Beruflich Dokumente
Kultur Dokumente
Dear colleagues,
I wish you a warm welcome to the sixth edition of the Symposium of Electronics and
Telecommunications.
I am delighted to announce you that, from its first edition of 1994, a continuous increase in both
quantity and quality of papers, submitted by well known researchers from universities and
industry, from Romania and abroad, can be clearly identified. With its 13 sections and over 170
papers, the Symposium Proceedings cover wide areas in its field.
Besides the scientific and informative value of the published volumes, another purpose of our
Symposium is to be, for all participants, a forum for exchange of ideas and socialization, for
emphasizing the feeling of belonging to a highly specialized and expert community.
Although research results are the focus of a scientific conference, we cannot forget that higher
education is in a process of profound transformations. Therefore, we took advantage of the
presence of an important number of high ranked personalities from universities to organize a round
table on the Impact of the Bologna Declaration on Electronics and Telecommunications
Education. We hope that the ideas exposed by the participants and the common conclusions will
constitute guidelines for our future common action.
Our region has been target of important investments in the last 14 years, which led to an
important industrial development in the field of electronics and telecommunications. Recently we
have celebrated the graduation of the 30th generation from our faculty. What these two facts have
in common is the highly trained human resource exiting our faculty and entering the industry.
Fortunately, a feedback mechanism has been created, as industrials understood the necessity of
supporting education. This symposium would not have been possible without the gracious help of
our sponsors, to whom we express our gratitude.
I am also honored to express my thanks to you, all the participants, for attending the Symposium,
to wish you a successful and profitable audience through the sessions, and a nice stay in Timisoara.
I am looking forward of meeting you again in 2006.
Dean,
Prof. Dr. Eng. Marius Otesteanu
Table of Contents
S. V. Halunga, O. Fratu
New interleaver design algorithms with enhanced B.E.R. performances.. 19
R. Radescu, R. Popa
On the performances of symbol ranking text compression method... 25
R. Stoian, L. A. Perisoara
Application of turbo principle to product codes 28
H. Balta, M. Kovaci
A study on turbo decoding iterative algorithms. 33
H. Balta, M. Kovaci
The Performances of Convolutional Codes used in Turbo Codes. 38
S. Popescu
A new approach on delay coding: the receiver.. 44
G. G. Fericean, M. Borda
Selective encryption of image with IDEA algorithm.. 50
L. Scripcariu, P. Duma
Analysis of simple inversable functions defined on Galois fields for cryptography use 55
R. Stoian, L. A. Perisoara
Parallel concatenated convolutional turbo codes: performance analysis for different interleaving
schemes.. 60
R. Radescu, I. Balasan
Evaluation of parameters used in lossless text compression with the Burrows-Wheeler transform. 65
Signal Processing
C. Partheniu
Gradient algorithms with improved convergence.. 75
1
E. Szopos, N. Toma, M. Topa
Adaptive filtering algorithms. 81
C. Chioncel, J. Gal
Parameter estimation of the chirp signal... 87
G. Budura, C. Botoca
Efficient implementation of the second order Volterra filter. 91
L. Grama
Phase approximation using signals affected by random perturbations. 96
L. Stanciu, L. Banu
Reconstruction methods for missing portions in signals... 102
D. Tarniceriu, V. Munteanu
Information theoretic approach of filterbanks performing energy compaction 106
M. A. Matin
Invisible watermarking for Copyright Protection.. 114
C. Nafornita
A wavelet-based watermarking for still images. 126
A. Vlad, A. Luca
The statistical behaviour of the chaotic signals: application to cryptography.. 132
R. Arsinte, C. Ilioaei
Considerations and results in Multimedia and DVB application development on Philips Nexperia
Platform. 138
I. G. Mocanu
Shape similarity measure for k nearest-neighbor queries. 142
I. G. Mocanu
Shape representation and retrieval using centroid radii and turning angle. 146
Microcontrollers
P. Duma
Software setting telephone links using ATMEL microcontrollers in time switching network PABX 154
S. Zoican
Variable step size affine projection adaptive algorithm implementation.. 160
2
M. Otesteanu, D. Criste
Precision electronic driver for pneumatic engines 175
P. Duma, L. Scripcariu
Development system equipped with AT89S8252 microcontroller. 187
T. Ionica, C. Balint
Telephone interface for remote control systems with network capabilities... 197
M. S. Crainic
Prepayment gas meter a new trends in natural gas metering technology.. 201
D. Belega
Accurate sinewaves implemented with a 16-bit fixed-point digital signal processor 225
S. Mischie
On frequency measurement by using zero crossings. 230
S. C. Ionel
A correlation analysis of measured CO-concentration signals. 240
S. Mereuta
Analysis of biomolecular sequences through spectral based methods.. 244
C. I. Dumitrescu
K - complex Detection using the Continuous Wavelet Transform. 247
3
D. Stoiciu, M. Lascu
PC-based system for automated calibration of a digital voltmeter... 253
A. Ignea, A. Mihaiut
The measurement of dc magnetic field... 255
A. Rosu-Niculescu, T. Petrescu
The simulation of the effect of the geometry of the Rectangular Double Barrier structure about the
transmission and reflection coefficient.. 259
S. Simion
MEMS based broadband phase shifters 265
F. Toadere
Optical coherent and incoherent systems frequency analyze in Cartesian coordinate. 271
F. Toadere
Cartesian coordinate optical filter analyses.. 275
F. Toadere
Space and frequency analyze of an LSI optical system with different input signals.. 280
E. Teodoru , S. Demeter
A discrete model for reference source noise in indirect frequency synthesis 300
V. Tiponut
A Physical Laboratory for Smart Transducers Education. 305
G. Sirbu, D. Aiordachioaie
On radio spectrum measurements with the ESVB Rohde & Schwarz test receiver... 309
G. Sirbu, D. Aiordachioaie
On GSM mobile phone measurements with the CTS-65 Rohde & Schwarz digital radio tester... 313
A. De Sabata, L. Matekovits
Scattering parameters of symmetrical networks 317
4
G. Oltean, E. Sipos, I. Oltean
A new approach of op-amp amplifier biasing 328
S. V. Tiponut
A toolkit for internet based distance laboratory development... 332
R. Popa
A complete laboratory on evolutionary electronics... 335
S. Ionel, M. Daneti
Low-cost electronic board improves electronics laboratory efficiency. 346
D. Stoiciu, C. Dughir
A web-based teaching tool for laboratory classes. 348
Wireless Communications
D. M. Dobrea, N. Cleju, A. T. Sechelea, A. Banar
Mobile accident warning system -The LoRD- 354
A. A. Enescu, S. Ciochina
An improved MIMO-OFDM channel estimator in the tracking phase.. 360
A. F. Paun, S. G. Obreja
A low complexity decision feedback equalization for sparse wireless channels... 370
S. G. Obreja, A. F. Paun
Terestrial digital video broadcasting (DVB-T). System performances simulation 378
I. I. Duma
Noise impulse generation with convenient characteristics in time and frequency domain... 398
M. Moise
Mobility concept for wireless ATM networks 409
5
M. Moise
Lossless handover scheme for Mobile ATM networks... 415
A. Sikora
Design rules for lightweight short-range wireless networks. 421
6
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Technical University Cluj-Napoca, Faculty of Electronics and Telecommunications, Communications Department,
str. G. Baritiu Nr.25, 400427, Cluj-Napoca Romania; e-mail: Vasile.Bota@com.utcluj.ro
7
Hvt = 0 (5) group of N bits v that is checked by means of
syndrome-computation; if the syndrome equals zero,
This approach has two major shortcomings: the algorithm considers v to be the correct codeword;
- for great values of parameters j and/or p, C otherwise it performs another iteration adjusting the
becomes large implying a significant computational values of the a posteriori probabilities by using some
load that increases the processing time and/or the internal values computed in the previous iteration.
hardware required by the implementation; The maximum number of iterations allowed, B, is a
- it requires all information bits il, l = 0,..., (k-j)p-l parameter of the algorithm. The values of the a
at the same time; this requirement induces a one- posteriori probabilities are previously extracted from
codeword additional latency in the system. the OFDM-demodulated coordinates I and Q of each
These shortcomings may be avoided by a simpler and QAM-symbol, by means of the soft-demapping
faster encoding method, described in [5].
procedure.
E. Bit-Mapping on the QAM Signal Constellation This algorithm does not search for the closest
codeword compared to the received sequence, but
When the LDPC-coded bits are to be modulated on a tries to correct every bit. Due to this property, the
b-bit/symbol QAM constellation, the b-tuple is number of error bits after the decoding is always
mapped on the I and Q coordinates of the QAM smaller than the one of error bits prior the decoding,
vector, by splitting the b-tuple into two groups of b/2 when the algorithm is convergent. Extensive
bits, each group being assigned to one axis. The bits simulation performed by the authors confirmed this
that are assigned to an axis are mapped to the property, which might lead to the decrease of error-
amplitude levels of that axis according to a Gray packet length that should be corrected by the RS code
encoding [1]. Since the transmission bit-loading might that follows the LDPC or convolutional codes in
involve non-coded information bits, they are also many applications.
mapped according to a separate Gray encoding, in
order to maximize the distance between levels having G. Soft-Demapping
the same non-coded bits. Therefore, the multibit Because the MP algorithm requires the a posteriori
assigned to an axis of the QAM constellation, coded
probabilities of each bit and the received vector
and non-coded bits, is mapped according to a 2-level
carries more bits, a soft-demapping [3] is required in
Gray encoding described in [3]. Fig. 1 presents an
order to provide the Fn0(0/r) and Fn1(1/r) probabilities
example of mapping b/2 = 4 bits on one axis (I or Q)
of a QAM constellation. of each bit mapped the received vector.
For multibit/symbol modulations, the two probabilities
-15 -13 -11 -9 -7 -5 -3 -1 1 3 5 7 9 11 13 15 of each bit are extracted, from the received level on the
I or Q branches, by using (7) that gives the probability
0000 0001 0011 0010 0100 0101 0111 0110 1100 1101 1111 1110 1000 1001 1011 1010
of bit bj to be 1 when the demodulated level on a
Coded bits Non-coded bits
branch equals r and the channel is AWGN [1]:
Distance between groups, dg Distance within a group dig
2b / 2
(r L(l)) 2
Fig. 1. Bit mapping on a QAM-constellation axis using the 2-level
Gray encoding
exp( 2 2
) b lj
(7)
Fj1 = l=12b / 2 ; j = 0,..., b / 2 1;
(r L(l)) 2
The amplitude levels employed on each axis belong to exp(
2 2
)
the set A defined by: l =1
A = {Al = 2l-(Lb-1); l = 0,1,Lb-1;}; Lb = 2b/2; (6) In (7) blj denotes the logical value of j-th bit of the l-th
modulating level of the I or Q branch of the
This bit-mapping method, which allows only for the demodulated vector. A similar expression is derived
employment of the square QAM constellations, is for Fj0 and the two values are normalized to their sum.
simpler because the mapping is identical on both The soft-demapping requires a previous estimation of
coordinates. the noise variance ; computer simulations run by the
F. Decoding the LDPC Codes authors showed that estimation errors of less than 2 dB,
between the actual channel noise variance and the one
The decoding of the LDPC codes is accomplished by stored in the soft-demapper, lead to insignificant
using the message-passing algorithm (MP), as decreases of the decoder performances.
presented in [2], which will not be described here.
This algorithm, based on the Bayes criterion, requires H. Soft Decision of the Non-Coded Information Bits
the previous computation of the a posteriori
probabilities for every bit of a codeword, Fn0(r/0) and The information non-coded bits mapped on a QAM
Fn1(r/1), where r denotes the received vector, n is the symbol can be decided by two methods, namely:
bit index and 0/1 denote the bit logical value. - hard decision, applying the Bayes criterion to the
Basically, the algorithm performs the decoding of a probabilities provided by the soft-demapping; this
codeword by using the a posteriori probabilities of a method does not employ the information provided by
8
decoding the coded bits placed on the same tone It displays the BER values and the BER vs. SNR
during the same symbol period. characteristic for the selected SNR range, the number
- soft decision, that considers the information of coded bits error after the decoding of each
provided by the decoding of the coded bits mapped on codeword, and the number of non-coded bits decided
the same QAM symbol and tone. by soft-demapping. It also displays the throughput of
Basically, the optimal decision memorizes the the transmission for a defined packet dimension.
received level r and, using the decoded bits provided The simulations were performed on a test of 106
by the LDPC decoder, selects the closest (in the dE information bits and the maximum number of B = 15
sense) level that was mapped with the same decoded iterations/codeword for the decoding algorithm.
bits, see fig. 2. This method provides lower BER of
B. Effects of the Coding Rate upon the BER
the non-coded bits, as resulted from simulations Performances of LDPC-Coded QAM Constellations
performed by the authors, but may error the non-
coded bits if the corresponding coded bits were As shown in (2) the coding rate might be changed,
wrongly decoded. without changing the codeword length, by changing
PL PL OD r HD PL PL the parameter j. Considering k=14 and p = 31, so that
-15 -13 -11 -9 -7 -5 -3 -1 1 3 5 7 9 11 13 15 a short codeword N = 434 bits (see (2)) is used, a
0000 0001 0011 0010 0100 0101 0111 0110 1100 1101 1111 1110 1000 1001 1011 1010
family F1 of LDPC codes with RC ranging from 0.78
Coded bits PL possible levels for 10 coded bits to 0.21 is displayed in table 1. The family includes the
Non-coded bits
OD optimal decision non-coded configuration for comparison.
HD hard (Bayes) decision
Distance between groups, dg Distance within a group dig
Table 1. Parameters of F1 LDPC codes; k = 14, p = 31
Fig. 2. Optimal decision of non-coded bits for 10 coded bits
F1 j N C R
I. Bit-loading on Coded-QAM OFDM Employing C11 Non-coded 1
Non-Coded Bits
C12 3 434 93 0.78
Denoting by T the number of available sub-carriers,
by DOFDM the OFDM symbolrate, by Nci and Nni the C13 7 434 217 0.50
numbers of coded and non-coded bits on the i-th C14 9 434 279 0.35
subcarrier and by RC the LDPC code rate, the nominal
payload Dn of the OFDM transmission would be (8). C15 11 434 341 0.21
The number of bits actually carried by each QAM-
symbol equals Nc + Nni. In order to evaluate the correction capability of these
codes, the BER vs. SNR performances were evaluated
D n = D OFDM T ( N ci R C + N ni ); (8) employing a 2-PSK constellation. And are shown in
fig. 3.
The rate of the coded QAM-modulation is computed
by:
N ci R C + N ni ; (9)
R CM =
N ci + N ni
9
Taking into account the fact that the MP decoder has C. Effects of the Codeword Length upon the BER
about the same implementation complexity as the 64- Performances of LDPC-Coded QAM Constellations
state Viterbi decoder we may say that the LDPC
codes provide about the same coding gain as the In order to evaluate the effects of the codeword length
convolutional code, at a higher coding rate. upon its performances we consider a family of codes
Comparing the coding gains with the ones of the turbo F2, see table 4, with RC = 0.214 (as code C15 of F1).
codes, the RC = LDPC code ensures BER =10-6 at The code word length N is modified by means of
a SNR = 2 dB, about 1 dB higher than the turbo-codes parameter p, ranging from 23 to 73.
[6], but it requires only one MAP decoder instead of Table 4. Codes of rate Rc = 0.214 and various
two decoders (MAP + MLD) and a de-interleaver. codeword length N family F2
By decreasing the code rate by changing the code Code k j p N [bits]
parameter j additional coding gains of up to 3.5 dB C151 14 11 23 322
can be obtained. The code C15 (RC = 0.21) requires a C151 14 11 31 434
SNR = 1dB to provide a BER = 10-6. C153 14 11 43 602
In order to evaluate how close these codes are to the C154 14 11 53 742
Shannon limit, we recall that the maximum spectral
C155 14 11 73 1022
efficiency wMi, which could be provided by a code of
rate Rci in an error-free transmission across an fs The BER vs. SNR performances of these codes are
bandwidth, is: displayed in fig.4.
1 S
wMi = lg 2 (1 + R ci ) [bps/Hz]; (10)
2 N
Assuming that the error-free transmission is
accomplished for SNRs > SNR0i, where:
BER (SNR0i) = 110-5; (11)
we may compute the maximum spectral efficiency,
wM0i, that could be accomplished by a code with rate
RCi, by using (10) and SNR0i obtained by simulations Fig. 4 BER vs.SNR of LDPC codes from table 4
(see fig.3). This value, compared to the actual spectral
efficiency wi, provided by the code (which equals The results in fig.3 indicate that by increasing the
Rci) indicates how far is the code from the theoretical codeword length, extra coding gain, can be provided
limit. at the same coding rate, but at the expense of a more
Another evaluation can be performed by computing difficult implementation. The extra coding gain may
the minimum signal/noise ratio, SNRmi, for which wi be as high as 1.5 dB, see codes C152 and C155 in fig.4
could be obtained. The difference SNRi = |SNRmi - on one hand and code C15 in fig.3 on the other. The
SNR0i| shows the quality of the code in terms of the code C155 provides a total 11 dB coding gain.
SNR. Increasing the codeword length would also bring the
The values of these parameters for the codes of table 1 code performances closer to the Shannon limit.
are shown in table 3 Computing the parameters defined in table 3 for the
codes of family F2 we get the values of table 5.
Table 3. Ideal and actual performances for codes of
family F1 Table 5. Ideal and actual performances for codes of
Code wi wMi SNR0i SNRmi SNRi family F2
Ci [bps/Hz] [bps/Hz] [dB] [dB] [dB] Code wi wMi SNR0i SNRmi SNRi
C12 0.785 1.59 4.1 -0.35 4.45 [bps/Hz] [bps/Hz] [dB] [dB] [dB]
C13 0.5 0.80 1.7 -0.82 2.52 C151 0.214 0.352 0.5 -1.26 1.76
C14 0.357 0.546 1.1 -1.04 2.14 C152 0.214 0.291 0.2 -1.26 1.46
C15 0.214 0.29 0.1 -1.26 1.36 C153 0.214 0.280 0.0 -1.26 1.26
C154 0.214 0.274 -0.1 -1.26 1.16
As expected, codes of a certain length come closer to
the theoretical limits as their rate decreases. Results of C155 0.214 0.268 -0.25 -1.26 1.01
tables 2 and 3 show that rather short codes, easy to The longest code of family F2, C155, is about 1 dB
implement, are close enough to the theoretical maxi- away from the Shannon limit, for BER < 110-5.
mum performances. Simulations performed by the authors showed that the
The performances of the LDPC codes might be coding gains provided by the LDPC-coded QAM
improved by increasing the maximum number of constellations are about the same, compared to the
iterations/codeword of the MP decoder; simulations same non-coded QAM constellations, regardless the
showed that increasing B = 25, leads to extra coding constellation employed.
gains of 0.5-1 dB, at the expense of a longer proces-
sing time required.
10
III. LDPC-CODED QAM CONFIGURATIONS curves of the coded and non-coded bits before the MP
EMPLOYING NON-CODED BITS decoding (for the coded bits) and before the soft
decision (for the non-coded bits), and after these
The employment of error-correcting codes leads to a
decoders. The BER prior to decoding was obtained
decrease of throughput provided by the coded QAM
constellation. This occurs due to the control bits by using a hard Bayes decision that employs the a
inserted by the code, and the throughput decrease is posteriori probabilities provided by the soft-
higher for low-rate codes that secure a good demapping. The curves correspond to configuration
correction capability. no.3 from table 6.
In order to reach a reasonable trade-off between the
correction capability and the coding rate, which 1
BER
affects significantly the throughput, each QAM
symbol is loaded with nci coded bits and with nni non- 2
coded bits. 4 3
Considering a LDPC code with an RC coding rate the SNR
coding rate of the configuration employing coded and
non-coded bits is expressed by (9). Fig.6. BER vs. SNR of the coded and non-coded bits of
The coded and non-code bits are mapped on the I and configuration 3 from table 6; line 1 coded bits Bayes decision;
Q coordinates (6) of the QAM symbol using the 2- line 2 non-coded bits Bayes decision; line 3 coded bits MP
decision; line 4 non-coded bits soft decision
level Gray mapping, see fig.1, on each axis.
A. BER Performances of the LDPC-Coded QAM As shown in figure 6, the non-coded bits have lower
Configurations Employing Non-Coded Bits BER than the coded bits loaded on the same QAM
symbols, both before their decoding (line 1 vs. line 2)
To evaluate the effects of employing non-coded bits, a and after it (line 3 vs. line 4). This is due to the 2-level
family F3 of possible LDPC-coded configurations Gray mapping of the coded and non-coded bits.
using non-coded bits based on the 256-QAM (Nci + Fig. 6 shows that the number of error bits is always
Nni = 8 bits) are presented in table 6 together with smaller after the decoding process, than before it.
their coding rates RCM and coding gains CG. The This, combined with additional simulations performed
LDPC code employed is (k = 14, j = 3, p = 31; RC = by the authors indicate that the two decoders might
0.78), which provided a 6.5 dB coding gain on a 2- require some smaller outer codes (small RS or even
PSK modulation (code C12 in table 1 and fig. 3). The BCH) in FEC schemes employing concatenated
BER vs. SNR performances of this family are codes.
presented in fig. 5. The spectral efficiencies of the configurations
employing non-coded bits are higher than the ones of
Table 6. Coded QAM Configurations of Family F3
the proximity to the Shannon limit of the
Cfg. Nci Nni RCM CG configurations from family F3 are displayed in table 7.
1 0 8 1 -
2 2 6 0.945 5 Table 7. Ideal and actual performances of
3 4 4 0.890 6 coded configurations from family F3
4 6 2 0.835 6.5 wi wMi SNR0i SNRmi SNRi
Cfg RCMi
5 8 0 0.780 7 [bps/Hz] [bps/Hz] [dB] [dB] [dB]
1 1 8 10.46 31.5 24 7.5
2 0.945 7.52 9.05 27.5 22.9 4.6
3 0.890 7.12 8.64 26.5 21.9 4.6
4 0.835 6.68 8.21 25.5 20.9 4.6
5 0.780 6.24 8.31 25.0 19.8 5.2
The spectral efficiencies of the coded configurations
carrying non-coded bits (lines 2, 3, 4 in table 7) are
higher than the ones of the coded configuration with
no non-coded bits (line 5). Also they are closer, in the
Fig. 5. BER vs. SNR for configurations of F3 defined in table 6
SNR sense, to the Shannon limit.
As shown in table 6, the employment of non-coded
C. Throughput Performances of the LDPC-Coded
bits leads to a significantly increase of the coding rate,
QAM Configurations Employing Non-Coded Bits
at the expense of a coding gain decrease of about 1-2
dB. The coding rate increases starting from 0.78 up to The employment of the non-coded bits decreases the
0.945, in terms of the proportion of non-coded bits coding gain leading to a higher BER at a given SNR,
within the 8 bits loaded on a QAM symbol. on one hand, and increases the coding rate leading to
The relatively small decrease of the coding gain more information bits transmitted, on the other. The
could be explained by the protection of the non- effects of these two factors upon the throughput
coded provided by the 2-level Gray mapping and by provided by such configurations are shown below.
their soft decoding. Fig. 6 shows the BER vs. SNR For throughput evaluation we considered an OFDM
11
transmission that is based on a user-bin of Ti tones The SNR ranges of optimum for each configuration
and F OFDM symbol periods, thus containing Tsi = Ti are separated by the thresholds Ui.
x F QAM symbols (packet length), out of which only The family F3 together with the thresholds Ui may be
Asi are active symbols being used for the payload. employed into an adaptive modulation scheme that
The cyclic prefix is denoted by G and represents a provides the best throughput according to the channel
fraction of the symbol period, the number of bits per current SNR.
QAM symbol is ni (it defines the QAM constellation This scheme is very simple since it changes only the
employed), the bin rate is Db and the CRC (required bit-loading and employs the same QAM constellation
for channel estimation-prediction) is t bits long. and LDPC code. Despite its simplicity, it provides a
reasonable throughput over a SNR range of about 10-
Using (8) the nominal payload for a non-coded
12 dB.
constellation i (ni), i.e. the maximum value for the
V. CONCLUSIONS
payload when SNR is very high, is:
The array-based LDPC codes employed in the present
1 As t paper allow for a simple encoding and a moderate
Dni = Db Ts n i (1 ); (12)
1 + G Ts As n i complexity decoding, compared to the turbo codes.
The LDPC decoder requires the a priori knowledge of
As for the coded configurations, their nominal pay-
the channels noise variance.
load is computed using (8) and is given by (13). There
The BER performances of the LDPC-coded QAM
should be noted that the number of control bits of a
modulations are close to the ones provided by the
codeword jp should be divided by a constant that
indicates the number of bins on which a codeword is similar modulations coded with turbo codes at the
loaded. same rate and the same number of iterations per code
1 A jp word. A BER = 10-6 at SNR = 2 dB in an AWGN
D ci = s D b Ts n i (1 ); (13) channel can be obtained by an R = 0.5 LDPC-coded
1 + G Ts ki As ni
2-PSK, ensuring a coding gain of about 9 dB.
Considering an adapted version of the values A very flexible rate changing LDPC-coded scheme
proposed in [7], namely Db = 1500 bins/sec, Ts = 120 can be obtained by using a bit-loading that combines
symbols, As = 108 symbols, G = 0.11, t = 8 bits and ni coded and non-coded bits. This approach allows for
= 8 the nominal payloads (12, 13) of the significant increases of the coding rate at the expense
configurations of table 7 are listed in table 8. The of rather small coding gain losses.
constants ki are respectively 2, 1, 2/3 and . Due to the behavior of the LDPC-decoding algorithm
and to the soft-decision of the non-coded bits, the
Table 8. Nominal payload of configurations from
authors estimate that small and high rate RS outer
table 6
codes should be employed in FEC schemes based on
Cfg 1 2 3 4 5
concatenated codes.
Dni (kbps) 1156 1104 1041 979 916
The LDPC-coded modulation scheme proposed in the
The non-coded throughput ni is computed paper ensures coding gains of about 6 dB, compared
considering only the correctly received bins given by to the correspondent non-coded scheme.
the error-bin probability BinERni, and is: It also provides an increased throughput and offers the
possibility of adaptive employment according to the
ni = D ni (1 BinER ni ) (14)
channel SNR.
The throughput ci of the coded configurations is REFERENCES (SELECTED)
computed using (14), but the error-rate of the coded [1] ITU-T, LDPC codes for G.dmt.bis and G.lite.bis,
bins BinERci is employed. Temporary Document CF-060.
The throughput ni or ci vs. SNR curves of the trans- [2] D.J.C. McKay, Good error-correcting codes based on very
missions employing the configurations of table 6 were sparse matrices, IEEE Trans. on Information Theory, vol. 45, No.
2, March, 1999.
obtained by simulations and are displayed in fig. 7. [3] R. Gallagher, Low-density parity-check codes IRE Trans.
Each configuration exhibits a range of SNR where it Information. Theory, vol. IT-8, January 1962.
provides the best performance, out the entire family. [4] ITU-T, Low-density parity-check codes for DSL
transmission, Temporary Document BI-095.
[5] V.Bota, `Zs.Polgar, M.Varga, BER Performance of the QAM
Modulation Coded with LDPC Codes, International Symposium
Etc. 2002, Buletinul Universitii Politehnica, Tom 47 (61),
2002, Fascicola 1, 2, 2002, pp. 104.
[6] - Cl. Berrou, A.Glavieux, Near Optimum Error Correcting
Coding And Decoding: Turbo-Codes, IEEE Transactions on
Communications, vol.44, pp. 1261-1271, October 1996
[7] - W. Wang, M.Sternad, T. Ottosson, A. Ahlen, A. Svensson, -
Impact of Multiuser Diversity and Channel Variability on
Adaptive OFDM, Proceedings of COST 289 Spectral and Power
Efficient Broadband Communications Seminar, Budapest 2004.
Fig. 7. Throughput vs. SNR of configurations from table 6
12
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
{( ( ) ( ) ( )) }
RSG (k ) = f 0 , f 1 , K, f n 1 , deg f < k ; f Fq [x ] (3) deg w x i y j = u i + v j ; w = (u , v ) (6)
There are two possible ordering rules, namely direct
Codes dual to the evaluation codes.
ordering (lex order) and reversed ordering (revlex
Defining the RS codes as evaluation codes, leads
order), defined by:
to the possibility of employing list-type algorithms for
their decoding, algorithms that provide higher per- lex order : xi1 y j1 < xi 2 y j2 if ui1 + vj1 < ui2 + vj2 or
formances than the classical ones, represented in this
paper by the BM algorithm. ui1 + vj1 = ui2 + vj2 and i1 < i 2 (7)
The list-type decoding algorithms [3] [5] operate rev lex order : the same order but for i1 > i2
1
Technical University of Cluj Napoca, Communications Department,
26-28 G. Baritiu Str., 400027 Cluj Napoca, e-mail: Zsolt.Polgar@com.utcluj.ro
13
A significant theorem that, together with other 2
n k+1 k+1 n
theorems, secures the existence of an interpolation Lmax= m(m+1) + <(m+0.5) (12)
polynomial is [3]: k1 2k 2 2k 2 k 1
Theorem 1: Let {m(,):(,)F2} be the
multiplicity function of the zeros of Q(x,y) and
0<1<.... an arbitrary monomial order. There always
exists a polynomial Q(x,y): n (k 1) m + 1 rd n 1 (k 1) m + 1 k 1
m m m
C (13)
Q(x, y) = a i i (x, y) (8)
i =0
III. ANALYSIS OF THE SIMULATION RESULTS
In (8), C is expressed by:
The main goal of this paper is to compare, by
m(, ) + 1 computer simulations, the correction capability and
C = (9)
, 2 processing time of the GS and BM (representative for
the classical algorithms) RS decoding algorithms. The
The complete proof of the existence of the analysis is intended to establish the optimum values of
interpolation polynomial is to be found in [4] and [5]. the parameters of the GS algorithm, for which a
The existence of a polynomial that could be maximum ratio correction capability/decoding time is
decomposed in (y-f(x)) factors is secured by theorem accomplished and to elaborate some thumb rules for
2 [3]: adapting these parameters, so that shorter decoding
Definition 1: For Q(x,y)F[x,y] and f(x)F[x] the Q- times could be attained.
score of f(x) is defined as: The software simulator, that can operate in the
S Q (f ) = ord zero(Q : , f ( )) (10) Galois fields GF(23), GF(24), GF(25), GF(26) and
GF(28), performs the following functions:
Theorem 2: generation of a symbol-sequence represented on
If f(x)Fk[x], Q(x,y)F[x,y] and SQ(f) > deg1,vQ the number of bits corresponding to the employed
(11) Galois field.
then y-f(x) is a factor of Q(x,y) ; v=k-1. RS encoding (cyclic code for the BM or evaluation
A thorough analysis of the factorization step is code for the GS), depending on the decoding
presented in [4] and [5]. algorithm employed.
One of the most efficient interpolation algorithms serialization of the coded bits, generation of the
is the Koetter algorithm [3], which is defined by the packet-errors and their insertion in the coded bits.
pseudo-code below: GS or BM decoding and computation of the
Koetter interpolation algorithm parameters of the simulated transmission, namely: bit
- input data: L number of code words in the list, (i and symbol error rates, the ratio of the correction
j) ni=1 interpolation points, (mi)ni=1 zeros multi- capability of the GS algorithm versus the correction
plicity order, (1,k-1) monomials weighted degree. capability of the BM algorithm, the numbers of words
in the decoding list and erasures, both for the GS
1. FOR j=0 to L algorithm.
gj=yj The generation of the packet-errors, which
2. FOR i=1 to n DO simulates the transmission channel, is performed
2. FOR (r,s)=(0,0) to (mi-1,0) DO /*lex order according to the impulse noise models employed for
4. . FOR j=0 to L DO the xDSL transmissions [6]. This model was adopted,
5. j=Dr,sgj(i,j) with several simplifications, since it is a representative
6. J={j: j 0} one for transmission systems employing RS codes as
7. IF J outer codes. The main features of algorithm that
8. j*=min_rank {gj:jJ} generates the packet-errors are:
9. f=gj* ; =j* the distance in symbols between two packet-errors
10. FOR jJ DO has a Poisson distribution, with a modifiable average
11. IF (jj*) value . In the simulations performed, the value of
12. gj=gj+jf equaled the number of symbols of two code words, for
13. ELSE IF (j=j*) each GF.
14. gj=(x+i) f the packet-error length, in bits, has a gaussian
15. Q0(x,y)=min_rank{gj(x,y)} /* the interpolation distribution, defined by the average value and
polynomial variance . The value of equaled tbq, tb denoting the
number of error-symbols that could be corrected by
One of the best factorization algorithms, the Roth- the classical decoding algorithms (e.g. BM) and q
Ruckenstein [3], was used in the present analysis. denoting the number of bits/character of the GF
The bounded values of two significant parameters employed. The value of was set according to the
of the GS algorithm, the number of code words in the estimated correction capability of the GS algorithm.
decoding list, L, and the decoding radius, rd, are given the positions of the errors inside the packet are
by [3]: random, being distributed according to a uniform law.
14
A. Decoding Capability of the GS Algorithm Fig.1.c shows that for RS codes defined in GF(23)
The performances of the GS algorithm were evaluated and GF(24), m has to be set to 3 or 4, for Rc close to
0.5, and to 1 or 2 for Rc close (or smaller) than 0.3.
for RS codes with the coding rate Rc[0.3, 0.65]. The
The increase of m above a certain limit does not bring
parameters that indicate the correction capability are
a performance improvement, but it might lead to a
the minimum and the maximum decoding radius,
decrease of performances (see Rc = 0.33). A more
computed using (12), and the correction rate Rd
complete evaluation of the GS decoder requires the
(obtained by simulations). The Rd parameter is defined
consideration of the rmin and rmax, as well; good
as the ratio of the number of error words after the GS
decoding performances should be accomplished when
decoding and the number of error words that have a
the two parameters take equal or close values. The
number of error symbols higher than tb (the decoding
optimum values of m can not be established by
radius of the classical algorithms), before the
considering only the rmin and rmax parameters of the
decoding. The codes with Rc < 0.3 were not
code, as shown by Rc = 0.46.
considered, since they are of low practical importance.
As for the codes with Rc > 0.5 - 0.6 (depending of the 14
Fig. 2.a Rc=0.29
employed GF), the correction capability of the GS Rc=0.35
12
r_min
the zeros in the GS algorithm; m = 0 is actually 6
7
=3.5 ; Rc=0.42
Rc=0.33 10
6
Rc=0.46 =2 ; Rc=0.48
r_min
9
5
=2 ; Rc=0.55
4 8
Rc=0.43*
3 7
0 1 2 3 4 5 6
m
2
0 1 2 3 4 5 6
m
1
9 Fig.2.c
=3.5 ; Rc=0.2 Fig. 1.b 0.9 Rc=0.55
8
0.8
7
Rc=0.48
=2.5 ; Rc=0.33 0.7
6
=2 ; Rc=0.46 0.6
r_max
Rd
5
0.5
4
=2 ; Rc=0.43* 0.4
3
0.3
2
0 1 2 3
m
4 5 6
0.2 Rc=0.42
1
0.1 Rc=0.29 Rc=0.35
Fig. 1.c
0.
Rc=0.43* 0
0 1 2 3 4 5 6
0. m
0.
Rc=0.33
Rc=0.2 The values of Rd, see fig.2.c, indicate that for RS
0.
codes defined over GF(25) the optimum values of m
0. are m = 3 - 4 for Rc close to 0.5, m = 2 - 3 for Rc
0. around 0.3 and m = 4 - 5 for Rc around 0.4. There
0.
should be noticed that for m=6, the performances of
0 1 2 3
m
4 5 6
the GS decoder exhibit a significant decrease,
Fig.1 Minimum, rmin (1.a), maximum decoding radius rmax
especially for high values of the coding rate Rc.
(1.b), correction rate Rd (1.c) in terms of m ; RS codes in For RS codes defined in GF(25) having the
Galois GF(23) and GF(24) ; * denotes codes defined in GF(23); mentioned Rc and for the optimum values of m, the
15
decoding radius of the GS lies between rmin tb, and The performance loss exhibited by the GS
rmax = rmin+1 or rmin+2. The values of parameter of algorithm for high values of m, regardless the coding
the error-packet for the considered rates are given in rate, could be explained by the incomplete
fig.2.b. factorization, see (11), the requirements for the
interpolation being ensured by a proper choice of the
25
Fig.3.a number of words within the decoding list.
Rc=0.33
Rc=0.4 A primary analysis of the interpolation algorithm
20
Rc=0.46 presented in Section II and of the properties of the
two-variable polynomials [3] leads to the following:
15 Rc=0.52
the number of iterations, nit, performed by the
r_min
Rc=0.59
10
interpolation algorithm for n-symbol code words and
multiplicity order of zeros equaling m, is :
5
m (m + 1)
n it = n (15)
2
0
0 1 2 3 4 5 6
m
the initial polynomials of Koetter interpolation
26 algorithm, for maximum L words in the final decoding
Fig.3.b
list, are:
24 =7 ; Rc=0.33
22
=6 ; Rc=0.4 p0(x, y) = 1, p1(x, y) = y, p2(x, y) = y2 ,K, p2(x, y) = yL (16)
20
=4 ; Rc=0.46 supposing that the values of , computed within
r_max
Considering the RS codes defined in GF(26), see from the factorization requirements we have:
figs. 3, they exhibit a clear separation of the optimum
r (L + 1)
values of m, in terms of the coding rate Rc. For 1
SQ min n 2
Rc 0.5, optimum m equals 3 or 4, but for Rc 0.45, >1 (20)
optimum m equals 2 or 3. Sometimes, see Rc = 0.4, deg min (m + 1) + L (L + 1) R
m = 4 provides better performances at the expense of a m
longer decoding time. The values of the ratio defined in (20), for the
The codes defined in GF(26) exhibit the same codes of figs. 2 and 3 and for various values of m, are
decrease of performance for higher values of m (e.g. smaller than 1 (approximately equal, but smaller).
m = 6), as the ones defined in GF(25); for the There should be noted that the considerations above
considered values of Rc, the performances secured by are not complete, since it did not considered that the
the GS become equal to the ones of the BM. The evolution of polynomials degrees within the
values of parameter of the error-packet for the interpolation algorithm would be different, mostly
considered rates are given in fig.3.b. because of the fact that might equal zero quite often,
changing the evolution of the polynomials degrees
16
(see the interpolation algorithm in Section II), and 1
Rd
accomplished by using different values of m for every Rc=0.32
0.4
interpolation point, values chosen depending of the
Rc=0.56
channel characteristics [5]. Obviously, this approach 0.3
would complicate the implementation of the decoding 0.2
Rc=0.36
GS algorithm. Rc=0.51
Unlike the previous cases, for optimal values of m, 0.1
Fig.4.c
the codes defined in GF(26) have rmin < tb, but the 0
difference rmax - rmin takes values between 3 and 6. The 0 1 2 3 4 5
m
difference rmax - tb takes values between 0 and 3. So, Fig.4 Minimum, rmin (4.a), maximum decoding radius rmax
for the codes defined in GF(26) the optimum values of (4.b), correction rate Rd (4.c) in terms of m ; RS codes in Galois
m cannot be evaluated only be considering the limit GF(28);
values of the decoding radius, rmin and rmax. Fig. 4.c shows three optimum values of m,
Note: the relation rmin < tb does not imply that the GS depending of the coding rate Rc, for the codes defined
algorithm could not correct tb symbol-errors (the in GF(28). For coding rates higher or equal to 0.6 the
declared correction capability of the code), so optimum value of m is 4, for Rc (0.6, 0.45) the
practically one should consider that rmintb. The values optimum value of m is 3, and for coding rates ranging
of rmin and rmax provided by (13) evaluate the between 0.3 and 0.45, the optimal m equals 2. The
possibility of the GS algorithm to correct more errors figure also shows that, similar to the codes defined in
than the classical algorithms. As for the RS codes GF(26), the maximum limit of m falls to 5, for the
defined in GF(28), see figs. 4, the considerations coding rates considered.
regarding rmin and rmax, presented above, are still valid. The comparison of the results presented in figs. 1.c
There should be mentioned that, for the optimal values - 4.c show that the coding rate for which the GS
of m, rmin tb, and the difference rmax - rmin takes values decoding algorithm provides better performances than
higher or equal than 20, and the difference rmax-tb lies the classical decoding algorithms increases with the
between 2 and 13. increase of dimension of the Galois field in which the
RS codes are defined.
100
Fig.4.a
Regarding the number of words in the decoding
Rc=0.32
90
Rc=0.36 list, the simulations performed by the authors show
80
Rc=041 that for the RS codes defined in GF(25) and in the
70 higher fields, the number of the words in the list
Rc=0.47
60
equals 1, with very few exceptions, when then list
Rc=0.51 contains more than one code word. As for the codes
50
Rc=0.56 defined in GF(23) and GF(24), there are more cases
r_min
40
Rc=0.61 when the decoding list contains more than one code
30 word, but their percentage is still small, about 1%. As
20
Rc=0.65 a general conclusion, if the GS decoding algorithm
10
can not correct a code word, this fact is owned to an
unsuccessful interpolation or factorization and, quite
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 seldom, to the presence of more than one code words
m
in the decoding list.
110
Fig.4.b =50 ; Rc=0.32
B. Evaluation of the GS algorithm decoding time
100 =40 ; Rc=0.36 The references [3] [4] [5] present some considerations
regarding the number of operations performed by the
90 =30 ; Rc=0.41 GS algorithm, which affect significantly the decoding
time, but these considerations do not include a
80 =20 ; Rc=0.47
comparison to the decoding time required by the
r_max
17
The evaluation of the decoding time implied the large average decoding time, even larger than the ones
measurement, for a certain number of code words, of presented in fig. 5, because even the correctly received
the simulation time tsim, and of the time required for code words would be decoded in a very long time.
encoding and error-pattern insertion taux ; for the the successive increase of the value of m, from 1 to
measurement of taux the decoding procedures were a maximum optimal value. The decoding is stopped
removed from the simulation program. There should when the decoding list contains at least one code
be noted that the time required to decode a correct word; this approach would require a smaller average
code word differs from the time required to decode an decoding time for packet-errors with relatively small
error code word for both algorithms, especially for the lengths, compared to the maximum packet length for
GS algorithm. The ratio between the average decoding which a successful GS decoding is accomplished.
times, tdec, of the two algorithms is expressed by: the employment of two values for parameter m,
t t t t namely 1 and an optimum value mopt. The correct code
t d = decGS = simGS auxGS simGS (21) words and the ones affected by a small number of
t decBM t simBM t auxBM t simBM errors (equal or higher than tb) would be decoded
Fig. 5 presents the variation of the ratio td using m=1, and the code words with more errors
(expressed on a logarithmic scale) between the would be decoded with m = mopt; this last option
average decoding times of the GS and BM algorithms should be employed if the decoding with m=1
in terms of m, for various coding rates and for codes generates no code word in the decoding list. This
defined in several Galois fields. variant of employing the GS algorithm provides a
smaller average decoding time for long error-packets,
4.5 compared the maximum packet length for which a
successful GS decoding is accomplished. The results
4 GF(28) ; Rc=0.32 displayed in fig. 5 were obtained using this decoding
variant.
3.5 IV. CONCLUSIONS
GF(26) ; Rc=0.59
GF(28) ; Rc=0.64 The computer simulations performed by the
3
authors showed that the GS decoding of RS codes,
lg(td)
18
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
19
with blocks of bits, the overall code can be
considered as a block code too
v0
x
RSC 1 Puncturing
c
v1 And
Multiplexing
Interleaver RSC 1
~
x w1
20
out column-wise. The interleaver has the following used to control the interleaver are long. One idea was
distance property to use a bloc even-odd interleaver structure, for which
()i, j A, i j MB 1 (10) the indexes have been pseudorandomly interleaved in
(i ) ( j ) M 1
advance. In this way, using 10 times shorter PN
codes, and the results are close to the ones obtained
using purely random interleavers.
C. Random Type Interleavers introduces an N
bits input block of data into a memory and reads it
out randomly, in accordance to the following N step IV. DECODER STRUCTURE
algorithm:
Step 1: choose index i1 from the set A {1,2, ... ,N}, The iterative decoder structure consists of two
in accordance to a uniform probability function component decoders, serially concatenated via an
p(i1 ) = ; the corresponding output index is (1);
1 interleaver identical to the one used in the encoder, as
N shown in figure 2. The first decoder uses the received
Step k: choose index ik from the set information bits r0 and the parity bits generated by the
Ak = {i A, i i1 , i2 , ... ik 1 } , in accordance to a first encoder r1 in order to produce a soft output,
denoted with 1e , which is interleaved and used to
uniform probability function p (i ) = 1 ; the
k improve the estimate of the apriori probabilities for
N k 1
the second decoder. The other two inputs of the
output index is (k); second decoder are also the received information
Since a pure random interleaver is sequence which are interleaved by the same algorithm
generally hard to implement, in practice are often as in the encoder, ~r0 and the received parity sequence
used pseudo-random interleavers, where the indexes
(i) are the outputs of a pseudonoise shift register, produced by the second encoder r2. This decoder
generated by a primitive polynomial. produces also a soft output, denoted with 2e , that is
de-interleaved and used by the first decoder to
D. Hybrid Interleavers improve its apriori probabilities. This iterative
feedback operation increase the performances of the
Simulations have shown that bloc / even-odd overall structure, especially in the first decoding steps.
interleavers are simple to implement, but their After a number of iterations the soft outputs from the
decorrelation properties are low and, therefore, the decoders will no longer affect significantly the
overall BER properties are poor. On the other hand, performances, and, therefore, a hard decision is
the random type (pseudorandom) interleavers have applied at the end in order to obtain the decoded data
the best decorrelation capabilities, so the BER sequence. Both decoders are Soft Input Soft Output
results are very good, especially when the PN codes (SISO) type
De-Interleaver
1e
Decoder
r0 1 Interleaver
r1
~
r0
Decoder 2
Interleaver 2 De-Interleaver Threshold
r2 2e
21
1 ; (ct ) 0
ct = (14)
0 ; (ct ) < 0
The log-likelihood ratio from (6) can be determined
using MAP, log-MAP, Max-Log-MAP and SOVA
algorithms ([5], [6]).
Using this backbone program, several important Fig. 4. Bit Error rate versus Eb/N0, different interleaver types,
aspects have been studied and compared one SOVA decoder, frame length L=400, punctured, 5 iterations, 10
errors to terminate de decoding
another from the BER point of view. The simulation
of the overall system leaded to a complex and time
The random interleaver gives the best performances
consuming program and the simulation process is
from all (about 10dB improvement at SNR=1dB), for
still under development in order to cover all the
both SOVA and MAP algorithms, and this
problems encountered, to compare all the possible
improvement does not depend on the random
configurations and obtain relevant and
interleaver realization. The hybrid interleavers
comprehensive results. However, from the results
performances are close to the ones obtained using the
obtained till now, several aspects have to be
pure hybrid one, especially when the interleaver
emphasize.
length is large (in figure 3, for frame length L=200,
the hybrid interleaver behaves slightly worse then the
The interleaver type effect: several interleaver random one while in figure 4, for frame length
structures have been studied: block (row-column), L=400, the two curves merely overlap).
even-odd, helical, random and hibrid. The results
are shown in figures 3 and 4. The block interleaver
The interleaver length effect: the interleaver
achieves the worst performances from all; the even-
length gives us the length of the data block that has to
odd and helical have similar performances, better
be processed by the decoder at a certain step in order
then the row-column, with both SOVA and MAP
to recover the data. As it can be seen form figures 3, 4
decoding algorithms (the difference becoming
and 5, as the interleaver length increases, the system
significant at high SNRs). The cyclic-shift (helical)
performances improve also, whatever decoding
interleaver has better performances then the all
algorithm is used.
block ones, especially at high SNRs (about 6dB
improvement at SNR=1dB).
22
Fig. 5. Bit Error rate versus Eb/N0, hybrid interleaver, SOVA & Fig. 6. Bit Error rate versus Eb/N0, SOVA decoder, different
MAP decoders, frame length L=200/ 400, punctured, 5 iterations, number of iterations, hybrid interleaver, frame length L=200,
10 errors to terminate de decoding punctured, 10 errors to terminate de decoding
However, the frame length increase determines also curves for 2, 3 and 4 degree generator polynomials,
an increase in structure complexity and decoding namely
delay; therefore, the solution has to be chosen as a
compromise between a certain threshold in deg 2 : g1 (D ) = 1 + D + D , g 2 (D ) = 1 + D 2
2
23
interleaver are in between, close to one another REFERENCES:
from BER point of view;
the new hybrid interleaver achieves BER [1] C. E. Shannon A mathematical Theory of Communications,
Bell Syst. Techn. Journal Vol. 27, pp. 379-423 (part I) & 623-656
performances close to the random ones, (part II), Oct 1948.
especially when the frame length is large; [2] C. Berrou, A. Glavieux and P. Thitimajshima, Near Shannon
as the frame length (block size) increases, the limit error-correcting coding and decoding: turbocodes" ICC-1993,
system performances improves also; however, Geneva, Switzerlend, pp. 1064-1070.
[3] J. J. Ramsey, Realization of Optimum Interleavers, IEEE
as the frame length increases, the system Trans. on Info. Theory, Vol.16, No.3, May 1970, pp. 338-345
complexity (and costs) and the delay increases [4] G. D. Forney, Jr., Burst Correcting Codes for Bursty
also; therefore a compromise has to be made Channels, IEEE Trans. Comm., Vol. 19, No. 5, Oct 1971, pp. 772-
between performances and costs; 781.
[5] P. Robertson, E. Villebrun, P. Hocher, A comparison of
the MAP decoding algorithm achieves better
Optimal and Sub-Optimal MAP Decoding Algorithms Operation in
performances then the SOVA one, especially the Log-Domain, Proc. ICC95, Seattle, June 1995
fol low SNRs; for large SNRs the difference is [6] B Vucetic, Iterative Decoding Algorithms, PIMRC97, Sept
no longer significant; however, the MAP 1997, Finland, pp. 99-120.
[7] S. A. Barbulescu, W. Farrell, P. Gray, and M. Rice, Bandwidth
algorithm complexity is 3 times larger then the Efficient Turbo Coding for High Speed Mobile Satellite
SOVA one, and therefore the decoder structure Communications", in Proc. International Symposium on Turbo
(and cost) and the associate delays are also Codes and Related Topics, Brest, France, pp 119-126, Sep. 1997.
higher; [8] S. S. Pietrobon, Implementation and performance of a
turbo/MAP decoder, International Journal on Satellite
the number of iterations in decoding algorithm Communications, Vol. 16, pp.23-46, 1998.
leads to an increase in BER performances, till [9] S. A. Barbulescu and S. S. Pietrobon, Interleaver design for
to a certain threshold which depends on the turbo codes, Electronics Letters, Vol 30, No 25, Dec. 1994.
frame length (as the frame length increases, the [10] W. Feng, J. Yuan and B. Vucetic, A code matched interleaver
design for turbo codes, Proceedings International Symposium on
number of necessary iterations decreases);
Personal, Indoor and Mobile radioCcommunications (PIMRC),
as the memory encoder / decoder polynomials Osaka, Japan, pp. 578-582, Sep. 1999.
degrees increase, the system BER performances [11] J. Vogt, K. Koora, A. Finger and G. Fettweis, Comparison of
improves, especially for large SNRs, while for Different Turbo Decoder Realizations for IMT-2000,
GLOBECOM99, pp.2704-2708, Dec. 1999
low SNRs the low order degree encoders [12] J. H. Kang and W. E. Stark, Turbo codes for noncoherent FH-
behaves better. SS with partial band interference, IEEE Transactions on
Communications, Vol. 46, pp. 1451-1458, Nov. 1998.
It has to be mentioned that simulations are [13] D. Roddy Satellite Communications, McGraw-Hill
TELECOM Engineering, 2001.
continuing, in order to determine the system
behavior for low (negative) SNRs. Other
interleaver structures are currently under study, as
well as different types of fading effects on BER
performances. Moreover, in order to obtain more
reliable results, the BER results has to be averaged
on several number of simulations
24
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract This paper presents an implementation of a The first block of the two mentioned above is
method for text compression, first described by suggesting which the next symbol could be, based on
Shannon in 1951 and later, in 1997, by Fenwick. Unlike the current context. It starts by searching for a string
other compressors, which exploit symbol frequencies in of the maximum permitted length matching the
order to assign shorter codes to more frequent symbols,
this technique prepares a list of probable symbols to
current context and as soon as it is found, the next
follow the current one, ordered from most likely to least symbol is offered. If the offer is rejected, the search
likely. Tests were conducted on various input sources continues until there are no possible strings of this
and the results are shown here. The implementation length left in the buffer. Then the length is
supports further optimizations, as it will be explained. decremented and the search starts over. The context's
Keywords: symbol ranking, context, number of tries 1
length can go as low as 1, meaning a symbol could be
the context, but no lower. When a string is found to
I. INTRODUCTION be equal to the current context, the symbol next to it
is first checked to see if it was not offered before and
Symbol ranking compression is a method initially rejected, in which case it is not used again.
described by Shannon in 1951 [1] and later, in 1996,
by Fenwick [2]. The algorithm is expecting to The second block performs as a validation. It reads a
encounter repeating strings in the input, which makes symbol and asks for suggestions, a kind of guess
it more suitable for repetitive text, such as human game. If it receives a good answer, it outputs a "1"
language. It consists of two blocks: a seeking bit, otherwise a "0" bit: this is where the compression
function, for both compression and decompression, occurs. It does not accept more than a certain number
which is used to suggest a symbol, and a processing of guesses, if, until that, the right symbol has not been
function, specific to the action being carried out. offered the search is aborted, the correct symbol is
output, and the scheme moves on. On decompression
Other similar methods do exist, by Bentley et al. [3], things work quite similar, but now the answers are
Howard and Vitter [4] or Burrows and Wheeler [5] read from the compressed file, be it a "1" or a "0" bit,
(the Burrows-Wheeler Transform), but with little or the correct symbol.
reference to Shannon's original work. BWT could be
regarded as a symbol ranking compressor, with the III. EXPERIMENTAL SOLUTION
Move-To-Front list acting as a good estimate of
symbol ranking. A software implementation of this compression
method was performed. The program uses a circular
II. ALGORITHM DESCRIPTION buffer to keep track of processed text, both on
compression and decompression. This means that
The algorithm extends one of Charles Bloom's when the first symbol is read it is inserted at the first
methods of offering possible symbols in the address in the buffer, and subsequent symbols follow
approximate order of probability of their occurring in it, until the end of the buffer is meet, then symbols are
the current context. Bloom [6] noted that the longest again inserted from the first position, and so on.
earlier context, which matches the current context, is
an excellent predictor of the next symbol. However, Obviously, this way some possible better contexts in
this implementation only uses contexts of a certain the past are lost, and an ideal approach would store in
maximal length, as it would be more time-consuming the memory all the previous text, instead of a bound
to leave the context unbound. length buffer. That is why the buffer's length can be
varied to some extent, to try to accustom to different
input texts.
1
Facultatea de Electronic i Telecomunicaii, Catedra de
Electronic Aplicat i Ingineria Informaiei, Bd. Iuliu Maniu
nr. 1-3, sector 6, Bucureti, e-mail: rradescu@atm.neuro.pub.ro
25
To find a matching context the search begins in the 0011100100000s1111101000100100000_00000o000
buffer at the previously inserted symbol with the 00f000f000e0001000010001,
maximum context length. It seeks backwards for a
symbol that is equal to the one at the end of the which means that for the 168-bit input, the coder
context. outputs 110 bits, giving a compression ratio of
0.65476. Of course, the result depends a great deal on
Having found one, it tries to match the symbol at the the parameters used in the algorithm and on the
half of the current context to the one at the half of the context (the previous symbols).
possible candidate context and then the first symbols,
in the context and in the candidate. If all these IV. COMPRESSION RESULTS
symbols match, then a complete string comparison is
performed, and, if it succeeds, the symbol is offered The following tests were conducted using 5 files of
and marked in the exclusion table. This approach is various content and dimension:
used to avoid unnecessary comparisons, and many
false contexts fail the half- or start- symbol test. codulpenal.txt Romanian text, legal;
book1.001 English text, fiction book
An example of how the compression sequence could (incomplete, in order to have about the same length as
look like is presented in the following. The text to be the first one a comparison between the two
coded is "the symbol is offered", the context is languages' compressibility was attempted);
"Having found one [...]", from the previous obj1 object code for VAX machine, binary file;
paragraph, all in the buffer. At most 5 unsuccessful pic black & white fax picture (it is supposed to
attempts are allowed. The maximum context size is 4. have a big redundancy and to be very compressible);
In Table 1, the first column is the original text and progl source code in LISP.
the columns to the right are the attempts of guessing
the next symbol. The rightmost cell on every row Every file (except the first one) is part of the Calgary
contains the sequence of output symbols for the Corpus collection [7]. The goal was to try different
corresponding input character, underlined characters scenarios for the use of this algorithm, as it is a
representing a binary value (a "1" bit or a "0" bit) and known fact that a compression method can yield
the numbers between parentheses the context length better results only for certain file types.
at which the right symbol was found, or (n/a) if no
guess was successful. White spaces were converted The program can be tuned by three parameters:
to underscores for reasons of readability.
(a) the length of the buffer;
Table 1 (b) the maximum context length;
Context: Having found one, it tries to match [] (c) the number of tries before outputting the
Text Attempts Output unchanged symbol.
t i a t 001 (2)
h h 1 (3) The following graphics represent the compression
e e 1 (4) ratio of the symbol ranking method when modifying
_ n s _ 001 (4) only one of the parameters, keeping the others to
s c f p h o s 00000s (n/a) some fixed best-ratio values.
y y 1 (4)
m m 1 (4) The columns grouped by five in the histograms below
b b 1 (4) (see Fig. 1, 2, and 3) represent the compression ratio
o o 1 (4) of the files in the above listed order.
l l 1 (4)
_ s _ 01 (4) 1
i a t s i 0001 (1)
s t f s 001 (2)
0.8
Ratio
26
As the buffer length is increased, the compression REFERENCES
ratio improves, except for the object code file; still,
the buffer should not be too large, as the running time [1] Shannon, C. E., "Prediction and Entropy of Printed English",
Bell System Technical Journal, Vol. 30, pp. 50-64, January 1951.
can reach high values. [2] Fenwick, P., "Symbol Ranking Text Compression with
Shannon Recordings", Journal of Universal Computer Science,
Vol. 3, No. 2, pp. 70-85, February 1997.
0.8 [3] Bentley, J. L., Sleator, D. D., Tarjan R. E., and Wei, V. K., "A
Locally Adaptive Data Compression Algorithm", Communications
0.6 of the ACM, Vol. 29, No. 4, pp. 320-330, April 1986.
Ratio
[4] Howard, P. G., and Vitter J. S., "Design and Analysis of Fast
0.4 Text Compression Based on Quasi-Arithmetic Coding", Data
0.2 Compression Conference, pp. 98-107, IEEE Computer Society, Los
Alamitos, California, 1993.
0 [5] Burrows, M., and Wheeler, D. J., "A Block-Sorting Lossless
Data Compression Algorithm", SRC Research Report 124, Digital
2 4 6 8 10 Systems Research Center, Palo Alto, California, May 1994,
available at: gatekeeper.dec.com/pub/DEC/SRC/research-
Maximumcontext size reports/SRC-124.ps.Z.
[6] Bloom, C., "LZP: A New Data Compression Algorithm", Data
Compression Conference, Vol. 3, No. 2, pp. 70-85, IEEE Computer
Fig. 2. Compression ratio as function of maximum context size Society, Los Alamitos, California, 1996.
[7] The Calgary Corpus can be found on the Internet at the address:
ftp://ftp.cpsc.ucalgary.ca/pub/projects/text.compression.corpus.
The maximum context size does not appear to make a [8] Fenwick, P. M., "Symbol Ranking Text Compressors: Review
lot of difference; it probably helps occasionally to and Implementation", Software Practice and Experience, Vol. 28,
have a larger context, but overall it seems to be less No. 5, pp. 547-559, April 1998.
important. The default value of this parameter should
be set to 4.
0.8
0.6
Ratio
0.4
0.2
0
3 6 9
Maximumnumber of tries
V. REMARKS
27
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Electronics and Telecommunications Faculty, Applied Electronics Departament, Iuliu Maniu 1-3, Bucharest, Romania
e-mail: rodicastoian2003@yahoo.com, lperisoara@yahoo.com
28
If the two codes can correct t1 = (d1 1) 2 , Lr- = Lec is used as a priori information in conjunction
with Y. The decoder generates the a posteriori log-
respectively t2 = (d 2 1) 2 errors, then the product likelihood ratios Lr+ for all bits. The extrinsic
code C is capable of correcting any combination of y
t = (d1d 2 1) 2 = 2t1t2 + t1 + t2 errors. Thus, we can
build very long block codes with large minimum
Hamming distance by combining short codes with Read entire block
small minimum Hamming distance. and place in array
The parity check matrix H of a product code C is Channel
computed using the parity check matrices H1 and H2 of reliability Y
individual systematic codes C1 and C2 as: computing Lc-
Columns
H T = I n2 H1T | H T2 I n1 (1) Lc
SISO decoder
Lec
where I n1 and I n2 are the unit matrices of order n1 and
n2, respectively. Lr -
Given the construction procedure, it is clear that Rows
(n2-k2) last columns of the matrix are the control bits of SISO decoder
C2. Also, all (n1-k1) last rows of matrix C are the
control bits of C1. Hence, all the rows of matrix C are Ler
Lr+
the codewords of C2 and all the columns of matrix C
are codewords of C1.
Hard
decision
B. The iterative decoder
29
def P (y | x = +1)
- L( y | x) = log (5)
P (y | x = 1)
SISO decoding of Le(i)
rows or columns of + According [9], for AWGN fading channel using
L-(i) matrix Y L+(i) L-(i+1)
- BPSK modulation we can write:
LcY
1 E 2
P( y | x = 1) = exp b ( y m a ) , (6)
N 0 N 0
Eb E Noted
( y a ) ( y + a ) = 4a b y = Lc y
2 2
In this case, the decoding procedure described L( y | x) =
N0 N0
above is generalized by cascading elementary decoders
illustrated in Fig. 3. The parameter i indicates the (7)
current decoding step of the iterative process. For the
implementation of SISO decoders, we can use the where Lc = 4a Eb N 0 is defined as the channel
MAP algorithm or the SOVA algorithm, which are reliability value. For non fading AWGN channels a=1
described below. and Lc = 4 Eb N 0 .
In [1], [9] the extrinsic information is defined as:
C. The Maximum A Posteriori algorithm
P( x = +1| y ) P (x = +1)
Bahl, Cocke, Jelinek and Raviv proposed the Le = log log
Maximum A Posteriori (MAP) decoding algorithm for P( x = 1| y ) P (x = 1)
convolutional codes in 1974 [6]. The iterative decoder P(y | x = +1)
developed by Berrou et al. [1] in 1993 has a greatly log (8)
increased attention. In their paper [1], the MAP P(y | x = 1)
algorithm was modified to minimize the sequence error = L+ L Lc y
probability instead of bit error probability for the
original MAP algorithm. Because of its increased In the iterative decoding procedure the extrinsic
complexity, the MAP algorithm was simplified and the information Le becomes the a priori information L- for
optimal MAP algorithm called the Log-MAP the next decoder. If L- is a large (or small) positive
algorithm was developed. number, then it would be difficult (or easier) to change
The decoder operates based on the Logarithm the estimated symbol decision from +1 to -1 between to
Likelihood Ratio (LLR) for the transmitted bits x consecutive decoding stages [10].
which is defined as: The term Lcy is the soft output of the channel for the
information symbol x. For high SNR, the channel
def P( x = +1) reliability value Lc will be high and this information
L( x) = log =L (3)
( = 1) symbol will have a large influence on L+. Conversely,
P x
for low SNR, the Lc is low and its influence on L+ is
insignificant.
where the sign of the LLR L(x) indicate whether the
each bit of x is more likely to be +1 or -1 and the D. The Soft Output Viterbi Algorithm
magnitude of the LLR gives an indication of the
correct values of x. In practical systems, we quantize the received
In channel coding theory we are interested in the channel symbols with one (hard decision) or a few bits
probability that x = 1 , based or conditioned on of precision (soft decision) in order to improve the
received sequence y. So, we use the conditional LLR: performances of the Viterbi decoder. For m-bit
quantization, one quantization bit is devoted to the sign
def P (x = +1| y ) We noted + of the decision and m-1 bits are devoted to the signal's
L( x | y ) = log = L (4)
P (x = 1| y ) magnitude. The larger the magnitude, the more
confidence that the sign bit is correct. Decoders that
exploit soft decisions can reduce S/N ratio requirements
The conditional probabilities P (x = 1| y ) are the
by approximately 2 dB over those that use hard
a posteriori probabilities of the decoded bits x and L+ decisions alone [11].
is the a posteriori information about x. The Viterbi algorithm finds the trellis path or state
Also, it is used the conditional LLR L( y | x) based sequence s so that the a posteriori probability p(s|y) is
on the probability that the receivers output would be y maximized. Accordingly to the Bayes rule, we can
when the transmitted bits x were either +1 or -1: equivalently maximize:
30
p(s j , y j ) = p (s j 1 , y j 1 ) p(u j ) p( y j | s ', s ) , (9) 1
0.0001
Obviously, nriter=3
(
p(s j , y j ) = exp M j (s j ) . ) (11)
0.00001
0 2 4 6
Eb/N0 (dB)
Substituting (9) into (10) gives:
Fig. 4. BER(Eb/N0) performance for MAP algorithm.
( ) (
M j (s j ) = M j 1 (s j 1 ) + log p(u j ) + log p( y j | s ', s ) )
(12) The performances of the SOVA algorithm are
illustrated in Fig. 5. In this case, for the same Eb/N0
(
where log p(u j ) ) is the a priori information of the of 3dB the BER is greater, 0.0867 for one iteration,
0.0273 at iteration 2, 0.0016 at iteration 3 and
source symbol uj and log p( y j | s ', s)( ) is the branch 0.0007 at iteration 5.
metric for the state transition ss given the received
signal yj. At time j, for each state s, the path metrics for From Fig. 4 and Fig. 5 we observe that the MAP
all possible paths terminating at state s are calculated. algorithm gives better results, in terms of BER,
compared with SOVA algorithm.
III. PERFORMANCE EVALUATION
1
To simulate the application of iterative decoding uncoded
Bit Error Rate
31
ANNEX 1. TRELLIS DESCRIPTION Conference on Communication, vol. 2, pp. 974-978, Dallas, Texas,
June 1996.
OF BLOCK CODES [5] A. Stefanov and T. M. Duman, Turbo-coded modulation for
wireless communications with antenna diversity, in Proc. IEEE
For a Hamming code with control matrix Vehicular Technology Conference, pp. 1565-1569, Amsterdam,
H = [h1 , h 2 ,..., h n ] , where hi is the ith column of H, any Netherlands, September 1999.
[6] L. R. Bahl, J. Cocke, F. Jelinek and J. Raviv, Optimal
codeword ci = (ci1 ,ci 2 ,..., cin ), ci C (n, k ), i = 1, n Decoding of Linear Codes for Minimizing Symbol Error Rate, IEEE
Trans. on Inf. Theory, vol. 20, pp. 284-287, 1974.
must satisfy the condition: [7] R. J. McEliece, On the BCJR trellis for linear block codes,
IEEE Trans. on Inf. Theory, vol. IT-42(4), pp. 1072-1092, 1996.
ci1h1 + ci 2 h 2 + ... + cin h n = 0 (13) [8] P. Elias, Error-free coding, IRE Trans. on Inf. Theory, vol.
IT-4, pp. 29-37, September 1954.
[9] C. Berrou and A. Glavieux, Near Optimum Error Correcting
where cij F2 and h j F2n k . Coding and Decoding: Turbo Codes, IEEE Trans. on Comm. vol.
44, no. 10, pp.1261-1271, October 1996.
For any codeword affected by errors the value of [10] R. Stoian, L.A. Perioar, The reliability of turbo decoders
the syndrome is: over Gaussian channels, in Proc. of Communications 2004,
Technical Military Academy, pp. 545-550, Bucharest, June 2004.
n
[11] J. Hagenauer and P. Hoeher, A Viterbi Algorithm with Soft
s n = yi h i (14) Decision Outputs and Its Applications, in Proc. of GLOBECOM
1989, pp. 1680-1686, Dallas, Texas, November 1989.
i =1
[12] R. M. Pyndiah, Near-optimum decoding of Product Codes:
Block Turbo Codes, IEEE Trans. on Comm., vol. 46, no. 8, pp.
where yi are the components of received vector y. 1003-1010, August 1998.
The BCJR trellis construction for linear block
codes is based on recursive computation of the
syndrome [6]:
si = si 1 + yi h i , s 0 = 0 (15)
111
110
101
100
011
010
001
000
symbol 0 symbol 1
REFERENCES
[1] C. Berrou, A. Glavieux and P. Thitmajshima, Near Shannon
limit error-correcting coding: Turbo codes, in Proc. of ICC 93,
pp.1064-1070, Geneva, Switzerland, May 1993.
[2] R. Stoian, L. A. Perioar, M. Crngeanu, A comparison of
MAP and SOVA decoding algorithms for Turbo Codes, in Proc. of
the 35th Symposium of ACTTM, Bucharest, March 2004.
[3] J. Hagenauer, E. Offer and L. Papke, Iterative decoding of
binary block and convolutional codes, IEEE Trans. on Inf. Theory,
vol. 42, pp. 429-445, March 1996.
[4] S. Benedetto, D. Divsalar, G. Montorsi and F. Pollara, Parallel
concatenated trellis coded modulation, in Proc. IEEE International
32
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1,2
Facultatea de Electronic i Telecomunicaii, Departamentul
Comunicaii Bd. V. Prvan Nr. 2, 300223 Timioara, balta@etc.utt.ro; kmaria@etc.utt.ro
33
A turbo code implies at least two coders, C1 and C2.
ln e xi max(x i ) (5) At each coder corresponds a trellis. Different trellis
i
i closing techniques can be used for these coders. In
this paper we investigate few trellis closing methods
So, the computing relations, in the MaxLogMAP for a turbo code (parallel). The table 1 presents these
algorithm case, are the following: methods.
Ak (s ) =
ln( k (s )) max( Ak 1 (s ) + k (s , s )) (6) Table 1
s Variant Start Final Coding rate
C1, C2 C1, C2
B k (s ) =
ln( k (s )) (B k (s ) + k (s , s )) (7) 01 0, 0 0, ? (N-M)/3N
11 0, 0 ?, ? 1/3
k (s , s ) =
ln( k (s )) = C Sx, Sy Sx, Sy 1/3
1 L n
= C + u k L(u k ) + c y ki x ki (8) 01. In this case, the first coder closes the trellis on
2 2 i =1 both extremities, it inserts M redundant bits after the
N-M information bits. The second coder can not do
the same final closing due to the interleaving of the
L(uk|y) max (Ak-1() + k(,s) + Bk(s))
(s,s ) input sequence. So, the second trellis is not closed.
u k = +1 (9) The first decoder initializes the alpha coefficient,
which corresponds at the front end of the trellis to the
max (Ak-1() + k(,s) + Bk(s))
(s,s ) zero state, with the probability 1, and the other alpha
u k = 1 coefficients with the probability zero. The first
decoder treats the beta coefficients in the same way at
the end of the trellis. The second decoder acts in the
The implementation simplification price of the
same way, like the first, for the alpha coefficients.
MaxLogMAP algorithm is the reduction of the
The beta coefficients of the second decoder, can be
performances (of the BER) with 0,2 dB versus the
initialized by the one of the following methods:
MAP algorithm.
01s. The beta coefficients are met equals with
LogMAP Algorithm, proposed by Robertson and al.
the values of the alpha coefficients obtained at the last
[3], corrects the approximation used by MaxLogMAP
iteration. This is called the soft initialization;
algorithm and is a little bit more complicated than it.
01h. This method initializes the beta coefficient,
that corresponds to the state with the highest alpha
x x
ln( e x1 + e x 2 )=max(x1,x2)+ln(1+ e 1 2 ) (10)
coefficient, at probability 1 and the others beta
= max(x1,x2) + c( |x1-x2| ) coefficients at the probability zero. This is called the
hard initialization;
01e. This method makes the beta coefficients to
II. THE TRELLIS CLOSING have the same probability. This is called the equal
probability initialization;
In function of the trellis closing the alpha and beta 11. None of the trellises is final closed. The
coefficients are initialized. The initial trellis closing advantage, in this case, is a higher coding rate. But
means the coder initialization with a predefined state. this coding rate increase can not be observed if
This state is also known by the decoder. So, the N>>M. Both decoders must initialize the beta
initialization of the alpha and beta coefficients can be coefficients in one of the three ways enounced above.
done. With the exception of circular coding, this In this paper we implement only the soft initialization.
initial state is zero. The final trellis closing is more The case of both trellis final closing is possible only
difficult to be realized. It is done (excepting the with some modifications of the interleaving between
circular code) with the price of insertion of the M (the the two coders.
code memory) redundant bits in the information C. There is the possibility, using a pre coding
sequence. This fact realizes the reduction of the technique, to find, for any data sequence x, an initial
transmission rate from 1/2 to the following value: state So of the coder identically with its final state. So
the coding becomes circular. The decoder does not
Rcc = (N M) / 2N. (11) know the state S0, but knows that it can use the final
state like initial state. So, it must to do at least a
The trellis closing gives the advantage of the initial forward recurrence. We are implemented and
state knowledge (and/or of the final state), fact which simulated the following variants of circular turbo
leads to the firm knowledge of the alpha coefficients code:
(at the beginning of trellis) and beta coefficients (at C1 the decoder realizes a forward recurrence and
the end of the trellis). In the case of the unclosed computes a final state, S0. The alpha and beta
trellis these coefficients can only be predicted coefficients are initialized with S0, the backward
probabilistically. recurrence is made and the forward recurrence is
34
remade. It memorizes the new state S0 for the starting fc(x). The values of the function gc(x) are in the set
of the next iteration. {0.6, 0.3, 0.14, 0.065, 0.03, 0.014, 0.005, 0.002, 0}.
C2 the decoder realizes the both recurrences in the The linear variant corresponds to an approximation of
soft variant and retains, for the next iteration, with the fc(x) of the form:
role of S0, the beta coefficients values from the end of
the backward recurrence. 0.7
0.7 x, x x o
C3 the decoder realizes the both recurrences in the hc (x ) = xo (13)
soft variant plus one for the alpha coefficients, only. 0, x > xo
The initial state for that second recurrence is done by
the finals values of beta coefficients of the last
iteration. The final coefficients of the second forward and by numerical approximation was obtained the
recurrence give the state to be stored for the next value xo = 2,347 for which hc(x) realizes the better
iteration. approximation of fc(x).
C4 the decoder realizes the forward recurrence and
build a S0 state in a hard decision (it searches the IV. EXPERIMENTAL RESULTS
alpha coefficients maximum). It retains this state for
the next iteration and also makes the backward In the figure 4 are presented the curves BER(SNR)
recurrence and remakes the forward recurrence. obtained with the three MAP algorithms variants 01
plus the MAP algorithm 11s. Despite the fact that
III. THE LOGMAP ALGORITHM. for signal to noise ratios inferior to 1 dB the
IMPLEMENTATION. performances are identical, up to this value the results
show that the variant 01e is better. It is followed, in
The variants of LogMAP algorithm differs by the order, by the variants: 01s, 11s and 01h. These results
correction term approximation way, described in show that at low signal to noise ratios the errors are
equation (10): produced exclusively by the bad selection of the path
in the trellis and up 1 dB the errors due to the trellis
fc(x) = ln(1+e-x) , x 0 (12) non closing have a higher weight.
In figure 5 are represented the BER(SNR) curves
The functions that approximate fc(x) must be easy to obtained with the four variants of the circular MAP
implement and they must reproduce the most exactly algorithm already defined in comparison with the best
possible the form of this function. Two MAP algorithm: 01e. The first three circular MAP
approximations were proposed in this paper, indicated variants have similar performances, inferior to the
in Fig. 2 and Fig.3. performances of the variant 01e.
Tacking into account all the results already presented
it results that the hard variant is not a good solution.
fc(x) The simulation results realized with 1/3 rate RSC
turbo code (parallel) with G=[1,5/7], which utilizes in
gc(x) the variant 00s the LogMAP algorithms are compared
in Fig.6 with the results obtained with the best MAP
variant: 01e. From figure results that all the two
LogMAP variants are better than the MAP at least for
values of the signal to noise ratio inferior to 1 dB. Up
this value the curves are not very accurate but
x obviously the performances are similar. The curves
reduced precision is due to the reduced volume of
Fig.2. The rectangular approximation way.
simulations.
- Despite the fact that practical implementations of the
LogMAP algorithm are faster than those of the MAP
fc(x)
algorithm the simulation programs work slower in the
case of the LogMAP algorithm.
hc(x) Between the two LogMAP algorithm variants the
results show that the linear one is better. These
conclusions must be verified also for other component
codes.
35
BER
uncoded
MAP01s
MAP01h
MAP01e
o MAP11s
SNR
Fig. 4 The BER curves obtained with: MAP01s, MAP01h, MAP01e, MAP 11s algorithms.
BER
uncoded
MAPC1
MAPC2
MAPC3
o MAPC4
+ MAP01e
SNR
Fig. 5 BER performances of C1, C2, C3, C4 algorithms versus MAP01e algorithm.
BER
uncoded
MAPC1
MAPC2
MAPC3
SNR
Fig. 6 BER performances of rectangular and linear LogMAP algorithms versus MAP01e algorithm.
36
V. CONCLUSIONS
VI. REFERENCES
37
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Transmission
transmission system, which utilizes the forward error
Channel
correcting by codes concatenation and iterative decoding Iex21
(turbo coding). There have been investigated all the I I I
DI
systematic convolutional codes having the constraint
C2 x2 y2 DEC2 Iex12
length K less or equal to 6, under three diffrent
concatenated forms: parallel PCCC (pure turbo code),
serial SCCC and hybrid HCCC, at the following rates: a)
u
1/3, 1/4 and 1/3, respectively, all unpunctured.
Two interleaver types were used: pseudo-random
and S-interleaver having the same length, 1784. C1 MUX x , x
1 2 y1, y2
The AWGN channel and the BPSK modulation LLR1
Transmission
were employed. I DMUX
channel
The used iteration number was eight. For
increasing the work speed an iterations stop criterion I DEC1
DEC2 DI DMUX
was used. When the resulting error number from the
decoding of a data block is zero, the remaining iterations C2 I MUX
x3 y3
are not effectuated, passing to the next block.
For decoding, the MAP, MaxLogMAP and Log b)
MAP algorithms were used. In all the cases, a tail off
LLR1
was employed for the first code, with the decreasing x0
u y0
transmission rate price. C1 x1 DEC1
y1
The transmitted data block numbers for a
Transmission
c)
I. INTRODUCTION Fig.1 a) Parallel concatenated convolutional codes, b) Serial
concatenated convolutional codes, c) Hybrid concatenated
convolutional codes
Two procedures which improve the performances of
convolutional codes, CCs, (and block codes), from Blocks I and DI realize interleaving and
point of view of error rate (BER), are concatenation deinterleaving functions. We used in this paper two
and iterative decoding. interleavers: pseudo-random [2] and S-interleaver [3].
Fig.1 illustrates the possible ways of convolutional Our simulations prove that the interleavers have an
codes concatenation. Concatenated convolutional essential influence on performances of Turbo codes.
schemes tend to fall into three categories: parallel DEC1 and DEC2 are iterative decoder blocks [4],
concatenated convolutional codes, PCCCs, (as in which implement algorithms like: the MAP algorithm,
Fig.1 a)), serial concatenated convolutional codes, the first and the must important, proposed by Bahl and
SCCC, (as in Fig.1 b)) and hybrid concatenated al. [5], the MAXLogMAP algorithm [6], and the Log
convolutional codes, HCCC, (as in Fig.1 c)). MAP algorithm, proposed by Robertson and al. [7].
The PCCC (or Turbo code) was introduced by Berrou The constitutive codes can be convolutional codes or
et al. [1], in 1993, and it was the beginning of turbo block codes. In this paper we studied the first
code era. exclusively. The general scheme of a recursive
1,2
Facultatea de Electronic i Telecomunicaii, Departamentul
Comunicaii Bd. V. Prvan Nr. 2, 300223 Timioara, balta@etc.utt.ro; kmaria@etc.utt.ro
38
systematic convolutional code, RSC, is shown in Fig. The two polynomials attached, a and b, define in
2 a) and an example of RSC is shown in Fig. 2 b). totality the convolutional code. The maximum degree
of a and b polynomials give the coders memory
(equal with K-1) and it is a measure of the complexity
systematic
and of the volume of the effectuated computation of
each component decoder. It increases exponentially
x[n] b1 b2 bM-1 bM with K or M [6].
+ D D D Practically, convolutional codes with K=36 are used.
input
Table 1 shows the generator polynomials with degree
a0 a1 a2 aM-1 aM
w[n] inferior or equal with 5 which can be selected like a
or b. Because of the restriction that a and b to be
+ prime the table shows also the possible divisors for
each polynomial.
y[n]
output Table1
Ireductibil
Primitive
Degree
a) Polynomials in octal and their
divisors
systematic
x[n] 0 * 1 1 0 0 0 0 0 0 0
+ D D 1 * * 3 1 0 0 0 0 0 0 0
input
2 5 1 3 0 0 0 0 0 0
+ * * 7 1 0 0 0 0 0 0 0
3 11 1 3 7 0 0 0 0 0
y[n] * * 13 1 0 0 0 0 0 0 0
output * * 15 1 0 0 0 0 0 0 0
b) 17 1 3 0 0 0 0 0 0
4 21 1 3 0 0 0 0 0 0
Fig.2 Recursive systematic convolutional code: a) general
* * 23 1 0 0 0 0 0 0 0
scheme, b) example
25 1 0 7 0 0 0 0 0
The following ecuations result: 27 1 3 0 0 15 0 0 0
* * 31 1 0 0 0 0 0 0 0
33 1 3 7 0 0 0 0 0
M
y[n] = a k w[n k ] (1) 35 1 3 0 13 0 0 0 0
k =0 * 37 1 0 0 0 0 0 0 0
5 41 1 3 0 0 0 0 0 37
M 43 1 0 7 0 15 0 0 0
w[n] = bk w[n k ] + x[n] (2) * * 45 1 0 0 0 0 0 0 0
k =0 47 1 3 0 13 0 0 0 0
* * 51 1 0 0 0 0 0 0 0
M 53 1 3 0 0 0 0 31 0
Y (D ) = a k D k W (D ) (3) 55 1 3 7 0 0 0 0 0
k =0 * * 57 1 0 0 0 0 0 0 0
61 1 0 7 13 0 0 0 0
M 63 1 3 0 0 0 0 0 0
W (D ) b k D k = X (D ) (4) 65 1 3 0 0 0 23 0 0
k =0 * * 67 1 0 0 0 0 0 0 0
71 1 3 0 0 15 0 0 0
M * * 73 1 0 0 0 0 0 0 0
k
ak D (5) * * 75 1 0 0 0 0 0 0 0
Y (D ) k = 0 77 1 3 7 0 0 0 0 0
=
X (D ) M k
bk D
k =0 II. EXPERIMENTAL RESULTS
The code generator matrix, G(D), is: Table 2 and Table 3 present the simulation results
obtained with parallel concatenated convolutional
a (D )
G (D ) = 1, (6)
code, Fig.1 a), rate R=1/3, RSC, with the pseudo-
b (D ) random interleaver (table 2), and S-interleaver (table
3).
39
Table 2 (pseudo-random interleaver)
1 3 5 7 11 13 15 17 21 23 25 27 31 33 35 37
1 0 2954035 3021300 1935725 2864349 1188340 1177130 1093049 3157698 1073326 1821748 710482 1242932 585061 560538 704876
3 3612356 0 0 1358637 0 527925 498583 0 0 243060 156083 0 318487 0 0 392376
5 4516335 0 0 100348 0 61305 62825 0 0 5311 1471412 0 5259 0 0 20042
7 1262331 919632 5710 0 0 5163 6663 378923 1419 1445 0 8216 1916 0 5840 289992
11 4684497 0 0 0 0 37148 37295 0 0 47924 0 0 37591 0 0 437
13 4684497 33594 6755 7043 1720 0 312 292 220 3144 243 8839 265 1192 0 155
15 108671 34188 9633 6457 1294 235 0 403 204 579 313 0 2910 1077 7302 229
17 445882 0 0 230633 0 9302 9040 0 0 8749 1741 0 10009 0 0 249017
21 4886023 0 0 13579 0 11594 11975 0 0 176919 159130 0 173697 0 0 8037
23 178171 20128 1102 521 2061 1629 230 691 6165 0 578 476 1032 1283 3005 586
25 1420029 5267 966445 0 0 751 1159 926 86082 34388 0 3327 1676 0 1904 1999
27 21498 0 0 6201 0 12060 0 0 0 345 1105 0 373 0 0 546
31 429482 6548 1855 600 2888 418 1823 344 10420 3162 207 828 0 776 534 1443
33 10949 0 0 0 0 2602 2471 0 0 417 0 0 450 0 0 1374
35 10609 0 0 6446 0 0 9675 0 0 241 558 0 368 0 0 270
37 135016 123269 1379 114297 1445 236 449 104996 956 757 10480 1418 737 901 731 0
41 4107943 0 0 2430 0 1991 1651 0 0 2242 2031 0 1643 0 0 0
43 424030 8621 7204 0 0 456 0 2488 6217 551 0 0 2140 0 6606 3117
45 299338 13935 5521 1638 3316 901 555 6442 1873 12670 4607 854 3971 5999 4711 13557
47 38916 0 0 10547 0 0 3771 0 0 1679 9612 0 13089 0 0 6179
51 205797 7216 6101 1502 3468 678 630 2024 1006 3910 5735 3119 17711 5515 3467 15763
53 9846 0 0 599 0 2125 6769 0 0 962 681 0 0 0 0 1890
55 16898 0 0 0 0 3796 3708 0 0 1940 0 0 13015 0 0 16663
57 5160 1513 2592 2879 913 748 4673 879 4271 1351 1768 8903 30764 20257 3587 2084
61 510340 12737 2421 0 0 0 607 721 6592 4919 0 18487 744 0 0 5890
63 125731 0 0 38990 0 29225 38412 0 0 39576 43539 0 36760 0 0 38388
65 28245 0 0 698 0 2550 3709 0 0 0 2436 0 2649 0 0 1497
67 2075 8583 2910 2383 8190 355 2117 607 22152 26281 20200 1009 2018 1191 3883 11581
71 21641 0 0 7941 0 18221 0 0 0 4869 8983 0 945 0 0 2437
73 2587 6523 3030 4425 12835 5061 471 313 114070 997 16231 1540 12697 977 424 3386
75 9336 4653 11238 1249 968 12662 961 2685 6851 104145 2325 7577 781 16290 1835 450
77 94043 0 0 0 0 1404 2169 0 0 1751 0 0 2123 0 0 73523
41 43 45 47 51 53 55 57 61 63 65 67 71 73 75 77
1 2558092 1291785 1193479 523168 1207492 480732 644618 277521 1286689 589732 508488 246833 586229 251741 312032 409436
3 0 145046 50884 0 41916 0 0 31688 119502 0 0 137957 0 139590 32654 0
5 0 6347 50865 0 52682 0 0 20946 5037 0 0 1560 0 1861 23048 0
7 348 0 262 13007 216 329 0 737 0 257 363 18076 11557 14284 351 0
11 0 0 43879 0 44683 0 0 401 0 0 0 416 0 445 662 0
13 7029 416 199 0 250 1679 1008 195 0 392 242 153 215 1689 265 424
15 1396 0 198 122 151 482 1639 499 370 282 1975 2524 0 466 235 359
17 0 8663 1597 0 5878 0 0 8822 9446 0 0 8446 0 8800 9174 0
21 0 155602 9816 0 9225 0 0 1372 159927 0 0 14240 0 12664 3965 0
23 1917 254 616 140 17637 261 604 368 3818 1717 0 1465 13100 471 14498 2881
25 1267 0 865 15090 359 2107 0 2429 0 4654 2191 17575 15184 9580 1931 0
27 0 0 586 0 1273 0 0 4623 23549 0 0 297 0 585 13015 0
31 6477 3924 21692 5640 332 0 490 27704 250 1406 185 379 253 3520 365 14592
33 0 0 13682 0 16959 0 0 6718 0 0 0 1994 0 2053 2180 0
35 0 12497 820 0 265 0 0 15250 0 0 0 547 0 251 3434 0
37 0 617 6125 912 13870 3399 21406 885 723 229 5265 1037 837 972 909 90250
41 0 52386 51113 0 54221 0 0 3930 44754 0 0 74738 0 11696 4972 0
43 12883 0 7156 5074 4128 6676 0 2928 0 8412 18832 18168 0 14370 14586 0
45 7550 7623 0 4591 23549 16302 4950 16167 4498 37402 2142 19979 28922 17354 20076 24917
47 0 29280 13511 0 24500 0 0 13531 0 0 0 13803 0 16293 113221 0
51 8407 4568 39701 11210 0 525 12873 24937 3149 14515 129624 24140 3039 15473 27847 69808
53 0 11421 16531 0 4511 0 0 23816 149866 0 0 13064 0 9364 7155 0
55 0 0 8543 0 6700 0 0 21371 0 0 0 11158 0 14395 11214 0
57 17131 27777 26594 1752 14839 10324 8576 0 23236 15364 10750 15603 6537 6747 23012 6439
61 10064 0 5965 0 2505 23247 0 35109 0 3933 3802 9639 3637 89620 14001 0
63 0 7429 35644 0 29862 0 0 19276 19300 0 0 6196 0 14240 18842 0
65 0 113080 5937 0 20680 0 0 14321 7637 0 0 19730 0 14989 4491 0
67 12965 25181 13423 15669 12916 7387 7400 13738 3109 13093 9074 0 6612 2022 20297 30407
71 0 0 15047 0 13372 0 0 14385 5929 0 0 14642 0 12603 7323 0
73 12240 1979 40387 16590 24624 14869 5216 19117 35211 7683 12892 3100 4207 0 6536 93671
75 15321 20015 23252 40156 9457 106484 6020 112287 78940 18715 2690 26072 2081 9212 0 18490
77 0 0 33231 0 26997 0 0 1144 0 0 0 1291 0 1185 1035 0
We used an AWGN noise and a BPSK modulation. The tables contain bit error rate (BER108) obtained
All the simulations were made for signal/noise ratio for each polynomial pair indicated in octal, at the
equal with 1 dB and for a number of 500 errors, at beginning of each row (the denominator, b(D), from
least. relation 6), or column (the numerator, a(D), from
An iterations stop criterion was used for each decoder. relation 6).
