high linearity must be maintained
at full output power levels.
8.4 Sky Wave Maximum Detection Range

The relative performance of the best known sky wave OTHR systems is shown
in Figure 8.12. One of the important performance parameters of the OTHR
is the maximum detection range. Figure 8.13 shows the OTH-B sky wave
target detection geometry. The received power at the radar receiver from the
target can be expressed as
PCW GT GR 2 T
PRT = (8.33)
(4)3 RT2 RR
2L
P 2 LF L
where PCW is the average transmitter power in watts, GT is the transmit

antenna gain, GR is the receive antenna gain, is the wavelength at the
carrier frequency fc ,
T is the targets radar cross-section, the term LP 2 is the two-way trans-
mission path loss and is on the order of 1020 dB [21], L is the system losses
which include the transmitter and receiver subsystem losses which are 15 dB
[23]. LF is the Faraday polarization loss of the ionosphere which is typically
3 dB [21], RT is the distance of the ray path traveling from the radar, re-
flecting o of the ionosphere down to the target and RR is the distance of
the return ray path traveling from the target to the receiver also reflecting
Figure 8.12: Relative performance of the best known sky wave OTHR systems
(after [22]).
o of the ionosphere. Understanding the ionosphere, and how it aects the

transmit and received waveform (from the target) is critical to predicting the
OTHR capability. For example, the ionosphere can modulate the targets
Doppler profile making it undetectable. This is discussed in greater detail in
the following sections.
The minimum input signal-to-noise ratio SNRRi is related to the receivers
sensitivity R . The receiver can detect and process an incoming target signal
at this signal level or higher. Substitution of the sensitivity for PRT in (8.34),
the maximum detection range (reflecting o of the ionosphere) of the radar
becomes w W1/4
PCW GT GR 2 T
RR max = (8.34)
(4)3 R LP 2 LF L
where RRmax is calculated by assuming that the OTHR transmitter and re-
ceiver are located at the same range from the target RT = RR . The sensitivity
R is the product of the minimum SNR required at the input SNRRi times
the noise power in the input bandwidth of the receiver or
where k is the Boltzmanns constant (k = 1.38071023 J/K), T0 is the stan-

dard noise temperature (290 K), FR is the receiver noise factor and includes
Figure 8.13: Target detection diagram.
the fact that the additional noise is 2050 dB larger than the receivers ther-
mal noise, and BRi is the receivers input bandwidth in Hertz. The maximum
detection range (reflecting o of the ionosphere) then can be expressed as
w W1/4
PCW GT GR 2 T
RR max = (8.36)
(4)3 kT0 FR BRi (SNRRi )LP 2 LF L
Consider the Chinese FMCW OTH-B radar characteristics with
PCW = 1.2 MW (61 dBW), GT = 18 dB and GR = 26 dB at 14.5 MHz,
FR = 40 dB, BRi = 30 MHz, LP 2 = 15 dB, L = 15 dB, and
LF = 3 dB [21, 40, 41]. The separation distance is typically 60200 km
for the Chinese OTH-B radar system [21]. We can calculate the modulation
bandwidth F from the published range resolution 15 km = R = c/2F
or F = 10 kHz. Figure 8.14 shows the radar maximum detection range
(reflecting o of the ionosphere) as a function of the required input SNR
(SNRRi ) for 1, 10, and 100 m2 at operating frequencies fc = 14.5 MHz [22].
Assuming a flat Earth situation, the detection range Rfootprint along the
Earth becomes
w W2
RR max
Rfootprint = 2 h2 (8.37)
2 F 2layer
where hF 2layer is the F2-layer height from the Earths surface. For this
example, assume the F2-layer height is about 240 km. The geometry diagram
for (8.37) is shown in Figure 8.15. The detection distance along the flat Earth
(Rfootprint ) is calculated from (8.36) and (8.37) as shown in Figure 8.16.
Figure 8.14: FMCW OTH-B maximum detection range (RR max ) for T = 1,
10, and 100 m2 [22].
From [21], the Chinese OTH-B radar has a skip zone or minimum detection
range of 700 km and a maximum detection range of 3,500 km. For a maximum
detection range along the flat Earth Rfootprint = 3,500 km, the RR max
is calculated as 3,532 km from (8.36). The minimum required input SNR
(SNRRi ) for fc = 14.5 MHz for T = 1, 10, and 100 m2 is 107, 97 and
87 dB respectively from Figure 8.16. The processing gain of the emitters
waveform used then provides the sucient SNR for target detection.
After target detection, coordinate registration is used for multipath track-
ing to convert the slant ranges and slant azimuth to surface coordinates.
Several methods based on planar and spherical models have been reported
recently [4244]. Furthermore, with sophisticated processing to eliminate the
coherent integration loss caused by irregular target motions, the transmit
power can be lowered considerably.
8.5 Sky Wave Footprint Prediction

PROPLAB PRO is a propagation software for the personal computer that
calculates the precise behavior of radio signals as they travel through the
atmosphere. It simulates accurately, radio transmissions into the ionosphere
using sophisticated ionospheric ray-tracing techniques [45]. One of the OTHR
Figure 8.15: Geometry diagram of the detection range calculation.
parameters that is of high interest to the noncooperative intercept receiver to

predict is the emitter footprint or coverage area. For the Chinese OTH-
B radar, the transmitter is assumed to be located at Nanjing (32.05o N,
241.22o W) and the target is located at coordinates 30o N, 204.3o W which
is 3,500 km distance away from the transmitter. A typical transmitter an-
tenna radiation pattern is shown in Figure 8.17 with a maximum transmit
gain of 18 dB at 14.5 MHz from a vertical dipole array pointed in azimuth
85o from true North (towards the United States).
Simulation results for the summertime SSN 200 are shown below. The
level of geomagnetic activity (A-Index, AI) = 5 on 21 July 2007. The electron
density profile (Ne ) at the midpoint ( 32.37o N, 222.55o W) of the transmitter
and the target along the great-circle path is shown in Figure 8.18. The D, E,
F2 and topside layers are shown.
The maximum usable frequency (MUF) profile (24-hour period) is shown
in Figure 8.19 with the F2-layer MUF (the top-most line), optimum working
frequency or FOT (the second line), average MUF of E and F-layer (the thin
line next to FOT line), and the E-layer MUF (the bottom line). The most
often used frequency range is between the FOT and the average MUF of E
and F layer.
From Figure 8.19, the maximum and minimum FOT values are 28 MHz at
08:00 UTC (coordinated universal time) and 20 MHz at 19:00 UT (universal
time). Based on the maximum and minimum FOTs, the ray tracing screens
can be generated as shown in Figures 8.20 and 8.21. To examine the footprint
coverage and the skip zones generated as a function of frequency and elevation
angle, the ray trace plots for fc = 6, 8, 10, 12, 14, 16, 18, 20 and 22 MHz were
generated for angles between 0o and 20o . For these results SSN = 200 and
the AI = 5 on 0:00 UTC 2007/7/31. Figure 8.20 shows the results for 10 MHz
and 14 MHz. Two eects can be noticed from the results. First, the increase
of frequency toward the MUF results in an extended range and increased
Figure 8.16: FMCW OTH-B maximum detection range (Rfootprint ) for T =

1, 10, and 100 m2 [22].
skip zone. In addition, when going from fc = 10 MHz to fc = 14 MHz,

the 20o elevation launch angle finally penetrates the ionosphere. In Figure
8.7, a summary of the ray trace plots is shown. The range coverage for each
frequency is shown (including the skip zone). Note that as the frequency
approaches the MUF, the range increases. In Figure 8.21, fc = 18 MHz and
fc = 22 MHz are shown. Considerable ionospheric penetration is occurring
for SSN = 200 and the AI = 5 on 0:00 UTC 2007/7/31 as the frequency
approaches the MUF.
8.6 Surface Wave OTHR

While the principles of HF surface wave or ground wave OTHR have been
known for decades, they still present challenges to remain covert. Figure 8.22
shows the surface wave OTHR concept that uses a spatial separation of the
transmit and receive system. The OTHR transmitter (XMTR) emits radio
waves that follow the surface of the sea extending over the horizon. Surface
wave radar works best when using vertically polarized antennas in contact
with salty conducting water. The sea water is a good conductor and the air
acts as the dielectric. As a result, the lowest layer of air and uppermost layer
of sea form a waveguide in which the HF radiation is constrained by internal
Figure 8.17: Example transmitter antenna radiation pattern for the OTH-B
radar system [22].
Figure 8.18: Electron density profile for SSN = 200 and AI = 5 on 2007/6/21
[22].
Figure 8.19: Maximum usable frequency for SSN = 200 and AI = 5

on 2007/6/21 [22].
reflection. The antenna polarization is always chosen to be vertically polarized

to avoid the higher attenuation associated with horizontal propagation. This
coupling of the radiation to the sea surface provides a means to detect targets
over the horizon beyond the line-of-sight limit experienced by conventional
microwave radar systems.
The surface wave method cannot be used over land, on freshwater lakes,
or where fresh water dilutes the sea such as in the Baltic or the Nile Delta
[23]. They do not require real-time knowledge of the ionospheres behavior.
These systems are most applicable to naval applications. For example, moni-
toring a countrys exclusive economic zone and providing surface combatants
with early warning of an attack by antiship cruise missiles. Being relatively
inexpensive they are also widely used for collecting good quality wave, current
and tidal information.
Surface wave OTHR can detect surface and air targets from 10 to 400 km.
The range is limited by the amount of power transmitted and the attenua-
tion incurred. The attenuation can be predicted by examining the modified
refractive index profile above the sea surface which is a function of the sea
state and the atmospheric conditions. Computer programs such as the ad-
vanced refractive eects prediction system (AREPS) [46] can give accurate
predictions of current attenuation anywhere in the world. The range reso-
Figure 8.20: Ray tracing results for SSN = 200 and the AI = 5 on 0:00 UTC
2007/7/31 for several elevation angles at (a) 10 MHz and (b) 14 MHz [22].
Figure 8.21: Ray tracing screen for SSN = 200 and the AI = 5 on 0:00 UTC
2007/7/31 for several elevation angles at (a) 18 MHz and (b) 22 MHz [22].
Figure 8.22: Surface wave radar concept.
lution varies widely from system to system and depends on the modulation
bandwidth used which can be on the order of a few megahertz [8]. The opti-
mum surface wave carrier frequency that is used depends critically on where
the target of interest will appear in Doppler space relative to the clutter.
8.6.1 Example Surface Wave OTHR System

Raytheon Canadas coastal surveillance radars are designed to detect and
track ships, aircraft and ice formations out to ranges of up to 400 km [47].
The radars are shelter-mounted for ease of installation in remote locations
and feature a comprehensive supervisory system for unattended operations.
The also include spectrum management to enhance performance reliability.
The SWR-503 (3.55.5 MHz) is optimized for long-range surveillance of tar-
gets out to a range of 407 km. The SWR-610 configuration (610 MHz) is
designed for medium-range applications and features detection and tracking
of smaller targets when compared with SWR-503. The transmit antenna
is a monopole log periodic array and is shown in Figure 8.23. The receive
array is a 16-element array on ground screen and is shown in Figure 8.24.
The bandwidth is 380 kHz but is typically about 20 kHz (R = 7.4 km).
Typical velocity resolution is 0.1 km/h (iceberg), 0.9 km/h (ship); 7.4 km/h
(aircraft). The average transmitter power is 3.2 kW with a peak power of

16 kW. The detection capability for the SWR-503 is shown in Figure 8.25 for
several classes of air and surface targets.
Figure 8.23: SWR-503 transmit antenna configuration [47].
8.7 Surface Wave LPI Waveform

Considerations
There are two problems with FMCW for surface wave systems. First, the
necessary isolation between the transmitter and receiver for continuous-wave
operation at HF frequencies is more dicult to achieve. Second, due to the
propagation losses, the signal levels decrease rapidly with distance and the
dynamic range of the signals exceeds that of any available receiver hardware
[7]. Due to heavy user congestion, surface wave OTHR systems are also
restricted to operating within narrow frequency bands.
8.7.1 FMICW Characteristics

To overcome these problems, the spectrum must be made discontinuous and
FMICW type signals are used [48]. The FMICW is an FMCW waveform that
Figure 8.24: SWR-503 receive antenna configuration [47].
is gated on and o with a well defined sequence which is either pseudorandom

(adding to the LPI nature of the emitter) or a deterministic sequence. The
on and o gating or interruption process involves breaking the transmitted
signal into an integral number of shorter bursts during the chirp period.
To generate the FMICW, a synthesizer is used to produce a linear fre-
quency swept waveform with frequency varying across the modulation band-
width F (chirp waveform). From Chapter 4
}w W ]
F F 2
s1 (t) = sin 2 fc t+ t (8.38)
2 2tm
for 0 t tm . This waveform is then gated on and o to produce the pulsed

waveform shown in Figure 8.26. The frequency sweep is pulsed P times by
multiplying it by the gate signal which can be expressed as
fT (t) = AT (t)s1 (t) (8.39)
where the gating function is

P
3 1
AT (t) = rect[(t nq R /2)/R ] (8.40)
n=0
Figure 8.25: SWR-503 detection capability for various classes of surface and
air platforms [47].
where P is the number of short bursts, R is the pulse width in seconds,

rect(t/R ) is a rectangular pulse of width R centered at the origin and q is
the gating period. The signal fT (t) = AT (t)s1 (t) represents a single FMICW
sweep [49].
In Figure 8.26(a), the FMCW signal used to generate the FMICW is
shown. In Figure 8.26(b, c) the transmitter gating sequence and the receiver
gating sequence (complement of transmit sequence) are shown respectively.
In Figure 8.26(d) the FMICW signal generated using the gating sequence in
Figure 8.26(b) is shown. The eect of target range on the received signal
is shown in Figure 8.27 where for the same transmitted signal, the received
signals are shown for two dierent target ranges [7].
The FMICW concept is to transmit for a specific period of time and
after the transmitter has been turned o, to receive during the quiet period.
Both the transmit and receive waveforms are subjected to dierent interrupt
sequences. The FMICW allows the weak return signals to be more easily
detected. With a duty factor of, for example, 50%, the average transmitted
power is reduced by 3 dB compared to the FMCW waveform and the improved
isolation typically decreases the system noise floor by more than 3 dB adding
to the LPI nature of the emitter. It is important to remember that to prevent
Figure 8.26: (ad) FMICW generation showing transmit and receive gating
(from [7]).
eclipsing, the receiver must always be o whenever the transmitter is on.

Further dierences between the two sequences are introduced by soft gating
which is used to suppress transients that would be caused if the interrupt
sequences were switched rapidly between on and o states [7].
The radar returns from the targets are delayed by the two-way travel time
td and are modified by the receiver gating sequence AR (t) as
fR (t) = AR (t)AT (t td )s1 (t td ) (8.41)
or
fR (t + td ) = AR (t + td )AT (t)s1 (t) (8.42)
With the FMICW, the receiver must always be o whenever the transmitter
is on so that the transmitter and receiver interrupt sequences are not identi-
Figure 8.27: FMICW radar echo from target with two-way travel time of
(a) t1 s and (b) t2 s (from [7]).
cal. This interrupted spectrum gives rise to high-range side lobes and conse-
quently a low dynamic range [50]. Dierences between the two sequences are
introduced by the soft gating used to suppress these transients that would
be caused if the interrupt sequences were switched rapidly between the on
and o states. Depending on the targets range, the transmitted waveform is
generally not received at the radar in its entirety.
There are three ways to extract the spectrum of the received echo from
the target. The first method is to weight the return signal bursts individually
and process the complete return as for a conventional FMCW receiver using
a single long FFT. The second method is to weight each burst individually
and pack the remainder of the signal with zeros and process the complete
return as for a conventional FMCW receiver. The third method is to weight
each burst individually and process each short section using an FFT matched
to its length. Here an FFT with a duration equal to the burst period is used
to produce a spectrum. The number of floating point operations required
to generate the FFT is N log2 N where N is the number of record points in

the FFT. Therefore it is more ecient to generate a number of shorter FFTs
and integrate them noncoherently than to generate a single long FFT. For
example, for six bursts, the noncoherent integration improvement factor is
about 6.2 dB compared with about 7.8 dB for coherent integration of the
complete interrupted signal [7].
Conventional Doppler processing of FMICW is constrained by the limited
time that a target is present in a range cell. Targets whose speed is outside the
range of conventional processing have their signal energy smeared in range due
to the target traversing several range cells within the processing period. This
smearing however, can be removed with velocity-matched filters that perform
phase corrections on the processed signal. The velocity-matched filters can
accurately extract the range and velocity of a low-flying aircraft whose signal
is aliased in velocity and smeared in range conventional processing.
8.7.2 FMICW Ambiguity Space

The relative time delays in the receiver gating sequences significantly change
the actual signals gated into the receiver. The ambiguity function will also
be dierent for these two ranges and must be carefully considered for long
range applications. Since the target Dopplers are very small, the zero Doppler
ambiguity is sucient to evaluate the performance of the waveform. The zero
Doppler ambiguity function for the FMICW can be computed using
8 p Q p Q p Q j2 t
(, , td ) = AT t + AR t + td AT t e dt (8.43)
2 2 2
where = F/tm is the frequency sweep rate.

The ambiguity function of the FMICW is evaluated in [7] and it is shown
that the gating on the received signal varies with the target range and intro-
duces a range dependence on the ambiguity function. That is, the performance
of the FMICW emitter is dependent on the relationship between the trans-
mit period and the round-trip time td and hence is range-dependent. The
evaluation of the ambiguity function is dicult for the FMICW due to the
finite extent of the linear frequency sweep as well as the eect of the interrupt
sequence used. Note the inverse problem of determining the parameters of an
FMICW waveform to realize a desired ambiguity function is more dicult.
Also, since the target velocities are typically very small, the zero-Doppler
ambiguity function is sucient to evaluate the performance of the waveform.
It is very dicult to find clear channels (30100 kHz) in the HF band for
use in FMICW. A sparse waveform design can help mitigate the interferences;
however, it also gives rise to high-range side lobes. The cochannel interference
that is present can be mitigated by using an adaptive coherent side lobe
cancelation algorithm based on forming subarrays that use negative frequency
range bin samples to calculate the covariance matrix and correlation vector as
described in [15]. Improvements in signal-to-interference ratio on the order
of 20 dB can be achieved. The sparse waveform approach is to find the
interference-free channels in the HF band and then transmit the signal in
these clear channels. The main problem caused by the spectrum discontinuity
is the range side lobes that are created [8]. Another interesting approach is to
use a multiparametric generalization of the nonuniform FMICW train and by
exploiting a factorization of the ACF based on genetic algorithms, an optimal
solution has been proposed [8].
A new FMICW waveform is presented in [51] and provides high-range
resolution using a larger modulation bandwidth and a narrow modulation
period to achieve a longer coherent integration time for high-speed targets.
To prevent the high speed targets from smearing into many range cells during
the longer integration periods, two frequency sweep bandwidths with dier-
ent sweep repetition intervals are used to determine the required coherent
integration time (CIT) and R independently. The target velocities are esti-
mated using one waveform and then applied to compensate the velocity phase
terms to account for the target movement. Due to the ability of higher order
correlation and spectral analysis methods to eectively suppress symmetrical
distributions such as Gaussian noise, these techniques have also been explored
for interference cancelation and signal detection in OTHR [52].
8.8 Surface Wave Maximum Detection Range

The HF radar system planning and implementation problem is compounded
by the influence of the conducting ground plane on the radiation resistance of
radar antennas and target backscattering cross-section. The potential error
resulting from basing a system design on a radar equation with an inconsistent
set of parameter definitions is high. Without careful attention to detail,
predicted received signal-to-noise ratios might be in error by 10 dB or more.
To determine the maximum detection range as a function of the input
required SNR, we develop the surface wave radar equation taking into account
three eects that do not occur in conventional radars [53]:
1. A doubling of field strength due to direct and ground reflected waves;
2. A mutual impedance between the antennas and their images in the
ground plane;
3. A coupling between the target and its image in the ground plane.
Eects (2) and (3) are only significant when the antenna or the target is within
approximately one wavelength of the ground. Due to the HF wavelength,
most antennas and some classes of target will be within one wavelength of
the ground and so the eects cannot be ignored. The coupling of the antennas
Table 8.1: RISP Gains for Some Example Antennas

Gain (dB)
Antenna RISP dBi
Isotropic in free-space 0 0
Isotropic on a perfect ground plane 0 3
Hertzian dipole in free-space 1.8 1.8
Hertzian dipole on a perfect ground plane 1.8 4.8
/4 monopole on a perfect ground plane 2.2 5.2
and target with their images in the ground modifies their radiation resistances
and this is the root cause of eects (2) and (3). In this section it is convenient
to describe the gain of an antenna as Relative to an isotropic antenna at the
same position (RISP). The RISP gain of some example antennas are presented
in Table 8.1 along with the dBi value.
To develop the surface wave equation we follow the development in [53]
and begin with the power flux Fi , incident on a target due to transmission
from a vertical Hertzian dipole over a perfectly conducting ground plane
E2 pt gt [2 sin cos(kht cos )]2
Fi = = (8.44)
Z0 4d2 (1 + t )
where E is the field strength, Z0 the impedance of free-space, pt the radiated
power in the presence of the ground, the transmitting antenna has RISP gain
gt and d is the distance between transmitter and target. The term (1+t ) is a
factor to allow for coupling between the antenna and its image in the ground.
The angle is the zenith angle of the target measured at the transmitter,
k = 2/, and ht is the dipole height above the ground. The term is given
by } ]} ]
3 sin(2kh)
= cos(2kh) (8.45)
(2kh)2 2kh
At h = 0, (1 + ) = 2. For h > , (1 + ) 1. The term in square brackets
in (8.44) gives the vector sum of direct and ground reflected signals. When
/2
pt gt
Fi = 2
4 (8.46)
4d (1 + t )
The trailing factor of 4 is due to the in-phase addition of the direct and
ground-reflected waves.
For the collecting aperture of a target consider that the target is a matched
antenna with RISP gain gx . The power available to the matched load is equal
to that which is reradiated. The target power collecting aperture is given by
gx 2 1
(8.47)
4 (1 + x )
This expression comprises:

1. the aperture of an isotropic antenna in free-space, 2 /4;
2. a factor for additional gain, gx ;
3. a factor for coupling between the target antenna and its image in the
ground (1 + x )1 .
The power reradiated Prerad by the target is found by combining (8.46)
and (8.47) as p 2Q

pt gt gx 4
Prerad = 4 (8.48)
4d2 (1 + t ) (1 + x )
The power flux back at the receiver antenna FRR can be found by reapplying
(8.46), to extend (8.48) as
p 2Q

pt gt g x 4 gx
FRR = 4 4 (8.49)
4d2 (1 + t ) (1 + x ) 4d2 (1 + x )
The second gx is the gain term for the power reradiated by the target, the
trailing 4 indicates that the direct and ground-reflected waves from the target
add in phase.
The power available from the receiving antenna with RISP gain gr , when
located above a perfectly conducting plane, is given by
p 2Q

gr 4
(8.50)
(1 + r )
Multiplying (8.49) and (8.50) gives the received power pr

p 2Q p 2Q

pt gt gx 4 gx gr 4
pr = 4 4 (8.51)
4d2 (1 + r ) (1 + t ) 4d2 (1 + x ) (1 + t )
and is the main HF surface wave radar equation and includes the ground plane
eects on antennas, target, and propagation. The target backscattering term
is evident only indirectly by the target gain gx . To be useful, it is necessary
to relate gx to a target backscattering cross-section T .
Three definitions of the target backscattering cross-section coecient are
considered.
1. Conventional microwave radar definition:
w 2W
4gx2
T = 2
(8.52)
(1 + x ) 4
and relates the scattering cross-section of a matched antenna of RISP

gain gx above a perfectly conducting ground plane with elevation h > .
The factor (1 + x )2 accounts for the antenna collecting and reradi-
ating properties.
2. Free-space backscattering cross-section definition:
w 2W
2
F S = gx (8.53)
4
and relates the backscattering cross-section of an isotropic antenna in
free-space (absence of a ground plane). If the isotropic antenna were
replaced by a target antenna with RISP gain gx , then its backscattering
cross-section in free-space is given by F S .
3. Shearmans definition [54]:
p Q
2
gx2 4
s = (8.54)
(1 + x )2
which departs from the conventional microwave radar definition by a
factor of 4 to allow for the ground reflection on the signals return to
the receiver.
By substituting for gx from (8.52), (8.53), and (8.54) into (8.49), we obtain
three dierent equations, all equally valid, but using dierent definitions of
the target cross-section:
p 2Q

pt g t 4 T (1 + x )2
pr = 2
4 D 2 i (8.55)
4d (1 + t ) (1 + x ) 4 4
p 2Q

1 g r 4
4
4d2 (1 + x ) (1 + r )
p 2Q

pt g t 4 F S
pr = 4 D 2 i (8.56)
4d2 (1 + t ) (1 + x ) 4
p 2Q

1 g r 4
4
4d2 (1 + x ) (1 + r )
p 2Q

pt g t 4 s (1 + x )2
pr = 2
4 D 2 i (8.57)
4d (1 + t ) (1 + x ) 4
p 2Q

1 g r 4
4
4d2 (1 + x ) (1 + r )
The propagation loss must now be addressed to be included in the ex-

pression for pr . Two definitions are given below and are ratios of transmitted
power to received power and thus are values greater than unity.
ITU-R Definition: The internationally recognized definition of basic trans-

mission loss Ib is the ratio of radiated power to power available from a matched
receiving antenna when the actual antennas are replaced by isotropic anten-
nas at the same location as the actual antennas. For a one-way path above a
perfectly conducting ground plane
w W2
2d
Ib = (1 + t )(1 + r ) (8.58)

For antennas on the ground plane t = r = 1 and Ib is the same as the
2
free-space value of (4d/) .
Barrick Definition: In 1971, Barrick [55] published curves of additional

groundwave attenuation for propagation over the sea when it is roughed by
wind waves. As part of the work, Barrick included curves showing basic
transmission loss over a smooth sea. Barricks theory and curves are very
widely used in HF groundwave radar design. Barricks basic transmission
loss, IBar , for a one-way path above a perfectly conducting ground plane is
given by
w W2
2d
IBar = (8.59)

Note the two propagation loss quantities Ib and IBar are related as
Ib = IBar (1 + t )(1 + r ) (8.60)
and Ib can be up to 6 dB greater than IBar (i.e., a possible dierence of

12 dB on the two-way radar path).
The three expressions for pr (8.55), (8.56), and (8.57) corresponding to
the three definitions of target cross-section can be combined with the two
expressions for propagation loss. This leads to six expressions for pr , as
presented below. All six are equally valid.
For the ITU-R version of loss Ib and the microwave definition of cross-section
T :
1 T (1 + x )2 1
pr = pt gt D 2 i gr (8.61)
Ib 4 4 Ib
For the ITU-R version of loss Ib and the free-space definition of cross-section
F S :
1 F S 1
pr = pt gt D 2 i gr (8.62)
Ib 4
Ib
For the ITU-R version of loss Ib and the Shearman definition of cross-section
s :
1 s (1 + x )2 ) 1
pr = pt gt D 2 i gr (8.63)
Ib 4
Ib
For the Barrick version of loss IBar and the microwave definition of cross-
section T :
1 T (1 + x )2
pr = pt gt D 2 i (8.64)
IBar (1 + t )(1 + x ) 4 4
1
gr
IBar (1 + r )(1 + x )
For the Barrick version of loss IBar and the free-space definition of cross-
section F S :
1 F S 1
pr = pt gt D 2 i gr (8.65)
IBar (1 + t )(1 + x ) 4
IBar (1 + r )(1 + x )
For the Barrick version of loss IBar and the Shearman definition of cross-
section s :
1 s (1 + x )2
pr = pt g t D 2 i (8.66)
IBar (1 + t )(1 + x ) 4
1
gr
IBar (1 + r )(1 + x )
The maximum detection range of the surface wave emitter can be deter-
mined by substituting the receivers sensitivity
I = kT0 FR BRi (SNRRi ) (8.67)
for the return power pr (given by the six equations) and then solving for the
distance d. Figure 8.28 shows the results of two simulations (fc = 5 MHz and
fc = 15 MHz) using the surface wave MATLAB code surface detect.m on the
CD. Other (default) parameters include pt = 100 kW, gt = gr = gx 3 dB,
ht = hr = 10m, hx = 5m. The emitter input bandwidth BRi = 30 MHz and
the receiver has a noise factor of FR = 10 dB with kT0 = 4 1021 W/Hz.
Figure 8.28: Surface wave emitter maximum detection range as a function of

the required SNRRi for fc = 5 MHz and fc = 15 MHz.

Both surface wave and sky wave radars experience ionospheric clutter (un-
wanted backscatter from the moving ionosphere) that can sometimes hide
the presence of the desired target echoes. Sky wave radars can also suf-
fer significant polarization losses and focusing/defocusing problems due to
ionospheric eects. Military HF radars are susceptible to deliberate jamming
and consequently they are incorporating LPI waveforms and techniques to
prevent their detection while fulfilling their early warning role while avoiding
electronic warfare measures (for example, see [4, 15]).
Another significant problem in OTHR is robust high-resolution Doppler
processing of accelerating or decelerating targets. The ionosphere often mod-
ulates the return signals and spreads the Doppler, which makes it dicult
to detect targets. This Doppler eect arises during aircraft and ship target
maneuvers and during observations of rockets in boost phase and mid-course
phase flight. Most OTHR systems use classical Doppler processing, where
one Doppler spectrum is computed using one full CIT. Typically, the CIT is
on the order of 1100 seconds in OTHR systems. Some systems use over-
lapped Doppler processing to provide a spectrogram analysis of time-varying
Doppler [6].
Today, more than ever, the prediction of OTHR performance is important
especially for naval systems that are trying to avoid being detected at a long-
range distance. Sky wave simulation results show that for Rfootprint = 3500
km, the minimum required input SNR (SNRRi ) for fc = 14.5 MHz for T =
1, 10, and 100 m2 is 107, 97 and 87 dB, respectively, from Figure 8.16.
PROPLAB results have been shown for one day. The results however will
change depending on the time of year as well as the time of day.
References
[1] Wise, J.C., Summary of Recent Australian Radar Developments, IEEE
Aerospace and Electronic Systems Magazine, pp. 810, Dec. 2004.
[2] Kolosov, A. A., Over-the-Horizon Radar, Artech House Inc., Norwood, MA,
1987.
[3] Headrick, J. M., and Skolnik, M. I., Over-the-horizon radar in the HF band,
Proc. of IEEE, Vol. 62, No. 6, pp. 664673, 1974.
[4] Green, S. D., Kingsley, S. P., and Biddiscombe, J. A., HF radar waveform
design, Proceedings of the HF Radio Systems and Techniques Conf., 47
July, pp. 202206, 1994.
[5] Musa, M., and Salous, S., Ambiguity elimination in HF FMCW radar sys-
tems, IEE Proc. Radar, Sonar and Navig. Vol. 147, No. 4, pp. 182188,
Aug. 2000.
[6] Musa, M., and Salous, S., Evaluation of multiple WRF-HF-FMCW radar
waveforms, Proc. of the IEE HF Radio Systems and Techniques Conf., pp.
207211, 2000.
[7] Khan, R. H., and Mitchell, D. K., Waveform analysis for high-frequency
FMICW radar, IEE Proc.-F, Vol. 138, No. 5, pp. 411419, Oct. 1991.
[8] Liu, W. X., Lu, Y. L., and Lesturgie, M., Optimal sparse waveform design
for HFSWR system, International Waveform Diversity and Design Conf.,
pp. 127130, 48 June 2007.
[9] Guo, X., Ni, J. L., Liu, G. S., Ship detection with short coherent integration
time in over-the-horizon radar, Proc. of the International Conf. on Radar,
pp. 667671, 35 Sept. 2003.
[10] Anderson, S. J., Mei, F. J., and Peinan, J., Enhanced OTHR ship detection
via dual frequency operation, Proc. of the CIE International Conf. of Radar,
pp. 8589, 2001.
[11] Wei, Y., and Liu, Y., New anti-jamming waveform designing and processing
for HF radar, Proceedings of CIE International Conf. of Radar, pp. 281284,
2001.
[12] Frazer, G. J., Abramovich, Y. I., and Johnson, B. A., Spatially waveform
diverse radar: perspectives for high frequency OTHR, Proc. of the IEEE
Radar Conf., pp. 385390, 2007.
[13] Topliss, R. J., Maclean, A. B., Wade, S. H., Wright, P. G., and Parry, J. L.,
Reduction of interference by high power HF radar transmitters, Proc. of
the IEE HF Radio Systems and Techniques Conf., No. 411, pp. 251255,
July 1997.
[14] Turley, M. D. E., FMCW radar waveforms in the HF band, Input to the
ITU-R JRG 1A-1C-8B meeting, November 2006.
http://www.its.bldrdoc.gov/meetings/itu-r/contributions4/jrg-73.pdf.
[15] Xianrong, W., Hengyu, K., and Biyang, W., Adaptive cochannel interference
suppression based on subarrays for HFSWR, IEEE Signal Processing Letters,
Vol. 12, No. 2, pp. 162165, Feb. 2005.
[16] Lu, K., Liu, X., and Liu, Y., Ionospheric decontamination and sea clutter
suppression for HF skywave radars, IEEE Journal of Oceanic Engineering,
Vol. 30, No. 2, pp. 455462, April 2005.
[17] Guo, X., Hongbo, S., and Yeo, T. S., Transient interference excision in over-
the-horizon radar using adaptive time-frequency analysis, IEEE Trans. on
Geoscience and Remote Sensing, Vol. 43, No. 4, pp. 722735, April 2005.
[18] Tarran, C. J., Operational HF DF systems employing real time superresolu-
tion processing, Proc. of the IEE HF Radio Systems and Techniques Conf.,
No. 411, pp. 311319, July 1997.
[19] Dawber, W. N., Pote, M. F., Turner, S. D., Graddon, J. M., Barker, D.,
Evans, G., and Wood, S. G., Integrated antenna architecture for high fre-
quency multifunction naval systems, Proc. of the CIE International Conf.
of Radar, pp. 15, 2006.
[20] Huang, G., Meng, J., and Yang, L., Time-delay estimation for sub-sampling
sinusoidal signals, Proc. of the International Conf. on Communications,
Circuits and Systems, pp. 761764, 2005.
[21] Xiaodong, T., Yunjie, H., and Wenyu, Z., Sky wave over-the-horizon backscat-
ter radar, Proceedings of the CIE International Conf. on Radar, pp. 9094,
1518 Oct. 2001.
[22] Liu, B-Y, Pace, P. E., and Knorr, J. B., HF skywave FMCW OTH-B systems
expected emitter footprint, Proc. of the IEEE System of Systems Engineer-
ing Conf., Monterey, CA, June 2008.
[23] Kingsley, S., and Quegan, S., Understanding Radar Systems, Scitech, Mend-
ham, NJ, 1999.
[24] Davies, K., Ionospheric Radio, IEE Electromagnetic Waves Series, Vol. 31,
IEEE Electromagnetic Waves Series, 1990.
[25] http://www.utdallas.edu/research/spacesciences/ionosphere.htm
(The Earths Ionosphere).
[26] Dyson, P. L., and Bennett, J. A., Exact ray path calculations using realistic
ionospheres, IEE Proc. - H, Vol. 139, No. 5, pp. 407413, Oct. 1992.
[27] Ong, C. Y., Bennett, J. A., and Dyson, P. L., An improved method of
synthesizing ground backscatter ionograms for spherical ionospheres, Radio
Science, Vol. 33, No. 4, pp. 11731185, 1998.
[28] Croft, T. A., and Hoogasian, H., Exact ray calculations in a quasi-parabolic
ionosphere with no magnetic field, Radio Science, Vol. 3, No. 1, pp. 6974,
Jan. 1968.
[29] http://nssdc.gsfc.nasa.gov/nssdc news/june01/iri.html Ionospheric Model
(IRI-2001).
[30] http://modelweb.gsfc.nasa.gov/models/iri.html IRI Model, Space Physics
Data Facility (SPDF).
[31] Jordan, E. C., and Balmain, K. G., Electromagnetic Waves and Radiating
Systems, Prentice Hall, Inc., Englewood Clis, NJ, 1968.
[32] Barnes, R. J., Automated propagation advice for OTHR ship detection,
IEE Proc. Radar, Sonar, Navig., Vol. 143, No. 1, pp. 5363, Feb. 1996.
[33] Gill, E. W., Howell, R. K., Hickey, K., Walsh, J. and Dawe, B. J., High fre-
quency ground wave radar measurement of ocean surface parameters during
the ERS-1 calibration-validation experiment, Proc. of OCEANS 93 Engi-
neering in Harmony with Ocean, pp. I55I60, 1821 Oct. 1993.
[34] Georges, T. M., and Harlan, J. A., New horizons for over-the-horizon radar,
IEEE Antennas and Propagation Magazine, Vol. 36, No. 4, pp. 1424, Aug.
1994.
[35] Jindalee Operational Radar Network (JORN), Janes Radar and Electronic
Warfare Systems, Land-Based Air Defence Radars, January 10, 2007.
[36] Gurgel, K-W, Essen, H-H, Schlick, T., HF surface wave radar for oceanography
a review of activities in Germany, Proc. of the International Radar Conf.,
pp. 700705, 2003.
[37] Hartnett, M. P. and Clancy J. T., Utilization of a nonrecurrent waveform to

mitigate range-folded spread Doppler clutter: Application to over-the-horizon
radar, Radio Science Vol. 33, No. 4, pp. 11251133, 1998.
[38] Zhang, Y., Amin, M. G. Amin and Frazer, G. J., High-resolution time-
frequency distributions for manoeuvring target detection in over-the-horizon
radars, IEE Proc. Radar Sonar Navig., Vol. 150, No. 4, pp. 299304,
Aug. 2003.
[39] Pace, P. E., Advanced Techniques for Digital Receivers, Artech House Inc.,
Norwood, MA, 2000.
[40] Chinese OTH Radar, Janes C4I Systems, Intelligence Systems in Direction
Finding, June, 2000.
[41] Li, N.-J., A review of Chinese designed surveillance radarspast, present
and future, Record of the IEEE International Radar Conf., pp. 288293,
Alexandria, VA, May 8-11, 1995.
[42] Krolik, J. L., Anderson, R. H., Maximum likelihood coordinate registration
for over-the horizon radar, IEEE Trans. on Signal Processing, Vol. 45, No.
4, pp. 945959, April 1997.
[43] Torrez, W. C., and Blasch, E., An application of generalized least squares
bias estimation for over-the-horizon radar coordinate registration, Proc. of
the Third International Conf. on Information Fusion, Volume 1, 1013 July,
2000.
[44] Min, K., Wang, G-H, and Wang, X-B, Coordinates registration and error
analysis based on spherical model for OTH radar, Proc. of the International
Conf. on Radar, pp. 14, Oct. 2006.
[45] PROPLAB-PRO version 2.0 Users Manual, High Frequency Radio Program
Laboratory, 1994-1997.
[46] http://areps.spawar.navy.mil/
AREPS (Advanced Refractive Eects Prediction System)
[47] SWR series High-Frequency Surface Wave Radars (HFSWR), Naval/Coastal
Surveillance and Navigation Radars, Canada, Janes Radar And Electronic
Warfare Systems, April 10, 2007.
[48] Yang, S., Ke, H., Wu, X., Tian, J., and Hou, J., HF radar ocean current
algorithm based on MUSIC and the validation experiments, IEEE Journal
of Oceanic Engineering, Vol. 30, No. 3, pp. 601618, July 2005.
[49] Kahn, R., Gamberg, B., Power, D., Walsh, J., Dawe, B., Pearson, W., and
Millan, D., Target detection and tracking with a high frequency ground wave
radar, IEEE Journal of Oceanic Engineering, Vol. 19, No. 4, pp. 540548,
Oct. 1994.
[50] Green, S. D., and Kingsley, S. P., Improving the range/time sidelobes of
large bandwidth discontinuous spectra HF radar waveforms, Proc. of the
IEE HF Radio Systems and Techniques, No. 411 00. 246250, July 1997.
[51] Yiying, S., Ning, Z., and Yongtan, L., New waveform with both high range
resolution and long coherent integration time in a HF radar, Proceedings of
CIE International Conf. of Radar, pp. 285288, 810 Oct. 1996.
[52] Zongchuang, L., Zingzhao, L., and Yongtan, L., A signal detection algorithm
based on higher-order statistics for HFSW-OTH radar, Proceedings of CIE
International Conf. of Radar, pp. 9961000, 2001.
[53] Milsom, J. D., HF groundwave radar equations, Proc. of the IEE HF Radio
Systems and Techniques, No. 411, pp. 285290, July 1997.
[54] Barrick, D. E., First order theory and analysis of MF/HF/VHF scatter from
the sea, IEEE Trans. on Antennas and Propagation, AP-20, pp. 210, 1972.
[55] Barrick, D. E., Theory of HF and VHF propagation across the rough sea,
Parts I and II, Radio Science Vol. 6, No. 5, pp. 517533, 1971.
Problems
1. A targets range is measured with a t0 = 200 ms and shows up in range
bin 15. If the target has a Doppler shift of 20 Hz, what is the corrected
range bin of the target?
2. Wideband clutter covers 200500 km. If a FMCW waveform with
F = 10 kHz is used, in what range bins will the clutter appear?
3. In the multiple WRF technique, if the F = 10 kHz and if the num-
ber of range bins covered by the clutter is NRB = 40 (a) what is the
maximum integer value for WRF2 ? and (b) what are the three WRFs?
4. Consider a multiple WRF FMCW radar with a bandwidth
F = 20 kHz and WRF1 = 7 Hz, WRF2 = 8 Hz, and WRF3 = 9 Hz
and an integration time of 150s (each block is 50s). For the first block
(WRF1 ), determine (a) the total number of range bins, (b) the range
resolution, and (c) the corresponding maximum unambiguous range.
(d) Repeat (a)(c) for WRF2 and WRF3 . (e) What is the maximum
clutter width (in km) in order that sea echoes and other unambiguous
targets can be detected?
Chapter 9
Case Study: Antiship LPI

Missile Seeker
In Chapter 1 we examined the characteristics that make a radar LPI, and in
Chapter 2 we looked at a number of important applications. A significant
advantage can be gained over the noncooperative intercept receiver when the
radar uses frequency, phase, and hybrid wideband waveform coding tech-
niques. In Chapter 3 the periodic ambiguity function was presented as a
means of quantifying the characteristics of the LPI waveforms, which are dis-
cussed in Chapters 48. In this chapter, we bring some of these concepts
together in a case study that examines the detection capability of a power-
managed LPI antiship cruise missile seeker. In the scenario examined, the
ASCM has an FMCW seeker that attempts to detect and track a low radar
cross section ship at the horizon. RCS values considered include 50, 100, and
500 m2 . To predict the target detection capability, sea clutter models are
developed, and the emitter is flown at 300 m/s in a scenario that starts at a
range of 28 km from the target. Each sea state (04) is characterized by a
second-order polynomial that describes the normalized mean sea backscatter
coecient as a function of the grazing angle. The emitter transmit power
level is adapted to be consistent with the RCS and range to the target, while
keeping the output signal to noise ratio SNRRo at 20 dB. A brief history of
ASCM seeker technology is given first.
9.1 History of ASCM Seeker Technology

Antiship cruise missiles have been a significant threat to navy surface ships
for many years. The first generation of ASCM threats (prior to 1969) used a
single-frequency RF pulse with a constant pulse repetition interval [1]. The
301
antenna scanned mechanically and was very susceptible to electronic attack.

Second generation ASCMs (19691979) were marked by an increase in their
electronic protection capability. Monopulse processing and the use of discrete
computer circuits allowed additional RF capability, using both staggered and
jittered PRI. In the third generation (19791989), the integration of multiple
sensors was introduced along with the use of complex RF modulations and
frequency agility. Coherent Doppler processing and a large amount of elec-
tronic protection were also incorporated. Computer-aided design also started
to play a significant role. In the fourth generation (1990 to the present), the
signal processing performed by the missile has been vastly improved. The
use of embedded computers allows the seekers to do imaging and interleave
modes such as track-while-scan. Dual-mode infrared (IR) and millimeter
wave seekers using wideband frequency agility demonstrate the capabilities
of these fourth-generation ASCMs. The most important trends under way
are the use of LPI seekers with hybrid combinations such as LPI/imaging IR,
LPI/antiradiation (ARM), and millimeter wave LPI/imaging IR.
9.2 The Future for ASCM Technology

Stealthy ship designs, such as the Lafayette-class frigate shown in Figure 9.1,
are a response to the ASCM threat. Future ASCM threat technology will be
the result of the balance between the available technology and required littoral
warfare capabilities, and the aordability and export sales potential that
exist. The number of development programs for cruise missiles has greatly
increased, following the publicity given to the use of the Tomahawk missiles
during the 1991 Gulf War [2]. Including the United States, 19 countries
now have cruise missile programs with missile ranges extending to 3,000 km
(e.g., the Chinese HN-3, and the Russian AS-15C and Kh-101). Ship-based
helicopters capable of firing ASCMs, such as the Saudi Dauphin II shown in
Figure 9.2, are already gaining popularity in the international market.
The capabilities that will be required for ASCMs in the future include
fire-and-forget and man-in-the-loop. Cruise missiles such as the Chinese C802
ASCM are all-weather, fire-and-forget missiles that have a range greater than
400 km. The seekers will have the capability to select a target very accurately,
ignoring any land clutter or other ships in the vicinity, and they will also
ignore any decoys that are used. To defeat the protection systems that could
possibly shoot down the missile on its way to the target, programmable way
points will be commonly used in antiship threats, such as the Swedish RBS-15.
The capability will also exist to come back around for a reattack in case
the ship was missed. In order to strike the ship at the most vulnerable point,
future ASCMs will have programmable aimpoints, along with the ability to
adjust their attack aspect. With high-G maneuverability, integrated EA and
self protection, and a stealthy cross section (with regard to both RF and IR),
Case Study: Antiship LPI Missile Seeker 303
Figure 9.1: Stealthy ship designs, such as this Lafayette-class frigate, are a
response to the ASCM threat [1]. (Source: Horizon House c 1998. Reprinted
with permission.)
Figure 9.2: Saudi Dauphin II firing an antiship cruise missile [1]. (Source:
Horizon House c 1998. Reprinted with permission.)
the future ASCM penetration capability will be ominous. The capability to

perform coordinated salvos (coordination of multiple launch platforms) with
simultaneous arrival times will also increase the probability of hitting the
ship. Several ASCMs approaching a ship from multiple aspects coordinated
as a stream raid seem to make the odds of ship survival remote at best.
A number of advanced electronic technologies such as field-programmable
gate arrays in GaAs, and application-specific integrated circuits, will be used
in future ASCMs. Low-probability-of-intercept, power-managed seekers op-
erating in the 8- to 20-GHz range, as well as the 35- and 96-GHz ranges will
use pulse-to-pulse spread-spectrum modulation with coherent range Doppler
processing to target the ships. Phased arrays and active arrays will refine the
targeting capability, allowing variable sectors to be scanned using multiple
beams. Exclusion zones will also provide the ability to reject any decoys that
might interfere with the target kill. The active seekers will also use other
sophisticated modes, such as Doppler beam sharpening, unfocused synthetic
aperture radar, and inverse synthetic aperture radar. Wideband ARM seekers
with phased arrays will be able to easily recognize and discriminate targets,
and will be especially robust against emissions control tactics (RF < 110
dBm). When faced with emissions-control tactics, these ARM seekers will
also use a technique called loitering, in which a parachute is deployed to slow
the missile until the radiation source comes back up.
Although not as prevalent today as RF seekers, infrared seekers are fast
becoming a force to be reckoned with. By 2010, large-scale InSb, HgCdTe,
and PtSi imaging arrays on the order of 1,024 1,024 will be available. New
cooling techniques and the development of detectors that do not require cool-
ing will improve the fidelity of the IR images. Improvements in IR dome
materials, and protection against laser jammers and interrogators, will allow
the seeker to be very robust in the terminal phase of attack. Future dual-mode
technologies that will appear include ARM/millimeter-wave and ARM/laser
radar using high-accuracy inertial systems for the terminal phase. For flight
guidance, most new missiles are using an inertial navigation system (INS)
together with global positioning system updates. Whereas a modern INS will
have an accuracy of about 2 km per flight hour at around Mach 0.85, a com-
bined INS/GPS would have an accuracy of around 50 to 100m circular error
probable (CEP)1 regardless of flight time. The use of fiber optics (instead of
copper wire) will also make the weapon lighter and increase the range that it
can travel.
The RBS-15 shown in Figure 9.3 is an example of an advanced missile
1 A circular error probable is determined by a series of flight tests, and is usually cal-
culated by taking the square root of the sum of the squares of the range and track errors.
The resultant CEP indicates the radius of the circle that encompasses half the impact
points during the flight tests. The other half of the results could spread out to many
times the CEP radius, and the CEP may be dierent for dierent ranges, flight profiles or
target sets [2].
Figure 9.3: Swedish RBS-15 ASCM in flight [3] ( c 2002 Janes Information
Group).
that is designed to operate in all surface attack roles, from littoral warfare to
blue-water situations, and includes a day-and-night, all-weather, land-attack
capability. Its long range (over 200 km) and flexible trajectory mean that
it can attack hostile vessels well beyond the horizon, but also at very close
ranges. The missiles, each individually prepared in a salvo, can be prepro-
grammed to enable attacks to be mounted from dierent directions, with a
preselected time of arrival for each missile to confuse air defenses. Using an
advanced missile engagement planning system, the missiles can make use of
terrain masking for a concealed approach, to minimize warning time.
9.3 Detecting the Threat

To adequately defend the ship, the ASCM must be detected before it comes
over the horizon. The detection of an incoming cruise missile seeker at the
horizon ( 24 km) is dicult with modern ES intercept receivers. Since the
missile usually flies just above the water surface, it is hard to detect and
extract it from the clutter using radar, since the RCS can be very small. The
ASCM is also hard to detect with infrared sensors. It can possibly be detected
when the seeker turns on, but that does not give much time for the ships self
defense.
Ships receive insucient warning against the missiles being developed to-
day. An even greater problem exists within the littoral theater, where any-
thing can be fired in short order [4]. Detection techniques being researched
today extend the first engagement of the cruise missile out to 300 km, ex-
panding the ASCM area-defense capability. To detect the missile at this
range requires an airborne adjunct system capable of 3-D surveillance with a
high-fidelity tracking capability. The concept of using an airborne platform
to guide ship-launched missiles for intercepting low-flying targets beyond the
ships horizon was considered over two decades ago. Since the completion
of the Mountain Top tests [5], infrared search and track (IRST) surveillance
technology has also been investigated for airborne ASCM detection. Long-
wave IR focal plane arrays oering 640 480 resolution were developed in
1994, with the resolution expected to exceed 960 1,280 by 2005. Sensor
platforms include, for example, the E2-C Hawkeye flying at altitudes above
7,600m, extending the engagement range out past 220 km.
9.4 ASCM Target Scenario

In this case study we have used an FMCW waveform to illustrate the power
management LPI technique, and assume that the problems of transmitter-
receiver isolation and transmitter phase noise can be solved satisfactorily in
the single-antenna seeker environment, even with the reflected power from
antenna and radome mismatch near the 0.1W level. The ASCM-target sce-
nario being evaluated in this case study was introduced in Chapter 4, and
is shown in Figure 9.4. The FMCW technique separates the target echo in
frequency from the transmission by a significant fraction of the modulation
bandwidth, while tolerating relatively high levels of transmitter leakage into
the receiver. Practical solutions to the leakage problems are becoming avail-
able to the seeker designer such as those discussed in Section 4.4. The missile
contains an FMCW LPI emitter, and flies at the ship starting at a range-
to-target of R = 28 km at a height of 70m o the surface of the water at a
speed of Mach 0.9 (300 m/s). Below the target model, the sea clutter model
and the emitter model are described. Simulation results are described to pre-
dict the detection performance of the emitter. Note that this is a first-order
analysis, and the results shown do not include any standard or nonstandard
propagation eects such as spherical spreading and ducting [6, 7].
9.4.1 Low RCS Targets

The future design of naval vessels will have a low RCS in addition to other
signature reduction techniques. For example, the HMS Visby is the first of
five Visby-class stealth corvettes under construction for the Swedish Navy
by HDW-owned Kockums shipyard in Karlskrona, Sweden. It is shown in
Figure 9.5 performing high-speed sea trials o the German Baltic coast [3].
The stealth corvette has a length of 72m, is constructed almost entirely
from carbon fiber-reinforced plastic material, and features a variety of inno-
vative signature-reduction techniques covering radar cross-section, infrared,
acoustic, magnetic, and hydrodynamics.
The most favorable situation for the seeker is when the ship is broadside
(largest return within a range bin). To detect a stealthy target, the CW
frequency should be between 30 and 960 MHz. This is also the frequency
range that contains public broadcasting and mobile communication systems.
At broadside, however, there is no Doppler separation between the ship and
the clutter, which is why most nonimaging missile seekers do not use Doppler
Figure 9.4: ASCM-to-target scenario.
Figure 9.5: The Swedish stealth corvette HMS Visby conducting high-speed
trials in the Baltic. (Source: Michael Nitz. c 2002 Janes Information
Group.)
Table 9.1: Normalized Mean Sea Backscatter Coecients for Grazing and 0.1
to 10.0 Degrees for Sea States 0 to 4 in Decibels Below 1 m2 /m2 for 9.3
SEA STATE
Grazing Angle
(degrees) 0 1 2 3 4
0.1 70 60 56 51 48
0.3 62 58 52 45 43
1.0 57 50 44 39 37
3.0 52 45 41 38 35
10.0 46 42 36 32 29
processing. Also note that since the emitter is at a height of 70m within the
model, the waterline of the ship is visible throughout the entire flight. If the
height of the emitter is lower, then the waterline only becomes visible at a
closer range. For example, with a sea-skimming missile at a height of 9m,
the waterline is visible at a max range of about 13 km. This is why most
seekers turn on at 713 km (in addition to minimizing the time the ship has
to react). For the detection analysis below, we choose 50, 100, and 500 m2 as
examples of low RCS values for the ship target.
9.4.2 Sea Clutter Model

To model the sea clutter, a set of normalized mean sea backscatter coecients
for low grazing angles and sea states is given in Table 9.1. This figure gives
the normalized mean sea backscatter coecients for grazing angles 0.110.0
degrees for sea states 0 to 4 in decibels below 1m2 /m2 for 9.3 GHz, vertical
polarization (adapted from [8]). Sea clutter exhibits very dierent spectral
characteristics at higher frequencies compared to those at low frequencies.
Some of the values have been estimated, since errors in the reported values
are not unlikely, and some values are not reported at all (especially for low
grazing angles) [8]. Other experimental sea clutter coecients as a function
of the grazing angle for various frequencies are reported in [9, 10].
To extract the correct value for the mean sea backscatter coecient 0i
as a function of the grazing angle, a polynomial was developed for each sea
state
0i = A2 + B + C (dB below 1 m2 /m2 ) (9.1)
where i is the sea state, is the grazing angle (in radians), and the coe-
cients A, B, and C are given in Table 9.2. Using these coecients, the value
of 0i for the five sea states (00 04 ) is shown in Figure 9.6. It is in-
Table 9.2: Polynomial Coecients A, B, and C to Determine 0i as a

Function of the Sea State (i)
i A B C
0 1.8289 (104 ) 1.1146 (104 ) 2.5296 (108 )
1 2.0882 (103 ) 7.3396 (104 ) 1.4661 (106 )
2 9.7730 (104 ) 1.5948 (103 ) 2.1903 (106 )
3 4.6285 (103 ) 2.6412 (103 ) 2.6779 (105 )
4 9.0787 (103 ) 5.3639 (103 ) 3.5646 (105 )
Figure 9.6: Normalized mean sea backscatter coecient 0i as a function of

the grazing angle.
teresting to note that the sea state of one backscatter coecient does not
increase as dramatically as the other sea states when the grazing angle gets
larger. Also note that these clutter calculations minimize the fact that the
clutter distribution becomes highly non-Gaussian at low grazing angles, due
to sea spikes. For increased accuracy, lognormal, Weibull, or K-distributed
analysis can be used and will typically increase the required SNR by 1020
dB. In the next section, the transmitter power management is discussed. The
backscatter coecient polynominals are then used to predict the detection
capability of the LPI emitter design for three RCS values (50, 100, 500 m2 ).
9.4.3 Linear FMCW Emitter Power Management

In a power-managed seeker, the emitter transmits a power level consistent
with the RCS and range of the target to be detected, thereby keeping the
SNR a constant. When the missile first enters the engagement envelope and
turns on the seeker, the most probable range cells are monitored, and the
transmitter output power is at a level where detection of the target can be
made on a single scan using only a few modulation periods. If the target is
not detected, the power can be increased gradually on the next scan. Once
the target is detected, the seeker shuts down and the missile moves up to a
new location, making any subsequent targeting or exploitation of the emitter
impossible. The emitter uses computer control so that the RF energy is only
emitted when it is necessary to measure the target characteristics and update
the track file.
From (1.28), the average transmit power of an FMCW emitter can be
written as
w W w 4 W
(4)3 kT0 FR Lf RT4 SNRRo RT SNRRo
PCW = =K (9.2)
G2t 2 T T
where L2 1, FR is the receiver noise factor, kT0 = 4.0 1021 W/Hz,

L = LRT LRR represent the system losses, SNRRo is the required output
signal-to-noise ratio required for target detection, f = 1/tm is the filter
bandwidth, RT is the range to the target (ship), and T is the ships RCS.
Continuing with the FMCW example discussed in Chapter 4,
fc = 9.3 GHz and tm = 1 ms. The modulation bandwidth is chosen as
F = 15 MHz in order to provide a R = 10m such that the ship return
all lies within a range bin. Also, FR = 10 and L = 10, which is reasonable
for a single antenna implementation. The antenna is a circular aperture with
diameter da = 0.3m with uniform illumination. This antenna diameter will
easily fit in the nose of an ASCM pod. The 3-dB beamwidth in the azimuth
plane depends on the aperture size and is approximately [11]

a = 1.29 rad (9.3)
da
Table 9.3: Summary of LPI Emitter Parameters Used in Simulation Results
Carrier frequency fc 9.3 GHz

Modulation period tm 1.0 ms
Coherent processing interval t0 0.81 ms
Modulation bandwidth F 15 MHz
Eective modulation bandwidth F 12.2 MHz
Range resolution R 10.0m
Eective range resolution R 12.3m
FFT size N 8,192
Time bandwidth product t0 F 9,922
Average transmitter power Pavg Adaptive
ADC sampling speed fs 10.1 MHz
Detection signal-to-noise SNR 20 dB
Noise factor F 10
Range-doppler gate f 1.23 kHz
System losses L 10
Antenna diameter da 0.3m
Antenna eciency 0.90
Beamwidth a 7.9
Antenna gain Gt 810
or 7.9 degrees. The gain Gt of a circular aperture antenna at X-band is

approximately
Gt d2a (9.4)
where da is in cm and is the aperture eciency. For a uniform illumination,
= 0.9 considering a 0.5-dB loss in an equal-level feed network. This results
in an antenna gain of Gt = 810.
Typically, the SNR for a CW emitter must be at least 6 dB (rather than
the 13-dB value required for detecting steady targets with pulse emitters;
see p. 449 in [8]). However, a more realistic value for ASCM seekers is
20 dB. The emitters PCW is adjusted as a function of the range-to-target in
order to keep the SNR = 20 dB. That is, the intelligent power management
automatically adjusts the transmit power to maintain a constant IF SNRRo =
20 dB. This value, of course, depends on the postdetection integration, which
can be calculated per the equations given in Chapter 1. A summary of the
LPI emitter parameters is given in Table 9.3 for a F = 15-MHz design
(Example 1 in Table 4.1).
The corresponding average power transmitted by the emitter as a function
of range-to-target is shown in Figure 9.7 for RCS values of 50, 100, and 500
m2 . Note that the transmitted power is adaptive, and calculated to keep
Figure 9.7: Average power transmitted by LPI emitter as a function of range

to the target and target RCS value.
the total echo power from the target a constant. Also note that only for the
50 and 100 m2 case does the transmit power exceed 100W. From (1.34) the
targets echo power from the CW emitter (with L2 = 1 and LRT = LRR = 1)
is
PCW G2t 2 T
PRT = W (9.5)
(4)3 RT4 L
and is a constant (PRT = 140 dBW). In summary, the seeker uses the
FMCW waveform and adaptive power management to achieve the LPI char-
acteristics. With higher RCS targets, the transmitted power can be reduced
even further.
9.4.4 Target-to-Clutter Ratio

Since the primary purpose of the FMCW emitter is to detect and track ship
targets in the presence of sea clutter, the target-to-sea clutter ratio within
a range bin is examined for sea states 0 to 4 and ship RCS values 50, 100,
and 500 m2 . Using the backscatter coecients 0i , the power of the clutter
within the targets range bin can be estimated as
PCW G2t 2 0i RT a R
PRC = W (9.6)
(4)3 RT4 L
Figure 9.8: Target-to-clutter ratio for R = 10.09m and T = 50 m2 .
Using (9.5) and (9.6), the target-to-clutter ratio (TCR) is

w W
PRT T 1
TCR = = (9.7)
PRC 0i R a RT
and is shown in Figure 9.8 for R =10.09m, and a target RCS of T = 50
m2 . Note that the TCR curve for sea state 1 is not quite accurate. If the
required TCR for detection and tracking is 20 dB, Figure 9.8 shows that
the target can be tracked throughout the flight of the missile in sea states 0
to 2 (TCR >20 dB). Detection becomes more dicult for higher sea states.
Detection is possible at R < 4 km for sea state 3, with no detection capability
in sea state 4. Figure 9.9 shows the target-to-clutter ratio for R = 10.09m
and T = 100m2 . Note that the detection capability is now possible in sea
state 3 when R < 15 km and sea state 4 when R < 6 km. Figure 9.10 shows
the target-to-clutter ratio for R = 10.09m and T = 500m2 . Here it is
shown that detection is possible in all sea states at all ranges with a required
TCR > 20 dB. The detection capability for the FMCW emitter is summarized
in Table 9.4. Note that only for sea state 4 and RCS = 50 m2 is the target
not detectable. Also, it is assumed that the targets RCS is totally contained
within the range bin. The maximum value of PCW required throughout the
flight is also summarized.
Note that the inclusion of propagation loss standard and nonstandard

Table 9.4: Summary of Detection Results for Power-Managed FMCW Emitter

TCR > 20 dB
Sea State RCS = 50 m2 RCS = 100 m2 RCS = 500 m2
0 Yes Yes Yes
1 Yes Yes Yes
2 Yes Yes Yes
3 Yes (R < 4 km) Yes (R < 15 km) Yes
4 No Yes (R < 6 km) Yes
Max Pavg 110 W 100 W 90 W
mechanisms will aect these first-order results. Inclusion of these eects

must also be included for an accurate prediction of the weapons system
performance. Assessment programs such as the advanced refractive eects
prediction system (AREPS) [12] can easily provide such results.
9.5 ASCM Ship Target Model

Recall that the two MATLAB programs from Chapter 4:
lpi fmcw design.m
receiver design.m
can be used to design and simulate the FFT processing of the FMCW radar,
to extract the range and range rate of a target. The results shown in this
chapter use the design parameters generated by lpi fmcw design.m and the
sea clutter backscatter coecients generated by clutter polynomial x.m. The
missile-to-target scenario results are derived by ascm.m. The variable inputs
are: (1) ship RCS, (2) sea state (0, 1, 2, 3, or 4), (3) missile velocity, and
(4) initial range of the missile. Variable inputs for the seeker include (1) the
required TCR for detection, (2) the diameter of the seeker antenna, (3) the
seeker noise factor, and (4) the seeker losses. All other input values for the
emitter come from lpi fmcw design.m as global variables.
References
[1] Pace, P. E., and Burton, G. D., Antiship cruise missiles: Technology, simu-
lation and ship self-defense, Journal of Electronic Defense, Vol. 21, No. 11,
Nov. 1998.
[2] Lennox, D. Cruise missile technologies and performance analysis, Janes
Strategic Weapons Systems 38, Nov. 2002.
[3] Janes International Defence Digest, Sept. 2002.
[4] Jurcheck, J., Visualizing the littoral battlespace, Surface Warfare, Vol. 22,
pp. 1015 Aug. 1997.
[5] Zinger, W. H., and Krill, J. A., Mountain top: Beyond-the-horizon cruise
missile defense, Johns Hopkins APL Technical Digest, Vol. 18, No. 4, pp.
501520, 1997.
[6] Hitney, H. V., Refractive Eects from VHF to EHF Part A: Propagation
Mechanisms, Propagation Modeling and Decision Aids for Communications,
Radar and Navigation Systems, NATO AGARD Lecture Series 196, Ottawa,
Canada, Oct. 1994.
[7] Hitney, H. V., Refractive Eects from VHF to EHF Part B: Propaga-
tion Models, Propagation Modeling and Decision Aids for Communications,
Radar and Navigation Systems, NATO AGARD Lecture Series 196, Ottawa,
Canada, Oct. 1994.
[8] Nathanson, F. E., Radar Design Principles, Second Edition, McGraw-Hill
Inc., New York, 1991.
[9] Paulus, R. A., Evaporation duct eects on sea clutter, IEEE Trans. on
Antennas and Propagation, Vol. 38, No. 11, pp. 17651771, Nov. 1990.
[10] Chan, H. C., Radar sea-clutter at low grazing angles, IEE Proc. Part F,
Vol. 137, No. 2, pp. 102112, April 1990.
[11] Barton, D. K., Modern Radar Systems Analysis, Artech House, Inc., Nor-
wood, MA, pp. 155, 1988.
[12] Defense information and infrastructure common operating environment
Users manual for advanced refractive eects prediction system, SPAWAR
Systems Command, METOC Systems Program Oce, Jan. 9, 2003.
Problems
1. It is an easy matter to modify the second-order polynomial describing
the normalized mean sea backscatter 0i as a function of grazing an-
gle for the five sea states (useful when better empirical data might be
obtained). Recall that these coecients were derived by curve fitting
the values given by Nathanson [8]. The polynomial coecients (p0p4)
are used in ascm.m and can be regenerated by adjusting the backscat-
ter coecients in clutter polynomial x.m (y0bs through y4bs). (a) Run
lpi fmcw design.m to design the F = 15 MHz LPI seeker discussed
in Section 7.4. (b) Edit the file clutter polynomial x.m and change the
sea state three normalized mean sea backscatter values to
y3bs = [53 46 40 39 37] (9.8)

(horizontal polarization). (c) Run the scenario discussed above using

ascm.m and compare the detection performance against the vertical po-
larization results given in the text (plot the results from both designs
on the same graph). (d) Summarize the detection range for this hori-
zontally polarized seeker.
2. Using the programs lpi fmcw design.m, clutter polynomial x.m and ascm.m,
(a) summarize the detection capability of the 9.3-GHz seeker discussed
in Section 9.4 if the modulation bandwidth is changed to F = 25
MHz. (b) summarize the detection capability if the seeker in (a) keeps
the SNR = 13 dB (instead of 20 dB).
Chapter 10
Network-Centric Warfare
and Netted LPI Radar
Systems
LPI radar systems can be networked together into a system of systems to
covertly gather and share surveillance and targeting data as part of a network-
centric warfare architecture. In this chapter, network-centric warfare concepts
are introduced including the information grid (network), the sensor grid, and
the weapons grid. A set of metrics is presented to quantify the value added
to an operation by the network. Electronic attack on the network is also con-
sidered. Advantages of netted LPI radar systems (part of the sensor grid) are
discussed, including the improvement in emitter sensitivity that is gained, and
a multiple-input multiple-output (MIMO) signal model is presented. Network
analysis and netted radar system analysis are presented. Simulation results
using LPIsimNet are shown. LPIsimNet is a MATLAB program included
with the CD that allows the user to evaluate any general netted radar con-
figuration and the operational performance of a sensor network. Orthogonal
PSK, FSK, and noise waveforms for netted LPI radar applications are also
presented. Use of MIMO techniques for OTHR is discussed.
10.1 Network-Centric Warfare

In a platform-centric naval architecture the aircraft carrier is the epicenter
of power. Each weapon has its own sensor and if that sensor is remote there is
a stovepipe communication system transmitting the data back to the shooter.
Ultimately, this makes necessary many platforms to eectively project the
319
power needed on the battlespace. Platform-centric command and control

(C2) also suers limitations in its ability to coordinate operations. For ex-
ample, in suppression of enemy air defense (SEAD) operations, the stando
jammer suers significant geometrical problems in alignment, making it di-
cult to detect and jam coherent threats. The geometrical limitations result in
an extended stando range being necessary and consequently an ineective
jammer management.
Over the past several years there has been a major shift away from
platform-centric warfare concepts. Currently we no longer have access to
masses of ships and large numbers of weapons. Also, the weapons today are
precise and must be employed at exactly the right time and place. No single
sensor has the precision to target these advanced technology eectors. They
require a dynamic knowledge of the target set and the integration of many
sensors and databases.
Evolving from platform-centric warfare, network-centric warfare (NCW)
integrates a distributed system of C2, sensors and weapons called a grid.
NCW can extend the capabilities of sensors and weapons across all the plat-
forms on the network to pursue the maximum eciency in mission execution.
The grid provides the capability to collect, process and disseminate an unin-
terrupted flow of C2, sensors and weapons information between nodes while
exploiting and denying the adversarys ability to do the same (information
superiority)[1]. In contrast to platform-centric warfare, which has an additive
eect on combat power (N nodes, total force value = N ), NCW has an expo-
nential eect (N nodes, total force value = N 2 ). The exponential advantage
in total force value also gives a maneuver and time-critical strike advantage
to the NCW nodes. A more formal definition of NCW is given below.
Definition 10.1
Network-centric warfare is military operations that exploit state-
of-the-art sensor information and networking technologies to in-
tegrate widely dispersed human decision makers, weapons, situ-
ational and targeting sensors and forces into a highly adaptive
comprehensive system to achieve unprecedented mission eective-
ness.
The NCW grid is composed of three subgrids; the global information grid,
the sensor grid and the shooter grid as shown in Figure 10.1 [1]. The global
information grid is a deployed tactical sensor and weapons network that pro-
vides the infrastructure for plug-and-play of sensors and shooters. It exists
in space, low- and high-Earth orbit, and at all altitudes on land and un-
dersea. It is a physical, permanent and fault-tolerant network that receives,
processes, transports, stores and protects the information. It makes available
communications and sensor data to the war fighter and is self-organizing,
self-monitoring and continuously available. Also provided are adaptive and
Network-Centric Warfare and Netted LPI Radar Systems 321
Figure 10.1: Network-centric architecture employing an information grid,

shooter grid, and a sensor grid. The sensor grid is composed of a netted
LPI radar system of systems.
automated decision aids. The number of nodes directly reflects a volume of

force on the network and the information processing capability required. The
network topology represents the configuration of the links to integrate the
sensor and weapons nodes.
The sensor grid is composed of air-, sea-, ground-, space-, and cyberspace-
based sensor nodes. Netted LPI radar systems play an important part to
provide at all levels a covert surveillance capability and high degree of situ-
ational awareness. In this type of netted radar systems, a number of transmit-
ters and receivers are spatially distributed with each receiver being capable of
processing signals scattered by the target from every transmitter. In addition,
weapons can also be targeted using the netted LPI emitters. An increase in a
radars target processing capability (local node processing) benefits the speed
of relaying the targeting information. The sensor grid is a transient grid and
exists only for the task at hand. It is reformed for every mission with the
collaborative C2 performing dynamic sensor tasking and data fusion. With
the benefits of new network assurance technology, topologies are now capable
of providing a more robust network for information fusion. Sensor network
protocols are described in [2] and secure routing techniques in the presence
of electronic attack are described in [3].
The shooter grid consists of both weapons and jammers. It enables the
joint war fighter to plan and execute operations in a manner that achieves
power projection at a precise time and place. By exploiting the battlespace
awareness, new operational capabilities are realized including the execution
of time-critical missions, and the rapid acquisition and execution of targets
in a timely manner. The shooter grid is also a transient grid where the piece
parts are physical. The grid exists for the task only and is reformed for every
mission. In Figure 10.1, the NCW architecture is used to track a low RCS
target using a netted LPI sensor grid (advantages of netting the distributed
radar systems together are discussed in Section 10.5). The target is disabled
with a missile from the weapons grid.
10.1.1 NCW Requirements

There are several requirements for NCW operations. To achieve a force that
is network-centric, a wideband RF transmit/receive capability is required to
compress and transport large amounts of data. Also required is a wideband lo-
cal area network which can process and transmit information locally between
the sensors and/or weapons. Eective information management or the abil-
ity to eciently use, process and apply information is also required. Finally,
a critical mass of platforms, sensors, and weapons that have the informa-
tion processing capability is required. When forming the NCW architecture,
questions to be answered include: How do dierent degrees of networking im-
pact the strategic, operational and tactical outcomes? What is the optimal
network topology (physical, virtual, arrangement of nodes)? How will the
network impact the C2? What is the correct balance of sensors, shooters and
network technology? Can we quantify how the network processing sustains
degradation from events such as an electronic attack?
The answer to these questions is dicult since there are complex rela-
tionships between the network space and the battlespace. For example, from
the information standpoint, the overall information processing capability is
mainly determined by the number of nodes, the individual node capability,
and the topology of the network as shown in Figure 10.2. Note that this
figure does not show information flow but shows an overall relationship de-
pendence. For example, the number and distribution of LPI emitters on a
network must be sucient to build a precise and timely picture of the bat-
tlespace taking into account the limited detection ranges available. The data
distribution and data association between nodes must correlate the data accu-
rately avoiding any misidentification. The increase in information processing
capability sequentially results in an enhancement in the situational aware-
ness and operational tempo that aect the maneuverability, decision speed,
lethality, and agility on the battlefield.
Figure 10.2: Relationship between network space and battlespace
10.1.2 Situational Awareness

Situational awareness is defined by the U.S. Armys Training and Doctrine
Command as the ability to have accurate real-time information of friendly,
enemy, neutral, and noncombatant locations; a common, relevant picture of
the battlefield scaled to specific levels of interest and special needs [4]. The
LPI radar sensor can provide a critical role in maintaining the required situa-
tional awareness since it is able to gather the information without injecting an
influence. That is, the radar sensor used to gather the information should be
required to have a low probability of intercept. If detection by enemy nonco-
operative intercept receivers occurs, an enemy response will ensue increasing
the characteristic tempo T uncontrollably. The characteristic tempo is de-
fined as the speed in which the situational awareness is processed in order to
orient (or adjust) the force to the current situation. In practice, situational
awareness is built by continuous snapshots that are gathered from the bat-
tlefield and transferred to the commander. A larger information gathering
ability (e.g., more nodes) results in a larger information volume enabling for
example, beyond line-of-sight targeting. Better information exchange ability
results in a quick refreshing of the snapshot. Consequently, the situational
awareness is mainly determined by the information processing capability [5].
10.1.3 Maneuverability
A far-reaching netted radar system of systems can also improve force maneu-
verability, which is the capability to perform a strategic or tactical movement.
To evaluate the maneuverability performance, we consider three of its prop-
Figure 10.3: Improvement in maneuverability.
erties: speed, safety, and cost. Maneuverability can be promoted through

the support of situational awareness. Figure 10.3 shows the improvement in
maneuverability when a network-enabled situational awareness is eective [6].
For example, better terrain awareness results in optimal route design. The
route design not only increases the speed of the maneuver, but it can also
reduce the probability of risk and possibly result in a lower cost. Another
example is better threat awareness which helps the preparation of a proper
oense and contributes to improvements in maneuverability. Furthermore,
better integration of coalition war fighters into battlespace actions can also
increase force maneuverability.
10.1.4 Decision Speed and Operational Tempo

The observation-orientation-decision-action (OODA) loop is important for
operations and has become a critical concept in military strategy. John Boyd
originally developed the concept to explain how to direct ones energies to
defeat an enemy and survive [7]. The OODA loop concept is shown in Fig-
ure 10.4. A war-fighting enterprise that can process the entire OODA cycle
quickly, observing and reacting to unfolding events more rapidly than an op-
ponent, can get inside the opponents decision cycle and gain a military
advantage. The LPI radar nodes and the characteristic tempo T play a sig-
nificant part in the observation-to-orientation phase of the OODA loop. The
decision tempo C2 is defined as the speed to make a decision to act. After
the decision to act is made, the speed at which action is taken is the sum of
the characteristic tempo and a deployment tempo d . After deployment, the
speed at which the situational response (or fighting) is made is the sum of
the characteristic tempo and a fighting tempo f . These individual tempos
can be used to quantify the maximum operational tempo of the network-an
Figure 10.4: Observation-orientation-decision-action (OODA) loop.
Figure 10.5: Operational tempo determined by the agility of a force.
important attribute when considering the fusion of netted radar data. The
maximum operational tempo OODA is the inverse of the maximum frequency
to complete the OODA cycle.
In the experiments and exercises of the Army Battlefield Command Sys-
tem, it has been verified that due to the promotion of information process-
ing capability, operational planning could be improved as the speed of order
preparation and the operational tempo is increased. The commanders intent
is then clarified more quickly [1, 8]. Note also that the OODA loop can be
scaled to dierent levels of an operation. For example, it could be used to rep-
resent the operation of targeting a missile or the operation of force movement
on a battlefield. We will come back to the OODA concept in our discussions
later.
10.1.5 Agility
Agility is defined as the ability of an organization to sense and respond to
advancement opportunities in order to stay ahead and competitive on a tur-
bulent battlefield quickly. The operational tempo is highly dependent on the
agility. Figure 10.5 shows the comparison of fast and slow operational tempo.
In a given time period, the upper force with low operational tempo (less agile)
can only respond to environment events a maximum of three times. The fast
operational tempo can react five times and represents better agility.
10.1.6 Lethality
Lethality is the ability to damage an enemy. Only with the sucient situa-
tional awareness and ecient operational tempo can the forces perform with
the best lethality. The radar sensor network plays a key role in the measure of
lethality. For example, the artillery can perform with high lethality with ac-
curate targeting information and timely approval of attack. Infantry attacks
also do well with enough intelligence and under quick and timely command.
It is important to note that if the enemy is aware of the targeting informa-
tion being gathered, the characteristic tempo is slowed due to the further
reconnaissance that is necessary. The lethality of the action is also severely
degraded emphasizing the need for the battlefield sensors to be LPI.
10.2 Metrics for Information Grid Analysis

A military sensor and weapons grid is sometimes assumed to have an infinite
number of nodes each with a similar capability. In the analysis, this often
leads to misleading results, especially when a small number of dissimilar nodes
are used [8]. This section examines the network theory and metrics that are
designed to quantify the general value inherent in the information network
topology. These include the connectivity measure, the network reach, and
the network richness. Combining these metrics with the operational tempos
previously discussed, we can quantify the maximum operational tempo of the
network. The presence of an electronic attack is also addressed.
10.2.1 Generalized Connectivity Measure

A time-dependent, generalized connectivity measure (CM ) of a network of
sensors and weapons is defined as the sum of the value of all the nodes and
their connections scaled by the lengths of the routes and their directionality.
The connectivity measure can be expressed as [8]
NT N N,
CM (t) = K (t) L, (d, t) (10.1)
=1 =1 =1
where NT is the number of nodes in the network, N is the total number of

nodes connected to the node , N, is the total number of possible routes
connecting the pair of nodes and , K (t) is the capability value of node
, and relates to the ability of the node to process and transfer information
quickly. L, is the information flow parameter of the route connecting
nodes and and depends on the length of the route d and is also a function
of time t. The capability and information flow parameters have the range
0 K (t), L, 1. For the examples below, we use a normalized value for
the route length d. That is, d = 1 from one node to another.
Figure 10.6: Three node example to demonstrate generalized connectivity

link calculations.
Table 10.1: List of All Available Links in Figure 10.6
Links
13
21
23
31
32
The term route is the possible connection from one node to another
node. The term link represents the direct connection between any two
nodes. One route contains at least one or more links. Figure 10.6 shows
three information nodes deployed with dierent capability values K . The
link from node = 1 to node = 2 is not available. A list of all available
links and routes are shown in Tables 10.1 and 10.2, respectively.
The functional dependence of L, on the length of the route d (number
of links) and time t can be simplified by separating it into a time-independent
component L, and a time dependent flow coecient F, (t), which is scaled
by the route length d raised to the power . The expression for CM (t) then
becomes [8]
NT N N,
,
F, (t)
CM (t) = K (t) L (10.2)
=1 =1 =1
(d )
The value of F, (t) is a minimum of zero and reaches a maximum of one

when the route is capable of supporting all information exchanges. Note the
Table 10.2: List of All Possible Routes

Start Node End Node Routes
1 2 132
1 3 13
2 1 21
231
2 3 23
213
3 1 31
321
3 2 32
Figure 10.7: Two nodes with a unidirectional link.
order of the node superscripts matters. For example, consider the two nodes
shown in Figure 10.7. The flow coecient 0 F, 1 however, F, = 0.
To illustrate these ideas, assume K (t) is time independent and that any
two nodes are either connected or not (F, (t) = 0 or 1). The directionality
of the information is also included. Also assume that the scaling exponent
= 1, and the time independent information flow parameter L, = 1 for
every route are identical. As a result, (10.2) can be simplified to
NT N N,
F, (t)
CM (t) = K (10.3)
=1 =1 =1
d
In summary, the following assumptions are held. First, the connectivity is

time-independent. That is, K (t) = K and F, (t) = F, . Also, = 1 and
any two nodes are either connected (or not). Table 10.3 demonstrates the
generalized connectivity CM calculation for Figure 10.6.
10.2.2 Reference Connectivity Measure

R
The reference connectivity measure (CM ) is defined to represent a fully con-
nected network configuration. The reference network has all nodes fully con-
nected with bidirectional links [8]. In addition, each node has a capability
value of K = 1. For example, Figure 10.8 shows a realistic four-node infor-
mation transfer network deployed with dierent capability values K . There
Table 10.3: Calculation of Connectivity Measure
CM
Route K d Contribution
132 1 2 0.500
13 1 1 1.000
21 0.75 1 0.750
231 0.75 2 0.375
23 0.75 1 0.750
213 0.75 2 0.375
31 0.25 1 0.250
321 0.25 2 0.125
32 0.25 1 0.250
CM = 4.375
is a unidirectional link from node 1 to node 2 and from node 3 to node 2.