When the resulting errors number for a data block is
zero, the remaining iterations are not effectuated,
passing to the next block.
40
Table 3 (S- interleaver)
1 3 5 7 11 13 15 17 21 23 25 27 31 33 35 37
1 0 2782307 2943825 1786715 3066143 1207492 1339953 1123232 2853139 1250243 1811238 611821 1159192 575409 691787 657002
3 3713565 0 0 1513452 0 480592 586229 0 0 188714 158108 0 235426 0 0 515695
5 4532351 0 0 38628 0 54530 45274 0 0 2482 1401345 0 3080 0 0 11095
7 1040109 1022982 3436 0 0 2994 2956 305910 832 817 0 6151 850 0 4025 241462
11 3979820 0 0 0 0 36253 41956 0 0 41470 0 0 35644 0 0 477
13 3979820 24035 5447 4684 3999 0 143 242 167 763 185 1311 479 182 0 100
15 118335 36556 7559 4703 1166 267 0 185 266 738 145 0 707 401 1421 228
17 394618 0 0 162101 0 332 365 0 0 302 3075 0 307 0 0 193675
21 4155989 0 0 594 0 781 830 0 0 42699 46773 0 48925 0 0 10501
23 186127 5812 539 658 1333 752 144 333 24686 0 313 743 6509 1428 2622 794
25 1198150 2598 1069026 0 0 539 162 640 2893 4516 0 731 2887 0 908 758
27 7770 0 0 2034 0 1874 0 0 0 258 732 0 506 0 0 432
31 180019 6177 1006 568 2092 220 939 501 5303 5826 437 1721 0 945 439 1285
33 5135 0 0 0 0 1273 1560 0 0 302 0 0 252 0 0 1161
35 9702 0 0 2503 0 0 3214 0 0 460 487 0 439 0 0 324
37 102566 114577 322 76369 275 385 317 81253 543 397 4446 565 421 570 404 0
41 4212043 0 0 1934 0 1139 824 0 0 639 1355 0 903 0 0 0
43 186387 5332 2908 0 0 1305 0 2480 9810 550 0 0 1105 0 15923 6397
45 359028 10573 6894 2182 7325 413 737 2858 1180 2040 3133 2808 4826 15890 3769 14232
47 9265 0 0 13877 0 0 2067 0 0 992 9972 0 7480 0 0 9044
51 228619 14099 4668 2885 2697 1039 1665 9634 1735 8191 3568 3668 8333 13754 1056 13976
53 12047 0 0 1625 0 8100 2027 0 0 777 1633 0 0 0 0 2386
55 29230 0 0 0 0 4696 6396 0 0 3125 0 0 1994 0 0 6292
57 8908 4020 1333 2419 612 1843 19325 949 4723 739 16047 1396 71654 16720 4934 4592
61 227980 4904 2968 0 0 0 846 703 6697 2943 0 10200 291 0 0 3813
63 10090 0 0 263 0 417 384 0 0 307 167 0 190 0 0 387
65 8205 0 0 817 0 2784 2383 0 0 0 1867 0 2581 0 0 1138
67 3536 4241 1694 2789 12315 190 2025 287 28519 130035 26804 1066 648 1183 1525 2540
71 6549 0 0 2588 0 2093 0 0 0 2634 2889 0 1310 0 0 24800
73 3793 12398 1442 43816 15414 2714 672 788 33805 1276 113225 12719 19350 2428 714 3802
75 12376 2481 3854 2155 1844 10375 687 377 11232 19005 8164 5724 329 15031 934 2206
77 70523 0 0 0 0 2842 2491 0 0 5198 0 0 3437 0 0 48073
41 43 45 47 51 53 55 57 61 63 65 67 71 73 75 77
1 2993273 1207492 1223435 556141 1188807 549547 717488 262103 1134529 583417 583417 260213 494650 250240 333025 367943
3 0 131135 31141 0 28824 0 0 39378 117391 0 0 104040 0 102492 33592 0
5 0 3172 41831 0 40679 0 0 9642 3278 0 0 358 0 727 8419 0
7 276 0 127 9802 74 197 0 373 0 258 125 8786 8039 11305 363 0
11 0 0 41854 0 49803 0 0 508 0 0 0 485 0 519 569 0
13 1649 83 80 0 61 386 473 127 0 147 935 196 121 396 193 287
15 1299 0 286 94 73 1313 499 208 186 128 518 350 0 342 122 575
17 0 333 532 0 533 0 0 335 335 0 0 488 0 782 402 0
21 0 45040 1133 0 588 0 0 1887 35638 0 0 1074 0 2472 2761 0
23 11315 233 1940 136 25743 437 1122 391 3591 2305 0 2710 4540 330 15135 2778
25 1331 0 538 13856 204 1010 0 1071 0 15702 878 13606 19005 9179 708 0
27 0 0 202 0 752 0 0 982 16096 0 0 520 0 249 152923 0
31 13931 2866 13021 16983 425 0 600 69932 269 1091 217 286 167 13009 207 4345
33 0 0 7457 0 14444 0 0 17526 0 0 0 919 0 1416 1690 0
35 0 23136 258 0 336 0 0 7477 0 0 0 218 0 335 2291 0
37 0 731 5433 390 22609 3707 14937 376 331 384 4299 1593 534 923 798 58997
41 0 36471 47788 0 40759 0 0 30444 46317 0 0 18534 0 401227 6071 0
43 9174 0 2550 2094 1518 4200 0 4059 0 5968 20458 20065 0 444 15899 0
45 10600 14247 0 23325 26952 22341 4008 22238 5201 31271 3309 16184 8513 98094 24113 10032
47 0 2094 5453 0 16476 0 0 3385 0 0 0 4066 0 30534 26549 0
51 20778 1979 25027 13750 0 2564 4858 21800 8835 52297 15758 99546 3659 14286 16219 68429
53 0 28429 44149 0 3700 0 0 4375 206608 0 0 14589 0 15480 13301 0
55 0 0 16959 0 14568 0 0 5918 0 0 0 15839 0 923983 6148 0
57 22967 36629 28408 616 8852 1526 3734 0 6688 29861 39126 15876 27662 17594 32804 18244
61 13068 0 3898 0 2653 23247 0 32727 0 2570 3929 3518 686 64571 4821 0
63 0 19832 53766 0 30534 0 0 8299 3021 0 0 7494 0 13151 30265 0
65 0 65910 14435 0 26699 0 0 8136 10792 0 0 33623 0 39829 3974 0
67 61580 63321 111671 5047 66312 7881 24005 29670 21168 10619 13226 0 18438 1958 19668 16879
71 0 0 27579 0 12537 0 0 192134 7742 0 0 22430 0 20300 5443 0
73 27606 3694 31173 17325 108604 14879 4281 17458 122750 99527 2136 3568 14395 0 61241 25116
75 24360 10319 17105 15974 18931 12671 2340 119031 27355 24284 3416 22923 1731 27310 0 48372
77 0 0 27526 0 27316 0 0 64 0 0 0 1055 0 1720 582 0
The zeros in the tables contain mark that the be remarked the performances obtained using, for the
respective codes have common divisors (see Table 1) Feed Back loop, the polynomials b1=7=111,
and can not be used together. With little exceptions, b2=13=1011 and b3=15=1101 as in combination with
the results obtained with S-interleaver are superior to the polynomials with the same degree (ex. 15/13 or
the results corresponding the pseudo-random 13/15) as in combination with the polynomials with
interleaver. We also remark the superior results superior degree (ex. 51/7, 51/13 or 51/15).
obtained with both interleavers in the case of the use Good performances are obtained using the Feed Back
at denominator, b(D), of the primitive polynomials, with b4=23=10011, indicated in [2] (25/23, 33/23,
comparing to non-primitive polynomials, at the same 37/23). The last two combinations (33/23 and 37/23),
constraint length. Despite of the fact that global correspond to the situation when the pseudo-random
performances increase proportionally with K, it must interleaver is superior to S-interleaver.
41
In the following are presented some practical results.
BER
BER
x uncoded uncoded
MAP, P MAP
o MAP, S MLMAP
MLMAP, P LogMAP
+ MLMAP, S
LogMAP, P
* LogMAP, S SNR
BER
BER
uncoded
MAP
MLMAP
uncoded LogMAP
5/7 code
15/13 code
37/13 code SNR
o 51/13 code
Fig.8 Simulation of rate 1/3 SCCC, 15/13 code, S-interleaver.
SNR
The results obtained with different decoding
algorithms (MAP, MaxLogMap and LogMAP), with
Fig.4 Simulation of rate 1/3 PCCC, P-interleaver, MaxLogMAP, two types of interleavers: pseudo-random and S-type
for 5/7, 15/13, 37/13 and 51/13 codes.
interleaver, with S=29, are comparted in Fig.3. In all
the cases we used the 5/7 code (the most performant
from all codes which have K=3) with data blocks size
equal with N=1784 bits. We can remark the
superiority of S-interleaver versus the pseudo-random
BER
42
codes can lose their supremacy. This verification [5] L.R. Bahl, J. Cocke, F. Jelinek, and J. Raviv, Optimal
Decoding of Linear Codes for Minimising Symbol Error Rate,
constitute the objective of a future study which we IEEE Transactions on Information Theory, Vol. 20, pp. 284-287,
propose us. March 1974.
The diagrams, for the last three figures, show the BER [6] L.Hanzo, T.H.Liew, B.L.Yeap, Turbo Coding, Turbo
performances obtained with different concatenation Equalisation and Space-Time Coding for Transmission over Fading
Channels, John Wiley & Sons Ltd, England, 2002
modes (parallel, hybrid and serial) and with different
[7] P. Robertson, E.Villebrun, P.Hoeher, A Comparison of
algorithms (MAP, MaxLogMAP and LogMAP). In all Optimal and Sub-Optimal MAP Decoding Algorithms Operating in
the cases we used the 15/13 code and the S- the Log Domain, Proceedings of the International Conference on
interleaver. Communications, Seattle, USA, pag. 1009-1013, iunie 1995
Obvious, the serial concatenation is less performant
than parallel and hybrid concatenations. Because of
multiplexing and restriction using of interleavers with
the same length, N=1784, in the case of SCCC code it
results a number of N=1784/2=892 information bits
per block, than N=1784 for PCCC and HCCC codes.
Moreover, because the SCCC transmission rate is of
1/4 versus 1/3 in the case of PCCC and HCCC codes,
for the equivalence of the ratio between the
transmitted energy in information bit (E0) and the
noise power spectral energy (N0/2), for the 1/4
transmission rate case, the channel noise power is
higher than the corresponding power for the 1/3
transmision rate case.
We remark the falling of the SCCC codes, for SNR of
2 dB, falling what its not find in the case of the
PCCC and HCCC codes.
Finally, we also notice the good behavior of the
LogMAP algorithm in the case of SCCC and HCCC,
fact which invite us to make an investigation more
detailed of this algorithm for the future.
III. CONCLUSIONS
REFERENCES
[1] C. Berrou, A. Glavieux, P. Thitimajshima Near Shannon limit
error-correcting coding and decoding: Turbo-codes, Proc.ICC93,
Geneva, Switzerland, May 1993, pp. 1064-1070.
[2] Consultative Committee for Space Data Systems Telemetry
Channel Coding, CCSDS 101.0-B-6, Blue Book, October, 2002.,
http://www.ccsds.org/documents/101x0b6.pdf
[3] P. Ha, A Fast Algorithm for Generating Random Interleavers,
www.sarim.changwon.ac.kr
[4] J.Hagenauer, E. Offer, L. Papke, Iterative Decoding of Binary
Block and Convolutional Codes, IEEE Transactions on
Information Theory, Vol 42 No 2, March 1996 pp 429-445
43
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
2. The contents
The writing has three sections.
In Section I one remember the obtaining of a VSB-
QAM signal using three level quadrature carriers and
two data flows: the first, divided data (DD) is
obtained from the (DCD) by dividing it in a T type
flip-flop; the second is the modulo-2 sum of this two
flows. The two operations result in a flow of Gray-
coded dibits. They possess a structure of Hilbert pairs.
In Section II we present three demodulation methods.
Fig.1. Data representation: a) On-Off Keying (OOK); b) DCD; c)
The first method uses a QAM procedure, but the Bit timing, is the carrier for the DBP code. d) Differential biphase
decision rule is with floating threshold. code, with binary 00/1800 jumps; e) DC Signal: a transition at half-
interval represents a "1" in DCD data ak. f) P is the In phase
carrier, with 00/1800 phases. g) AMI (bipolar) code with "0 Volt"
for logic zero and alternating "1 V" for logic 1;
1
Facultatea de Electronic i Tc., UPB, Catedra de
Telecomunicaii, Bd. Iuliu Maniu 3-5 Bucureti e-mail:
sorin.popescu @comm.pub.ro
44
The OOK data will be denoted bk and called absolute
coded data (ACD). DCS is a quaternary DPSK signal.
Data bk must be differentially coded, leading to the
data sequence ak (DCD). The ACD produce for P a
phase change of -90 when bk=1 and no change when
bk=0. We will have 00/1800 when ak=0 and -900/-2700
when ak=1.
Fig.5 - QAM ternary carrier: carriers with phases 1800 and 2700
are obtained by inverting P and Q. Fig.10c. The resulting
schematic; N1, N2 are BCD numbers. G1, G2 are Gray coded bits
Fig2 - Standard Miller detector .The detector works in two steps. It
decides Mimod: a. the DCD a k and b. the bit differences ACD b k . The synthesis of the bit streams G2,G1 is obtained
based on the constellation from fig.5 and the
If the phase is e.g. 90/270, it changes during T and the modulation steps kT. The phases l900 with l=0,1,2,3
sum will be ak=1 (Fig.4, Mimod signal). Then one written in BCD provide the N2N1 bits. By writing
obtain bk=akak-1. The drawback is that it needs a G2G1 as being the coordinates of the projections of
clock with double frequency, 2vs . By division it is the signal phasors on the axes, one can obtain Gray
possible equally to obtain an inversed clock, =1800, coded data (GC), as illustrated by the QAM scheme in
sampling at the points denoted by xx in Fig.4. The so Fig.6, right (see10c). Projections of the phasors k on
decided data are erroneous and the recovery is the axes and K take the values on the constellation;
difficult, after few hundred of erroneus bits.. This is then G2 G1 also takes the necessary values. Fig.6
the phase ambiguity we mentioned previously illustrates the PSD of the VSB signal.
I.3 DC modulation is a VSB -QAM with ternary Fig.6- The VSB spectrum; spectra of the Gray quadrature data
carrier (VSB-QAM-TC) flows are visible as odd and even functions
Section II New demodulation methods
DCS can be built by modulating two quadrature
carriers by two data bit streams at the vs rate, which II.1 Coherent quadrature detection floating
are then summed up. Summing displaced binary threshold decision
carriers result in a ternary one. So it is necessary these a. Receiver model
carriers to be ternary to resulting in a binary signal. In The channel is simulating a cooper pair in a city
Fig.4 the constellation 4QAM is rotated by 450 w.r.t. network. The noise is placed at the channel entry or
P and Q. In Fig.5 we have the ternary carriers and output. The carriers P, Q have fs/2 frequency and are
K; their sum and difference are P and Q signals. rectangular. The two low pass post detection filters
45
have a cut-off frequency of fs/2 too, at the Nyquist error rate is presented in the tables below together
limit. The idea of the method is becomes clear by with standard Miller, denoted Mold and a base band,
observing the three level demodulated signals, NRZ code, Mnyq. In the first hypo thesis the noise is
amazingly the same, in fact complementary. If one of placed at the channel issue, in the second is at the
them is big (1 or 1) the other is 0. The succesion is entry. All upper mentioned names receive a 2 at the
1,0.-1,0,1,-1.. end (e.g. Misnd2).
Table 1
ak differ.data 00 1 0 1 1 000 1 0 1
k sent phases 0 0 90 180 90 90 0 0 0 90 180 270
p=inphase sign. 1 1 0 -1 0 0 1 1 1 0 -1 0
q=quadr. signal 0 0 1 0 -1 -1 0 0 0 1 0 -1
|u| |v| = -a k,+dc 1 1 -1 1 -1 -1 1 1 1 -1 1 -1
|u|-|v| s c = -a k 11 0 1 0 0 111 0 1 0
In the example from Table 1 a k are arbitrary. The
modules difference is finally DCD a k. Let +/-A be the
amplitude of p and q. The decision may be effected by
simple rectifying one of them and comparing it with a
threshold of A/2 value. The noise can be at most A/2.
Fig.7 The transmission system: Miller encoder sends DCS to a Table 2 BER for noise at the issue
low pass filter as a channel with additive white Gaussian noise. Snr/ber 4 dB 6 dB 8 dB 9 dB
Two multipliers translate the line band in the base band, aided by
two l.p.f. abs(p,q) represent modules, is an adder/substractor. Msnd 5e-3 .9e-4 1.6e-4 8.6e-6
The name of this scheme is Misnd. The signals are thresholds each
other and are noisy. is used as an adder only for computations Mold 2.8e-1 1.3e-1 3.5e-2 1.3e-2
purposes.
In our solution (fig.7) a threshold circuit decide Mnyq 1.8e-1 8.6e-2 2.7e-2 1.1e-2
between a voltage A and an electric 0 on the threshold snr/ber 10 dB 11 dB 12 dB 13 dB
entry, usually set at 0 voltages. But herein this entry is
hot because always there exists the noise. In fact it Mold 5.2e-3 1.4e-3 2.6e-4 1.2 e-5
happens on both entries. The sum will be of double Mnyq 4.9 e-3 1.5 e-3 4.2 e -4 7.5e-5
power (3dB) wile the permitted noise can be at most
-4
A. These get the advantage of 6 dB for the signal. The Mnyq has 12dB (11.4) for BER=10 . Figure 4.2 is the
noise increases with only 3 dB. It can be expected a effect of the descrambler. Miller old is equivalent
large advantage, of 3 dB. with a 4 PSK. Remember the penalty of double BER
Our simulations implied a plain old Miller receiver, for DPSK. So, our method is 5 dB superior to the
denoted Milvec(no) for old or new structure concern classic Miller decoder. See for all that the detection
ing the placement of the noise source in the transmi has included two LPFs.
ssion chain. In fig.8 it is put at the receiver entry, at
the end of the channel. The eye diagrams are plotted Table 3 BER for noise at the entry
on the two branches in fig.8. The three levels are snr/ber 4 dB 6 dB 8 dB 9 dB
visible and the sampling points are too. The Msnd e
2.2 -3 e
5.8 -5 e
6.4 -6 -
transitions in the above part figure are producing only
e e
between the neighbouring levels. The intersymbol Mold 2.5e-2 3.4 -3 3.3 -4 2.3e-5
interference ISI is not present. In the figure below Mnyq 4e-3 6.6e-4 8.1e-6 -
there exist transitions between extreme levels. But the
ISI is not them imputable. At a careful insight one
observe on the two figures two adjacent and not When the noise is inserted at the entry (table3) it
identical eyes in every bit interval. They are displaced undergoes a loss of 9.5DB. The hierarchy remains
each other and the second is sampled in a wrong but the differences are smaller. However, an effective
moment. So the two branches are not equally noise consistent advantage of 2.5dB is real. 1dB is due to
protected. This method will be referred as Msnd. The the post detection filters. A 1.5 value results from the
46
peculiar probability density function of the module subtracting the signal is doubled but the noise is only
difference (see later). increased by a sqrt2 factor.
Snr/ber 5 dB 6 dB 7 dB 8dB
e e e
Mddds 8 -5 3.6 -5 3.3 -6 <1e-6
Mold 7e-3 3.4e-3 7e-4 3.3e-4
dk,++dk,-=ak 00 1 0 1 1 000 1 0 1
47
p= cos(k+ ), q=-sin(k+ ); f 2
x2
2 f / 2
d( f ) =
2 2
4
e e dx ; (8)
f / 2
2
f
2
(9)
d( f ) = erf ( f / 2 );
2
4
e
h = u v,v > 0,u > h;d(h) = g(v)g(h + v)dv; (10)
0
h2 (v + h / 2) 2
2
d (h) =
2 2
4
e e dv (11)
2 0
h2
2
d (h) = erfc ( h / 2 )
2
4 (12)
e
Figure 12 PLL for data aided carrier recovery 2
e y dy = F ( ) F ( x ); (13)
2
erfc ( x ) =
x
2
h2
1 Figure 14 Statistic distributions for the nonlinear processed noises
d (h ) = e 4 erfc ( h / 2 ); for h 0
2
2
(4) The error probability is to be approximated by:
2 x2 / 2 1
pe,new = d(h)dh
2
e dx =. 2 e y / 2 (14)
Both u and v has normal, Gaussian distribution: y y x y
2
u2
2
e 2 2
for u 0 (5) erfc ' ( x ) = F ' ( x ) = e x (15)
2
g ( u ) = d ( u ) = 2
0 for u < 0
We retain the first term from the Laurent series:
2
ex a1 a2 a3
erfc ( x ) = ( + ...) (16)
x x3 x5
where ak are obtained from derivative of erfc and
identifying:
2
ex 1 13 135
erfc ( x ) = (1 + + ....) (17)
x 2x2 4x4 8x 6
48
1 c) A graphic image for sums and difference of
lg = lg p e , gauss lg p e , new = 0.196 + n signal / noise ; (21)
10 powers of the rectified noises as a.c. phasors; there
virtual angle has peculiar proprieties.
III.2 How to explain some numbers
Bibliography
Now, lets expose a peculiar result. If we denote by:
[1] N.D. Alexandru, Dae Young Kim: Spectral Shaping via
2 = f 2 = f 2d( f )df = 3.27325.175d B; =1.809 (22) Coding , Ed. Cermi, Iasi 2003
0 [2] N.D. Alexandru, G. Morgenstern: Digital Line Codes and
Spectral Shaping Ed. Matrix, Buc. 1998
s = h = h2d(h)dh= 0.7271.39dB;s =0.8525
2 2
; [3] W.R.Bennett, J.Davey: Data Transmission Ed. Mc-Graw Hill
(23) Book Co. 1965
[3] Gilbert Held,: Understanding Data Com munications, Sams
2
= g2 = x 2 g ( x ) dx = 1 0 dB ; = 1; (24) Publishing 1996
[4] M. Stein, Les modems pour transmissions de donnees, Ed.
the ms values of u+v, u-v and p,q respectively, then Masson, Paris, 1987
(see fig.15) the effective values for this noises are [5] S. Popescu, Transmisia Datelor, Ed. Matrix, Buc. 2003
summing as sinusoidal signals and the angle between [6] Popescu, S: "Phase-Precession Modulation, a New Approach for
SSB-AM" - IV International Conference on Reliability of
them is obtained from: cos =2/=0.6366. Semiconductor Devices and Sistems RSDS'96 Chiinu, June 5-7,
2 = 2 + 2 + 2 2 cos = 2 (1 + cos ) = 3 .2732 ; (25) 1996, pp. 182-187
[7] S. Popescu, Modulaia cu precesie de faz, o abordare nou a
s 2 = 2 (1 cos ) = 0 .7268 ; modulaiei de amplitudine cu BLU - Telecomunicaii nr. 3/1997,
5 1 (26) pp. 9-16
cos = 2 / = 0 .6366 !; = 50 0 40 ' a cos ; [8] S. Popescu: A non-Interference Criterion for FSK Data
2 Signals Proceedings of the Symposium on Electronics and
The value for it is: =50040, very close near to: Telecommunications ETc. 2000, Timioara, November 23-24,
5 1 2000, Vol. I pp. 227-231
= a cos = 51 0 50 ' (27) [8] Popescu, S.: Delay Coding a new perspective Proceedings
2 of the IEEE International Conference Communications 2004
This is right the angle with the basement for the faces June 3-5, 2004, vol.2, pp 115-122, Bucharest, Academia Militar
of Keopss great pyramid. And is in visible relation Tehnic, Romnia
with a root of Fibonacis series characteristic
equation: x2-x-1=0.
49
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract In this paper, a study of selective image Confusion is achieved by mixing three different
encryption using IDEA algorithm is presented. operations. Each operation is executed on two 16-bit
Experimental results show a good image security inputs. These operations are:
using the proposed selective encryption. Exclusiveor (XOR)
Addition by integer modulo 216, inputs and
Keywords: IDEA, image encryption, selective output are unsigned 16 bit integer
encryption, cryptography, encryption. Multiplication of integers modulo 216+1,
inputs and output is unsigned 16-bit integer.
In this case the blocks of all zeros is treated
I. INTRODUCTION as representing 216.
Using this three operations we provide a complex
The strongest solutions for security are offered by transformation of the input, making cryptanalysis
computational cryptography. Cryptography is used for much more difficult than DES algorithm, which uses
insuring the communication confidentiality in military just XOR operation.
and diplomatic fields for a long time. During last In IDEA, diffusion is provided by the basic
years, cryptography has known a spectacular progress, building block of the algorithm, known as
many services and devices of security, which are used multiplication/addition (MA) structure (figure 2.).
in Internet is a proof of this fact. Image encryption is This structure has as inputs two 16-bit values derived
among the last applications of cryptography. It looks from the plaintext and two 16-bit subkeys derived
to have in the near future many applications in from the primary key and produces 16-bit outputs.
Internet, taking into account that fingerprints and This particular structure is repeated eight times in the
retina images will replace the numeric passwords. algorithm (figure 1.).
IDEA uses a primary key of 128 bits long. This
II. IDEA ALGORITHM primary key produces 52 subkeys with 16-bit long.
Encryption and decryption makes on 64-bit blocks.
Xuejia Lai and James Massey of the Swiss
Federal Institute of Technology developed the Subkeys Generation
International Data Encryption Algorithm (IDEA) in First 8 subkeys are taken directly from the
1999 year. The main application for IDEA is PGP primary key through segmentation in 16-bit segments.
(Pretty Good Privacy). This program is the most Then a circular left shift of 25 bit position is applied
secure and fast encryption system nowadays. to the primary key and the next eight subkeys are
Cryptographic strength of IDEA is given by: extracted. This procedure is repeated until all 52
Block length subkeys are generated.
Key length is long enough to prevent
exhaustive key searches Encryption
Confusion the cipher text should depend on Encryption schema for IDEA has two inputs
the plain text and key in a complicated way. plaintext (64b) and primary key (128b). IDEA is
Diffusion each plaintext bit should making up for 8 rounds and one output
influence every chipertext bit, and each key transformation. These algorithms divide plaintext in 4
bit should influence every chipertext bit. blocks of 16 bits. Output transformation achieves 4
outputs of 16 bits, which is concatenated and makes
chipertext of 64 bits. Each round uses 6 subkeys of
16-bits, but output transformation uses 4 subkeys
(figure 1.)
1
Facultatea de Electronic i Telecomunicaii, Departamentul
Comunicaii Str. Daicoviciu Nr. 2, Cluj-Napoca, e-mail gabifericean@yahoo.com
2
Facultatea de Electronic i Telecomunicaii, Departamentul
Comunicaii Str. Daicoviciu Nr. 2, Cluj-Napoca, e-mail Monica.Borda@com.utcluj.ro
50
inverse modulo 216 of corresponding second
and third encryption subkeys.
For the first 8 rounds, the last two subkeys of
decryption round i are equal to last two
subkeys of encryption round 9-i. [5],[6]
51
number. For this reason the number of all zero
subkeys is very high, so we looked for another
solution: we generated eight subkeys with values
between 1 and 216, than we transformed in binary and
concatenated them. After these operations we
obtained primary key with 128 bits dimension.
Next we present the results obtained for color
plane encryption on different images.
For a correct function, after the program
compilation, we must import an image and generate a Figure 6 Encryption R plane
primary key. These operations achieve on File menu
(figure 3).
52
The proposed method has as disadvantage the
necessary time for all pixels encryption. Figures 11 to
18 present results obtained for bitplanes encryption.
Each of these figures is followed by the representation
of the initial image without bitplanes encryption. The
most significant bit (msb) plane is considered the
plane 1 and the last significant bit (lsb) plane is
considered the plane 8.
Figure 16 Original image without first,
second and third planes
53
[4] Bajenescu T., Borda M.: Securitatea in informatica si
telecomunicatii. Editura Dacia, Cluj-Napoca, 2001
[5] Schneier B.: Applied Cryptography, Second Edition, Editura
John Wiley& Sons, Inc.1996
[6] Stallings W.: Cryptography and Network Security: Principles
and Practice Second Edition, Editura Prentice Hall, New Jersey,
1999
[7] Borko Furth, Darko Kirovski, Multimedia Security
Handbook, February 17, 2004
[8] Vlaicu A., Curs multimedia, 2002 (manuscris)
[9] Gibson D. Jerry, Berger T., Lookabaugh T., Lindbergh D.,
Figure 22 b. Encryption first plane Baker R., Digital compresion for multimedia: Principles and
(msb) Standards, Editura morgan Kaufmann, 1998
[10] Biryukov A., Nakahara J., Preneel B., Vandewalle, J, New
Weak Key Classes of IDEA, Advances in Cryptology, Eurocrypt
1998
[11] Daemen J., Govaerts R., Vandewalle J.: Weak Keys for
IDEA, Advances in Cryptology, Crypto93, LNCS 773, D.R.
Stinson, Ed., Springer-Verlag, 1994
[12] Marc Van Droogenbroeck and Raphael Benedett.
Techniques for a selective encryption of uncompressed and
compressed images. In Proc. Advanced Concepts for Intelligent
Vision Systems (ACIVS2002), pages 9097, 2002.
Figura 22 c. Original image without [13] Martina Podesser, Hans-Peter Schmidt, and Andreas Uhl.
first plane Selective bitplane encryption for secure transmission of image data
in mobile environments. In Proc. 5th IEEE Nordic Signal
Processing Symposium (NORSIG2002), 2002.
Through selective encryption we can offer more [14] Roland Norcen, Martina Podesser, Andreas Pommer, Hans-
security planes, the minimum for an acceptable Peter Schmidt, and Andreas Uhl. Confidential storage and
security being represented by the first plane transmission of medical image data. Computers in Biology and
encryption (msb plane). This affirmation is sustained Medicine, 33(3):277292, 2003.
by the experiment presented in figure 22.
V. CONCLUSIONS
REFERENCES
[1] http://www.byte.ro/byte95-03/vic.html [2004]
[2] Vasiu Ioana, Criminalitatea informatic, Editura Nemira,
2001
[3] Patriciu Victor, Monica Pietroanu-Ene, Ion Bica, Costel
Cristea, Securitatea informatic n UNIX i INTERNET, Editura
Tehnic, 1998
54
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Facultatea de Electronic i Telecomunicaii Iasi, Romania, Departamentul
Telecomunicaii Bd. Copou Nr. 11, Iasi, E-mail lscripca@etc.tuiasi.ro
2
***, E-mail pduma@etc.tuiasi.ro
55
The performances of the encryption algorithm will be Some coefficients combinations do not generate all
improved using algebraic functions to generate the the symbols of the GF and therefore the decoding
permutation order. process becomes catastrophic. These sequences could
not be used as encryption keys because there is no
inverse function in these cases. It is necessary to find
out which combination of coefficients ensures the
permutation of the GF symbols. The original sequence
Addition of two symbols is made modulo-2 bit-by-bit. These functions have only three coefficients which
The null element (0) does not change the result of an compose the transmission key of an encryption
addition. system.
The opposite element is the element itself. The inversed functions, defined on the same GF, are:
Multiplication of two elements is defined based on the
polynomials product of the two elements and an
irreducible m-degree polynomial p(x):
[
E k 1 ( c ) = k 1 1 ( c + k 0 ) ]
q
(12)
c = a b c( x) = a( x)b( x) mod[ p ( x)] (6)
The unit element (1) does not change the result of a The integer exponent q is the inverse key component
multiplication. which verifies that:
If the product of two elements is equal to one, than k2
(a ) q = a . (13)
they are named inversed elements:
and
a b = 1 a 1 = b, b 1 = a. (7)
(k 2 q ) mod(2 m 1) = 1 (14)
The substraction and the division are defined based on
the opposite and the inversed elements.
The existence of q for any value of k2 is quaranteed
a b = a +b (8)
1
only if m is a prime number and all the GFs elements
a / b = a b (9) have the maximum order equal to 2m-1. In fact, m and
Polynomial functions defined on GF(2m) could be 2m-1 should be simultaneously prime numbers to
used as permutation transform for different encryption ensure the maximum number of simple encryption
algorithms: functions defined on a GF(2m). We deduce some
M 1 optimum values of m: 3, 5, 7, 13, 17, 19, and 31.
k a , M = 2
i
E k (a) = i
m
1 For example, on GF(8) these couples (k2, q) are:
i =0 (10)
(2 4), (3 5), (4 -2), (5 3) and (6 6).
The sequence of coefficients from the
GF, k = [k 0 k1 ... k m ] , represents the encryption key. Other Galois fields, such as GF(16), GF(64),
GF(256), do not allow any combination of
56
coefficients for simple polynomial inversable 2
c = E1 (a) = 1 + 5a
functions because the order of some elements is less
then 2m-1. But there are some values of the
coefficients which produce inversable functions and After value encryption, it results:
these GFs could be used with few constraints.
C1 = [7 3 5 2 4 1 0 3]
III. ENCRYPTION ALGORITHMS
For position encryption let us use other functions
The polynomial inversable functions defined on GFs defined on GF(4):
could be used for symbol permutation.
2 2
For optimum m, the number of simple polynomial E 2 (a) = 3a , E 3 (a) = 2a + 2
functions defined on GF(2m) is equal to the number of
the generated permutations (except the identity one)
First function permutes the reference sequence of 4
and it is given by:
symbols (0 1 2 3) into (0 3 2 1).
The second function permutes the reference sequence
M = 2 m (2 m 1) ( 2 m 2) 1 (15) into (2 0 3 1).
Each block of four symbols will be permuted
The coefficients of the 3-coefficients functions could according to a different function. The final symbol
be randomly generated to change the permutation sequence is:
order in a fast way. C2 = [7 2 5 3 0 4 3 1].
The function could be applied directly on the data The transmitted bits stream is:
symbols to permute the bits of a symbol or indirectly,
on the position index of each data symbol from a C = [1 1 1 0 1 0 1 0 1 0 1 1 0 0 0 1 0 0 0 1 1 0 0 1].
block of 2m elements, resulting a permutation of
symbols. In this case, if the encryption functions coefficients
We call the direct method the Value Encryption are not changed, the CVPEA permutes 24 bits.
Algorithm (VEA). In a similar way, longer permutation length could be
The indirect method is called the Position obtained.
Encryption Algorithm (PEA). If VEA uses a GF with 2v elements and the PEA uses
Both algorithms could be applied simultaneously on another GF with 2p symbols, then the permutation
the data with different encryption functions, defined length of the CVPEA is:
on different GFs. The last case represents the
Combined Value-Position Encryption Algorithm L = v 2 p (bits ) (16)
(CVPEA) which is robust against the differential
attacks.
For longer permutation length, the GF dimension of
The direct method could be applied in a fast way with
the PEA has to be increased first because the
different encryption functions for short sequences of
dimension of the GF used for VEA affects harder the
symbols.
encryption algorithm complexity then those used for
VEA has no constraints but PEA is constrained to be
PEA.
applied on a sequence of exactly 2m elements.
For a higher diversity of the coded sequence, both
The coefficients of the encryption key could be fast
VEA and PEA must use larger GFs.
and randomly generated to ensure great value
The transmission key contains the GFs dimensions
diversity.
and the coefficients of the encryption functions or the
For high GFs dimensions the efficiency of the
parameters of the coefficients generator.
algorithms is increased but the processing time of the
Fast and random generation of the key components
algorithm does not become very high because only
ensures a large diversity of the encrypted sequence.
arithmetical operations defined on GFs are used.
A pseudorandom sequence generator could be used by
Example:
the VEA for faster permutation of the composing bits
Let us consider the binary data sequence:
of each symbol. In this case, a high dimension GF
should be used.
A = [1 0 1 0 1 0 1 1 1 1 0 0 0 0 1 0 0 0 1 1 0 0 1 0]
IV. NUMERICAL RESULTS
If the GF(8) is chosen for VEA, the binary data
stream is transformed into a sequence of symbols
Different GFs are analyzed to establish the number of
expressed on three bits:
permutations obtained with the simple polynomial
B = [5 2 7 4 1 0 6 2]
functions.
Small dimensions of GFs are sufficient if
A simple inversable function is applied. for the value
combinations of GFs are used to generate high-length
encryption of data:
permutations with CVPE which is very efficient, very
57
fast and hard to attack with an acceptable
computational complexity. The encryption key type is specified:
For example, if both VEA and PEA use functions S symmetric;
defined on GF(16) then the minimum permutation A asymmetric.
length of CVPEA is about 64 bits, but if we change There are 9 symmetric keys different from the identity
randomly the coefficients of the encryption functions, and 14 asymmetric keys.
then longer binary sequences are permuted. For the asymmetric encryption system, it is easy to
For higher GFs dimensions, longer sequences deduce the inverse polynomial functions coefficients
permutation is made but the computational from Table 1. There are 7 couples of direct and
complexity and the processing time are both inverse keys:
increased. (0,2,3,1) - (0,3,1,2);
A. GF(4) (1,2,0,3) - (2,0,1,3);
This is a small algebraic field with 2-bits elements so (1,2,3,0) - (3,0,1,2);
it is not efficient for value encryption but it could be (1,3,0,2) - (2,0,3,1);
used by the PEA. Position permutation is made on 4- (1,3,2,0) - (3,0,2,1);
symbols vectors. (2,1,3,0) - (3,1,0,2);
There are 4!-1 = 23 possible permutations without the (2,3,1,0) - (3,2,0,1).
identity one (0 1 2 3) (Table 1).
All these permutations could be generated using We are not interested to store the inverse function
simple polynomial functions with the maximum coefficients because the inverse algorithm depends
degree equal to 2: only on k0, k1 and q equal to k2.
k2 A decimal identifier of each permutation could be
E k (a) = c = k 0 + k1 a , k1 0,
used as the encryption key.
k 2 {1; 2}, (k 0 , k1 , k 2 ) (0, 1, 1) B. GF(8)
All these functions could be inversed: This field has eight 3-bits symbols of order 7 and it
could be used by VEA and PEA.
[
E k 1 ( c ) = k 1 1 ( c + k 0 ) ]
q
= q 0 + q1 c
q2 There are 8!-1 = 40 319 possible permutations without
the identity one (0 1 2 3 4 5 6 7) but not all these
permutations are generated using simple polynomial
The inverse functions are simple polynomial functions functions defined on GF(8).
with another set of coefficients. There are 335 simple 3-coefficients functions:
On GF(4) we use two couples (k2, q): (1, 1) and (2, 2).
k2
E k (a) = c = k 0 + k1 a , k1 0, k 2 0,
Table 1.
Encryption Key Permutation Inverse Key k 2 7, ( k 0 , k1 , k 2 ) (0, 1, 1)
k0 k1 k2 Permutation Type
0 1 1 (0,1,2,3) (0,1,2,3) Identity On GF(8) the couples (k2, q) are: (1, 1), (2, 4), (3, 5),
1 1 1 (1,0,3,2) (1,0,3,2) S (4, 2), (5, 3) and (6, 6).
2 1 1 (2,3,0,1) (2,3,0,1) S Other polynomial inversable functions defined on
3 1 1 (3,2,1,0) (3,2,1,0) S GF(8) have the following expression:
0 2 1 (0,2,3,1) (0,3,1,2) A 2 3
E k (a) = k 0 + k1 a + k 2 a + k 3 a , k1 k 2 0, k 3 = k16 k 22
1 2 1 (1,3,2,0) (3,0,2,1) A
2 2 1 (2,0,1,3) (1,2,0,3) A These functions generate different permutations then
3 2 1 (3,1,0,2) (2,1,3,0) A those obtained with the simple functions but the
0 3 1 (0,3,1,2) (0,2,3,1) A inverse function is difficult to deduce.
1 3 1 (1,2,0,3) (2,0,1,3) A C. GF(16)
2 3 1 (2,1,3,0) (3,1,0,2) A This field has sixteen 4-bits symbols and it could be
3 3 1 (3,0,2,1) (1,3,2,0) A used by VEA, PEA or CVPEA .
0 1 2 (0,1,3,2) (0,1,3,2) S Only some elements of this field have the maximum
1 1 2 (1,0,2,3) (1,0,2,3) S order 15.
There are 8 couples (k2, q) which can be used:
2 1 2 (2,3,1,0) (3,2,0,1) A
(1, 1), (2, 8), (4, 4), (7, 13), (8, 2), (11, 11), (13, 7)
3 1 2 (3,2,0,1) (2,3,1,0) A
and (14, 14).
0 2 2 (0,2,1,3) (0,2,1,3) S
1 2 2 (1,3,0,2) (2,0,3,1) A There are 16!1 21 1012 possible permutations of
2 2 2 (2,0,3,1) (1,3,0,2) A 16 symbols.
3 2 2 (3,1,2,0) (3,1,2,0) S There are 1920 simple polynomial inversable
0 3 2 (0,3,2,1) (0,3,2,1) S functions defined on GF(16).
The experimental analysis of these functions showed
1 3 2 (1,2,3,0) (3,0,1,2) A
that only the linear and the square functions generate
2 3 2 (2,1,0,3) (2,1,0,3) S
unique permutations. For higher exponents, the
3 3 2 (3,0,1,2) (1,2,3,0) A
58
permutations are repeated. So we can use 479
functions for permutation on GF(16):
k2
E k (a) = c = k 0 + k1 a , k1 0,
k 2 {1, 2}, (k 0 , k1 , k 2 ) (0, 1, 1)
D. GF(32)
This field has all 5-bits symbols of order 31 and it is
optimum to define permutation functions for CVPEA.
There are 32!1 2.6 10 35 possible permutations of
32 symbols except the identity one.
29759 permutations are generated using simple
polynomial functions defined on GF(32):
k2
E k (a) = k 0 + k1 a , k1 0,
k 2 0, k 2 31; (k 0 , k1 , k 2 ) (0, 1, 1)
V. CONCLUSIONS
REFERENCES
[1] R.E Blahut, Digital Transmission of Information, Addison-
Wesley Publishing Co., 1990.
[2] A.Menezes, Handbook of Applied Cryptography, CRC Press,
Inc., 1997.
[3] Scripcariu L., Duma P, About Some Cryptography Functions
Defined on Galois Fields, Buletinul Institutului Politehnic Iasi,
Romania, Sect. III Electrotehnica, Energetica, Electronica, Tom
L(LIV), Fasc.1-2, 2004, pp.65-70.
59
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Electronics and Telecommunications Faculty, Applied Electronics Departament, Iuliu Maniu 1-3, Bucharest, Romania
e-mail: rodicastoian2003@yahoo.com, lperisoara@yahoo.com
60
RSC 1 coder u
D D
u
x=
c1
Interleaver
(u,c1,c2)
D D
u
c2
RSC 2 coder
Fig. 1: Parallel Concatenated Convolutional Code with rate 1/3 and G ( D ) = [ 1 1 + D 2 1 + D + D 2 ] .
61
1 1
L
2 (K-1)L
2L 2L
3 K-2
in out in out
L
K-1
(K-1)L
K K
(a) (b)
Fig. 3: Block diagram of the convolutional interleaver (a) and convolutional deinterleaver (b).
Interleaver Block
Convolutional
Parameters LR/ TB LR/BT RL/TB RL/BT
KL ( K 1)
Interleaver memory (U-1)(V-1) (U-1)V U(V-1) UV-1
2
Period T T T T K
62
waiting, the first decoder makes an estimate of the
transmitted information, interleaves it to match the
format of parity c2, and sends it to the second decoder.
The second decoder takes information from both the
first decoder and the channel and re-estimates the
information. This second estimation is looped back to
the first encoder where the process starts again.
This cycle will continue until certain conditions are
met, such as a certain number of iterations or stop
criterion are performed. When the decoder is ready, the
estimated information is finally kicked out of the cycle
and the decisions are made in the threshold block. The
result is the decoded information sequence.
V. SIMULATION RESULTS
The interleavers are specified by the design of a Fig. 6: The output frame of LR/TB block interleaver U=100, V=100.
permutation on the integers {0, 1, 2, , T-1}, where
T=10000 is the frame size. For a LR/TB block
interleaver, we set the matrix dimensions to U=100,
V=100, with U*V=T. The total memory used by the
interleaver and deinterleaver is 20000 and the delay of
interleaving is also 20000 time units. The spreading
factors are s=99, t=100 or s=100, t=99. For this
interleaver, Fig. 6 shows the scatter plot when the input
frame is the identity permutation (i)=i (Fig. 5).
The dispersion for a convolutional interleaver is
shown in the scattered plot (Fig. 7 and Fig. 8), for
different values of the K and L parameters. Setting
K=10 and L=15 we obtain the convolutional interleaver
with memory M=675 and a delay of 1350 time units.
The spreading factors (s,t) are (10, 149), (141, 10) or
(151, 9).With this parameters, the convolutional
interleaver has the scatter plot represented in Fig. 7. If
we change the K and L parameters (K=20, L=30), the
dispersion is bigger, as we see from the scatter plot in Fig. 7: The output frame of convolutional interleaver with K=10
the Fig. 8. The spreading factors (s,t) are (20, 599), and L=15.
(581, 20) or (601, 19). Also, the memory size and the
delay increases at M=5700 and D=11400 time units.
The number of lines from Fig. 7 and Fig. 8 is
given by the number of lines K of the interleaver.
The distance between them varies with the KL
product. It is obvious that the best performance is
obtained for the block interleaver.
63
dimensions (U,V): (2,1000), (10,200), (16,125), 1
(40,50). We consider that the transmission was done
0.01
0.1
0.001
0.01 uncoded
0.0001 conv
uncoded
block
I1 (2,1000)
0.001 I2 (10,200) 0.00001
I3 (16,125) 0 1 2 3 4
I4 (40,50) Eb/N0 (dB)
0.0001 Fig. 11: BER(Eb/N0) for different types of interleavers.
0 1 2 3 4
Eb/N0 (dB)
64
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract This paper presents a study of parameters acceptable complexity. The transformation groups
involved in the lossless text compression methods using similar symbols, so the probability of finding a
the Burrows-Wheeler Transform (BWT). BWT (also character close to another instance of the same
known as Block Sorting) is one of the most efficient character increases substantially. The resulting text
techniques used in data compression. Its purpose is to
preprocess the text before applying a compression
can be easily compressed with fast locally adaptive
algorithm, thus providing a better use of the inner algorithms, such as Move-to-Front coding combined
redundancy of the text. BTW converts the original with Huffman or arithmetic coding.
blocks of data into a format that is extremely well suited
for compression. This paper deals with the choice of the The Block Sorting algorithm transforms the original
range for the block length for different types of text string S of N characters by forming all possible
files, evaluating the compression ratio and compression rotations of those characters (cyclic shifts), followed
time. by a lexicographical sort of all of the resulting strings.
Keywords: Block Sorting, Move-to-Front (MFT), Run The output of the transform is the last character of the
Length Encoding (RLE), Huffman coding, arithmetic
coding
strings, in the same order they appear after sorting.
All these strings contain the same letters but in a
1
The transform divides the original text into blocks of Since BWT groups closely together symbols with a
the same length, each of them being processed similar context, the output can be more than two
separately. The blocks are then rearranged using a times smaller than the output obtained from a regular
sorting algorithm. This is why it is also called Block compression. Compressing a text file with the
Sorting. [1] The resulting block of text contains the Burrows-Wheeler Transform can reduce its size while
same symbols as the original, but in a different order. the compression without the transform gave a weaker
Sorting the rows will be the most complex and time output. The compression method used in both cases
consuming task in the algorithm, but present consists of the three stages following BWT: Move-to-
implementations can perform this step with an Front, Run-Length Encoding and arithmetic coding.
1
Facultatea de Electronic i Telecomunicaii, Catedra de
Electronic Aplicat i Ingineria Informaiei, Bd. Iuliu Maniu
nr. 1-3, sector 6, Bucureti, e-mail: rradescu@atm.neuro.pub.ro
65
III. MOVE-TO-FRONT
66
Table 2
Block length (103) Compressed file (kB) Therefore, it is obvious that the required time will
0.1 671 increase, depending on the dimension of the block of
1 256 processed data.
2 213
3 196 The compression time as function of block length is
5 180 shown in Figure 4. For large values of the block
10 165 length, the compression time substantially increases
50 145 in the case of applying the BWT. This result could be
explained not only by the presence of the BWT but
100 142
also by the adaptive arithmetic compression, which
250 139
supposes a two-stage processing of the block and an
286 134
adjustment of the codewords, depending on their
300 132 frequencies and, eventually, on the number of
325 131 symbols.
350 132
400 134 V. COMPARISON WITH STANDARD
600 138 COMPRESSORS
800 138
WinRAR, WinAce and WinZip were chosen among
the usual compression programs, in order to compare
the performances of the algorithm presented above,
applied on the same 5-file test set. The dimensions of
the compressed files (using the 3 standard
compressors and the BWT algorithm) are shown in
Figure 5, for the considered test files.
67
length. From this point of view, one have to take into
account both the compression performances and the
required computing resources. To get a good
compression ratio and an acceptable compression
time, the block length could be situated around
200,000 bytes. For block length less than 100,000
bytes, the compression ratio is sensibly decreased, as
well as the compression time. An excessive
increasing of block length (over 800,000 bytes)
produces an unacceptable compression time.
REFERENCES
[1] M. Burrows and D. J. Wheeler, "A Block-Sorting Lossless
Data Compression Algorithm", 1994, report available at:
http://gatekeeper.dec.com/pub/DEC/SRC/research-
reports/abstracts/src-rr-124.html.
[2] M. Nelson, "Data Compression with the Burrows-Wheeler
Transform", September 1996, available at:
http://dogma.net/markn/articles/bwt/bwt.htm.
[3] M. A. Maniscalco, "A Run Length Encoding Scheme for Block
Sort Transformed Data", 2000, available at:
http://www.geocities.com/m99datacompression/papers/rle/rle.html.
[4] P. M. Fenwick, "Block Sorting Text Compression", 1996,
available at: ftp.cs.auckland.ac.nz.
[5] T. C. Tell, J. G. Cleary and I. H. Witten, Text Compression,
Prentice Hall, Englewood Cliffs, NJ, 1990.
[6] R. Rdescu, Compresia fr pierderi: metode i aplicaii,
Matrix Rom, Bucharest, 2003.
68
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
{ }
we propose a version of FTF algorithm based on a 1 n
E { J ( n )}
2
modified form of the cost function, using an E e ( n) (4)
1
asymptotically unbiased estimator of the mean square
error [3], [4]. The reduced dynamics of the modified
1 n
algorithms parameters could lead to facility for fixed- E { ( n )} R (5)
point implementation. 1
1
Politehnica University of Bucharest, Electronics and Telecommunications Faculty, 1-3, Iuliu Maniu Bvd., Bucharest, Romania,
e-mail: pale@comm.pub.ro, silviu@comm.pub.ro, aenescu@comm.pub.ro
69
where R is the correlation matrix of input data. We III. LOW DYNAMICS FTF ADAPTIVE
can see that J(n) is a biased estimate of E{|e(n)|2} and ALGORITHM
similarly (n) is a biased estimate of R. It will result
Let us consider a forward linear predictor of order N
E { J ( n )}
n
1
1
E e ( n)
2
{ } (6)
with the vector coefficients at time n denoted by
a N (n) . The forward a posteriori prediction error
produced at the output is
1
E { ( n )} R (7)
n 1 eNf (i ) = a H
N ( n) x N +1 (i ) (13)
Some classes of applications [5] require a high where x N +1 (i ) is the N+1-by-1 the tap-input vector,
memory algorithm, which means that the value of the
with 1 i n .
exponential weighting factor is very close to unity.
The cost function is the sum of weighted forward a
In this case very large values for these parameters can
posteriori prediction-error squares in the modified
result, causing unwanted finite precision effects in a
form according to (8):
practical implementation.
Taking into account the previous discussion we
n 2
propose an unbiased estimator of the matrix (n). So J Nf ( n ) = (1 ) n i eNf ( i ) =
that, we will modify the cost function from equation i =1 (14)
(1) as follows: 2
= J Nf ( n 1) + (1 ) eNf (n)
n 2
J ( n ) = (1 ) n i e (i ) =
i =1 (8) The corresponding backward linear prediction-error
2 filter with the vector coefficients denoted by c N ( n)
= J ( n 1) + (1 ) e ( n )
will produced the backward a posteriori prediction
error:
In this case
ebN (i ) = c H
N ( n) x N +1 (i ) (15)
E { J ( n )} (1 n ) E { e (n) 2
} (9)
In this case, the cost function is the sum of weighted
backward a posteriori prediction-error squares:
is an asymptotically unbiased estimator of the mean
square-error. n
2
Following this idea we have to perform the same ( ) ( ) ni ebN ( i ) =
b n = 1
JN
modification in equations (2) and (3) obtaining i =1 (16)
2
n = JN ( ) ( ) N( )
b n 1 + 1 eb n
( n ) = (1 ) ni x ( i ) x H ( i ) =
i =1 (10)
Let N +1 ( n ) denote the N+1-by-N+1 correlation
= ( n 1) + (1 ) x ( n ) xH (n)
matrix of the tap-input vector x N +1 (i ) , where
n 1 i n , f ( n ) denote the m-by-1 cross-correlation
(n) = n i x (i ) (i ) =
d*
i =1 (11) vector between x(i ) and x N (i 1) , and b ( n )
= ( n 1) + (1 ) x ( n ) d * ( n ) denote the N-by-1 cross-correlation vector between
x N (i ) and x(i N ) . According to (10) and (11)
According, these parameters have the following forms:
( )
n
E { ( n )} 1 n R (12) N +1 ( n) = (1 ) ni x N +1 (i )x H
N +1 (i ) =
i =1
is an asymptotically unbiased estimator of the = N +1 ( n 1) + (1 ) x N +1 (n)x H
N +1 ( n)
correlation matrix. (17)
Most of the expressions in the following section may n
look familiar to readers acquainted with the theory of f (n) = (1 ) ni x m (i 1) x* (i ) =
least-squares transversal filters. However, the i =1 (18)
derivation that follows is developed according to the
= f (n 1) + (1 ) x m ( n 1) x* (n)
new approach.
70
n g ( n ) = g ( n 1) + k N ( n ) N ( )
b* n (28)
b (n) = (1 ) ni x N (i ) x* (i N ) =
i =1 (19)
= b (n 1) + (1 ) x N (n) x* (n N ) k N ( n ) = (1 ) N1 ( n ) x N ( n ) (29)
Nf ( n ) = a H
N ( n 1) x N +1 ( n ) (23) It can be demonstrated that the inverse of the
correlation matrix may be expressed as follows:
and k N ( n 1) is the modified gain vector:
0 0 1
N1+1 ( n ) = 1 n 1
+ a ( n) aH
N (n)
N ( ) J N min ( n ) N
0 f
k N ( n 1) = (1 ) N1 ( n 1) x N ( n 1) (24)
(34)
Taking these into account we may write the recursion Using the previous relation we get the following
for updating the tap-weight vector of the prediction- recursion for the modified extended gain vector:
error filter:
0 eNf ( n )
0 f k N +1 ( n ) = (
+ 1 ) N( ) f
a n
a N ( n ) = a N ( n 1) * N (n) (25) k N ( n 1) J N min ( n )
k N ( n 1) (35)
Similarly, using an alternative expression for the
Finally, we get the following recursion for updating inverse of the correlation matrix:
the minimum value of the sum of weighted forward
prediction-error squares: 1 ( n ) 0 1
N1+1 ( n ) = N + b cN ( n) cH
N (n)
0 0 J N min ( n )
J Nf min ( n ) = J Nf min ( n 1) + (1 ) Nf ( n ) eNf * ( n )
(36)
(26) we get the second recursion for the modified extended
In a similar manner we obtain a set of relations for the
gain vector:
backward prediction part of the algorithm:
k ( n ) ebN ( n )
0 k N +1 ( n ) = N + (1 ) c N ( n ) b
N +1 ( n ) c N ( n ) = b (27) 0 J N min ( n )
J N min ( n )
(37)
71
The definition of the modified gain vector from J Nf min ( n 1)
relation (29) may also be viewed as the solution of a N +1 ( n ) = N ( n 1) (46)
special case of the normal equations. It defines the J Nf min ( n )
tap-weight vector of a transversal filter that contains N
taps and that operates of the input data x N ( n) to
min ( n 1)
b
JN
produce a least-squares estimate of a special desired N +1 ( n ) = N ( n ) (47)
min ( n )
b
JN
response:
1, i=n
d (i ) = (38) Finally, we have to put together four distinct tasks
0, 0 < i < n (forward linear prediction, backward linear prediction,
computation of the gain vector and estimation of the
The estimation error (modified conversion factor) is desired response) in order to obtain our modified FTF
defined as follows: adaptive algorithm.
First, let us define the normalized gain vector:
N ( n ) = 1 (1 ) x H
N (n) N (n) xN (n)
1 (39)
k N (n)
k N (n) = (48)
Taking into account the expression of the inverse of N (n)
the correlation matrix from the standard recursive
least-squares estimation problem [1], [2] we get:
According, some simplified recursions can be
obtained:
1
N (n) = (40)
1 + (1 ) 1x H
N (n) N (n) xN (n)
1
(1 ) N ( n )
f
0
k N +1 ( n ) = + a N ( n 1)
k N ( n 1) J N min ( n 1)
f
Three useful interpretations of the conversion factor
are known [1], [2]: (49)
- for recursive least-squares estimation: 0 f
a N ( n ) = a N ( n 1) * eN ( n ) (50)
eN ( n ) k N ( n 1)
N (n) = (41)
N (n)
Nb ( n ) = b
JN min ( n 1) k N +1, N +1 ( n ) (51)
where eN (n) is the a posteriori estimation error (1 )
and N (n) is the a priori estimation error;
N +1 ( n )
- for adaptive forward linear prediction: N (n) = (52)
1N
b
( n ) N +1 ( n ) k *N +1, N +1 ( n )
eNf ( n )
N ( n 1) = (42)
Nf ( n ) where k *N +1, N +1 ( n ) is the last element of the vector
k N +1 ( n ) .
- for adaptive backward linear prediction:
Similarly, in the case of backward prediction we get:
ebN ( n ) k N ( n )
N (n) = (43)
Nb ( n ) = k N +1 ( n ) (1 )k N +1, N +1 ( n ) c N ( n 1)
0
(53)
Taking these into account, the following recursions
k
*
for updating the conversion factor can be obtained: c N ( n ) = c N ( n 1) ebN ( n ) N (54)
0
2
eNf ( n )
N +1 ( n ) = N ( n 1) (1 ) (44) In order to complete the algorithm it is necessary to
J Nf min ( n ) update the tap-weight vector of the adaptive filter as
follows:
2
ebN ( n ) N (n) = d (n) wH
N ( n 1) x N ( n ) (55)
N +1 ( n ) = N ( n ) (1 ) (45)
min ( n )
b
JN
w N ( n ) = w N ( n 1) + *N ( n ) k N ( n ) =
(56)
= w N ( n 1) + e*N ( n ) k N ( n )
72
In this manner we obtain our Low-Dynamics FTF procedure presented in [1] and [2] and we obtain the
(LD-FTF) adaptive algorithm. It is summarized algorithm as follows.
below.
Initialization of LD-FTF adaptive algorithm
LD-FTF adaptive algorithm
a0 (1) = c0 (1) = 1 , k 0 (1) = 0 , 0 (1) = 1
Predictions
d (1)
Nf ( n ) = a H
N ( n 1) x N +1 ( n ) w1 (1) =
x (1)
eNf ( n ) = N ( n 1) Nf ( n) 2
f
J 0min ( n ) = (1 ) x (1) , x (1) 0
J Nf min (n) = J Nf min ( n 1) + (1 ) Nf (n) eNf * (n)
for n = 2:N+1
J Nf min ( n 1)
N +1 ( n ) = N ( n 1) nf 2 ( n ) = aTn 2 ( n 1) x n 1 ( n )
J Nf min (n)
a n 2 ( n 1)
(1 ) N ( n )
f
k N +1 ( n ) =
0
+ a N ( n 1) a n 1 ( n ) = nf 2 ( n )
k N ( n 1) J N min ( n 1)
f
x (1)
0 f enf 2 ( n ) = n 2 ( n 1) nf 2 ( n )
a N ( n ) = a N ( n 1) * eN ( n )
k N ( n 1)
J nf1min ( n ) = J nf 2 min ( n 1)
Nb ( n ) = b
JN min ( n 1) k N +1, N +1 ( n )
(1 ) J nf 2 min ( n ) = J nf1min ( n ) + (1 ) nf 2 ( n ) enf2 ( n )
N +1 ( n ) J nf1min ( n 1)
N (n) = n 1 ( n ) = n 2 ( n 1)
1N ( ) N +1 ( n ) k *N +1, N +1 ( n )
b n
J nf 2 min ( n )
ebN ( n ) = N ( n ) N ( )
b n
0 (1 ) nf2
k n 1 ( n ) = + an 2 ( n 1)
k n 2 ( n 1) J n 2 min ( n 1)
f
min ( n ) = J N min ( n 1) + (1 ) N ( n ) eN ( n )
b
JN b b b*
if n = N+1
k N ( n )
= k N +1 ( n ) (1 )k N +1, N +1 ( n ) c N ( n 1)
cn 1 ( n ) = ( ) n 1 ( ) n 1 ( )
x 1 n k n
0
1
k *
c N ( n ) = c N ( n 1) ebN ( n ) N J nb1min ( n ) = (1 ) n 1 ( n ) x (1)
2
0
Filtering end
else
The initialization of the algorithm, i.e. 1 n N + 1
period, is quite complex and requires a lot of paper w n 1 (n 1)
w n (n ) = n (n )
space in order to be deduced. The most common
initialization is for the case when the initial condition
is zero. At time n = N, initialization of both the gain x (1)
vector and the adaptive filter is completed. However,
the forward and backward prediction-error filters are end
both one unit longer. So, their initialization is end
completed at time n = N + 1. We have introduced our
modifications into the standard initialization
73
IV. SIMULATION RESULTS It can be noticed that the performances of both
adaptive algorithms are the same. Hence, the LD-FTF
For the experimental results we consider a system algorithm keeps the fast rate of convergence and
identification configuration. In this class of specific to the family of fast LS algorithms.
applications dealing with system identification, an Moreover, the reduced dynamics of the modified
adaptive filter is used to provide a linear model that algorithms parameters could lead to facility for fixed-
represents the best fit (in some sense) to an unknown point implementation.
system. The adaptive filter and the unknown system
are driven by the same input. The unknown system
V. CONCLUSIONS AND PERSPECTIVES
output supplies the desired response for the adaptive
filter. These two signals are used to compute the
In this paper we have proposed a modified version of
estimation error, in order to adjust the filter
the FTF adaptive algorithm, named LD-FTF, with low
coefficients.
dynamics of the parameters, as a result of a different
In our experiments we compare the classical FTF
approach of the least squares estimation problem.
algorithm and the proposed LD-FTF algorithm. The
The basic idea was to use a modified form for the
input signal is a random sequence with an uniform
algorithms cost functions in order to obtain
distribution on the interval (1;1). The length of the
asymptotically unbiased estimators for the mean
adaptive filter is N = 5. The results are presented in
square errors. In this manner we reduce the dynamic
Fig. 1 and Fig. 2, using an exponential weighting
range of the algorithm parameters, preventing the
factor = 0.999.
unwanted overflow or stalling phenomena which may
appear when such an algorithm is implemented using
fixed-point arithmetic.
The simulation results prove that LD-FTF adaptive
algorithm keeps the fast rate of convergence specific
to the family of fast LS algorithms.
This paper represents only the first step of our
research. Future work will focus on fixed-point DSP
implementation of this algorithm. Also, a careful
analysis of numerical stability of LD-FTF algorithm
could be considered in perspective.
REFERENCES
74
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Facultatea de Electronica si Telecomunicatii, Catedra de Comunicatii, Bucuresti, Bd.Iuliu Maniu 1-3, phone: 021.331.18.17,
e-mail: cezarpart@yahoo.com
75
To that cause, the parameter from (6) is made A.Adaptive system configurations
gradient adaptive as
There are four adaptive system configurations
e[n + 1] = e[n] [ n 1]e[n] (7) defined by the function realized.
System identification (Fig. 1). We want to create a
model for an unknown system. This system and the
Using the chain rule, the gradient [ n 1]e[n] can adaptive filter have the same test signal x. The output
signal of the unknown system is the desired signal for
be evaluated as the adaptive filter. When y and d are close, the transfer
function of the unknown system is approached with the
E[n] E[n] [n] y[n] transfer function of the adaptive filter. The dynamics of
= * the system determine a time variability for the model.
[n 1] [n] y[n] w[n]
w[n] [n 1]
* = (8)
[n 1] [n 1]
e[n]e[n 1]x H [n]x[n 1]
= 2
( x[n 1] + [n 1]) 2
Fig. 1 System identification
The GNGD algorithm is therefore described by
Reverse modelling (fig. 2). The model also identifies an
unknown system. When the error is zero,the global
y[n] = x [n]w[n]
H
(9)
transfer function of both unknown system and adaptive
e[n] = d [n] y[n] (10) filter is reduced to a delay. The transfer function of the
adaptive filter is the reverse transfer function of the
w[n + 1] = w[n] + [n]e[n]x[n] (11) unknown system with a small difference caused by the
unavoidable noise. The model can also eliminate the
w[n] = 2
(12) result of an unknown function (eg. Automate
x[n] + [n] equalisation of communications channels).
76
Interferences cancelling (fig. 4). The primary signal is order of the filter ord and the specific parameters ,
the useful signal. It has an unuseful perturbing signal for RLS, and .
overlapped. There must be created a similar signal
which will be substracted from the primary signal using
a reference. This signal results from the adaptive filter.
Linear prediction
y[n] = 1.79y[n - 1] - 1.85y[n - 2] + Fig 6. Convergence of GNGD, Mathews and Benvenistes algorithms
(15)
1.27y[n - 3] - 0.41y[n - 4] + x[n]. We set N=1000, ord=7, =0.001 and =0.9. The mean
square error for the algorithms is shown in Fig. 7.
where x[n], a white noise with a zero average and
unitary variance, is passed through a AR filter.
We observe in Fig.5 that GNGD converges faster
than NLMS with 500 iterations. This improved
convergence results from the gradient adaptive in the
denominator of the learning rate of the algorithm.
In [4] we find a comparison between GNGD and
Mathews and Benvenistes algorithms. Using usual
values of the parameters, it is evaluated that GNGD has
faster convergence than Mathews and Benvenistes
algorithms. This result is shown in Fig.6.
System identification
77
d[n] = sin(n 0 ), 0 = 0.05 * (16)
The optimization of the parameters results in a
faster convergence for GNGD comparing to NLMS.
and a perturbation with the following recursive relation
Interferences cancelling
78
A noise factor n=0.5 leeds to an inpossible
interferences cancelling for all the algorithms (Fig. 12).
Reverse modelling
The reverse modelling configuration cancels the
results of an unknown transfer function (eg. Automate
equalization of communication channels). For this
configuration we use an adaptive channel equalizor with
a general design as Fig.14 describes.
The input signal has the 1 values randomly
distributed. The channel transfer function H c [z ] is
given by (15). The output signal has a white noise
overlapped and the adaptive filter realizes the
equalization. The switch is used in position 1 with
training sequence. Fig.16 NLMS channel equalization
79
REFERENCES
III. CONCLUSION
APPENDIX
Mathews algorithm:
y[n] = x H [n]w[n] (20)
Benvenistes algorithm:
80
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Technical University of Cluj-Napoca, Str. C-tin. Daicoviciu 15, 400020 Cluj-Napoca, Romania,
Phone-fax: +40 264 591340, E-mail: erwin@bel.utcluj.ro
81
= E ek2 = E yk2 2W T E [ yk X k ] + W T E X k X kT W . (5) the probabilistic operator of the square error function
with respect to the vector of weights.
The SME function can be expressed more suitable by
substituting the term E[XkXkT] from (5) with the
autocorrelation matrix RXX. More, the term E[ykXk] e 2 e
= E k = E 2ek k =
can be substituted with the intercorrelation matrix W W
RyX. Thus, the SME can be expressed as:
= 2E { X k yk } + 2E { X k X kT }W = (8)
= E {ek2 } = E { yk2 } 2W T RyX + W T RXX W . (6) = 2RyX + 2RXX W .
From (6) we can observe that is a quadratic function When the vector of weights (filter coefficients) has
of the weights of vector W (filters coefficients). When the optimum value Wopt, the SME will be minimum.
equation (6) is expanded, the elements of W will be So, the gradient will be zero ( = 0). Equating (8)
only of first and second order. This equation is valid with zero results
when the input components and the desired response
1
are stochastic (random) variables. Wopt = RXX RyX . (9)
W p + 1 = W p p , (10)
82
ek = yk X kT W (11) 2.4.1 Discrete Kalman filter
= e e e
2 2 2
k
k k
= 2ek X k . (12)
k 2.4.1.1 Process estimation
w(0) w(1) w(N - 1)
The Kalman filter tries to estimate a state x n
This gradient estimation can be replaced in equation belonging to a controlled process, time discrete,
(10) and we obtain: described by the finite difference linear equation:
83
To obtain the equations for the Kalman filter, we two groups: equations for re-update in time and
have to find an equation which computes the a equations for re-update of measurement. The
posteriori estimated state xk as a linear combination equations for re-update in time are responsible for
time designing of the current state and the estimated
of the a priori estimated state xk and of the weighted
error covariance to obtain the a priori estimations for
difference between the current measurement zk and the next instant. The equations for re-update of the
the measurement prediction Hx k : measurements are responsible for the realization of
the inverse feedback the inclusion of new
xk = xk + K( zk Hx k ) . (22) measurement in the a priori estimation to obtain an
improved a posteriori estimation. The equations for
re-update in time are also called prediction equations
The term ( zk Hx k ) from equation (22) is called and the equations for re-update of measurement are
the measurement innovation or residue. This also called the correction equations. So, the final
difference reflects the discrepancy between the estimating algorithm is a predictor-corrector
measurement prediction Hx k and the current algorithm, shown in figure 3.
measurement zk. If the residue is zero then the two
quantities are equal.
The matrix K[nm] from (22) represents the gain Re-update in time Re-update of measurement
(prediction) (correction)
or the interference factor and its role is to minimize
the a posteriori estimated error covariance Pk from
(21). This can be achieved by substituting equation
Fig. 3. Discrete Kalman filter cycle.
(22) in the definition equation of ek; the obtained error
will be replaced in (21) and the probability operator
The specific time and measurements re-update
will be computed. The obtained result will be derived
equations are presented below:
with respect to K, equated with zero and solved to
compute K. The most encountered form of the
Kalman gain, which minimizes equation (21), is: xk = Axk1 + Buk
(26)
Pk = APk 1 A + Q
T
Pk H T
K k = Pk H T ( HPk H T + R )1 = (23) K k = Pk H T ( HPk H T + R ) 1
HPk H T + R
xk = xk + K ( zk Hxk ) . (27)
When the covariance of the measurement noise R P
k = (1 K k H ) Pk
tends to zero, the gain K weights better the residue:
From (27) we notice that the first step consists in
lim K k = H 1 . (24) Kalman gain Kk computation. The next step consists
Rk 0
in measurement updating to obtain zk and generating
the a posteriori estimated state. The final step consists
When the a priori estimated error covariance Pk is
in obtaining the a posteriori estimated error
achieving zero, the gain K weights very slightly the covariance.
residue: After each re-update of the time and
measurements pairs, the process is repeated using the
lim K k = H 1 . (25) a posteriori estimations to compute the new a priori
Pk 0
estimations. That recurrence is one of the most
interesting property of the Kalman filters, that makes
So, when the measurement noise covariance R is its practical implementation to be more easier to
achieving zero, the current measurement zk matters realize relative to implementation of the Wiener filter,
more and the measurement prediction Hx k matters designed to work directly on all data for each
less. On the other side, when the a priori estimated estimation [4]. Instead, the Kalman filter computes
error covariance Pk is achieving zero, the current recursively the current estimation based on all
measurement zk matters less and the measurement measurements performed.
prediction Hx k matters more. 3. IMPLEMENTATION AND SIMULATION
2.4.1.3 Description of the Kalman algorithm In the last years, LabVIEW and Matlabs Simulink
became the most well known software packages used
The Kalman filter estimates a process using a in education and industry for modeling and simulation
feedback control: the filter estimates the state of a of dynamic systems.
process at a moment and then obtains a feedback in An example for processing signals using the LMS
the form of the measurements (in noisy conditions). algorithm is shown in figure 4. The LMS algorithm
Thus, equations of the Kalman filter are divided in structure with 4 coefficients is shown in figure 5.
84
A _ n o is e
signal is an audio wave signal on 16 bits, with 8 KHz
S a m p le s frequency, on single channel (mono).
S in + N o is e
111
-1
A _ s in
F ilt e re d s ig n a l
S in
d e lt a
100
Hz
S in
S in + Z g
w _ 0 (n )
w _ 1 (n )
w _ 2 (n )
w _ 3 (n )
d e lta
F ilte r_ O u tp u t
d e lta * e (n )
e (n )
w _ 0 (n + 1 ) w _ 1 (n + 1 ) w _ 2 (n + 1 ) w _ 3 (n + 1 )
a)
Fig. 5. The LMS algorithm structure.
b)
O u tp u t filte r
X _ k -1
P _ k -1 + Q + R
P _ k -1 Q
P _ k -1 + Q
R * (P _ k -1 + Q ) P_k
g a in K
K_k
Z _ k - X _ k -1
85
more efficient algorithms than the gradient algorithm
can be developed. The gradient algorithm can be
improved, for example, using different coefficient
step variations, which are obtained from statistical
estimations of the signal characteristics. However, due
to implementation imperfections, applying these
algorithms and sensitivity problems can be more
difficult.
The specific noise cancellation case was already
Fig. 11. Input signal; Filtered signal; Desired signal (Sine wave). studied since 4 decades but lately hardware
implementation possibilities of the theoretical systems
In the simulation example with Kalman algorithm with signal processors are loomed.
(implemented in Simulink), an input audio signal of 8 This paper tries to achieve some important sides
KHZ frequency, represented on 16 bits, and Q=110-8; of the adaptive systems in noise cancellation.
R=110-2 were considered. The obtained results are B. Conclusions concerning the simulation results.
shown in figure 12. From simulations, we can observe when the noise
amplitude is growing up, the filtered signal is visible
distorted, that reduces the respectively algorithm
performances. Otherwise, increasing the number of
filter coefficients, the accuracy of the filtered signal
increases.
Another important problem is choosing the
optimum parameters of the adaptive filters. For
example, in case of the LMS algorithm, choosing the
a)
step (), which determines the convergence rate is
critical. If is too large, the algorithm will converge
very rapidly but will present oscillations until stability
limit is reached, or, the effect of inverse error
minimizing - another drawback - appears. If a too
small step is chosen, oscillations will not appear
during the convergence process, but the convergence
b) speed is slower.
The optimum Wiener filter theory was made for
random stationery processes. When the statistical
properties of the random processes are changing in
time, the above description becomes more difficult.
Due to the permanently modifying of the error surface
of which minimum is to be searched, the adaptive
algorithm must ensure not only the convergence to the
c) optimum solution, but also to follow the continuous
Fig. 12. a) Audio signal with noise; b) filtered changing of this optimum value. The Kalman filter
audio signal; c) desired audio signal. theory that allows a model, for the considered
application, based on state equations gives the
solution. The obtained recursive algorithm is more
4. CONCLUSIONS rapid than the LMS algorithm and less dependent by
the static characteristics of input data, but presumes
A. General conclusions. more complex computations.
In this paper, several techniques for designing and
implementing adaptive filters were presented. These 5. REFERENCES
techniques were based on the gradient algorithm,
being the simplest and the most efficient instrument [1] Adelaida Mateescu, Neculai Dumitriu, Lucian Stanciu,
for varying the coefficients. Semnale i Sisteme. Aplicaii n filtrarea semnalelor, Ed. Teora,
2001;
The gradient algorithm leads to slowly modifying [2] Maurice Bellanger, Digital Processing of Signals. Theory and
filter coefficients values, when it requests a reduced Practice, Ed. John Wiley&Sons, 1984;
residual error and it is used in the simpler form (the [3] J.G. Proakis, Advanced Digital Signal Processing, Ed.
sign algorithm). To find the most rapid adaptation McGraw Hill , 1992;
[4] Greg Welch, Garry Bishop, An Introduction to the Kalman
rate, all of the coefficients can be re-computed Filter, Ed.ACM, 2001; (http://www.cs.unc.edu/~welch/kalman);
periodically using rapid iterative procedures. [5] Yifen Tu, Multiple Reference Active Noise Control, March
It is possible to consider some other criteria, 1997;
which, for other precision applications, are more [6] Digital Signal Processing Chapter 7: Adaptive Filtering
(http://www.staff.ncl.ac.uk/oliver.hinton/eee305/Chapter7.pdf);
suitable than the minimizing SME criterion, thus, [7] Mathlab Help. Signal Processing Toolbox, Simulink.
86
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Facultatea de Inginerie, Universitatea E.Murgu, P-a Tr.Vuia 1-4 Reia, e-mail c.chioncel@uem.ro
2
Facultatea de Electronic i Telecomunicaii, Departamentul
Comunicaii Bd. V. Prvan Nr. 2, 300223 Timioara, e-mail gal@etc.utt.ro
87
k = e Ft = 1 + (Ft ) +
(Ft )2 + ...
(6) after setting the derived equal to zero, is called the
Kalman gain:
2!
III. KALMAN FILTER The covariance matrix, associated with the optimal
estimate, is now:
The Kalman filter is essentially a set of mathematical
equations that implements a predictor-corrector type Pk = (I K k H k )Pk (14)
estimator that minimizes the estimated error
covariance. ^
The random process that has to be estimated can be The update estimates x k can be projected ahead via
modeled in form of equation (1). The process
the transition matrix:
measurement, at discrete points is:
^ ^ (15)
z k = H k xk + vk (7) x k +1 = k x k
Table 1
E [v v ] = R
T
k i k
(9) Predict Correct
(time update) (measurement update)
The estimation error is defined as the difference (1) Project the state (1) Compute the Kalman
between the state and his a priori estimate (- best ahead gain
estimate):
^
x k +1 = k x k
^
(
K k = Pk H kT H k Pk H kT + R k ) 1
88
1 d (t ) (20)
f i (t ) =
2 dt
where =
[( )
log f t g f (0 ) ]. After the first eight steps, the filter settles down to a
tg steady state condition where the Kalman filter gain
is about 0.8274. Fig. 4 shows the evolution of the
V. INSTANTANEOUS FREQUENCY AND PHASE phase estimation corresponding to the first part of the
Kalman algorithm.
The chirp signal to be dealt with is a linear, quadratic
one, given by
= ( f 1 f 0 )t1 p (17)
1+ p (18)
y = cos 2 t + f 0 t + phi / 360
1 + p
Fig. 4 First part of the phase estimation
where p is the polynomial order and phi de initial
phase. The expression of the instantaneous frequency (17)
The model of the signal that is here considered can be was computed based on equation (16).
expressed as:
2 f (21)
y (t ) = A cos (t ) (19) F= t + 0
1+ p 2
with A constant. The instantaneous frequency, fi(t), of
Using the same Kalman filter parameters like those in
the signal is
the phase estimation, but for a greater number of
89
points, 101, the simulation for the frequency estimation will follow too much the measurement and
parameter estimation is shown in Fig. 5. the noise that affects it.
The importance of the correct determination of the
filter parameters k state transition matrix, Hk
measurement relationship to x, the noise sequence Qk
and the measurement error Rk, have been evidenced in
the plots form Fig. 6 and 7. The estimation, after a
higher number of steps, doesnt settle down to an
optimum estimation but increases or decreases from
the actual process x(t).
After the first ten steps the Kalman filter stabilizes by [1] R.G. Brown, Y.C. Hwang, Introduction to random signals and
applied Kalman filtering, Second edition, John Wiley &Sons, Inc,
0.618. If we repeat the simulation for other filter 1992
parameters like the state transition matrix (scalar in [2] G.Welch, G. Bishop, An Introduction to the Kalman Filter,
this case) k, we can observe a different behavior of ACM, INC, 2001
[3] J.Gal, M.Slgean, M.Bianu, I.Nafornita, The Instantaneous
the estimated result. For k=1.45, the phase estimation Frequency Determination for Signals with Polynomial Phase using
will have the following allure, Fig 6. Kalman Filtering, Buletinul Universitii Politehnica, Seria
Electronic i Telecomunicaii, Tom 47(61), 2002, Fascicola 1-2.
[4] A. Quinquis, A erbnescu, E,Rdoi, Semnale i sisteme.
Aplicaii n MatLab, Ed. Academiei Tehnice Militare, Bucureti,
1998
[5] MatLab R12 Tutorial
[6] S.Saha, S.Kay, A noniterative Maximum Likelihood parameter
estimator of superimposed chirp signals, Department of Electrical
and Computer Engineering, University of Rhode Island
[7] O.Besson, M.Ghogo, Parameter estimation for random
amplitude chirp signals, IEEE Transactions on signal processing,
Vol.47, No.12, December 1999
[8] Chung-Chieh Lin, Petar M. Djuric, Bayesian estimation of
chirplet signals by MCMC sampling, Department of Electrical and
Computer Engineering, State University of New York at Stony
Fig.6 Phase estimation with k > 1
Brook
CONLUSIONS
90
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Facultatea de Electronic i Telecomunicaii, Departamentul
Comunicaii Bd. V. Prvan Nr. 2, 300223 Timioara, e-mail getab@etc.utt.ro
91
(i) The use of orthogonal transform can H H 12
improve the performance of LMS H 21 = 11 (9)
adaptive filters. H 12 H 11
(ii) The WHT is a fast orthogonal transform
which only involves the addition where the size of matrices H11 and H12 is half that of
operation. H 2' . In that case the transformed kernel is a block
The WHT matrix is an NxN matrix (N=2k , diagonal matrix:
k=1,2,3,) usually defined recursively using a block-
1 1 1 a b 0 0
matrix decomposition as follows: W1 =
2 1 1 b c 0 0
H 21W = (10)
1 Wn 1 Wn1 0 0 d e
and Wn = . The WHT matrix is
2 Wn 1 Wn 1 0 0 e f
denoted by W in the following discussion. An
important property of the WHT matrix is given in where the variables a,b,c,d, e and f represent the
eq.4: independent elements. The output of the new second
order filter now becomes:
W W T = 1 (4)
' '
y 2 [ n] = X nW H 2W X nW
T
= X nW H 2W ( X nW )T (11)
For the new implementation we consider the second
order kernel given in eq.3 and the isotropic property: This new implementation raises two problems:
-Is the reduction of the computational complexity
h11 h12 h13 h14 accomplished? We have calculated the number of
operations required for a direct implementation and
h12 h22 h23 h13
H2 = (5) those required for the new implementation. The
h13 h23 h22 h12
results listed in Table 1 show that the new
h14 h13 h12 h11 implementation requires less multiplication
operations.
It can be easily seen that the matrix H2 is symmetric
according to its both diagonals. The nonlinear filter Tabel I
produces the output signal y2[n]: Multiplications Additions
Direct N [(3 / 4) N + 1] (3N 2 + 2 N 4) / 4
implementation
y 2 [ n] = X n H 2 X nT = X nWW H 2WW
T T
X nT =
T
X nW H 2W X nW New [N(N/2)+1] N2/2+Nlog2N-1
implementation
(6)
where: X nW = X nW is the Walsh-Hadamard
transform of the input vector and H 2W = W T H 2W is - Can the TWH of the input vector( X nW = X nW ) be
the WHT of the second order Volterra kernel. substituted by the WHT of the rearranged input vector
If we rearrange the input vector as: '
( X nW = X n' W , X n' = [ x(n) x(n 1) x(n 3) x(n 2)] )?
X n = [x[ n], x[ n 1], x[n 3], x[ n 2]] ,
'
then the We easily find the relationship by examining X nW '
corresponding isotropic kernel H 2' is: and X nW . For example, we have the following
relationships between the two transformed vectors
h11 h12 h14 h13 '
X nW '
= [ X nW '
(1) X nW '
( 2) X nW '
(3) X nW ( 4)] and
H2 =
h12 h22 h13 h23
(7) X nW = [ X nW (1) X nW ( 2) X nW (3) X nW ( 4)] :
h14 h13 h11 h12
'
h13 h23 h12 h22 X nW (1) = X n (1)
'
X nW (2) = X n (4)
We can easily demonstrate that: (12)
'
X nW (3) = X n (3)
X n H 2 X nT = X n' H 2' ( X ' ) Tn (8) '
X nW (4) = X n (2)
This new second order kernel can actually be Similar relationships can be established for the input
decomposed into four sub matrices of the form: vectors of size 8, 16, etc.
92
III. EXPERIMENTAL RESULTS filter weights so that the system output y 1[n] tracks
the desired signal y[n] . The error signal e[n] is
The new implementation was applied to a typical
nonlinear system modeling problem, as shown in defined as the difference between the desired signal
fig.1. An adaptive second-order Volterra filter is used y[n] and the predicted signal y 1[n] , as indicated in
to identify the nonlinear system having the input- fig.1. The input process x[n] is assumed to be a zero
output characteristic given in fig.2. mean independent sequence with covariance, a
positive definite matrix. This simplifying assumption
is often made in literature [1,3].
y[n] The LMS type adaptive algorithm is a gradient search
nonlinear
system algorithm which computes a set of filter weights
+ H 1[ k + 1], H 2 [ k + 1] , at the time moment k+1, that
x[n] e[n]
+ seeks to minimize the error function, E[(e[k ]) 2 ] ,
cosidered at the time moment k.
-
The update equations for the Volterra adaptive filter
1
y [n] weights are well known in the literature[3] and are
adaptive Volterra given in eq. 15:
filter
H 1 [k + 1] = H 1[ k ] + 21e[k ] X [k ]
(15 )
lms H 2 [ k + 1] = H 2 [ k ] + 2 e[k ] X T [k ] * X [ k ]
adaptive
algorithm
where 1 and 2 are in both cases two small positive
Fig1.Nonlinear system identification using adaptive constants (referred to as the step size) that determine
quadratic filter the speed of convergence and also affect the final
error of the filter output.
The nonlinear system output is given in Eq.13. In our simulation we have used the efficient
implementation for the quadratic kernel. In this case
y[n] = A * X n + X nT * B * X n (13) the update equation for that kernel is:
0.54 3.72 1.86 0.76 where H 2W and X W" are the WHT of the quadratic
3.72 1.62 0.76 1.86 kernel respectively the WHT of the rearranged input
B= (14) vector.
1.86 0.76 1.62 3.72
Finally the output of the adaptive Volterra filter,
0.76 1.86 3.72 0.54 y 1[ n] , is:
This system is a slight modification of that used in
[1]. y 1[ n] = H 1 * X n + ( X nW
'
) T * H 2W * X nW
T
(17)
For a linear input signal x[n], the resulted output
signal is plotted in Fig.2. The linear kernel is a 1x4 vector and the quadratic
kernel is a 4x4 matrix. The input sequence is a
random gaussian zero-mean sequence having 1500
values.
The majority of papers examine the LMS algorithm
with a constant step size. The choice of the step size
reflects a tradeoff between misadjustement and the
speed of adaptation. The approximate expressions
derived in [3] showed that a small step size causes
small misadjustement, but also a longer convergence
time constant.
For adaptive Volterra filters the problems seem to be
much more complicated. In [3,5] the problems of step
size for different order kernels are well discussed.
Fig. 2 Input-output characteristic of the nonlinear system
The maximum step size bound is related to the
maximum eigenvalue of the autocorrelation matrix of
The adaptive filtering or system identification the input vector. Because we consider a second order
problem being considered is to try to adjust the set of Volterra filter without DC component included in the
93
estimation algorithm, the step size bounds for 1 and
2 are those given in [3]:
2 2
0 < 1 < ; 0 < 2 < (18)
3tr[ R XX ] 3(tr[ R XX ]) 2
where 0 < min < max . The initial step size 1[0] Fig.4 The adaptive variable step size filter error with white
input signal
was chosen to be 1 max although the algorithm is not
sensitive to this choice. As can be seen from eq.20,
the step size 1 is always positive and is controlled by
the size of the prediction error and the parameters
and . A large prediction error increases the step
size to provide faster tracking. If the prediction error
decreases, the step size will be decreased to reduce the
misadjustements. The constant 1 max is chosen to
ensure that the mean-square error(mse) of the
algorithm remains bounded . A sufficient condition
for 1 max to guarantee bounded mse is:
2
1 max < (21)
3tr ( R XX ) Fig.5 The adaptive variable step size filter error with
colored input signal
Usually 1 min will be near the value that would be
The colored input signal was generated as specified in
chosen for the fixed step size algorithm. In our eq.22.
simulation the value is 1 min = 10 7 .
Parameter must be chosen in the range (0,1) to x[n] = 0.25 * r[n 1] + r[n 2] + 0.25 * r[n 3] (22)
provide exponential forgetting. A typical value of
that was found to work well in simulations is where r[n] is a random, normal distributed sequence.
= 0.97 . The parameter is usually small In Fig.6 we have compared the mean-squared error of
the proposed filter (represented with solid line ) with
( 4,8 * 10 4 was used in our simulations.) those of a classic second order LMS adaptive
filter(represented with doted line ) . We also consider
the case of the colored input signal for the new filter
(represented with dashed line).
94
Fig.6 The mean-squared errors for the compared filters
For high level input signal the filter with variable step size
still adapts, as can be seen in fig.7.
Fig.7. The adaptive variable step size filter error with high
level input signal
IV. CONCLUSIONS
REFERENCES
[1] V.J. Mathews, Adaptive Polinomial Filters, IEEE Signal
Processing Magazine, No.7, pp. 10-23, July, 1991.
[2] M. Schetzen, The Volterra and Wiener theories of nonlinear
systems, Wiley and Sons, New York, 1980.
[3] J. Tsimbinos, Identification And Compensation Of Nonlinear
Distortion, Thesis, http://www.unisa.edu.au/html
[4] G.Budura, Contributions to the nonlinear systems study using
Volterra series, Thesis, Politehnica University of Timisoara, 30
Sept.,1999.
[5] G.Budura, C. Botoca, Applications of the Volterra Models in
Nonlinear Systems Identification, Buletinul Universitatii
Politehnica, Seria Electrotehnica, Electronica si Telecomunicatii,
Tom 47(61),Fascicola 1-2, pp.196-201, 2002.
[6] R.H. Kwong, E.W. Johnston, A Variable Step Size LMS
Algorithm, IEEE Transactions on Signal Processing, Vol.40, No.7,
pp.1633-1642.
95
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
96
2. BAYARD-BODE RELATIONSHIPS AND
( )
pN
p [ ( p ) ( p )],
(6)
HILBERT TRANSFORM >1
The Bayard-Bode relations method is based on the Using quadrature formulae, several approximations
fact that the transform results. Here we shall consider for study that one
derived from Simpson approach (the parabolic rule):
H ( j ) = R ( ) + jI ( ), (1)
1
of a causal function h(t) is uniquely determined in S ( ) = () +
terms of R() or I() (subject to an arbitrary
reactance value if determined from R() and to an 2 ln ( ) ( 1 )
+ +
arbitrary real value, if determined from I()) [1]. 3 1
Proofs based on Cauchys residue theorem [13] or on
convolution [6] establish ( 2 ) ( 2 ) ( 3 ) ( 3 )
+4 2
+2 + (7)
2
3 3
1
I ( y) ( k 1 ) (1 k )
R( ) = R()
y
dy = + + 4
k 1 1 k
+
(2)
2
yI ( y ) I ( ) ( k ) ( k )
+2
= R ( )
0 y2 2
dy k k
or
1 R( y ) 2 R ( y ) R( )
I ( ) =
y
dy =
y2 2
dy (3)
0 S ( ) = S p ( p ),
pZ
One can easily obtain the gain-phase relationships (or
the Bayard-Bode relations) from (2) and (3) directly
by taking logarithms [6], after fulfilling the 2 ln
1 1 + 3 , p = 1
requirements needed to satisfy the right half plane 1
analyticity conditions of the Hilbert transform, i.e. the
stable and minimum phase conditions. Under the 8ln
, p = 2, 2m
3 ( )
assumption that H ( s ) is not only analytic, but has no p p
2 ( y ) ( )
0 y 2 2
( ) = dy = 4. PHASE APPROXIMATION IN LINEAR
FREQUENCY DOMAIN
2
( c eu ) ( c )
eu e u
du = (5)
The formula between the imaginary and real parts of a
complex function of real frequency as expressed in
1 d u |u| equation (3) can be rewritten in many ways [4]. By
= ( c e ) ln coth du
0 du 2 integrating the right member of (3) by parts one can
find:
3. PHASE APPROXIMATION IN LOGARITHMIC 1 y +
FREQUENCY DOMAIN
I ( ) =
R '( y) ln
y
dy (9)
97
R( y ) (v) = (v + 1) ln | v + 1 | + (v 1) ln | v 1| 2v ln | v |
lim =0 (10)
y y
Remarks:
Alternatively, we can continue by integrating the right 1. The an numbers are determined by a broken-
member of (9) by parts, i.e. a double integration by line approximation to the gain-versus-
parts of the right member of (3) and the integrand will arithmetic-frequency characteristic.
be: 2. This procedure cannot be employed when the
gain characteristic has slopes different from
y + zero at zero and at high frequency.
R ''( y ) 2 ln 2 y 2 y ln (11) 3. The non-compact support gain method can
y
be easily extended to broken-parabolic (or
higher order curve approximation.
provided
R( y )
lim =0 (12) 5. MODIFIED BODE TRANSFER FUNCTIONS
y y
Previous attempts to test the phase approximations
and approaches have used the Bode transfer functions [1]
1 AK 2 s 3 + ( AH + B ) K 2 s 2 + ( A + BK 2 ) Hs + BH
( ) a ( ) =
a n n (16)
n n K 2 s 3 + K 2 Hs 2 + ( H + 1) s + H
[ BH ( AH + B ) K 2 2 ] + j[( A + BK 2 ) H AK 2 3 ]
where
( H K 2 H 2 ) + j[( H + 1) K 2 3 ]
2
An extended form of ( ) can be found in [11]
98
[ BH ( AH + B ) K 2 2 ]2 + [( A + BK 2 ) H AK 2 3 ]2 K test = K + noise _ real (24)
( H K 2 H 2 ) 2 + [( H + 1) K 2 3 ]2
6. SIMULATIONS
Thus
Now we are going to compare the given approaches.
1 1
( ) = ln U ( ) ln V ( ) , A. Logarithmic Frequency Domain
2 2
For logarithmic frequency domain, the selected
where transfer function is:
1
U ( ) = [ BH ( AH + B) K 2 2 ]2 + H ( s) = (25)
1
+[( A + BK 2 ) H AK 2 3 ]2 ; s+
1
4s +
s
+1
V ( ) = ( H K 2 H 2 ) 2 + [( H + 1) K 2 3 ]2 2
where we used the Bode transfer function (17),
considering K = H = 2 . The phase of the selected
The gain slope is given by3:
transfer function (i) is almost constant for < 0.01
U '( )V ( ) V '( )U ( ) and > 10 [1], consequently the interval of interest
'( ) = = in our experiments was [ 0.01,10] . We select
2U ( )V ( )
(19)
(2 A2 + B 2 K 2 ) 9 + + ( ) H 2 = 2 as sample ratio. Three plots are shown for
= different number of samples: k = 5 (ii), k = 9 (iii)
2 A2 K 8 12 + + B 2 H 4
and k = 17 (iv).
Now, If noise is present, then it can affect the quality of
1. From lim '( ) = 0 , we need phase approximation. The phase (v) and phase
approximations for different number of samples:
A2 K 4 K 4 0 ;
k = 5 (vi), k = 9 (vii) and k = 17 (viii) using as test
2. From lim '( ) = 0 , it follows
0 signal one that is affected by random perturbations are
H 2 B2 H 2 0 . also ploted.
Consequently, the modified Bode transfer functions 1. signal affected by a complex random perturbation
should satisfy the requirements:
A B K H 0 (20)
3. parameters of the Bode transfer function, Fig. 1. Phase (i) and phase approximation (ii), (iii), (iv) for the
respectively of the modified Bode transfer transfer function (25); phase (v) and phase approximation (vi), (vii),
function affected by a real random (viii) using signal affected by a complex random perturbation
perturbation
2. system function affected by a real random
perturbation
H test = H + noise _ real (23)
3
An extended form of ( ) can be found in [8]
99
frequencies varying from 0 to 10. Outside this
interval, both gain and phase of the transfer function
do not exhibit important variations. The phase and the
approximated phase are also shown (ii). If noise is
present, then it can affect the quality of phase
approximation. The gain (iii) and phase
approximation (iv) using as test signal one that is
affected by random perturbations are also ploted.
1. signal affected by a complex random perturbation
Fig. 2. Phase (i) and phase approximation (ii), (iii), (iv) for the
transfer function (25); phase (v) and phase approximation (vi), (vii),
(viii) using system function affected by a real random perturbation
Fig. 4. Gain (-) and gain samples (*) (i) used in linear
approximation of gain for (26); phase (-) and phase approximations
(.) with this linear approximation of gain (ii); gain (-) and gain
samples (*) (iii), respectively phase (-) and phase approximations
(.) (iv) for signals affected by complex random perturbations
Fig. 3. Phase (i) and phase approximation (ii), (iii), (iv) for the
transfer function (25); phase (v) and phase approximation (vi), (vii),
(viii) using parameters affected by a real random perturbation
100
REFERENCES
[1] H. W. Bode, Network analysis and feedback amplifier design,
D. Van Nostrand, Princeton, NJ, 1945.
[2] B. D. O. Anderson and M. Green, Hilbert transform and
gain/phase error bounds for rational functions, IEEE Transactions
on Circuits and Systems, vol. 35, no. 5, pp. 528535, May 1988.
[3] M. Green and B. D. O. Anderson, On the continuity of the
Wiener-Hopf factorization operation, J. Austral. Math. Soc. Ser. B,
vol. 28, pp. 443461, 1987.
[4] G. C. Newton, L. A. Gould, and J. F. Kaiser, Analytical
design of linear feedback controls, John Wiley & Sons, Inc.,
London, 1957.
[5] C. Rusu, P. Kuosmanen, and A. Burian, 1-D non-minimum
phase retrieval by gain sampling, in Proc. ECCTD99, Stresa,
Italy, Sept. 1999, vol. 2, pp. 755758.
[6] A. Papoulis, The Fourier integral and its applications,
McGraw-Hill, 1962.
[7] Corneliu Rusu and Pauli Kuosmanen, Phase approximation
by logarithmic sampling of gain, IEEE Transactions on Circuits
and Systems II: Analog and Digital Signal Processing, vol. 50, no.
2, pp. 93101, Feb. 2003.
[8] Lacrimioara Buzan, Phase approximation by gain samples,
M.S. Thesis, Technical University of Cluj-Napoca, 2003.
Fig. 6. Gain (-) and gain samples (*) (i) used in linear
[9] Peter Henrici, Applied and Computational Complex Analysis,
approximation of gain for (26); phase (-) and phase approximations
vol. I: Power Series - Integration - Conformal Mapping - Location
(.) with this linear approximation of gain (ii); gain (-) and gain
of Zeros, John Wiley & Sons, New-York, 1974.
samples (*) (iii), respectively phase (-) and phase approximations
[10] A. Papoulis, Signal analysis, McGraw-Hill, 1977.
(.) (iv) for signals whose parameters are affected by real random
[11] Lacrimioara Buzan, Phase Approximation in Linear and
perturbations
Logaritmic Frequency Domain, M.S. Thesis, Technical University
of Cluj-Napoca, 2004.
[12] Lacrimioara Buzan, Corneliu Rusu, Radu Ciprian Bilcu, Pauli
7. CONCLUSIONS Kuosmanen, Phase Approximation in Linear and Logarithmic
Frequency Domain, Proceedings of the First International
Symposium on Control, Communications and Signal Processing,
We have presented two methods for phase Hammamet, Tunisia, 21-24 March 2004, pp.709-712, ISBN: 0-
approximations: non-compact gain technique for 7803-8380-X
linear frequency domain and the approach based on [13] L. A. Zadeh and C. A. Desoer, Linear System Theory,
McGraw-Hill, New York, 1969.
logarithmic sampling of gain for logarithmic
frequency domain, using signals affected by random
perturbations. A comparison of the behavior of these
algorithms, considering signals affected by
perturbations, respectively signals that are not
affected by perturbations, was also presented.
Unlike the non-compact gain technique where we
need only the gain samples, the logarithmic sampling
of gain requests for two parameters: which
describes the frequency sampling, and k used to
provide a satisfactory approximation of the integral in
(7). There is a trade-off between and k [7].
Indeed, as k increases, has to tend to one more
rapidly. For this reason we have considered = 2m ( q ) ,
with m(q ) = 22 q for q = 1, ,9 . The gain samples
for non-compact gain technique were available by
sampling with the frequency interval
[ 0.01,10] .
From the experimental results we can conclude that
best achievements can be obtained using the linear
frequency domain approximation, unless when the
number of gain samples is low, where the logarithmic
sampling of gain is superior, but we can also see that
both methods for phase approximation behave well
when we have considered as test signals, signals
affected by random perturbations.
101
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
h
proposed for the signal reconstruction method. A
method for the detection of the impulsive noise, using x n'' = ''
k xn+k (6)
two thresholds, is presented. k =1
Keywords: extrapolation, disturbance, reconstruction The impulse response vectors h and h are
obtained by Burg method [1]. The forward and
I. INTRODUCTION backward extrapolated signal is given by
xn = wn xn' + (1 wn )xn'' . (7)
There are many situations when the long portions
in the recorded audio signals are removed or affected where wn is a weighting function
by the impulsive noise. A method for the restoration 1 1
1 (2u )a, un n ns
based on the separation of autoregressive processes is 2 n 2 and un = .
proposed. The corrupted samples are replaced by a wn =
1
(22u )a, u > 1 ne ns
weighted average of the signals extrapolated from 2
n n
2
areas preceding and corrupted area. For the detection An other proposed weighting function is:
of the impulsive noise a method with two thresholds
is used. wn1 = 2u n 3 3u n 2 + 1 (8)
This function is obtained using the polynomial:
II. RESTORATION METHOD FOR LONG g ( x) = ax 4 + bx 3 + cx 2 + dx + e
PORTIONS IN AUDIO SIGNAL
with the conditions: g (0) = 1, g (1) = 0, g ' (0) = 1,
The impulse response function is obtained from g ' (1) = 0 . The solution is : a = 0 , b = 2 , c = 3 ,
M = 2 N samples of the known signal by using the
d = 0 and e = 1 .
equation:
In Fig.1 the original and the bilateral extrapolated
signal are given for 1000 extrapolated samples using
Xh'= x (1)
N=1000 coefficients of extrapolation.
where The general weighting functions wn, 1-wn, for
h' = [h1 ' , h2 ' ,....., hN ']T , (2) a=3, and wn1, 1-wn1 are illustrated in Fig. 2.
and The relation for signal to noise ratio is:
W 1
x = [ x N +1 , x N + 2 ,......, x2 N ]T . (3)
x ' 2
n
The matrix X contains the shifted samples of the RSZ = 10 lg W 1 n=0 (9)
signal:
xN xN 1 xN 2 L x1 (x n xn' ) 2
n =0
x xN xN 1 L x2 where xn is the original signal and the term x n x n'
X = N +1 . (4)
M M M M represents the error of extrapolation. The mean square
x L xN
2 N 1 x2 N 2 x2 N 3 error (MSE) between the estimate and the desired
signal is given by:
The one-step forward extrapolation equation is:
1
University Politehnica of Bucharest, Department of Telelcommunications,
Str. Iuliu Maniu 1-3, Bucharest, Phone: 0214024815, e-mail: lucians@comm.pub.ro
102
W 1
1
MSE =
2W (x
n =0
n xn' ) 2 (10)
b)
Fig.3. RSZ and MSE as functions of number of impulse response
coefficients N, by using comparatively the weighting functions wn
and wn1 [a) signal 1; b) signal 2)]
a)
Fig. 4 illustrates the variation of Msd as function
of number of impulse response coefficients N, for the Fig. 4. (to be continued)
signal 1 [a)] and signal 2 [b)]. For the signal 2 we
have better restoration results.
103
b)
Fig.4. Msd as function of number of impulse response coefficients Fig.6. RSZ and MSE as functions of extrapolation length
N, for the signal 1 [a)] and the signal 2 [b)]. measured for last 500 extrapolated samples, for a bidirectional
extrapolation
III. CANCELLATION OF THE IMPULSIVE NOISE e(k ) with the amplitude greater than the threshold
which is given by the relation:
For the classical method, the audio signal affected
by perturbations, y (k ) , is segmented in frames of N = K e (14)
samples. Every frame is modeled as an autoregressive where e is the estimated value of the variance for the
process (AR) of order p: signal e(k ) . The affected samples will be replaced by
y (k ) = x(k ) + d (k ), an interpolation algorithm.
where x(k) is affected by the additive impulsive noise,
d(k), and
p
x(k ) = a ( j ) x ( k j ) + e( k ) ,
j =1
k = p, N 1. (12)
104
In the relations (14) and (15) r is the reduction factor, real-time applications. The information in the
f is a parameter that controls the reduction speed of b preceding and in the following data sections can be
and i max is the maximum number of iterations to pass used to recover the lost information. This recovering
to the next step (Fig. 9). The choice of the parameters is not perfect. The parameters RSZ, MSE, Msd and its
imax , f and r is based on the experimental proposed graphical representations, for different
situations, give the appreciation of this recovering.
observations. The final judgement is determined by the human ear,
With this method more iterations could be because audio signals do not exactly satisfy the
necessary to detect all the noise pulses from a frame requirements of being fully perdictable, and therefore
(especially for the small amplitude noise pulses). This they cannot be perfectly extrapolated.
algorithm contains also stop criterions for the
processing of a frame. Tab. 1: Comparative measurements between the results of the
The Table 1 give comparative measurements modified (MOD) and conventional (CONV) methods
between the results of the modified (MOD) and
conventional (CONV) methods. The parameters were Undetected pulses Wrong detections
the same for all the six signals. [%] [%]
MOD CONV MOD CONV
Signal 1.56 16.41 3.26 1.17
1
Signal 2.1 20.39 1.56 0.68
2
Signal 1.56 10.71 1.83 0.87
3
Signal 4.21 21.52 2.40 0.93
4
Signal 1.63 14.86 6.65 3.54
5
Signal 1.85 12.11 8 4.45
6
105
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Facultatea de Electronic i Telecomunicaii, Bd. Carol I, Nr. 11, 213737, Iasi, 6600, e-mail tarniced@etc.tuiasi.ro
106
H k ( e j ) H m ( e j ) = ( k m) ( 3) direct x2
M G SBC ( M ) = (7)
SBC M 1 2
1/ M
M 1
H [e
2
j ( 2k / M )
] = M for all . (4) For fixed input power spectrum density S x (e j ) , the
k =0
variance xi2 depends only on the analysis filters
An aliasfree(M) filter is defined to be one whose H i ( e j ) .
output can be decimated by a factor of M, without The subband coder is said to be optimal, if the coding
aliasing. gain is maximized. This is equivalent to minimizing
the product of subband variances.
A. Basic Model for Signal and Quantization Noise
The input x[n] to the subband coder is assumed to be C. Total Decorrelation of Subbands
real valued and wide sense stationary (WSS) with This property implies
zero mean and power spectral density S x (e j ) . The
total bit budget is R bits per sample. The quantizer in { }
E vi [n]v k* [m] = 0 , for i k and for all m,n. (9)
each channel is scalar and uniform and it has allocated
a budget of Ri bits, so that E{} stands for the statistical mean. This condition is
necessary and sufficient for optimality in orthogonal
M 1
1 transform coding, but not for orthonormal subband
R=
M R
i =0
i (5) coders.
D. Majorization Property
We assume that the quantization noise qi [n] is Let S k (e j ) denote the power spectrum density of
additive and independent of the signal, and the noise 2
source in different channels are wide sense stationary the k-th decimated subband signal v k [n] and vk , its
and zero mean. This model does not require each variance. Assume that the subbands have been
q i [n] white or uncorrelated with the others (for the numbered such that
case of biorthogonal filterbanks, the white,
uncorrelated noise model is required in coding gain v20 v21 ... vM
2
1 (10)
derivation). This is the standard model used at high bit
rates. Each quantizer is assumed to have the variance
of the form
{
The set of subband power spectra S k (e j ) has the }
majorization property if
direct
G(M ) = (7)
Fig. 3
Using the standard noise model, an expression for the
coding gain GSBC(M) of the orthonormal subband The optimum energy compaction problem is to
coder in Fig. 1 is derived in [1]: maximize the variance
107
solution. Given an arbitrary paraunitary matrix
2
2 d U (e j ) , we can always write E opt (e j ) as:
H (e
j
y2 = ) S x ( e j ) (12)
2
0
E opt (e j ) = F (e j )U (e j ) (16)
2
subject to the constraint that H (e j ) is Nyquist
where U(e j ) is unitary and therefore, nonsingular
(M).
x[n] is assumed to be zero mean WSS input, having matrix and F (e j ) is a diagonal matrix. Due to
relation (16), the filter bank in Fig. 2 can be redrawn
the power spectrum density S x (e j ) .
as in Fig. 4. Since U(e j ) is unitary, the errors
The compaction gain is given by
e[n] = x[n] x[n] and e y [ n] = y[n] y[ n] have the
y2 same mean square value.
Gcomp = (13)
x2
III. CODING GAIN MAXIMIZATION Now we can design F (e j ) to be optimal for its input
y[n] , i. e., it minimizes the mean square value of
In order to maximize the coding gain in rel (7), we
have to minimize the variance of the reconstruction e y [n] . If U (e j ) is optimal orthonormal, we can
error. In biorthogonal filter banks, according to Fig. 1, always split the design of the optimal biorthogonal
it can be writen as:
E(e j ) into design the optimal orthonormal system
1
M 1 U (e j ) which is a principal - component filterbank
E{ x[n] x[n]
2
= }=
M and design the biorthogonal F(e j ) for y[n] .
n=0
M 1 In Fig. 5a one branch of a filter bank is considered
1 2 d
c2
2
G i ( e j )
2 Ri
vi2 = and P(e j ) is choosen to satisfy
M 0 2
i =0 (14)
M 1
1 2 R
2 d P(e j ) = F 1 (e j )
c2
2
(17)
j j
H i (e ) S x (e )
M i =0 0 2
2 2 d In Fig. 5b the quantizer is modeled using the model
0
G i ( e j )
2 described in Section II.
M 1 2 d
2
= c 2 2 R H i ( e j ) S x ( e j ) Fig. 5
2
i =0
0
(15)
1/ M Let e y [ n] = y[n] y[ n] be the reconstruction error. Its
2 d
2
j
Gi (e ) variance is
0 2
2
2 2 d
P (e
By minimizing the product in the right hand side over e2y = E{ e y [n] } = q2 j
) =
the analysis/synthesis filters, they result signal 2
0
adapted. (18)
2 2
For the filter bank in Fig. 2, let d2 2d
S P (e
2 R j j j
j j c2 yy (e ) F (e ) )
E(e ) = E opt (e ) be an optimal biorthogonal 2 2
0 0
108
Taking into account rel. (17) and making use of the IV. THE TWO CHANNEL CASE
Chauchy Schwartz inequality, the previous relation
satisfies the following inequality The conditions to maximize the coding gain are not
the same as those to get the maximum compaction
2 gain (see rel.13).
d
e2y
0
S yy (e j )
2
(19) For the M=2 case, the maximization of the coding
gain is equivalent to the design of an optimum
compaction filter [3].
with equality only when The coding gain is
F ( e j ) = P * ( e j ) (20) x2
G SBC (2) = (23)
x20 x21
where is an arbitrary nonzero constant. Since
P(e j ) = F 1 (e j ) , must be of the form = K 2
But, due to the orthonormality, x20 + x21 = 2 x2 , so
and
that, relation (23) becomes
F (e j ) = KS yy
1 / 4
(e j ) (21) x2
G SBC (2) = (24)
x20 (2 x2 x20 )
This choice of F (e j ) depends on the input signal
statistics and assures the minimum value for the or, in terms of the compaction gain
variance of the reconstruction error and therefore, the
maximum coding gain. 1
In general, the optimal F (e j ) and its inverse may G SBC ( 2) = (25)
Gcomp (2)(2 Gcomp (2))
not be realizable filters. They may be replaced with
causal, stable approximations, which result in
suboptimum coding gain. The denominator in relation (24) is minimized by the
Considering the spectral flatness measure of the input choice of H 0 ( e j ) subject to filterbank
signal defined as [2] orthonormality condition, that is equivalent to
Nyquist(2) condition
2
d
exp{ ln S x (e j )
2
} 2 2
H 0 ( e j ) + H 0 ( e j ) = 2 (26)
x2 = 0
(22)
x2
If H 0 (e j ) is designed to be the optimum energy
we observe that the coding gain is large for small compaction filter for the given input power spectrum
values of x2 , that is, for a peaky spectrum and density S x (e j ) , that is, its output variance x20 is
small for values of x2 closed to unity, that is for a maximized under the Nyquist constraint, the other
relatively flat spectrum. filters H 1 (e j ) , G0 (e j ) and G1 (e j ) are chosen
In general, for fixed number of subbabds M, as in relation (1) to get perfect reconstruction.
biorthogonal filter banks provide better gain than
orthonormal filter banks. But, for a special shape of REFERENCES
power spectrum density of the input signal, when it
consists of contiguos constant frequency segments of [1] P. P. Vaidyanathan, Multirate Systems and Filter Banks.
length 2 / M , the performances of biorthogonal and Englewood Cliffs, NJ: Prentice Hall, 1993.
[2] Jayant N. S. and Noll P., Digital Coding of Waveforms.
orthonormal filter banks are the same. This is due to
Englewood Cliffs, NJ: Prentice-Hall, 1984.
the fact that when we use brickwall filter banks with [3] P.P. Vaidyanathan, Theory of optimal orthonormal
contiguos staking, the subband signal xi [n] has a subband coders, IEEE Trans. Signal Processing, vol. 46,
constant power spectrum density in the passband. pp. 1528 1543, June 1998.
Therefore, the signals {v k [n]} satisfy the
[4] M. Unser, On optimality if ideal filters for pyramid and
wavelet bases for signal decomposition, IEEE Trans.
decorrelation as well as the majorization properties Signal Processing , vol. 41, pp. 3591 3596, Dec. 1993.
and the filter bank is optimal orthonormal. [5] M. K. Tsatsanis, G. B. Giannakis, Principal component
filter banks for optimal multiresolution analysis, IEEE
Trans. Signal Processing, vol. 43, pp. 1766 1777, Aug.
1995.
[6] A. Kirac, P.P. Vaidyanathan, Theory and design of
optimum FIR compaction filters, IEEE Trans. Signal
Processing, vol. 46, pp. 903 919, Apr. 1998.
109
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract - This paper presents a new denoising method for 2. A non-linear filtering is applied in the wavelet
over-sampled constant within intervals signals corrupted by domain:
additive noise. The novelty of this paper is a special MAP
where t is a threshold. This system is called soft
filter, called composed bishrink. A complete statistical
analysis of this filter is reported. Some simulations are
presented. The results obtained are compared with the ( )
sgn { yi [k ]} yi [k ] t , yi [k ] > t
y0 [k ] =
( 2)
results of other denoising methods and with other state-of- 0, if not
the art filtering techniques.
Keywords: wavelets, denoising, soft thresholding, bishrink. thresholding filter. Because the noise ny is Gaussian, if
t>3n, the probability P(ny>t) is very little (the rule of
I. INTRODUCTION 3 sigmas). So the noise is quasi entirely suppressed.
This is the reason why the signal y0 is a denoised
In recent years, the techniques that use multiscale and version of the signal yi. This is a non-linear adaptive
local transform-based algorithms have become filter whose statistic analysis was presented in [4].
popular in noise filtering applications. In particular The adaptability is due to the selection of the
the use of non-linear filters in the DCT domain was threshold value in function of the noise power.
studied, [1]. In this paper, we consider local transform 3. Taking the inverse DWT (IDWT) of the signal y0,
based denoising. We propose such an algorithm, the denoised version of the signal s, x0, is obtained.
based on the discrete wavelet transform, (DWT). The principal disadvantage of the already described
Section II deals with the local DWT-based denoising. denoising method is due to the fact that it is based
In section III, a statistical analysis of the new only on the estimation of the noise variance (the
denoising method is presented. The use of local filters useful part of the input signal is ignored) and on
in the DWT domain is described in the following hypotheses confirmed only asymptotically. This is the
section. In section V, numerical simulation results are reason why in the following another denoising
presented and discussed. The last section is dedicated
strategy, based on the use of a Maximum a Posteriori,
to some concluding remarks.
MAP, filter, will be described.
II. LOCAL DWT-BASED DENOISING
III. A STATISTICAL ANALYSIS OF THE DWT
The following model of the observed signal corrupted
The probability density function, (pdf), of the wavelet
by additive noise is considered in this paper:
coefficients at the mth scale, (after m iterations),
x [k ] = s [k ] + n [k ]
k
(1) x Dm (k being equal with 1 for detail coefficients and
with 2 for approximation coefficients) is given by the
where s and n represent the useful part and the noise. following relation:
The problem is to estimate s starting from x. The N (k ) M0
noise is usually considered to be an uncorrelated with f k (a) = ...
s, stationary random process, with a null mean and a x Dm
variance n 2 . To estimate the signal s, Donoho, [2],
r1 = 1 q2 = 1
(3)
proposed the following method: M0
1. The Discrete Wavelet Transform (DWT) of the f d ( k , r1 , q2 ..., qm , a )
signal x is computed. The result is the signal yi = y +
ny. The noise ny converges asymptotically to a qm = 1
Gaussian white one, with the same variance, [3]. where:
110
f d ( k , r1 , q2 ..., qm , a ) = k = k + k (9)
x Dm s Dm n Dm
= G ( k , r1 , q2 ..., qm ) (4) If the input noise is a zero mean white Gaussian, the
correlation of its wavelet coefficients becomes:
f x ( G ( k , r1 , q2 ..., qm ) a )
and:
k
n Dm
[ n1 ] = n2 [ n1 ] (10)
111
position at the next scale (named the parent of the 2 mm
considered coefficient). This correlation can be l = 1 2 (24)
exploited to construct adaptive filters acting at a given and:
scale and using for the estimation of their parameters g, g > 0
information obtained at the next scale, [5]. Using the ( g )+ = (25)
parent and child wavelet coefficient of the input signal 0, if not
it is possible to estimate the child coefficients of the and for the models in (20) and (22) the solution of the
DWT of the useful part of the input signal, with the maximization problem in (19) is:
1m m
2
aid of a bishrink filter, [5]. Let 1 yi be the considered 1l
y= 1 yi (26)
m m 2
detail coefficient and 2 yi its parent. The statistical 1 2 m
+n
parameters of the child coefficients can be determined So, the input-output relations of the composed
using their parent coefficients and the neighbor child bishrink filter are (23) and (26). The noise variance is
coefficients, located in a window with a length of 3, estimated using the details obtained after the first
centered on the current child coefficient. It can be m m
written: iteration and the variances 1 and 2 are estimated
yi = y + n y (17) in moving windows centered on the current child and
parent coefficients. First the means are estimated in
where: each window and second the variances. But, applying
( ) ( )
y i = 1 yi , 2 yi ; y = 1 y , 2 y ; n y = 1n y , 2 n y (18) ( ) the relation (16), a different estimation of the local
variance of the child coefficients can be obtained:
The MAP estimation of y, realized using the m
2
observation y i , is given by: 1n
d = (27)
{( )}
2
y ( y i ) = arg max ln fn y ( y i y ) f y ( y ) (19) To profit of these two estimations of the local
y variances, obtained at two successive scales, it can be
In the following, we will consider that the DWT of written:
the noise is distributed following a zero mean m
2
Gaussian: 1m
m +
( 1ny ) +( 2ny ) 1m
2 2 2
= (28)
2
( )
fn y n y =
1
2 n
e 2 n2
(20) This estimation will be used in (24) and (26),
m
substituting 1 , for the input-output relations of the
Concerning the model of the DWT of the useful composed bishrink filter.
component, in the case of the composed bishrink
filter, for the first Nu2 iterations, a Laplace V. SIMULATION RESULTS
distribution will be considered (like in the case of the
bishrink filter, [5]): A useful input signal constant within intervals was
considered. This is a data sequence, specific for the
( 1 y ) +( 2 y )
3 2 2
3 communication in the base-band. This sequence has a
fy ( y ) = e (21) number of 16384 symbols, each having b=128
2 samples. A portion of this signal is represented in
and for the other iterations, a Gaussian distribution figure 1. Taking into account the waveform of this
will be considered: signal, the Haar mother wavelets must be used. The
( 1 y ) +( 2 y )
2 2 DWT was computed on blocks, each having a length
of 4096 samples. The maximal number of iterations
1
fy ( y ) = e 21 2 (22) (equal with 12) was used for the computation of each
DWT. For the implementation of the composed
2 1 2
bishrink, a value of Nu2 =8, was used.
(like in the case of the Wiener filter, [6]). For the
models in (20) and (21) the solution of the
maximization problem in (19) is:
2
3mn
( yi ) + ( yi )
2 2
1 2
l
1m + 1 Figure 1. The waveform of the useful component of the
y= yi (23) input signal.
( 1 yi ) + ( 2 yi )
2 2
112
Figure 2. The dependence of the output SNR of the input
SNR. Figure 3. A comparison between the use of an adapted filter
and a denoising system, in a base-band communication
denoising methods. The filter used in the wavelets application.
domain gives the difference. It can be observed that
these dependencies are linear. The curves describing adapted filter (ad.filt.part.sync 1, the dash dot line in
the bishrink filter and the composed bishrink filter are figure 3) for input SNRs superior to 10 dB. Also, in
superposed. These filters give the better results. These the third hypothesis, the denoising system is superior
are superiors to the results obtained using the soft- to the adapted filter (ad.filt.part.sync 2, the dotted line
thresholding filter. The poor results are obtained using in figure 3), for input SNRs superior to 10.47 dB.
the Wiener filter. Finally, a comparison between the Practically in this case the adapted filter cannot be
use of an adapted filter and a denoising system, into a used.
communication application is discussed. Each of the VI. CONCLUSION
two systems are connected at the output of a
communication channel, that adds a zero mean white In this paper is proposed a new denoising method
noise to the data sequence, which beginning is based on the use of the composed bishrink filter in the
represented in figure 1. The first system is a filter wavelets domain. This method takes into account also
adapted to a rectangle, having a duration equal with b. the statistics of the useful part of the input signal. That
Six experiments are made, with different noise makes that this method to perform better than the
variances. At the output of the investigated system denoising method using the soft thresholding filter for
(adapted filter or denoising system), an ideal sampling input SNRs superior to 10 dB. It can be used in
system is connected. Three hypotheses, concerning communications, replacing the adapted filter, when
the synchronization, are used. The first hypothesis the synchronization is difficult.
supposes a perfect synchronization. The second
hypothesis accepts a little loss of synchronization (the REFERENCES
sampling moments are delayed with 3 b / 8 ) and the
[1] K. O. Egiazarian, V. P. Melnik, V. V. Lukin, J. T.
third hypothesis accepts a more important loss of Astola, Local transform-based denoising of radar image
synchronization (the sampling moments are delayed processing, Nonlinear Image Processing and Pattern
with b / 2 ). The output of the sampling system is Analysis XII, Edward R. Dougherty, Jaakko T. Astola,
connected to the input of a comparator. The output of Editors, Proceedings of SPIE vol. 4304, 2001.
[2] D. L. Donoho, "De-noising by Soft Thresholding",
this comparator represents the output of the simulated Technical Report no.409, Stanford University,
receiving unit. The denoising system uses a soft December 1992.
thresholding filter when the input SNR is inferior to [3] M. Borda, D. Isar, "Whitening with wavelets",
16.65 dB and a composed bishrink when the input Proceedings of ECCTD'97 Conference, Budapest,
August 1997.
SNR is superior to 16.65 dB. The better result is [4] A. Isar, A. Cubichi, Miranda Naforni, Algorithmes et
obtained with the adapted filter with perfect techniques de compression, Editura Orizonturi
synchronization (ad.filt.perf.sync, the continuous line universitare, Timisoara, 2002.
in figure 3). For the other hypotheses, an analysis, [5] L. Sendur and I. W. Selesnick, Bivariate shrinkage
functions for wavelet-based denoising exploiting
taking into account the value of the input SNR must interscale dependency, IEEE Trans. on Signal
be made. The synchronization losses do not affect the Processing. 50(11): 2744-2756, November 2002.
performances of the denoising system [6] H. Zhang, A. Nostratinia, R. O. Wells Jr., Image
(denois.perf.syn, the dashed line in figure 3, Denoising via Wavelet-Domain Spatially Adaptive FIR
Wiener Filtering, IEEE ICASP, Istanbul, June 2000, vol.
denois.part.sync.1 and denois.part.sync. 2) but affect 5, 2179-2182.
the performances of the adapted filter. In the second
hypothesis, the denoising system is better than the
113
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract Due to the rapid expansion of the Internet and factors for the maximum amount of watermark bits that
the overall development of digital technologies, millions of can be stored in a data object.
users, who are scattered all over the world, are able to use a
vast number of multimedia products. Every participant in
this process wants to assert their rights, which are given by III. BUILDING WATERMARKING
their role in the business string. Naturally, solutions to
digital copyright protection are required urgently to tackle It consists of two parts:- The first part is concerned
the problem of unauthorized copying and distribution. The with insertion strategy i.e. where in the host signal shall
aim of this paper is concerned with inserting copyright
we place the information?. The second one is watermark
information into host image. In this paper, discrete cosine
transform (DCT) domain watermarking technique for
structure -how shall we place the additional information
copyright protection of still digital images is analyzed. The into the signal?. It is often necessary to utilize Human
DCT is applied in blocks of 8 8 pixels as in the JPEG Visual System (HVS) models for adaptively embedding
algorithm. The watermark can encode information to track the watermark. This can reduce the impacts of
illegal misuses concerned with the protection of copyright modifications on image quality or for the same visual
information contained in digital images. quality a much stronger watermark can be embedded.
The human eye is sensitive to the following
I. INTRODUCTION characteristics of image-contrast, frequency, luminance
sensitivity, edges and texture area[2]. One can combine
In general, digital images and digital video-streams the above four properties to construct a perceptual mask
can be easily copied one way or another. Even which determines the amount of modification permitted
though such copying may violate copyright laws, it is on each image cover data (pixels, transform coefficients)
widespread. The ease with which electronic images value. Using perceptual masks, energy can be added
may be copied without significant loss of content locally in places where the human eye cant notice it.
contributes to illegal copying. One of the goals of This increases robustness and hence capacity
digital watermarking is authentication for copyright
protection. To prove the ownership of an image, a
perceptually invisible pattern (a watermark) is IV. WATERMARK EMBEDDING APPROACH
embedded into the image and ideally stays in the
image as long as the image is recognizable. There are two general approaches to embedding a digital
watermark. One approach is to transform the host image
into its frequency domain representation and embed the
watermark data therein. The second is to directly treat
II. REQUIREMENTS OF WATERMARKING the spatial domain data of the host image to embed the
watermark. Bruyndonckx et al. in [3] proposed a spatial
Digital watermarking, particularly digital image domain scheme for copyright labeling of digital images
watermarking, has several conflicting requirements. based on pixel region classification.
The three most important requirements are
perceptibility robustness, and capacity[1]. For The advantage of spatial techniques is that they can
example: a very robust watermark can be obtained by be easily applied to any image, regardless of subsequent
highly modifying the host data for each bit of the processing (whether they survive this processing
watermark by increasing the watermark strength. however is a different matter entirely). A possible
However, this large modification will be perceptible. disadvantage of spatial techniques is they do not allow
As a second example, increasing the number of for the exploitation of this subsequent processing in
embedded bits increases the capacity but decreases the order to increase the robustness of the watermark.
robustness. Therefore, the maximum amount of
modification that can be acceptable for the quality of In addition to this, adaptive watermarking
the media and robustness are the two determining techniques are a bit more difficult in the spatial domain.
1
Department of Computer Science and Engineering, BRAC University,
Bangladesh, e-mail: rumel120@yahoo.com
114
Both the robustness and quality of the watermark could
be improved if the properties of the cover image could
DCT and IDCT are linear transformations and all
similarly be exploited. For example, it is generally
DCT coefficients are real. Any image block can be
preferable to hide watermarking information in noisy
represented as a superposition of scaled DCT
regions and edges of images, rather then in smoother
transformed images scaled with DCT coefficients.
regions. The benefit is two-fold: degradation in smoother
regions of an image is more noticeable to the HVS, and
secondly becomes a prime target for lossy compression A. Selection of DCT coefficient
schemes.
The low frequency components of an image are
Taking these aspects into consideration, working in a perceptually the more significant ones and any
frequency domain of some sort becomes very attractive. modification on them deteriorates the image fidelity.
Frequency domain watermarking was introduced by Cox Therefore, watermarking shouldnt be applied on low
et al.[4]. Coxs approach uses spread spectrum frequency components. On the other hand, the high
communication techniques to embed a bit in the image. frequency components are the ones, which are usually
However, it needs the original image to decode the less significant in terms of fidelity. As a consequence,
watermark and Smith et al.[10] refer to these approaches compression techniques utilize this property and
(when the original image is needed in the decoding suppress the high frequency components first to reduce
process) as of limited interest because of their narrow the size of images. Therefore, the watermarking
range of practical applications. The classic and still the techniques that modify high frequency coefficients
most popular domain for image processing is that of cannot be robust carriers of watermark. This leaves us
Discrete-Cosine-Transform, or DCT. Koch et al.[5] with the choice of mid frequency coefficients.
reported an efficient DCT domain watermarking
techniques resisting to JPEG compression. But our B. DCT based techniques
proposed approach is robust also against attacks such as
filtering, cropping, Scaling and geometric rotation. One such technique utilizes the comparison of
middle-band (FM ) DCT coefficients to encode a single
The DCT allows an image to be broken up into
bit into a DCT block. Suppose two locations Bi(u1,v1)
different frequency bands, making it much easier to and Bi(u2,v2 ) are chosen from the FM region for
embed watermarking information into the middle
comparison. Rather then arbitrarily choosing these
frequency bands of an image. The middle frequency
locations, extra robustness to compression can be
bands are chosen such that they avoid the most visual achieved if we base the choice of coefficients on the
important parts of the image (low frequencies) without
recommended JPEG quantization shown below in table
over-exposing themselves to removal through
2. If two locations are chosen such that they have
compression and noise attacks (high frequencies) [6]. identical quantization values, we can feel confident that
any scaling of one coefficient will scale the other by the
V. FREQUENCY DOMAIN TECHNIQUE same factor preserving their relative size.
One such technique utilizes the comparison of Table 1 Definition of DCT regions
middle-band DCT coefficients to encode a single bit into
a 88 DCT block. We first divide the NxN image into
(N/8)*(N/8) = N2/64 non overlapping 8x8 blocks; then
take DCT on each block and embed the watermark
FL
middle-band DCT coefficients 8x8 Discrete Cosine
Transform (DCT) is defined as: FM
I(u,v)=
m(u ) n(v) 7 7 ( 2 k + 1) u ( 2 l + 1) v
. X ( k , l ). cos( ) cos( )
2 2 k = o l = 0 16 16
115
Table 2 - Quantization values used in JPEG The watermarking procedure can be made somewhat
compression scheme [7] more adaptive by slightly altering the embedding process
0 0 7 to the method shown in equation 2.
16 11 10 16 24 40 51 61 I * (1 + k * W x , y ) In FM region
I W x, y = x, y -
12 12 14 19 26 58 60 55 Ix, y In FL and FH region
14 13 16 24 40 57 69 56 (2)
116
that their difference >= k. Finally the block is
transformed back into spatial domain.
For detection, the watermarked image is broken up
into those same 8x8 blocks, and a DCT is performed The
same PN sequence is then compared to the middle
frequency values of the transformed block.
117
(which is blurred also compared to the original host alignment. The bilinear interpolation can be
image). The reconstructed watermark is also still better approximated as an averaging filter.
in median filter.
The above figure shows watermarked image compressed The scaling experiment was done by scaling the
using lossy index-100 JPEG and index-25 JPEG watermarked image down to one quarter of its original
compression. The index ranges from 0 to 100, where 0 is size (256x256) and rescaled back to 512x512 using
the best compression and 100 is the best quality. The bilinear interpolation The algorithm requires the pixels
reconstructed watermark is a good reproduction in our in the watermarked image to be in the corresponding
experiment. location as the original host image in order to extract the
watermark correctly.
VII. CONCLUSION
118
[10] J. Smith and B. Comiskey, Modulation and information hiding in
images, , in Proc. First International Workshop on Information
Hiding, Lecture Notes on Computer Science, Cambridge, UK, pp. 207-
226, June 1996.
[11] A. Netravali, B. Haskell, Digital Pictures Representation and
Compression, Plenum Press, New York, 1988.[12] S.
Voloshynovskiy, S. Pereira, A. Herrigel, N. Baumgartner, T. Pun,
A generalized watermark attack based on stochastic watermark
estimation and perceptual remodulation, in: P.W. Wong, E.J.
Delp (Eds.), IS&T/SPIE's 12th Annual Symposium, Electronic
Imaging 2000: Security and Watermarking of Multimedia Content
II, SPIE Proceedings, Vol. 3971, San Jose, CA, USA, 23} 28,
January 2000.
119
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Digital medical images require that information these types of images. This can be extended to any
quality cannot be comphromised. With increasing image with a critically important region. As this type
demands to communicate these types of images of watermarking specifically survives JPEG
over wireless systems, the need to verify image compression, transmission of a small image file is
integrity is mandatory. This paper develops a possible without sacrificing image quality in the ROI
technique using embedded watermarking to verify where no watermark is placed. This scheme can be
that important detail has not been comphromised applied in one of two ways illustrated in Fig. 1.
as a result of incidental degradation or unexpected
compression.
Keywords: Semi-Fragile Watermarking, Authentication,
Medical Images.
1. INTRODUCTION
1
Centre for Biomedical Engineering (CBME)
The University of Adelaide, SA 5005, Australia e-mail dosborne@eleceng.adelaide.edu.au, dabbott@eleceng.adelaide.edu.au
2
School of Electrical and Electronic Engineering
The University of Adelaide, SA 5005, Australia, e-mail matthew.sorell@adelaide.edu.au, derek.rogers@unisa.edu.au
120
2. PROBLEM STATEMENT of schemes are viewed with suspicion by many
members of the scientific and medical community
Adding small amounts of noise to corrupt the who believe that image alteration may lead to loss of
bitstream of an image file that has been channel-coded diagnostic or scientific value.
does not usually affect the importance of the
diagnostic features present in the image after 3. PREVIOUS WORK IN ROI WATERMARKING
transmission has taken place. Incidental distortions
that are not corrected through channel decoding [4] The concept of ROI watermarking was first proposed
may slightly distort the file structure of the By A. Wakatani [8] who placed signature information
compressed image file without any noticeable change into the ROB. A progressively compressed version of
to perceptual quality. This could involve a loss of a signature image is used and the most significant
diagnostic feature information, which for medical information is embedded into the region closest to the
images is detrimental as detailed density information ROI. This method allows for the signature image to be
is mandatory. Hence it is critical to authenticate image detected with moderate quality from a clipped version
quality prior to any diagnosis that is made [5]. A of the image that included the ROI. This system was
classic example of this type of feature information is intended for use over web-based medical image
shown in the Infant's Fracture of Fig. 2. database systems with primary focus placed on
ensuring copyright and intellectual property
protection. The ROI area in the original image is
specified prior to compressing the signature image
using a progressive encoding algorithm to generate a
bitstream. This allows for increasing visual detail as
the extracted bitstream is followed. The payload is
embedded into pixels around the ROI in a spiral way
as depicted in Fig. 3
121
The robust watermark proposed is designed to survive Multiple embedding can give the receiver additional
after acceptable levels of low-pass filtering and JPEG- confidence in the unlikely event that both a watermark
2000 compression and not to survive malicious and signature are corrupted in an identical way and
attacks. This signature is based on wavelet coefficient the watermark is falsely detected as authentic. It may
properties of the ROI, where features are extracted also be of benefit if one watermark is corrupted.
based on absolute differences between corresponding Semi-fragile (or robust) watermarking is specifically
coefficients in the LH3 and HH3 subbands on 8 8 designed to withstand application specific
blocks. Similarly in this work a signature is based on transformation operations, such as lossy compression
the absolute differences between corresponding and geometric distortions, but is designed to be
coefficients in adjacent DCT tiles from inside the corrupted as a result of undesirable alterations
ROI, which have been uncorrelated as part of the including malicious manipulations and incidental
signature extraction process. It is mentioned in [9], degradation over mobile links which may or may not
that the procedure degrades the ROB significantly, be perceptible to the receiver. Semi-fragile robust
however this is not a primary concern as the ROB signature embedding ensures that the watermark
area is typically encoded at a low quality and gains survives JPEG compression or slight degradation up
minimal attention from users. The main focus of the to a point where the value of the work is lost. Because
works by [9] and [8] is copyright protection and ROI compression has been successfully subjectively
assurance that malicious attacks on the embedded evaluated in ROB of diagnostic medical images [10],
watermarks are prevented. Our focus is primarily the radiologist can have greater confidence that the
concerned with integrity verification as images are to diagnostic value of the image has not been
be transmitted in error-prone lossy transmission comphromised.
channels, such as those encountered in mobile phone
telephony to degradation experienced in Wireless The basis of singular semi-fragile watermark
Local Area Networks (LANs.) The most useful extraction and embedding was initially developed by
contribution in our work is assurance of ROI image C. Y. Lin [11]. Standard lossy image compression
content integrity after image files are subject to systems involves converting an image into some
incidental degradation in these environments. This is transform domain, such as wavelet or block DCT
made possible with extraction of DCT signature domain and quantizing the coefficients in order to
coefficients from the ROI and embedding multiply in reduce their entropy. Coefficients are quantized to a
the ROB. level proportional to how easy it is to perceive
changes in them and the property of quantization of
4. WATERMARKING TECHNIQUE USED coefficients is exploited to remove redundancy in the
image. Let x q be the result of quantizing x to an
If the signature information is lumped and localized integral multiple of a quantization step size, q.
within the ROB it is possible to authenticate and
verify the diagnostic integrity of such images. A x
simple method to multiply watermark involves x q = q + 0.5 (1)
q
embedding in the same shape of the ROI in the eight
regions surrounding the ROI or fewer regions if the
Consider s to be a real valued scalar quantity and q1
space in the ROB is unavailable. A visual impression
of this method is shown in Fig. 4 and q2 as quantization step sizes with q2 q1 , then:
((s q1 ) q2 ) q1 = s q1 (2)
122
as a semi-fragile watermark, which is illustrated on a The greater the embedding strength, the more
block level in Fig. 5. compression the image can survive and the more
perceptible the watermark will be. This is not a
problem as removal of the watermark can be
performed easily at the receiving end.
5. SYSTEMATIC IMPLEMENTATION
123
channel. The watermarked image is permitted to performance that is identical to JPEG. If degradation
undergo types of lossless compression, which will not of the ROI through JPEG quantization is not
degrade the image pixels or lossy JPEG, which can be permissible and hybrid coding is preferred as
applied up to a threshold specified by the user by the illustrated in the flow diagram of Fig.7, the bulk of the
embedding strength. Robustness to varying levels of bit budget will be stored in the ROI. This is because
JPEG compression took place on 100 grayscale quantization does not take place in this region and all
images of arbitrary types and varying resolutions from ROI transform coefficients must be encoded, which
256256 to 12801280 pixels. The ROI was specified are typically non-zero. As the ROB can undergo
to occupy a sufficiently small area at the center of compression through quantization, the majority of
each image so that 8 watermarks could be embedded coefficients will be zero. This will result in a file size
around this region. Results are shown in Fig. 8. that is dependent on the size of the selected ROI. The
larger this region is, the more near-lossless
compression is required, the larger the file size. For a
typical fracture or tumor, the area of the ROI does not
usually extend beyond 20% of the entire image. This
is also verified in work in [14] and [10], where ROI
Maxshift JPEG-2000 compression was utilized to
compress these types of medical images. Strom et al
[6] also validated the effectiveness of combined lossy
and lossless JPEG compression with these types of
ROI sizes. It was shown through extensive subjective
testing that the diagnostic value of the medical image
did not degrade for very low bit rate coding. These
approaches reinforce that the ROI is exactly the area
where all diagnostic information is located. Bit-rate
performance was evaluated in Fig. 9 with and without
the use of watermarking and with sizes of ROI
varying from complete lossy compression, where the
peripheral regions were the entire image to the
Fig. 8. Testing the robustness of 100 images with the semi-fragile extreme of having the entire image encoded near-
ROI watermarking scheme designed to withstand JPEG
compression. As expected, the system fails the authentication test losslessly as a ROI.
consistently after each of the three watermark embedding levels.
124
in comparison with complete near-lossless JPEG REFERENCES
encoding that provides minimal compression of 5 bpp
[1] D. Osborne, D. Abbott, R. Coutts and D. Rogers, An Overview
of the original grayscale image. After baseline JPEG
of Wavelets for Image Processing for Wireless Communications,
compression, which corresponds to using a SPIE Smart Structures, Devices and Systems, vol 4935, 2002, pp
quantization multiplier of 1, it is typical for the 427-435.
resulting quantized 88 DCT coefficient ROB blocks [2] D. Clunie, Lossless Compression of Grayscale Medical Images
Effectiveness of Traditional and State of the Art Approaches
to be reduced from 64 to an average of 2 coefficients. [3] G. K. Wallace The JPEG Still Picture Compression Standard,
This confirms the majority of the bit-budget is Communications of the ACM 34, 4th April, 1991, pp 30-45.
devoted to the ROI, which makes the compression [4] A. J. Viterbi, Error Bounds for Convolutional Codes and an
scheme useful for low bandwidth mobile Asymptotically Optimum Decoding Algorithm, IEEE Transactions
on Information Theory, IT-13, April, 1967, pp 260-269.
communication. This does assume that the ROI [5] D. Osborne, D. Abbott, M. Sorell and D. Rogers, Embedded
chosen is relatively small in comparison to the rest of Importance Watermarking for Image Verification in Radiology,
the image. SPIE BioMEMS and Nanotechnology , vol 5275, December, 2003,
pp 383-390.
[6] J. Strom and P. Cosman, Medical Image Compression With
8. CONCLUSION Lossless Regions of Interest, Signal Processing, vol 3, 1997, pp
155-171.
The use of watermarking can be used to verify the [7] N. Ahmed, T. Natarajan and K. R. Rao, Discrete Cosine
integrity of digital images. A method is developed Transform, IEEE Transactions on Computer Theory, vol c-23,
January, 1974, pp 90-94.
which is designed to be used on images that have [8] A. Wakatani, Digital Watermarking for ROI Medical Images
critically important regions. The scheme is designed Using a Compressed Signature Image, Proceedings of the 35th
to perform multiple watermarking around the ROI Hawaii International Conference on System Sciences. IEEE
containing features from this region. This technique Computer Society, vol 6, 2002, pp 157-163.
[9] W. N. Lie, T. L. Hsu and G. S. Lin, Verification of Image
can provide assurance that this region has not been
Content Integrity by Using Dual Watermarking on Wavelets
corrupted as a result of incidental degradation Domain, IEEE International Conference on Image Processing
resulting from a wireless link or that the ROI has not (ICIP), vol 2, 2003, pp 487-490.
exceeded a compression level threshold. As [10] G. K. Anastassopolous and A. Skodras, JPEG2000 ROI
Coding in Medical Imaging Applications, Proceedings of the 2nd
watermarked images can be compressed, a smaller file IASTED International Conference on Visualization, Imaging and
size is achievable. This facilitates verification of ROI Image Processing (VIIP2002.)
integrity as well as wireless communication. The [11] C. Y. Lin, A Robust Image Authentication Method
method used allows the user to evaluate quality of this Distinguishing JPEG Compression From Malicious
Manipulations, IEEE Transactions on Circuits and Systems for
region in a received image without the need of a Video Technology, vol 11, no 2, 2001, pp 153-168.
reference image. This is most useful for transmitted [12] D. A. Huffman, A Method for the Construction of Minimum
medical images where high levels of quality assurance Redundancy Codes, Proceedings of the IEEE, vol 40, 1962, pp
are mandatory prior to making any diagnosis. The 1098-1101.
[13] I. J. Cox, M. L. Miller and J. Bloom, Digital Watermarking
technique used can be systematically designed in two
1st Edition, The Morgan Kaufmann Publishers, 340 Pine St, Sixth
ways, one which fits in a modular way into the JPEG Floor, San Fransisco, CA 94104-3205, USA, 2002.
standard resulting in minimal changes and hybridly [14] D. H. Foos, E. Muka, R. M. Slone, B. J. Erikson, M. J. Flynn,
coding the ROI and the ROB resulting in superior D. A. Clunie, L. Hildebrand, K. Kohm and S. Young, JPEG2000
Compression of Medical Imagery Proceedings of the SPIE: PACS
ROI quality. Alternatively this method can operate Design and Evaluation for Engineering and Clinical Issues, vol
extraneously to the standard providing greater 3980, 2000, pp 85-96.
compliance with improved bit-rate performance. This
results in a degraded ROI when lossy JPEG is used on
the watermarked image pixels or improved picture
quality if lossless picture encoding techniques are
used. This method could also be used to monitor
image or video quality for quality control systems or
benchmarking image/video processing systems and
algorithms. A limitation is that authentication based
on watermarking cannot replace classical
cryptographic authentication protocols that protect
communication channels. Embedded robust
watermarking for ROI integrity verification can allow
for compression and provision of image integrity.
This can be most useful for medical images that must
be transmitted quickly in a wireless environment.
125
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Politehnica University of Timisoara, Communications Dept.
Bd. V. Prvan Nr. 2, 300223 Timioara, e-mail corina@etc.utt.ro
126
image. We take into account the fact that the HVS is To extract the mark from the watermarked possibly
not sensitive to small changes in high frequencies of distorted work, X
w
, we make use of the wavelet
the image, but is rather sensitive to changes affecting
the smooth parts of the image, that is, the coarsest coefficients d s , j (m, n ) , that should contain a
resolution level of the image. Therefore, we place the watermark bit:
mark into the wavelet domain, specifically, into the
d s , j (m, n ) d s , j (m, n )
w (m, n ) = sgn , (7)
HH, LH, and HL subbands, selecting only part of
these coefficients, leaving the LL subband d s , j (m, n )
unmodified.
A random guess is made for the watermark bit in the
A. Insertion procedure
location (m, n ) if d s , j (m, n ) = d s , j (m, n ) or if
Let X be the original gray-level image and the d s , j (m, n ) = 0 .
watermark W a pseudo random sequence, with binary
If the mark has been embedded in different locations
() { }
values: w i 1,1 and length N w . The basic several times, the most common bit value is assigned
steps for embedding the mark are: for the recovered watermark bit.
We make use of the correlation coefficient to compare
(a) Wavelet decomposition of the original image by L the original and the extracted mark:
levels to obtain a multiresolution decomposition:
w(n )w (n )
Nw
Y = DWT ( X ) c(w, w ) = n =1
(8)
{ } w (n ) w (n )
Nw 2 Nw 2
= LL , HL , LH , HH , HL
x
L
x
L
x
L
x
L
x
L 1 ,..., HH 1
x
n =1 n =1
127
of PSNR for each watermarked image, as a measure observer isnt very high. By embedding the
of the distortions introduced by the watermark: watermark bits into the edges and textures of the
image we make use of the human visual system. One
PSNR, proposed PSNR, Cox et al can see that both methods, proposed in [6] and ours
method method are image-dependant. Apparently, the Cox method is
Lenna 45.39 dB 27.19 dB superior for AWGN attack, comparable with the NC2
detector in the case of JPEG compression, and inferior
Boat 44.35 dB 25.35 dB for median filtering. However if we take into account
Barbara 44.18 dB 26.44 dB the fact visibility of the mark, an essential aspect of a
watermarking system, it is possible that our methods,
Peppers 45.55 dB 25.75 dB with the two proposed detectors (NC1 and NC2) to be
considered comparable or better than the Cox method
We present for each image the detector response as a in the given situation.
function of the filter size M, compression ratio and Future work will concentrate into the study of coding
signal-to-noise ratio, in case of median filtering, JPEG the watermark bits for a better performance.
compression and additive white noise, respectively.
The detector response was computed as a mean value ACKNOWLEDGEMENT
of 32 responses for 32 uncorrelated watermarks (Fig.
2-5). This work was supported by a grant from the
The plots marked with the o and + symbols are the Consiliul National al Cercetarii Stiintifice din
results from the proposed method, with the detector Invatamantul Superior, Romania, cod CNCSIS 47
NC1 and NC2 respectively, while the remaining plots TD.
are from the method proposed in [6]. REFERENCES
Setting the threshold value in the detection process at [1] G.Voyatzis, I. Pitas, Problems and Challenges in Multimedia
0.5 we have the followings. Networking and Content Protection, TICSP Series No. 3, Editor
Iaakko Astola, March 1999.
Median filtering attack: [2] I. Cox, M. Miller, J. Bloom, Digital Watermarking, Morgan
Kaufmann Publishers, 2002.
For all watermarked images, except Boat, the attack [3] A. Sequeira, D. Kundur, Communications and Information
by median filtering with filter size larger than M=3 Theory in Watermarking: A Survey, Multimedia Systems and
leads to a correlation smaller than 0.5. In fact, only Applications IV, A. G. Tescher, B. Vasudev, and V. M. Bove, eds.,
the detector NC2 allows filtering with filter size M=3. Proc. SPIE (vol. 4518), pp. 216-227, Denver, Colorado, August
2001.
For Boat watermarked image, not even the NC2 [4] M. Borda, I. Nafornita, Digital Watermarking Principles and
detector is successfully used in finding the mark. Applications, Proc. Of Int. Conf. Communications 2004, pp.41-54.
[5] S. Craver, N. Memon, B. Yeo, M. Yeung, Resolving Rightful
JPEG compression: Ownerships with Invisible Watermarking Techniques: Limitations,
Attacks, and Implications, IEEE Journal On Selected Areas In
For Lenna, the correlation is smaller than 0.5 at a Communications, Vol. 16, No. 4, May 1998.
compression rate of 16 (detector NC2 and Cox) and [6] I. Cox, J. Killian, T. Leighton, T. Shamoon, Secure Spread
10 (NC1), respectively. Spectrum Watermarking for Multimedia, IEEE Transaction On
For Boat and Barbara, the correlation is smaller than Image Processing, 6, 12, pp.1673-1687, 1997.
[7] B. Chen, G. W. Wornell, Quantization Index Modulation: A
0.5 at a compression rate of 13 for NC2, 10 for Cox Class of Provably Good Methods for Digital Watermarking and
and 7 for NC1. Information Embedding, IEEE Trans. On Information Theory,
For Peppers, the compression rate values for which Vol. 47, No. 4, May 2001.
the correlation is smaller than 0.5 is 15 (NC2, Cox) [11] D. Kundur, D. Hatzinakos, Diversity and Attack
Characterization for Improved Robust Watermarking, IEEE
and 8 (NC1). Transactions on Signal Processing, Vol. 49, No. 10, pp. 2383-2396.
[12] C. Nafornita, A. Isar, Digital Watermarking of Still Images
AWGN attack: using the Discrete Wavelet Transform, Buletinul tiinific al UPT,
For Lenna and Peppers, the detector response in the Tom 48(62), Fascicola 1, 2003, pp. 73-78.
[13] C. Nafornita, M. Borda, A. Kane, A Wavelet-Based Digital
Cox et al method is above 0.5 at a signal-to-noise Watermarking using Subband-Adaptive Thresholding for Still
ratio of 5 dB, having a considerably better Images, microCAD 2004 International Scientific Conference,
performance than detector NC1 (12 dB) and NC2 (15 University of Miskolc, 18-19 March 2004, pp.87-92.
dB). [14] N. Nikolaidis, I. Pitas, Robust Image Watermarking in the
Spatial Domain, Signal Processing, Vol. 66, No. 3, pp. 385-403,
For Boat and Barbara, the detector values are 1998.
approximately the same for each method: 3 dB (Cox),
around 14 dB (NC2) and 7 dB (NC1).
IV. REMARKS
128
(a) (b)
(c) (d)
Fig. 1: Original images used for simulations: Lenna (a), Boat (b), Barbara (c) and Peppers (d).
129
(a)
(a)
(b)
(b)
(c)
(c)
Fig. 3: Detector response to attacks against watermarked Boat
Fig. 2: Detector response to attacks against watermarked Lena: (median filtering, JPEG compression, AWGN). The plots marked
median filtering (a), JPEG compression (b), AWGN (c). The with the o and + symbols are the results from the proposed
plots marked with the o and + symbols are the results from the method, with the detector NC1 and NC2 respectively, while the
proposed method, with the detector NC1 and NC2 respectively, remaining plots are from the method proposed in [6].
while the remaining plots are from the method proposed in [6].
130
(a)
(a)
(b)
(b)
(c)
(c)
Fig 5: Detector response to attacks against watermarked Peppers:
Fig 4: Detector response to attacks against watermarked Barbara: median filtering (a), JPEG compression (b), AWGN (c). The
median filtering (a), JPEG compression (b), AWGN (c). The plots marked with the o and + symbols are the results from the
plots marked with the o and + symbols are the results from the proposed method, with the detector NC1 and NC2 respectively,
proposed method, with the detector NC1 and NC2 respectively, while the remaining plots are from the method proposed in [6].
while the remaining plots are from the method proposed in [6].
131
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
I. INTRODUCTION
1
Faculty of Electronics and Telecommunications, POLITEHNICA University of Bucharest
2
The Research Institute for Artificial Intelligence, Romanian Academy
corresponding address: adriana_vlad@yahoo.com
132
started the study with Q = 27 intervals having in view texts.
some immediate applications in enciphering natural
Table 1. First order statistical description of the random process modelling the chaotic systems.
k = 200 k = 300 k = 500 k = 2000
p r p r p r p r
I1 0.0000; 0.0370 0 0 0 0 0 0 0 0
I2 0.0370; 0.0741 0 0 0 0 0 0 0 0
I3 0.0741; 0.1111 0.05000 0.08543 0.05380 0.08219 0.05493 0.08130 0.05160 0.0840
I4 0.1111; 0.1481 0.03980 0.09626 0.04330 0.09212 0.04450 0.09082 0.04590 0.08935
I5 0.1481; 0.1852 0.03500 0.10291 0.03540 0.10231 0.03590 0.10156 0.03510 0.10276
I6 0.1852; 0.2222 0.02810 0.11526 0.02770 0.11612 0.03030 0.11087 0.02450 0.12367
I7 0.2222; 0.2593 0.02500 0.12116 0.02320 0.12717 0.02600 0.11996 0.02290 0.12802
I8 0.2593; 0.2963 0.02020 0.13650 0.02240 0.12948 0.02200 0.13067 0.02250 0.12918
I9 0.2963; 0.3333 0.02170 0.13160 0.02190 0.13098 0.02140 0.13253 0.02090 0.13414
I10 0.3333; 0.3704 0.07290 0.06989 0.06990 0.07149 0.07000 0.0750 0.07110 0.07084
I11 0.3704; 0.4074 0.04290 0.09257 0.04380 0.09155 0.03910 0.09716 0.04250 0.09303
I12 0.4074; 0.4444 0.04990 0.08552 0.04750 0.08776 0.04840 0.08690 0.04800 0.08728
I13 0.4444; 0.4814 0.02500 0.12240 0.02800 0.11164 0.02740 0.11677 0.02410 0.12472
I14 0.4815; 0.5185 0.02670 0.11833 0.02990 0.11164 0.03000 0.11145 0.02900 0.11341
I15 0.5185; 0.5556 0.02420 0.12445 0.02650 0.11870 0.02500 0.12240 0.02720 0.11721
I16 0.5556; 0.5926 0.02380 0.12552 0.02310 0.12745 0.02200 0.13067 0.02330 0.12689
I17 0.5926; 0.6296 0.02350 0.12634 0.02400 0.12498 0.02360 0.12606 0.02830 0.11484
I18 0.6296; 0.6667 0.02580 0.12043 0.02310 0.12745 0.02350 0.12634 0.02650 0.11879
I19 0.6667; 0.7037 0.02620 0.11949 0.02460 0.12341 0.02310 0.12745 0.02290 0.12802
I20 0.7037; 0.7407 0.02680 0.11810 0.02430 0.12419 0.02360 0.12606 0.02300 0.12774
I21 0.7407; 0.7778 0.02665 0.11879 0.02370 0.12579 0.02470 0.12326 0.02680 0.11810
I22 0.7778; 0.8148 0.02940 0.11261 0.02630 0.11925 0.02885 0.11443 0.02870 0.11402
I23 0.8148; 0.8519 0.03260 0.10676 0.03080 0.10994 0.02850 0.11443 0.03000 0.11144
I24 0.8519; 0.8889 0.06710 0.07308 0.07170 0.07052 0.06680 0.07325 0.06360 0.07520
I25 0.8889; 0.9259 0.07270 0.06999 0.07040 0.07122 0.06780 0.07267 0.07210 0.07031
I26 0.9259; 0.9630 0.10590 0.05690 0.11020 0.05569 0.11000 0.05575 0.11090 0.05549
I27 0.9630; 1.0000 0.07780 0.06747 0.07450 0.06908 0.08300 0.06514 0.07860 0.06710
Fig. 2. Histograms for the frequencies distribution of Table 1: Q = 27 intervals, at k = 200 (left) and at
k = 500 (right); on the vertical axis the occurrences of the intervals.
133
We started the investigation with verifying the first Thus, we continued the study with applying the test
order stationarity of the chaotic signal by considering on the equality between two probabilities (see
the Q discret intervals. This implies a comparative Appendix ).
study of the discrete random variables sampled at We succesively compared the two data sets (one for
different iterations. k = 200 and another one for k = 500 ) for each Ii
We determine the probablity that at the k iteration interval ( i = 1 27 and j = 1 27 ) in Table 1.
the chaotic signal passes through a certain Ii interval Table 2 presents the results only for I10, I12, I18 and I22
(chosen from the Q possible intervals). Then, we try intervals. All the four tests were passed; the test
to verify if this probability depends or not on the values and the decisions are shown in Table 2. As a
k iteration (the sampling time) while preserving the conclusion, the stationarity assumption is again
same Ii interval. For the statistical inferences used the sustained.
experimental data should comply with the i.i.d. model
(i.e. observations coming up from independent and Table 2. Experimental values for the test on the
identically distributed random variables). Moreover, equality between two probabilities
for the statistical tests we compared independent data I10 I12 I18 I22
sets. Test: T1
The first experimental results are presented in Table 1 z 0.0819 1.2248 1.2910 1.4523
where we considered four different iterations:
H 0 / H1 H0 H0 H0 H0
k = 200 , k = 300 , k = 500 , k = 2000 and
N = 10000 trajectories for each sampling time. Note 2*10-11 9.8*10-6 0.1532 0.0328
that for each k sampling time we used N = 10000
different trajectories, generated by different initial Because all tests are passed, the probability of type
conditions (randomly chosen from (0; 1) interval)).
II statistical error (that means H 0 accepted, although
Hence, for the four iterations k = 200 , k = 300 , the two compared probabilities are not equal) is
k = 500 , k = 2000 we had at our disposal four important. It was computed according to (4) (see
independed i.i.d. data sets (that means we generated Appendix).
40000 different trajectories of the chaotic signal). Fig. 3 shows the values as a function of p1 .
For example for the I10 = (0.3333; 0.3704) interval,
There are three plots for values: = 0.1 ,
at k = 200 , the estimated value of the p probability = 0.15 , = 0.2 , N1 = N 2 = N = 10000 , = 0.05 .
that the chaotic signal passes through I10 is Table 2 presents the values for the corresponding
p = m / N = 0.07290 (m is the occurrence number of intervals, when = 0.20 . For = 0.10 , values
the investigated interval in the considered i.i.d. data are much larger. For a better accuracy (low values for
set). while < 0.15 ) we need to resume the experiment
Each time we experimentally checked-up the de generating much more trajectories of the chaotic
Moivre-Laplace conditions in the form
signal.
Np (1 p ) 14 , [4]-[7]. As a consequence we can
say that the p true (unknown) probability lies inside
the ( p * (1 r ); p * (1 + r )) = (0.06780; 0.07799)
interval computed with 1 = 0.95 statistical
confidence level;
r = z / 2 * p (1 p ) / N = 0.06989 is the relative
experimental error, where z / 2 = 1.96 is the
/ 2 point value corresponding to the standard
Gaussian law (of 0 mean and 1 variance).
For the same I10 = (0.3333; 0.3704) interval, but at
k = 500 iteration, the estimated value is p = 0.07000 ,
the 95% confidence interval is (0.06499; 0.07500)
and the relative experimental error r = 0.0750 . It Fig. 3. The type II error size for the test on equality between
probabilities. On the horizontal - p1 probablity ; on vertical -
can be noticed that the two confidence intervals for
the probability overlap; this brings some evidence in values. The curves corresponding to:
the favor of the stationarity assumption. = 0.1 - +, = 0.15 - and = 0.2 -
The fact that the confidence intervals overlap
encouraged us to a more detailed investigation.
134
Fig. 4. Frequency distribution representation: the histograms if we discretize in Q=6 intervals at k = 200
(left) and at k = 500 (right); on the vertical axis the occurrences of the intervals.
Fig. 2 shows histograms coresponding to the probability of the test was the temporal value for the
frequency distribution from Table 1. Instead of the corresponding Ii interval in Table 3.
relative frequencies of the intervals, the histograms
are constructed on the basis of intervals occurences. Table 3. Temporal description
The study was resumed for Q=6 intervals. This x0 = 0.31 x0 = 0.456 x 0 = 0.758
number of Q=6 intervals could be of some interest in
the cryptographic field when two iterations are I10 0.0725 0.0711 0.0708
simultaneously considered and assigned to an I12 0.0481 0.0487 0.0483
alphabet character of the natural language (for
exemple letters, punctuation marks). Fig. 4 presents I18 0.0214 0.0219 0.0213
histograms for the random process modelling the
I22 0.0253 0.0273 0.0234
chaotic signal.
For a temporal description we generated several
individual trajectories of the chaotic signal for Thus, the null hypothesis H 0 has the form
L = 10000 iterations. We measured how many times H 0 : p = p 0 where p 0 denotes the temporal
the investigated trajectory (randomly chosen from the
ensamble) passes through a certain interval of values; probability obtained for a certain trajectory.
~ the occurrence number. The relative occurrence We successively applied this test (Table 4) for I10, I12,
be m
I18 and I22 interval considering the i.i.d. data sets
number ~ p=m ~ / L was computed for each I interval of
i
obtained at k = 500 . The theoretical p 0 probabilities
values (Ii is the same from Table 1 where we
are those from Table 3 and the trajectory with initial
discretized the (0; 1) interval in Q = 27 non-
condition x 0 = 0.31 . All the tests were passed, thus
overlapping intervals of equal length).
sustaining again the ergodicity assumption.
Another issue was if ~ p=m ~ / L (the temporal relative
frequency of the investigated interval) lies inside of Table 4. Test of probability
the confidence interval for the probability Test I10 I12 I18 I22
corresponding to the same Ii investigated interval (at
k iteration). p0 0.0725 0.0481 0.0214 0.0253
As an illustration we used three curves with initial 0.9641 0.1402 0.8983 0.7642
z
condition: x 0 = 0.31 , x0 = 0.456 si x 0 = 0.758 and
four investigated intervals: I10, I12, I18 and I22 (see H0/H1 H0 H0 H0 H0
Table 3). We computed the temporal relative
frequency ~ p=m ~ / L of the investigated interval for
III. SECOND ORDER STATISTICAL
each trajectory. For example for the trajectory with DESCRIPTION
x 0 = 0.31 the temporal relative frequency
corresponding to I10 interval is ~
p = 0.0725 . Here, we again consider the (0; 1) interval of values of
Looking at Table 1, the 95% confidence interval for the chaotic signal discretized in Q non-overlapping
the probability assigned to I10 interval at k = 500 was intervals of equal length. For the second order
(0.06499; 0.07500) . We can see that ~ p lies inside statistical description we shall consider
this confidence interval for the probability. simultaneously two iterations ( k1 and k 2 ).
We resumed this type of investigation for each Ii This leads to the noisy information channel shown in
interval and several trajectories; all the numerical Fig. 5. Fig. 5 illustrates our procedure of
results sustained the ergodicity assumption of the first investigation considering Q = 6 intervals. As a
order distribution function. consequence, the input space X = {x1 ,..., xi ,..., x 6 }
We continued the verify this type of ergodicity by corresponds to the Ii interval at k1 iteration and the
using a test of probability [4], [6], [7]. In this test the
i.i.d. data sets is the same we used in Table 1 for a { }
output space Y = y1 ,..., y j ,..., y 6 corresponds to the
fixed k (the considered iteration) and the theoretical Ii interval at the k 2 = k1 + k iteration.
135
Table 6. Noise matrix estimation (proportions)
X Y p(yj/xi)
y1 y2 y3 y4 y5 y6
x6 I6
I1 y1 0.1020 0.0977 0.2239 0.1011 0.1236 0.3518
x1
x5 I5
x2 0.1229 0.1066 0.2051 0.1192 0.1048 0.3415
xi Ii
x3 0.1184 0.1055 0.2129 0.1144 0.1100 0.3388
p( y j / xi ) Ij yj
x4 0.1020 0.1067 0.2229 0.1199 0.1265 0.3220
The Table 6 shows the conditional probabilities Table 8. The mutual information of the channel
p ( y j / xi ) for k1 = 200 and k 2 = k1 + 50 = 250 . k1 k2 I ( X ;Y )
The mutual information corresponding to Table 6 is k = 10 300 310 0.136066
very low: I ( X ; Y ) = 0.001642 . This suggest the
independence between the input and the output (also k = 20 300 320 0.011214
revealed in Table 6 by the equality between k = 50 300 350 0.008182
probabilities estimates p( y j / xi ) p( xi ) , i = 1 6
and j = 1 6 ). k = 500 300 800 0.008017
We resumed this procedure of verifying the statistical k = 1000 300 1300 0.007521
independence. Table 7 shows some results that
indicate a k = k 2 k1 distance for which we can
speak about independence; this happens for k 30 . We also computed the conditional probability and the
The investigation was carried out on N = 10000 mutual information I ( X ; Y ) for different k1 and
trajectories. k = k 2 k1 . All results sustain the second order
This procedure based on the noisy information stationarity for the discrete random process assigned
channel assigned to discretized chaotic signal was to the chaotic signal.
further resumed for Q = 27 intervals. Some results
are presented in Table 8. IV. CONCLUSION AND OPEN PROBLEMS
136
This paper suggests how to obtain from the chaotic depends on the p1 and p 2 = p1 (1 ) value for fixed
signal a stationary discrete information source (and , N1 and N 2 . It is denoted by ( p1 , p 2 ) and is
according to case practically zero-memory) having the
same symbols as a printed natural language. For computed according to equation (4):
example we can generate an information source with
Q = 27 symbols that may correspond to printed
*z / 2
Romanian (the alphabet whitout blank and 1 (x ( p1 p2 ))2
punctuation marks), where we omit some very low
( p1, p2 ) = 2
exp(
2 2
)dx
frequency characters. The message generated by this *z / 2 2
information source (provided by the chaotic signal) (4)
may be a key in a various enciphering methods.
An immediate example that can be further used in where:
different variants is to make a summation modulo = p(1 p )(1 / N1 + 1 / N 2 ) ;
Q (successively for each character) between the
plaintext and the key. On the basis of the entropy
(redundancy) of the information source corresponding = p1 (1 p1 ) / N1 + p 2 (1 p 2 ) / N 2 .
to the key and also using some knowledge about the
entropy of natural language (the plaintext) we can
evaluate the performance of the cipher. REFERENCES
[1] Al. Serbanescu, Aplicatii ale sistemelor haotice in comunicatii,
ACKNOWLEDGEMENT Editura Academiei Tehnice Militare, Bucuresti, 2004.
[2] M.S. Baptista, Cryptography with chaos, Physics Letters, A
The autors wish to thank to Professor Jean-Pierre 240, (1998), pp. 50-54.
[3] Al. Spataru, Fondaments de la theorie de la transmission de
Barbot, the head of the ECS research group from linformation, Presse Polytechniques Romandes, 1987.
ENSEA, Cergy, France where this study was started, [4] J. Devore, Probability and Statistics for Engineering and the
for his assistance and support. Sciences, second edition, Brooks/Cole Publishing Company,
Monterey, California, 1987.
[5] A. Vlad, B.Badea, Metode statistice in prelucrarea datelor.
APPENDIX Estimarea statistica, Editura Paideia, Bucuresti, 2002.
[6] A. Vlad, B. Badea, M. Mitrea, Metode Statistice in Prelucrarea
In the appendix we briefly present the test on the Informatie. Compendiu si Aplicatii, Editura Metropol, Bucresti,
equality between two probabilities: 1999.
[7] V. Craiu, Verificarea Ipotezelor Statistice, Editura Didactica si
Be there two samples each complying with i.i.d. Pedagogica, Bucuresti, 1972.
statistical model with the sample size [8] A. Vlad, M. Ferecatu, M. Mitrea, Teoria transmisiunii
N1 = N 2 = N = 10000 . Denoting by m1 the number informatiei II: elemente teoretice ilustrate in MathCAD, Ed.
Paideia, Bucuresti 2002.
of successes of the event in the first data sample, the
probability estimate is p 1 = m1 / N1 . Similarly, in the
second data sample, the probability estimate is
p 2 = m 2 / N 2 . The two statistical hypotheses (null
hypotheses H 0 /alternative hypotheses H 1 ) are:
H 0 : p1 = p 2 and H 1 : p1 p 2 . We have to verify
whether the two estimates p 1 and p 2 derive from the
same theoretical probability. We apply the test based
on the z test value defined in (3):
z = ( p 1 p 2 ) / p1 (1 p1 ) / N1 + p 2 (1 p 2 ) / N 2 ,
where p1 = p 2 = p (m1 + m 2 ) /( N1 + N 2 ) (3)
137
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
communications datastreams.
PNX1300 processors are ideal building blocks for
devices required to process several types of PCI slot Keyboard Boot Receiver IR
multimedia datastreams simultaneously, including the Interface EEPROM
1
Facultatea de Electronic i Telecomunicaii Cluj-Napoca, Departamentul Comunicaii, Str. G.Baritiu 26-28,
e-mail radu.arsinte@com.utcluj.ro
2
Tedelco SRL, Calea Turzii 42, Cluj-Napoca
138
Few details regarding this block schematic. III. SOFTWARE DEVELOPMENT IN NEXPERIA
1. Processor PNX1302 ENVIRONMENT
- offers data processing capabilities
2. Tuner A professional application is constructed within the
- RF processing of incoming TV signal framework of software architecture optimized for
- re-encodes the audio/video information into RF streaming multimedia data. This framework allows
channel software modules to be developed independently
3. Channel decoder (TDA10046) because it clearly defines the interface between these
- COFDM demodulation components. A programmer can easily integrate
- outputs TS (Transport Stream) to Nexperia diverse modules as they connect in a common way.
4. CI Connector (Common Interface) This software architecture is known as the TriMedia
- Links the receiver module with CI (Conditional Streaming Software Architecture (TSSA) [6]. Several
Access) module dozen TSSA components are now available, and they
- Transmits a scrambled signal and receives the are used extensively in the design of the complete
descrambled signal product. TSSA uses a data driven design. The RTOS
5. Video encoder provides a foundation that allows the system to be
- transforms video output stream (VO) of PNX1302 in factored into independent tasks that communicate
CVBS PAL/SECAM/NTSC using queues and semaphores. A given task will sleep
- video data transfer is performed using ITU656 until data is available, process the data, send it along,
standard and sleep again.
6. Audio encoder The architecture of a TSSA application is given in
- Converts audio I2S in analog audio Fig.2. [5]
7. PCI slot The priority-based scheduler of the RTOS kernel
- standard PCI interface used to add (interface) of handles scheduling. Priorities are generally set using a
compatible devices rate-monotonic rule. A priority-based scheduler is
8. Flash memory chosen over a deadline scheduler because of its
- stores the executable program predictable behavior in overloaded conditions. High
9. RAM memory priority tasks continue to meet their deadlines, while
- temporary stores data and settings low priority tasks are deferred.
10. MS (Micro Stick) interface TSSA has several features. Some may be useful for a
- used to store data (removable peripheral) given application, some may not. TSSA brings in all
11. Keyboard Interface aspects of the TriMedia software architecture. It
- used to read local keys describes a method of constructing and connecting
12. JTAG interface autonomous, task-based components that
- used in debugging processes stream data between them.
Most of the components are common for a TSSA provides a framework for components, whether
microprocessor system. The core is the Nexperia streaming or not, that includes:
processor with features adapted to real-time - A standard Application Programmer Interface (API)
processing of audio-video data. - Common data formats (as defined in header files
All TriMedia (Nexperia) processors consist of an that are contained on the CD)
internal 32-bit high-speed data bus, which is
connected to external SDRAM. Attached to this
highway are chip internal DMA interface blocks, the
VLIW CPU, and coprocessor blocks.
A heavily simplified diagram is presented in Fig.2 [5].
139
kit). IADK contains the libraries of all the 4.AAC decoder/player: this application decodes and
components needed in applications and the NDK plays the encoded AAC files
(Nexperia Development Kit). We had also software Characteristics:
support for the stand-alone systems (SAS). - program was tested on Philips_ATV1 board
The environment has the following folders: - using non-streaming architecture
-audio: the libraries for the audio software - limitation given by RAMDISK size
components like Audio Digitizer, Audio Renderer or - doesnt work in real time(at this moment)
MP3 Decoder. - compiled with nohost option for our board
-video: libraries for the video components like Video
Digitizer, Video Renderer, Mpeg decoder etc. 5.AAC encoder :this application encodes .wav and
-tssa: libraries for some components that make some .pcm files to AAC format
actions like File Reader or Copy IO Characteristics:
-mdm: libraries for Transport Stream Demux and - program was tested on Philips_ATV1 and works
Programme Stream Demux components satisfactory
-net :libraries for HTTP network communication - is using non-streaming architecture
support and for RPC sockets - compiled with nohost option for our board
-build: the directory where we built components file
libraries and the applications for our board. 6.AVI decoder :this application decodes
-sas: contains the SAS environment support. uncompressed AVI files and displays the images
Characteristics:
IV. RESULTS - program was tested on Trimedia Zapper board
- uses non-streaming architecture
The development system was used to test some - limitation given by RAMDISK size
original applications useful in laboratory works and - the program works fine for the small AVIs
demonstrations. Most applications avoid the copyright - compiled with nohost option for our board
problems using original implementations for
algorithms and code sequences. This was a 7.MPEG-2 video decoder: this application implements
requirement of the project, making the applications the MPEG-2 decoding algorithm (IDCT, Huffman,
independent of IADK, which costs about 10.000$ for etc) and displays the images
the smallest configuration. It is necessary to have only Characteristics:
NDK, which has an affordable cost. - program was tested on Philips_ATV1 board and
Here are some of the applications and brief results of it works
tests. - almost reaches real-time(93-95%)
- using non-streaming architecture
1.PCM Player: this application plays PCM files - compiled with no-host option for our board
(*.pcm), which contain audio PCM sequences
Characteristics: 8.Image processing: some standard operations like
- program was tested binarization, edge detection were implemented over a
- using non-streaming architecture picture.