There are also two bidirectional links from node 3 to node 1 and from node 3
to node 4. Figure 10.9 shows the corresponding reference network for Figure
10.8. Note that all the nodes are homogeneous and connected to one another
(fully connected) and that all capability values are K = 1 and F, = 1 for
all , and . The reference connectivity measure only depends on the total
number of nodes NT and is calculated as
R NT 2 (NT 2)(NT 3) 2 1
CM = NT (NT 1) 1 + + + (10.4)
2 NT 1
The term outside the square brackets in (10.4), NT (NT 1), represents the
number of possible connections in a given network with NT nodes. The
numerator in each term inside the square brackets is the number of possible
routes of the length given in the denominator. The reference network has
the highest connectivity measure of any network with same number of nodes.
R
Table 10.4 shows the value of CM for 3 NT 8 and shows the exponential
R
increase in CM with a linear increase in the number of nodes.
10.2.3 Network Reach

R
The reference connectivity measure CM provides a means to normalize the
connectivity measure (10.3) resulting in the network reach IR as [8]
CM
IR = R
(10.5)
CM
which is a dimensionless quantity. Normalization by the reference network
allows us to investigate varying degrees of network connection, nonidentical
Figure 10.8: Realistic four-node sensor network configuration.
Figure 10.9: Reference network for the radar information network shown in
Figure 10.8.
Table 10.4: List of Reference Connectivity Measures

R
Node Number CM
3 9
4 32
5 120
6 534
7 2,905
8 18,976
nodes/links and the eect of broken symmetries due for example, to electronic
attack of the network.
10.2.4 Suppression Example

A network-enabled SEAD example with NT = 5 is considered in Figure 10.10.
In this scenario, we are concerned with the capability of the network to trans-
fer sensor data eciently. To suppress the Eagle (Ka-band battlefield sur-
veillance and missile control radar), an EA-6B is networked with an RQ-1
Predator, an AC-130 Gunship and an EA-18G. The EA-6B, the gunship and
the EA-18G are all data linked to the special operations forces on the ground
(note the unidirectional flow of information).
The capability value K is assigned as shown for each asset. The capability
value depends on the platforms level of networking and ability to participate
in the data and information exchange needed. For the example, both the
EA-6B and the EA-18G have a value K = 1.0 since they are in control of
the suppression mission and are fully network capable. The RQ-1 Predator
is assigned a value of K = 0.5 since it cannot relay directly to the special
operation forces, any of its images and data that are gathered. The AC-
130 gunship is assigned K = 0.85 since it is somewhat less capable than the
EA-18G and EA-6B due to its multimission characteristics. Note that the
special operation forces are assigned K = 0.3 and is a low value since they
only receive images and data and do not transmit (otherwise they would give
their position away).
The reference connectivity measure is calculated first. For NT = 5,
R 3 32 321
CM = 5(4) 1 + + + (10.6)
2 3 4
R
or CM = 120. Note that this value only depends on the number of nodes
participating in the network. Using this value, the expression for the network
reach is
NT =5 N N, ,
1 F
IR = K L, (10.7)
120 =1 =1 =1
d
The values for N = {N1 = 4, N2 = 3, N3 = 4, N4 = 4, N5 = 3}. Also,

N11 = 0. The value of N12 = 5 can be verified from Figure 10.10. That is,
there are five routes from node 1 to node two. The five routes are shown in
Table 10.5. With L12 = 1, the summation of the flow coecients scaled by
the route lengths is
F 1,2 1 1 1 1 1
= + + + + = 2.67 (10.8)
d 1 2 2 3 3
Continuing on, the other node to node flow coecients scaled by the route
Figure 10.10: Eagle radar suppression using NT = 5 nodes.
lengths can be calculated. The network reach is then

1
IR = 120 (1.0 [3(2.67) + 4.33] + 0.5 [3(2.67) + 5] +
(10.9)
0.85 [3(2.67) + 4.33] + 1.0 [3(2.67) + 4.33] + 0) = 0.3473
The first four terms in brackets are the EA-6B, RQ-1A, AC-130 and the
EA-18G respectively. The zero term (last term) is due to the fact that this
node contributes nothing to the overall information transfer capability of the
network. The low value of network reach is due to the reduction in node
capability values and the loss of sensor information rerouting options. Note
that if the capability value K = 1.0 for all nodes in the above example, then
Table 10.5: Five Routes Identified from Node 1 to Node 2

1 2 3 4 5
1 1 1 1 1
2 3 4 4 3
2 2 3 4
2 2
IR = 0.4167.
10.2.5 Extended Generalized Connectivity Measure

We can generalize (10.3) by considering the case where 0 < F, < 1 exists
(partial eciency of route). For instance, if a traversed node on one route
has a low capability (K 1), this route will not be able to maintain full
capability in the flow of sensor information [6]. Consider for example the
network shown in Figure 10.6. The route node 1 node 3 node 2 is
evaluated as K1 /d = 1/2 = 0.5. However, the traversed node K3 = 0.25
gives indication that the sensor information flow from node 1 to node 3 cannot
be eciently exchanged via node 3. Taking the limitation of the traversed
intermediate nodes into account, from (10.3) we get an extended definition of
connectivity measure as
NT N N,
K F,
CMe (t) = (10.10)
=1 =1 =1
d
where K represents the K with the lowest capability value (bottleneck) in

route . Note the fact that K in the route only considers the starting node
and exchangers; the receiving node is not included. This consideration is due
to the fact that many nodes in military networks only accept the information
without an equivalent information processing capability in transmitting. For
instance, in route node 1 node 3 node 2, only the transmitter (node
1) and exchanger (node 3) are available for assignment to K to reflect the
bottleneck of the information flow. For the same route, shown in Figure
10.6, the extended CM is recalculated as shown in Table 10.6. Comparing to
Table 10.3, notice the value of CM decreases from 4.375 to 3.75 due to the
consideration of the bottlenecks in route 1 3 2 and 2 3 1 reflecting
a more realistic capability.
In summary, the robustness of the network can be quantified by comparing
R
CM to CM , and by disabling nodes in the reference and the real network
R
and recalculating CM and CM . By comparing these values, a representation
appears of the real network behavior while under attack. In fact, the rate of
change (degradation) in the value of the connectivity measure as a function
of the number of links severed and nodes being attacked can provide good
insight into the robustness of the sensor network.
10.2.6 Entropy and Network Richness

At each node or source, the rate at which information is sent has a direct
impact on the operational tempo of the grid. Consider the set of J possible
sample values (or source symbols) by S = {x1 , . . . , xJ }. We assume the
Table 10.6: Extended Connectivity Measure Calculation
Bottleneck CM
Route Node K d Contribution
132 3 0.25 2 0.125
13 1 1 1 1.000
21 2 0.75 1 0.750
231 3 0.25 2 0.125
23 2 0.75 1 0.750
213 2 0.75 2 0.375
31 3 0.25 1 0.250
321 3 0.25 2 0.125
32 3 0.25 1 0.250
CM = 3.750
probability of the source output xj is known as pj . The amount of information

sent from a digital source when the jth message is transmitted is
I(j) = log2 (pj ) (10.11)
where pj is the probability of transmitting the jth message. Shannon defined
the primary information-related measure of each message H as a function of
the probability of transmission of each message [9]. This entropy (or uncer-
tainty) of the source is
J
H(S) = E{I(j)} = pj log2 (pj ) (10.12)
j=1
and is measured in information bits per source symbol.1 The information

rate of the source is then
H
= bits/s (10.13)
T
where T is the time required to send the message. A related measure is the
channel capacity or
C = B log2 (1 + SNR) bits/s (10.14)
where B is the channel bandwidth (in Hertz) and SNR is the signal-to-noise
power ratio (not in decibels) at the receiver input [9].
The channel capacity can be used as a unifying principle for EA and EP
actions in EW. Every EA measure (except exploitation) is an attempt to re-
duce the bandwidth of an adversary signal and/or to reduce the SNR. Every
1 Since the entropy is measured in bits per sample, the binary logarithm must be used.
EP action (except protection against exploitation) is an attempt to increase

bandwidth and/or increase SNR.
Example 1: The LPI emitter uses frequency hopping as an EP measure and

uses a large total bandwidth to protect against jamming, but a small instanta-
neous bandwidth to protect against interception and exploitation. The large
total bandwidth in this case makes it dicult for the jammer to set on the
transmission frequency, thus limiting the reduction in SNR to that provided
by barrage jamming.
Example 2: Repeater or gate stealing EA techniques must achieve a certain

reduction of SNR within the bandwidth of the victims receiver to be eec-
tive. The corresponding EP technique might utilize a combination of guards
and filters to recognize and eliminate the unwanted jamming signal, thereby
protecting the SNR.
Example 3: Against exploitation, a LPI emitter uses a very large bandwidth

with low average power density. The low average power reduces the probabil-
ity of intercept, but the energy over the bandwidth can be summed to extract
the information from the signal. Therefore, the transmitter compensates for
the low SNR with increased bandwidth to transmit the information at a fast
enough rate. The jammer can only achieve high SNRs over small portions of
the bandwidth.
Each node within the sensor network is able to process the information at
a certain rate. The information processing rates of each node can be combined
to quantify the networks richness. The information rate, , of a node , is
the rate at which the network information is processed by the node (in Hz).
The minimum information rate, min , of the node is the minimum rate that
information must be processed for generating decision-level knowledge from
the sensor network data. From Shannons information entropy theory, the
knowledge function is defined as [8]

0, if < min
min
Q ( ) = ln min
, if < < e min
(10.15)

emin
ln
= 1, if e min
min
Using the knowledge function, the network richness RQ is defined to represent

the average rate at which information entropy (or knowledge) is generated
from the sensor data shared through the network or [8]
NT
=1 Q( )
RQ = s1 (10.16)
NT
Figure 10.11: Time spent in each phase in OODA cycle. (After [8].)
From this equation, if a node cannot provide the knowledge at a rate above its
minimum value, the nodes contribution Q( ) degrades the overall value
RQ . In addition, there is little advantage to generating sensor data faster
than knowledge can be generated and absorbed.
10.2.7 Maximum Operation Tempo

A network has a maximum information exchange rate that is determined by
the number of nodes, the communication and sensor technologies employed,
the information data transfer rates, and the network topology. To quantify
this rate within an OODA, a characteristic tempo (T ) is defined and relates
the network topology and its ability to gather the situational awareness. The
characteristic tempo for the network is the product of the network reach IR
and the network richness RQ
T = IR RQ Hz (10.17)
and relates the information exchange capability of the sensor network. In

addition, for every command and control structure (and associated doctrine),
there is a characteristic decision-making rate (C2 ) or speed at which decisions
are made using the transferred sensor data being processed [10].
Figure 10.11 shows the tempo parameters of the sensor network OODA
loop. The variable t1 represents the time from observation to orientation
and is limited by the information exchange time, t2 is the time from ori-
entation to decision and is dominated by the decision speed, t3 is the time
from decision to action and must be greater than the information exchange
time (command time) and deployment time, and t4 is the time from action
to observation and is always greater than the sum of information exchange
time and fighting time.
Using the OODA tempo parameters, the maximum operation tempo of the
network (OODA ) is of interest and represents the maximum tempo of the
network to perform an entire OODA including responding to events as
1
1 1 1 1 1 1
OODA + + + + + (10.18)
T C2 T d T f
or after some algebra

C2
OODA (10.19)
1 1 3C2
1+ d + f C2 + T
Note that C2 in the numerator emphasizes that the fact that while technol-
ogy can help increase the network and action tempos, the C2 tempo plays a
limiting role not helped by technology alone [8]. Also note that in practice the
operational tempo is not a fixed value. The operational tempo calculated here
represents the maximum value due to the limitation of the network topology
and nodes capabilities. It provides a direct link between the internal metrics
of the network and the operational outcome of a sensor and weapons network
through a single equation. It is also significant in that it enables direct evalua-
tion of the networks capability to collect, process and disseminate information
(information superiority) to the combat outcome (battlespace superiority).
10.3 Electronic Attack

Jamming of the information grid is a form of electronic attack and can take
on many forms such as partial band jamming, and tone jamming [11]. The
eectiveness of the jamming waveform depends on the signaling format used
to transfer the data and the type of jamming used. When a jammer is taken
into consideration, the jam-to-signal ratio (JSR) at the victim nodes receiver
causes a link failure to occur if the jam-to-signal ratio is greater than a par-
ticular threshold causing the bit error rate to be unacceptable. Figure 10.12
shows a jammer (node 4) added into the previous example shown in Figure
10.6.
The JSR is determined by many factors including jamming and signal
power, jammer range, jamming strategy, RF waveform bandwidth, and prop-
erties of the receiver. To simplify the calculation, considering only power and
range, the JSR in a single information link can be written as
2
ERPJ /4(RJ )2 ERPJ RC
JSR = = (10.20)
ERPC /4(RC )2 ERPC RJ
where ERPJ is the eective radiated power of the jammer (node 4), ERPC is
the eective radiated power of the data/communication signal emitted from
Figure 10.12: Addition of a jammer into the node configuration shown in

Figure 10.6.
node 3 to node 1, RC is the range from node 3 to node 1, and RJ is the range
from the jammer at node 4 to the receiver at node 1.
The capability values can be assigned and used within the JSR calculation
to quantify the jammer eectiveness. The capability values are defined as
follows. The value of K is defined as the information-processing capability
of the receiving node and its importance to the network. Assuming the
importance of the information transferred through each node is not dierent,
we see that the K is related to the information exchange capability. Also,
KJ is the jamming capability of the hostile jammer and is defined as the
information link jamming capability at node . Similar to K , 1 KJ 0
and is determined by factors such as its eective radiated power, jammer
waveform type, and jamming strategy. Without a loss in generality, the
ratio of KJ to K is set equal to the ratio of the eective radiated powers.
Therefore, (10.20) can be written as
2 2
ERPJ RC KJ RC
JSR = = (10.21)
ERPC RJ K RJ
The JSR is used to represent the eect of the jamming on an existing infor-
mation exchange link. When the JSR is higher than a given threshold, the
information link is regarded as unavailable.
10.4 Information Network Analysis Using

LPIsimNet
The MATLAB folder LPIsimNet (see Appendix D) provides the tools to cal-
culate the metrics discussed above and generates a visual summary of the
simulation results for any user-defined global information grid configuration.
In this section, several simulations are presented to illustrate the metrics dis-
cussed above including the eects of an electronic attack. The label notation
used is in the form of (i, j) XYZ. The i represents the node type and can be
any one of the following:
N: Friendly information/data transfer communication node;
R: Friendly LPI radar node (discussed in the next section);
NR: Friendly node with both information/data transfer capability and
radar capability;
JN: Hostile communication jammer node;
JR: Hostile radar jammer node;
JNR: Hostile communication and radar jammer node.
The j indicates the index of the node and ranges from 1 to the number of
nodes utilized NT . The XYZ represents the name of the node (e.g., EA-6B,
E2C).
The first simulation considers a sensor network with three nodes as shown
in Figure 10.13. In this simulation, the communication between an E-2C, an
F-16 and an AC-130 are being studied. Figure 10.14 shows the scenario setup
used to generate the simulation. Note that the user can control the number
and characteristics of each node within the scenario (including the placement
and movement). Top-level properties are in rows 2 through 5. Rows 6 through
10 show the characteristics of the individual nodes. The last section shows
the node connectivity. For the simulation shown in Figure 10.13, there are
two bidirectional links and one unidirectional link indicated by the direction
arrows. The simulation is run and the results are summarized in Table 10.7.
The sensor network simulation results can be generated for any number of
nodes and connectivity but can take a significantly longer period of time for
simulations with a large number of nodes. The details of the connectivity
measure CM and network richness RQ are shown in Tables 10.8 and 10.9,
respectively.
To quantify the eect of an electronic attack, a jammer onboard a Russian
Su-34 is added to the sensor network. The sensor network under attack is
shown in Figure 10.15. The Russian Su-34 is located at the bottom right
corner and is represented by a hollow circle. The jamming connection is shown
by a dashed line to E-2C. The initial scenario configuration is shown in Table
10.16. The total time index row represents the number of time indexes that
are calculated in the simulation. This oers the ability to include movement
of all assets. For the setup shown in Figure 10.15, total time indexes is set
to 3. When the simulation is run, the jammer moves closer to the E-2C at
each time index and all metrics are recalculated (total of 3 times). Position
Figure 10.13: Three communication nodes.
Table 10.7: Simulation Results of Scenario Shown in Figure 10.13
Results Values
Reference connectivity measure 9
Connectivity measure 3.75
Network reach 0.42
Network richness 271.60
Characteristic tempo 113.16
Operational tempo 26.78
Table 10.8: Analysis Detail of Connectivity Measure for Simulation Shown in

Figure 10.13
Bottleneck CM
Route Node Contribution
132 3 0.125
13 1 1.000
21 2 0.750
231 3 0.125
23 2 0.750
213 2 0.375
31 3 0.250
321 3 0.125
32 3 0.250
CM = 3.750
Table 10.9: Analysis Detail of Network Richness for Simulation Shown in

Figure 10.9
Node Q (/m ) Q (/m )

1 200 0.69315 138.630
2 200 0.69315 138.630
3 300 1.7918 537.540
814.800
RQ = 814.8/3 = 271.600
Figure 10.14: User setup of the scenario shown in Figure 10.13.
refers to the initial position of the node and velocity indicates the movement
of each node per time index (km/time index).
A summary of the simulation results for the three time instances is shown
in Figure 10.15 (time index 1), Figure 10.17 (time index 2), and Figure 10.18
(time index 3).
Note that for each time instant, the jammer is moving closer to the E-2C.
Notice on the second time index, the jammer is close enough to disable one of
the links. The node 3 to 1 link is disabled due to the JSR > 1. On the third
index, the link from node 2 to 1 is also suppressed. Consequently, several
trends in the network metrics can be noted across the three time indexes. As
the jammer moves towards the network, the measure of connectivity decreases
as does the network reach. Also, a noticeable decrease in the characteristic
tempo and maximum operational tempo is shown.
10.5 Netted LPI Radar Systems

Despite recent advances in monostatic radar systems (colocated single trans-
mitter and receiver), two major disadvantages are inherent. They oer little
Figure 10.15: Simulation of network jammer at time index 1.
to counter stealth technology and they only oer a single perspective for each
radar [12]. The development of stealth technology has primarily been aimed
at defeating the monostatic radar by the use of absorbing materials and non-
reflective structural designs that minimize the scattered energy reflected into
the hemisphere from which the signal arrives. The limited energy that is
returned to the emitter from the stealth target, makes it very dicult to
detect the target. In addition, due to terrain obscuration, ground-based or
low-flying monostatic radar systems often do not have a line of sight to the
target and therefore cannot provide detections.
Due to this single perspective, the information contained in the multiple
perspectives is missed. Consequently, if a number of cooperative radar sys-
tems are distributed spatially and networked together, they can provide the
opportunity to view the target from a number of dierent aspect angles. In
multifrequency radar networks each radar performs a significant amount of
local preprocessing. Outcomes of the local preprocessing can then be deliv-
ered to a central processor through a communication link. The preprocessing
limits the amount of information that needs to be passed on to make a fi-
nal detection decision. These systems use dierent frequencies to cope with
interference rejection but each receiver is unable to process the information
from all transmitters.
Figure 10.16: User setup of the scenario shown in Figure 10.15.
Netted radar systems sometimes referred to as spatial multiple-input

multiple-output (MIMO) radar systems, consist of a number of distributed
radar systems (transmit and receive sensors) each having the ability to trans-
mit independent orthogonal waveforms (to avoid interference) and the ability
to receive and process synchronously all waveforms that are transmitted. Fig-
ure 10.19 shows an example of a netted radar system with three radar nodes
connected by a network. All three radars have already acquired and are track-
ing the target with their antenna beams. The radar systems R1, R2, and R3
each transmit a dierent waveform but receive and process all three waveforms
that are collected from the target. The use of the network allows each system
to share its target information noncoherently (using orthogonal waveforms)
or coherently where each radar has a common precise knowledge of space and
time. The implementation of networked radar systems has become feasible
due to recent advances in large bandwidth wireless networks, high-capacity
transmission lines, multichannel electronically scanned antennas, high-speed
low-cost digital processors and precise synchronization systems [13].

Figure 10.19: Example of a sensor network connecting three LPI emitters

(R1, R2, R3).
10.5.1 Advantages of the Netted LPI Radar Systems

There are two important characteristics of a netted radar system; the spa-
tial dispersion of the nodes (i.e., transmitter and receiver locations) and the
data fusion processing (i.e., processing performed in the receiver at a node
to combine multiple receiver outputs). These characteristics lead to several
advantages of a netted LPI radar system.
The spatial distribution (or geometry) of the nodes enable the surveil-
lance area to be tailored according to the specific mission objective [14]. The
multisite emitters can be used to form a specially designed surveillance area
to more eciently detect targets based on known patterns of military be-
havior. The network also allows a multiperspective SAR or ISAR image to
be generated. By using a number of distributed transmitters and receivers
to collect the echos from the target at dierent aspects or directions, the
independent angular samples provide spatial diversity of the targets RCS.
With widely separated antennas, netted radar systems also have the ability
to handle slow moving targets by exploiting Doppler estimates from multiple
directions. If coherent processing is used, high-resolution target localization
can be achieved with a resolution that far exceeds that supported by the
radars waveform [15]. This however, comes with a price. The receiving
and processing requirement for such a coherent summation is highly demand-
ing. For example, if each waveform produces 1,000 resolvable range cells and
10 Doppler cells, integration would be required simultaneously in 10, 000N 2
cells (possibly reduced by excluding regions not mutually covered). Given the
high-resolution in both range and Doppler, the numbers used here may in-
crease by orders of magnitude, straining the capability of most modern signal

processors.
One of the inherent properties of the netted radar system is the increase
in the number of degrees of freedom due to the spatial diversity [14, 16].
The scattered reflections can be captured to take advantage of the target
scintillation providing a gain in detection performance. The spatial diversity
can be used to separate scattering centers from one another that otherwise
cause a glint signature. In addition, targets can be more easily separated from
the clutter in clutter-limited detection scenarios [17]. The targets RCS spatial
variations can also be exploited to obtain a diversity gain for estimation of
various parameters (e.g., angle of arrival, Doppler). Target classification and
recognition can be improved as a result of the extra information retrieved
from the dierent perspectives [18]. With more degrees of freedom, flexible
time-energy management modes can be utilized to minimize the amount of
energy that is radiated. Since more of the scattered energy from the target is
collected (from dierent directions), the sum of the ERP from all the radar
systems can be made approximately equivalent to that of a single monostatic
radar [19]. As a result, the detection performance of the system is superior
while also utilizing a minimum ERP. A better system sensitivity results due
to the additional transmitters and receivers which augment the total received
signal power leading to an increase in overall SNR. Further, every node is less
vulnerable to physical and electronic attackincreased survivability. That is,
the destruction of the sensor network by an ARM is less likely. The probability
of being coherently jammed is also less likely since the probability of intercept
is lowered even further.
The likelihood of obtaining a line-of-sight to the target is also greatly
improved due to the spatial dispersion of the radar nodes [20].2 Having several
radar systems will add confusion to the noncooperative intercept receiver
that has to cope with the increased number of signals. Another significant
advantage is the increase in the reliability of the netted radar system. The
loss of one or even several nodes may not destroy the surveillance capability
but more of a graceful degradation will take place as there are still other
nodes available [14].
There are also technical challenges to be addressed. The most important
is the time and frequency synchronization for coherent operation. By using
GPS as a reference timing signal, the network can be made coherent. Another
important challenge is the data fusion and registration of the various data
streams, which requires reliable and high-capacity communication links in
the network [14].
2 One of the disadvantages of bistatic and multistatic radar systems is that more than
one line-of-sight path is required. For low-altitude targets, the network of monostatic radars
has a much higher probability of having at least one unusable path. The more nodes that
require simultaneous line-of-sight paths, the lower the probability of success.
10.5.2 Netted LPI Radar Sensitivity

The spatially distributed and networked LPI radar system of systems can
be broken down into a set of M N transmitter-receiver pairs each with a
bistatic component contributing to the entirety of the netted radar sensitivity
[13]. Figure 10.19 can be considered as a connected series of bistatic radar
systems where the transmitter and receiver are separated. With this, it is
necessary to calculate the target-to-transmitter range and target-to-receiver
range separately. These range values then replace the single range term in
the monostatic radar equation. In addition, a separate bistatic RCS value for
each bistatic radar pair must be computed. Thermal noise at each receiver
can be assumed to be statistically independent. The overall netted radar SNR
can be calculated by summing up the partial SNR of each transmitter-receiver
pair as [12, 19]
M N
PCWi Gti Grj T ij 2i
SNRnet = 2 R2 F L (10.22)
i=1 j=1
(4)3 kT0ij BRi Rti rj Rj ij
where PCWi is the ith average CW transmitter power, Gti is the ith transmit
antenna gain, Grj is the jth receive antenna gain, T ij is the RCS of the target
for the ith transmitter and jth receiver, i is the ith transmitted wavelength,
BRi is the bandwidth of the matched filter for the ith transmitted waveform,
k is Boltzmanns constant, T0ij is the receiving system noise temperature at a
particular receiver, FRj is the noise factor for each receiver, Lij is the system
loss for the ith transmitter, jth receiver (Lij > 1), Rti is the distance from
the ith transmitter to the target and Rrj is the distance from the target to
the jth receiver. Note that this assumes that all signals can be separately
distinguished at each receiver and that all antenna beams are pointed at the
target. Also note that with i = j = 1, (10.22) reverts to the monostatic case.
An important characteristic of netted radar systems can be identified when
we consider each radar to be identical with every transmitter-receiver com-
bination the same [12]. In this case the netted radar SNR equation can be
written as
M N
PCW Gt Gr T 2 1
SNRnet = 3 2 2 (10.23)
(4 )kT0 BFR L i=1 j=1 Rti Rrj
Insight is gained if we group all of the range independent parameters together

into a constant K, then the netted radar SNR can be expressed as
M N
K
SNRnet = 2 R2 (10.24)
i=1 j=1
Rti rj
Normalizing (10.24) by the SNR for a single monostatic radar (10.37) we have
M N
SNRnet 1
= R4 2 R2 (10.25)
SNR1 i=1 j=1
Rti rj
which shows that the SNR of the system is related to the power received
from the contributing transmit and receive paths. Further, if Rti = Rrj and
M = N it follows that [19]
SNRnet
N2 (10.26)
SNR1
That is, the system SNR is a function of the square of the number of nodes for
coherent operation. This represents an N -fold improvement over the nonco-
herent case ( = 0.5). For noncoherent netted radar processing, the number
of transmit antennas serves as a factor in the number of diversity paths. For
coherent processing the number of transmit antennas contributes to reduc-
ing the spurious peaks. In either mode, the processing at the receiver scans
through all the possible target locations.
It must be pointed out that for each of the N radars to receive and process
the N dierent waveforms transmitted by those radars all with antenna gains
Gt and Gr , achieving SNRnet applies to the sum of N 2 coherently combined
signals. Since antenna gain G 4/b where b is the solid angle within the
half-power beam contour, there are in the hemisphere visible to each radar
2/ = G/2 beam positions. Unless the target has been acquired and placed
in track by a single radar, using the single-radar (monostatic) SNR available
to that radar, and used to point the other radars, the probability that all
radars illuminate the target simultaneously is extremely small for Gt Gr 1.
This implies that near omni-directional antennas must be used to achieve
initial detection based on SNRnet . If designation from a monostatic radar
is used, then the other radars must each place a transmitting and receiving
beam on each target for which SNRnet is to be obtained, implying either near
omnidirectional or multiple directional beams that require splitting transmit-
ter energy amongst multiple targets. In summary, for most cases SNRnet will
only be available for tracking or identifying a target that is first detected by
a single monostatic radar in the network.
10.5.3 Signal Model

To develop the netted radar (spatial MIMO) signal model, a distributed tar-
get with Q independent isotropic scatterers is considered. Figure 10.20 shows
four such scatterers located in a 2-D plane along with the M LPI trans-
mitters Tk = (xtk , ytk ), k = 1, . . . , M that illuminate the target and the N
receivers Rl = (xrl , yrl ), l = 1, . . . , N that collect the scattered energy. We let
E/M sk (t) be the set of transmitted waveforms where E is the total trans-
mitted energy. Normalization by M makes the total energy independent of the
number of transmitters used to illuminate the target [18]. The target reflec-
tivity can be expressed in a diagonal QQ matrix with = diag(1 , . . . , Q ).
T
The target average RCS is E[tr( )] = 1 and is independent of the num-
ber of scatterers. For the netted radar case, spatial diversity is achieved with
the distributed antenna positions. The M waveform generators (W Gk ) trans-
mit orthogonal (noncoherent) waveforms in order that the energy from the
dierent transmitters may be easily separated at the receiver where each of
the receive antennas has M matched filters (M Fi1 . . . M FiM with one corre-
sponding to each orthogonal waveform).
Neglecting the path loss and summing over all of the scatterers, the total
signal received can be expressed as
Q
E
zlk (t) = q sk [t tk (Xq ) rl (Xq )]ej2fc [tk X(q)+rl (Xq )] (10.27)
M q=1
where tk (Xq ) = d(Tk , Xq )/c is the propagation time delay between the
kth transmitting sensor and the scatterer at Xq . The distance d(Tk , Xq ) =
(xtk xq )2 + (ytk yq )2 . The propagation time delay rl X(q) is defined
analogously. The two exponential terms in (10.27) reflect the phase shift due
to the propagation from transmitter k to scatterer q and the phase shift due
to the propagation from the scatterer q to the receiver l.
The channel components of (10.27) are often collected as [18]
(q)
hlk = q ej2fc [tk (Xq )+rl (Xq )] (10.28)
and can be interpreted as the equivalent channel between transmitter k,
scatter q and receiver l. The channel element (10.28) consists of ej2fc tk (Xq )
which is the phase shift due to the propagation from transmitter k to scatterer
q. Similarly, ej2fc rl (Xq ) is the phase shift due to the propagation from the
scatterer q to the receiver l. The reflectivity of the scatterer is q . With
(10.28), (10.27) can be expressed as
Q
E (q)
zlk (t) = hlk sk [t tk (Xq ) rl (Xq )] (10.29)
M q=1
If the target has an RCS center of gravity at X0 = (x0 , y0 ) and we assume

that sk (t tk (Xq ) rl (Xq )) sk (t tk (X0 ) rl (X0 )) for all q = 1, . . . , Q
then
E
zlk (t) = hlk sk [t tk (Xq ) rl (Xq )] (10.30)
M
where
Q
(q)
hlk = hlk (10.31)
q=1
Figure 10.20: Netted radar with M transmit antennas each with a separate
orthogonal waveform generator. Receive array consists of N antennas each
with a parallel set of M matched filters. Target is shown with distributed
scatterers located at Xq with reflectivity . Targets RCS center of gravity is
located at X0 .
Accounting for additive noise, the observed waveforms at the receive antenna
l can be expressed as
E
rl (t) = hlk sk [t tk (X0 ) rl (X0 )] + wl (t) (10.32)
M
where wl (t) is the additive circularly symmetric, zero mean, complex Gaussian
noise that is spatially and temporally white with autocorrelation function
2
w ( ).
Properties of the elements hlk of the channel matrix and the conditions
for spatial decorrelation E[hlk hli ] 0 are further addressed in [18, 21]. In
essence, the spatial decorrelation means that dierent receive antennas mea-
sure a dierent value of the RCS. Also discussed is the relationship of the
model to other types of emitters such as phased arrays, adaptive radar STAP
and multistatic radar. Properties of the MIMO radar ambiguity functions are
given in [2224]. As a final point, we point out that the maximum number
of targets Kmax , that can be uniquely identified simultaneously by a phased
array with N receive antenna elements is
2N
Kmax = (10.33)
3
while the maximum number of targets that can be uniquely identified simul-
taneously by a MIMO radar is [25]
2M N
Kmax = (10.34)
3
That is, the maximum number of targets that can simultaneously be uniquely
identified by a MIMO radar is up to M times its phased array counterpart.
10.5.4 Netted Radar Electronic Attack

The JSR as defined by the jamming power and signal (radar echo) power is
given by
jamming power
JSR = (10.35)
signal power
Unlike communication antennas that often use dipole antennas for omni-
directional communication, radar antennas frequently use highly directional
antennas that can identify the target angle in azimuth and elevation. The
shape of the radar antenna pattern (pencil beam) results in degradation of
the jamming signal when the jamming signal is not incident on the main lobe.
The jam-to-signal ratio is
2

ERPJ 4 (RJ ) 2
JSR = cos = ERPJ RT cos (10.36)
ERP 4 (R )2 ERPR RJ
R T
Figure 10.21: Example of jamming with incident angle.
where ERPJ = Pj Gj is the eective radiated power of the jammer, ERPR =

PCW Gt is the eective radiated power of the LPI radar, RJ is the range
from the jammer to the radar, RT is the range from the radar to the target,
and is the incident angle of jamming. Figure 10.21 provides an example of
the jamming signal incident with = 60 degrees that results in cos = 0.5
degradation in the jamming power.
10.6 Netted Radar Analysis Using LPIsimNet

The LPIsimNet MATLAB tools (see Appendix D) are used in this section
to demonstrate the SNR advantages of a netted-radar configuration. Any
user-defined netted radar configuration can be analyzed [26]. Results are also
shown when a jammer is included in an electronic warfare topology. We start
by examining the SNR contour tools for a monostatic LPI emitter.
10.6.1 Monostatic LPI Emitter and the SNR Contour

Chart
A contour chart represents an important analysis tool to quickly quantify the
advantages of any netted radar sensor network and jammer configuration. We
start by examining the monostatic LPI emitter to present the SNR contour
chart that is generated by the MATLAB tools LPISimNet. In a monostatic
LPI radar system, the transmitter and receiver are co-located and can only
intercept a very small portion of the electromagnetic energy scattered from
the target. Much of the signal and its information is lost.
To introduce the SNR contour analysis, we revisit the monostatic radar
configuration. The SNR for a monostatic configuration can be written as
PCW Gt Gr T 2
SNR1 = (10.37)
(4)3 kT0 FR BR RT4 L
To develop a useful analysis tool for an LPI emitter, a contour chart is con-
structed with the RCS T = 1m2 . This normalized contour chart can easily
be scaled for any RCS. In addition, the SNR is only dependent on the radar
properties and target range. By plotting the results in a 2-D geometric map,
the SNR of the radar can be read as shown in Figure 10.22.
Figure 10.22: Example of SNR contour chart for a monostatic LPI emitter.
This chart illustrates the SNR contour generated by the MATLAB soft-
ware contained on the CD (LPIsimNet.m). For this simulation, the Pilot
radar is used with an ERPR = PCW Gt = 1,000W, Ae = 0.0815 m2 and is
the eective receiving aperture area (equal to Gr 2 /2), and noise power
kT0 FR BRi = 7.5 1013 W. For any target position selected, the value of
SNR can be read from the figure.
10.6.2 Three Netted LPI Emitters

To demonstrate the advantages of a netted LPI radar configuration, three
radar systems are simulated within a 2,500-km2 region using the MATLAB
Figure 10.23: Sensor network containing three LPI emitters.
system of systems software (LPIsimNet). The objective of this simulation is

to display and compare the SNR contour chart with network synchronization
and without network synchronization. The sensor network shown in Figure
10.23 indicates the three radar nodes within an area of 2,500 km2 . The
black asterisk at the position (15, 25) indicates the target, which has an RCS
= 1m2 . In this normalized presentation, the contour analysis can easily be
scaled to any target RCS value. The properties of each LPI emitter used
in the simulation are shown in Table 10.10. The detailed analysis report
provided by LPIsimNet, is referenced to the user selected target position.
For the sensor network displayed in Figure 10.23, the simulation results
are shown in a contour map in order to quantify the SNR quickly and eval-
uate the benefits of the sensor network configuration [26]. Figure 10.24
shows the contour chart of the three emitters when the sensor network is
disabled. The SNR values for each emitter are independent and can be
read directly on the map. For the no network configuration, the SNR =
48.7 dB at the target (node 4) as shown in Figure 10.24. For the sensor
networkenabled configuration, the contour chart is shown in Figure 10.25.
The SNR = 43.2 dB. That is, the netted radar configuration increases the
SNR 5.71 dB over the no network configuration as shown in Figure 10.24.
Table 10.10: LPI Sensor Network: Parameters for the Three Emitter Nodes
Node Index 1 2 3
Type Blue Force Blue Force Blue Force
Name Radar1 Radar2 Radar3
ERP (W) 1000 100 10
Ae (m2 ) 0.0815 0.0815 0.0815
Noise Power (W) 7.5 1013 1 1012 1.5 1012
Position (Km) (15, 40) (15, 15) (30, 25)
Figure 10.24: SNR contour chart for three emitters without sensor network.
Figure 10.25: SNR contour chart for three emitters with sensor network.
Figure 10.26: Jammer attack on a sensor network containing two LPI emit-
ters. Target is 1 m2 at position (15, 25) km.
10.6.3 Two Netted LPI Emitters with Jammer

To quantify the eects of an electronic attack on a netted radar system, two
LPI emitters are placed in a topology with the 1-m2 target. A jammer to
attack both emitters is added as shown in Figure 10.26. The parameters
for the two LPI emitters and the jammer (onboard an Su-34) are shown in
Table 10.11. By comparing the signal-to-jam ratio (SJR) contour chart, with
and without networking, the advantages of a sensor network in an electronic
warfare configuration can be identified.
The contour results for the SJR when no network is used are shown in
Figure 10.27. The contour results for the SJR when the LPI emitters use
a sensor network are shown in Figure 10.28. The SJR improvement in the
sensor network case is 5.75 dB.
10.7 Orthogonal Waveforms for Netted Radar

The increased area of coverage using a system of netted radar systems, each
diverse and independent, make netted radar sensing and the development
of appropriate waveforms an important area of investigation. Multiradar
systems can operate in both monostatic and multistatic modes simultaneously
Table 10.11: LPI Sensor Network: Parameters for the Two Emitters and One
Jammer
Node Index 1 2 3
Type Blue Force Blue Force Hostile Jammer
Name Radar1 Radar2 Su-34
ERP (W) 1000 100 10
Ae (m2 ) 0.0815 0.0815 -
Noise Power (W) 7.5 1013 1 1012 -
Position (Km) (15, 40) (15, 15) (30, 25)
Figure 10.27: SJR for jammer attack on two LPI emitters. Target is 1 m2 at
position (15, 25) km.
Figure 10.28: Jammer attack on a sensor network containing two LPI emit-
ters. Target is 1m2 at position (15, 25) km.
and can retain the favorable features of both a monostatic radar system and
a multistatic radar system if each system transmits a distinct signal from a
set of orthogonal signals in which any two signals are not cross-correlated and
each receiver uses multiple matched filters. As discussed in Chapter 5, in order
to achieve high range resolution and multiple target resolution, the periodic
autocorrelation function of any transmitted code sequence should have a low
peak side lobe level (PSL). For moving targets, Doppler loss occurs at the
matched filter output of the correlation receiver. For a sequence of length Nc ,
with Doppler shift fd and signal duration T the excessive phase increments
from one sequence to the next is 2fd T /Nc .
Netted LPI radar systems require a code sequence with low PSL, resis-
tance to Doppler loss and the use of orthogonal waveforms that have a low
cross-correlation between them. This is to avoid interference and to provide
independent information about the target at various angles. The concept of
orthogonal netted radar systems is dierent than the traditional netted radar
systems. Consider the multiradar system shown in Figure 10.19 consisting of
L LPI radar systems where each system transmits a distinct low power CW
signal using an orthogonal code set {sl (t), l = 1, 2, 3, . . . , L}. Any two signals
in the set are uncorrelated or
C sp (t)sq (t + )dt = 0 (10.38)

t
for p = q and p, q = 1, 2, . . . , L. For high-range resolution, the aperiodic

autocorrelation function of any code sl (t) in the code set should be close to
an impulse function or
1
A= sl (t)sl (t + ) = 1 (10.39)
E t
for = 0 and A = 0 otherwise.

If the radar stations in the multiradar system transmit signal {sl (t), l =
1, 2, 3, . . . , L}, respectively, any radar station can choose to receive and process
any of the L signals by including a matched filter that correlates to the trans-
mitted signal only. The target echoes from the other signals generate near-
zero outputs at the matched filter because it does not correlate with any of
them. If a radar system is equipped with only a matched filter that correlates
to its own transmitted signal, then the system will only operate in the mono-
static mode. If there are multiple parallel matched filters at a receiver that
correlate to the signals that are transmitted by other radar stations, multiple
detection results of targets are available for integration processing (coherently
or noncoherently).
The waveforms used by netted radar systems must be carefully designed to
avoid the self-interference and detection confusion. An orthogonal waveform
set is a group of waveforms in which each of the waveforms has the nearly ideal
noise-like aperiodic autocorrelation property and any two of them have no

cross-correlation. If the emitter transmits the waveforms from an orthogonal
coding waveform set, it can adaptively operate based on the environments and
real-time needs in regular monostatic mode or in a multistatic mode with the
same carrier frequency.
10.7.1 Orthogonal Polyphase Codes

Orthogonal netted radar systems require a set of orthogonal transmit signals
with properties as outlined in (10.38) and (10.39) [27]. Orthogonal polyphase
codes have several advantages for the emitter over the use of binary phase
codes. The orthogonal polyphase codes have a larger main lobe-to-side lobe
ratio than binary signals with the same code length Nc . They also have a
more complicated signal structure making the signal harder to detect and
analyze by a noncooperative intercept receiver.
The orthogonal polyphase code set consists of L signals with each sig-
nal containing Nc subcodes and can be represented by the complex number
sequence
sl (n) = ejl (n) (10.40)
where n = 1, 2, . . . , Nc and l = 1, 2, . . . , L where l (n), (0 l (n) < 2) is
the phase of subcode n of signal l in the signal set. If the number of distinct
phases available to be chosen for each subcode in a code sequence is Mc ,
the phase for a subcode can only be selected from the following admissible
values [27]
2 2 2
l (n) 0, ,2 , . . . , (Mc 1) (10.41)
Mc Mc Mc
or
l (n) = {1 , 2 , . . . , Mc } (10.42)
Considering a polyphase code set S with code length Nc , set size of L and
distinct phase number Mc , the phase samples of S can be represented with
the L Nc phase matrix
1 (1) 2 (2) 1 (Nc )

2 (1) 2 (2) 2 (Nc )
S(L, Nc , Mc ) = .. .. .. .. (10.43)
. . . .
L (1) L (2) L (Nc )
where the phase sequence in row l is the polyphase sequence of signal l and
all the elements in the matrix can only be chosen from the phase set in
(10.41), (10.42). From (10.38) and (10.39), it can be shown that the aperiodic
autocorrelation of the polyphase sequence sl and cross-correlation properties
of orthogonal polyphase codes sp and sq will satisfy or nearly satisfy [27]

Nc k

1
exp j[l (n) l (n + k)] = 0 0 < k < Nc
Nc
n=1
A(l , k) = Nc

N1c exp j[l (n) l (n + k)] = 0 Nc < k < 0
n=k+1
(10.44)
for l = 1, 2, . . . , L and
Nc k

1
exp j[q (n) p (n + k)] = 0 0 < k < Nc
Nc
n=1
C(l , k) Nc

N1c exp j[q (n) p (n + k)] = 0 Nc < k < 0
n=k+1
(10.45)
for p = q and p, q = 1, 2, . . . , L where k is the discrete time index.
To design the polyphase code set with the properties given in (10.44) and
(10.45), the minimization of a cost function that is based on the total autocor-
relation side lobe energy and the cross-correlation energy is performed. This
minimization then leads to uniformly distributed autocorrelation side lobe
and cross-correlation energies among all possible locations thus minimizing
the autocorrelation side lobe peaks and cross-correlation peaks. Given values
of Nc , Mc and L, the energy-based cost function used is
L Nc 1 L1 L Nc 1
E= |A(l , k)|2 + |C(p , q , k)|2 (10.46)
l=1 k=1 p=1 q=p+1 k=(Nc 1)
where is the weighting coecient between the autocorrelation function and

the cross-correlation function in the cost function. Minimization of this func-
tion with Nc = 40, L = 4 and Mc = 4 generates a group of polyphase values
that are orthogonal as shown in Table 10.12. Minimization of the energy
cost function (10.46) was accomplished with a simulated annealing statistical
optimization algorithm [27] that was chosen for its ability to avoid becoming
trapped in a local optima during the search process. The autocorrelation side
lobe peaks (diagonal terms) and cross-correlation peaks (o diagonal terms) of
the polyphase code set are shown in Table 10.13. These results were obtained
using = 1. A larger value of means that the cross-correlation energy
is weighted more in the cost function and leads to smaller cross-correlation
peaks [27].
The four orthogonal polyphase Nc = 40 sequences were generated with a
carrier frequency fc = 1,000 Hz, fs = 7,000 Hz, cpp = 1. The power spectrum
magnitude of the signal with Code 1 is shown in Figure 10.29. The polyphase
shift for the Nc = 40 orthogonal codes for Code 1 are shown in Figure 10.30.
The Nc = 40 orthogonal polyphase shifts for Code 2 are shown in Figure
10.31 and Code 3 are shown in Figure 10.32. The polyphase shift for the
Table 10.12: Phase Sequences of a Polyphase Code Set with Nc = 40, L = 4,

and Mc = 4 (from [27])
No. Code 1 Code 2 Code 3 Code 4

1 /2 3/2 3/2 /2
2 0 0 /2
3 0 /2 0
4 3/2 /2 0
5 /2 /2 3/2 0
6 3/2 0 /2 /2
7 3/2 0 3/2 3/2
8 /2 /2
9 /2 /2 /2 3/2
10 3/2 /2
11 /2 0 /2 0
12 /2 /2 /2 3/2
13 3/2 /2 3/2
14 3/2 /2 0
15
16 /2 /2 0
17 /2 /2 0 0
18 3/2 3/2 /2 /2
19 0 0 3/2 3/2
20 /2 3/2
21 0 3/2 /2 /2
22 3/2 /2 3/2 3/2
23 /2 /2 0 0
24 /2 3/2 /2
25 /2 /2 0
26 3/2 0 0 3/2
27 3/2 0 0 0
28 3/2 3/2 0
29 /2 /2 0 /2
30 3/2 /2 /2
31 3/2
32 3/2 /2 3/2
33 3/2 /2
34 /2 /2 /2
35 3/2 3/2 0
36 /2 /2 3/2
37 3/2 0 /2
38 /2 0
39 3/2 3/2
40 /2 /2 0
Table 10.13: Autocorrelation Side Lobe Peaks (Diagonal Terms) and Cross-
Correlation Peaks (O-Diagonal Terms) of Orthogonal Polyphase Code Set
with Nc = 40, L = 4 and Mc = 4. (from [27])
Code 1 Code 2 Code 3 Code 4

Code 1 0.1521 0.2062 0.1904 0.2121
Code 2 0.2062 0.1414 0.2064 0.1768
Code 3 0.1904 0.2064 0.1346 0.2016
Code 4 0.2121 0.1768 0.2016 0.1820
Nc = 40 orthogonal codes for Code 4 are shown in Figure 10.33. Note the
cross-correlation between any two of these four codes is approximately zero.
The ACF and PACF for the Code 1 sequence is shown in Figure 10.34. Note
the PSL = 16 dB. The PAF is shown in Figure 10.35. The characteristics
of the other three codes are very similar. Note the low Doppler side lobes in
PAF. The polyphase code sequences described in the section can be generated
using ortho40.m in the LPIT.
10.7.2 Addressing Doppler Shift Degradation

For moving targets, the polyphase sequences above degrade severely in the
presence of small Doppler shifts. The Doppler loss results in a degradation
of the autocorrelation and cross-correlation properties at the matched filter
outputs of the correlation receiver. In [28], an algebraic design method for
generating polyphase orthogonal sequences with good Doppler tolerance is
presented. The method uses a Hadamard matrix construction technique with
circulant matrices3 based on polyphase complementary sequences (sum of
their aperiodic autocorrelation functions equals zero except for the zero shift).
In [28], Frank complementary sequences are used to create the Hadamard ma-
trix. For their Nc = 36 length sequence, a Doppler tolerance of |fd T | = 1.7
was achieved compared to 0.7 for the Deng sequences above with Nc = 40. De-
pending on the allowable reduction in output SNR, this implies that the tol-
erant waveform, when detecting a subsonic target (v = 300m/s) at S-band
( = 0.1m), for which fd 6 kHz, would be limited to T < 1.7/12, 000 =
0.141 ms. Code lengths beyond this value would require multiple Doppler
filters (or correlators) to retain sensitivity to subsonic targets of unknown
velocity. The mean autocorrelation PSL = 16 dB compared to the Deng
sequences with PSL = 16.3 dB. Although the length of the codes that can
be developed is constrained, the waveform design methodology addresses all
three issues (autocorrelation, cross-correlation and Doppler tolerance).
3 A circulant matrix is a special type of Toeplitz matrix where each row vector is shifted
one element to the right relative to the preceding row vector.

Figure 10.29: Power spectrum magnitude of Code 1 with fc = 1,000 Hz.
Figure 10.30: Orthogonal polyphase shifts for Code 1 Nc = 40 phase shifts.


Figure 10.34: Orthogonal polyphase Code 1 ACF and PACF.

Figure 10.35: Orthogonal polyphase Code 1 PAF.
In [29], a new set of polyphase sequences is presented with good corre-

lation properties as well as resilience to Doppler shifts. The sequences are
built using a numerical cross entropy optimization based on correlation prop-
erties and a structural constraint is also imposed on the optimized polyphase
sequences so that Doppler tolerance is maintained. Deng [27] suggested im-
proving Doppler tolerance by using an ambiguity based optimization cost
function to include reciprocals for the main lobe peaks, side lobe peaks and
the cross-correlation peaks for all possible Doppler frequencies. This method
however, is computationally costly for even short code lengths.
As another method, recall the polyphase code sets described in
Chapter 5 of Frank, P1P4. For these codes, there is a harmonic relationship
of phases from one sequence element to the next that aids in the ability of
the code to resist the Doppler loss. In [29], an algorithm is described where
this harmonically related structure is applied as a constraint and added to
the correlation cross-entropy optimization algorithm to improve the Doppler
tolerance. The technique can be used to construct arbitrary length sequences
for an arbitrary number of transmitters. Table 10.14 lists three polyphase
sequences of length Nc = 40. Table 10.15 shows the autocorrelation PSLs
and cross-correlation peaks for these sequences. The mean PSL for these se-
quences is 17.3 dB and the mean cross-correlation is 14.3 dB; these figures
improve on the Deng sequences of the same length discussed above. The op-
timized polyphase code sequences described in the section can be generated

using ortho40CE.m in the LPIT.
10.7.3 Orthogonal Frequency Hopping Sequences

Due to the flexible locations of the multitransmitters and multireceivers, the
netted radar system of systems has much stronger capability in target de-
tection, tracking, recognition and electronic protection compared with the
conventional emitter. In Chapter 6, an algebraic method was presented to
derive a single discrete frequency-coding Costas sequence with NF frequencies
for use in a monostatic frequency hopping LPI emitter. Although the Costas
arrays have nearly an ideal periodic autocorrelation property, for any two or
more constructed Costas arrays there is no guarantee that any two sequences
will have a nearly zero cross-correlation property.
In [30], a set of NF discrete frequency hopping waveforms with good auto-
correlation and nearly zero cross-correlation were derived by using a numerical
optimization technique. In this technique, the result is achieved through min-
imizing a cost function that measures the degree to which a result satisfies the
design requirements. In frequency hopping waveform design, the cost function
is chosen as in the previous section as the sum of the total autocorrelation
side lobe energy for each waveform in the set and the total cross-correlation
energy for all distinct combinations of two waveforms sp (t), sq (t) in the set.
Thus, the cost function to be minimized for the discrete frequency hopping
waveform design is
L L1 L
E= |A(sl , )|2 + |C(sp , sq , )|2 d (10.47)
l=1 p=1 q=p+1
Details on the minimization algorithm can be found in [30]. Table 10.16 lists
the three frequency hopping sequences of the designed waveform set with
NF = 32 and L = 3. The autocorrelation side lobe peaks and cross-correlation
peaks of the designed frequency hopping sequence sets in
Table 10.16 are given in Table 10.17.
The discrete frequency hopping sequence Code 1 was generated with a
base frequency multiplier of 1,000 Hz. The NF = 32 codes were sampled
with fs = 100 kHz with tp = 0.001 s. The power spectrum magnitude of
the discrete frequency hopping sequence Code 1 is shown in Figure 10.36.
The ACF and the PACF are shown in Figure 10.37 and the PAF is shown in
Figure 10.38. Note the extremely well behaved time and Doppler side lobe
levels. The PSL = 20 dB. The discrete frequency coding waveforms can be
generated using dfc32.m in the LPIT.
Table 10.14: Optimized Cross-Entropy Sequences with Nc = 40 (from [29])
No. Code 1 Code 2 Code 3

1 8/5 3/5 4/5
2 /5 2/5
3 6/5 0
4 4/5 6/5 0
5 9/5 9/5
6 2/5 9/5 9/5
7 27/20 3/20 21/20
8 21/20 3/5 9/20
9 3/10 27/20 6/5
10 3/20 3/10 9/20
11 27/20 27/20 21/20
12 27/20 6/5 27/20
13 /5 /2 4/5
14 0 4/5 9/10
15 /10 /2 /10
16 4/5 0 /5
17 0 7/10 9/10
18 3/5 4/5 9/10
19 /20 0 0
20 /5 /20 /20
21 /20 /10 /20
22 /4 /20 /4
23 9/10 4/5 0
24 9/10 /10 9/10
25 /5 4/5 0
26 7/10 /10 4/5
27 0 9/10 4/5
28 /5 3/5 /5
29 6/5 6/5 27/20
30 27/20 9/20 3/20
31 0 9/20 27/20
32 0 0 27/20
33 6/5 27/20 3/10
34 3/10 6/5 3/4
35 7/5 2/5 6/5
36 3/5 6/5 3/5
37 /5 7/5 3/5
38 6/5 3/5 6/5
39 6/5 0 7/5
40 8/5 9/5 7/5
Table 10.15: Autocorrelation Side Lobe Peaks (Diagonal Terms) and Cross-
Correlation Peaks (O-Diagonal Terms) for the Cross Entropy Sequence Set
of Length Nc = 40 (from [29])
Code 1 Code 2 Code 3

Code 1 0.1365 0.1820 0.1799
Code 2 0.1820 0.1303 0.1840
Code 3 0.1799 0.1840 0.1413
Table 10.16: Discrete Frequency Hopping Sequences with NF = 32 and L = 3

(from [30])
No. Code 1 Code 2 Code 3

1 31 2 31
2 28 11 13
3 1 12 18
4 30 14 20
5 3 29 11
6 8 16 7
7 10 6 29
8 12 1 27
9 7 9 8
10 25 21 22
11 11 24 0
12 15 23 1
13 23 5 21
14 13 26 14
15 27 19 9
16 22 3 17
17 26 7 5
18 17 30 25
19 5 13 26
20 21 8 10
21 6 20 19
22 2 17 16
23 29 18 30
24 14 4 15
25 16 0 12
26 19 28 23
27 9 31 4
28 20 15 3
29 18 27 6
30 0 10 2
31 24 25 28
32 4 22 24
Table 10.17: Autocorrelation Side Lobe Peaks and Cross-Correlation Peaks

of the Discrete Frequency Hopping Sequences in Table 10.16 (from [30])
Code 1 Code 2 Code 3

Code 1 0.0764 0.0979 0.1250
Code 2 0.0979 0.0881 0.1068
Code 3 0.1250 0.1068 0.0855
Figure 10.36: Power spectrum magnitude of the orthogonal discrete frequency

hopping Code 1 with fs = 100 kHz.
Figure 10.37: ACF and PACF of the orthogonal discrete frequency hopping
Code 1 with fs = 100 kHz.
10.7.4 Noise Waveforms

The concept of a multiuser, wireless netted LPI radar system using random
noise is proposed in [31]. The proposed system uses noise signals for radar
surveillance and a multiradar communication network for transferring the
sensor data to a central command center where informed decisions can be
made. Due to the spectral characteristics of the UWB random noise wave-
form, an LPI capability is provided while also eciently sharing the frequency
spectrum with other users. A number of UWB random noise radar systems
can operate over the same frequency band with minimal cross-interference
since each transmitted noise waveform is uncorrelated with the others. It
is this property that allows a number of the UWB noise radars to be inte-
grated into a NCW architecture [31]. The bandlimited noise (12 GHz) is
also notch filtered (1.21.3 GHz) to provide room for the intrasensor network
communications among the dierent emitters. The spectral fragmentation
for the embedded communication causes no distortion if the gap in the noise
Figure 10.38: PAF of the orthogonal discrete frequency hopping Code 1 with
fs = 100 kHz.
Figure 10.39: Noise radar concept with (a) element-space approach and
(b) beam-space approach (after [32]).
band is not excessive ( 30%). The sensor data to be shared is modulated

onto a CW signal whose frequency lies within the notch of the UWB noise
signal. Orthogonal frequency division multiple access modulation is used for
transporting the target data between sensors.
The development of a netted noise radar is also presented in [32]. Two
transmission approaches are compared as shown in Figure 10.39. The first
approach shown in Figure 10.39(a), is the element space approach where mul-
tiple channels (antennas) of independent noise are transmitted. K incoherent
noise sources are transmitted. Ignoring the angular variation in the targets
RCS, the received power is independent of the angle of the scatterer from
the transmit array. The second approach shown in Figure 10.39(b), is the
beam-space approach where each independent noise source is fed into each
antenna but is delayed i (or phase shifted) so as to form a beam illuminat-
ing a selected sector of the radar field of view (FOV). This eectively codes
each sector in the FOV according to a particular noise source. The direc-
tion of each sector is determined by the delay (or phase shift) and the width
is determined by the beamwidth of the array. Comparison of element- and
beam-space approaches to the netted noise radar indicate that when oper-
ating the transmit array at frequencies such that d/ < 0.5 where d is the
receiver spacing, the beam-space approach is a more ecient method of con-
centrating the wave number spectrum of the transmit signal in the radiating
region and minimizes the problem of nonradiating waves.
10.8 Netted Over-the-Horizon Radar Systems

Future OTHR systems are expected to deliver dramatically improved capa-
bility in every performance dimension including LPI. There are three direct
benefits to applying MIMO radar concepts to OTHR [33]. First it provides
a means to implement radar management trade-os between radar sensitiv-
ity and surveillance footprint coverage allowing a more ecient use of re-
sources (surveillance area requirements, sensitivity, target dynamic behavior,
and the interaction of the target characteristics with the data processing al-
gorithms). MIMO radar is also a convenient method to implement adaptive
processing algorithms on transmit for clutter mitigation. By changing the
illumination source at the transmit array, the clutter is more eectively sup-
pressed. Consequently, orthogonal waveforms have also found application in
netted OTHR systems [33]. The use of multiple simultaneously transmitted
orthogonal waveforms permit better sensitivity and more flexible trade-os
in footprint coverage. It also allows for adaptive management of the trans-
mitted beam to minimize clutter and simplifies propagation mode selection
for improved clutter rejection.
In an OTHR, both the transmitter and receiver subsystems can be con-
sidered as M and N dimensional digital arrays. The transmit subsystem
consists of one waveform generator per transmit power amplifier and trans-
mit antenna element. The receive subsystem consists of one digitizing receiver
per array element. Achieving full orthogonality with the CW waveform set
over the space-time ambiguity of concern is not possible. Space-time adaptive
processing using multiple transmitters and receivers allows using one wave-
form generator per transmit element and enabling the transmit and receive
beamforming to be performed entirely at the receive site [34]. The diversity
of target scattering leads to better detection performance using lower power
waveforms.
Orthogonal waveforms that can be used in OTHR include time-staggered
FMCW, Doppler oset FMCW and noise waveforms. The time-staggered
FMCW uses a time oset between dierent FMCW waveforms to exploit the
fact that the range interval of interest is frequently limited by ionospheric
propagation. In surface modes (high Doppler resolution mode), low WRFs
are used over extended CITs. It is therefore possible to provide orthogo-
nality between a number of waveforms after range correlation with a single
reference waveform. The approach maintains the attractive power eciency
and spectral occupancy of the FMCW waveform. For the Doppler oset
FMCW a small frequency oset between FMCW waveforms provides orthog-
onality after slow-time Doppler processing (slow-time MIMO). In this case,
the Doppler extent of the clutter and expected target Doppler shifts limit
the number of concurrent orthogonal waveforms that can be supported. This
waveform approach is more appropriate to the air-mode for aircraft detection.
The band-limited noise waveforms also provide an orthogonal waveform choice
despite the large peak to average power ratio and poor Doppler sensitivity.
This approach provides a broad class of orthogonal waveforms.
References
[1] Cebrowski, A. K., Implementation of Network-Centric Warfare, Oce of
Force Transformation, 2004.
[2] Ahmed, A. A., Hongchi, S., and Shang, Y., A survey on network protocols
for wireless sensor networks, Proc. of the IEEE International Conf. on
Information Technology: Research and Education, pp. 301305, Aug. 2003.
[3] Karlof, C. and Wagner, D., Secure routing in wireless sensor networks: at-
tacks and countermeasures, Proc. of the IEEE International Workshop on
Sensor Network Protocols and Applications, pp. 113127, 2003.
[4] Stein, F., Garska, J., and McIndoo, P.L., Network-centric warfare: Impact
on Army operations, EUROCOMM 2000 Information Systems for Enhanced
Public Safety and Security, IEEE/AFCEA, pp. 288295, 2000.
[5] Kruse, J., Adkins, M. and Holloman, K. A., Network centric warfare in the
U.S. Navys fifth fleet, Proc. of the IEEE 38th Hawaii International Conf.
on Systems Sciences, 2005.
[6] Chen, Y-Q, and Pace, P. E., Simulation of Information Metrics to Assess
the Value of Networking in a General Battlespace Topology, Proc. of the
IEEE International Conf. on System of Systems Engineering, June 2008.
[7] Boyd, J.R., A discourse on winning and losing, Title Boyd gave to his
collection of briefings on competitive strategy (widely known as the Green
Book, with apologies to Wittgenstein), 1987.
[8] Ling, F.M., Moon, T., and Kruzins, E., Proposed network centric warfare
metrics: From connectivity to the OODA cycle, Military Operations Re-
search, vol. 10 No. 1, pp. 513, 2005.
[9] Shannon, C. E., A mathematical theory of communication, Bell Systems
Technical Journal, July 1948.
[10] Phister, P.W. Jr., and Cherry, J.D., Command and control concepts within
the network-centric operations construct, Proc. of the IEEE Aerospace
Conf., pp. 19, Mar. 411, 2006.
[11] Posiel, R. Modern Communications Jamming Principles and Techniques, Artech
House Inc., 2004.
[12] Hume, A. L. and Baker, C. J., Netted radar sensing, Proc. of the CIE
International Conf. on Radar, pp. 110114, Oct. 2001.
[13] Teng, Y., Griths, H. D., Baker, C. J., Woodbridge, K., Netted radar sen-
sitivity and ambiguity, IET Radar and Sonar Navig., Vol. 1, No. 6, pp.
479486, 2007.
[14] Derham, T.E., Doughty, S., Woodbridge, K., Baker, C.J., Design and eval-
uation of a low-cost multistatic netted radar system, IET Radar, Sonar &
Navigation, Vol. 1, No. 5, pp. 362368, October 2007.
[15] Lehmann N. H., Haimovich, A. M., Blum, R. S., and Cimini, L., High
resolution capabilities of MIMO radar, Record of the Fortieth Asilomar Conf.
on Signals, Systems and Computers, pp. 2530, 2006.
[16] Bliss, D. W., and Forsythe, K. W., Multiple-input multiple-output (MIMO)
radar and imaging: degrees of freedom and resolution, Record of the Thirty-
Seventh Asilomar Conf. on Signals, Systems and Computers, Vol. 1, pp.
5459, Nov., 2003.
[17] Sammartino, P.F., Baker, C.J., and Griths, H.D., Target model eects
on MIMO radar performance, Proc. of the IEEE International Conf. on
Acoustics, Speech and Signal Processing, pp. V-1129V-1132, May, 2006.
[18] Haimovich, A. M., Blum, R. S., and Cimini, L. J. Jr., MIMO radar with
widely separated antennas, IEEE Signal Processing Magazine, pp. 116129,
Jan. 2008.
[19] Hume, A. L. and Baker, C. J., Netted radar sensing, Proc. of the IEEE
Radar Conf., pp. 2326, 2001.
[20] Baker, C. J., and Hume, A. L., Netted radar sensing,IEEE Aerospace and
Electronic Systems Magazine, Vol. 18, No. 2, pp. 36, Feb. 2001.
[21] Fishler, E., Haimovich, A., Blum, R., Cimini, L. J., Chizhik, D., and Valen-
zuela, R. A., Spatial diversity in radars - models and detection performance,
IEEE Trans. of Signal Processing, Vol. 54, No. 3, pp. 823838, 2006.
[22] Chen, C-Y, and Vaidyanathan, P. P., Properties of the MIMO radar ambi-
guity function, Proc. of the IEEE International Conf. on Acoustics, Speech
and Signal Processing, pp. 23092312, 2008.
[23] Teng, Y., Baker, C. J., and Woodbridge, K., Netted radar sensitivity and
the ambiguity function, Proc. of the International Conf. on Radar, Vol. 1,
No. 6, pp. 14, Dec. 2006.
[24] Papoutsis, I., Baker, C. J., and Griths, H. D., Netted radar and the ambi-
guity function, Proc. of the IEEE International Radar Conf., pp. 883888,
2005.
[25] Li, J., Stoica, P., Xu, L., and Roberts, W., On parameter identifiability of
MIMO radar, IEEE Signal Processing Letters, Vol. 14, No. 12, pp. 968971,
Dec. 2007.
[26] Chen, Y-Q, and Pace, P. E., Simulation of Network-Enabled Radar Systems
to Assess the Value of Jamming in a General Radar Topology, Proc. of the
IEEE International Conf. on System of Systems Engineering, June 2008.
[27] Deng, H., Polyphase code design for orthogonal netted radar, IEEE Trans.
on Signal Processing, Vol. 52, No. 11, pp. 31263135, Nov. 2004.
[28] Khan, H. A., and Edwards, D. J., Doppler problems in orthogonal MIMO
radars, Proc. of the IEEE Conf. on Radar, pp. 244247, 2006.
[29] Khan, H. A., Zhang, Y., Ji, C. Stevens, C. J., Edwards, D. J., and OBrien, D.,
Optimizing polyphase sequences for orthogonal netted radar, IEEE Signal
Processing Letters, Vol. 13, No. 10, pp. 589592, 2006.
[30] Deng, H., Discrete frequency-coding waveform design for netted radar sys-
tems, IEEE Signal Processing Letters, Vol. 11, No. 2, pp. 179182, Feb.
2004.
[31] Surender, S. C., and Narayanan, R. M., Covert netted wireless noise radar
sensor: OFDMA-based communication architecture, Proc. of the Military
Communications Conf., pp. 17, Oct. 2002.
[32] Gray, D. A., and Fry, R., MIMO noise radarelement and beam space com-
parisons, Proc. IEEE Waveform Diversity & Design Conf., pp. 344347,
2007.
[33] Frazer, G. J., Johnson, B. A., Abramovich, Y. I., Orthogonal waveform
support in MIMO HF OTH radars, Proc. IEEE Waveform Diversity &
Design Conf., pp. 423427, 2007.
[34] Frazer, G. J., Abramovich, Y. I., and Johnson, B. A., Spatially waveform
diverse radar: perspectives for high frequency OTHR, Proc. IEEE Radar
Conf., pp. 385390, April 2007.
Problems
1. A netted LPI radar transmits the target parameters using 1.0 and 0.0-
V levels with a probability of 0.2 each and 3.0- and 4.0-V levels with a
probability of 0.3 each. Determine the average information being sent.
2. A C2 operator uses a numerical keypad that has the digits 0, 1, 2, 3,
4, 5, 6, 7, 8, and 9. Assume that the probability of sending any one
digit is the same as that for sending any of the other digits. Calculate
how often the operator must press the buttons in order to send out
information at the rate of 2 bits/s.
3. An army field computer has 110 characters on the keyboard and each
character is sent using binary words. (a) What is the number of bits
required to represent each character? (b) How fast can the characters
be sent (characters/s) over a channel if the channel bandwidth is 3.2
kHz and the SNR=20 dB? (c) What is the entropy of each character if
each is equally likely to be sent?
4. A 480-by-500 pixel range-Doppler image is to be transmitted from a
netted LPI radar where each pixel can have one of 32 intensity values.
The emitter sends 30 images/s. If all image elements are assumed to
be independent and all 32 intensity levels are assumed to be equally
Figure 10.40: Destruction of terrorist weapon system.
likely, (a) determine the source rate in (bits/s). (b) If the image is to
be transmitted over a channel with a 4.5-MHz bandwidth and a 35-dB
SNR, find the capacity of the channel (bits/s).
5. The USS Enterprise (capability K = 1.0) has launched a Tomahawk
missile to destroy a terrorist weapon system as shown Figure 10.40. To
follow up with a damage report, a predator UAV follows the Tomahawk.
To provide an intelligence, surveillance and reconnaissance (ISR) com-
ponent, a Global Hawk (K = 0.8) is also used. The Tomahawk missile
capability is given as K = 0.3 due to its limited connectivity and signal
rerouting options (note its unidirectional link to the Global Hawk). The
Predator also has a limited signal rerouting capability and is given the
capability value K = 0.5. Consider each link to have a flow component
and value component of either 1 or zero (i.e., F = L = 1, 0). (a) Find
R
the reference connectivity measure CM . (b) Determine the network
reach IR . (c) Determine the network reach if a jammer takes out the
link between the USS Enterprise and the Tomahawk. (d) To examine
the impact of the rerouting options of the original network configura-
tion (unjammed), let each node capability be K = 1.0 and determine
the network reach.
6. Consider problem 5. (a) If the rate that information is processed by

node = 1 is 1 = 50 kHz, node = 2 is 2 = 18 kHz, node =
3 is 3 = 260 kHz, and node = 4 is 4 = 3 kHz, determine the
network richness RQ (average rate that knowledge is generated through
the network). (b) Using the network richness calculated in Problem 5
above, determine the networks characteristic tempo T . (c) Consider
that the command and control (C2) decision making speed to target
the Tomahawk missile is C2 = 0.5 kHz. After the decision is made, it
takes 0.5 ms to send the retargeting command to the missile (1 = 2
kHz). After the missile receives the retargeting command, the missile
guidance takes 0.25 ms to initiate the turn (2 = 4 kHz). Calculate the
OODA operational tempo OODA for this retargeting command.
7. Behavior of the OODA tempo: The action tempos of a force will vary
vastly depending on the specific net-centric EW situation; it is not
possible to make general statements about their scale. Therefore, the
action tempos can be treated as adjustable parameters in OODA . To
illustrate the behavior of the OODA tempo with respect to the network
tempo, we can normalize both by the C2 tempo. Using MATLAB,
plot on the same graph the OODA /C2 (vertical axis) versus T /C2
(horizontal axis) for the three action tempos: (a) 1 = 0.5C2 and
2 = 0.25C2 (low-action tempo), (b) 1 = C2 and 2 = C2 (medium-
action tempo), and (c) 1 = 2C2 and 2 = 4C2 (high-action tempo).
What can you conclude about the OODA tempo and the C2 tempo as
the network tempo increases?
8. Using the LPIsimNet tools (a) complete the tutorial in Appendix D for
the sensor network analysis and the netted radar analysis. (b) Construct
a sensor network with four nodes, and a jammer. Movement of one or
more nodes should be included with five time index steps. Build an
analysis summary of the results. (c) Construct corresponding netted
radar scenario with four LPI emitters, a target and a jammer. Build
an analysis summary of the results. (d) What insights are you able to
gather from your simulation study?
9. (a) Add to the LPIT, four CW orthogonal PSK signals each with one
of the polyphase codes shown in Table 10.12. Include five periods of the
code set for each signal. (b) Generate the four polyphase signals with
fc = 1 kHz, and B = 500 Hz. (c) Compute the ACF, PACF and PAF
of each signal with N = 1 and N = 3. (d) How do the PSL levels of the
orthogonal PSK waveforms compare with the polyphase codes Frank,
P1P4 ? (e) Compute the cross-correlation of the four PSK signals to
verify Table 10.13.
10. (a) Repeat the problem above for the cross-entropy orthogonal sequences
given in Table 10.14 with correlation properties given in
Table 10.15. (b) Repeat for the discrete frequency hopping sequences
given in Table 10.16 and correlation properties given in Table 10.17
using a scale factor of 102 Hz.
PART II:
INTERCEPT RECEIVER STRATEGIES AND

SIGNAL PROCESSING
Chapter 11
Strategies for Intercepting

LPI Radar Signals
The LPI radar characteristics discussed in the first part of this book (Chap-
ters 110) pose a particular challenge to the noncooperative intercept receiver.
Modern electronic warfare (EW) intercept receivers must perform the tasks
of detection, parameter identification, classification, and exploitation in a
complex environment of high noise interference and multiple signals. The
wideband nature of the LPI emitter signal can force the intercept receiver
to have a significant processing gain by implementing sophisticated receiver
architectures and signal processing algorithms (time-frequency, bifrequency)
in order to determine the waveform parameters. In this chapter, modern
network-centric strategies for EW receiver architectures are discussed, and
a contrast is drawn to the traditional platform-centric approach. This in-
cludes the use of swarm intelligence. In addition, the look-through problem
is discussed in the framework of suppression of enemy defense. Digital re-
ceiver architectures are briefly discussed including the direct RF sampling
approach. EW intercept receiver problems are also emphasized.
11.1 EW Intercept Receiver Techniques

11.1.1 Traditional Approach
Electronic warfare intercept receivers are used to process threats on the mod-
ern electronic battlefield, and consequently, must cover extremely wide bands
from 300 MHz to 100 GHz and above, since they do not know the charac-
teristics of the signal that they are attempting to intercept. The wideband
nature of LPI threat signals presents a significant challenge to the intercept
387
receiver design. The interception of LPI radar signals has been a topic of
investigation for over a decade [13]. Traditionally, EW receivers have been
divided into three categories: radar warning receivers, electronic support re-
ceivers, and electronic intelligence receivers. RWRs are designed to passively
intercept enemy radars in time to enable the pilot to react quickly through
maneuvering or employing appropriate electronic attack techniques. Their
use on the battlefield is time-critical, and combat action is taken directly
from their threat information output. Electronic support receivers encom-
pass all actions necessary to provide the information required for immediate
decisions involving EW operations, threat avoidance, targeting, and homing.
Although not as time-critical as RWRs, information operations rely heavily
on ES receivers for intelligence updates and important operational decisions.
For electronic intelligence receivers, the information provided is extracted
from detailed analysis of radar signals and other noncommunication emit-
ters in a timely manner. Although their operation is the least time critical,
their threat identification is used to update national databases. Examples
of U.S. collection ELINT assets include the U2 Senior Ruby, the Armys
Guard Rail, and the Air Forces RC-135 Rivet Joint. These high-value stand-
o assets typically operate hundreds of kilometers from the emitter and at
a high altitude. Together, these receivers provide the underlying intelligence
needed for weapon systems deployment. In a platform-centric configuration,
each weapon system traditionally had its own receiver system and, if that
receiver was remote, there was a stovepipe communication system providing
the intercept data back to the shooter.
There are limitations to the use of intercept receivers in a platform-centric
configuration. Geometrical limitations include extended stand-o ranges and
alignment problems, which make it especially dicult to detect and jam LPI
emitters. Also, the intercept receiver is limited by look-through. The look-
through process allows the jammer to observe its eectiveness on the LPI
emitter by stopping the jamming assignment to listen periodically. This re-
sults in inecient jammer management, and limited coordination during a
mission.
11.1.2 The Look-Through Problem

To emphasize the look-through problem, consider a frequency-hopping LPI air
defense radar used for targeting a surface-to-air missile (SAM) against an in-
coming strike aircraft. To protect the strike aircraft against these SAM sites,
platforms such as the EA-6B Prowler provide stando (long-range) jamming
of the enemy radar, in what is normally called a suppression of enemy air de-
fense operation.1 The aircraft has a fully integrated electronic warfare system
combining long-range, all-weather receive and jamming capabilities, mainly
1 The EA-6B is also used to protect ground troops and ships, by jamming enemy elec-
tronic data links and communications.

Strategies for Intercepting LPI Radar Signals 389
supplied by the Litton LR-700 intercept receiver, working together with the
AN/ALQ-99 tactical jamming system (TJS). The Litton LR-700 is the new
intercept receiver system in the Increased Capability III (ICAP III) version
of the EA-6B, and gives the Prowler a selective-reactive jamming capability
with the TJS. During a SEAD operation, the LR-700 receiver on board the
suppression aircraft must detect the threat emitters and manage the TJS in
order to prevent the radar from detecting the inbound strike aircraft it is
trying to protect.
During the jamming process, a certain amount of look-through is required.
For example, with an EA-6B reactively jamming a frequency-hopping radar,
the jamming must stop in order to sense the radars transmit frequency. Of
course, the duty cycle of the intercept receiver look-through process must be
less than the time necessary for the radar to sense it is being jammed, and
switch frequencies. The bottom line is that any amount of look-through is
not desired, since this allows the threat radar a window in which to detect
the strike aircraft.
If however, the EA-6B integrates threat parameters from an electronic
order-of-battle database, a reconnaissance aircraft with near real-time on-
scene intelligence collection, analysis, and dissemination capabilities (e.g.,
Rivet Joint), and frequency data from an o-board stand-in sensor (e.g., a
UAV) to cue the on-board intercept receiver (tip and tune), a fast reactive
electronic attack can be performed that eliminates the need for look-through.
For the reactive jamming assignments to be eective, however, the data link
used to provide the cueing data must not induce a delay time of any signifi-
cance to the reactive assignment. That is, if the frequency-hopping radar can
switch frequencies faster than the cueing data can arrive from the o-board
intercept receiver, then the eectiveness is significantly degraded.
11.1.3 Modern Network-Centric Concepts Arriving

Due to the complex emitter modulations now available, and the speed with
which information is shared, the distinction between the roles of RWRs, ES,
and ELINT receivers is fast disappearing, and all capabilities are being in-
tegrated within a single EW receiver system, in order to provide a complete
situational awareness for ships, helos, and high-value aircraft. In addition,
these receivers must now include precision direction finding, countermeasures
control, cueing of weapon systems, enhanced radar warning, fusion of o-
board sensors and databases, and full integration with the electronic combat
system. Other capabilities will include emitter classification and identifica-
tion, emitter-to-platform correlation, detailed analysis, and signal recording.
Eliminating the limitations inherent in a platform-centric configuration
comes from a distributed system of systems. A distributed system of systems
provides significant geometric flexibility, and can reduce or eliminate the need
for look-through. In addition, coordinated jammer responses and improved
Figure 11.1: Disabling an LPI emitter within a network-centric architecture.
jammer management can be achieved due to better information being avail-

able. The ability for EW receivers to acquire, track, and locate conventional
threat emitters and targets, and share this information among stand-o plat-
forms (e.g., for weapons targeting), is an example of a network-centric archi-
tecture, and represents a fundamental shift from a platform-centric approach.
As emphasized in Chapter 10, in a network-centric architecture, the net-
work acts as a force multiplier by networking sensors (e.g., EW receivers),
decision makers, and shooters (e.g., weapons systems), to achieve shared
awareness. The network requires sucient bandwidth for all users to take
advantage of data mining in appropriate databases afloat and ashore. The
architecture is determined mostly by the mission altitudes, signal densities,
reaction times, and modulation analysis that must be performed. Figure 11.1
demonstrates the detection and jamming of an LPI emitter using a network-
centric architecture. See also Figure 10.1 for comparison. The LPI emitter
is detected using a number of sensors that relay the information to both a
command and control point, and the proper shooter. The command and
control then allows the shooter to apply the appropriate electronic attack to
disable the LPI emitter. The shooter also relays its information concerning
the jammed emitter back to both the sensors and the command and control.
That is, instead of each platform making decisions on information received by
only its own intercept receiver (the platform-centric approach), modern EW
receivers integrate information from many sensors and databases for target-
Figure 11.2: Predator UAV with a Hellfire missile.
ing. Through a superior battlespace awareness, forces can employ the best
weapons on the right targets to greatly reduce risk to themselves, and increase
the opportunity for a successful LPI emitter engagement.
11.2 Detecting the LPI Radar with UAVs

The network-centric approach to intercept receiver integration is an important
trend. This capability requires the platforms to be available in the correct
location, and the data links between those platforms to be jam resistant. Un-
fortunately, LPI emitters are becoming increasingly dicult to detect, locate,
and track from stand-o platforms, due to their low peak power (< 1W),
low side lobes (40 dB down), short on-times (ms), high mobility, and use of
terrain masking. As a result, the stand-o platforms with EW receivers are
augmented by specialized receivers that can go to the emitters (stand-in plat-
forms). These specialized receivers are mounted, for example, in unmanned
aerial vehicles such as the Predator, shown in Figure 11.2.
The use of swarm intelligence technology is fast becoming an important
concept in network-centric sensor configurations. Swarm intelligence allows
the design of EW receiver networks to detect LPI emitters, and is inspired by
the behavior of social insects [4]. In a swarm sensor architecture, the signal
collection capability is defined by the group behavior and not the individual
behavior. One advantage of using a UAV swarm of EW receivers is the
ability to behave autonomously, using digital information pheromones (DIPs;
see page 106 in [4]). In what follows we use the analogy of the EW receiver as
being an ant insect that exists within a swarm. The idea is to use another
EW receivers (ants) experience in prior LPI emitter searches. For example,
ants that are looking for particular LPI emitters, access information that has
been left in the form of DIPs (ratings) from previous detections by other
colony members. As LPI emitters are detected, an ant DIP is published

on the network by this ant server. This lets other ants know how many
previous detections of this emitter were found, and the characteristics of the
emitter. Ant DIPs that are continuously detected are continually reinforced,
while those that are disabled (see Figure 11.1) evaporate. Only information
that is regularly verified and reinforced is conserved.
This type of behavior requires only a small number of operators to control
many UAVs. Another advantage is the ability of the UAVs to behave coopera-
tively. Cooperative behavior allows the UAVs to form a robust, self-organizing
and self-adapting sensor architecture, while retaining the intercept function
even in the presence of a loss. The swarm LPI detection architecture requires
only low-cost medium-endurance airframes (expendable), existing wideband
intercept receivers (e.g., R-300A, highly integrated microwave receiver), and
the use of swarm logic [4] with intraswarm communications, using, for ex-
ample, an 802.11 link. With the swarm approach, LPI radars run the risk
of detection (and classification), especially when the intercept receiver incor-
porates advanced signal processing techniques that take advantage of time-
frequency, bifrequency processing. According to the 2002 Defense Acquisition
Board, EA from the UAV platform will become a significant capability.
11.3 Noncooperative Intercept Receivers

The EW community has long debated and ranked many dierent intercept
receiver architectures based on their ability to process signals [5, 6]. The
comparisons, however, have limited usefulness, since dierent mission scenar-
ios require dierent capabilities. What is certain, however, is that future EW
receivers will be digital, and will incorporate various technologies as discussed
below.
11.3.1 Comparison of Classic Receiver Architectures for

Detecting LPI Waveforms
There are many variations of intercept receivers. These passive receivers can
be used to detect the LPI emitter emissions over considerable distances. In
this section, three popular intercept receiver architectures are compared in
terms of their ability to detect several types of LPI emitter waveforms. The
receivers that are compared include the square-law, wideband and channelized
receivers [7]. These receivers are relatively inexpensive, readily accessible and
are shown in Figure 11.3. The square-law receiver is an energy detector.
The parameters of the square-law receiver are given in Table 11.1. The
wideband crystal video receiver is characterized by a wide RF bandwidth to
account for the uncertainty in the intercepted signal parameters. The spec-
ifications of the wideband receiver are given in Table 11.2. The channelized
Figure 11.3: Block diagram of receiver architectures being compared with

(a) square-law, (b) wideband crystal video, and (c) channelized.
Table 11.1: Square-Law Receiver Parameters [7]
Receiver Feature Specification

Noise figure 8 dB
Instantaneous bandwidth 1 GHz
Noise floor 76 dBm
System loss 12 dB
Video bandwidth 60 kHz
Integration time period matched
Detection threshold SNR = 12 dB
Local oscillator 12 MHz
IF bandwidth 12 MHz
Table 11.2: Wideband Crystal Receiver Parameters [7]

Noise figure 8 dB
Noise floor 70 dBm
System loss 10 dB
Video bandwidth 1 MHz
Integration time 100 ns
Local oscillator fLO 12 MHz
Table 11.3: Channelized Receiver Parameters [7]

Number of channels 40
Noise figure 10 dB
Channel bandwidth 50 MHz
Noise floor 92 dBm
System loss 5 dB
Video bandwidth 1.25 MHz
Integration time 100 ns
receiver contains a large number of parallel narrowband receivers. The RF

band is divided into 40 contiguous front-end channels (N = 40). The channel
outputs are all folded into a common baseband and passed through 40 IF
subchannels (M = 40). The folding is done with the local oscillators at 50
MHz frequency increments. The 40 RF channels span 2 GHz and each chan-
nel is 50 MHz wide. The 40 IF subchannels each span 1.25 MHz giving a final
50 MHz coverage of the spectrum. The parameters for the channelized re-
ceiver are shown in Table 11.3.
LPI Waveforms Used

To compare each receivers detectability performance, it is important that the
emitter waveforms have equal bandwidth and equal energy. This is because
many of the compression waveforms oer the same range resolution but not
all have the same LPI/LPID properties [8]. Six waveforms were compared
including a rectangular pulse, an FMCW, P1, P2, P3 and P4 all with a
range resolution of 50m. The polyphase codes each have a subcode period of
tb = 333 ns (B = 3 MHz) and a code period of T = 64tb = 21.3 s. The
FMCW has a modulation period tm = 64tb or 21.3 s and the modulation
bandwidth is F = 3 MHz (starting at dc). The rectangular pulse has a pulse
duration R = tb = 333 ns with a pulse repetition interval PRI = 64tb = 21.3
s. The average power transmitted by each emitter is PCW = 100W and
= 3 102 m. The emitter transmit antenna gain in the direction of the
intercept receiver is Gr = 0 dB [8].
Intercept Range and Sensitivity Comparison

A MATLAB simulation was run first to determine the sensitivity. The sen-
sitivity I is the minimum signal at the back end of the receiver that is
detectable given the noise floor associated with that particular receiver. The
sensitivity was determined by iteratively adjusting the front-end signal am-
plitude that achieves the required back-end SNR. Since the noise floor, noise
figure, bandwidths and integration times are defined for each receiver, the
amplitude was adjusted until detection was achieved [8].
Models were developed in MATLAB for each receiver in order to find the
front-end signal strength necessary to satisfy the minimum back-end SNR.
A voltage gain of 20 dB was used for each receiver. After determining the
sensitivity for each receiver, the intercept range (direct path) was calculated
from (1.41) as 5
PCW Gt GI
RI max = (11.1)
4 I
where LRT = LIR = L1 = 1. Also, the propagation loss was assumed to
be zero. Each receiver uses a 12-dB detection threshold. In the case of the
channelized receiver, the detection threshold applies to each channel. By
keeping the threshold the same across all receivers, the relative detection
capability can be quantified. A 1-hour false alarm interval was assumed for
each receiver and for each channel of the channelized receiver. The sensitivity
and interception range for each intercept receiver is given in Table 11.4.
The square-law detector was most eective and detected all signals at
essentially the same range ( 25 km). The least eective receiver was the
wideband receiver. The most detectable waveform in the wideband and chan-
nelized receiver was the rectangular pulse. The P1, P2 and P4 perform signif-
icantly better than the rectangular pulse, P3 and FMCW in the channelized
receiver. The P2 waveform represents a factor of 2.3 reduction in range over
the rectangular pulse in the channelized receiver [8].
Table 11.4: Sensitivity and LPI Interception Range for Three Intercept
Receivers (After [8])
Square-Law Wideband Channel

I RI max I RI max I RI max
Waveform (dBm) (km) (dBm) (km) (dBm) (km)
Pulse 80.30 24.70 59.00 2.10 64.90 4.20
FMCW 80.57 25.49 52.96 1.06 59.98 2.38
P1 80.38 24.94 52.20 0.97 56.37 1.57
P2 80.39 24.97 52.79 1.04 56.18 1.54
P3 80.34 24.83 52.54 1.01 60.36 2.49
P4 80.36 22.88 52.54 1.01 57.14 1.72
11.3.2 Digital EW Receivers

Radio receivers that perform the analog-to-digital conversion process close to
the antenna and do most of the signal processing in the digital domain are
known as digital receivers. Digital receivers, often called software radios, place
a high performance burden on the ADC, but allow a good deal of flexibility
in postdetection signal processing.
EW receiver parameters of interest include sensitivity, dynamic range, res-
olution, simultaneous signal capability, complexity, and cost. Figure 11.4(a)
shows a block diagram of a wideband digital EW receiver. The input signal
from the antenna is first amplified by a wideband LNA. Most digital EW
receivers use frequency conversion before digitizing the signal. That is, the
signal is first downconverted in frequency, and then digitized by an ADC.
The digital signal is then processed by a spectrum analyzer that extracts the
frequency information. Using this frequency information, the signal is sorted,
and a parameter encoder then forms a pulse descriptor word (PDW). For LPI
CW emitters, the PDW contains the center frequency fc , the signal coding
details such as the modulation period and bandwidth (FMCW), the code
period and subcode period details (PSK), and frequency-hopping frequencies
(and order), as well as the signals angle of arrival.
In a network-centric architecture, the PDWs are sent to a fusion processor
that integrates other EW receiver information, in order to perform emitter
identification (e.g., using a neural network [9]), develop a situational aware-
ness, and form a corresponding response. When the receiver is used to manage
a coherent jammer, wideband/narrowband digital RF memories (DRFMs) are
also employed. To process all of the emitter information in a timely manner,
100 millions of instructions per second (MIPS) processors must be employed.
When downconverting the signal in an EW receiver, two approaches can
Figure 11.4: Block diagram of (a) a wideband digital EW intercept receiver,

(b) two-stage heterodyne down conversion process, and (c) homodyne down
conversion process.
be used, and are shown in Figure 11.4(b, c). The first (heterodyne) approach
(b) downconverts the signal, first to IF and then to baseband, using two or
more bandpass filter-local oscillator-mixer stages in series. Since the LPI
signals are phase- and frequency-modulated, both in-phase and quadrature
components are required at baseband. If the signal bandwidth is B, and I and
Q are available with each channel containing an ADC, the sampling frequency
fs > B. The advantage of this approach is that by driving the mixer with
a frequency-agile LO, the frequency of the desired signal or channel is con-
verted to a fixed frequency. Once converted to a fixed IF, it can be processed
by highly selective narrowband filtering (e.g., using surface-acoustic wave de-
vices or high-temperature superconductors). Also, all subsequent frequency
translations can be done using fixed-frequency LOs. Also performed is signal
amplification using fixed gain LNAs (at RF), and variable gain amplifiers (at
IF). The distribution of gain across the IF stage prevents instabilities in the
amplifiers, and reduces the chance of saturation.
A direct conversion (homodyne) downconversion can also be used, as
shown in Figure 11.4(c). This two-channel approach uses only a single lo-
cal oscillator, and translates the signal of interest to zero frequency (zero-IF).
Due to the elimination of the IF stages, all signal conditioning must be per-
formed either at RF or baseband. The direct conversion approach oers a
Figure 11.5: Block diagram of a digital EW receiver with an ADC at the

antenna (no downconversion).
higher degree of integration at the front end with fewer components, allowing
most of them to be monolithically fabricated on a single chip [10]. The direct
conversion receiver performance still does not match the IF receiver, due to
filter saturation and distortion caused by the dc osets and self mixing at
the mixer inputs. To take advantage of both receiver topologies, a low-IF
receiver is now an alternative (a few hundred kilohertz). The low-IF receiver
has a high degree of filter integration, and is also insensitive to dc osets and
LO-to-RF crosstalk. In all cases, the signal is downconverted to a baseband
frequency that depends on the analog-to-digital converter technology that is
available. A direct conversion receiver at Ka-band is described in [11].
11.3.3 Direct RF Sampling

The trend in EW digital receivers is to push the ADC as far towards the
antenna as possible, and to eliminate the downconversion stage, as shown in
Figure 11.5. The receiver is made up of three sections: the RF front end,
which amplifies and bandpass filters the antenna signal before it is sampled;
the ADC; and the digital signal processing. That is, the ADC is used directly
on the RF signal after appropriate preconditioning by means of amplification
and filtering.
ADC technology has improved to the point where direct sampling and
digital signal processing in the microwave spectrum is possible. Although the
development of ADCs have made considerable advancements in the last 10
years, more wideband solutions are required using electro-optics (extremely
wideband) and superconductivity (high sensitivity). Bandpass sampling does
not use any tuner or mixers to downconvert the antenna signals but instead
takes advantage of digital aliasing to down convert a Nyquist band. The ad-
vantage here is that the gain fluctuations and noise sources due to the analog
mixers and local oscillators that are used in a conventional receiver are elimi-
nated. Other advantages include a simplified hardware approach (fewer com-
ponents) that allows the integration of the receiver onto a multi-chip module
or single chip monolithic microwave integrated circuit. Also, the LPI signal
Figure 11.6: Block diagram of a reconfigurable direct RF sampling architec-

ture [13].
information is captured prior to significant analog distortion and mixer non-

linearities. Direct RF sampling also allows the receiver to be reconfigurable
to support software-defined detection and classification algorithms.
If a single fixed clock is used (partially reconfigurable), multiple bands can
be covered as long as the bands alias to the same intermediate frequency band.
Direct sampling works well for low RF signals but places severe constraints
on the ADC for higher analog input frequencies due to the eect of clock
jitter or clock uncertainty [12]. To achieve arbitrary tuning over a wide RF
range, however, the RF sample clock must be tunable or selectable because
of the problems with signal recovery on the boundaries of the Nyquist bands
generated by a fixed RF sample clock [13].
A reconfigurable direct RF sampling architecture that oers flexible tun-
ing to cover high RF bands is shown in Figure 11.6. The RF input from the
LNA is filtered by the antialiasing filter H() for the band of interest. The
filter output signal is then sampled by pulses2 at a rate of fS1 and filtered
by a continuous time interpolation filter which also serves as an antialiasing
filter for the ADC that is sampling at a rate of fS2 . Sampling is achieved
in two stages. In the first stage, the RF signal is bandpass filtered and sam-
pled using an impulse sampling device without quantization. After tunable
pulse sampling, the signal is continuous time lowpass or bandpass filtered to
generate an IF signal that is then sampled by a conventional ADC.
By using continuous time filtering after the first stage sampling, the ADC
sample clock may be completely decoupled from the RF sample clock to al-
low arbitrary tuning without impacting the ADC sample rate. That is, by
separating the sampling and quantization processes into multiple stages, the
2 Pulse sampling is a technique that can be used for direct RF sampling at much higher
frequencies than track and hold based sampling. The basic requirements for high RF pulse
sampling are narrow pulse width and low pulse amplitude jitter, in addition to the low
time jitter required in any direct RF sampling scheme [13].
jitter and clock speed requirements on the ADC can be relaxed. A high de-
gree of reconfigurability in tuning range and bandwidth is achieved by using
a tunable (or selectable) antialiasing filter before the first stage of sampling
and by using a tunable sample clock in the first stage of sampling. Extension
of this receiver architecture to an analog-to-information receiver is described
in [14]. Motivated by recent developments in compressed sensing the receiver
performs frequency modulated pulsed sampling at sub-Nyquist rates to com-
press a broadband RF environment into an analog interpolation filter and
samples the signal at the information rate rather than using the Shannon
bandwidth criteria. The receiver uses structured nonuniform sampling to im-
plement a direct analog-to-information receiver that is eective at recovering
signals that have a sparse frequency domain representation [14].
11.4 Demodulation of the LPI Waveform

LPI signals attempt to make the detection and demodulation process impos-
sible. The EW intercept receiver requires a large processing gain to detect the
LPI emission, and extract the parameters of the signal. This is followed by
the task of classification. Classification requires sorting the signal into groups
having similar parameters (clustering). Parameters such as carrier frequency,
bandwidth, modulation period, modulation bandwidth, and time of arrival
are a few of the parameters that distinguish one signal from another. Cor-
relation with existing signals in a database (identification) can then aid in
signal tracking and response management.
To identify the emitter parameters, Fourier analysis techniques using the
FFT have been used as the basic tool. From this basic tool, more com-
plex signal processing techniques have evolved, such as the short-time Fourier
transform, in order to track the signal parameters over time. More sophisti-
cated techniques have also been developed, called time-frequency and bifre-
quency distributions in order to identify the exotic modulation schemes used
by the LPI radar [15]. These techniques include the Wigner distribution,
Choi-Williams distribution, quadrature mirror filtering, and cyclostationary
processing. The use of these techniques to extract the parameters from some
of the well-known LPI modulations is the subject of the remaining chapters.
11.5 EW Receiver Challenges

The steady increase in sophistication of radar systems has resulted in an elec-
tromagnetic environment where very few pulses can now be expected. In
addition, pulse-to-pulse PRI agility and frequency agility now make it ex-
tremely dicult to identify a specific emitter, especially when only a few
pulses are intercepted (e.g., from a track-while-scan or LPI radar). LPI CW
radar, as well as digital-pulsed radar, now transmit with an enormous number

of new, complex modulations. These complex modulations can result in many
reports for a single diverse emitter, making correct signal identification dif-
ficult. Furthermore, these modern radars will essentially blind existing EW
designs, resulting in a situation where the warning receivers can no longer
handle even those radars they were designed to intercept. For example, an
existing EW might be subject to performance degradation, due to possible
interference from modern pulse-Doppler radars transmitting from friendly
platforms.
Other problem situations include the fact that many communication CW
signals are now within the radar spectrum. A significant problem can also
occur when the intercept receiver processes signals with parameters outside
its bounds. This can often cause resets that bring the system o-line for
several minutes. Resets can also occur when the receiver processes signals
that are near the internal thresholds. Finally, a significant problem for the
receiver is a failure to intercept a threat emitter that is present (including its
modulation parameters). This is more likely to happen with LPI emitters.
Todays modern EW receivers must have the ability to intercept both pulse
radar signals and CW signals within a wide bandwidth (e.g., 0.5100 GHz).
The trend is to share a common aperture, and combine the communications,
the EW, and radar functions requiring less antenna apertures.
Another serious problem for the EW receiver is the presence of ultra-
wideband sources such as spread spectrum communication signals, impulse
jammers, and impulse radar. The impact of ultrawideband synthetic aperture
radar (SAR), and inverse SAR (or ISAR) imaging radar, and high range
resolution profiling sensors, must also be considered. These sources (whether
intercepted intentionally or not) can significantly raise the noise floor of the
receiver, disabling the ability of the EW receiver to see the important threats
of interest. Consequently, the ability to reject unwanted signals is now just as
important as the ability to process the signals of interest. Wideband receivers
require adaptive notch (band reject) filters at the front end, to exclude these
unwanted signals. YIG (yttruim iron garnet) filters are often used. Adaptive
thresholds can also be used to increase sensitivity. The EW receiver must
also have high power detection and protection circuits at the front end, to
protect itself from deliberate destruction by microwave weapons and other
directed energy weapons. One high-power microwave pulse at the front end
of the intercept receiver can destroy the EW receiver function, causing total
failure of the ES/ELINT system onboard the aircraft.
The EW receiver and associated EA must also be able to provide the quick
reaction mode necessary to counter the new modern range-Doppler imaging
missiles. These missiles will use FMCW modes such as SAR, ISAR and high
range-Doppler imaging in order to improve target aimpoint accuracy, and to
reject decoys that are launched. As we discussed in Chapter 1 and demon-
strated in Chapter 9, power-managed seekers adjust the transmitter power
such that the received power at the EW receiver is kept constant (or is de-
creasing). EW receivers that prioritize incoming threats based on a change of
the received signal amplitude will be vulnerable to the power-managed LPI
seekers. These power-managed emitters must also be detected and processed
correctly. Advanced missiles will also use dual mode seekers (e.g., an antira-
diation missile seeker combined with an active millimeter wave LPI seeker)
that must also be identified and countered.
Finally, the EW receiver must be able to disseminate all onboard detec-
tions in real time sometimes referred to as real time out of the cockpit (RTOC).
RTOC data is critical in a network-centric architecture in order to provide
multiplatform targeting and geolocation. Also the ability to accept real time
data in the cockpit (RTIC), and utilize the oboard (multispectral) sensor
data, is an important capability in order for the platform to precisely target
its weapons or electronic attack. In this way the intercept receiver can be
a major player in a multiplatform time dierence of arrival (TDOA)-based
geolocation network with, for example, Rivet Joint. The data fusion also
exploits any oboard and multispectral signals intelligence (SIGINT) data
received.
Specific emitter identification (SEI) attempts to fingerprint the emitters
that are intercepted. SEI can also be used for improved tracking and dein-
terleaving. A number of algorithms have been investigated for doing SEI,
but their details remain classified. More importantly, however, is the fact
that future SEI systems must be standardized for interoperability between
platforms and organizations (especially since the dividing line between RWR
and ELINT is fast going away). That is, the success in SEI will ultimately
lie in the infrastructure (dissemination of databases, correlation of absolutes,
organized collection of targets), and not so much in the algorithms that are
used.

The trend in intercept receivers is toward digital receivers and the concept
of digital antennas (ADC at the antenna). The future digital receiver will
incorporate optical technologies for speed and bandwidth, and will also in-
corporate high-temperature superconductors for sensitivity. Networking the
EW receiver within an information, sensor, and shooter grid will allow the
sharing of intercept data. The use of swarm architectures will also become
more prevalent. In the following chapters we assume that the signal is digi-
tized in the receiver, and we mainly focus on the signal processing methods
used to extract the LPI waveform parameters to classify the signals.
References
[1] Schrick, G., and Wiley, R. G., Interception of LPI radar signals, IEEE
[2] Wiley, R. G., Electronic Intelligence: The Interception of Radar Signals,
Artech House Publishers, Dedham, MA, 1985.
[3] Lee, J. P. Y., Interception of LPI radar signals, Defence Research Estab-
lishment Ottawa, Technical Note 91-23, Nov. 1991.
[4] Bonabeau, E., Dorigo, M., and Theraulaz, G., Swarm Intelligence From Nat-
ural to Artificial Systems, Oxford University Press, New York, 1999.
[5] Tsui, J. B. Y., and Stephens, J. P. Sr., Digital microwave receiver technol-
ogy, IEEE Trans. on Microwave Theory and Techniques Vol. 50, No. 3, pp.
699705, March 2002.
[6] Rodrigue, S. M., Bash, J. L., and Haenni, M. G., Next generation broadband
digital receiver technology, The 15th Annual AESS/IEEE Symposium, pp.
1320, 1415 May 2002.
[7] Gross, F. B., and Chen, K., Comparison of detectability of traditional pulsed
and spread spectrum radar waveforms in classic passive receivers, IEEE
Trans. on Aerospace and Electronic Systems, Vol. 41, No. 2, pp. 746751,
April 2005.
[8] Gross, F. B., and Connor, J., Comparison of detectability of radar compres-
sion waveforms in classic passive receivers, IEEE Trans. on Aerospace and
Electronic Systems, Voltt . 43, No. 2, pp. 789795, April, 2007.
[9] Shieh, C-S, and Lin, C-T., A vector neural network for emitter identifi-
cation, IEEE Trans. on Antennas and Propagation, Vol. 50, No. 8, pp.
11201127, Aug. 2002.
[10] Pekau, H., and Haslett, J. W., A comparison of analog front end architec-
tures for digital receivers, Proc. of the IEEE CCECE/CCGEI, Saskatoon,
May 2005.
[11] Tatu, S. O., et al., Ka-band direct digital receiver, IEEE Trans. on Mi-
crowave Theory and Techniques, Vol. 50, No. 11, pp. 24362442, Nov. 2002.
[12] Chalvatzis, T., Gagnon, E., and Wight, J. S., On the eect of clock jitter
in IF and RF direct sampling systems, 3rd International IEEE NEWCAS
Conference, pp. 6366, 1922 June 2005.
[13] Fudge, G. L., Chivers, M. A., Ravindran, S., Bland, R. E., and Pace, P.
E., A reconfigurable direct RF receiver architecture, Proc. of the IEEE
International Symposium on Circuits and Systems, May 2008.
[14] Fudge, G. L., Bland, R. E., Chivers, M. A., Ravindran, S., Haupt, J. and
Pace, P. E., A Nyquist folding analog-to-information receiver, Proc. of the
Asilomar Conf. on Signals, Computers and Signal Processing, Nov. 2008.
[15] Stephens, J. P., Advances in signal processing for electronic warfare, IEEE
Aerospace and Electronic Systems Magazine, pp. 3138, Nov. 1996.
Chapter 12
Wigner-Ville Distribution
Analysis of LPI Radar
Waveforms
In Chapter 11, it was shown that tomorrows digital intercept receiver must
incorporate a time-frequency analysis capability in order to identify the LPI
modulation types and also extract the LPI signals parametric data. The
Wigner-Ville Distribution (WVD), introduced by Wigner in 1932 as a phase
representation in quantum statistical mechanics [1] and separately by Ville
in 1948 addressing the question of a joint distribution function [2], simul-
taneously gives the representation of a signal in both time and frequency
variables. The WVD has been noted as one of the more useful bilinear time-
frequency analysis techniques for signal processing. In this chapter, the WVD
is presented and used to analyze the signals discussed in Part I. Extraction
of the signal parameters is also emphasized. The main objective is that by
studying the results and correlating the signal parameters that are revealed,
the user can learn to determine the presence of a particular LPI signal and
to recognize the LPI modulation characteristics under various signal-to-noise
ratios. We also show how well we can distinguish among several waveforms
that have similar time and frequency characteristics. Multiple signal analysis
is left as an exercise for the reader. By using the Wigner analysis tools, an
intercept receiver can come close to having a processing gain near the LPI
radars matched filter processing gain. The WIGNER folder on the CD pro-
vides the MATLAB tools that can be used to re-create any of the figures
presented, as well as generate new and useful results.
405
12.1 Wigner-Ville Distribution

The WVD has been used in many fields of engineering. These include optical
implementations of the WVD [3], medical applications [46], image analysis
[7, 8], target detection [9, 10], and the analysis of nonstationary (LPI) signals
[1114].
The WVD exhibits the highest signal energy concentration in the time-
frequency plane for linearly modulated signals, but has drawbacks in the
case of nonlinear frequency modulated signals. To improve the concentration
where nonlinear modulations are present, various higher-order time frequency
representations have been investigated [15]. The WVD also contains inter-
fering cross terms (or ghost terms) between every pair of signal components.
As illustrated in the examples that follow, the presence of the cross terms
sometimes make it dicult to determine the LPI modulation parameters.
A good review of bilinear transforms and their use in signal analysis is
given in [16]. The influence that the cross term interference has on the WVD
is analyzed in [17, 18]. The extension of the WVD to discrete time signals
has been discussed in [19, 20] and a formulation to remove the cross terms
has been reported in [21, 22]. Below, we begin with the definition of the
WVD, and then present a windowed version of the WVD, the pseudo WVD
(PWVD) which is useful in the signal processing of the digital signals within
the receiver.
12.1.1 Continuous WVD

The WVD of a continuous one-dimensional function (or input signal) x(t) is
given by [23]

j
Wx (t, ) = x t+ x t e d (12.1)
2 2
where t is the time variable, is the angular frequency variable (2f ), and the
indicates a complex conjugate. The WVD is a three-dimensional function

describing the amplitude of the signal as a function of time and frequency.
Since the LPI emitter modulations vary the compression of the CW waveform
as a function of time, these types of time-frequency distributions give a higher
probability of detecting the modulation parameters. The WVD can also be
defined from the Fourier transform X() of x(t) by

1 0 0 j0 t
WX (, t) = X + X e d0 (12.2)
2 2 2
From (12.1) and (12.2), the following relation is obtained:
Wx (t, ) = WX (, t) (12.3)
Wigner-Ville Distribution Analysis of LPI Radar Waveforms 407
That is, the WVD of the spectra of a signal can be determined simply from
that of the time functions by an interchange of the frequency and time vari-
ables. This shows the symmetry between space and frequency domain defin-
itions [23].
Equation (12.1) implies that evaluation of the WVD is a noncausal oper-
ation. As such, this expression does not lend itself to real-time evaluation.
This limitation is overcome by first applying the WVD analysis to a sampled
time series x( ), where is a discrete time index from to . The discrete
WVD is defined as

W ( , ) = 2 x( + n)x ( n)ej2n (12.4)
n=
Windowing the data results in the pseudo-WVD and is defined by [17]

N 1
W ( , ) = 2 x( + n)x ( n)w(n)w(n)ej2n (12.5)
n=N+1
where w(n) is a length 2N 1 real window function with w(0) = 1. Using

f (n) to represent the kernel function
f (n) = x( + n)x ( n)w(n)w(n) (12.6)
the PWVD becomes

N 1
W ( , ) = 2 f (n)ej2n (12.7)
n=N+1
The choice of N (usually a power of 2) greatly aects the computational

cost, as well as the time-frequency resolution, of the PWVD output. A large
N gives a higher time-frequency resolution since it influences the frequency
resolution in (12.7). When the continuous variable in (12.7) is sampled
to produce a suitable form of the discrete Fourier transform (DFT), a larger
N also gives more output samples, yielding a smoother result [24]. The
maximum value of N is limited by
M +1
N (12.8)
2
where M is the data length.
Once N is chosen, the kernel function can be generated. Since
f (n) = f (n) (12.9)
only f (n) needs to be computed for n 0. A block diagram of the PWVD

kernel generation for N = 8 is shown in Figure 12.1, where v(n) = w(n)w(n).
Figure 12.1: The computational structure for an N = 8 PWVD kernel gen-

eration [24] ( c 1989 IEEE).
Here the input signal enters the buer register from the left and shifts to the
right after each kernel generation. The right-most element is disposed after
the next shift.
The PWVD can detect the presence of LPI signals, as well as extract the
signals modulation characteristics. For an intercept receiver, it is important
that the computation be done in real time or near real time. From the
PWVD expression in (12.7), we notice that it is computationally expensive to
directly compute the PWVD. Boashash et al. [25] have presented an ecient
algorithm to compute the discrete PWVD. The algorithm is presented below.
To begin, the continuous frequency variable is sampled as
k
= (12.10)
2N
where k = 0, 1, 2, , 2N 1 (2N samples). The kernel indexes are modified
to fit the standard DFT:
N1
k j2nk
W , =2 f (n) exp (12.11)
2N 2N
n=N+1
or
2N1
k j2nk
W , =2 f (n) exp (12.12)
2N n=0
2N
where

f (n), 0nN 1
f (n) = 0, n=N (12.13)

f (n 2N ), N + 1 n 2N 1
Since the kernel is a symmetric function, the DFT of the kernel is always real.
The resulting PWVD using 2N samples is
2N 1
jkn
W ( , k) = 2 f (n) exp (12.14)
n=0
N
Equation (12.14) is the algorithm implemented, and several examples are

shown in the next section to illustrate the properties of the computation.
12.1.2 Example Calculation: Real Input Signal

Consider an example using a real input signal
x( ) = {2, 4, 3, 6, 1, 7} (12.15)
where N = 3 and the length of the input signal x( ) is 2N = 6. Here
= 3, 2, 1, 0, 1, 2 and is the discrete time index in the range N to
N 1. Note that x = 0 for 4 or 3. From (12.13), with N = 3,

f (n), 0n2
f (n) = 0, n=3 (12.16)

f (n 6), 4n5
From (12.6) f3 (n) ( = 3), for input signal x( ) is computed as follows:
f3 (n = 0) = x(3) x (3) = 2 2 = 4
f3 (n = 1) = x(2) x (4) = 4 0 = 0
f3 (n = 2) = x(1) x (5) = 3 0 = 0
f3 (n = 3) = 0
f3 (n = 4) = x(1) x (7) = 1 0 = 0
f3 (n = 5) = x(2) x (8) = 7 0 = 0
So, from 12.13, f3 = {4, 0, 0, 0, 0, 0}. Similarly for f0 , ( = 0),
f0 (n = 0) = x(0) x (0) = 6 6 = 36
f0 (n = 1) = x(1) x (1) = 1 3 = 3
f0 (n = 2) = x(2) x (2) = 7 4 = 28
f0 (n = 3) = 0
f0 (n = 4) = x(2) x (2) = 4 7 = 28
f0 (n = 5) = x(1) x (1) = 3 1 = 3
Figure 12.2: The kernel f (n) matrix for the real six input example.
and so, f0 = {36, 3, 28, 0, 28, 3}. Repeating the above procedure, the kernel
matrix for all values = 4 to 3, and n = 0 to 5 is as shown in Figure 12.2.
The second step after the kernel transformation is to use (12.14) to calcu-
late the Wigner distribution. As an example of the calculation, one can pick
any and k to examine the values inside the PWVD matrix. For example,
choose = 1, k = 2, with N = 3. The PWVD is
2N1
kn
W ( = 1, k = 2) = 2 f (n) exp j
n=0
N
231
2n
= 2 f (n) exp j
n=0
3
5
2n
= 2 f (n) exp j (12.17)
n=0
3
From the kernel matrix in Figure 12.2, the kernel function for = 1 is f1 (n) =
{1, 42, 0, 0, 0, 42}. So from (12.17), the PWVD for = 1, k = 2 (6 terms) is
20 21
W (1, 2) = 2f1 (0) exp j + 2f1 (1) exp j
3 3
22 23
+ 2f1 (2) exp j + 2f1 (3) exp j
3 3
Figure 12.3: The PWVD matrix W ( , k) for the real six input example.
24 25
+ 2f1 (4) exp j + 2f1 (5) exp j
3 3
= 2 1 (0) + 2 42 (0.5000 0.8660i)
+ 2 0 + 2 0 + 2 0 + 2 42 (0.5000 + 0.8660i)
W (1, 2) = 82
Repeating the above procedure gives the PWVD matrix at each discrete time
index = 4 to 3 for each discrete frequency index k = 0 to 5. The result
is a symmetric matrix about k = 3, as shown in Figure 12.3. An important
feature of the PWVD is that all the components in the matrix are real. Other
important properties of the PWVD are given in [3, 4, 7].
12.1.3 Example Calculation: Complex Input Signal

To demonstrate the PWVD computation for a complex input, consider the
signal
x = I + jQ (12.18)
where
I = cos(2fc t) (12.19)
Q = sin(2fc t) (12.20)
If the carrier frequency fc = 1 kHz, sampling frequency, fs = 7 kHz, and

t {0, 1/fs , 2/fs , . . . , 7/fs }, then the first eight input points for the discrete
Figure 12.4: The kernel matrix for the complex eight input example.
time index = 4 to 3 is
x( ) = {1 + 0i, 0.62 + 0.78i, 0.22 + 0.97i, 0.90 + 0.43i,
0.90 0.43i, 0.22 0.97i, 0.62 0.78i, 1 + 0i} (12.21)
Consider the value when = 0, n = 3. Using (12.13) with an input length
2N = 8 or N = 4. The kernel is

f (n), 0n3
f = 0, n=4 (12.22)

f (n 8), 5n7
or
f (n) = {f (1), f (2), f (3), 0, f (3), f (2), f (1)} (12.23)
Since f (n) = x( + n) x ( n), the kernel at = 0, n = 3 is f0 (3) =
x(3) x (3) = 1 (0.6235 + 0.7818i) = 0.6235 0.7818i. Repeating the
same procedures as discussed in the real input case, the kernel matrix for
the complex eight input example is shown in Figure 12.4. Referring to
Figure 12.4, we can calculate the PWVD when = 1. The kernel is
f1 (n) = {1.00, 0.22 + 0.97i, 0.90 0.43i, 0.62 0.78i,
0, 0.62 + 0.78, 0.90 + 0.43i, 0.22 0.97i} (12.24)
Figure 12.5: The PWVD matrix for the complex eight input example.
Consider the case when = 1, k = 4. From (12.14), the PWVD for N = 4

is
2N 1
kn
W ( = 1, k = 4) = 2 f (n) exp j
n=0
N
241
4n
= 2 f1 (n)
n=0
4
7
= 2 f1 (n) exp(jn) (12.25)
n=0
From (12.24) and (12.25)

7
W ( = 1, k = 4) = 2 f1 (n) exp(jn) = 3.2073
n=0
Again, the PWVD matrix of the complex eight input samples is real. The
complete PWVD matrix is a symmetric 2N 2N matrix. Figure 12.5 shows
the PWVD matrix of the complex eight input samples. Note this important
feature: the PWVD is always real whether the input signal is real or complex.
Figure 12.6(a) shows a 3D mesh plot of the PWVD for the complex sig-
nal example with eight inputs. This plot shows the magnitude in both the
time domain and the frequency domain. Note that it directly correlates with
Figure 12.5. The peak corresponds to the 1-kHz carrier frequency.
Figure 12.6(b) shows the corresponding PWVD contour plot. The contour
plot is a 2D time-frequency plot that is useful for characterizing the time-
frequency behavior of the signal. The magnitude is represented by a dierent
gray scale, as shown in the legend bar.
To see the marginal details of the PWVD, Figure 12.7(a) shows a plot of
the PWVD obtained by rotating the mesh plot in Figure 12.6(a) to show
just the time axis with the eight samples. The time resolution is 1/fs .
Figure 12.7(b) shows the marginal details in the frequency domain, and is
obtained in the same manner as Figure 12.7(a). The carrier frequency is rep-
resented by the peak in this plot, and shows up at 900 Hz, very close to the
real value 1 kHz. The frequency resolution fs /2/# samples is also indicated.
In summary both the real signal example and the complex signal exam-
ple illustrate the mechanics of the PWVD calculation. The PWVD time-
frequency results, when presented in the four dierent plots, give a variety
of aspects so that the LPI signal and its modulation characteristics can be
determined.
12.1.4 Two-Tone Input Signal Results

Now we consider the PWVD for a two-tone input (two carrier frequencies)
with fc1 = 1 kHz and fc2 = 2 kHz. Now I = cos(2fc1 t) + cos(2fc2 t) and
Q = sin(2fc1 t) + sin(2fc2 t).
Figure 12.8(a) shows the PWVD results for the two-tone signal in a 3D
time-frequency mesh plot. In this plot the cross terms are stronger than the
signal terms, and show up with many peaks. Figure 12.8(b) is the 2D PWVD
time-frequency contour plot and shows the time dependence of the real signal
and the cross terms.
Figure 12.9(a) shows the marginal time domain plot obtained by rotating
the 3D mesh plot in Figure 12.8(a), to show only the time axis. This reveals
the cross terms as a series of positive and negative magnitude components in
the time domain. Figure 12.9(b) shows the frequency domain plot obtained
in the same manner, and reveals the two-carrier frequencies and the cross
term. Note that the shape and magnitude of the cross term is not like the
two-carrier frequency components, and can be easily identified.
Figure 12.6: PWVD for the eight input complex example: (a) 3D mesh plot,
and (b) time-frequency domain.
Figure 12.7: PWVD for the eight input complex example: (a) 2D mesh in
time domain, and (b) 2D mesh in frequency domain.
Figure 12.8: PWVD for the two-tone example, showing the (a) 3D time-
frequency domain mesh plot, and (b) 2D time-frequency contour.
Figure 12.9: PWVD for the two-tone example, showing the (a) marginal time
domain plot, and (b) marginal frequency domain plot.
Figure 12.10: Diagram of a triangular FMCW waveform.
12.2 FMCW Analysis

In this section, extraction of the signal parameters for two FMCW waveform
examples (see Chapter 4) are investigated. When measuring the parameters
of the LPI modulations, in all cases the absolute value of the relative error
should be reported. That is, if a is a measurement value of a quantity whose
exact value is a, then the absolute value of the relative error r is defined by
a a Error
r = = (12.26)
a True value
The time-frequency diagram of a triangular waveform is shown in

Figure 12.10. For the examples, both signals are sampled by the ADC at
a rate of fs = 7,000 Hz. The first signal examined is an FMCW wave-
form with a center frequency of fc = 1,000 Hz, a modulation bandwidth of
F = 250 Hz, and a modulation period of tm = 20 ms. Figure 12.11(a)
shows the PWVD frequency plot of the FMCW waveform. This graph can
also be compared with the PSD of the waveform (see Chapter 4). Note the
presence of the additional structure due to the triangular modulation. The
carrier frequency can easily be identified and measured. The bandwidth of
the signal can also be estimated. Figure 12.11(b) shows the time-frequency
distribution. Although cross terms are present in the output image, the
modulation bandwidth and the modulation period can be accurately iden-
tified, as well as the carrier frequency. Figure 12.12(a) shows the PWVD
image for SNR = 0 dB and Figure 12.12(b) shows the PWVD image for
SNR = 6 dB. The SNR is defined over fs /2. It is interesting to note that
the carrier frequency, the modulation period, and the modulation bandwidth
can all be extracted, even with this amount of interference present in the
Figure 12.11: PWVD for an FMCW with F = 250 Hz, tm = 20 ms (sig-

nal only), showing the (a) marginal frequency domain plot, and (b) time-
frequency plot.
signal. Also note that this information is not available if only the PSD is cal-
culated in the receiver. In summary, the PWVD technique works extremely
well for FMCW waveforms. The results for the F = 500-Hz signal are
shown in Appendix E.
12.3 BPSK Analysis

In this section we apply the PWVD and use it to investigate the properties
of the BPSK signal discussed in Chapter 5. Parameters varied include the
length of the Barker code (Nc = 7 or 11), the number of carrier cycles per
Barker subcode (cpp), and the SNR (signal only, 0 dB, and 6 dB). We also
investigate how many parameters of the signal can be extracted from the
PWVD results. All signals demonstrating the concepts have an fc = 1 kHz
carrier frequency, and a sampling frequency of fs = 7 kHz. Both frequency
domain and time domain plots are shown for the BPSK signals after the
PWVD processing. The time-frequency domain results are the most useful.
The first signal, examined in Figure 12.13(a), has a carrier frequency of
1 kHz and can be clearly identified by the location of the highest or lowest
peak value. The carrier frequency can also be identified as the center of the
symmetric frequency distribution in Figure 12.13(b). Also, the relative peak
magnitude in Figure 12.13(a) is about 600, so the 3-dB bandwidth (or subcode
rate) B, is the frequency range lying at 300 on both sides, which extends from
500 Hz to 1,500 Hz. Since B = fc /cpp = 1,000 Hz, this correlates well. In
Figure 12.13(a), if we look closely within the 3-dB bandwidth, one can find
that there are 15 peaks within the bandwidth. That is, there are 14 intervals
in the range from 500 Hz to 1,500 Hz. This number is always two times the
Barker code length. In Figure 12.14, the 7-bit Barker code is examined with
an SNR = 0 dB. Figure 12.14(a) shows the frequency domain where the 15
intervals can be counted within the 3-dB bandwidth. Figure 12.14(b) shows
the corresponding time-frequency domain. In this particular case, not much
information is revealed.
Figure 12.15(a) shows the results for an 11-bit signal. In this case there
are 23 peaks within the 3-dB bandwidth (Barker code length Nc = 11).
Figure 12.15(b) shows the time-frequency results centered about the carrier
frequency of fc = 1,000 Hz.
Figure 12.12: PWVD for an FMCW with F = 250 Hz, tm = 20-ms time-
frequency plot for (a) SNR = 0 dB and (b) SNR = 6 dB.
Figure 12.13: PWVD for BPSK with 7-bit Barker code, cpp = 1, signal only,
showing the (a) marginal frequency domain plot and (b) time-frequency plot.
Figure 12.14: PWVD for BPSK with 7-bit Barker code, cpp = 1, SNR =
0 dB, showing the (a) marginal frequency domain plot and (b) time-frequency
plot.
Figure 12.15: PWVD for BPSK with 11-bit Barker code, cpp = 1 (signal
only), showing the (a) marginal frequency domain plot and (b) time-frequency
plot.
12.4 Polyphase Code Analysis

In this section we begin with two CW signals that are phase coded with
a Frank code (see Chapter 5), and examine them using the PWVD. Both
signals are sampled in the receiver by an ADC with a sampling frequency
fs = 7,000 Hz. The first signal examined has a carrier frequency of fc =
1,000 Hz, 16 phase codes Nc = 16, (M = 4), and a cpp = 1 or one cycle
per subcode. That is, each subcode has a length of tb = 1 ms, resulting in a
phase-coded signal with a code period of
Nc Nc cpp
T = Nc tb = = = 16 ms (12.27)
B fc
Using cpp = 1 results in the maximum bandwidth that can be achieved with
any particular carrier frequency.
Identifying the signal parameters within the PWVD image is considered
again. Figure 12.16(a) shows the PWVD frequency plot and Figure 12.16(b)
shows the PWVD time-frequency image. Note that the carrier frequency
can be identified by the largest peak value. Also note the presence of the
harmonics that appear as modulation spikes every n/T Hz (or 62.5n Hz).
The Frank code shows up as a series of unique evenly spaced parallel lines.
The bandwidth B can also be identified in the image. When measuring
the Frank code B within the PWVD image, it is necessary to skip one of the
lines due to the presence of the cross terms. The slope of each line has a
magnitude of
B Bfc fc2
S= = = = 62,500 Hz s1 (12.28)
T Nc cpp Nc cpp2
The code period T is measured through the major cross term, and is also
illustrated. This measurement stresses the fact that the PWVD integration
period must be at least larger than the signals code period, in order to
provide an accurate estimate (# samples/fs > T ).
Figure 12.17(a) shows the PWVD time-frequency image for SNR = 0 dB.
Although the addition of noise is clearly present in the output, the phase
code parameters can be easily determined. In Figure 12.17(b) the SNR = 6
dB and it becomes a bit more dicult. Preprocessing the image with a
lowpass filter can help reduce the presence of the high frequency noise that
hinders the extraction of the important signal parameters. A second Frank
signal example with M = 8 (Nc = 64) is given in Appendix F. The PWVD
results for the P1, P2, P3, and P4 polyphase codes are similar and given in
Appendix G.
Figure 12.16: PWVD for Frank code with B = 1,000 Hz, T = 16 ms, signal
only, showing the (a) marginal frequency domain plot, and (b) time-frequency
plot.
Figure 12.17: PWVD for Frank code with B = 1,000 Hz, T = 16 ms, time-
frequency plot for (a) SNR = 0 dB, and (b) SNR = 6 dB.
12.5 Polytime Code Analysis

In this section an analysis of the polytime codes (see Chapter 5) using the
PWVD is presented. The structure of the polytime codes (T1 through T4)
within the PWVD are significantly dierent, even though they were derived
from both step frequency and linear FM waveforms (as are the Frank code
and P1P4).
The first signal examined is the T1 code with fc = 1,000 Hz and T = 16
ms. The number of stepped frequency segments used is k = 4 (zero beat
at the leading segment), and the number of phase states n = 2, T1(2).
The sampling frequency of the ADC is fs = 7,000 Hz. Each segment is
4 ms in duration, resulting in the overall code period of 16 ms. The fre-
quency step between adjacent segments is 1/4 ms = 250 Hz with a total fre-
quency excursion of 1,000 Hz. Figure 12.18(a) shows the PWVD frequency
domain. Compared to the Frank and P1 through P4 codes, the T1(2) energy
is more evenly distributed within the (approximately) same bandwidth. Also
note that the harmonics are not uniformly spaced, due to the time modula-
tion of the binary phase change. Figure 12.18(b) shows the time-frequency
distribution of the T1(2) code. The signal shows up as a set of vertical roof
tops stacked next to each other, separated by T /2. Note that the carrier is
easily identified, as well as the measurement of the bandwidth B and code
period T (about the cross term).
To understand the bandwidth characteristics shown in Figure 12.18, the
phase shift for the T1(2) is shown in Figure 12.19. Here the smallest phase
change shown is 4 samples long (0.571 ms). This results in a bandwidth
excursion of B = 1,750 Hz, which can be identified in Figure 12.18(b). Figure
12.20(a) shows the signal for an SNR = 0 dB. The signal can still be identified
as a T1(2) and the parameters can still be extracted. In Figure 12.20(b) with
an SNR = 6 dB, no signal identification can be made and no parameters
can be extracted. The T2(2), T3(2), and T4(2) code examples are examined
in Appendix H.
Figure 12.18: PWVD for polytime code T1(2) with B = 1,750 Hz, T =
16 ms, signal only showing the (a) marginal frequency domain plot, and
(b) time-frequency plot.
Figure 12.19: T1(2) phase shift showing a minimum subcode width of four
samples (0.571 ms), resulting in a bandwidth excursion of B = 1,750 Hz.
12.6 Distinguishing Between Phase Codes

The main objective of the previous sections was to examine the PWVD for
each of the important LPI phase modulations discussed in Part I. The inter-
cept receiver running the PWVD must also be able to distinguish between
these phase modulations, in addition to extracting the signal parameters (as
described above).
To illustrate the similarities and dierences, the phase modulations are
compared together in Figure 12.21. The Frank code, P1, P3, and P4 have the
same slope sign but, although similar, have dierent time-frequency charac-
teristics that can be used to identify the particular phase modulation. The
P2 has a dierent slope. The distinguishing features, of course, depend on
the sampling period of the ADC, and any receiver nonlinearities that might
be present.
The T1(2) (as well as T2T4) are unique, since they contain time-frequency
components with both slope signs. Figures 12.2212.25 show the PWVD
time-frequency results, and the corresponding phase states for the T1 through
T4, for comparison. Examination of these results shows that it is easy to dis-
tinguish between the polytime codes, and also to distinguish them from the
Frank, P1, through P4, codes.
Figure 12.20: PWVD for T1(2) code with B = 1,750 Hz, T = 16 ms, time-
frequency plot for (a) SNR = 0 dB, and (b) SNR = 6 dB.
Figure 12.21: PWVD for (a) Frank code, (b) P1, (c) P2, (d) P3, and (e) P4
codes.
Figure 12.22: PWVD for (a) T1(2) code, and (b) phase code showing mini-
mum subcode with 18 samples (2.57 ms).
mum subcode with 36 samples (5.14 ms).
mum subcode with six samples (0.86 ms).
mum subcode with seven samples (1 ms).
12.7 FSK and FSK/PSK Analysis

In this section we examine the PWVD results for an FSK Costas signal and
a hybrid PSK/FSK signal. The PWVD is used first to investigate the time-
frequency characteristics of a Costas sequence {3, 2, 6, 4, 5, 1} kHz. The
signal was generated with the costas.m program within the LPIT, using a
frequency duration of tp = 0.005s and a sampling frequency of fs = 15,057
Hz. This resulted in single frequency-hopping sequence consisting of 520 sam-
ples. The PWVD marginal frequency domain is shown in Figure 12.26(a).
Note that the frequencies in the sequence are clearly present, as well as the
cross terms. The PWVD time-frequency image is shown in Figure 12.26. The
arrows indicate the positions of the Costas frequencies. Note that the time
axis is reversed. That is, the frequency order begins at the right side of the
figure. The cross terms present tend to make the identification of the frequen-
cies intricate, especially when the cross terms lie about one of the frequencies
in the code. In Figure 12.27(a, b), the time-frequency image is displayed for
both 0 dB and 6 dB, respectively. As expected, the identification becomes
more dicult.
With the FSK (Costas)/PSK signal, each frequency selected is phase
shifted with a 5-bit Barker code with cpp = 5 (five cycles per phase code).
The results are shown in Figure 12.28(a, b) for the signal only, and SNR = 0
dB, respectively. Note that although the cross terms are again significant, the
frequencies in the Costas code can be identified. For the FSK/PSK target
signal described, the PWVD does not give good results, and no parameters
can be determined.
12.8 Summary
The PWVD theory was presented in this chapter and several examples were
used to demonstrate generating an ecient kernel function and the subsequent
calculation of the PWVD time-frequency results. Whether the signals are real
or complex, the kernel and PWVD matrix are always real and symmetric.
This is an important feature for the Wigner distribution and a good reason
why the PWVD can be used for accurate signal analysis (in spite of the cross
terms present).
Apart from the parameters listed in the table, other signal characteristics
can be measured or estimated. For example, for the FMCW signal, knowing
F and tm , the range resolution R and the unambiguous range Ru may be
estimated. For the FSK Costas code, the identification of the signal is dicult
due to the ghost terms present. This is also the case for the FSK/PSK (binary
phase code) signal. For the FSK/PSK (target) signal reported, the PWVD
was not able to identify any meaningful signal parameters. This should not
be a surprise, considering the PACF, PAF results shown in Chapter 6.
Figure 12.26: PWVD for Costas code sequence {3, 2, 6, 4, 5, 1} kHz, showing
the (a) marginal frequency domain, and (b) time-frequency image.
Figure 12.27: PWVD time-frequency image for Costas code sequence {3, 2,
6, 4, 5, 1} kHz, showing (a) SNR = 0 dB, and (b) SNR = 6 dB.
Figure 12.28: The PWVD for the FSK/PSK signal, showing the (a) signal
only, and (b) SNR = 0 dB.
The presentation of the PWVD results to a trained operator will allow the
signal parameters to be extracted, and can enable good classification results
in moderate amounts of noise. A modern intercept receiver/analyzer would
implement a set of parallel processors; each designed to recognize, within a
particular frequency band, a particular class of waveforms that might occupy
that band. The outputs would consist of pulse descriptor words containing
estimates of the signal parameters. How well the PWVD processing performs
this task, as a function of its bandwidth (relative to the actual signal band-
width) and the SNR in that bandwidth, is of primary concern and must be
investigated. Autonomous classification and parameter extraction within an
intercept receiver is a significantly harder problem, and is addressed further
in Chapters 17 and 18.
References
[1] Wigner, E. P., On the quantum correction for thermodynamic equilibrium,
Physics Review, Vol. 40, pp. 749759, 1932.
[2] Ville, J., Theorie et applications de la notion de signal analytique, Cables
et Transmission, Vol 2A, pp. 6174, 1948.
[3] Li, Y., Eichmann G., and Conner, M., Optical Wigner distribution and
ambiguity function for complex signals and images, Optics Communication,
Vol. 67, No. 3, pp. 177179, July 1988.
[4] Clayton, R. H., and Murray, A., Comparison of techniques for time-frequency
analysis of the ECG during human ventricular fibrillation, IEE Proc. on Sci-
ence and Measurement Technology, Vol. 145, No. 6, pp. 301306, Nov. 1998.
[5] Darvish, N., and Kitney, R. I., Time-frequency and time-scale methods in the
detection and classification of non-stationarities in human physiological data,
Record of the 28th Asilomar Conference on Signals, Systems and Computers,
Vol. 2, pp. 10851158, Oct. 31Nov. 2, 1994.
[6] Millet-Roig, J., et al., Time-frequency analysis of a single ECG: to discrimi-
nate between ventricular tachycardia and ventricular fibrillation, Computers
in Cardiology, pp. 711714, 1999.
[7] Cristobal, G., Bescos J., and Santamaria, J., Image analysis through the
Wigner distribution, Applied Optics, Vol. 28, No. 2, pp. 262271, Jan.
1989.
[8] Gonzalo, C., et al., Space-variant filtering through the Wigner distribution
function, Applied Optics, Vol. 28, No. 4, pp. 730735, Feb. 1989.
[9] Haykin, S., and Bhattacharya, T., Wigner-Ville distribution: an important
functional block for radar target detection in clutter, Record of the 28th
Asilomar Conference on Signals, Systems and Computers, Vol. 1, pp. 6872,
Oct. 31Nov. 2, 1994.
[10] Kumar, P. K., and Prabhu, K. M. M., Simulation studies of moving target-
detection: a new approach with the Wigner-Ville distribution, IEE Proc. on
Radar, Sonar and Navigation, Vol. 144, No. 5, pp. 259265, Oct. 1997.
[11] Milne, P.R., and Pace, P. E., Wigner distribution detection and analysis of
FMCW and P-4 polyphase LPI waveforms, IEEE International Conference
on Acoustics, Speech, and Signal Processing, Vol. 4, pp. 39443947, 2002.
[12] Taboada, F., et al., Intercept receiver signal processing techniques to de-
tect low probability of intercept radar signals, Proc.of the Fifth Nordic Sig-
nal Processing Symposium, Hurtigruta Tromso-Trondheim, Norway, 47 Oct.
2002.
[13] Barbarossa, S., Parameter estimation of undersampled signals by Wigner-
Ville analysis, IEEE International Conference on Acoustics, Speech, and
Signal Processing, ICASSP-91, Vol. 5, pp. 32533256, April 14-17, 1991.
[14] Gau, J-Y, Analysis of low probability of intercept (LPI) radar signals using
the Wigner Distribution, Naval Postgraduate School Masters Thesis, Sept.
2002.
[15] Katkovnik, V., and Stankovic, L., Instantaneous frequency estimation using
the Wigner distribution with varying and data-driven window length, IEEE
Trans. on Signal Processing, Vol. 46, No. 9, pp. 23152325, Sept. 1998.
[16] Chen, V. C., and Ling, H., Time-Frequency Transforms for Radar Imaging
and Signal Analysis, Artech House, Inc., Norwood, MA, 2002.
[17] Stankovic, L., and Stankovic, S., On the Wigner distribution of discrete time
noisy signals with application to the study of quantization eects, IEEE
Trans. on Signal Processing, Vol. 42, No. 7, pp. 18631867, July 1994.
[18] Stankovic, L., Algorithm for the Wigner distribution of noisy signals reali-
sation, IEE Electronics Letters, Vol. 34, No. 7, pp. 622623, April 1998.
[19] ONeill, J. C., Flandrin, P., and Williams, W. J., On the existence of discrete
Wigner distributions, IEEE Signal Processing Letters, Vol. 6, No. 12, pp.
304306, Dec. 1999.
[20] Claasen, T. A. C. M., and Mecklenbrauker, W. F. G., The Wigner
distributiona tool for time-frequency signal analysis, Part II: Discrete-time
signals, Phillips Journal of Research, Vol. 35, No. 4/5, pp. 276300, 1980.
[21] Kadambe, S., and Adali, T., Application of cross term deleted Wigner rep-
resentation (CDWR) for sonar target detection/classification, Record of the
32nd Asilomar Conference on Signals, Systems & Computers, Vol. 1, pp.
822826, Nov. 14, 1998.
[22] Kadambe, S., and Orr, R., Comparative study of the cross term deleted
Wigner and cross biorthogonal representations, Record of the 31st Asilomar
Conference on Signals, Systems & Computers, Vol. 2, pp. 14841488, Nov.
25, 1997.
[23] Claasen, T. A. C. M., and Mecklenbrauker, W. F. G., The Wigner
distribution a tool for time-frequency signal analysis, Part I: Continuous-
time signals, Phillips Journal of Research, Vol. 35, No. 3, pp. 217250,
1980.
[24] Sun, M., et al., A Wigner spectral analyzer for nonstationary signals, IEEE
Trans. on Instrumentation and Measurement, Vol. 38, No. 5, pp. 961966,
Oct. 1989.
[25] Boashash, B. and Black, P. J., An ecient real-time implementation of the

Wigner-Ville distribution, IEEE Trans. on Acoustics, Speech and Signal
Processing, Vol. ASSP-35, No. 11, pp. 16111618, Nov. 1987.
Problems
1. (a) Using the LPIT, generate the default FMCW waveform and the
default P4 waveform. Load both signals into MATLAB, and truncate
such that they both have the same size (be sure to at least include one
to two code periods of each signal). (b) Add the two signals together
and save as a new signal (e.g., fmcw p4.mat). (c) Using the PWVD, an-
alyze each signal and extract the waveform parameters that are evident.
(d) Repeat (b) and (c) for SNR = 0 dB.
2. Using the PWVD, compute the Wigner-Ville distribution for the ran-
dom noise radar waveform, random noise plus FMCW waveform, ran-
dom noise FMCW plus sine, and random binary phase modulation dis-
cussed in Chapter 7. For each waveform, which modulation parameters
cant be extracted from the Wigner-Ville distribution?
3. Using the PWVD, compute the Wigner-Ville distribution of the (a)
polyphase signal that uses one of the orthogonal sequences given in
Table 10.12, (b) polyphase signal that uses one of the Doppler tolerant
orthogonal sequences in Table 10.14, and (c) frequency hopping signal
that uses one of the orthogonal frequency sequences given in Table 10.16.
4. To help identify the capability of the PWVD as a tool for identifying the
LPI modulation, extracting the modulation parameters, and to aid in
deciding on what signal processing algorithm performs best, construct
a table to show the PWVD measurement results for the LPI signals
contained in the Test Signals folder. For each parameter of interest,
show the actual value, the measured value, and the absolute value of
the relative error [see (12.26)]. Although the cross term interference
makes things particularly dicult, the measured results should tend
to coincide well with the actual values. The relative error depends on
how closely the PWVD results are examined. With noise added, the
measurement ability degrades slowly as the reader can document and
verify.
5. Generate the two orthogonal polyphase codes using ortho40.m and or-
tho40CE.m and the discrete frequency coding waveforms using dfc32.m
with the same signal parameters illustrated in Chapter 10. Use the
PWVD tools to examine the waveforms and determine if any coding
structure can be extracted from the time-frequency images or their mar-
ginal distributions.
Chapter 13
Choi-Williams Distribution
Analysis of LPI Radar
Waveforms
The pseudo Wigner-Ville distribution (PWVD) is useful for identifying the
LPI waveform modulation parameters due to the time-frequency characteris-
tics that are calculated. The PWVD time-frequency images however, contain
large cross terms, which can sometimes make identification of the modu-
lation, and extraction of the modulation parameters dicult especially in
low SNR situations. This chapter examines the Choi-Williams distribution
(CWD), which uses an exponential kernel in the generalized class of bilinear
time-frequency distributions to minimize the cross term components that are
so prevalent in the PWVD. The CWD is used to examine the LPI modu-
lations for comparison to the results in the PWVD chapter. By using the
CWD analysis tools, the intercept receiver can increase its processing gain
approaching that of the LPI emitter. The absence of strong cross terms in
the time-frequency plane allows the modulation type to be more readily de-
termined and also makes the extraction of the modulation parameters easier.
The CHOI folder on the CD provides the MATLAB tools that can be used
to re-create any of the figures presented, as well as generate new and useful
results.
445
13.1 Mathematical Development

The general class of time-frequency distributions introduced by Cohen is given
by
1
Cf (t, , ) = ej( t) (, )A(, )dd d (13.1)
2
where (, ) is a kernel function and

A(, ) = x + x (13.2)
2 2
and x() is the time signal, and x () is its complex conjugate. This repre-
sents a generalized class of a bilinear transformation that satisfies the mar-
ginals and has good resolution in both time and frequency spaces. The
Wigner-Ville time-frequency distribution, discussed in Chapter 12, is based
on (13.1) where the kernel function (, ) = 1. For multicomponent signals,
the cross terms that are present in the Wigner-Ville distribution were demon-
strated to be quite large. The cross terms cause interference that can obscure
physically relevant components of the LPI signals modulation.
Choi and Williams [1] realized that by choosing the kernel in (13.1) care-
fully, the calculation can minimize the cross terms and still retain the de-
sirable properties of the self-terms. The Choi-Williams distribution (CWD)
uses an exponential weighting kernel in order to reduce the cross term compo-
nents of the distribution. The kernel function that gives the Choi-Williams
distribution is 2 2
(, ) = e / (13.3)
where ( > 0) is a scaling factor. By substituting this kernel into (13.1)
the continuous CWD of the input signal x(t) is given as [1]

CWDx (t, ) = ej G(, )A(, )d d (13.4)
= = 4 2
where 2
/(4 2 )
G(, ) = e(t) (13.5)
and t is the time variable, is the angular frequency variable (2f ), and is a
positive-valued scaling factor. The bracketed term in (13.4) is the estimation
of the time-indexed autocorrelation. Just as for the WVD, the CWD can be
defined from the Fourier transform X() of x(t) by
2
1 ()
CWDX (t, ) = ejt e 42 /
2 = = 4 2

X + X dd (13.6)
2 2
Choi-Williams Distribution Analysis of LPI Radar Waveforms 447
and in discrete form, the Choi-Williams distribution is

1
CWDx ( , ) = 2 ej2
= = 4n2 /
( )2 /(4 2 )
e x( + )x ( ) (13.7)
For computational purposes it is necessary to apply the weighting windows

WN ( ) and WM () for the summations in (13.7) before evaluating the distri-
bution at each time index . The windowed Choi-Williams distribution can
then be expressed as

CWDx ( , ) = 2 WN ( )ej2 WM ()
= =
4 2
2
2
e 4 x( + + )x ( + ) (13.8)
where WN ( ) is a symmetrical window which has nonzero values for the range
of N/2 N/2, and WM () is a rectangular window which has a value
of 1 for the range of M/2 M/2. The parameter N , is the length
of the window WN ( ). The length N along with the shape of the window
determines the frequency resolution of the distribution. The parameter M ,
which is the length of the window WM (), determines the range from which
the time indexed autocorrelation is estimated.
The CWDx in (13.8) can also be expressed as
L
CWDx ( , ) = 2 S( , n)ej2n (13.9)
n=L
where the kernel is

M/2
(2 )2
1
S( , n) = W (n) e 4n2 / x( + n)x ( n) (13.10)
=M/2
4n2 /
where W (n) is a symmetrical window (such as Hamming), which has nonzero

values on the interval L to L, and W () is a uniform rectangular window
that as a value of 1 for the range of M/2 and M/2. The choices of N and
M on these windows, respectively, determine the frequency resolution of the
CWD and the range at which the function will be defined. Choi and Williams
state that decreasing the size of W (n) reduces the oscillatory fluctuations of
the cross terms, which at the same time decreases the frequency resolution of
the distribution. In other words, there is a trade-o between the reduction of
the cross terms and the frequency resolution obtained from the distribution.
When the above kernel function in (13.10) is compared to the one given
for the Wigner-Ville distribution
f (n) = x( + n)x ( n)w(n)w(n) (13.11)
the reader will notice that the CWD contains parameters similar to the
Wigner-Ville distribution, but includes an exponential term and introduces a
new summation. The reader will also notice that the CWD kernel function is
a series of Gaussian distributions. Barry points out that these distributions
are aligned diagonally and that the mean and variance of each distribution is
1 and 2n2 /, respectively [2].
As with the Wigner-Ville distribution, the discrete CWD can be modified
to fit the standard DFT by setting = k/2N . Substituting this result into
(13.9) and (13.10) above, and adding the window limits, we obtain [3]
2N1
k
CWDx , =2 S ( , n)ej2kn/N (13.12)
2n n=0
where the kernel function S ( , n) is defined as

S( , n), 0nN 1
S ( , n) = 0, n=N (13.13)

S( , n 2N ), N + 1 n 2N 1
13.2 LPI Signal Analysis

Next we review how the Choi-Williams distribution interprets various LPI
signals generated by the LPIT. The greatest advantage to the CWD is the
reduction of cross terms such as those in the WVD. Several types of LPI
signals will be evaluated to determine the eectiveness of the cross term
reduction and the overall suitability of using the CWD detection of the LPI
modulations. The MATLAB algorithm for the Choi-Williams distribution
uses the same type of kernel transformation as described for the Wigner-Ville
distribution. As the utility of these types of algorithms becomes increasingly
popular for signal analysis, there is a strong interest to execute the code as
fast as possible. Porting the MATLAB algorithms to C++ for execution on a
reconfigurable computing architecture can provide significantly faster results
than running them on a personal computer. In the reconfigurable computer,
the code execution is divided up between the microprocessors and the field
programmable gate arrays (FPGAs) [4]. Due to the dierent execution speed
of the FPGA processing elements compared to the microprocessor processing
elements, significant improvement in run time can result if the code division
is done correctly [5].
13.2.1 FMCW Analysis

In this section, extraction of the signal parameters for an FMCW waveform
example is investigated. The signal is sampled at a rate of fs = 7,000 Hz. The
signal has a center frequency of fc = 1,000 Hz, and a modulation bandwidth
of F = 500 Hz, and a modulation period of tm = 20 ms. Figure 13.1(a)
shows the CWD marginal frequency results of the FMCW and highlights the
carrier frequency. The time-frequency plot in Figure 13.1(b) clearly shows the
modulation period (tm ) and the modulation bandwidth (F ). The absence of
cross terms presents a clear picture of the modulation. Figure 13.2(a) shows
the CWD image for SNR = 0 dB and Figure 13.2(b) shows the CWD image
for SNR = 6 dB. The signal parameters are clearly visible in the 0 dB and
6 dB SNR. In summary, the CWD technique works well for the FMCW
waveforms, and it reduces the cross terms observed by the WVD.
13.2.2 BPSK Analysis

In this section we apply the CWD and use it to investigate the properties of
the BPSK CW signal. The parameters varied include the length of the Barker
code (number of subcodes Nc = 7 or 11) and the SNR (signal only, 0 dB,
6 dB). All signals have a fc = 1,000 Hz carrier frequency and a sampling
frequency of fs = 7,000 Hz. Both frequency domain and time domain plots
are shown for the BPSK signals after the CWD processing.
With the first signal examined shown in Figure 13.3(a), in the absence
of the cross terms, the carrier frequency of 1,000 Hz is suppressed with the
CWD. It can, however, be identified as the center of the 3-dB bandwidth.
The carrier frequency can also be identified in Figure 13.3(b) as the center of
the symmetric frequency distribution. In this case, there are 7 peaks within
the 3-dB bandwidth which correspond to the 7 subcodes in the Barker code.
Note also, that the Barker subcodes cannot be identified within the WVD.
In Figure 13.4, the 7-bit Barker code is examined with an SNR = 0 dB.
Figure 13.4(a) shows the frequency domain where the 7-bit code is still vis-
ible about the carrier frequency. With the time-frequency plot shown in
Figure 13.4(b), the subcodes about the carrier frequency are still clearly vis-
ible as well as the bandwidth. These results also indicate that the CWD also
appears to suppress the noise better than the WVD.
Figure 13.5(a) shows the marginal frequency domain for an Nc = 11-bit
Barker code BPSK signal only. All 11 peaks can be identified within the 3-dB
bandwidth. In Figure 13.5(b), the bandwidth and the carrier frequency can
still be identified.
Figure 13.1: CWD for an FMCW with F = 250 Hz, tm = 20 ms (signal

only), showing (a) the marginal frequency domain plot and (b) the time-
frequency plot.
Figure 13.2: CWD for an FMCW with a F = 250 Hz, tm = 20 ms time-

Figure 13.3: CWD for BPSK with 7-bit Barker code, cpp = 1, signal only,
showing (a) the marginal frequency domain plot and (b) the time-frequency
plot.
Figure 13.4: CWD for BPSK with 7-bit Barker code, cpp = 1, SNR = 0 dB,
plot.
Figure 13.5: CWD for BPSK with 11-bit Barker code, cpp = 1, signal only,
plot.
13.2.3 Polyphase Code Analysis

In this section, we examine two CW polyphase Frank code signals with the
CWD. Both signals are sampled in the receiver by an analog-to-digital con-
verter (ADC) with a sampling frequency of fc = 1,000 Hz. The CW signal
is generated with Nc = 16 phase codes (M = 4), with a cpp = 1 or one cycle
per subcode. Each subcode has a length of tb = 1 ms, resulting in a Frank
code signal with a code period of
Nc Nc cpp
T = Nc tb = = (13.14)
B fc
or 16 ms. Figure 13.6(a) shows the CWD marginal frequency results. The
carrier frequency can be identified by the largest peak value. The harmonic
spikes appear every n/T Hz (or 62.5n Hz). Figure 13.6(b) shows the CWD
time-frequency image. Unlike the WVD, the Frank code shows the exact
number of code periods intercepted. The bandwidth B can also be identified
in the image. When measuring the Frank code bandwidth B within the WVD
image, it was necessary to skip one of the modulation lines due to the presence
of cross terms. With the CWD there are no cross terms and the bandwidth
can be measured directly. The slope of each line has a magnitude of
T f2
S= = 2 c 2 = 62,500 Hz s1 (13.15)
B Nc cpp
The code period T is also measured directly as illustrated. This measurement
stresses the fact that the CWD integration period must be at least larger
than the signals code period in order to provide an accurate estimate of the
modulation parameters (# samples/fs > T ).
Figure 13.7(a) shows the CWD time-frequency image for SNR = 0 dB.
Although the addition of noise is present in the image, the phase code pa-
rameters can be easily determined. In Figure 13.7(b), the SNR = 6 dB.
The presence of the signal can be identified but the parameter extraction is
becoming more dicult.
13.2.4 Polytime Code Analysis

The structures of the polytime codes (T1 through T4) within the CWD are
significantly dierent than the polyphase codes such as the Frank code. For
the T1 code examined, fc = 1,000 Hz and T = 16 ms. The number of
frequency segments used is k = 4 and the number of phase states is n = 2.
This signal is referred to as T1(2). The sampling frequency of the ADC is
fs = 7,000 Hz. With a code period of 16 ms and 4 frequency segments, each
segment must be 4 ms in duration. The frequency step between adjacent
segments is 1/4 ms = 250 Hz with a total frequency excursion of 1,000 Hz.
Figure 13.8(a) shows the CWD marginal frequency domain. Note that the
Figure 13.6: CWD for Frank code with B = 1,000 Hz, T = 16 ms, signal only,
plot.
Figure 13.7: CWD for Frank code with B = 1,000 Hz, T = 16 ms,
time-frequency plot for (a) SNR = 0 dB and (b) SNR = 6 dB.
harmonics are uniformly spaced due to the time modulation of the binary
phase change. Figure 13.8(b) shows the time-frequency distribution of the
T1(2) code. The signal shows up as a set of vertical triangles stacked next to
each other. They are separated by the code period T . Figure 13.9(a) shows
the signal for an SNR = 0 dB. The signal can still be identified as a T1(2) and
the parameters can still be extracted. In Figure 13.9(b), with an SNR = 6
dB, no signal identification can be made and no parameters can be extracted.
13.2.5 FSK and FSK/PSK Analysis

In this section, we examine the CWD results for the FSK Costas signal and
a hybrid FSK/PSK signal. The CWD is used first to investigate the time-
frequency characteristics of the Costas frequency hopping sequence
{3, 2, 6, 4, 5, 1} kHz. The signal was generated with the LPIT using a fre-
quency duration of tp = 0.005 seconds and a sampling frequency of fs =
15,057 Hz. The CWD marginal frequency domain is shown in Figure 13.10(a).
Note that the frequencies in the sequence are present and there are no cross
terms, as were present in the WVD. The CWD time-frequency image is shown
in Figure 13.10(b). The positions of the six Costas frequencies are shown
clearly.
In Figure 13.11, the time-frequency image is displayed for both the 0-dB
and 6-dB SNR signals, respectively. As expected, the identification becomes
more dicult with increasing noise levels but the presence of the signals can
still be identified. With the FSK (Costas)/PSK signal, each frequency is
overlayed with a 5-bit narrowband Barker phase code with cpp = 5 (five
cycles per phase code). The results are shown in Figure 13.12(a, b) for the
signal only, and SNR = 0 dB, respectively. The frequencies in the Costas
code can be identified along with the phase code overlay. For the FSK/PSK
target signal described, the CWD generates excellent results and gives a clear
picture without cross terms.
13.3 Summary
The CWD theory was presented in this chapter. Several examples were used
to demonstrate generating an ecient kernel function and the subsequent
calculation of the CWD time-frequency results. The CWD was compared
to the WVD and the usefulness of the CWD to reduce the cross terms was
demonstrated for key LPI signals. The MATLAB folder CHOI contains the
file choi.m, which can be used to generate the results in this chapter or any
other LPI waveform results.
Figure 13.8: CWD for polytime code T1(2) with B = 1,750 Hz, T = 16 ms,
signal only, showing (a) the marginal frequency domain plot and (b) the
time-frequency plot.
Figure 13.9: CWD for T1(2) code with B = 1,750 Hz, T = 16 ms, time-
Figure 13.10: CWD for Costas code sequence {3, 2, 6, 4, 5, 1} kHz, showing
(a) the marginal frequency domain and (b) the time-frequency image.
Figure 13.11: CWD for time-frequency image for Costas code sequence
{3, 2, 6, 4, 5, 1} kHz, showing (a) SNR = 0 dB and (b) SNR = 6 dB.
Figure 13.12: CWD for the FSK/PSK signal, showing (a) signal only and (b)
SNR = 0 dB.
References
[1] Choi, H.I. and Williams W. J., Improved Time-Frequency Representation
of Multicomponent Signals Using Exponential Kernels, IEEE Trans. on
Acoustics, Speech, and Signal Processing, Vol. 37, No. 6, pp. 862871, June
1989.
[2] D. T. Barry, Fast calculation of the Choi-Williams time-frequency
distribution, IEEE Trans. on Signal Processing, Vol. 40, No. 2 pp. 450455,
Feb. 1992.
[3] Cardoso, J.C., Fish, P. J., and Ruano M. C., Parallel Implementation of
a Choi-Williams TFD for Doppler Signal Analysis, Proceedings of the 20th
Annual International Conference of the IEEE Engineering in Medicine and
Biology Society, Vol. 20, No. 3, 1998.
[4] Harkins, J., El-Ghazawi, T., El-Araby, E., and Huang, M., Performance
of sorting algorithms on the SRC 6 reconfigurable computer, Proceedings
of the IEEE International Conference on Field-Programmable Technology,
pp. 295 - 296, 1114 Dec. 2005.
[5] Upperman, G. J., Upperman, T. L. O., Fouts, D. J., and Pace, P. E.,
Ecient time-frequency and bi-frequency signal processing on a reconfig-
urable computer, Proceedings of the IEEE Asilomar Conference on Signals,
Systems & Computers, 2629 Oct. 2008.
Problems
1. Using the Choi-Williams distribution, (a) generate the time-frequency
results for the random noise radar waveform, the random noise plus
FMCW, random noise FMCW plus sine, and random binary phase mod-
ulations discussed in Chapter 7. For each waveform, which modulation
parameters can be extracted from the Choi-Williams distribution?
2. Using the Choi-Williams distribution, calculate the time-frequency dis-
tribution of the (a) polyphase signal that uses one of the orthogonal se-
quences given in Table 10.12 (ortho40.m), (b) the polyphase signal that
uses one of the Doppler-tolerant orthogonal sequences in Table 10.14
(ortho40CE.m), and (c) the frequency hopping signal that uses one of
the orthogonal frequency sequences given in Table 10.16 (dfc32.m).
3. To help identify the capability of the Choi-Williams distribution as

a tool for identifying the LPI modulation, extracting the modulation
parameters, and to aid in deciding on what signal processing algorithm
performs best, construct a table to show the Choi-Williams measure-
ment results for the LPI signals contained in the test signals folder. For
each parameter of interest, show the actual value, the measured value,
and the absolute value of the relative error [see (12.26)]. Compared
with Wigner-Ville distribution, the absence of the cross terms should
help considerably in identifying the LPI modulation and extracting the
modulation parameters.
Chapter 14
LPI Radar Analysis Using

Quadrature Mirror
Filtering
In Chapter 13, it was shown that the Choi-Williams distributions time-
frequency characteristics are useful for identifying LPI waveform parameters
and oered an improvement over the Wigner-Ville analysis due to the sup-
pression of the cross-terms which sometimes gave misleading results. In this
chapter, we investigate an LPI intercept receiver, based on a linear decom-
position of the received waveform through a quadrature mirror filter bank
(QMFB) tree, using wavelet filters. In this approach, the input signal is bro-
ken down into a series of time-frequency layers, with each subsequent layer
providing a trade-o in time and frequency resolution.
By examining the correct layers, the QMFB time-frequency receiver ap-
proach provides good estimates of the LPI signal parameters, making it easy
to distinguish between the dierent modulations, and extract the parameter
values. Parameters such as bandwidth, center frequency, energy distribution
within a tile (region in the time-frequency plane that contains most of the
wavelet basis functions energy), phase modulation, signal duration, and loca-
tion in the time-frequency plane can be determined. In addition, the number
of transmitters present and the types of LPI emitters can be determined. In
this chapter, the QMFB theory is presented first, followed by a discussion
of the mathematical waveform decomposition using wavelets. The QMFB is
then used to analyze and extract the parameters for the LPI signals discussed
in Part I. The QMFB folder on the CD provides the MATLAB tools that can
be used to re-create any of the figures presented, as well as generate new and
useful results.
467
14.1 Time-Frequency Wavelet

Decomposition
Various methods of decomposing a waveform on the time-frequency plane
have recently been investigated. The most common methods use orthogonal
basis functions, and can be divided into linear and bilinear transforms. The
short-time Fourier transform (STFT) and the wavelet transform (WT) are
examples of linear transforms. The Wigner transform discussed in the previ-
ous chapter is an example of a bilinear transform.1 After a discussion of basis
functions, the STFT and the WT are discussed, along with their advantages
and limitations.
14.1.1 Basis Functions