- has a video interface, coming from YUV files
- limitation given by the RAMDISK size Every application has a complex structure, and a
- compiled with nohost option for our board thoroughly implementation. For exemplification we
are presenting some details of the MP3 encoder
2.YUV Player: this application displays images that implementation.
are read from the .y, .u and .v files
Characteristics: Digital
Audio
Bit or
Quantized
Bitstream
Encoded
Filter Samples Bitstream
- program was tested on DVB T board and works Input
Bank Noise Formatting
128kb bitrate
Characteristics: Fig.4. Block diagram of the MP3 encoder
- program was tested on DVB-T board
- doesnt works in real time(for the moment) This simple diagram (fig.4), extracted from MPEG
- limitation given by the RAMDISK size standard [7] results in a lot of tables and associated
- compiled with nohost option for our board files, and presenting them is far beyond the space
allocated for this article.
140
The source files for the encoder have at least 4000 Future work will be concentrated on program
program lines (including tables). optimization, and extension of the application base.
After compilation this results in an .out file of about For the moment the goal remains educational
10MB. This could look a large file, but the file application development, but some applications could
includes the source file (MP3), the input and output be ported on embedded processors from Nexperia
buffers. This is a small amount of the RAM available Family (PNX85xx series) for commercial usage.
(32MB) in the DVB-T system used in our
experiments. REFERENCES
V. CONCLUSION AND FUTURE WORK [1] Texas Instruments - TMS320DM641/TMS320DM640 Video/
Imaging Fixed-Point Digital Signal Processors Data Manual,
Our activity brought us the following achievements: June 2003 .
[2] ST Microelectronics - STi5518 Single-Chip SET-TOP BOX
1. Extended work with the compiler and with other Decoder with MP3 and Hard Disk Drive Support data sheet -
Trimedia tools; 2001
2. Unterstanding how makefiles work; [3] Radu Arsinte, Ciprian Ilioaei - Some Aspects of Testing Process
3. Creating the executables (*.out) for some specific for Transport Streams in Digital Video Broadcasting Acta
Technica Napocensis, Electronics and Telecommunications, vol.44,
applications ; Number 1, 2004
4. Simulating those executable files with tmsim ; [4] Radu Arsinte - A Low Cost Transport Stream (TS) Generator
5. Building the support for all given platforms: Used in Digital Video Broadcasting Equipment Measurements
Foxbox (ATV), DVE, Trimedia Zapper ; Proceedings of AQTR 2004 (THETA 14) - 2004 IEEE-TTTC-
International Conference on Automation, Quality and
6.Generating the library files (*.a) for all the software Testing,Robotics May 13-15, 2004, Cluj-Napoca, Romania
Trimedia components that came with the two support [5] Chuck Peplinski, Torsten Fink - A Digital Television Receiver
CDs. Constructed Using A Media Processor- Philips Semiconductor
1999
[6] *** - TriMedia Programmers Reference Manual (SDE 2.0)
Using low level and complexity software tools offers 1999, Philips Semiconductors
the possibility to obtain results comparable with the [7] *** - International Organisation for Standardisation - ISO/IEC
TSSA architecture, in small applications or JTC1/SC29/WG11 Coding of Moving Pictures and Associated
educational environments. Audio - ISO/IEC JTC1/SC29/WG11 - November / 1994
Obviously, the total efficiency is lower than a
professional application, using libraries optimized
many times, but for most applications efficiency is
less important than simplicity and affordability.
141
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Universitatea Politehnica Bucuresti - Facultatea de Automatica si Calculatoare, Sectia Calculatoare
Spl. Independentei Nr. 313, 060032, Bucuresti, e-mail irinam@cs.pub..ro
142
numbers [1]. In the above case the difference between Further, two more orientation are possible due to the
the shapes is 27 by XOR operation on the two sets. vertical and horizontal flips of the original region, like
Hence, in grid method two objects are similar in in Fig. 2 c) and e).
shape, if and only if the difference between their Both the shape size and the grid size affect the binary
binary representations is less than a prespecified number derived for a boundary. This problem is
threshold, and they have similar eccentricities. handled by choosing a fixed length of the major axis
However, it must be noted that the binary number (the standardized major axis) and then scaling the
obtained for the same shape with a different shape in a manner that the major axis of the shape
orientation in space or with a different scale will be equals the standardized major axis. Scaling
different. However, it must be noted that the binary normalization is thus achieved by scaling along the
number obtained for the same shape with a different major-axis so that the major axis of the shape
orientation in space or with a different scale will be becomes equal to the length of the standardized major
different. The criteria for invariance of indices is not axis. The shape is scaled along the minor axis
met and hence it is required to normalize the shape to proportionally in order to maintain the perceptual
achieve scale, rotation and translation invariance. The similarity of the shape, like in Fig. 3 a), b) and
normalization process involves three steps: (i) shape respectively in Fig. 3 c).
boundaries are normalized for rotation, (ii) they are
normalized for scale, (iii) they are normalized for
translation. The principals steps of computing grid
descriptor are: (i) binary image, (ii) major axis, (iii)
rotation normalization, (iv) scale normalization, (v)
translation normalization, (vi) scan grid cells.
The following definitions are needed to perform the
normalization process [3]:
Major axis: is the straight line segment
joining the two points on the boundary
farthest away from each other (in case of Fig. 3. a) and b) two similar shapes before scale normalization. c)
the shapes after scale normalization
more than one, select any one);
Minor axis: is perpendicular to the major axis To improve the efficiency of this method, another
and of such length that a rectangle with sides shape feature, eccentricity was used [1]. Eccentricity
parallel to major and minor axes that just of shape is the ratio of the major axis to the minor
encloses the boundary can be formes using axis. Therefore, for two objects to be similar, their
the lengths of the major and minor axis; sequences of numbers and their eccentricity values
Basic rectangle: the above rectangle formed should be similar [1]:
with major and minor axis as its two sides. a) If two normalized shapes have the same
A shape after rotation will have a different binary basic rectangle, the distance between them is
number. This is because rotation changes the spatial equal to the number of positions having
relationships between the grids and the shape. This different values in their corresponding binary
problem can be solved by normalizing the shape for sequences;
rotation. The purpose of rotation normalization is to b) If two normalized shapes have very different
place shape regions in a unique common orientation. basic rectangles (i.e., they have very different
Hence the shape region is rotated such that its major minor axis lengths), there is no need to
axis is parallel to the x-axis. There are still two calculate their similarity, because the shapes
possibilities as shown in Fig. 2 b) and d), caused by are very different. The difference threshold
1800 rotation. between minor axes depends on applications
and cell size. Normally, if the lengths of the
minor axes of two shapes differ by more than
3 cells, these two shapes are considered quite
different;
c) If two normalized shapes have slightly
different basic rectangles, it is still possible
these two shapes to be perceptually similar.
It is added 0s at the end of the index of the
shape with shorter minor axis, so that the
extended index is of the same length as that
of the other shape. The distance between
these two shapes is calculated as in the first
case a).
Fig. 2. a) a shape before rotation normalization b), c), d), e) the The grid descriptor algorithm is applicable for
shape after rotation normalization contour-based shape and it assumes shape boundary
coordinates have been known. This algorithm is
143
extended into describing region-based shape. The If we want to extract all the images which are
main extension to the grid method is the method of similarly with the shape from a query, this method of
finding the major axis and region interpolation after calculating distances between shapes doesnt influnce
scale rotation. the result. In the knearest-neigbor query [4], the users
In the case of region shape, boundary information is query is specified by a vector and an integer k. The k
not known. That is not practical to find the major axis objects whose distances from the query vector are the
of a region shape by traversing all the points in the smallest are retrieved. Using the grid method for
shape region, the computation would be O(N2), where evaluating the k nearest-neighbor query, the results
N is the number of pixels in the shape region. will not be conforming to human intuition. For
Therefore, the major axis for a region shape is found example, if we want to perform a k nearest-neighbor
by searching the outer border point pairs on the shape query to extract shapes similar with shape A from
boundary in a number of directions (e.g. 360 shapes of Fig. 4, for k=2, the result will be shapes D
directions). The algorithm for calculating the major and E. Shape B is the most similar with shape A,
axis involves three major steps: (i) find the bounding conform with human intuition. To obtain this, a new
box of the shape; (ii) find the pair of boundary points method to calculate shape differences was proposed.
in a number of directions; (iii) find the two points at Instead of calculating the difference between two
the furthest distance in the found boundary points. shapes like in grid method, we associated a weight to
The algorithm is described bellow: each cell in the grid that are covered by one shape and
1. Find the bounding box of the shape; not the other,. At a first stage, this weight is chosen to
2. Start from a line segment d0 which passes be inverse proportionally with the number of cells
through the shape center, trace from the two neighbors which are covered by the two shapes. The
end points towards the center along the line difference between two shapes will be the sum of
segment. If a shape point is found, it is a these associated weights. For example d(A, B)= 5-
boundary point. For every tracing, two 2+5-2+5-2+5-2=8, d(A, C) =8-3+8-3+5-0+5-0=20,
boundary points are found; d(A, D) = 8-4+8-5+8-7=8 and d(A, E) = 8-4+8-6+8-
3. Increase the angle of d0 by an increment of 7=7. In this case the shapes D and E are more similar
2 /n (n is the number of directions to trace), to shape A than shape C, too. But this is not sufficient.
repeat step 2; We must differentiate between a shapes peek
4. Repeat step 3 until boundary points at all produced by noise (shapes B and C) and a hole in a
orientations are found; shape (shapes D and E). The cells neighbors
5. Find the two points p1, p2 with the furthest considered for determining the associated weight of
distance in the above boundary points, then the respective cell may form a continuous sequence
p1p2 is the major axis. (like in case of shape B and C) or not (like in case of
shapes D and E). At a second stage, the weight
III. A NEW SHAPE SIMILARITY MEASURE associated with a cell will be multiply with a
predefined factor in case that the considered cells
Consider the following five shapes A, B, C, D, E, F neighbors will not form a continuous sequence. For
with similar eccentricities (7/4, 7/4, 7/6, 7/4, 7/4) from example if =2, (A, D) = (8-4)* +(8-5)* +8-
Fig. 4. If we notate with d(x, y), the distance between 7= 15 i d(A, E) = (8-4)* +(8-6)* +8-7=13 and
shapes x and y conform with the grid method, then d(A, B) = 8, d(A, C) = 20. Therefore, using this
d(A, B) = 4, d(A, C)=4, d(A, D)=3, d(A, E)=3. method for calculating distances between shapes, for a
Hence, shapes D and E will be more similar with k nearest-neighbor queries with k=2, the result consist
shape A than shapes B and C, which is not of shape B amd E, which is more similar with human
conforming to human intuition. intuition than grid descriptor method.
144
polygons. The average precision and recall of the V. CONCLUSIONS
shapes used as k nearest-neighbor queries is given
in Fig. 5. This paper presents a new shape similarity method
based on the grid descriptor representation. and
1.2
retrieval method based on the shapes contour
1
which has a better retrieval performance compared
0.8 to the distance histogram method. The method is
Precision
145
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Universitatea Politehnica Bucuresti Facultatea de Automatica si Calculatoare, Sectia Calculatoare,
Spl. Independentei Nr. 313, 060032 Bucuresti, email: irinam@cs.pub.ro
146
i [0, R-1] is the number of distances belong to
1 n 1 this distance range.
A = x i y i +1 x i +1 y i (1)
2 i =0 4. Normalize the distance histogram;
Two similar shapes at different scales will have
The area computed by (1) is a signed value, where different values of distances calculated at step 2. To
a negative sign indicates that the vertices are in make this representation invariant to scale these
clockwise order and a positive sign indicates that distances must be normalized for scaling. The
the vertices are in counter clockwise order. process of distance normalization consists of
The centroid coordinates are: dividing the value of all distances by the value of
the maximum distance. After that the value of all
_
x _ y distances will be in [0, 1]. And because the sample
x= ,y = (2) points are chosen based on the length of the edge,
A A evenly spread on it, two similar shapes with
different sizes will generate the same normalized
where distances. Therefore the method is invariant to
scale after normalization.
1 n 1
(x i+1 + x i )(x i y i+1 x i+1 y i )
This approach is invariant to translation because
x = (3) the distance set will not change after translating the
6 i =0 shape. The sample points are chosen in each edge
proportionally with the edges length and they are
and spread evenly on it. The location of the sample
points will not change after rotating the shape.
1 n 1
( y i+1 + y i )(x i y i+1 x i+1 y i )
Therefore this method is invariant to rotation
y = (4) because the distance set will not change after
6 i =0 rotating.
After representing the shapes by distance
2. Select a set of sample points in the boundary of histograms, the similarity among them can be
the polygon, and calculate the distances between calculated by Euclidean distance between their
the sample points and the centroid of the shape; distance histograms. For example, for two shapes
The number of sample points is variable. It can be with the distance histograms D1: (d10, d11, d12, ,
changed for different situations. But the sample d1R-1) and D2: (d20, d21, d22, , d2R-1) the distance
points are not uniformly distributed around the between them is:
boundary. Each edge has assigned a number of
sample points proportionally to its length. For
R 1
example, if the length of an edge is Li, the
perimeter of the shape is L, and the total number of
d(D1, D2) = (d1
i =0
i d2i ) 2 (7)
sample points is N, then the number of sample
points in that edge Ni will be:
III. A NEW SHAPE REPRESENTATION
L
Ni = i N (5)
L For the two different shapes in Fig. 1, applying the
distance histogram method, their histograms will
These points are evenly spread on the edge. be similar as it is shown in Fig. 2. Therefore, using
the distance histogram method the two shapes will
The sample points and the centroid of the shape are be similar, although they look different. This is
used to calculate the distances. For example, given because the distance histogram method discards
a sample point si = (xi, yi) and the centroid c = (xc, spatial information.
yc), the distance between them is:
d(s i , c) = ( x i x c ) 2 + ( y c y i ) 2 (6)
147
formed with a reference axis by the counter- Representing the shapes as above, one shape can
clockwise tangent to the boundary of a shape which have more edge directions that the other. But it
goes from a boundary point of a shape to the next cannot say that the shapes are different because the
one. In Fig. 3 the edge directions of the shape are shapes can be affected by noise. To make this
represented by their turning angles, where the supposition, first it is necessary a preprocessing
reference axis is considered to be the x axis. stage to approximate the boundary of a shape. The
Therefore, a shape will be represented by its edge boundary approximation is a process that
directions, each edge direction having associated a eliminates insignificant shape features and reduces
list of corresponding radii. the number of data points.
This preprocessing stage is used to reduce the
influence of noise and to simplify the shapes by
removing irrelevant features and keeping only
which are relevant. The boundary of the shape is
analyzed in a number of evolution steps [8]. On
every evolution step, a pair of consecutive line
segments s1, s2 is substituted with a single line
segment joining the endpoints of s1 and s2. The key
property of this evolution is the order of the
substitution. The substitution is done according to a
relevance measure K given by:
148
and does not care about the starting point. If an stage eliminates the noise and reduces the number
equality of all these differences is found, then the of edges, making easier to obtain the sample points.
shapes may be similar, otherwise the shapes are not
similar. If an equality between the two lists of V. CONCLUSIONS
directions exists, suppose the order of these lists of
directions is (dA1, dA2, , dAn) and (dB1, This paper presents a new shape representation and
dB2,, dBn). retrieval method based on the shapes contour
Compare the radii associated with the lists of which has a better retrieval performance compared
edges directions obtained in step 1. For example, to the distance histogram method. The method is
it must compare the radii corresponding to invariant to translation, scale and rotation. The
direction dA1 with the radii corresponding to distance histogram method discards spatial
direction dB1, and so on. In fact it must be information to obtain rotation invariant. In the
compared the lists of radii. To compare these proposed method, the radii together with the edges
values it is sufficient to compare the standard directions associated with them are used for shape
deviation of the radii associated with each representation.
direction. The distance between the two shapes will In the presented method a preprocessing step for
be the Euclidean distance between the standard reducing noise is necessary. However this step
deviations of radii corresponding to each direction. reduces the number of edges for a shape and
Consider that the standard deviations of radii computing the sample points from the shapes
corresponding to each direction are (stdA1, stdA2, boundary is easier.
, stdAn) and (stdB1, stdB2, , stdBn). Then
the distance between the two shapes will be: REFERENCES
[1] Ming-Kuei Hu: Visual Pattern Recognition by Moment
n
(stdA' stdB' )
Invariants. IRE Transactions on Information Theory 8 (1962)
d(A, B) = i i
2
(9) 179-187
i =1 [2] Charles T. Zahn and Ralph Z. Roskies: Fourier Descriptors
for Plane closed Curves. IEEE Trans. On Computer 21 (1972)
269-281
IV. RETRIEVAL EXPERIMENTS 3. Michael Reed Teague: Image Analysis Via the General
Theory of Moments. Journal of Optical Society of America 70
To test the retrieval performance of the proposed (1980) 920-930
[4] E. M. Arkin, L. P. Chew, D. P. Huttenlocher, K. Kedem, and
method compared with the distance histogram J.S.B. Mitchell An efficiently computable metric for
method, a retrieval framework has been comparing polygonal shapes. IEEE Transactions on PAMI,
implemented on a database with synthetic shapes. 13(3), March 1991.
The performance has been evaluated using [5] F. Mokhtarian, S. Abbasi, J. Kittler Efficient and Robust
Retrieval by Shape Content through Curvature Scale Space. Int.
precision and recall [6]. Workshop on Image DataBases and Multimedia Search,
Precision is defined as the ratio of the number of Amsterdam (1996) 35-42
similar shapes retrieved to the total number of [6] Lu, G.: Multimedia Database Management Systems. Artech
shapes retrieved. Recall is defined as the ratio of House Publishers, Boston (1999)
[7] Lu, G. and A. Sajjanhar: Region-based shape representation
the number of similar shapes retrieved to the total and similarity measure suitable for content-based image
number of similar shapes in the whole database. retrieval. Multimedia System 7, 2, (1999) 165-174 Volume 7 ,
Precision indicates accuracy of the retrieval and Issue 2
recall indicates the robustness of the retrieval [8] L. J. Latecki and R. Lakamper: Polygon Evolution by Vertex
Deletion. Proc. 2nd Int. Conf. on Scale-Space Theories in
performance.
Computer Vision, Corfu, Greece, Springer-Verlag (1999) 398-
The database used consists of approximate 3,000 409
polygons. The average precision and recall of the [9] http://www.efg2.com/Lab/Graphics/PolygonArea.htm, June
shapes used as queries is given in Fig. 4. 2001
[10] D. S. Zhang and G. J. Lu: Shape Retrieval Using Fourier
Descriptors. Int. Conference on Multimedia and Distance
1.2
Education, Fargo, ND, USA (2001)
1
[11] Shuang Fan: Shape Representation and Retrieval Using
0.8
Precision
0.6
Distance Histograms. Technical Report TR 01-14, department
0.4 of computer science, University of Alberta, Edmonton, Canada
0.2 (2001)
0 [12] Zhang, D. S. and G. Lu: An Integrated Approach to Shape
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Recall
Based Image Retrieval. The Fifth Asian Conference on
Computer Vision (2002) 652-657
Distance Histogram Proposed Method
[13] G. Iannizzotto and L. Vita: A New Shape Distance for
Fig. 4. Average retrieval performance of the two methods Content Based Image Retrieval -
jada1.unime.it/~ianni/publications/mmm96.pdf
The method proposed in this paper outperforms the [14] K-L. Tan and L.F.Thiang: Retrieving similar shapes
effectively and efficiently Multimedia Tools and Applications.
distance histogram method by precision and recall. Kluwer Academic Publishers (2001)
This method by the preprocessing stage for
boundary approximation has a supplementary step
compared to distance histogram method. But this
149
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Communications and Electronic Systems Department, Military Technical Academy, Bucharest, Address: 81-83, Bd. George
Cosbuc, Sector 5, Bucuresti, Romania, Tel. 0213354660 / 0197, e-mail: astoica@mta.ro, vic@mta.ro
150
If f a still increase, and f a f s 2 , theoretically the filter may be programmable by software to optimize
fi1m = f s f a component falls inside the first Nyquist the circuit respond and the filtering characteristics.
For practical signal processing the ideal sampler can
zone, and certainly this gives an unwanted signal at be obtained with an ADC followed by a FFT
the output of ADC. This case is similar to the analog processor, which only provide an output in the
mixing process and the alias signal fi1m = f s f a is frequency interval [0, f s 2] , even for useful signal
like an intermediary frequency for radioreceiver. or alias one. For this reason it is necessary to use the
For undersampling we consider the signal bandwidth filtering ahead the sampler circuit in order to remove
Ba = f 2 f1 = 2 f , which is symmetrical around the frequency components which are outside the first
carrier frequency, noted f cr . In fig. 2 is shown a Nyquist zone whose aliased components fall inside it.
The anti-aliasing analog filter is a good choice to
signal in the third Nyquist zone, centered around a
assure an unambiguously high speed sampling process
carrier frequency.
into the ADCs both for baseband sampling and for
Attenuation ML f cr MH undersampling.
a( f ) 2 f s f1 f1 f2 Generally, for analog filter circuits, the
3 fs f2
transition bandwidth is smaller if more poles are used
Band of in the filter design. In these conditions it is necessary
Image Image Image Image
DR
interest
to use filters with high complexity, and for most of
f ADC producers the Elliptic filters with more than 10
fs
poles are a popular choice. The Bessel filters or the
NZ = 1 NZ = 2 NZ = 3 NZ = 4 NZ = 5 Chebyshev filters (with the ripple error under 1 dB),
0 0.5 1 1.5 2 2.5
both with minimum 8 poles, represent another options
Fig. 2 The aliases signals for undersampling process
to implement the analog anti-aliasing filter. The
In this case the signal bandwidth must be Ba < f s , Butterworth filters, which achieve 6dB attenuation per
octave for each pole of filter transfer characteristic (or
and f = f 2 f cr = f cr f1 given from the Nyquist 20 dB per decade), is the most usual type of filter used
criteria with the centered carrier frequency. in low cost practical applications.
For the baseband sampled signals are used the low-
III. THE ANTI-ALIASING FILTERS pass analog anti-aliasing filter (LP-AAF), but for the
undersampled signals must use the band-pass analog
A generally block diagram for a data conversion anti-aliasing filter (BP-AAF).
system is shown in the fig. 3, with the analog filtering
at input and digital filtering of outputs in order to IV. THE DESIGN RECQUIREMENTS
reject all unwanted signals.
In the concrete applications it can be used only the frequency (the corner frequency), simply noted f a ,
analog filtering, or the digital filtering (Fast Fourier which is equal to the highest frequency in analog
Transform FFT, Finite or Infinite Impulse Response input signal spectrum, and the transition band, which
FIR or IIR), or both in very sophisticated systems. is the interval [ f a , ( f s f a )] .
There are some differences between the filtering in Only the frequencies from transition band have the
the analog domain and this filtering process in the aliases signals in the band pass [0, f a ] , but the
digital domain.
aliases components have the levels under the limit of
The analog filtering is more suitable for high
dynamic range, regarding fig. 4.
frequency, and can remove the extraneous noise
The band pass of LP-AAF is lower than the width of
spikes which can saturate the sample-and-hold
first Nyquist zone, and in these conditions the aliased
amplifier (SHA) at the input of ADC, and the alias
components between f a and f s 2 are not interest
signals which can produce distortion in the conversion
process. and do not limit the desired dynamic range for ADC.
The digital filtering has a proper action after analog- The dynamic range of LP-AAF is chosen based on the
to-digital conversion and for this reason cant reduce requirements for signal fidelity and ADC resolution.
the influence of analog noise in the conversion Usually the stop band attenuation is between 60 dB
process, but it can remove the noise injected in ADC and 80 dB, and the transition band is the interval
parts during the conversion. In addition, the digital [0.45 f s , 0.6 f s ] .
151
1
Attenuation fc = f s . (4)
a( f ) fa fs fa 2DR [ dB ]
20n
aPass
2 10 1
SNR = [ 6.02 N + 1.76] dB . (7)
DR <
where G is the desired gain for filter circuit, usually For ADC with 8, 12, or 16 bits it can obtain the
G = 1 , and the coefficients a0 , a1 , an 1 , an are given in desired SNR at 50 dB, 74 dB, or 98dB, which mean
reference [1], [2]. the desired dynamic range DR at 60 dB, 80 dB, or
The amplitude response in dB is given by relation: 100dB. If Ba is six times less than f s 2 , the
correction term in (8) increase SNR with 7.8 dB.
G In reference [1] are presented some interesting results
a = 20 log H ( j ) = 20 log n
, (3)
2 of LP-AAF simulation with Microchip FilterLab
1+ software, and the possible dynamic range with usually
c filters order is presented in table I.
Table I The dynamic range with various LP-AAF
where the cut off frequency is f c = c 2 f a , and Filter Dynamic range / Maximum attenuation
n is the order of filter, the number of poles for the order amax [dB]
transfer function. n Butterworth Bessel Chebyshev
In these conditions we have a single equation and two 4 80 66 90
unknown variables: n and c . 5 100 79 117
Usually the design process is started with the 6 120 92 142
attenuation a DR , and with n 4 . If the sampling 7 140 104 169
frequency is imposed in the digital system, it must to
calculate the corner frequency of low pass filter from The implementation of analog anti-aliasing filter may
the relation (3) as well: use the active cells with op amp, like successive
Salen-Key low pass filters or Multiple Feedback low
pass filters (MFB cells). These solutions are most
152
popular and offer a good satisfaction for ADCs which Flash (Parallel) ADC;
process signals in the first Nyquist zone. Successive approximation ADC;
Sigma-Delta ADC;
Pipeline ADC;
Bit per stage (Serial) ADC.
Based on our study and applications result a
recommendation for pipeline ADC in various
applications, because the producers offer numerous
variants of these ADCs, and their price isnt so high..
V. CONCLUDING REMARKS
153
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract The work briefly presents the resources of a II. THE RESSOURCES OF THE LOW CAPACITY
low capacity telephone exchange with temporal switching, ELECTRONIC TELEPHONE EXCHANGE WITH
using PCM encoding for the conversation signal, then the TEMPORAL SWITCHING
stages for performing local, outgoing and incoming
telephone links. Based on these stages, macro-states graphs
The general structure of a low capacity electronic
were designed to model the processes that perform
telephone links and then used to write the command
telephone exchange is presented in figure 1, where the
software using the assembly language for ATMEL family notes have the following meanings:
microcontrollers. I - interface to the external medium;
Keywords: macro-states graph, local telephone link, CIA - individual subscriber circuits;
outgoing telephone link, incoming telephone link, CJ - junction circuits;
command software. RC - temporal connection network;
MT - tones and calling signal machine;
I. INTRODUCTION C - microcontroller;
UC - command unit;
The process of establishing a telephone connection PC - personal computer.
through a low capacity electronic telephone exchange
(PABX) is a complexe sequential process. A graph of
states must be used in order to describe it. The
transitions between states are induced by certain events,
which are a generalisation of the input signal, meaning
that they can be input signals and/or the result of system
internal processing.
It is practically proven that the graph describing a
software real-time telephone connection, using periodic
interrupt operation, includes very many states.
Moreover, using this description, complex problems
emerge while implementing the services.
Analysing the steps of performing local, outgoing and
incoming telephone links, it results that certain steps are
repeated several times, having different initial or final Fig. 1.
data.
Based on these facts, there were defined macro-states, The external medium includes the subscribers' lines
which perform the steps of establishing a telephone link, (LAb) through which the subscribers' telephone sets are
such as: receiving the pulse selection information, connected to the exchange, and the junction lines (LJ)
determining the type of the call, checking the selection through which the exchanges themselves are connected.
information, establishing (interrupting) the connection, This interface performs the following functions:
transmitting (interrupting) a tonality, a vocal message monitoring the lines, receiving and transmitting the
etc. selection information, supply, adapting, galvanic
The macro-states are designed so that they can be called separation, protection, testing and processing the
from any program sequence and they return to any conversation signal which consists of sense separation,
program sequence also. filtering, sampling and holding, analogic-digital and
This last facility allows a very simple succession of digital-analogic conversion. These last functions are
macro-states and a simple development of telephone implemented with TCM29C13N integrated PCM
services. codecs.
The interface to the external medium consists of the
154
individual subscriber circuits and the junction circuits. that communicates serially with the main processor.
CIA performs the interfacing of the subscribers lines to
the exchange and the following functions, from the point III. STEPS OF PERFORMING A
of view of the command software: LOCAL TELEPHONE LINK
- detects the apparition of a call from a subscriber when
closing the DC loop; Performing a local telephone link between subscribers A
- receives the pulse selection information by converting and B (assuming that A is the calling subscriber while B
the line pulses in TTL-leveled voltage pulses; is the called subscriber) includes the following steps:
- transmits the calling signal to the subscribers. - in the first moment, subscriber A is standing by;
CJ performs the interfacing of the urban telephone - subscriber A picks up the receiver in order to initiate a
exchange lines to the low capacity exchange and the telephone conversation; the DC loop is closed, which
following functions, from the point of view of the performs the subscriber's A call to the exchange;
command software: - after this call, the exchange connects the calling
- detects the apparition of a call from a distant exchange; subscriber to the command unit, so that the selection
- receives the DTMF selection information from a information can be received;
subscriber connected to a distant urban exchange for - subscriber A receives the dial tone which constitues an
automatic processing the incoming calls; invitation to the transmission of the pulse or tone
- transmits the calling signal to the distant exchange; selection information;
- transmits the pulse and DTMF selection information to - the selection information transmission is initiated;
a distant urban exchange for automatic processing the - the dial tone transmission is interrupted;
outgoing calls. - the first digit of the selection information is received
For the reception and the transmission of the tone and processed;
selection information, specialized integrated circuits - after receiving the first digit, the type of telephone link
(M8880) are used as they are performing these functions is determined (local, outgoing or service);
with a minimal need for supplementary hardware. - if the first digit received is different from 9 and 0, it is
The temporal connection network has no internal a local telephone link and the reception of the other
blocking and performs the switch of any channel from digits in the selection information is continued;
any input PCM line in any channel in any output PCM - after receivig the selection information, the number is
line. verified, which means to determinate whether the called
The network is implemented using digital integrated subscriber exists, if the connection between the two
temporal switch MT8980 which features 8 PCM input subscribers can be performed and if the called
and 8 PCM output lines, each consisting of 32 channels, subscriber is free;
so that the switch has a capacity of 256 x 256 channels. - if the called subscriber's number does not exist, then an
The tones machine sends towards the subscribers all the unexisting number tone is transmitted towards the
signals required for the exchange's operation (dialing calling subscriber;
tone, busy, reverse call, unexisting number, warning, - if the telephone connection cannot be performed, then
false call etc.), through the exchange's connection a busy exchange tone is transmitted towards the calling
network, as well as the calling signal. MT can also send subscriber;
to the subscribers various information vocal signals. - if the called subscriber is busy, then the busy tone is
The digital switch which implements connection transmitted towards the calling subscriber;
network includes, from the point of view of the - if the number transmitted exists, the connection can be
command software, a control register, a data memory performed and the called subscriber is free then the
(256x8 bits) and a command memory (256x11 bits) reverse call tone is transmitted towards the calling
which includes a low section (256x8 bits) and a high subscriber and the calling signal is transmitted towards
section (256x3 bits). By programming the control the called subscriber;
register, the data and command memory can be written - if, after a certain time interval, the called subscriber
and read and, therefore, any link between any two answers, then the transmission of the reverse call tone
subscribers can be performed, as well as any tone or and the calling signal is interrupted;
vocal message can be transmitted to any subscriber. - afterwards, the connection between the two subscribers
The distributed command unit of the low capacity is commanded;
electronic telephone exchange is implemented using - after performig the connection, the subscribers start
several ATMEL family microcontrollers hierarchized on their conversation, and the system awaits for one of the
two levels of priority. subscribers to hang up;
On the lower level are the peripheral microcontrollers - after the conversation is finished (one of the
that control the individual subscriber circuits, the subscribers hangs up) the connection between the
junction circuits, the temporal connection network, the subscribers is interrupted;
tone machine and perform a large amount of simple but - the system returns to the initial status.
critical in time command and control operations. On the From each step of this process, the selection information
upper level is placed a main processor that performs the for a certain telephone service can be received and also
connection settings and telephone services. The user has each subscriber can hang up, and the process returns to
access to the command unit through a personal computer the initial status.
155
outgoing, service);
IV. THE MACRO-STATES GRAPH FOR THE M2 - macro-state for processing the selection
PROCESS OF ESTABLISHING A LOCAL information;
TELEPHONE CONNECTION M3 - macro-state for establishing the connection
between the two subscribers;
In fig.2 is shown the macro-states graph for the process M4 - macro-state for interrupting the connection
of establishing a local telephone connection through a between the two subcribers which were in conversation;
low capacity electronic telephone exchange, using the M5 - macro-state for transmitting to a subscriber a tone
steps described above. signal or the calling signal;
M6 - macro-state for interrupting the transmission of a
tone signal or of the callig signal;
M7 - macro-state for waiting the standing-by calling
subscriber;
M8 - macro-state for waiting the calling subscriber
which picked up the receiver has to initiate an action;
M9 - macro-state for waiting the calling subscriber
which has picked up the receiver has to initiate an action
and the standing-by called subscriber to answer the call;
MA - macro-state for conversation; it is a waiting macro-
state while the two suscribers have the receivers picked
up;
MB - macro-state for waiting the interruption of the
telephone connection.
The events are represented on the graph's arches and
have the following meanings:
a - subscriber A has picked up the receiver;
p - specific processings for the current macro-state;
i - the reception of the selection information;
t - time counting;
c1 - first digit received;
ni - unexisting number;
l - the connection can be performed;
b - subscriber B has picked up the receiver;
AE - outgoing call;
S - telephone service;
AL - local call.
Using the macro-states and the events defined above, the
process of performing a telephone connection through a
low capacity electronic telephone exchange can be
followed on the graph. Every macro-state includes
severals states and the evolution between these is
depicted also by a graph of stages.
156
- the calling subscriber receives a dialling tone from the
distant urban exchange;
- the pulse selection information is received from the
calling subscriber and it is DTMF transmitted to the
exchange;
- the state of the calling subsriber is monitored;
- if the calling subscriber hangs up, the conversation link
is interrupted and the junction line is released.
From each step of this process, the selection information
for a certain telephone service can be received and the
calling subscriber can hang up, and the process returns
to the initial status.
157
processed;
- if the called subscriber's number does not exist, then an
unexisting number tone is transmitted towards the
calling subscriber;
- if the telephone connection cannot be performed, then
an busy exchange tone is transmitted towards the calling
subscriber;
- if the called subscriber is busy, then the busy tone is
transmitted towards the calling subscriber;
- if the number transmitted exists, the connection can be
performed and the called subscriber is free then the
reverse call tone is transmitted towards the calling
subscriber and the calling signal is transmitted towards
the called subscriber;
- if the subscriber B does not answer in a specified time
interval (e.g. 40 seconds) the transmission of the reverse
call tone and the calling signal is interrupted and, after a
vocal warning message, the junction is released;
- if, after a certain time interval, the called subscriber
answers, then the transmission of the reverse call tone
and the calling signal is interrupted;
- afterwards, the connection between the two subscribers
is initiated;
- after performig the connection, the subscribers start
their conversation, and the system awaits for one of the
subscriber B to hang up;
- after the conversation is finished (subscriber B hangs
up) the connection between the subscribers is
interrupted;
- the system returns to the initial status.
During this process, the calling subscriber cannot access
any service of the low capacity exchange, while the
called subscriber can only access the telephone services
after the conversation starts.
158
designed the command program of the main processor [2]. GRINSEC: La commutation electronique, Ed. Eyrolles, Paris,
1988.
that controls a low capacity telephone exchange with
[3]. Hintz J.K., Tabak D., Microcontrollers. Arhitecture,
temporal switching, using PCM encoding for the Implementation and Programming, McGaw Hill, 1993.
conversation signal. [4]. Radulescu T.: Telecomunicatii, Media Publishing, Bucureti,
There were also written the command programs of the 1994.
peripheral processes for the individual subscriber [5]. XXX: ATMEL Microcontroller Family, Data Book, 1990.
[6]. XXX: MITEL, Data Book, 1992.
circuits, the junction circuits, the temporal switching [7]. XXX: Telecom Design Solutions. Teltone Component Data Book,
network and the tone machine. 1990.
The command programs, written in the assembly [8]. XXX: TEXAS INSTRUMENTS, Data Book, 1992.
language of ATMEL family microcontrollers, require an
amount of around 1-2 Kb of program memory for each
peripheral process and of 8 Kb for the main process of
setting telephone links.
The use of macro-states in the manner described above
allows easily re-sequencing them for implementing also
the telephone services.
REFERENCES
159
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract - The paper presents a new variable step- plant filter of length M, ui - the (1 x M) input
size adaptive algorithm for affine projection (VSS- vector, Ui = [ ui ui-1 ui-K+1]T the (K x M)
AP) with a faster convergence and lower input matrix ( T denote transpose), di = [ d(i) d(i-1)
misadjustment. The VSS-AP algorithm is compared d(i-K+1)]T a K x 1 vector , ei the residual
with the LMS algorithm. The possibility of real time
implementation of this algorithm is investigated,
error (K x 1) vector.
using a DSP microcomputer (ADSP 21161 Analog
Devices). The VSS-AP algorithm is the following:
1
POLITEHNICA University of Bucharest, Electronic and Telecommunications Faculty
Telecommunication Department, sorin@elcom.pub.ro
160
III. THE IMPLEMENTATION AND Interval timer
MAIN RESULTS On-Chip SRAM (1 Mbit)
SDRAM Controller for glueless interface to
The new algorithm was implemented on an SDRAMs
evaluation kit with DSP processor (ADSP21161 External port that supports:
EZ-KIT). Interfacing to off-chip memory peripherals
The ADSP-21161 SHARC DSP is the first Glueless multiprocessing support for six ADSP-
low-cost derivative of the ADSP-21160 featuring 21161N SHARCs
Analog Devices Super Harvard Architecture. Like Host port read/write of IOP registers
other SHARCs, the ADSP-21161 is a 32-bit DMA controller
processor that is optimized for high performance Four serial ports
DSP applications. The ADSP-21161N includes a Two link ports
100 MHz core, a dual-ported on-chip SRAM, an SPI-compatible interface
integrated I/O processor with multiprocessing JTAG test access port
support, and multiple internal busses to eliminate 12 General Purpose I/O Pins
I/O bottlenecks.
The ADSP-21161 offers a Single-Instruction- The main characteristics of the ADSP21161
Multiple-Data (SIMD) architecture. Using two are presented below:
computational units the ADSP-21161 can double High performance 32-bit DSP
cycle performance on a range of DSP algorithms. applications in audio, medical, military,
Figure 1 shows a block diagram of the ADSP- wireless communications, graphics,
21161, illustrating the following architectural imaging, motor-control, and telephony
features: Super Harvard Architecture-four
Two processing elements, each made up of an independent buses for dual data fetch,
ALU, Multiplier, Shifter and Data Register File instruction fetch, and nonintrusive, zero-
Data Address Generators (DAG1, DAG2) overhead I/O
Program sequencer with instruction cache
PM and DM buses capable of supporting four 32-
bit data transfers between memory and the core
every core processor cycle
161
and transmit pins, which supports up to Peripheral Interface (SPI)
16 transmit or 16 receive channels of interface
audio 2. 64-bit background DMA
Integrated peripheralsintegrated I/O transfers at core clock speed, in
processor, 1 Mbit on-chip dual-ported parallel with full-speed
SRAM,SDRAM controller, glueless processor execution
multiprocessing features, and I/O ports 3. 800 Mbytes/s transfer rate over
(serial, link, external bus, SPI, & JTAG) IOP bus
ADSP-21161N supports 32-bit fixed, 32- 4. Host processor interface to 8-,
bit float, and 40-bit floating point 16- and 32-bit microprocessors,
formats. the host can directly read/write
100 MHz (10 ns) core instruction rate ADSP-21161 IOP registers.
Single-cycle instruction execution, 32-bit (or up to 48-bit) wide synchronous
including SIMD operations in both External Port provides glueless connection
computational units to asynchronous, SBSRAM and SDRAM
600 MFLOPS peak and 400 MFLOPs external memories
sustained performance Memory interface supports programmable
1 Mbit on-chip dual-ported SRAM (0.5 wait state generation and wait mode for
Mbit block 0, 0.5 Mbit block 1) for off-chip memory
independent access by core processor and 24-bit address, 32-bit data bus. 16
DMA additional data lines via multiplexed link
400 million fixed-point multiply and port data pins allow complete 48-bit wide
accumulation operations (MACs) data bus for single-cycle external
sustained performance instruction execution
Dual Data Address Generators (DAGs) 32-48, 16-48, 8-48 execution packing for
with modulo and bit-reverse addressing executing instruction directly from 32-bit,
Zero-overhead looping with single-cycle 16-bit, or 8-bit wide external memories
loop setup, providing efficient program 32-48, 16-48, 8-48, 32-32/64, 16-32/64,
sequencing 8-32/64, data packing for DMA transfers
IEEE 1149.1 JTAG standard test access directly from 32-bit, 16-bit, or 8-bit wide
port and on-chip emulation external memories to and from internal
Single Instruction Multiple Data (SIMD) 32-, 48-,or 64-bit internal memory
architecture provides: SDRAM Controller for glueless interface
1. Two computational processing to low cost external memory
elements Extended external memory banks (64 M-
2. Concurrent execution--Each words) for SDRAM accesses
processing element executes the Multiprocessing support provides:
same instruction, but operates on glueless connection for scalable DSP
different data multiprocessing architecture and
Parallelism in busses and computational distributed on-chip bus arbitration for
units allows: parallel bus connect of up to six ADSP-
1. Single-cycle execution (with or 21161s, global memory and a host
without SIMD) of: a multiply Two 8-bit wide link ports for point-to-
operation, an ALU operation, a point connectivity and array
dual memory read or write, and multiprocessing be-tween ADSP-21161
an instruction fetch Serial Ports provide:four 50 Mbit/s
2. Transfers between memory and synchronous serial ports with
core at up to four 32-bit floating- companding hardware, 8 bi-directional
or fixed-point words per cycle, serial data pins, configurable as either a
sustained 1.6 Gigabytes/second transmitter or receiver, TDM support for
bandwidth T1 and E1 interfaces, and 128 TDM
3. Accelerated FFT butterfly channel support
computation through a multiply
with add and subtract The VSS-AP algorithm was implemented
DMA Controller supports: and compared with the LMS algorithm under the
1. 14 zero-overhead DMA following circumstances: the input signal was
channels for transfers between chosen as a colored signal (that is, filtering a white,
ADSP-21161N internal memory zero-mean, Gaussian random sequence through a
and external memory, external second order IIR filter), K=4, M=17, = 0,995, C
peripherals, host processor, = 10-4, max = 1.0 .
serial ports, link ports or Serial
162
The main results, illustrated in figure 2, reasonable. Using an adequate technique, such as
show a significantly performance improved for the switching buffers, a real time implementation is
VSS-AP ( both for convergence speed and realizable.
misadjustment). The execution time is quite
0.4
0.4
0.3
0.2 0.2
0.1
0
0
-0.2 -0.1
-0.2
-0.4
-0.3
-0.6
-0.4
-0.8 -0.5
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
0.6 0.4
0.3
0.4
0.2
0.2
0.1
0
0
-0.2
-0.1
-0.4 -0.2
-0.6 -0.3
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
163
residual error - VSS (K =8) residual error - VSS (K =10)
0.5 0.6
0.4 0.5
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
-0.1
-0.1
-0.2 -0.2
-0.3 -0.3
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Figure 3. The VSS adaptive filter performance for various values of parameter K
180000
Number of processor cycles
152393
160000
140000
120000
100000 87419
80000
60000
40000 17447
20000 1382 4289
0
1 2 4 8 10
K
The VSS-AP implementation takes an This algorithm woks very well for ill-
execution time between and 0.2 ms and 16 ms conditioned input signal (e.g. speech signal).
witch is a quite reasonable execution time if
the switching buffer technique is involved. V. REFERENCES
This techniques requires as a timing condition
that the acquisition time for the whole signal
window (about 20 ms) must be greater than [1] Shin, H.C, et. al, Variable Step Size NLMS
total execution time for signal processing in and Affine Projection Algorithms , IEEE Signal
the current window. Processing Letters, vol.11, no. 2, february 2004,
pp. 132-135
[2] S. Haykin, Adaptive Filter Theory, 3rd ed.,
IV. CONCLUSIONS Prentice Hall, 1996
[3] ADSP2116x Hardware Reference Manual,
A new adaptive algorithm was presented. Analog Devices 2000
The performance of the new algorithm is better [3] ADSP2116x Software Manual, Analog Devices
than classical algorithm (similar with RLS 2000
algorithm) but the computational effort is
smaller than RLS.
164
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
I. INTRODUCTION
II. AN OVERVIEW OF SC140 ARCHITECTURE
It has been shown in [2], [3] that QRD-LSL algorithm
has great performances when used in echo canceller As we have chosen to implement the algorithm on
configuration. However, in the original version, the SC140, it is necessary to describe the features of this
complexity is quite large, making the algorithm processor first. The specific features of this
impossible to be implemented in real-time. In [1], a architecture, described in [6] are the following:
modified version of the algorithm is presented. A
comparison between the two algorithms is presented
High level abstraction of the Application
Software
in Table 1:
Applications development in C language
Hardware supported integer and
Table 1.
fractional data
Adaptive
algorithm QRD-LSL MQRD-LSL Scalable performance
multiplications 25M+11 22M+10 4 ALUs (Arithmetic logic Units) and 2
AAUs (Address Arithmetic Units)
divisions 4M+2 4M+2 4 MMACS (million multiply and
additions/ subtractions 8M+3 8M+3 accumulate operations per second) for each
square-root operations 4M+2 0 megahertz of clock frequency
High Code Density for Minimized Cost
It is to be noticed that the square root operations 16-bit wide instruction encoding
require a large computing time, as they must be
approximated by another technique, e. g. Taylor The core important features are:
series. Either way, the computing time increases
significantly and the application area becomes
Up to 10 RISC MIPS for each megahertz
of clock frequency
restricted due to a reduced sampling frequency. Thus,
the main goal is to minimize the number of A true (16*16) + 4040-bit MAC unit
instructions within the implemented algorithm. in each ALU
1
Facultatea de Electronic i Telecomunicaii, Catedra de
Telecomunicaii Bd. I. Maniu 1-3, Bucureti, e-mail: {aenescu, pale, silviu}@comm.pub.ro
165
A true 40-bit parallel barrel shifter in However, the main feature that we have already
mentioned is the C compiler and the ability to convert
each ALU
C source code into assembly code. The complexity of
16 x 40-bit data registers for fractional QRD-LSL algorithm is quite large and therefore the
and integer data operand storage need for flexibility is important, since programming in
16 x 32-bit address registers (8 can be C code is much easier than implementing the
used as 32-bit base address registers) algorithm direct in assembly code. The C compiler
supports ANSI C standard and also intrinsic functions
4 address offset registers and 4 modulo for ITU/ETSI primitives. Assembly code integration is
address registers also possible, which optimizes supplementary the
Unified data and program memory space code.
(Harvard architecture) The block diagram of SC140 core, as presented in [2],
is described in Fig. 1.
Byte addressable data memory
166
f , m (n) = f , m 1 (n) b , m 1 (n 1) *f , m 1 (n 1) At the price of losing from finite precision, some of
the bits used in quantization can be used for
scaling, regarded in binary arithmetic as a simply
f ,m1 (n) = cb,m1 (n 1) f , m1 (n 1) + sb,m1 (n 1) f , m1 (nright-shifting
* *
) by the same number of bits.
m (n 1) = cb, m 1 (n 1) m 1 (n 1) 450
2
Fm 1 (n) = Fm 1 (n 1) + f , m 1 ( n 1) f , m 1 ( n) 400
350
Fm 1 (n 1) 300
c f , m 1 (n) =
Fm 1 ( n) 250
s f , m 1 (n) = f , m 1 (n 1) 150
Fm 1 (n)
100
b , m (n) = b , m 1 (n 1) f , m 1 (n) b*, m 1 (n 1)
50
167
1 [ J m 1 , m , m , em , m 1 , pm 1 ] =
lm (6)
1 = prediction( J m 1 , m 1 , m 1 , m 1 ,
em 1 , m 1 , pm 1 )
As shown in [3], a small residual error in an echo
2
cancelling configuration is achieved with a RLS aux = J m 1 + m 1 m 1 2 b
adaptive algorithm by setting the forgetting factor
as closer to 1 as possible. All the same, if we set J m 1
c m 1 =
to the maximum possible represented number on aux
short format of 16 bits, i.e. 1-2-15, then lm is limited J m 1 = aux
by 215. The number of bits used for scaling should m = c m 1 m 1
be log2 (lm)=15, which is unacceptable, since it is
exactly the precision used for a short format m 1
s m 1 = 2 b
variable. Thus, a trade-off is required between echo J m 1
cancellers theoretical performances and the
precision used for cost functions. m = m 1 m 1 m 1
In order to test echo cancellers performances, the m = cm 1 m 1 + s m 1 m 1
algorithm has been implemented using Code if flag
Warror C Compiler for SC140. A simulation on the
evaluation board was run, with a sinusoid of a em = em 1 m 1 p m 1
normalized frequency 0.05 as near-end signal and a p m 1 = c m 1 p m 1 + s m 1 em 1
scaling of 10 bits. The signals are described in Fig. end
3:
Fig. 3. Far-end signal. Near-end signal. Output Fig. 5. Block diagram for one cell prediction
signal. Residual error.
The filtering part is included in backward
prediction part and is performed if a flag is set.
IV. OPTIMIZING TECHNIQUES USED FOR This flag is set before the backward prediction and
QRD-LSL reset before the forward prediction. Then, iteration
is described in Table 4.
In this paragraph, we evaluate the number of cycles
needed by the algorithm per iteration. The goal is Table 4.
to minimize this number, in order to lower the
computational time per iteration under the = 1
sampling time of the CODEC. for m = 0 , M
If we take advantage of the fact that the structure is flag = 1; // backward prediction
symmetrical, because of the similarities between
the forward prediction structure and the backward [ Bm 1 , b , m , f , m , em , f , m 1 , pm 1 ] =
prediction structure, then we can use two identical = prediction( Bm 1 , b , m 1 , m 1 , f , m 1 ,
blocks for each lattice cell, thus we can call twice a
function in C language during one iteration. em 1 , f , m 1 , pm 1 )
A behavioral description of the block is given in flag = 0 ; // forward prediction
Table 3. [ Fm 1 , f , m , b , m , em , b, m 1 , pm 1 ] =
Table 3 = prediction( Fm 1 , f , m 1 , f , m 1 , m 1 ,
em 1 , b , m 1 , pm 1 )
= b
168
end
REFERENCES
Another optimization technique, accomplished
[1] C. Paleologu, S. Ciochina, A. A. Enescu, Modified
using this procedure is that all the transformations versions of QRD-LSL Adaptive Algorithm with Lower
are made in-place, regardless of the iteration (i.e. Computational Complexity, Rev. Roum. Sci. Techn.
moment of time), saving a large amount of Electrotechn. et Energ., vol. 46, no.3, 2001.
memory. Choosing an appropriate level of [2] C. Paleologu, S. Ciochin, A.A. Enescu, A Network Echo
Canceller Based on a SRF QRD-LSL Adaptive Algorithm
optimization from the C compiler, Code Warrior Implemented on Motorola StarCore SC140 DSP, IEEE
(0-3), makes further optimization. As well, the Int. Conf. ICT 2004, Fortaleza, Brasil, 2004
proper use of intrinsic functions from C compiler [3] S. Ciochina, C. Paleologu, On the Performances of
can further reduce the number of cycles. QRD-LSL Adaptive Algorithm in Echo Cancelling
Configuration, Proc. IEEE ICT 2001, Bucharest,
A very good approximation on computing time per Romania, vol.1, 2001, pp. 563-567.
iteration shows that it is proportional to adaptive [4] S. Haykin, Adaptive Filter Theory, Third Edition, Prentice
filters order: Hall International, Inc. Englewood Cliffs, 1996.
tc M (7) [5] C. Paleologu, S. Ciochina, A. Enescu, A Simplified
QRD-LSL Algorithm in Echo Cancelling Configuration,
We shall refer to from now on as proportionality Proc. IEEE ICT 2002, Beijing, China, vol.1, 2001, pp.
constant. During implementation on StarCore, the 563-567.
[6] SC140 DSP Core Reference Manual, Revised 1, 6/2000
evolution of this constant was most relevant and it
[7] St. Gay, An Efficient, Fast Converging Adaptive Filter
is described in Fig. 6. for Network Echo Cancellation, Proc. Asilomar, Pacific
Grove, CA, Nov. 1998, pp 394-398.
[8] Ph. Regalia, Numerical Stability Properties of a QR-
The evolution of proportionality constant
Based Fast Least Squares Algorithm, IEEE Trans. Signal
1000 Processing, vol. 41, no. 6, June 1993, pp 2096-2109.
900 [9] M. Hartenek, R. W. Stewart, J. G. McWhirter, I.K.
800 Proudler, Algorithmic Engineering Applied to the QR-
700
RLS Adaptive Algorithms, Proc. 4th International
600
Conference on Math. Signal Proc., Warwick, U.K. 1996.
500
[10] Regalia P., Bellanger G. On the Duality Between Fast QR
400
Methods and Lattice Methods in Least Squares Adaptive
300
200
Filtering, IEEE Trans. Signal Processing, vol. 39, no. 8,
100
April 1991, pp. 879-891.
0
[11] Ciochin S., Negrescu C., Adaptive Systems, Ed. Tehnic,
1 2 3 4 5 6 7 8 9 Bucharest, 1999.
Steps
[12] Liu J. "A Novel Adaptation Scheme in the NLMS
Fig. 6. Evolution of proportionality constant Algorithm for Echo Cancellation", IEEE Signal Processing
Letters, vol. 8, no. 1, January 2001, pp. 20-22.
[13] J.G. Proakis, C. M. Rader, F. Ling, C. L. Nikias, Advanced
In Fig. 6, on X axis, optimization steps are Digital Signal Processing Algorithms, Macmillan
represented in time. The optimization process also Publishing Company, 1992.
included taking advantages of the parallel [14] W.M. Gentleman, Least Squares Computations by Givens
Transformations without Square-Roots, J. Inst. Math. Its
architecture and can be found in [2]. Appl., vol. 12, 1973, pp. 329-336.
We describe the stages from Fig. 2: [15] ITU-T Recommendation G.168, Digital Network Echo
1-3. Intrinsec optimizations from procedure Cancellers, 2000, Draft 3.
prediction (use auxiliary value for J, rearrange the [16] ITU-T Recommendation G.711, pulse code modulation
(PCM) of voice frequencies, CCITT-Blue Book, Volume
cos factor computation) III, Fasc. III. 4, pp. 175-184.
4. Modularization by using the procedure [17] A. Andronache, C. Anghel, S. Pop, A Novel Adaptation
prediction Scheme in the NLMS Algorithm for Digital Network
Echo Canceller Implemented on Motorola StarCore
5-6. Levels 0-1 of optimization
SC140, Int. Conference COMM 2002, Dec. 2002,
7. Use in-place transformation Bucharest, Romania
8-9. Levels 2-3 of optimization
V. REMARKS
169
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
I. INTRODUCTION
1
Faculty of Electronics, Telecommunications and Information Technology - University Politehnica of Bucharest
1-3 Iuliu Maniu Bd., Bucharest, Romania
e-mail: danny_silion@yahoo.com, dpanaitopol@yahoo.com, sorin252000@yahoo.com
170
of singletalk is desirable that d (n) y (n) = 0 and in from below by the performance of the NLMS,
case of doubletalk d ( n) y ( n) = z (n) 0 . whereas it converges in environments where NLMS
diverges. The GNGD is shown to be robust to
significant variations of initial values of its
A. LMS (Least mean square algorithm) parameters. Simulations in the prediction setting
support the analysis.
The least mean square algorithm is a simple, yet most The proposed GNGD algorithm is described by:
frequently used, algorithm for adaptive finite-impulse
response (FIR) filters. It is described by the following:
y (n) = x T (n) w(n) (5)
y (n) = x (n) w(n)
T
(1)
e( n) = d ( n) y ( n) (6)
e( n ) = d ( n ) y ( n ) (2)
w(n + 1) = w(n) + (n)e(n) x(n) (7)
w( n + 1) = w( n) + e( n) x( n) (3)
( n) = (8)
x ( n) 2 + ( n)
2
The parameter is the step size (learning rate) that
defines how fast the algorithm is converging.
Ideally, we want an algorithm for which the speed of e(n)e(n 1) x T (n) x(n 1)
convergence is fast and the steady-state (n) = (n 1) (9)
( x(n 1) 2 + ( n 1)) 2
2
B. NLMS (normalized LMS) The step-size governs the rate of convergence and
the steadystate excess mean-square error. To meet the
The step size of NLMS was found to conflicting requirements of fast convergence and low
be (n) = x(n) 2 , 0 < < 2 , where || . || 2 denotes misadjusment, the step-size needs to be controlled. In
2
VSS-LMS:
In theory, value = 1 provides the fastest (n) = (n 1) + e 2 (n) (10)
convergence, whereas in practice, the step size of the
NLMS algorithm needs to be considerably smaller. RVS-LMS:
To preserve stability for close-to-zero input vectors, (n) = (n 1) + p 2 (n) (11)
the optimal NLMS learning rate is usually modified
as x(n)
2
( x(n) 2 + ) , where is a small
2
p(n) = p( n 1) + (1 )e(n)e(n 1) (12)
2
positive constant.
VS-NLMS:
C. GNGD (generalized normalized gradient descent x2 (n)
algorithm) ( n) = ( n) (13)
e2 (n) +
The GNGD represents an extension of the normalized
least mean square (NLMS) algorithm by means of an x2 (n) = 1 x2 (n 1) + (1 1 ) x 2 (n) (14)
additional gradient adaptive term in the denominator
of the learning rate of NLMS. This way, GNGD e2 (n) = 2 e2 (n 1) + (1 2 )e 2 (n) (15)
adapts its learning rate according to the dynamics of
the input signal, with the additional adaptive term
e 2 ( n)
compensating for the simplifications in the derivation (n) = (n 1) + (1 ) (16)
of NLMS. The performance of GNGD is bounded x ( n) +
2
171
To preserve stability for close-to-zero input vectors,
two very small positive constants and were
introduced as a necessity in the experiments.
Table 1
As it is shown in figures presented below, some In Fig. 6 are presented the test signal (CSS_ST), the
algorithms are proved to have a lower speed of echo signal and the error signal (the minimized echo)
convergence and others a lower attenuation of the for GNGD algorithm. For single-talk case, z (n) = 0 ,
echo, than GNGD. so in this simulation, CSS_DT is not used. In this
case, d ( n) = ec( n) .
172
e(n)e(n 1) xT (n) x(n 1)
(n) = (n 1) (9)
( N x2 (n 1) + (n 1)) 2
V. IMPLEMENTATION ON DSP
VI. CONCLUSION
173
been shown to be robust to the initialization of its REFERENCES
parameters.
[1] ITU-T Recommendation G.168, Digital Network Echo
Cancellers, 2000.
[2] ITU-T Recommendation P.56, Objective measurements of
VII. ACKNOWLEDGMENT active speech level.
[3] Danilo P. Mandic, A Generalized Normalized Gradient
The authors wish to thank Freescale Semiconductor - Descent Algorithm, IEEE Signal Processing Letters, Vol. 11,
No. 2, February 2004.
Romania for their support to accomplish this study. [4] Hyun-Chool Shin, Ali H. Sayed, Woo-Jin Song, Variable
Implementations of algorithms presented above were Step-Size NLMS and Affine Projection Algorithms, IEEE
performed using software and equipment donated by Signal Processing Letters, Vol. 11, No. 2, February 2004.
Freescale Semiconductor - Motorola Romania to our [5] Silviu Ciochina, Cristian Negrescu, Adaptive systems,
Editura Tehnica, Bucharest, 1999.
laboratory.
[6] Jianfeng Liu, A Novel Adaptation Scheme in the NLMS
Algorithm for Echo Cancellation, IEEE Signal Processing
Letters, Vol. 8, No. 1, Jan 2001.
[7] Tomas Gnsler, Steven L. Gay, M. Mohan Sondhi, Jacob
Benesty, Double-Talk Robust Fast Converging Algorithms for
Network Echo Cancellation, IEEE Transaction on Speech and
Audio Processing, Vol. 8, No. 6, November 2000.
[8] SC140 DSP Core Reference Manual, Revision 3, November
2001.
[9] SC100C Compiler Users Manual, Revision 2, November
2001.
[10] StarCore SC140 Application Development Tutorial, Revision
0, March 2003.
174
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Politehnica University of Timisoara, Bv. V. Parvan 2, 300223 Timisoara, Romania
marius.otesteanu@etc.utt.ro
175
The forward and backward propulsion control signals evacuation. Unfortunately, such devices offer poor
and the left and right steerage control signals are positioning precision at high costs.
obtained from two joysticks, which can be either Another approach was oriented on the possibility to
digital, or proportional resulting either digital or modify the air circuits by introducing some variable
analog signals. The type of signals used, digital or volume elements. This would result in reducing the
analog, is determined by the particular application. speed and therefore the mechanisms inertia in the
positioning zone. By these means, the precision may
III. PNEUMATIC SYSTEMS BEHAVIOUR be substantially improved, but the main disadvantage
is the necessity of a control equipment which should
The pneumatic motors used in transporters are decide the moment and time period when the
characterized by inertia. Compared to precise electric admission volume is modified. A possible solution
motors, pneumatic motors, because of the inherent air would be the use of proportional electrovalves, also
compression, have a slow response to impulse known as boosters, but this solution implies very high
commands. Thats why digital control can only be costs, and is only used in systems with very high
used in applications where small trajectory or performance requirements.
positioning errors are accepted. For higher accuracy The third, and last approach, was considered the most
driving, the proportional control is recommended. adequate for the solution to be implemented. This
Precision driving, in forward backward propulsion, approach tries to find a compromise between the
means quick start, relative high speed motion and stop problems appeared in the previous approaches, by
at a definite marker. Usually, positioning errors of 5 controlling the air pressure, and the air admission in
10 mm are obtained with pneumatically driven the equipment, obtaining a quasi-variable pressure,
wheels. which can be controlled between some limits, enough
For steerage, the propulsion wheel is left right to obtain steerage angular errors as low as 0.25.
driven with a second pneumatic engine. Precision
steering means forward motion (0), cross motion IV. NECESSITY OF FEEDBACK CONTROL
(90) or any definite angle. Usually, angular errors of
3 5 are obtained. The traditional on / off driving pneumatic equipments,
With the traditional on / off driving pneumatic even controlled by electronic command modules,
engines, because of the air compression, higher following continuously the parameter indicating the
accuracy cannot be obtained. The inertia results very current position of the mechanism (figure 2) have
high, especially with hundreds of tones charge. great inertia and the positioning error is also great. In
Alternatively, PWM driving is not proper for moving this case the system may even oscillate, because it is
mechanical devices, as electrovalves. not able to stop within the required error window.
Because the pneumatic mechanisms used depend on Electronic Pneumatic
control Command driven
the pressure of the compressed air, on the compressed system mechanism
airs behavior in the pneumatic circuits and, not less
meaningful, on the control mode of the admission and Control value
evacuation of the air, all those factors have to be taken
into account in order to develop a proper control Fig. 2 Pneumatic equipment driven by electronic control system.
solution. In order to stop the mechanism with a very small
The first approach was oriented on the pressure of the error, close to the target position, it is necessary to
air inserted in the pneumatic circuits. It was reduce the movement speed of the mechanism
determined that for high pressures of 5 6 Bar, the dynamically and almost linearly, by controlling the air
mechanism has a very good behavior, excepting the pressure in the pneumatic circuit to a value which
zone close to the target position, where the movement ensures the equivalent of a mechanical braking until
continues with the same high speed. When the target the target position. Figure 3 emphasis the possibility
position is assumed reached and the decision to stop of instability when high pressure (high force and
the air admission is taken, due to the high air pressure, rotation speed) is used, associated with narrow error
a very long relaxation process appears until reaching window (for high accuracy), e/2.
the normal atmospheric pressure. During this
relaxation process, the mechanism continues to move Position
after the target position is reached, producing a high 5-6 bar 2-3 bar 0-0.1 bar
176
V. CONTROL ALGORITHM This solution offers a good compromise between
positioning precision and equipment costs, the
Because there are no pressure sensors to allow the pneumatic and mechanical equipment being the same,
electronic system to monitor this parameter, the only just the proposed electronic control module having to
way to get information from the pneumatic- be added.
mechanical equipment is through angular position
transducers. These transducers provide information on VI. HARDWARE IMPLEMENTATION
the mechanical position of the pneumatic driven
equipment, by means of electrical signals. Using this Considering the possibility to integrate on a low cost
data, an intelligent electronic system is able to single chip multiple modules as processor, memory,
calculate the distance between the current position A/D and D/A converters, counters, comparators,
and the target position and to determine the necessary PWM modules and so on, a microcontroller is
time and position for the appropriate braking, in order indicated for the implementation of the solution.
to stop the rotation on the target position. The Position
10-bit
electronic system has to modify the air pressure in the Transducers
ADC
pneumatic circuit, according to the computed values.
The solution to implement the control above is to Pneumatic
command the electrovalves, which control the air driven engine Microcontroller
admission in the pneumatic circuit, by variable width
Electrovalves
electrical pulses, as represented in figure 4. Turning speed I/O I/O
Position P P
Direction
O O
p1 R R
Propulsion
PWM T T
p2
control
p3
+e/2 User interface
target - position command
- parameters setting
Control
signal
Time
Fig. 5 Microcontroller based electronic control system.
177
control of the electrovalves, the minimum pulse width then all the 8 time slots will be active. The
was set at 15 ms, and the working frequency at 25 Hz information provided by the position transducer is
(i.e. 40 ms period). Further practical tests proved that sampled each 5 ms and if it is assumed that the
only when the pulse width changes with at least 5 ms, mechanism reached the required position, the control
a visible effect on the movement speed can be signal is disabled, whether the PWM cycle is
observed. completed or not.
Starting from the above considerations, an intelligent The algorithm relies on the principle of minimizing
control algorithm was implemented to control the the error between the reference position (target
admission electovalves. The feedback is based on the position) and the current position of the mechanism
information obtained from the transducer, placed on (indicated by the transducer). The pulse width is
the rotating mechanism. The signal provided at the modified as the mechanism gets closer to the target
control output has a frequency of 25 Hz and a width, position. The change of the pulse width depends also
adjustable in steps of 5 ms, according to the on some adjustable parameters, which can be set by
movement speed and the distance to the target an operator:
position. By decreasing the pulse width, while the braking shape,
position of the mechanism gets closer to the target braking duration,
position, the movement speed is decreased and, at the error window.
target position, the mechanism can be stopped within For large error window, the movement is stable, but
the required precision window of 0.25. The main the final error might be too large. For narrow error
purpose of the braking is to minimize the mechanical window, if the braking system is not precise enough,
inertia at the moment, when the air admission in the oscillations occur, which can become dangerous for
pneumatic installation is stopped. charges of tones. The set of parameters has to be
The software implementation of the PWM algorithm carefully chosen from a wide range of possibilities
is running on the microcontroller in real-time, because (with 3 variables) in order to achieve the best
all the signals have a slow time variation. Therefore a performance.
5 ms period was chosen for the systems clock These parameters must take into account the working
(diagram in figure 6), enough to allow the conditions and the external factors which may
microcontroller to perform all the computations influence the working parameters of the pneumatic
required by the control algorithm for up to 3 control equipment. Various working modes may be chosen,
loops and other computations for processing from abrupt braking shape in a small time window
additional external information. with up to 2 steps for the pulse width, to slow braking
with 5 steps for the pulse width, involving a longer
braking time, as represented in figure 7.
NO 100%
5ms timer 87.5%
YES
STOP
Decide pulse
DAC PWM
width signal
Sample position
Ts 5 ms 40 ms Tmin 15 ms
Tpwm
Legend:
Ts sampling period
Compute Estimate remaining Timer
... ... Tpwm PWM signal period
interrupts Tmin minimum pulse width
current error distance
Fig. 7 Pulse widths for 5-steps braking.
178
VIII. CONCLUSIONS
ACKNOLEDGMENT
REFERENCES
179
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Facultatea de Electronic i Telecomunicaii, Departamentul
Electronic Aplicat, Bd. V. Prvan Nr. 2, 300223 Timioara, e-mail {aurel.gontean; mircea.babaita; roxana.jibleanu}@etc.utt.ro
180
Although evolved, modern IDE software is not A software PWM generator is almost impossible to be
hardware oriented it is impossible to simulate evaluated under MPLAB control. Once built, the hex
continuous voltage values, currents, and transistors file is loaded in the microcontroller (Fig. 2) and the
not to mention LEDs, LCDs, buzzers or motors. There dynamical behavior may be analyzed with virtual
are areas not covered at all by the IDE software instruments (Fig. 3).
RS232, I2C, or SPI communication, or difficult to use The combination of the two software packages allows
PWM, soft delays. Therefore there is a high demand the user to design most or even all of the schematics
for an appropriate simulation tool in both design and and to edit, debug and verify the software with little
teaching embedded systems. or even no need for prototyping, shortening the
overall design time.
II. PROTEUS SIMULATOR
A. Hardware interaction
Proteus software offered by Labcenter Electronics is a
solution allowing for mixed analog and digital Beyond pure software simulation, Proteus offers also
simulation, along with models for Microchip hardware interaction. A microcontroller model with a
midrange microcontrollers. digital sounder may work together in order to generate
audible sounds. Fig. 4 demonstrates a typical case
where Proteus VSM is running a simulation of a PIC
program which generates audio tones in real time. The
sounder model picks up the transitions on port A,
RA3, and converts them into a 44 kHz data stream
which is sent to the sound card. On a Pentium II or
better PC, the simulated PIC will run fast enough to
generate audio tones in real time.
B. Using displays
181
Another useful feature is the LCD simulation for the
standard LM032 display (fig.6). The simulation
supports both 4 bit and 8 bit modes, while busy flag
can be read on D7, exactly like in the normal use.
C. Serial transmission
III. LIMITATIONS AND ERRORS programmable warnings for stack under -or overflow,
and incorrect jumping computation.
MPLAB IDE is a powerful tool which is continuously
improved by Microchip. Some peripherals are not
supported at all (such as the USART), and others are
only partial simulated (the ADC conversion does not
change the ADRESH and ADRESL registers). There
are differences in timing Timer1 has a constant 4 s
error compared with the real behavior.
Proteus simulator has also some errors, for example
the 74LS148 model is wrong the I 0 input does not
trigger the EO nor the GS lines, as it should (Fig.
8). Another drawback is that only single file code can
be source code debugged there is no possibility to
link files within Proteus. However difficult items such
Fig. 8. There is no reaction for the I0 input line stimulus.
as banking are accurate simulated there are
182
IV. COMBINED ACTION and to edit, debug and verify the software with little
or even no need for prototyping, shortening the
For educational and design purposes, the best result is overall design time.
accomplished working using both packages. Changes
For teaching purposes, the overall performance is
in MPLAB source code are reflected in the .hex file
excellent, the learning curve being dramatically
and loaded in background in the Proteus model each
reduced.
time a new simulation is started (there is no need to
restart the simulator). Working with an ICSP
programmer is a low cost debug solution, suitable for REFERENCES
students; while an ICD2 debugger is a professional
approach at a moderate price, with the advantage of [1] http://www.microchip.com.
[2] http://www.labcenter.com.
stepping through the code, breakpoints, real timing, [3] * * * MPLAB IDE Simulator, Editor. Users guide,
the Proteus simulation may replace it for most cases. Microchip, 2001.
[4] * * * MPLAB ICD2Debugger. Users guide, Microchip,
V. CONCLUSIONS 2002.
[5] A. Gontean, The PIC16F84A Microcontroller, Editura
Orizonturi Universitare, 2004 (in Romanian).
The combination of the two software packages allows
the user to design most or even all of the schematics
183
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
The board is compatible with both Microchips hardware characteristics of the system (usage of shift
MPLAB IDE and JDM programmer, thus enabling the registers for input and output, multiplexing digits,
student with means for editing, simulating, building and external bus line) and also advanced
programming and testing his own routines. Being software topics (interrupt driven RS232 routines,
FLASH based, the PIC16F84A microcontroller may LCD interfacing, SPI communication with ADC).
be erased and program without removal from the All the hardware and software involved with this
socket via the JDM programmer and ICSP port. The board was developed in Timisoara, at the Electronics
board comes with specific demos highlighting the and Telecommunication Faculty.
1
Facultatea de Electronic i Telecomunicaii, Departamentul
Electronic Aplicat, Bd. V. Prvan Nr. 2, 300223 Timioara, e-mail {mircea.babaita; aurel.gontean; roxana.jibleanu}@etc.utt.ro
184
II. BOARD DESCRIPTION
III. RESULTS
185
displaying with the multiplexed digits (fig. 5), SPI has limited peripheral capabilities, with the extensions
interfacing for the shift registers and ADC, complete provided buy the board the student or the designer
RS232 communication software. may experiment, debug and verify specific projects.
REFERENCES
IV. CONCLUSIONS
[1] * * * MCP3204/MCP3208. 2.7V 4-Channel/8-Channel 12-Bit
A/D Converters with SPI, Microchip, 2001.
The novel approach in this development board is the [2] A. Gontean, The PIC16F84A Microcontroller, Editura
possibility of in circuit programming the Orizonturi Universitare, 2004 (in Romanian).
microcontroller, using the ultra-low cost JDM [3] * * *, EPIC-1 User Guide, Mikroelektronika, 2001.
[4] http://www.microchip.com.
programmer. Various displays possibilities and [5] http://www.mikroelektronika.co.yu.
flexible architecture for the input switches make the
debug process an easy task. Although the PIC16F84A
186
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
187
Fig. 1.
The microcontroller has separate command signals for the signal ALE, the lower part of the address bus is
accessing the program memory (/PSEN- addresses 64 demultiplexed by the data bus (fig.4).
Kbytes of program memory) and the data memory (/RD,
/WR also address 64 Kbytes of data memory). In order
to execute user programs stored in the data memory, the
program memory space and data memory space must be
joined.
Fig. 4.
188
Fig. 5.
189
contents of 16 memory bytes in hexadecimal. LP P1 .
DP P1 _ P2 . Loads from a personal computer through the serial
DD P1 _ P2 . interface, the user program listing file (max. 26-28 Kb)
DE P1 _ P2 . into the external RAM memory starting from address
P1. The machine code user program is extracted from
Displays on the console the contents of a program
memory area (DP command), external data memory area this file.
(DD command), and internal EEPROM data memory MD P1 _ P2 _ P3 .
area, respectively (DE command) from address P1 until ME P1 _ P2 _ P3 .
address P2. MI P1 _ P2 _ P3 .
DP command displays any internal program memory Transfers the contents of the external data memory area
area within the memory space between 0000 1FFFH (MD command), internal EEPROM data memory area
and any external program memory area within the (ME command), and internal data memory area,
memory space between 2000H FFFFH, if /EA= + 5V, respectively (MI command) from address P1 to address
or even any external program memory area whatsoever P2 into the memory area starting at address P3 .
if /EA=Gnd.
N .
DE command validates the access to the internal
Clears the console display.
EEPROM data memory, then displays the specified area
P .
and, in the end, re-blocks the access to this area.
Authorizes the password access of the user to the SFR
DS .
peripheral registers.
Displays on the console the internal memory zone of the
R P1 _ P2 .
microcontroller, addressable directly between the
addresses 80H and FFH; this memory zone consists of Receives a hexadecimal data block through the serial
the special function registers (SFR) area. interface, from a computer or another development
DT P1 _ P2 . system and stores it in the external RAM data memory
from address P1 to address P2 .
Displays on the console the characters corresponding to
ASCII codes stored in the external data memory from SD P1 _ ... .
address P1 to address P2. Displays and/or substitutes the contents of an external
The console display can be suspended if key A is data memory area that starts at address P1. The
pressed, resumed if key C is pressed, or aborted if key command displays every byte of any external data
O is pressed. memory area, while the substitution is performed only in
E P1 _ P2 _ P3 _... . RAM external data memory locations.
Extracts break-points in the user program from For every substitution commands, the use of BLANK
addresses P1, P2, P3, ... separator displays and/or substitutes the contents of the
following memory location, while the use of COMMA
FD P1 _ P2 _ P3 .
separator displays and/or substitutes the contents of the
FE P1 _ P2 _ P3 . previous memory location. The command is closed with
FI P1 _ P2 _ P3 . ( . ) terminator.
Fills the external RAM data memory (FD command), SE P1 _ ... .
the internal EEPROM data memory (FE command) and Displays and/or substitutes the contents of an internal
the internal RAM data memory, respectively (FI microcontroller EEPROM data memory area that starts
command) from address P1 to P2 with P3 data value. at address P1. When the command is issued, the access
G P1 _ P2 _ P3 _... . to the internal EEPROM data memory is validated, then
Determines the execution of a user program from the memory locations are displayed and/or substituted,
address P1 until one of the addresses of the break-points and in the end, the access to this memory is blocked.
SI P1 _ ... .
P2, P3, ... . After executing the user program segment,
Displays and/or substitutes the contents of an internal
the break-points assigned by this command are removed.
microcontroller RAM data memory area that starts at
I P1 _ P2 _ P3 ... .
address P1.
Introduces break-points in the user program at the
SP P1 _ ... .
addresses P1, P2, P3...
Displays and/or substitutes the contents of a program
LC .
memory area that starts at address P1. The command
Loads from a personal computer through the serial
interface, a machine code user program into the external displays every byte of any internal program memory
RAM program memory and runs the program. The area within the space between 0000H and 1FFFH and of
address used for loading the user program, the length of any external program memory area within the space
the program, the execution address and the break-point between 2000H and FFFFH if /EA= + 5V and,
addresses are transmitted using the header of the file that respectively, any external program memory area if
includes the user program. /EA=Gnd. The substitution is performed only in the
190
RAM external program memory locations. WMCON
SS P1 _ ... . EEMEN, EEMWE, DPS, RDY/BSY
Displays and/or substitutes the contents of an internal XF .
microcontroller data memory area destined to the special Displays the names and the contents of the flag register,
function registers (SFR) that starts at address P1. Certain then the names of the flags and their binary values.
PSW
microcontroller peripheral circuits provide password-
C, AC, F0, RS1, RS0, OV, PSW.1, P
protected access.
XI .
ST P1 _ ... .
Displays the names and the contents of the registers
Substitutes the contents of an external RAM data used for maskable interrupt system validation and for
memory area with the ASCII codes of the characters setting the priority level of an interrupt source, then the
issued by the user, starting from address P1. The names of the flags and their binary values.
command is closed when CTRL+P key is pressed. IE, IP
T P1 _ P2 . EA, ET2, ES, ET1, EX1, ET0, EX0, PT2, PS, PT1,
The command executes the user program from address PX1, PT0, PX0, IE1, IT1, IE0, IT0
P1 and performs P2 instructions. XP .
Displays the names and the contents of the parallel
W P1 _ P2 . input-output ports on byte and bit levels:
Transmits a hexadecimal data bloc from the external P0, P1, P2, P3
data memory from address P1 to address P2 towards a P0.0, P0.1, P0.2, P0.3, P0.4, P0.5, P0.6, P0.7
computer or another development system through a P1.0, P1.1, P1.2, P1.3, P1.4, P1.5, P1.6, P1.7
serial interface. P2.0, P2.1, P2.2, P2.3, P2.4, P2.5, P2.6, P2.7
X . P3.0, P3.1, P3.2, P3.3, P3.4, P3.5, P3.6, P3.7
Displays the names and the contents of the user registers XS .
of the AT89S8252 microcontroller. The systems Displays the names and the contents of the control,
console displays the following registers: status and data registers of the peripheral serial interface
R0, R1, R2, R3, R4, R5, R6, R7 (bank 0) SPI, then the names of the control flags and their binary
R0, R1, R2, R3, R4, R5, R6, R7 (bank 1) values.
R0, R1, R2, R3, R4, R5, R6, R7 (bank 2) SPCR, SPSR, SPDR
R0, R1, R2, R3, R4, R5, R6, R7 (bank 3) SPIE, SPE, DORD, MSTR, CPOL, CPMA, SPR1,
ACC, B, PSW, SP, DPTR0, DPTR1, PC SPR0, SPIF, WCOL
P0, P1, P2, P3 XT0 .
TCON, TMOD, TH0, TL0, TH1, TL1 Displays the names and the contents of the data and
T2CON, T2MOD, RCAP2H, RCAP2L, TH2, TL2 control registers of counter T0, then the names of the
WMCON control flags and their binary values.
IE, IP TCON, TMOD, TH0, TL0, T0
PCON GATE0, C/T0, M10, M00, TR0, TF0
SCON, SBUF XT1 .
SPCR, SPSR, SPDR Displays the names and the contents of the data and
XA . control registers of counter T1, then the names of the
Displays the names and the contents of the data and control flags and their binary values.
control registers of the serial asynchronous interface TCON, TMOD, TH1, TL1, T1
used for UART data transmission and reception, then GATE1, C/T1, M11, M01, TR1, TF1
the names of the control flags and their binary values. XT2 .
SCON, SBUF Displays the names and the contents of the data and
SM0, SM1, SM2, REN, TB8, RB8, TI, RI, SMOD control registers of counter T2, then the names of the
XB . control flags and their binary values.
Displays the names and the contents of the basic T2CON, T2MOD, RCAP2H, RCAP2L, TH2, TL2,
registers of the microcontroller. RCAP2, T2
ACC, B, PSW, SP, DPTR0, DPTR1, PC TF2, EXF2, RCLK, TCLK, EXEN2, TR2, C/T2,
XC . CP/RL2, T2OE, DCEN
Displays the name and the contents of the low power XTW .
consumption mode register, then the names of the Displays the name and the contents of the control
control flags and their binary values. register of the watchdog timer, then the names of the
PCON control flags and their binary values.
PD, IDL WMCON
XE . WDTEN, WDTRST, PS2, PS1, PS0
Displays the name and the contents of the command and Y .
control register for access to the internal data EPROM Determinates the software initialization of AT89S8252
memory, then the names of the control flags and their microcontroller.
binary values.
191
IV. CONCLUSIONS REFERENCES
192
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Freescale Semiconductor Romania, tirbei Vod 26-28, phone 3052449, e-mail: ENED001@freescale.com
2
Politehnica University of Bucharest Faculty of Electronics and Telecommunications, Iuliu Maniu 1-3,
e-mail: mihnea@comm.pub.ro, ionut_pirnog2001@yahoo.com, cristina_sucholotiuc@yahoo.co.uk
193
Fig. 1. Transmitter block diagram
In the modulator system (Fig. 1), the following III. IMPLEMENTATION ASPECTS
processes shall be applied to the data stream: transport
multiplex adaptation and randomization for energy A. Motorola MSC8101 DSP Performances
dispersal, outer coding (Reed-Solomon code), outer
interleaving (convolutional interleaving), inner coding The family of the MSC8101 processor implements a
(punctured convolutional code), inner interleaving, new model of instructions execution called VLES
mapping and modulation, Orthogonal Frequency (Variable Length Execution Set), which allows in
Division Multiplexing (OFDM) transmission. general the usage of more parallels addressing and
Since the system is designed for digital terrestrial computing units, during the same cycle.
television services to operate within the existing VHF The next lines present the most important features of
and UHF (Very and Ultra-High Frequency) spectrum the present processor as following:
allocation for analogue transmissions, it is required - up to 10 MIPS (Million Instructions Per Second)
that the system provides sufficient protection against for every Mhz frequency;
high levels of Co-Cannel Interference emanating from - 4 ALU (Arithmetic Logic Unit) which include
existing PAL/SECAM/NTSC services. dedicated circuits for addition, multiplication and
It is also a requirement that the system allows the bit operating units;
maximum spectrum efficiency when used within the - in every ALU there a presented MAC (Multiply
VHF and UHF bands; this can be achieved by using and ACcumulate) and shifter units.
Single Frequency Networks (SFN) operation. Concerning the registers, the MSC8101 processor
In fact, two modes of operation are defined in the presents next features:
OFDM technique with two options in the number of - 16 registers of 40 bits length used for integers and
carriers: a "2K mode" and an "8K mode". The "2K fractional operations;
mode" is suitable for single transmitter operation and - 16 register for addressing of 32 bits length, from
for small SFN networks with limited transmitter which 8 bits can be used for generating base
distances. The "8K mode" can be used both for single addresses in buffers;
transmitter operation and for small and large SFN - 4 offset registers for addressing and 4 registers for
networks. circular addressing.
As far as bandwidth requirements are concerned, the Some other features of this DSP are about the
preferred channel spacing is 8 MHz, but if desired, 7 presence of the orthogonal instruction set coded on 16
MHz or 6 MHz spacing is also possible by scaling bits; the possibility of executing up to 6 instructions in
down all system parameters. one cycle. Another important feature of the MSC8101
The error correction can be separated in two blocks: is that it has a CMOS logic which allows reduced
the outer coding and outer interleaving that are power consumption.
common to the Satellite and Cable Baseline
Specifications (DVB-S and DVB-C) and the inner B. Transmitter blocks implementation
coding is common to Satellite Baseline Specifications.
The use of inner interleaving is specific to the DVB-T DSP implementation consists on processing each
system. MPEG-2 transport packet of 188 bytes. Outer coding
To accommodate different transmission rates, in is formed by a RS encoder (204,188) followed by a 12
addition to five code rates, three types of non- stages interleaver. Inner coding is a convolutional
differential modulation schemes can be selected: code of rate . The output from convolutional
QPSK, 16-QAM and 64-QAM. The 16-QAM and 64- encoder can be optionally punctured. Accordingly, the
QAM can also be used in combination with uniform overall convolutional code rate is 1/2, 2/3, 3/4, 5/6 or
or non-uniform mapping rules and thus input data 7/8. Interleaving is performed here both bit-wise and
streams can be separated in a low and a high priority symbol-wise. Former is done depending on the
data stream with different error protection for symbol rate (bits/symbol), based on typical
hierarchical transmission purposes. These two permutation operators. The symbol-wise interleaver
bitstreams are mapped into the signal constellation by maps the bit words onto the active subcarriers. Every
the Mapper and Modulator. This feature allows the carrier is modulated by a modulation symbol. QPSK,
simultaneous broadcasting of different programmes 16-QAM and 64-QAM are used as modulation
with different error protection and coverage areas. methods, e.g. 2, 4 or 6 bits per modulation symbol.
194
The bits are assigned to the particular points in the
phase space according to the so called Gray-code
mapping. The advantage of this mapping is the fact
that closest constellation points differ only in one bit.
Each frame is formed by 68 OFDM symbols. Four
frames constitute one super-frame. Each symbol is
formed by 6817 samples in 8K mode and 1705
subcarriers in 2K mode. The symbol is formed by
applying a IFFT operator to the 8K/2K samples
resulted as the data samples, intercalated with pilot Fig. 2. Convolutional encoder
samples and zero subcarriers. A computational
efficient radix 4 IFFT algorithm was implemented to into account for the output buffer the maximum length
be used both in 2K or 8K modes in order to achieve which is about 1632 elements. This number of bits is
time requirements for a real-time processing. obtained from the next relation:
195
here this value was considerate lower than 1. This is
about the fact that the medium energy for 4 signals is:
2 2 2 2
1 + j + 1 j + 1 + j + 1 j 8
E{c c*} = = = 2 (2)
4 4
IV. PERFORMANCES
Regarding the performance of the implemented 8K: execution time 348 500 cycles out of 384 000
modulator it must be underlined first of all that there
was needed to be processed a great amount of data. Fig. 3. Cycles burned for the modulator 2K mode
That is way the cycles burned for performing all the (right) and 8K mode (left)
operation contained in the transmission chain have big
values for both operating mode of 2K or 8K. The target of the cycles count were achieved on the
With all of these, it may be said that this modulator one hand thanks to the MSC8101 capabilities and on
work in real time, meaning that it was possible to the other hand thanks to the optimization technique
obtain a number of cycles close to the limit but lower used in C such as: loop merging, split computation
that this. Next figure presents these results together and so on. Even if the implemented solution for the
with the cycle burned for each group of functions. DVB-T modulator, work in real time, due to the fact
In fig. 3 we denoted with F1...F6 next 6 group of that we are at the time limit the future development of
functions: this work will consists in developing the big cycles
F1 Iner Coding Function; burning module in assembling language.