Linear transforms of a continuous time signal f (t) have the following form
ak = f (t)k (t)dt (14.1)
where (t) is the basis set, t is the time index, and k is the function index.
The Fourier transform, for example, has a basis set consisting of sines and
cosines of frequency 2k that oscillate forever. The basis functions are said
to be orthogonal if
E if k = 0
(t)(t k) = E(k) = (14.2)
0 otherwise
where E stands for the energy of(t) [1]. If (t) is normalized by dividing
by the square root of the energy E, then the basis functions are said to be
orthonormal defined by [2]
1 if k = 0
(t)(t k)dt = (k) = (14.3)
0 otherwise
If the basis functions are orthonormal, there is no redundancy in the repre-

sentation of the signal f (t). If the signal is sampled at or above the Nyquist
rate, all of the signals information is retained. In this case, the time variable
t in (14.1) and (14.3) can be considered to be discrete t = nT where T is the
sampling period and the integral should be replaced with summations.
1 Wigner transforms are called bilinear because the input waveform appears twice in the
development of the transform. Better resolution occurs in the time-frequency plane than
with linear techniques; however, the computational burden is greatly increased and the
cross terms can be bothersome for some applications.
LPI Radar Analysis Using Quadrature Mirror Filtering 469
14.1.2 Short-Time Fourier Transform Decomposition

The Fourier transform uses complex sinusoids as basis functions to perform
the analysis of signals. This approach is dicult, due to the infinite extent of
the basis functions as any time-local information (such as an abrupt change
in the signal) is spread out over the entire frequency axis [3]. This problem
was addressed by Gabor by introducing windowed complex sinusoids as basis
functions. This leads to the doubly indexed windowed Fourier transform:

XW F (, ) = ejt w(t )x(t)dt (14.4)

where w(t ) constitutes an appropriate window, and XW F (, ) is the

Fourier transform of x(t) windowed with w() shifted by . The function of
the window is to extract a finite-length portion of the signal x(t) such that the
spectral characteristics of the section extracted are approximately stationary
over the duration of the window. Also, if w(t) = 1 then the STFT reduces to
the conventional Fourier transform of x(t). In most applications, the magni-
tude of the STFT is of interest, and the display of the STFT magnitude is
usually referred to as a spectrogram [4].
The major advantage of the windowed transform or STFT is that if a
signal has most of its energy in a given time interval [T, T ] and frequency
interval [, ], then its STFT will be localized in the region [T, T ][, ]
and will be close to zero in time and frequency intervals, where the signal has
little energy [3]. A limitation of the STFT is that, because a single window
is used for all frequencies, the resolution of the analysis is the same at all
locations in the time-frequency plane. The possibility of having arbitrarily
high resolution in both time and frequency is thus excluded.
14.1.3 Wavelets and the Wavelet Transform

Wavelets are localized basis functions for time-frequency analysis of a signal.
That is, the wavelet basis function is eectively nonzero for only a finite
time interval, and is designed to satisfy the orthonormality condition (14.3).
From a signal processing point of view, a wavelet is a bandpass filter. In
the time-frequency analysis, the wavelet filter occurs most often in pairs (a
lowpass filter and a highpass filter), and includes a resampling function that
is coupled to the filter bandwidth as shown in the two-band analysis bank
in Figure 14.1. Here, H0 (z) is the highpass filter and H1 (z) is the lowpass
filter. Like the design of conventional digital filters, the design of a wavelet
filter can be accomplished by using a number of methods including weighted
least squares [5, 6], orthogonal matrix methods [7], nonlinear optimization,
optimization of a single parameter (e.g., the passband edge) [8], and a method
that minimizes an objective function that bounds the out-of-tile energy [9].
Figure 14.1: Two-band analysis bank.
A quadrature mirror filter (QMF) is an iteration of filter pairs with re-

sampling, to generate the wavelets. By varying the window used, resolution
in time can be traded for resolution in frequency. To isolate discontinuities in
signals, it is possible to use some basis functions, which are very short, while
longer ones are required to obtain a fine frequency analysis. One method to
achieve this is to have short high-frequency basis functions, and long low-
frequency basis functions [3]. The WT makes this possible by obtaining the
basis functions from a single prototype wavelet h(t) using translation, dila-
tion, and contraction as
1 tb
ha,b (t) = h (14.5)
a a
where a is a positive real number and b is a real number. For large a, the basis
function becomes a stretched version of the prototype wavelet (low frequency
function). For small a, the basis function becomes a contracted wavelet
(short high-frequency function). This basis function concept is shown in
Figure 14.2(a).
The WT is defined as

1 tb
XW (a, b) = h x(t)dt (14.6)
a a
The WT divides the time-frequency plane into tiles as shown in Figure 14.2(b).
Here, the area of each tile represents (approximately) the energy within the
function (rectangular regions of the frequency plane). Note that not all of
the signals energy can be located in a single tile because it is impossible to
concentrate the functions energy simultaneously in frequency and time. The
WT can be interpreted as constant-Q filtering with a pair of subband filters
(a lowpass filter and a highpass filter), followed by a sampling at the respec-
tive Nyquist frequencies corresponding to the bandwidth of the particular
subband of interest.
Figure 14.2: Basis functions and time-frequency resolution of the wavelet

transform: (a) basis functions and (b) coverage of time-frequency plane [3]
( c IEEE 1992).
Figure 14.3: Haar lowpass FIR filter.
14.1.4 Wavelet Filters

Finite impulse response (FIR) filters are the popular choice for the wavelet
filter. To meet the requirements for a wavelet filter, the coecients must
ensure an orthogonal decomposition of the input signal, such that the energy
at the input will equal the energy at the output from each filter pair [10]. The
filter pairs are designed to divide the input signal energy into two orthogonal
components based on the frequency. The filter should also pass as much
energy within its tile with a flat passband, and reject as much energy outside
the tile as possible.
Haar Filter
A classic example of a wavelet filter is the Haar basis function, which is not
continuous but is of interest because of its simplicity. The Haar basis function
is
1 for 0 t < 1/2
h(t) = 1 for 1/2 t < 1 (14.7)

0 otherwise
and is shown in Figure 14.3. The Haar basis function can serve as a wavelet
lowpass FIR filter and has two coecients, both with values of 1/2. The
Haar filter meets the wavelet requirements with the orthonormality being
easily verified since, at a given scale, the translations are nonoverlapping [3].
Because of the scale change by 2, the basis functions are orthonormal across
scale. Unfortunately, the Haar function is discontinuous. Although the filter
meets the wavelet requirements and perfectly tiles the input energy in time,
it does not tile well in frequency, and is not appropriate for signal processing.
Consequently, a continuous set of basis functions (or filters) is needed that
Figure 14.4: Sampled sinc filter impulse response.
best approximates the perfect time-frequency tiling, by minimizing the out-

of-time and out-of-frequency energy.
Sinc and Modified Sinc Filter

A function that tiles the energy perfectly in frequency would have a flat
magnitude response across the passband, an infinitely narrow passband-to-
stopband transition, and a zero across the stopband. From the time-domain
description (inverse Fourier transform), the function is called a sinc filter.
While it has an infinite number of coecients, this condition can be modified
by windowing [11]. The sinc filter can be expressed as
sin(k)
k=0
sinc(k) = k (14.8)
1 k=0
Since the passband ranges from /2 < < /2 or 0.25 < f < 0.25, the
nulls of the sinc function will be at 2T for a sampling period of T [12]. To
obtain the filter coecients, the sinc function is sampled at the normalized
sampling period of T = 1 for a situation similar to that shown in Figure 14.4.
One way to sample the function would be to let the main tap sample occur
at the center of the main lobe. However, two main taps are needed, and their
sum needs to be as large as possible. This occurs for the sinc function if both
main tap samples are equally spaced about the center of the main lobe [10].
The sum of the square of the coecients must be unity also, which is achieved
by scaling the sinc by 1/ 2, giving
1 n + 0.5
h(n) = sinc (14.9)
2 2
where n is an integer.
This filter meets the criteria of wavelet filters. The only problem is
that there is an infinite number of coecients. A small amount of non-
orthogonality will occur when this filter is truncated. Some cross correlation
will also take place between both highpass and lowpass filters. If the ends of
the filter are simply truncated (a rectangular window in the time domain),
some ripples in the passband of the frequency response will appear (Gibbs
phenomena).
One solution is to use a nonrectangular window, and one whose Fourier
transform has a narrower main lobe and smaller side lobes than the sinc
function. The Hamming window is one that is commonly used. Multiplying
the coecients from (14.9) by this window, and using the results in an FIR
filter, the frequency response needed is generated. Energy will be lost at the
filter transitions, which is primarily the result of the loss of orthogonality
from truncating the filter [11].
For detection, instead of losing the energy at those frequencies, a better
trade-o would be a small amount of cross correlation between the filters,
so that some energy appears in more than one tile. To achieve this type of
prototype filter, the impulse response can be modified to have a passband
that is slightly greater than /2. Thus, the lowpass and highpass filters
are squeezed together slightly. This can be achieved by compressing the
sinc envelope of (14.9) slightly. At the same time, it is desirable to rescale
the coecients slightly, so the sum of the squares equals one. With these
modifications, a modified sinc filter results as [9, 10]
S n + 0.5
h(n) = sinc w(n) (14.10)
2 C
where N/2 n (N 2)/2, C is the compression variable, S is the

scaling variable, N is the number of coecients, and w(n) is the Hamming
window to suppress the Gibbs phenomena. For these filters, the greatest
cross correlation occurs between tiles in the same frequency band, and ad-
jacent in time, when N = 512 (the number of coecients), with values
C = 1.99375872328059, S = 1.00618488680080, and a Hamming window with
a cross correlation of less than 0.001 results.2 Note the number of coecients
N can be changed using the MATLAB file tsinc su.m.
14.2 Discrete Two-Channel

Quadrature Mirror Filter Bank
Digital two-channel QMFB structures have found applications in many areas,
including modems, data transmission, image, and video coding. Figure 14.5
2 Personal communication between P. Jarpa and T. Farrell, March 20, 2002.
Figure 14.5: Two-channel quadrature mirror filter bank [5] ( c IEEE 2000).
Figure 14.6: Typical frequency response of the analysis filters.
shows the basic two-channel QMFB structure consisting of an analysis filter

bank and a synthesis filter bank. In the two-channel analysis filter bank, a
discrete time signal x[n] is first split into two subbands {vk [n]} by means of the
wavelet or analysis filters H0 (z) and H1 (z). The two-band analysis filter bank
containing the filters H0 (z) and H1 (z) typically have lowpass and mirror-
image highpass frequency responses, respectively, with a cuto frequency at
/2 [5, 13]. The typical frequency response of the analysis filters is shown
in Figure 14.6. After filtering, each subband signal is downsampled by 2, to
form the outputs of the analysis stage. These signals can then be analyzed
or processed in various ways, depending on the application. The signals
are then transmitted to the signal synthesis section, where the signals are
upsampled by a factor of 2, and passed through a two-band synthesis filter
bank composed of the filters G0 (z) and G1 (z), whose outputs are then added
yielding y[n]. The purpose of the synthesis filters is to eliminate the images
that are formed in the analysis stage.
It follows from the figure that the sampling rates of the input signal x[n]
and output signal y[n] are the same. The reconstructed signal y(n) diers,
however, from the input x(n) due to aliasing, amplitude distortion, and phase
distortion [5, 13]. Consequently, the analysis and the synthesis filters in the
QMFB are chosen so as to ensure that the reconstructed output y[n] is a
Figure 14.7: Wavelet filter bank tree filtering the lowpass component (wavelet
tiling) [9, 12].
reasonable replica of the input x[n]. Moreover, they are also designed to
provide good frequency selectivity, to ensure that the sum of the power of the
subband signals is reasonably close to the input signal power.
14.3 Tree Structure to Filter the Lowpass

Component
Finite impulse response filters and downsamplers can also be arranged in
a tree structure, as shown in Figure 14.7, to eect an orthogonal wavelet
decomposition of a signal [9, 12]. This structure filters the lowpass output
(H filter) from each stage. The discrete input waveform is denoted as the
sequence {c0 } and the output sequences of each branch are as shown in the
figure. Since each branch of the tree downsamples by 2, each sequence will
have half as many elements as the preceding sequence. A filter tree using the
same orthogonal pair of filters throughout and with equal length branches, as
in Figure 14.7, yields a rectangular tiling diagram. The time-frequency tiling
diagram shown in Figure 14.8 is one method that can be used to describe
this decomposition. The time-frequency tile is the region in the plane that
contains most of that functions energy. However, not all of a functions
energy can be located in a tile, because it is impossible to fully concentrate
energy simultaneously in time and frequency. The tiles are of dierent shapes,
but have a constant area and trade frequency resolution for time resolution,
and vice versa.
Figure 14.8: Time-frequency diagram for the lowpass wavelet filter bank tree
[9] ( c IEEE 1999).
Figure 14.9: Wavelet filter bank tree filtering the highpass component.
14.4 Tree Structure to Filter the Highpass

Component
The last section demonstrated that by cascading filters and filtering the low-
pass component of the previous output, a tiling with finer frequency resolution
at lower frequencies was achieved. Now consider the cascading filter diagram
in Figure 14.9 where, instead of filtering the lowpass output of each stage, the
highpass filter (G filter) output is filtered. Again, the input sequence is split
at each stage into high-frequency and low-frequency orthogonal sequences.
The tiling diagram is shown in Figure 14.10.
Notice that the second and third layers seem to be flipped in Figure 14.9.
The figure is drawn so that the output sequence at the top of the drawing
contains the highest frequency components of the input sequence. To un-
derstand why they are flipped, consider the aliased frequency spectrum of
Figure 14.10: Time-frequency diagram for the highpass wavelet filter bank
tree.
the filters, shown in Figure 14.11. The output from the G filter in the first
layer contains the higher frequency components of the original sequence, but
shifted, so that it is actually the dc component of the output of G. The result
is that the output of G is frequency reversed, much like the lower sideband of
a single sideband communication system. A similar structure farther down
the cascade will unflip the signal.
It is possible to create another tiling scheme by combining the wavelet
filter bank and wavelet tiling, as demonstrated in Figure 14.12. In order to
keep the higher frequency outputs of each branch above the lower frequency
outputs, the construction rule for this figure is to count the number of G
filters up to the branch. If the number is even, the next G filter will output
the high frequencies. If odd, the next H filter will output the high frequencies.
14.5 QMFB Tree Receiver

Orthogonal wavelet decomposition of the unknown signal can be implemented
using QMFs, by designing filter pairs to divide the input signal energy into two
orthogonal components, based on frequency. The tiles are used to refer to the
rectangular regions of the time-frequency plane containing the basis functions
energy. By arranging the QMF pairs in a fully developed tree structure, it
is possible to decompose the waveform in such a way that the tiles have the
same dimensions within each layer. Thus, every filter output is connected to
a QMF pair in the next layer, as shown in Figure 14.13 [10]. Each QMF pair
divides the digital input waveform into its high-frequency and low-frequency
components, with a transition centered at . A normalized input of one
sample per second is assumed, with a signal bandwidth of [0, ]. Since each
filter output signal has half the bandwidth, only half the samples are required
Figure 14.11: Frequency response of filters H and G.
to meet the Nyquist criteria; therefore, these sequences are downsampled by

two. The same number of output samples is returned. For example, if 100
samples appear at the input of the first QMF pair, 100 samples appear at
the output. Each of the two resulting sequences is then fed into QMF pairs,
forming the next layer, where the process is repeated, and so on down the
tree.
Within the time-frequency plane, the WT is sharper in time at high fre-
quencies. At low frequencies, the WT is sharper in frequency. That is, the
tiles become shorter in time and occupy a larger frequency band, as the fre-
quency is increased. Since the WT is linear, there is a fundamental limit on
the minimum area of these tiles. However, the nature of the QMFB config-
uration is such that each layer outputs a matrix of coecients for tiles that
are twice as long (in time) and half as tall (in frequency) as the tile in the
previous layer. The outputs from each layer of the tree in Figure 14.13 form
a matrix whose elements, when squared, approximately represent the energy
contained in the tiles of the corresponding time-frequency diagrams shown in
the figure.
The block diagram of a receiver that uses the QMFB structure is shown
in Figure 14.14. A received waveform is bandpass filtered and sampled at the
Nyquist rate. The digital sequence is then fed to the QMFB tree where it
is decomposed. Matrices of values are output from each layer, and are then
squared to produce numbers representing the energy in each tile.
Wavelet decomposition has been investigated as a tool for pattern recogni-
Figure 14.12: Combining the wavelet filter bank and wavelet tiling. (Down-
sampling by 2 is included in each filter box.)
Figure 14.13: Quadrature mirror filter bank tree [10] ( c IEEE 1996).
Figure 14.14: Quadrature mirror filter bank tree receiver [10] ( c IEEE 1996).
tion and target detection [14], and also as a means for identifying signal mod-
ulations [15]. The architecture discussed above has been used to investigate
the detection of LPI signals in [1619], and is used below to investigate the
parameters of the signals in Part I. We start with two example calculations,
to become familiar with the QMFB processing and its output waveforms.
14.6 Example Calculations

In this section two example calculations are shown for a complex input. A
complex single-tone example is shown first, followed by a two-tone signal.
These examples serve to demonstrate the dierent QMFB output layers, and
show the trade-o in time-frequency resolution as a function of the layer
number being examined. The lower the layer number, the smaller (better)
the resolution in time, and consequently the larger (poorer) the resolution in
frequency. As the layer number gets larger, the resolution in time gets larger,
and the resolution in frequency gets smaller.
14.6.1 Complex Single-Tone Signal

To demonstrate the results available from the QMFB signal processing, we
again consider a complex, single-tone signal as in Chapter 9. The signal has
a carrier frequency fc = 1 kHz and is sampled by the ADC at a rate of 7 kHz.
The results, shown in Figure 14.15 show layers 2, 3, and 4, respectively, in
the time-frequency domain using gray scale plots. Figure 14.16(a, b), show
layers 5 and 6. One of the important objectives of showing the five layers
of the QMFB is to demonstrate how each layer results in a matrix of energy
values, and the fact that the tiles are twice as long (in time) and half as tall
(in frequency) as the tile in the previous layer. That is, as the layer number
increases, the frequency resolution gets smaller, and the time resolution gets
larger. This adds quite a bit of flexibility to the analysis of nonstationary
signals. Several dierent layers can be examined and compared, and the
parameters of the signals can be extracted with high fidelity. Also, since the
first and last layer in the QMFB are a single row of data, it is not useful to
display them in a time-frequency format.
The input signal is zero padded with z zeros, such that the resulting
number of data points is a power of 2. This resulting power of 2 is the number
of layers L within the QMFB that result. That is, Np = 2L . The QMFB
output resolution depends on the layer number. The frequency resolution of
a layer l is [19]
fs fs
f = l
= (14.11)
2(2 1) 2(NF )
where NF is the number of tiles displayed in frequency. For example, for layer
2 in Figure 14.15, f = 7,000/2(3) = 1,166.67 Hz. The resolution in time is
Figure 14.15: Time-frequency layers for the 1-kHz single-tone signal, showing
(a) layer 2, (b) layer 3, and (c) layer 4.
Figure 14.16: Time-frequency layers for the 1-kHz single-tone signal, showing
(a) layer 5, and (b) layer 6.
determined by how many samples are integrated within the QMFB. For layer
l < L,
Np Np
t = Ll
= (14.12)
fs (2 1) fs NT
where L is the total number of layers, and NT is the number of tiles in
time. Also, Np = 2L . For example, for layer 2 in Figure 14.15, t = 128
(1/7,000)/31 = 590 s. Also note that the lower layers (e.g., layers 2 and
3) can be used to identify how many samples of the signal were collected
(excluding zero padding). Since the sampling period for this example is T =
0.143 ms, from layer 2 we see that 105 samples were collected, and that 23
zeros were used to pad the signal.
Referring to layer l = 6 in Figure 14.16(b), the tiles have a frequency
resolution of f = 55.5556 Hz and a resolution in time of t = 18.286 ms.
Layer 6 shows the signal between 944.445 Hz and 1055.56 Hz, and from 0 to
18.286 ms. That is, we can say that fc 1,000 Hz with the accuracy limited
by the tile resolution. Note that if more detailed time information is required,
a lower layer could be examined.
14.6.2 Complex Two-Tone Signal

The second example consists of a signal with two frequencies fc1 = 1 kHz
and fc2 = 2 kHz, with a sampling frequency fs = 7 kHz. In this example, a
contour plot is used. Although the gray scale plot illustrated above quantifies
the energy within each tile, the contour plot is useful for other types of in-
formation (such as time-domain characteristics), as illustrated in the results
below. The number of signal samples collected, the time resolution t, and
frequency resolution f for each layer within the QMFB are the same as for
the single-tone example above. Figure 14.17 shows the contour plot for layers
2 through 4. Figure 14.18 shows the contour plot for layers 5 and 6. As be-
fore, layers 1 and 7 are not displayed, since they have only a single row vector
of data. This example illustrates the important concepts that are evident
using a contour image. First, for lower layers such as layer 2 and layer 3, the
time domain characteristics of the signals can be clearly identified. In layer 2,
the complex phase interaction in time, of the two signals within a single filter,
can also be identified. The high-frequency resolution layers [such as layer 6
shown in Figure 14.18(b)] reveal the frequencies contained in the input signal
with a good amount of accuracy (f = 55 Hz).
Figure 14.17: Time-frequency layers for the two-tone signal (1 kHz, 2 kHz),
showing (a) layer 2, (b) layer 3, and (c) layer 4.
Figure 14.18: Time-frequency layers for the two-tone signal (1 kHz, 2 kHz),
showing (a) layer 5, and (b) layer 6.
14.7 FMCW Analysis

In this section, the extraction of FMCW signal parameters (discussed in
Chapter 4) is demonstrated using the QMFB, and the appropriate layers
for parameter extraction are identified.
The first example is shown in Figure 14.19. In Figure 14.19(a), the l = 2
layer is shown. For the number of signal samples collected, L = 11 (Np =
2, 048). For layer l = 2, f = 1,166.67 Hz and t = 572.5 s. For this
signal, it appears that four periods of the triangular FMCW waveform were
captured. The waveforms also have the general appearance of a linear FM
modulation. The modulation period tm can be measured as tm = 20 ms, for a
total signal length of 160 ms. Notice that the concentrations of energy within
each period of the FMCW waveform are not centered on the carrier frequency
fc = 1,000 Hz, but contain a 300-Hz bias. This is due to the fact that there
are only three filters in this layer being used to calculate the results.
In Figure 14.19(b), layer l = 6 is shown with f = 55.55 Hz and t =
8.866 ms. This layer is appropriate, since the time resolution t < tm /2. Note
that the modulation period can be easily identified and measured. Since this
layer also has 63 filters, each with a narrow filter width, the bandwidth F
can be measured with good accuracy, as F = 250 Hz.
Figure 14.20 shows the QMFB layer 5 performance against the FMCW
Figure 14.19: QMFB contour images for FMCW F = 250 MHz, tm =

20 ms (signal only), showing (a) layer 2, and (b) layer 6.
Figure 14.20: QMFB contour image for FMCW F = 250 MHz, tm = 20 ms

(signal only), showing layer 5 with SNR = 0 dB.
signal with SNR = 0 dB. Note that the parameters of the signal can still be
measured satisfactorily. Appendix I presents the results for an FMCW signal
with F = 500 Hz.
14.8 BPSK Analysis

In this section, the QMFB is used to investigate the properties of the binary
phase shift keying signal discussed in Chapter 5, and also investigated in
Chapter 9 with the PVWD. The signal parameters that were changed include
the length of the Barker code (7 or 11), and the SNR (signal only, SNR = 0
dB, SNR = 6 dB).
In the first example, fc = 1 kHz, fs = 7 kHz, and cpp = 1, and a
7-bit Barker code is used. The total number of layers is L = 11 (Np =
2,048), and the layers investigated for this signal are l = 3 and l = 6. Figure
14.21(a) shows the l = 3 layer (f = 500 Hz, t = 1.15 ms), and the 25
BPSK code periods captured. The number of subcodes within a code period
T cannot, however, be distinguished with this particular QMFB layer and
scaling. A closer look however is illustrated in Figure 14.21(b). In this figure
(f = 500 Hz, t = 286 s), a concentration of energy can be located near
the beginning of each code period (along the carrier frequency fc = 1 kHz).
This is due to the three contiguous subcodes, all with the same phase. The
code period can be measured between these energy concentrations as T = 7

ms. Note that care must be exercised in this measurement, since a secondary
feature could be used, resulting in T 14 ms. The total length of the signal
is 25 times 7 ms = 0.175s and can also be identified in Figure 14.21(a).
Figure 14.22(a) examines the l = 6 layer (f = 55 Hz, t = 9.44 ms)
to further evaluate the frequency content of the signal at a higher resolution.
The first important feature is the frequency bands, due to the various lengths
of subcode groups within the code period. The overall bandwidth of the signal
B = fc /cpp is also shown, and depends on the length of a single subcode.
Figure 14.22(b) shows the same layer 6, except that the SNR = 0 dB, and
demonstrates the eects of noise on the QMFB output.
To further illustrate the QMFB capability to analyze the microstructure
of the phase-modulated signal, cpp is increased from 1 to 5. That is, the
bandwidth of the signal is narrowed from 1 kHz to B = fc /cpp = 0.2 kHz
(code period of T = cpp/fc = 35 ms). Figure 14.23(a) shows the QMFB layer
2 for two code periods of the 7-bit signal. The important feature illustrated
here is the presence of a null at each BPSK phase shift. After the total
code period T is determined, the measurement of the smallest subphase
code is performed (time is measured between the two closest nulls). After
this, the number of subphase codes contained within each section of the code
is determined, which then uniquely identifies which BPSK code has been
intercepted. Figure 14.23(b) shows the QMFB layer 6. Note that the code
period T is clearly identified, as well as the bandwidth. The QMFB results
for an 11-bit code are shown in Appendix J.
Figure 14.24 shows the QMFB layer 6 performance against the BPSK sig-
nal, with SNR = 0 dB and cpp = 5. The parameters of the signal such as the
code period, carrier frequency, and bandwidth, can still be measured satis-
factorily. For layer 2, however, the noise severely distorts the time-frequency
results, and identification of the phase nulls is not possible, making it dif-
ficult to identify the exact binary phase code used, without further image
processing.
Figure 14.21: QMFB layer 3 contour plot for BPSK with 7-bit Barker code,
cpp = 1 (signal only), showing (a) the complete captured signal, and (b) a
close-up view showing the code period T = 7 ms.
Figure 14.22: QMFB layer 6 for BPSK with 7-bit Barker code with cpp = 1,
showing the (a) contour image (signal only), and (b) contour image with
SNR = 0 dB.
Figure 14.23: QMFB contour images for BPSK with 7-bit Barker code, cpp =
5 (signal only), showing (a) layer 2, and (b) layer 6.
Figure 14.24: QMFB layer 6 contour image for BPSK cpp = 5 signal, with
SNR = 0 dB.

In this section we examine the polyphase codes with the QMFB, beginning
with the Frank code. In Chapter 5 we saw that the Frank code is derived from
a linear FM waveform. The phase modulation is applied both to the I and Q
signals, which are 90 degrees out of phase. In this and the following sections,
it is shown that the QMFB can be used to not only identify a particular type
of phase modulation, but also to extract the important parameters of the
signal.
The Frank phase code signal is generated with Nc = 64 (M = 8). The
phase codes for M = 8 are shown in Figure 14.25. This is demonstrated
in the QMFB l = 2 layer shown in Figure 14.26. The number of layers in
this example is L = 12 (Np = 4,096). For this layer f = 1,166.67 Hz,
and t = 571.99 s (small dierence). In Figure 14.26(a), the additional
48 subcodes within a code period results in a longer duration signal. The
five code periods have a total length of 320 ms. Figure 14.26(b) shows a
close-up of the frequency characteristics within a code period. The linear
frequency modulation characteristics are viewed in the QMFB l = 5 layer
in Figure 14.27(a, b). Here, the bandwidth can be clearly identified, as well
as the code period T = 64 ms. Note the wraparound characteristic within
the bandwidth, similar to the 16-subcode example above. Correlation of the
Figure 14.25: Frank code phase values for M = 8.
occurrence in time of the eight major energy concentrations within T , with the
M = 8 Frank phase modulation waveform sections, can be easily made. The
distribution of the signal energy within the nine frequency tiles within B helps
in identifying the phase code, and in distinguishing between the modulation
characteristics. The frequency characteristics for the M = 8 Frank code is
shown in Figure 14.28. The energy is distributed about the carrier frequency
in a Gaussian-type distribution, with the carrier frequency fc centered about
tile nine (the tile with the largest energy content). In fact, from Figure 14.28
the five largest energy tiles (in order from largest to smallest) are 9, 10, 7,
11, and 8. Figure 14.29 shows the l = 6 layer results for the M = 8 Frank
code with the SNR = 0 dB. Note that the parameters can still be measured
quite accurately. The QMFB results for the Nc = 16 Frank signal (M = 4)
are given in Appendix K and the results for the P1, P2, P3, and P4 are given
in Appendix L.

In this section the T1(2) polytime code is analyzed using the QMFB.
Figure 14.30 shows the QMFB contour images for the polytime T1(2) code
with a resulting B = 1,750 Hz, T = 16 ms (signal only). The QMFB for
this signal has l = 10 layers. Figure 14.30(a) shows the layer 2 output, and
Figure 14.30(b) shows a close-up of layer 2, showing the frequency changes
due to binary phase code that varies as a function of time. Figure 14.31(a)
shows the fourth layer QMFB contour images for the polytime T1(2) code.
Figure 14.31(b) shows a close-up of layer 4, showing the resulting linear fre-
Figure 14.26: QMFB contour images for M = 8 Frank code with B = 1,000
Hz, T = 64 ms (signal only), showing (a) layer 2 output, and (b) close-up of
layer 2, showing detailed frequency changes due to phase codes.
Figure 14.27: QMFB contour images for M = 8 Frank code with B = 1,000
layer 5, showing resulting linear frequency modulation.
Figure 14.28: QMFB layer 5 frequency profile for Nc = 64 Frank code.
Figure 14.29: QMFB layer 6 contour image for Frank Nc = 64 signal with
SNR = 0 dB.
quency modulation. Note that the resulting bandwidth is easily identified.

The QMFB results for the T2(2) to T4(2) are shown in Appendix M.
14.11 Costas Frequency Hopping Analysis

In this section, five code periods of a Costas frequency hopping signal are
examined. The Costas code within each period is 4, 7, 1, 6, 5, 2, and 3 kHz.
In Figure 14.32(a), the QMFB layer 4 shows that the frequencies can easily
be identified. The close-up in Figure 14.32(b) reveals that no modulation is
present on the frequency. Figure 14.33 shows layer 6 with a finer frequency
resolution. Note the spread in time, demonstrating the trade-o in resolution
from one layer to the next.
14.12 FSK/PSK Signal Analysis

When a binary phase code modulation is added to the frequency hopping
signal, the bandwidth about the carrier is increased. The QMFB layer 4
results for the FH code above, with a binary 5-bit Barker code added, is
shown in Figure 14.34. Note that in the close-up figure, the phase modulation
is clearly present. In Figure 14.35, the QMFB layer 6 is shown. Due to the
decrease in time resolution, the phase modulation is not as distinct as in
layer 4.
14.13 Noise Waveform Analysis

The random noise waveforms discussed in Chapter 7 can also be examined
with the QMFB. In Figure 14.36, a random noise waveform that has a band-
width of B = 300 MHz, carrier frequency of fc = 350 MHz and a code period
of T = 4 s is examined with the QMFB layer 6 (out of 13). Note that the
bandwidth can be easily identified as well as the carrier frequency and code
period. The random noise plus triangular FMCW waveform with tm = 1 s,
F = 300 MHz, B = 300 MHz, and fc = 350 MHz is shown in Figure 14.37.
Note in layer 6 the noise bandwidth and the FMCW modulation are easily
identified. The use of the QMFB to examine the random noise FMCW plus
sine waveform is left as an exercise for the reader.
The random binary phase modulation waveform is shown in Figure 14.38
using layer 7. For this waveform, fc = 300 MHz, cpp = 3, and
tb = 10 ns. With Nc = 64 and 5 code periods included, T = 3.2 s. These
signal parameters can be easily identified from layer 7. For the signal only,
layer 2 is examined in Figure 14.39 and shows the direct correlation of the
phase modulation parameters used.
Figure 14.30: QMFB contour images for polytime T1(2) code with resulting
B = 1,750 Hz, T = 16 ms (signal only), showing (a) layer 2 output, and
(b) close-up layer 2, showing detailed frequency changes due to phase codes.
Figure 14.31: QMFB contour images for polytime T1(2) code with resulting
B = 1,750 Hz, T = 16 ms (signal only), showing (a) layer 4 output, and
(b) close-up of layer 4 showing resulting linear frequency modulation.
Figure 14.32: QMFB contour images for layer 4 for FSK code using Costas
sequence, showing (a) layer 4 output, and (b) close-up of layer 4 showing
frequency resolution.
Figure 14.33: QMFB contour images for layer 6 for FSK code using Costas
sequence, showing (a) layer 6 output, and (b) close-up of layer 6 showing
frequency resolution.
Figure 14.34: QMFB contour images for layer 4 for FSK/PSK code using
Costas sequence plus 5-bit Barker code cpp = 5 (signal only), showing
(a) layer 4 output, and (b) close-up of layer 4 showing Barker phase modula-
tion.
Figure 14.35: QMFB contour images for layer 6 for FSK/PSK code using
Costas sequence plus 5-bit Barker code cpp = 5 (signal only), showing (a)
layer 6 output, and (b) close-up layer 6 showing Barker phase modulation.
Figure 14.36: QMFB contour image for the random noise waveform showing
the extraction of the bandwidth and carrier frequency from l = 6.
14.14 Summary
The QMFB theory was presented in this chapter and several examples were
used to demonstrate the time-frequency results. To extract the unknown
signal parameters, several layers must be examined to determine those that
provide the best information. The phase changes can be identified from the
lower layers, while the frequency information is best obtained from the higher
layers.
Contrary to the Wigner-Ville distribution, and Choi-Williams distribu-
tion, the QMFB performs remarkably well for the FSK Costas code and
FSK/PSK (binary phase code) signal. Not only were the frequency hops
identified, but the frequency duration could also be indentified, as well as
the binary phase modulation if present. For the FSK/PSK target signal re-
ported, the QMFB was not able to identify any meaningful signal parameters
for the same reason that the PWVD could not. The use of the QMFB was
also demonstrated to work remarkably well with the random noise modula-
tions. The main drawback is the fact that even if the most useful results are
from, for example, layer 7, all of the other layers prior to layer 7 must still be
computed.
The presentation of the QMFB results to a trained operator will allow the
signal parameters to be extracted, and can enable good classification results
Figure 14.37: QMFB contour image for the random noise plus FMCW wave-
form showing the extraction of the noise bandwidth, FMCW modulation, and
carrier frequency.
Figure 14.38: QMFB contour image layer 7 for the random binary phase
modulation showing the extraction of the bandwidth, and carrier frequency.
Figure 14.39: QMFB contour image layer 2 for random binary phase modu-
lation showing the direct correlation of the phase modulation used to create
one code period.
when the information from several layers is combined. The use of the QMFB
in noisy environments, however, gives problems in parameter identification,
and further image processing is necessary.
References
[1] Burrus, C. S., Gopinath R. A., and Guo, H., Introduction to Wavelets and
Wavelet Transforms, A Primer, Prentice Hall, Upper Saddle River, NJ, 1998.
[2] Chui, C. K., Wavelets: A Mathematical Tool for Signal Analysis, First Edi-
tion, SIAM, Philadelphi, PA, 1997.
[3] Vetterli, M., and Herley, C., Wavelets and filter banks: Theory and design,
IEEE Trans. on Signal Processing, Vol. 40, No. 9, pp. 22072232, Sept.
1992.
[4] Mitra, S., Digital Signal Processing, A Computer-Based Approach, Second
Edition, McGraw-Hill, Boston, MA, 2001.
[5] Al-Namiy, F., and Nigam, M. J., On the design of 2-band FIR QMF filter
banks using WLS techniques, Proc. of the Fourth IEEE International Con-
ference on High Performance Computing in the Asia-Pacific Region, Vol. 2,
pp. 772776, May 2000.
[6] Goh, C. K., and Lim, Y. C., A WLS algorithm for the design of low-delay
quadrature mirror filter banks, Proceedings of the IEEE International Sym-
posium on Circuits and Systems, Vol. 1, pp. 615618, May 2000.
[7] Zahhad, A., and M. A. Sabah, Design of selective M-channel perfect recon-
struction FIR filter banks, IEE Electronics Letters, Vol. 35, No. 15, pp.
12231225, 1999.
[8] Zhang, Z., and L. Jiao, A simple method for designing pseudo QMF banks,
Proceedings of the IEEE International Conference on Communication Tech-
nology, Vol. 2, pp. 15381541, Aug. 2000.
[9] Farrell, T., and Prescott, G., A Method for Finding Orthogonal Wavelet
Filters with Good Energy Tiling Characteristics, IEEE Trans. on Signal
Processing, Vol. 47, No. 1, pp. 220223, Jan. 1999.
[10] Farrell, T., and Prescott, G., A Nine-Tile Algorithm for LPI Signal Detec-
tion Using QMF Filter Bank Trees, Proceedings of the IEEE Conference on
Military Communications MILCOM 96, Vol. 3, pp. 974978, 1996.
[11] Proakis, J., and Manolakis, D., Digital Signal Processing. Principles, Algo-
rithms, and Applications, Third Edition, Prentice Hall, Upper Saddle River,
NJ, 1996.
[12] Herley, C., et. al., Tilings of the time-frequency plane: Construction of
arbitrary orthogonal bases and fast tiling algorithms, IEEE Transactions on
Signal Processing, Vol. 41, No. 12, pp. 33413359, Dec. 1993.
[13] Shang, Y., Longzhuang, L., and Ho, K. C., Optimization design of filter
banks for wavelet denoising, Proceedings of the 5th International Conference
on Signal Processing, Vol. 1, pp. 306310, Aug. 2000.
[14] Chan, L. A., and Nasrabadi, N. M., An application of wavelet-based vector

quantization in target recognition, Proceedings of the IEEE International
Joint Symposium on Intelligence and Systems, pp. 274281, Nov. 1996.
[15] Ho, K. C., Prokopiw W., and Chan, Y. T., Modulation identification of
digital signals by the wavelet transform, IEE Proceedings Radar, Sonar and
Navigation, Vol. 147, No. 4, pp. 169175, Aug. 2000.
[16] Farrell, T., and Prescott, G., A Low Probability of Intercept Signal Detection
Receiver Using Quadrature Mirror Filter Bank Trees, IEEE International
Conference on Acoustics, Speech and Signal Processing, Vol. 3, pp. 1558
1561, March 1996.
[17] Copeland, D. B., and Pace, P. E., Detection and analysis of FMCW and
P-4 polyphase LPI waveforms using quadrature mirror filter trees, Proceed-
ings of the IEEE International Conference on Acoustics Speech and Signal
Processing, Vol. 1, May 2002.
[18] Taboada, F., et al., Intercept receiver signal processing techniques to de-
tect low probability of intercept radar signals, Proceedings of the 5th Nordic
Signal Processing Symposium, Hurtigruta Tromso, Norway, Oct. 2002.
[19] Jarpa, P., Quantifying the dierences in low probability of intercept radar
waveforms using quadrature mirror filtering, Naval Postgraduate School
Masters Thesis, Sept. 2002.
Problems
1. (a) Using the LPIT, generate the FMCW waveform (signal only, SNR =
0 dB, and SNR = 6 dB) with a carrier frequency of 2 kHz, tm =
5 ms, and F = 500 Hz. (b) Process the signals with the QMFB
algorithm. (c) For each useful layer, diagram your estimates of all the
signal parameters for signal only, SNR = 0 dB and SNR = 6 dB.
2. (a) Using the LPIT, generate a P4 waveform (signal only, SNR =
0 dB, and SNR = 6 dB) with a carrier frequency of 2 kHz, Nc = 128,
tb = 1 ms, and fs = 7,000 Hz. (b) Process the signals with the QMFB
algorithm. (c) For each useful layer, diagram your estimates of all the
signal parameters for signal only, SNR = 0 dB, and SNR = 6 dB.
such that they both have the same size (be sure to at least include 1 to 2
code periods of each signal). (b) Add the two signals together and save
as a new signal (e.g., fmcw p4.mat). (c) Using the QMFB processing,
analyze the signal and extract the waveform parameters for each signal
that is evident. (d) Repeat (b) and (c) for SNR = 0 dB. (e) Repeat (b)
and (c) for SNR = 6 dB.
4. Using the QMFB tools, (a) add a tic and tock command to the program
and then (b) examine the random noise, random noise plus FMCW,
random noise FMCW plus sine, and random binary phase modulations
discussed in Chapter 7. Be sure to record the amount of time it takes
to compute the results.
5. The MATLAB program tsinc su.m, allows you to change the number of
filter coecients used in the wavelet pairs. Use the program to change
the number of filter coecients from 512 to 128. Work the problem
above with this new filter configuration and note the amount of time
needed.
6. (a) For the 512 filter coecients, calculate the group delay of the filter.
(b) Since the number of filter coecients is the same for all wavelet pairs,
the group delay of each layer is also the same. If the ADC sampling
rate is fs = 7,000 Hz calculate how many layers can be computed if the
results must be calculated in real time (time to gather the input signal
record equals the time to process the signal record).
7. To help identify the capability of the quadrature mirror filter bank
analysis as a tool for identifying the LPI modulation, extracting the
modulation parameters, and to aid in deciding on what signal process-
ing algorithm performs best, construct a table to show the quadrature
mirror filter bank measurement results for the LPI signals contained in
the test signals folder. For each parameter of interest, show the actual
value, the measured value, and the absolute value of the relative error
[see (12.26)]. Be sure to include the layer that is used for identifying
the dierent parameter values being measured.
Chapter 15
Cyclostationary Spectral
Analysis for Detection of
LPI Radar Parameters
The Wigner-Ville distribution (Chapter 12), the Choi-Williams distribution
(Chapter 13) and the quadrature mirror filter bank processing (Chapter 14)
together give time-frequency results that allow the LPI parameters to be de-
termined with good accuracy. This chapter presents an additional bifrequency
spectral analysis technique, known as cyclostationary processing, that oers
some additional capability in the detection and classification of LPI modu-
lations. Instead of examining the LPI signals in the time-frequency domain,
cyclostationary processing transforms the signal into the frequency-cycle fre-
quency domain. Two ecient methods for computing the cyclostationary
spectrum are presented: the time-smoothing FFT accumulation method, and
the direct frequency-smoothing method. The cyclostationary signal process-
ing is then used to extract the parameters from the LPI radar waveforms
discussed in Part I. The folder titled CYCLO contains the signal processing
MATLAB files that allow the user to re-create any of the results presented,
as well as new results of interest.
15.1 Introduction
Since the development of the theoretical concepts by Gardner in the early
1980s, much work has been carried out to investigate the potential of cyclo-
stationary spectral analysis for many signal processing tasks. Cyclostationary
processing has been investigated for use in the detection and identification of
513
weak spread-spectrum communication signals [13] and SATCOM signals [4].

Cyclostationary signal models have also been used in array signal processing
for estimating the direction of arrival of multiple narrowband signals [5], as
well as building adaptive arrays [6, 7]. The estimation and detection of radar
signal parameters have been investigated [810]. In this chapter, a thorough
treatment of this application is presented.
Cyclostationary spectral analysis is based on modeling the signal as a
cyclostationary process rather than a stationary process. A signal is cyclo-
stationary of order n if and only if one can find some nth order nonlinear
transformation of the signal that will generate finite-strength additive sine
wave components that result in spectral lines [11]. For example, an n = 2
or quadratic transformation (like the product of the signal with a delayed
version of itself, often used to detect BPSK signals) will generate spectral
lines. That is, a signal x(t) is cyclostationary with cycle frequency , if and
only if at least some of its delay product waveforms, z(t) = x(t )x (t)
for some delays , exhibit a spectral line at frequency , and if and only if
the time fluctuations in at least some pairs of spectral bands of x(t), whose
two center frequencies sum to , are correlated. In contrast, for stationary
signals, only a spectral line at frequency zero can be generated. For signals
with periodic features (e.g., LPI radar signals), the advantage of using a cy-
clostationary model is that nonzero correlation is exhibited between certain
frequency components when their frequency separation is related to the pe-
riodicity of interest. Applications that use cyclostationary spectral analysis
include time dierence of arrival estimation, signal detection, identification,
and parameter estimation. Many useful characteristics of LPI radar signals
can be determined, and are reflected in the cyclic autocorrelation function
and the spectral correlation density, which form the basis for cyclic spectral
analysis. These concepts are discussed below.
15.1.1 Cyclic Autocorrelation

To discuss the cyclic autocorrelation, we begin with the definition of the
correlation integral. The correlation integral is defined as
8
Rc (x) = f (u)g(x + u)du (15.1)

Applying the FFT to both sides gives

F {Rc (x)} = F (s)G (s) (15.2)
If f (x) and g(x) are the same function, the integral above is called the auto-
correlation function and cross correlation if they dier. The autocorrelation
function is a quadratic transformation of the signal, and may be interpreted
as a measure of the predictability of the signal at time t + based on knowl-
edge of the signal at time t. When considering a time series of length T , the
Cyclostationary Spectral Analysis for Detection of LPI Parameters 515
autocorrelation function becomes the time-average autocorrelation function

given by
8
1 T /2 p Q p Q
Rx ( )=
lim x t+ x t dt (15.3)
T T T /2 2 2
The cyclic autocorrelation of a complex-valued time series x(t) is then defined
by [12]
8 p
1 T /2
Q p Q j2t
Rx ( )=
lim x t+ x t e dt (15.4)
T T T /2 2 2
and can be interpreted as the Fourier coecient of any additive sine wave
component with frequency that might be contained in the delay product
(quadratic transformation) of x(t). The nonzero correlation (second-order
periodicity) characteristic of a time series x(t) exists in the time domain, if
the cyclic autocorrelation function is not identically zero. That is, the signal
x(t) is said to be cyclostationary if Rx ( ) does not equal zero at some time
delay (any real number) and cycle frequency = 0.
15.1.2 Spectral Correlation Density

Recall that the power spectral density is defined as the Fourier transform of
the autocorrelation function
8
Sx (f ) = Rx ( )ej2f d (15.5)

In the same manner, the spectral correlation density (SCD), or cyclic spectral
density, is obtained from the Fourier transform of the cyclic autocorrelation
function (15.4) as [12]
8 p
1 Q p Q
Sx (f )=
Rx ( )ei2f d = lim XT f + XT f (15.6)
T T 2 2
where is the cycle frequency and

8 T /2
XT (f )=
x(u)ej2f u du (15.7)
T /2
which is the Fourier transform of the time domain signal x(u). The additional
variable (cycle frequency) leads to a two-dimensional representation Sx (f );
namely, the bifrequency plane (f , ) [12]. Measurement of (15.4) and (15.6)
in signal analysis constitutes what is referred to as cyclic spectral analysis.
Good insight is gained if we examine a second-order cyclostationary process
and compare the time-domain implementation and the frequency-domain im-
plementation. In Figure 15.1 it is shown that the time-domain implementa-
Figure 15.1: Time-domain implementation of a second-order cyclostationary

process.
tion consists of a delay (i ) and multiply operation followed by the multiplica-

tion by the exponential cycle frequency term. The expected value then gives
the cyclic autocorrelation function Rx ( ) and the subsequent FFT gives the
spectral correlation density Sx (f ). With this perspective, it is easy to see
that if the signal x(t) contains a periodic component and the delay is chosen
properly, a strong sinusoid will be present at the output.
A frequency-domain implementation of a second-order cyclostationary
process is shown in Figure 15.2. The input signal x(t) with spectral represen-
tation X(), is split into two channels and multiplied by the two exponential
factors that are a function of the cycle frequency and are complex conjugates
of each other. The time-domain output signals are u(t) and s(t) which have
spectral representations of U () and S() respectively. This time-domain
multiplication results in a spectral shift of u(t) by /2 and a spectral shift
of s(t) by /2. Figure 15.3 shows the spectral representations X(), U (),
and S() and illustrates the narrowband spectral components of x(t) being
aligned at = f . Both u(t) and s(t) are filtered with a bandpass filter with
bandwidth B and center frequency f . Note that this captures the narrowband
spectral components of x(t) centered at f + /2 and f /2. The Fourier
transform is taken of both filter outputs and then the correlation of the two
spectrums is computed. The expected value of the correlation output is then
the spectral correlation density function Sx (f ).
15.2 Spectral Correlation Density Estimation

Estimates of the cyclic spectral density or SCD can be obtained via time or
frequency-smoothing techniques. Since the signals being analyzed are defined
over a finite time interval t, the cyclic spectral density is only an estimate.
An estimate of the SCD can be obtained by the time-smoothed cyclic peri-
odogram given by
8 t+(t/2)
1
Sx (f ) SxT (t, f )t = SxTW (u, f )du (15.8)
W t t(t/2)
Figure 15.2: Frequency-domain implementation of a second-order cyclosta-

tionary process.
Figure 15.3: Frequency-domain representation of (a) x(t) [X()], (b) modu-

lation of x(t) by /2 [U ()], and (c) modulation of x(t) by /2 [S()].
Figure 15.4: Cyclic spectral density estimation using short-time FFTs.
where p
1 Q p Q
SxTW (u, f ) = XTW u, f + XTW u, f (15.9)
TW 2 2
with t being the total observation time of the signal, TW is the short-time
FFT window length, and
8 t+(TW /2)
XTW (t, f ) = x(u)ej2f u du (15.10)
t(TW /2)
is the sliding short-time Fourier transform. Figure 15.4 shows the SCD es-
timation graphically for a signal x(t). Here the frequency components are
evaluated over a small time window TW (sliding FFT time length), along
the entire observation time interval t. The spectral components generated
by each short-time Fourier transform have a resolution, f = 1/TW . In
Figure 15.4, L is the overlap (sliding) factor between each short-time FFT.
In order to avoid aliasing and cycle leakage on the estimates, the value of L
is defined as L TW /4 [13].
Figure 15.5 shows the spectral components of each short-time FFT be-
ing multiplied according to (15.9), providing the same resolution capability
f = 1/TW , for the cyclic spectrum estimates [4, 14]. Note that the dummy
variable t has been replaced by the specific time instances t1 . . . tp . Within
each window (TW ), two frequency components centered about some f0 and
separated by some 0 are multiplied together, and the resulting sequence of
products is then integrated over the total time (t), as shown in (15.8).
Figure 15.5: Sequence of frequency products for each short-time Fourier trans-
form.
The estimation Sx (f ) SxT (t, f )t can be made as reliable and accu-

W
rate as desired for any given t and f , and for all f by making t suciently
large, provided that (15.4) exists within the interval t and that a substantial
amount of smoothing is carried out over t. This leads to the Grenanders
uncertainty condition tf 1 [14]. This uncertainty condition means that
the observation time (t) must greatly exceed the time window (TW ) that
is used to compute the spectral components. A data taper window is also
used to minimize the eects of cycle and spectral leakage (estimation noise),
introduced by frequency component side lobes [14]. The spectral components
obtained from the short-time FFT have a resolution of
1
f = (15.11)
TW
The cycle frequency resolution of the estimate is related to the total observa-
tion time by
1
= (15.12)
t
The estimation of some (f0 , 0 ) represents a very small area on the bifre-
quency plane, as shown in Figure 15.6, and, since one needs a significant
number of estimates to represent the cyclic spectrum adequately, it follows
that obtaining estimates becomes very computationally demanding, and e-
cient algorithms are required [15].
Figure 15.6: Frequency and cycle frequency resolutions on the bifrequency

plane (adapted from [3, 15]).
15.3 Discrete Time Cyclostationary

Algorithms
Cyclostationary signal processing can be used to extract the parameters from
the sampled LPI signals in an intercept receiver, when moderate to large
amounts of additive noise are present. With the signal displayed on the bifre-
quency plane (frequency-cycle frequency) the intercept receiver or operator
can examine and compare the modulation characteristics, using several al-
gorithms that estimate the SCD. Computationally ecient algorithms for
implementation of time- and frequency-smoothing techniques are discussed
in [16]. These are the FFT accumulation method (FAM), a time-smoothing
algorithm, and the direct frequency-smoothing method (DFSM), a frequency-
smoothing algorithm, as described below. The temporal and spectral smooth-
ing equivalence is also addressed in [9].
15.3.1 The Time-Smoothing FFT Accumulation Method

The time-smoothing FFT accumulation method was developed to reduce the
number of computations required to estimate the cyclic spectrum [3]. This
technique divides the bifrequency plane into smaller regions called channel
pair regions, and computes the estimates one block at a time using the fast
Fourier transform. Describing the estimated time-smoothed periodogram
Figure 15.7: Block diagram of the FAM (adapted from [3, 11]).
from (15.8) and (15.9), in discrete terms, yields

N 1 } p ]
1 3 1 Q p Q
SX (n, k) = XN I n, k + XN I n, k (15.13)
NI N n=0 N 2 2
where I
N
3 1
I
XN I (n, k) =
w(n)x(n)e(j2kn)/N (15.14)
n=0
is the discrete Fourier transform of x(n), w(n) is the data taper window (e.g.,
Hamming window), and the discrete equivalents of f and are k and ,
respectively. A block diagram of the FFT accumulation method is shown in
Figure 15.7.
The algorithm consists of three basic stages: computation of the complex
demodulates (divided into data tapering, sliding N point Fourier transform,
and baseband frequency translation sections), computation of the product
sequences, and smoothing of the product sequences. Table 15.1 shows the
relationship between the variables in (15.8), (15.9), and (15.13). The parame-
ter N represents the total number of discrete samples within the observation
time, and N represents the number of points within the discrete short-time
(sliding) FFT. In the FAM algorithm, spectral components of a sequence,
x(n), are computed using (15.14). Two components are multiplied (15.13) to
provide a sample of a cyclic spectrum estimate representing the finite channel
pair region on the bifrequency plane, as shown in Figure 15.8. There are N 2
Table 15.1: Comparison of the Estimated Time-Smoothed Periodogram

Expressed in Continuous and Discrete Time
Name Continuous Time Discrete Time

SCD SX (t, f )t SX I
(n, k)N
TW N
Short FFT size TW NI
Observation time t N
Time t n
Frequency f k
Cycle frequency
Grenanders Uncertainty M = (f /) ( 1 M = (N/N I ) ( 1
Condition
channel pair regions in the bifrequency plane. Note the 16 small channel pair
regions corresponding to a value of M = 4 in Figure 15.8.
A sequence of samples for any particular area may be obtained by multi-
plying the same two components of a series of consecutive short-time sliding
FFTs along the entire length of the input sequence. After the channelization
performed by an N -point FFT sliding over the data with an overlap of L
samples, the outputs of the FFTs are shifted in frequency in order to obtain
the complex demodulate sequences (see Figure 15.7) [4]. Instead of comput-
ing an average of the product of sequences between the complex demodulates,
as in (15.8), they are Fourier-transformed with a P -point (second) FFT. The
computational eciency of the algorithm is improved by a factor of L, since
only N/L samples are processed for each point estimate. With fs the sam-
pling frequency, the cycle frequency resolution of the decimated algorithm is
defined as res = fs /N (compare to = 1/t), the frequency resolution
is kres = fs /N (compare to f = 1/TW ), and the Grenanders Uncertainty
Condition is M = N/N 1 (compare to tf 1).
Figure 15.9 reveals that the estimates toward the top and the bottom
(shaded areas) of the channel pair region do not satisfy the Uncertainty Con-
dition. In order to minimize the variability of these point estimates, we can
retain only those cyclic spectrum components that are within = kres /2
from the center of the channel pair region [15]. A solution to resolve the
entire area of the channel pair region without leaving gaps is to apply a data
taper window on the frequency axis (such as a Hamming window), to obtain
better coverage.
15.3.2 Direct Frequency-Smoothing Method

Direct frequency-smoothing algorithms first compute the spectral components
of the data, and then execute spectral-correlation operations directly on the
spectral components. Generally, the direct frequency-smoothing method is
Figure 15.8: Channel pair regions within the bifrequency plane (Adapted
from [3, 15]).
Figure 15.9: Cycle frequency and frequency resolutions of the Grenanders

Uncertainty Condition (adapted from [3, 15]).
Figure 15.10: Block diagram of the direct frequency-smoothing algorithm

(adapted from [3, 11]).
computationally superior to indirect algorithms that use related quantities

such as the Wigner-Ville Distribution, but DFSM is normally less ecient
than a time-smoothing approach [13].
The basis for the DFSM is the discrete time frequency-smoothed cyclic
periodogram represented by
N 1
1 3 p Q p Q

SX (n, k)k = XN n, k + XN n, k (15.15)
N
N n=0 2 2
where
N1
3
XN (n, k) =
w(n)x(n)e(j2kn)/N (15.16)
n=0
is the discrete Fourier transform of x(n), w(n) is the rectangular window

of length N that is the total number of points of the FFT related to the
total observation time, t, is the cycle frequency discrete equivalent, the
frequency-smoothed ranges over the interval |m| M/2, and k M fs /N
is the frequency resolution discrete equivalent [9].
The block diagram in Figure 15.10 illustrates the implementation of the
DFSM. In order to provide full coverage of the bifrequency plane with mini-
mal computational expense, (15.15) is computed along a line of constant cycle
frequency, thus spacing the point estimates by k = M fs /N . This method
is easier to implement, and is generally used to validate the time-smoothing
approach, but may become more computationally demanding. This is espe-

cially true in the last block in which the complex demodulate product se-
quences are summed. Considerations on the parallel processing of both time
and frequency algorithms are discussed in [11]. Finally, we note that combi-
nations of both time-smoothing and frequency-smoothing methods may also
be advantageous for certain applications.
15.4 Test Signals

In order to grasp a good understanding of how the signals appear on the
bifrequency plane, this section examines several test signals used in previ-
ous chapters. The first test signal examined is a tone composed of a sin-
gle carrier frequency with fc =1 kHz, and is sampled with sampling fre-
quency fs = 7 kHz. The time-smoothing technique to estimate the SCD
is demonstrated first using the real part of the input signal. Figure 15.11
shows the time-smoothing SCD results. Figure 15.11(a) shows the bifrequency
plane, and reveals that the signals frequency shows up at four separate loca-
tions. The (, k) frequency pairs are (2fc , 0), (0, fc ), (0, fc ), and (2fc , 0).
Figure 15.11(b) details a close-up of the time-smoothing estimation charac-
teristics for the signal outlined in the box in Figure 15.11(a). For these results
the frequency resolution is k = 128 Hz. With the Grenanders uncertainty
value of M = 2, the cycle frequency = 64 Hz. The overlap parameter
is fixed at L = 4. The number of points in the first FFT N is the next
largest power of 2 value of fs /k or N = 64. The number of points in the
second FFT P is the next largest power of 2 value of 4fs /N or P = 8. The
total number of signal samples integrated into the SCD are N = P L = 128.
Note that in Figure 15.11(b), the = 2fc cycle frequency position lies at the
resolved signals centroid.
The frequency-smoothing SCD results for the single-tone signal are shown
in Figure 15.12. Figure 15.12(a) shows the bifrequency plane, and
Figure 15.12(b) details a close-up of the frequency-smoothing estimation char-
acteristics. The results serve to demonstrate the dierences between the time-
smoothing and frequency-smoothing techniques for estimating the SCD. For
the frequency-smoothing results, k = 128 Hz. The number of samples in-
tegrated into the FFT is the next largest power of 2 value of fs M/f = 109
or N = 128. Note in Figure 15.12(b), the = 2fc cycle frequency position
does not lie at the resolved signals centroid.
Next, the time-smoothing technique is used to estimate the SCD of a
two-tone signal (fc1 = 1 kHz and fc2 = 2 kHz). Figure 15.13 shows the
time-smoothing SCD results. Figure 15.13(a) shows the bifrequency plane,
and reveals that the two tones show up in the four separate quadrants, along
with the cross terms. Figure 15.13(b) details a close-up of the time-smoothing
estimation characteristics for the signal outlined in the box in Figure 15.13(a).
Figure 15.11: Time-smoothing SCD for a single frequency fc = 1 kHz tone,

showing the (a) bifrequency plane, and (b) close-up of the time-smoothing
estimation characteristics.
Figure 15.12: Frequency-smoothing SCD for a single frequency fc = 1 kHz

tone, showing the (a) bifrequency plane, and (b) close-up of the frequency-
smoothing estimation characteristics.
For these results, the frequency resolution k = 128 Hz, the Grenanders
uncertainty value M = 2, and the cycle frequency = 64 Hz. Also, N , N ,
and P are the same as for the single-tone signal.
The frequency-smoothing SCD results for the two-tone signal are shown in
Figure 15.14. Figure 15.14(a) shows the bifrequency plane, and
Figure 15.14(b) details a close-up of the frequency-smoothing estimation char-
acteristics, including the cross terms. As for the single-tone results, k = 128
Hz and N = 128. Note that in Figure 15.14(b), the = 2fci cycle frequency
positions do not lie at the signal centroids.
15.5 BPSK Analysis

In this section, two Nc = 7-bit binary phase shift (BPSK) signals are used
to present the method of measuring and determining the BPSK signal pa-
rameters in the bifrequency plane. Both a wideband modulation (cycles per
subcode period cpp = 1) and a narrowband modulation (cpp = 5) are used to
illustrate the technique. The ability to extract the BPSK parameters under
various SNR conditions is also quantified. For these results, the frequency
resolution k = X and the Grenanders uncertainty value M = 2 (cycle
frequency = X Hz).
Figure 15.15 shows the narrowband BPSK signal. In Figure 15.15(a), the
complete SCD bifrequency plane is shown. Note that the cycle frequency
extends from fs to fs and the frequency extends from fs /2 to fs /2 (see
also Figure 15.6). The BPSK modulation shows up in the four quadrants
centered on = 2fc = 2 kHz. A closer look at the boxed section in Figure
15.15(a) is shown in Figure 15.15(b). The most important parameters of the
BPSK signal can be identified clearly. These are the bandwidth B = 1/tb ,
the code rate Rc = 1/tb Nc , the subcode period tb , and the number of bits
N used in the Barker code. The bandwidth can be measured in both the
frequency dimension and the cycle frequency dimension. The measurement
in the cycle frequency dimension is the width from the centroid (C) to the
edge of the pattern, where the SCD peaks on the bifrequency plane start to
fall o in amplitude. The spots to exclude in the calculation have a lower
intensity. This is especially noticeable in the k dimension. The bandwidth
is measured as B = 1 kHz, giving a subcode period of 1 ms. The code rate
Rc is measured in the cycle frequency domain, and is the width between any
two spots within the BPSK modulation pattern. Here, Rc = 142.8 Hz. The
number of Barker bits is then Nc = B/Rc =7 bits.
At first, the sensitivity of the extracted parameter values to the measure-
ment of the bandwidth might seem critical. This is not true, however, since
the number of bits N for a BPSK signal can only take on a select set of values.
For example, if the next set of spots was included in the calculation, and the
bandwidth was measured to be B = 1,142.8, then N 8, which we know is
Figure 15.13: Time-smoothing SCD for a two-tone signal (fc1 = 1 kHz, fc2 =
2 kHz), showing (a) the bifrequency plane, and (b) a close-up of the time-
smoothing estimation characteristics.
Figure 15.14: Frequency-smoothing SCD for a two-tone signal (fc1 = 1 kHz,

fc2 = 2 kHz) showing (a) the bifrequency plane, and (b) a close-up of the
frequency-smoothing estimation characteristics.
not correct. Therefore, too many spots were included in the measurement.
Figure 15.16 shows a contour plot that illustrates the BPSK pattern and
measurement technique when the SNR = 0 dB. The white Gaussian noise
added to the signal is distributed over a bandwidth equal to fs /2. The com-
plete bifrequency plane is shown in Figure 15.16(a), and the boxed region is
examined closer in Figure 15.16(b). Note that the noise is suppressed sig-
nificantly. This is due to the noise being uncorrelated. The centroid (C) is
identified, and the bandwidth is measured out to the row of spots just be-
fore the crosshatch ends. The bandwidth in the frequency k dimension is
also shown. The bandwidth is measured as shown, and extends between the
highest and lowest corners of the crosshatch region. Note that since the noise
has enhanced the details of the crosshatch, the measurement is easier to take
in the k dimension. The code rate is also easily measured between adjacent
spots as shown in Figure 15.17.
The next signal examined is the narrowband 7-bit BPSK (cpp = 5). Since
the carrier frequency is fc = 1 kHz, the subcode period tb = 5 ms, resulting
in a bandwidth of B = 200 Hz. The frequency-smoothing SCD is shown in
Figure 15.18. Figure 15.19 illustrates the extraction of the narrowband BPSK
signal parameters when noise is present. The bandwidth B is measured in
the same manner as is the code rate Rc . This important example shows
that significant amounts of noise in the bifrequency plane can still give good
results.
15.6 FMCW Analysis

In this section, a method is presented for extracting the parameters from a
triangular FMCW radar signal. The signal shows up at four positions within
the bifrequency plane, as illustrated in Figure 15.20 and Figure 15.21. We
first examine a F = 250 Hz, fc = 1 kHz FMCW signal using frequency-
smoothing. For these results, the frequency resolution k = 16 Hz and
the Grenanders uncertainty value M = 4 (cycle frequency = 4 Hz). The
frequency-smoothing SCD for one of the four positions is illustrated in Figure
15.20, and is a unique pattern for the FMCW modulation. For this result,
N = 1,024 and k = 16 Hz.
The modulation centroid lies at a cycle frequency of = 2 kHz, indicating
that the fc = 1 kHz. Note that the centroid lies to the right of the spot
(as is the case for the frequency-smoothing SCD results). To determine the
modulation bandwidth F from the SCD, the width from the centroid out
to the last large set of spots is measured on the cycle frequency axis. Note
also that F can also be determined from the frequency axis by measuring
the total extent of the modulation as shown. The modulation period tm is
determined by measuring Rc in the cycle frequency domain. In the SCD,
Rc = 1/2tm for the FMCW signal. From Figure 15.20, Rc = 25 Hz, giving
Figure 15.15: Frequency-smoothing SCD patterns for an Nc = 7-bit BPSK

signal, with fc = 1 kHz and cpp = 1, showing (a) the complete bifrequency
plane, and (b) a close-up illustrating the method of parameter measurements.

signal with fc = 1 kHz, cpp = 1, and SNR = 0 dB, showing (a) the complete
bifrequency plane, and (b) a close-up illustrating the method of parameter
measurements.

signal with fc = 1 kHz, cpp = 1, and SNR = 6 dB, showing (a) a partial
bifrequency plane, and (b) a close-up illustrating the method of parameter
measurements.
Figure 15.18: Frequency-smoothing SCD patterns for an Nc = 7-bit narrow-

band BPSK signal with fc = 1 kHz and cpp = 5, showing the bifrequency
plane illustrating the method of parameter measurements.
tm = 20 ms.
Figure 15.21 illustrates the extraction of the FMCW parameters when
noise is present. In Figure 15.21(a) the SNR = 0 dB. Note that the pattern
is still recognizable as being unique to the FMCW waveform. The noise
present actually aids in identifying the centroid of the modulation. The
modulation bandwidth F is measured in the same manner as shown in
Figure 15.20 also with good fidelity. The measurement of Rc is also easily
made. In Figure 15.21(b) the SNR = 6 dB. With this level of noise, a sig-
nificant degradation in the contour image results, and makes the modulation
bandwidth dicult to measure in the cycle frequency dimension. However,
the F measurement can still be easily made, with only a slight bit of er-
ror. Here F = 240 Hz. The Rc value, however, can still be made with a
good degree of accuracy. The extraction of the FMCW parameters from a
wideband signal F = 500 is given in Appendix N.

In this section we look at the bifrequency plane results for the polyphase
codes, and demonstrate the bifrequency extraction techniques. We begin
with the Frank code example. The analysis of the wideband Frank signal
Figure 15.19: Frequency-smoothing SCD patterns for an Nc = 7-bit narrow-

band BPSK signal with fc = 1 kHz and cpp = 5, showing (a) the bifrequency
plane with SNR = 0 dB, and (b) SNR = 6 dB.
Figure 15.20: Frequency-smoothing SCD patterns for a F = 250 Hz, tm =

20-ms triangular FMCW signal with fc = 1 kHz.
with a long code period using the time-smoothing SCD is presented. For these
results, the frequency resolution, k = 16 Hz, and Grenanders uncertainty
M = 4. With a longer code period T = 64 ms (Nc = 64 subcodes, and
cpp = 1), the Frank code signal converges to a more well-defined insect shape
on the bifrequency plane as shown in Figure 15.22(a). Interestingly enough,
all the longer phase codes derived from linear FM waveforms have this type
of shape, using the time-smoothing SCD technique. Figure 15.22(b) shows a
close-up of one of the four modulation patterns. Note the position of the head,
abdomen, and wings that provide a convenient reference for measurements
of the signal parameters. Also note that the insect points to the right. The
direction of the insect is important to help distinguish between the dierent
phase codes. The centroid (c) is symmetrically located within the pattern
characteristic of the time-smoothing SCD. The bandwidth can be measured
as the width from the centroid to the head, on the cycle frequency axis.
A correlation can also be made using the bandwidth measurement on the
frequency axis and is the width between the wing tips.
Also indicated in Figure 15.22(b) is a box that is examined in closer
detail to illustrate the Rc measurement. Figures 15.23(a, b) illustrate the
measurement of Rc = 1/T . Figure 15.23(b) indicates Rc = 15.5 Hz, giv-
ing a modulation period of 64 ms. Since the number of subcodes used by
LPI radar are most often a power of 2 (e.g., 64 = 26 ), an accurate result
Figure 15.21: Frequency-smoothing SCD patterns for a F = 250 Hz, tm =

20-ms triangular FMCW signal with fc = 1 kHz, showing (a) the bifrequency
plane SNR = 0 dB, and (b) SNR = 6 dB.
Figure 15.22: Time-smoothing SCD insect patterns for the Frank code with
Nc = 64, fc = 1 kHz, and cpp = 1, with (a) the complete bifrequency
plane, and (b) a closer examination of one of the four modulation patterns
illustrating the bandwidth measurements.
can be obtained, even from bifrequency planes with small SNR with Nc =
B/Rc = 64. Appendix O examines the Frank code with a shorter code pe-
riod. The cyclostationary results for the P1, P2, P3, and P4 codes are given in
Appendix P.