F2 1st Interleaving and Convolutional Encode
F3 Bit Interleaving Function VI. ACKNOWLEDGMENT
F4 - Symbol Interleaving Function
F5 Symbol Mapping Function Implementations of algorithms presented above were
F6 Inverse Fast Fourier Transformer Function performed using development software and DSP
boards donated by Freescale Semiconductor Romania
V. CONCLUSIONS to our University.
196
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract This paper presents the design and standard. In this case, the functionality of the interface
development on a telephone interface based on M8880 is made by a master module, which sends commands
DTMF circuits and vocal memory ISD33120, controlled and receives data based on communication protocol.
with an 80C552 microcontroller. The interface is The schematic diagram of the proposed home
connected like slave in a RS-485 multipoint network that
use balanced lines. The master initializes interface, send
automation system is depicted in figure 1.
commands and read data and status information.
The voice generator circuit is used for two functions. RS RS 232/ RS 485
PC
First, after recognition of telephone call, send a vocal RS 485
Master 232
message indicating the valid operation. Second, for Convertor
transmission, after a initiated call, send to called station
the information like vocal messages. Phone Other
Keywords: telephone interface, DTMF, voice memory, interface
(slave)
slave
call progress, microcontroller.
1
Department of Automation and Industrial Informatics, University Politehnica of Timisoara Blvd.V. Parvan 2, 300223 Timisoara,
Romania, Phone: +40256-403245, E-mail: tionica@aut.utt.ro
2
Department of Communications, University Politehnica of Timisoara, Blvd. V. Parvan 2, 300223 Timisoara, Romania, Phone:
+40256-403310, E-mail: cbalint@etc.utt.ro
197
Multipoint bus TxD / RxD
RS 485 Microcontroller
Interface (80C522)
RS 485
DTMF Transceiver
M8880
Phone line
Hybrid
(2/4 wire) SPI
interface
Speech memory
ISD 33120
Speaker Microphone
198
control software. Master-slave connection is P2 enable transmission and number of
implemented according to fig. 1. retransmission
The master initializes the interface and send Enable transmission ( b7 )
DTMF codes which is transmitted in line. Also, Number of retransmission at no dial tone (0
master read DTMF codes received from telephone 7) (b4 b6)
line and decoded by M8880 circuit and status Number of retransmission at line busy (1
information indicating the progress of a call on 15) (b0 b3)
telephone network. P3 waiting times
The authors build a network structure using the waiting time at line busy (1 15sec.) (b4
Open System Interconnection of ISO Reference b7)
model (OSI-ISO model) [6] adapted for field bus waiting time at ring back tone (1 15sec.)
systems, which use only layers 1, 2 and 7. In this (b0 b3)
model, each layer implements a specified protocol and P4 number of digits for telephone call
provides services to the layer above and also uses P5 list of the digits for the number in ASCII format
services from the layer below. 2. Master read the code (DTMF tone) received by
1. Physical layer defines signal voltages and slave from the telephone line:
physical connections for sending bits across
a physical transmission media. We adopt RS CMD1 0
485 standard for multipoint systems that use
balanced lines.
2. Data link layer handles transmission of data 3. Master sends the code, which will be transmitted by
packets between station of the network, slave on the telephone line:
checking for errors, control internal data flow
and access control on channel. We CMD2 N Codes (N bytes)
implement this layer entirely in software.
3. The application layer contains the specific 4. Master read the status byte, which identifies the
functions for the functionality of each progress of an initiated telephone call:
module.
The interface can be connected, like a slave CMD3 0
module, to a serial multipoint bus using RS 485
standard. In this case, the functionality of the interface At the data link layer the software adds to previous
is made by a master module, which sends commands messages an address byte and calculates a CRC-16
and receives data based on communication protocol. sum for errors detection, building a packet will be
transmit on the bus:
B. Communication Protocol
ADR Previous Master commands CRC-16
Access to the bus is controlled by the master/slave
technique, which is the simple and therefore efficient The addressed slave (telephone interface) receives
bus protocol. the packet (at data link layer), evaluates the CRC-16
Communication is based on the principle where a field, and, if no errors are detected, sends the
master sends a request and the addressed slave returns extracted message to application layer.
an immediate response. Requests can be a read or a The application executes commands and prepares
write type. an answer for the master with the following structure:
The command send by the master at application
layer contain one command code, a number of DATA LENGTH DATA
parameters that follows and the parameters list. CMD STATUS
(0 N) (1 N)
199
2. Send remote commands in DTMF signaling
CMD1 STATUS
DATA LENGTH CODE (e.g. light on/off)
(= 1) RECEIVED Receive a vocal messages that informs about
3. state of home system
DATA LENGTH Records vocal message or play a previous
CMD2 STATUS
(= 0) stored messages (answer machine functions)
Stop the call and disconnect from line
4. To avoid unauthorized access to remote system,
DATA LENGTH the interface can accept commands only after the user
CMD3 STATUS STATE
(= 1) provides an access code (DTMF). In same sense was
implemented and tested a function (provided by the
where STATE can take the following value : public exchange) that identifies the calling number
EV_OK telephone call successfully finished (CLIP Call Line Identity Presentation) that allows
EV_NOT_ANSWER the called station not answer system to accept commands only from a list of
EV_TONE_NOT_OK no dial tone possible calling numbers.
EV_REV_CALL_BUSY busy tone
B. Remote Data or Message Transmission
C. CRC Generation
In case of a system event, the program
A multi-byte CRC error checking protocol should automatically sends a call to one pre-programmed
be used on all data transmissions between the master numbers from a list of numbers (e.g.: owner - mobile,
and slave nodes of an RS-485 communications police, fire brigade etc.).
system. After connected to line, the interface uses call
The sending system will calculate a CRC and progress information in order to complete the call. If
append it to the message. The receiving system will dial tone is present, the program dials the number and
calculate a new CRC based on the entire message, analyzes ring-back tone. If called station answers, the
including the appended CRC bytes. If the CRC information about produced event is transmitted like
calculated is not equal to zero, then an error occurred vocal message.
in the transmission and all data should be ignored.
The CRC-16, 16 bits polynomial used is V. CONCLUSION
X16+X15+X2+1, and will detect all single/double
errors, all errors with odd bits and all burst errors The phone line interface with network capabilities
shorten than 16 bits [6]. is a necessary connection for automated home system
with external networks. The user can send orders from
IV. FUNCTIONAL DESCRIPTION remote location using DTMF codes poll the state of
system or receive information about any event
The application program is written in C language produced in home using prerecorded vocal messages.
and tested using a development system with 80C552 An other way to communicate with external
microcontroller (DSM) [2]. networks is an Internet connection that is presented by
The DSM allows the user to develop application authors like a future improvement.
programs (in C or assembly language) with full access
to the resources of microcontroller and system. REFERENCES
At application level the program is implemented in
many independent modules. [1] K. Wacks, Home Systems Standards: Achievements and
For the new introduced voice memory functions Challenges, IEEE Communications Magazine, April 2002.
[2] T. Ionic, C. Balint, Telephone Interface for Home Automation
was implemented following library functions: Systems, Buletinul Stiintific al Universitatii Politehnica din
Rec_Mic recording a message from microphone, Timisoara, Periodica Politechnica, Transactions On Automatic
to a specified address. Control And Computer Science Vol.48 (62), 2003.
Rec_Line recording a message from phone line [3] G. Niculescu, L. Ioan, Tehnici si Sisteme de Comutatie, Ed.
Matrix Rom, Bucuresti, 2000.
to a specified address. [4] http://www.winbond-usa.com/products/isd_products/
Play_Mes play a recorded message from a chipcorder/datasheets/33120/33120.pdf
specified address. [5] *** M8880 Data Sheet, www.teltone.com.
[6] A. S. Tanenbaum, Retele de calculatoare, Ed. Byblos,
Bucuresti, 2003
A. Telephone Call Analysis and Data Reception
200
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
AEM Luxten Lighting Company SA, Gas and Water Meter Research Department, 26 Calea Buziasului 300693
Timisoara, Romania, Tel: 4-0256-222200, Fax 4-0256-490928, e-mail sales@aem.ro
201
III GAS METER WITH SMART CARD consuming energy. So that it saves the energy in the
battery. The valve is able to evaluate the
The gas meter with smart card comprises opening/closing data sent from the electronic
mechanical gas meter with pulse output, valve group, hardware. Other performance of the pressure pulse
electronic hardware, battery group electronic valve are presented in table 2.
software and smart card.
3.3 ELECTRONIC HARDWARE
3.1 MECHANICAL GAS METER WITH PULSE
OUTPUT This unit controls the mechanical meter and
the valve and establishes the communication of the
These are volumetric dry, diaphragms gas system with the gas administration or distribution
meters [3-4] meant for measuring domestic natural company by means of smart card. The electronic
gases consumption. They comply with OIML R6, hardware is a very unique one and all components are
R31 and to SR 6681-98 provisions. Some technical gathered in one PBC. The socket switch locks on
characteristics are presented in Table 1 them are used in the circuit to ensure the rigidity of
Their cases are cupped steel bodies with the system and facilitating the assembly process. The
electrostatic spray paint with epoxipolyesteric electronic components used are chosen with a great
powder. The rotation of gear is transferred via a care and of the best quality.
magnetic coupling.
The gas meter with pulse output converts the 3.4 DISPLAY
data obtained from mechanical meter into an
electrical signal by means of a reed switch group that Thanks to the Alphanumeric LCD used, the
is activated by a permanent magnet. It makes a characters can be read easily. Without any difficulty
sensitive reading and gives 2 pulses per one liter. and hindering the eye look. When the display is
active, its power consumption is so lowly by the help
3.2 VALVE GROUP of display driver.
202
Table 2 [6] the type of consumer,
the number of consumer,
PERFORMANCE
Shrouded coil Intrinsically safe potted the number of meter,
construction and dual the information about the credits,
zener circuit the information about the spare credits, that will
Coil winding resistance be determined by the administration,
ELECTRICAL
Operating pulse (flat top)* 7,0 5% at 20 C the charge condition of the battery,
(with actuator seat
mounted uppermost)
the last data at which the credits has been loaded
to the meter,
100 ms minimum
duration or capacitor the information about the valve malfunctioning,
discharge pulse 2,5 V- the information about the meter malfunctioning
peak and
Energy typically 60m the information about the consumption.
Joule The datas and messages that maintain the
Diaphragm /spindle stroke 7,0 mm nominal. Shut system to be controled are: the alarm messages about
off against flow valve and meter malfunctioning, electronic hardware
Overtravel (shut-off)** pressure and valve battery charge information and a message
1,5 mm maximum/ 500 about meter out of credit.
MECHANICAL
mm diameter range
the valve closes itself automatically for the safety
Spindle to diaphragm
reasons. The valve can be opened and meter can be
Spindle connection spring; ball joint or
put in service again by using a card. This card is used
clipped
by administration officer. The duration 10 days can
Indication for end of
be adjusted up to request.
Switch stroke closure; no
Checking the level of battery is made by the
volts reed switch
electronic hardware continuously. In case the battery
circuit
level is lower than the level it should be, all the datas
Compatible with
are stored in the EEPROM memory. By means of the
Materials natural and
smart card the information about the battery levels
manufactured gas
are carried to the Credit Sales Point letting the
* Typical value. Other combinations are possible to administration be informed too. In case of
suit various and applications malfunctioning of the valve a message is displayed.
** Typical values. Other flow rates and working The meters goes on metering the consumption and
pressure possible up to 10 psi recording the consumption higher than the amount of
203
credit as the consumers debt. In the case of the meter billing system, the reading of meters and the
box is opened by the unauthorized people or administration of the revenue collection. Pre-payment
interfered deliberately, the electronic control unit benefits both utilities and consumers. Utilities benefit
terminates the gas flow by shutting off the valve and because payment is received on average 45 days
records the date in smart card. After the maintenance early than with a conventional billing system. This is
is performed by the administration, the system is not, however the only advantage. Pre-payment
update by the authorization card metering offers improved customer service, no meter
The time and the data adjustment of the readers required, eliminate of bad debts,
meter is made by means of a card. The real time disconnection and reconnection fees, ensure a hand
clock placed in the electronic hardware can be control and eliminate inaccurate meter reading.
programmed to detect the working hours, working The typical user of pre-payment gas
days the month and even the number of days in metering system is a member of lower income groups
February adding one day to the year every 4 years. in the population [7-8]. He will appreciate the fact
If the consumer uses the spare credit and if that he now has direct control over his budget and
the meter goes out of credit not in the working hours, often his acceptance of pre-payment is much higher,
valve will not close itself until the new working days because there is a direct link between the money, he
starts letting the gas being consumed by lending spends and the value he gets. Also pre-payment
credit. When the consumer goes to Credit Sales Point metering system require no cost for
to purchase credit, the landed credits will be disconnection/reconnection and no waiting
decreased from the number of new credits loaded. reconnection and offer ability to payback debts. To
implement a pre-payment metering system, means a
3.7 SMART CARD change of mind set, a change in the way to revenue
collection is managed, a change in IT procedures, a
The cards used in control system conforms change in customer service, a change in metering and
with the ISO7816-2 and they have secure memory. If a change consumer behavior. Because pre-payment
a wrong card is inserted into the meter, the main gas meter is much more expensive than a
control unit identifies that card and warns the conventional meter to be able to reap the benefits as
consumer displaying a message. The names of the expressed above, all parties need to be into the
cards used in the system are: system and appreciate the benefits they themselves
The consumer card by which the data transfer will receive.
data displaying and credit purchasing processes
can be executed. REFERENCES
The authorization card which executes all the [1] *** Improvements in or relating to coin-feed meters GB
operations that the consumer card does with the Patent no 191505193 from 30 March 1916
exception of the credit loading. In necessary, one [2] *** Improvements in prepayment gas meters GB Patent no
can change all the datas and reload them with 191403216 from 8 July 1915
[3] *** Omega Transactions: Technical Reference Series vol. 4
this card. Flow & Level Measurement 2001
Redundant Credit Card by which in case the [4] Dane Enrich A guide to metering technologies ASHRAE
consumer inquires his redundant credits he Journal October 2001 p.33
[5] *** Catalogue AEM Luxten Lighting Company SA
makes back loading by using this card taking [6] *** Pressure pulse valve actuator BLP Components Ltd
finished credit loaded in the meter. After having Catalogue http://www.blpcomp.com
finished this operation, he delivers this card to [7] *** Fuel Pouerty: Low Income, Prepayment Meters and
the Credit sales Point and can take his money Social Obligations Center for Management under Regulation
back. University of Warwick and Center for Competition and Regulation
University of East England March 2001
Switching off Card which is used by gas [8] Roger D Colton Prepayment meters and the Low-Income
administration and executes the meter switching Utility Consumer Fischer, Sheehan & Colton Public Finance and
off process by shutting off the valve. General Economics October 1998
204
Buletinul tiinific al Universitii Politehnica din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract - The program who is the subject of this paper the block diagram (actually the source code this one
develops many possibilities in order to facilitate reading contains the corresponding instructions, constants,
and analysing electrical quantities (voltage, current) for functions and pointers from front panel). Flowing
analogy and digital signals. Using acquisition boards for data is determined in block diagram using links
analogical or digital data from various tranducers,
signals can be analysed or conditioning and
represented by lines between icons [3].
measurements instruments can be created or simulated The hardware-software system used contains the
(virtual instrumentation). following components and programmes:
Keywords: acquisition board, rezistive inductive load, - Toshiba Laptop
one phase rectifier - Windows XP
- LabVIEW 6i, Measurement & Automation
I. INTRODUCTION DAQ Card - 6024E (National Instruments) for
PCMCIA adapters.
In order to increase the works productivity, data
acquisition and data analyse system are used in II. THEORETICAL CONSIDERATIONS
automations systems. One of them is LabVIEW
System [1-5]. In order to do the aquisitions we will use the
In the following lines we will present some of its next theoretical considerations. The resistive voltage
possibilities and performances. We will especially divider is made with coiled resistors and it could
describe data acquisition using DAQ 6024 E 5 6
reach a high level precision ( 10 ...10 ) or it is
acquisition board.
realized with metal resistors and it has a low precision
LabVIEW represents a graphical alternative to 2 3
the conventional programming design for ( 10 ...10 ), but good enough for analogical and
instrumentation It is equipped with all necessary tools digital instrumentation. The divider is used for D.C.
for testing the measurement systems. LabVIEW is a or low frequency voltage measurement.
graphical developed environment designed in order to
create flexible and scalable test, to measure and to
control more rapidly the applications, at a minimal
price. The fastness of this program is high, due to the
introduction of an intuitive graphical interface.
LAbVIEW uses a generally graphical language
for programming called G, containing wide
libraries with proper functions. The LabVIEW
programs are called virtual instruments and are made
from two parts, distributed in two windows:
-the front panel (containing the necessary elements Fig.1 Voltage divider
for interactive operations and the display of the
results) The input, known measure is the D.C. voltage
1
Electrical Engineering Research Institute, Bucharest, Romania,
U 1 and the output measure is the D.C. voltage U 2 .
simona492273@yahoo.com If the divider works no load, it will result an output
2
University Polytechnic Bucharest, Bucharest, Romania, voltage:
dfaur@electro.masuri.pub.ro
205
R2
U 2 = R2 I = U 1 (1)
R1 + R2
U2 R2 1
D= = = (2)
U 1 R1 + R2 R
1+ 1
R2 Fig.3 Electrical diagram for one phase rectifier
The shunt is an input current-voltage and the output voltage for a resistive load
convertor. It is used for currents measurements in
D.C. circuits. The used D.C. shunt is made from Sampling Theorem (Shannon, 1949): any
manganin and it is included into devices in case the signal in continuous time, with a limited spectrum,
currents are less than 20-30A or it is external, as a can be represented without loosing information
separate piece, for a 1000A current. through a sample series of the original signal, or in
other words, through a discrete signal.
The data aquisition systems principal
components are sampled circuits, the memorized ones
and the analogical-numerical convertors. The
numerical and analogical signals caused by
processing can be used to memorize and give back
the information or to command the execution
elements (motors, relay), which control the physical
processes.
To operate with discrete amplitude signals
Fig. 2 The shunt
means a special attention. The result might be often a
sum of quantification noises, with a statistical
The shunt resistance RS is defined between the characterization factors and consequences.
voltage terminals. We can write the following
equations:
IV. DAP APPLICATION
I = I S + I T
(3) In DAP application (the programme is named
RI S = RI R Data_Aquisition_Programme) we will read,
memorize and compute analogical and digital signals,
The shunt establishes the next factor between the particular currents and voltages, in order to analyse
the behaviour of certain system in stationery or
output measure I R and the input one I :
permanent mode.
The programme allow to operator to record data
I R +R R simultaneously, on maximum 16 analogical channels
n= = S =1+ . (4)
IR RS RS (ACH) and 8 digital channels (DIO). The aquisition
board DAQ Card-6024E doesnt admit data reading
synchronization if the aquisition is made
We obtain the computed relation for the shunt
simultaneously for analogical signals in ACH socket,
resistance:
respectively digital signals in DIO socket. Because of
this reason we use both signals type, analogical and
R digital, in ACH socket. We are interesting in data
RS = . (5)
n 1 reading synchronization for at least one digital
channel.
The rectifier has a converting function for the We can use also the special DIO sockets when
electrical energy form A.C. into d.c. His working is the object isnt the perfect synchronization between
depending on the load type, connected at its output. different channels. In this case, a late of
This dependence is shown in a very simple diagram, approximatively 1 second might appear between
see fig.3. The diode should be an ideal one ( u D = 0 signals.
The aquisition board DAQ 6024E can operate
for working state; i D = 0 for blocking state). with a maximum analogical scan rate of 200000
scans/second, meaning a maximum scan rate for each
206
channel of 200000/16 = 12500 scans/second.
Considering that a scan rate of 1000 readings/second
is equal with a millisecond data reading, we can
affirm that the technological possibilities of this board
are properly [4-5].
DAQ 6024E board allows signal aquisition
between 10Vd.c. limits. Its obviously that we need
an intermediate electronic board to adapt the real
acquired signals to the specified interval (10Vd.c.),
with suitable scan factors. The board admits the
independently scanning for each channel.
The two applications windows are described in
fig. 4 and fig. 5. In DAP programme we used many
specific functions:
- each channel has his own configuration; Fig. 5. Bloc Diagram
- START/STOP for the acquisition, controlled
by the user or from one digital channel command Data are displayed in two different ways:
(this allows to display data with a seted number of - in real time (one second constantly updated);
seconds before/after 0/1 passing on that very - historically (displaying the entire interval
channel) ; ordered by the user).
- many possibilities for changing the parameters Data reading is made as long as the programme
(scaling factors for each channel, delay factors on the goes on. One digital channel can order START and/or
OY axis for each channel, zoom on OX axis, the STOP recording, having the possibility to extend and
memory size used by the programme, the scan rate, set a gap in seconds or milliseconds before START
the channels number for reading and displaying, the and after STOP, equal or different periods.
cursor for reading the exact acquisitioned values) ; The advantage appear when we want to record a
- the mean for diagrams (we use the arithmetical transitory phenomenon about whom we dont know
mean, with a setted number of points); exactly the moment it will be happen. Data reading is
- the tangent is computed with the help of two permanently, but data recording has a controlled
selected points coordinates, for the data to be start/stop, given by the operator or presetted.
analysed in every particular mode. This is the way to avoid a useless loading of
memory or even an overcharge of hardware system.
The diagrams allow to simultaneously displaying
all channels for reading or only a few of them after
selection.
The utility of this programme is the possibility to
use it for tracing and visualization of electrical
quantities, any deviation from the normal behaviour
is unliked and it must be eliminated without any
delay. (Example: hydroelectric power
stations, power stations the entire national circuit of
electrical and thermical energy).
The possibility of change the principal
zoo m parameters, which interfere in the acquisition and in
the recording, and also the filtering of the data are
very important programme performances. We can
print all acquired diagrams.
In Fig. 6a we choose to show an 35 seconds
recording, with a 1000 readings/second scan rate, for
the Excitation System supply voltage (SRAT), in case
of changing it with the Backup Supply (AAR).
Through a 10 data meaning we will obtain the Fig. 6b
diagram.
207
52 IV. CONCLUSIONS
50
48
The numerical computing techniques are
limitated from the maximum frequency for analogical
Amplitudine (V)
46
input signals and also from numerical computed
44 speed point of view.
42 In an application these limitations are
depending on the data acquisition system
40
characteristics, on the work speed of the numerical
38 computing systems and on the numerical computing
36 algorithms complexity.There is applications in which
SELECTIE 0 5000 10000 15000 20000 25000 30000 35000
CANAL Timp (citiri/sec) a real time data computing is demanded, meaning that
49 the computing algorithms are correlated with the data
48 access speed. Because of the time axis discretization
47
the analogical signals become discrete. The signal
becomes discrete if we also divide the OY axis. One
Amplitudine (V)
46
condition for the signal to be a good approximation
45
for the analogical one is that the sampling frequency
44
must be big enough reported to the maximum
43
frequency from the sampled signal spectrum
42 (Shannon Theorem). The most insignificant bit signal
41 level must be small enough (the scales must be small
40
15000 20000 30000
on OX and also on OY).We are interested in these
0 5000 10000 25000 35000
MEDIERE Timp (citiri/sec) requirements because the final purpose of this
research paper is to simulate an industrial process, in
Fig. 6 a) Original Data; b) Mean Data. the aim to know it better, to control and to predict it.
The useful signal, representing the physical
Another example of aquisition is the voltage phenomenon or systems behaviour, is mixed with
wave for the current and the voltage obtained at the perturbations, at aquisition and through the
output of a one phase rectifier with an inductive- transmition channel. The discretization introduces a
resistive (RL) load. noise too. The perturbations and noises are
continuous time phenomena, like the useful signals.
Between them is a subjective difference, the
specialists point of view. Because of the high
mathematical level, it is hard to analyse and to
separate them.
The virtual instrumentation utilization
advantages are in the decreasing expenses with new
instruments (the system acquisitioning price, the
expenses with the development and the maintenance)
and increasing performances (flexibility,
reutilization, and reconfiguration).Low prices and
Fig.9 The dropping voltage in case of resistive-inductive load high performances are the desired qualities customers
for one phase rectifier.
expect from their delivers.
REFERENCES
208
Buletinul tiinific al Universitii Politehnica din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
209
B. Calibration method C. Determining the value of the measurand
Methods used for the calibration [5] are: The true values of the calibrators are determined in
- substitution the reference conditions.
- direct comparison In the absence of other indications, the zero
- indirect comparison adjustment is made on the smallest interval (if
The selection of method is made according to the possible).
accuracy of the calibrator to be calibrated and the Calibrator will be calibrated to the best accuracy,
measured quantity. corresponding to long-term stability.
For the calibration of calibrators on the function to We use standards ensuring:
generate electrical direct voltage, we can use two - stable values of generated quantities
methods: - required resolution
- substitution, for calibrators with high accuracy - measurement intervals to cover
- direct comparison, for calibrator with lower generation intervals of the calibrators to
accuracy. be calibrated
For calibration of high-accuracy calibrators on the - required accuracy/uncertainty
function to generate direct voltage, the multiple- Determining the value of the measurand and data
function calibrator, is used as standard, and the digital processing, for the calibration of calibrators on the
multimeter with 8 digits or 7 digits is used as functions of generating direct voltage using the
measuring mean. substitution method.
The assembly method is shown in Fig. 1. The calibrator generates direct voltages for each
measuring point; they can be read on the digital
multimeter display.
We conduct n readings for each point, in conditions of
+ + repeatability. We calculate the average X X .
Standard Calibrator to For the same measuring points, standard calibrator
calibrator be measured
_ _ generated voltages measured with the same
multimeter. We conduct n readings for each point, in
repeatability conditions. We calculate the average X E .
We determine the true value [2] according to formula:
K
X = XN + ( X X - X OX +XTX ) ( X E - X OE -XE - XDE
+ XTE) + XR+ XX (1)
Digital
multimeter + where:
XN the nominal value displayed on the
_ calibrator to be calibrated
XX average value for n direct voltages
values generated by the calibrator
Fig. 1 Assembly in case of substitution and read on a digital multimeter
with resolution appropriate for
Most often, the calibration process may include three measuring
approaches of the calibration, depending on the X OX average value for n zero direct
customers requests and level of information:
voltages generated by the calibrator
a) calibration before adjustment, in case we
and read on the same digital
notice errors more significant than the
multimeter having the same
measurement errors provided in the technical
resolution
specification of the apparatus to be calibrated
-adjustment (manual or using a software) XTX correction of calibrator due to
-calibration after adjustment environment temperature
b) calibration XE average value for n direct voltages
c) adjustment, when the Customer knows that generated by the standard calibrator
the apparatus must be adjusted and informs the read on the same digital multimeter,
laboratory thereof with resolution appropriate to
- calibration measuring
If errors are more significant than the ones admitted X OE average value for n zero direct
by the technical book, the calibrator will be adjusted, voltages generated by standard
if possible, and the errors determined prior and after calibrator and read on the same
adjustment will be specified in the calibration digital multimeter having the same
certificate. resolution
210
XE the correction of the value of the the value in the calibration certificate of the standard
standard (the values in the used.
calibration certificate)
Uncertainty of element XE.
XDE correction due to time drift of the
standard The value of the measurement uncertainty U(XE)
XTE correction of standard calibrator provided in the calibration certificate of the standard
due to environment temperature calibrator,.
XR correction due to variations of the e) Correction due to drift in time of standard
supply network XDE is taken into account depending on the data
XX correction due to instability of the provided by the history of the standard used.
value displayed by the digital
Uncertainty of element XDE.
multimeter
In formula (1), the first bracket represents the basis Based on the history of the standard, a time drift a is
for the calculation of the value indicated by the evaluated. The associated uncertainty is calculated as:
calibrator to be calibrated.
Each element in formula (1) has a certain value and a
related uncertainty [3, 4]. a
XDE) = (6)
a) average value X X , for direct voltage 3
generated by the calibrator, for n readings made on
the digital multimeter, in conditions of repeatability, f) Correction of standard calibrator, due to
is calculated according to the formula: environment temperature XTE, is applied only when
measurements are conducted at a temperature
n differing from the reference temperature and the
X Xi technical book provides temperature correction
XX = i =1 (2) coefficients cTE, for standard calibrator
n
n
a
(X Ei X E )2
u(XTX) = (9)
u( X E ) = i =1
n(n 1)
(5) 3
211
XR= 0. Table 1
a XN 1V 0 0
u(XR) = (10)
6 0.9999740 V 0.086 V Normal 1 0.086 V
XX
212
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract This paper presents a polarization method bring about his anisotropy, meaning that the refractive
for electrical current measurement using optical fiber. index is depending on the polarization plane of light.
Classical polarization method has the advantage of being The effects used are: electrical optical effect (Kerr,
simple but for getting a good performance the optical Pockels) or magneto-optical effect (Faraday).
components have to be carefully assembled due to their
sensitivity to disturbance (temperature, vibration). In II. THE FARADAY EFFECT
the field of power systems the most common sensors are
those used for the measurement of electrical quantities The current measuring is based on Faraday effect in a
and to oversee the electrical equipment under voltage. single-mode fiber, which is wrapped around a
Many times these sensors are located in hardly
conductor in a closed loop [2]. The light waves
accessible places (vacuum spaces, with oil, gas, etc.) or
connected at a higher potential than earth potential. polarization direction is rotated by the magnetically
fields is parallel to the light waves path (fig. 1).
Keywords: Polarization Method, Current Measurement,
Optical Fiber, Optical Fiber Sensor, Faraday Effect The Faraday effect is not reciprocal, meaning it
doesnt vary with the light propagation distance, while
I. INTRODUCTION optical activity is a reciprocal phenomenon. This
solution led to decrease sensitivity to temperature
Regularly, for the measurement of electrical currents variations. It must be avoided to establish a fixed
are used devices and sensors, which allow galvanic point in the fiber and the twist sense changing must be
separation from the conductor where the current to done carefully.
measure passes, like current transformers or sensors From a macroscopic point of view, the Faraday effect
based on Hall or Faraday effects. Other important appears like a rotation of the polarization plane of a
requirement is the immunity to the electromagnetic polarized light. This light is propagating in optical
perturbations encountered in electrical power stations. fiber that is the magnetic field. This rotation of the
The research of new sensors called unconventional polarization plane, due to Faraday effect is
it was quickly oriented to optical measurement proportional to the magnetic field circulation ds along
methods and especially for using Faraday effect). This the luminous trajectory (path) according to the
method was first investigated using a glass bar, expression:
developed later due to the apparition of optical fibers.
Optical fibers sensors can perform all these F = V H (s )ds (1)
requirements because galvanic separation and L
electromagnetic interferences are inherent properties where:
of optical fibers [1]. S curvilinear abscissa along the luminous
In the field of power systems the most common trajectory;
sensors are those used for the measurement of L - total length of luminous trajectory;
electrical quantities and to oversee the electrical V Verdet constant (a property of propagation
equipment under voltage. Many times these sensors
medium).
are located in hardly accessible places (vacuum
spaces, with oil, gas, etc.) or connected at a higher
potential than earth potential
The measurement of electrical quantities (current,
voltage) using optical fibers is based on the
polarization modulation. This phenomenon supposes
that the quantity to measure produce modifications of
the birefringence at the optical fiber level. The
birefringence of the active medium can be realized to
_____________________________________________________________
1
Faculty of Electrical Engineering, Bucharest Politehnica University, 313, Splaiul Independentei Street,
phone. +4021-4029258, email: vsimion@electro.masuri.pub.ro
2
Faculty of Computer Science, Bucharest Politehnica University, 313, Splaiul Independentei Street, phone: +4021-4029351,
Email: cstefan@cs.pub.ro
213
IV. THE ANSWER OF OPTICAL FIBER
CURRENT SENSOR
214
Ttot = Trefl Tat T fin Tld fo (11)
The total transmittance when using a light analyzer in
system is:
TP tot = T pol Ttotala (12)
215
will increase, meaning that the rapport signal vs. noise
will be better.
We could use a Wollaston prism instead the analyzer,
allowing the compensation of certain influence
quantities, but this solution is expensive.
These types of optical current sensor have the small
dimensions and it offered the immunity of the
electromagnetic perturbations and the galvanic
isolation.
REFERENCES
[1] Royer, P., Capteur de courant et traitements associes,
SEE Limoges, 14-15 oct 1995;
Fig. 4 The transfer function for the alternating current
[1] Dakin, J., Culshaw, B., Optical Fiber Sensors: Principles
and Components, Artech House, Boston, London, vol. I,
1988;
[2] Chai, Y.,Handbook of Fiber Optics. Theory and
Applications Academic Press, Inc., San Diego, California,
1980;
[3] www.oxford-electronics.com;
[4] Stanciu, M., Senzori cu fibre optice, Ed. Secorex, Bucureti,
2001;
[5] Ionescu, A., Traductoare pentru automatizri industriale,
Ed. Tehnic, Vol. I, Bucureti, 1985;
[6] Simion, V., Pantelimon, B., tefnescu, C., Aspecte privind
pierderile din sistemele de msurare cu fibre optice,
Proceedings of International Metrology Conference, (ISBN:
973-99385-5-8), vol.3, pag. 855-858, Bucharest, Romania,
September 18-20, 2001;
VII. CONCLUSION
The current sensor described above works employing
the Faraday effect. The major difficulty in the
construction of such a sensor consists of using a
classical detection configuration.
The used solution attests the stability and the good
linearity of the method and allows the projection for a
very large scale of current by a good choosing of the
inductor. Our solution confers a good linearity and
that will permit the design for a large current domain.
When the currents have large values, the sensibility
216
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract: The paper presents the efforts made commercial transactions, work safety and
in order to accurately measure electrical energy, environment protection, (), are subjected to the
n activity belonging to the public domain. mandatory metrological control of the state.
The electric energy meters are instruments submitted The manufacturers of measuring instruments in our
to the mandatory metrological control of the state,
in order to assure the consumer protection in this
country such as AEM and Luxten Lighting from
field. Some matters of concern for the state regarding Timisoara, Electromagnetica from Bucharest, are
the transposing of European and international norms interested in providing quality products, that meet the
in view of the accession of our country to the European requirements of international standards, and especially
Union are also presented. of the European ones, in order to comply with the
rigours of the European Union.
Key words: static meter
1
Romanian Bureau of Legal Metrology, National Institute of Metrology, Electrical Measurements Laboratory, sos. Vitan Birzesti
No. 11, Bucharest, Romania, e-mail: buzac@inm.ro
2
Romanian Bureau of Legal Metrology, National Institute of Metrology, Interdisciplinary Metrology, sos. Vitan Birzesti No. 11,
Bucharest, Romania, e-mail: urdea@inm.ro
3
Politechnica University Bucharest, Faculty of Electrotechnics, Measurement Apparatuses and Converters Department,
sp. Independentei No. 331, Bucharest, Romania, e-mail: costin@electro.masuri.pub.ro
217
e) metrological verifications after the repair or IEC 13251:1996:
modification of an instrument; International vocabulary of basic and general
f) metrological surveillance of measuring terms in metrology
instruments.
and specifies the metrological and technical
According to the provisions of OG No. 20/1992, the conditions for the following types of control: pattern
Romanian Bureau of Legal Metrology (BRML), a approval, initial verification, verification after repair
specialised body of the state central public and periodic verification. The norm applies to
administration, responsible with the co-ordination of induction and static single phase and three-phase
the metrology activities in Romania, identifies the active electric energy meters of accuracy classes 0.2;
measuring instruments used in the public domain and 0.5; 1 and 2, used in AC networks.
nominates them in the Official List of measuring
instruments submitted to the mandatory metrological III. MODERN INSTRUMENTS FOR THE
control of the state. MEASUREMENT OF THE ELECTRICAL
This list is published, and periodically updated, in the ENERGY
Official Journal of Romania.
Among the measuring instruments submitted to the There is a variety of instruments that may be used to
mandatory metrological control of the state, there are: measure electric energy. Beside the classical
watt-hour meters for active electric energy, whose
- measuring rules and tapes;
operation is based on the electromagnetic induction
- meters for cold water up to DN 800 and hot water
principle, a wide variety of electronic meters has been
up to DN 400;
developed in the latest years at a rapidly increasing
- electronic converters for gas volume;
rate, along with the development of microelectronics,
- programmable clocks for watt-hour meters;
based on the following measurement principles:
- load cells;
- single phase active/reactive electric energy
- the double amplitude and duration modulation
meters;
principle;
- gas analysers;
- the Hall multiplier principle;
- medical monitoring equipment for patients, etc.
- the thermoelectric multiplier principle.
Using its specialised departments, BRML establishes
The rapid modernization of the electric energy meters
the appropriate metrological control mechanisms,
is also a result of the manufacturers efforts to meet
applicable to each type of measuring instrument, as
the current requirements of the consumers regarding
well as the maximum permissible interval between
the technical, structural and metrological
two subsequent metrological verifications [1].
characteristics of these measuring instruments used in
The measurement of electric energy belongs to the
the public domain.
public domain and, therefore, electric energy meters
Due to the fact that these measuring instruments are
are instruments submitted to the mandatory
involved in the commercial transactions between the
metrological control of the state.
suppliers and the consumers, they are submitted to
The metrological assessment of these instruments is
the mandatory control of the state and, therefore, they
carried out according to the Legal Metrology Norm
are part of the regulated area, with specific aspects
(NML) No. 5-02-97, currently in force.
related to the consumer protection that are to be dealt
This norm has been prepared in compliance with the
with.
international regulations applicable to this field,
Based on its role and competence and accumulated
namely:
experience, at the National Institute of Metrology,
within the AC Measurements workgroup of its
IEC 60687:1992:
Electrical Measurements Laboratory, a large number
Alternating current static watt-hour meters for
of electric energy meters for low energy consumers
active energy (classes 0.2 S and 0.5 S)
were assessed, within their type testing for the pattern
approval certificates granted by the Pattern Approval
IEC 60387:1992:
Department (SAM) of BRML, as well as within their
Symbols for alternating current electricity meters
initial verification, their subsequent periodic
verification and their verification after repair.
IEC 60521:1988:
The active electric energy meters for household and
Class 0.5, 1 and 2 alternating-current watt-hour
industrial consumers are tested in metrology
meters
laboratories where all the tests required by the
specialized norms in force are carried out in order to
IEC 61036:1996:
assess the meters.
Alternating current static watt-hour meters for
A new issue that has to be currently dealt with is the
active energy (classes 1 and 2)
quality of the electric energy.
218
Tests such as:
CALIST 3 Programme, nine European norms were
- impulse voltage tests for circuits and between the translated and adopted as Romanian standards within
circuits; TC 164, such as:
- tests for electromagnetic compatibility (EMC);
- tests of immunity of electrostatic discharges; IEC 61358:1996:
- tests of immunity to electromagnetic HF fields; Acceptance inspection for direct connected
- fast transient burst test; alternating current static watt-hour meters for
- radio interference measurement; active energy (classes 1 and 2)
219
All these actions, that were carried out and are still REFERENCES
going on at the institutional level of metrology, as [1] *** Ordinance of the Government No. 20/1992 regarding
well as the continuous efforts of the manufacturers to Metrology.
improve the quality of their products intended for use
[2] *** SR 13251/1996, International vocabulary of Basic and
in the public domain, together with the programmes General Metrological Terms.
initiated by the Romanian government, in order to
harmonise the Romanian laws and regulations with
the EU legislation, are all aiming at facilitating the
accession to the European Union of our country.
220
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Assoc. Prof.dr.eng. Faculty of Electrical Engineering, University Valahia Targoviste, Romania18-24, Bdul. Unirii, Targoviste,
Dambovita, Romania, Phone/fax: +40-245-217 683; email: dogaru@valahia.ro
2
Prof.dr.eng. Faculty of Electrical Engineering, University Politehnica Bucharest, Romania, 313, Splaiul Independentei, 77 206,
Phone/fax: +40-21-410 04 00/ +40-1-410 43 55; email: costin@electro.masuri.pub.ro
3
Prof.dr.eng. Faculty of Electrical Engineering, University Valahia Targoviste, Romania, 18-24, Bdul. Unirii, Targoviste, Dambovita,
Romania, Phone/fax: +40-245-217 683; email: handrei@valahia.ro
4
As..drd.eng. Faculty of Electrical Engineering, University Valahia Targoviste, Romania18-24, Bdul. Unirii, Targoviste, Dambovita,
Romania, Phone/fax: +40-245-217 683; email:ahusu@valahia.ro
5
stud. Faculty of Electrical Power Engineering, University Politehnica Bucharest, Romania, 313, Splaiul Independentei
6
Eng. Faculty of Electrical Engineering, University Valahia Targoviste, Romania18-24, Bdul. Unirii, Targoviste, Dambovita, Romania
221
range range refers to the minimum and sources contain a signal that is not connected to an
maximum voltage levels that the ADC can absolute reference. Some common examples of
quantize floating signals are batteries, thermocouples,
Before you begin developing measurement transformers, PV cells.
applications, you must install and configure the A measurement system can be placed in one of three
measurement hardware. The software drivers need the categories: differential, referenced single-ended,
hardware configuration information to program the nonreferenced single-ended. In a differential
hardware properly. When measuring a physical measurement system, you do not need to connect
phenomena, a transducer must convert this either input to a fixed reference. DAQ devices with
phenomena into a measurable electrical signal. instrumentation amplifiers can be configured as
Common types of signal conditioning include differential measurement systems. An ideal
amplification, linearization, transducer excitation and differential measurement system, reads only the
isolation. Some signal conditioning can be performed potential difference between its two terminals inputs.
in the software in the Data Acquisition function A referenced single-ended measurement system
palette. measures a signal with respect to building ground.
DAQ devices often use a nonreferenced single-ended
measurement system, wich is a variation of the
referenced single-ended measurement system. In these
case, all measuremnts are made with respect to a
common reference, because all of the input signals are
already grounded (AISENSE is the common reference
for taking measurements and all signals in the system
share this common reference. AIGND is the system
ground).
LabVIEW for Windows installs a configuration utility
for establishing all board and channel configuration
parameters. This utility is known as the Measurement
& Automation Explorer MAX. After installing a
DAQ board in computer, MAX utility reads the
information the Device Manager records in the
Windows registry and assigns a logical device number
to each DAQ board. The configuration of channels for
this application is presented in figure 1.
222
(which connects to P-Type material) and the where q is the electronic charge and is equal to
negative voltage is applied to the cathode lead 1.602 x 10-19 coulombs, k is the Boltzmann
(which connects to N-Type material). constant with a value of 1.381 x 10-23 J/K and T is
A reverse bias results in no current flow through the temperature in Kelvin.
the diode (diode blocks). A diode is reverse biased The zener diode uses a p-n junction in reverse bias to
when the anode lead is made negative and the make use of the zener effect, which is a breakdown
cathode lead is made positive. phenomenon which holds the voltage close to a
The P-N Junction region has three important constant value called the zener voltage.
characteristics:
1) The junction is region itself has no charge carriers
and is known as a depletion region.
2) The junction (depletion) region has a physical
thickness that varies with the applied voltage. A
forward bias decreases the thickness of the depletion
region; a reverse bias increases the thickness of the
depletion region.
3) There is a voltage, or potential hill, associated with
the junction. Approximately 0.3 of a volt is required
to forward bias a germanium diode; 0.5 to 0.7 of a Fig. 5. Zener diode
volt is required to forward bias a silicon diode.
The social strong involvement of the energy systems
Semiconductor diodes are made by joining two
and the complex boundaries between these system
different types of semiconductor materials in a special
and all other technical systems, or the environment,
way so that when a proper polarity voltage is applied,
have imposed the development of some researches on
electrons readily pass through one material to the
unconventional process of producing electric energy.
other. However, if the voltage is reversed, there is
Nowadays there are many unconventional methods
very minimal electron flow.
for obtaining electric energy, based on much or less
In other words, a semiconductor diode allows current
studied physical or chemical phenomena. Photovoltaic
to pass through when in forward bias, and blocks
systems convert sunlight energy into electric energy
current when in reverse bias. They also have
and they are characterized by modularity, functional
properties or characteristics that enable them to
autonomy and long function period.
perform many different electronic functions.
To ensure the accuracy of the measurement, the
operating parameters of the photovoltaic system and
the configuration of the acquisition system are taken
into account and have imposed the signal conditioning
and the setting of the signal source, of the field and of
the channels.
223
(current-voltage, power-voltage, power-charge photovoltaic panels. After shutting off the experiment,
resistance), to present the measured parameters the investigators apply appropriate curve fitting
(during the data acquisition) in tables, continuous techniques to determine the functional relationship
acquisition, to save data into files for future I=f(U) manifest in their data. Curve fitting represent a
processing. technique for extracting a set of curve parameters or
Building integrated photovoltaics, the integration of coefficients from the data set to obtain a functional
photovoltaic cells into one or more of the exterior description of the data set. LabVIEW provides built-in
surfaces of the building envelope, represents a small VIs that perform a least-squares fit of data to
but growing photovoltaic application. In order for commonly used equations including a strainght line,
building owners, designers, and architects to make an exponential curve and a mth order polynomial. As
informed economic decisions regarding the use of an illustrative example of how these curve-fitting VIs
building integrated photovoltaics, accurate predictive function, see the image shown in fig.8.
tools and performance data are needed. At Valahia
University of Targoviste, enclosed in the ICOP
DEMO 4080-90 European research program, a
photovoltaic system has been realized, with an
installed power of 10 kWp, composed by 66
OPTISOL SFM 72 Bx photovoltaic modules made by
Pilkington Solar International and 24 ST 40 modules
produced by Siemens. These modules are connected
to Sunny Boy invertors.
IV. CONCLUSIONS
REFERENCES
[1] R. Bishop, Learning with LabVIEW 6i Pretince Hall PTR,
Upper Saddle River, New Jersey, 07458, 2001
[2] J. Essick, Advanced LabVIEW labs, Pretince Hall PTR,
Upper Saddle River, New Jersey, 07458, 1999
[3] R. Jamal, LabVIEW applications and solutions, Pretince Hall
PTR, Upper Saddle River, New Jersey, 07458, 1999
[4] V. Maier, LabVIEW in calitatea energiei electrice, Ed,
Albasta, Cluj-Napoca, 2001
[5] *** National Instruments - Data Acquisition Basics Manual
[6]*** National Instruments Measurements Manual
224
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract In this paper the performances of two Moreover, the accuracy of the frequency of the
methods in the implementation of a sinewave with a sinewave implemented is determinate.
16-bit fixed-point digital signal processor (DSP)
TMS320C5x are evaluated and compared. The methods II. THE METHODS USED FOR IMPLEMENTING
used are the recurrence formula method and the look-
up-table (LUT) method. By each method the sinewaves
THE SINEWAVES
are generated by simulation and with the TMS320C5x
Starter Kit (DSK) board. The performances of the A discrete-time sinewave can be implemented by the
methods are appreciated by means of the dynamic following two methods:
parameters of the sinewaves implemented: signal-to-
noise and distortion ratio (SINAD) and total harmonic A. The recurrence formula method
distortion (THD). The accuracy of the frequency of the We consider a continual-time sinewave x(t)
sinewaves implemented is also determinate. characterized by amplitude A and frequency fin. After
Keywords: implementation of the sinewave, digital the sampling with frequency fs (fs > 2 fin) the discrete-
signal processor, accurate estimate of the sinewave
dynamic parameters.
time sinewave x[n] is obtained. x[n] is given by
f
x[n] = A sin 2 in n , n = 0, 1, 2, K (1)
I. INTRODUCTION f s
The sinewaves are the easiest to generate in practice at
Based upon (1), after some simple algebra, the
the frequencies with adequate accuracy. From these
following recurrence relationship is achieved
reasons the sinewaves are used in many applications,
such as communication, instrumentation and control.
x[n] = 2ax[n 1] x[n 2] n = 2, 3, 4, K (2)
The sinewaves can be implemented with analog or
digital circuitry. Due to its advantages concerning the
accuracy, speed and the easiest in establishing the f
parameters the microprocessors and the digital signal where: a = cos 2 in ;
fs
processors (DSP) are mostly employed for
implementing the digital sinewave generators [1]-[3]. x[0] = 0 (from (1));
The objective of this paper is the determination and f
x[1] = A sin 2 in (from (1)).
the comparison of the performances of two methods fs
in the implementation of a sinewave with a 16-bit
Thus, from (2) it follows that the discrete-time
fixed-point DSP-TMS320C5x. The methods are ones
sinewave x[n] can be implemented by means of a
of the most used in the implementation of a digital
recurrence formula.
sinewave: the recurrence formula method [1], [4] and
the look-up-table (LUT) method [3].
B. The look-up-table (LUT) method
Using each method the sinewaves are generated by
If N samples of the x[n] (n = 0, 1, 2, , N-1) are
simulation and by means of the TMS320C5x Starter
acquired, then the relationship between the
Kit (DSK) board. The accuracy of the sinewave
frequencies fin and fs is given by
implemented was determinate by its dynamic
parameters: signal-to-noise and distortion ratio
(SINAD) and total harmonic distortion (THD). f in J +
= (3)
fs N
1
Dept. of Measurements and Optical Electronics,
Faculty of Electronics and Telecommunications,
e-mail: daniel.belega@etc.utt.ro
225
generated with 15.625 kHz sampling frequency. The
where J is the number of sinewave cycles (J is an sampling frequency of the acquisition process was
integer) and 0 < 1. equal to 15.625 kHz. Were acquired N = 1024
For = 0 the sampling process is coherent with the samples.
sinewave and (3) represents the coherent sampling
relationship between the frequencies fin and fs. In this
case, based upon (1), x[n] is periodically with the
period N, i.e. x[n] = x[n + N]. So, x[n] is completely
represented by its first N samples (n = 0, 1, 2,, N-1).
The LUT method consists in the following two steps:
step 1: The first N samples of x[n] (n = 0, 1, 2,,
N -1) are stored in a memory (table).
step 2: The sinewave is generated by stepping the
table, wrapping around at the end of the table
whenever n N, i.e. the following sequence is
obtained
Sinewave generator
RS232 OUT
IN
DSK5X board #1
PC #1
Acquisition
(b)
RS232 OUT
IN
DSK5X board #2
PC #2
(b) (b)
(c) (c)
Fig. 4. The performances of the sinewaves implemented by the Fig. 5. The performances of the sinewaves implemented by the
recurrence formula method at two sampling frequency of the LUT method at two sampling frequency of the acquisition process:
acquisition process: fs = 7.95 kHz (star) and fs = 15.625 kHz (circle).
fs = 7.95 kHz (star) and fs = 15.625 kHz (circle).
By comparison the performances SINAD and THD of
Fig. 5 presents the performances of the sinewaves the sinewaves implemented with TMS320C5X DSK
generated by the LUT method with the algorithm board at the acquisition sampling frequency
proposed at the same frequencies as in simulation. fs = 15.625 kHz with the ones obtained from
There were used, also, two sampling frequencies of simulated sinewaves it follows that the performances
the acquisition process fs = 7.95 kHz and are severely degraded because of the poor
fs = 15.625kHz and were acquired N = 1024 samples. performances of the low-pass output reconstruction
filter of the AIC [9]. Due to the behavior of the AIC
output filter the dynamic performances of the
sinewaves implemented by both methods are very
close. However, the dynamic performances of the
sinewaves are, in many cases, superior to the ones
obtained with the analog sinewave oscillators.
228
From Figs. 4(c) and 5(c) it follows that the sinewaves ACKNOWLEDGEMENT
frequencies are more accurate at small frequencies
when the LUT method with the proposed algorithm is The results reported here were partially obtained in
employed than when the recurrence formula method is the framework of the CNCSIS Grant AT 32940
used. number 1, dedicated to the analog-to-digital
Another important conclusion drawn from Figs. 4 and converters dynamic testing in multi-tone mode.
5 is that the performances of the sinewaves are not
affected by the value of the sampling frequency of the
acquisition process. RFERENCES
229
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
230
t m = ( N 1)Ts , (3) 0.5 1
f = + fs = + fs , (7)
Np Nz
where N is the number of acquired data samples, and
where fs represents the relative accuracy of sampling
using equations (1) and (2), it follows that the
frequency and depends on the manner of practical
frequency is
implementation of the sampling process (usually, the
N p fs on-board counter-timer circuit give the clock for this
f = . (4)
N 1 process). It follows that the single way to minimize f
is represented by choosing a higher value for Np, or
equivalently, for Nz. For a given frequency, Np is
direct proportional with N (see equation (4)).
Also, based on (6) it follows that the frequency
resolution (also the minimum measurable frequency)
is
fs
f = . (8)
2( N 1)
231
2( N 1) increases and become higher than fs/2, the number of
N za = kf s f , (10) zero crossings results from equation (10) with k = 1
fs
and will decrease. Thus follows the false conclusion
that the frequency would decrease. It is specify that in
where k is a positive integer, and its value is that the all presented cases, the sampling frequency is
expression |kfsf | to be minimum. It can be shown constant.
that for k = 0, equation (10) is identically with (9). In order to avoid this false measurement, it is
Thus, for a correct measurement, namely for f<fs/2, desired the detecting of the aliasing phenomenon, that
when N is constant, the increasing of f is equivalent is the case when f > fs/2. One way to do this is by
with the increasing of number of zero crossings, and making the data acquisition with two different
the decreasing of f is equivalent with the decreasing of sampling frequencies.
number of zero crossings (Nz in equation (9) ). If f
For this purpose, first in fig.3 is presented the detected, but the transition from aliasing free to
variation of number of zero crossings function on aliasing can be detected.
sampling frequency for a constant number of samples, In order to materialize above statements, one
if also the signal frequency is constant.. Thus, two proposes the following idea.
statements result from this figure: On each measurement, further on current data
1. If sampling theorem is satisfied, fs 2f, the acquisition with sampling frequency denoted as fs1
number of zero crossing decreases as sampling and corresponding number of zero crossings Nz1, it
frequency increases. makes an additional data acquisition with fs2 = kfs1,
2. If sampling theorem is not satisfied, fs<2f, the where k is higher than 1, followed by the computing
number of zero crossings can increase or can decrease of the new value of the number of zero crossings,
as sampling frequency increases. denoted as Nz2. In fig.4 are presented the variations of
From these statements, results that the aliasing number of zero crossings function on signal
phenomenon can be not detected. frequency, for both values of sampling frequency, fs1
However, if sampling theorem is satisfied, and then and fs2. Also, the number of samples N is constant.
the signal frequency increases, that f>fs/2 but f<fs, in From fig.4 it can be seen that, if f<fs1/2, then Nz2 <Nz1.
this case number of zero crossing increases as This is the normal case. If f > fs1/2, i.e. the aliasing
sampling frequency increases, in opposite with above occurs, the inequality Nz2 <Nz1 is satisfied, but only for
first statement. In conclusion, the aliasing can be not a little interval, namely for f (fs1/2, fc). f=fc (denoted
as critical frequency) represents the point for which
232
the two curves one intersects. If f > fc, one obtains that Thus, the proposed methods were implemented on a
Nz2 >Nz1, this being the element which signals that data acquisition board of type National Instruments
aliasing was occured. PCI 6023.
It is noted however that this idea for aliasing Mainly, this algorithm does the following:
detecting can be applied only if the measurement -the acquisition of data samples begin at maximum
process begin in a case for which the aliasing not value of sampling frequency, in accordance with
occur and the increasing of frequency to be that its features of the used data acquisition board; if it is
value to be not greater than sampling frequency. necessary, the sampling frequency is decreased until a
In order to computes the frequency fc beginning the approximate value of frequency can be computed;
aliasing is detected, the expressions of both zero - based on previously computed frequency, the
crossings, for f > fs/2 must be equals, as next is optimum sampling frequency and the optimum
presented. number of samples are choose;
-the permanent frequency measurement is achieved:
2( N 1)( f s1 f c ) 2( N 1) f c the frequency is computed and displayed, and it is
= (11) checked if the aliasing occurs; if it there are, an
f s1 f s2
message is displayed and the sampling frequency is
It results that increased to the next value.
In order to achieve the measurement ranges, it is
f s1 f s 2 necessary the settling of the sampling frequency (i.e.
fc = (12)
f s1 + f s 2 the highest limit of range is one half the sampling
frequency). Further, the number of samples must also
be settled because, so was pointed, these two amounts
together establish the frequency resolution. Next, in
the table 1, the sampling frequency, the resolution, the
number of samples and the measurement time for
each range are presented. The values correspond to
equations (8) and (3).
Table 1
Range fs Hz Resolution N tm
233
computing the frequency with an acceptable accuracy IV. EXPERIMENTAL RESULTS
(Np=50).
3. Acquire N1 data samples with fs/2. Go to 2. The presented algorithm was experimental tested in
4. Compute the frequency f using (6). Set fs= 2f. order to achieve the desired frequency meter, and the
5. Choose the sampling frequency function of fs: the obtained results will be presented in this section.
less value from table 1 which is higher than fs. Denote Thus, a HM 8130 function generator was used in
this value by fs1. Also, function on selected sampling order to apply the signal to be measured. The type of
frequency, from table 1 results the number of samples waveform was mainly sine, but also square, triangle
N. or sawtooth were used. This generator displays the
6. It is make the current frequency measurement, value of generated frequency and will be used as
with the sampling frequency to fs1. reference for comparison with the implemented
The next two steps are executed continuously. instrument.
6.1. Compute Nz1 at fs1 for N samples. Compute Two categories of experiments have been achieved.
frequency f using (6) and display its value. First, the accuracy of measurement was verified, for
6.2. Compute Nz2 at fs2=kfs1 for N/10 samples. If all ranges of the instrument. Second, the manner of
Nz2>Nz1/10 the aliasing occurs; display the message detecting the aliasing was tested.
Aliasing and utter a sound; set fs1 to the next value Thus, in order to verify the precision of the
accordingly tab.1. instrument, different frequencies were settled to the
In first three steps, the frequency is computed with HM 8130 function generator. For each frequency, a
an accuracy of 2 percent, because the thereshold Nt number of 50 measurements were made with the
has been imposed as 50. The value of 501 for number implemented instrument and the obtained results were
of data samples allows as the sampling frequency to stored.
be decreased to 1/20 from signal frequency. In table 2, for each frequency from HM 8130 are
In order as the needed time for aliasing detecting to presented the distribution of the obtained values fm
be as low as possible, in the second data acquisition (the value and the number of achievements).
only N/10 samples has been acquired.
If is necessary, for instance when the measurement Table 2
signal is changed, the step.6 can be executed only f (from HM8130) fm
once, and then the algorithm can be resumed 239.30 Hz 239.3 Hz 239.4 Hz
beginning with step.1, but in this mode the needed 49 1
time is higher than that for executing of step.6. 729.00 Hz 729.0 Hz 729.1 Hz
Based on previous presented two restrictions, the 43 7
constant k has value of 1.05. 2207.0 Hz 2207 Hz 2208 Hz
Also, next is presented the way for determination of 48 2
number of zero crossings, where e(i) represents the 8554.0 Hz 8554 Hz 8555 Hz
sample at moment i and Nz represent the number of 45 5
zero crossings: 23930 Hz 23930 Hz 23931 Hz
For i=2,,N
38 12
If e(i-1)e(i)<0 or e(i-1)=0, Nz=Nz+1.
41040 Hz 41040 Hz 41050 Hz
It is specify that before of determination the number
47 3
of zero crossings, the dc component of signal is
removed by using the adaptive LMS algorithm [7]. 91670 Hz 91670 Hz 91680 Hz
This algorithm was implemented by a program in C 40 10
language. For this purpose, the low level functions 239.28 kHz 239.28 kHz 239.29 kHz
from NI-DAQ software was used in BorlandC 5.0. 36 14
The algorithm is an off-line real time algorithm (see
[6]), because it processes a group of N data samples, From the results from table 2 it can be seen that the
that previous are stored in the memory. This way was quantization error from numbering of periods brings
imposed because the functions from NI-DAQ give about the frequency resolution. For all frequency, the
groups with a pre-established number of data samples, number of achievements of the results without error is
which can be processed only when the acquisition is higher than that of the results affected by quantization
ready. For this reason, the measurement time is a little error.
amount higher then that from table 1, namely with the In order to verify the detection of aliasing, on each
required time for computing the number of zero measurement range, the frequency of measured signal
crossings and frequency. However this increasing is was modified beginning from values a little amount
visible only for last three ranges, thus the new value less than one half the sampling frequency, to values a
of the measurement time for these ranges is about of little amount higher than one half the sampling
0.07s. frequency, and then the value for which the
instrument signals aliasing was kept in mind. This
value has been compared with the theoretical value,
which is obtained using equation (12).
234
Thus, in table 3 these data that allow the detection
of aliasing are presented. fcexp represents the
experimental value which has been obtained for
signaling of aliasing.
Table 3
fs1 fc fcexp
500 Hz 256.09 Hz 256.3 Hz
2000 Hz 1024.39 Hz 1025.5 Hz
5000 Hz 2560.97 Hz 2564.3 Hz
20000 Hz 10243.9 Hz 10210 Hz
50000 Hz 25609.7 Hz 25643 Hz
100000 Hz 51219.5 Hz 51.405 kHz
200000 Hz 102439 Hz 102.05 kHz
500000 Hz 256097 Hz 257.15 kHz
V. REMARKS
REFERENCES
235
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
University of Applied Sciences, Informatics Department, Neidenburgerstr. 43, D-45877, Gelsenkirchen, Germany,
e-mail toma@informatik.fh.ge.de, shu@informatik.fh-ge.de, werner.neddermeyer@informatik.fh-ge.de
2
Facultatea de Electronic i Telecomunicaii, Departamentul de Masurari si Electronica Optica,
Bd. V. Prvan Nr. 2, 300223 Timioara, e-mail alimpie.ignea@etc.utt.ro
236
2, one can understand better the definition of this the stereo sensor is realized, we are able to measure
angle. 3D coordinates of different points. These coordinates
are measured with respect to the stereo sensor frame.
The stereo sensor frame is defined in the calibration
procedure and remains fixed to the sensor after the
calibration, reference [11].
The light projector can be also seen in figure 3, just
below the cameras. Our goal was to be able to create
on the tire surface, using light, some marks, which
could be further measured with the stereo sensor.
In figure 4, one can see the shape of the structure light
created by the light projector on the tire surface.
3
III. STEREO SENSOR AND LIGHT PROJECTOR
In this section we present some important aspects Fig. 4. Structured light projected on the tire surface
about the stereo sensor and the light projector we
used, in order to be able to describe our system, in There are two possibilities to make use of this
section 4, shortly and efficiently. structured light. First one is to use as marks the
The stereo sensor, which was used in our intersections between the light and different forms
measurement system, is the result of three years of existing on the tire surface. As one can see in figure 4,
work. One can understand all our steps concerning the in this category are included points 1 and 3. The
building, the calibration and the verification of the second one is to use as marks the crosses defined by
stereo sensor by consulting from the reference list the the structured light itself on the tire surface. To this
following: [7], [8], [9], [10] and [11]. With this sensor category belongs point 2.
one can measure the 3D position of a certain point
with an absolute accuracy of 0.1 mm. The distance IV. DESCRIPTION OF THE MEASUREMENT
between the point to be measured and the stereo SYSTEM
sensor can be adjusted between 150 mm and 500 mm
without any influence to the absolute accuracy. The A. Presentation of the system
measurement area is defined by a square with a side
of 100 mm. As we said at the beginning of this paper the goal is to
build a vision system for measuring Camber and Toe.
Having now the explanations presented in section 2
and 3, we are able to define a mathematical model in
order to reach this goal.
First of all we define a coordinate frame for the
wheel. We call this, the wheel frame. The origin of
this frame is situated in the middle of the tire. Axe z is
perpendicular to the tire so that the plane determined
by axes x and y is parallel to the tire. Axe x is
horizontal. One can see all these details in figure 7.
With these notations, Camber is determined by
measuring the rotation of the wheel frame around x
axe and Toe is given by the measured value of the
rotation of the wheel frame around y axe.
In this moment we know what we have to measure so
Fig. 3. Stereo sensor and light projector
the next problem, which must be solved, is how we
In figure 3, one can see the stereo sensor. It is have to measure. In order to explain way we built our
composed from two cameras mounted in a parallel measurement system in the way we did, it is necessary
configuration, reference [8]. After the calibration of to present here some details about the measurement
237
procedure. A detailed description of the measurement of the stereo sensors frames with respect to the
procedure will be presented in part C of this section. reference frame. We denoted with SR the reference
The angle information we need is obtained by frame situated in the middle of the plate and with SS1,
knowing the orientation of the tire plane (the plane SS2 and SS3 the stereo sensors frames. In figure 6, it is
defined by axes x and y) relative to a reference plane. drown only one sensor frame, because the situation is
So, the task is to measure this tire plane. It is known similar for the other two. The mathematical
that a plane is determined by at least three points, explanation, which follows for one sensor, will be
which are not all situated on the same line. Starting applied in the same way for the other two sensors.
from the plane definition we decided to use three With T(SR-SSi) we denoted the transformation from the
stereo sensors placed on a circle at equal relative reference frame to one sensor frame.
distances between them, as one can see in figure 5. The calibration plate, we used, has 121 points and we
know very precisely their position with respect to the
reference frame. We denote the coordinates of one
point from this plate with xR, yR and zR. The same
point will be measured with the stereo sensor and we
obtain the coordinates xSi, ySi, and zSi. According to the
reference [1], between these coordinates we have the
1200
1200 following relation:
1200
238
The second method is based on identifying marks of In figure 9, one can see the distribution of the errors
type noted with 2, as one can see in figure 4. The idea for Toe. They are situated between -0.38 and 4.86
is to use the light crosses for identifying which pixel minutes.
from the image obtain with one camera of the sensor
corresponds to a certain pixel from the image obtained 5.00
l e 0.00
e -3.000 -2.000 -1.000 0.000 1.000 2.000 3.000
Profile 1 y
-1.00
ang le in d eg r ee
239
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
A Correlation Analysis
of Measured CO-Concentration Signals
Sabin C. Ionel1
Abstract Ideal low-pass filtering can be used as a 600 cycles/hour. The temporal length covered by each
useful pre-processing method in order to eliminate the signal is 5h50' . The concentrations are measured in
measuring noise from signals representing CO-
concentration. Thereafter, the correlation coefficient [ mg / m3 N ]. Such signal pairs were measured, by
proves to be a suitable tool in a comparative analysis of day and by night, in several parks and crossroads of
the data. If the signals are measured simultaneously and Timioara city.
in the same location, using different instruments, the
Signals CDH1,2,3 Signals CDS1,2,3
correlation analysis allows a comparison of the 20 10
measuring systems. Applied to signals measured with a
single instrument, the correlation analysis can put into
evidence diurnal repeatability of CO-concentrations. 10 5
Keywords: CO-concentration, ideal filtering, correlation
coefficient. 0 0
1000 2000 3000 1000 2000 3000
20 20
I. INTRODUCTION
10 10
CO-concentration signals have a nonstationary,
random character. Consequently, certifying the 0 0
1000 2000 3000 1000 2000 3000
validity of CO-concentration data delivered by 40 40
different measuring devices can be a difficult task.
In this paper, a method allowing comparative 20 20
analysis of CO-concentrations measured with two
fully different instruments is presented. Thus, the 0 0
classic HORIBA instrument measures the CO- 1000 2000 3000 1000 2000 3000
concentrations locally, in a certain point, while the Sample Number Sample Number
Siemens-HAWK analyser delivers CO-concentrations Fig. 1. Typical signals measured with Horiba (CDH*) and with the
Siemens-Hawk (CDS*) instrument, respectively.
spatially averaged over a distance of 10 100m .
Obviously, the two instruments do not measure the All concentrations are measured in [ mg / m3 N ].
same quantity, but a comparison of the measured data
is still meaningful since both signals are representing
the pollution level in the same location [1], [2], [3]. III. IDEAL FILTERING
1
Facultatea de Electronic i Telecomunicaii, Departamentul Electronic Aplicat
Bd. V. Prvan Nr. 2, 300223 Timioara, e-mail: sabin.ionel@etc.utt.ro
240
presented in fig. 1 were filtered using f c = 2 cycles / measured with the Horiba and Siemens devices,
hour. The low-pass (LP) part of the measured signals respectively. Correlation coefficients of the LP part of
can be seen in fig. 2. The rapid fluctuations of the the air pollution signals represent also a good
initial signals, including measuring noise, were measure of the diurnal repetition of the CO-
elliminated through low-pass filtering. concentration level. This correlation is meaningful
since the pair of signals were measured on
Signals CDH*-low Signals CDS*-low consecutive days or nights, starting at the same hour.
10 10
8
5
IV. CORRELATION ANALYSIS
6
0
1000 2000 3000 1000 2000 3000 The measured CO-concentration samples can be seen
10 10 as realizations of some random variables. The
correlation coefficient of two random variables
8 5 x[n] and y[n] , defined by [5].
6 0 C
1000 2000 3000 1000 2000 3000 rxy = (2)
12 20 x y
where C and x , y are the covariance respectively
10 10
the variances of x[n] and y[n] , can be used as a
8 0 comparison tool between two or several data series.
1000 2000 3000 1000 2000 3000
Since -1 rxy 1 , through comparative analysis one
Sample Number Sample Number
can distinguish three different cases:
Fig. 2. LP signal pairs CDH*-low and CDS*-low,
in [ mg / m3 N ]. positive correlated data for 0.33 < rxy 1;
uncorrelated data for - 0.33 rxy 0.33 (3)
In the same manner, using ideal high-pass
filtering, one can obtain the high-pass (HP) part of the negative correlated for - 1 rxy 0.33;
measured signals. Thus, the signals CDH*-high and
CDS*-high represented in fig. 3, contain only For example, the correlation coefficients matrix for
frequencies greater then 2 cycles / hour. the six LP signals represented in fig. 2 are shown in
Table 1. Numbering the LP signals according to Table
Signals CDH*-high Signals CDS*-high
20 2 2, the correlation coefficient rij = r ji characterizes the
0 0 resemblance between signals i and j .
-20 -2
1000 2000 3000 1000 2000 3000 Table 1
20 5 1.0000 -0.0774 0.0448 -0.2851 -0.0144 -0.2620
-0.0774 1.0000 -0.4630 0.6906 -0.5757 0.7735
0 0
0.0448 -0.4630 1.0000 -0.1601 0.1942 -0.5015
-0.2851 0.6906 -0.1601 1.0000 -0.4156 0.5531
-20 -5
1000 2000 3000 1000 2000 3000 -0.0144 -0.5757 0.1942 -0.4156 1.0000 -0.6787
50 10 -0.2620 0.7735 -0.5015 0.5531 -0.6787 1.0000
0 0
Table 2
-50
1000 2000 3000
-10
1000 2000 3000
Signal number i, j LP Signal
Sample Number Sample Number 1 CDH1-low
2 CDS1-low
Fig. 3. HP signal pairs CDH*-high and CDS*-high
3 CDH2-low
in [ mg / m3 N ]. 4 CDS2-low
5 CDH3-low
The HP part of the signals can be analysed in 6 CDS3-low
order to evaluate the measuring noise, to optimise the
sampling period, the cut-off frequency etc. However, For an easier interpretation, the correlation matrix is
in the following, we are interested only in the LP part graphically represented in fig. 4. Obviously, the
of CO-concentration signals. Using the correlation signals measured with Horiba instrument are not
coefficients between the LP part of the signals correlated (absence of diurnal repetition). On the
measured under similar conditions (the same place contrary, the signals measured with the Siemens
and time interval for one signal pair; however, instruments are highly and positive correlated,
different days for different signal pairs), one can denoting a diurnal repetition of CO-concentration.
compare the tendencies of the CO-concentrations
241
Correlation Coefficients for the LP Signals, CDH*-low and CDS*-low Correlation Coefficients for the LP Signals, PDH*-low and PDS*-low
1 1
1 1
S S
D 0 DFig.0 6. A graphical representation of the correlation coefficients
C; P;
1 1 for the signals PD**-low
H H
D -1 D -1
C 1 2 3 4 5 6 P 1 2 3 4 5 6
1 1
2 2
S S
D 0 D 0
C; P;
2 2
H H
D -1 D -1
C P 1 2 3 4 5 6
1 2 3 4 5 6
1 1
3 3
S S
D 0 D 0
P;
C;
3
3 H
H D -1
D -1 P 1 2 3 4 5 6
C 1 2 3 4 5 6
Channels: 1,3,5 ---> PDH1,2,3-law; 2,4,6 ---> PDS1,2,3-law
Channels: 1,3,5 ---> CDH1,2,3-law; 2,4,6 ---> CDS1,2,3-law
Fig. 6. A graphical representation of the correlation coefficients
Fig. 4. A graphical representation of the correlation coefficients
for the signals PD**-low
for the signals CD**-low
242
REFERENCES
[1] * * * ROSE (Remote Optical Sensing Evaluation) EU Project
G6RD/CT/2000/00434, (2001-2004).
[2] Bisorca D., Ionel Ioana, Popescu F., Ionel S., Ungureanu C.,
Air Quality investigation by means of remote sensing, with
application to CO thermodynamic measurements in the city of
Timioara, 13-th International Conference on Thermal
Engineering and Thermogrametry (THERMO), Budapest, 18-20
June 2003, pp.274-279.
[3] Ionel Ioana, Ionel S., Beurteilung von Luftqualitt mittels
optischen Fernmessystemen, in Vergleich zu der ND-IR Methode,
VDI Optische Technologien, 4. Konferenz ber Optische
Analysenmesstechnik in Industrie und Umwelt, 7-8 Oktober 2004,
Dsseldorf.
[4].Hoffmann J., Ionel S., Signakonditionierung mit Wavelet-
Techniken, Horizonte, Nr.24, Juli, 2004, pp.16-19.
[5] A. Papoulis, Probability, Random Variables and Stochastic
Processes, Third Edition, McGraw-Hill, Inc., New York, 1991.
Acknowledgement
The research has been carried out in the frame of the
ROSE Project, contract NB GR6D-CT2000-00434,
founded by the European Commissions Competitive
and Sustainable Growth Program.
243
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Facultatea de Electronic i Telecomunicaii, Catedra de Telecomunicaii
Bd. Carol I nr.11, 700506, Iai, e-mail: smereuta@etc.tuiasi.ro
244
U A [k ] , U T [k ] , U C [k ] and U G [k ] provide a four- For example, we may apply a sliding window
dimensional representation of the frequency of length L to a sequence of length N, where N > L ,
spectrum of the character string. The quantity: resulting in a sequence of DFTs. Each of these
DFTs provides a localized measure of the frequency
2 2 2 2 content, and is an example of a location-dependent
S[k ] = U A [k ] + U T [k ] + U C [k ] + U G [k ] (5)
Fourier transform, known as the short-time Fourier
transform (STFT).
can be used as a measure of the total power spectral
content of the DNA character string at frequency k.
From (2) and (4) it follows that: II. DNA SPECTROGRAMS
2
x r [ n] = (2uT [n] u C [ n] u G [n])
3
6
x g [ n] = (u C [n] u G [n]) , (7)
3
1
x b [ n] = (3u A [n] uT [n] u C [n] u G [n])
3
B. Short Time Fourier Transform The vertical axis corresponds to the frequencies
k from 1 to 30, while the horizontal axis shows the
Instead of evaluating the DFT of a full-length relative nucleotide locations, starting from nucleotide
sequence, we have the option of evaluating the DFTs 858001. The genomic annotations establish that the
of several of its subsequences. This strategy makes DNA stretch contains three regions (C. elegans
sense particularly in the case of long sequences telomere-like hexamer repeats) at relative locations
consisting of several segments with different (953-1066), (1668-1727), and (1807-2028) [4]. These
characteristics. three regions are well depicted as bars of high-
intensity values corresponding to the particular
245
frequency k = 10 (because hexamers -period 6-
N
correspond to = 10 ). Furthermore, the frequencies
6
k = 6 (corresponding to a periodicity of 10) and its
multiples, appear to play a prominent role in the
whole region of the 4000 nucleotides.
For comparison purposes, Fig. 2 shows the
texture of a spectrogram coming from a sample of
totally random DNA, i.e., in which each type of
nucleotide appears with probability 0.25 and
independent of the other nucleotides.
REFERENCES
[1] R. Voss, Evolution of long-range fractal correlations and 1/f
noise in DNA base sequences, Physical Review Letters, vol.
68(25), p. 3805-3808, 1992.
[2] W. Li, T.G. Marr, K. Kaneko, Understanding long-range
correlations in DNA sequences, Physica D., vol. 75, p. 392-416,
1994.
Fig. 2. Color spectrogram of totally random DNA.
[3] B.D. Silverman, R. Linsker, A measure of DNA periodicity,
Journal of Theoretical Biology, vol. 118, p. 295-300, 1986.
[4] http://www.ncbi.nlm.nih.gov/entrez
III. PROTEIN CODING DNA REGIONS [5] J.-M. Claverie, Computational methods for the identification of
genes in vertebrate genomic sequences, Human Molecular
Genetics, vol. 6(10), p. 1735-1744, 1997.
Protein synthesis is governed by the genetic code [6] J.W. Fickett, Recognition of protein coding regions in DNA
which maps each of the 64 possible triplets (codons) sequences, Nucleic Acids Research, vol. 10, p. 5303-5318, 1982.
of DNA characters into one of the 20 possible amino [7] B. Alberts, D. Bray, A. Johnson, J. Lewis, M. Raff, K. Roberts,
P. Walter, Essential Cell Biology, New York, Garland Publishing,
acids (or into a punctuation mark, like a stop codon, 1998.
signaling termination of protein synthesis).
One of the most relevant and yet unsolved
problems in bioinformatics is to accurately and
automatically annotate sequences by identifying such
regions using gene prediction [5], [6]. It is clear [7]
that the total number of nucleotides in the protein
coding area of a gene will be a multiple of three.
N
The frequency k = is of particular
3
importance for protein coding DNA regions because it
corresponds to a period of three samples, equal to the
length of each codon (triplet of nucleotides).
We now show how frequency-domain analysis of
DNA sequences can be a powerful tool for
specifically identifying protein coding regions in
DNA sequences. In Fig. 3 we have plotted the
sequence S [k ] , as defined in (5), for a coding region
of length N = 1320 inside the genome of the baker's
yeast (formally known as S. cerevisiae),
demonstrating a peak at frequency k = 440 ( = N 3 ).
This peak confirms the genetic findings reported for
S. cerevisiae [4].
246
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
graduate degree of Doctor of Electronics Engineering, University Politehnica of Bucharest.
Address: 1, Polizu St. sector 1, 011061, Bucharest, Romania, phone 0722-539019, mail catalindumi@ yahoo.com
247
In the last fifteen years wavelet has been The study of EEGs has a long and fruitful history,
widely used in EEG analysis as much as epilepsy and and I knew that due to my time and equipment
Alzheimer diagnosis as sleep stage classification. constraints I could not tackle the general problem of
The main of this work is to extract EEG interpretation, so I restricted myself to a narrow
information from sleep EEG raw data about the scope just to get a feel for the problem. The question I
presence of K-complexes. We decided to work in the posed for myself was this: is it possible to devise a
time-frequency domain instead of either pure time computer program which will analyze an EEG signal
domain or pure frequency domain as previous works and detect a particular waveform pattern? Sleep in
in this field. In order to implement a time-frequency humans can be divided into two major categories:
analysis the Continuous Wavelet Transform will be Rapid Eye Movement (REM) sleep, and non- REM
employed because it has been probe to be an efficient (abbreviated NREM) sleep. REM sleep is
tool in extraction of transient characteristics from a characterized by coordinated, darting movements of
collection of raw data. Therefore, the problem the eyes as if scanning a scene, and is most correlated
statement of this work is to build and evaluate a K- with dreaming. NREM sleep on the other hand is
complex detection system using the wavelet transform distinguished by its lack of eye activity. NREM sleep
and, posteriorly, evaluate the algorithm performance is subdivided into four stages (Figure 1) with stage 1
trying to find out possible important faults that may being the lightest stage of sleep, sometimes
affect the system. experienced by nighttime drivers who suddenly
realize theyve been driving for a few seconds in the
1. Relevant Theory wrong lane, and stage 4 being the deepest stage of
This chapter will try to cover all the necessary sleep, characterized by total muscle paralysis and
theoretical background in order to give the reader a insensitivity to external stimuli. The different stages
better approach to the sleep stage classification and of sleep are distinguished from each other by the
time-frequency analysis using wavelet transform. It predominant EEG waveforms at a given time in the
begins with the basic concepts of sleep classification recording (Figure 2). Thus stage 1 is characterized by
and a brief description of the bioelectrical signal so-called theta waves (between 4 and 7.75 Hz), stage
involved, particularly the electroencephalogram 2 is composed of sleep spindles (14-15Hz) and K-
(EEG). Then, an explanation of the relevant EEG complexes, and stages 3 and 4 are composed of
waveforms is given. As a first step toward a process primarily delta activity (mainly 4Hz). In the waking
of EEG transient signal detection, the Fourier adult, alpha activity is characterized by waves
Transform and the Short Time Fourier Transform are between 8 and
explained. Finally, a review of the definition and basic 13Hz, and beta rhythm is characterized by waves
proprieties of the Continuous Wavelet Transform, greater than 15 Hz. I chose to study stage 2 of NREM
with the corresponding example and reason of why sleep, because the K complex can be easily
this Transform will be used for time-frequency distinguished from the spindle signals, and because
analysis are given. data for stage 2 was already available.
248
presence of one or more. This EEG waveform have a energy calculated and second before and after the K-
well-outlined negative sharp wave, immediately complex. The first criterion, about the frequency
followed by a positive component. Before and after a range, was settled using the LabView Based on
K-complex there is a period of low amplitude which literature [Mallat, 1998], [Kaiser, 1994], [Polikar,
is useful to distinguish the K-complex from Delta 1996] the most used wavelets for time-frequency
activity [Bankman 1992], [Didier 1994]. analysis have been Mexican Hat and Morlet wavelet.
Consequently, these two wavelet were chosen for
1.3 Time-Frequency Analysis further analysis. The Mexican hat function is the
Short-Time Fourier Transform (STFT) t2
The STFT is a time-frequency tool that consists second derivative of the Gaussian function e 2
and
of a Fourier transform with a sliding time window. is:
The time localization of frequency components is 1 t2
obtained by suitably pre-windowing the input signal. 2
(t )dt = 0
L2 ( R ) (2)
After determine which wavelet use, the next step
The discrete synthesis operation can be presented as was to settle the location in time of the K-complex
follows: within its respective 10 seconds epoch signal and its
respective time duration T. The K-complex interval T
1 t b
, f (a,b) = , f (a,b) = f (t) is the value which must be equal or greater that 0.5
*
CWT dt (3)
a a seconds and equal or lower than 1.5 seconds (see fig.
3).
where, l , k (a, b) = f , a ,b (t ) [Oppenheim and Posteriorly, the CWT was computed and from the
Shafer, 1989]. absolute values of the obtained coefficients matrix,
the highest value in amplitude and its respective
2. Methods and Implementation frequency value were looked assuming that this
2.1 Wavelet Selection frequency is the corresponding spectral component of
In order to choose the wavelet that will be the K-complex. The wavelet coefficients
employed in the K-complex detection algorithm, corresponding only to this spectral component will be
criteria based on how the wavelet spreads the signal called line of frequency. Consequently, using the
energy in time was developed. Thus, the chosen signal extracted from this line of frequency, as it is
criteria were based on two main points: depicted in the right illustration on figure 4.
1. The K-complex frequency range is from 0.5
Hz to 3.5 Hz.
2. A K-complex has to have a notorious
amplitude difference between the K-complex
energy and the energy registered and second Figure 3. K-complex time period T.
before the K-complex and one second after
it. This criterion tries to make the distinction
between a K-complex and the burst of delta
activity.
Based on these criteria, the best wavelet for the
detection algorithm will be that which give the biggest
difference the energy of the K-complex and the
249
corresponding to the highest absolute value in the
CWT matrix, will be the K-complex spectral
component. This was probed by comparing the
Fourier transform of the original signal with the
Fourier transform of the frequency line corresponding
to the maximum value found in the CWT matrix. As
is illustrated in figure 5 we can see that the CWT
pseudo-frequency line obtained, the energy per on
second was computed having a result of ten energy
value per epoch. To calculate the energy per one
second E, intervals of 200 samples were taken
(because the original signal is sampled at 200 Hz, 1
Figure 4. Continuous wavelet transform (absolute value) of the K-
second contain 200 samples) computing the energy
complex shown where the maximum amplitude correspond to the as:
pseudo frequency content of the K-complex.
200 2
2.2 Algorithm Design E = si , si = i location sample (7)
As K-complex are transient phenomena from i =1
EEG an algorithm will be developed in order to Using the K-complex database an attempt to find
achieve an automatic detection of these transient a common behavior of the energy in the presence of a
signals. The algorithm will be based on time- K-complex was tried.
frequency analysis searching the manner of how
quantifies the energy distribution of K-complex in the
time-frequency plane. To develop this algorithm the
CWT will be employed because this tool has
demonstrated a good performance in transient
detection and feature extraction in several previous
works [Bailey, 1998], [Schiff, 1994]. Employing
some of the same parameters used in the wavelet
selection process, the design of this K-complex
detection algorithm will be based on the Energy
Distribution of the K-complex in the time-frequency Figure 5. Left top: K-complex in a ten seconds epoch from
plane using the CWT. The wavelet employed in this EEG; Right top: CWT for scale 57.00 that correspond to the
algorithm will be the Mexican Hat wavelet function. pseudo-frequency of 0.88 Hz, it can be seen how the wavelet try to
As in the wavelet selection procedure, the assimilate the shape of the K-complex. From this signal the energy
value was computed; Left bottom: Fourier transform of the K-
frequency criterion was based on theory assuming that complex, the highest, amplitude correspond to 0.88 Hz; Right
a K-complex has a frequency range between 0.5 and bottom Fourier transform of the CWT pseudo-frequency line.
3.5 Hz. The pseudo-frequency range obtained was
splitted into 17 pseudo-frequency values which were
used to calculate the CWT. The scale and pseudo-
frequency range are in table 2. The number selected to
split the pseudo-frequency range was established
basically in order to obtain an acceptable resolution in
the time-frequency representation, without
compromises the time performance of the algorithm.
Figure 6. Energy distribution
250
the algorithm. The summarized results can be seen in in account the morphology, frequency content, time
Table 3 and in Figure 7 and 8. duration and power spectrum of the K-complex. From
this test, the most important conclusion we could
extract was that the wavelet capability in the detection
of K-complex has a strong dependence on the wavelet
waveform. Since the waveform of the wavelet has
probed to be an importance parameter for transient
Table 3. Results of the algorithm performance signal detection we would like to left this field open
for further analysis based on other different wavelet
depending on the application they will be used. The
way to use the CWT was a precise bandpass filter
we could obtain a very narrow frequency band or only
one pseudo-frequency line without big distortion in
the signal shape.
We achieved a very good separation of
frequencies in a range 0.5 3.5 Hz (17 frequency
lines) and very good signal suppression in the exterior
Figure 7. Pie chart plot that shows the percentage distribution from this frequency range. This feature of CWT was
of table 3 (discrimination between K-complexes and other transient implemented in both algorithms to detect K-complex
signal).
signals and was achieved a good results to detect
them. To know the real capacity of the algorithms to
detect K-complex, they were tested using a single
channel from eight hours EEG signal. From the
indices specificity, sensitivity and validity we
obtained very different results. The performance of
the algorithm based on the energy distribution was
relatively poor to make a good discrimination
between real K-complexes and false K-complexes.
Figure 8. Pie chart plot that describe the percentage The lack of enough criteria for K-complex detection
distribution achieved in the detection of real K-complexes only. could be the answer of this poor performance. During
our experience we realized that the decision regarding
The algorithm performance has a good capacity the detection of a K-complex may need to be
to exclude false K-complexes, but the main idea of corroboration by a single consideration that we did
obtaining a good K-complex detection algorithm, and not take in account. This consideration is concerning
at the same time, trying to minimize the number of to the vicinity of sleep spindles and K-complexes.
criteria used for the detection was to much restrictive Another interesting point to mention was the fact that
in the criteria number. detection of K-complexes was based on the research
of only real K-complexes since from the results
4. Conclusion obtained we realized that a more difficult task to carry
In this report we tried to cover the necessary out would be the develop of accurate criteria in order
theoretical and practical topics in order to develop to achieve a better recognition between Delta activity
different algorithms based on the Continuous Wavelet and K-complex. When looking in the false K-
Transform for K-complex detection on EEG signals. complexes detected as K-complexes we realized that
A description of the sleep stage classification, Fourier is possible to find real K-complexes in this set of
Transform, Short Time Fourier Transform and signals. Almost all these signals are out of stage two,
Continuous Wavelet Transform was given. The STFT and some of them just in the edge of a particular stage
and the CWT are two different tools with the same two. This makes to use think that we found real K-
aim: time-frequency analysis. Are their performance complexes in these signals, and a deeper investigation
are also different. Therefore, when time-frequency should be made on this field. One possible reason for
analysis is required, we should be very careful about this problem is that we only looked for K-complexes
the features of the signal to analyze, since for some in stage 2, since we did not find one single reference
signals the STFT could be more appropriate than the about the existence of K-complexes out of stage 2.
CWT and also in the other direction. For example in Another reason is a possible not proper stage
signals with no transient content and a limited band classification. Even when all signals in question were
width, the STFT has a good performance and the real K-complexes, the performance of the algorithm
computation time is not large, but when there are will not be good enough, therefore, a criterion for
transient signals involved, the CWT becomes make the difference between K-complexes and Delta
necessary, and the computation time increase. Two waves is highly necessary in order to improve the
wavelets function were tested with the purpose to validity of the algorithms.
obtain a quantitative description about how these two
different wavelets, Mexican hat and Morlet, are
capable to achieve a good K-complex detection taken
251
REFERENCES
252
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
RS232
DVM DVM (standard)
PC
Power supply
IEEE 488
253
A part of the block diagram is shown in fig. 3. The block diagrams of the main VI and of the subVIs
have been supplied with error handling blocks. These
blocks are not necessary if the system operates
properly, but if the application stops, it is very
difficult to find out why. The general structure of
such a block is presented in fig. 4.
254
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
255
the generated ac fluxes generated by these to be equal It is worth to make some observations
and of contrary signs. concerning relationship (5).
The measurement windings N1and N2 are 1. The non-symmetries between the winding lead
phased series connection; with such a configuration, to the fundamental and third order harmonic
the odd harmonics will be reciprocally canceled, and component appearance that can be eliminated using a
the even will add together, which is enhancing the band pass filter centered on the second harmonic.
signal to noise ratio, also doubling the probe 2. The sensitivity of the method is rising with the
sensitivity. number of turns belonging to the two inductors,
If we consider that the dependency between the exciting ca signal frequency and amplitude and
magnetic induction (for a saturated core, fig. 2) and obviously depends on the used core, taking into
the magnetic field intensity is: account the section S and the coefficient a3, that
characterizes the core saturation.
B = a1 H a 3 H 3 (2) 3. It is known from the electrical transformer
theory [4] that the core section can be saturated if it is
I determinated by the following formula:
Pa
3
S<K (6)
B = a1H a3 H f
256
sinusoidal, without distortions or dc component, sensitivity stay near constant because the loss in
because the presence of this dc component could lead
U2[mV] car frecv car frecv
to an additionally magnetization of the core and
therefore to the appearance of measurement errors.
The selective microvoltmeter V, has at the signal 200
input a band-pass filter with a rejection factor with
respect to the central frequency, that can be adjusted 150
two level, 25 dB respectively 40 dB, concerning the
desired selectivity extraction of the measured signal. 100
The input voltmeter signal can be measured directly
or passed through the filter that can be adjusted for the 50
desired frequency. As well, at the output of the f[Hz]
0
voltmeter, an oscilloscope can be connected for the
visualization of the measured signal. 0 500 1000 1500 2000 2500
In this schematic, the selective voltmeter Fig. 4. The dependency between the output voltage and
frequency is adjusted to a double frequency in respect the frequency.
to the generator frequency, choosing in this way the
magnetic core is grown
measurement of the second order harmonic
In fig. 5 is represented the output voltage
component.
dependency with respect to the excitation current.
The measurement schematic calibration has been
done using the Helmholtz inductor; the Frster probe f=1500 f=1500
has been placed in the middle of the Helmholtz 150
inductors, with an east-west orientation, so that the dc
terrestrial magnetic field that operates upon the probe
to be canceled. The Helmholz inductors are supplied 100
by a known dc current, a constant magnetic field will
be generated, and the voltmeter tuned on the
frequency second order harmonic component, will 50
indicate a proportional value of the measured I[mA]
magnetic field.
In the calibration experiment have been used: 0
The magnetic field generator (Helmholtz 0 50 100
inductors), which has a constant kB = 596 T/A
known with an error = 0,5%. Fig. 5. The dependency between the output voltage and
Digital multimeter type M3650D, accuracy class the excitation current.
0,5 1 digit
Selective nanovoltmeter Unipan, type 237, It is possible to observe that around 50 mA, the
accuracy class 1,5. curve is a third order characteristic and when the
After the calibration the dc magnetic field current is greater than 50mA the curve becomes
measurement constant has been obtained: saturated.
The explanation of this effect consists of the
kB I fact, that at high magnetic field, the core is saturated
kc = = 0,476 T/mV (8)
U and the hysterezis curve can be considered like a
breaked line. In this case, for a sinusoidal magnetic
The measurement uncertainty will be equal with [5]: field, the magnetic induction becomes trapezium.
For a trapezium signal (fig.6), the second
I k k
2 2
cl I k I cl U
2 harmonic component has the amplitude [6]:
B = B B + B A + B2 V =
U 100 3 U 100 3 U 100 3 (9)
sin (2tc / T0 )
A T0
U2 = (10)
k I 1 2 2 tc
= B 2B + cl A2 + clV2 = 0,005 T/mV
U 100 3
Where, A is the amplitude of the trapezium signal, T0
That corresponds to an approximately deviation the period of the signal and tc rising time.
of 1%. If we have tcT0, the sinusoidal function is
In fig. 4 is represented the dependency between approximated with its argument and the expression
the output voltage and the frequency for a supply (10) becomes:
current and an external constant induction, revealing
the sensitivity dependency with respect to the square A
root frequency, but if the frequency is large, the U2 = (11)
257
We can see that the second harmonic IV. CONCLUSIONS
amplitude is independent of period and rising time.
The measurement of a dc magnetic field,
A especially for low intensities, based on the Frster
probe is a very sensitive method.
To obtain a large sensitivity and small errors it
is necessary that the core of the probe is saturated.
tc T0 t We observe that the frequency of the
sinusoidal magnetic field has to be of some kHz,
because at larger frequencies, the loss of the
Fig. 6. Trapezium signal magnetic material is bigger.
Now, we suppose that we have two Referencies
complementary trapezium signals, one of them
positive and the second one negative and a delay [1] ***, EN 62052-11: General requirements test
of T0/2 between them, the first with an amplitude of and test conditions for electricity metering
A+ and the second with an amplitude of A-. equipment (AC),
In this case, the amplitude of the second [2] ***, 1999/519/EC, Recomandarea Consiliului
harmonic component will be: din 12 iulie 1999 privind reducerea expunerii
publicului la cmpuri electromagnetice(0-300
2 GHz), www.acero.ro,
U 2 = (12)
[3] Wiener, U., Masurari electrice, vol II, Ed.
Tehnica, 1969,
As a conclusion it can be said that the [4] Richter, R., Transformatorul electric, Ed.
sinusoidal field applied to the Frster probe is large Tehnic, Bucureti, 1960,
enough and the output voltage is dependently only [5] Ignea, A., Stoiciu, D., Msurri electronice,
on the continuous component. senzori i traductoare, Ed. Politehnica, Timioara,
2003,
[6] Ignea, A., Introducere n compatibilitatea
electromagnetic, Ed. de Vest, Timioara, 1998.
258
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
ORANGE S.A., Departamentul Transmisiuni, Bucure~ti, e-mail ana.rosu@orange.ro
2
Facultatea de Electronica si Telecomunicatii, Dept. Telecomunicatii, Bucuresti, e-mail teodor.petrescu@munde.ro
259
1* ( x ) = A1+ (e 1 x + e 1x ), x<0, (1) The method can be used for the arbitrary potential
barrier structures, so that we consider a
2* ( x ) = A2+ e 2 x , x > 0, (2) symmetrical double barrier structure with a
where rectangular quantum well (RDB), as shown in
figure 2:
i = j (2mi / = 2 )( E Vi ) , (3)
mi and Vi (i=1,2) are the propagation constant,
GaAs AlAs GaAs AlAs GaAs
the effective mass and the potential,
respectively, for the ith region; is the wave 1 2 3 4 5
amplitude reflection coefficient.
Obs.: For to obtained a better accurate of the E2 Vb
calculation, the effective mass mi is assumed to be the E1
different in both materials, depending of the Al
concentration, X, from the structure. a b c d
Based on this observation, mi is calculated thru:
mb = (0.0667 + 0.083 X ) m0 (4) l2 l3 l4
m w = 0.0667 m0 (5)
Fig. 2. Schematics of the resonant tunneling effect and resonant
where
transmission obtained for E = En, when electrons tunnel
m0 is the effective mass of the free electron; resonantly into the nth bound state of the well
mw - the effective mass of the electron in the
quantum well; The characteristic impedance for the rightmost section
mb - the effective mass of the electron in the serves as the load and, because the region 5 is
potential barrier. theoretically infinite, Z l = Z 0,5 . Having the load
Vi function of Al concentration, X, is: impedance at point d, the input impedance Zc at the
V b = 0.57 (1.155 X + 0.37 X 2 ) . (6) point c is calculated using relation (13), with
Using equation (1) we obtain relation (7): Z 0 = Z 0, 4 and l = l 4 . With Zc as the load at point c,
(
i ( x) = j (= / mi ) (d i /d x) = Ai + e i x e i x w ) the input impedance at point b is computed and this
ith process is repeated until the point a is reached. The
Z 0 i = j mi / i = . (8) reflection coefficient is calculated using the equation
(11) with Za as the load impedance and Z 0,1 as the
Let us now write the equation for voltage (U) and
current (I) used in the transmission lines with characteristic impedance. The transmission coefficient
generalized distributed impedance. These are: T (E ) is given by:
U ( x ) = I + Z 0 (e x + e + x ) , (9) 2
T ( E ) = 1 ( E ) . (14)
+ x + x
I ( x ) = I (e e ), (10)
We have used this method in the PHP script from the
where simulator of paper, for to obtained the better RDB
Zl Z 0 quantum structures which lead at electronic devices
= (11)
Zl + Z 0 with good transmission performances. Also, for to
realize a complete study, we have utilized the Smith
is the voltage reflection coefficient, Zl and Z 0 are the
chart by using an interactive Java applet, in witch you
load and the characteristic impedances of the can set parameters, obtained from PHP calculation
transmission line, respectively. If we compare the script. The theory and the mode of work of this
equations for i* and i and the corresponding interactive Java applet will be presented in the next
expressions for U and I for a transmission line we see subsection.
that these are analogous equations. Thus, we can
regard Z0 as the characteristic impedance of a region. B. The description of the interactive Java
Also, the ratio of i* and i (analogous to the ratio of applet
the voltage and current), will define, at any plane x,
the quantum mechanical wave impedance In any transmission system, a source sends energy to a
Z i ( x) = * i ( x) / i ( x) (12) load, such as an antenna. Ideally, we design the
transmission network such that the characteristic
Thus, the input value of impedance, Zi = Z (l ) , impedances of the source, the transmission line and
the load are all identical. Unfortunately, many real-
may be expressed in terms of the load impedance
world situations prevent the match from being perfect.
Zl = Z (0) [1] as: For example, we might want an antenna (the load) to
be useful over a broad range of frequencies. But the
Z l cosh( l ) + Z 0 sinh( l )
Zi = Z0 (13) characteristic impedance of an antenna is unlikely to
Z 0 cosh( l ) + Z l sinh( l )
260
stay constant with frequency, especially if the
frequency span is great.
When the transmission line impedance does not match
that of the load, part of the transmitted waveform is
reflected back towards the source. The reflected wave,
which varies in phase and magnitude, adds to the
incident (transmitted) wave and the sum is called a
Standing Wave. The reflected wave causes the
amplitude to vary as a function of position along the
transmission line. The Standing Wave Ratio (SWR),
which is the ratio between the maximum and
minimum amplitudes of the total waveform, will in
this case be greater than one. If there is no reflected
wave, i.e., if the impedance match is perfect, the F
ig. 4. Drawing the VSWR circle.
amplitude of the total waveform (incident plus
reflected wave) will be the constant, regardless of In general, only the horizontal line (diameter) is
where we measure it along the transmission line. The labeled with (normalized) resistance values and only
result is a SWR of 1. SWR = 1 indicates maximum the unit (outer) circle is labeled with (normalized)
power transfer to the load. SWR can be inferred by reactance values. To read the desired values, it is
measuring the reflection coefficient of the circuit. The necessary to follow the appropriate circle of constant
network analyzer is a tool that enables us to do just resistance to the diameter line, and to follow the
that. If we know the reflection coefficient, we can appropriate arc of constant reactance to the unit circle.
determine the characteristic impedance of the load by Hit Play again, and the program will display the
using a Smith Chart. The Smith Chart has circles of constant-resistance circle and the constant-reactance
constant resistance and arcs of constant reactance. The arc for you (fig. 5). The actual values are calculated
relationship between reflection coefficient and and shown at the left side of the screen.
characteristic impedance is shown in the diagram (fig.
3).
261
These coefficients are obtained from PHP script. The
last step is the graphic representation of the
transmission coefficient function of the incident
electron energy (using another PHP script) and the
representation of the reflection coefficient on the
Smith Chart by utilizing the interactive JAVA applet
witch receives the dates from the first PHP script.
By realize more experiments with this simulator we
can anticipate the RDB geometry witch leads to
electronic devices with wishing transmission
performances.
Thus, we can observe that, if we chose an asymmetric
RDB structure (l2 = 20,
l3 = 100, l4 = 35 and X = 0.45 =>
Vb = 338.96 meV) it will obtained a devices with Fig.7. Reflection coefficient for the asymetric RDB structure
small transmission coefficient (T < 0.7
- fig. 6) and with grate reflection ( Ox axe of the
Smith Chart fig.7).
262
Table 1
Thickness Height of Resonant Transmission
of the potential frequency coefficient
potential barriers [THz]
barriers [meV]
[]
10 193.12 5.803 0.99999
10 297.08 30.225 0.99971
10 313.7 74.718 0.99994
10 305.37 30.467 0.99999
10 322.08 74.718 0.99999
10 330.5 74.959 0.99995
10 338.96 74.959 0.99999
10 103.49 4.594 0.99937
10 147.55 28.049 0.99996
10 232.26 29.258 0.99987
11 305.37 73.75 0.99999
12 338.96 73.75 0.99999
13 297.08 30.225 0.99983
13 313.7 72.783 0.99986
13 103.49 4.836 0.99892
14 170.15 27.807 0.99987
15 272.46 29.742 0.99977 Fig.10. The dependences T (Vb), Frez (Vb), T (Gb), Frez (Gb), for
15 313.7 71.574 0.99994 symmetric RDB structure with Gw = 100
15 322.08 71.816 0.99999
15 103.49 5.078 0.99963 IV. CONCLUSIONS
16 297.08 30.225 0.99998
In conclusion, making more experiments with this
17 338.96 30.951 0.99981 simulator you can predict the geometry of the RDB
17 216.48 28.533 0.99997 witch leads of the electronic devices with wishing
17 82.038 4.8361 0.99731 transmission performances. In this way, our present
18 103.49 24.906 0.99823 study clear proved that, for to obtained the circuits
20 272.46 29.742 0.99884 with good transmission power coefficient, it is
20 103.49 24.906 0.99999 necessary to use a symmetric RDB structure with the
thickness of the potential barriers equal with them and
23 272.46 29.742 0.99273
the thickness of the quantum well much grater of
24 103.49 24.18 0.9998 them. Thus, the transmission performances of the
25 193.12 27.565 0.99635 electronic devices can by easy modified by the
26 103.49 23.938 0.99993 changing the geometry size of the RDB structures.
27 118.01 5.8033 0.99998 Much more, for a easy identification of the geometry
28 216.48 28.049 0.99892 structure with leads at the wishing performances,
29 103.49 23.697 0.9995 using MySQL, in this paper we created a Data Base,
in witch it stockade the values of the transmission
32 67.944 4.836 0.99611 coefficients and resonant frequencies of the
32 67.944 4.836 0.99611 submilimetric filters obtained with symmetric RDB
34 103.49 22.971 0.99866 structures, for different height and thickness of the
35 67.944 4.836 0.99982 potential barriers. This Data Base can be anytime up
38 67.944 4.836 0.99761 gradated and interrogated using PHP scripts.
41 103.49 22.487 0.9995 So, our paper present an original and modern
simulator, realized with PHP scripts and interactive
JAVA applet (two language programs very much used
in Internet, in the last time) witch permit us to
determinate the transmission performances of the
electronic devices obtained from RDB quantum
structures.
263
REFERENCES
264
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Academia Tehnica Militara, George Cosbuc 81-83, 75275, Bucharest, Romania, e-mail: simions@mta.ro
265
1.40
1.35
1.30
1.25 W=140m
1.20
Cs/Co
1.15
1.10
W=90m
1.05
Fig. 1. The MEMS cross section.
1.00
0 5 10 15 20 25 30 35 40
that g z has a very sharp variation for the applied Voltage [V]
voltage around the pull-down value. Then, for values (b)
of the applied voltage close to the pull-down voltage,
the position of the bridge is not stable. In this case, for Fig. 2. The distance between the bridge and the dielectric layer, (a)
and the normalized capacitance (b), versus the applied voltage.
MEMS applications such as phase shifters, in order to
control the phase shift value, the applied voltage must
The MEMS is not a fast device. If a sinusoidal signal
not exceed ~25V, for W = 140 m and ~30V, for
is applied on the MEMS bridge, arround a DC
W = 90 m. It is not difficult to show that the
voltage, then the bridge position cannot follows the
normalized equivalent MEMS capacitance, is: fast variation of the signal. Therefore, if the signal
frequency is high enough, the bridge position is
Cs 1 possible to be given by the DC voltage only. To prove
= ,
Co
1
z this, for the MEMS geometrical dimensions given
zo above, the simulation results concerning the MEMS
where: behaviour to the RF signal are shown in Fig. 3, for
W = 140 m and in Fig. 4, for W = 90 m, where the
t wW amplitude of the sinusoidal signal, Vampl , is equal to
z o = g + d and Co = C s V =0 0
r t 10V and the DC voltages are equal to 25V and 15V,
g+ d respectively. The RF frequency is 10KHz, 30KHz and
r
200KHz. From these figures, it is observed that for a
microwave signal applied on the bridge, the MEMS
is the MEMS capacitance for V = 0 (the minimum
equivalent capacitance depends on the DC voltage
capacitance value). The normalized capacitance, only, even if this is a nonlinear capacitance, due to the
C s / C o , versus z / z o is shown in Fig. 2 b, for the small time response of this kind of device. As a result,
same data as for Fig. 2 a. Taking into account the even for high power microwave signal, the MEMS
maximum value for the applied voltage given above, may be seen as a linear device. This observation is
from Fig. 2 b, it is drawn the conclusion that the important for a phase shifter based on a MEMSs
maximum capacitance ratio is ~1.2, a similar array, because the phase shift will depend only by the
estimation as in [8]. apllied DC voltage, without other small signal
constrains.
266
of the MEMS series inductance may be neglected
1
( Ls << ). Also, because the losses due to the
C s
CPW transmission lines and due to the MEMSs are
small, so Rl 0 , Gl 0 and R s 0 . Therefore,
taking into account that the circuit consists of n
identical cells, the Bragg frequency and the input
impedance of the circuit are given by:
1 Ll
fb = and Z in ,
Ll (C + Cl )
s
C s + Cl
(a) (a)
(b)
(b)
267
applied for a maximum operating frequencies, f max ,
of 17GHz for the first circuit and 25GHz for the
second one, in the both cases f b / f max =0.15. For
the two circuits, they were obtained =11.46deg,
w = 50 m, s = 125m, while the CPW lengths and
the CPW equivalent inductance and capacitance,
between two consecutive MEMSs, are l =223m,
Ll = 0.141nH, Cl = 25fF, for the first circuit and
l = 150 m, Ll = 0.094nH, Cl = 16.7fF, for the
second one. Imposing W = 140 m and W = 90 m
(geometrical dimensions for the MEMSs analyzed in
section II), for the both circuit, wb = 60 m. Also,
Cs =31.3fF and Cs =20.9fF, corresponding to
VDC =25V and 15V, respectively. Taking into
account the others geometrical and electrical MEMS
parameters (given in section II), it is obtained
(c) Ls = 50.75 pH. The MEMS series resistance depends
on the frequency. For these values and n = 38, the
Fig. 4. The distance between the bridge and the dielectric layer, maximum phase shift is 385deg, for the both
for the MEMS operating at VDC = 15 V, Vampl = 10 V, W=90m phase shifters, computed at f max = 10GHz for the
and f = 10 KHz (a), 30 KHz (b) and 200 KHz (c). first circuit and f max = 15GHz, for the second one.
The two phase shifters have been numerically
decrease the CPW electrical length, (so to decrease analyzed, in order to obtain de magnitude and the
the length of the circuit), but, on the other hand, it phase for S 21 and also the magnitude for S11 (the
must be lower in order to reduce the CPW losses,
CPW losses effect are included). The results are
( Z c must be chosen to minimize ). For shown in Fig. 5, wherefrom, for the phase shift, it is
Z in = 50 , f max / f b = 0.15 , r = 11.9 (the observed a good agreement between the analytical
substrate dielectric constant for silicon), t = 1 m (the and simulated values (see Fig.5 a). Also, a return loss
thickness of the CPW gold metallization), better than 30dB (see Fig. 5b), has been obtained for
w + 2 s = 300 m ( w and s are the width and the slot the both circuits, while the insertion loss is smaller
of the CPW between MEMSs), is minimized if than 0.5dB for the first circuit and 0.4dB for the
second one, up to the maximum operating
Z c is 60 - 80 [7]. In this paper, Z c =75.
frequencies. The second phase shifter is lossless
Combining the expression for f b , Z in and Z c , they compare to the first one because is shorter (the two
are obtained, Ll = Z in / ( f b ) , (
Cl = Z in / f b Z c2 ) circuits have the same number of cells).
[ ]
and C s = (Z c / Z in ) 1 C l . The electrical length
2
The phase shift introduced by the circuit may be also
of the CPW transmission line which connects two analyzed in the time domain. Fig. 6 shows the results
consecutive MEMSs may be obtained with for the second phase shifters and two values of the DC
= 2f max Cl Ll and then, by using a commercial voltage, VDC , 10V and 30V. The frequency of the
software, the CPW length, l , it is easily computed. input signal is 15GHz. The delay time is different in
the two cases because the MEMS equivalent
The distances g and t d are usually imposed by the
capacitance depends on the DC voltage. From this
technological constraints. Assuming that the values figure, the time delay introduced by the circuit is
for g , t d and r ,d are known, the formula for w W ~74ps for VDC =10V and ~79ps for VDC =30V, these
g r ,d + t d results being in good agreement (5% error) with those
is: wW = C o . For n cells, the maximum
o r ,d obtained by using formula n Ll (Cs + Cl ) . The 5ps
phase shift introduced by the circuit, at the maximum difference between the two time delay values at
operating frequency, may be computed as 15GHz means a phase shift difference of 27deg. In
2f max n Ll ( C s + Cl ) . This formula for may some applications which ask for a larger value of the
phase shift difference, the number of cells must be
be used to compute the number of cells, n , if the increased or/and to minimize the influence of the
value for is imposed. CPW equivalent capacitance.
Two phase shifters have been designed, the first one In order to evaluate the broadband characteristic of
for W = 140m and the second one for W = 90m. this phase shifter, the next simulation has been
The design formulas introduced above have been performed for a square pulse applied to the input of
268
the second circuit. For V DC =10V and pulse width of
10ps, 30ps, 50ps, the output waveforms presented in
Fig. 7 show that if the pulse width decreases, for a
given Bragg frequency, the dispersive characteristic
of the circuit contributes to the pulse shape distorsion.
This is because for short pulses, the condition
f b / f max =0.15, which assures the less dispersive
character of the circuit, is not true for the spectral
components having important amplitudes above
f max 25GHz. From Fig. 7, it is observed that for
pulse width greater than 30ps, the dispersive character
of the circuit may be neglected. For an input pulse
width of 50ps, the delay time introduced by the circuit (b)
is the same as for a sinusoidal wave (see Fig. 6), for
V DC =10V as well as for V DC =30V (see Fig. 8),
showing the broadband characteristic of the circuit.
IV. CONCLUSIONS
(a)
(a)
(b)
Fig. 6. The input (a) and the output (b) waveforms for the second
phase shifter, for VDC =10V and VDC =30V (the input signal
has the frequency equal to 15GHz).
269
Fig. 7. The output waveforms for pulses of different widths, applied
to the input of the second phase shifter ( VDC =10V).
270
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
I i = h(u , v ) I o ( , )dd
Abstract-Our goal in this paper is to make a study about 2
optical coherent and incoherent systems frequency (4)
response. We begin from the definition of an optical
system then we define transfer function for coherent and AMPLITUDE TRANSFER FUNCTIONS TYPICAL
incoherent systems. We find the response of this system
to a step indices stimulus, then we generalize for an
FOR COHERENT CASE.
image and finally we make a comparison between two
different systems. We define input and output frequency spectrum
Go ( f x , f y )
INTRODUCTION (5)
= U o ( x, y ) exp{ j 2 ( f x u + f y v)}dvdu
Generally speaking optical systems can be seen as a
Gi ( f x , f y )
black box with an input and an output, at the input we
(6)
have object plane and at the output image plane which = U i ( x, y ) exp{ j 2 ( f x u + f y v)}dvdu
is obtained from convolution between object image
and transfer function of optical systems (the black box We define transfer function
can contain one ore more optical elements). Here we Hi ( fx , f y )
will have a diffraction element. (7)
U i (u , v) = h(u ; v )U o ( , )d d (1) = h(u , v) exp{ j 2 ( f x u + f y v)}dvdu
We apply convolution to (3) and we obtain:
h(u , v; , ) optical system impulse response Gi ( f x , f y ) = H ( f x , f y )Go ( f x , f y ) (8)
but system response to optical impulse is Fourier This is the relation between image and object plane in
transform (Fraunhoffer diffraction) of diffraction frequency.
element aperture. But transfer function is Fourier Transform of impulse
h(u , v) = response system. Then we will have;
A 2 (2) H ( fx , f y ) =
z i P( x, y) exp{ j z (ux + vy )}dxdy
A 2
P ( x, y ) exp{ j (ux + vy )}dxdy}
i
F{
Next we will try to calculate h(u,v) for coherent and zi zi
incoherent case.
What do we understand by coherent and incoherent = ( A zi ) P( zi f x , zi f y ) (9)
illumination? If we put A zi =1 then
Coherent illumination is made by lasers.
Incoherent illumination is made by diffuse source like H ( f x , f y ) = P( zi f x , zi f y ) (10)
sun or gaze lamp. As a conclusion for coherent illumination Amplitude
For coherent illumination the system is described by Transfer Function is the aperture trough which the
amplitude convolution equation. light passes and the diffraction is made. For a square
U i (u , v) = h(u ; v )U o ( , )d d (3) aperture we will have:
x y
For incoherent illumination the system is described by P(x,y)= rect ( )rect ( )
intensity convolution equation.
2w 2w
1
Facultatea de Elcetronica si Telecomunicati, Departamentul Bazele
Electronicii, str. Baritu nr.26, Cluj Napoca , tflorin@bel.utcluj.ro
271
Next we will study optical coherent systems response
The transfer function will be: to a step indices stimulus for a square aperture. We
zi f x zi f x will study 2D and 3D case Fig. 1
H ( f x , f y ) = rect ( )rect ( )
2w 2w
Fig. 1. first line present 2D case; second line present 3D case; first column present step indices stimulus;
second line present square aperture; third line present response
=
I o (u , v) exp[ j 2 ( f x u + f y v)]dudv
We apply convolution to (4) and we obtain:
(13)
=
I (u, v) exp[ j 2 ( f u + f
i x y v)]dudv function on their definition imply function h (optical
system impulse response) so there is a relation
I (u, v)dudv
i
between this two function. Optical transfer function is
the normalized autocorrelation of amplitude transfer
(12) function.
272
z i f x z i f x
P( x + 2 , y + 2
)
H( fx, fy )=
P( x, y)dxdy
z i f x z i f x
P( x 2 , y 2
)
dxdy (15)
P( x, y)dxdy
This represent area of superposition of two apertures
z i f x z i f x
of the same shape one at , the other at
2 2
z i f x z i f x Fig. 2 H( f x , f y ) for incoherent case.
, divided at total area of the two
2 2 When this area is normalized with total area 4w2
apertures as in Fig. 2 have
Mathematical relation of common area is: fx fy
(2 w z f x )(2 w z f y ) H ( f x , f y ) = tri ( )tri( )
A( f x , f y ) = i i 2 f0 2 f0
0 w
f0 = cutoff frequency for coherent case
2w 2w zi
fx ; fy
zi zi Next we will study optical incoherent systems
response to a step indices stimulus for a triangular
aperture. We will study 2D and 3D case Fig. 3.
Fig. 3. first line present 2D case; second line present 3D case; first column present step indices stimulus; second
line present triangle aperture; third line present response .
273
CONCLUSION oscillation at the end (Gibb phenomenon) and a phase
difference from the axe of symmetry. Optical
Comparing response in Fig. 1 and Fig. 3 we see a incoherent systems do not have oscillation at the end
great difference between optical coherent and and phase difference. Finally to have a clear view we
incoherent system response for a step indices. So for will put an image instead of step indices stimulus and
optical coherent systems we have a response with well see how acts in the two cases. Fig.4
Fig. 4. first line present coherent case; second line present incoherent case; first column present input image;
second line present square aperture and triangle aperture; third line present output image.
REFERENCES
1. A. Papoulis: System and transform with application in
optics McGraw Hill, New York 1968.
2. J.W.Goodman: Introduction to Fourier optics McGraw
Hill, New York 1968.
3. E. Hecht: A. Jajac: Optics Addison Wesley, Reading,
MA 1974.
4. M. Born, E. Wolf: Principle of Optics Pergamon New
York 1964.
5. Keigo Izuka Engineering Optics Springer, Verlang
Berlin, Hailderberg 1985
6. J.D.Gaskill Linear System Fourier transform and Optics
John Wiley and Sons New York 1978
7. L.Boas Mathematical Methods In the Physical Science
Wiley New York 1966
8. B.E.A Saleh, M.C.Teich Fundamentals of Photonics
John Wiley New York, NY, 1991
274
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Facultatea de Electronica si Comunicatii, Catedra Bazele
Electronicii, str.Baritu nr.26, Cluj Napoca, tflorin@bel.utcluj.ro
D 0 is a specific nonnegative quantity and D(u,v) is
275
the distance from H(u,v) to the origin of frequency LPF have the next characteristics: high frequency
plane. The point of transition between 1 to 0 is called reduction but they make blur, noise reduction because
the cutoff frequency; here cutoff frequency is D 0 . noise is installed at high frequency.
Next we will study how LPF acts for different cutoff
We have the LPF 2D representation in Fig. 3 and 3D frequency for an image with 256x256 dimensions we
representation in Fig. 4
will use the next cutoff dimension: 128x128, 80x80,
and 40x40. The result is illustrated Fig. 5
Fig. 5 The graphics present the same image filtered with filters having different cutoff; as the cutoff frequency
decrease we will have more blur on the image and oscillation. Oscillations around image are caused by
transitory regime passing from 1 to 0.
276
A HPF has inverse properties like LPF, and is defined representation in Fig. 7
like HPF=1-LPF which we can write like this: HPF has the next characteristics: block low pass
frequency and let to pass high frequency. As effect it
0 if D (u , v) D0 enhances the noise, reduces basic characteristics of an
H(u,v)= image and is used for edge detections. Next we will
1 if D(u , v) > D0 study how HPF acts for different cutoff frequency for
an image with 256x256 dimensions we will use the
D 0 is cutoff frequency
next cutoff dimension: 128x128, 80x80, and 40x40.
The result is illustrated in Fig. 8
Fig. 8 the graphics present the same image filtered with filters having different cutoff; as the cutoff frequency
increase we will have edge detection more pronounced on the image and oscillation. Oscillations around image
are caused by transitory regime passing from 1 to 0.
BPF high pass filter
277
with different cutoff frequency BPF=H1(u,v)-H2(u,v) It acts like a median filter and has the characteristics
1 if D(u , v) D01 that we can reduce noise without blurring image and
H(u,v)= we can cheep edge characteristics. Next we will study
0 if D (u , v) > D01 how BPF acts for different cutoff frequency for an
image with 256x256 dimensions we will use three
windows with dimensions: 40x220, 100x160, 40x220
1 if D(u , v) D02 for first LPF and other three different windows with
H(u,v)=
0 if D(u , v) > D02 dimensions 80x180, 100x160, 125x135 for second
filter as in Fig. 11
with the 2D representation in Fig. 9 and 3D
representation in Fig. 10
Fig. 11 the graphics present the same image filtered with filters having different dimensions. It keeps specific
characteristics of LPF and HPF and is a mediator filter between this two filter
278
CONCLUSION
REFERENCES
279
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract-Our goal in this paper is to develop a calculus Knowing the impulse system response we can
algorithm beginning from Optical Fourier Transform calculate in space with convolution relation between
and then to generalize to optical linear shift invariant
systems. We begin from the premise that an optical
input and output signals:
system can be seen like a black box with an input and an
output. Our system property analyze can be made in
y( u1 , u 2 ) = x(v , v
1 2 )h(u1 v1 ; u 2 v 2 )dv1 dv 2
space with the help of convolution and in frequency by (4)
multiplying specters, these operation are If in (4) we apply Fourier Transform then we will
complementarily. Optical system also can be seen from have Y( 1 , 2 ) =X( 1 , 2 ) H( 1 , 2 ) were:
the perspective of components which is made. So in our
example we choose components which can made Optical X( 1 , 2 ) =
x(u , u
Fourier Transform: a square aperture and a convergent
lens. 1 2 ) exp[ j (1u1 + 2 u 2 )]du1 du 2
input spectrum (5)
INTRODUCTION
H( 1 , 2 ) =
We consider an optical system which is characterized
by input signal x( u1 , u 2 ) and output signal
x(u , u
1 2 ) exp[ j (1u1 + 2 u 2 )]du1 du 2
frequency response (6)
y( u1 , u 2 ) . We will find a relation between input In conclusion convolution in space is:
signal and output signal by the means of system y( u1 , u 2 ) =f( u1 , u 2 ) *g( u1 , u 2 ) (7)
impulse response and transfer function. In optics we And equivalently in spatial frequency by the means of
utilize linear shift invariant systems. By linear optical Fourier transform we have:
Y( 1 , 2 ) =X( 1 , 2 ) H( 1 , 2 )
systems we understand that the system has a linear
(8)
relation between input signal and output signals by the
relation: In optical linear systems (7) and (8) can be
generalized:
F[a x1 ( u1 , u 2 ) +b x 2 ( u1 , u 2 ) ]=
y( u1 , u 2 ) =f( u1 , u 2 ) *g( u1 , u 2 ) **j( u1 , u 2 ) (9)
aF[a x1 ( u1 , u 2 ) ]+bF[ x 2 ( u1 , u 2 ) ] (1)
Y( 1 , 2 ) =X( 1 , 2 ) H( 1 , 2 ) ...J( 1 , 2 )
If at the input we apply a 2D optic impulse
(10)
(u1 1 , u 2 2 ) then output image will be In conformity with result in eq.9 and eq.10 we will
called system response to optical impulse with the make a simulation of an optical system made by a
notation h (u1 , u 2 ; 1 2 ) so we will have the square aperture and a convergent lens.
relation:
FIRST EXAMPLE
F[ (u1 1, u 2 2 ) ]=h (u1 , u 2 ; 1 2 ) (2)
The system is linear shift invariant if system response In table 1 we will se mathematical relation in parallel
to optical impulse is independent of impulse input for the two situations. We have as input signal a
position or more clearly an input impulse variation harmonically plane wave. Results are presented in
produce the same variation at the output. Fig. 1
F[ (u1 1, u 2 2 ) ]= h(u1 1, u 2 2 ) (3)
1
Facultatea de Elcetronica si Telecomunicati, Departamentul Bazele
Electronicii, str. Baritu nr.26, Cluj Napoca , tflorin@bel.utcluj.ro
280
Spatial domain analyses Frequency domain analyses
Input signal: A(x,y)=sin(x)sin(y) ( ) 2 [ (1 1 ) (1 2 )]
A( 1 , 2 ) =
[ ( 2 2 ) ( 2 + 2 )]
Aperture transfer B(x,y)=rect(x)rect(y) B( 1 , 2 ) =sinc( 1 )sinc( 2 )
function
convolution C(x,y)=A(x,y)*B(x,y) C( 1 , 2 ) =A( 1 , 2 ) B( 1 , 2 )
Lens transfer x2 y2 1 2 22
function D(x,y)=pi/2exp(- ) D( 1 , 2 ) =exp(- )
2 2 2 2
convolution E(x,y)=C(x,y)*D(x,y) E( 1 , 2 ) =C( 1 , 2 ) D( 1 , 2 )
Table 1
Fig. 1 The first column present analyses in space domain for the case considered in Table 1, on the x and y axes
we have spatial dimensions. The second column present frequency domain analyses on x and y axes we have
281
spatial frequency dimension.
SECOND EXAMPLE This means that we have at the input 16 harmonics
plan waves and equivalent spectrum consisting in 64
We will make a more complex analysis with the help points. The rest of the calculus is in conformity with
of an optical binary signal, made by the next optical situation in Table 1
Fourier string:
A(x,y)=[sin(x)+sin(3x)/3+sin(5x)/5+sin(7x)/7]
[sin(y)+sin(y)/3+sin(5y)/5+sin(7y)/7]
Fig. 2 The first column present analyses in space domain, on the x and y axes we have spatial dimensions. The
second column present frequency domain analyses on x and y axes we have spatial frequency dimension.
282
Fig. 3 The first column present analyses in space domain, on the x and y axes we have spatial dimensions. The
second column present frequency domain analyses on x and y axes we have spatial frequency dimension.
283
REFERENCES
284
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
I. INTRODUCTION
The 3-D Finite-Difference Time-Domain (FDTD) A field component such as E xn +1 / 2 (i, j , k ) at time step
method [1] was employed due to its accuracy and n+1/2 is not calculated directly in FDTD scheme, but
versatility. The method selects from all four Maxwell it may be approximated by the arithmetic average
equations the two curl equations and solves them between the components calculated at time steps n
numerically by finite differences in time-domain. The and n+1. The final reult is an explicit expression of
grid points where the electric and magnetic fields are the component at time step n+1 when the components
calculated alternate in space forming the FDTD cell as at previous time steps are already known.
(i, j, k )t t n +1 2 1 n +1 2 1 n +1 2 1 n +1 2 1
1 H z i, j + , k H z i, j , k H y i, j , k + H y i , j , k
2 (i, j, k ) (i, j, k ) 2 2 2 2
E
n +1
(i, j, k ) = E x (i, j, k ) + (i, j, k )t
n
(2)
x
(i, j, k )t y z
1+ 1+
2 (i, j, k ) 2 (i, j, k )
1
National Institute of Materials Physics, Atomistilor 105 bis, Bucharest-Mgurele, e-mail gbanciu@infim.ro
2
Politehnica University of Bucharest, Faculty of Electronics Telecommunications and Information Technology, Iuliu Maniu 1-3,
061071, Sector 6, Bucharest, e-mail george.lojewski@munde.pub.ro
285
The devices were designed on a 0.635 mm height progression ratio. The absorption profile of the non-
Rogers substrate with 10.8 dielectric constant. The homogeneous PML was established after empirical
FDTD grid was x = y = 0.15 mm and investigations. Satisfactory results were found for
z = 0.127 mm. The time step was chosen of 0 = 1 mS, g = 2.3 for a layer thickness of 13 cell.
t = 0.27 ps, in order to satisfy the stability criterion
Courant-Friedrichs-Levy. The field excitation was
chosen as a Gaussian pulse IV. FDTD SIGNAL PROCESSING
The accuracy of the FDTD method depends very where the functions j 0,k (t ) = 2 j0 / 2 (2 j0 t k ) and
,
much on the quality of the absorbing boundary
j ,k (t ) = 2 2 j0 / 2 (2 j t k ) , (t) is the mother
j /2
conditions. The perfectly matched layer (PML) idea
is to use a lossy material to match the incident waves. wavelet and (t ) is the scaling function.
However, for an isotropic lossy material, the match The wavelet analysis rejects the noise caused by the
occurs only at normally incidence, therefore such a finite precision; it also clears the effects of the FDTD
material has only a limited application for absorbing signals of too high frequency, which cannot be
boundary condition [2]. A PML should match waves accurately computed. An accurate FDTD simulation
of arbitrary incidence, polarization, and frequency. of the propagation of signals of too high frequency,
The Brengers innovation consits in a derivation of a greater than ~30 GHz in our case, is unnecessary and
split-field formulation of Maxwells equations; it would involve a very fine mesh and extensive
namely, each vector field component is split into two computer resources.
orthogonal components. The PML technique The time step required by the stability criterion is too
decomposes each field projections in two, and the small and a conventional Fast Fourier Transform
wave incident to PML is attenuated via electric () would provide a low-resolution signal in frequency
and magnetic (*) conductivity. The extremely small domain. Therefore, the FDTD signal is not only de-
reflection is satisfied by the impedance matching noised but also de-sampled. The first nmin points of the
condition perpendicularly to the PML layer. FDTD signal, are used as a training set to the signal
In order to analyze the new microstrip devices, a Non- estimation technique. In Fig. 2, nmin = 100 and the
Homogeneous Perfectly Matched Layer (NH-PML) estimated signal y4(t) is a close approximation of the
was developed. For such a non-homogeneous FDTD signal y3(t). The total FDTD number of
boundary, the values of electric () and magnetic (*) iteration time was reduced to a fifth of the initial
conductivities should satisfy the impedance matching number.
condition.
0.05
*
j j = const. , (4)
Signals y3(n) and y4(n)
2f 2f y3
y4
where is the magnetic permeability, is the electric
permittivity and f is the frequency. The conductivities
increase with the depth into the PML. The efficiency 0
of the NH-PML increases with its thickness. For a
given thickness best profile for the conductivity is the
geometric series profile such as
0 r ( g 1) l
l* = g (5) -0.05
2 ln g 100 200 300 400 500 600
Time step n
where l = 0, 1, 2 represents the grid point index
Fig. 2. The comparison between the the FDTD signal y3(t) and the
inside the NH-PML, 0 is the electrical conductivity estimated signal y4(t)
at the PML interface and g is the geometric
286
V. FILTER DESIGN Two-pole filter
0
Resonators and filters were developed on the substrate
with the characteristics mentioned above. In the first -10
0.07
0.06
Coupling coeficient
0.05
0.04
0.03
(a)
(b)
0.02
0.01
The proposed designs allow direct couplings with the -10 S21 measured
external circuit. The variation of the external quality
S11 and S21 (dB)
S21 FDTD
S21 narrow band model
factor Qext with the coupling line position was also -20
S11 measured
analyzed by employing the FDTD method.
-30
Several two-pole filters were developed by employing
different couplings between resonators. For the two- -40
pole filters in Fig. 4 the coupling between resonators
is the mixed. Each resonator in the insert of Fig. 3 has -50
same substrate. Fig. 6. Measured versus simulated responses of the filters in Fig. 5.
Despite the fact that filter shown in the insert (b) of
Fig. 3 exhibits two transmission nulls at each side of An example of a cross-coupled filter developed with
the pass-band, it is very hard to control the positions new compact planar resonators is shown in Fig. 5.
of these nulls. For an improved filter response, it was Each resonator is 10.64 mm in size. This filter is
already shown that a full control of these nulls can be characterized by a coupling matrix
provided by filters with cross-coupled resonators [3].
The extra negative couplings result in a sharper filter 0 0.0360 0 0.0081
roll-off for an increased filter selectivity. The newly 0.0360 0 0.0259 0 , (7)
developed square planar resonator can be effectively M =
employed for cross-coupled filter design [4, 5]. 0 0.0259 0 0.0360
0.0081 0 0.0360 0
287
and an external quality factor Qext= 17.86. As it is Two-pole and four-pole cross-coupled filters were
shown in Fig. 6, the measured response of the four- developed for 900 MHz wireless systems such as
pole cross-coupled filter follows closely the simulated GSM and GPRS for the 900 MHz. However, the same
response. design technique can be applied for devices of the 3G
communications standards. The devices are cost-
effective; they do not require via-holes and any
VI. CONCLUSIONS
additional lumped elements.
A 3-D FDTD method was developed in order to
REFERENCES
accurately design compact resonators and filters. A
non-homogeneous perfectly matched layer (NH-PML) [1] A. Taflove (Editor), Advances in Computational
with a geometric series profile for the electric and Electrodynamics The Finite Difference Time-Domain method,
magnetic absorption was developed as absorbing Norwood, MA, Artech House, 1998
[2] J.-P. Brenger, Perfectly Matched Layer for the FDTD
boundary conditions. The FDTD signal processing by Solution of Wave-Structure Interaction Problem, IEEE Trans.
using wavelet packets and signal estimation Antennas and Propag., vol. AP-44, 1996, pp. 110-117
techniques resulted in a reduction by up to five times [3] G. Lojewski, Computer Aided Design of Some Pass-Band
of the computation time. Microstrip Compact Filters of Quasi-Elliptic Type (in Romanian),
Telecomunicaii, No. 2, 2003, pp. 43-51
The FDTD method was successfully applied to small- [4] M. G. Banciu, G. Lojewski, T. Petrescu, A. Ioachim,
size filter design. The proposed resonators occupy L. Nedelcu, R. Cacoveanu, N. Militaru, D. Brinaru,
down to 32% of the surface area of a folded half- D. Ghimpeteanu, Small-Size Filters with Improved Characteristics
wavelength resonator designed on the same substrate for Wireless Communications, Proceedings of the 35th Scientific
METRA Symposium, Bucharest, May 27-28, 2004, pp. 437-440
for the same frequency. [5] M. G. Banciu, G. Lojewski, Proceedings of the International
Conference, COMMUNICATIONS 2004, Bucharest, June 3-5,
2004, p. 341-346, 2004
288
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1 2 3
Croatian Telecommunication d.o.o. Mostar, Transmission Department
Kneza Branimira bb, 88000 Mostar, Bosnia and Herzegovina
e-mail: ivan.rados@ht.ba, tanja.sunaric@ht.ba, pero.turalija@ht.ba
289
III. FAILURE ANALYSIS AND AVAILABILITY - failures which simultaneously break all fibers in the cable,
CALCULATION unless the network is ring, the operator has to repair all
fibers in the cable with no regard to the existence of some
In order to calculate the availability of optical cables, data on protection mechanisms (optical modules). According to the
failures and time to repair of optical cables are used. The collected data in 100% of causes happens the break of all
collected data referred to the period from spring 1994 to May fibers, either caused by digging or by fire or by vehicle.
01, 1999 and period from May 01, 1999 to May 01, 2001. All
cables are installed sub-surface, in polyethylene pipes and above Generally, two measures to repair are being used [2]:
them a warning tape was installed as a supplementary way of
protection. Fibers of all cables are standard, single-mode with a - temporary repair time,
diameter of 9/125m for the use of 1310 nm and 1550 nm - permanent repair time.
wavelengths.
According to the collected data the main cause of most Temporary repair time is the time needed for service restoration
failures is outside interference (86,48%), where digging after the failure.
participates in 72.97% of cases [3]. The vehicle owing to the This time to repair includes:
improper depth of installed cable causes two failures (5.40%) - time needed to report the failure to the maintenance team
and the fire causes three failures (8.10%). Four failures and their arrival to the telecommunication center,
(10.80%) are the consequence of the planned works by the HT - time needed for the preparation of splicing material (cable)
d.o.o. Mostar. These failures lasted relatively a short time and vehicles,
because of previously well-done preparations. - way to the failure location,
- laying of the new piece of cable (if needed) and its splicing,
From the point of view of the optical failures availability we - final measurements.
distinguish [4]:
- failures which break individual fibers in the cable, if there is
no automatic protection, the operator has manually to direct
the traffic to the correct fibers or to repair the faulty once,
According to the experience, the time needed to report to the failure location is short, the influence of the arrival time to the
maintenance team and their arrival to the telecommunication failure location in relation to the total time to repair is
center is less than an hour. As there is no data on exact distance insignificant.
from the maintenance centers to the failure location, the time The greatest influence on the time to repair has the type of
needed to arrive to the failure location is no special analyzes. failure, for example: difficult access to damaged cable, necessity
But when the distance from the maintenance team center to the for digging and installing the new piece of cable, cable capacity
290
and splicing of fibers of different manufacturers, unfavorable - repair of all fibers in the cable.
weather conditions.
On the area covered by telecommunication network of HPT
If only two fibers on one cable are actively used, regard to the d.o.o. Mostar actively exist more transmission systems via the
availability there are exist two cases: single mode cable, so the time needed to repair all
- repair of active fibers wherewith the system becomes
available,
fibers in the cable (or in more cables) is taken for the calculation Unavailability of optical cables per km is obtained as a
of the time to repair. product of failure rate per km of cable and the mean time to
Permanent repair time includes, in addition, final storage of repair as shown in Table 3.
new splicing closures, final construction works and final For the unavailability calculation of the optical cable besides
protection of a new cable segment. the mean time to repair and failures rate per km it is necessary to
In this article the temporary repair time is used as mean time to know the failure rates of the splices on the fiber and failure rate
repair for the availability calculation owing to its influence on of connectors on the optical distribution frame. Data on failure
availability. rates of splices (30 FIT) and connectors (100 FIT) are taken
Until May 01, 1999 HT d.o.o. Mostar had only one team with from the [5] and [6]. The total length of the cable stage consists
three members for maintains of optical cables. Two members of delivered cable from factory with an avarage length of 4 km.
out of three do the splicing and the one do finally According to this length the number of splices is calculated as:
measurements. The maintenance team had only one splicer and
one OTDR, what practically mean that they be able do splicing Length of cable / 4.
only on one side of optical cable (if is necessity for installing the
new piece of cable). It was 60 % of the exact documentation of The number of connectors on the optical distribution frame is 2
installing optical cables. Average time for repair during this for the average stage length which is used in this calculation.
period was 15,70 hours. The result analysis in Table 4 shows that unavailability
From May 01, 1999 HT d.o.o. Mostar took precautions with aim increase almost linearly to the cable length and depends on
to decrease average time to repair of optical cable (detail failure rate and mean time to repair of optical cables.
explanation in chapter 4). The results of this precaution were For SDH network HT d.o.o. Mostar, mean time to failure
decrease average time to repair on 13,43 hours. (MTTF) is obtained as follows:
That we on the best way see influence to decrease average time
to repair on availability of optical cables in this chapter we are 1 h
presume that no decrease its that mean we are use for MTTFnetwork =
460.61 1251 failures
calculation average time to repair 15,70 hours.
= 1735 h 73 days
Table 3 Failure rate (), unavailability (U) and mean down time (MDT) calculated for optical cables
Table 4 Failure rate (-total), unavailability (U) and mean down time (MDT) calculated for different optical link lengths
Table 5 Failure rate (-total), unavailability (U) and mean down time (MDT) calculated for different optical link lengths
(n=32 failures)
291
IV. SUGGESTIONS FOR AVAILABILITY IMPROVEMENT cable characteristics caused by the outside interference. For the
preventive failure protection against outside interference (long-
From the availability expression (1) can be seen that the term exploitation) it will be necessary to install the surveillance
availability depends on the mean time between failures and the system in order to foresee a failure. Costs for installation of such
mean time to repair of optical cables. Availability improvement a system would be slight compared to with failure losses on the
can be obtained by increasing the mean time to failures and cable.
decreasing the mean time to repair of optical cables [4].
B. Decreasing of mean time to repair
A. Increasing of mean time to failure
From the unavailability expression can be seen that
The increase of mean time to failures, relatively, the decrease availability depends on the mean time to repair. In order to
of the number of failures can be achieved by preventive decrease the MTTR it is essential to have a maintenance plan
protection of optical cables against digging and by using the which should contain the following components [4]:
surveillance system for preventive maintenance. As most - exact documentation,
failures on the optical cables are caused by digging it is - maintenance team,
necessary to attract special attention to it. Although most - training,
countries have laws for preventive protection of underground - equipment,
cables there are still unsatisfactorily defined punishments (fees) - plan of action,
for their infringements (digging without previous consent). The - practice,
law must have to most rigid punishments (invitation prior to - continued process of improvement.
digging). While digging belong to the category of
instantaneous breaks, the others belong to the category Exact documentation on optical cables is one of the most
preventive because they are caused by complete loss of cable significant components for diminishing the MTTR. It includes:
characteristics owing to the outside interference. cable traces, number of failures in cable, fibers attenuation,
splicing points, cable lengths, trace marking and outer-metal
In our country still no have law about preventive protection of shield condition. Additional 34% of documentation made during
underground cables, what is a possible conclude on the base of period from May 01, 1999 to May 01, 2001, that mean than till
number of failure during both monitoring periods, most of the now we made documentation for 94% of installing optical cable.
failures were caused digging without previous consent and Also it procured one mobile computer for frequent modification
transgressor has no adequate punishment. That we are show and bring up to date. Only two persons have access to that base
influence mean time to failure on availability we will suppose and they are responsible for its processing. Besides the exact
that we already have the legal regulations regarding the data base for diminishing the MTTR it is necessary to exactly
protection of underground cables during the failure monitoring know where are the tools and material needed to repair, as well
period and that, through change of that law, the mean time to as the key to the entrance of the building. Plan of action contains
failures increased from 73 to 84 days, which means that the instructions on who is calling whom and when, as well as the
number of failures decreased from 37 to 32, representing a numbers of fixed and mobile telephones. Now, maintenance
decrease of about 13 %, and that the mean time to repair team has seven members: five on the failure location and two at
remained same, i.e. 15.70 hours. As the failure rate is just terminals (one in each). Four members out of five do the
proportional to the number of failures, so, decreasing the splicing (two teams of two members) and the fifth have the
number of failures also decreases the failure rate by 13.5 % or to radio connections with the members at terminals.
398.37 FIT. Maintenance staff has to know to use the splicers and measuring
equipment. Owing to the ever-improving measuring and
As seen in the Table 5, the decrease of the number of failures connecting equipment for different types of cables, regular
resulted in the availability improvement, relatively, the decrease training of the maintenance team is very important because each
of MDT, for example, for d=20 km from 78.89 to 68.64 improvement which leads to diminishing the time to repair
min/year or by 12.99%. Surveillance system for preventive increases the availability of optical cables. In HT d.o.o. Mostar
maintenance can foresee the possible failure location using the they have training two times a year at least in order to acquire
metal protective layer on the cable as a sensor. Surveillance new knowledge. Training in the field is more purposeful
system alarms when the entirety of the outer sheath or the measure than the classroom teaching. Well planned and sudden
splicing point is being breaked, indicating that potential failure exercise is the best way of the emergency staff training (one per
should be removed. As optical cables installed within the HPT year). The aim of each exercise is to achieve better results each
d.o.o. Mostar have been exploited a short time (the first one time.
about seven years), there were no deterioration as yet of the
292
Quantity and kind of equipment depend on geographical improved considerably, as shown in Table 6. In the concrete, the
spreading, network size, and the number of skilled staff. If MTTR decrease of 14.45 % results in the availability
network is too large and geographically spread there must be improvement of 14.46 % or to, the decrease MDT from 78.89 to
more maintenance teams. As network of HT d.o.o. Mostar no 67.48 min/year.
geographically spread, for the now is sufficient one maintenance Every greatest availability improvement would be achieved
team of optical cable. Our maintenance team has one by the simultaneous decrease of the number of failures (32) and
reflectometer (OTDR) for measuring at 1310 and 1550 nm, two the decrease of the mean time to repair of the cables (13.43
splicers with tools (cutter, air, screwdrivers...), optical power hours), as shown in following table. In the concrete, mean down
meter, voltmeter and the car. time of failure is decrease for 25.57%, or to decrease from 78.89
Using above mentioned suggestions, the MTTR of the cable is to 58.71 min/year.
obtained to 13.43 hours. The availability would also be
Table 6 Failure rate (-total), unavailability (U) and mean down time (MDT) calculated for different optical link lengths
(MTTR=13.43 h)
Table 7 Failure rate (-total), unavailability (U) and mean down time (MDT) calculated for different optical link lengths
(n=32 failures, MTTR=13.43 h)
V. CONCLUSION REFERENCES
Data on failures and time to repair, which are analyzed in this [1] C. Coltro, Evolution of Transport Network Architectures, Alcatel
article, refer to the 7 years exploitation of optical cables within Telecommunication Review, 1st Quarter 1997, pp.10-18,
the HT d.o.o. Mostar transmission network. The analysis show [2] I. Jurdana and B. Mikac, An Availability Analysis of Optical Cables,
that the most frequent cause of the optical cables break is Proceedings WAON`98, pp. 153-160,
digging (72.97%) and regardless to the break cause there has [3] T. H. Victor, Update on Interim Results of Fiber Optic System Field
Failure Analysis, Bellcore, New Jersey,
been breaks of all fibers in the cable. The analysis of temporary
[4] I. Rados, Availability analysis of Synchronous Digital Hierarchy
repair time shows that it mostly depends on the type of failure Network, Master thesis, University of Zagreb, 2000, (Croatian).
and cable capacity. Availability improvement of optical cables
[5] D. Gardan, Availability analysis of the fibre optic local loop,
can be achieved by increasing of mean time to failure, Optical Access Networks, EFOC & N, pp. 28-32, 1994.
relatively, decreasing the number of failures as the most [6] P. N. Woolnough and N. E. Andersen, FITL System Recommendations
frequent case of break, and decreasing the mean time to repair. Final Report, Deliverable FIRST 1.0.11, pp. 1-42, 1 May 1996.
The law on underground cables protection and monitoring
system for preventive maintenance would be the cause of
decreasing the number of failures. The mean time to repair cable
is decreased considerably by using the plan of maintenance. An
approximate unavailability can be used to evaluate availability
of different structures.
293
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
294
h ( k ) (t ; S , R ) =
III. OPTIMIZATIONS
= h ( 0) (t ; S , {r , n, FOV , dr 2 }) h ( k 1) (t ; {r , n,1}, R )
S The enormous volume of calculus renders
(4) spatial numerical algorithms slow even on powerful
machines so any amelioration in speed is more than
But, following John Barrys method, if we welcome. We have considered the two possible
consider the walls discrete, made up of N indivisible solutions proposed in [1]: a direct implementation of
elementary surfaces the function becomes: (4) and an implementation using look up tables. The
first is more suitable for inferior k where high
resolution is needed. The second works better for the
superior reflection orders where a high spatial
h ( k ) (t ; S , R ) h (0) (t; S , i ) h (k 1) (t; i , R) (5) resolution (high N) is needed.
Certain aspects are to be considered during the
implementation. Computing the vectors absolute
and more explicitly: value not by using the Euclidian distance should do
for example calculating the distance between a
source-receiver pair, which is needed to calculate the
n +1 delay.
h( k ) (t; S , R)
2 But we can gain more in terms of speed if wee
consider physical reality carefully. Light undergoing k
N
r cosn ( ) cos( )
( d2
rect( / FOV ) (6) bounces between point A and point B will always
travel slower than light undergoing k-1 bounces under
i =1
h ( k 1)
(t d / c;{r, n,1}, R)A) the same points A and B. Therefore the first non-zero
sample that appears when calculating hk will appear
after the first non-zero sample hk-1. It therefore
where A is the area of the receiver and the becomes useless to calculate a certain number of
rectangular function is defined above. points as shown in figure 2.
h(0)
h(1)
h(2)
h(3)
295
A variant of Phongs law was also used by Yang
PHONG
and Lu to calculate infrared illumination diagrams in a
room, [4].
Echelle linaire
We have used a similar variant to compute
impulse response. The Lamberts law (1) is replaced
by Phongs law:
1
RS (, ) = [rd cos() (1 rd ) cosm ( )]
[ / 2, / 2] (8)
[ / 2, / 2]
where rd represents a coefficient indicating the
percentage of diffuse radiation and m is a parameter s *10
-11
REFERENCES
[1] J.R. BARRY, J.M. Khan & all Simulations of Multipath
Impulse response for Indoor Optical Channels
[2] Y. A. Alquad, Mohsen Kaverhad, MIMO Characterisation
of Indoor Wireless Optical Link Using a Diffuse-Transmission
Configuration IEEE Transactions on Communications vol
-11
51 No 9 September 2003
s *10
[3] Bui Tong Phong Illumination for Computer Generated
a) Pictures Communications ACM 1975
[4] H. Yang, C. Lu Infrared Wireless LAN using Multiple
Optical Sources IEEE proceedings 2000
296
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
I. INTRODUCTION
297
x 2 y 2 z 2 x 2 ( R) y 2 ( R) z 2 ( R)
R= = = = = = =
y (t ) = x(t ) Rh(t ) + n(t ) (1) x1 y1 z1 x1 ( R ) y1 ( R) z1 ( R )
Where R is the detector responsivity, n (t) is signal- x 2 (S ) y 2 (S ) z 2 (S )
independent additive noise and h(t) is the channel = = = (3)
impulse response, [4]. x1 ( S ) y1 ( S ) z1 ( S )
The channel can be described in terms of frequency A2
R2 = ( 4)
response: A1
h(t )e
j 2ft
H( f ) = dt (2) 1.5
which is the Fourier transform of h(t).
h0[arbitrary units]
1
Channel response can be determined experimentally
using three methods:
1. The direct method
0.5
A short impulse is emitted and the response is
room
RMC
measured directly. Although this method is
successfully employed in low frequency applications,
for example in acoustics. It is hard to use in infrared. 00 1 2 3 4 5 6 7 8 9 10
2. The spread spectrum method t[ns]
A pseudo-random binary sequence is transmitted and Fig.3. The impulse response for the direct link
the received signal is cross-correlated with the input
sequence to yield the impulse response. For the multi path components a correction is needed:
3. The frequency sweep method R scales the time axes and the function amplitude is
Signals are transmitted by sweeping through a finite
scaled by R . In Fig.4 we show the high orders
range of frequencies and the complex channel
response captured for the entire frequency range. The impulse response for R=2.
broader the range of frequencies considered, the
closer is the approximation of the measurements to 0.1
0.07
RMC
h1[arbitrary units]
0.06
III. MESUREMENT OF THE IMPULSE
0.05
RESPONSE FOR A REDUCED SIZE MODEL
0.04
0.03
Here we show that the impulse response function of
0.02
an actual room can be estimated once we measured
the impulse response function of a reduced model
0.01
room
0
0 10 20 30 40 50 60 70
channel (RMC). This is our most important theoretical
result. It offers a high degree of flexibility to our t/R[normed
experiment allowing us to test real environment time]
conditions in a laboratory. It allows us to anticipate
technological advances in infrared 1550nm receivers Fig.4. The impulse response for the reflected links
by performing experiments with existent low-cost
photodiodes. The measurement resolution
In order to estimate the response of a room we must If we reduce the room dimensions we reduce both
build our model keeping certain proportions. For this spread distance and time, so we need a wider band for
we used a link simulator of impulse response for the measurement system.
diffuse indoor optical wireless channels, [2]. c
f max = (5)
Let x1, y1, z1 be the dimmensions of the room, x1(R), d
y1(R), z1(R), and x1(S), y1(S), z1(S) the coordonates of d c
the reciever and of the source respectivlly. We will x = = (6)
3 3 f max
have x2, y2, z2, x2(R), y2(R), z2(R), x2(S), y2(S), z2(S)
for the model. A1 is the size of the reciever used in We find the relationship between the spread distance
the real room and A2 is the size of the reciever used (d) and the band with of the measurement system
in the model. (fmax) in equation (5), where c is the speed of light.
We have empirically shown that if the relations (3) Consequently the spatial resolution (x) of the model
and (4) exist between the model and the real room is given by equation (6).
than the impulse response given by the direct optical For example in a 1mx1mx0.60m model for 1ns
path care of identical amplitude and differ only by a temporal resolution one needs 30cm of spread
time offset, as shown in Fig. 3. distance, which corresponds to a spatial resolution of
298
17.3 cm and needs 1 GHz of the band with of the V. RESULTATS
measurement system. Other values are shown in Table
1.It's obviously that a smaller temporal of resolution is Here we present the measured impulse response for a
demanded than a greater measurement frequency is direct link in a 1mx1mx0.60m reduced model room.
needed. The distance between the emitter and the receiver is
62 cm. The measurement frequency is 2 GHz which
Table 1 assures a temporal resolution of 0.5 ns and a spatial
Measurement Temporal Spatial resolution resolution of 8.66 cm (see table 1).
frequency resolution [cm]
0.1
[GHz] [ns]
1 1 17.3
0.08
2 0.5 8.66
h0 max 0.102544
2.5 0.75 13 0.06
10 0.1 1.73
0.04
tr= 2.003ns
IV. EXPERIMENTAL SET-UP 0.02
299
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
where:
- damping factor of the loop;
n natural pulsation of the undamped system.
With the substitution s = jm in (2) we have the
response of the reference oscillator at a noise
component of frequency m:
0 i ( j m ) 2j n m + 2n
= N (3)
0 ( j m ) 2m + 2n m + 2n
The transfer function of the reference oscillator noise
is:
1
Academia Fortelor Terestre Sibiu, Catedra de Stiinte tehnice
Bd. Revolutiei 1-3, e-mail: eteodoru@actrus.ro
300
II. EQUATIONS 2
T T
m 2 = n + 2 n +1
If some elements (phase detector, frequency dividers) 2 2
2
of the system are digitally ones, a digital approach is T (9)
possible, so that the noise transfer function is more m 1 = 2 + n
2
useful with z transform: 2
m 0 = n T 2 n T
+1
0 i ( z ) 2
n z + n 1z + n 0 2 2
= 2 (4)
i ( z ) m 2 z 2 + m 1 z + m 0
respectively:
The equation with finite differences for the input
i(n) and output 0i(n) samples is: T
2
n 2 = n T + n N
2
1
0i ( n ) = [ n 2 i ( n ) + n 1 i ( n 1) + ( n T ) N
2
m2 n 1 = (10)
1 2
+ n 0 i ( n 2 )] [ m 1 0 i ( n 1) + (5) T
2
m2 n 0 = n + n N
+ m 0 0i ( n 2) 2
With bilinear transform method the transfer function The relations (9) and (10) allow the obtaining of the
in z is: noise levels of the output versus the reference noise,
using Simulink models.
0i (z)
=
i (z)
2 2 2
a2 + a1 T + a0 T z2 + 2a2 + a0 T z + a2 a1 T + a0 T
2 4 2 2 4 III. DIAGRAMS
=
2 2 2
b2 + b1 T + b0 T z2 + 2b2 + b0 T z + b2 b1 T + b0 T We realize two situations:
2 4 2 2 4
a) = 0,707; N = 1000; n T = 2 0,1 (11)
(6) generating the following sets of coefficients m and n:
m 2 = 1,045
where:
m1 = 1,998 (12)
m = 0,957
a 2 = 0 0
a1 = 2n N (7) n 2 = 45,409
2
a 0 = n n 1 = 1,974 (13)
n = 43,445
and: 0
b2 = 1
b1 = 2n (8) The Simulink model is given in fig. 2, and the
2 diagrams of the input/output noise in fig. 3,4, in time
b0 = n and frequency domains, respectively.
The noise is simulated with a gaussian noise generator
T is the period of the reference signal. and we search to observe the response of the system
Making the identification between the relations (6),(7) regarding this signal. So we use both a spectral
and (8), the coefficients m and n are: analyser and a oscilloscope to the output.
301
Periodogram
10^(-5)
Generator zgomot
gaussian Periodogram 1
Constant Product1 Periodogram1
1 1
Scope1
0.956 z z
-43.435
Product3 Unit Delay Unit Delay1
1/m2
Product4
n0
Sum
1.974
n1 Product2
Periodogram
45.409
Periodogram Frequency
Vector Scope1
n2 Product5
-1.998
m1
Scope2
Product6 Sum2
0.956 z
Product8
Unit Delay4
1
0.957
z
m0
m 2 = 1,543
m1 = 1,803 (15)
m = 0,654
0
n 2 = 542,917
n 1 = 197,392 (16)
n = 345,525
0
Fig.3. Noise diagrams of the system
input in time and frequency domain. The Simulink model and time and frequency diagrams
are in fig. 5, 6, and 7, respectively.
302
Periodogram
10^(-5)
Generator zgomot
gaussian Periodogram 1
Constant Product1 Periodogram1
1 1
Scope1
0.91513 z z
-84.857
Product3 Unit Delay Unit Delay1
1/m2
Product4
n0
Sum
7.88
n1 Product2
Periodogram
92.74384
Periodogram Frequency
Vector Scope1
n2 Product5
-1.9960
m1
Scope2
Product6 Sum2
0.91513 z
Product8
Unit Delay4
1
0.91514
z
m0
IV.CONCLUSIONS
303
system and normate frequency nT. The slow phase
variations are transmitted to the output.
The synthesizer works like a phase tracking system
with gain of N. Because the reference source noise
has a contribution in the spectral density of the output
noise in [6]:
2n s + 2n
Si () N 2 (17)
s 2 + 2n s + s 2
REFERENCES
304
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Facultatea de Electronic i Telecomunicaii, Departamentul Electronica Aplicata
Bd. V. Prvan Nr. 2, 300223 Timioara, e-mail: virgil.tiponut@etc.utt.ro
305
includes both sensors and actuators. A Smart communications, based on the SPI (Serial Peripheral
Transducer is A transducer that provides functions Interface) protocol. The NCAP usually initiates a
beyond those necessary for generating a correct measurements or action by means of triggering the
representation of a sensed or controlled quantity. This STIM, and the STIM responds with an
functionality typically simplifies the integration of the acknowledgement once the requested function is
transducer into applications in a networked completed.
environment. A Smart Sensor is A sensor version of After this short review of the most important elements
a smart transducer [4]. of the IEEE1451 Standard, we are capable now to
We highlighted the above mentioned terms because develop a curricula for a Smart Transducer course.
they are important entities in IEEE1451 Standard, and The planned material could be highlighted from the
the proposed content of sensors/actuators course and following six important aspects that are the six main
the corresponding laboratory works will be developed chapters of this course.
around this standard.
Smart Transducer Structure and Caracteris-
II. PLANNED EDUCATIONAL MATERIAL tics: Smart Transducer Interface Module
(STIM), Network Capable Application pro-
The IEEE1451 is a family of standards for connecting cessor (NCAP), Transducer Independent In-
ST (sensors and actuators) to networks. These terface (TII), Electronic Data Sheet (EDS),
standards will enable network-capable but network- signal conditioning circuitry, A/D and D/A
independent plug-and-play transducers for use in converters, digital processors;
embedded products, distributed data acquisition and Distance Access of Smart Transducers: wired
control systems, and networked controllers. and wireless busses (Radio Frequency,
The key elements of a ST, according to the IEEE1451 Mobile Phone and Infrared support), Internet
Standard, are depicted in Fig.1 [4] [5]. access;
Sensor Technologies: semiconductor techno-
logies, polymer films, thin and thick-films,
Microelectromechanical Systems (MEMS)
technology;
Sensors (sensing effect, sensor structures,
characteristics):force/moment, visual, tactile,
ultrasonic, global positioning system (GPS),
biosensors;
Actuators and micro actuators;
Application of Smart Transducers: industrial
applications, robotics including mobile and
personal robots, biomedical applications.
306
For actual implementation of the laboratory works, we
have to choose now appropriate hardware and
software resources. The most convenient solution for
STIM implementation is to use highly integrated
circuits, containing on the same chip signal
conditioning circuitry, A/D and D/A converters,
digital I/O, and some control logic(a microcontroller
core with memory resources and some peripherals).
Moreover, in some cases (ultrasonic, tactile ST) a
high speed A/D converter is needed, while in other
cases (force/moment transducer) high resolution
converters are necessary. The most recommended
circuits, which meet these requirements, are the
family of microconverters developed and
manufactured by Analog Devices [6]. These circuits
have been design according to the philosophy system Fig.2. A top view of the STIM module for fast
on chip and ready to be used for STIM processes.
implementation [7].
The software applications are an important task in the In this way, we can easily implement the hardware
development process of the laboratory works. In order resources for all of the above mentioned laboratory
to successfully complete this step, appropriate works, using only a limited number of modules. This
software development tools have to be used. strategy keeps down the cost of the whole system
In universities, MATLAB and LabVIEW are the without sacrificing the educational content.
most preferred software tools for simulations, data
acquisition and control, data analysis and data III. EXAMPLE OF LABORATORY WORK
presentation. LabVIEW, a graphical programming
language, is a easy to use and a very efficiently tool, In order to evaluate the efficiency of the proposed
especially when a Graphical User Interface (GUI) strategy, in this chapter will be presented, as an
have to be developed; on the other hand, MATLAB is example, a laboratory work on a Smart Transducer
recommended for process simulation and data Implementation. Actually, the purpose of the practical
presentation. activity is to study the structure, behavior and some
In our application, where each laboratory work has its characteristics of the hardware/software components
own GUI, we will prefer to use LabVIEW for included in a ST. A force/moment sensor will be
software development. connected to the input of the STIM for demonstration.
The whole system includes, as hardware resources, a The structure of the smart sensor is depicted in Fig. 3.
number of modules, each of them having a well
defined functionality. Up to date, the following RS-232
modules have been developed, build around
microconverters from Analog Devices: F/M STIM NCAP
307
microconverter chip is simple, low cost and has both STIM and NCAP, allow in a simple manner to
excellent performances, being the most appropiate load the modified/improved versions of the software
solution in this application. applications, using appropriate loaders.
In the present laboratory work NCAP play a minor The GUI is a very flexible application, which allow
roll, because STIM can be actually directly connected the user to send control strings to the NCAP and to
to the serial interface of the personal computer visualize the acquired data. The analogical signals are
(UART within microconverter should be used for this presented on a waveform graph while the digital I/O
purpose). However, NCAP has been included in the signals are displayed on array of LEDs.
structure of the smart sensor in order to maintain the A snapshot of the developed GUI is shown in Fig. 5.
compatibility with IEEE1451 Standard and to allow
the students to experiment the TII between STIM and
NCAP.
NCAP is implemented on a 8052 microcontrolller
system, having a RS-232 compatible serial interface,
which allow the interconnection with the host
computer.
Some digital I/O lines of Port 0 are used for TII
implementation, as is represented in Fig. 4. The
corresponding lines from the STIM belong to the
308
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
309
- High-speed measurement with
external triggering.
In order to make further signal evaluation and for
driving or feeding additional devices, the ESVB
features the following interfaces [1]:
Coding and supply socket (ANTENNA
CODE) for active antennas and for coding
of transducer factor.
74.7 MHz IF output for connecting a
spectrum analyzer.
10.7 MHz IF output for evaluating the IF Figure 1 IEEE 488.2 serial link used in both applications.
signal e.g. with an oscilloscope.
Controlled in phase and quadrature signal With this configuration all data obtained with the
output for evaluating signals of any Test Receiver is transferred along an HP serial
modulation. cable to a computer. Via this bus almost all
Envelope detector output (VIDEO instrument settings can be effectuated,
OUTPUT) for evaluating the rectified IF measurements triggered and test results read in by
signal e. g. with an oscilloscope the PC.
User interface with: Basically, both software applications have the same
- 6 TTL ports for controlling logical structure, shown in the Figure 2.
external devices
- Input for external trigger signals GPIB initialization,
- Outputs for analog display start
voltage with and without meter communication
simulation with ESVB
- RS-232 interface for firmware
updates by reprogramming the
built-in flash EPROMs by means
of an IBM-compatible PC
- Parallel interface (PRINTER Setup the ESVB: start
INTERFACE) for connecting a freq., stop freq.,
printer incremental freq.
IEC/IEE-bus interface.
Connector for MF2-compatible keyboard
for text entry.
Output for internal oven-controlled crystal
reference frequency (10 MHz). Start
Battery input for independent powering scans
310
In the Figure 3 are shown the elements of the concatenation. Finally, the concatenated data is
LabView application, which represent the main converted into ASCII string (in the main program),
program. and it can be further processed (not implemented).
311
- Third column: Status word,
which is 0 for error free.
5. CONCLUSIONS
ACKNOWLEDGEMENTS
REFERENCES
312
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
313
10. All coding schemes CS1- 4 Off ->On
the mobile
GSM measurement functions are [1]:
Power versus time (Power time
template).
Timing error.
Echo test (voice test, includes also
testing of loudspeaker and Make a
microphone). call from
Function test of mobile's keypad mobile
through display of dialled number
Display of
o IMSI (international mobile
subscriber identity)
o IMEI (international mobile
equipment identity)
Sensitivity Handshake
o Bit error rate BER and mobile with
RBER CTS
o RxLev, RxQual and BLER
Phase and frequency error
3. SET-UP OF THE GSM MOBILE ANALYSIS The structure and the components of the experiment
EXPERIMENT are shown in Figure 2.
The experiment we have made consists in a RS-232
upload/download link between a PC and CTS, with
a serial cable DTE <-> DTE. For the analyse
purpose a Motorola mobile was used, with a 50
ohms RF coaxial cable link to the CTS. The
operations diagram is shown in Figure 1:
Make physical
connections
PC <->CTS +
Figure 2 The structure and the components of the experiment:
CTS <-> mobile desktop computer, CTS-65 and a mobile phone
314
programmed tests and trace an analysis report
moments obtained in a simple test experiment
concerning a Motorola mobile phone.
315
5. CONCLUSIONS
ACKNOWLEDGEMENTS
REFERENCES
316
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Facultatea de Electronica si Telecomunicatii, Departamentul de Masurari si Electronica Optica, Bd. V. Parvan Nr. 2, 300223
Timisoara, Romania, e-mail aldo.desabata@etc.utt.ro
2
Politecnico di Torino, Dipartimento di Elettronica, C.so Duca degli Abruzzi, 24, I-10129 Torino, Italy, e-mail
ladislau.matekovits@polito.it
317
v
be considered: first, the "even" case, e g 2 = e g 3 = ,
2
A
(for which "+" will be used), and second, the "odd"
2
v2 A 2 i2 1 E
r r 1
1 r
1 r + eg2
r r rin,2
eg1 +
-
rin, r r 1 r D r
- C
3 1 v1 r r
B
v3
1 v3 F
(a)
B 3 i3 1 + eg3
(a) -
1 1 r r
v3+
2i+
eg1 +
rin,1 r/2 r/2
- 1/2 +
v/2
r/2 r/2
-
1 r r
1/2 v2=v3
1
(b) (b)
1 1 v1-
2
2v 2 8
S 21 = S 31 = 2 = = . (c)
e g1 r r 1 11r + 12
1 + r + || r + +
2 2 2
(3) Fig. 3: a) The network in Fig. 1 with additional
components for the computation of the Si2, Si3
For the rest of the parameters, the network in Fig. 3 coefficients; b) equivalent circuit, "even" excitation; c)
(a) can be used. In order to apply the definition, for equivalent circuit, "odd" excitation.
calculating S22, S12, and S32 it must be
imposed e g 3 = 0 , while for S33, S13, and S23 one must
v
case, e g 2 = e g 3 = (denoted by "-"). In the even
choose e g 2 = 0 . Of course, in this case S12 = S13 due 2
to symmetry, and the fact that S 23 = S 32 results not case, points A and B in Fig. 3 (a) are at the same
potential, and can be electrically connected, and
only from symmetry, but also from reciprocity. similarly points E and F. The resulting network is
Advantage will be taken of symmetry for avoiding represented in Fig. 3 (b). In the odd mode, points D
again wye-to-delta transformations. Inspiration comes and C in Fig. 3 (a) are at the ground potential, and the
from the even and odd modes on coupled transmission network schematic can be redrawn as in Fig. 3 (c).
lines [2], or common mode and differential mode in By inspection, it can be written
differential amplifiers [3]. The following method is
also used in [4] and tackled in [5]. Two situations will
318
v1+ =
v 1
=
4v
, v1 = 0 , z 2 ' z1
|| + z 2 '
2 1 r r 11r + 12 z 2 2 ,
+ || + r + r + 1 v 2 = e g1 in
2 2 2 1 + z in z1 z 2 ' z1
+ || + z 2 '
r r 2 2 2
|| + r + r + 1
v v 2 2 v 4 z2 '
v 3+ = + = , v3 = v 2 .
2 2 1 r r 2 11r + 12 z1
+ || + r + r + 1 + z2 '
2 2 2 2
r
v v r
v 3 = 1 2 = ,
2 r 2 r+2
1+
2
v2
v/2 4v z2 1
2i + = = ,
1 r r 11r + 12
+ || + r + r + 1 2
2 2 2 z1 z1
A v3
v 1 1 3
2 v
i = = . z2
r r+2 +
1+ eg1 zin,1 z1
2 z2 z1 B
- 1
4
Due to linearity, for e g 2 = v and e g 3 = 0 , the values
z2 1
for the electrical quantities defined in Fig. 3 (a) can be
obtained from v1 = v1+ + v1 , v 3 = v 3+ + v 3 ,
v
i 2 = i + + i , and rin , 2 = 1 . Finally, the following (a)
i2
scattering parameters result:
1 z1/2 z1/2
1
rin , 2 1 11r + 8r 8
2
S 22 = = ; (4) eg1 +
rin , 2 + 1 11r 2 + 34r + 24 zin,1 z2 z2'/2 v2 z2' v3
-
2v1 8
S12 = = ; (5)
v 11r + 18
2v 18r + 16
S 32 = 3 = . (6) (b)
v (11r + 12)(r + 2)
Fig. 4: a) Symmetric four-port reactive circuit with additional
The second example is the network represented with components for S parameters computation (dashed line); b)
solid line in Fig. 4(a). The circuit contains the equivalent circuit.
normalized impedances z1 and z2, which are
respectively an inductance and a capacitance in [1].
Now, the formulas for the scattering parameters [1]
Due to symmetry, it is sufficient to find the S
can be applied:
parameters at only one of the four ports. The elements
it must be completed with in order to calculate Si1,
z in ,1 1
i=14 are drawn with dashed line in Fig. 4 (a). The S11 = =
same symmetry allows for the electrical connection of z in ,1 + 1
points A and B, and the equivalent network ( z1 z 2 z1 2 z 2 )( z1 + 3z 2 + z1 z 2 )
represented in Fig. 4 (b) is obtained, where z 2 ' = 1 || z 2 . = + (7)
( z1 z 2 + z1 + 2 z 2 )( z1 + 4 z 2 + z1 z 2 )
There results at a glance
z 2 ( z 2 1)
z z ' z +
z in ,1 = z 2 || 1 + 2 || 1 + z 2 ' , ( z 2 + 1)( z1 + 4 z 2 + z1 z 2 )
2 2 2
2v 2 2 z 22
S 21 = = , (8)
e g1 (1 + z 2 )( z1 + 4 z 2 + z1 z 2 )
2v 3 2z2
S 31 = = S 21 . (9)
e g1 z1 + 2 z 2 + z1 z 2
319
A general purpose, non circuit- or microwave-
oriented software package, such as Matlab, can be
used for plotting the insertion loss, return loss or 2 i2 1
zc, , L
isolation for various values of the elements in the
circuit. 1
1
v2 + e
In figs. 9 and 10 of [1], the S parameters seem to be r g2
rin,2
evaluated at a few frequency points and the response -
+ zc, , L v3
in the whole frequency range is obtained by an eg1 rin,1
interpolation. Usually, this approach is used when - i3 1
determination of the response in one point requires 3 + e
2' g3
long computation time. -
The proposed method can be used both for the
determination of the response of the circuit at few 1'
sampling points followed by an interpolation scheme,
or to compute the response in the whole range with a 3'
relative large number of points. (a)
The computation of the S11, S12, S13 was implemented
in a Matlab script. Its run on a Pentium III with
1GHz clock required 0.01 s for 1000 sampling 1
1 zc/2, , L
points.
As a third example, a symmetric Wilkinson power eg1 + 1/2 v2= v3
divider [1]is schematically represented with solid line zin,1
-
in Fig. 5 (a); on the same figure, with dashed lines
(b)
the circuits needed for the calculation of the S
parameters, are also represented. All the impedances zc/2, , L 2i2+=2i3+ 1/2
are normalized to a given reference impedance, here 1
considered identical for the three ports. The v +
1+ 1 v2+= v3+ eg2
transmission lines in the circuit are supposed
=eg3
identical, with (normalized) characteristic impedance 1' - =v/2
zc, propagation constant = + j and length L. (c)
For the calculation of S11, S21, and S31, one must
take e g 2 = e g 3 = 0 . Due to symmetry, the voltages zc, , L i2-=-i3- 1
over the lines are equal at equally spaced points from 1 eg2
the generator; consequently, there is no current flow v1-=0 r/2 v2-=- v3- + =-eg3
through the resistance r, and the two transmission =v/2
1' -
lines can be conceptually connected in parallel, as in
Fig. 5 (b). The result is a transmission line with the (d)
same propagation constant, but with a characteristic
impedance zc/2.
(Indeed, consider a standard equivalent circuit of a Fig. 5. (a) The Wilkinson power divider (solid line) and the
differential length of transmission line, [2, p. 86]. components needed for the calculation of the scattering
Conceptually connecting two such circuits in parameters (dashed line); (b) equivalent circuit for the
calculation of the scattering parameters at port 1; (c) "even"
parallel, node by node, in the case of identical,
excitation; (d) "odd" excitation.
corresponding voltages and currents, reduces by a
factor of two the line resistance and the line z in ,1 1 ( z c2 2) sinh(L) z c cosh(L)
S11 = = . (11)
inductance, and doubles the line capacitance and the z in ,1 + 1 ( z c2 + 2) sinh(L) + 3 z c cosh(L)
line conductance. The substitution of these values in
the expressions of the characteristic impedance and
the propagation constant [2, p. 88] yields the above
In the following, it will be needed the use of an
stated result.)
expression for the voltage on a transmission line of
The normalized input impedance in Fig. 5 (b) is
normalized characteristic impedance zc, propagation
constant , and length L, terminated on a normalized
1 zc
+ tanh(L) load impedance zL, and connected to a generator eg, of
zc 2 2
z in ,1 = , (10) normalized internal impedance zg. Such an expression
2 zc 1
+ tanh(L) can be found in most textbooks on transmission lines;
2 2 the variant adopted here is
and, consequently
320
z c e ( L x ) + L e ( L + x ) zc
v( x) = e g , (12) v 2 (1 + L + )e L
z g + zc 1 L g e 2L v1+ = ,
2 1 z c 1 L + g + e 2 L
+
z L zc 2 2
where x is the distance to the load, and L = , zc
z L + zc
v 2 1 + L + e 2 L v
z g zc v 2+ = v 3+ = .
g = 2 1 z c 1 L + g + e 2L 2
are the Fresnel reflection coefficient to +
z g + zc 2 2
the load and to the generator respectively.
The application of (12) at the load end for the Finally, from Fig. 5 (c) there results directly
transmission line (x=0) in Fig. 5 (b) gives i2+ = i3+ = v 2+ = v3+ .
In the odd case, by taking into account Thvnin's
zc theorem applied to the generator, the reflection
e g1 L
v 2 = v3 = 2 (1 + L )e r
z 1 L g e 2L
. (13) 1 || 1
1+ c coefficients are L = 1 , g = 2 , and
2 r
1 || + 1
2
1 zc consequently, from (12) and Fig. 6 (d)
v1 = 0 ,
In (13) the reflection coefficients are L = 2 2 ,
1 zc r
+
2 2 v 2 zc 1 e 2 L v
v2 = v3 = 2 L
,
z 2 r r
1 + 1 || + z c 1 + e 2
1 c g
321
z in , 2 1 2( v 2 + + v 2 ) REFERENCES
S 22 = = 1+ =
z in , 2 + 1 v [1] M. N. O. Sadiku, "Deficiencies in the Way Scattering
2 cosh(L) + z c sinh(L)
Parameters Are Taught", IEEE Transactions on Education, Vol. 46,
= zc + (16) No. 3, pp. 399-404, August 2003.