In this section the frequency-smoothing SCD is used to examine the poly-
time codes. We begin with the T1(2) code. The T1(2) code has an fc = 1
kHz, and has a time-modulated binary phase shift (of various widths). Fig-
ure 15.24(a) shows the bifrequency plane and the four modulation patterns.
Figure 15.24(b) shows one of the four unique patterns, and illustrates how
the bandwidth of the signal can be measured. Recall that the bandwidth
B is measuring the largest excursion in frequency, due to the shortest time
phase code. For the case shown in Figure 15.24(b), B = 1,750 Hz, which
agrees with the results as measured by the time-frequency tools earlier. Note
that this value is not the modulation bandwidth F of the linear FM signal
used to derive the polytime phase modulation. Figure 15.25 shows a closer
examination of the bifrequency plane, and the measurement of Rc = 1/T =
62.5 Hz. This gives the estimate for the code period as T = 16 ms. Note
also that an SCD spot exists at ( = 2fc , k = 0). The results for the T2(2),
T3(2), and T4(2) codes are given in Appendix Q.
15.9 Costas Frequency Hopping Results

In this section the time-smoothing SCD is used to determine what SCD prop-
erties a Costas frequency hopping signal takes on. The time-smoothing SCD
was created using = 16 and N = 2,048. The Costas sequence in this
example is S = {4, 7, 1, 6, 5, 2, 3} and is used since this sequence is discussed
in Chapter 6. The sampling frequency of the ADC is 15 kHz, and each Costas
frequency is generated with 20 cycles per frequency. That is, the time spent
at each frequency is not a constant. Figure 15.26(a) shows the complete bifre-
quency results. One of the four quadrants is shown in Figure 15.26(b). The
frequencies within the sequence fci show up at 2fci and are outlined along
the k = 0 axis. Also note the presence of the cross terms k = |fci fcj |/2.
Although the SCD analysis does not let us determine the time sequence of
information, we can, however, see that frequencies = 6 kHz and = 5 kHz
are fired next to each other, as is the case for frequencies = 2 kHz and
= 3 kHz, but we cannot determine the order.
Figure 15.23: Close examination of the time-smoothing SCD for the Frank
code with Nc = 64, fc = 1 kHz, and cpp = 1, with (a) modulation cycles and
(b) the measurement of Rc .
Figure 15.24: Frequency-smoothing SCD patterns for the polytime T1(2)

code with fc = 1 kHz, showing (a) the complete bifrequency plane, and
(b) a closer examination of one of the four modulation patterns illustrating
the bandwidth measurement.
Figure 15.25: Close examination of the frequency-smoothing SCD pattern for

the polytime T1(2) code illustrating the Rc measurement.
15.10 Random Noise Analysis

In the first example, a random noise waveform with a bandwidth of
B = 300 MHz and carrier frequency of fc = 350 MHz was examined with
the time-smoothing SCD. Figure 15.27 shows that the bandwidth and carrier
frequency can easily be identified as expected.
In the second example, a random noise plus FMCW waveform is exam-
ined. The FMCW waveform parameters used to modulate the noise are
F = 300 MHz (from 200 to 500 MHz) and the carrier frequency is
fc = 350 MHz. The noise bandwidth being modulated is B = 300 MHz.
Since in this case the noise bandwidth and the modulation bandwidth over-
lap, the total noise FMCW bandwidth transmitted is B = 300 MHz. The
time-smoothing technique was chosen to estimate the SCD.
From Figure 15.28, the diamond is centered at 1,200 MHz, which is twice
the center frequency of the modulated signal. It might be expected that the
center frequency would appear at 1,400 MHz, or twice the center frequency
of the two modulated signals (350 MHz for the noise and 350 MHz for the
FMCW signal). The dierence of 200 MHz is observed as an oset. Several
noise plus FMCW waveforms were examined with dierent modulation band-
widths as summarized in Table 15.2. The relationship of the observed center
Figure 15.26: Time smoothing SCD patterns for the Costas sequence S =
{4, 7, 1, 6, 5, 2, 3} showing (a) the complete bifrequency plane, and (b) a closer
examination of one of the four modulation patterns illustrating the frequency
cross terms (k).
Figure 15.27: Time smoothing SCD for the random noise waveform showing
the carrier frequency and the bandwidth measurements.
frequency oset as a function of the FMCW modulation bandwidth is left as

an exercise for the reader. Other signal characteristics, such as the signal
bandwidth of 300 MHz, can be measured along the cycle frequency axis, as
expected. The bandwidth can also be measured along the frequency axis as
well. This sweep bandwidth shows nicely in the QMFB results.
15.11 Summary
The cyclostationary signal processing was presented in this chapter, and
several examples were used to demonstrate the bifrequency results. To ex-
tract the unknown signal parameters, the bifrequency plane (frequency-cycle
frequency) is examined to determine directly (and indirectly) parameters
such as the carrier frequency, code rate, bandwidth, and modulation pe-
riod. Information not available includes any parameters that change as a
function of time (such as the signals phase). These phase changes, however,
can be identified from the other signal processing tools that are included
(such as the PWVD, CHOI, and QMFB). Measurement of the LPI signal
parameters agree well with the actual values. With moderate amounts of
noise added, however, the measurement ability using the bifrequency analy-
sis, does not degrade significantly but remains fairly robust since symmetrical
Figure 15.28: Time smoothing SCD for the random noise plus FMCW wave-
form showing the bandwidth measurement and the carrier frequency oset
that appears in the bifrequency domain.
Table 15.2: Summary of Time Smoothing SCD for the Random Noise Plus
FMCW Waveform Showing the Bandwidth Measurement and the Carrier
Frequency Oset That Appears in the Bifrequency Domain
F (MHz) Oset (MHz) Center Frequency Observed

500 250 1,150
300 200 1,200
200 100 1,300
100 50 1,350
1 0 1,400
white Gaussian noise is not correlated and is suppressed in this spectral cor-
relation technique.
The cyclostationary processing, however, does not perform well with the
FSK Costas code and FSK/PSK (binary phase code) signal. This is mainly
due to a lack of temporal information needed, in order to identify the code
sequence in time. The presentation of the cyclostationary results to a trained
operator will allow the signal parameters to be extracted, and can enable
good classification results for the signals that are appropriate. The use of the
cyclostationary processing in noisy environments is particularly good.
References
[1] Spooner, C. M., and Gardner, W. A., Robust feature detection of signal
interception, IEEE Trans. on Communications, Vol. 42, No. 5, pp. 2165
2173, May 1994.
[2] Gardner, W. A., Signal interception: A unifying theoretical framework for
feature detection, IEEE Trans. on Communications, Vol. 36, No. 8, pp.
897906, Aug. 1988.
[3] Gardner, W. A., and Spooner, C. M., Signal interception: Performance
advantages of cyclic feature detectors, IEEE Trans. on Communications,
Vol. 40, No. 1, pp. 149159, Jan. 1992.
[4] Tom, C., Cyclostationary spectral analysis of typical SATCOM signals using
the FFT accumulation method, Defence Research Establishment Report No.
1280, Ottawa, Canada, Dec. 1995.
[5] Xin, J., and Sano, A., Linear prediction approach to direction estimation of
cyclostationary signals in multipath environment, IEEE Trans. on Signal
Processing, Vol. 49, No. 4, pp. 710720, April 2001.
[6] Yu, S-J., and Ueng, F-B., Implementation of cyclostationary signal-based
adaptive arrays, Elsevier Signal Processing, Vol. 80, pp. 22492254, 2000.
[7] Lee, J-H., and Lee, Y-T., A novel direction-finding method for cyclostation-
ary signals, Elsevier Signal Processing, Vol. 81 pp. 13171323, 2001.
[8] Gini, F., Montanari, M., and Verrazzani, L., Estimation of chirp radar
signals in compound-Gaussian clutter: A cyclostationary approach, IEEE
Trans. on Signal Processing, Vol. 48, No. 4 pp. 10291039, April 2000.
[9] Skinner, B. J., Ingels, F. M., and Donohoe, J. P., The eect of radar signal
construction on detectability, Proc. of the 26th Southeastern Symposium on
System Theory, pp. 147150, March 1994.
[10] Gillman, A. M., Non-cooperative detection of LPI/LPD signals via cyclic
spectral analysis, Air Force Institute of Technology, Masters thesis, March
1999.
[11] Gardner, W. A., Statistical Spectral Analysis: A Nonprobabilistic Theory,
Prentice-Hall, Englewood Clis, NJ, 1987.
[12] Gardner, W. A., Exploitation of spectral redundancy in cyclostationary sig-

nals, IEEE Signal Processing Magazine, pp. 1436, April 1991.
[13] Lima, A. F., Jr., Analysis of low probability of intercept radar signals using
cyclostationary processing, Naval Postgraduate School Masters thesis, Sept.
2002.
[14] Roberts, R. S., Brown, W. A., and Loomis, H. H., Jr., A review of digital
spectral correlation analysis: Theory and implementation, Cyclostationarity
in Communications and Signal Processing, IEEE Press, 1994.
[15] Roberts, R. S., Brown, W. A., and Loomis, H. H., Jr., Computationally ef-
ficient algorithms for cyclic spectral analysis, IEEE Signal Processing Mag-
azine, pp. 3849, April 1991.
[16] Brown, W. A., III, and Loomis H. H., Jr., Digital implementations of spectral
correlation analyzers, IEEE Trans. on Signal Processing, Vol. 41, No. 2,
pp. 703720, Feb. 1993.
Problems
1. (a) Using the LPIT, generate the FMCW waveform (signal only, SNR =
0 dB, and SNR = 6 dB) with a carrier frequency of 2 kHz,
tm = 5 ms, and F = 500 Hz. (b) Process the signals with both
the time-smoothing and frequency-smoothing algorithm. (c) For each
useful algorithm, diagram your estimates of all the signal parameters
for signal only, SNR = 0 dB, and SNR = 6 dB.
2. (a) Using the LPIT, generate a P4 waveform (signal only,
SNR = 0 dB, and SNR = 6 dB) with a carrier frequency of 2 kHz,
Nc = 128, tb = 1 ms, and fs = 7,000 Hz. (b) Process the signals with
both the time-smoothing and frequency-smoothing algorithm. (c) For
each useful algorithm, diagram your estimates of all the signal parame-
ters for signal only, SNR = 0 dB, and SNR = 6 dB.
such that they both have the same size (be sure to at least include 1
to 2 code periods of each signal). (b) Add the two signals together
and save as a new signal (e.g., fmcw p4.mat). (c) For each signal,
use the cyclostationary processing to analyze and extract the waveform
parameters that are evident. (d) Repeat (b) and (c) for SNR = 0 dB.
(e) Repeat (b) and (c) for SNR = 6 dB.
4. To help identify the capability of the cyclostationary signal processing

as a tool for identifying the LPI modulation, extracting the modulation
parameters, and to aid in deciding on what signal processing algorithm
performs best, construct a table to show the bifrequency measurement
results for the LPI signals contained in the test signals folder. For each
parameter of interest, show the actual value, the measured value, and
the absolute value of the relative error [see (12.26)]. Be sure to include
which frequency axis (frequency, cycle frequency) is used for identifying
the dierent parameter values being measured.
Chapter 16
Antiradiation Missiles
A Wild Weasel provides a Warsaw Pact SAM operator the maximum

opportunity to give his life for his country Relic quote of the Wild Weasel
This chapter gives a brief account of the concept of suppression of enemy

air defenses. The beginnings of SEAD and the development of antiradiation
missiles (ARMs) are presented. The use of ARMs in Vietnam and post-
Vietnam is also presented. The design of ARM seekers is addressed and
the concept of dual-mode ARMs is discussed including ARM performance
metrics. The important ARMs from around the world are then reviewed and
their performance given. Anti-ARM techniques that can be employed other
than the use of LPI emitters are also presented.
16.1 Suppression of Enemy Air Defense

Suppression of enemy air defense (SEAD) is defined by the U.S. Depart-
ment of Defense as that activity that neutralizes, destroys or temporarily
degrades surface-based enemy air defenses by destructive and/or disruptive
means [1]. SEAD includes the use of air-to-ground missiles against an en-
emys integrated air defense system (IADS). An IADS is an integration of air-
and ground-based sensors and the communication that links them together
551
with the weapons and command and control. SEAD capabilities fall within
the traditional discipline of electronic warfare (EW), which includes electronic
attack (EA), electronic protection (EP), and electronic warfare support (ES).
SEAD is also an interdisciplinary construct that integrates EW as an activity
of information operations with the use of EA capabilities such as antiradiation
missiles (ARMs) against the enemys IADS in an eort to obtain information
superiority. In suppression, the ARMs are fired and home in on the enemys
surface-based radar systems that are used to target their surface-to-air mis-
siles (SAMs) against any incoming strike aircraft. In modern network-enabled
warfare, there is a dedicated aircraft assigned that specializes in the hard-kill
of enemy guidance radars by deploying ARMs [2]. The ARM relies on pas-
sive detection of the radiation emitted from the radar. The ARM mission or
sortie is an aircraft strike capability against radar directed/radar controlled
missile and gun system sitesthe greatest threat to eective air operations.
ARMs can also be fired preemptively in order to prevent the SAM radar from
coming up. The use of ARMs contributes to information superiority by pre-
venting and reducing the enemys use of the electromagnetic spectrum while
protecting our own spectrum vulnerabilities. That is, SEAD actions increase
an air forces ability to conduct air operations by reducing their vulnerability
to air defense missiles and guns.
Below, the U.S. Armys description of the various forms of SEAD are given
highlighting their potential use in suppression of an enemys IADS [3].
Campaign SEAD: SEAD operations that are preplanned, theaterwide
eorts conducted concurrently over an extended period against air de-
fense systems normally located well behind enemy lines.
Complementary SEAD: Those operations that involve continuously seek-
ing enemy air defense system targets to destroy them.
Localized SEAD: Those operations that support tactical air operations,
Army aviation operations, reconnaissance, and the establishment of cor-
ridors for ingress and egress routing for air force and army assets.
Joint SEAD: Broad term that includes all suppression of enemy air de-
fense activities provided by one component of the joint force in support
of another.
Nonlethal SEAD: Aims to neutralize or degrade enemy IADS rather
than destroy them. While nonlethal SEAD is most commonly associ-
ated with the electronic jamming of IADS sensors and command, control
and communications (C3) links, this is not the only form of nonlethal
SEAD. Other forms of nonlethal SEAD include the use of specialized
tactics to exploit known air defense system limitations and the use of
stealth technology, or false targets, to deceive enemy air defenses. Note
Antiradiation Missiles 553
that: the threat of destruction alone may degrade an air defense sys-
tems eectiveness by forcing its operators to employ defensive measures
that would result in suboptimal system performance.
Lethal SEAD: Measures taken to physically destroy one or more com-
ponents of an IADS. Most hard-kill SEAD options involve specialized
weapons such as ARMs, precision guided munitions (PGMs), and stand-
o weapons (SOWs); the successful use of which will damage the enemys
IADS and possibly inflict casualties among their crews. Lethal SEAD
can suppress enemy air defenses for a potentially longer period than
can nonlethal measures. However, the success of lethal SEAD missions
depends on adequate EW support to provide both accurate targeting
and protection of the SEAD platform.
Preemptive SEAD: This concept diers from lethal-SEAD practices
by preemptively disrupting enemy air defenses before they can engage
friendly aircraft. This is typically done by firing an ARM in the air
in the direction of a SAM that is suspected to exist but which has not
turned on its radar in the aim of preventing the SAM radar from coming
up. Although an eective and necessary tactic, it is not ecient.
16.1.1 The Beginning of SEAD

Since the introduction of radar in World War II (WWII), radar was used
by both sides to alert ground-based air defenses and fighters of an impend-
ing raid. The SEAD role originated in WWII when the German Luftwae
bombed elements of the British CHAIN HOME early warning radar net dur-
ing the Battle of Britain. This first SEAD attempt heralded a series of
Allied SEAD developments made in response to the very eective German
radar-based air defenses encountered during the Allied bombing oensive of
Germany [4].
One of the first SEAD measures employed by the Allies was to drop cha,
consisting of small strips of metal foil, from lead bombers or pathfinder air-
craft to disrupt the German radar picture by creating thousands of false
targets. If the location of the radars were discovered, they could be attacked,
generally by bombing from a large formation. The British developed a radar
homing device (dubbed Abdullah) that would locate the enemy radar. The
aircraft that were equipped with the Abdullah were not armed and flew only
with escort fighters. In addition, Ferret aircraft were used to detect the op-
erating frequencies of German radars and radios so they could be jammed by
EA equipment or physically attacked by the Ferret or other aircraft. Similar
SEAD measures were used in the Pacific theater during WWII and again
during the Korean conflict [4].
Figure 16.1: The Corvus antiradiation missile.
Figure 16.2: The Corvus antiradiation missile being loaded.
16.1.2 Early ARM Developments

In 1955, the U.S. Navy had a requirement for a long-range nuclear-armed
heavy stando air-to-surface missile to be employed by carrier-based attack
bombers. In April 1955, the ASM-N-8 Raven project was initiated to develop
such a missile [5]. In the same year, a parallel project for an anti-radar
missile was redefined to cover the Raven requirements, and therefore the
Raven project was canceled and the ASM-N-8 designation transferred to the
Corvus. A development contract for the ASM-N-8 Corvus shown in Figure
16.1 was awarded to Temco in January 1957, and the first flight test of an
XASM-N-8 prototype occurred in July 1959. Figure 16.2 shows the Corvus
missile being loaded with a special purpose cart. The Corvus was designed
as an ARM, and had a passive radar seeker to home in on the emission of

enemy radars at a speed of Mach 0.8 [6]. The Corvus seeker could also home
on nonradiating targets when they were illuminated by a compatible radar in
the launching aircraft. In this mode there was also a data-link between the
missile and the launching aircraft, which could provide mid-course command
guidance until the missiles seeker could detect the radar reflections from the
target [7]. The missile could be launched from high or low altitudes, and
maximum ranges for high-altitude (15 km) launches were 315 km in ARM
mode and 185 km in semiactive homing mode. Corvus was to be armed with
a light-weight W-40 nuclear fission warhead (10 kT yield) [5].
By March 1960, the XASM-N-8 test program had progressed to fully
guided flights, but in July that year the Corvus program was terminated. The
reason was that overall responsibility for long-range nuclear air-to-surface mis-
siles had been transferred to the U.S. Air Force, which regarded the Corvus
as unnecessary [6].
16.1.3 Vietnam
The greatest SEAD advances were made during the Vietnam War in 1965
when the eectiveness of the North Vietnamese IADS caused significant
losses. The initial U.S. response of launching conventional air strikes against
the SAM sites resulted in heavy friendly losses [8]. The Soviet SA-2 SAMs
killed at least 83 aircraft and forced the USAF to increase its SEAD capa-
bility by building the F-100 Super Sabre Wild Weasel (an evolution of the
Ferret aircraft) in 90 days pairing experienced fighter pilots with electronic
warfare ocers from the Strategic Air Command. The F-100 was followed by
the F-105G Wild Weasel and the F-4G Wild Weasel [9].
The Wild Weasels were free-roving hunters that baited SAM sites at point
blank range. The Wild Weasel mission was developed by the U.S. Air Force
in 1965, during the Vietnam War era. Its primary concept was the use of
two-seat aircraft, to counter hostile radar-controlled surface-to-air weapons.
They were able to detect and locate the SAM radars and attack them with
bombs, napalm or rockets. However, attacking air defense systems with short-
range weapons proved to be very hazardous. In 1966, the Weasels received a
stand-o attack capability in the form of the AGM-45 Shrike ARM.
With early ARM technology, when the enemy air defenses shut o their
emitters, the already launched ARM could only fly about without guidance
for a short time until it ran out of fuel and crashed. Shutting down a radar
to evade detection protected the radar from destruction but it did not do
much good for the radar operator. Although suppression was achieved the
most preferred solution in most cases is the destruction of enemy air defense
(DEAD) and the destruction of their command and control in order to reduce
the number of SAM shooters. For DEAD, a precise knowledge of the enemy
location is required.
Other significant SEAD developments of the Vietnam War included the

use of compact EA pods that allowed fighter aircraft to conduct self-defense,
support jamming and escort jamming of enemy radars [9]. Self-defense jam-
mers are relatively low powered jammers that are primarily designed to
counter missiles targeted at the jamming aircraft. Support and escort jam-
mers are somewhat more powerful systems that can also provide protection
to aircraft in the immediate vicinity of the jamming aircraft. This type of
jamming helps reduce the number of SAM shots that are taken. Most combat
aircraft were also fitted with radar warning receivers (RWRs) that allowed the
timely employment of defensive EA and evasive maneuvers to avoid enemy
missiles. A stand-o jamming capability was also developed that allowed spe-
cialist aircraft to suppress enemy air defenses at long range while electronic
reconnaissance aircraft were used to determine the enemys electronic order of
battle and to locate the enemy air defense radars and radios [4]. Note that
stand-o jamming does not make the jammer or other aircraft invisible.
16.1.4 Post Vietnam

SEAD developments continued after the Vietnam War, notably during the
Israeli Operation Peace for Galilee during which remotely piloted vehicles
(RPVs) were used to detect, locate and decoy the Syrian IADSs [8]. The
USAF also established the Wild Weasel School. Developments also continued
during Operation Desert Storm, with the coalition not losing a single strike
aircraft to a radar threat while an armed Wild Weasel was on station [9].
Army AH-64 Apache helicopters were also used in the lethal-SEAD role and
EC-130H Compass Call aircraft were used to jam air defense communications.
The F-4G units were disbanded after 1996 and the USAF retired the last
of its EF-111 radar and communications jamming aircraft in May 1999. This
left a critical hole in USAF capabilities. The USAF then replaced its F-4G
Wild Weasel fleet with the combination of the F-16CJ Falcon and HARM tar-
geting system (HTS) using multirole squadrons to partially fill the gap. The
EF-111s mission was transferred to the United States Navys EA-6B which
requires one pilot and three electronic warfare ocers (EWOs). The U.S. Gen-
eral Accounting Oce (GAO) believed the success of air operations during
Operation Desert Storm depended heavily on SEAD aircraft [10]. Moreover,
the GAO predicted SEAD would continue to be important to air operations,
even those involving stealth aircraft, and criticized the U.S. Department of
Defense (DoD) for reducing its traditional SEAD capability. In 2001 the DoD
released the Joint Airborne Electronic Attack (AEA) Analysis of Alternative
(AOA) that examined the options available for replacing the aging EA-6B
including the F-22, the F/A-18, and the Joint Strike Fighter [11]. The de-
cision was to eventually replace the EA-6B by the carrier-based EA-18G (or
Growler)a multirole two-seat aircraft for jamming (AN/ALQ-99), SEAD
and preemptive SEAD.
Table 16.1: Estimates of Combat Aircraft Losses (from [1])
Conflict Combat Sorties Combat Losses Attrition Rate

World War II 2,498,283 19,030 0.76%
Korea 591,693 1,253 0.2%
Vietnam (AF only) 219,407 1,437 0.65%
Desert Storm (Iraq) 68,150 33 0.04%
Bosnia (Coalition) 30,000 3 0.01%
Kosovo 21,111 2 0.009%
N./S. Watch 268,000 0 0.0%
Iraqi Freedom 20,733 1 0.004%
Today the Army has the primary responsibility for suppressing ground-
based enemy air defense weapons to the limits of observed fire. The USAF
has responsibility from beyond the limits of observed fire out to the range
limits of the Army weapons systems; the Army has secondary responsibility.
Even if the USAF can target or observe, the Army may still have to attack the
target. Beyond the range limits of Army weapons, the USAF is responsible.
Although the U.S. DOD is pursuing a new approach to SEAD, the United
States continues to recognize the important requirement for SEAD [1113].
It should also be noted that ARMs are not only used for suppression
of air defense SAM networks. Other targets include airborne early warning
systems, shipboard radar systems, battlefield surveillance systems and any
other radiating RF sensor that can be intercepted. Although some military
experts question the need for ARMs, countries continue to develop and build
ARMs, including countries new to developing ARMs such as Germany, Tai-
wan, China, and Brazil. This is in response to the double-digit SAM radars
such as the SA-10, SA-11, SA-12, SA-15 and SA-17 that can be integrated
into a formidable IADS [14].
To address future SEAD/ARM requirements, three measures of eective-
ness have been proposed: combat attrition, eort expended, and eciency
[1]. The combat attrition measures how many aircraft have been shot down
in recent conflicts. Table 16.1 shows that the loss of U.S. combat aircraft has
steadily declined both in absolute terms and relative to the number of com-
bat sorties flown. This identifies that SEAD is an important contributor to
aircraft survivability. The amount of eort that is expended to protect U.S.
aircraft can be used to assess SEAD capabilities. As shown in Table 16.2,
2030% percent of all combat sorties in recent conflicts were devoted to SEAD.
That is, SEAD continues to be a growing mission area of concern. While
suppressing enemy air defenses through EW or intimidation can eectively
protect U.S. aircraft, destroying enemy air defenses is generally preferred to
suppressing them because of the enduring eect that destruction has on the
enemys air defense. Table 16.3 shows that the USAF has had mixed results
Table 16.2: Estimates of Combat Aircraft Losses (from [1])
Conflict Combat Sorties SEAD Sorties %

Vietnam 219,407 11,389 5.2
Desert Storm (Iraq) 68,150 4,326 6.3
Bosnia 2,451 785 32.0
Kosovo 21,111 4,538 21.5
N./S. Watch 268,000 67,000 25.0
Table 16.3: Destructive SEAD: Some Estimated Results (from [1])
Conflict Estimated Results

Desert Storm 35 of 120 fixed SAM batteries destroyed
Bosnia 52 of 70 air defense targets destroyed
Kosovo 3 of 25 SA-6 batteries destroyed, 10 of 41 SAM radars destroyed
N./S. Watch 33 of 35 air defense targets damaged, but many rebuilt and improved
in recent conflicts destroying enemy air defense targets. In cases like Iraq,
DEAD eorts have been somewhat successful. In Kosovo however, the SAM
threat to NATOs aircrew proved far more pronounced and harrowing than
originally depicted [14, 15]. Even though only two aircraft were shot down
(one of them a stealth F-117 by an SA-3), SEAD eorts were comparatively
less successful.
16.1.5 Miniature Air-Launched Decoys

The SAM always has the first shot and they start the fight knowing where
the target is. That is, the target is always attacked from ambush. Even if a
reactive ARM times out, it is, at best a revenge weapon [9]. Consequently,
ARM shooters are always looking for a way to stimulate the threat and force
it to reveal its position. The best way to do this is to stimulate the threat
with jamming or decoys such as the miniature air-launched decoy (MALD)
as shown in Figure 16.3. The second best way is to stimulate the threat with
the SEAD aircraft. The not so good option is to let the strike aircraft
stimulate the threat.
The MALD is an expendable air launched vehicle that serves as a decoy
for fighter aircraft and bombers mimicking their radar signatures and flight
characteristics to distract the attention of enemy air defense systems. The
MALD is a turbojet-powered decoy, configured as a swing-wing missile that
can be launched from an F-16 or B-52 bomber. After launching, MALD flies
a preprogrammed flight path into hostile air space to stimulate enemy air
defenses, presenting itself to enemy radar as a real aircraft. Once radars and
Figure 16.3: Miniature air-launched decoy.
air defenses are activated they are intercepted by high speed antiradiation
weapons
16.2 Antiradiation Missile Seeker Design

The ARM seeker competes with the warhead for valuable finance, weight and
volume and presents a significant trade-o to the ARM designer. After all,
in the endgame, its the warhead that counts [16]. However, if the seeker
does not guide correctly to the target, more serious fratricide problems can
occur. Protected by a wideband RF transparent radome, ARM seekers use
passive RF homing with an antenna and microwave receiver. The antenna
and receiver are tuned to the frequency of the threat radar to acquire and
provide location data that can be processed to derive guidance commands.
The guidance commands are passed to the missiles autopilot that filters
the signals to produce guidance and control commands that are sent to the
control surfaces. Stability and control of the flight path are then provided by
the control surfaces. An ARM may have a mid-course as well as a terminal
phase of the flight.
16.2.1 Antenna Design

The ARM must detect and track the radar radiation over a very broad band-
width. It must also have a wide beamwidth in order to detect the emitters
at large angles o boresight. Spiral antennas are frequency independent an-
tennas that can be used to obtain dual-polarized, rotationally symmetric,
multiple-mode patterns over a very broad frequency range [17]. Frequency
independent antennas are antennas whose geometries are specified by angles
and their radiation pattern, impedance and polarization remain virtually un-
changed over a large bandwidth. These features make antennas such as the
logarithmic and conical spirals a good choice for the ARM seeker. They can
also be implemented with printed circuit technology with a simple microstrip

feed network topology that facilitates ease of fabrication [18]. Being confor-
mal with the skin of the airframe and not consuming the internal volume of
the nose cone can also be an asset when dual-mode seekers are considered.
They can also be fixed in place within the nose, do not require a high-precision
gimbal and provide a low cost approach [16].
The antenna performance is not constant for all frequencies. There are
physical bounds that limit the band over which the performance can be held
almost constant [19]. The performance varies over the bandwidth in a manner
that is periodic with the logarithm of the frequency. Consequently, they are
called logarithmically periodic or log-periodic antennas. The most popular of
those have shapes prescribed by logarithmic spiral curves or log-spirals. As an
example, a logarithmic uniplanar spiral antenna that covers a 9:1 bandwidth
with a return loss better than 10 dB from 0.4 GHz to 3.8 GHz is described
in [20].
The logarithmic spiral antenna has N arms interleaved in a spiral pattern
about the center. Their electrical dimensions however scale with frequency.
The feed network at the center of the antenna acts as a beamformer and
splits the power into the arms with a linear progression in phases and equal
amplitudes to produce the various modes of radiation. Radiation occurs from
the points on the spiral which have a half-wavelength in transmission line
between them. Therefore higher frequencies radiate from the antenna near the
central feed point and lower frequencies from the edges. The mutual coupling
between arms relates the excitations at the feed ports to the eective pattern
radiated by each arm. An N = 2-arm spiral consists of two interleaved arms
wound in a spiral, each terminated in a resistance. N arm spiral antennas
are rotationally symmetric such that the rotation of an arm about its axis by
2/N does not change the spiral structure [17]. An N arm spiral can radiate
or receive N 1 independent (characteristic) modes.
The shape of right-handed planar log-spiral antenna is shown in
Figure 16.4 and is based upon the logarithmic spiral curve defined by the
generating equation [21]
= 0 ea(0 ) (16.1)
where is the radial distance from the origin in the direction given by the
angle , and 0 is the radius for = 0 . For a logarithmic spiral, a is a
constant that controls the flare rate of the spiral (1/a is the spiral rate or rate
of expansion of the spiral). We note that the spiral constant
1 d
a= = cot (16.2)
d
where is the angle between a tangent to the curve at any point and a line
to the origin at that point. Since is constant for a given logarithmic spiral,
an alternative name is equiangular spiral [22].
Figure 16.4: Equiangular spiral curve with = 0 ea(0 ) .
In wavelengths, (16.1) can be written as
0 ea(0 )
= = (16.3)

which shows that changing the wavelength is equivalent to varying 0 , which is
just a rotation of the infinite structure pattern and thus results in a frequency
independent antenna.
The total length L of the spiral can be calculated as [22]
8 ^ w W2 1/2
1
2 d
L= +1 d (16.4)
0 d
which can be reduced to

5
1
L = (1 0 ) 1 + (16.5)
a2
where 0 and 1 represent the inner and outer radii of the spiral shown in
Figure 16.4.
The design of a planar logarithmic (equiangular) spiral antenna can be
accomplished using
1 = 0 ea (16.6)
Figure 16.5: Two-arm equiangular spiral plate (adapted from [21]).
as edge number 1 (see Figure 16.4). Edge number 2 has the same spiral curve
but is rotated through a rotation angle as
2 = 0 ea() (16.7)
The other half of the antenna has edges that make the structure symmetric
(opposite configuration) [21]. Rotating one spiral arm by one-half turn brings
it into congruence with the other arm. This assures the antenna can receive
signals of either right-hand polarization or left-hand polarization. A two-arm
spiral has 0 = 0, .
The generating equations for the congruent spiral are
3 = 0 ea() (16.8)
and
4 = 0 ea() (16.9)
The structure is shown in Figure 16.5 and is self-complementary containing
a good degree of pattern symmetry with = /2.
The frequency of the spiral antenna at the upper end of the operating
band fu is determined by the feed structure [21]. For a = 0.221, the minimum
radius 0 /4 at fu . A nearly equivalent criterion is that the circumference
in the feed region Cu = 20 = c/fu . The circumference of a circle just
Figure 16.6: Archimedean spiral antenna with left hand circular polarization
(radiation out of the page) [21] ( c John Wiley & Sons, Inc. 1997).
enclosing the spiral can be used to set the low frequency limit fl as Cl =
2 = c/fl . The low frequency limit set by the overall radius is approximately
a quarter wavelength at fl . For example, consider a spiral with one and one
half turns with a = 0.221. Here the maximum radius R = ( = 3) = 8.030
which is c/4fl . At the feed point R = (0 ) = 0 = c/4fu . The bandwidth is
then fu /fl = 8 which indicates an 8:1 bandwidth (a typical value).
To maintain a large bandwidth, the antenna must also be fed by an elec-
trically and geometrically balanced line. This feed is often referred to as an
infinite balun and has an impedance of Z 120 [21]. The radiation pattern
of the self-complementary planar equiangular spiral antenna is bidirectional
with two wide beams broadside to the plane of the spiral. The filed pattern
is approximately cos when the z-axis is normal to the plane of the spiral.
The half-power beamwidth is thus approximately 90o and the polarization is
near circular over wide angles. Spiral antenna gain values range between 2
and 4 dB. The frequency limit is typically 500 MHz to 18 GHz.
In the Archimedean spiral antenna, named after the third-century B.C.
Greek mathematician Archimedes, the radial distance is linearly propor-
tional to the polar angle in the generating equation (rather than exponen-
tially related). It flares more slowly as shown in Figure 16.6. The generating
Figure 16.7: Cutaway view of the programmable wideband British alarm

showing the seeker and spiral antennas.
equations for the Archimedean spiral antenna are

= 0 (16.10)
and
= 0 ( ) (16.11)
The successive turnings of the spiral have a constant separation distance
(equal to 20 if is measured in radians). The arms are fed 180o out-of-
phase at F1 and F2 giving the antenna circular polarized radiation which is
frequency independent. The radiated fields created by the currents are or-
thogonal, equal in magnitude and 90o out of phase. A unidirectional beam
can be created by backing the spiral with a metallic cavity behind the spi-
ral. The pattern of the cavity-backed Archimedean spiral can be modeled
empirically by [21]
F () = cos5.8 (0.53) (16.12)
and has a half power beamwidth of 74o . Figure 16.7 shows a schematic
cutaway drawing of the British alarm ARM and shows the use of four planar
Archimedean spirals to cover the quadrants of interest.
Conical equiangular spirals are also often used in ARM seekers since they
generate a unidirectional pattern. Figure 16.8 shows the schematic of a conical
equiangular spiral. The edges of one conical spiral surface are defined as
Figure 16.8: The conical equiangular spiral antenna [21] ( c John Wiley &
Sons, Inc. 1997).
= e(a sin h ) (16.13)

of which the planar spiral is a special case with h = 90o . Larger values of
h in 0 h /2 represent less tightly wound spirals [22]. The generating
equations for the conical equiangular spiral are
1 = e(a sin h ) (16.14)
(a sin h )()
2 = e (16.15)
and = /2 for the self-complementary configuration. The edges of the arms
maintain a constant angle with a radial line for any cone half-angle h
a = cot (16.16)
The advantage to the conical spiral is that it provides a unidirectional ra-
diation (single lobe) toward the apex of the cone with the maximum along
the axis. It also preserves the circular polarization and relatively constant
impedances over the large bandwidths required [22]. Typical patterns for
h 15o and 70o have half-power beamwidths of 80o . For the band de-
sign, the apex diameter determines the upper frequency d = c/4fu . The lower
frequency of the antenna is determined by the base diameter B = 3c/(8fl )
[21]. As an example of a conical spiral, the Russian Kh-31 ARM uses an array
of seven conical elements as shown in Figure 16.9.
Figure 16.9: Photo of the Kh-31 seeker that uses a series of seven conical
spiral elements.
Several patents have been issued that address the shortcomings of spiral
antenna designs for ARM seekers. Methods to counteract the threat of the
radar shutting down and to also improve the boresight error of the antenna
system and also reduce the radome error slope are reported in [23, 24]. Here
the antenna system includes a parabolic reflector dish having a dielectric
substrate and a conductive material coating on the substrate in order to
provide a narrowbeam high-gain radiation pattern. The parabolic dish also
has a conductive material coating on the reflector substrate defining the spiral
antenna for a low-gain, wideband radiation pattern. To address coupling
between the antenna and the missile body in the VHF band, a broadband
polarization diverse monopulse spiral antenna with a body cancelled current
array and radial arm-coupled log periodic loop antenna is described in [25].
To provide a novel nonobvious solution to the problem of fitting a number of
spiral antennas having dierent configuration senses into the space of a single
spiral, the spirals can be symmetrically arranged about a point at the center
of a circle. Each spiral antenna is deformed to occupy substantially all of the
area within a sector of the circle [26].
16.2.2 Receiver and Signal Processing

A block diagram of an ARM including its seeker is shown in Figure 16.10
[27]. In this example, the receiver (RX) accepts signals from four antennas
that are used to intercept and direction find the RF emission from the radar.
Many designs include both a high-band and low-band antenna to improve the
direction finding accuracy (smaller beamwidth). The frequency synthesizer is
used to scan the instantaneous bandwidth through the operating bandwidth
in order to search for the target radiation. The intercepted emissions are
down-converted, and filtered with a passband filter.
Logarithmic amplification is used in the RX and applied to the passband
filter output. Logarithmic amplifiers are used widely in antiradiation seek-
ers and can be classified into two primary families, the logarithmic IR/RF
amplifiers and the detector logarithmic video amplifiers (DLVA) [28]. The
logarithmic IF/RF amplifier obtains the logarithmic transfer function at the
IF (or RF) frequencies, while the DLVA obtains the logarithmic transfer func-
tion in the video frequency domain. Advantages of the logarithmic IF ampli-
fiers over the DLVA include an easily obtainable CW response (important in
ARMs attacking LPI emitters), excellent pulse recovery time, fast rise time
and wide instantaneous dynamic range. The DLVA, however, generally has
a smaller logarithmic error over the temperature range and frequencies of in-
terest. The DLVA has superior dual-channel tracking characteristics and is
usually the choice for ARM monopulse direction finding. This is because it
is easier to produce matched nonlinear circuits at video frequencies than at
IF/RF frequencies [28].
The amplifier output is lowpass-filtered and the monopulse azimuth (AZ)
and elevation (EL) error signals are digitized. Also digitized are the RF
frequency, the pulse time of arrival (TOA), the pulse width (PW), and the
amplitude of the pulse. The signal processing then gates these error signals
and uses pulse discriminant logic, deinterleaving, and a PRF correlator in or-
der to sort the various radar signals being intercepted including their angular
location. A Kalman filter is then used to derive the command acceleration
from the seeker line of sight rate on the selected target signal. The autopilot
then filters these commanded accelerations, which are then applied perpen-
dicular to the airframe velocity vector to guide the missile to the target.
16.2.3 Dual-Mode Design

There are two major problems that aect ARM performance. The first consid-
eration is that air defense radars, when anticipating an ARM attack, usually
operate in a blink mode turning on just long enough to obtain tactical infor-
mation then shutting down to avoid attracting ARMs. The second problem
is that if the radar stays active allowing the ARM to be launched against
it, some attacks will fail because some of the radar energy bounces o the
ground creating a false (multipath) target [29]. If the multipath is not too
severe, the problem can be addressed using leading-edge track. To address
more significant multipath situations and also counter the blink mode the
concept of the dual-mode ARM seeker has been explored.
Figure 16.10: Block diagram of an antiradiation missile seeker (from [27]).

Dual-mode ARM seekers can increase the capability of the missile when
the RF radiation source shuts down. A unique dual-mode guidance scheme
devised by the U.S. Army researchers was key to developing a more eective
ARM with a minimal increase in cost and weight. The dual-mode seeker
would retain the traditional passive mode RF homing capability but would
also have either an imaging infrared or active millimeter wave (MMW) radar
mode. These latter modes do not require radar emissions from the target
and can deal with blink tactics. They can also give higher accuracy when
used in the final stage of the attack. The addition of a second guidance
mode involves additional hardware increasing the manufacturing costs of the
missile. Note that the design and production of ARMs represent a balance
between the technology that is incorporated into the missile and the ability
of the manufacturer to sell the weapon system.
The solution proposed by the U.S. Army is based on an active MMW
seeker, but adds to that seeker antiradiation homing antennas and down
conversion elements as shown in Figure 16.11. The antiradiation homing an-
tennas intercept signals emitted by enemy air defense radar but instead of
passing these signals to a dedicated receiver and guidance system, the an-
tennas pass them to the conversion elements. The task of these conversion
elements is to convert the intercepted signals to the IF that is also used by the
active MMW processing. This IF could then be handled by the signal proces-
sor that already exists as part of the MMW seeker. Use of the same processor
to handle both passive and active-mode radar signals greatly reduces the cost
penalties of providing the second guidance mode.
At least three or four passive detector channels should be used. The as-
sociated antennas would be mounted on the exterior of the missile at regular
intervals around the circumference of the fuselage. With proper phase rela-
tionships between detector channels the azimuth and elevation direction find-
ing (DF) information can be provided. The antennas should have a broad
beamwidth so phase comparison monopulse techniques can be used rather
than the alternative amplitude comparison DF technique. The angle of ar-
rival of the enemy radar signal would be determined by comparing the phase
of the emission signals from the individual antennas.
During the initial and mid-course portions of flight the dual-mode mis-
sile would use its passive-radar mode to home in on the emissions from the
hostile radar. During the terminal phase of the attack, it would switch to
the active MMW mode, acquire the hostile radar and conduct an accurate
attack that does not depend on the target remaining active. The distance
from the target at which the missile switches modes is a function of its speed
and maneuverability and is typically 24 km.
Another example of a dual-mode ARM seeker is Alliant Techsystems
advanced antiradiation guided missile (AARGM) shown in Figure 16.12. De-
veloped under the Navys Quick Bolt program, the passive conformal array
antenna provides high accuracy wideband DF capability [30]. Autonomous
Figure 16.11: MMW/Passive detector channels diagram.
target detection, identification, tracking and location ranging of the target

are provided. The large field of view, sensitivity, frequency and DF accuracy
and processing enable the weapon to be successful without an independent
targeting system.
In addition to homing in on the signal emitted by the hostile radar, the
AARGMs dual-mode guidance includes an active MMW radar. This section
of the seeker provides terminal target acquisition, tracking, guidance and
fusing to find its target even if the hostile radar is no longer radiating (anti-
ARM tactics). If the radar shuts down, the MMW will go into a search
mode. Since the seeker is in the MMW band it can detect RF scattering from
the radar antenna, radar platform and missile launcher [30]. An integrated
GPS/INS navigation suite is also included in the seeker to provide mid-course
guidance and supply the rough coordinates to fly to after the radar shuts
down. This prevents the ARM from missing the target in the terminal phase
and most importantly, keeps the weapon from landing in the wrong country
(preventing fratricide) [13]. It also enables a sensor fusion and autonomous
ranging capability.
These features extend the AARGM capability providing a long-range,
time-critical strike weapon against other than the traditional ARM radar
targets. That is, the warfighter survivability is increased with the addi-
tional speed, range and targeting capability. The AARGM weapon also has
a network-enabled capability with a receiver that links the weapon to na-
Figure 16.12: Advanced Anti-Radiation Guided Missile (AARGM) seeker.
tional systems targeting data. A burst transmitter that transmits critical

target data into the warfighters information warfare network prior to missile
impact on the target is also provided [30].
16.2.4 Signal Processing

To learn and recognize the dierent threat radar emissions being intercepted,
traditional if-then-else constructs have been traditionally applied when the
radar parameters being sorted (e.g., pulse width and time of arrival) are
known. If a new radar emission is intercepted, the deinterleaving process
for example may have problems. Artificial neural networks have also been
applied with some success due to their ability to learn [31]. To understand
this, consider that the threat radar signal features can be quantified. Signal
angles of arrival (AZ, EL) and SNR are examples of extrinsic features having
to do with where the emitter is located. Carrier frequency, pulse width, pulse
repetition pattern and sophisticated RF modulation are intrinsic features
having to do with what particular radar is active. These features can be
used eciently in a neural network application. An adaptive network sensor
processor was designed and implemented in an ARM software application in
[31]. They demonstrate that noisy binary inputs could be characterized with
respect to noise amplitude and shape by a recall procedure that was tuned
to enhance the analog output feature shapes. They also demonstrate that
analog input shapes could be stored and recalled and that unstored states
could be discriminated against without an output nonlinearity.
In summary, modern microelectronic technology has made it possible to

put immensely sophisticated detection and analysis systems into the ARM.
Each missile carries its own radar seeker, signal analysis equipment and a
threat file, which enables it to identify virtually any radar by its signal
structurefrequency, pulse length, modulation and pulse repetition frequency.
The ARM can leave the launch aircraft and search for radar signals, compar-
ing received pulses with a comprehensive threat file, in order to find and then
home onto the greatest hostile threat. With the loiter capability (e.g., HARM
and Alarm), the ARM can climb to altitude, deploy a parachute and then
search for target signals as it descends slowly.
Once a target has been identified, a high-speed attack phase is initiated,
which enables the missile to get to the target in a very short time. Both types
have inertial navigation systems which can store positions from which signals
have been detected, and then guide the missile to the selected target without
further signal inputs. Switching o the radar provides little protection against
this sort of capability. In any case, radar is a fundamental part of many
air defense systems, often providing direct control of anti-aircraft missiles in
addition to detecting and tracking targets. Such radar systems must have
a high priority in the threat file carried by any antiradiation missile and
switching o the radar for self-protection will render the complete missile
system ineective. Partly as a result of this, there is a steadily increasing
emphasis on highly mobile radar systems although mobility is unlikely to
provide a defense against an imminent attack. If an ARM has already been
launched from an aircraft, the amount of warning will be measured in seconds.
If the launch aircraft pops up over the radar horizon, detects a hostile radar
emission and then launches an ARM, the radar will have its first warning
of the attack when it is only about 15 km away. This would probably give
it between 10 and 20 seconds to reorganize. As shown in Figure 16.13 the
F-4G Wild Weasel can carry an air-to-ground missile (AGM)AGM-88 high-
speed ARM (HARM); the AGM-65 Maverick, which has an infrared seeker;
the ALQ-119 electronic attack pod; the AGM-78 Standard ARM; and the
AGM-45 Shrike ARM.
16.2.5 Future ARMsAddressing the LPI Emitter

Until recently, almost all radars were designed to transmit short-duration
pulses with a high peak power. This type of signal is easy to detect using
relatively simple, traditional ARM seeker receivers as shown in Figure 16.10
making the radar source vulnerable to an ARM attack. With the arrival of the
LPI requirements, the ARM seeker and signal processing methodology must
be revisited. The use of very low peak power (e.g., PCW = 1 mW) requires
the ARM seeker to have a much higher sensitivity in order to detect these
types of signals. With the increasing number of radars using LPI techniques,
the ARM is now required to measure and characterize conventional pulsed
Figure 16.13: F-4G Wild Weasel carrying an AGM-88 HARM, AGM-65 Mav-
erick, ALQ-119 electronic attack pod, AGM-78 Standard ARM, and an AGM-
45 Shrike ARM.
radar signals as well as detect and characterize the LPI signals.

Detection of LPI signals can be accomplished using a number of dierent
receiver architectures. The signal processing functions can be quite intensive
if all of the received data (pulsed waveforms and LPI waveforms) is digitized
at the IF band and analyzed using signal processing. Figure 16.14 shows an
example of an LPI signal, a pulsed emitter signal and the presense of thermal
noise within the ARM seeker.
A solution to the problem of overloading the signal processor when LPI
signals are present along with pulsed waveforms can be addressed with an
LPI signal discriminator. A block diagram of the discriminator is shown in
Figure 16.15. The LPI signal discriminator is operatively coupled to the down
converter and produces a trigger signal that is used when the incoming signal
is above a predetermined threshold, to thereby transfer the digitized signal
to a special digital signal processor for analyzing the incoming LPI signal.
The architecture suppresses the high-peak power, short duration signals and
triggers a data buer for gating the digitized LPI data to the digital signal
processor [32]. The LPI signal discriminator in an analog processor uses a
cascade of IF amplifiers with a pad between the amplifiers to distribute the
signal power evenly throughout the chain and soft-limit the strong pulsed
signals that are received. The amplified signal is detected by a large dynamic
Figure 16.14: LPI and pulsed signal amplitudes within an ARM seeker as a
function of time (from [32]).
range, successive detector log video amplifier. The logarithmic video output
is further compressed using a follow-on video logarithmic amplifier before the
signal is time-integrated by one or more integrators. A threshold comparator
receives the output from the integrators to provide a trigger pulse output
when the input crosses a predetermined threshold value set by the signal
processor [32].
The LPI emitter can also be detected with a sucient amount of inte-
gration of the intercepted energy. For each direction of arrival, an optimal
detector is able to integrate the energy of the emitted signal of which the
parameters are unknown. Using a multichannel detector, the dierent chan-
nels can be tuned onto dierent durations and passbands for the noncoherent
integration. The output of each channel depends on the time of arrival and
the starting spectrum frequency. Searching for the emitter in time and fre-
quency is most conveniently done in the time-frequency domain which is easily
calculated as shown in the previous chapters.
A multichannel detection algorithm recently suggested for time-frequency
domain LPI detection is given as [33]
0 +T f3
t3 0 +B
(p, q) = |G(t, f )|2 (16.17)

t=t0 f =f0
which uses the Gabor time-frequency distribution G where t0 is the signals

time of arrival, f0 is the initial spectrum frequency, T is the signal duration
and B is the signal bandwidth. This noncoherent integration is done using
two-dimensional sliding windows as shown in Figure 16.16 [34]. In Figure
16.16(a) a model of the time domain signals within the receiver are shown
Figure 16.15: Block diagram of an analog LPI signal discriminator (adapted

from [32]).
Figure 16.16: (a) Time domain and frequency domain model of the thermal
noise and LPI emitter signal and (b) two examples of a noncoherent integra-
tion sliding window for detection of the LPI signal (adapted from [34]).
Figure 16.17: Block diagram of a digital FMCW signal discriminator (adapted

from [35]).
and include the LPI signal x(t) and the receivers thermal noise n(t) in the
time domain and the magnitude spectrum |X(f )| and |N (f )| in the frequency
domain.
Another architecture for detecting FMCW LPI waveforms employs a tech-
nique described as deramping which forms an adaptive matched filter to
the linear FMCW signal in order to achieve the processing gain that is equal
to the LPI signals time-bandwidth product [35]. A block diagram of this
technique is shown in Figure 16.17. The deramping process mixes the input
signal with a locally generated linear FM signal to produce an output signal
with a reduced FM slope in comparison with the input signal. To construct a
matched filter, the carrier frequency, modulation period and the modulation
bandwidth must be known. To determine these parameters, the matched
filter must be adaptively formed. A multichannel arrangement is proposed
by examining the output of an FFT filter bank using a CFAR scheme that
sets the threshold for determining the channel hit. The channel that yields
the first detection has its matched filter parameters adaptively changed to
achieve optimal processing gain.
16.3 ARM Performance Metrics

There are 10 metrics that may be used to determine the capability of an ARM
missile. These are presented below from [36].
Maximum Range: Measure of how distant a radar can be successfully
engaged by an aircraft at a given altitude.
Speed: Measure of how quickly the missile can reach its target. Slow
missiles provide the radar operator time to react and shut down.
Frequency coverage: Measure of how many dierent types of radar can
be identified, tracked, and engaged by the missile. Low band coverage
is important since it allows engagement of Early Warning (EW) and
Ground Control Intercept radar. High band coverage is also important
since it allows engagement of SAM fire control radar and illuminators.
Pulse density, CW limitations: Measure of seekers ability to identify
specific radar in a high pulse density threat environment. The seeker
must be capable of de-interleaving pulse trains from many radars in
order to select a specific target. This also includes the ability to identify
the LPI emitters that are present.
Electronic protection capability: Measure of weapons ability to resist
seduction by dummy emitters and decoys. This also includes the ability
of the ARM to withstand a directed energy attack on the seeker.
Lethality: Determined by accuracy and warhead eectiveness, a mea-
sure of what kill probability can be achieved. If the ARM has poor
lethality, more rounds must be fired on average per killed radar.
Deliverable Payload: Measure of the delivering aircrafts payload of
ARMs.
Flexibility: Measure of how many dierent modes exist for the weapon.
The more delivery modes, the more dicult it is for the opponent to
devise defensive measures.
Cost: Measure of how many weapons can be delivered per dollar ex-
pended.
Integration with the launch aircraft: Measure of how weapons can ex-
ploit the launch aircrafts onboard RWRs, radar homing and warning
system, or emitter locating system.
As pointed out in [36], the decision to select a particular type of ARM is not
trivial and the ultimate system level metric of usefulness is that of how many
hostile radars can you take down for what dollar investment in ARMs, ARM
support costs, and aircraft integration is required.
16.4 Former Soviet Union and Warsaw Pact

Allies
The former Soviet Union and the current Russian Federated States have been
developing ARMs since the 1960s. Unlike the United States, the Soviet Union
did not develop aircraft specialized for SEAD missions. That is, they didnt
treat SEAD as independent air operations [37]. The Soviets development
of its ARMs was driven by doctrine of a massive frontal air assault on the
most common threat at the time, Western Europe. Since this assault would
have numerous aircraft, these ARMs did not have to be sophisticated and no
dedicated SEAD aircraft were fielded.
SEAD missions were flown to support air-ops on a tactical level (by non-
specialists). They relied on signals intelligence (SIGINT) and other recon
assets and preferred preplanned strikes on known enemy IADS rather than
targets of opportunity [37]. With the theater of war in western Europe, the
western SAM threats (late 1950s, early 1960s) consisted of the U.S. Nike
Ajax, Nike Hercules, U.K. Thunderbird, and the Bloodhound systems. With
the Ilyushin Il-28 Beagle (nuclear weapon platform) having a cruising speed
of only 500 mph, no low-level penetration was attempted. Following this,
the Yakovlev Yak-28 (Brewer) supersonic (1960) and Sukhoi Su-7 (Fitter-A)
supersonic (1960) also provided no solution to the SAM problem. Instead,
the Soviet Union relied on mass nuclear weapons employment. Not until 1963
did work begin on the first tactical ARM. Table 16.4 lists the ARM weapons
developed along with their NATO name and year the ARM entered service.
Table 16.4: Russian ARM Development
Missile NATO code Russian name Entered Service

AA-10 Alamo R-27P 1989
AS-4 Kitchen Kh-22MP 1974
AS-5 Kelt KSR-2P 1962
AS-6 Kingfish Kh-26MP 1969
AS-9 Kyle Kh-28, Kh-28E 1973
AS-11 Kilter Kh-58 1977
AS-12 Kegler Kh-25MP, Kh-25MPU 1981
AS-16 Kickback Kh-15P 1988
AS-17 Krypton Kh-31P 1991
16.4.1 AA-10 Alamo

The AA-10 Alamo is an air-to-air missile fitted with a passive seeker that
homes in on the emissions from a threat fighters radar. It was fielded in 1989
Figure 16.18: AA-10 Alamo (from [39]).
and is now being possibly exported to India and China [38]. The missile can
be used in two modes. One mode is in conjunction with the launch aircrafts
radar. The other mode is one in which the launch aircraft does not use its fire
control radar at all. Instead the passive seeker 9B-1032 is used to detect the
most powerful radar emission which is then reported to the pilot. A photo of
the AA-10 is shown in Figure 16.18 [39].
The Alamo is intended for use against enemy fighters at long range, when
the launch aircraft may still be beyond the maximum range of the targets
radar. Since the weapon uses passive homing, it will give the target no warn-
ing that a launch has been made. Radars that are fielded on F-15 and F-16
aircraft along with other western fighter radars are the R-27Ps primary tar-
gets with the main aim being to stop the threat aircraft radar from emitting.
The ARM seeker is capable of detecting emissions from a threat radar at
ranges up to 120 km. The homing head is however, capable of detecting a
target from a range of more than 200 km, but the R-27EP cannot carry out an
interception at such distances. 1 The flight time would exceed the operating
duration of the missiles onboard power supply. Vympel is working on ways
of increasing the operating time of the power supply in order to allow R-27EP
engagements at up to 200 km. The deployment is believed to have an eect
on NATO tactics spurring radar upgrades to more LPI emitters and the use
of towed radar decoys. There is also thought of producing a passive-seeker
variant of the R-77 AA-12 Adder as a successor to the R-27P [38].
16.4.2 AS-4 Kitchen

The Kh-22MP with NATO code name AS-4b Kitchen was built by Raduga is
launched from the Tu-22M Backfire B, Tu-20 Bear G, and Tu-22 Blinder
B aircraft. The missile entered service in 1974 and is shown in Figure 16.19
1 Vympel oers two versions of the missile: the standard R-27P with a maximum range
of 72km and the R-27EP version with a bigger rocket motor which gives a maximum range
of 110km.
Figure 16.19: Kh-22 Kitchen (from [40]).
[40]. It is used against ship- or ground-based radars and has a 930-kg HE

blast fragmentation warhead with an active laser fuse. Guidance is inertial
with a passive radar terminal seeker.
The passive guidance (PG) radar seeker can lock on to a target from a
distance of 250270 km. The inertial guided versions use the PSI Doppler
radar to compute distance covered. The passive radar homing missiles have
the PGP passive radar seeker with a range of 380350 km [41]. After launch
the missile climbs to an altitude of 22,500m for cruise flight stabilized by the
APK-22A autopilot and then dives into the target at an angle of 30 degrees.
Maximum launch range depends on the speed and altitude of the launching
aircraft: from a speed of 950 km/h and an altitude of 10 km it is 400 km
and from a speed of 1,720 km/h and an altitude of 14 km it is 550 km. The
initial versions of the Kh-22 had a maximum speed of 3,600 km/h, which was
increased to 4,000 km/h with the Kh-22M series that entered service in 1974
and 1976. The Kh-22B experimental version reached a speed of Mach 6 and
an altitude of 70 km during tests in the 1970s [41].
16.4.3 AS-5 Kelt

The KSR-2P with NATO code name AS-5 Kelt was built by OKB MiG and
could be launched from the Tu-16 Badger C mod, Badger G and had a
maximum range of 220 km and a maximum speed of Mach 1.2. The Kelt
shown in Figure 16.20 entered service in 1962, but was phased out by the late
Figure 16.20: An AS-5 Kelt air to surface missile loaded on the wing pylon of
an Egyptian Air Force Tu-16 Badger aircraft [42] ( c 2007 Janes Information
Group).
1980s [42]. The operational ceiling of the missile was 9.1 km. About 25 Kelts
were used by the Egyptian Air Force in 1973 against Israeli forces, and five
are reported to have been successful. It is believed that 12 of these missiles
were anti-radar versions of the AS-5.
16.4.4 AS-6 Kingfish

The KSR-5P with NATO code name AS-6 Kingfish was built by Raduga and
was launched from the Tu-16 Badger G mod. It had a maximum range of
400 km with maximum speed of Mach 3.0 [43]. The AS-6 missile as shown
in Figure 16.21 entered service in 1969, with a second version in 1973 for
carriage on the Tu-95 Bear, and a third version in 1976 for carriage on the
Tu-95M. There were reported to be around 100 missiles in service in Russia
in 1990, but modified missiles were oered for export as air targets and it
is believed that all AS-6s had been removed from operational service by the
end of 1994. Some AS-6 missiles were retained by Ukraine, but by 2002 it is
believed that these missiles had been destroyed. The missile is launched from
an altitude of 0.511 km after which it climbs to an altitude of 20 km for
cruise flight. The missile starts its terminal dive 60 km from the target. The
most recent version, the KSR-5NM also included the ability to loiter which
added more flexibility to attack SAMs that shut down once the missile was
in the air.
Figure 16.21: AS-6 Kingfish [37] ( c 2001 Horizon House).
16.4.5 AS-9 Kyle

Along with the Yak-28N supersonic bomber, the Kh-28 missile with NATO
designation AS-9 Kyle, and the radar target and acquisition system, the K-
28P was their sole attempt to build a dedicated SEAD weapon system com-
plex. The P comes from Russian word protivradiolokatsyonny meaning anti-
radar. At the same time the aircraft also became an EW jamming platform,
the Yak-28PP Brewer-E. The jammer was taken from the Tupolev Tu-16PP
and split among three of the Yak-28PP. Their most eective suppression was
when each with a dierent jammer component, operated in one formation on
each side of the ingress, egress corridor [37].
The development of the Kh-28 ARM with NATO code name AS-9 Kyle
(built by MKB Raduga) began in January 1963 and was based on the Kh-
22 and KSR-5 missile technology. It entered service in 1973 and is shown in
Figure 16.22 being loaded onto an Su-22M3. The Kh-28 seeker had a conically
scanning antenna. Its range was 110 km with a speed of 800 m/s with launch
altitude of 20011,000m. From low altitude the launch range is reduced to
45 km. For guidance the missile originally used the APR-28 passive radar
seeker developed by NPO Avtomatika. The seeker was later carried on the
Su-24 Fencer-A (tactical bomber) and the Su-17M Fitter-C (tactical fighter-
bomber). It was tuned to the frequencies of the Nike Hercules shown in Figure
16.23, the Thunderbird, and the Bloodhound which is shown in Figure 16.24.
Later the Kh-28M missile received a new passive radar guidance PRG-28M
seeker that could be used also against the radars of the HAWK SAM system.
After cancellation of the K-28P complex in 1967, NATO adopted a new
Figure 16.22: Ground crews prepare to load a AS-9 Kyle (Kh-28) on an

Su-22M3 Fitter-F [37] ( c 2001 Horizon House).
Figure 16.23: Nike Hercules SAM.
Figure 16.24: Bloodhound Mk 2 SAM deployed on its Type 202 launcher.

defense doctrine: No Nukes. In changing their SEAD doctrine, for air

defense targets 300400 km deep, the Tu-22K (Blinder-B), Su-24M (Fencer-
D), and Tu-22M (Backfire) were employed to lay down cha corridors 4050
km wide using two to three strike groups. The corridors were separated by
10 km and 2 minutes; two corridors were often oset. One corridor was also
used to suppress SAMs that had been located by SIGINT. Still there were no
dedicated SEAD aircraft with the corridor task groups having 2 to 4 aircraft
armed with Kh-28s. The Kh-28 has now been withdrawn from service in
Russia. However, a small number may remain in operational use in other
countries. It is believed that it was exported to Afghanistan, Azerbaijan,
Belarus, Bulgaria, Georgia, Hungary, Iraq, Kazakhstan, Libya, Poland, Syria,
Ukraine, and Vietnam [44].
16.4.6 AS-11 Kilter

The Kh-58 with NATO code name AS-11 Kilter uses a new seeker to engage
surveillance pulse-radar systems such as the AN/TPS-43 and the AN/TPS-
44 used for ground controlled fighter intercept and the AN/MPQ-53 radar
of the Patriot. Aircraft platforms include the Su-24M Fencer-D, MiG-25BM
and Su-17M4. The missile shown in Figure 16.25, has a range of 120 km from
a height of 10,000m and 160 km when fired at 15,000m. The missile entered
service in 1977 and had a speed of 900 m/s. An upgrade to the missile (Kh-
58U) extended the range to 250 km [45]. The missile also has an improved
seeker allowing lock-on after launch mode. The MiG-25BM Foxbat-F or Su-
24M with an upgraded Kh-58U missile under the wings is the closest thing
the Russians have to a dedicated SEAD aircraft. They are concentrated at
the 98th Reconnaissance Wing on the Kola Peninsula [37]. The passive radar
seeker of the Kh-58 itself can target various surveillance radars in addition to
the Nike-Hercules, HAWK, I-HAWK, and Patriot SAM systems. The missile
performs a pop-up maneuver in the terminal phase to hit the target at a
2030 degree angle which improves the eectiveness of the warhead. The
kill probability of the antiradiation version is claimed to be 80% within a
20m radius of the target radar. The Kh-58U was designed primarily for the
MiG-25BM SEAD aircraft. It has extended range, improved aerodynamic
characteristics and a guidance system allowing lock-on after launch mode,
and a new rocket engine. The Kh-58E is an export version of the Kh-58U
without the lock-on after launch feature. The missile can be launched at
speeds of Mach 0.47 to 1.5 from altitudes of 0.122 km, but from low altitude
the launch range of the Kh-58 is reduced to 36 km, and that of the Kh-58U to
80 km. The Kh-58 was exported to Bulgaria, Czechoslovakia, East Germany,
and Poland [46].
Figure 16.25: AS-11 Kilter (from [45]).
16.4.7 Kh-27
The Kh-27 antiradiation missile (no NATO designation) has a speed of
850 m/s and was conceived as a replacement for the heavy Kh-28. Conceived
as an ARM variant of the earlier Kh-23 (AS-7 Kerry), the Kh-27 entered
service in 1977 targeting the Nike Hercules, Thunderbird, Bloodhound and
HAWK SAM systems. The missile was carried on aircraft platforms MiG-27
and Su-17M3 and had a range of 60 km when launched from 15,000m and
40 km from 5,000m. The new seeker was developed with a highly sensitive
receiver with five antennas in the PRGS-1 type guidance version and with six
antennas in the PRGS-2 version. Direction finding to the target was based on
phase dierence interferometry and was much more accurate than the coni-
cal scanning antenna in the Kh-28. The weapon saw limited service in the
late 1970s and early 1980s, and was replaced by the ARM member of the
modular Kh-25M AS-12 Kegler family.
16.4.8 AS-12 Kegler

The Kh-25MP and improved Kh-25MPU with NATO code name AS-12 Kegler
(built by Zvezda) are dedicated antiradiation variants of the Kh-25 (AS-10)
air-to-surface missile. The missile was given a new NATO code name because
of its specific defense suppression role. The Kh-25MP eectively replaced the
Kh-28 (AS-9 Kyle) in Russian service. The missile entered service in 1981
[47]. The missile shown in Figure 16.26 had two interchangeable seeker heads
with antennas tuned to the radar frequencies of the HAWK and Nike Her-
cules SAM systems respectively [48]. Targeting was done with the Vyoga ES
Figure 16.26: AS-12 Kegler (from [48]).
pod. The missile was programmed to perform a hump maneuver in the

terminal phase to hit the target at a 2030 degree angle and thus improve
the eectiveness of the warhead. The seeker had tracking rates of 6 to 8o /s,
azimuth coverage of 30o , and elevation cover from +20 to 40o . Most aircraft
are reported to carry an APK-8 radar emitter locator pod with the KH-25MP
missiles. The concept was also new since it had an interchangeable guidance
system in the nose and tail modules attached to a common missile core [37].
It was carried on aircraft platforms MiG-27, Su-17M3 and M4. The missile
has a range of 60 km and had a speed of 850 m/s at a launch altitude of
10015,000m. It is reported as still being used in the Russian Air Force al-
though in limited numbers. The Kh-25MP has a maximum range of 60 km
when launched from medium altitude (30,000 ft), and a range of 25 km when
launched from low altitude. The minimum range is 3 km [47].
The improved Kh-25MPU version is optimized for use against X-band
surface-to-air missile engagement radars, and has a weight increased to
320 kg. The minimum range is 3 km, and the maximum range is reduced
to 40 km. A successor antiradiation weapon to replace the Kh-25MP/MPU
is under long-term development by Zvezda Strela now the Tactical Missiles
Corporation.2 Very little is known about the status of this weapon, the Kh-38
[47]. The new missile is likely to be fitted with a dual-mode passive radar
and imaging infrared seeker. Maximum speed: 3,100 km/h. The Kh-25MP
is a later version based on the universal Kh-25M design. For threats in the
A-waveband a PRGS-1VP seeker is used, whereas for A1-waveband threats
a PRGS-2VP seeker is utilized. The Kh-25MPU is a modernized version de-
signed to defeat also the Roland and Crotale SAM systems. The missile can
be launched at altitudes of 50m to 10 km and has speeds of 600 to 1,250
km/h. The Kh-25MP was exported to East Germany, Czechoslovakia and
2 In March 2003 the Zvezda-Strela State Research and Production Centre transformed
itself into the Tactical Missiles Corporation JSC (Joint Stock Company) following the
incorporation of the various engine, seeker, electronics and other equipment concerns that
were associated with its missile development programs.
Figure 16.27: AS-16 Kickback [49] ( c 2007 Janes Information Group).
Poland [47].
16.4.9 AS-16 Kickback

Little was known about the existence of this antiradar missile until the visit
in 1988 by the U.S. Secretary of Defense to Kubinka airbase to see the Tu-160
Blackjack bomber. NATO has given this missile shown in Figure 16.27 the
designator AS-16 Kickback and it has the Russian designator Kh-15. AS-16
is reported to be able to cruise at altitudes between 30m and 22 km. There
is also a report that states that the missile can have a ballistic trajectory,
reaching a maximum altitude of 40 km [49].
The missile has a range of 150 km when released from medium altitude
(30,000 ft) against a large ship target, and about 100 km when released from
low level. The AS-16 has a minimum range of 40 km. The missile can be
released at altitudes between 300m and 22 km. Maximum speed is Mach 5.0.
After launch from a speed of 1,000 to 2,100 km/h and an altitude of 0.3 to
22 km, the missile climbs to a height of about 40 km. After having acquired
its target, the missile dives, reaching a speed of Mach 5 [49].
16.4.10 AS-17 Krypton

The Kh-31P with NATO name AS-17 Krypton entered service in 1991 and is
shown in Figure 16.28. The development program began in the late 1970s, as
a follow-on to the Kh-25MP (AS-12 Kegler). Guidance for the Kh-31P is by
passive radar homing, with an inertial system to enable homing to continue
even if the target radar is switched o. Three interchangeable seeker head
options are available for the Kh-31P, each one tailored to a specific range of
radar frequencies (Avtomatika L-111, L-112, and L-113) [50]. These seekers
were tailored for use against the Nike Hercules/Improved Hawk and Patriot
SAM systems plus the maritime SPY-1 Aegis phase array radar system.
Figure 16.28: AS-17 Krypton (from [51]).
The Kh-31PM upgrade, now in development, will produce a single inte-

grated wideband seeker (L-130, developed by the Avtomatika CKBA plant at
Omsk) to replace the L-111, L-112 and the L-113. The MiG-27 Flogger usu-
ally carries an APK-8 radar emitter locator pod if equipped with the Kh-31P
[50].
The missile has a cruise speed of 700 m/sec and a maximum speed of
1,000 m/sec. The Kh-31 can fly high- or low-level cruise profiles and can be
launched at altitudes of between 100 and 15,000m. Launch speeds range from
600 km/h (Mach 0.65) to 1 m 250 km/h (Mach 1.5). The high-level cruise
can be made at up to 15 km (50,000 ft), with a speed of M3.0. It is reported
that the Kh-31 missiles can fly at low level, down to 200 m altitude, cruising
at M 2.5 and can maneuver at 10g. The missiles can be programmed to climb
at a distance of 2 km from the target, and to dive down onto the target. The
missile is designed to be fired in salvos by one or several launch platforms.
Due to the threat represented by the Kh-31, a unique U.S.-Russian accord
was struck, for the acquisition of missiles by the U.S. Navy for test and trial
purposes. Beginning in 1994, under a joint program with McDonnell Douglas
(now Boeing), 13 MA-31 test vehicles were supplied for flight trials. The U.S.
has launched its MA-31 targets from QF-4 Phantoms. Follow-on batches
of MA-31 targets have since been delivered but U.S. access to the missile
has now been suspended. In 1998, it was reported that Zvezda-Strela had
developed an improved antiradar missile, using the export designation KR-
1, with a range increased to 400 km. This missile has been strongly linked
with China. Several photographs have now emerged of Kh-31s in Chinese

hands. These weapons have appeared in Chinese research facilities and have
also been mounted on a full-scale engineering mock up of the Xian JH-7
strike fighter. Chinese sources report that a small numbers of Kh-31Ps have
been purchased and used as the basis for an indigenously produced variant,
the YJ-91. It remains unclear how many Russian-built Kh-31s have been
supplied to China and how many (if any) have been assembled in China.
Some sources believe that China has now established its own Kh-31 (YJ-91)
production capability [50]. During the 1990s there were reports that an air-
to-air version of the Kh-31, designated Kh-31PD, was under development.
This was thought to be an antiradiation missile, for use against important
targets such as AWACS and JSTARS aircraft and others that use airborne
early warning radar.
The Kh-31 has been cleared for carriage on MiG-27 Flogger, MiG-29K
and MiG-29SMT Fulcrum, Su-17 Fitter, Su-24M Fencer, Su-25TM
Frogfoot, and the Su-30/Su-34 Flanker family aircraft. In June 2005 a
new upgrade program for the Su-27 to make it compatible with all versions
of the Kh-31 (and other advanced weapons) was revealed and is known as
the Su-27M1. The Kh-31 is also being integrated on Russian air force Su-27s
under the Su-27SM multirole upgrade program [50]. In 2006 an improved
version of the Kh-31P was under test at Russias Akhtubinsk weapons test
center (the Valery Chkalov State Fight Test Center, or 929 GLITs). This
weapon was identified as the Kh-31PMK. The designation of the Kh-31PMK
marks it as an export-dedicated program (K, Kommercheskaya, commercial).
In practice the K in export programs also stands for China (K, Kitai). The
Kh-31PMK is longer than a standard Kh-31P and its range is extended to
approximately 200 km, indicating the incorporation of additional fuel for the
ramjet. Integration of the Kh-31PMK on the Su-27SM would inevitably make
it available for China, and Chinas Su-30MKKs are already operating with
the Kh-31P [50]. Special versions to attack AWACS aircraft such as the E-3A
Sentry and an antiship version are also available.
16.5 United States

16.5.1 Shrike
After the time of the Korean conflict, the development of radar-guided surface-
to-air missiles (SAMs) added a new and lethal threat to U.S. aircraft. To
combat these new threats, the U.S. Navy started development of an ARM in
1958, with the designation ASM-N-10. This missile program would become
the AGM-45A Shrike, named for the predatory songbird Butcher Bird. The
Shrike was based on the airframe of the AIM-7 Sparrow and had a top speed
of Mach 2. The Shrike first saw combat in Vietnam in 1966 [4, 8].
Figure 16.29: Shrike missile AGM-45A (from [52]).
The concept behind the Shrike missile, shown in Figure 16.29 was that
the host aircrafts radar warning receiver was used to activate the AGM-45s
seeker head that, in turn, notified the pilot (by an audio tone in his headset)
that it had achieved a positive lock on the target signal. Following launch,
the weapons guidance section continuously monitored the threat signals di-
rection of arrival and generated the appropriate steering commands for the
missiles four mid-body steering surfaces [52]. This enabled the missile to
follow the radar beam down to the emitter and destroy it. This would disable
the SAM site, making it possible to destroy the SAMs themselves or to allow
a strike package to pass through the SAM sites airspace. The first Shrikes
were equipped with seekers optimized for E/F band emitters (24 GHz). As
other emitters arose, 10 additional seekers were developed to cover the dif-
ferent emitter bands including G-band (46 GHz) and I-band (810 GHz)
[53].
The Shrike first saw combat in Vietnam in 1966. The Shrike was used
by the Wild Weasels to suppress enemy air defense (SEAD). The Shrike had
better range than the gravity weapons being used and did not require the
aircraft to overfly the SAM sites to identify and destroy them. There were
however, limitations to employing the Shrike missiles. In order to lock on
to the target, the aircraft would have to fly directly at the SAM site. The
aircraft must have had the correct AGM-45 loaded. Also, the maximum oper-
ational launch range of the AGM-45A was limited to 16 km but progressively
increased to 46 km. Also, SAMs such as the SA-2 had a maximum speed of
Mach 3.5 compared to the AGM-45 maximum speed of Mach 2. This would
allow the SAM site to launch, guide to intercept, and turn o the emitter
before the AGM-45 could strike the site. Furthermore, the AGM-45 needed
to track the emitter until impact. If the emitter shut o, the AGM-45 would
not be able to guide to it [37].
Figure 16.30: Standard antiradiation missile (from [54]).
16.5.2 Standard ARM

The limited range and warhead of the AGM-45 led to the development of
a larger AGM-78 standard ARM as shown in Figure 16.30 [54]. The de-
velopment of AGM-78 combined the standard missile airframe (designed for
shipboard use) with the Shrike seeker. The standard missile was a radar-
guided surface-to-air missile used for ship defense by the Navy. The standard
ARM had top speed of Mach 2.5, with a maximum range at an altitude of 56
km. Due to its size, only two standard ARMs could be carried on an EA-6.
The Navy then decided to integrate an improved broadband gimbaled seeker
that allowed the aircrew to avoid flying directly at the target in order to fire
the missile. The standard ARM was also able to remember the elevation and
azimuth to the target if the emitter shut down. This did not guarantee the
emitter would be destroyed since a small amount of drift in the navigation
system would result in a miss. Although the AGM-78 was an improvement
over the Shrike, its large size limited it to being carried on large aircraft. In
addition, the standard missile airframe was complicated and costly to operate
and maintain.
16.5.3 HARM
The U.S. Navy began development of the AGM-88 high speed antiradar mis-
sile (HARM) shown in Figure 16.31 which was light weight allowing it to be
carried on U.S. fighters [55]. The top speed is described as over Mach 3. The
maximum range at altitude is 65 miles. The HARM has only one seeker,
which uses a broadband antenna to engage the emitters. The features of the
HARM are shown in the cutaway view in Figure 16.32 [54]. The HDAM
(HARM destruction of enemy air defense attack module) missile variant was
developed partly to address the LPI emitter. It successfully engaged a simu-
lated radar system that was radiating at a low power level. The new HDAM
variant adds inertial navigation system/global positioning system (INS/GPS)
Figure 16.31: High speed antiradiation missile (from [55]).
Figure 16.32: Cutaway drawing showing the HARM features (from [54]).
capability to the existing HARM.
16.5.4 AARGM
The AGM-88E advanced antiradiation guided missile (AARGM) demonstrates
a dual-mode guidance section on a HARM airframe (see Section 16.2.3). The
issue of shutdown is a major shortcoming in the SEAD element of the oen-
sive counter-air mission. The AARGM development is to produce an eective
and aordable lethal SEAD capability against mobile, relocatable, or fixed air
defense threats even in the presence of emitter shutdown or other anti-ARM
countermeasures.
The AARGM can be employed in the oensive counterair/SEAD role in
direct support of strike warfare, amphibious warfare, antisurface ship war-
fare, command and control warfare, and information warfare. The missile
design provides a new multimode guidance section and modified control sec-
tion mated with existing HARM propulsion and warhead sections. The new
guidance section is designed to have a passive antiradiation homing receiver

and associated antenna, a GPS/INS, and an active millimeter wave radar
for terminal guidance capability enabling the missile to engage and destroy
enemy air defenses in the event that these systems shutdown or employ other
electronic protection. AARGM is projected to have the capability to transmit
terminal data via a weapons impact assessment transmitter to national satel-
lites just before AARGM impacts its target. Also incorporated is a provision
to receive o-board targeting information, via the integrated broadcast sys-
tem. The AARGM acquisition objective is 1,750 missiles [56]. It will provide
USN and U.S. Marine Corps F/A 18 Hornet and EA-6B Prowler aircrews
with a significantly improved capability to search for, identify, and destroy
enemy air defense targets.
16.5.5 Aordable Reactive Strike Missile

The aordable reactive strike missile (ARES) is a derivative of the AGM-
88E advanced antiradiation guided missile (AARGM) under development for
the U.S. Navy and was expected to enter its inventory after 2008 [57]. ARES
would have a range greater than 50 nm and a speed of around Mach 3.0, mak-
ing it the only supersonic, tactical, GPS-guided strike weapon available.
The ARES concept features the same control section going into AARGM and
presents an attractive option for those nations that cannot aord to upgrade
their HARMs to the AARGM configuration. Additional options include a
semiactive laser seeker for human-in-the-loop control. In missions like pin-
point strikes in an urban setting this could allow parts of buildings to be
targeted without bringing down the entire structure and injuring civilians
and friendly troops. As with the AARGM, the new control section on ARES
will enable pilots to program exclusion zones in which the missile will not
strike. Aircraft operating the HARM or the AARGM, like the EA-6B, EA-
18, F-16, F/A-18, and Tornado, will not require software upgrades to carry
ARES. The new missile retains the HARMs rocket motor, airframe and war-
head, but adds the new control section and an all-digital passive seeker and
millimeter-wave active terminal seeker to the missiles front. This allows it to
accurately engage fixed-site and mobile air-defense radars that are emitting
or have shut down to avoid detection.
16.5.6 Sidearm
The United States has also developed ARMs for more defensive roles by
taking advantage of obsolete AIM-9Cs to develop the AGM-122 Sidearm
(Sidewinder ARM) as shown in Figure 16.33. The AIM-9C seeker was adapted
to be a broadband passive radar seeker. It also has a gravity bias function
added to its autopilot in order to facilitate loft launching from low-altitude,
low-speed platforms. The AGM-122A was developed for use on Marine he-
Figure 16.33: Sidearm ARM being loaded onto an aircraft (from [58]).
licopters (Bell AH-1W SuperCobra) to suppress air defense threats. The

AGM-122 was used as a defensive weapon, rather than in an oensive role
and entered service in 1989 [58]. AGM-122A is noted as having a 10.2 kg
high explosive and fragmentation warhead that is triggered by active laser
and impact fusing. Range is given as being approximately 8 km and func-
tionally, the weapon alerts the pilot of its host aircraft to lock-on via direction
of lock-on symbology on his head-up display and an audio tone in his head-
phones [59]. The AGM-122 has a 10-mile range. Once the stock of AIM-9Cs
was depleted, the Navy considered building new AGM-122s, but the program
never happened. While Sidearm is less capable than modern antiradiation
missiles (like AGM-88 HARM), it is still a cost-eective alternative against
low-tech threats [60].
16.5.7 Rolling Airframe Missile

The U.S. Navy has developed the rolling airframe missile (RAM) for ship
self defense against incoming antiship cruise missiles (ASCMs). The RAM,
designated the RIM-116, was a joint venture between the United States and
Germany to develop a low-cost self defense system. The RIM-116 shown in
Figure 16.34 uses the 5 inch rocket motor and warhead technology from the
AIM-9 Sidewinder. The missile is capable of maneuvers up to 20g in any di-
rection. The RIM-116 seeker is a dual-mode, passive radio frequency/infrared
seeker. Initial guidance is provided by the passive RF seeker on the ASCMs
RF emissions. If the ASCMs IR radiation is acquired, RAM transitions
to IR guidance. Originally, the missile was cued by the ships ES suite or
radar [61].
More recently, the RAM has replaced the 20 mm Gatling gun of the
Phalanx system. This new marriage of RAM and the Phalanx system is
Figure 16.34: Rolling airframe missile launch [61] ( c 2007 Janes Information
Group).
called the SeaRAM and combines RAMs superior accuracy, extended range,
and high maneuverability with the Phalanx high resolution radar systems.
SeaRAM shown in Figure 16.35, is essentially a Phalanx Block 1B but with
the gun replaced by the 11-cell launcher and is intended to extend ship self-
protection to ranges of 4 km and can enable prosecution of low Doppler
targets. In addition to providing an on-mount J-band (12 to 18 GHz), digital
MTI search radar and pulse Doppler monopulse tracker radar, there is also
the electro-optical sensor used in the latest Phalanx systems for surface target
detection.
16.5.8 Army UAVs

The U.S. Army is exploring potential requirements for an antiradiation mis-
sile for carriage by its larger unmanned air vehicles (e.g., AAI RQ-7B Shadow
200 UAV) to counter hostile UAV systems [62]. The concept calls for the an-
tiradiation missile to target a hostile UAV systems ground control station by
following its command datalink. The option is one of a number of new weapon
concepts being studied by the service as it prepares for the introduction of its
new General Atomics Sky Warrior and Northrop Grumman RQ-8B Fire Scout
UAVs. The development of an anti-UAV capability anticipates that UAV sys-
tems will continue to be an important operational target. Strategies such as
targeting unmanned aircraft to take away the intelligence, surveillance and
Figure 16.35: SeaRAM missile launch [61] ( c 2007 Janes Information

Group).
reconnaissance capability is a significant tactical response. Other anti-UAV

capabilities being studied include the use of electronic warfare techniques such
as countertargeting and jamming of the command datalink.
16.6 France
The French introduced the Armat (Anti-Radar Matra) in 1984 which was an
evolved variant of the antiradiation version of the French-British BAe-Matra
AS-37 Martel missile. By using the AS-37 airframe and replacing the seeker
and associated electronics with new and improved versions with added elec-
tronic protection, the missile shown in Figure 16.36 was given the capability
to overcome decoys and jamming techniques including long radar switch-o
periods. The microprocessor based seeker homes on to a programmed emitter
and uses inertial midcourse guidance. Several interchangeable homing heads
are used to cover the wide spectrum of target radar frequencies including L-
band (5002,000 MHz), S-band (24 GHz), C-band (48 GHz) and X-band
(812 GHz) [63]. The missile can be launched from high or low levels and
will home onto the radar or jamming transmissions of the pre-selected target
radar. After lock-on of the missile seeker, the location, radar parameters and
launch success zones are displayed to the aircrew who can then select the
best launch time. With its high launch weight, heavyweight warhead and
long range, the Armat is primarily an oensive strategic ARM designed to
Figure 16.36: AS-37 Martel.
destroy early warning and ground control intercept radars. This is where it
diers fundamentally from the HARM and the ALARM, which are built to
also perform as defensive ARMs carried as part of a mixed weapon load. The
missile has been cleared for carriage on Jaguar, Atlantique, Mirage F1, Mi-
rage III and Mirage 2000 aircraft. The missile is believed to have a maximum
range of 100 km when launched from high altitude. An improved version has
also been produced, the Armat-D, which is fitted with an updated passive
homing seeker.
16.7 United Kingdom

In the early 1980s, the British conducted a study to replace their aging AS-
37s. The British decided on the Alarm for several reasons. The United
States developed HARMs for specially equipped aircraft, such as the F-4G
Wild Weasel and the EA-6. Both of these aircraft are equipped with sophisti-
cated and complex emitter locating systems. Since the British have a smaller
number of aircraft than the U.S., they did not want to limit the role of any
aircraft by specially equipping them. With the smaller numbers, the British
required that all of their fighters be able to conduct any mission. The British
also did not wish to incur the cost and in-service support overheads of deploy-
ing and maintaining these sophisticated and complex systems. Additionally,
the British thought that adopting the Weasel operational model would expose
SEAD aircraft to attack more frequently, thus incurring high loss rates. The
Alarm can be carried on all British fighters, such as the Harrier and Tornado,
which carry up to nine Alarms. Figure 16.37 shows the launch of an Alarm
from a Tornado [63].
The Alarm has a range of 28 miles and has five launch modes: direct,
dual, loiter, universal, and area suppression. A unique feature to the Alarm
is it parachute system for loitering over the target area. In the dual and
loiter modes, the missile climbs to a high altitude above the target area and
Figure 16.37: Alarm missile firing from an RAF Tornado (from [63]).
searches for the enemy emitter. If one is detected, the missile dives on the
target. If not, it deploys a parachute and listens for the enemy emitter to
come up. Once the radiation is detected, the missile jettisons the chute and
dives on to the emitter. The parachute system allows the ALARM to loiter
for several minutes. With ALARM-equipped aircraft in the area, enemy SAM
sites would not know when it was safe to turn on their radars, thus suppressing
their ability to deter attacking aircraft.
Figure 16.38 shows the modes used by the Alarm. The Alarms seeker is
similar to that of the HARM with a microprocessor controlled passive homing
receiver, designed to locate and identify the characteristic Pulse Repetition
Frequencies (PRF) of programmed threat emitters [63]. The Alarm has a
wideband RF antenna/receiver and a conventional quartet of cavity backed
spiral antennas, forming a fixed two axis interferometer with lower mid-band
to hi-band coverage. Like the HARM, the Alarm has logic to select the highest
value alternate target, should the primary target go o the air.
16.8 Taiwan
Taiwan is intensely concerned about Chinas growing air power dominance
of the Taiwan straits. Having an eective ARM capability hinders Chinas
ability to conduct air operations from secure bases on the mainland. Since
no country will export ARMs to Taiwan, its air force has developed its own
ARM, known as the Tien Chien IIA. Tien Chien translates to sky sword.
The Tien Chien IIA replaces the active radar seeker of the Tien Chien II
air-to-air missile with a passive antiradar seeker and guidance section. The
seeker is housed in a reprofiled, notched fairing, giving rise to speculation that
it may use a dual-mode design, incorporating both passive RF and infrared
sensors [64].
The Tien Chien IIA equips the Republic of Chinas Air Forces AIDC F-
CK-1 Ching Kuos (otherwise known as the IDF, Indigenous Defense Fighter).
Figure 16.38: Alarm modes of attack (from [63]).
Figure 16.39: Tien Chien 11A being carried by an indigenous defense fighter.
According to the Chung-Shan Institute of Science and Technology (CSIST),

any Ching Kuo can carry the new ARM, as it requires no modifications to
the launch aircraft. The ROCAF does not intend to field a dedicated Wild
Weasel defense suppression force armed with the Tien Chien IIA. Instead,
the capability will be introduced across the front-line fleet. Up to four Tien
Chien IIAs can be carried by one aircraft on individual pylonswithout losing
the existing hardpoints for two Tien Chien IIs plus two wingtip-mounted Tien
Chien Is. Tien Chien IIA can be carried on any aircraft equipped to carry
the Tien Chien II as shown in Figure 16.39.
Figure 16.40: German ARMIGER [65] ( c 2007 Janes Information Group).
16.9 Germany
Germany, a longtime user of the HARM, is developing the ARMIGER (antira-
diation missile with intelligent guidance and extended range). The Germans
have been participating in the development of the international HARM up-
grade program, the AGM-88D. The improvements consist of software and
hardware upgrades including replacing the original mechanical gyros with a
state-of-the-art GPS/IMU. The addition of GPS to the HARM would correct
the long-standing problem of ARMs of what to guide on if the emitter shuts
down. GPS allows you to fly to a certain point when the target is not emitting
at all. However, the U.S. Navy decided not to proceed with the project. The
Germans decided to proceed with the ARMIGER program as a replacement
for the HARM. Due to concerns over whether it is wise to develop a single
purpose weapon, the German Luftwae has decided to proceed slowly with
the ARMIGER [65].
The ARMIGER is roughly the same weight as the HARM. The ARMIGER
will have a GPS/IMU, as would the AGM-88D, to overcome the ARM
problem when the emitter shuts down. In addition to the GPS/IMU, the
ARMIGER will have a new technology passive radar/high-resolution imag-
ing infrared dual-mode seeker (called ARAS). This dual-mode seeker will also
combat the problem of an emitter shutting down while the missile is in flight.
Typically, the ARMIGER would be launched using the passive radar and
switch over to the infrared for terminal tracking if the targeted emitter shuts
down. In addition to the seeker, the ARMIGER will have improved range
over the HARM. The GPS will bridge the gap between the time when the
passive radar seeker loses track of the radar emitter and the point when the
enemy radar is detected by infrared. Combined GPS information collected
by multiple aircraft can be used to find the position of enemy radar with the
information then passed to the ARM prior to its launch from the aircraft [65].
16.10 Israel
16.10.1 Harpy
Dedicated for the SEAD mission, Harpy is an operational loitering attack
weapon. The current version of Harpy is deployed as a fire and forget weapon.
In order to verify the drones operational capability, its seeker head is being
tested by a special radar simulator just before launch, to ensure that all sys-
tems are working. It patrols the assigned area, and will attack any hostile
radar activated in its vicinity. When used in appropriate numbers, Harpy
can be launched into a target area to support continuous operations, or time
limited strike packages. Unlike antiradar missiles such as HARM, whose
speed, range and direction of approach are predictable, the killer-drone de-
ployment is more flexible and unpredictable, and therefore, conventional coun-
termeasure techniques are not useful against it. The Harpy system shown in
Figure 16.41 is designed to operate multiple munitions simultaneously over
a specific area, to eectively cover the target. Each drone is deployed au-
Figure 16.41: Harpy antiradar UAV being launched from a truck canister [66]
( c 2007 Janes Information Group).
tonomously, without interference and overlapping the other drones [66].

The Harpy mission is planned and programmed in the ground control cen-
ter, as an independent mission, or planned in accordance with other manned
or unmanned systems. Prior to launch, individual weapons are programmed
and tested, to verify their operational readiness. After the rocket-assisted
launch, the drone flies autonomously in route to its patrol area, predefined
by a set of navigational waypoints. Due to its low speed and economical fuel
consumption, the drone can sustain a mission of several hours over the target
area. Its radar seeker head constantly searches for hostile radars. Once an
enemy radar is acquired, Harpy compares the signal to the library of hostile
emitters, and prioritizes the threat. If the target is verified, the drone enters
an attack mode and a near vertical dive homing in on the signal. The attack
sequence is shown in Figure 16.42. The drone is set to detonate its warhead
Figure 16.42: Harpy UAV attack sequence against an emitter [66] ( c 2007
just above the target, to generate the highest damage to the antenna, and sur-
rounding facilities. If the radar is turned o before Harpy strikes, the drone
can abort the attack and continue loitering. If no radar was spotted during
the mission, the drone is programmed to self-destruct over a designated area.
Follow-on systems are calling for a combination of seeker and killer drones
that will enable visual identification and attack of targets even after they turn
o their emitters.
Current Harpy canisters are installed on trucks, and can be carried by C-
130 transport aircraft. Each truck carries 18 weapon launchers. Each battery
of Harpy is composed of three trucks, capable of deploying up to 54 drones
for simultaneous, coordinated attack. The battery also has a ground control
station and logistical support element. The system can also be deployed
from the decks of assault landing ships, in support of marine or amphibious
operations. Harpy is currently operational with the Turkish, Korean, Chinese
and Indian Armies, in addition to the Israel Air Force. In December 2004
China was reported to be interested in an upgrade of its systems to a more
advanced version.
Figure 16.43: STAR-1 antiradiation missile [67] ( c 2005 Janes Information

Group).
16.10.2 STAR-1
Israel Military Industries (IMI)formerly TAAShas developed its Delilah
air-launched decoy into a long-range, lightweight cruise missile. Described
by its manufacturer as an advanced air-to-ground stando powered UAV,
the Delilah has a range capability that takes it out of the tactical category
and into the realms of the cruise missile. Furthermore, the ocial maximum
range quoted for the system is 250 km. The original Delilah decoy was derived
from the US MQM-74 Chukar aerial target, that entered service in the mid-
1960s. The first reports that Israel had developed the Delilah air vehicle as
an oensive weapons system emerged in 1995 [67]. Since then, the Delilah has
evolved into a modular air strike weapon with a range of possible applications.
Driving the design of the Delilah system was an emphasis on single-pilot
operations. The weapon is programmed on the ground with key parameters
such as waypoints and flight altitudes, but a datalink gives the launch aircraft
the ability to retask the missile in flight. During the mid-1990s a long-range
antiradar defense suppression variant of the Delilah was actively marketed
under the designation STAR-1. This program was linked with China in several
reports but has since disappeared from view and is no longer included in IMIs
ocial product portfolio. To give the Delilah a SEAD capability, a broadband,
218-GHz, passive radar seeker with an INS/GPS mid-course update system
was incorporated into its basic design. The STAR-1 shared the same size and
weight of the Delilah, and used the standard 30 kg warhead. The STAR-1 as
shown in Figure 16.43, would have both a direct attack mode and the ability
to loiter over a target area waiting for hostile radars to start transmitting, or
to reattack radars if they restart transmissions after an attack. A STAR-1
variant (or equivalent) is believed to be in service in Israel and was oered
for export as far back as 1995. In 2004 an IMI representative said that a
dual-mode antiradiation seeker was available for the Delilah, but that it was
not being oered as a product yet. It is understood that the STAR-1 can be
both air-launched and ground-launched [67].

IMI has now developed several versions of the Delilah, which can be
launched from land, sea and air. The standard Delilah missile can be fit-
ted with interchangeable FLIR (forward looking infrared) or electro-optical
seekers, in separate seeker assemblies. The seekers have a target auto-tracking
capability but not yet an automatic target recognition function. In tests the
Delilah has proved its ability to hit a target moving at 50 km/h, and IMI
is promoting the system for use against time-critical targets such as mobile
SSMs or SAMs. The existing seekers are capable of identifying targets at
ranges of up to 10 miles. Operators have found that moving targets are
easier to locate and identify than static ones. The missiles datalink capa-
bility allows for man-in-the loop control, to confirm the final target. If this
target is not confirmed, or if datalink communications are lost, the missile
has a default navigation mode to fly around the target and reengage. The
Delilah carries enough onboard fuel to fly for a maximum of 22 minutes, so
the weapon is optimized for high-altitude straight line cruise profiles. A 30-kg
high explosive warhead is currently fitted, but IMI confirms it is working on
a new penetrating warhead option for hardened target attacks [67].
16.11 China
The China National Precision Machinery Import and Export Corporation
(CNPMIEC) oers the FT-2000 (Chinese Fei Tung = FT), which is an ex-
port variant of the antiradar surface-to-air system specifically for use against
airborne early warning, command and control, and EA aircraft. It is believed
that the missile is an upgrade of the SA-10 design. The missile system has the
Chinese designator Hong-Qi-12 (HQ-12)[68]. The missile contains a wideband
surface-to-air passive seeker to engage either single or multiple radiating air-
borne targets that radiate in the 218 GHz band such as the Airborne Early
Warning and Command System (AWACS) or Suppression of Enemy Air De-
fense (SEAD) EA-6B/EA-18G EA aircraft. The missiles are also capable of
detecting and locking on to random electronic interference and jamming.
With primarily Russian technology and with minimal use of imported
components, China has mastered the production of air defense missile systems
such as the HQ-15 missile. This missile is shown in Figure 16.44 along with
its transporter-erector launcher vehicle. A photograph of the passive radar
seeker assembly for the HQ-15 missile is shown in Figure 16.45. A typical
HQ-12 battalion has three batteries. Each battery has four ES vehicles, three
transporter erector launchers (TELs), one command launch center, and three
transporter/loader vehicles. The ARMs are vertically cold launched from the
TEL Taian TAS5380 that is also an 8 8 transport vehicle as shown in Figure
16.45. The TEL has four missile canisters that are raised to the vertical for
launching. The four ES stations are deployed at distances of up to 30 km in a
Figure 16.44: Outline drawing and launch system for the Chinese HQ-15
missile.
Figure 16.45: Photograph of the passive radar seeker assembly for the Chinese
HQ-15 missile.
triangle with a central fourth unit acting as command and control. The four
ES receiver vehicles associated with each battery can track up to 50 targets.
The complete FT-2000 system includes a wideband passive radar detecting
station, the specially developed ARM vertically launched missile and a four-
round launcher platform. A first test launch was reported in September 1997.
The passive seeker has a memory, for use if the target radar is switched o,
and a home-on-jam capability. The missile has a maximum speed of 1.2 km/s.
Targets can be intercepted at altitudes between 3 and 20 km. The proximity
fuse is activated 5 km from the target with a range of 35m.
The second version, known as FT-2000A in its export version, was re-
ported to have a passive radar seeker covering the 26-GHz (S and C-band)
range, that has its frequency selected on the ground before launch. This mis-
sile has a maximum range of 60 km, and can be used as part of the HQ-2
system but requires separate launchers and fire-control units. This system is
still in the developmental stages but is expected to be a static weapon system
[69]. The FT-2000B version has been designed for use as an upgrade to the
HQ-12 missile system, with the maximum range increased to 120 km and with
a new 118-GHz passive radar seeker. The system has been oered for export
with the potential first customer Pakistan. During discussions between China
and Pakistan in February 2004, the oer was made by China to supply the
FT-2000/FT-2000A to counter the Indian threat to Pakistan of the Indian
Agni missile systems. Batteries have been reported around Beijing and in
Fujian province.
16.12 Anti-ARM Techniques

The earliest form used to defeat the ARM (and still the most common
method) is radar position flexibility. In most cases the radar position be-
ing attacked is provided by an electronic intelligence (ELINT) system prior
to aircraft takeo. The ability of the radar to set up, tear down, and move
to a new location within a few minutes can help to hide the radar position.
The latest generation of SAMs have put a much higher premium on system
mobility. That is, the ability to leave in a hurry is closely linked to survival
in modern warfare. Most land-based surveillance, ballistic missile detection
and weapon-related radars currently in development claim to have relatively
high mobility, including some of the very powerful long range systems such as
Marconis latest version of Martello, the S 753, and the Israeli Arrow Green
Pine antiballistic missile radar.
Rapid relocation however, has its limits. If a surveillance radar is attacked
by an ARM fired from below the radar horizon, or by an aircraft approaching
low over the horizon, the radar has approximately 20 seconds to move to a
safe location. Consequently, manufacturers are looking to increase mobility,
with multifunction radars using planar-array antennas, mounted on a truck
(including all of the electronics), and with a microwave link to relay target
data back to command and control. They will also use highly sophisticated
land navigation systems to provide an accurate position reference. This is
important if their target data is to be tied into an overall battlefield awareness.
Radar methods include twinkle (or blink) transmission. In this technique,
the o time is much greater than the on time. This makes it dicult for the
ARM to keep track of the emitters signal (and location). In this method
of protecting a pulse radar from an ARM missile attempting to home in
on interrogating pulses emitted, a number of decoys at dierent locations are
deployed in the vicinity of the pulse radar. Each one of the decoys are adapted,
when activated, to emit pulses of a given amplitude and duration; activating,
when each one of the interrogating pulses is generated in the pulse radar. A
selected decoy is chosen to lead the remaining decoys for a period of time
determined by range and range rate measurements of a pulse Doppler missile
warning sensor. The decoy selected is changed to lead the remaining decoys
at a time determined from the range and range rate measurements of a pulse
Doppler missile warning sensor thereby defining a blink rate associated with
the decoys, to form a covering pulse overlapping the then emitted one of the
interrogating pulses, and adaptively changing the blink rate. Two examples
of radar using this transmission control technique for anti-ARM include the
AN/APY-1 Cosmic Shield and the AN/MPQ-53 Patriot [70].
Another technique often used is called the snap-and-shoot method. In this
technique, a fire control radar is assigned to intercept and track targets. The
remaining fire control radar systems receive the target flight path parameters.
When the tracked targets enter the fire range, the unit snaps open and makes
the response. Other methods that may be used include using other types
of radiation to detect and track the targets. This may include using visual
television with infrared measurements of the targets range. Examples of this
include the Swedish GLV200 and the Swiss Air Guard. Also, the use of very
high frequency (VHF) band or ultra high frequency (UHF) bands can be used
to avoid the ARM attack [70]. The reason for this is that the diameter of the
ARM body is limited with the aperture of the ARM antenna greater than the
wavelength. For example, with a diameter of 40 cm, it is dicult to target
radar with frequencies below 1 GHz.
16.12.1 Decoys
Dispensing of active decoys is also an important method for anti-ARM [71].
The ARM decoy has characteristics that are coincident with the radar. For
example, the decoy has the same eective radiated power (or ERP) and carrier
frequency and the transmission waveform is synchronous with the radar. The
decoy also emits a decoy pulse 0.10.2 s ahead of the radar pulse so that the
ARM triggers its guidance on the wrong waveform. Typically, the distance
between the radar and the decoy is 100300m with the spatial angle between
Figure 16.46: Flaps technology for decoying an ARM (from [72]).
radar and decoy smaller than the ARM track angle resolution.
An ARM decoy antenna that uses low-windload FLAPS (flat parabolic
surface) technology is shown in Figure 16.46 [72]. Once the incoming ARM is
detected, the radar is turned o and the reflectors are illuminated by a remote
feed. Since the frame is staked to the ground, it can survive an ARM blast
from any direction. The blast travels through the aperture then the reflector
springs back to is operating position. The antenna uses FLAPS technology
to enable a geometrically flat surface as shown in Figure 16.47(a) to behave
electromagnetically as though it were a parabolic reflector as shown in Fig-
ure 16.47(b). The FLAPS reflector is a thin (planar or conformal) surface
consisting of an array of dipole scatterers. The elemental dipole scatterer as
shown in Figure 16.47(c), consists of a dipole positioned approximately 1/8
wavelength above a ground plane. Here, a crossed shorted dipole configu-
ration is shown with each dipole controlling its corresponding polarization.
Incident RF energy causes a standing wave to be set up between the dipole
and the ground-plane [72]. The dipole itself possesses an RF reactance which
is a function of its length and thickness. This combination of standing-wave
and dipole reactance causes the incident RF to be reradiated with a phase
shift, which can be controlled by a variation of the dipoles length [72].
The integrated ARM warning radar and decoy deployment method is shown
in Figure 16.48 and consists of an integrated system of advanced ARM de-
tection radar and general purpose distributed decoys to protect the ground
air-defense radar [73]. The ARM detection radar is used to detect and iden-
Figure 16.47: FLAPS antenna technology showing (a) a thin planar surface
consisting of an array of radiating and phase shifting elements, (b) a conven-
tional reflector, and (c) the schematic of an elemental dipole scatterer used
in the FLAPS antenna technology (from [72]).
tify the attacking ARM (RCS = 0.1 m2 ), which triggers a shutdown of the
sensors and cues the crew manning the site to leave. Studies also indicate
that by using changeable sample ratios within the radar receiver, a higher
probability of ARM detection can result [74].
ARM detection radars with multiple antennas using VHF have also been
reported [75]. The use of VHF enhances the ARM RCS significantly increas-
ing the probability of detection. Extended coherent integration and dedicated
signal processing can also be used. At the same time, the ARM messages are
sent to a series of decoys. The signal radiated by the decoys guide the ARM
(speed 2 to 4 Mach) to a preset safety area. If the ARM fails to continue
the attack, a cancellation of the alarm is made and protected radar triggered
to restart. The anti-ARM warning radars frequency band selection (UHF,
VHF) is to give an antistealth capability, ground clutter and weather sup-
pression and to also control the deployment of the decoys. It must be highly
mobile to operate in the sometimes rough terrain and also have good target
identification capability. The PRF of the warning radar should be as large as
possible with a compressed pulse width as small as possible to decrease the
energy in the range-Doppler detection cell.
With high-speed ARM detection (>1.5 Mach), the separation of the mis-
sile from the launch aircraft can be detected with the warning radar using
pulse-Doppler waveforms and frequency agility [73]. For slow ARM targets
such as UAV ARMs and cruise ARMs, an accurate radial velocity and unam-
biguous range of the target must also be reported.
To avoid turning the radar o early and deploying the decoys, two trans-
mitters T1 and T2 can be used that are placed far away from the protected
Figure 16.48: Deployment of the integrated anti-ARM system (adapted from

[73]).
radar with the receiver placed near the protected target as shown in Fig-
ure 16.48. They should be connected by an RF optical fiber communication
link. When the ARM approaches, T1 and T2 can work together to protect
each other. Their use of polarization and frequency diversity can provide the
means to reduce the power of each solid state transmitter by 35 dB. The
two transmitters are noncoherent and can be placed at dierent heights to
prevent lobe splitting [73]. The antenna should be a small foldable, nonrota-
tion antenna array. Because the distance between T1 and T2 and the receiver
cannot be too far, the use of othogonal waveforms such as those discussed in
Chapter 10 are required.
Another type of decoy is the simulated modelin eect, a cardboard cut-
out that looks like a radar. These can be very eective, and it is possible
to metalize them to provide a radar return that looks like a gun, a tank, or
a radar station. This type of decoy has been used very eectively against
ground attack aircraft threats, and could contribute toward the protection
from the active radar homing phase in a dual-mode ARM.
16.12.2 Gazetchik
The Iraqi newspaper Al-Qabas Daily, in July 2000, reported that Iraq had
acquired from Russia a jamming device that was capable of neutralizing U.S.
ARMs during the enforcement of the no-fly zones, making the missiles miss
their targets. The first two systems were reported to be a gift to Baghdad
by the Russian ultra-nationalist leader Vladimir Zhirinovsky. The Gazetchik
anti-ARM system built by the All-Russian Radio Engineering Research Insti-
Figure 16.49: Russian Gazetchik anti-ARM system.
tute, Moscow, is designed to protect radar emitters such as the GAMMA-DE

(67N6E) (a mobile 3-D solid-state phased array surveillance radar) from at-
tack by ARMs.
The Gazetchik system shown in Figure 16.49 consists of a stand-alone
detector, active radio frequency decoys, a set of passive countermeasure dis-
persers (not shown), and an interface with the radar or radars being defended.
The ARM detector unit alerts the system to the approach of an incoming
ARM weapon. Then a warning is initiated that activates a host of responses
including interruption of the protected emitters transmission, transmission
of RF decoy signals on the protected emitters operating frequency and the
firing of passive decoys from the equipments cha and aerosol launchers [75].
With coverage of up to 90 degrees elevation and 360 degrees azimuth, the
system operates autonomously and draws its power from the radar systems
being protected. Gazetchik is reported to be available in a number of variants
and has an automatic operating mode if required.
16.12.3 AN/TLQ-32 ARM-D Decoy

AN/TLQ-32 ARM-D is a lightweight, ruggedized, tunable magnetron that
imitates the AN/TPS-43E and AN/TPS-75 radar signals. It is designed to
protect radars in the field from ARMs that are guided by homing in on
the radars own transmission signals. The ARM-D built by ITT Gilfillan,
provides protection to the radar by emulating the transmission characteristics
of the host radar, thereby deceiving and confusing the incoming missile. The
decoys are placed on the ground as shown in Figure 16.50 in triangulation
to attract enemy antiradiation missiles, and ultimately saving the radar site.
Figure 16.50: AN/TLQ-32 anti-ARM radar transmitter [76] ( c 2004 Janes

Information Group).
Features of ARM-D include its capability to emulate frequency-agile radars;

360o coverage; protection of both the radar and the decoy assets against
ARMs; lightweight fiber optic interface between the radar and decoy emitter
groups and low prime power operation. It also features rugged, lightweight
modular packaging, extensive built-in test capability and rapid set up and
tear down. It is claimed that the decoy can be transported by two people
with individual decoys being deployable within 15 minutes. In operational
use, three decoys are allocated to each radar system. The surveillance decoys
are designed to be capable of protecting the radar site from multiple missile
launches, whether simultaneous or consecutive [76].
The AN/TLQ-32 ARM-D was selected by the USAF in March 1989, with
a contract for two first article examples being awarded during the following
September. Testing of these began in May 1992 and full-scale production of
14 systems to protect USAF AN/TPS-75 radars began in December 1992.
During 1996, additional TLQ-32 systems were delivered to the U.S. Air Na-
tional Guard.
References
[1] Bolkcom, B., Military suppression of enemy air defenses (SEAD): Assessing
future needs, CRS Report for Congress, RS21141, Sept. 23, 2004.
[2] Farroth, A., and Krishnamurthy, V., Optimal threshold policies for hard-
kill of enemy radars with high-speed anti-radiation missiles (HARMs), Proc.
of the International Conference on Acoustics, Speech and Signal Processing,
14-19 May, 2006.
[3] Department of the Army, Field Manual 101-5: Sta organization and opera-
tions, May 1997.
[4] Davis, L., Wild Weasel: The SAM Suppression StoryVietnam, Squadron
Signal Publications, July 1993.
[5] Friedman, N., US Naval Weapons, Conway Maritime Press, 1983.

[6] Parsch, A., Corvus, Encyclopedia Astronautica,
www.astronautix.com/lvs/corvus.htm
[7] Gunston, B., The Illustrated Encyclopedia of Rockets and Missiles, Salaman-
der Books Ltd, 1979.
[8] Streetly, M., Airborne Electronic Warfare: History, Weapons and Tactics,
Janes Information Group, 1988.
[9] Pietrucha, M., Starbaby: A quick primer on SEAD, Defense IQ Airborne
Electronic Warfare Conference, London, England, Aug. 2006.
[10] Levin, R. E., Electronic WarfareComprehensive strategy needed for sup-
pressing enemy air defenses, United States General Accounting Oce Report
to Congressional Requesters, GAO-01-28, Jan. 2001.
[11] Kopp, C., Support jamming and force structure, The Journal of Electronic
Defense, May, 2002.
[12] Lum, Z., Hardcore Hard Kill: Seeds of a New SEAD, Journal of Electronic
Defense, February 1997.
[13] McKenna, T., Poisoned Arrows, The Journal of Electronic Defense, March
2004.
[14] Zaloga, S. J., The evolving SAM threat: Kosovo and beyond, The Journal
of Electronic Defense, May, 2000.
[15] Lambeth, B. S., Kosovo and the continuing SEAD challenge, Air Power
Journal, Summer 2002.
[16] Martin, N. M., Nandagopal, D., Kara, M., Tran, V. N., and Hamilton, S.
Body fixed antenna options for seekers, Proc. of the IEEE International
Conference on Radar, pp. 272275, Oct. 1992.
[17] Cencich, T. and Human, J., The analysis of wideband spiral antennas using
modal decomposition, IEEE Antennas and Propagation Magazine, Vol. 46,
No. 4, Aug. 2004.
[18] Muller, D. J., and Sarabandi, K., Design and analysis of a 3-arm spiral
antenna, IEEE Trans. on Antennas and Propagation, Vol. 55, No. 2, pp.
258266, Feb. 2007.
[19] Mayes, P. E., Frequency-independent antennas and broad-band derivatives
thereof, Proc. of the IEEE, Vol. 80, No. 1, pp. 103112, Jan. 1992.
[20] Thaysen, J., Jakobsen, K. B., and Appel-Hansen, J., A logarithmic spiral
antenna for 0.4 to 3.8 GHz, Applied Microwaves & Wireless pp. 3245, 2001.
[21] Stutzman, W. L., and Thiele, G. A., Antenna Theory and Design, John Wiley
& Sons, Inc. New York, 1997.
[22] Balanis, C. A., Antenna Theory Analysis and Design, Harper & Row, Pub-
lishers, New York, 1982.
[23] Salmond, W. E., High accuracy broadband antenna system, U.S. Patent
4,095,230, June 13, 1978.
[24] Salmond, W. E., Common aperture dual mode seeker antenna, U.S. Patent
4,348,677, Sept. 7, 1982.
[25] Corzine, R. G., Bolstad, B. E., and Johantgen, J. S., Broadband polarization
diversity monopulse antenna, U.S. Patent 5,021,796, June 4, 1991.
[26] Bohlman, W. A., and Schuchardt, J. M., Dual polarized ambidextrous mul-
tiple deformed aperture spiral antennas, U.S. Patent 5,227,807, July 13,
1993.
[27] Neri, F., Introduction to Electronic Defense Systems, 2nd ed., Artech House,
2001.
[28] Hughes, R. S., RF detector logarithmic video amplifier, Microwave Journal,
vol. 32, no. 8, pp. 137148, Aug. 1989.
[29] Janes Information Group, Two-for-one guidance could steer future anti-
radar missiles, Janes Missiles and Rockets, Jan. 2005.
[30] Klass, P. J., New anti-radar missile uses dual-mode seeker, Aviation Week
and Space Technology, pp.60 Oct., 26 1998.
[31] Penz, P. A., Katz, A., Gately M. T., Collins, D. R., and Anderson J. A.,
Analog capabilities of the BSB model as applied to the anti-radiation homing
missile problem, IEEE Conference, pp. II-7 II-11.
[32] Lee, J. P. Y., Circuit for LPI signal detection and suppression of conventional
pulsed signals, U.S. Patent 6,388,604, issued May 14, 2002.
[33] Shirman, Y. D., Orlenko, V. M., and Seleznev, S. V., Passive detection of
the stealth signals, Proc. of the European Radar Conf., Amsterdam, pp.
321324, 2004.
[34] Shirman, Y. D., Orlenko, V. M., and Seleznev, S. V., Present state and ways
of passive anti-LPI radar implementation, Proc. of the International Radar
Symposium, pp. 14, 24-26 May, 2006.
[35] Jie, S., Xiao-ming, T. and You, H., Multi-channel digital LPI signal detec-
tor, Proc. of the International Conf. on Radar, pp. 14, Oct. 2006.
[36] Kopp, C. Texas Instruments (Raytheon) AGM-88 HARM, Air Power In-
ternational, Vol. 4, No. 1, Dec. 1998.
[37] Fiszer, M. and Gruzczynski, J., Crimson SEAD, The Journal of Electronic
Defense, pp. 44 56, Oct. 2001.
[38] Barrie, D., Silent Hunter, Aviation Week and Space Technology, pp. 36,
July 26, 2004.
[39] http://www.fas.org/man/dod-101/sf/missile/row/aa-10.htm
[40] http://www.fas.org/nuke/guide/russia/bomber/as-4.htm
[41] Janes Strategic Weapon Systems, Kh-22 (AS-4 Kitchen/Burya), Sept.
2007.
[42] Janes Strategic Weapon Systems, KSR-2P (AS-5 Kelt), Sept. 2007.
[43] Janes Strategic Weapon Systems, KSR-5P (AS-6 Kingfish), Sept. 2007.
[44] Janes Air Launched Weapons Kh-28 (AS-9 Kyle), May, 2006.
[45] http://www.testpilot.ru/russia/raduga/kh/58/images/kh58 1.jpg.

[46] Janes Strategic Weapon Systems, Kh-58 (AS-11 Kilter), April, 2006.
[47] Janes Air Launched Weapons, Kh-25MP/MPU (AS-12, Kegler), Aug.
2006.
[48] http://www.fas.org/man/dod-101/sys/missile/row/as-12.htm
[49] Janes Strategic Weapon Systems, Kh-15 (AS-16 Kickback/RKV-15), Sept.
2007.
[50] Janes Strategic Weapon Systems, Kh-31P (AS-17 Krypton), Sept. 2007.
[51] http://www.military.cz/russia/air/weapons/rockets/agm/ch31/Ma-31.jpg.
[52] http://www.raf.mod.uk/falklands/images/bbmartel.jpg.
[53] Janes Electronic Mission Aircraft, Raytheon Systems AGM-45 Shrike Anti-
Radiation Missile, Dec. 2002.
[54] http://www.fas.org/man/dod-101/sys/missile/agm-78.htm.
[55] http://common.wikimedia.org/wiki/image:agm-88 harm on f-4g.jpg.
[56] http://www.globalsecurity.org/military/systems/munitions/aargm.htm.
[57] Sirak, M., Aordable Reactive Strike Missile (ARES), Janes Defence Weekly,
June 29, 2005.
[58] Parsch, A., Sidearm, www.astronautix.com/lvs/sidearm.htm.
[59] Friedman, N., World Naval Weapons Systems, 1997/98, Naval Institute Press,
1997.
[60] Ozu, H., Missile 2000 - Reference Guide to World Missile Systems, Shinki-
gensha, 2000.
[61] Janes Electro-Optic Systems, Raytheon/RAM Systems RIM-116 Rolling
Airframe Missile, June, 2007.
[62] La Franchi, P., Army looks to battle unmanned threat,
http://www.flightglobal.com, Sept. 8, 2007.
[63] Kopp, C., Matra/BAe Alarm, Australian Aviation, June 1997.
[64] Janes Air Launched Weapons, Tien Chien IIA Anti-Radiation Missile, July
2006.
[65] Janes Air Launched Weapons, ARMIGER, Oct. 2007.
[66] Janes Electronic Mission Aircraft, Israel Aerospace Industries (IAI) MBDA
Raytheon CUTLASS/Harpy/Horop/White Hawk, Sept. 2007.
[67] Janes Electronic Mission Aircraft, Israeli Miltitary Industries (IMI) STAR-1
Anti-Radiation Missile, Oct. 2005.
[68] Janes Land Based Air Defense, HQ-9/FT-2000 surface-to-air anti-radiation
missile system, Sept. 2007.
[69] Janes Land Based Air Defense, FT-2000A anti-radiation seeking air defense
missile, Mar. 2004.
[70] Neng-Jing, L., Radar ECCM new area: anti-stealth and anti-ARM, IEEE
Trans. on Aerospace and Electronic Systems, Vol. 31. No. 3, pp. 11201127,
July 1995.
[71] Fan, W., RuiLong, H. and Xiang, S., Anti-ARM technique: distributed
general purpose decoy series (DGPD) pp. 306309, 2001.
[72] http://www.maliburesearch.com/technology.htm.
[73] Fan, W., RuiLong, H., and Xiang, S., Anti-ARM technique: Feature analysis
of ARM warning radar Proc. of the International Conference on Radar, pp.
293296, Bejing, China 2001.
[74] Wang, S., and Zhang, Y., Detecting of anti-radiation missile by applying
changeeable-sample ratios technology in the AEW, Proc. of the Interna-
tional Conference on Radar, pp. 289292 , 2001.
[75] Streetly, M., Gazetchik Anti-Anti-Radiation Missile (ARM) system, Janes
Radar and Electronic Warfare Systems, Jan. 2004.
[76] Streetly, M. AN/TLQ-32 ARM-D anti-radiation missile decoy, Janes Radar
and Electronic Warfare Systems, Jan. 2004.
Problems
1. An ARM seeker (fixed) antenna is being considered for use in the AGM-
88 (missile body diameter of 0.25 m, length of 4 m). It uses four cavity-
backed spiral antennas arranged in a phase comparison monopulse con-
figuration (protected by a radome) as shown in Figure 16.51. Assuming
that the antennas are on a flat disc and the spirals are nearly touching,
(a) what is the equation for the external radius of each spiral in terms
of the disc diameter? One of the properties of the spiral antenna is that
the longest useable wavelength L = 4r. (b) Estimate the frequency
coverage of the AGM-88 HARM missile if the bandwidth coverage is
10:1. (c) If the gain of each antenna is 3 dB (az = el = 80o ) over the
10:1 frequency band, what is total gain of the ARM antenna?
2. For the ARM seeker above, now assume a pulsed emitter with a carrier
frequency fc = 9 GHz whose transmitted peak power is Pt = 1 MW.
Its one-way main lobe antenna gain is Gt = 25 dB with a general side
lobe level of 30 dB with respect to the main lobe, giving a side lobe
gain Gt = 5 dB. The ARM missile is aimed at the radar with the
gain of the array as calculated above. Assume the range to the radar is
25 km. Also assume that the receiver front-end uses a superhetero-
dyne configuration with a bandwidth BIR = 250 MHz with a linear
detector to feed a bank of 250 video filters each with a bandwidth of 1
MHz. For this wideband receiver, a reasonable value of noise figure is
N F = 20 dB. (a) Determine the expression for the single-pulse signal-
to-noise ratio at the ARM receiver. (b) Do you think the seeker will
Figure 16.51: Cavity backed spiral antennas.
have any problem acquiring the emitter? (c) If the ARM RCS is 0.03
m2 , the noise figure of the emitter is 10 dB, the transmitted pulsewidth
is just sucient to enclose the ARM within a range bin and the min-
imum single pulse SNR required by the emitter to detect a target is
13 dB, calculate the emitters maximum detection range for this ARM
target. (Assume T0 = 290K.)
3. The expansion ratio for an equiangular spiral antenna can be expressed
as
( + 2) 0 ea(+2)
= = = ea2 (16.18)
() 0 ea
For = 4 for a two turn spiral ( = 4), determine the bandwidth ratio.
Chapter 17
Autonomous Classification
of LPI Radar Modulations
In this chapter, autonomous (no human operator intervention) feature extrac-
tion and classification algorithms that can be used for identifying LPI radar
modulations using time-frequency (T-F) detection images are presented. The
multilayer perceptron network and the radial basis function network are pre-
sented to identify the type of LPI modulation present in the intercepted
signal. These nonlinear classification networks use an input feature vector
that is generated from the T-F images (preprocessing). In the first feature
extraction algorithm, the modulation energy is cropped from the T-F image
using the marginal frequency distribution to determine the cropping region.
An adaptive binarization algorithm is then used to build the feature vector
in order to preserve the high-resolution detail that emphasizes the dierences
between modulation classes without overwhelming the classification networks.
Initial classification results show that the cropping region is sensitive to high-
frequency noise contained in the marginal frequency distribution. In a second
feature extraction algorithm, lowpass filtering of the T-F image is used prior
to calculation of the marginal frequency distribution. Wiener filtering of the
marginal frequency distribution is also added to improve the stability of the
cropping region. The use of principal components analysis to construct the
feature vector is investigated. An extended database is developed and the
classification results for simulated LPI radar modulations are shown as a
function of both SNR variations and variations in the modulation parameters
(most dicult, but realistic case).
619
17.1 Classification Using Time-Frequency

Imaging
Although the automatic recognition of LPI radar modulations is a new area
of investigation, the automatic recognition of communication signal modu-
lations has been of interest for many years [13]. In general, there are two
methods for autonomous classification of signal modulations: decision theo-
retic techniques and pattern recognition techniques. In particular, research
on this topic is typically applicable to military systems. Now with the advent
of software radios, research on autonomously recognizing communication sig-
nal modulations has resulted in the realization of reconfigurable and adaptive
wireless transceivers. The use of neural networks [4], wavelet transforms [5],
higher order statistics [6], and hidden Markov models [7] have been explored.
In a general sense, the autonomous recognition of communication modula-
tions is an easier problem than the autonomous recognition of LPI radar
modulations due to the fact that there are only a finite number of modula-
tion techniques used for communication. On the other hand, there are an
infinite number of modulation techniques that can be used for the LPI radar.
In fact this is why the noncooperative intercept receiver has such a dicult
time!
Classification using T-F imaging has received considerable attention in
such diverse fields as humpback whale signal recognition [8, 9], biomedical
engineering [10, 11], underwater acoustic target detection [12], radar target
classification [13], power grid analysis [14], and radar transmitter identifi-
cation [15]. With the high degree of detail contained in the image, train-
able autonomous classifiers can easily be overwhelmed by the complexity of
the T-F input representation and many eorts have been examined to re-
duce this problem. Smoothing the T-F images can be used to reduce the
density of the features but will most often remove the class-distinction de-
tail that the representation was intended to resolve. Quantizing the T-F
representation in a class- or signal-dependent manner can also preserve the
needed high-resolution detail that highlights the dierences between classes.
A vector quantization technique that is a modified version of a Kohonens
self-organizing feature map is applied to the T-F representation in [16].
Class-dependent smoothing can also be accomplished by optimizing the
T-F transformation kernel [15]. This approach eliminates the need to make
a priori assumptions about the amount and type of smoothing needed and
also allows for a direct classification without the need for preprocessing to
reduce the dimensionality. Optimizing the T-F kernel parameters based on
the Fisher criterion objective function is also examined in [8]. The Fisher
criteria however, assume the classes have equal covariance. In [9], the T-F
representation is used to construct a quadratic discriminant function, which
is evaluated at specific times to form a set of statistics that are then used in a
Autonomous Classification of LPI Radar Modulations 621
multiple hypothesis test. The multiple hypotheses are treated simultaneously

using a sequentially rejective Bonferroni test to control the probability of
incorrect classification.
A method based on T-F projection filtering is presented in [12]. In this
approach the decision strategy about which target is present depends on
the comparison of a reference target and the filter output signal. In [13], a
reduction in the feature vector dimensionality using the geometrical moments
of the adaptive spectrogram is investigated. A principal components analysis
is then used to further reduce the dimension of the feature space. This involves
calculation of the covariance matrix and its eigenvectors. The feature vector
is then formed using the eigen vectors associated to the highest eigen values,
and then it is applied to a multilayer perceptron for automatic recognition.
17.2 Classification Authority and Automation

The LPI emitter has established itself as the premier tactical and strategic
radar in the military spectrum. In addition to surveillance and navigation,
the LPI emitter also operates in the time-critical domain for applications such
as fire control and missile guidance. In the EW battle, the noncooperative
intercept receiver is a significant element in the detection and classification of
the LPI radar in a complex environment of multiple emitters and high noise
interference. The LPI radar modulations force the intercept receiver to in-
crease its processing gain by implementing T-F signal processing algorithms.
With these detection techniques a human operator can examine the result-
ing T-F image on a human-computer interface (HCI) and identify the type
of signal modulation present (classification) as well as quantify (or extract)
the modulation parameters. The development of a corresponding jammer
response, when required, is almost always a time critical event.
17.2.1 Human-Computer Interface Considerations

In an embedded system, cost, size, power, and complexity are limited, so
the HCI must be easy to use without sacrificing accuracy in the analysis ca-
pability. Human operators are often one of the biggest sources of error in
any embedded system and many operator errors are attributed to a poorly
designed HCI [16]. Electronic intelligence (ELINT) receiver designers must
insure that the HCI is easy and intuitive for human operators to use, but not
so simple that it lulls the operator into a state of complacency and lowers
their responsiveness to vital situations. The ELINT receiver HCI must give
appropriate feedback to the operator to allow well informed decisions to be
made based on the most up-to-date information on the electromagnetic envi-
ronment. High false alarm rates will make the operator ignore a real alarm
condition. If the human operator is out of the control loop in an automated
Table 17.1: Sheridan Levels of Authority (after [18]).
Level Computer Task Human Task

1 No assistance Does all
2 Suggests alternatives Chooses
3 Selects way to do task Schedules response
4 Selects and executes Must approve
5 Executes unless vetoed Has limited veto time
6 Executes immediately Informed upon execution
7 Executes immediately Informed if asked
8 Executes immediately Ignored by computer
task, the operator will tend to adapt to the normal operation mode and not
pay close attention to the system (operator drop out). When an emergency
condition occurs, the operators response will be degraded and they will tend
to make more mistakes. For example, the operator might unexpectedly have
to manage a proper EW response to the intercepted emitter.
17.2.2 Automation and the Human Operator

The need for human analysis of the T-F results limits these techniques to
ELINT receivers where the emitter information derived is not time-critical.
High-level automation of the classification decision, parameter extraction and
response management are however justified in highly time-critical situations
in which there is insucient time for a human operator to respond and take
appropriate action [17]. This is the case for ES receivers and RWRs. Human
beings are often still needed to be the fail-safe in an otherwise automated
system.
The Sheridan level shown in Table 17.1 is a system of eight levels to
indicate the amount of automation that is incorporated in the response, its
level of autonomy and whether the response execution authority is assigned to
the system or to the operator [18]. The Sheridan levels or levels of authority
(LoA) vary from level 1: Computer oers no assistance, human does all to
level 8: Computer selects method, executes task and ignores human. In
levels 1 to 4 the operator has authority over function execution; in levels 6
to 8 authority has moved to the system. In level 5 the authority is shared
between the system and the operator.
Figure 17.1 shows an example where the intercept receiver calculates the
T-F results from an intercepted LPI signal and must then administer a jam-
ming waveform response. The figure shows the EW response management
detailing the interaction between automation, autonomy and authority for
Figure 17.1: Interaction between automation, autonomy, and authority

(adapted from [17]).
the jamming waveform. Depending on the Sheridan level of the response, the
T-F data is presented to the operator (arrow 1a) or used by the system part
Autonomous decision making to decide what LPI modulation is present
and what the modulation parameters are (1b), given these T-F inputs [17].
Then, the system can suggest the particular modulation type to the operator
(2a), who then schedules the jamming response execution (4a) or the system
can select and schedule an automated response (2b). Whether the execution
of the scheduled jamming response must be acknowledged by the operator
depends on the LoA assigned to the response (4b). This is realized by the
Authority filter and the switch below the filter that determines whether
the scheduled response is executed (3). Depending on the setting of the Au-
thority filter, the operator does or does not receive feedback upon response
execution (5). The interaction can be summarized as: Autonomy schedules
automated responses, while authority allows or blocks response execution [17].
17.2.3 Autonomous Modulation Classification

Figure 17.2 shows the steps that can be used to autonomously classify the
LPI radar signal modulations. The LPI radar signal is intercepted with a
digital receiver that digitizes the intercepted signal. The signal is processed
by both T-F and B-F detection techniques. After the T-F, B-F detection
processing, the resulting image planes are used by the autonomous decision
making process to identify the modulation type. The autonomous decision
making consists of a feature extraction algorithm that is used to derive the
feature vector from the T-F, B-F image plane. A nonlinear classification
Figure 17.2: Noncooperative intercept receiver for autonomous detection and

identification of LPI radar modulation.
network is then used to recognize the LPI signal modulation type from the
feature vector or pattern. The most important segment of this pattern recog-
nition scheme is how the feature vector is formed and how it is presented
to the nonlinear classification network. Note that if a high performance re-
configurable computer is used, several T-F and B-F detection/classification
algorithms can be executed quickly and in parallel [19]. Below we first discuss
the nonlinear classification networks that are used to identify the modulation
type. This includes the multilayer perceptron (MLP), and the radial basis
function (RBF) network. Feature extraction image processing techniques are
then discussed and results are shown.
17.3 Nonlinear Classification Networks

Nonlinear classification networks use a set of processing elements (or nodes)
loosely analogous to neurons in the brain (hence the name, artificial neural
networks). The nodes are interconnected in a network that can then identify
patterns in data as it is exposed to the data. In a sense, the network learns
from experience just as people do. This distinguishes neural networks from
traditional computing programs that simply follow instructions in a fixed
sequential order. The architectures are specified by: (1) the network topology,
(2) the node characteristics, and (3) the training or learning rules used to
configure the weights on each connection [20].
The classification networks can be either static or dynamic. Static net-
works are characterized by node equations that are memoryless. That is,
their output is a function of only the current input and not of past or fu-
ture inputs or outputs. Dynamic networks are systems with memory. The
Figure 17.3: Rosenblatt perceptron (static network).
Figure 17.4: Hard limiting nonlinearity.
dynamic neural networks are characterized by dierential equations or dier-

ence equations [21].
17.3.1 Single Perceptron Networks

An example of a static network is the Rosenblatt perceptron as shown in
Figure 17.3. Here Xn represents the n-dimensional input vector and Wn
represents the n-dimensional weighting vector. The Rosenblatt perceptron
forms a weighted sum of n-components of the input vector and adds a bias
value, . The result y is passed through a nonlinear activation function to
give the output value u. The activation function shown in Figure 17.3 is a
hard-limiting nonlinearity fHL . An example of a hard-limiting nonlinearity
is shown in Figure 17.4 where
1 y>0
fHL (y) = (17.1)
0 y0
Another popular activation function is the sigmoid. The sigmoid nonlinearity

is given by the expression

1 1
fs (y) = 1 + ey = (17.2)
1 + ey
and is continuous. The nonlinearity varies monotonically from 0 to 1 as y
varies from to . The value represents the gain of the sigmoid.
One of the key attributes of the sigmoid nonlinearity fs (y) is that it is a
dierentiable function. This also makes it well suited to our application of
pattern recognition since the output is between 0 and 1. Note that this can
be interpreted as a probability distribution. The value of the output y is a
weighted sum and is the inner product between the augmented input vector
and the weight vector or [21]
T
y=W X (17.3)
or
1
X1

X2
y = [W0 , W1 , , Wn ] (17.4)
..
1n .
Xn
n1
and then the output

u = fHL (y) (17.5)
A single Rosenblatt perceptron can be used to build several important logic
units. One example is the AND function as shown in Figure 17.5 [21]. With
the weights shown the summation output y = 2X1 +2X2 3. The output u for
binary values of X1 and X2 and the value of y are as shown in the truth table.
The binary logic unit OR can also be implemented with one perceptron as
shown in Figure 17.6. The summation is y = 2X1 +2X2 1. The complement
or NOT function can also be implemented with a single perceptron with one
input as shown in Figure 17.7 [21]. The equation for the NOT summation is
y = 2X1 + 1. Note that a single perceptron cannot implement an exclusive
OR (XOR) or an exclusive NOR (XNOR).
To recognize how the perceptron can be used to recognize patterns, we ex-
amine the general two input (three weights) perceptron shown in
Figure 17.8. A critical threshold occurs when the linear output y = 0 or
y = X1 W1 + X2 W2 + W0 = 0 (17.6)
Therefore, in slope intercept form we have

W1 W0
X2 = X1 (17.7)
W2 W2
Figure 17.5: Binary logic unit: AND [21].
Figure 17.6: Binary logic unit: OR [21].

Figure 17.7: Binary logic unit: NOT [21].
Figure 17.8: General two-input perceptron.
Figure 17.9: General two-input perceptron as a linear separable function.

Figure 17.10: Three-layer perceptron model.
which is a linear separable function. That is, a linear line is formed to separate
two regions of a plane as shown in Figure 17.9. With each additional weight,
a new dimension is added to the separation boundary. That is, with four
weights, the separation boundary becomes a plane, and with five weights, the
separation boundary becomes a hyper-plane.
17.3.2 Multilayer Perceptron Networks

In an MLP network the perceptrons (neurons or nodes) are the information
processing units and they are cascaded in layers to create the complex decision
regions. The inputs propagate through the network in a forward direction, on
a layer by layer basis. Most often the input set of nodes is not considered a
layer. A model of a three-layer perceptron network is shown in Figure 17.10.
In this model there are four neurons at the input, two hidden layers with five
and four neurons respectively and an output layer with two neurons.
Within the MLP is a set of synapses or connecting links, each of which is
characterized by a weight of its own. Each neuron has an adder for summing
the input signals, weighted by the respective synapses of the neuron. The
activation function then limits the amplitude of the output of each neuron.
The neuron may also include the externally applied bias which has the eect
of increasing or lowering the net input to the activation function, depending
on whether it is positive or negative, respectively. The network exhibits a high
degree of connectivity, determined by the synapses of the network. Most often
the nodes are fully connected with every node in layer i connected to every
node in layer i + 1.
In an MLP network the inputs propagate through the network in a forward
direction, on a layer by layer basis. Training algorithms include gradient
search, backpropagation and temporal dierence. The measure of how well
the network performs on the actual problem, once training is complete, is
called generalization. It is usually tested by evaluating the performance of
the network on new data that is outside the training set. Parameters that
can aect the generalization are: (a) the number of data samples and how
well they represent the problem at hand, (b) the complicity of the underlying
problem, and (c) the network size. In general, a large number of weights
adversely aects generalization and the time required to learn the solution.
It is also worth noting that the feature vector derived from the T-F and B-F
images has a significant impact on both (a) and (b).
An MLP with I input nodes, and H hidden layers can be described in
general as [22]
H I
yk ( ) = fs wkh fs whi xi ( ) (17.8)
h=1 i=1
where yk is the output, xi is the input, is the sample number, i is the

input node index, h is the number of hidden layers index and k is the output
node index. Here wkh and whi represent the weight value from neuron h to k
and from neuron i to h respectively and fs represents the sigmoid activation
function. All weight values in the MLP are determined at the same time in
a single, global (nonlinear) training strategy involving supervised learning.
The activation function fs may vary for dierent layers within the net-
work. The activation function can be any type of function that fits the action
desired from the respective neuron and is a design choice which depends on
the specific problem. Log sigmoid and hyperbolic tangent sigmoid functions
are commonly used in multilayer neural networks since they are dierentiable
and can form arbitrary nonlinear decision surfaces [23]. The network activa-
tion function, fs , that is popular for pattern recognition classification is the
log-sigmoid discussed previously defined as
fs (y) = 1/(1 + ey ) (17.9)
When supervised learning is used, the input-output examples are used to

train the network and derive the network weights. Since the network design
is statistical in nature we can improve the network generalization during the
supervised learning process by minimizing the trade-o between the reliability
of the training data and the goodness of the model. This trade-o is realized
during the supervised learning process through the network regularization R
R = gMSE + (1 g)MSW (17.10)

where g is the Tikhonovs regularization parameter 0 < g < 1 [23]. The term
MSE is a performance measure and is the mean sum of squares of the network
errors. The performance measure depends on both the network design and
the training data. The term MSW is the mean sum of squares of the network
weights and biases and is sometimes referred to as the complexity penalty.
From (17.10), the regularization parameter g influences directly the trade-o
between the complexity penalty and the performance measure. The optimum
values to minimize R are found and the process is carried out for all the
training examples on an epoch-by-epoch basis. Note that if g = 1, the network
design is unconstrained with the solution depending only on the input-output
training examples.
For most applications, a three-layer network with H = 2 hidden layers
should sucient. Note that when more hidden layers are included, the con-
vergence of the weight values becomes more dicult and significantly more
time is required to complete the global training. Further, there is a much
larger chance that an overgeneralization will be provided which degrades the
ability of the network to identify correctly the modulation type present. The
number of output neurons reflects the number of modulation types that are
expected. For example, if 12 modulation types were expected in the theater
of operations, then the output layer should have 12 neurons each of which
corresponds to a modulation type. The output neurons can be hard lim-
iting (0 or 1) or can be sigmoidal which gives more of a modulation type
probability. The input feature vector is extracted from the T-F or B-F de-
tection processing image. The feature vector dimension D 1 is determined
by feature extraction signal processing.
The supervised training of the feed-forward MLP network uses the gra-
dient of the performance function to determine how to adjust the weights.
The gradient is determined using a technique called backpropagation [24].
The backpropagation algorithm is a generalization of the least mean square
algorithm used for linear networks, where the performance index is the mean
square error. Basically, a training sequence is passed through the multilayer
network, the error between the target output and the actual output is com-
puted, and the error is then propagated back through the hidden layers from
the output to the input in order to update weights and biases in all layers.
Dierent modifications of training algorithms may improve the convergence
speed of the network. One of these modifications is the variable learning
rate. With the standard steepest descent algorithm, the learning rate is held
constant throughout the training. The performance of the algorithm is very
sensitive to the proper setting of the learning rate. When a variable learning
rate is used and the learning rate is allowed to change during the training
process, the performance of the steepest descent algorithm is improved [24].
Figure 17.11: Block diagram of a radial basis function neural network.
17.3.3 Radial Basis Function

The radial basis function (RBF) network is a static feed-forward, two-layer
network originally proposed by Broomhead and Lowe [25]. A block diagram
of the RBF is shown in Figure 17.11. Each element of the input vector x is
applied to the hidden layer which is composed of J basis functions . The
N output nodes form a linear weighted (Wnj ) summation of the basis func-
tion outputs that are computed [26]. Unlike the MLP, the RBF uses a linear
adaptive algorithm in the training of the network coecients. This makes the
RBF network appropriate for real-time applications since it can be designed
in a fraction of the time it takes to train an MLP. Since LPI signal modula-
tions are nonstationary, this on-line learning ability allows the classifier to be
adaptive to the electromagnetic threat environment. In addition, the RBF
network is a universal approximator which has the capability of approximat-
ing a decision boundary of any shape providing a major advantage for the
noncooperative intercept receiver.
Each basis function in the hidden layer produces a nonzero response to
the input data when the input falls within the basis functions small localized
region. The RBF determines the similarity between a new input vector and
a number of stored vectors representing the basis function centers by using
the concept of Euclidean norm or distance. For example if we let xi denote
an n-by-1 vector
xi = [xi1 , xi2 , . . . , xiN ]T (17.11)
all of whose elements are real, the vector xi defines a point in n-dimensional
space called Euclidean space. The Euclidean norm between a pair of m-by-1
vectors xi and xj is defined as

n 1/2
2
||xi xj || = (xik xjk ) (17.12)
k=1
where xik and xjk are the kth elements of the input vectors xi and xj respec-
tively [23]. Correspondingly, the similarity between the inputs represented by
xik and xjk is defined as the reciprocal of the Euclidean distance ||xi xj ||.
The output of the RBF can be expressed as
J
ypn = Wn0 + Wnj (||xp cj ||) (17.13)
j=1
where ypn is the output of the nth modulation node in response to the pth
input pattern, (||xp cn ||) is the output of the hidden node n in response
to the pth input vector xp and the vectors cn , n = 1, . . . , N are referred to as
the centers of the radially symmetric basis functions . The weighting matrix
Wnj represents the synaptic weights from the jth radial basis function to the
nth modulation output node and Wn0 is the bias or threshold assigned to the
nth modulation output node.
One symmetrical choice for the radial basis function is the Gaussian
function
J
(xj cnj ) 2
(||xp cn ||) = exp 2 (17.14)
2nj
j=1
where nj are the elements of a covariance matrix (or spread), which is taken
here to be diagonal.
The set of hidden units consist of a set of functions which constitute an
arbitrary basis for the feature vector patterns to be classified when expanded
into hidden unit space. These are referred to as radial basis functions. The
expansion of input vectors into a hidden unit space of relatively high dimen-
sion (many radial basis functions) will result in a greater likelihood of the
classification problem becoming linearly separable. One approach for an ef-
ficient RBF network design is by iteratively creating the RBF one neuron
at a time. Neurons are added to the network until either the sum-squared
error falls beneath an error goal or a maximum number of neurons has been
reached [23]. The two parameters used to optimize the RBF to obtain a bet-
ter probability of correct classification are the goal and spread . The spread
constant should be larger than the distance between adjacent input vectors,
so as to get a good generalization, but smaller than the distance across the
whole input space. The training is accomplished in two stages. The basis
functions are determined by unsupervised techniques using the input data
while the second layer weights are found by a fast linear supervised method.
Hence the training is fast and ecient.
17.4 Feature Extraction Signal Processing

In Figure 17.2, the LPI signal is detected with three T-F signal processing
methods. Each T-F image contains features that identify the modulation type
(and its parameters). Methods to autonomously extract these features from
the T-F distribution have recently received much attention (see for example
[27, 28]). Two ecient feature extraction methods that build a feature vector
from a T-F image are described below.
17.4.1 Marginal Frequency Adaptive Binarization

An autonomous (no human operator intervention) T-F feature extraction
image processing approach that can be used for classification of LPI radar
modulations is examined. The feature extraction image processing uses the
marginal frequency distribution, or instantaneous energy [29], derived from
the T-F representation in order to isolate the location of the modulation au-
tonomously. In order to isolate the modulation, a histogram of the normalized
marginal frequency distribution is first computed to choose a threshold for
comparison against the mean energy value calculated from the convolution
of the normalized marginal frequency distribution with an averaging kernel
(sliding window) of length n. This convolution gives the start and stop fre-
quency of the modulation energy. With the location of the modulation energy
known it is possible to crop the energy from within the T-F image and com-
pute a feature vector for input into the classification network. In order to
preserve the high-resolution detail that emphasizes the dierences between
modulation classes without overwhelming the classifier, an adaptive binariza-
tion algorithm is used to generate a vector of 1s and 0s that represent the
modulation [30].
A block diagram of the autonomous feature extraction image processing
technique is shown in Figure 17.12. This technique uses the T-F marginal
frequency distribution and an adaptive binarization algorithm to form the
feature vector. The Choi-Williams detection of a T1(2) signal modulation is
used to demonstrate the marginal frequency adaptive binarization (MFAB)
feature extraction algorithm.
Deleting No-Signal Region

One of the characteristics that may be present in the T-F image is when
the signal does not extend for the entire distribution. The T-F image will
then show the presence of a black column starting where the signal ends and
covering all frequencies. Since this no-signal region within the T-F image does
not contain useful information for classification, this region must be removed
(see the detect & delete no-signal region block in Figure 17.12). Figure 17.13
Figure 17.12: Autonomous feature extraction using the marginal frequency

adaptive binarization (MFAB) algorithm [30].
shows the no-signal region within a Choi-Williams distribution for a T1(2)

modulation. Also shown is the image with the no-signal region removed.
Marginal Frequency Distribution to Determine Cropping

The marginal frequency distribution oers a way to examine a T-F image in
intercept situations where there is a low SNR [30]. The marginal frequency
distribution gives the instantaneous energy of the signal as a function of
frequency. The marginal frequency distribution is generated by
CWDx () = CWDx ( , ) (17.15)
or summing the time values for each frequency in the T-F image and then stor-
ing the sums in an array. Each marginal frequency distribution is a unique
representation of the T-F image it was generated from. The marginal fre-
quency distribution is normalized by dividing the sums by the largest sum
in the array. The normalized marginal frequency distribution of the T1(2)
modulation is shown in Figure 17.14.
The normalized distribution is used to extract a threshold that is used
later to isolate and crop the modulation energy within the T-F image. The
threshold is determined by generating a histogram of 100 bins of the nor-
malized marginal frequency distribution and then taking the value from the
histogram bin which generates the best probability of correct classification
(Pcc) results. An example of the T1(2) histogram is shown in Figure 17.15.
Once the threshold is determined (n = 9) the convolution of the aver-
aging kernel with the normalized marginal frequency distribution is used to
Figure 17.13: Choi-Williams T-F image for the T1(2) modulation showing
(a) presence of the no-signal region and (b) no-signal region removed.
Figure 17.14: Normalized marginal frequency distribution of the T1(2) mod-

ulation shown in Figure 17.13.
Figure 17.15: Histogram of the normalized marginal frequency distribution

for T1(2) modulation used to determine threshold autonomously.
determine the start and stop cropping locations of the modulation energy.
The convolution operation calculates the mean of the corresponding cells in
the normalized distribution and compares it to the threshold identified from
the histogram. If the average of the convolution is greater than the threshold
the start of the modulation energy is found. To find the stop location the
same convolution algorithm is used but the kernel is initialized at the end
of the normalized distribution. With the location of the modulation energy
known it can be cropped from the original image as shown in Figure 17.16.
Adaptive Binarization
An adaptive binarization algorithm is used to generate a binary image that is
then resized to form a feature (column) vector containing ones and zeros. A
block diagram of the adaptive binarization process is shown in Figure 17.17
[30]. The intensity image |I(t, )| is normalized I (t, ) with respect to the
largest value in the image, where the pixel values 0 I (t, ) 1.
A histogram of the intensity level content, h(n), is then generated using
N = 50 bins. The cumulative distribution function is computed using this
histogram as
n
h(i)
cdf (n) = Ni=1 (17.16)
n=1 h(n)
A cdf threshold, C, is chosen and the intensity bin n where cdf (n) C is
Figure 17.16: Autonomous cropping with (a) original resized image and
(b) cropped image.
then identified. For example, if the cdf threshold is chosen experimentally

to be C = 0.8, this means that only 20% of the brightest pixels above the
threshold are retained. With this intensity bin, a corresponding normalized
intensity threshold, T = n/N , is calculated. This adaptive threshold is then
used to convert the intensity image into a binary image. That is,
1 (black) I (t, ) T
I (t, ) = (17.17)
0 (white) I (t, ) < T
which eectively removes much of the noise and weak interference.

The final step is to resize the image. The binarized image is resized
to Nr Nc by lowpass filtering, resampling and then applying a bilinear
interpolation. The lowpass filter is used to reduce the eect of Moire patterns
and ripple patterns that result from aliasing during the resampling operation
[31]. The Nr Nc image is then converted into a feature vector of size (Nr Nc )
1 for processing by the multilayer perceptron classification network where
Nr = Nc = 50.
17.4.2 Classification Results with Multilayer

Perceptron
For testing the MFAB feature extraction, five modulation types are used. The
modulation types include BPSK, FMCW, Frank, P4, and T1. To classify the
signal modulations, a multilayer perceptron is used with two hidden layers
and an output layer. The signals used in the training of the classification
Figure 17.17: Adaptive threshold binarization process [30].
network contain signal only, SNR = 0 and SNR = 3 dB. Each signal is
corrupted with additive white Gaussian noise (WGN) before input to the T-
F detection transforms. The SNR is defined as SNR = A2 /2 2 where A is the
amplitude of the signal and 2 is the WGN power. WGN is used since this
model most generally reflects the thermal noise present in the IF section of
an intercept receiver. Two dierent carrier frequencies were also used (fc = 1
kHz and fc = 2 kHz). To test the classification algorithm, noise variations
(TestSNR) and modulation variations (TestMod) were used. Note that the
TestMod testing is the most dicult case. The database description is given
in Table 17.4.2.
To optimize classification with the MFAB algorithm it is necessary to
pick an accurate threshold using the histogram derived from the marginal
frequency distribution in order to find the start and stop frequencies of the
modulation energy. In order to do so, a loop that cycles through each of the
histogram bins can be used to determine the bin that gives the best threshold
for optimum classification. Once the bin that gives the best Pcc is identified,
that threshold can be used to generate the classification results.
Table 17.2: Database Description for MFAB Feature Extraction and

Classification (fs = 7 kHz)
Signal Variable TestSNR TestMod Description

modulation values values
BPSK cpp 1, 4, 7 2, 3, 5, 6 Cycles per subcode
Nc 7, 11 7, 11 No. of subcodes
fc (kHz) 1, 2 1, 2 Carrier frequency
FMCW F (Hz) 250, 500 350, 450 Mod. bandwidth
tm (ms) 20, 50 35, 45 Mod. period
Frank cpp 1, 4, 7 2, 3, 5, 6 Cycles per subcode
Nc 16 16 No. of subcodes
P4 cpp 1, 4, 7 2, 3, 5, 6 Cycles per subcode
Nc 16 16 No. of subcodes
T1 n 2, 6 3, 4, 5 No. of phase states
k 4 4 No. of code segments
The MLP used to generate the results was executed for a total of 5,000
epochs, with an error goal of 1 106 . Thirty-five neurons were used in both
the first and second hidden layers (S1 = S2 = 35) for the Choi-Williams
results and the Wigner-Ville distribution results. For the QMFB, due to the
dierent size of the layers, S1 = 20 neurons were used in the first hidden layer
and S2 = 35 neurons were used in the second hidden layer. The output layer
for all MLP configurations contained 5 neurons which matches the number of
modulations that were expected.
Optimum classification for the Choi-Williams distribution occurred when
bin 16 was used as shown in Figure 17.18 (testing with modulation variation)
and Figure 17.19 (testing with noise variation). No classification results were
obtained using thresholds from histogram bins greater than 72 because the
feature extraction algorithm could not isolate the modulation. Table 17.3
shows the classification results in the form of a confusion matrix for the Choi-
Williams distribution. The diagonal terms represent the Pcc percentage. The
o-diagonal terms are the percentages for the modulation being misclassified.
Classifying signals with variations in their modulation (TestMod) is a more
dicult case than classifying signals with only variations in noise (TestSNR).
This fact is present in all of the classification results.
Figures 17.20 and 17.21 are the optimization tables for the Wigner-Ville
distribution. Optimum classification occurs when bin 31 (n = 31) is used.
No classification results were obtained using thresholds from histogram bins
Figure 17.18: Choi-Williams: MLP optimization (TestMod).
Figure 17.19: Choi-Williams: MLP optimization (TestSNR).

Table 17.3: Choi-Williams MLP Classification Confusion Matrix (n = 16)
TestMod BPSK FMCW Frank P4 T1

BPSK 0.93 0.0 0.08 0.35 0.08
FMCW 0.00 1.0 0.01 0.00 0.00
Frank 0.01 0.0 0.53 0.05 0.07
P4 0.05 0.0 0.35 0.60 0.02
T1 0.01 0.0 0.04 0.00 0.83
TestSNR BPSK FMCW Frank P4 T1
BPSK 0.96 0.00 0.22 0.11 0.12
FMCW 0.00 1.0 0.00 0.00 0.00
Frank 0.01 0.0 0.74 0.02 0.03
P4 0.03 0.0 0.01 0.85 0.05
T1 0.00 0.0 0.03 0.01 0.81
greater than 66 because the feature extraction algorithm could not isolate the
modulation. Table 17.4 shows the classification results for the Wigner-Ville
distribution when bin 31 is used.
Figures 17.22 and 17.23 are the optimization tables for the QMFB detec-
tion technique. Optimum classification occurs when bin 9 (n = 9) is used.
No classification results were obtained using thresholds from histogram bins
greater than 18 because the feature extraction algorithm could not isolate
the modulation. Table 17.5 shows the classification results for the QMFB
distribution when bin 9 is used.
17.4.3 Classification Results with Radial Basis

Function Network
To produce optimum results with the RBF it is necessary to pick an ac-
curate threshold using the histogram derived from the marginal frequency
distribution in order to accurately find the start and stop frequencies of the
modulation energy similar to what was done for the MLP. The optimum clas-
sification occurs when bin 16 (n = 16) is used. Table 17.6 shows the classifica-
tion using the RBF and the Choi-Williams for n = 16. For the Wigner-Ville
distribution, the optimum classification occurs when bin 58 (n = 58) is used.
Table 17.7 shows the classification using the RBF and the Wigner-Ville distri-
bution for n = 16. For the quadrature mirror filtering, the optimum classifica-
tion occurs when bin 14 (n = 14) is used. Table 17.8 shows the classification
using the RBF and the Wigner-Ville distribution for n = 16.
Figure 17.20: Wigner-Ville distribution: MLP optimization (TestMod).
Figure 17.21: Wigner-Ville distribution: MLP optimization (TestSNR).

Table 17.4: Wigner-Ville MLP Classification Confusion Matrix (n = 31)

BPSK 0.96 0.0 0.08 0.12 0.20
FMCW 0.00 1.0 0.00 0.00 0.00
Frank 0.02 0.0 0.58 0.22 0.27
P4 0.02 0.0 0.33 0.65 0.04
T1 0.01 0.0 0.00 0.01 0.49
BPSK 0.95 0.00 0.19 0.06 0.10
FMCW 0.00 1.0 0.00 0.00 0.01
Frank 0.02 0.0 0.76 0.01 0.02
P4 0.01 0.0 0.03 0.91 0.01
T1 0.02 0.0 0.03 0.02 0.86
Figure 17.22: Quadrature mirror filtering: MLP optimization (TestMod).

Figure 17.23: Quadrature mirror filtering: MLP optimization (TestSNR).
Table 17.5: Quadrature Mirror Filtering MLP Classification Confusion

Matrix (n = 9)

BPSK 0.82 0.40 0.48 0.35 0.23
FMCW 0.14 0.48 0.21 0.21 0.00
Frank 0.01 0.01 0.13 0.14 0.05
P4 0.03 0.11 0.18 0.30 0.02
T1 0.00 0.0 0.01 0.00 0.70
BPSK 0.95 0.03 0.08 0.11 0.09
FMCW 0.01 0.94 0.03 0.01 0.01
Frank 0.02 0.02 0.81 0.01 0.03
P4 0.02 0.02 0.03 0.86 0.04
T1 0.00 0.00 0.05 0.01 0.85
Table 17.6: Choi-Williams RBF Classification Confusion Matrix (n = 16)

BPSK 1.00 0.00 0.0 0.00 0.17
FMCW 0.00 1.00 0.00 0.00 0.00
Frank 0.00 0.00 0.50 0.14 0.05
P4 0.00 0.00 0.38 0.88 0.00
PT1 0.00 0.00 0.13 0.03 0.83
BPSK 0.95 0.00 0.07 0.07 0.00
FMCW 0.00 1.00 0.00 0.00 0.00
Frank 0.00 0.00 0.87 0.00 0.00
P4 0.02 0.00 0.03 0.90 0.10
PT1 0.03 0.00 0.03 0.03 0.90
Table 17.7: Wigner-Ville Distribution RBF Classification Confusion Matrix

(n = 58)

BPSK 0.75 0.00 0.00 0.00 0.00
FMCW 0.00 1.00 0.13 0.00 0.00
Frank 0.00 0.00 0.50 0.38 0.00
P4 0.00 0.00 0.25 0.50 0.00
T1 0.25 0.00 0.13 0.13 1.00
BPSK 0.83 0.00 0.07 0.03 0.00
FMCW 0.03 1.00 0.00 0.00 0.00
Frank 0.00 0.00 0.70 0.10 0.05
P4 0.02 0.00 0.07 0.83 0.10
T1 0.13 0.00 0.17 0.03 0.95
Table 17.8: Wigner-Ville Distribution RBF Classification Confusion Matrix

(n = 58)

BPSK 0.81 0.25 0.75 0.63 0.67
FMCW 0.19 0.50 0.13 0.25 0.00
Frank 0.00 0.00 0.13 0.00 0.00
P4 0.00 0.25 0.00 0.13 0.00
T1 0.00 0.00 0.00 0.00 0.80
BPSK 0.97 0.00 0.10 0.10 0.20
FMCW 0.03 1.00 0.17 0.07 0.00
Frank 0.00 0.00 0.73 0.00 0.00
P4 0.02 0.00 0.00 0.83 0.00
T1 0.00 0.00 0.00 0.00 0.80
17.4.4 Discussion of Classification Results

The marginal frequency adaptive binarization (MFAB) algorithm was able to
be applied to a complex database of LPI signals that closely resembles the
types of signals that are found operationally. The ADC sampling frequency
(fs = 7, 000 Hz) and LPI signal frequencies and bandwidths are considerably
lower than those within the actual intercept receiver hardware due to compu-
tational considerations. The classification results however, are representative
of the type of results that would be obtained using actual LPI emitters and
intercept receiver hardware. The success of the MFAB feature extraction al-
gorithm is due in part to the users ability to choose the cropping threshold
from an optimization analysis. Choosing the optimum threshold ensures that
the algorithm is able to isolate the modulation energy and produce the best
possible results for the signal types that are expected.
The best classification results were produced with the Choi-Williams dis-
tribution. It might be suggested that the success of using the Choi-Williams
distribution is attributed to its lack of cross terms (compared to the Wigner-
Ville distribution), which can confuse the classification network. That is, the
large cross-terms are not conducive to producing unique T-F features that
can help the classifier distinguish the dierent signal modulations from one
another.
The MLP consistently produced better results than the RBF. The MLP
was successful because it had more variables that could be configured to pro-
duce optimum results. The only variable in the RBF that could be changed
was the spread. While the results from the RBF were not as good as those
seen coming from the MLP the RBF does have an advantage in terms of its
application to an operational environment. Its training phase is fast and e-

cient. Because it trains faster than the MLP, in an operational environment
a user would not have to wait long before receiving classification results for
an LPI signal. This feature of the RBF is important and should be investi-
gated in an attempt to improve the RBFs results so that it can be applied
operationally. In the next section, several signal processing changes are made
in order to improve the classification results.
17.5 Modified Feature Extraction Signal

Processing
Calculation of the cropping region using the marginal frequency distribution
allows the low frequency LPI modulation to be retained and the remaining
T-F regions to be discarded. It is important that the size of the cropping
region be adaptive and only contain the modulation energy so that the de-
rived feature vector is consistently correlated with the modulation type. In
the cropping technique described above, the presence of high frequency noise
within the T-F image however, can vary the size of the cropping window.
To minimize this eect, the use of a lowpass filter prior to the calculation of
the marginal frequency distribution is investigated to help achieve the most
consistent cropping of the modulation energy. To make the threshold calcu-
lation more robust, the marginal frequency distribution is smoothed using a
Wiener filter before normalization. The Wiener filter takes the form of a linear
adaptive filter which adjusts its free parameters in response to the statistical
variations in the marginal frequency distribution. Also as an alternative to
directly using the feature vector as input to the classifier, this section also
examines the use of principal components analysis (PCA) in order to develop
a lower dimensional feature vector for use by the classifier. A block diagram
of the modified autonomous T-F cropping and feature extraction algorithm
is shown in Figure 17.24 [32].
17.5.1 Lowpass Filtering for Cropping Consistency

The detect and delete no-signal region is followed by a low pass filter (LPF)
applied to the T-F image. This insures that the low frequency LPI modulation
energy is preserved and the high frequency noise is removed. The filtering
can easily be performed in the frequency domain. Frequency domain filtering
using the 2-D Fourier transform is fast and ecient. Let f (k1 , k2 ) for k1 =
0, 1, 2, . . . , M 1 and k2 = 0, 1, 2, . . . , N 1 denote the M N T-F image.
The 2-D discrete Fourier transform (DFT) of f denoted by F (u, v) is [33]
M1 N 1
F (u, v) = f (k1 , k2 )ej2(uk1 /M+vk2 /N) (17.18)
k1 =0 k2 =0
Figure 17.24: Modified T-F autonomous cropping and feature extraction

algorithm.
for u = 0, 1, 2, . . . , M 1 and v = 0, 1, 2, . . . , N 1. The M N rectangu-

lar region F (u, v), defined by u and v, is often referred to as the frequency
rectangle and is the same size as the input image. Note that the frequency
rectangle can also be defined by digital frequencies as shown in Figure 17.25
where 1 = 2u/M and 2 = 2v/N .
Given F (u, v), f (k1 , k2 ) can be obtained by means of the inverse DFT.
Both DFT and inverse DFT are obtained in practice using a fast 2-D Fourier
transform (FFT) algorithm [33]. The convolution theorem, which is the foun-
dation for linear filtering in both spatial and frequency domains, can be writ-
ten as follows
f (k1 , k2 ) h(k1 , k2 ) H(u, v)F (u, v) (17.19)
and conversely,
f (k1 , k2 )h(k1 , k2 ) H(u, v) F (u, v) (17.20)
Filtering in the spatial domain consists of convolving an image f (k1 , k2 ) with
a filter mask, h(k1 , k2 ). According to the convolution theorem, the same
result can be obtained in the frequency domain by multiplying F (u, v) by
H(u, v), which is referred to as the filter transfer function. A block diagram
of the frequency domain filtering process is shown in Figure 17.26.
The filter transfer function H(u, v) can be obtained in three steps. First,
the desired frequency response (ideal lowpass filter) Hd (u, v) is created as a
Figure 17.25: Frequency rectangle for F (u, v).
Figure 17.26: Frequency domain filtering.

matrix. An ideal lowpass filter has the transfer function [33]
1 if D(u, v) D0
Hd (u, v) = (17.21)
0 if D(u, v) D0
where D0 (cuto parameter) is a specified nonnegative number and D(u, v)

is the distance from point (u, v) to the center of the filter. D0 can also be
defined as the normalized value of digital frequencies 1 , 2 by . Second,
a two-dimensional Gaussian window is created with a standard deviation
= N D0 /8 where N is the number of columns in the image. The standard
deviation of the window is related to D0 , and the structure becomes adaptive
to the changes in the desired frequency responses. For the detection of LPI
emitter modulations, both the frequency response matrix and the Gaussian
window have dimensions of M N which is equal to the image dimension
f (k1 , k2 ) and the 2-D FFT output dimension F (u, v). The last step is to
multiply Hd (u, v) by the Gaussian window.
The transfer function of the Gaussian lowpass filter obtained by this mul-
tiplication process is then given by [34]
2
(u,v)/2 2
H(u, v) = eD (17.22)
These steps are illustrated in Figure 17.27. Figure 17.27(a) shows the desired
frequency response with D0 = 0.3 (where |D0 | [0, 1]) or 1 = 2 = 0.3,
Figure 17.27(b) shows the Gaussian window with = N D0 /8 = 33.825.
The dimension of both the frequency response matrix and Gaussian window is
M = 1,024, N = 902. Figure 17.27(c) shows the resultant Gaussian lowpass
filter and Figure 17.27(d) shows the Gaussian lowpass filter as an image.
Several values of 1 , 2 can be tested during the simulation process to find
an optimum value for each distribution. For each trial the digital cuto
frequencies should be set to 1 = 2 . After obtaining the lowpass filter,
the frequency domain filtering can be implemented by multiplying F (u, v) by
H(u, v). This operation is followed by shifting the frequency components back
and taking the inverse FFT of the filtered image. The last step is obtaining
the real part of the inverse FFT.
17.5.2 Calculating the Marginal Frequency Distribution

After the LPF is used to eliminate the high frequency noise, the marginal
frequency distribution of the T-F image is calculated. The marginal frequency
distribution gives the instantaneous energy of the signal as a function of
frequency. The steps for determining the modulation frequency band from
the T-F plane are shown in Figure 17.28. The operations are applied to the
MFD of the T-F plane. The MFD gives the instantaneous energy of the
signal as a function of frequency. This is obtained by integrating the time
values for each frequency in the T-F image resulting in an M 1 vector A.
Figure 17.27: Implementation of filter function (a) desired frequency response,

(b) Gaussian window, (c) Gaussian lowpass filter, and (d) Gaussian lowpass
filter as an image [32].
Figure 17.28: Modified method for determining the cropping region [32].
Figure 17.29: Frank code signal with Nc = 36 (a) MFD and (b) MFD after
thresholding [32].
As an example, the marginal frequency distribution of a Frank coded signal

with fs = 7 kHz, fc = 1,495 Hz, Nc = 36, and cpp = 1 (B = 1, 495 Hz)
with an SNR = 0 dB is shown in Figure 17.29(a). The higher energy interval
corresponds to the frequency band of interest and contains the modulation
energy. The goal is to isolate and crop the LPI modulation as accurately
as possible. This is done by computing the threshold from the histogram as
before. As the noise level changes however, the cropping window set by the
threshold may change as a function of noise (from one SNR to another). In
order to minimize this eect, a smoothing operation is applied on A [32].
The smoothing of the marginal frequency distribution can be applied in
a number of dierent ways. One of the most ecient methods is to apply
a linear adaptive filter to attenuate the noise followed by a moving average
filter to smooth the edges and local peaks. The smoothing operation is then
followed by a normalization.
An adaptive filter is a filter that changes behavior based on the statistical
characteristics of the input signal within the filter. A Wiener filter is a good
choice. The Wiener filter is applied to A using the local neighborhood of size
m-by-1 to estimate the local image mean and standard deviation. The filter
estimates the local mean and variance around each vector element. The local
mean is estimated as [34]
1
= A(n) (17.23)
m n
and the local variance is estimated as

1
2 = A2 (n) 2 (17.24)
m n
where is the m-by-1 local neighborhood of each element in the vector A.