(2 + z c2 ) sinh(L) + 3z c cosh(L)
[2] R. Collin, Foundations for Microwave Engineering, 2nd ed.
rz c (r + 1)(1 e 2L ) New York: McGraw-Hill, 1992.
+ 1.
r + (r + 2) z c r + 1 e 2L [3] P. L. Gray, C. L. Searle, Electronic Principles. Physics, Models,
and Circuits, New York: John Wiley & Sons, 1969.
A Matlab script was implemented again for the
[4] R. Badoudal, Ch. Martin, S. Jacquet, Les Micro-ondes, Vol., 1,
computation of the S11, S22, and S23 parameters. Its run Paris: Masson, 1993.
on a the same Pentium III with 1GHz clock required
less than 0.1 s for 5000 sampling points. [5] D. M. Sazonov, A. N. Gridin, B. A. Mishustin, Microwave
Circuits, Mir, Moscow, 1982.
III. CONCLUSION
322
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
323
Adaptive Platform-Independent Mobile Learning,
[7]. Here the SMS-based interaction is supplemented C. Architecture
by a voice-based one, using the H323 Gateway and an
Interactive Voice Response. The proposed system architecture is described in
figure 1. It contains several components:
C. Nomadic access architectures - The rendering components: HTML, WML,
VoiceXML. A renderer [4] transforms a
The nomadic access architecture is an architecture that UIML or SunML user interface description
allows user access to a service from different into a concrete user interface. This gives us
terminals and through different networks. The most the possibility to have only one interface
important problem to solve here is the integration of description for several platforms and we are
different constrains and technologies. using it for the publishing of the examination
Our attention was focused on user interface results.
integration solutions because we want to offer access - The J2ME client is a MIDlet, dedicated to
for the same service using different user interfaces. absences and observation management. The
The idea is to have a unique description for the user professor/ teacher is the one who uses it. The
interface that can be automatically transformed into connection can be wired or wireless.
platform specific user interfaces. Several languages - The SMS Notify is an interface to the GSM
have been developed to allow a platform independent that allows the system manager to send
description of the user interface: UIML (User notifications to the parents. The parent must
Interface Markup Language) [8], XIML (eXtended subscribe to this service first.
Interface Markup Language) [1], SunML (Simple - The Database contains basically all the
Unified Natural Markup Language) [4] and others. persistent information related to the users:
These languages allow us to describe the user user account information, examination
interface in a platform independent manner. Several results, absences, etc. Different users are
renderers create a concrete terminal dependent user authenticated using this database, and they
interface. have access to services based on their status.
- The User Interface description contains the
III. OUR PROPOSAL unique description for the examination
results and absences publishing service. This
A. Objective fact reduces the design effort for all the
required interfaces.
As we have seen in the previous section, there are
several field specific solutions for educational
J2ME
services, mobile learning services and nomadic access Client
services. In our case we need something that
SMS
integrates all this solutions. In this paper we intend to
propose an integrated solution for educational
nomadic services, in particular we want to develop a
system that will help the evaluation and publishing of
students grades. For high school or gymnasium it is
possible to involve the parents by sending them
instant notifications about the pupils activity.
Professors Parents Students
B. Functions interface interface interface
324
- The System Manager links all the other using the HTTP protocol. This client
components and contains the business logic. application allows the professor to mark the
It uses the user interface description and the absences and to introduce comments for each
renderers in order to interact with the user student. The user interface is based on the
according to his particular interface. At this java.microedition.lcdui package. A
level we have the application algorithm and screenshot of this client is shown in figure 2.
the task (user machine dialog) description
- The SMS Notify component is based on
and succession. The manager implementation
JSMSEngine package. This component needs
can be based on a workflow engine for
a physical interface that is a mobile phone or
instance.
a model connected by a serial interface with
the machine that runs the SMS Notify
IV. IMPLEMENTATION
component.
A. Technologies
- The Database is partially based on MySQL
Several technologies were used in order to build this and partially on Excel. It was easier to use
prototype: MySQL in order to store information such as
username, password, status and absence
- J2ME: represents a highly optimized Java
number. Figure 3 shows a screenshot of the
runtime environment, which specifically
MySQL database interface. The advantage of
addresses the vast consumer space, which
using Excel was the compatibility with the
covers the range of extremely tiny
existent list used by the professors and it is
commodities such as smart cards, phones or
also easy to configure the algorithm to
a pager all the way up to the set-top box, an
compute the final result for a student. A
appliance almost as powerful as a computer.
dedicated connector is used in order to
- Java Servlet: a Java technology that offers a connect Java with Excel.
fast, powerful, portable environment for
- The User Interface description is a UIML
creating dynamic content for all the XML
file. This represents the user menus and they
based technologies as HTML, WML,
will be transformed into different concrete
VoiceXML.
implementations by the renderers.
- JSMSEngine: is an API package, written in
- The System Manager is a complex Java
Java, which allows sending or receiving the
application that implements the systems
SMS messages from PC, by using a mobile
workflow. For instance, the first user action
phone or a GSM modem.
is the authentication then he should see the
- UIML: is an XML language for defining user main menu, the functions and so on.
interfaces. It can be used to define buttons,
C. Screenshots
menus, lists and other controls that allow a
program to function in a graphical interface. In figure 2 we show the mobile user interface that
It also defines actions to take when certain allows the professor to mark the absences.
events take place.
- kXML-RPC: is a RPC (Remote Procedure
Call) middleware implementation for the
mobile phone. The advantage of using this
communication protocol is the high
abstraction level. Aversion for J2ME mobile
phones exists. This protocol encodes the
method calls using the XML syntax and send
this messages over the HTTP protocol.
B. Prototype description components and use
cases
A prototype was implemented using the technologies
described above. The architecture components were
implemented as follows:
- The HTML, WML and voice rendering
components are UIML renderers based on
LiquidUI [8] software distribution
- The J2ME client is a small MIDlet
application with a simple user interface that Figure 2. The mobile client
communicates with the System Manager
325
The mobile client was tested with the J2ME Wireless a solution may be to try to automatically generate the
Toolkit simulation environment. The client size is platform dependent elements.
about 56Kbytes and the kXML-RPC connector takes
Another problem is the ergonomics of the
about 24Kbytes.
automatically generated user interface. Anyway this is
The communication can be based on GPRS, another research direction and we intended just to
WLAN, IrDA, Bluetooth or even cable. Anyway, reuse this existent technology in order to prove our
because of the price, it is not recommended to use architecture.
GPRS.
In figure 3 we show the MySQL database interface.
In figure 4 we show an UIML code sequence from our The records correspond to the absences and
application. observation statistics. The grades are stored as one
Excel file for the reasons explained in the section IV.
<?xml version="1.0"?>
<!DOCTYPE uiml PUBLIC
"-//Harmonia//DTD UIML 2.0 V. CONCLUSIONS
Draft//EN""UIML2_0g.dtd">
<uiml>
<interface>
<structure> This paper has presented an integrated system that
<part id="Main" class="Wml"> allows actors involved in an educational process -
<part id="Welcome" class="Card"> professors, students, pupils and parents - to manage
<style> the student results and absences. We consider this
<property name="title">Bine ati venit
!</property> proposal a contribution to the e-learning and m-
</style> learning fields.
<part id="newP" class="Paragraph">
<part id="titletxt" class="Text"> We are focusing on the nomadic access for the
<style> educational services because, as we have seen in
<property name="content">Ati
accesat pagina catalog al sectiei de
chapter II, this field is less approached in literature.
comunicatii al unversitatii tehnice din cluj The proposed system is available not only on the Web
napoca</property> but also on mobile devices and even on very simple
</style> terminals like fixed phones. This can be done by using
</part>
</part>
an automatically generated user interface.
</part> A prototype was implemented in order to test our
proposition. Numerous software technologies were
Figure 4. UIML code sequence.
used because the nomadic access implies extended
platform diversity.
Even if UIML wants to be platform independent it has The implementation has also some limitations
platform dependent elements. This is a drawback and that come basically from the user interface automatic
326
generation, which is not optimised, and, as we have
seen, the platform independence is not always
complete.
REFERENCES
[1] A. Puerta, J. Einstein, XIML: A Common Representation for
Interaction Data, In proceedings of the International Coference on
Inteligent User Interfaces, IUI 2002, pp. 214 225, ISBN: 1-
58113-382-0, San Francisco, USA, 2002.
[2] A. Stone, Blended Learning, Mobility, and Retention:
Supporting First Year University Students with Appropriate
Technology, In the Proceedins of the third European Conference
on Mobile Learning, MLEARN 2004, Rome, Italy, 2004, ISBN 1
85338 855 6.
[3] C. Martel, La modlisation des activits conjointes: rles,
places et positions des participants, PhD Thesis, University of
Savoie, September 1998.
[4] J. Fierstone., A-M. Pinna-Dery, M. Riveill, Architecture
logicielle pour l'adaptation et la composition d'ihm - mise en
oeuvre avec le langage sunml. Technical Report I3S/RR-2003-02-
FR, Laboratoire I3S, Universit de Nice, ESSI - BP145 - F-06903
Sophia Antipolis, Januay 2003.
[5] J. Colley, G. Stead, Mobile learning = collaboration, In the
Proceedins of the third European Conference on Mobile Learning,
MLEARN 2004, Rome, Italy, 2004, ISBN 1 85338 855 6.
[6] M. F. Vaida. JAVA 2 Enterprise Edition (J2EE) Aplicatii
multimedia, Ed. Microinformatica, Cluj, 2002
[7] N. Capuano, M. Gaeta, S. Miranda, L. Pappacena, A System
for Adaptive Platform-Independent Mobile Learning, In the
Proceedins of the third European Conference on Mobile Learning,
MLEARN 2004, Rome, Italy, 2004, ISBN 1 85338 855 6.
[8] UIML 3.0 Specifications: http://www.uiml.org/specs/uiml3/,
February 2002.
327
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
0 vI [V] Comparing the Fig.1 and Fig.2 one can observe that
VI the VTC in the second figure results by horizontally
Fig.1. The VTC of a single supply op- translation of the initial VTC with an appropriate dc
amp amplifier voltage, lets name it VBIAS. This translation is made
1
Technical University of Cluj Napoca,, Str. C. Daicoviciu 15, 400020, e-mail Gabriel.Oltean@bel.utcluj.ro
2
Electrotechnical School Group Edmond Nicolau, Cluj-Napoca
328
on the x axis and is obtained by series connection of a
VPS
+VPS
R
VBIAS
A B A C1 B
~
vi vI vI Av vO vi R
So, beside the small signal gain also results a dc gain. Fig.7. The final block scheme of the biased amplifier
To improve the operation of the amplifier (reduce the
effects of finite input dc offset voltage) it is often a
III. METHOD ILLUSTRATIONS
great idea to roll off the gain to unity at dc,
especially if the amplifier has large voltage gain [3].
A. Non-inverting amplifier
This can be done by using a capacitor in the negative
feedback path of the amplifier. This capacitor should
The circuit with the VTC presented in Fig.1
be placed so that it can be able to play a double role:
corresponds to a non-inverting amplifier that looks
first, to set the desired a.c. gain (Av) (by its short-
like the one in Fig. 8. To be able to use this circuit to
circuit equivalence in the small signal regime) and
amplify the vi voltage we apply all the steps presented
second, to set a unitary d.c. gain (by its open-circuit
in the previous paragraph.
equivalence in the d.c. regime).
7 +VPS
The voltage output equation of the amplifier became:
+
vO = VBIAS + vo Av (9) 6
vO
R1
2
To avoid the use of an extra dc voltage source, VBIAS -
vi 1K 4
can be obtained by a resistive voltage divider from
VPS to the ground. For equal resistances the value of R2
VBIAS results VBIAS = VPS/2. To prevent the dc current
to flow through the input signal source we use a 10K
capacitor to insert to the input the VBIAS dc voltage - Fig.8. The non-inverting amplifier
see Fig. 5. In the steady state regime (after the ending
of the transient regime) this capacitor is charged up
329
First we translate the VTC using the VBIAS voltage The equations in the circuit are:
series connected at the input of the amplifier. The
VBIAS voltage is obtained through the voltage divider R2 R
from the power supply voltage VPS. The series v O = (v in + V BIAS )(1 + ) 2 V BIAS (10)
R1 R1
connection is realized by the C1 capacitor, resulting
the intermediary circuit depicted in Fig. 9. v O = v in Av + V BIAS (11)
R
4 Av = 1 + 2 (12)
12Vdc VPS R1
R3
1k
C1 0 The VTC and the waveforms of the input voltage with
3 zero dc level vin(t) and the final vO(t) obtained after
+
10u 6 vO
simulation are presented in Fig. 12.
R4
2
1k
0
R1 R2
vin 0
1k 10k
(a)
0 0
12Vdc (b)
VPS
R3
1k
C1 0
3
+
4
10u 6
vO
R4
2
1k -
0
R1 R2
vin 0 Fig.12. The simulation results
1k 10k (a) VTC; (b) waveforms
C2
10u B. Inverting amplifier
0 0
Fig. 10. The final circuit for non-inverting amplifier In Fig. 13 is presented the circuit and in Fig. 14 the
VTC of an inverting amplifier.
330
vO [V]
The simulation results are presented in Fig. 17: the
VPS VTC in Fig 17.(a) and the waveform of the input
voltage vi(t) and the final voltage vO(t) in Fig 17.(b).
VO=VPS/2 The simulation results showed that our method is
correct.
0 vI [V] 12V
VI
Fig.14. The VTC of the inverting amplifier
8V
VBIAS 10V
3
+
vO (b)
6
C2 R1 5V
4 vi 2
-
R4 0 0V
R2
0s 2.5ms 5.0ms
0 V(VO) V(Vin:+)
10k
Time
Fig.15. The final circuit for inverting amplifier
Fig.17. The simulation results for inverting amplifier
(a) VTC; (b) waveforms
The equivalent circuit that shows how the VBIAS
voltage appears in the circuit is the one in Fig.16. The
equivalence is made for the steady state regime, where IV. CONCLUSIONS
the capacitor can be replaced with a d.c. voltage
source and also the voltage divider can be replaced We presented an intuitive method to deal with the
with a d.c. voltage source. amplification of a zero d.c. level variable signal using
an op-amp operated from a signal power supply. The
examples presented here (for a non-inverting and for
12Vdc VPS an inverting op-amp amplifier) demonstrates the
usefulness of the method, especially for the students
in the struggle with electronics. This method can
7
save classroom instruction time and help the students
3 5 0 to understand and easily solve this kind of problem.
+
VBIAS=6Vdc 6 vO This method can be further use to try new way to bias
R1
2
some amplifiers in a desired operating point, even
- transistor amplifiers.
vin 1k
VBIAS=6Vdc 4 0
R2 REFERENCES
0 0 10k
[1]. Sedra, A.S., Smith, K.C., Microelectronic
Fig.16. The equivalent circuit for inverting amplifier Circuits, Holt, Rinehart and Winston, Inc, 1987
[2]. Oltean, G., Dispozitive si circuite electronice.
The equation of the amplifier is:
Dispozitive electronice, Risoprint, Cluj-Napoca, 2003
[3]. Horowitz, P, Hill, W., The Art of Electronics,
R2 R
vO = (vin + VBIAS ) + (1 + 2 )VBIAS (13) Cambridge University Press, 1997
R1 R1 [4]. Oltean, G., Gordan, Mihaela, Oltean, Ioana, A
R2 new method to deduce the voltage transfer
v O = v in + V BIAS (14) characteristic for some two-port network, Acta
R1
Tehnica Napocensis Electronics and
R2 Telecommunications, Vol. 40, Nr. 2, 2000, pp 17-20.
Av = (15)
R1
331
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Inginer, S.C. Spidernet S.R.L. Calea Bogdanesti 6A
Timioara, e-mail seba.tiponut@spidernet.co.ro
332
undergo several implementations while keeping the communication between the graphical user
information exchangeability and compatibility. interface and the experiment; the second is a very
simple file transfer protocol used to retrieve the page
containing the list of the available experiments. The
II. GENERAL PRESENTATION OF THE SYSTEM former protocol has double purpose: it transmits the
commands issued from the graphical interface and it
The proposed distance laboratory system consists of transmits back the data from the server.
seven elements which are interacting using a set of
predefined, well-known, interfaces. The core of this
system is a client-server structure. The client is
represented by a graphical user interface which
communicates with the server using predefined
network routines; its function is to display collected
data. To mediate the concurrent access to the physical
experiment of a large number of clients, the server
part has been designed. It comprises of networking
routines, decision code and a user-written driver
whose purpose is to interface the server with various
data acquisition systems which collect data from and
control the experiment.
To make the graphical interface more adaptable to Fig.1 General view of the system
different types of experiments, a predefined, custom
set of widgets has been designed. This widget set -
which is supposed to be extended and even The file format for the page listing the experiments
standardized - contains from simple primitives like along with a short description is an XML page which
buttons, frames to complex objects in form of gauges, contains tags which specifies the location of the
2D or 3D displays, sliders etc. In order to facilitate the graphical interface description on the server. It
construction of various graphical interfaces in a very specifies for each experiment the experiment
simple and platform independent manner, the identification and an arbitrarily long description of the
graphical interface is described in XML. An XML experiment itself.
parser is then building the interface based on this file,
instantiating the widgets from a library. One can At the core of creating the user interface stays the
associate an experiment with an XML description of a XML file containing its descriptions. It is composed
graphical interface, customizing the latter to suit of tags, each tag describing a widget along with its
experiment's needs. characteristics. Some widgets rely upon the presence
of others; thus, the XML parser is doing a check to
One server can host multiple experiments, possibly assure no mistakes have been made during file
different experiments. Because of that they must be editing. The library of widgets content will be
controlled using distinct graphical interfaces. Thus a described later on, in the section dedicated to
need arises to associate a particular experiment with implementation.
the correspondent XML description of the graphical
interface. This association was done using a very The physical experiment is connected to the server via
simple HTML-like page where, modeled after the a data acquisition system and a driver which is
World Wide Web. The path of the graphical interface mediating between the two. Because the data
description is associated with a hyperlink which, if acquisition system can be connected to the computer
clicked, will start the downloading process of that using a serial, parallel, I2C or other type of
particular XML description. It also supports some connection, the driver must be designed to handle
primitive form of text formatting. A special browser is these particularities. Another issue the driver must
used to display the page downloaded from a separate solve is the communication protocol with the data
server. The user can click on the hyperlink, have the acquisition system.
XML description downloaded and then the graphical
interface built from it. Once the graphical interface is
generated the experiment may start. A general view of III. PRESENTING AN IMPLEMENTATION
the system is depicted in Fig.1.
The author has chosen to implement a distant
As stated in the Introduction, this design emphasizes laboratory using the design presented above. The
portability and standardization. Thus an programming language in which the whole distant
implementation must comply with the specified laboratory was written is Python. It has been selected
protocols, file formats and interfaces. There are two because it is a very powerful object oriented
communication protocols: the first one is mediating programming language, supporting well designed
XML and general purpose libraries. Configuration
333
files, written in XML (for ease of parsing), have been goals the author had in mind when conceiving the
designed for all the components. structure of this system was to make it versatile for a
wide range of experiments. The widget library can be
The heart of the system is the parser which is extended with new types of widgets which can be
generating the graphical interface by interpreting an used to broaden the types of experiments for which
XML file containing the interfaces' description, as this model can be used.
presented above. This parser is a standalone program
which is usually called within the browser once an Another goal which was targeted was to make the
XML file with the interface description has been system as portable as possible. That is why it was
downloaded; it can also be ran from console, passing designed as an open project the interfaces and the
the file to be parsed as its first argument. A minimal protocols between the components are presented in
set of widgets have been designed to allow the detailed so that one can implement a compatible
construction of a graphical interface suitable to be system or just a part of it. While the system could
used for basic experiments. Among the widgets have been designed platform independent and
included in this set are: distributed as a software bundle we rather wish to
a graphical XY display; one can customize emphasize the benefits of a known protocol and an
the number of ticks and the start and stop open design schematic.
values separately for each of the two axes;
a sliding bar which can be used to input Last but not least, the system was thought to be open
values; the start/stop values and the to improvements and modifications while maintaining
increment are customizable; backward compatibility. This makes this design very
an amplifier which multiplies the values different from existing implementation of distance
outputted by the sliding bar with a certain, laboratories which are - the vastness majority - close
configurable, factor; source and very hard to develop and to integrate.
buttons which can be used for a variety of
operations; The implementations decisions are also making this
frames, used to group the other widgets distance laboratory different from other distance
together. laboratories. It is using XML as its core for both
creating a graphical interface for a given experiment,
The server which is serving the user interface XML communication between components and
descriptions is called the cache_manager. The server configuration files. The servers are designed such way
which is used by the user interface to mediate the that they can mediate the access of a large mass of
communication with the hardware is called the users to the available experiments. Using the Python
gateway. Both servers have XML configuration file. language for all the components offers some degree of
portability. The system was conceived to be used on a
IV. THE PILOT EXPERIMENT Unix machine but with minimal changes it can be ran
on other platforms as well.
This is a simulated experiment which is supposed to
test the distant laboratory whose implementation was REFERENCES
presented above. A piece of software is meant to [1] T. Chang, D. Hunt, Web-Based Distance Experiments: Design
substitute the data acquisition system and the physical and Implementation' 2000 International Conference on
experiment. This is simulating the characteristic of the Engineering Education, Taiwan.
diode which can be obtained by applying an [2] J. Henry, LabVIEW Applications in Engineering Labs, ASEE
Conference Anaheim, California, June, 1995.
increasingly high voltage between the anode and the [3] J. Henry, Controls Laboratory Teaching via The World Wide
cathode of the silicon diode. The use inputs different Web, ASEE Paper, June 1996.
voltage values which will be sent to server for [4] J. Henry, Internet Laboratory Server in Engineering Systems
evaluation. The simulation plugin will compute the Laboratory, Conference at M.I.T., June, 1997.
[5] H. Shen, Z. Xu, V. Kristiansen, O. Strom, M. Shur, Conducting
associated current value and it will reply with pair of Laboratory Experiments over the Internet, IEEE Transaction on
coordinates which represents a new point on the Education, Vol. 42, pp. 180-185, August, 1999.
graphic. [6] R. Berntzen, J. Strandman, T. Fjeldly, Advanced Solutions for
Performing Real Experiments over the Internet, 2001 International
Conference on Engineering Education, Norway.
After the graphic has been drawn it can be saved for [7] B. Aktan, et. al. Distance Learning Applied to Control
later use. Saving a graphic is also saving the Engineering Laboratories, IEEE Transactions on Education, v37,
environment (the number of ticks and the values from No. 3, pp. 320-326.
both axis and the measurements units). [8] D. Knight, S. DeWeerth, A Distance Learning Laboratory for
Engineering Education
[9] J. A. del Alamo, et. al. Educational Experiments with an
Online Microelectronics Characterization Laboratory
V. DISCUSSION [10] E. Bobkov, Y. Sheynin, The Web-based Technology in
Laboratories for Distance Learning and Training
[11] J. C. Piower, et. al. Web-Based Educational Experiments
This paper presents the design of a distant laboratory
[12] http://www.ni.com/academic/live_experiments.htm
along with a possible implementation. One of the
334
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
335
Fig. 1. A GUI for the search of a test vector which Fig. 2. A GUI for the generation of an optimum
points a stuck-at 0 fault in the marked node set of test vectors by hibridation of a GA
II. GENETIC LEARNING IN EHW to its fitness, while keeping the population size
constant. The least fit individuals are deleted. This is
All the developed algorithms are based on GAs, an the survival of the fittest part of the GA.
adaptive searching technique for solving optimisation
problems based on the mechanics of natural genetics The next step is crossover, where individuals are
and natural selection. The success of the application chosen two at a time, as parents. They are converted
of GAs to an optimisation problem depends on the into two new individuals, called offsprings, by
representation of chromosomes, fitness function, exchanging parts of their structure. Thus, each
method of crossover, mutation operation, and on the offspring inherits a combination of features from both
diverse information from the chromosomes. When the parents. We have obtained the best results with one
diversity is lost before the global optimum solution is point crossover, with a probability of 80%. This
found, the performance of GAs deteriorates and their operator may be used more times on different selected
solution processes converge prematurely. Moreover, pairs of chromosomes in a generation.
the mutation operation is important. While the
mutation operation adds new information to a The next step is mutation. A small change is made to
chromosome, it can also destroy useful information each resultant offspring, with a small probability.
held in the chromosome. After mutation is performed on an individual, it no
longer has just the combination of features inherited
In GAs the search is conducted using information of a from its two parents, but also incorporates the
population of candidate solutions, called additional change caused by mutation. This ensures
chromosomes, so that the chance of the search being that the algorithm can explore new features that may
settled in a local optimum can be significantly not yet be in the population. It makes the entire search
reduced. Four essential components need to be space reachable despite the finite population size. The
designed in applying a GA for an optimisation whole process is repeated for several generations, and,
problem: chromosomes representation, crossover if the best chromosome in population will have the
operator, mutation operator and fitness function. fitness of 100%, then this bit string represents a good
solution for our function.
In a reconfigurable circuit, each bit of a chromosome
represents usually the state of a programmable switch. III. EXPERIMENTS WITH EXTRINSIC EHW
The entire chromosome represents the state of all
switches, that is a complete circuit, which may be The first set of experiments show the generation of
good or bad, according with his fitness. The initial complexity with very simple rules in unidimensional
population of chromosomes (bit strings) is generated and bidimensional Cellular Automata (CA), and the
randomly. All these potential solutions are evaluated solving of some complex NP-problems (the finding of
using a fitness function. In our case, for a single the global minima in a multimodal function, or the
boolean function, fitness is the ratio between the solving of the TSP) with GAs. These experiments
number of the correct values of the function and the have been ample described in [7].
number of all possible values (which is 2 , if the
n
boolean function has n input variables). A well- Another set of experiments have been prepared for the
designed circuit will be obtained only when the value purpose of automated generation of test vectors in
of fitness is 100%. An approximately value of the digital circuits. If we want to generate a test to detect
fitness is unacceptable here. a stuck-at 0 fault in the marked node of the circuit
represented in the Fig.1., the required vector is
The next step is selection and reproduction. For each 1111111111000000000, a combination of bits nearly
individual, a number of copies are made, proportional impossible to find using a random approach ([6]). As
336
x1
x2
x1
f
x3
we can see in the Graphic User Interface (GUI) from idea given in [2]. Each combinational circuit is
the Fig.1, the GA used to solve this problem has represented as a rectangular array of logic gates. Each
found the correct solution in 40 generations. The of these gates has two inputs and one output, and the
algorithm uses a population of 32 chromosomes and a logic operator may be selected from a list. At the
mutation rate of 3%. Fitness was calculated as the beginning of the search, all the gates from the matrix
sum of (1 if fault is excited or 0 otherwise) + (fraction are disposable to implement a functional circuit. Once
of inputs in AND gate set to 1) + (fraction of inputs in a functional solution appears, then the fitness function
OR gate set to 0). The maximum value of the fitness is modified such that any valid designs produced are
defined in this way is 3. rewarded for each gate which is replaced by a simple
wire. The algorithm tries to find the circuit with the
By using the GUI from the Fig.2., we can solve the maximum number of gates replaced by wires that
Fault Coverage Code Generation Problem for a more performs the function required.
complex combinational logic. The problem consists in
finding of a given number of test vectors that The chromosome defines the connection in the
maximizes the fault coverage of the circuit. We have network between the primary inputs and primary
chosen two ways of hibridation of the standard GA: outputs. We have used a network of 4 gates, a
by using the inductive search, like in the Fig.2., or by population of 32 chromosomes, 10 of them beeing
using the simulated annealing algorithm. The example changed each generation, a single point 100%
from the Fig.2. shows that only 6 test vectors could crossover and 5% rate mutation. A feasible solution
cover more than 75% from the total number of stuck- has been obtained in less than 50 generations, as we
at 0 faults in the circuit. All these GUIs (and also can see in the Fig.3. and in the Fig.4. The cost is given
those from the first set of experiments) have been now by 3 inverting gates and 6 inputs (one of the
developed in Matlab 5.3. gates in the network is useless), and this solution has
the minimum delay time between any input and the
A. The Implementation of a Boolean Function output of the circuit, in a gate level implementation.
We have considered a boolean function represented in B. The Implementation of a Finite State Machine
a minimal disjunctive form by using a Karnaugh map:
The Finite State Machine (FSM) represented in the
f = x1 x 2 x3 + x1 x3 + x 2 x3 (1) Fig.5. is a sequence detector with one-input, one-
output and 6-internal states. When the input sequence
011 occurs, the output becomes 1 and remains on this
This representation has a cost of 7 gates and 13
logic value until sequence 011 occur again. In this
inputs, including inverters. By applying some
case, the output returns to 0, and maintain this value,
switching-algebra theorems our function may be
until a new sequence 011 appears.
written in the next form:
Initially a GA has been used to find optimal state
f = x3 x1 x 2 (2) assignment. The chromosome represents the FSM as a
list of states. The initial population is generated
Now, the cost of implementation is of only 3 gates randomly. The goal of the GA is to extract the
and 5 inputs. Unfortunately, there is no algorithm to optimum state assignment, which requires the least
find this convenient form of the function, only the number of logic gates. For that reason the number of
heuristics and experience of the human designer. 2-inputs AND/OR logic gates are used to define the
fitness function. The optimum state assignment is
We have tried to find another representation of this given in the Fig.5. A more detailed description of this
function by evolutionary design. We have used the problem is presented in [1].
337
1
X
S/Y S0/0
0
X
0 S1/0
1 0
S2/0
1 D Q Y
1 2
Q
1 S3/1
S0: 000 0 D Q
S1: 010 1
S2: 001 0 S4/1 Q
S3: 100 1 0
S4: 110 D Q
S5: 101 S5/1 0
CLK Q
Fig. 5. A sequence detector described as state
transition graph and GA state assignment Fig. 6. Evolved optimal circuit solution of the
sequence detector
The circuit XCR3064XL, is a Xilinx CPLD with 64 In sequential circuits, the optimal state assignment is
macrocells and 1500 usable gates, providing low- crucial. The sequence detector from the subsection B,
power and very high speed, and beeing in-system implemented with the equations 3,4,5 and 6, has used
programmable through JTAG IEEE 1149.1 Interface. only 3/64 macrocells, 3/224 product terms, and 3/160
Unfortunately, this circuit has only 1000 erase/ function block inputs. The same circuit, implemented
programming cycles guaranteed, so it can not be used with a non optimal state assignment has used 4/64
with intrinsic EHW. macrocells, 9/224 product terms, and 4/160 function
block inputs. Even the combinational time delay is
This programmable circuit is mounted on a board, different for these circuits (4.7ns in the first case and
called Digilab XCRP, delivered by Digilent, Inc. This 7,2ns in the second case). Its true that the main
low cost platform can be used to implement a wide differences in the complexity of these circuits are
variety of digital circuits. The programming pins of given by the state assignment, but it seems that
the circuit are directly connected to the parallel port evolutionary design is more efficient even for the
pins of the computer. combinational part of a FSM.
338
1 0 1 R sw R sw
OUT +5V
1 1 R sw R sw
IN
OUT
1 1
R sw
R sw
+5V GND IN
GND
Fig. 7. Instantiation of the NOT gate on the
evolvable testbed Fig. 8. The equivalent circuit diagram for the NOT
evolved gate
339
Q2 1 Q Q2
1 J
Q2 0
Q1 K Q
Q1 MUX
CLK
Q0
Q0
4053 4027 Q J Q Q1
i
J Q
4051
K Q
CLK
K Q
CLK
4081
CLK J Q Q0
340
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Technical University of Cluj Napoca, 26-28 G. Baritiu street, Cluj Napoca,
0264-401309, fax. 0264-591689, Bogdan.Orza@com.utcluj.ro
341
types of layer implementations without writing any mechanism and reference counter. In this case the
code writing. This way an application that uses TCP server receives the messages from the client at
can be very easily modified so that it will use HTTP predefined intervals of time. When it stops receiving
as a transport channel, thus having a better scalability. messages, the server will release the resources.
342
Classes WClientModel and WServerModel object that will be finally sent to the graphical
correspond to the models from Model View interface.
Controller, WConnection to the controller,
WhiteboardGui and ChatGui to the interface. B. CLASSES AND OBJECTS USED FOR
COMMUNICATION
The communication between the client and the server
IV. COMMUNICATION MECHANISM BETWEEN is made using the WObject objects. From this generic
THE CLIENT AND THE SERVER type we have derived all the other objects which are
used for drawing or text display.
A. COMMUNICATION PROTOCOL
In the first phase the client sends to the server an
authentication request. The server takes the request
and interrogates the database. If the name and WObject
password are correct, the server will return to the
client a connection object that contains some
information for the user (the user type, the name, the
database id etc.). If the authentication is not made the WLine WEllipse WIntMess .....
343
mandatory that these objects inherit the The object storing is made on the server side but also
MarshalByRefObjects class. Instead of transferring a on the client side. We chose this method because we
value that points to such an object, in the network it wanted to keep a low traffic when we want to save the
will be transferred only one type of objects: ObjRef, drawings. The storing is made in lists of ArrayList
objects contain the name of the server/ip and an type, where we can store WObject objects and
identifier, indicating uniquely an object on the server. WObject objects. The user subscribed at the virtual
In the case of IelCom we use both transfer types. The university can open a chat and/or a white board
graphical objects and text are serialized in the XML session for every course that is running, and he can
format and they are sent from the client to the server communicate with any other user through a private
where they are stored on the corresponding server list chat.
and then depending on the destination they have they
For each of these communication methods there is a
are sent to the appropriate client. Every object has two
list where the text and objects of the drawings are
addresses. The first one is the user name of the user stored. The server keeps the corresponding models of
(this name is unique in the data base) and the second every opened course with active users; the client
one is a combination between identifiers representing keeps only the ones where the user is active at that
the address of the client and the address of the moment of time.
graphical interface to which the object is sent. All the
other identifiers are attributes of WObject class, in To distribute the objects from the server to the
such a way that all the derived classes will inherit destination clients it was implemented an algorithm
them. WLogin Class makes the authentication and based in the listener concept. This concept presume
creates the connection object (WConnection) for each that once a connection is made, it is added on a list on
the server (WServerModel) and at the moment an
client. This class inherits the MarshalByRefObjects,
object appears on the server, the list will be read and it
so the transfer is made through the interface. Such a will be sent to the appropriate connections.
method is WConnection doLogin (String user, String From here through a thread the objects will be taken
pass, String type) which checks if the user is in the by an object corresponding to the WClientModel
database and if the answer is affirmative it will create client model and the temporary list will be emptied.
the connection object. The following code sample is an example of the way
Client 1 objects are distributed on connections through the
Whiteboard GUI function wakeUpListners (WObject obj).
doLogin() ChatGUI
Server
whiteboard The listener concept is also used in the case of
sendData() Client 2 graphical interfaces (whiteboard and chat) which are
getData() Whiteboard GUI registered as listeners to WClientModel. So the
Login
ChatGUI
Connection 1 objects that come will be redirected depending on the
doLogin() destination application (the text for chat and the
graphical objects for the whiteboard).
Connection n sendData() Client n
getData() Whiteboard GUI The WClientModel class as the WServerModel class
ChatGUI implements the Singleton pattern, in such a way that
there will be only one instance of every class on the
Fig. 4. The logging mechanism whole application. To obtain the private class
So all the clients will communicate with the server constructor is declared, such that we cannot instantiate
through their communication object, thus the the class outside. First is declared a static attribute of
application is more secure because the client does not the class type, the value of this attribute will be set
have direct access to any data on the server, this initially to null. It will be declared also a static
makes possible the implementation of a more method that will initiate the attribute just once. The
complex security system that doesnt exist at this at uniqueness of the object corresponding to the models
this point in our application. The connections are is very important because it is necessary to be able to
object uniquely identified by two attributes: the client obtain references to them from different points of the
user name and a number generated by the server. application, more then this they offer flexibility
The connections are instances of WConnection class because we can add new modules without doing
having two important methods: sendData (WObject important changes of the application.
obj) that allows transfer of objects from the client to
the server, and addNewObj (WObject obj) through So any graphical interface, or any other module that
which a client receives objects from the server. needs the user identification data, can access them
through the reference to the WClientModel provided
C. CLIENT AND SERVER SIDE OBJECT by the method getInstance().
STORING
344
For a better management of the application we have
defined the WIntMess class. Through objects of that The application was developed using .NetRemoting
class the messages are sent and received from the technology, because it is a good compromise between
server. This objects are not stored on the server, they the bandwidth needed to communicate and the ease of
are used just for the management of the application. implementation. Although it requires a larger
There are two attributes of the class: query and bandwidth than the use of sockets, .NetRemoting
response. When the client wants to obtain some provides the programmer an advanced
information from the server, for instance the number implementation environment, which abstracts the
of student in a course, he will not point directly the transport layer from the OSI model, allowing the
data base because that can cause security problems. transmission of objects through the network and the
The solution is the creation of a WIntMess object with calling of remote methods. In the case of
the query attribute set with the proper message, which .NetRemoting as in the case of Java RMI, due to
is then sent to the server. The server will take the actual security demands, the application configuring
object and analyze the request and after that it will set process is hard enough, because often there are added
the response attribute of the object with the object that new security levels from one version of the frame
holds the information desired by the client. The object work to the other, thus the need of adding new
will be then sent on the connection that came from. information in the configuration files.
This type of object is also use to signal if a client
connects on another machine, if he left the The use of these technologies allows an easier
application, or if he wants to create a new drawing. implementation of the object oriented programming
concepts (polymorphism, inheritance, etc.), it adds
An interesting advantage of the application is that of scalability plus to applications so that one can add
undo. Although for stand-alone application this is
more easily new modules and facilities. This is also
quite a simple thing to implement, in the case of
distributed applications this arise some problems. The the case of IeL Com, the developer can add new
first problem appears when we want to establish the graphical objects, deriving the appropriate classes
way we want to make the undo. There are at least to belonging to WObject, without worrying about the
possibilities: the first one is that the user is able to transmission through the network.
make undo only to the objects that he created, but this
contrast the principle of shared whiteboard, because Bibliography
the users must be able to modify also the work of
others. The second possibility is that the user can [1] I. Rammer, Advanced .Net Remoting (C# edition), APress
make undo on all the objects on the drawing, this is 2002.
[2] A. Turtschi, C# .Net, Syngress Publishing, Inc.
also what we choose for our implementation. So it [3] A. Vlaicu, V. Dobrot, S. Iacob, Tehnologii multimedia,
was created a WIntMess object that sends the undo sisteme, reele i aplicaii , UTCN.
message to the server every time a user hits the Undo [4] B. Orza, M. Givan, S. Cristea, A. Vlaicu, "IEL 2 - an
button from the graphical interface. This message Integrated Solution for Management, Evaluation and
Communication in E-Learning" , International Conference
reach the server, the server will update the model Advanced tools for E-learning in the Environmental Education, 12-
using the updateModels (WIntMess mess) method, 13 February, Napoli, Italy.
which will eliminate the last added object. The [5] B. Orza, M. Givan, S. Cristea, A. Vlaicu, "Integrated solution
message is also sent on the connections corresponding for management, evaluation and communication in distance
education systems ", Optimization Of Electrical And Electronic
to the online users through the wakeUpListeners Equipment Optim 04, May 20-22, 2004, Brasov, Romania
(WObject obj).
V. CONCLUSIONS
Taking into account the continuous growing of the use
of computers in the academic environment, such an
application (chat and whiteboard) is a very useful tool
that can be successfully used for distance education.
At the moment the application doesnt need a large
bandwidth for transferring the information from the
client to the server, so it can be used even with poor
internet connections (e.g. dial-up).
345
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract A low-cost electronic board (Laborplatine) reduced program according to the Bologna process
connected to a PC can be used to simulate the determines an imperative efficiency improvement
facilities of a digital oscilloscope, function generator, of all teaching activities, including practical
voltmeter etc. Using this board in order to achieve training. In this paper the utilization of the
real signals from electronic circuits, one can organize
frontal and/or guided experiments which improve
electronic board in teaching electric and electronic
efficiency of electronics laboratory This can be circuits through frontal experiments is suggested.
important in the context of a reduced teaching
program for disciplines like Electric Circuits or II. THE ELECTRONIC BOARD
Fundamentals of Electronics.
Keywords: electronic board, electronics laboratory, The electronic board presented in fig.1 was
frontal experiments developed at the University of Applied Sciences
Karlsruhe, Germany [1], [2]. The size of the board is
I. INTRODUCTION 160mm x 100mm and it can be connected for data
transmission to the serial port of the PC. The
The basic idea in developing the electronic board
minimal requirements for the computer are: CPU
presented in this paper was to offer a cheap
frequency greater than 200MHz, free memory more
solution, so that students interested in electric and
than 20MB, 17 CRT monitor or 15 Notebook with
electronic circuits can do their own experiments at
1024 x 768 pixels, Windows 95, 98, XP or NT.
home, as individual study [1]. However, a future
1
Facultatea de Electronic i Telecomunicaii, Departamentul Electronic Aplicat,
Bd. V. Prvan Nr. 2, 300223 Timioara, sabin.ionel@etc.utt.ro; marlene.daneti@etc.utt.ro
346
Fig. 2 A representation of the electronic board on the PC monitor
The reduced price (100) of the electronic board area network. Combining problem solving (specific
implies certain constraints. For example, the sampling seminary activity) with PSpice simulation and
frequency of the two input signals is only 2 MHz. frontal experiments using the electronic board, one
This is, however, more than enough for experiments can assure better understanding of theoretic and
on basic electric and electronic circuits. The supply practical issues regarding electric and electronic
voltages for the electronic board are +12V (150mA) circuits. This approach can accelerate the learning
and -12V (60mA). The software for the PC (written in and save time. Groups of two or three students
the HP-VEE language) as well as the latest update can guided by the laboratory assistant in their work
be downloaded from internet [2]. with the electronic board will also realize the
importance of cooperation and teamwork. The
III. THE IMPLEMENTED DEVICES feedback from the students is also important, since
they can raise interesting problems. Our experience
The electronic board itself can deliver a dual supply shows that students are better motivated to do
voltage for the circuit under experimentation: (0 to 12) simulations and experiments using the PC as main
V and (-0 to -12) V. tool, than working with several different measuring
A DC-voltmeter can measure constant voltages instruments. Certainly, the frontal and guided
in ranges from 2V to 20V. A DC-ampmeter is also experiments under the control of the laboratory
implemented for constant currents in ranges from assistant must be continued with individual
60A to 200mA. assessments and hands-on exercises developed by
The implemented function generator (quartz the students in their own free time.
controlled PLL synthesizer) delivers, via BNC,
usual signals (sinus, rectangle, triangle and positive V. CONCLUSIONS
or negative sawtooth) as well as customized
waveforms (programmable function). The presented electronic board is a cheap and
Certainly, the most important instrument suitable solution for experiment-based teaching and
implemented on the electronic board is the digital learning of electric and electronic circuits. It can be
oscilloscope. The impedance of the two Y1, Y2 inputs utilized not only for individual study, but also in
(via BNC) is 1M||13pF and the bandwidth for each organized laboratory and classroom activities.
channel is 2MHz (-3dB). Both inputs have DC, GND Especially, frontal and guided experiments based
and AC (time constant 0.1 sec) facilities. One can choose on this electronic board can improve teaching
the sensitivity in the following steps: 10, 20, 50, 100, efficiency, in comparison with classical laboratory
200, 500mV/Unit and 1, 2, 5V/Unit. Important functions training utilizing expensive measuring instruments.
like external triggering (BNC output), FFT and signal
averaging are also provided. Fig. 2 shows a REFERENCE
representation of the electronic board on the PC monitor.
[1] R. Koblitz, Neue Laborplatine, fh-magazin WS2003/2004,
24 Jahrgang, Nr.48, pp.72.
IV. FRONTAL AND GUIDED EXPERIMENTS [2] http://www: /fbeit.fh-karlsruhe.de/laborplatine/
347
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
RS232
DVM DVM (standard)
PC
Power supply
IEEE 488
348
The planning and handling part of the application has the professor, by e-mail, a complete report concerning
been developed in HTML and PHP, and the data base the experiment carried out.
for the data about the users has been implemented in
MySQL. When the user accesses the e-learning site, a
window appears and asks the user for identification
(fig. 3).
III. CONCLUSION
349
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract - Because it is very important to have a proper understanding of the educational process of each stu-
image of all activities dealing with the field of electronic dent. [2]
packaging, the aim of the current paper is to present a As part of our efforts to increase the quality of educa-
model of how to monitor educational and research activi- tion in Electronic Packaging, our team has developed
ties in an Integrated CAD Environment. The system
described in this paper is a client/server application, has
a flexible software platform which consists of two
an easy to use interface, and a secure database. Such a operational programs: a server program and a client
system is a necessity in an academic environment be- program and the client program has three modules: for
cause is a tool for a better evaluation, for making statis- monitoring laboratory activities, applications opened
tics with the most accessed sites, for avoiding troubles on workstations and web traffic.
made by students who speculate systems or networks
weak points. II. GENERAL ASPECTS OF THE SOFTWARE
PLATFORM
Keywords: monitoring application, client/server arhitec-
ture
Why would be necessary such a monitoring instru-
ment? A few reasons could be: the pursuit to under-
I. INTRODUCTION
stand the assimilated information, the improvement of
student evaluation system, research activity not to be
The problems of electronic packaging may cover a
disturbed by other unwanted activities (file deletion,
large area and represents "The engineering discipline
file coping, database altering, etc.).
that combines the engineering and manufacturing
technologies required to convert an electrical circuit
The monitoring is not a restrictive instrument for the
into a manufactured assembly. These include at least
activity or the research direction chosen by students; it
electrical, mechanical and material design and many
is an early correction and rectifying instrument of
functions such as engineering, manufacturing and
eventual errors in the training process.
quality control." [1]
The software application MA-ICADE (Monitoring
Shaping the students for designing and manufacturing
Activities Integrated CAD Environment) is a cli-
of electronic packaging involves an interdisciplinary
ent/server application which can be accessed and used
and multidisciplinary approach. In this context, there
from school laboratories. A client is a computer that
are at least two important aspects: one of them is the
queries, through a message, another computer, named
teaching and the other one is the evaluation of the ac-
server, for information or services. The server,
quired knowledge. These two activities concern all the
through another message, delivers the information or
students involved in educational activities, especially
services to the client, the whole operation being trans-
in CAD-CAM-CAE because they define the high per-
parent to the user.
formance in human resources. All this implies persons
responsible for education with a lot of qualities as
Programming languages as Delphi, PHP (recursive
high qualification, creativity, patience and talent. Be-
acronym for "PHP: Hypertext Preprocessor") and
yond this, sometimes the educational process involves
HTML and also database free tools as MySQL and
a great amount of work routine, which could lead to
Apache Server were used in order to develop MA-
professional one. Many of these problems could be
ICADE Software. In order to monitor the web traffic
avoided by using a Computer Added Instruction
was used SQUINT a free application from the Inter-
(CAI). Through such a software application the com-
net because it offers daily, weekly and monthly re-
puter can take over a large part of
ports of the Internet activity in the network.
the routine and consequently teachers could afford to
spend more time for creativity and for a better
Fig.1 presents a logical diagram of the MA-ICADE
SOFTWARE APPLICATION that emphasizes the
1
University Politehnica of Bucharest.
350
main parts of the application and the relationships workstation is on, or if the server program is on. If
among module. not, an error message will be post on the professor
display.
IV.CLIENT PROGRAM
III. SERVER PROGRAM To better understand the way in which this program
works, we have to look at the functional diagram of
For the tests, the server program was installed on 16 the client program.
workstations. It starts with the operating system and
runs in the background. Students can see an icon in
the task-bar and they know their activity is monitored.
351
Fig.5 Functional diagram of the client program
B. AOW MODULE
This program had been done using Delphi, PHP and
HTML: contains Delphi written code for laboratory Another functionality of the client program is referring
activity and local stations supervision, contains PHP to the applications opened on each workstation. It
written code for database administration, contains shows application lists opened in 16 list boxes. It has a
HTML written code for the applications help file. popup menu for saving the application list. Not Con-
nected or Error messages are printed. If students are
A. MLA MODULE working with application which are not designated for
training activities they will be seen and in the future
The students knowledge evaluation is one of the most the teacher will have the possibility to close the un-
important activities of the didactic process. The MLA necessary applications. The friendly interface permits
(Monitoring Laboratory Activities) module is particu- the switching between modules. Buttons from the pre-
larly useful for teachers and lecturers during this activ- vious windows can be accessed as well as other func-
ity, as it helps, prevent time losses, subjectivism dan- tionalities of the application
ger and occurrence of routine and redundancy, during
the students' knowledge evaluation.
352
C. WT MODULE
The application is supported by any Windows version.
For monitoring web traffic, it was chosen the Squint The activity can be monitored from any computer in
program. It presents daily, monthly and yearly reports the network which has the client program installed.
of Internet activity. The following information is re- The screen resolution is unimportant.
corded: the number of downloaded bytes, the time on-
line, the longest session of a user and a table with the REFERENCES
time of Internet access. It presents a complete image
of the accessed sites as well as links to those pages, [1] Charles A. Harper and Martin B. Miller - "Electronic Packaging
Microelectronics and Interconnection Dictionary", McGraw-
time and downloaded pages Hill, 1993, page 67
[2] Glaser R. - Programmiertes Lernen und
Unterrichstechnologie, Franz Cornelsen Verlog 1971.
[3] A. Drumea, P. Svasta, V.D. Ene, Some Aspects of Educational
Environment in Packaging, XXV International Conference IMAPS
Poland 2001, Rzeszow Polanczyk, 26-29 September 2001, pp.
31-36
[4] Paul Svasta, Virgil Golumbeanu, Ciprian Ionescu, Norocel Dra-
go Codreanu, Daniel Leonescu, Marian Vldescu, Dan Tudor Vuza
- Training Program in Electronic Passive Components Education,
The 52nd Electronic Components and Technology Conference, May
28-31, 2002, San Diego, California USA, pp. 780-786
[5] PHP Documentation, www.php.net/docs.php
[6] Mysql Documentation, www.mysql.com/documentation
V. COURSE OF DEVELOPMENT
VI. CONCLUSIONS
353
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
One of the major problems with respect to the design II.1. The way it works
and management of any emergency system is the
ability to quickly handle any distress call. The faster The overall system is composed of mobile LoRD
the better is the main principle that governs the (Location Retrieval Device) units located in cars and
functioning of any emergency service. Only a quick a central base station that receives all the emergency
response to fires, car crashes or accidents can notifications. They communicate by means of SMS
diminish the risks of human injuries or fatalities. (Short Message Service) messages. The functioning
According to the US Bureau of Transportation of the system is presented schematically in Fig.1.
Statistics, car crashes are an important issue of public
safety around the world, with a number of 42,815
fatalities and 2,925,758 injured people in 2002, in the
US [1]. The victims chances of survival can be
greatly improved with an accurate response, implying
that the paramedic units should be notified as soon as
possible about the precise location where an accident
took place. The lack of precision means wasting
extremely precious time.
1
Technical University "Gh. Asachi", Faculty of Electronics and Telecommunications, Iasi, Romania, e-mail: mdobrea@etc.tuiasi.ro
2
Student, Technical University "Gh. Asachi", Faculty of Electronics and Telecommunications, Iasi, Romania, e-mail: nikcleju@gmail.com
3
Student, Technical University "Gh. Asachi", Faculty of Electronics and Telecommunications, Iasi, Romania, e-mail: astek@personal.ro
4
Student, Technical University "Gh. Asachi", Faculty of Electronics and Telecommunications, Iasi, Romania, e-mail: alex584@apropo.ro
354
Format of the Configuration Message
Security Code Time Interval (x 10 seconds) Temperature Limit
Before using any LoRD module, it must be remotely II.2. Security and reliability features
configured by the central base station. This is done by
sending a message to it that contains the necessary As the system handles critical data, the performance
information: the phone number to which the LoRD requirements are very high, and we had to consider
will reply, the time interval between two consecutive reliability and security issues from the early stages of
messages and the temperature limit which will trigger the project.
the thermal alarm. This process is allowed only if one
specific jumper is set. After a successful The module performs numerous self-tests to ensure
configuration, unsetting this jumper causes the LoRD the reliability of the information supplied. Any error
to reject any other configuration attempt. detected is reported in the SMS sent to the base. Any
phone error is visually signaled to the driver.
In normal mode, the LoRD module continuously
monitors for a possible abnormal situation. The events To ensure a prompt reaction in case of an accident the
that trigger the alarm are the car security systems alarm signals and the entire serial communication are
(opening of the airbags, car alarm, or panic button) or processed through the microcontrollers system of
the detection of a fire, when the temperature exceeds interrupts. The alarm signals have the highest priority,
the previously programmed limit. Therefore the so as soon as a collision occurs the LoRD begins to
LoRD gathers information about the car status and send the emergency messages.
location and uses the attached mobile phone to send it
through a SMS to the base station. The message The SMS messages by which the LoRD modules and
contains the coordinates of the car, obtained with a the base station communicate with each other contain
GPS Receiver, the temperature, in order to detect a a certain security code, computed from the senders
possible fire, and information about the extent of the own number. By checking this security code against
cockpit damage, Fig.2. If it is not damaged, the LoRD the senders phone number, as it is reported by the
continues to send messages periodically, allowing the phone network, the receiving device is able to detect
base station to monitor the evolution of the situation. and reject fake messages.
There is also a remote control mode, as an alternative The whole system is designed to be infrastructure-
function of the LoRD. The base station may send a independent in order to have the highest level of
request to receive alert messages, although no generality and to be easy extendable. The
accident was reported. This can be useful, for communication process is independent of the type of
example, in tracking down the LoRD in scenarios like phone network, because all of the GSM, CDMA and
mountain rescues, fire-fighting or retrieving stolen TDMA technologies support SMS technology. The
vehicles. The base station sends a START message, LoRD module, the mobile phone, and the GPS
which will enable the LoRD to send back notifying Receiver have their own power supplies, which are
messages. The analogue command STOP will switched on only in case of a crash; the rest of the
disable any message sending. time they are powered by the car battery.
The central station runs the software that manages all Using the dedicated MySQL database server adds to
the incoming messages. It is developed in the the overall independence and security. The server can
LabWindowsTM /CVI version 5.5 integrated reside on a remote host (MS Windows, Unix etc.), and
environment and it uses a dedicated database server to communicate through SSL secure TCP/IP protocols.
securely store the information. Upon receiving an We decided upon a database server because its
emergency message, the software reads it from the security, platform-independence and the capacity to
attached mobile phone and presents it to the human handle and store data with minimal risks of failure
operator. make it an excellent choice.
355
III. HARDWARE IMPLEMENTATION OF THE maximum 10 s with the frequency given by the
LORD MODULE formula:
356
IV. THE SOFTWARE OF THE LORD MODULE error is visually signaled using a LED.
It verifies the temperature sensor by writing a
IV.1. Overview certain bit pattern to a location in its memory.
Receiving acknowledge bits assures of the proper
The microcontroller software was developed in functioning of the I2C bus, and the correct
assembly language, as it is faster than a similar retrieval of the pattern validates the sensor circuit
program developed in a C environment. The program itself.
flow is represented in Fig.3. It verifies the presence of the GPS Receiver by
polling commands [4] that ensure the device is
Upon startup, the LoRD runs an initialization routine, connected and working properly, as it must have
and then it normally remains in a loop where it acquired at least three satellites in order to be able
performs various self-tests and component-tests. to function. In case the Receiver acquires no
Whenever an external interrupt is activated, the signal, it returns null instead of the coordinates,
hardware interrupt routines are executed and they set in which case an error flag is set.
an acknowledgement semaphore. The software
verifies the status of these acknowledge flags and if This continuous testing allows the fastest response in
an alarm is indicated it runs the message sending case of an emergency. Whenever an external interrupt
routine. is activated, signaling an emergency, the LoRD does
not waste any more time on tests. The current state of
IV.2. The Initialization Routine the device is already determined, so it can start
sending the message as soon as possible.
The initialization routine starts with a self-test it
erases all the locations in the internal memory and IV.4. Processing Hardware Interrupts
verifies if they are valid. Then it enables internal
timers, interrupts, serial communication and initializes There are two causes that will determine the
external devices. microcontroller to enter the alarm state: either a signal
from the car interface (the opening of the airbags or
The AT89C4051 microcontroller has two timers. One other alarm systems) activates External Interrupt 0, or
of them, TIMER 0, is used to set the time interval the temperature exceeds a critical level, activating the
between two consecutive messages, and the other one, External Interrupt 1.
TIMER 1, must set the default 9600 bps baud rate
used by the microcontroller to communicate through We used the entire interrupt system in a way that
the serial port. enables the LoRD to react as promptly as possible to
these situations.
To ensure the priority of alarm operations, we used all
of the microcontrollers interrupts in the following We had to analyze how to obtain the fastest response
way: to an emergency, still doing all the necessary tests in
order not to compromise the reliability degree of the
Two external interrupts: they have the highest
provided information. (i.e. the LoRD should detect
priority [3], and the activation of any of them will
and report any temperature sensor error, because the
determine the microcontroller to send alarm
base station should know that the temperature
messages.
readings may be incorrect). The following issue,
Two timer interrupts [3]: they increment some
which cannot be avoided, is very important: if the
time counters each time the interrupts are
interrupt is requested during any serial
activated;
communication, the communication process must be
The serial interrupt: it has the lowest priority and allowed to finish. Sending the emergency SMS
it is activated each time a character must be sent requires serial communication with the phone, so
or received [3]. simply cutting off the previous communication may
issue scrambled commands to the phone or the GPS
IV.3. Testing Routines receiver. Our solution is presented in Fig.4.
While it remains in a continuous loop, the When an interrupt takes place during a test that
microcontroller performs periodical tests to check the involves serial communication with a device, the
system integrity: interrupt routine is executed but it only sets an
It checks for the presence of a mobile phone by acknowledge semaphore. The software is therefore
sending the Hayes AT command and verifying allowed do terminate the current communication
that the answer received was the expected OK process. After every test, the program checks these
(see section IV.5. Serial communication semaphores and, if it determines that an interrupt took
routines). This test is the most important of all, place, runs the routine which sends the emergency
because without the phone the LoRD cannot send message. This way, the maximum possible delay
any emergency message. That is why any phone between signaling an emergency and the actual start
357
Fig. 4. Processing the interrupt requests
of the message sending routine is the maximum time messages received from the LoRD units. It also
spent in a testing subroutine. allows easy configuration of the new units and
reconfiguration of the existing ones. The software
The other alternative means doing all the tests after controls one mobile phone attached to the serial port
the interrupt request, to ensure the reliability of the of the PC, which is used to send and receive
information. This implies that the delay until the messages. Created in the LabWindowsTM /CVI
sending of the message would be the sum of all tests version 5.5 integrated environment, the software
durations. This clearly increases the response time and displays information about received messages, while
hence the risks of failure during a crash. in the background it handles receiving and checking
the arrived messages, extracting critical information
IV.5. Serial Communication Routines from them and inserting it into a database. The human
operator is asked to assign the best rescue unit
The communication with the mobile phone is based available to answer the distress call.
on the assumption that the phone supports the Hayes
commands. These are ASCII commands that are The database server that stores and manages the
transmitted over a RS232 connection and allow using information is the MySQL 4.0.18 server. The software
the phones capabilities. We were interested mostly in communicates with the server using the MySQL C
the ability to send and receive SMS messages, but API, sending SQL commands for retrieving or
also to read phone numbers stored in the SIM card. inserting data.
Sending and reading a SMS message also implies
converting the text to and from a special format, the
PDU (Protocol Description Unit) format, which the
phone uses.
358
VI. SUMMARY
V.2. Communication with the phone
In this paper we present a general overview of the
This is accomplished using the Hayes commands system we have developed, as well as provide detailed
that the phone supports. On startup, the software information about its functioning parameters.
opens the serial port of the PC and checks if a phone The LoRD is designed to meet all requirements an
is present. If so, it attempts to do a necessary emergency system has. Providing important, real-time
configuration of the phone. This essentially tells it to data about the site of any kind of accident, the
send a short notifying line of text over the serial warning system we built instantly notifies the rescue
connection whenever a new SMS message is received. units of the extent of the danger someone could be in.
Then the software enters an idle state, when it It also has many built-in security and user-friendly
permanently checks the phones status while capabilities.
monitoring both the serial port for notification alerts
and the interface for the human operators actions. Our system was built to prove the fundamental
concept of our project: gathering remote information
When the phone receives a SMS from the network, it and sending it to a base station the instant an accident
sends a notifying text over the serial connection. Then occurs. Further developments, like integrating more
the software obtains the message after a series of different types of sensors car integrity indicators,
stages, which include decoding it from the PDU smoke detectors etc. may be carried out. In the end,
format and checking the security bytes. it is even possible to fully embed the LoRD module
into the car computer system using phone and GPS
To send a SMS, the software uses the same process OEM modules.
mentioned above, only reversing the stages. When the
human operator wants to configure / reconfigure a Amateur hikers, mountain rescue units, firefighters on
LoRD unit or to send a start / stop command to it, the a mission, all of them can benefit from using our
software sends the necessary message to the unit, LoRD module without many modifications. Traffic
containing the configuration options or the command. jams and chain collisions can also be detected early
and partially avoided. Furthermore, the remote
V.3. Communication with the database server activation mode supports tracing applications like
retrieving stolen vehicles, detecting cars exceeding
To ensure proper storing and handling of the critical speed limits using the facilities of the GPS, or
emergency information, as well as fast access and tracking people who may be lost. The social
sorting, we decided to use a dedicated database server. usefulness of the LoRD system is based on the large
Using an existing database server ensures a very low number of different applications that call for such an
rate of possible failures, as its strong and weak spots emergency management system, being implied by the
are already known and the server is used by many potential to minimize injury extent and to save human
other applications and web services. lives.
359
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
The OFDM (Orthogonal Frequency Division II. TIME DOMAIN AND FREQUENCY DOMAIN
Multiplexing) systems are known to provide good CHARACTERIZATION OF MIMO-OFDM
spectral efficiency along with mitigating the effects of CHANNEL
frequency-selective fading by transmitting the useful
information over separate subchannels. Most of the If no intersymbol interference occurs, i.e. the cyclic
communication standards state that the same power is prefix length is assumed to be greater than the channel
transmitted over the subcarriers, without any impulse response length and any other
supplementary power control. At the channel output, synchronization errors are neglected (carrier offset,
some of the subcarriers will be more affected by sampling clock offset, phase noise etc.), the input-
fading than others and thus, after equalization the output relation of the MIMO channel may be directly
subcarriers will feature different values of signal-to- expressed in time domain as a cyclic convolution
noise ratio (SNR). product:
This is the main reason for which space diversity has
been used for the past years in order to combat the Nt
effect of fading in different forms. ri ( n, l ) = hij ( n, l ) t j ( n, l ) + ni ( n, l ) (1)
The most common and attractive method because of j =1
its simplicity is space-time coding, which uses a
certain way to introduce redundancy, somehow as in where
the linear binary channel codes. ri(n,l) represents the received sample for symbol
The MIMO channel can be used in order to increase
n, at receive antenna i, at lth time instant
the spectral efficiency in spatial multiplexing systems
context, when different information is sent over tj(n,l) represents the transmitted sample for
different antennas (layers). symbol n, at receive antenna i, at lth time instant,
In both cases, the channel estimator has the same hi,j(n,l) is the propagation path impulse response
structure and it is initially based on training symbols, lth tap, sampled at time n, from antenna j to
either preambles (first one or two OFDM symbols are antenna i,
known at both Tx and Rx) or pilots (a number of ni(n,l) is the noise impinged on receive antenna i,
known subcarriers modulator per each OFDM on symbol n and time l.
symbol). In [2] and [3], some classic channel
1
Facultatea de Electronic i Telecomunicaii, Catedra de
Telecomunicaii Bd. Iuliu Maniu Nr. 1-3, Bucureti, e-mail {aenescu,silviu}@comm.pub.ro
360
Nt is the number of transmit antennas and Nr the Let us define the vectors
number of receive antennas.
ri ( n ) = Ri ( n, 0 ) L Ri ( n, K 1)
T
( )t ( n, l p )
K 1
hij ( n, l ) t j ( n, l ) = hij n, p h ij ( n ) = hij ( n, 0 ) L hij ( n, K 0 1)
T
K0 j K0
p =0
where where
Ri(n,k) represents the received sample on
subcarrier k, for symbol n, at receive antenna i, d 0,1 ( n ) L d 0, Nt ( n )
Tj(n,k) represents the transmitted sample on
D (n) = M O M (8)
subcarrier k, for symbol n, at transmit antenna j, d
Hi,j(n,k) are the propagation path transfer K 1,1 ( n ) L d K 1, Nt ( n )
function, sampled at time n, normalized
frequency k, from antenna j to antenna i, and
Ni(n,k) is the noise impinged on receive antenna
i, on symbol n and subcarrier k. WK ( l ) = WKl 0 WKl 1 L WK (
l K0 1)
(9)
Nt is the number of transmit antennas and Nr the
number of receive antennas.
In (3), k takes values from 0 to K-1, where K is with
the total number of subcarriers.
Obviously, the signals from (3) are strictly dl , q ( n ) = Tq ( n, l ) WK ( l ) , l = 0,K , K 1, q = 1,K , N t (10)
related through a Discrete Fourier
Transformation with the samples from (1). j
2
In (9), we denoted by WK = e the Nth root of the
T j ( n, k ) = DFT {t j ( n, l )} ( k )
K
unit value.
Ri ( n, k ) = DFT {ri ( n, l )} ( k ) (4) The problem of estimating the channel taps can be
H ij ( n, k ) = DFT {hij ( n, l )} ( k )
equivalently formulated to finding the optimum set of
taps h i ( n ) , which minimizes the least squares
The channel estimator must thus compute a total
(LS) cost function
number of KN t N r values of channel transfer
function coefficients, or a number of K 0 N t N r 2
J i ( n ) = ri ( n ) D ( n ) h i ( n ) (11)
taps of the channel impulse response in time
domain.
Based on the above relation, the paper develops From [2], we get the solution of the problem,
an improved version of a least squares based given by the normal matrix equation
time-domain channel estimator, that makes use
of a conveniently chosen set of training symbols, Q ( n ) h i ( n ) = pi ( n ) (12)
by extending the estimation procedure in the
tracking phase, when no training symbols are where we denoted
available.
Q ( n ) = DH ( n ) D ( n ) (13)
III. TIME-DOMAIN CHANNEL ESTIMATOR
361
1 Nt
(23)
H
is the Hermitian operator (T*). h i( r ) ( n ) = p(i ) ( n ) Q r , j ( n ) h i( j ) ( n 1) , r = 1, N
r
K t
j =1
The complexity of the estimator derived in (12) j l
may be further reduced when choosing an One should notice that the matrices p, Q are not yet
appropriate modulation scheme for the training known at the current iteration, since they depend on
symbols [2] i.e. current transmitted symbols, according to (13), (14).
The transmitted symbols will be detected using the
previous channel taps from the equation (3),
T j (n, k ) = T1 (n, k )WK ( j 1) k , N , j = 2, N t (15) expressed by matrix operands
K r% ( n, k ) = H
% ( n, k ) t% ( n, k ) + n% ( n, k ) (24)
= (16)
Nt
in which we denoted
where takes the integer part of the number.
r% ( n, k ) = R0 ( n, k ) L RNr 1 ( n, k )
T
(25)
(12) assumes a constant modulus constellation.
In this case,
t% ( n, k ) = T0 ( n, k ) L TNt 1 ( n, k )
T
(26)
Q ( n ) = KI N t K 0 (17)
H 0,0 ( n, k ) L H 0, Nt ( n, k )
and the estimated solution becomes
H ( n, k ) = M O M (27)
H ( n, k ) L H
1 Nr ,0 N r , Nt ( n, k )
h i (n) = p i (n) (18)
K The decision for t% ( n, k ) may be taken in a
similar LS matrix equation solving, where fast
Tracking phase
algorithms (e.g. QR decomposition) may be
The tracking phase updating relation is used.
t% ( n, k ) = H % ( n, k ) 1 r% ( n, k )
% H ( n, k ) H (28)
Nt
Q ( n ) h ( n ) = p( ) ( n ) ,
r, j
( j)
i i
r
r = 1, N t (19)
j =1 After solving (28), we compute all the matrices p, Q
and solve (23).
If the symbols are of constant modulus, without In Fig. 1, we present the estimator block diagram in
generally fulfilling the constraint (15) then training phase, without explicitly showing the
decision block [5].
Q j , j ( n ) = KI K 0 (20)
IV. IMPROVED CHANNEL ESTIMATOR
and
Our idea arises from the natural observation that (16)
is a linear system that may be iteratively solved, once
h i( r ) ( n ) =
1 (r) Nt
p i ( n ) Q r , j ( n ) h i( j ) ( n ) , r = 1, N (21) having a primary estimated solution set. Suppose that
K we solve the linear system at a rate Tsys and that the
t
j =1
j l OFDM symbol period is Tsym. Then the number of
affordable iterations will be
If when computing pi( r ) ( n ) and Q r , j ( n ) we use
the detected appropriate symbols Tj(n,k) and if Tsym
I = (29)
we assume that the channel varies slowly enough Tsys
as to fulfill
Obviously, this number can be increased up to
hi( j ) ( n ) hi( j ) ( n 1) (22) any value in a pipeline based structure. However,
the price to be payed is reflected in the
then complexity increasing with a factor equal to I.
362
Fig. 1. Channel estimator in the training phase
363
The channel model is the one used in [5], model
3.2.1 and 3.2.2, with TRMS/Ts=2, the Doppler offset
fd=1kHz.
We intend to study the variation of MSE with
respect to the number of iterations.
364
improved, especially when the signal to noise ratio
is high. Theoretically, when the number of
iterations becomes larger, MSE in tracking phase
approaches MSE in the training phase.
The estimator itself features poor performances
when using non-constant modulus constellations
(QAM-type).
The next step in our study will be simulating the
algorithm in finite precision and elaborating an
efficient digital structure with low speed and low
area.
REFERENCES
[1] R. Narasimhan, Performance of Diversity Schemes for
OFDM Systems with Frequency Offset, Phase Noise and
Channel Estimation Errors, IEEE Transactions on
Communications, vol. 50, no. 10, Oct. 2002
[2] Ye Li, Nambirajan Seshadri, Sirikiat Ariyavisitakul, Channel
Estimation for OFDM Systems with Transmitter Diversity in
Mobile Wireless Channels, IEEE Journal on Selected Areas in
Communications, vol.17, no.3, Mar. 1999
[3] S. Coleri, M. Ergen, A. Puri, A. Bahai, A Study of Channel
Estimation in OFDM Systems, Globecom 01, Vol. 1, pp 136-140
[4] W. Bai, C. He, L. Jiang, H. Zhu , Blind Channel Estimation
in MIMO-OFDM Systems, IEEE Transactions on Broadcasting,
vol. 48, no. 3, Sep. 2002
[5] A.A. Enescu, Space-time coded OFDM communication
systems, B. Sc. Thesis, UPB, June 2003
[6] W. Bai, C. He, L. Jiang, H. Zhu , Blind Channel Estimation
in MIMO-OFDM Systems, IEEE Transactions on Broadcasting,
vol. 48, no. 3, Sep. 2002
[7] S. Zhou, G. Giannakis, Finite Alphabet Based Channel
Estimation for OFDM and Related Multicarrier Systems, IEEE
Transactions on Communications, vol. 49, no. 8, Aug. 2001
[8] Ye Li, Leonard Cimini, Nelson Sollenberger, Robust
Channel Estimation for OFDM Systems with Rapid Dispersive
Fading Channels, IEEE Transactions on Communications,
vol.44, no.9, Jul. 1998
[9] IEEE 802.16 Broadband Wireless Access Working Group,
Channel Models for Wireless Fixed Applications
365
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Military Technical Academy, Bucharest
2
Politehnica University of Bucharest
366
II. REVIEW OF DCSK
367
Fig.4. Transmission scheme for the MA-DCSK communication system
E
1 2
variance N 0 / 2 . For each user, the signal received E 2
+4 N b + 2 b
(7)
during a reference sample slot will correlate with N0 N 0
the signal at the corresponding data sample slot.
Depending on whether the output is larger or where Eb denotes the average bit energy,
smaller than the threshold, a +1or -1 is
Eb = 2 Ps , Ps = E ( xk ) and
2
decoded. Such a correlator-based DCSK receiver is
shown, also, in Fig. 5. Note that the sampling
switch only operates during the second half of each
var xk
( j) 2
( )
frame and the threshold detector produces a ( j) = 2
(8)
Ps
decoded symbol after each slot during that period
of time. Clearly, the performance of the MA-
Consider the j th user and the received signal DCSK system is affected by
1. spreading factor 2 ;
{( x ) } for a given P ;
during the l th time slot. Suppose the slot 2
( j)
corresponds to a reference-sample slot for the j th 2. variance of k s
368
1. Change the spreading factor 2 until the
optimal BER is obtained;
( x( ) )
2
2. Minimize the variance of k
j
for a
fixed Ps ;
3. Reduce the number of users N .
V. CONCLUSIONS
REFERENCES
369
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
AbstractIn this paper, a low complexity technique for reducing ISI [2], [3]. Moreover, it has
decision feedback equalizer (DFE) appropriate for been shown that the DFE structure is particularly
channels with long and sparse impulse response (IR) is suitable for multipath channels, since most part of ISI is
studied. Such channels are encountered in many high- due to the long postcursor portion of the impulse
speed wireless communications applications. It is shown
that, in cases of sparse channels, the feedforward and
response (IR). Recall that an important feature of the
feedback (FB) filters of the DFE have a particular DFE is that the postcursor ISI is almost perfectly
structure, which can be exploited to derive efficient cancelled by the feedback (FB) filter, provided of
implementations of the DFE, provided that the time delays course that the previous decisions are correct. Since the
of the channel IR multipath components are known. This postcursor ISI is cancelled by the FB filter, a relatively
latter task is accomplished by other technique, which shorter feedforward (FF) filter is adequate to reduce the
estimates the time delays based on the form of the channel remaining ISI. Moreover since noise is involved only in
input-output cross-correlation sequence in the frequency the output of the FF filter, the DFE exhibits less noise
domain. A distinct feature of the resulting DFE is that the enhancement effects as compared with linear equalizers.
involved FB filter consists of a reduced number of active
taps that implies some computational savings than
In high-speed wireless applications, of the type
conventional DFE. The resulting DFE also exhibits, described above, the implementation of a DFE
improved tracking capabilities and faster convergence as algorithm becomes a difficult task for two main reasons.
compared with the conventional DFE, that implies a First, due to the small intersymbol interval, the time
shorter training sequence . Moreover, the new algorithm available for real-time computations is very limited.
has a simple form and its steady-state performance is Second, due to the long span of the introduced ISI, the
almost identical to that of the conventional DFE. DFE must have a large number of taps, which implies
heavy computational load per iteration.
Index Terms Adaptive equalizers, decision During the last decade there have been many
feedback equalizers (DFEs), multipath channels.
efforts in different directions toward developing
efficient implementations of the DFE. As such
directions, we mention IIR methods, block adaptive
I. INTRODUCTION implementations, efficient algebraic solutions, modified
DFE schemes, etc. [4][12]. As mentioned above, in the
In many wireless communication systems the applications of interest, the involved multipath channel
involved multipath channels exhibit a long time has a discrete sparse form. Efficient DFE schemes
dispersion, and delay spreads of up to 40 s are often which exploit the sparseness of the channel IR have
encountered. A typical application of this is high been derived in [13][15].
definition television (HDTV) signal terrestrial In this paper, a new DFE algorithm,
transmission, where the involved channels consist of a appropriate for sparse multipath channels is studied,
few non-negligible echoes, some of which may have proposed in [19]. The algorithm consists of two steps.
quite large time delays with respect to the main signal In the first step, the time delays of the multipath
(see for instance the HDTV test channels reported in components are estimated in a novel way by properly
several ATSC documents and summarized in [1]). If the exploiting the channel IR form [16]. In the second step,
information signal is transmitted at high symbol rates the DFE is applied, with the FB filter having a
through such a dispersive channel, then the introduced significantly reduced number of taps. These taps are
intersymbol interference (ISI) has a span of several tens selected so as to act only on time positions associated
up to hundreds of symbol intervals. This in turn implies with the estimated time delays of the involved multipath
that quite long adaptive equalizers are required at the components. A distinct feature of the novel approach
receivers end in order to reduce effectively the ISI followed in this paper is that the required channel
component of the received signal. Note that the parameters are the locations of the multipath
situation is even more demanding whenever the channel components. This is opposed to most of the existing
frequency response exhibits deep nulls. works [14], [15], in which the whole channel IR has to
The adaptive decision feedback equalizer be initially estimated. Moreover, the relation between
(DFE) has been widely accepted as an effective the active FB tap positions and the echo time delays is
it offers an additional saving in bandwidth. Note that its The overall channel IR, including the combined
overall complexity is of the order of the number of
multipath components and hence it is, in practice,
transmitter and receiver filters response, say , p ( )
several times lower as compared with the conventional can be written as
h ( ) = i e ji p ( i )
DFE.
(3)
The paper is outlined as follows. In Section II,
i
the multipath channel is described and the problem is
formulated. In Section III, the proposed efficient As mentioned in the introduction, in this paper, we deal
method for estimating and tracking the time delays is with sparse multipath channels having a relatively long
presented. The new DFE algorithm is developed in IR. Due to the sparseness of the multipath channel IR
Section IV and relevant computational issues are and the form of the pulse shaping function p ( ) ,
discussed. In Section V, the new algorithm is tested and
some indicative experimental results are provided. which decreases rapidly, the overall symbol spaced
Section VI concludes the work. channel IR remains sparse and can be expressed as
L
h ( nT ) = hnl ( nT nlT ) (4)
II. PROBLEM FORMULATION l =0
hc ( t , ) = i ( t , ) e ( c i i ) ( i ) (1)
j 2 f ( t ) + ( t , )
where {d } is an independent identically distributed
(i.i.d.) symbol sequence with variance d and {w} is
i 2
where i ( t , ) , i (t ) are the real amplitudes and zero-mean complex white Gaussian noise uncorrelated
excess delays, respectively, of the multipath component with the input sequence, with variance w2 . Note that
at time t. The phase term 2 f c i ( t ) + i ( t , ) symbol period T has been omitted for reasons of
371
simplicity. Obviously, { x} suffers from intersymbol an appropriate partitioning of both channel input and
output sequences and is described below.
interference due to the presence of undesired multipath
Let us first formulate the following 2N-DFT
components and in most cases equalization is necessary
for reliable reception. sequences for k = 0,1,..., 2 N 1
As mentioned in the introduction, the DFE N + p 1 2
jmk
structure is particularly suitable for equalizing multipath
channels. The LMS-based adaptive DFE is given by the
D (k ) =
m= p
d (n + m) e 2N
(12)
following set of equations:
2 N 1 2
jmk
X (k ) = x ( n + m) e
0 N
ck ( n ) x ( n k ) + bk ( n ) d% ( n k )
2N
d ( n ) = (13)
m =0
k = M +1 k =1
(7) where p is assumed to be an overestimated value of the
noncausal size of the channel IR (i.e p > k1 ). The
d% ( n ) = f {d ( n )} (8)
same is presumed for the quantity N-p as far as the size
e ( n ) = d ( n ) d% ( n ) (9)
of the causal part of the channel IR is concerned. If
these facts hold true, the method which is described
ck ( n + 1) = ck ( n ) 2 c x* ( n k ) e ( n ) below detects the positions of all precursor and
(10) postcursor components. Note that X ( k ) in (13) is
k = M + 1,..., 0
based on a 2N-length output sequence, while D ( k ) in
bk ( n + 1) = bk ( n ) 2 d% * ( n k ) e ( n )
b
(12) results from an N-length input sequence padded
(11) with zeros. As it will become evident from the
k = 0,..., N subsequent derivation, this is done in order for all
samples of the cross-correlation sequence to be equally
where { x} and {u%} denote the equalizers input and weighted. Indeed, if we consider the expected value of
decision sequences, respectively, ck are the the product of the above sequences, we obtain
coefficients of the Mlength FF filter, and bk are the E { X ( k ) D* ( k )} =
coefficients of the N-length FB filter ( N is taken at least 2 N 1 N + p 1 2
j (i m)k
equal to the channel span [10]). f {} stands for the = E { x ( n + i ) d * ( n + m )} e 2N
i =0 m= p
decision device function, , are the step sizes and
c b
(14)
* denotes complex conjugation. It is assumed that a
training sequence of appropriate length is available where E {} denotes the expectation operator. If we
ensuring convergence of the equalizer. That is the now substitute(6) to (14), we get
equalizer operates initially in a training mode and then
L
E { X ( k ) D* ( k )} = hnl
switches to a decision directed mode. In the following
sections, first, a frequency domain procedure is
l =0
proposed for detecting the time delays of the multipath
components of the channel IR. Then, a new efficient 2 N 1 N + p 1 j (i m)k
DFE structure is derived, which takes advantage of the E {d ( n + i nl ) d ( n + m )} e * N
special properties of the multipath channel. i =0 m= p
(15)
Since p is larger than the noncausal part of the
III. ESTIMATION OF THE ECHO DELAYS
channel IR, it is easily shown that for every l the indices
of d and d* in (15) are identical for N combinations of m
A well-established nonparametric procedure
for estimating the time delays of the multipath and i (with m = i nl ). Therefore, due to the i.i.d.
components is based on a proper cross-correlation of property of the input sequence (15) is written as
the input symbols with the corresponding channel
L
E { X ( k ) D* ( k )} = N d2 hn e
output samples. In a time domain implementation, the jnl k
N
(16)
estimation of the cross-correlation sequence for N lags l
l =0
requires O ( N ) operations per sample. It is shown
that, an appropriate frequency domain expression of the for k = 0,1,..., 2 N 1 . That is, we end up with a sum
cross-correlation sequence can be viewed as a sum of nl
complex harmonics, with the unknown time delays of complex harmonics at normalized frequencies .
interpreted as frequencies. Thus, to estimate the time
2N
delays, we suggest an FFT-based scheme of complexity Applying the 2N - IDFT to the resulting sequence, the
372
locations nl of the multipath components are Determination of Dominant Components: We
determined at the nonnegligible points of the IDFT. see from (17) that samples RN of { x} and {d } are
Obviously, in a practical situation, time used to compute C
( R)
( k ) . The L+1 IDFT points of
averaging is used instead of E {} in order to
DX
(17) having the highest amplitude are then chosen as the
implement (16). In cases where the channel is assumed desired locations. The number L of the dominant
stationary, the above procedure can be done once during undesired can be computed by setting a threshold and
the training phase and then the obtained time delays can select the locations of the IDFT points of (17) having
be used in the algorithm as described in the next amplitudes which exceed this threshold.
section. Of course, in most situations in practice, the
channel exhibits variations and, thus, the required time
delays have to be tracked continuously. During IV. THE NEW METHOD
tracking, the frequency domain expression of the cross-
correlation sequence is formed using the decisions In the proposed algorithm, we focus our
provided by the equalizer (which operates in a decision attention to the demanding FB part and reduce the
directed mode). computational load by properly selecting O(L) number
Exponentially fading memory is imposed on of taps out of N taps. The main idea behind the
the estimation procedure by including a forgetting derivation of the algorithm is that due to the channel
factor in the frequency domain expression of the sparseness, the FB filter also possesses a specific sparse
cross-correlation sequence as follows: form. After exploiting its sparse form, the FB filter is
R 1 built so as to act only to a restricted set of tap positions.
CDXR = Rr ( k ) Dr* ( k )
( ) ( R 1 r ) N As a result, the algorithm offers significant
(17)
r =0
computational savings while its steady-state error
performance is similar to that of the conventional DFE.
where 0 < 1 and
N + p 1 2 A. Derivation of the Algorithm
jmk
Dr ( k ) =
m= p
N + p m 1
d ( n + rN + m ) e 2N
(18)
It is well known [2], [4] that in the minimum
mean-squared error (MMSE) DFE, the FF and FB
2 N 1 2 coefficients can be expressed in terms of the channel IR
jki
Xr (k ) = x ( n + rN + i ) e
i =0
2N coefficients. Indeed, based on the assumption that
previously detected symbols are correct, the
minimization of the mean-squared error (MSE)
for k = 0,1,..., 2 N 1 . Note that if factor were
E{ e (n) } leads to the following set of equations for
2
included only in (17), then the exponential weighting
would be applied on a block-by-block basis, thus the FF filter c M and the FB filter b N
affecting the tracking capabilities of the new
algorithms. However, additionally including in (18)
1
2
is equivalent to applying an exponential window in the c M = H1H1H + w2 I M H1e M + k1 (20)
time-domain sample-by-sample computation of the d
cross-correlation lags. When a new N-length block of
H H c
input and output samples is available, CDX ( k ) is
( R)
bN = 2 M (21)
updated as 0( N k2 )1
CDXR +1 = N CDXR + RR ( k ) DR* ( k ) where ( )
( ) ( ) H
stands for the conjugate transpose
(19)
k = 0,1,..., 2 N 1 operation, IM is the M M identity matrix,
e M + k1 = [ 0 ... 0 1] M ( k1 + M ) ,
T
and the
Recall that quantity C
( R)
( k ) can be
DX
M k2 matrices are given as shown in (22) and (23),
interpreted as a sum of complex harmonics with
unknown frequencies and complex amplitudes. Indeed,
as can be easily seen by inspecting (16), the frequency
bin k corresponds to the sequence index while the time
delay nl corresponds to the unknown frequency.
373
h k1 h k1 +1 L h1 h0 h1 L L hM 2 hM 1
0 h L h2 h1 h0 L L hM 3 hM 2
k1
H1 [ H11 | H12 ] =
M M M M M M M M M M
(22)
0 0 L h k1 h k1 +1 L L L L hM k1
M M M M M M M M M M
0 0 L 0 0 L h k1 L h1 h0
1 k1 +1
374
H
last column of H 2 and matrix H 2H has a Toeplitz Table I
Comparison in terms of numbers of complex
form, we deduce from (21) and (29) that the FB filter multiplications
possesses approximately the following structure.
1) There are first-order (primary) nonzero taps at the Conventional
2M + 2 N
positions nl where nl > 0 is a position of a causal DFE - LMS
component in the channel IR. SDFE-2 2 M + 3 log 2 ( N ) + 2 L1 ( L2 + 1) + 5
2) For each primary tap at nl > 0 , there are second- SDFE-3 2 M + 3 log 2 ( N ) + 2 L1 ( L2 + S + 1) + 5
order nonzero taps at the positions nl + ni > 0 , where
ni < 0 are positions of the anticausal components in the V. SIMULATION RESULTS
channel IR.
3) For each primary nonzero tap at nl > 0 , there are The low complexity DFE algorithms have
been tested for different sparse channels (including
third-order terms located at nl + ni + n j > 0 , where measured microwave channels ) and various noise
ni , n j is any combination of component locations specifications. Their performance has been evaluated
for time invariant channels and also for slow time
with ni + n j < 0 . varying channels. Some simulation result are described
Thus, it turns out that the FB filter has a below.
sparse form and, hence, can be restricted to act to the Fig. 1 shows a typical terrestrial HDTV channel IR.
above positions only. In case strong echoes are not The channel IR is the convolution of test channel D of
present in the channel IR, a second-order [17] with a square-root raised cosine filter with 11.5%
approximation of the FB filter [points 1) and 2) above] rolloff. Note the presence of four postcursor
seems to be sufficient for the proposed algorithm to components, including a strong far echo, and one
achieve a performance similar to that of the precursor component of relatively low magnitude. The
conventional DFE. However, when there are strong input to the channel is a 16-quadrature amplitude
components in the channel IR (especially strong modulation (QAM) sequence, while complex white
precursor components), a higher number of taps should Gaussian noise is added to the channel output, resulting
be considered for the FB filter as dictated by point 3). in an SNR of 25 dB.
This results in a slight increase of the computational 0.5
complexity of the proposed algorithm. In any case, the
FB filter comprises a small number of taps and the
0
novel sparse equalizer offers considerable
Real
B. Complexity Issues
The main feature of the algorithm described in -1
0 20 40 60 80 100 120 140
Section IV-A is that instead of a long FB filter, it uses a
small number of nonzero FB taps. As a result, it is
expected that its computational load will be equally 0.3
375
and steady-state error performance. The red curve
1
correspond to the conventional DFE and the blue and 10
black curves-to the SDFE with second- and third-order DFE-LMS
approximation of the FF and FB filter, respectively. We SDFE(2)
see that both SDFEs have greater steady-state 0
SDFE(3)
10
performance and faster convergence than conventional
-3
10
-4
10
0 2000 4000 6000 8000 10000 12000
Number of QAM-16 symbols
Fig.2. Convergence and steady-state MSE curves of DFE-
LMS and SDFEs for the channel IR of Fig.1.