The processed image within the local neighborhood can be expressed as
max( 2 2 , 0)
b(n) = + (A(n) ) (17.25)
2
where v is the noise variance estimated using the average of all the local
estimated variances. When the variance is large, the filter performs little
smoothing and when the variance is small, it performs more smoothing. For
PWVD and CWD images a local neighborhood of = 10 is used and for
the QMFB images = 4 is used. Figure 17.29(b) shows the output of the
adaptive filter for the input MFD of the Frank signal with Nc = 36. Note the
considerable noise attenuation.
Although the adaptive noise attenuation gives promising results, the
threshold determination may be aected by the local noise peaks that could
not be reduced by the adaptive filter. To avoid this problem a moving average
filter is applied to the output of the adaptive Wiener filter. As a generaliza-
tion of the average filter, an averaging over N + M + 1 neighboring points
can be considered. The moving average filter is represented by the following
dierence equation [35]
M
1
y(n) = x(n k) (17.26)
N +M +1
k=N
where x(n) is the input and y(n) is the output. The corresponding impulse
response is a rectangular pulse.
For PWVD and CWD images a window length of N + M + 1 = 10 is used
and for QMFB images N + M + 1 = 4 is used. The moving average filter
output, Aavg is then normalized as
Aavg
An = (17.27)
max(Aavg )
where An is the normalized smoothed MFD. After normalization a histogram

of 100 bins is generated for PWVD and CWD images and a histogram of 30
bins is generated for QMFB images. Using these histogram bins a threshold
is determined. Threshold determination is illustrated in Figure 17.30(a) using
the histogram of An shown in Figure 17.30(b) for n =30 bin. Note that the
corresponding value to the 30th bin Th = 0.2954 is selected as the threshold.
For the simulation purposes the histogram bin numbers are optimized using a
Figure 17.30: Threshold determination showing (a) normalized energy values,

(b) histogram of energy values and (c) cropped frequency band of interest
using n = 30.
Figure 17.31: (a) LPF output, (b) cropped region, and (c) contour plot of
the cropped region showing the Frank modulation.
range of values for each detection technique and each network. The bin num-
ber that provides the best Pcc is selected. Once the threshold is determined,
the values of An below the threshold are set to zero. Then the beginning and
ending frequencies of the frequency band of interest are determined as shown
in Figure 17.30(c). Using the lowest and highest frequency values from the
frequency band of interest the modulation energy can now be cropped from
the image.
After the determination of the modulation band of interest the energy is
autonomously cropped from the LPF output containing the noise filtered im-
age. The cropping was illustrated in Figure 17.31. Figure 17.31(a) shows the
LPF output that is obtained previously, Figure 17.31(b) shows the cropped
region and Figure 17.31(c) shows the contour plot where the signal energy
can easily be seen. Once the LPF output is cropped, the new image is resized
to 50 400 pixels for the PWVD and CWD images. The QMFB images are
resized to 30 120 pixels. Resizing is done in order to obtain as much simi-
larity as possible between the same modulation types. Following the resizing
operation the columns of the resized image are formed with the feature vector
of size 50 400 = 20, 000 for PWVD and CWD images, and of size 30 120 =
3,600 for the QMFB images.
17.5.3 Principal Components Analysis

Principal components analysis (PCA) is mathematically defined as an or-
thogonal linear transformation that transforms the data to a new coordinate
system such that the greatest variance by any projection of the data comes to
lie on the first coordinate (called the first principal component), the second
greatest variance on the second coordinate, and so on [36]. In other words,
PCA is a rotation of the existing axes to new positions in the space defined
by the original variables, where there is no correlation between the new vari-
ables defined by the rotation. PCA is theoretically the optimum transform
for a given data set in least square terms. That is, the method projects the
high-dimensional data vectors onto a lower dimensional space by using a pro-
jection which best represents the data in a mean square sense. Using PCA
the given data vector is represented as a linear combination of the eigenvec-
tors obtained from the data covariance matrix. As a result, lower dimensional
data vectors may be obtained by projecting the high-dimensional data vectors
onto a number of dominant eigenvectors [37].
PCA can be used for dimensionality reduction of the feature vector by
retaining those characteristics of the cropped modulation that contribute
most to its variance, by keeping lower-order principal components and ig-
noring higher-order ones. This assumes of course, that the low-order compo-
nents contain the most important features of the LPI modulation within the
cropped (and resized) T-F data. To facilitate the PCA, we form a training
matrix X as shown in Figure 17.32 where N is the length of the feature vec-
tor and P is the number of training signals, which is 50 for our results. It is
important to note that the mean has been subtracted from the data set.
The PCA maps the ensemble of P N-dimensional vectors
X = x1 , x2 , , xp onto an ensemble of P D-dimensional vectors Y =
y 1 , y 2 , , y p where D < N using a linear projection. This linear projection
can be represented by a rectangular matrix A so that [37]
Y = AH X (17.28)
where A has orthogonal column vectors, i = 1, 2, , P and H is the Her-

mitian operation. The matrix A is selected as the P D matrix containing the
D eigenvectors associated with the larger eigenvalues of the data covariance
matrix XH X. With this choice of transformation matrix A, the transformed
data vectors Y have uncorrelated components.
The matrix X is obtained first to form the training data set. The feature
extraction algorithm is applied to the images in the Training folder for
each detection technique. The cropped images are resized and a column
vector is formed to represent the signal modulation. These column vectors
are stacked together to form the training data set matrix. The mean of the
training matrix is calculated column wise and the mean is subtracted from
the training data set matrix giving the matrix X. This operation is illustrated
in Figure 17.32 where P is the number of training signals which is 50 for this
example, and N is the length of the feature vectors. For PWVD and CWD
X is of dimension 20,000 50 (50 training signals) and for the QMFB X is
of dimension 3,60050.
Figure 17.32: Forming the training matrix X.
Figure 17.33 shows a block diagram of the PCA signal processing. In

order to obtain the eigenvectors of X, singular value decomposition (SVD)
may be performed. SVD states that any N P matrix X can be decomposed
as [37]
X=U VH (17.29)
where U is the N N unitary matrix, V is the P P unitary matrix and
is the N P matrix of nonnegative real singular values. Note that
H H
XH X = V (U)H U VH = V( )VH (17.30)
Figure 17.33: Principal components analysis.

indicates that the eigenvectors of XH X are contained in the V matrix and

the eigenvalues of XH X are the squared singular values of X, which are the
diagonal elements of the matrix H . It can similarly be shown that the
eigenvectors of XXH are contained in the U matrix.
If p = min(P, N ), both XXH and XH X will have the same p nonzero
eigenvalues. The product of X and V gives
XV = U VH V = U (17.31)
since V is unitary and the eigenvectors associated with nonzero eigenvalues

can be extracted by
1
U = XV (17.32)
As a result, the nonzero eigenvalues of the higher dimensional covariance
matrix XXH may be computed by computing the SVD of the smaller dimen-
sional covariance matrix XH X.
Following the SVD of the data matrix and determination of the eigenvector
matrix U, dimensionality reduction is performed using the projection (trans-
formation) matrix A. The matrix A is composed of D eigenvectors selected
from the eigenvector matrix U corresponding to D largest eigenvalues. In or-
der to find the D largest eigenvalues, the biggest eigenvalue is multiplied by a
threshold constant and the eigenvalues above the product are taken. Let Th
be the eigenvalue selection threshold constant. In our example, three values
are used as Th = [0.001, 0.005, 0.01]. For each case, once the eigenvalues are
found, four variations of eigenvector selection are used. Let these variations
be i , where i = 0, 1, 2, 3. The variations are defined by the i index as follows:
0 : All the eigenvectors corresponding to the eigenvalues above Th are

used to form the matrix A.

selected initially; all of them except the eigenvector corresponding to
the eigenvalue with the highest value are used to form the matrix A.

selected initially; all of them except the two eigenvectors corresponding
to the two eigenvalues with the highest values are used to form the ma-
trix A.

selected initially; all of them except the three eigenvectors correspond-
ing to the three eigenvalues with the highest values are used to form
the matrix A.
Once the projection matrix A is generated, both the training matrix X

and the test signals are projected onto a smaller dimensional feature space.
The dataset is reduced in dimension to D using the projection process. The
projected data is then used for classification.
17.5.4 Classification Using Modified Feature Extraction

The classification results in this section use an extended database to determine
the performance of the modified feature extraction technique as a function
of the SNR. After the database is described, steps to optimize the MLP
and RBF are discussed. Classification results are then shown for both the
TestSNR signals (same signals used in training but with varying SNR) and
the TestMod signals (dierent modulations and varying SNR).
Extended Database
To investigate the detailed performance of the modified feature extraction and
classification process, a more extensive database is developed that consists of
12 LPI modulation techniques each having 21 SNR levels
(10 dB, 9 dB, ,9 dB, 10 dB). The LPI modulation techniques include
Costas frequency hopping, Costas frequency hopping plus a Barker phase
shift keying, FMCW, PSK and FSK. PSK signals include polyphase (Frank,
P1, P2, P3, P4) and polytime (T1, T2, T3, T4) codes. This database allows
a detailed look at the Pcc as a function of the SNR. The signals are generated
using the LPIT and placed in the Input folder within the proper subfolder
(TestSNR, TestMod, Training, Signals). Note that the Signals folder should
contain only one signal from each modulation type being used. This folder
is used to correlate the modulation prefix (F for FMCW, FR for Frank and
so forth) to build the confusion matrix. The output T-F and B-F images
from the detection signal processing (Wigner-Ville, Choi-Williams, quadra-
ture mirror filtering, cyclostationary processing) are automatically placed in
the corresponding output folder (e.g., QMFB output). Before the feature ex-
traction and nonlinear classification signal processing algorithms are run, the
detection output signals within the TestMod and TestSNR folders that
have the same SNR must be collected and put into a folder that designates
the SNR (e.g., TestMod-10, and TestSNR4). The folder structure should be
as shown in Figure 17.34. Note that the SNR = 10 dB signals for each modu-
lation are used for training. This is a choice that the user can make. Training
the LPI feature extraction and classification networks with only signal only
waveforms however, is not realistic since any received signal will have a noise
component related to the thermal noise present in the intercept receiver and
the range of the LPI emitter.
Figure 17.34: Folder structure for TestSNR, TestMod, and Training (10-dB
TestSNR only) [32].
Table 17.9: Costas Frequency Hopping Modulation Parameters for TestSNR

(fs = 7,000 Hz)
Signal Frequency Frequency

modulation sequence (Hz) duration tp (ms)
Costas {3, 2, 6, 4, 5, 1} 200 5
{2, 4, 8, 5, 10, 9, 7, 3, 6, 1} 150 3
Signals Used for TestSNR and TestMod

The signals used to test the performance of the feature extraction and clas-
sification signal processing for various values of SNR are described below.
This database is used to generate the results shown in this section. Super-
vised training of the autonomous classification process is done with the signal
modulations below using SNR = 10 dB. The parameters for the Costas codes
and the Costas codes with a Barker PSK used for testing the performance
of the classification signal processing as a function of the SNR are shown in
Tables 17.9 and 17.10, respectively. The FMCW signal parameters are shown
in Table 17.11. The polyphase signals (Frank, P1P4) used for testing the
performance as a function of the SNR are as shown in Table 17.12. Tables
17.13 and 17.14 are the polytime signals T1, T2, T3, and T4 respectively.
These signals are used to evaluate the performance of the autonomous clas-
sification Pcc when the received signal has the same modulation parameters
but dierent SNR.
After supervised training of the classification network using the
SNR = 10 dB signals from the TestSNR database, the performance using
the signals from the TestMod database are evaluated. It is important to
Table 17.10: Costas Plus Barker PSK Frequency Hopping Modulation

Parameters for TestSNR (fs = 7,000 Hz)
Signal Frequency Barker subcode

modulation sequence (Hz) period tb (ms)
Costas + PSK {3, 2, 6, 4, 5, 1} 150 1 (Nc = 5)
{5, 4, 6, 2, 3, 1} 300 0.3 (Nc = 13)
Table 17.11: FMCW Modulation Parameters for TestSNR

Signal Carrier Modulation Modulation
modulation frequency bandwidth period
fc (Hz) F (Hz) tm (ms)
FMCW 1,495 250 15
2,195 800 15
point out that this is a more dicult (and realistic) situation. The TestMod
signals model the interception of a waveform with a modulation that is not
within the training set. The signals in TestMod are also tested as a function
of the SNR. The Costas frequency hopping modulation parameters and the
Costas frequency hopping plus Barker PSK modulation parameters used for
TestMod are shown in Tables 17.15 and 17.16, respectively. The FMCW sig-
nals used for testing the performance of the signal processing as a function of
the SNR are as shown in Table 17.17. The polyphase signals (Frank, P1P4)
used for testing the performance as a function of the SNR are as shown in
Table 17.18. The polytime signals are shown in Tables 17.19 and 17.20.
Optimizing the Feature Extraction and Classification Network

Using the initial nonlinear network parameters two feature extraction
parameters, LPF cuto frequency and histogram bin, must be optimized.
Using the optimum values derived, the PCA network parameters are then
optimized. The Pcc results shown are with the final optimum values. The
optimization is performed using the test signals with SNR = 10 dB. The
optimum parameter selection is based on the highest average Pcc.
For each detection technique, the MLP network configuration starts with
a default set of values for the epochs, the number of neurons in the first and
second hidden layers S1 , S2 , the eigenvalue selection threshold constant T h
and eigenvector selection variations i . Once the initial values are set, an
optimization is performed to determine optimum values for the LPF digital
frequencies 1 = 2 and histogram bin number. After these two values are
Table 17.12: Polyphase Modulation Parameters for TestSNR
Signal Carrier Code Cycles per

modulation frequency fc (Hz) length Nc subcode cpp
Frank 1,495 9 5
25 2
36 1
............ ................... .......... .............
2,195 16 6
25 3
P1 1,495 9 5
25 2
36 1
............ ................... .......... .............
2,195 16 4
16 5
P2 1,495 16 3
36 1
............ ................... .......... .............
2,195 16 5
36 3
P3 1,495 9 4
9 5
36 1
............ ................... .......... .............
2,195 16 6
25 3
P4 1,495 9 5
25 2
36 1
............ ................... .......... .............
2,195 16 4
16 5
Table 17.13: Polytime T1, T2 Modulation Parameters for TestSNR
Signal Carrier Code No. phase No. code

modulation frequency period states segments
fc (Hz) T (ms) n k
T1 1,495 30 2 5
30 3 4
............ ........... ....... ........... ..........
2,195 30 2 3
30 2 4
30 4 3
T2 1,495 30 4 3
30 8 4
............ ........... ....... ........... ..........
2,194 30 4 3
30 4 4
30 6 3
Table 17.14: Polytime T3, T4 Modulation Parameters for TestSNR
Signal Carrier Modulation Modulation No. phase

modulation frequency period bandwidth states
fc (Hz) tm (ms) F (Hz) n
T3 1,495 25 300 4
30 900 9
............ ........... ............. ............. ...........
2,195 25 400 2
30 1000 7
35 800 6
T4 1,495 25 400 2
30 550 3
30 850 7
............ ........... ............. ............. ...........
2,194 30 600 5
30 900 9
Table 17.15: Costas Frequency Hopping Modulation Parameters for TestMod

(fs = 7,000 Hz)
Signal Frequency Frequency

modulation sequence (Hz) duration tp (ms)
Costas {5, 4, 6, 2, 3, 1} 400 5
Table 17.16: Costas Plus Barker PSK Frequency Hopping Modulation

Parameters for TestMod (fs = 7,000 Hz)
Signal Frequency Barker subcode

modulation sequence (Hz) period tb (ms)
Costas + PSK {3, 2, 6, 4, 5, 1} 200 0.4 (Nc = 11)
{5, 4, 6, 2, 3, 1} 250 0.7 (Nc = 7)
Table 17.17: FMCW Modulation Parameters for TestMod

Signal Carrier Modulation Modulation
modulation frequency bandwidth period
fc (Hz) F (Hz) tm (ms)
FMCW 1,495 500 20
2,195 400 20
Table 17.18: Polyphase Modulation Parameters for TestMod
Signal Carrier Code Cycles per

modulation frequency fc (Hz) length Nc subcode cpp
Frank 1, 495 9 4
............ ................... .......... .............
2, 195 16 4
16 5
P1 1,495 9 4
............ ................... .......... .............
2,195 16 6
25 3
P2 1,495 16 2
............ ................... .......... .............
2,195 16 4
P3 1,495 25 2
............ ................... .......... .............
2,195 16 4
16 5
P4 1,495 9 4
............ ................... .......... .............
2,195 16 6
25 3
Table 17.19: Polytime T1, T2 Modulation Parameters for TestMod
Signal Carrier Code No. phase No. code

modulation frequency period states segments
fc (Hz) T (ms) n k
T1 1,495 30 4 4
30 6 3
............ ........... ....... ........... ..........
2,195 30 3 3
T2 1,495 30 6 4
30 4 5
............ ........... ....... ........... ..........
2, 194 30 8 3
Table 17.20: Polytime T3, T4 Modulation Parameters for TestMod
Signal Carrier Modulation Modulation No. phase

modulation frequency period bandwidth states
fc (Hz) tm (ms) F (Hz) n
T3 1,495 30 500 5
35 700 8
............ ........... ............. ............. ...........
2,195 30 600 3
T4 1,495 35 700 6
............ ........... ............. ............. ...........
2,194 25 450 4
35 750 8
Table 17.21: Optimum Feature Extraction and MLP Parameters
Detection 1 = 2 Bin S1 = S2 T h i epochs

Wigner-Ville 0.1 45 80 0.001 1 5,000
Choi-Williams 0.1 15 80 0.001 0 6,000
Quadrature mirror 0.4 18 60 0.005 0 5,000
Table 17.22: Optimum Feature Extraction and RBF Parameters
Detection 1 = 2 Bin goal T h i

Wigner-Ville 0.2 55 2000 0.9 0.001 0
Choi-Williams 0.5 55 3500 0.9 0.001 0
Quadrature mirror 0.6 4 25 0.8 0.001 0
found, a second optimization for epochs, S1 , S2 , Th and i is performed.

Once all the values are found and set the classification network is tested. For
the classification of PWVD images the initial values used are epochs = 6,000
S1 = S2 = 50, Th = 0.001 and i = 0 . After optimization 1 = 2 = 0.1
and the histogram bin number is 45. Using these values, the remaining para-
meters giving optimum Pcc are S1 = S2 = 80, Th = 0.001, 1 and epochs =
5,000. The optimization is repeated for the Choi-Williams, the quadrature
mirror filtering, and the Wigner-Ville distribution detection techniques. Ta-
ble 17.21 shows the resulting optimum parameters using the MLP. A similar
optimization is also run for the RBF. The optimum parameters are shown in
Table 17.22 for the RBF.
17.5.5 Classification Results with the Multilayer

Perceptron
The classification results are presented for comparison of the three T-F detec-
tion techniques including the Choi-Williams distribution (CWD), the pseudo
Wigner-Ville distribution (PWVD) and the quadrature mirror filter bank
approach (QMFB). The MLP classification results for the Costas frequency
hopping code are shown in Figure 17.35 and the results for the Costas plus
Barker PSK are shown in Figure 17.36. The MLP classification results for the
FMCW are shown in Figure 17.37. The MLP results for the Frank, P1P4,
T1T4 are shown in Figures 17.3817.46.
All the detection techniques show similar results on the TestSNR case.
Most of the modulations are classified with more than 80% classification rate
for SNR > 0 dB. There is a considerable stability in classification of signals
with SNR > 0 dB. This stability indicates that the autonomous modulation
Figure 17.35: Costas code classification results using the MLP for (a) Test-
SNR and (b) TestMod.
Figure 17.36: Costas frequency hopping plus PSK classification results using
the MLP for (a) TestSNR and (b) TestMod.
Figure 17.37: FMCW classification results using the MLP for (a) TestSNR
and (b) TestMod.
Figure 17.38: Frank classification results using the MLP for (a) TestSNR and
(b) TestMod.
Figure 17.39: P1 classification results using the MLP for (a) TestSNR and
(b) TestMod.
(b) TestMod.
(b) TestMod.
(b) TestMod.
Figure 17.43: T1 classification results using the MLP for (a) TestSNR and
(b) TestMod.
(b) TestMod.
(b) TestMod.
(b) TestMod.
energy isolation and cropping becomes more sensitive to noise variations be-
low 0 dB. The Pcc of Frank, FSK/PSK, FMCW, T1, T2, and T4 modulations
with PWVD and CWD techniques exhibit 100% for most of the SNR levels
above 0 dB.
Concerning the TestMod case, the best results are obtained in the clas-
sification of FMCW, Costas, FSK/PSK, P2, and T2 modulations while the
worst results are obtained in the classification of polyphase codes. Note that
most of the results for Frank, P1, P3, and P4 modulations are below Pcc =
0.4. Classification of Costas, FSK/PSK, FMCW, P2, P4, T1, T2, T3, and
T4 modulations with PWVD and CWD techniques exhibit similar results.
Overall, the classification results with the PWVD technique outperform the
other detection techniques. Overall the QMFB technique performs worse
than the other techniques but it does well in the classification of T2 and T4
modulations for SNR > 5 dB. Recall that the QMFB images have a very
low resolution compared to the PWVD and CWD images, which becomes a
disadvantage for modulation discrimination.
One interesting result is observed on Costas modulation classification.
While the Pcc for TestMod is 100% with all detection techniques, the Pcc
for TestSNR is not. This is an unexpected result. It is expected that the
TestSNR results would outperform the TestMod results since the signals used
in TestSNR have the same parameters as the training signals. In this sense
the TestSNR results can be used as a measure of reliability. This shows
that, although the Costas results seem very good for TestMod case, they
may not be reliable. Further, it is shown that the classification of Costas
code is best performed with CWD detection technique for SNR > 4 dB.
Note also that it is not necessarily true that the TestMod results perform
better if the TestSNR results perform well. The Pcc for TestMod depends
on the modulation discriminative power of the feature extraction algorithm
implemented.
17.5.6 Classification Results with the Radial Basis

Function
The RBF classification results for the Costas frequency hopping signals are
shown in Figure 17.47. The RBF classification results for the Costas plus
PSK frequency hopping signals are shown in Figure 17.48. The results for
the FMCW signals are shown in Figure 17.49. The results for the Frank
polyphase signals are shown in Figure 17.50. The RBF classification results
for the P1P4 polyphase signals are shown in Figures 17.5117.54. The results
for the T1 and T2 polytime signals are shown in Figures 17.5517.58.
For the TestSNR classification, the Frank, FMCW, P2, T1, T2, T3 and T4
modulations are classified with greater than Pcc = 80% probability of correct
classification for SNR > 2 dB. The autonomous modulation energy isolation
and cropping becomes more sensitive to noise variations below SNR = 2 dB.
Figure 17.47: Costas classification results using the RBF for (a) TestSNR and
(b) TestMod.
The FMCW modulation is classified with 100% for SNR > 4 dB, and the
P2 modulation is classified with 100% for SNR > 4 dB with all detection
techniques.
Concerning the TestMod case, the best results are obtained in the clas-
sification of FMCW, Costas, P1, P2 and T2 modulations while the worst
results are obtained in the classification of P4, T1 and T3 modulations. The
FMCW modulation is classified 100% with PWVD detection technique for
SNR > 10 dB and 100% with CWD detection technique for SNR > 1 dB .
The T4 modulation is classified 100% with PWVD detection for SNR > 2 dB
and the P2 modulation is classified 100% with CWD detection for SNR > 3
dB. Overall, the classification results with the PWVD technique outperform
the other detection techniques. The QMFB technique performs worse than
the other two detection techniques. It outperforms however, the other tech-
niques in the classification of the P1 modulation with a Pcc above 66%.
MLP and RBF Comparision

Both the MLP and RBF networks are examples of nonlinear layered feedfor-
ward networks. The important trait that is illustrated in the figures shown
is that the RBF classification Pcc results are not as stationary as those for
the MLP. This is due in part to the fact that the RBF has a single hidden
Figure 17.48: Costas plus PSK classification results using the RBF for
(a) TestSNR and (b) TestMod.
Figure 17.49: FMCW classification results using the RBF for (a) TestSNR
and (b) TestMod.
Figure 17.50: Frank classification results using the RBF for (a) TestSNR and
(b) TestMod.
Figure 17.51: P1 classification results using the RBF for (a) TestSNR and
(b) TestMod.
(b) TestMod.
(b) TestMod.
(b) TestMod.
Figure 17.55: T1 classification results using the RBF for (a) TestSNR and
(b) TestMod.
(b) TestMod.
(b) TestMod.
(b) TestMod.
layer, whereas the MLP has two hidden layers. Also, the computation nodes
of the MLP within a hidden or output layer share a common neuronal model
whereas the computation nodes in the hidden layer of the RBF network have
a significantly dierent purpose from those in the output layer of the network.
The argument of the activation function of each hidden unit in the RBF net-
work computes the Euclidean norm (distance) between the input vector and
the center of that unit. For the MLP the activation function of each hid-
den unit computes the inner product of the input vector and the synaptic
weight vector of that unit. Finally, the MLP constructs a global approxima-
tion to the nonlinear input-output mapping of the LPI modulations while the
RBF network uses exponentially decaying localized nonlinearities (Gaussian
functions) to construct local approximations to the nonlinear input-output
mappings. For the approximation of a nonlinear input-output mapping, the
MLP requires a smaller number of parameters than the RBF network for the
same degree of accuracy [23].
17.6 Summary
Autonomous (no human operator intervention) feature extraction and classifi-
cation algorithms that can be used for identification of LPI radar modulations
using time-frequency (T-F) images are presented. The first approach uses a
histogram processing of the marginal frequency distribution to identify the
modulation within the T-F image. After the modulation is cropped from the
image, an adaptive binarization process is used to develop a feature vector
for classification of the modulation contained in the signal. Classification
techniques evaluated include the multilayer perceptron and the radial basis
function neural networks. To evaluate the performance of the feature ex-
traction processing, the classification results for five LPI modulations were
investigated. The algorithms were trained using an SNR = 10 dB. To eval-
uate the classification performance of the algorithms, a database containing
the LPI signals with varying SNR was used (TestSNR database). A second
database containing the same modulations but with varied parameters (Test-
Mod database) was also used. The percent of correctly classified modulations
for this considerably more dicult (but more realistic) database set of signals,
were much lower than the TestSNR signals.
Due to the poor stability of the cropping region and the large size of the
feature vector, a modified feature extraction method was also presented. The
modified approach included the addition of a filtering process (to reduce the
presence of high frequency noise) and the use of principal components analysis
(to reduce the dimensionality). Results showed good improvement and the
performance of the modified feature extraction technique was evaluated as a
function of SNR. Both the TestSNR and TestMod results were shown. By
eliminating the need for a human operator to examine the T-F results, real-
time signal analysis is possible, which can allow a faster response management
to the intercepted threat signals.
References
[1] Azzouz, E., and Nandi, A. K., Automatic Modulation Recognition of Com-
munication Signals, Kluwer Academic Publishers, 1996.
[2] Nandi, A. K., and Azzouz, E., Algorithms for automatic modulation recog-
nition of communication signals, IEEE Trans. on Communications, Vol. 46,
No. 4, pp. 431436, April 1998.
[3] Azzouz, E., and Nandi, A. K., Automatic identification of digital modulation
types, Signal Processing, Vol. 47, No. 1, pp. 5569, 1995.
[4] Louis, C., and Sehier, P. Automatic modulation recognition with a hierar-
chical neural network, Record of the IEEE Military Communications Con-
ference, MILCOM 94, Vol. 3, pp. 713717, October 1994.
[5] Lin, Y-C., and Kuo, C-C. J., Modulation classification using wavelet trans-
form, Proc. of the SPIE, Vol. 2303, pp. 260271, Wavelet Applications in
Signal and Image Processing II, Andrew F. Laine, Michael A. Unser; Eds.,
Oct. 1994.
[6] Reichert, J., Automatic classification of communication signals using higher
order statistics, Proc. of the IEEE International Conference on Acoustics,
Speech and Signal Processing, Vol. 5, pp. 221224, 2326 Mar. 1992.
[7] Schreyogg, C., Kittel, K., Kressel, U., and Reichert, J., Robust classification
of modulation types using spectral features applied to HMM, Record of the
IEEE Military Communications Conference, MILCOM 97, Vol. 3, pp. 1377
1381, Nov. 1997.
[8] Breakenridge, C. Nonstationary signal classification using time-frequency
optimization, Proc. of the 10th IEEE International Conference on Electron-
ics, Circuits and Systems, ICECS, pp. 132135, 1417 Dec. 2003,
[9] Roberts, G., Zoubir, A. M., and Boashash, B., Time-frequency classifica-
tion using a multiple hypothesis test: an application to the classification of
humpback whale signals, Proc. of the IEEE International Conference on
Acoustics, Speech, and Signal Processing, Vol. 1, pp. 563566, 1997.
[10] Breakenridge, C., and Mesbah, M., Minimum classification error using time-
frequency analysis, Proc. of the 3rd IEEE International Symposium on Sig-
nal Processing and Information Technology, ISSPIT, pp. 717720, 1417 Dec.
2003.
[11] Wang, T. Deng, J., and He, B., Classification of motor imagery EEG pat-
terns and their topographic representation, Proc. of the International Con-
ference on Engineering in Medicine and Biology Society, IEMBS 04, pp.
43594362, 15 Sept. 2004.
[12] Gache, N., Chevret, P., and Zimpfer, V., Target classification near complex
interfaces using time-frequency filters, Proceedings of the IEEE International
Conference on Acoustics, Speech and Signal Processing, Vol. 4, pp. 2433
2436, 1215 May 1998.
[13] Kim, K.-T., Choi, I.-S., and Kim, H.-T., Ecient radar target classification
using adaptive joint time-frequency processing, IEEE Trans. on Antennas
and Propagation, Vol. 48, No. 12, pp. 17891801, Dec. 2000.
[14] Chilukuri, M. V., Dash, P. K., and Basu, K. P., Time-frequency based pat-
tern recognition technique for detection and classification of power quality
disturbances, Proc. of the IEEE Region 10 Conference, Vol. 3, pp. 260
263, 2124 Nov. 2004.
[15] Gillespie, B. W., and Atlas, L. E., Optimizing time-frequency kernels for
classification, IEEE Trans. on Signal Processing, Vol. 49, No. 3, pp. 485
496, March 2001.
[16] Shelton, C. P., Human Interface/Human Error, em Dependable Embedded
Systems, Carnegie Mellon University pp. 18849b Spring 1999.
[17] De vries, M. F. L., Koeners, G. J. M., Roefs, F. D., Van ginkel, H. T. A.,
and Theunissen, E., Operator Support for Time-Critical Situations: Design
and Evaluation, Proc. of the IEEE/AIAA 25th Digital Avionics Systems
Conference, Delft Univ. of Tech., Netherlands, pp. 114, Oct. 2006.
[18] Sheridan, T. B., Humans and Automation: System Design and Research Is-
sues, John Wiley & Sons, Inc., 2002.
[19] Upperman, G. J., Upperman, T. L., Fouts, D. J., and Pace, P. E., Ecient
time-frequency and bi-frequency signal processing on a reconfigurable com-
puter, IEEE Asilomar Conference on Signals, Systems and Computers, Nov.
2008.
[20] Lippmann, R. P., An introduction to computing with neural nets, IEEE
ASSP Magazine, pp. 422, April 1987.
[21] Hush, D. R., and Horne, B. G., Progress in supervised neural networks,
IEEE Signal Processing Magazine, pp. 839, Jan. 1993.
[22] Wong, M. L. D., and Nandi, A. K., Automatic digital modulation recognition
using artificial neural network and genetic algorithm, Signal Processing, Vol.
84, No. 2, pp. 351365, February 2004.
[23] Haykin, S., Neural NetworksA Comprehensive Foundation, Second Ed., Up-
per Saddle River, New Jersey: Prentice Hall, 1999.
[24] Theodoridis, S., and Koutroumbas, K., Pattern Recognition, Third Ed., San
Diego, CA: Academic Press, 2006.
[25] Broomhead, D. S., and Low, D., Multi-variate functional interpolation and
adaptive networks, Complex Systems, Vol. 2, pp. 321355, 1990.
[26] Husain, H., Khalid, M., and Yusof, R., Nonlinear function approximation
using radial basis function neural networks, Student Conference on Research
and Development, pp. 326329, July 2002.
[27] Atlas, L., Owsley, L., McLaughlin, J., and Bernard, G., Automatic feature-
finding for time-frequency distributions, Proceedings of the IEEE-SP Inter-
national Symposium on Time-Frequency and Time-Scale Analysis, pp. 333
336, 1821 June 1996.
[28] Zilberman, E. R., and Pace, P. E., Autonomous time-frequency morpho-
logical feature extraction algorithm for LPI radar modulation classification,
Proc. of the IEEE International Conf. on Image Processing, 2006.
[29] Cohen, L., Time-frequency distributionsA review, Proc. of the IEEE, Vol.
77, No. 7, p. 941981, 1989.
[30] Zilberman, E. R., and Pace, P. E., Autonomous cropping and feature
extraction using time-frequency marginal distributions for LPI radar clas-
sification, Proc. of the IASTED International Conf. on Signal and Image
Processing, Aug., 2006.
[31] Van De Ville, D.,Van de Walle, R., Philips, W., and Lemahieu, I., Image
resampling between orthogonal and hexagonal lattices, Proc. International
Conf. on Image Processing, Vol. 3, pp. III-389III-392, 2428 June 2002.
[32] Gulum, T. O., Autonomous Non-Linear Classification of LPI Radar Signal

Modulations, Naval Postgraduate School Masters Thesis, 2007.
[33] Gonzales, R. C., Woods R. E. and Eddins, S. L., Digital Image Processing
Using MATLAB, Upper Saddle River, NJ, Prentice Hall, 2004.
[33] Lim, J. S., Two-Dimensional Signal and Image Processing, Englewood Clis,
NJ Prentice Hall, 1990.
[34] Oppenheim, A. V., Willsky, A. S., and Nawab, S. H., Signals and Systems,
Upper Saddle River, NJ: Prentice Hall, 1997.
[35] Jollie, I. T., Principal Component Analysis, Series: Springer Series in Sta-
tistics, 2nd ed., Springer, New York, 2002.
[36] Fargues, M. P., Investigation of Feature dimension Reduction Schemes for
Classification Applications, Naval Postgraduate School Technical Report,
Monterey, CA, NPS-EC-01-005, June 2001.
[37] Therrien, C. W., Discrete Random Signals and Statistical Signal Processing,
Prentice Hall, Englewood Clis, NJ, 1991.
Problems
1. Setting Up the Database:
a. Begin by reading the Readme.doc file in the LPI Class folder on the
CD.
b. Using the LPIT, generate with a sampling frequency of fs = 7,000 Hz,
and SNR = 10 dB, a BPSK signal, a polyphase Barker signal, a P1
signal, a P2 signal, a P3 signal, a P4 signal, a Frank code signal, a
FMCW signal, T1(2) and a T3(2) signal. Make sure you record the
parameters for each of your signals.
c. Go to the Input Folder. The Input Folder should have four subfolders.
They are the Signals, TestSNR, TestMod and Training folders. Copy
the 10 signals into the Training folder and the Signals Folder.
d. Copy the 10 signals into the TestSNR folder.
e. For each signal generated in b., generate the same signal but with
SNR = 10, 9, . . . , 0, . . . , 8, 9 dB. Copy these signals into the
TestSNR folder.
f. For each signal generated in b., generate the signal with dierent mod-
ulation parameters and SNR = 10, 9, . . . , 0, . . . , 10 dB. Be sure to
record the modulation parameters for each signal. Copy these signals
into the input TestMod folder.
2. Running the Detection Signal Processing:
a. Go to the LPI class folder and examine and run the MATLAB script file
SCRIPT detection ALL.m. This runs the Wigner-Ville distribution,
Choi-Williams distribution, quadrature mirror filtering, and cyclosta-
tionary detection signal processing algorithms on all of the files within
the Input folder database. Note that these detection algorithms may be
run separately by commenting out all but one of the detection process-
ing calls. The output files from the detection processing are put in
their respective output folders. Each detection algorithm should be flow
diagrammed by the student.
b. Collecting the TestSNR, TestMod Output Files: Within each
TestSNR folder, collect all of the signals that have the same SNR and
put them into a separate folder named as, for example, TestSNR10
(10 dB files), TestSNR-8 (8 dB files) and so on. Repeat this for the
TestMod files. When this is complete, under each output folder you
should have folders Signals, Training, TestSNRxx and TestModxx. You
should also have a folder named TestSNR and a folder named TestMod
that are empty.
3. Running the Feature Extraction and Classification Process
In the LPI class folder, examine and flow diagram the MATLAB scripts for
computing the feature vector (feature extraction) and running the classifica-
tion processing. These scripts are named:
SCRIPT FE Classification PWVD.m;
SCRIPT FE Classification CHOI.m;
SCRIPT FE Classification QMFB.m;
SCRIPT FE Classification CYCL.m.
Note the diary files that are initiated. You should name the diary files so that
you can track the results that you generate. From these diary files, the results
can be extracted and put into the EXCEL spreadsheet included. This enables
the Pcc (probability of correct classification) to be plotted as a function of
the SNR. Be sure to normalize your results to one by dividing by the number
of times the network runs through the classification algorithm (max test).
Include with your classification results, the flow diagrams of the detection,
feature extraction and classification algorithms. Note the software architec-
ture is structured to be flexible enabling any new modulations of interest can
be included in the analysis.
Chapter 18
Autonomous Extraction of
Modulation Parameters
In the previous chapter, autonomous classification techniques were investi-
gated to identify the LPI modulations present on the intercepted signal. In
this chapter, postclassification signal processing techniques are used to au-
tonomously extract the modulation parameters. Algorithms to autonomously
extract the parameters from the time-frequency plane are presented and in-
clude extracting the polyphase modulation parameters from the QMFB and
Wigner-Ville distribution. Autonomous extraction of parameters from the
bifrequency plane (cyclostationary signal processing) is also presented.
18.1 Emitter Clustering

The noncooperative intercept receiver attempts to detect the LPI emitter
waveform and determine the angle of arrival. The intercepted waveform
is detected most eectively using time-frequency techniques (i.e., Wigner-
Ville distribution, quadrature mirror filtering, Choi-Williams distribution)
and bifrequency signal processing (cyclostationary processing). Noncoop-
erative intercept receivers must classify the modulation type autonomously
across a broad spectrum in the presence of noise and multipath. The de-
tection processing results in a two-dimensional image that is preprocessed
in order to produce a feature vector for purposes of classifying the modula-
tion. The detection results are also used for direction finding to determine
the angle-of-arrival (AOA) and geodetic location of the emitter.
The detection T-F and B-F images can also provide the means for increas-
ing the receivers processing gain by autonomously extracting the modulation
parameters. Figure 18.1 shows a block diagram of the autonomous detection,
687
Figure 18.1: Block diagram of autonomous classification and parameter ex-

traction processing for a noncooperative intercept receiver.
and classification including the parameter extraction process. That is, the im-
ages can provide details about the modulation parameters that are unavailable
using power spectral density techniques. The need for human interpretation
of the T-F and B-F results to determine the parameter values however limits
the extraction process to nonreal-time electronic intelligence receivers.
The autonomous parameter extraction of the LPI emitter modulations can
eliminate the need for a human operator and enable near real-time coherent
handling of the threat emitters being intercepted. Parameter extraction fol-
lowed by correlating the modulation parameters of the intercepted waveform
with a database of previously detected emitter parameters or clustering can
then aid in signal tracking and coherent EA response management.
18.2 Polyphase Parameters Using Wigner-Ville

DistributionRadon Transform
This section presents an ecient algorithm to autonomously extract the
polyphase modulation parameters from an intercepted waveform using a novel
Wigner-Ville distributionRadon transform [1]. The modulation parameters
include the bandwidth B, carrier frequency fc , cycles of the carrier frequency
Autonomous Extraction of Modulation Parameters 689
per subcode cpp, code length Nc , and code period T . Results show that the
method results in a small relative error in the extracted parameters for signal-
to-noise ratios as low as 6 dB. The Wigner-Ville distributionRadon trans-
form approach is particularly useful for this time-frequency signal processing
task since the majority of polyphase modulations are developed by approxi-
mating a linear frequency modulation waveform.
We evaluate the sensitivity of the algorithm using the five polyphase
modulations Frank, P1, P2, P3, and P4 for signal-to-noise ratios (SNRs)
of 0 dB and 6 dB. To illustrate the algorithm, a Frank code is used with
Nc = 36 subcodes, a carrier frequency of fc = 1,495 Hz and an analog-to-
digital converter (ADC) sampling frequency of fs = 7 kHz with SNR = 0
dB. The number of carrier frequency cycles within a subcode is cpp = 1, giv-
ing a transmitted bandwidth B = fc /cpp = 1,495 Hz and a code period of
T = 24.1 ms.
18.2.1 Time-Frequency Algorithm Description

A block diagram of the autonomous PWVD-Radon transform algorithm is
shown in Figure 18.2. The carrier frequency fc is extracted by finding the
location of the maximum intensity level within the PWVD image. In order to
extract the code length T and bandwidth B, the Radon transform is computed
from the T-F PWVD image. The Radon transform is the projection of the
image intensity along a radial line oriented at a specific angle. It transforms
the 2-D image with line-trends into a domain of the possible line parameters
and , where is the smallest distance from the origin and is its angle
with the x-axis. In this form, a line is defined as [1].
= x cos + y sin (18.1)
Using this definition of a line, the Radon transform of a 2-D image f (x, y)
can be defined as
8 +
R(, ) = f ( cos s sin , sin + s cos ) ds (18.2)

where the s-axis lies along the perpendicular to as shown in Figure 18.3.
Here s can be calculated as
s = y cos x sin (18.3)
Note and s can be calculated from x, y, and using (18.1) and (18.3)
[2].
In this work the projection of the images are computed as line integrals
from multiple sources along parallel paths in a given direction. The beams
are spaced 1 pixel unit apart. Figure 18.4 shows the Gray-scale image from
Figure 18.2: Block diagram of the Wigner-Ville distributionRadon transform

technique.
the PWVD illustrating the parameters to be extracted (i.e., signal bandwidth

B and polyphase code period T ). The algorithm measures B and T by im-
plementing the Radon transform to find and d. Here d is the perpendicular
distance between consecutive linear energy lines at the modulation angle s
[3, 4].
Once s and d for the modulation are determined, B and T can be cal-
culated using geometrical relations [3, 4]. The Radon transform is imple-
mented so that the parallel-beam projections of the image are taken between
[0 , 179 ]. Once the transform is completed it is normalized. In some cases the
maximum intensity on the transform may occur around = 90 , which cor-
responds to the marginal frequency distribution (MFD) and around = 0 ,
which corresponds to the time marginal. In order to avoid the detection
of the angle corresponding to the MFD and marginal time distribution, it
is assumed that the slope of linear energy lines are not between [10 , 10 ]
and between [85 , 95 ] and the projections on angles between = [80 , 100 ],
[0 , 5 ], [175 , 179 ] are masked, and set to zero. After masking, the location
of the maximum intensity level of the transform is found. The corresponding
projection angle at this location gives s . Once s is found, the projection at
angle s is cropped from the masked Radon transform and a projection vector
is obtained. Figure 18.5 illustrates the cropping of the projection vector A
Figure 18.3: Geometry of the Radon transform.
Figure 18.4: Radon transform geometry on PWVD image.

Figure 18.5: Radon transform and projection cropping on angle s = 156 .
at angle , from the masked Radon transform of the Frank code.

From Figure 18.5 the number of modulation energy lines contained in the
PWVD image (number of code periods intercepted) can easily be detected
from both the Radon transform and the projection vector at angle s . The
ripples between each modulation energy component correspond to the addi-
tive noise and the cross term integration at angle s . The projection vector
is then smoothed with a Wiener filter.
max( 2 2 , 0)
b(n) = + (A(n) ) (18.4)
2
where n is an index into the local neighborhood of size , is the estimated
local mean, 2 is the estimated local variance and 2 is the estimated noise
variance obtained by using the average of all the estimated local variances.
A local neighborhood of = 10 is used in the adaptive filter.
Following smoothing, the projection vector is thresholded with a threshold
equal to one half of the maximum value of the projection vector.
Figure 18.6(a) shows the filtered projection vector and Figure 18.6(b) shows
the thresholded projection vector after filtering. After thresholding several
distances can be found between the nonzero values in the projection vector
which correspond to the consecutive modulation energy components. The
final distance d (pixels) can be determined by finding the mean value of these
distances. In Figure 18.4, once d is found the modulation code period can
now be found using [3, 4]
Figure 18.6: (a) Filtered projection vector and (b) threshold projection vector
after filtering.
} ]
1 d
T = (18.5)
fs cos (s )
and the bandwidth B can be found using the relation
} ]
d
B = f / tan(s ) (18.6)
cos (s )
where f is the frequency resolution of the PWVD image. Note that (18.5)
is not applied to P2 coded signals since the modulation has an opposite T-F
slope. For P2 code modulation, the following relationship applies:
} ]
1 d
T = (18.7)
fs cos (s )
Figure 18.7: Parameters of polyphase modulation signals.
Once fc , T , and B are obtained, the code length Nc can be found using
Nc = T B and the number of carrier frequency cycles per subcode cpp can
be obtained using the relation cpp = fc /B.
18.2.2 Testing the Algorithm

The parameter extraction algorithm is tested with 6 LPI polyphase signals
as shown in Figure 18.7. The parameters used to generate the polyphase
LPI signal modulations are: fs = 7,000 Hz for the noncooperative intercept
receiver ADC sampling frequency, fc = 1,495 Hz (signals 1 to 3), 2,495 Hz
(signals 4 to 6) for the carrier frequency, Nc = 9, 16, 25, 36 for the number of
subcodes, number of cycles of carrier frequency per subcode of cpp = 1, 2, 3, 4,
5, 6 and SNRs of 0 dB and 6 dB. Figure 18.7 also shows the corresponding
code periods T and modulation bandwidths B that range from 299 Hz to
1,495 Hz.
Recall that if a is a measurement value of a quantity whose exact value

is a, then the absolute value of the relative error r is defined by
e e
e a a e ee error ee
r =e
e e=e e (18.8)
a e truevalue
The relative error is plotted in Figure 18.8 for the carrier frequency fc , code
period T , bandwidth B and number of subcodes Nc for each signal number
under test. The carrier frequency error is very small for 0 dB but for 6 dB
higher errors occur for small values of Nc . If the frequency resolution of the
PWVD is increased (integration of more samples from the ADC), the error
in estimating fc is expected to decrease and can be easily investigated with
the software contained with the textbook. The error in the estimation of Nc
is related to algorithm results for T and B since Nc = T B. The overall errors
are reasonably small for 0 dB. For SNR = 6 dB the largest errors occur
for Nc = 9, 16. That is, the simulation shows the important result that for
smaller values of SNR, the error in the extracted parameters are smaller for
larger values of Nc . That is, due to the larger processing gain obtained by the
intercept receiver (larger numbers of subcodes), a better estimation of the B
(18.6) and T (18.7) can be obtained. Note that another important advantage
to this approach is that the extraction algorithm is not aected by the cross
terms present within the PWVD images. The reason is that integration of the
cross term projections is very small compared to the modulation projections
obtained.
18.3 Polyphase Parameters from Quadrature

Mirror Filtering
In this section the polyphase modulation parameters are extracted using the
middle quadrature mirror filter bank (QMFB) time-frequency layer. The
parameters to be extracted are carrier frequency (fc ), the code period (T ),
number of subcodes within a code period (Nc ), the signal bandwidth (B) and
the number of carrier cycles within a subcode (cpp). Polyphase modulations
include the Frank, P1, P2, P3, and P4 codes.
18.3.1 Wavelet Decomposition Algorithm Description

This algorithm depends on the quadrature mirror filter bank (QMFB) tech-
nique to examine and analyze polyphase signals. The QMFB generates time-
frequency layers and the number of layers depends on the signal length. The
initial layers oer high time resolution with the final layers oering higher fre-
quency resolution. The middle layer, however, provides the best compromise
of both time and frequency, which makes it suitable to extract the parameters
Figure 18.8: Relative error results for polyphase parameter extraction using
Wigner-Ville distribution.
of the signals. The accuracy of extracted parameters depends on frequency

as well as time resolution of the layer. A flow diagram of the algorithm is
shown in Figure 18.9. The algorithm works as follows:
a. First qmfb gui.m is executed to input the name of signal (without file
extension), directory of signal, sampling frequency (Hz), number of layer
and 0 for first time or 1 for another time computation of the layer.
b. QMFB is executed by clicking on the Run tab on the graphic user
interface (GUI) by calling startpoint.m which reads inputs from the GUI
and calls the qmfb.m function for formatting and filtering of the signal.
This function also computes the total number of time-frequency layers
(N ). As mentioned above the center layer provides good compromise
of time and frequency resolutions.
c. For extraction of the parameters the time-frequency layer is selected
(N/(21) for N even and (N 1)/2 for N odd) and user is prompted on
MATLAB command window to input this layer number in the choose
layer block of GUI and 1 in examining another layer block. After
these inputs the Run tab is executed on the GUI.
d. The data of selected time-frequency layer is saved as
QMFB signalfile.mat for further input to the algorithm.
e. The main algorithm file Ext Para.m is called within startpoint.m
to extract the requisite parameters of the poly phase LPI signal.
f. The carrier frequency (fc ) is calculated by finding the maximum inten-
sity point of the time-frequency matrix of the selected layer.
g. The 3-dB bandwidth is computed by picking the signal intensity points
greater than 0.5 in frequency dimension.
h. To calculate time period (T ) the time slice is taken on carrier frequency
and MATLAB command movavg is used to reduce the noise eects
along time axis.
i. The number of sub codes and number of cycles per sub code are calcu-
lated with already computed parameters (carrier frequency, bandwidth
and time period).
Figure 18.9: Block diagram of QMFB parameter extraction algorithm.


The information contained in the time-frequency QMFB layers can be used
to extract the parameters with reasonable accuracy using the algorithm de-
scribed in this section. The middle time-frequency layer that is computed
provides a good compromise of time and frequency resolutions. for further
analysis of LPI signals. For extracting the carrier frequency and bandwidth of
the signal, the signal processing could be restructured to examine the higher
layers for more accurate results. Similarly, for time period measurement, the
initial layers can yield better results. These accurate results will be at the
cost of more computations and processing time. The algorithm should be
tested with a number of polyphase signals including Frank, P1, P2, P3, and
P4 polyphase codes. We leave this an exercise for the student.
18.4 FMCW Parameters from Cyclostationary

Bifrequency Plane
This section demonstrates an algorithm to autonomously extract the mod-
ulation parameters of a triangular FMCW signal using the spectral corre-
lation density function. The signal processing uses the DFSM bifrequency
plane where the presence of the FMCW modulation has been identified. The
parameter extraction process determines the modulation period tm , modu-
lation bandwidth F and the carrier frequency fc . Extraction for low SNR
bifrequency images gives reasonable results due to the denoising capability of
the spectral correlation processing since noise is not correlated.
The DFSM algorithm first computes the spectral components of the signal
and then executes the spectral correlation operations directly on the spectral
components. One important consideration in obtaining accurate results with
the FMCW extraction algorithm is to insure that the DFSM frequency res-
olution, cycle frequency resolution selected, is small enough to measure the
parameters accurately (for example, the code rate Rc ).
As an example, an FMCW waveform with F = 500 Hz, tm = 30 ms,
fc = 1 kHz and SNR = 0 dB is processed through the DFSM algorithm.
The frequency resolution was set to 16 Hz and the value of the Grenanders
uncertainty condition is selected as M = 2 (N = 1,024). Due to the quadrant
symmetry of the bifrequency plane, the FMCW modulation shows up at four
distinct locations as shown in Figure 18.10. Each of these four positions
contains a geometrical shape representative of the modulation from which all
the parameters can be extracted.
Two of the modulation parameters of interest are shown on the one se-
lected quadrant of the bifrequency plane shown in Figure 18.11. Both the
modulation bandwidth and carrier frequency are easily recognized from this
result. To extract the modulation period we closely examine the details re-
Figure 18.10: Direct frequency-smoothing method for cyclostationary extrac-

tion of FMCW parameters with F = 500 Hz, tm = 30 ms and fc = 1 kHz
using a frequency resolution of 16 Hz and M = 2.
vealed in any one of the four quadrants. What we can measure easily is the
modulation code rate Rc as shown in Figure 18.12. The modulation period
for a triangular FMCW waveform is related to the code rate as
1
tm = (18.9)
2Rc
where the factor of 2 in the denominator accounts for the triangular waveform
extending for 2tm .
18.4.1 Cyclostationary Algorithm Description

A block diagram of the FMCW extraction processing algorithm that uses
the DFSM bifrequency plane results is shown in Figure 18.13. The first
step in the algorithm is to crop one of the quadrants within the area of
support matrix. A dierent angle of the bifrequency plane is shown in
Figure 18.14 and is a contour plot of the bifrequency matrix S from the
DFSM processing. The figure shows the area of support that is cropped for
parameter extraction processing (left upper corner of the figure). After crop-
ping the area of support, an adaptive threshold operation is performed to
reduce the amount of noise present in the image. This is followed by creating
windows for detection of local maximums. The local maximums are used to
Figure 18.11: Bifrequency plane showing the measurement of FMCW modula-

tion parameters F = 500 Hz (frequency), and fc = 1 kHz (cycle frequency).
Figure 18.12: Bifrequency measurement of FMCW code rate on the cycle

frequency axis showing Rc = 16.6 Hz (=1/2tm ).
Figure 18.13: Block diagram of FMCW extraction processing using DFSM

bifrequency plane.
calculate the code rate Rc and subsequently the modulation period tm by

finding the minimum cycle frequency dierence for the local maximums.
Determining the boundaries of the modulation allows the carrier frequency
and modulation bandwidth information to be calculated. The boundaries of
the signal are defined in Figure 18.14. The modulation bandwidth of the
signal is
F = f2 f1 (18.10)
and the carrier frequency is
1 2
fc = (18.11)
2
where the goal is to find the correct values of 1 , 2 , f1 and f2 . To find
the appropriate index values corresponding to the correct values for this
computation, the image is scanned in two dierent directions as shown in
Figure 18.15. An approximation to the pdf is constructed by scanning both
the i- and j-axis and finding the magnitude corresponding to each index. The
horizontal scan gives the i-index values as
N
3
xi = S(i, j) (18.12)
j=1
and xi=(1:N )
pdfi = (18.13)
xj=(1:N )
and the vertical scan similarly the j-index values. This energy distribution
enables the location of the signal on the bifrequency plane to be determined.
The algorithm scans from the lower index values to the higher index values to
look for the signal power above the threshold of 6 dB. The first index with
a level above 6 dB gives i1 and j1 . Starting from the higher index values
and going toward the lower index values gives i2 and j2 resulting in
F = fi2 ,j2 fi1 ,j1 (18.14)
Figure 18.14: Contour plot of DFSM matrix S showing the region of support
being cropped for parameter extraction.
and
i1 ,j2 i2 ,j1
fc = (18.15)
2
The same algorithm is used to extract the parameters from a noisy signal.
The scan method described above however, will smear the signals pdf so the
procedure for determining the index values must be modified to avoid possible
errors (due to the high energy levels in the skirts of the pdf). An adaptive
noise filter is also used here.

To evaluate the algorithms accuracy, the program is run for 12 dierent
signals, which are listed below. The algorithm is also run for a set of higher
frequency signals. The program checks the signal for the existence of the
noise and uses the appropriate method to extract the parameters. The al-
gorithm also checks the extracted parameters and does a closest match with
the possible threat signals in a database. This gives the exact parameters of
Figure 18.15: Scan geometry for DFSM matrix S to determine the correct
index values i and j.
the signal. The signals are analyzed using cyclo gui.m. The signals with no
noise and with 0 dB noise are analyzed with 16 Hz of frequency resolution but
the signals with 6 dB noise are analyzed with 32 Hz of frequency resolution.
The high resolution selected for the 6 dB signals is due to the computational
diculties when using a personal computer. Hence the results for 6 dB sig-
nals do not have a modulation period solution because the solutions are not
reliable for the 32 Hz of resolution that is used. As discussed in Chapter 15,
the resolution of the cycle frequency and frequency should be smaller than
the largest parameter being measured or extracted.
F 1 7 250 20 s.mat (fc =1 kHz, fs =7 kHz, F = 250 Hz, tm = 20 ms, signal only)
F 1 7 250 20 0.mat (fc =1 kHz, fs =7 kHz, F = 250 Hz, tm = 20 ms, SNR = 0 dB)
F 1 7 250 20 -6.mat (fc =1 kHz, fs =7 kHz, F = 250 Hz, tm = 20 ms, SNR = 6 dB)

In this chapter, we have shown that we can extract the parameters for the
signal only measurements very accurately as shown in Figure 18.16. The
parameters extracted from the noisy signals are also fairly accurate, but the
algorithm for the noisy signals can be improved to get better performance.
The threshold set for the noisy measurements were set to a fixed intuitive
level, which is somewhat subjective. One may think of adding an adaptive
threshold that changes according to the noise level in the signal. Here we took
advantage of the cyclostationary analysis to get rid of the noise. Note that
the cyclostationary analysis has an inherited noise reducing process within its
spectral correlation algorithm. Once the signal is classified correctly, the exact
parameters of the intercepted signal can be determined.
References
[1] Hejazi, M.R., Shevlyakov, G., and Ho, Y-S., Modified discrete radon
transforms and their application to rotation-invariant image analysis,
IEEE 8th Workshop on Multimedia Signal Processing, pp. 429434,
Oct. 2006.
[2] Minsheng, W., Chan, A.K., and Chui, C.K., Linear frequency modu-
lated signal detection using Radon-ambiguity transform, IEEE Trans.
on Signal Processing, Vol. 26, No. 3, pp. 571586, March 1998.
[3] Gulum, T. O., Pace, P. E. and Cristi, R. Extraction of Polyphase
Radar Modulation Parameters Using a Wigner-Ville DistributionRadon
Transform, IEEE International Conf. on Acoustics, Speech and Signal
Processing, Las Vegas, NV, March 2008.
[4] Gulum, T. O., Autonomous Non-linear Classification of LPI Radar
Signal Modulations, Naval Postgraduate School Masters Thesis, Sept.
2007.
Problems
1. (QMFB) The files in the Part II folder Extract\POLY FROM QMFB
perform the quadrature mirror filtering but have been modified to imple-
ment the extraction routines in Section 18.3. (a) Generate a Frank code
signal with fc = 1 kHz, fs = 7 kHz, M = 16 and cpp = 2 for SNRs be-
tween 10 dB and 10 dB (in steps of 1 dB). (b) Run poly from qmfb.m
to extract the parameters for the Frank code. Be sure to follow the in-
structions on the command line that are displayed. (c) Plot the relative
error as a function of the SNR for each of the extracted parameters. (d)
Repeat (a)(c) for the P4 code and compare your results.
Figure 18.16: Extraction results using the cyclostationary signal processing

bifrequency plane.
2. (QMFB) Edit the m-files contained in the Part II folder

Extract\POLY FROM QMFB folder to analyze a dierent layer (other
than the middle layer). Repeat Problem 1 and compare your results.
3. (QMFB) Write a MATLAB procedure similar to Ext Para.m to extract
the parameters from an FMCW signal. Use the FMCW files in the
Test Signals folder to evaluate your results.
4. (CYCLO) The files in the Part II folder Extract\FMCW FROM CYCLO
perform the direct frequency smoothing spectral correlation technique
to derive the bifrequency domain results. The file cyclo gui.m calls
DFSM.m which has been modified to include the parameter extraction
algorithm described in Section 18.4. (a) Generate an FMCW signal
with fc = 1 kHz, fs = 7 kHz, F = 500 Hz and tm = 30 ms for SNRs
between 10 dB and 10 dB (in steps of 1 dB). (b) Run cyclo gui.m to
extract the parameters of the FMCW waveforms. (c) Plot the relative
error as a function of the SNR for each of the extracted parameters. (d)
Repeat (a)(c) for an FMCW signal with twice the modulation band-
width and compare your results.
5. (CYCLO) Copy the FAM.m file contained in the Part II folder CYCLO
to the Extract\FMCW FROM CYCLO folder. Edit the file to include
the parameter extraction algorithm (see Problem 4). Repeat Problem
4 (a)(c) and compare your results.
6. (PWVD) The files in the Part II folder Extract\POLY FROM WVD
perform the extraction of the polyphase parameters using the PWVD
Radon transform described in Section 18.2.1. The main file
poly from WVD.m uses a routine that cycles through the subfolder
polyfiles to extract the parameters for all of the files contained in
the folder. (a) Copy one example of each polyphase modulation for
signal only from the TestSignals folder (Part I) to the polyfiles folder.
Edit the main file and change the diary file to represent your case un-
der study. Run the parameter extraction algorithm and then compute
the relative error for each result. (b) Repeat (a) for a SNR = 0 dB.
(c) Repeat (a) for a SNR = 6 dB. (d) How does the noise aect the
relative error for the extraction results?
Appendix A
Low Probability of
Intercept Toolbox
A.1 Introduction to the LPIT

The low probability of intercept toolbox is a collection of MATLAB files that
give the user the quick capability to generate a number of LPI complex signals
easily. The user can change the parameters of the signal, plot out the signals
time domain and power spectral density characteristics, and save the time
domain signal to the current directory for further analysis, using the time-
frequency and bifrequency classification programs discussed in Part II.
The LPIT software is contained on the CD provided with this book, in
the folder titled LPIT. To install the toolbox, simply copy this folder to your
computer. When the LPI signals are generated, they are saved to the same
folder where the program files reside. After folder has been copied to your
computer, the following steps should then be followed to generate the signals:
1. Start MATLAB and change the current directory to the LPIT folder.
2. Type lpit on the command line to start the toolbox.
3. Choose one of the LPI signals on the menu and enter the parameters
requested.
4. Choose whether (or not) to plot the signal.
5. Choose whether (or not) to save the signals to the current directory.
Note that if the user chooses to save the signal, two .mat files are created:
one for the signal only, and one for the signal-plus-noise for SN R = 0 dB.
709
SNRs other than 0 dB can also be generated when supplying the requested
parameters. For each file, the complex signal is saved in a .mat file, with
the I and Q variables in two separate column vectors. The number of code
periods generated by the LPIT is five (four for FMCW) although that can
easily be changed.
The 14 signal types available from the LPIT, and the corresponding chap-
ter where they are discussed are shown below:1
1. FMCW (Chapter 4);

2. BPSK (Chapter 5);
3. Polyphase Barker codes (Chapter 5);
4. Polyphase Frank code (Chapter 5);
5. Polyphase code P1 (Chapter 5);
9. Polytime signals (Chapter 5).
10. Costas frequency hopping code (Chapter 6);
11. FSK/PSK (Costas) (Chapter 6);
12. FSK/PSK (Target) (Chapter 6);
13. Noise waveforms (Chapter 7);
14. Test signals (Chapters 9, 10, and 11);
A.2 Naming Convention and Example

For all signals, the name is automatically assigned to the .mat file, and reflects
the signal parameters that were supplied by the user. The naming conven-
tion is the same for each signal, but varies slighty due to dierent parameters
required for dierent types of signals. The first character in the file name
1 Taboada, F. L., Detection and classification of low probability of intercept radar
signals using parallel filter arrays and higher order statistics, Naval Postgraduate School,
Masters Thesis, Sept. 2002.
Appendix A: Low Probability of Intercept Toolbox 711
Figure A.1: LPIT file naming convention.
always indicates the type of signal. The second character indicates the car-
rier frequency (in kHz). The third character indicates the selected sampling
frequency (in kHz). The remaining characters are dierent, depending on the
type of signal generated. The file naming convention is summarized in the
tree diagram shown in Figure A.1
For example, for an FMCW, the fourth character is the modulation band-
width (in Hz), and the fifth character is the modulation period (in ms). The
sixth character is either an s indicating signal only, or a number indicating
the SNR for a noisy signal. Consider the signals F 1 7 250 20 s.mat (signal
only) and F 1 7 250 20 0.mat (SNR=0 dB). Here the F indicates FMCW.
The 1 indicates an fc = 1 kHz carrier frequency, the 7 indicates fs = 7 kHz
sampling frequency, the 250 indicates the modulation bandwidth F = 250
Hz, and the 20 represents the modulation period tm = 20 103 s. The s
in the first file name indicates that the .mat file contains only the signal. The
0 in the second file name indicates the .mat file with SN R = 0 dB. For
example, if the signal is generated with a 6 dB SNR, then this value would
be 6. When the signals are saved, the names are always displayed, so the
parameters chosen can be recognized.
Appendix B
Generating PAF Plots

Using the LPIT Files
For CW signals, the Web site by Levanon (www.eng.tau.ac.il/nadav/) in-
cludes the files to calculate and display the autocorrelation function, the pe-
riodic autocorrelation function, and the periodic ambiguity function.1 These
files can be downloaded and used easily with the time domain signals gener-
ated by the LPIT.
To calculate and display the ACF, PACF, and PAF, download the files
into the LPIT folder. After generating the .mat files containing the LPI signal
(see Appendix A), follow the steps below:
1. Run ambfn7.m (Levanons PAF code from the Web site).

2. Select User Defined in the first block at the top.
3. Deselect the Frequency Radio button that follows.
4. Next go to the command line and run the paf preprocess.m file distrib-
uted with the LPIT. This file will ask the user to supply:
Name of the signal file (.mat file that resides in the LPIT direc-
tory);
Sampling frequency fs (in Hz);
Carrier frequency fc (in Hz);
1 The MATLAB code to calculate the ambiguity function is described in Mozeson, E.,
and Levanon, N., MATLAB code for plotting ambiguity functions, IEEE Trans. on
Aerospace and Electronic Systems, Vol. 38, No. 3, 10641068, 2002.
713
Number of code periods to include in PAF calculation N (1) for

PSK and (2) for FSK.
5. Return to the ambfn7.m graphical user interface (GUI) and, next to
the sliders, enter in the five suggested values for the signal analysis.
6. On the GUI select Cal & Signal Plot first, then ACF and SPEC plot or
PAF.
Note that the number of code periods N to include in ACF, PAF calculation
is independent of the number of periods generated using the LPIT.
Appendix C
Primitive Roots and

Costas Sequences
In this appendix, a concise description of prime numbers, residues, reduced
residues, the Euler function (n), and primitive roots is given. The signifi-
cance of this appendix is to present the concepts in a framework that lends
itself to the derivation of Costas frequency-hopping sequences. These con-
cepts are useful for understanding the construction of Costas sequences using
the Welch method.1
C.1 Primes
To begin we give the definition of a prime number.
Definition 1
An integer p > 1 is called a prime number, or a prime, in case
there is no divisor d of p satisfying 1 < d < p.
For example, the numbers 2, 3, 5, and 7 are prime numbers and there is
an infinite number of primes. Although the numbers 4, 5, and 7 are not all
primes, they are all relatively prime with respect to each other, in that none
have a common factor.2
1 Note that mathematical concepts such as the division algorithm, the Euclidean algo-
rithm for finding the greatest common divisor (gcd), and the solutions of simultaneous
congruencies using the Chinese remainder theorem are not described here, but a complete
treatment is presented in Pace, P. E., Advanced Techniques for Digital Receivers, Artech
House, Inc., Norwood MA, 2000.
2 Niven, I., Zuckerman, H. S., and Montgomery, H. L., Introduction to the Theory of
Numbers, 5th Edition, John Wiley and Sons, New York, 1991.
715
C.2 Complete and Reduced Residue Systems

If a congruence involves only addition, subtraction, and multiplication, we
may replace the integers with congruent integers. To help in this, the follow-
ing definitions describing residue systems are given, followed by an example
to illustrate the concept.
Definition 2
If h and j are two integers and h j(mod m), then we say that
j is a residue of h modulo m.
Definition 3
The set of integers {r1 , r2 , . . . , rs } is called a complete residue
system modulo m if ri = rj (mod m) and for each integer n there
corresponds an ri such that n ri (mod m).
If s dierent integers r1 , r2 , . . . , rs form a complete residue system modulo m,
then s = m. If m is a positive integer, then {0, 1, . . . , m 1} is a complete
residue system modulo m. For example, for m = 7, the smallest positive
integer values within the modulus are {0, 1, 2, 3, 4, 5, 6}.
Definition 4
The set of integers {r1 , r2 , . . . , rs } is called a reduced residue sys-
tem modulo m if (a) the gcd(ri , m) = 1 for each i, (b) ri =
rj (mod m) whenever i = j, and (c) for each integer n relatively
prime to m there corresponds an ri such that n ri (mod m).
For example, the set {0, 1, 2, 3, 4, 5} is a complete residue system modulo 6,
but {1, 5} is a reduced residue system modulo 6. That is, we can obtain a
reduced residue system from a complete residue system by simply deleting
those elements of the complete residue system that are not relatively prime
to m.
Example 1 The sets {1, 2, 3}, {0, 1, 2}, {1, 0, 1}, and {1, 5, 9} are all com-
plete residue systems modulo 3.
When working with congruences modulo m, we can replace the integers in

the congruences by elements of {0, 1, 2, . . . , m 1}. This can make many
complicated problems much easier.3
3 Andrews, G. E., Number Theory, Dover Publications Inc., New York, 1971.
Appendix C: Primitive Roots and Costas Sequences 717
Example 2 Find an integer n that satisfies the congruence

325n 11(mod 3) (C.1)
Since
325 1(mod 3) (C.2)
and
11 2(mod 3) (C.3)
the problem is reduced to finding an integer n such that [15]
n 2(mod 3) (C.4)
The obvious answer here is the integer 2.
Definition 5
The function (m) denotes the number of positive integers less
than or equal to m that are relatively prime to m. This function
(m) is called the Euler function, and represents the number
of integers that form the reduced residue system modulo m.
Example 3 We know (6) = 2 and {1, 5} is a reduced residue system modulo
6. Note the set {5, 52 } is also a reduced residue system modulo 6, since
5 5(mod 6) and 25 1(mod 6).
C.3 Primitive Roots

We have examined the concept of the reduced residue system modulo p where
p is a prime number. In this section, we present an integer g such that
g, g 2 , . . . , g (p) constitutes a reduced residue system modulo p. The integer
g is called a primitive root. Primitive roots are fundamental to how Costas
frequency sequences can be formed. An algorithm for deriving these types of
sequences is presented, along with an example illustrating the method.
A few properties of reduced residue systems must now be given. The first
is that if h is the smallest positive integer such that
ah 1(mod m) (C.5)
we say that the order of a is h modulo m. If g is an integer, and the order
of g is (m) modulo m, then g is called a primitive root modulo m. Further,
if g is a primitive root modulo m, then g, g 2 , . . . , g (m) makes up a reduced
residue system modulo m (sometimes referred to as a cyclic group).
Since the goal of this development is to derive a Costas sequence, the first
step is to determine a prime number p, remembering that the number of fre-
quencies in the FH sequence will be NF = p 1. Also, since p is prime, we
have the following definition:
Definition 6
If p is prime, then there exist {{p}} = {p 1}, primitive roots
modulo p.
This number can be determined by first eliminating those elements in the
reduced residue set that are not relatively prime to p 1, and then counting
the remaining entries. Note also that the order of a in (C.5) must be a divisor
of (p) = p 1.
Example 4 Consider the case for which p = 11, a prime number. The
number of frequencies in the Costas array will be N = p 1 = 10. The
complete residue set is
{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10}
There are p 1 = 10 elements in the reduced residue set modulus 11
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10}
and there are {p 1} = {10} = 4 primitive roots. The questions we have
now are what are the primitive roots and what are the corresponding Costas
sequences?
To determine the primitive roots by which we can derive the Costas se-
quences, we start with g = 1 (the first element in the reduced residue set).
For g = 1 we have
{11 , 12 , 13 , . . . , 1{11} } = {1}
and we say the order of 1 is 1. Consequently, 1 is not a primitive root. For 2,
{21 , 22 , 23 , . . . , 2{11} }
or
{2, 4, 8, 5, 10, 9, 7, 3, 6, 1}
and the order of 2 is 10, indicating that g = 2 is a primitive root. For 3 we
have,
{3, 9, 5, 4, 1}
and the order of 3 is 5. That is, 3 is not a primitive root. Continuing on for
the rest of the integers within the reduced residue system for 4 we have,
{4, 5, 9, 3, 1}
and the order of 4 is 5 (not a primitive root). For 5 we have,
{5, 3, 4, 9, 1}
Appendix C: Primitive Roots and Costas Sequences 719
and the order of 5 is 5 (not a primitive root). For 6 we have,

{6, 3, 7, 9, 10, 5, 8, 4, 2, 1}
so the order of 6 is 10, and consequently, 6 is a primitive root. For 7 we have,
{7, 5, 2, 3, 10, 4, 6, 9, 8, 1}
so the order of 7 is 10, and 7 is also a primitive root. For 8 we have,
{8, 9, 6, 4, 10, 3, 2, 5, 7, 1}
so the order of 8 is 10, and therefore 8 is a primitive root. Since we now have
our four primitive roots, we know that there are no more. For completeness
however, we verify that for 9 we have,
{9, 4, 3, 5, 1}
and the order of 9 is 5 (not a primitive root), and for 10 we have,
{10, 1}
therefore the order of 10 is 2, and 10 is not a primitive root. In summary, the
four primitive roots are {2, 6, 7, 8}.
One question still remains. How do we know, short of multiplying out, if a
in (C.5) is a primitive root? That is, is there a way to find the primitive roots
and Costas sequences without having to calculate the order of each integer
value within the reduced residue set modulo p? Fortunately, the answer is
yes, and we can use the following definition
Definition 7
If a is a primitive root modulus p then ar is a primitive root
modulus p if and only if gcd(r, (p))=1 (r is relatively prime to
(p) = p 1).
In this case we can reduce our work by starting with the smallest value (e.g.,
a = 2), and first determining if this is a primitive root. Once the smallest
primitive root is found, the others may be found easily by using the definition
above.
In our example, since we have confirmed that a = 2 is a primitive root, we
know from the above definition that 2r is a primitive root modulus 11, if and
only if gcd(r,10)=1. So for r = 1, 21 = 2, for r = 3, 23 = 8, for r = 7, 27 = 7,
and for r = 9, 29 = 6. In summary, the primitive roots are {2, 6, 7, 8} and the
corresponding sequences are Costas arrays. Also note that the Costas arrays
for a = 2 and a = 6 are reverse ordered (except for the 1 on the end). This is
also true for a = 7 and a = 8. This symmetry can be used to further simplify
the sequence calculations.
Appendix D
LPIsimNet
LPIsimNet is a collection of MATLAB files that let the user easily evaluate
the information network metrics and the SNR advantages of general netted
LPI radar topologies that were discussed in Chapter 10 including the presence
of an electronic attack or jammer. The objective of this tutorial is to have the
student work several examples to become familiar with the program set. The
program set is organized as shown in Figure D.1. The ScenarioEditor.m file
lets the student open a Graphical User Interface (GUI) and create a new Sce-
nario File or modify an existing one. A Result File is generated after the
student confirms the Scenario File and executes the simulation calculation
with the assistance of Calculator.m. The SimulationViewer.m is used to
review the Result File by examining the results grid. The Painter.m file
supports the drawing of the two GUI figures.
D.1 Overview of LPIsimNet Architecture

The LPIsimNet architecture can be used to set up a sensor network with
any configuration and number of communication nodes. Evaluation of the
information exchange capability and the operational tempo is presented to
the user using the sensor network metrics. LPI radar detection performance
and the SNR values of a network enabled configuration of emitters across an
operational landscape containing targets is also presented to the user includ-
ing the ability to have the sensor nodes and jammer nodes move in time with
any velocity.
721
Figure D.1: Program flow diagram.
D.1.1 Loading the Default Sensor Network

Start MATLAB and change the Current Directory to the folder where the
LPIsimNet program resides. Run ScenarioEditor.m to open up the GUI
battlespace grid. You should see the grid as shown in Figure D.2. The
right side of the GUI is designed for displaying a schematic of the network
topology. Click the Refresh Figure to load the default network topology.
The default network consists of NT = 3 nodes: NR1 (node-1); NR2 (node-
2); and NR3 (node-3) which are capable both in information processing and
synchronized, coherent target detection. Note the links between these nodes
are bidirectional. Click on the Legend and see the legend as shown in
Figure D.3. The legend describes the symbols on the grid and can be brought
up at any time.
D.1.2 Building a Scenario File and Running the

Simulation
Go back to the ScenarioEditor. In the top left corner is the Top Level
Properties Panel containing several generic simulation properties that must
be set including the number of nodes, total time index(s) for including plat-
form and target movement, decision tempo, deployment tempo, and the fight-
ing tempo. The boundary of the X axis, and the boundary of the Y axis must
also be set to model the battlespace landscape. For this tutorial, modify these
properties according to the values shown in Table D.1.
Appendix D: LPIsimNet 723
Figure D.2: ScenarioEditor battlespace grid.
Figure D.3: Symbol legend for LPIsimNet.

Table D.1: Top Level Properties For Scenario
Properties Values Description

Number of nodes 3 Total number of nodes in network
Total time indexes 1 Number of time steps in simulation
Decision tempo 200 C2 decision tempo
Deployment tempo 400 Tempo of deployment in OODA
Fighting tempo 300 Tempo of fighting in OODA
Boundary of X axis (km) 0, 100 Left, right battlespace boundary
Boundary of Y axis (km) 0, 100 Upper, lower battlespace boundary
The panel below the Top Level Properties Panel is the Node Properties
Panel. The node properties panel contains the following scenario
information:
Current node index;
Type;
Name;
Initial position;
Velocity;
Availability of links to each node;
Capability value K of information or jammer;
Information rate ;
Minimum information rate, min
;
ERP of radar or jammer;

Eective antenna area Ae ;
Noise power.
The node properties for the Blue Force E2-C in the default simulation are
shown in Figure D.4.
Figure D.4: Node properties.

Table D.2: Link Condition of 001

Node 1 Node 2 Node 3
Availability of link N N Y
Table D.3: Parameters for Node 2 and Node 3

Properties Node2 Node3
Type Blue Force Blue Force
Name F-16 AC-130
Initial position 20, 70 60, 70
Velocity 0, 0 0, 0
Links 101 110
Cap. of info. or jamming 0.75 0.25
Info. rate 200 300
Min. info. rate 100 50
ERP (W) 0 0
E. antenna area 0 0
Noise power 0 0
D.2 Setting the Node Properties

At the bottom of the Node Properties Panel, try switching between the
properties of the dierent nodes. Note the Node Index that indicates the
current node. Set the properties of node 1 to the following values shown in
Figure D.4. The Availability of Links to Each Node represents the link
condition to each node. For example, for node 1, 001 represents the link
configuration as shown in Table D.2. After setting node 1 availability, set
node 2 and node 3 to the values shown in Table D.3.
After setting all node properties needed, click Refresh Figure to see the
layout and the overall connection of this scenario. The topology should look
like that shown in Figure D.5.
Click Save Scenario and save the scenario file as Sce-3C.mat. Config-
ure the MATLAB command line analysis to be visible along with the Sce-
narioEditor. Click Run Simulation to activate the calculation of the sim-
ulation results file. The MATLAB command line shows the tracking message
of the four phases in the calculation. Wait until a Save As dialog appears
and save the simulation results as Sim-3C.mat. Now, we have successfully
finished creating a scenario file (Sce-3C.mat) and generated the simulation
results file (Sim-3C.mat).
Figure D.5: Simulation topology.

D.3 Viewing the Simulation Results

Go to the MATLAB environment and launch the SimulationViewer.m. The
SimulationViewer GUI grid appears as shown in Figure D.6. Click Load to
load the simulation result file, Sim-T1(3C).m that was just generated. After
loading, the simulation result file is displayed as shown in Figure D.7. The
values for the simulation properties are now shown in the top left Information
Network Analysis Panel. This panel consists of:
Number of links suppressed;
Reference Connectivity Measure;
Connectivity Measure;
Network Reach;
Network Richness;
Decision Tempo;
Deployment Tempo;
Fighting Tempo;
Characteristic Tempo;
Max Operational Tempo.
Observe the simulation results in the Information Network Analysis Panel.
Click the Detail after the Reference Connectivity Measure. The detailed
analysis data is shown in the MATLAB command line as shown in Figure
D.8. Click the Detail after the Connectivity Measure. The detailed analy-
sis data is shown in MATLAB command line as shown in Figure D.9. Click
the Detail after the Reference Network Richness. The detailed analysis
data is shown in the MATLAB command line as shown in Figure D.10.
Figure D.6: SimulationViewer GUI.
Figure D.7: GUI.

R
Figure D.8: Reference connectivity measure CM for number of sensor network
nodes NT = 3 to 20.
Figure D.9: Detailed results for connectivity measure.
Figure D.10: Detailed results for network richness.

Table D.4: LPI Radar Network: Scenario Setup for Three Nodes
Properties Node1 Node2 Node3 Node4

Type Blue Force Blue Force Blue Hostile
Force Jammer
Name E-2C F-16 AC-130 Su-34
Initial position 40,40 20,70 60,70 80,40
Velocity 0,0 0,0 0,0 -10,0
Link configuration 0010 1010 1100 1000
Cap. of info. or jamming 1.0 0.75 0.25 0.3
Info. rate 200 200 300 0
Min. info. rate 100 100 50 0
ERP (W) 0 0 0 0
E. antenna area (m2 ) 0 0 0 0
Noise power (W) 0 0 0 0
D.4 Adding a Moving Jammer to the Scenario

Go back to the ScenarioEditor.m (if you have closed it, re-launch it) and
load the scenario file Sim-3C.mat. In the Top Level Properties Panel,
change the number of nodes to 4, and the total time index(s) to 3. That
is, the scenario can evolve over time and the platforms within the scenario
can have movement. Each platform can also have a dierent velocity (by
including larger movements over a time index) in any general direction. Click
Refresh Figure and see a fourth node, (NR4 ) node 4, that was added into
the network. Go to the Node Properties Panel and set the properties as in
Table D.4. Note the jammer is identified as being onboard an Su-34 Flanker
fighter-bomber 2-seat strike aircraft. After refreshing, the figure should look
like Figure D.11.
Save this scenario as Sce-3C+J.mat and run the simulation calculation
and save the result file as Sim-3C+J.mat. Go to SimulationViewer and
load Sim-3C+J.mat. Your figure should look like Figure D.11. Note the
two links to E-2C survive even with the Su-34 jammer. All the simulation re-
sults in Information Network Analysis are identical to those of the previous
simulation. Click the double right arrows in the lower right section to switch
the time index to 2. Note that the Su-34 jammer moves closer to E-2C and
the link from AC-130 to E-2C is now not available (the arrow is missing). At
this index the sensor information changes. The number of links suppressed in
now one as shown in the Information Network Analysis panel. Click De-
tail and review the detailed data in the MATLAB command line as shown
in Figure D.12.
Figure D.11: Sensor network with jammer added.
Figure D.12: Command line analysis of sensor network with link suppression.
Table D.5: LPI Radar Network: Scenario Setup for the Three Nodes
Properties Node1 Node2 Node3 Node4

Type Blue Force Blue Force Hostile Radar
Force Force Jammer Target
Name Radar1 Radar2 Su-34 Target
Initial position 15,40 15,15 30,25 15,25
Velocity 0,0 0,0 0,0 0,0
Link configuration 0000 0000 1100 0000
Cap. of info. or jamming 0 0 0 0
Info. rate 0 0 0 0
Min. info. rate 0 0 0 0
ERP (W) 1000 100 10 0
E. antenna area (m2 ) 0.0815 0.0815 0 0
Noise power (W) 7.5 1013 1.0 1012 0 0
Now click the double right arrows to increment the time index to 3. Note
that now two links are not available due to the new closer position of the
jammer. The Trend buttons provide the ability to review the trend of the
results as a function of time.
D.5 Netted Radar with a Jammer

To examine how a jammer influences a netted radar configuration, go to the
ScenarioEditor and change the Number of nodes to 4, Number of Time
Indexes to 1. Set the node properties according to Table D.5. Click Refresh
Figure to see Figure D.13.
Save this scenario as Sce-2R+J+T.mat and run the simulation
calculation. After the simulation completes save the result file as
Sim-2R+J+T.mat. After saving the file go to SimulationViewer and load
Sim-2R+J+T.mat. The figure should look like Figure D.14.
At the bottom left corner is the Netted Radar Analysis Panel. Figure
D.15 describes several options that are applied to control the contour chart
display. Leave the Enable Nework Synchronization box unchecked and
select the SNR(dB) radial button. Click Refresh to see the SNR contour
chart. It should look like Figure D.16. Note that this may take a few seconds.
Click Detail for the SNR and the detailed analysis data is displayed in the
MATLAB command line as shown in Figure D.17.
For a network-enabled configuration, check the Enable Network Synchro-
nization in the Netted Radar Analysis Panel and click Refresh again.
The SNR contour chart with the network synchronization should appear as
Figure D.13: Topology of simulation: two emitters, one target, one jammer.
Figure D.14: Radar network properties with jammer added.

Figure D.15: Description of netted radar analysis panel.
shown as Figure D.18. Click Detail for the SNR, and the detailed analysis
data is displayed in the MATLAB command line as shown in Figure D.19.
To examine the eects of the jammer, uncheck Enable Network Simula-
tion and select the S/N+J Ratio (dB) radial button. Then click Re-
fresh. The eect of hostile jamming on the netted radar systems and sensor
network can be examined by reviewing the SNJR contour chart as shown in
Figure D.20. In this figure, the network connecting the radar sensors is dis-
abled. Click the Detail of S/N+J and view the detailed analysis as displayed
in Figure D.21. The contour chart and the detailed analysis show that with-
out the network, the S/N+J = 70 dB at the target. If the sensor network is
enabled however, the S/N+J increases as shown in the contours displayed in
Figure D.22. The command line analysis shows the S/N+J = 64 dB when
the network is enabled.
Figure D.16: SNR contour chart without network synchronization.
Figure D.17: Command line analysis of SNRno network.

Figure D.18: SNR contour chart with network synchronization.
Figure D.19: Command line analysis of SNRwith network.

Figure D.20: S/N+J contour chart: sensor network with jammer and without
network synchronization.
Figure D.21: Command line analysis of netted radar systems with jammer
and without network synchronization (S/N+J = 70 dB at target).
Figure D.22: S/N+J contour chart: netted radar systems with jammer and
with network synchronization.
Figure D.23: Command line analysis of sensor network with jammer and
without network synchronization (S/N+J = 64 dB at target).
Appendix E
PWVD for FMCW with

F = 500 Hz
In Figures E.1 and E.2, a signal with fc = 1,000 Hz, tm = 20 ms and
F = 500 Hz is examined using the PWVD. In Figure E.1(a) the increase
in bandwidth is noticeable and in the time-frequency distribution in Figure
E.1(b), the important parameters can all be extracted. In Figure E.2(a), the
SN R = 0 dB, and the signal is still visible and the parameters can be ex-
tracted. In Figure E.2(b) the increase in noise is apparent (SN R = 6 dB),
but the parameters can still be extracted.
741
Figure E.1: PWVD for FMCW with F = 500 Hz, tm = 30 ms (signal only),
showing the (a) marginal frequency domain, and (b) time-frequency plot.
Appendix E: PWVD for FMCW with F = 500 Hz 743
Figure E.2: PWVD for FMCW with F = 500 Hz, tm = 30 ms, time-
frequency plot for (a) SN R = 0 dB, and (b) SN R = 6 dB.
Appendix F
PWVD for Frank Code

with T = 64 ms
In the second example of a Frank signal using the PWVD, the signal has a
carrier frequency of fc = 1,000 Hz, 64 phase codes (M = 8), and a cpp = 1.
The signal has a code period of T = N 2 tb = 64 ms. Figure F.1(a) shows the
PWVD frequency domain plot. As expected, due to the longer code length
(Nc = M 2 = 64), the modulation spikes that were clearly visible in Figure
9.16(a) (Nc = M 2 = 16) are now hard to distinguish. Figure F.1(b) shows
the PWVD time-frequency plot, and indicates the bandwidth measurement
and code period measurement. Figure F.2(a) shows the time-frequency plot
with SN R = 0 dB. All parameters can still be extracted. In Figure F.2(b),
however (SN R = 6 dB), identification of the major crossterm and the
measurements of the signal parameters become more dicult.
745
Figure F.1: PWVD for Frank code with B = 1,000 Hz, T = 64 ms (signal
only), with (a) frequency domain, and (b) time-frequency contour.
Appendix F: PWVD for Frank Code with T = 64 ms 747
Figure F.2: PWVD for Frank code with B = 1,000 Hz, T = 64 ms, time-
frequency contour, for (a) SN R = 0 dB, and (b) SN R = 6 dB.
Appendix G
PWVD Results for P1, P2,

P3, and P4 Codes
G.1 P1 Code Analysis

In this section, the PWVD is used to extract the parameters from a P1 phase-
coded CW signal. The intercepted signal shown in the following example has
a carrier frequency of fc = 1,000 Hz, a cpp = 1 (B = 1,000 Hz), 64 phase
codes (Nc = 64), and is sampled by the ADC at a rate of fs = 7,000 Hz.
The marginal frequency domain result is shown in Figure G.1(a). Here
the harmonics are not as evident, without zooming in on the signal in the
frequency domain. Figure G.1(b) demonstrates the P1 modulation in the
time-frequency domain, and reveals that a longer code period makes it easier
to identify the major crossterm in order to make the signal measurements.
Notice that the slopes of each line are negative. It is interesting to note that
if the crossterms were deleted, the parameter measurements might not be
easier to extract. Figure G.2(a) shows the SN R = 0 dB results. Extraction
is still possible, but in Figure G.2(b) with SN R = 6 dB, this capability
again disappears quickly.

The P2 code shows up in the PWVD in a similar manner as the P1 and
Frank code, except the slope is positive. Since the code period T = Nc2 tb
both (12.27) and (12.28) apply. The signal examined has fc = 1,000 Hz,
a cpp = 1 and 64 phase codes (Nc = 64). Figure G.3(a) shows the phase-
coded signal with a code period of 64 ms. Note the eect the additional code
749
Figure G.1: PWVD for P1 code with B = 1,000 Hz, T = 64 ms (signal only),
Appendix G: PWVD Results for P1, P2, P3, and P4 Codes 751
Figure G.2: PWVD for P1 code with B = 1,000 Hz, T = 64 ms, time-
frequency plot, for (a) SN R = 0 dB, and (b) SN R = 6 dB.
length has on the frequency domain modulations. The carrier frequency is

also easily identified. In Figure G.3(b), the measurement of the code period
and the bandwidth are illustrated. Figure G.4(a) shows the P2 signal with
an SN R = 0 dB. The signal parameters can be identified, although the
crossterms make this somewhat dicult. Figure G.4(b) shows the signal
with an SN R = 6 dB. Without further processing, the identification of the
signal parameters is not possible in this case.

In this section, a P3 signal with fc = 1,000 Hz, a cpp = 1 (B = 1,000 Hz),
and 64 four phase codes Nc = 64 is examined using the PWVD.
The characteristics of the P3 code are evident as shown in Figure G.5(a)
which shows the frequency domain. Figure G.5(b) shows the time-frequency
domain, and clearly shows the slope characteristics previously demonstrated
in the Frank and P1 code. Extraction of the signal parameters within these
figures can also be compared to Figures G.3(a) and (b).
Figure G.6(a) shows the signal with SN R = 0 dB, and Figure G.6(b)
shows the signal with SN R = 6 dB. Detection of the signal parameters
here can be compared with Figure G.4(a) and (b). Notice that we have not
discussed distinguishing between the various phase codes; this is an important
consideration that is discussed further below.

The P4 code signal is very similar to the P3 code signal in the way it shows
up in the PWVD. The P4 signal examined in this section has fc = 1,000 Hz,
a cpp = 1, and a code length of 64 (Nc = 64). Note that the equations for
the parameter measurements given for the Frank code also apply for the P4
code (as well as for the P1, P2, and P3).
Figure G.7(a) shows characteristics of the P4 code in the frequency domain
and Figure G.7(b) shows the corresponding time-frequency domain. Extrac-
tion of the signal parameters within these figures can be compared to Figure
G.5(a) and (b).
Figure G.8(a) shows the signal with SN R = 0 dB, and Figure G.8(b)
shows the signal with SN R = 6 dB. Detection of the signal parameters
here can be compared with Figure G.6(a) and (b).
showing the (a) marginal frequency domain, and (b) time-frequency domain.
Appendix H
PWVD Results for

Polytime Codes T2, T3,
and T4
H.1 T2(2) Polytime Code

The T2(2) signal examined with the PWVD has the same parameters as the
T1(2) signal investigated above, except that it has a zero beat at its carrier
frequency. Figure H.1(a) shows the PWVD frequency domain for the T2(2).
The energy is not as evenly spread out as the T1(2) previously shown, and a
strong negative component at the carrier frequency is evident. Figure H.1(b)
shows the time-frequency distribution, and shows a unique pattern of Xs
centered about the carrier (due to the zero beat at the carrier frequency). The
measurement of the bandwidth B and code period T are also shown, although
this is somewhat more dicult without any post-PWVD image processing.
To understand the bandwidth characteristics shown in Figure H.1, the
phase shift for the T2(2) is shown in Figure H.2. Here, the shortest phase
change is eight samples long, or 1.143 ms. This results in a bandwidth ex-
cursion of 875 Hz, as shown. Figure H.3(a) shows the signal for an SNR = 0
dB. The signal can still be identified as a T2(2), but it is much more dicult
to do compared to the T1(2). Errors can also occur when the parameters are
extracted. In Figure H.3(b) with a SNR = 6 dB, no signal identification
can be made, and no parameters can be extracted without any post-PWVD
image processing.
759
Figure H.1: PWVD for polytime code T2(2) with B = 875 Hz, T = 16
ms (signal only), showing the (a) marginal frequency domain, and (b) time-
frequency plot.
Appendix H: PWVD Results for Polytime Codes T2, T3, and T4 761
Figure H.2: T2(2) phase shift showing a minimum subcode width of eight
samples (1.143 ms), resulting in a bandwidth excursion of 875 Hz.
Figure H.3: PWVD for polytime code T2(2) code with B = 875 Hz, T = 16
ms, showing the time-frequency plot, for (a) SNR = 0 dB, and (b) SNR = 6
dB.

The T3(2) is examined next, and represents an approximation of a linear
FM waveform with modulation bandwidth F . As discussed in Chapter 5,
the T3(2) is generated from the quadratic linear FM phase trajectory, with
a zero beat at its beginning. Figure H.4(a) shows the PWVD frequency
domain for the signal-only case, with F = 600 Hz and modulation period
tm = T = 16 ms. Figure H.4(b) shows the time-frequency image with the
modulation bandwidth and modulation period indicated. Note the similarity
to Figure 9.18(a) and (b).
To understand the bandwidth characteristics shown in Figure H.4, the
phase shift for the T3(2) is shown in Figure H.5. With six samples making
up the shortest phase change (0.857 ms), the bandwidth excursion is 1,167
Hz (approximately 2F ). The PWVD results for the T3(2) with SNR = 0
and SNR = 6 dB are shown in Figure H.6(a) and (b), respectively.

The T4(2) code for a F = 600 Hz and tm =16 ms is shown in Figure H.7
To understand the bandwidth characteristics shown in Figure H.7, the phase
shift for the T4(2) is shown in Figure H.8. The smallest phase change here
is again six samples resulting in a bandwidth excursion of 1,167 Hz. Figure
H.9 shows the PWVD for the T4(2) code with Figure H.9(a) showing the
SNR = 0 dB case and Figure H.9(b) showing the SNR = 6 dB case (F =
600 Hz, B = 1,167 Hz).
Figure H.4: PWVD for polytime code T3(2) with F = 600 Hz, B = 1,167
Hz, and tm = T = 16 ms (signal only), showing the (a) marginal frequency
domain and (b) time-frequency plot.
Figure H.5: T3(2) phase shift showing a minimum subcode width of six sam-
ples (0.857 ms), resulting in a bandwidth excursion of 1,167 Hz.
Figure H.6: PWVD for T3(2) code with F = 600 Hz, B = 1,167 Hz, and
tm = T = 16 ms, showing the time-frequency plot for (a) SNR = 0 dB, and
(b) SNR = 6 dB.
Figure H.7: PWVD for T4(2) code with F = 600 Hz, B = 1,167 Hz,
tm = T = 16 ms (signal only), showing the (a) marginal frequency domain,
and (b) time-frequency plot.
Figure H.8: T4(2) phase shift showing a minimum subcode width of six sam-
ples (0.857 ms) resulting in a bandwidth excursion of 1,167 Hz.
Figure H.9: PWVD for T4(2) code with F = 600 Hz, B = 1,167 Hz, and
T = 16 ms, time-frequency plot, for (a) SNR = 0 dB, and (b) SNR = 6 dB.
Appendix I
QMFB Results for FMCW

with F = 500 Hz
In Figure I.1, the tm is increased from tm = 20 ms to tm = 30 ms, and the
bandwidth is increased from F = 250 Hz to F = 500 Hz. Figure 10.19(a),
shows layer l = 2. The same general characteristics are shown as in Figure
I.1(a) except that fewer zeros are needed to pad the signal to Np = 2,048. In
Figure I.1(b) the l = 5 layer is shown with F = 112.9 Hz and t = 4.64 ms
< tm /6. That is, this layer provides a finer detail in time, and less detail in
frequency. The bandwidth of the signal can easily be estimated as shown.
771
Figure I.1: QMFB time-frequency contour images for FMCW F = 500

MHz, tm = 30 ms (signal only), showing (a) layer 2, and (b) layer 5.
Appendix J
QMFB Results for 11-Bit

BPSK
An 11-bit Barker code (cpp = 1) is examined with the QMFB in Figure J.1.
In Figure J.1(a), the full contour QMFB output for layer l = 3 is shown. In
Figure J.1(b) a close-up view of the QMFB layer 3 output is shown, illustrat-
ing the time changing frequency detail. Note the similarity in the structure
of the frequency information when compared to Figure 10.21. Also note the
major dierences in the frequency structure.
Figure J.2 shows the QMFB output contour of the 11-bit signal for both
layers 3 and 6 for an SN R = 0 dB. Note that some of the features such
as bandwidth can be estimated but, even with a closer view of the results,
information such as the phase is hard to determine.
In Figure J.3, the bandwidth of the 11-bit signal is narrowed from 1 kHz
to B = 0.2 kHz (code period of T = 55 ms). Figure J.3(a) shows the QMFB
layer 2 for two code periods. The presence of a null at each BPSK phase shift
is evident. With T = 55 ms, the measurement of the smallest subphase code
is performed, to determine the number of subphase codes contained within
each section of the code. Figure J.3(b) shows the QMFB layer 6. The code
period T and bandwidth B are clearly identified. Comparison with Figure
10.23 indicates a slightly dierent form, due to the additional subphase codes.
773
Figure J.1: QMFB layer 3 for BPSK with 11-bit Barker code, cpp = 1 (signal
only), showing (a) the full contour image, and (b) a close-up view showing
frequency details of Barker code.
Appendix J: QMFB Results for 11-Bit BPSK 775
Figure J.2: QMFB contour image for BPSK with 11-bit Barker code, cpp = 1
(SN R = 0 dB), showing (a) layer 3, and (b) layer 6.
Figure J.3: QMFB contour images for BPSK 11-bit Barker code, cpp = 5
(signal only), showing (a) layer 2, and (b) layer 6.
Appendix K
QMFB Results for Frank

Signal with Nc = 16
The Frank code for M = 4 is shown in Figure K.1. The M = 4 Frank code
signal has Nc = 16 subcodes, fc = 1 kHz, and a cpp = 1. The QMFB
l = 2 layer for this signal is shown in Figure K.2 (f = 1,166.67 Hz, t =
573.67s). The QMFB has 10 layers (Np = 1,024). Figure K.2(a) shows the
entire signal (80 ms) and reveals that five code periods have been captured.
Note the unique structure of the phase modulation. A closer examination is
shown in Figure K.2(b). The code period (T = 16 ms) is shown, along with
the phase modulation characteristics due to the 16 subcodes. Correlation of
the frequency characteristics within the code period shown in Figure K.2(b),
with the phase modulation waveform shown in Figure K.1, can be made
directly (four sections).
In Figure K.3(a) and (b), the QMFB l = 4 layer is shown (f = 233.33
Hz, t = 2.32 ms), and reveals the linear frequency modulation that results
from the Frank phase code. The bandwidth is also indicated. Note that
the Frank code phase modulation results in the linear frequency modulation
wrapping around for the i = 3 and i = 4 segment, starting at t = 40 ms.
777
Figure K.1: Frank code phase values for M = 4 (Nc = 16).

Appendix K: QMFB Results for Frank Signal with Nc = 16 779
Figure K.2: QMFB contour images for M = 4 (Nc = 16) Frank code with
B = 1,000 Hz, T = 16 ms (signal only), showing (a) layer 2 output, and (b)
close up of layer 2, showing detailed frequency changes due to phase codes.
Figure K.3: QMFB contour images for M = 4 Frank code with B = 1,000
Hz, T = 16 ms (signal only), showing (a) layer 4 output, and (b) close up of
Appendix L
QMFB Results for P1, P2,

P3, and P4
L.1 P1 Analysis
The P1 signal is also derived from a linear FM waveform. In this section it
is shown that the P1 QMFB time-frequency characteristics are dierent from
the Frank code, and these dierences can be used for waveform identification.
In Figure L.1 the phase code for a P1 Nc = 64 signal is shown for reference.
The QMFB l = 2 layer is shown in Figure L.2. The total number of points
shown are Np = 4,096 with L = 12 QMFB layers, f = 1,166.67 Hz, and
t = 571.99 s. Since the number of phase codes is Nc = 64, the five code
periods extend for 320 ms. In the close-up view of the code period in Figure
L.2(b), the variation in phase modulation characteristics can be identified and
correlated with the phase waveform given above. This is especially evident
for the last four sections given in Figure L.1.
In Figure L.3(a) and (b), the l = 5 layer is shown, demonstrating the
linear frequency modulation resulting from the P1 phase codes. For this layer,
f = 112.9 Hz and t = 4.61 ms. Note the immediate frequency wraparound
at the beginning of the code period. Also evident is the nonuniform spacing
of the energy concentrations, due to the nonlinear phase modulation. The
code period (T = 64 ms) and bandwidth B = 1,000 Hz are also shown.
The marginal frequency characteristics for the Nc = 64 P1 code are shown
in Figure L.4. Due to the nonlinear phase modulation characteristics, the
energy is not symmetrically distributed about filter 9 (the carrier frequency
fc = 1 kHz). From Figure L.4 the four largest energy tiles (in order from
largest to smallest) are 7, 8, 10, and 9. That is, the carrier has the smallest
amount of energy among the four largest components.
781
Figure L.1: P1 code phase values for Nc = 64.
L.2 P2 Analysis
The P2 phase modulation diagram for Nc = 64 phase codes is shown in Figure
L.5. This phase diagram has a particularly interesting shape, giving rise to
some unique characteristics in the time-frequency domain.
The QMFB for the signal shown in this example has L = 12 layers. In
Figure L.6, the contour images for QMFB layer 2 are shown with f =
1,166.67 Hz and t = 571.99 s. The pattern of the frequency characteristics
changes form toward the middle of the code period (T = 40 ms). The result
is a linear frequency modulation that has a negative slope. This is illustrated
in Figure L.7, which shows the l = 5 layer. For this layer, f = 112.9 Hz and
t = 4.61 ms. Note from Figure L.7(b), the four energy concentrations are
located toward the center of the code period. Dierent signal modulations
will have distinct levels of energy concentration. Consequently, the marginal
frequency distribution can be used to identify the signals modulation type.
Figure L.8 shows the QMFB layer 5 marginal frequency profile for the Nc = 64
P2 code. From Figure L.8 the four largest energy tiles (in order from largest
to smallest) are 7, 9, 8, and 10. The dierence between the P1 and P2 code
can be identified by comparing marginal frequency distributions shown in
Figure L.8 and Figure L.4.
L.3 P3 Analysis
The P3 phase modulation for Nc = 64 (64 phase codes) is shown for reference
in Figure L.9. The QMFB l = 2 layer for the P3 signal is shown in Figure
L.10. The contour images are shown with f = 1,166.67 Hz and t = 571.99
Appendix L: QMFB Results for P1, P2, P3, and P4 783
Figure L.2: QMFB contour images for Nc = 64 P1 code with B = 1,000 Hz,
T = 64 ms (signal only), showing (a) layer 2 output and (b) close up of layer
2, showing detailed frequency changes due to phase codes.
T = 64 ms (signal only), showing (a) layer 5 output, and (b) close up of layer
5, showing resulting linear frequency modulation.
Figure L.4: QMFB layer 5 marginal frequency profile for Nc = 64 P1 code.

s. The shape of the time-frequency characteristics is unique from those of

the Frank, P1, and P2 codes.
Figure L.11 shows the l = 5 layer indicating the bandwidth and code
period. Here f = 112.9 Hz and t = 4.61 ms. A closer examination of the
layer 5 frequency details reveal a dierent distribution of the energy peaks
within the bandwidth. The marginal frequency distribution for the P3 code
is shown in Figure L.12. From Figure L.12 the four largest energy tiles (in
order from largest to smallest) are 7, 9, 10, and 8. Again, this is distinct
from the Frank, P1, and P2 codes and can be used to classify the detected
P3 signal.
L.4 P4 Analysis
In this section, an Nc = 64 P4 code is examined with the QMFB. The P4
phase modulation for Nc = 64 is shown in Figure L.13. The QMFB for
this P4 signal has L = 12 layers. The l = 2 layer (f = 1,166.67 Hz and
t = 571.99 s) is shown in Figure L.14. The close-up view examines the
time-varying frequency characteristics of the signal as it appears within two
of the three filters. Figure L.15 shows the l = 5 layer (f = 112.9 Hz and
t = 4.61), demonstrating the P4 linear frequency modulation. The close-up
view shows the major energy peaks about the carrier frequency (ninth filter).
For the P4 code, the marginal frequency distribution shown in Figure L.16,
reveals that the largest energy concentration is at the carrier frequency. the
four largest energy tiles (in order from largest to smallest) are 9, 7, 10, and
8.



Appendix M
QMFB Results for T2, T3,

and T4
Figure M.1(a) and Figure M.1(b) show the QMFB contour images for the
polytime T2(2) code with B = 875 Hz and T = 16 ms, showing the layer 2
output. The code modulation period is indicated, and the number of code
periods can also be identified. The close-up in Figure M.1(b) shows the unique
time-frequency pattern due to the T2(2) phase modulation. Figure M.2(a)
and Figure M.2(b) show the QMFB contour images for the polytime T2(2)
code with B = 875 Hz and T = 16 ms, showing the layer 4 output. Note the
bandwidth and code period can easily be indentified.
In Figure M.3(a) and (b), the QMFB contour images for the polytime
T3(2) code with B = 1,167 Hz, T = 16 ms (signal only) are shown. In
Figure M.3(a) the layer 2 output is shown, and in Figure M.3(b) a close-up of
layer 2 is used to indicate the frequency changes due to the phase codes. In
Figure M.4, layer 4 is examined, and shows the bipolar frequency modulation
characteristic of the T3(2) code. Also indicated are the bandwidth and code
period.
Layer 2 for the T4(2) code is shown in Figure M.5. All five code periods
are shown, as well as the code period of 16 ms. Layer 4 is shown in Figure
M.6. Interestingly enough, all of the polytime codes have a large frequency
spike within each code period, due to the recycling of the phase modulation.
797
Figure M.1: QMFB contour images for polytime T2(2) code with B = 875
Appendix M: QMFB Results for T2, T3, and T4 799
Figure M.2: QMFB contour images for polytime T2(2) code with B = 875
Figure M.3: QMFB contour images for polytime T3(2) code with B = 1,167
Appendix N
Cyclostationary Processing
Results with FMCW
F = 500 Hz
The extraction of the parameters from a wideband FMCW signal F = 500

Hz using cyclostationary processing is not significantly dierent, and is shown
in Figure N.1. The modulation period for this example tm = 30 ms. Figure
N.1(a) shows the modulation pattern with centroid at 2fc = 2 kHz. Note
that the arrowhead pattern unique to the FMCW is still present, but has
a bit more structure. With this resolution, F can easily be measured;
however, Rc cannot be measured. The closer examination shown in Figure
N.1(b) reveals the unique modulation characteristic of the FMCW waveform.
The value of Rc = 16.7 Hz is now easily identified giving a tm = 30 ms.
The SNR = 0 dB case is shown in Figure N.2(a) and (b). Note that the
cyclostationary results are fairly robust in significant amounts of noise.
805
Figure N.1: Frequency-smoothing SCD patterns for a F = 500 Hz, tm = 30-

ms triangular FMCW signal with fc = 1 kHz, showing (a) complete FMCW
modulation as one part of the bifrequency plane, and (b) closer examination
showing Rc measurement.
Appendix N: Cyclostationary Results with FMCW F = 500 Hz 807
Figure N.2: Frequency-smoothing SCD patterns for a F = 500 Hz, tm = 30-

ms triangular FMCW signal with fc = 1 kHz, showing (a) the bifrequency
plane with SNR = 0 dB, and (b) closer examination showing Rc measurement.
Appendix O
Results with Frank Signal,
Nc = 16
For the signal examined, fc = 1 kHz, cpp = 1, fs = 7 kHz, and Nc = 16

phase subcodes. The frequency-smoothing SCD is generated using k = 16
Hz with N = 1,024. One of the four modulation patterns generated in the
SCD is shown in Figure O.1(a). As in previous examples, the pattern centroid
is located at = 2fc . For the Frank code this is a bit more dicult to identify,
due to the dierent slopes of phase shift within each single code period T .
Generally, it lies in the center of the cross-hatch area inside the lesser SCD
amplitude (faded) regions. That is, the cross-hatch region of interest is not
symmetrical about the centroid. Location of the centroid can be used to
determine the bandwidth on the cycle frequency axis, but there is a more
straightforward method for determining the bandwidth in the bifrequency
plane. To do this we first outline the cross-hatch region (larger amplitude
SCD) with a parallelogram (shown by dashed lines). The top and bottom
corners of the parallelogram are colocated at the same cycle frequency, and
the left and right corners are colocated at the same frequency (f = 0). This
information helps identify the parallelogram position, and the bandwidth B
can be estimated more accurately. For the Frank code, the Rc measurement
is related to the code period as Rc = 1/T = 1/Nc tb . In Figure O.1(b), a closer
examination shows that the Rc measurement is straightforward, and for this
signal Rc = 62.5 Hz. The number of subcodes within a code period is then
Nc = B/Rc = 16. Figure O.2 shows the bifrequency analysis for the Frank
code in the presence of noise. In Figure O.2(a), the Frank code modulation is
809
Figure O.1: Frequency-smoothing SCD patterns for the Frank code with Nc =
16, fc = 1 kHz, and cpp = 1, showing (a) one of four Frank code modulation
patterns and measurement parallelogram, and (b) closer examination with
Rc measurement.
Appendix O: Cyclostationary Results with Frank Signal, Nc = 16 811
Figure O.2: Frequency-smoothing SCD patterns for the Frank code with Nc =
16, fc = 1 kHz, cpp = 1, and SN R = 0 dB, showing (a) one of four Frank
code modulation patterns on the bifrequency plane with the measurement
parallelogram, and (b) closer examination illustrating the Rc measurement.
shown with an SN R = 0 dB. Note the position of the parallelogram to enclose

the cross-hatch region. Also note the corner locations and the bandwidth
measurement. Figure O.2(b) shows the Rc measurement. In summary, the
unique pattern of the wideband Frank code lets us determine all of the signal
parameters using the SCD bifrequency plane.
Appendix P
Results for P1, P2, P3,
and P4
P.1 P1 Code Analysis

In this section we investigate the time-smoothing SCD for the P1 signal,
and illustrate the corresponding extraction technique. The signal examined
is a P1 phase-modulated signal with fc = 1 kHz, Nc = 64 subcodes, and
cpp = 1 (wideband). Figure P.1(a) shows the complete bifrequency plane, and
reveals that the P1 code also presents itself in an insect pattern. Compared
to the Frank code signal, however, note that the insect is pointed to the left.
This is expected, since the time-frequency slope as measured by the Wigner
distribution and quadrature mirror filtering is opposite to that of the Frank
code. As illustrated in Figure P.1(b), the bandwidth is measured on the cycle
frequency axis in a similar fashion to the Frank code, except that the head
is on the left. Correlation with the frequency axis measurement of B is also
illustrated.
Also indicated in Figure P.1(b) is a box that is examined in closer detail to
illustrate the Rc measurement. Figure P.2(a) and (b) illustrate the measure-
ment of Rc = 1/T . Figure P.2(b) indicates Rc = 15.5 Hz, giving a modulation
period of 64 ms. Again, since the number of subcodes used by LPI radar are
most often a power of 2 (e.g., 64 = 26 ), an accurate result can be obtained
even from a bifrequency plane with small SNR. Here Nc = B/Rc = 64.
813
Figure P.1: Time-smoothing SCD insect patterns for the P1 code with Nc =
64, fc = 1 kHz, and cpp = 1, with (a) the complete bifrequency plane, and
(b) closer examination of one of the four modulation patterns illustrating the
bandwidth measurements.
Appendix P: Cyclostationary Results for P1, P2, P3, and P4 815
Figure P.2: Close examination of time-smoothing SCD for the P1 code with
Nc = 64, fc = 1 kHz, and cpp = 1, with (a) modulation cycles, and (b) the
measurement of Rc .

The time-smoothing SCD results for the P2 code are illustrated in Figure
P.3(a). The P2 modulation has the expected insect shape, however, it is
pointing to the right (similar to the Frank code). Note also the distinct nulls
present in the bifrequency. The bandwidth is measured in a similar manner
to the P1 and Frank code; however, the bifrequency nulls must be used as
illustrated in Figure P.3(b). One advantage of using the bifrequency plane
for measuring Rc for the subcode period tb and number of subcodes N is the
nonspecific position in the bifrequency plane, where Rc can be measured. As
shown in Figure P.4(a) and (b), any (, k) region can be used to estimate the
value of Rc .

The time-smoothing SCD results for the P3 code are shown in Figure P.5.
The signal has fc = 1 kHz, Nc = 64 (64 subcodes), and cpp = 1. The insect
modulation pattern points to the right and the measurements of B and Rc ,
as well as N and cpp, are the same as the above cases.

The time-smoothing SCD results for the P4 code are shown in Figure P.6.
The signal has fc = 1 kHz, Nc = 64, and cpp = 1. The insect modulation
pattern points to the right and the measurements of B and Rc , as well as N
and cpp, are the same as above.
64, fc = 1 kHz, and cpp = 1, with (a) the complete bifrequency plane, and
Figure P.4: Close examination of time-smoothing SCD for the P2 code with
Nc = 64, fc = 1 kHz, and cpp = 1, with (a) modulation cycles, and (b) the
measurement of Rc .
64, fc = 1 kHz, and cpp = 1 with (a) the complete bifrequency plane, and
64, fc = 1 kHz, and cpp = 1 with (a) the complete bifrequency plane, and
Appendix Q
Results for T2, T3, and T4
Polytime Codes
Q.1 Polytime T2(2) Code Analysis

In this section, the frequency-smoothing SCD is used to examine the polytime
T2(2) code. The T2(2) code also has an fc = 1 kHz, and has a time-modulated
binary phase shift (of various widths). Figure Q.1(a) shows the bifrequency
plane and the four modulation patterns. Figure Q.1(b) shows one of the
four unique patterns, and illustrates how the bandwidth of the signal can
be measured. Comparison with the T1(2) code shows that the bandwidth is
exactly one-half as large as in the T1(2) code. For the T2(2) signal shown
in Figure Q.1(b), B = 875 Hz. Figure Q.2 shows a closer examination of the
bifrequency plane, and the measurement of Rc = 1/T = 62.5 Hz. This gives
the estimate for the code period as T = 16 ms. Note also that an SCD spot
does not exist at = 2fc , k = 0 and is an additional method to distinguish
between the two bifrequency patterns.

In this section, the frequency-smoothing SCD is used to examine the polytime
binary phase shift (of various widths) across a modulation bandwidth F =
600 Hz. Figure Q.3(a) shows the bifrequency plane and the four modulation
821
Figure Q.1: Frequency-smoothing SCD patterns for the polytime T2(2) code
with fc = 1 kHz, showing (a) the complete bifrequency plane, and (b) closer
examination of one of the four modulation patterns illustrating the bandwidth
measurement.
Appendix Q: Cyclostationary Results for T2, T3, and T4 Codes 823
Figure Q.2: Close examination of the frequency-smoothing SCD pattern for

the polytime T2(2) code, illustrating the Rc measurement.
patterns. Figure Q.3(b) shows one of the four unique patterns. The distance
from the centroid to the corner of the dot pattern on the k = 0 axis is 2F .
For the T3(2) signal shown in Figure Q.3(b), F = 600 Hz. Figure Q.4
shows a closer examination of the bifrequency plane, and the measurement of
Rc = 1/T = 62.5 Hz. This also correctly gives the estimate for the code period
as T = 16 ms. Note also that an SCD spot does exist at ( = 2fc , k = 0).

In this section the frequency-smoothing SCD is used to examine the polytime
binary phase shift (of various widths) across a modulation bandwidth F =
600 Hz. Figure Q.5(a) shows the bifrequency plane and the four modulation
patterns. Figure Q.5(b) shows one of the four unique patterns. The distance
from the centroid to the corner of the dot pattern on the k = 0 axis is also
2F . For the T4(2) signal shown in Figure Q.5(b), F = 600 Hz. Figure Q.6
shows a closer examination of the bifrequency plane, and the measurement of
Rc = 1/T = 62.5 Hz. This also correctly gives the estimate for the code period
as T = 16 ms. Note also that an SCD spot does exist at ( = 2fc , k = 0).
with fc = 1 kHz showing (a) the complete bifrequency plane, and (b) closer
measurement.

with fc = 1 kHz, showing (a) the complete bifrequency plane, and (b) closer
measurement.

List of Symbols
There are not enough symbols in the English and Greek alphabets to allow the
use of each letter or symbol once. Consequently, some symbols may be used to
denote more than one variable, but their use should be clear from the context.
Symbols
a multiplying coecients for envelope approximation detector
A complex leakage signal
A signal amplitude
A Albersheim SNR coecient
A coecients for seastate model
A continuous aperiodic autocorrelation function
A feature vector from time-frequency image
Aavg moving average filter output
An normalized, filtered marginal frequency distribution
A2 power reflection coecient of target
An excitation coecients
Ae eective area of the radar receive antenna
b spiral rate constant
b multiplying coecient for envelope approximation detector
b(n) Wiener filter output
bsc number of samples per subcode
B Albersheim SNR coecient
B coecients for seastate model
B feedthrough signal under vector modulator control
B positive real parameter for Taylor array
B absolute signal bandwidth
829
B conical spiral base diameter

Ba bandwidth after demodulation
BIR intercept receiver front-end RF bandwidth
BIV intercept receiver video bandwidth
BI intercept receiver bandwidth
BRi radar receiver input bandwidth
c speed of light
c window function selection variable
cdf cumulative distribution function
cn radial basis function centers
cpp cycles of the carrier frequency per subcode
C Hamming window coecient
C channel capacity
C coecients for seastate model
Cf (t, , ) Cohens general class of time frequency distributions
C(td , ; Tr , r ) cross-correlation
C(x) cosine Fresnel integral
Ci weighting factor
CD ( ) output of digital correlation receiver
CDA ( ) output of digital-analog correlation receiver
CI ( ) output of ideal analog correlation receiver
Cl spiral circumference outer diameter
CM connectivity measure
R
CM reference connectivity measure
Cu spiral circumference in feed region
CWDx ( , ) Choi-Williams distribution
d antenna element spacing
d conical spiral apex diameter
d length of route
da aperture dimension in azimuth
de aperture dimension in elevation
da diameter of circular antenna
dc duty cycle
D delay
Df angular frequency per volt of FMCW
e(t) transmitted noise plus FMCW signal
ec (t) noise FMCW plus sine transmitted signal
eR (t) noise plus FMCW echo of moving target
List of Symbols 831
E amplitude of noise plus FMCW signal

E energy
ER amplitude of noise plus FMCW echo
ERPR eective radiated power of radar
ERPJ eective radiated power of jammer
ERPC eective radiated power of data communication node
f1b upslope beat frequency
f2b downslope beat frequency
fb beat frequency
fc carrier frequency
fl lower frequency limit of spiral antenna
fu upper frequency limit of spiral antenna
fu unambiguous Doppler frequency
fclk clock frequency
f (t) continuous signal
fd max random binary phase modulation Doppler tolerance
fbmax maximum beat frequency
fHL hard-limiting nonlinearity
fl (n),flI (n) kernel function for Wigner-Ville distribution
fm video modulation signal
fj FSK transmitted frequency
fs (y) sigmoid nonlinearity function
fcr critical frequency
fs sampling frequency
fc1 first frequency of a two-tone signal
fc2 second frequency of a two-tone signal
f (k1 , k2 ) time-frequency image
fI single sawtooth FMCW signal
f0 frequency constant
F (u, v) frequency rectangle
Fi flux power incident on a target
F chirp rate
F I angular frequency increment per sample
FI intercept receiver noise factor
FR radar receiver noise factor
F (u) array pattern
F, time dependent flow coecient
F1 , F2 spiral antenna feed points
g response of the receiver

g Tikhonovs regularization parameter
G antenna gain
G filter transfer function
G Gabor time-frequency distribution
G(x) antenna taper function
Gr antenna receive gain along boresight
Gt antenna transmit gain along boresight
GIt antenna transmit gain in side lobe
GI intercept receiver antenna gain
gn ( ) modulation function
h height
ha,b (t) single prototype wavelet
h correlation integral
h impulse response
h(k1 , k2 ) filter mask
h(n) histogram
(q)
hlk received signal from scatterer q transmitter l, rcvr k
hF 2layer F2-layer height from Earths surface
ht dipole height above ground
H number of hidden layers in multilayer perceptron
H entropy
H(u, v) 2D filter transfer function
Hd (u, v) ideal lowpass filter transfer function
H() antialiasing filter transfer function
H0 (z) highpass filter
H1 (z) lowpass filter
i number of the sample in a given frequency
I inphase term
I neural network input nodes
I(t, ) intensity image
Ib transmission loss
I(j) measure of information in jth message
Ic circulator isolation
IBar Barricks transmission loss
IR network reach
List of Symbols 833
j frequency index
k Boltzmanns constant
k discrete frequency index
kres frequency resolution
K normalizing constant
K (t) capability value of node
Kmax maximum number of targets simultaneously identified
KJ jammer capability value
l quadrature mirror filter layer index
L total length of spiral antenna
L system losses
L total number of wavelet layers
L overlap sliding factor between each short time FFT
L(n) sum of the square of the coecients
Lc circulator loss
LIR losses between intercept receiver antenna to receiver
LP 2 two way transmission path loss through ionosphere
LRR losses between the antenna and receiver
LRT losses between the transmitter and the antenna
Lx transmission line loss
L,
information flow parameter
L1 one-way atmospheric transmission factor
L2 two-way atmospheric transmission factor
M number of pulse compressors (random binary phase modulation)
M square root of number of subcodes for Frank, P1, P2
M number of LPI transmitters in MIMO configuration
M sampled data length
M number of channel pair regions on the bifrequency plane
M Grenanders uncertainty condition
MSE mean sum of squares of network errors
MSW mean sum of squares of network weights and biases
n cyclostationary order
n time index
n number of reference cells in CFAR
n width of data path in the accumulator
n discrete index of Doppler frequency
n number of continuous antenna beams in elevation stack
n(t) receiver thermal noise
nr number of resolution elements in scan volume
N number of code periods used in correlation receiver

N number of arms interleaved within spiral antenna
N number of statistically independent noise samples
N FFT size
N number of receivers in MIMO configuration
N total number of discrete samples of time
NI discrete short time FFT size
Nc number of subcodes, processing gain
Nc feature vector dimension
Nd number of sections in subdivided scan volume
Ne electron density
Nm maximum electron density
Np total number of quadrature mirror samples
Nr feature vector dimension
NI integration intervals
NB number of Barker phase codes
NF number of transmitted continuous frequencies
NF number of tiles displayed in frequency
NI number of noncoherent integration intervals
NRB number of range bins containing clutter
NT number of tiles displayed in time
NT total number of phase slots in the FSK/PSK waveform
NT number of nodes connected in the network
NP nepers
NX beginning target state in Markov chain
NY ending target state in Markov chain
NF coherent processing interval
NI FFT record length
NF record length
N total number of nodes connected to node
N, total number of routes between node and
p(t) aperiodic rectangular window function
p( ) autocorrelation of the transmitted waveform
pj probability of transmitting jth message
pr surface wave received power from target
pt surface wave radiated power in presence of ground
pdf probability density function
List of Symbols 835
P number of polyphase code periods transmitted

Prad radiated power from the antenna
Pin total input power to the antenna
P DD power density at a range R
P GR processing gain
Pavg adaptive average transmitted power
Pd probability of detection
Pf a probability of false alarm
Prerad surface wave power reradiated by target
Pt peak power of pulsed emitter
Ptot total power at receiver using circulator
PN (Z) square law detector output with no target present
PCW average power of CW transmitter
PRC clutter power within a range bin
PRT received signal power of the radar from the target
PT R signal power to the intercept receiver
P GI intercept receiver processing gain
Q quadrature term
Q( ) knowledge function
r discrete time index
r radial distance
rb radial distance to layers base
re Earths radius (6,3781.1 km)
rk aperiodic autocorrelation coecients
rl received signal at antenna l
r1 spiral antenna generating function
R range
R neural network regularization
R maximum spiral radius
Rc code rate
Rc continuous cross-correlation function
Rd range error due to Doppler shift
Rfootprint OTHR range along Earths surface
RI range between LPI radar and intercept receiver
RImax maximum intercept range for intercept receiver
Rk target range or path length in kilometers
Rmax maximum detection range of the LPI radar
Rcancel,dB reflected power cancelation depth
Ru unambiguous range
Rx ( ) time average autocorrelation function

Rx ( ) cyclic autocorrelation of a complex time series x(t)
R(rtb ) periodic autocorrelation function
R(, ) Radon transform
RC range of communication node
RJ range of jammer
RQ network richness
RT target range
s(f ) single sweep FMCW spectrum
s(f ) roll-o rate single sweep FMCW spectrum
s(t) complex signal
sl (n) orthogonal polyphase complex signal
s1 (t) transmitted signal upslope
s2 (t) transmitted signal downslope
S slope of polyphase modulation in time-frequency plots
S(t) complex stationary process
S(x) sine Fresnel integral
I
SXN (n, k) estimated timed smoothed periodogram
Sx (f ) power spectral density
Si,j target scattering coecient with i, j = V or H
Sv (f ) noise plus FMCW correlation output spectrum
Sx (f ) power spectral density
Sx (f ) spectral correlation density

SX TW
continuous time-smoothed cyclic periodogram

SX NI
discrete time-smoothed cyclic periodogram
Sd output of up converter MXR1 in noise radar
S1,2 number neurons in first, second hidden layers in MLP
SNRIo intercept receiver output SNR from signal processor
SNRIi intercept receiver input SNR to signal processor
SNRnet SNR for netted radar
SNRRi minimum SNR required at the radar input
SNRRo output, Albersheim SNR
SNR1 SNR for single monostatic radar
t time
tb subcode period or duration
td round trip propagation time delay
tm modulation period
tI smallest radar coherent integration time
tp transmitted frequency duration
List of Symbols 837
t0 coherent processing interval

t0 signal time of arrival
T code period
T threshold multiplier
Tf frame time
Th noise radar threshold
Th eigenvalue threshold
Tint measurement time in noise receiver
Tp time for target to pass through range cell
TA lower limit angle threshold
TB upper limit angle threshold
TN short time FFT window
TN pulse repetition interval
TR delay of RF delay line
T0 standard temperature in Kelvin
TW short-time FFT window length
u(t) periodic complex envelope
v, V target velocity
i eigenvector selection
Vt threshold voltage
Vt maximum closing velocity of the target
w(t) weighting function
wST (f ) frequency domain taper function
wCT (t) time domain cosine-Tukey amplitude taper function
wl (t) additive symmetric zero mean Gaussian noise
We (f ) power spectrum of transmitted noise plus FMCW signal
Wn perceptron weighting vector
Wx (t, ), WX (, t) Wigner-Ville distribution
W () Fourier transform of window
WN , WM window functions for Choi-Williams
x0 cyclic frequency constant
X training matrix for principal components analysis
Xn n-dimensional perceptron input vector
Xq position of scatterer
XT Fourier transform of x(t)
XW F (, ) windowed Fourier transform of x(t)
Xr (f ) Fourier transform of the cyclic autocorrelation function
XTW (t, f ) continuous short time Fourier transform
XN I (n, k) discrete Fourier transform

XW (a, b) wavelet transform
y1 , y2 CFAR noise power levels
ym layer semithickness
ypn output of radial basis function
z(t) delay product waveform of x(t)
Z square law detector response
Z0 free space impedance (377 )
cycle frequency (continuous time)
k linear transform of continuous time signal
|(, )| ambiguity function magnitude
|N T (, )| ambiguity function for N reference correlators
|T (, )| single period ambiguity function
sensitivity ratio
spiral antenna rotation angle
range dierence between direct and multipath echoes
(k) Kroenekers delta function
F bandwidth increment
phase error from perfect quadrature
I intercept receiver sensitivity
R radar receiver sensitivity
i,j cell value in the dierence triangle
cycle frequency resolution
change in Doppler oset
f frequency resolution
t time resolution
tmin minimum time delay that can be detected
k points spacing in frequency
F FMCW modulation bandwidth
F I eective modulation bandwidth
R range resolution
RI eective range resolution
v first blind speed
w coherent processing interval
target extent in azimuth
spectral width of the beat frequency
aperture eciency
r relative error
discrete time cycle frequency
List of Symbols 839
route index
noncoherent integration eciency
FMCW flyback factor
res cycle frequency resolution
(t) target reflectivity profile
voltage reflection coecient
discrete time index
QMFB layer number
wavelength
information rate of source
C2 decision tempo
d deployment tempo
f fighting tempo
min minimum information rate
T characteristic tempo
OODA maximum operational tempo
node index
local mean
refractive index of ionosphere
Doppler frequency oset
node index
radian frequency
r scan rate
IF intermediate radian frequency
LO local oscillator angular frequency
frequency boundary
a scan coverage in azimuth
e scan coverage in elevation
s scan volume
(, ) kernel function for time-frequency distribution
c phase modulation
i P3, P4 phase sequence
i incidence angle
i,j Frank, P1, P2 polyphase sequence
k general phase modulation function
l orthogonal polyphase sequence
r phase shift
1 FMCW phase
T 1 T1(n) polytime sequence

0 initial angle of spiral antenna
(t) basis set
(p) Euler function of positive p
grazing angle
|(, )| ambiguity function of mismatched receiver
b solid angle within the half-power beam contour
spiral antenna generating function
spiral antenna generating function in wavelengths
ej complex correlation coecient
(t) periodic rectangular window
V voltage standing wave ratio
spread of radial basis function
1 total received power in noise radar
2 total power in delayed replica in noise radar
2 white Gaussian noise power
0 incremental backscattering coecient of the sea
0i mean sea backscatter coecient for seastate i
nj elements of radial basis function covariance matrix
min minimum detectable radar cross section
s transmitted noise signal power
s Shearmans definition of backscattering cross-section
sr power in the noise radar received signal
F S free-space backscattering cross-section
T target radar cross section
oset time delay
d dwell time
R pulse width
tk , rl propagation time delay
a 3 dB beamwidth in azimuth
e 3 dB beamwidth in elevation
n target azimuth angle
s Radon projection angle for maximum intensity
start beginning of targets position in azimuth
stop end of targets position in azimuth
0 angle of main lobe peak
(p, q) multichannel time-frequency LPI detector
i target reflectivity
Glossary
AARGM Advanced antiradiation guided missile
ACF Autocorrelation function
ADC Analog-to-digital converter
AEA Airborne electronic attack
ALCM Air launched cruise missile
AMRFC Advanced multifunction RF concept
AO Acousto-optic
AOA Analysis of alternatives
AOA Angle of arrival
AREPS Advanced Refractive Eects Prediction System
ARES Aordable reactive strike missile
ARM Antiradiation missile
ARMIGER Antiradiation missile with intelligent guidance and extended range
ARSR Air route surveillance system
ASCM Antiship capable missile
ATR Automatic target recognition
AWACS Airborne warning and control system
AZ Azimuth
B-F Bifrequency
BMEWS Ballistic missile early warning system
BPF Bandpass filter
BPSK Binary phase shift keying
C2 Command and control
CARA Combined Altitude Radar Altimeter
CCD Charge coupled device
CFAR Constant false alarm rate
CMRA Cruise Missile Radar Altimeter
COSPAR Committee of Space Research
CSIST Chung-Shan Institute of Science and Technology
841
CW Continuous waveform
CWD Choi-Williams distribution
DAC Digital-to-analog converter
DARPA Defense Advanced Research Projects Agency
DC Direct current
DDS Direct digital synthesizer
DFT Discrete Fourier transform
DFSM Direct frequency-smoothing method
DIP Digital information pheromones
DLVA Detector logarithmic video amplifiers
DoD Department of Defense
DRFM Digital radio frequency memory
DSP Digital signal processing
EA Electronic attack
EL Elevation
ELINT Electronic intelligence
EP Electronic protection
ERP Eective radiated power
ES Electronic support
EW Electronic warfare
EWO Electronic warfare ocer
FAM FFT accumulation method
FET Field eect transistor
FFT Fast Fourier transform
FH Frequency hopping
FIR Finite impulse response
FLAPS Flat parabolic surface
FLIR Forward looking infrared
FMCW Frequency modulation CW
FOT Optimum working frequency
FOV Field of view
FSK Frequency shift keying
GAO General accounting oce
GCS Ground control station
GDA Great deluge algorithm
GOCFAR Greatest-of constant false alarm rate
GPS Global positioning system
GUI Graphical user interface
Glossary 843
HARD Helicopter and Aircraft/Radar Detection

HARM High-speed antiradiation missile
HCI Human computer interface
HDAM HARM destruction of enemy air defense attack module
HEMT High electron mobility transistor
HTS High temperature superconductor
HTS HARM targeting system
HPM High power microwave
IADS Integrated air defense system
ICAP Increased capability
IF Intermediate frequency
IFF Identification friend or foe
IG Ionosphere index
IIR Imaging infrared
IMU Inertial measurement unit
INS Inertial navigation system
IR Infrared
IRI International Reference Ionosphere
IRST Infrared search and track
ISAR Inverse synthetic aperture radar
ISL Integrated side lobe level
JCC JORN coordination center
JORN Jindalee over-the-horizon network
JSR Jam-to-signal ratio
JSTARS Joint Surveillance and Target Attack Radar System
LAMPS Light Airborne Multipurpose System
LAN Local area network
LANTIRN Low-Altitude Navigation and Targeting Infrared for Night
LCM Least common multiple
LNA Low noise amplifier
LO Local oscillator
LPF Lowpass filter
LPI Low probability of intercept
LPID Low probability of identification
LPIT Low probability of intercept toolbox
LPRF Low pulse repetition frequency
LUT Lookup table
MALD Miniature Air Launched Decoy
MALI Miniature Air Launched Interceptor
MATLAB Matrix Laboratory
MF Matched filter
MFAB Marginal frequency adaptive binarization

MIMO Multiple input multiple output
MIP Millions of instructions per second
MLP Multilayer perceptron
MMIC Monolithic microwave integrated circuit
MMW Millimeter wave
MRSR Multirole Survivable Radar
MTD Moving target Doppler
MTI Moving target indication
MUF Maximum usable frequency
NATO North Atlantic Treaty Organization
NCW Network centric warfare
OLPI Omnidirectional LPI radar
OODA Observation-orientation-decision-action
OTH Over-the-horizon
OTHR Over the horizon radar
PACF Periodic autocorrelation function
PAF Periodic ambiguity function
PAGE Portable air-defense guard equipment
PALS Precision Approach and Landing System
PANDORA Parallel array for numerous dierent operational research activities
PCA Principal components analysis
PDW Pulse descriptor word
PG Passive guidance
PGM Precision guided munitions
PLL Phase-locked loop
PPI Planned position indicator
PRF Pulse repetition frequency
PRI Pulse repetition interval
PSD Power spectral density
PSK Phase shift keying
PSL Peak side lobe level
PWVD Pseudo Wigner-Ville distribution
QMF Quadrature mirror filter
QMFB Quadrature mirror filter bank
RAAF Royal Australian air force
RAM Rolling airframe missile
RBF Radial basis function
RBPC Random binary phase code
RCS Radar cross section
RF Radio frequency
Glossary 845
RISP Relative to isotropic antenna at same position

RNR Random noise radar
RNFR Random noise plus FMCW radar
RNFSR Random noise FMCW plus sine radar
RPC Reflected power canceler
RF Radio frequency
RPM Revolutions per minute
RPV Remotely piloted vehicles
RX Receiver
RTIC Real time into the cockpit
RTOC Real time out of the cockpit
RWR Radar warning receiver
SAM Surface-to-air missile
SAR Synthetic aperture radar
SATCOM Satellite communication
SAW Surface acoustic wave
SCD Spectral correlation density
SCR Signal-to-clutter ratio
SEAD Suppression of enemy air defense
SEI Specific emitter identification
SFDR Spurious free dynamic range
SIGINT Signals intelligence
SJR Signal-to-jam ratio
SLR Side lobe ratio
SNR Signal-to-noise ratio
SSBM Single-sideband modulator
SSN Sun spot number
STAP Space time adaptive processing
STC Sensitivity time control
STFT Short-time Fourier transform
SVD Singular value decomposition
TALS Tactical Automatic Landing System
TCR Target-to-clutter ratio
TEL Transporter erector launcher
T-F Time-frequency
TJS Tactical Jamming System
TOA Time of arrival
TR Transmit and receive
UAV Unmanned aerial vehicle
UCARS UAV Common Automatic Recovery System
UHF Ultra high frequency

U.K. United Kingdom
URSI Union of Radio Science International
U.S. United States
USAF United States Air Force
UT Universal time
UTC Coordinated universal time
UWB Ultra wideband
VCO Voltage controlled oscillator
VHF Very high frequency
VLSI Very large scale integrated circuits
VSTOL Vertical stationary take-o and landing
WGN White Gaussian noise
WRF Waveform repetition frequency
WT Wavelet transform
WVD Wigner-Ville distribution
WWII World War II
XNOR Exclusive not OR
XOR Exclusive OR
About the Author
Phillip E. Pace is a professor in the department of electrical and computer

engineering at the Naval Postgraduate School (NPS). He received B.S. and
M.S. degrees from Ohio University in 1983 and 1986, respectively, and a Ph.D.
from the University of Cincinnati in 1990 all in electrical and computer
engineering. Prior to joining NPS, he spent 2 years at General Dynamics
Corporation, Air Defense Systems Division, as a design specialist in the Radar
Systems Research Engineering Department. Before that, he was a member of
the technical sta at Hughes Aircraft Company, Radar Systems Group, for 5
years. He was selected for the Outstanding Research Achievement Award in
1994, 1995, and 1998 for his work at NPS in electronic warfare, and received
the 1995 Association of Old Crows (AOC) Academic Training Award.
Dr. Pace directs the NPS Center for Joint Services Electronic Warfare, has
been the chairman of the Navys Threat Missile Simulator Validation Work-
ing Group since October 1998, and was a participant on the Navys NULKA
Blue Ribbon Panel in January 1999. He is the author of the textbook
Advanced Techniques for Digital Receivers (Artech House, 2000), and has
been a principal investigator on numerous research projects in the areas of
receiver design, signal processing, electronic warfare, and weapon systems
analysis. Dr. Pace invented the concept of symmetrical number systems,
has five patents, over thirty journal publications and is a senior member
of the IEEE.
847
Index
A Swiss Air Guard, 607

twinkle transmission, 607
Abdullah, 553 warning radar and decoy,
Activation function, 630 608609
AD1990 altimeter, 45 Antiradiation missile (ARM), 4,
Adaptive binarization 637639 AARGM, 569571, 592593
Agility, 325326 Alamo, 578579
Alarm ARM, 564 Alarm, 564, 597599
Altimeter, 4145 antenna design, 559566
AD1990, 45 ARES, 593
AHV-2100, 45 ARMIGER, 600
CARA, 42 Corvus, 554
CMRA, 43 dual-mode, 567, 569
GRA-2000, 44 FT-2000, 604606
HG-9550, 43 HARM, 591592
PA-5429, 44 Harpy, 601602
Ambiguity function, 68 Kegler, 585586
Amplitude weighting, 77 Kelt, 580581
AMRFC, 13 Kh-27, 585
Amplifier Kickback, 587
predetection, 28 Kilter, 584585
postdetection, 28 Kingfish, 581582
AN/APG-77, 5657 Kitchen, 579580,
AN/APS-147, 56 Krypton, 565566, 587589
AN/APQ-181, 56-57 Kyle, 582584
Analog processor, 20 LPI processing, 572576
Analog to digtal converter, 20 Martel, 596597
Anti-ARM, 606612 millimeter, 569570
AN/TLQ-32 ARM-D decoy, 611 performance metrics, 577
Cosmic Shield, 607 Rolling airframe missile, 594595
flaps decoy, 608609 seeker, 566571, 605
Gazetchik, 610611 Shrike, 555, 589590
VHF/UHF, 607 Sidearm, 593594
Patriot, 607 signal processing, 571572
position flexibility, 606 Standard, 591
Swedish GLV200, 607 STAR-1, 603604
849
summary (Russian), 578 training, 631

targets, 557 Sheridan levels, 622-623
Tien Chien IIA, 598599 Autonommous parameter extraction
Antiship missile Wigner-Ville with Radon, 688
RBS-15, 5860 696
seeker technology, 301305 AHV-2100 altimeter, 45
Antenna, 5
aperture distribution, 7 B
bandwidth, 6
beamwidth (half-power), 5 Back lobe, 5
dimension, 6 Bandwidth
eective area, 25 fractional, 210
eciency, 5 instantaneous fractional, 211
gain, 5 intercept receiver, 2829
isotropic, 24 modulation (FMCW), 1920
isolation (using circulator), phase code, 127
9697 radar, 26
isolation (using RPC), 9799 Barker phase codes
main lobe, 5 binary, 128133
nonscanning, 12 polyphase, 133139
pattern, 510 Basis functions, 468
pencil beam, 6, 12 Beamforming
phased array, 67 digital, 14
radiation intensity, 8 Beamwidth, 6
scan pattern, 1013 Beat frequency, 18, 103104
side lobe, 510 bifrequency, 513,523524, 699
side lobe ratio, 5 Blind speed, 102
simultaneous transmit, 13 for OTHR, 264
spiral design, 559566 Boltzmanns constant, 26
stacked beam, 12 Bottleneck, 333, 341
taper functions, 9
Atmospheric C
absorption, 17
transmission, 25 Capability value, 326327
Autocorrelation function, 22 jammer, 338
Autonomous classification CARA altimeter, 42
authority, 621622 CHAIN HOME, 553
feature extraction, 634639 Channel capacity, 334
multilayer perceptron, 624, 629 unifying principal in EW, 334-
632 335
radial basis function, 624, 632 Chinese OTH-B, 272276
633 Chinese remainder theorem, 267269
results with MLP, 638645, 667 Choi-Williams dist.
674 BPSK analysis, 449, 452454
results with RBF, 642, 646648, classification, 634637
674681 comparison to WVD, 446
time-frequency, 620621 Costas analysis, 458, 461462
Index 851
demodulation, 400 noise analysis, 543, 545546

FMCW analysis, 449451 time-domain implementation, 516
hybrid analysis, 458, 463 polyphase analysis, 535, 537,
polyphase analysis, 455457 539541, 809820
polytime analysis, 455, 458460 polytime analysis, 540, 542543,
Classification, see Autonomous 821827
classification spectral correlation density,
CMRA altimeter, 43 515520
Coherent integration, 12
Combat losses, 557 D
Communications, 13
Compass Call, 556 Database, 638640
Complete residue, 716717 Database extended, 660666
Compound Barker code, 128129 Decision speed, 324
Compression ratio, 127 Decoy, 558559
Conical scan, 13 Demodulation, 400
Connectivity measure Detection,
extended, 333 homodyne, 1820
generalized, 326328 maximum range, 26
reference, 328329 range, 2426
Continuous waveform, 16 Detector
Correlation processor, 2124 envelope approximation, 84
coherence, 22 linear 29
for noise, 238243 square-law, 2829, 395
Fourier transform,21 Dierence triangle, 191195
Corvus ARM, 554555 Digital RF memory, 396
Cosmic Shield, 607 Digital-to-analog converter, 9293
Costas codes, 191195 Direct digital synthesizer, 20,
Critical frequency, 257 generating linear FM using,
Cyclostationary processing 9194,
BPSK analysis, 528531, 532534 RPC using, 9899
Costas analysis, 540, 544 Direct RF sampling, 398400
cycle frequency, 514 Distruction, 558
cycle frequency resolution, 519 Doppler,
cyclic autocorrelation, 514515 clutter spectrum, 259261
cyclic spectral analysis, 515 matrix correlation, 23
demodulation, 400 side lobe reduction, 23, 7578,
direct frequency smoothing, 110113
522525 zero, 22
FFT accumulation method, Down conversion, 397398
520522 Duty cycle, 1415
FMCW analysis, 531, 535537, Dwell time, 1113
805807
FMCW parameter extraction, E
699704
Frank analysis, 809812 Eagle, 4849
Grenanders uncertainty, 519 Eective radiated power, 18
Electronic attack range resolution, 189191

FMCW, 115 orthogonal codes, 370375
unifying principal, 334335 FSK see Frequency hopping
joint airborne, 556 FSK/PSK see Hybrid techniques
on information grid, 337338
on netted radar, 352360 G
Electronic protection,
unifying principal, 334335 Generalization, 630
Electronic warfare, 13 Gibbs phenomena, 474
receivers, 387388 Global information grid, 320
Electronic warfare ocer, 556 GOCFAR, 8485, 104
ELINT, 37 GRA-2000 altimeter, 44
Emitter clustering, 687688 Grating lobes, 14
Entropy, 333336 Growler EA-18G, 556
Espenschied, Lloyd, 41
Extinction coecient, 25 H
F Hamming window, 77, 447
Hann window, 77
Fan beam, 6 Harr fiter, 472473
Field of view, 10 HARD, 4849
resolution elements, 10 HG-9550 altimeter, 43
FM interrupted CW (OTHR), Homodyne detection, 18
282287 Human computer interface, 621623
FMCW radar Hybird techniques, 195198
advantages of, 8182 target matched, 199204
block diagram, 1820
modulation period, 102104 I
out-of-band emission, 270271
range-Doppler cross coupling, 102 Information rate, 334
received signal, 100101 Integration, 1820
waveform design, 8689, 9194 coherent, 28
waveform nonlinearities, 105106 LPI detection, 574576
waveform spectrum, 8991 netted radar, 348349
Fourier transform, 7, 18 noncoherent eciency, 29
FFT accumulation method, 520- noncoherent, 19, 20
522 postdetection, 8586
FMCW range profile, 84 Intercept receiver, 4
LPI detection, 576 challenges, 400402
Frame time, 1113 maximum range, 28
Frank code Interrupted CW, 16
example for PAF 7175 Ionospheric eects (HF), 253261
peak side lobe, 75
polyphase, 139, 143148 J
Frequency hopping
advantages, 187189 Jam-to-signal ratio
transmitted signal, 189 communication node, 337338
Index 853
radar node, 352353 M

Jamming, 556, 621
JY-17a, 5355 Main lobe, 5
MALD, 558559
K MALI, 59
Maneuverabilty,, 323324
Kh-31 ARM Markov chain, 85
seeker, 565566 Marginal frequency, 634638, 651656
Klipper (bandwidth), 28 Matched filter 22
Knowledge function, 335-336 Maximum usable frequency, 257
MIMO, 349352
L Missile systems, 5862
MMIC, 116
Landing systems, 4547 Modified feature extraction, 648660
PALS, 46 Modified sinc filter, 473474
TALS, 46 Modulation, 16
UCARS, 46 bandwidth (FMCW), 20
LANTIRN, 58 period (FMCW), 20, 102104
Lethality, 326 triangular,, 8691
Look-through, 388389 LPI, 3637
Low noise amplifier, 84 Moving target indication, 107108
Low probability of detection, 3 MRSR, 55
definition, 4 Multifunction aperture, 13
Low probability of intercept radar, Multilayer perceptron, 629632
antenna characteristics, 514 generalization, 630
definition, 3, 31
deramping detection of, 576 N
discriminator, 573574
multichannel detection, 574576 Narayanan noise radar, 215219
netted, 342 PAF, PACF, 219222
origin of, 29 Netted LPI radar, 342-345
transmitter characteristics, advantages, 346347
1418 LIPsimNet analysis, 353360
requirement, 4 MIMO, 349352
sensitivity definition, 3031 orthogonal codes, 362372
summary of characteristics, 18 signal-to-noise ratio, 348349
Lowpass filtering, 648651 signal model, 349352
Lookup table, 92 use of noise in, 374, 376377
LPI toolbox (LPIT), 709711 Network centric warfare
LPIsimNet, block diagram, 321
information network, 338345 definition, 320
radar network, 353360 entropy, 333-336
tutorial, 721739 global information grid, 320
information rate, 334
jam-to-signal ratio, 337338
LPIsimNet, 338345
metrics 326337
network reach, 329331 P3 code, 152, 156157

network richness, 333-336 P4 code, 157, 160162
receivers, 389391 PA-5429 altimeter, 44
requirements, 322 PAGE, 5152
sensor grid, 321 PALS landing system, 46
shooter grid, 322 PANDORA, 113114
Network reach, 329331, 336 Patriot, 607
Network richness, 333-336 Peak power, 14
Noise radar Perceptron
correlation receivers, 238243 single, 625628
MALI, 59 Perfect codes, 128
millimeter wave, 238 PACF, 69, 7475
Narayanan design, 215222 Periodic ambiguity, 6778
netted radar, 374, 376377 definition, 69
principles, 212215 generating results, 713714
random noise FMCW plus sine, periodicity, 70
227, 229234 Periodic autocorrelation, 6778
random noise plus FMCW, definition, 6869
222227 Pencil beam, 6
ultrawideband, 210212 Phased array, 67, 12
Phase code, 21
O advantages, 125126
binary, 128133
OLPI, 13, 179182 period, 127
OODA, 324325, 326 rate, 127
Operational tempo, 324 range resolution, 127
maximum, 336337 transmitted signal, 126127
Origin of LPI, 29 Pilot radar, 3136
Orthogonal waveforms, 358, 361362 technical characteristics, 33
frequency hopping, 370375 sensitivity, 34
MIMO, 350352 Platform centric, 319320
OTHR, 377378 Pointer, 5051
polyphase codes, 362370 Polarization, 6
OTHR systems Polyphase codes, 134162
Doppler spectrum, 259261 Barker, 134142
ionosphere, 253261 Frank, 139, 143148
netted configurations, 377378 orthogonal, 362372
sky wave, 252280 P1, 148151
sky wave waveforms, 265271 P2, 152155
surface wave, 276, 281294 P3, 152, 156157
surface wave waveforms, 282288 P4, 157, 160162
OTHR waveforms, 249251 Polytime codes, 163178
T1(n), 163165
P T2(n), 165, 168171
T3(n), 169, 172174
P1 code, 148151 T4(n), 169, 175178
P2 code, 152155 Position flexibility, 606
Index 855
Power two channel analysis, 474476

attenuation coecient, 25 wavelet decomposition, 468
average, 15 wavelet filters, 472474
density, 2425 wavelet transform, 469471
peak, 15 Quiet radar, 30
received, 25 Quiet naval radar CRM-100, 53
Power management, 1618
in seeker, 17 R
using constant SNR, 310312
Prime number, 715 Radar
Primitive roots, 717719 pulsed conventional, 14, 35, 42
Principal components analysis, 648, warning receiver, 37
656660 Radar cross section
Probability of detection, 19 free space backscattering, 291
Probability of false alarm, 19 low values, 306307
Processing gain, 19 microwave 290
FMCW, 20 Shearmans definition, 291
intercept receiver, 2829 Radial basis function, 624, 632633
phase code, 20 Euclidean norm, 632
random binary phase modulation, Gaussian basis, 633
211 Radon transform, 689692
PSK see polyphase code Random binary phase, 234237
Pulse compression, 1516, 1824 PAF, PACF, 236237
Pulse descriptor words, 396
Pulsed radar, 14 Random noise plus FMCW, 222227
PAF, PACF, 225228
Q Random noise FMCW plus sine, 227,
229234
Quadrature mirror filtering PAF, PACF, 230234
BPSK analysis, 489494, 773776 Range
complex input example, 482487 detection, 2426
Costas analysis, 499, 502503 interception, 2729
demodulation, 400 maximum detection, 26, 30, 32
FMCW analysis, 487489, 771 maximum interception, 2728, 30
772 resolution, 102
Frank analysis, 777780 RBS-15, 5860
Harr filter, 472473 Receiver (intercept)
hybrid analysis, 499, 504505 challenges with, 400402
noise analysis, 499, 506508 sensitivity, 2728
polyphase analysis, 494498, 777 Receiver (radar)
780, 781795 bandwidth, 26
polytime analysis, 495, 499501, comparison, 392396
797803 correlation, 23, 238243
polyphase parameter extraction, matched, 22, 108110
695699 mismatched, 24, 7578, 110113
short-time Fourier transform, 469 noise factor, 26
tree structure, 476482 sensitivity, 2526
Reduced residue, 716717 output (radar), 26

Reflected power canceler, 83 Pilots required, 35
Reference signal, 2124 processing gain, 19
Regularization, 630631 time-bandwidth, related to 19
Tikhonovs parameter, 631 Sinc filter, 473474
Resolution Single sideband modulator, 9294
cycle frequency, 519 Situational awareness, 323
FMCW range, 102 Sky wave OTHR, 249251
sky wave OTHR, 264 critical frequency, 257
phase code range, 127 detection range, 271276
velocity, 102 distance coverage 259
footprint, 274276
S Doppler clutter, 259261
ionosphere eects, 253259
Samples per subcode, 73 JORN, 261263
Scan pattern, 11-13 LPI considerations, 265271
confusion, 13 maximum usable frequency, 257
raster scan, 11 waveform repetition frequency,
Sea clutter, 308310 266269
Sea state, 308 SMART-L, 11
Search mode processing Spearfish, 61
FMCW, 101105 Spiral antenna, 559566
Seekers Archimedean, 563564
airborne, 5861 equiangular, 560562
torpedo, 6162 logarithmic, 559
Sensitivity conical, 559560
comparison, 29 conical equiangular, 564565
intercept receiver, 2728 STC, 82
Pilots, 35 Subcode, 20
radar, 2526 number of, 20
receiver comparison, 395396 period, 2022
Sensor grid, 321 Superconductor, 118119
Sheridan levels, 622623 Suppression
Short-time Fourier transform, 469 beginnings, 553
in cyclic spectrum estimation, definitions, 552553
518519 Eagle example, 331-333
Side lobes, 510 enemy integrated air defense,
ACF, PACF, PAF, 7071 551553
definitions of peak, 7071 look-through, 388389
OTHR 265266 Surface wave OTHR, 249251, 276,
Side lobe ratio, 5 278, 281
Taylor, 910 detection range, 288294
Sigmoid, 625626, 630 LPI considerations, 282288
Signal-to-noise ratio, SWR 503, 281283
Albersheim relation, 19 Surveillance systems
input required (radar), 26 airborne 5658
for netted radar, 348349 ground based, 4855
Index 857
Swarm, 391392 V
Swedish GLV200, 607
Swiss Air Guard, 607 Variant, 5253
Vietnam, 555556
T
W
TALS landing system, 46
Taper, 79, 12 Wavelet filters, 472474
Tapped delay line, 2224 Wavelet transform, 469471
Target-to-clutter ratio, 312315 Weighting
Taylor distribution, 810, 12 Hamming window, 77, 104
Temperature, 26 Hann window, 77
Time-bandwidth, Welsh construction (Costas), 193195
FMCW, 19 Wiener filter, 692
Time-frequency processing Wigner-Ville dist. (WVD), 405442
Choi-Williams distribution, BPSK analysis, 421426
445463 comparing polyphase results,
Gabor distribution, 574576 431433
quadrature mirror filtering, complex input example, 411414
467509 continuous 1-D, 406
Wigner-Ville distribution, discrete WVD, 407
405442 demodulation, 400
Time-on-target, 1113 FMCW analysis, 419421,
Time smoothing, 520522 741743
Track mode processing Frank analysis, 745747
FMCW, 104105 FSK, FSK/PSK results, 438441
Training, 631 kernel generation, 408
Transmission line, 9596 operator (ELINT), 442
Transmit, polyphase analysis, 426428,
multiple simultaneous, 13 745747, 749758
peak power, 14 polyphase parameter extraction,
power management, 16 688695
Transmitter polytime analysis, 429437,
continuous waveform, 15 759769
solid state, 15 pseudo WVD, 407
Twinkle transmission, 607 real input example, 409411
two-tone input example, 414418
U Wild Weasel, 555, 573, 597
UCARS landing system, 46 Y

Ultra-low (side lobes), 8
Uniform window, 77 YGBSM, 551
Unmanned aerial vehicle, 391392,
595596
Recent Titles in the Artech House
Radar Series
David K. Barton, Series Editor
Adaptive Antennas and Phased Arrays for Radar and

Communications, Alan J. Fenn
Advanced Techniques for Digital Receivers, Phillip E. Pace
Advances in Direction-of-Arrival Estimation, Sathish Chandran,
editor
Airborne Pulsed Doppler Radar, Second Edition, Guy V. Morris and
Linda Harkness, editors
Bayesian Multiple Target Tracking, Lawrence D. Stone,
Carl A. Barlow, and Thomas L. Corwin
Beyond the Kalman Filter: Particle Filters for Tracking Applications,
Branko Ristic, Sanjeev Arulampalam, and Neil Gordon
Computer Simulation of Aerial Target Radar Scattering,
Recognition, Detection, and Tracking, Yakov D. Shirman, editor
Design and Analysis of Modern Tracking Systems, Samuel Blackman
and Robert Popoli
Detecting and Classifying Low Probability of Intercept Radar, Second
Edition, Phillip E. Pace
Digital Techniques for Wideband Receivers, Second Edition,
James Tsui
Electronic Intelligence: The Analysis of Radar Signals, Second
Edition, Richard G. Wiley
Electronic Warfare in the Information Age, D. Curtis Schleher
ELINT: The Interception and Analysis of Radar Signals,
Richard G. Wiley
EW 101: A First Course in Electronic Warfare, David Adamy
EW 102: A Second Course in Electronic Warfare, David L. Adamy
EW 103: Tactical Battlefield Communications Electronic Warfare,
David Adamy
Fourier Transforms in Radar and Signal Processing,
David Brandwood
Fundamentals of Electronic Warfare, Sergei A. Vakin, Lev N. Shustov,
and Robert H. Dunwell
Fundamentals of Short-Range FM Radar, Igor V. Komarov and
Sergey M. Smolskiy
Handbook of Computer Simulation in Radio Engineering,
Communications, and Radar, Sergey A. Leonov and
Alexander I. Leonov
High-Resolution Radar, Second Edition, Donald R. Wehner
Introduction to Electronic Defense Systems, Second Edition,
Filippo Neri
Introduction to Electronic Warfare, D. Curtis Schleher
Introduction to Electronic Warfare Modeling and Simulation,
David L. Adamy
Introduction to RF Equipment and System Design, Pekka Eskelinen
Microwave Radar: Imaging and Advanced Concepts,
Roger J. Sullivan
Millimeter-Wave Radar Targets and Clutter, Gennadiy P. Kulemin
Modern Radar Systems, Second Edition, Hamish Meikle
Modern Radar System Analysis, David K. Barton
Modern Radar System Analysis Software and User's Manual, Version
3.0, David K. Barton
Multitarget-Multisensor Tracking: Applications and Advances
Volume III, Yaakov Bar-Shalom and William Dale Blair, editors
Principles of High-Resolution Radar, August W. Rihaczek
Principles of Radar and Sonar Signal Processing,
Franois Le Chevalier
Radar Cross Section, Second Edition, Eugene F. Knott et al.
Radar Evaluation Handbook, David K. Barton et al.
Radar Meteorology, Henri Sauvageot
Radar Reflectivity of Land and Sea, Third Edition, Maurice W. Long
Radar Resolution and Complex-Image Analysis, August W. Rihaczek
and Stephen J. Hershkowitz
Radar Signal Processing and Adaptive Systems, Ramon Nitzberg
Radar System Analysis, Design, and Simulation, Eyung W. Kang
Radar System Analysis and Modeling, David K. Barton
Radar System Performance Modeling, Second Edition,
G. Richard Curry
Radar Technology Encyclopedia, David K. Barton and
Sergey A. Leonov, editors
Range-Doppler Radar Imaging and Motion Compensation,
Jae Sok Son et al.
Signal Detection and Estimation, Second Edition, Mourad Barkat
Space-Time Adaptive Processing for Radar, J. R. Guerci
Theory and Practice of Radar Target Identification, August W.
Rihaczek and Stephen J. Hershkowitz
Time-Frequency Transforms for Radar Imaging and Signal Analysis,
Victor C. Chen and Hao Ling
For further information on these and other Artech House titles, includ-
ing previously considered out-of-print books now available through our
In-Print-Forever (IPF)
program, contact:
Artech House Artech House

685 Canton Street 46 Gillingham Street
Norwood, MA 02062 London SW1V 1AH UK
Phone: 781-769-9750 Phone: +44 (0)20 7596-8750
Fax: 781-769-6334 Fax: +44 (0)20 7630-0166
e-mail: artech@artechhouse.com e-mail: artech-uk@artechhouse.com
Find us on the World Wide Web at: www.artechhouse.com

Philip E. Pace-Detecting and Classifying Low Probability of Intercept Radar (2009) PDF

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Philip E. Pace-Detecting and Classifying Low Probability of Intercept Radar (2009) PDF

Hochgeladen von

Copyright:

Verfügbare Formate

Detecting and Classifying

Low Probability of Intercept Radar

For a listing of recent titles in the

British Library Cataloguing in Publication Data

Cover design by Igor Valdman

2009 ARTECH HOUSE

PART I: FUNDAMENTALS OF LPI RADAR DESIGN 1

1 To See and Not Be Seen 3

2 LPI Technology and Applications 41

2.2.2 Fielded LPI Landing Systems . . . . . . . . . . . . . . 46

3 Ambiguity Analysis of LPI Waveforms 67

4.12 PANDORA FMCW Radar . . . . . . . . . . . . . . . . . . . 113

5 Phase Shift Keying Techniques 125

6 Frequency Shift Keying Techniques 187

7 Noise Techniques 207

8 Over-the-Horizon Radar 249

8.6 Surface Wave OTHR . . . . . . . . . . . . . . . . . . . . . . . 276

9 Case Study: Antiship LPI Missile Seeker 301

10 Network-Centric Warfare and Netted LPI Radar Systems 319

10.5.3 Signal Model . . . . . . . . . . . . . . . . . . . . . . . 349

PART II: INTERCEPT RECEIVER STRATEGIES

11 Strategies for Intercepting LPI Radar Signals 387

12 Wigner-Ville Distribution Analysis of LPI Radar

12.5 Polytime Code Analysis . . . . . . . . . . . . . . . . . . . . . 429

13 Choi-Williams Distribution Analysis of LPI Radar

14 LPI Radar Analysis Using Quadrature Mirror Filtering 467

15 Cyclostationary Spectral Analysis for Detection of LPI

16 Antiradiation Missiles 551

16.4.10 AS-17 Krypton . . . . . . . . . . . . . . . . . . . . . . 587

17 Autonomous Classification of LPI Radar Modulations 619

17.5.3 Principal Components Analysis . . . . . . . . . . . . . 656

18 Autonomous Extraction of Modulation Parameters 687

A Low Probability of Intercept Toolbox 709

B Generating PAF Plots Using the LPIT Files 713

C Primitive Roots and Costas Sequences 715

D.4 Adding a Moving Jammer to the Scenario . . . . . . . . . . . 731

E PWVD for FMCW with F = 500 Hz 741

F PWVD for Frank Code with T = 64 ms 745

G PWVD Results for P1, P2, P3, and P4 Codes 749

H PWVD Results for Polytime Codes T2, T3, and T4 759

I QMFB Results for FMCW with F = 500 Hz 771

J QMFB Results for 11-Bit BPSK 773

K QMFB Results for Frank Signal with Nc = 16 777

L QMFB Results for P1, P2, P3, and P4 781

M QMFB Results for T2, T3, and T4 797

N Cyclostationary Processing Results with FMCW,

O Cyclostationary Processing Results with Frank Signal,

P Cyclostationary Processing Results for P1, P2, P3,

Q Cyclostationary Processing Results for T2, T3, and T4

List of Symbols 829

About the Author 847

In the foreword of Detecting and Classifying Low Probability of Intercept

pulse. Antiradiation missiles and other approaches to suppression of enemy

Richard G. Wiley, Ph.D.

The second edition of Detecting and Classifying Low Probability of Intercept

Overview of the Book

In Chapter 1, an introduction to LPI radar is presented which provides the

In Chapter 5, phase shift keying (PSK) LPI radar is discussed. Details

To begin Part II, Chapter 11 takes a look at (noncooperative) digital intercept

Chapter 14 investigates the use of quadrature mirror filter banks (QMFBs)

FUNDAMENTALS OF LPI RADAR DESIGN

To See and Not Be Seen

1.1 The Requirement for LPI