In order to investigate the tracking ability of amplitudes are kept fixed. The phase rotation step is
the new algorithm in a time-varying environment, we (0.1/360) rad per iteration. In Fig.4. we see that the new
consider the following scenario. After 7500 iterations, algorithm tracks the change in the environment and as a
the phases of all postcursor components of the channel result the misajustment error is small, lower than in the
of Fig. 1 start continuously rotating, while their case of conventional DFE. The convergence and track
376
capability of SDFE depend on the convergence of [6] P. Crespo and M. Honig, Pole-zero decision feedback
equalization with a rapidly converging adaptive IIR algorithm,
estimation of echo delay algorithm.
IEEE J. Select.Areas Commun., vol. 9, pp. 817829, Aug. 1991.
[7] N. Al-Dhahir and J. M. Cioffi, Stable pole-zero modeling of long
1 FIRfilters with application to the MMSE-DFE, IEEE Trans.
10
Commun., vol.45, pp. 508513, May 1997.
SFDE-3
[8] K. Berberidis, T. A. Rontogiannis, and S. Theodoridis, Efficient
DFE
block implementation of the DFE, IEEE Signal Processing
0
10 Lett., vol. 5, pp.129131, June 1998.
[9] K. Berberidis and P. Karaivazoglou, A block adaptive DFE in
the frequency domain based on tentative decisions, in Proc.
EUSIPCO,Rodos, Sept. 1998.
Mean Square Error
-1
10 [10] N. Al-Dhahir and J. M. Cioffi, Fast computation of channel-
estimate based equalizers in packet data transmission, IEEE
Trans. Signal Processing,vol. 43, pp. 24622473, Nov. 1995.
[11] S. Ariyavisitakul and L. J. Greenstein, Reduced-complexity
-2
10 equalization techniques for broadband wireless channels, IEEE
J. Select. AreasCommun., vol. 15, pp. 515, Jan. 1997.
[12] R. Gupta, Kiran, and E. Lee, Computationally efficient version
of the decision feedback equalizer, in Proc. IEEE ICASSP,
-3
10 Seattle, 1999, pp.12571260.
[13] S. Ariyavisitakul, N. R. Sollenberger, and L. J. Greenstein, Tap-
selectable decision feedback equalization, IEEE Trans.
-4
Commun., vol.45, pp. 14981500, Dec. 1997.
10 [14] I. J. Fevrier, S. B. Gelfand, and M. P. Fitz, Reduced complexity
0 0.5 1 1.5 2 decision feedback equalization for multipath channels with
Number of 16-QAM symbols 4
large delay spreads,IEEE Trans. Commun., vol. 47, pp. 927
x 10
937, 1999.
Fig.4. MSE curve under slow time-varying CIR [15] S. F. Cotter and B. D. Rao, Matching pursuit based decision-
feedback equalizers, in Proc. IEEE ICASSP, Istanbul, Turkey,
June 2000.
[16] K. Berberidis and A. A. Rontogiannis, Efficient decision
VI. CONCLUSION
feedback equalizer for sparse multipath channels, in Proc.
IEEE ICASSP, Istanbul, Turkey, June 2000.
In this paper, the sparse form encountered in [17] I. Kang, M. P. Fitz, and S. B. Gelfand, Blind estimation of
many high-speed wireless communications channels is multipath channels: A modal analysis approach, IEEE Trans.
Commun., vol. 47, pp. 11401150, Aug. 1999.
properly exploited in two ways. First, the time delays of [18] A. A. Rontogiannis, A. Marava, K. Berberidis, and J. Palicot,
the dominant multipath components are efficiently Semiblind estimation of multipath channel parameters via a
separable least squares approach, in Proc. 14th Digital Signal
Processing Conf., Santorini, July 2002, pp. 4952.
[19] A. A. Rontogiannis and K. Berberidis, Efficient decision
estimated and continuously updated using a frequency feedback equalizer for sparse wireless channels, IEEE Trnas.
domain approach. Second, an approximate form of the Wireless Commun, vol. 2, May 2003.
FB filter of the MMSE-DFE is derived. Based on the [20] M. H. Hayes, Statistical Digital Signal Processing and
above, a new efficient DFE algorithm is proposed, Modeling. New York: Wiley, 1996.
whose FB filter comprises a reduced number of active
taps. Depending on the multipath channel conditions,
there may exist a tradeoff between the number of FB
taps used and the performance of the proposed
algorithm. In general, the new algorithm offers
considerable computational savings, faster
convergence, and acceptable tracking capabilities while
exhibiting almost identical steady-state performance in
most practical cases, as compared with the conventional
DFE. The features of the new algorithm have been
confirmed through extensive simulation tests.
REFERENCES
[1] M. Ghosh, Blind decision feedback equalization for terrestrial
television receivers, Proc. IEEE, vol. 86, pp. 20702081, Oct.
1998.
[2] J. G. Proakis, Digital Communications, 3rd ed. New York:
McGraw-Hill, 1995.
[3] T. S. Rappaport, Wireless Communications: Principles and
Practice.Englewood Cliffs, NJ: Prentice-Hall, 1996.
[4] J. Salz, Optimum mean-square decision feedback equalization,
BellSyst. Tech. J., vol. 52, no. 8, pp. 13411373, Oct. 1973.
[5] G. Young, Reduced complexity decision feedback equalization
for digital subscriber loops, IEEE J. Select. Areas Commun.,
vol. 9, pp.810816, Aug. 1991.
377
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract: This paper analizes the performance of a shortened Reed-Solomon code (204,188,t=8). This
Terrestrial Digital Television system. The performances code can correct up to eight erroneous bytes in a
are estimated using a Matlab Simulink model for the frame of 204 bytes. The coded bits are interleaved by
DVB-T system. a convolutional interleaver that interleaves byte- wise
Keywords: DVB_T system, terrestrial broadcasting,
OFDM, QAM modulation, bit error rate, channel coding
with a depth of 12 bytes and then again coded by a
rate 1/2, constraint length 7 convolutional code with
I. INTRODUCTION generator polynomials (171,133 octal). The rate of
this latter code can be increased by puncturing to
DVB-T is the standard for Digital Television 2/3,3/4,5/6, or 7/8. The convolutionally encoded bits
Terrestrial Broadcasting defined for Europe. The are interleaved by an inner interleaver mapped onto
DVB family standards allows for digital video and QPSK, 16QAM, or 64QAM symbols.
audio broadcasting as well as transport of multimedia To obtain reference amplitude and phase to
services. For terrestrial broadcasting the system was perform coherent QAM demodulation, pilot
designed to operate within the existing UHF spectrum subcarriers are transmitted. For the 8k mode, in
allocated for analogue television. The system was each symbol there are 768 pilots, so 6,048
developed for 8MHz channels but it can be subcarriers remain for data. The 2k mode has
reconfigured also for 7or 6MHz channels. The net bit 192 pilots and 1,512 data subcarriers. The
rate available in the 8MHz channel ranges between 4 position of the pilots varies from symbol to
and 32 Mbit/s, depending of channel coding symbol with a pattern that repeats after four
parameters, modulation type and guard interval. OFDM symbols. The pilots allow receiver to
The Coded Orthogonal Frequency Division estimate the channel both in frequency as well as
Multiplexing modulation system was chosen, being in time, which is important as for mobile receivers
suitable for the multipath propagation environment of there can be significant channel changes within a
terrestrial radio channels. The system uses a large few OFDM symbols.
number of carriers per channel allowing the reduction Outer Outer Inner Inner
of symbol rate over one carrier. In this way the Coder Interleaver Coding Interleaver
symbol interval is increased and a better protection to
multipath propagation is obtained. The OFDM may
operate with two modes: 8k FFT mode and 2k FFT D/A
Frame OFDM Guard
mode. The system can select between different levels Adaptation interval Front end
of QAM modulation and different inner code rates
and also allows two level hierarchical channel coding
Pilot &
and modulation. Moreover, a guard interval with TPS
selectable width separates the transmitted symbols,
which allows the system to support different network Fig. 1 DVB-T transmission system block diagram
configuration: 8k mode for large single frequency
networks and 2k mode for small or mobile networks Terrestrial DVB use OFDM with two possible
modes, using 1,705 and 6817 subcarriers,
II. DVB-T System respectively [1]. These modes are known as 2k
and 8k modes, respectively, as these are the size
Figure (1) shows block diagram of a DVB-T
of FFT/IFFT needed to generate and demodulate
transmitter. The input data are divided into groups of
188 bytes, which are scrambled and coded by an outer all subcarriers.. Basically, the 2k system is a
1
Politehnica University of Bucharest,Electronics and Telecommunications Faculty, Telecommunications Departament,
Bd. Iuliu Maniu Nr. 1-3, 061071,Bucuresti, e-mail serban@radio.pub.ro
378
simplified version which require an FFT/IFFT Table2
that is only a quarter of the size that is needed for 2K mode
the 8k system. Because the guard time is also Guard 1/4 1/8 1/16 1/32
four times smaller, the 2k system can handle less Interval
Duration of 2048xT
delay spread and less propagation delay symbol part 224s
difference among transmitters within a single TU
frequency network but is less sensitive to the Duration of 2512xT 256xT 128xT 64xT
Doppler effect. The FFT interval duration for the guard 56 s 28 s 14 s 7 s
8k system is 896s while the guard time can interval
Symbol 2560xT 2304xT 2176xT 2112xT
have four different values from 28 to 224 s. The
duration 280 s 252s 238 s 231 s
corresponding values for the 2k system are four TS=+TU
times smaller.
The transmitted signal is organized in frames. Each
frame has duration of TF, and consists of 68 OFDM The emitted signal is described by the following
symbols. Four frames constitute one super-frame. expression:
Each symbol is constituted by a set of K = 6 817
carriers in the 8K mode and K = 1 705 carriers in the
67 K max
2K mode and transmitted with a duration TS. It is s (t ) Re e j 2f ct c m,l ,k m,l ,k (t )
composed of two parts: a useful part with duration TU m =0 l = 0 k = K min
and a guard interval with duration . The guard
interval consists in a cyclic continuation of the useful where
part, TU, and is inserted before it. Four values of guard
intervals may be used according to table 5. The m,l ,k (t ) =
symbols in an OFDM frame are numbered from 0 to k'
67. All symbols contain data and reference j 2 TU (t lTS 68mTS )
information. = e pTS < t < ( p + 1)TS
0 else
Since the OFDM signal comprises many separately-
modulated carriers, each symbol can in turn be
considered to be divided into cells, each k - the carrier number
corresponding to the modulation carried on one l OFDM symbol number
carrier during one symbol.
m transmission frame number
In addition to the transmitted data an OFDM frame
K number of transmitted carriers
contains:
- Scattered pilot cells; TS symbol duration
- Continual pilot carriers; TU inverse of the carrier spacing
- TPS carriers. guard interval
The pilots can be used for frame synchronization, fc central frequency
frequency synchronization, time synchronization, (K + K min )
k ' = k max
channel estimation, transmission mode identification 2
and can also be used to follow the phase noise. cm,l,k complex symbol
The carriers are indexed by k [Kmin; Kmax] and The apparent complexity of these equations can be
determined by Kmin = 0 andKmax = 1 704 in 2K simplified if it is noted that the waveform emitted
mode and 6 816 in 8K mode respectively. The spacing during each transmitted symbol period depends solely
between adjacent carriers is 1/TU while the spacing on the K complex values cm,l,k which define the
between carriers Kmin and Kmax are determined by complex amplitude of the K active carriers for that
(K-1)/TU. The numerical values for the OFDM period. Each symbol can thus be considered in
parameters for the 8K and 2K modes are given in isolation; for example, the signal for the period from
tables 1 and 2. t= 0 to t=TS is given by:
Table 1
8K mode
K max
s (t ) Ree j 2f ct c 0,0,k e j 2k (t ) / TU
'
Guard 1/4 1/8 1/16 1/32
Interval k = K min
Duration 8192xT
of symbol 896s
part TU There is a clear resemblance between this and the
Duration 2048xT 1024xT 512xT 256xT inverse Discrete Fourier Transform (IDFT):
of guard 224 s 112 s 56 s 28 s 1 N 1
interval xn = X q e j 2nq / N
N q =0
Symbol 10240xT 9216xT 8704xT 8448xT
duration 1120 s 1008s 952 s 925 s Since various efficient Fast Fourier Transform
TS=+TU algorithms exist to perform the DFT and its inverse, it
379
is a convenient form of implementation to use the We have 1704 carriers with only 1512 useful carriers.
inverse FFT (IFFT) in a DVB-T modulator to The symbol period is Ts=280s. In every symbol we
generate N samples xn corresponding to the useful have 1512 useful carriers and 4 bits per carrier (one
part, TU long, of each symbol. The guard interval is 16QAM symbol). We have a convolutional code with
added by taking copies of the last N/ TU . of these rate and a Red Solomon code with the ratio
samples and appending them in front. This process is 204/188. It results that the useful information rate is
then repeated for each symbol in turn, producing a rD=4/2*188/204*1512*106/280=9,9529 Mbii/s.
continuous stream of samples which constitute a In a superframe (4 frames of 68 symbols each) there is
complex baseband representation of the DVB-T an integer number of Red Solomon block (204) so
signal. A subsequent up-conversion process then there is no need for bit stuffing.
gives the real signal s(t) centered on the frequency fc The spectrum of the simulated 2k OFDM signal is
The OFDM symbols constitute a juxtaposition of given in figure 3 and time representation in figure 4.
equally-spaced orthogonal carriers. The amplitudes
and phases of the data cell carriers are varying symbol
by symbol according to the mapping process. The
power spectral density Pk (f) of each carrier is given
by the following expression:
sin ( f f k )TS
2
Pk ( f ) =
( f f k )TS
380
[3] J. A. C. Bingham, "Multi-carrier modulation
for data transmission: An idea whose time has
come", IEEE Communications Magazine, vol.28, no.
5, pp. 5-14, May 1990.
[4] A. V. Oppenheim and R. W. Schafer,
Discrete-Time Signal Processing, Englewood Cliffs,
NJ: Prentice Hall, 1989
[5] Yiyan Wu, William Y. Zou, Orthogonal
Frequency Division Multiplexing: A Multi-Carrier
Modulation Scheme, IEEE Transaction on
Consumer Electronics, Vol. 41,No. 3, August 1995,
pp. 392 - 399
[6] William Y. Zou, Yiyan Wu, COFDM: An
Overview, IEEE Transactions onBroadcasting, Vol.
41, No. 1, March 1995, pp. 1 8
Fig. 5 The received 16 QAM constelation [7] R. R. Mosier and R. G. Clabaugh, Kineplex, a
bandwidth-efficient binary transmission system,
AIEE Transactions, Vol. 76, January 1958,
We have simulated the system behavior for different
[8] Robert Chang, Synthesis of Band-Limited
S/N ratios. The resulted BER is presented in the
Orthogonal Signals for Multichannel Data
diagram in the figure 6.
Transmission, The Bell System Technical Journal,
December 1966, pp. 1775 -1796
[9] Robert Chang, Orthogonal frequency division
multiplexing, US. Patent 3,488445, filed November
14, 1966, issued January 6, 1970
[10] S. B. Weinstein, Paul M. Ebert, Data
Transmission by Frequency-Division Multiplexing
Using the Discrete Fourier Transform, IEEE
Transactions on Communication Technology, Vol.
COM-19, No. 5, October 1971, pp. 628 - 634
[11] Carl Magnus Frodigh, Perols Leif Mikael
Gudmundson, Adaptive channel allocation in a
frequency division multiplexed system, US. Patent
5,726,978, June 22, 1995, Issued: March 10, 1998
[12] S. OLeary, F. Ryan, B. Wynne, and C. Gilliam,
Interactive digital terrestrial televisionThe wireless
Fig. 6 BER(S/N) return channel and the EU sponsored WITNESS
project, IEEE Trans. Broadc., vol. 47, pp. 160163,
June 2001.
V. CONCLUSIONS [13] R. Mhiri, D. Masse, and D. Schafhuber,
Synchronization for a DVBT receiver in presence of
We have implemented in Simulink an DVBT co-channel interference. IEEE PIMRC-02, (Lisbon,
communications system and we have analyzed its Portugal), Sept. 2002. submitted.
performance for 2k mode and 16 QAM modulation. [14] P. Hoeher, S. Kaiser, and P. Robertson, Two-
We intend to complete the model by implementing dimensional pilotsymbol-aided channel estimation by
also the 8k mode, and the 64QAM and QPSK Wiener filtering, in Proc. IEEE ICASSP-97,
modulations. (Munich, Germany), pp. 18451848, April 1997.
We intend to adapt this system for the DVBH (DVB [15] Y. Li, Pilot-symbol-aided channel estimation for
for mobile communications) standard. The DVB OFDM in wireless systems, IEEE Trans. Veh.
system performance study permits the Technol., vol. 49, pp. 12071215, July 2000.
[16] M. Sandell, Design and Analysis of Estimators
for Multicarrier Modulation and Ultrasonic Imaging.
REFERENCES PhD thesis, Lulea University of Technology, Lulea,
Sweden, 1996.
[1] ETS 300 744, "Digital broadcasting systems for
television, sound and data services; framing structure,
channel coding, and modulation for digital terrestrial
television,. European Telecommunication Standard,
Doc. 300 744, 1997.
[2] R. V. Nee and R. Prasad, OFDM Wireless
Multimedia Communications, Norwood, MA: Artech
House, 2000.
381
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
382
N 1 Considering the channel to be fixed over the OFDM
s (t ) = sn (t ) = dk ,nk (t nTS ) (5) symbol interval, denoting it by ch( ) and taking into
n = n = k = 0
account the orthogonality condition expressed by (9),
we obtain after some mathematical operations the
output data, given by (10).
T
TCP l (t ) k (t ) = (k l )
*
(9)
ek = hk d k + nk , where
B
TCP j 2 k
hk = ch( ) e N d and (10)
0
TS
nk = n(TS t ) k* (t ) dt
TCP
383
pilot subcarriers, the rotation can be estimated and the new estimate is calculated using long training
compensated for [11], [12], [14]. symbols.
384
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Gh. Asachi Technical University of Iai, Telecommunications Department, 11 Carol I Blvd., Iai, 700506, Romania, e-mail:
ccomsa@etc.tuiasi.ro
385
Fig. 2 illustrates the baseband OFDM model Also, we can ignore the time index n when calculating
mathematically described bellow [4], [5]. Every nth the sampled output at the kth matched filter (7).
OFDM symbol of the transmission stream can be
* (T t ) , t [ 0, TS TCP )
written as a set of modulated carriers transmitted in k (t ) = k S (6)
parallel. Relations (2) express the waveforms used in 0 , otherwise
modulation.
ek = ( r * k )(t ) t =T = r (t ) k (TS t ) dt (7)
1
e j 2 f k (t TCP ) , t [ 0, TS ) S
k (t ) = TS TCP , where Considering the channel to be fixed over the OFDM
0 , otherwise symbol interval, denoting it by ch( ) and taking into
N 1 1 account the orthogonality condition expressed by (8),
f k = fC + k , k = 0,..., N 1 , for passband or (2)
2 T we obtain after some mathematical operations the
k
f k = , k = 0,..., N 1 , for baseband echivalent output data, given by (9).
T T
TCP l (t ) k (t ) = (k l )
*
Note that nonzero term of k (t ) has the period (8)
If d n,0 ,..., d n, N 1 denotes the complex symbols, where N is the number of subcarriers. To make
obtained by QAM mapping of the input data stream, OFDM more robust against multipath and timing
the nth OFDM symbol sn (t ) is expressed by (4) and offset, each symbol is extended with a cyclic prefix
(CP). The CP is constructed by copying the last Ng
the infinite sequence of OFDM symbols transmitted is
samples of the OFDM symbol (Tes being the sampling
obtained by juxtaposition of the individual ones.
N 1
period) at the beginning of it. So, the OFDM symbol
s (t ) = sn (t ) = dk ,nk (t nTS ) (4) transmitted is x N g x N g +1 ...... xN 2 xN 1 . Finally,
n = n = k = 0 the time domain signals are D-A converted, mixed
Assuming the impulse response ch( ; t ) of the with a carrier, filtered and transmitted through the air.
physical channel (possibly time variant) is restricted In the receiver, the opposite operations are performed
to the length of cyclic prefix [ 0, TCP ) , the using A-D conversion and DFT calculation.
received signal becomes (5), where n(t ) is the III. PULSE SHAPING AND WINDOWING
complex, additive and white Gaussian (AWGN)
channel noise. The complex envelope of one N-subcarrier OFDM
r (t ) = ( ch * s ) (t ) =
TCP
ch( ; t )s (t ) + n(t ) (5) block with pulse-shaping is expressed as (11), where
0
j = 1 , fc is the carrier frequency, fk is the
The filter from the receiver is matched to the last part
subcarrier frequency of the k-th subcarrier, p ( t ) is
[TCP , TS ) of the transmitter waveform (6), the CP
being this way effectively removed in the receiver. the time-limited pulse shaping function and
Since the cyclic prefix contains the ISI, the sample ak,k=0,1,,N-1 is a complex-valued data symbol
output from the receiver filter bank contains no ISI. transmitted on the k-th subcarrier.
386
N 1 Fig. 3. Frequency functions Pr ( f ) , Prc ( f ) and Pbtrc ( f ) .
x ( t ) = e j 2 fc t ak p ( t ) e j 2 f k t (11)
Another family of pulses has recently been reported
k =0
[11], than are intersymbol interference-free. It is about
Equation (12) expresses the condition that the Fourier
conjugate root pulses, which are not linear phase,
transform of the pulse p ( t ) should have spectral whereas the root raised-cosine (RC) pulses are linear
nulls at the frequencies 1 T , 2 T , to ensure phase. The first-order conjugate-root pulse phase is
subcarrier orthogonality [8]. piecewise linear, while the fourth-order pulse phase is
not piecewise linear. First-order conjugate-pulses are
j 2 ( f f )t 1 , k = m
p ( t ) e k m dt = 0 , k m (12) expressible in the time domain in closed-form (17),
while fourth-order pulses have more complicated
k m forms. Both first-order conjugate-pulses and fourth-
fk fm = (13) order pulses together with root raised-cosine pulses
T are plotted in the time domain for = 0.35 in Fig. 4.
There are considered here three time-limited-pulses,
which are Pr ( f ) , Prc ( f ) and Pbtrc ( f ) denoting the t t
sin cos
rectangular pulse, the raised-cosine pulse (in the time- p (t ) = Tes Tes (17)
domain) and the better than raised cosine (BTRC) t
pulse (in the time-domain), defined as (14), (15) and ( 2 t + T es )
Tes
(16), where is the roll-off factor and 0 1
[12]. When = 0 both raised-cosine and the BTRC
pulse coalesce into the rectangular pulse. The Fourier
transforms are denoted by Pr ( f ) , Prc ( f ) and
Pbtrc ( f ) respectively. Fig. 3 shows the frequency
functions of these pulses for = 0.2 and = 1 .
1 , T t T
pr ( t ) = T 2 2 (14)
0 , otherwise
1 T (1 )
, 0 t
T 2
1 T (1 ) T (1 ) T (1 + )
prc = 1 + cos
T
t , t (15)
2T 2 2 2
0 , otherwise
1 T (1 )
, 0 t
T 2
2 ln 2 T (1 )
1 e T
t
2
T (1 ) T
T , t
pbtrc ( t ) = 2 2 (16) Fig. 4. Time-domain plots of the RC and conjugate-root pulses for
2 ln 2 T (1+ )
t
T (1 + ) = 0.35
1 1 e T 2 ,
T
t
T 2 2
IV. PERFORMANCES OF PULSE SHAPING
0 , otherwise
Frequency offset, f ( f 0 ) , and phase error, ,
are introduced during transmission because channel
distortion or receiver crystal oscillator inaccuracy.
The received signal after multiplication by
e(
j 2 ( f c +f )t + )
becomes (18).
N 1
r (t ) = e (
j 2ft + )
ak p ( t ) e j 2 fk t (18)
k =0
r (t ) e
j 2 f m t
a$ m = dt =
= am e j p ( t ) e dt +
j 2ft
(19)
N 1
+e j ak p ( t ) e j 2 ( f k f m +f )t
dt
k =0
k m
387
The m-th subchanel correlation demodulator, thus, types of pulses and it concluded that the employment
gives the decision variable for transmitted symbol am of the better than raised-cosine pulse rather than the
raised-cosine pulse gives a substantial improvement in
in (19), where the first term contains the desired
signal component, and the second term is the ICI. the reduction of ICI caused by frequency offset in an
OFDM system. However, further work has to be done
Combining (13) with (19) gives (20), where P ( f ) is in comparison those pulse shapes with another
the Fourier transform of p ( t ) . Hence, the power of windowing functions, reported to the distortion
the desired signal is (21) and the ICI power is (22). introduced and to the implementation complexity.
N 1 Also, another pulse shapes may be found using a
mk
a$ m = am e j P ( f ) + e j ak P f (20) method of parametric construction of ISI-free pulses,
k =0 T based on some optimization criteria (as Lagrange
k m
method) of some properties of interest, like ICI.
2
P ( f )
2
m = am (21)
REFERENCES
N 1 N 1
k m nm
ICI
m
= ak an* P
T
+ f P
T
+ f (22)
[1] Beaulieu N. C. and Cheng J., Precise error rate analysis of
k =0 n =0 bandwidth efficient BPSK in Nakagami fading and co-channel
k m nm
interference, IEEE Trans. Communications., vol. 52, pp. 149
The ICI power depends not only on the desired 158, Jan. 2004.
symbol location, m, and the transmitted symbol [2] Beaulieu N. C. and Damen M. O., A parametric construction of
sequence, but also on the pulse-shaping function and ISI-free pulses, Proc. of 8th Canadian Workshop on Information
Theory, 2003, pp. 121124
the number of subcarriers. However, (22) gives the [3] Beaulieu N. C., Tan Ch. C., Damen M. O., A Better Than
average ICI power, averaged across different Nyquist Pulse, IEEE Communications Letters, Vol. 5, No. 9,
sequences as September 2001
N 1 2 [4] Coma C. R., Bogdan I., System Level Design of Baseband
k m
ICI
m
= P
T
+ f
(23) OFDM for Wireless LAN, International Symposium on Signals,
Circuits and Systems, July 10-11, Iai, 2003, Proceedings, pp.
k =0
k m 313-316
[5] Edfors O., Sandell M., Van de Beek J. J., An Introduction to
For the same value of the BTRC pulse outperforms Orthogonal Frequency-Division Multiplexing, 1996
the others, including raised cosine (RC) pulse, as [6] Intini A. L., Orthogonal Frequency Division Multiplexing for
shown in Fig. 5 for = 1 [12]. This interesting Wireless Networks, Santa Clara University of California, 2000
[7] Lawrey E. Ph., Adaptive Techniques for Multiuser OFDM,
behavior occurs despite the fact that the tails of Thesis submitted in December 2001 for the degree of PhD,
Pbtrc ( f ) and Prc ( f ) decay as f 2 and f 3 , James Cook University, Australia
[8] Proakis J. G., Salehi M., Communication Systems Engineering -
respectively. Second Edition, Prentice Hall, 2002
[9] Ramasami Vijaya Chandran, Orthogonal Frequency Division
Multiplexing, University of Kansas, USA, May 2001
[10]Sun Yi, Bandwidth-Efficient Wireless OFDM, IEEE Journal on
Selected Areas in Communications, vol. 19, no. 11, Nov. 2001
[11]Tan Ch. C., Beaulieu N. C., Transmission Properties of
Conjugate-Root Pulses, IEEE Transactions on
Communications, Vol. 52, No. 4, April 2004
[12]Tan P., Beaulieu N. C., Reduced ICI in OFDM Systems Using
the Better Than Raised-Cosine Pulse, IEEE Communications
Letters, Vol. 8, No. 3, March 2004
[13]Van Nee R. and Prasad R., OFDM for Wireless Multimedia
Communications, Artech House Publishers, Boston, 2000
V. CONCLUSIONS
388
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
389
caused by the multipath fading is indicated by the
II. SYSTEM DESCRIPTION AND FADING Doppler spectrum. In the second case, the channel
CHANNEL MODEL multipath intensity profile and the length of the
OFDM symbol indicate the phase change rate due to
In OFDM, the available bandwidth is partitioned fading conditions. Therefore, the two methods exhibit
into N subchannels. The desired high-rate symbol essentially different behaviors although both encode
stream is achieved by simultaneously transmitting N the data differentially. Another major difference
slower rate substreams using N orthogonal between the two methods (that gives them their
subcarriers. The binary data to be transmitted is specific name) is that the consecutive OFDM symbols
differentially encoded using a DBPSK modulation are interconnected through differential encoding in the
scheme, obtaining a sequence of complex data first case (inter-frame differential modulation), while
symbols (fig. 1). no successive OFDM symbols connection is realized
by the second method (in-frame differential
modulation), where the differential encoding is
DBPSK si , n Cyclic prefix
IFFT performed on the samples belonging to the same
Binary Modulation insertion frame (or to the same future OFDM symbol), as
Information illustrated in the fig. 2b.
Channel
After the differential encoding of the binary
s i,n Cyclic prefix message using one of the two methods presented
Decision FFT removal above, the sequence si,n is obtained, si,n denoting the n-
Estimated
Information TS th symbol of the i-th frame,
k
N where (0 n N 1, < i < ) . The n-th carrier is
Fig.1: The OFDM transceiver model
modulated by the samples {s i,n , < i < } and the
There are two possibilities to perform differential
modulation in the presented OFDM scheme. Data can modulated carriers (orthogonal one-another) are
be encoded in the relative phase of adjacent symbols added together to form the OFDM symbol to be
in each subchannel (correspondent samples in two transmitted. In a practical implementation, the N
consecutive OFDM symbols) or in the relative phase samples of the OFDM symbol corresponding to the i-
of samples transmitted in adjacent subchannels, that is th frame are generated by processing {si,n }stream
consecutive samples of an OFDM symbol (see fig. 2). using the fast implementation of the Inverse Discrete
Since the IFFT block accepts N parallel samples to its Fourier Transform (IDFT) (see fig. 1). In order to
entry, the whole difference of the two methods can be combat the inter-symbol and inter-carrier interference
thought as follows: if the phase modulation is introduced by the frequency selectivity and the time
separately achieved on each of the N parallel streams selectivity of the radio channel, each OFDM symbol
that constitute the entry to the IFFT block, then we are is preceded by a cyclic prefix of L samples. The
in the case of the first presented modulation type, cyclic prefix is a circular extension of the time
namely an inter-frame modulation is performed (see domain samples, being obtained by copying the last L
fig. 2a). If the modulation is made on the serial samples of the OFDM symbol in the front of it. The i-
stream, prior to the parallel conversion required by th transmitted symbol (including the prefix) contains
IFFT, then an in-frame modulation is chosen, since N N+L time domain samples, of which the m-th sample
consecutive serial samples will simultaneously is given by the equation below:
modulate N orthogonal carriers, forming an OFDM
nm
symbol (fig 2b). E S N 1 j 2
g i ( m) = si ,n e N
, m = L,..., N 1
Serial stream, N + L n =0
consecutive frames S/P (1)
IFFT
conversion
Correspondent Assuming the data symbols are statistically
samples OFDM symbol
a) Relative phase independent and having a unit average energy, the
coding transmitted average energy per symbol equals ES. The
transmitted signals can be expressed in complex form
I as:
S/P F
conversion
F s(t)= p(t iTS ) g i (t ) (2)
T i =
Relative phase
coding
b) where gi(t) represents the analogical waveform
OFDM symbol
corresponding to the OFDM symbol, obtained after a
Fig.2: (a) Inter-frame modulation (b) In-frame modulation DAC conversion of the sequence {gi(m)},
Both methods have an irreducible error rate because m=0,1,,N-1. p(t) is the pulse-shaping waveform of
of the random change of the relative phase, caused by each symbol, defined as:
the fading channel. In the first method the distortion
390
1, for t t s power of the two multipath components are
p(t ) = (3) considered for channel simulation. The BER
0, otherwise
computation was averaged over 20000 transmitted
OFDM symbols. Neither channel coding nor further
equalization to the receiver were considered at this
TS =+ts stands for the total duration of an OFDM stage. A comparison of the two methods is made,
symbol, composed by the cyclic prefix period () and studying the influence of the block length N, of the
by the observation period (ts). The fading channel channel multipath delay spread and of the Doppler
(assuming Rice conditions) can be modeled as a 3-ray shift introduced by the time-variant character of the
tapped delay line with one line-of-sight (LOS) path channel on the BER performance in both in-frame and
and two multipath components. If h(t,) denotes the inter-frame DBPSK-OFDM system. We emphasize
channel impulse response at time t-, it can be the essential different behaviors of the two methods
expressed as: with respect to the parameters presented above.
The BER performances of the DBPSK-OFDM
h(t , ) = 2 PS ( ) + P1 a1 (t ) ( 1 ) + system in a Rayleigh fading channel, as a function of
(4) the normalized delay of the second multipath 2/T is
P2 a 2 (t ) ( 2 )
illustrated in the figures 3,4.
t S + iTS
1 j 2f ( t iTS )
rm ,i =
tS
{ s (t )h(t , )d + n(t )} e D
dt
iTS 0
(6)
Finally, the differential detector decides what symbol
was transmitted.
391
a sensitivity of this method to the time-variant channel method to the variation of the maximum Doppler shift
character. The same observation becomes more parameter comparing to the inter-frame modulation
obvious regarding the figure 4, where another (whose performance is indicated by the two outer
difference between the two Doppler shifts is taken curves). If at low maximum Doppler shift inter-frame
into account. modulation performs better, once the value of this
The BER performance of the inter-frame parameter grows, the in-frame modulation method
DBPSK-OFDM system for three different Doppler becomes more efficient. It can also be observed that at
shifts is plotted in the figure 5, in order to stress the significant Doppler shifts, the two methods exhibits a
effectiveness of this parameter. It is shown that the very poor improvement of BER performance, with
maximum Doppler shift has a significant influence on respect to SNR, especially for the inter-frame
the BER, especially when the delay of the second modulation.
multipath is small. For large delays of the second The influence of the block length N on the BER
multipath the main amount of errors is brought by the performance in both modulation types is illustrated in
ISI introduced by the multipath components, which the figure 7, where maximum Doppler shift is
confirms the conclusion in [5], respectively in [4] for considered to be constant. As stressed in [6], the in-
an in-frame DBPSK modulation. frame OFDM-DBPSK system significantly improves
its performance when the block (or, equivalently, the
OFDM symbol) length increases, considering the
IV. ABOUT REFERENCES same multipath delay spread of the channel impulse
response. On the other hand, it turns out that the BER
References should be numbered in a simple form [1], performance of inter-frame modulation is almost
[2], [3], and quoted accordingly [1]. References are identical for the different values chosen for the
not allowed in footnotes. It is recommended to parameter N. In the three simulated situations
mention all authors; et al. should be used only for (N=16,32,64), the system performs to within 1-2dB
more than 6 authors. spread of the results. It can be asserted that the
performance obtained using this modulation type is
Table 1 very little sensitive to the OFDM symbol duration.
Parameter Value Unit
I 2.4 A
U 10.0 V
392
frame length and the BER performance can be IV. CONCLUSIONS
asserted, since the system performs better for N=16
than for N=64. In this paper we have studied the BER performance of
an OFDM-DBPSK system with two distinct phase
modulation types. The principles of both in-frame and
inter-frame modulation in an OFDM transmission
scheme were briefly exposed, accentuating on their
differences. The essential different behavior in
multipath fading conditions was emphasized by
means of computer simulation. The inter-frame
modulation system, while generally performing better
has though shown to be more sensitive at the variation
of the Doppler shift parameter. The in-frame
modulation method allows significant performance
improvement by increasing the data-block length. The
multipath delay spread degrades the BER
performance of both studied modulation types. Even
if at delay spreads that significantly exceed the cyclic
prefix duration the performance is only slightly
Fig. 8: BER performances of inter-frame and in-frame DBPSK improved increasing the signal-to-noise ratio, the
modulation in a Rayleigh fading channel, SNR=40dB, P1/P2=0dB, inter-frame modulation proved to be more resistant to
fD=0.0001
the inter-symbol interference introduced by the
In the figure 9, the effect of multipath delay multipath delayed components.
spread on the both methods is studied. The maximum
Doppler shift was kept constant while the BER REFERENCES
performance against signal-to-noise (SNR) ratio was
plotted for two different values of the normalized [1] B. Sklar, Rayleigh Fading Channels in Mobile Digital
Communication Systems- Part I: Characterization, IEEE Commun.
delay of the second multipath component. Mag., July 1997;
Considering a small value of the mentioned [2] B. Sklar, Rayleigh Fading Channels in Mobile Digital
parameter, the in-frame OFDM-DBPSK system Communication Systems- Part II: Mitigation, IEEE Commun.
performs to about 18dB better than the inter-frame Mag., July 1997;
[3] J.A.C. Bingham, Multicarrier Modulation for Data
system. To lie in such a situation, a correspondent
Transmission, An Idea Whose Time Has Come, IEEE
value was chosen for the normalized Doppler shift Communication Magazine, Vol. 31, No. 5, May 1990
(fD*TS=0.025). When a three times bigger value was [4] Lu, J. , Tjhung, T.T., Adachi, F., BER performance of OFDM
considered for the normalized delay of the second system in frequency-selective Rician fading with diversity
reception, available online:
multipath component inter-frame modulation http://www.cwc.nus.edu.sg/~cwcpub/zfiles/ict97_2.pdf.
performed better, despite the significant value of the [5]Lupea, E., Bianu, M., Slgean, M., Oltean, M., Naforni, M.,
Doppler shift. One can conclude that in-frame BER Performance of Frequency Selective Channels with Cyclic
modulation is more sensitive to the multipath delay Prefix Based Equalizers, Buletinul tiinific al Universitii
Politehnica Timioara, Tom 47 (61), Fascicola 1-2, 2002
spread introduced by the inherent dispersive nature of [6]Chini, A., Analysis and Simulation of Multicarrier Modulation in
the radio channel. Frequency Selective Fading Channels ,Ph. D. Thesis, 1994, Chapter
3
393
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Politehnica University of Bucharest, Electronics and Telecommunications Faculty, 1-3,
Iuliu Maniu Bvd., Bucharest, Romania; Phone: +4021 4024765; E-mail: calin@comm.pub.ro
2
Technical University of Timisoara, Electronics and Telecommunications Faculty, 2,
Vasile Parvan Bvd., Timisoara, Romania; Phone: +40256 403325; E-mail: lucaciu@etc.utt.ro
3
Military Technical Academy, 81-83, George Cobuc Bvd., Bucharest, Romania;
Phone: +4021 3354665/ext. 0308; E-mail: dandrei@mta.ro
394
corresponding to the interfering user k, and K is the increases the number of users accommodated for the
total number of users. same mean BER, by
The interference term rk ,i is written in terms of the
autocorrelation function C k as [3], [4]: K optimum
1.1547 , (7)
K white
N 1
rk ,i = 2C k (0)C i (0) + 4 C k (l )Ci (l ) + for large numbers of users. It is obvious from (7) that
l =1 the optimum case increases the number of users by
, (2)
N 1 more than 15% than the white sequences case, for the
+ [C k (l )Ci (l + 1) + Ci (l )C k (l + 1)] same mean BER.
l =0
The output signal of a Rician nonselective fading
channel is the sum of a non-faded version of the input
It is known that when perfectly random sequences signal (specular component) and a non-delayed
(white noise-like sequences) are employed the overall Rayleigh faded version of the input signal (scatter
interference variance from (1) can be written as [3], component). All communications links are assumed to
[6]: fade independently. We also assume that all users
have the same faded power ratio 2 .
PT 2 (K 1)
A2 , white (i ) = , (3) Under the SGA assumption, the average BER for any
6N user i over the Rician fading channel is given by [6],
[7], [8]:
where the normalization Ci (0) = 1 was considered.
According to [6] and [7] assigning the same
interference variance to every user can attain the P
T
lower bound of average BER for all the users. This BERA (i ) = Q 2
, (8)
can be done with C k (l ) = C i (l ) = C (l ) , for all k,i, and
( )
2
2 PT + 1 + 2 2 (i ) + 2
l. With the normalization Ci (0) = 1 , the solution that
A n
4
minimizes the interference power in (1), considering
the expression (2) for each term in the sum, leading to where A2 (i ) is the overall (non-faded) interference
minimum BER under the SGA assumption is given power for the desired ith user from all other users in
by: N T
an A-DS-CDMA, n2 = 0 is the variance for the
4
r l N r N l
C k (l ) = ( 1)l , l = 0,1, 2, ... , N 1, k (4) additive Gaussian noise with two-sided PSD (Power
r N r N Spectral Density) N0/2 [3], [5], and the numerator is
the is the useful component (the desired contribution
where r = 2 3 . from any user i). The Q-function is given by
Note that when l << N , C k (l ) ( r )l , which decays 1
2
Q( z ) = ey
2
dy .
exponentially with alternate sign. By introducing z 2
relation (4) into (2), the minimum interference power The theoretical BER for both optimum and white
is obtained for user i: sequences is presented in Fig. 1 for the following
system parameters values: K {3, 5, 7, 9}, N =31 and
PT 2 (K 1) 3 r 2 N r 2 N 2 = 0.1 .
A2 , optimum (i ) = , (5)
12 N r 2 N + r 2 N 2
III. OPTIMAL SETS OF CHAOTIC SEQUENCES
For large values of N, the second term in (5) rapidly FOR THE A-DS-CDMA SYSTEM
decreases to 1. It is important to note that the second
term in (5) is very close to 1 even for N=5 (it differs A second-order time-averaged statistic of the
from 1, starting with the sixth decimal value). With spreading sequences is needed for the sequence
this approximation the minimum interference variance design and performance analysis. According to the
is: ergodic theory the autocorrelation function of a
sequence generated by a measure-preserving and
ergodic transformation can be estimated statistically
PT 2 (K 1) 3
A2 , optimum (i ) = , (6) [6].
12 N An example of ergodic transformation is the nth
degree Chebyshev polynomial defined by:
Comparing the optimum case with the case when
white sequences are employed, the first one offers an Tn (x ) = cos(n arccos(x )) , (9)
increase in the system BER performances which
395
where x takes values from the interval [-1, 1]. It was
shown that the Chebyshev polynomials of degree 1 N N 1
n 2 are mixing and thus ergodic and they have an C (0 ) = P 2 (xi ) P (x ) (x )dx =
2
N i =1
invariant measure. 1
(11)
Performance of asynchronous DSCDMA (N=31), over a fading AWGN channel, 2 = 0.1
=
2
1 r 1 ( r 2 N not )= A
( )
-1
10
2 1 r2
C (l ) 1 1 N N
R
E A
=
A N l
P(xi )P(xi+l )
B i =1
( )
Th. wh. fad. BER K=3 1
-3
1
P(x )P T p l (x ) (x )dx =
10
Th. wh. fad. BER K=5
Th. wh. fad. BER K=7 (12)
Th. wh. fad. BER K=9 A 1
(r ) , for finite l
Th. opt. fad. BER K=3
lN
r N l
( 1)l
Th. opt. fad. BER K=5
(r )
Th. opt. fad. BER K=7
N
10
-4 Th. opt. fad. BER K=9 r N
0 5 10 15 20 25
Eb/N0(dB)
It is easy to note that the average value of the left-side
Fig.1. The theoretical BER performances for optimum
sequences and white sequences A-DS-CDMA system over C (l )
frequency-nonselective Rician fading channel with AWGN.
term, E x0 , is equal to the right-side term in
A
The common faded power ratio is = 0.1 .
2
396
value of the energy-per-bit to noise DSP ratio DS-CDMA system using optimal sequences offers a
Eb/N0=18dB. The resulting BER values as a function capacity increase of about 15% than when white
of the number of users K is depicted in Fig. 3. sequences or Gold codes are used.
The sensitive dependency of chaotic maps on the
Performance of asynchronous DSCDMA (N=63), over a fading AWGN channel, 2 = 0.1 initial condition offers both a greater number of
-1
10
available sequences and security increase.
The simulation results show that optimal Chebyshev
sequences are better than Gold sequences in terms of
the average BER per user, which is consistent with the
analytical result presented Section II and depicted in
-2 Fig. 1.
10
There are some differences between the simulation
R
E and analytical results given the fact that Gold
B
sequences are not perfectly white, Chebyshev
sequences are in fact pseudo-optimal, and the SGA
approximation is not quite valid for a small number of
10
-3 Th. wh. fad. BER K=10 users.
Th. opt. fad. BER K=10
Sim. Gold fad. BER K=10
Sim. Cheby. fad. BER K=10 REFERENCES
0 5 10 15 20 25 30
Eb/N0(dB) [1] E. H. Dinan, B. Jabbari, Spreading Codes for Direct Sequence
CDMA and Wideband CDMA Cellular Networks. IEEE
Fig.2. Theoretical and simulated BER for A-DS-CDMA Communications Magazine, no.9, pp. 48-54, September 1998.
system over frequency-nonselective Rician fading channel [2] O. Feely, Nonlinear Dynamics of Discrete-Time Electronic
with AWGN, using Chebyshev and Gold sequences (N=63, Systems, IEEE Circuits and Systems Society Newsletter, vol. 11,
K=10). The common faded power ratio is = 0.1 . 2 no. 1, pp. 1-12, March 2000.
[3] M. B. Pursley, Performance Evaluation for Phase-Coded
Spread-Spectrum Multiple Access Communication Part I: System
Analysis, IEEE Transactions on Communications, vol. COM-25,
Performance of asynchronous DSCDMA (N=63), over a fading AWGN channel, 2 = 0.1 no. 8, pp. 795-799, August 1977.
[4] M. B. Pursley, D. V. Sarwate, Performance Evaluation for
Phase-Coded Spread-Spectrum Multiple Access Communication
Part II: Code Sequence Analysis, IEEE Transactions on
-2 Communications, vol. COM-25, no. 8, pp. 800-803, August 1977.
10
[5] T. S. Rappaport, Wireless Communications Principles and
Practice, Prentice-Hall, 1996.
[6] C.-C. Chen, K. Yao, K. Umeno, E. Biglieri, Design of Spread-
Spectrum Sequences Using Chaotic Dynamical Systems and
Ergodic Theory, IEEE Transactions on Circuits and Systems I:
R
E
B
Fundamental Theory and Applications, vol. 48, no. 9, pp. 1110-
1114, September 2001.
[7] G. Mazzini, R. Rovatti, G. Setti, Interference Minimization by
Autocorrelation Shaping in Asynchronous DS-CDMA Systems:
Chaos-Based Spreading is Nearly Optimal, IEE Electronics
Th. wh. fad. BER EbN0 = 18 dB Letters, vol. 35, no. 13, pp. 1054-1055, June 1999.
Th. opt. fad. BER EbN0 = 18 dB
Sim. Gold fad. BER EbN0 = 18 dB
[8] E. Geraniotis, Direct-Sequence Spread-Spectrum Multiple-
Sim. Cheby. fad. BER EbN0 = 18 dB Access Communications over Nonselective and Frequency-
-3
10 Selective Rician Channels, IEEE Transactions on
12 14 16 18 20 22 24 26 28 30
Communications, vol. COM-34, no. 8, pp. 756-764, August 1986.
K (no. of users)
[9] C. Vladeanu, D. Andrei, BER Performance Evaluation for
Asynchronous DS-CDMA System using Optimal Chaotic
Fig.3. Theoretical and simulated BER for A-DS-CDMA Spreading Sequences, IEEE International Conference
system over frequency-nonselective Rician fading channel COMMUNICATIONS 2004, Proceedings, vol. 1, pag. 153-159,
with AWGN, using Chebyshev and Gold sequences (N=63, Bucharest, Romania, 3-4 June 2004.
K{12, , 30}, Eb/N0=18dB). The common faded power ratio
is = 0.1 .
2
V. CONCLUSIONS
397
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract - In the paper "Impulse generation with Telekom networks were performed. The authors
appropriate amplitude, length, inter-arrival, and abandoned the Henkel-Kesler (HK) model which was
spectral characteristics," by I.Mann et.al., [1] the proven to be more realistic for the modeling practice
authors present what the guest editors believe is the in favour of Weibull distribution in order to facilitate
most recent statistical model of nonstationary impulse the use of the results of Tough and Ward [3] on
noise.
The Henkel/Kessler (HK) model discussed in that paper
random noise generation with prescribed amplitude
proved to be a good fit for all measured impulse noise and spectral characteristics. In the following, we shall
voltage amplitude distributions collected in both the apply the Tough and Ward method to the HK model.
networks of Deutsche Telekom (DT) and British
Telecom (BT). Nevertheless, in order to facilitate the use II. STATISTIC MODELING FOR THE
of the results of Tough and Ward [3] on random noise PROBABILITY DENSITY OF IMPULSES
generation with prescribed amplitude and spectral AMPLITUDE
characteristics, a Weibull type density was investigated
as a possible alternative since it simplfies an
approximate realization of the stochastically varying The original model known as HK (Henkel- Kessler)
spectral properties. The authors recognized that HK model was proposed in [2]
model is a better fit than the Weibull density and can be Probability density function of the impulses
considerd as more realistic while suggesting that in amplitudes is given by:
further studies the Tough- Ward method for the DT
data sets will be finalized. u
This paper proposes a suitable method for simulating ( ) 1/ 5
1
impulse noise with Henkel/Kessler amplitude probability f i (u ) = e u0 , u0>0 (1)
240u 0
density function and impulse length according a stable
probability density function.
It tells that this is a probability density symmetric to
Keywords- impulse noise, nonstationary noise, xDSL, - the origin.
stable distribution. It can be demonstrated that the transformation of
1
u5
variable y = leads to a gamma symmetric
I. INTRODUCTION u0
1 y
distribution = y 4e .
Telecommunication companies and equipment 2(5)
manufacturers are interested in modeling the impulse The model was proven to be appropriate for the
noise that is disturbing the xDSL systems. I. Mann et information gathered from both networks (BT and
al. [1] present a statistical model considered to be the DT). Nevertheless, in order to facilitate the use of the
most recent nonstationary noise impulse model. A results of Tough and Ward on noise generation, a
method for the simulation of noise impulses with symmetric Weibull probability density was used
given amplitude, length and spectral density
characteristics is proposed. 1 b y
Impulse noise is considered to be one of the main P( y ) = 12 b y e (2)
causes of signals degradation in xDSL systems. That
is why companies are interested in modeling this The parameters for theWeibull and HK models in BT
noise. A noise impulse model must describe, in a and DT network measurements are given in table 1:
statistical sense, both time domain and frequency
domain impulses properties.
In Mann & Henkel model, the parameters are chosen
according to the empirically obtained statistics when
measurements in the British Telecom and Deutsche
398
Table 1
Weibull HK 1 y 1
Il = P(( )1 / 5 ,5) + (7)
a b u0 2 u0 2
BT(CP) 0,263 4,77 9,12 V
DT(CP) 0,486 44,4 23,23 nV The right side integral is known
DT(CO) 0,216 12,47 30,67 nV
1 x 1
Ir = erf ( )+ (8)
CP Customer Premises 2 2 2
CO- Central Office
u
2 t 2
In order to test the xDSL systems, synthetic impulses where erf (u ) = e dt is the error function.
are generated using the Tough-Ward method that 0
combines the amplitude probability density function The memoryless nonlinear transform (MNLT)
with the correlation function model to produce y = g (x ) can be numerically obtained in Matlab from
impulses with appropriate time domain and frequency
domain properties. the incomplete gamma function gammainc(y,a) and
First of all, this method assumes to find a memoryless the reverse of the error function erfinv(w). The result
nonlinear transform (MNLT) that maps between a is a symmetric with respect to the origin
zero-mean, unit variance Gaussian probability density function y = g (x ) (Fig 1 left)
and the required probability density function. This is In the following the correlation function of the
then used to calculate the relationship between process y is evaluated. This can be expressed in the
correlation coefficients of the two processes. Once form:
this relationship is found, then it is possible to impose
n =
a correlation onto the input Gaussian sequence of 1 RG (t )
given length by filtering with a FIR having a spectrum (0), (t ) =
2
n=0 2 n n!
that corresponds to the input correlation function.
The Gaussian filtered sequence is fed to the 2
x2 x
memoryless nonlinear transform in order to generate exp( ) H n g ( x)dx
impulse with given amplitude and spectral density
2 2
characteristics.
(9)
To find the memoryless nonlinear
transform y = g ( x ) , the cumulative distribution where H n are the Hermite polynomials of nth degree.
functions for the normal pdf and for the required pdf Once we have evaluated the integrals
are equated
x2 x
) H n g ( x)dx
u t2
( ) 1/ 5
exp( (10)
2
1 1
e u0 du = e 2 dt (4)
2 2
240u 0
y x
we have a power series representation of the mapping
Left side integral can be calculated: between the correlation functions of the input
Gaussian and the output non-Gaussian processes. This
( )
1 4 series are rapidly convergent.
Il = v + 4v 3 + 12v 2 + 24v + 24 e v (5) The integral (10) is numerically evaluated and the
48 resulting polynomial is used to generate a lookup
table relating the input and output correlation
1/ 5
y coefficients.
where v = . So far, we have established a readily evaluable and
u0 invertible mapping between the correlation functions
For the purpose of this paper it is recommended to use of the input Gaussian and the output non-Gaussian
the incomplete gamma function given by: processes related by the nonlinear
transformation y = g (x ) .
x
Using this we can tailor the correlation properties of
1
P ( x, a ) = t a 1e t dt , x R, a R+ (6)
(a ) the input Gaussian process through the methods
0 described in [1].
399
6
x 10
6 1
4
0.5
2
0
y 0 Ry
-0.5
-2
-1
-4
-6 -1.5
-4 -2 0 2 4 -1 -0.5 0 0.5 1
x Rg
Fig. 1. The mapping between the input and the output correlation functions under the MNTL for HK density
Frequency
-3
x 10
3
2.5
III. A NEW MODEL FOR LENGTH
DISTRIBUTION 2
1 t 0.5
ln 2
1 2 s 12 t1
f l (t ) = B e
2 s1t 0
0 500 1000 1500
(11)
1 t 2
time
2 ln
1 2s 2 t2
+ (1 B) e
2 s 2 t
The stable law is a direct generalization of Gaussian
distribution and in fact includes the Gaussian as a
The typical parameters of the model are given in limiting case.
table 2 The main difference between the non-Gaussian stable
distribution and the Gaussian distribution is that the
Table 2 tails of the stable density function decay less rapidly
B S1 T1 S2 T2 than the Gaussian density function. This characteristic
(s) (s) of the stable distribution is one of the main reasons
BT(CP) 0,45 1,25 1,3 21,5 129 why the stable distribution is suitable for modeling
DT(CP) 1 1,15 18 - - signals and noise of impulsive nature.
DT(CO) 0,25 0,75 8 1,0 125 The stable distribution is very flexible as a modeling
tool in that it is determined by four parameters: 1) the
In this paper we propose an alternative model for location parameter a 2) the scale parameter b, also
length probability density function, a stable called dispersion, 3) the index of skewness and 4)
distribution the characteristic exponent. For more information
about the stable distribution, we refer the reader to
0 x<0 appendix.
b 2 3
pb ( x) = b
2x x 2
x>0
2x e IV. CONCLUSIONS
We presented the Tough-Ward procedure in the case
of Henkel Kessler model for the amplitude probability
density function, an unsolved problem in June 2002
when the paper [1] was published. A new model for
length distribution is proposed. This is an -stable
distribution with = 0.5 .
400
APPENDIX and F has the Pearson density
e
f := E[e itX ] is also called stable. A d.f. F is stable 1 y2
N ( x) = 2
dy (A.7)
if and only if (iff) for every collection X 0 , X 1 , ..., X n 2
401
distribution. An important consequence of (iv) is the (at points of continuity) where G is a proper
nonexistence of the second order moment (except for distribution not concentrated at a single point.Then:
the case = 2 ).
(i) There exists a function L that varies slowly
Proposition 5 Let X be an -stable random variable. at infinity and a constant with 0 < < 1
If 0 < < 2 , then such that
(ii)
EX
p
= if p x L( x)
1 F ( x) (A.15)
(1 )
and
(iii) Conversely if F is of the form ( A.15) it is
p
EX < , if 0 p < . possible to choose a n such that
nL(a n )
1 (A.16)
If = 2 , then a n
and in this case (A.14) holds with G = G
p
EX < for all p 0 . (from the theory of regular variation a positive
function L defined on (0, ) varies slowly (at ) if
All non-Gaussian stable distributions have infinite L( sx) s
variance. for all x > 0 , 1 ).
L( s )
Let X 1 , ..., X n be a collection of i.i.d. r.v. and X ( n )
the largest among them. If the X j have the stable REFERENCES
density (A.4) then
[1] I. Mann, S. McLaghlin, W. Henkel, R. Kirkby,and T. Kessler, "
{ }
Impuse Generation with appropriate amplitude length inter-arrival,
2 / x
P n 2 X ( n ) x e b (A.9) and spectral characteristics," IEEE J. Select. Areas Commun., vol
20 pp. 901-911, June 2002
[2] W. Henkel, T. Kessler and H.Y. Chang, Coded 64-CAP
Proof: If a limit distribution G exists we have ADSL in an impulse noise environment - -Modeling of impulse
noise and first simulation results, IEEE J. Select. Areas Commun,
F n (n 2 x) G ( x) at all points of continuity. Passing vol 13, no.9, 1995, pp. 1622-1633.
[3] R. J. A. Tough , K. D. Ward, " The correlation properties of
to logarithms we get gamma and other non-Gaussian processes generated by
memoryless nonlinear transformation," J. Phys. D: Appl. Phys. vol
n[1 F ( n 2 x)] log G ( x) (A.10) 32 pp.3075-3084, Dec. 1999.
[4] W. Feller, An Intoduction to Probability Theory, vol II, New
We have for n York: Wiley, 1965.
[5] H.- J. Rossberg, B. Jesiak and G. Siegel Anallytic Methods of
Probability Theory, Berlin, Akademie-Verlag, 1985
b
n 2 N 1 b 2 = log G ( x)
xn
2 x
(A.11)
1
x [1 G ( x)] x (A.12)
(1 )
x 0
e x G ( x) 0 . (A.13)
F *n ( a n x ) G ( x ) (A.14)
402
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
tx (t ) = ( )d ,
pulse position modulation (PPM). Typically an UWB is the received pulse will be
defined as any signal in which the 3 dB bandwidth of the
signal is at least 25 percent of its center frequency. UWB t
signals are using previously allocated RF bands, by hiding A (t ) + n(t ) , where the constant A and are the
signals under the noise floor. attenuation and, respectively, propagation delay
Keywords: spread spectrum, time hopping, ultra-
experienced by the signal. The noise n(t ) is modeled as
wideband modulation,
AWGN with two-sided power density N 0 2 Watts/Hz.
I.INTRODUCTION
In this paper we consider that a UWB pulse is
Recent development in the area of wireless modulated by the second derivate of a Gaussian
communications systems indicates that ultra wideband 2
(UWB) technology is an attractive solution for short- t
function exp 2 properly scaled. In this case
range multiple-access communications due to a number tn
of attractive characteristics.
the transmitted pulse is:
The UWB signal is obtained using the impulse radio
technique, obtained by combining the Pulse Position t
2
Modulation (PPM) with Time Hopping Spread tx (t ) = t exp 2 (1)
Spectrum (TH-SS) technology. Impulse radio tn
communicates using baseband impulses of very short
duration, typically on the order of a nanosecond, thereby and the received pulse is :
spreading the energy of the radio signal very thinly from
near d.c to a few gigahertz. Even when those pulses are
applied to appropriate designed antennas, they t
2 t
2
1
(1)
University POLITEHNICA of Bucharest,
( ) =
E (t ) (t )d > -1 (3)
Electronics, Telecommunications and Information Theory Faculty,
Telecommunications Department,
1-3 Iuliu Maniu Blvd, 061071,Bucharest 6, Romania where:
Phone: (4021) 402 4996; Fax: (4021) 410 2379
ofratu@elcom.pub.ro, shalunga@elcom.pub.ro
403
+ The bandwidth for (t ) is determined by :
2
E = ( ) d (4) ( ft )2 ( ft )2
F ( f ) = 2tn n
exp n (6)
2
2
is the energy of the signal.
1 2
The final relation is given by: which has a maximum at f = = 1.7842 GHz .
tn
2
4 2
4 2
( ) = 1 4 + exp (5)
tn 3 tn tn
transmitted pulse received pulse autocorrelation signal bandwidth for the elementary pulse
0.1 0.25
0.08 1 1
-0.1 -1 -1 0
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4 0 2 4 6
time [ns] time [ns] time [ns] frequency [GHz]
Figure 1. The waveforms for transmitted pulse, received pulse, autocorrelation signal and the bandwidth for the elementary pulse
404
be easity seen that the frequency response of this vulnerable to occasional catastrophic collisions in
equally spaced pulse trains include both continuous and which a large numbers of pulses from two signals are
discrete spectral lines at regular intervals, so multiple- received at the same time instant.
access signals composed of uniform spaced pulses are
Uniform spaced pulse train
1
0.5
e
d
ut
i
pl
m 0
a
-0.5
0 20 40 60 80 100 120 140 160 180 200
time [ns]
Power Spectrum
0
-10
]
B
d[
r -20
e
w
o
P
-30
-40
0 0.5 1 1.5 2 2.5 3
Frequency [GHz]
To eliminate the catastrophic collisions that may multiple-access interference in many situation can be
occur in multiple access systems, each link, (indexed modeled as a Gaussian random process.
by ) is assigned to a distinct pulse shift pastern
{c } which is referred refer to as a time hopping
( )
k
Because the hopping code is periodic with period N p , the
code. These pseudorandom codes are periodic with
405
Randomised pulse train
1
0.5
e
d
ut
i
pl
m 0
a
-0.5
0 20 40 60 80 100 120 140 160 180 200
time [ns]
Power Spectrum
0
-10
]
B
d[
r -20
e
w
o
P
-30
-40
0 0.5 1 1.5 2 2.5 3
Frequency [GHz]
( )
(1 d m M ) symbol stream that convey the m = [mN s T f + (1) , (m + 1) N s T f + (1) ] (10)
information offered by the in some form to digital data.
This information is transported over the radio channel and
Nu
using the above pulse train, by modifying the pulse
positions. It should be noted that the pulse time delay ntot (t ) = A( ) x ( ) (t ( ) ) + n(t ) (11)
introduced due to modulation is relatively small =2
compared with the time delay resulting from pulse
spreading with the spreading code (pulse When the receiver is perfectly synchronized to the first
randomization), so the effects of pulse position user signal (e.g. having learned the value of (1) ), the
modulation on the power spectrum are insignificant. receiver is able to determine the sequence { m } of time
The received signal can be modeled as: intervals, with interval m containing the waveform
(1)
Nu representing data symbol d m . In this case the detection
r (t ) = A ( ) x ( ) (t ( ) ) + n(t ) (8) problem reduces to coherent detection of M equally-
=1 energy, equally-likely signals in the presence of
where multipleaccess interference in addition with AWGN,
- A( ) is the attenuation of user s signal over the and therefore the optimal receiver is a complicated
radio channel, structure that takes advantage of all of the receivers
knowledge regarding the characteristics of multiple-
- ( ) represents time asynchronism between the access interference.
clocks of the user s transmitter and receiver, and
- n(t) represents non-multiple-access interference Due to the complexity of the analysis, the multi-user
modeled as AWGN. detector will not be considered here. Instead we will
assume that ntot (t ) is a zero-mean Gaussian random
If the receiver wants to demodulate a particular user
signal (lets say user 1), representing the m-th data process. Hence, the detection problem becomes
(1) (1) coherent detection of M equal-energy, equally-like
symbol d m , where d m is one of M equally-likely signals in presence of a mean-zero Gaussian
symbols, then the received signal r(t) is: interference in addition to AWGN.
406
In this paper we consider data been carried by if AWGN any TOR > T will perform identically, so we
orthogonal signals. choose TOR = 2T .
To detect all the M signals we will need to correlate the
III. ORTHOGONAL SIGNALS input signal with all the M reference signals. In order to
design the receiver for the OR PPM signals we
Orthogonal signals (OR) represents a particular case of consider:
PPM TH-SS signals, for which the data is given by :
x(t ) = S j (t (1) C 01 (t (1) ) + n(t ) (15)
dik = [(k + i 1) mod M ]TOR (12)
The construction of orthogonal signals is therefore where S j is one of the signals in (14) ,
given by : ( m +1) N s 1
Si (t ) =
N s 1
( )
t kT f [(k + i 1) mod M ]TOR , i = 1,2,..., M (13)
( )
Cm (t ) = Tc ck( ) p(t kT f ) (16)
k = mN s
k =0
where TOR > T . In this paper a simulation has been and n(t) is AWGN. In this case, each of the M channel
correlation output can be written :
performed for Nu = 1 , and therefore the time hopping
{ } and the delay
N sT f
ck( ) (1)
x(t )S i (t
(1)
sequence have no effect in the yi = C 01 (t (1) ))dt
correlation properties of the PPM signals, and they were 0 (17)
omitted in this analysis. Ns 1 M 1
= k =0 q=0 q,[(k +i1) mod M ] z(k , q)
For the OR PPM signals the normalized correlation where:
coefficients are given by :
kT f + (1) + ck(1)Tc + ( q +1)TOR
1, if i = j
i, j =
x(t ) (t kT f ck(1)Tc qTOR )dt
(14) (1)
z (k , q) =
0, if i j
kT f + (1) + ck(1)Tc + qTOR
hence the normalized correlation matrix AOR is a
(18)
M M identity matrix.
and q,q ' is the Kronecker delta. From the expression
For a fixed impulse waveform (t ) , N s and T f , the for y i , i = 1,2,...M it is clear that the receiver neerds
orthogonal signal depends only on TOR . In the presence only one corelator and M store and sum circuits. This is
illustrated in figure 4.
y1
Store and Sum
kT f + (1) + ck(1)Tc + ( q +1)TOR
r(t)
kT f + (1) + ck(1)Tc + qTOR
z(k,q) y2
Store and Sum
(t (1) kT f ck(1)Tc )
(t (1) kT f ) (1)
Figure 4. Receiver block diagram for the reception of the first users signals
407
In figures 5 and 6 can be observed the benefits of using radio modulation using orthogonal signals is potentially
block waveform modulation. By using values for M able to support thousands of users.
higher than 2, it is possible to increase the number of
users for a fixed probability of error as it is shown in
figure 5. REFERENCES:
0
[1] R.A. Scholtz Multiple Access with Time-Hopping Impulse
Modulation MILCOM93, Bedford, MA, Oct. 11-14, 1993, pp. 447-
-5 450.
[2] F. Ramirez-Mireles, R. A. Scholtz, Multiple access with time
hopping and block waveform PPM modulation, ICC Conference, pp.
-10 775-779, June 1998.
e
)
rr [3] Vicenzo Lottici, Aldo DAndrea, Umbro Mengali Chanel
b
P(
g
Estimation for Ultra Wideband Communications, IEEE Journal, vol.
ol
0
1 -15 20, no. 9, December 2002.
[4] Moe Z. Win, Robert A. Scholtz Impulse Radio: How it works
IEEE Communication Letters, vol. 2, nr. 2, Februarie 1998.
-20 M=2
[5] Paul Hansell, Selcuk Kirtay Ultra Wideband Compatibility
M=4
M=6
Final Report to the Radiocommunications Agency, January 2002
M=8
M=16
-25
0 100 200 300 400 500 600 700 800 900 1000
Nu
-5
-10
)
rr
e
b
P (
g
ol
0
1 -15
-20 M=2
M=4
M=6
M=8
M=16
-25
0 50 100 150 200 250 300 350 400 450 500
Nu
V. CONCLUSION
In this paper we have shown that for applications
requiring high data rate (1024 Kbps) combined with low
probability of bit error (10-8 ), impulse radio modulation
using orthogonal signals is potentially able to support
hundred of users.
408
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Abstract In this paper, the Wireless ATM rerouting can offer full-scale mobility together with all the range
procedures are analyzed and categorised, based on a of ATM service capabilities existent also in the fixed
standard network topology, derived from the Wireless ATM networks [8].
ATM reference model. A new operational concept for a Mobility management has two distinct
mobile ATM network called Mobile Network
Architecture based on Virtual Paths (MNAVP), in which
components: location management dealing with the
the network nodes are connected to each other via correspondence between the subscribers data and his
preestablished permanent virtual path connections with current location and handover management which
fixed capacity assignments, is being proposed and controls the dynamic re-routing and transfer of
described. Finally, the handover hysteresis concept is connection for the terminals crossing cell boundaries.
introduced and a hysteresis gain is defined and The frequency-domain supposed to be used for
calculated as the factor by which the handover rate is Wireless ATM, situated in the Ghz range, will imply
reduced through the use of the hysteresis. the existence of small size cells, to cope also with the
Keywords: WATM, handover, routing, hysteresis. increased demands regarding system capacity. This
will lead, in conjunction with a higher terminal
I. INTRODUCTION mobility, to a very large number of handover of virtual
connections. Furthermore, smaller cells have tighter
Over the last few years, one of the major delay constraints, as the overlapping distances of the
commercial successes in the telecommunications cells are smaller. The more complex handover
world has been the widespread diffusion of cellular procedure has higher requirements regarding radio
mobile telephone services, whose provision relies on ressource management functions for the air interface
sophisticated algorithms implemented by state-of-the- paired with network signalling and control functions
art dedicated computer equipment. Lately, the for handover control, Quality of Service (QoS)
challenge resides in upgrading the service offer to management and rerouting of the connection to the
mobile users to include high-speed data new network access point. Exactly these rerouting
communication services. A natural approach in this procedures are the subject of this paper. A new
direction is to adopt the Asynchronous Transfer Mode operational concept based on Virtual Paths is
(ATM) in the wireless environment, resulting in the introduced and the principles of handover hysteresis
so-called wireless ATM (WATM) network. However, are analysed using a discrete Markov chain model.
ATM was developed for fixed networks and mobility
management functionality had to be added to the II. WIRELESS ATM NETWORK ARCHITECTURE
traditional set of capabilities. Mobile or Wireless
ATM consists of two major components: the radio A various number of reference architectures can be
access part which deals with the extension of ATM taken into account when we talk about Wireless ATM,
services over a wireless medium and the mobile ATM ranging from simple mobile terminals to complex
part wich addresses the issue of enhancing ATM for systems containing mobile ATM switches built in in
the support of terminal and service mobility in the ships, planes or satellites. One standard reference
fixed portion of the WATM network. Wireless ATM scenario contains a broadband wireless access system
started as a technology designed to be used for LAN providing unrestricted roaming capabilities within a
or fixed wireless access sollutions, where low mobility certain area of continuous radio coverage (Fig.1). The
constraints are encountered. Further research projects base stations (Radio Access Point, RAP) are of
and standardisation activities coordinated by the ATM picocellular size and implement the physical transport
Forum demonstrated the feasibility of broadband radio medium, multiple access control, data link control and
access networks based on ATM technology, which basic radio ressource management capabilities. The
1
Ph.D. Student, Politehnica University of Timisoara, Faculty of Electronics and Telecommunications Engineering
Private: Pilsenseestr.5, 82229 Seefeld, Germany
Phone: +49-179-2960468, Fax: +49-179-332960468, E-Mail: marius_moise@yahoo.com
409
RAP does not necessarily have to provide ATM-based reestablishment of the virtual connection, in order to
physical transport, it could use as well any other reach their current point of access to the network. This
access technology, as for example CDMA, also implies, beyond signaling and handover control, a
because the error detection and correction capability process of rerouting of the connection in the ATM
of the ATM stack is typically low, since it was network. QoS control based on requirements coming
designed for a reliable network. For this paper, we from the connection itself has to be provided in order
assume though the existence of a ATM radio interface to ensure the lossless and in sequence delivery of the
capable of transmitting ATM cells over the wireless ATM cells during the handover process.
medium. Special Mobile ATM switches (MAS) are
positioned at the border of an ATM network, III. CONNECTION REROUTING IN WATM
supporting end-system mobility by possesing the NETWORKS
necessary extensions in the signalling and control
planes to provide functions for mobility management We can categorise the approaches for connection
and also connection handover. rerouting in four basic categories: full reestablishment,
All the RAPs associated with a particular MAS connection extension, incremental reestablishment and
form a so called zone of continuous coverage. multicast reestablishment. They are schematically
Terminal mobility inside a certain zone and the presented in Fig.2 and Fig.3, showing the connection
handover associated with it (intra-zone handover) is phases during two handover steps for the different
handled locally by the MAS itself. Neighbouring basic methods.
zones with uninterrrupted radio coverage can form an The most simple method is the complete
area in which, at any time, a RAP can be found to reestablishment of the connection. For each change of
hand a connection over to while the terminal is a RAP coverage area, due to terminal mobility, a
moving without restrictions. The size of such an area completely new VC connection is being set up
is not limited, it could take the size of the entire between the mobile terminal and its peer. This can be
network. done in absence of any defined handover control
functions, only by the interaction of the two end
ACS
systems. The major disadvantages consist in the very
long duration of the procedure and the complete
interruption of the service. Quite opposite to this
procedure, the connection extension handover keeps
ATM Switch the impact on a local scale.
Full Connection
MAS MAS MAS MAS reestablishment extension
RAP
Zone 1 MAS MAS MAS MAS MAS MAS
Zone 3 Zone 4
Zone 2 Intra-zone
handover Inter-area handover
Inter-zone handover
It is not mandatory that all the switches are able of Fig. 2: Connection reestablishment and connection extension
supporting end-system mobility, therefore we
introduce an hierarchically superior instance, called Each handover is only prolounging the connection
Area Communication Server (ACS), providing from the old RAP to the new RAP, by this achieving a
mobility control for a specific area. The ACS very high speed, with the cost of a high routing
represents a mobility supporting ATM switch in inefficiency, due to the fact that no rerouting of the
charge of processing the protocoll requests in case of a connection is performed. Loops can easily occur if the
inter-zone handover. It also serves as anchor point terminal is moving back and forth in a limited area,
(AP) for the active connections of the terminals inside between only few neighbouring cells. This method has
this area. By using the ACS, the impact of the end- to be combined with a routing optimisation algorithm,
system mobility on the network can be significantly otherwise ressources are waisted. An example
reduced, because there is no need anymore for illustrating this scheme is presented in [9].
mobility specific functionality outside the ACS area. The multicast concept is also dealing with
The disadvantage consists in the fact that connectivity inefficiencies regarding the utilisation of network
cannot be guarateed for terminals leaving this area. ressources. The multicast tree is established at
A consequence of the high mobility of the connection setup time and can remain static or be
terminals is the requirement of a permanent dinamically updated for the duration of the connection
410
[2],[16]. All routes leading to the RAPs which will be route, which facilitates fast handover connection
presumly used by the mobile terminal during the setup. Second, call admission control decisions only
connection are precalculated and assigned as a have to be taken in the switches terminating the virtual
complete set to the connection at call setup time. This path connections (MAS and ACS), again reducing
leads to a very fast handover procedure with a connection setup complexity. Further, the
minimum of signalling load because no extra routing establishment of a virtual mobile network is ideally
and call processing is necessary due to the preparation suited for QoS-management and QoS-guarantees in a
work done. multi-operator fixed network environment. The
heterogenous nature of multiservice WATM virtual
Incremental Multicast
reestablishment extension
connections with a broad variety of QoS-constraints
and requirements is paid attention to by separating
traffic with different QoS-characteristics onto different
VPs as proposed for fixed ATM networks [4], i.e.
connections with similar QoS-requirements are
MAS MAS MAS MAS MAS MAS aggregated in one VP-subnetwork. Connections
carrying multirate services can be aggregated in single
VPCs, as long as they can be statistically multiplexed.
Handover 1 Handover 2 Handover 1
Handover 2 Different types of services, e.g. CBR and VBR, which
reduce the statistical multiplexing gain when
Fig. 3: Incremental reestablishment and multicast extension transported together within a VPC, are separated onto
different, parallel VP subnetworks.
Last but not least, the incremental reestablishment
represents a more complex and therefore efficient ACS ACS
Physical VC-tree Logical VP-tree
scheme during which a rerouting decision is made for topology topology
each individual handover. This decision affects only a
portion of the connection, namely the one between the
new RAP and some Cross Over Switch (COS) inside MAS MAS MAS
the current ACS area [13]. The high efficiency and
handover speed are due to the calculation of the MAS MAS MAS
optimal path to the new destination RAP for each
handover. The probability of reusing the longest part
of the connection is quite high, which enables a fast
handover, without going through the loop of routing Fig. 4: MNAVP topology
decision.
The obvious advantages of the MVPA concept are,
however, achieved at the cost of losing bundling gain
IV. MOBILE NETWORK ARCHITECTURE
and a somewhat less efficient statistical multiplexing
BASED ON VIRTUAL PATHS
on the physical links, resulting in a reduced utilization
of physical resources. The VP-based virtual topology
Basically, there are two alternatives to operate the
networking concept in the fixed ATM subsystem is
mobile ATM network as described in Fig.1: either VP-
peered by VPC operated RAP-links (Fig.4), where
based or VC-based. We choose a VP-based mode,
again VPCs are used for multiservice traffic
defining a concept called the Mobile Network
management between the MAS and the radio access
Architecture based on Virtual Path (MNAVP). In the
points. With this two-staged approach in the MNAVP
MNAVP, the MAS are net-worked with their
design, handovers can be handled in a partly
corresponding ACS over the fixed ATM network via
distributed fashion, i.e. intra-zone handovers are
preestablished permanent Virtual Path Connections, as
handled locally by the MAS, whereas the ACS is
shown in Fig.4. The VPs of this architecture have
involved only in inter-zone handovers. To further keep
fixed capacity assignments defining a virtual mobile
part of the handover processing within a zone, the
network topology over the fixed ATM infrastructure.
zone can be extended virtually by the use of VPs
In this VP-based network, all intermediate ATM
connecting RAPs of neighbouring zones to a MAS so
switches between ACS and MAS are only performing
that actually in the MNAVP virtual network the zones
VP-switching (cross-connect functionality). Two
overlap to some extent (Fig.4, Fig.6). By that, the
VPCs carried on the same physical link are not
number of inter-zone handovers can be reduced, and
statistically multiplexed. Inside a single VPC
the MAS has to do most of the work in handover call
statistical multiplexing is being applied.
processing.
This virtual networking approach has several
Consequently, a VC connection in a wireless ATM
advantages. First of all, the preestablished VP
network consists of two different segments (Fig. 5):
topology eliminates the need for complex call routing
functions and switching table updates along the VC-
411
the fixed segment from the ACS into the fixed
network with the ACS operating as a Cross Over Anchor Domain A Anchor Domain B
Switch (COS). This segment ist established at call ACS ACS ACS ACS
setup time and doesnt change during the lifetime of a
call.
the mobile segment, which follows a
MAS MAS
predetermined VP-route in the mobile network from MAS MAS MAS
the COS to the current RAP. This segment has to be MAS MAS MAS
rerouted during a connections lifetime due to user
mobility. Anchor Domain A Anchor Domain B
MAS COS
At call setup, the connection is switched through
MT RAP
the ACS belonging to the actual physical domain. In
the situation of a inter-zone handover, when the MT is
crossing physical domain boundaries, it still remains
Fig. 5: Fixed and mobile segment of a VC connection
in the anchor domain, due to the logical structure of
The most complex handover situation occurs, the network. A rerouting of the connexion over a new
when the mobile moves to a zone belonging to a ACS is necessary only in the case of a repeated inter-
different ACS. This inter-zone handover situation zone handover. This applies also for the movement in
generates the highest call processing load. The ACS the backward direction. By thus, a routing hysteresis
responsible for the connection during call setup which prevents the frequent occurence of handover in
remains the anchoring point (COS) for that case of limited geografical mobility, is introduced.
connection, switching between the its fixed and The hysteresis helps preserving one of the significant
mobile segment. The mobile segment has to be advantages of the MNAVP-based rerouting procedure,
extended from that COS to the new RAP via the namely that inside an anchor domain, the mobile
second ACS. Therefore, each ACS establishes VPCs segment between ACS / COS and MAS consists of
with its neighbouring ACS, reserved for carrying the only one VC segment, which would be lost in case of
inter-ACS handover connections. For best support of handover to another domain.
wide area mobility, this inter-ACS backbone The analysis of the hysteresis procedure is made
network could have a full mesh topology. The based on an discrete Markov chaining model. Lets
MNAVP connection anchoring mechanism can consider, for the start, the case without hysteresis,
produce non-optimal routes, which could be optimized shown in Fig. 7 and 8.
Physical border
by allowing the anchoring point COS to change during
the lifetime of a connection. To reduce the number of
inter-zone handover, each MAS can be virtually
linked to more than one ACS. At connection setup
time, one ACS has to be chosen as the connections 2 5 8
COS. This decision can be based on load-balancing 1 10
considerations as well as mobility prediction analysis 3 9
6
in order to minimize the number of inter-zone Anchor domain 4 7 Anchor domain
A B
handover.
V. HANDOVER HYSTHERESIS IN MNAVP Fig. 7: Handover between neighbouring domains without hysteresis
412
p1,1 1 p3,1 p16,1
1 1 1
p2,3 1 p 4,3 p2,3
n
=1
p4,5 1 1 p6,5
i ,j
(1)
1 1 1
p5,6 1 p7,6 p5,6
j =1
p7,8 1
1 1
P =
H p9,9 1 p10,9 p8,9
1
The equation (2) describes the matrix of transition-
p11,10 p12,10 1
p11,10 1 p12,10 1
states
p12,14 p13,13 1
1 1 1
p14,13 1 p15,13 1
p1,1 p3,1 p15,16 0
p p3,2 (2)
1,2
p2,3 p4,3
p2,4 p4,4 p6,4 Now, the effect of the handover hysteresis can be
PH =
p4,5 p6,5
quantified. The hysteresis gain can be defined as being
p5,6 p7,6 the factor by which the frequency of occurance of an
p5,7 p7,7 p9,7
inter-zone handover between anchor domains can be
p7,8 p9,8
reduced.
p8,9 p10,9
p8,10 p10,10
p5 + p6 (7)
GH =
p 3b + p 8a
To be able to calculate the probability of
occurrence of an handover, the state-probabilities of
the Markov-chain are needed Physical border
4a 7a 8a
p1,1 1 p3,1
(4)
1 1
p2,3 1 p4,3
3b
p4,5 p6,4
5b 8b
1 1
PH = p5,6 1
p7,6 6b 10b
p7,8 p9,7 9b
1 1 4b 7b
p8,9 1 p10,9
Physical border
0
and with the condition p e = 1 for the sum of the state p1,2
p2,4
p4,5
p5,7 p7,8
p1,1 2a p4,4 p6,5 5a p7,7
probabilities, the equation is transformed in
1a p3,2 4a p5,6 7a 8a
p2,3 p6,4 p8,9
n 3a 6a
i =1
pi = 1 (5) p3,1
p4,3
p15,13
p7,6
p12,10
p8,11
413
Table 1: Parameter for the numerical example of handover signalling and call admission load, maintaining a low
hysteresis. handover connection setup latency and facilitating
Parameter Value high handover rates. The concept of handover
p1,1 , p10,10 0,8 hysteresis is being introduced, with the benefit of
decreasing the number of inter-zone handover, for
p4 , 4 , p 7 , 7 0,1
terminals showing limited geografical mobility. The
p2 ,3 , p8,9 0,3 frequency of occurrence of this cost-intensive
handover type can be decreased with beneficial effects
on the ressource budget, and this is demonstrated
GH using a Markov decision process. A hysteresis gain is
defined and calculated as the factor by which the
handover rate is being reduced through the use of the
25
hysteresis.
20
REFERENCES
15
[1] Acharya, A.; Rajagopalan, B.; Raychaudhuri, D.: Mobility
10 Management in Wireless ATM Networks, IEEE Communications
Magazine, vol. 35, no. 11, November 1997, pp. 100-109.
5 [2] Acampora, A. S.; Naghshineh, M.: Control and Quality-of-
Service Provisioning in High-Speed Microcellular Networks, IEEE
Personal Communications Magazine, vol. 1, no. 2, 2nd Quarter
0 p
5,6
1994.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
[3] Akyol, Bora A.; Cox, Donald C.: Rerouting for Handoff in a
Wireless ATM Network, IEEE Personal Communications
Fig.11: Hysteresis gain Magazine, vol. 3, no. 5, October 1996, pp. 26-33.
[4] ATM-Forum: ATM User-Network Interface Specification,
The main parameter which describes the hysteresis The ATM Forum, Version 4.1, November 2002.
[5] ATM-Forum: Domain-based rerouting for active point-to-
gain is the probability of occurrence of consecutive point-calls, The ATM Forum, Version 1.0, August 2001.
inter-zone handover. If this probability equals zero, [6] ATM-Forum: Wireless Mobile Terminal/Network Anchor
the MT crosses the border, in one direction, only once. Switch Handover Model, The ATM Forum Technical Committee
If it is high, the the movement area is very restricted Wireless ATM Working Group, Contribution ATMF 97-0265.
[7] Chen, T.S.; Liu, S.S.: Management and Control Functions in
and the frequency of the border crossing high. Fig.11 ATM Switching Systems, IEEE Network, July/August 1994, pp.
shows this dependency and it can be observed that 27-40.
even in the case of a non-local movement, when [8] Dellaverson, L.: Reaching for the new frontier, 53 Bytes
p 5,6 = 0 , the probability of handover-occurrence is The ATM Forum Newsletter, vol. 4, no. 3, October 1996.
[9] Eng, K.Y.; Karol, M.J.; Veeraraghavan, M.; Ayanoglu, E.;
reduced to almost one third. This is based on the fact Woodworth, C.B.; Pancha, P.; Valenzuela, R.A.: BAHAMA: A
that, due to the hysteresis, one intra-zone handover at broadband ad-hoc wireless ATM local-area network, ICC 95,
Conference Proceedings, IEEE, 1995, S. 1216-1223.
the border domain (handover 4 and 7 in Fig.7) is not [10] Marsan, M.A.; Chiasserini, C.-F.; Lo Cigno, R.; Munaf, M.:
immmediately followed by an inter-zone handover, Local and Global Handovers for Mobility Management in Wireless
because, as shown in Fig.9, in this case of restricted ATM Networks, IEEE Personal Communications Magazine, vol.
mobility area, only inter-zone handover are performed 4, no. 5, October 1997, pp. 16-24.
[11] Minoli, D.; Golway, T.: Planning & Managing ATM
(handover 4a and 7b). Networks. Greenwich: Manning, 1996.
[12] Mitts, H.; Hansn, H.; Immonen, J.; Veikkolainen, S.:
Lossless handover for wireless ATM, MONET Mobile Networks
CONCLUSIONS and Applications Volume 1 (1996), no. 3, pp. 299-312.
[13] Toh, C.-K.: A hybrid handover protocol for local area
In this paper a Wireless ATM reference wireless ATM networks, MONET Mobile Networks and
architecture has been presented, consisting of a Applications Volume 1 (1996), no. 3, pp. 313-334.
[14] Vgel, H.-J.: Handover switching in mobile ATM networks,
wireless access system coupled with the support of Conference Proceedings European Personal Mobile
mobile end-systems within the ATM network. The Communications Conference EPMCC97, Bonn, 30. Sept. 2. Okt.
network is subdivided in several areas served by one 1997. ITG-Fachbericht 145, Berlin, Offenbach: VDE-Verlag, 1997,
area server ACS, areas which are in fact handover S. 375-381.
[15] Walke, B; Petras, D.; Plassmann, D.: Wireless ATM: Air
domains providing uninterrupted radio coverage and Interface and Network Protocols of the Mobile Broadband System,
full mobility support. After a brief description of the IEEE Personal Communications Magazine, August 1996.
rerouting schemes proposed already, a new [16] Yu, Oliver T.W.; Leung, Victor C.M.: Connection
networking concept called MNAVP has been architecture and protocols to support efficient handoffs over an
ATM / B-ISDN personal communications network, MONET
proposed, establishing a virtual mobile subnetwork Mobile Networks and Applications Volume 1 (1996), no. 2, pp. 123-
over the fixed ATM infrastructure, based on 139.
preestablished virtual paths with fixed capacity
assignments. This network design minimizes
414
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Ph.D. Student, Politehnica University of Timisoara, Faculty of Electronics and Telecommunications Engineering
Private: Pilsenseestr.5, 82229 Seefeld, Germany
Phone: +49-179-2960468, Fax: +49-179-332960468, E-Mail: marius_moise@yahoo.com
415
area in which, at any time, a RAP can be found to (backward handover) or across the target RAPs air
hand a connection over to, while the terminal is interface (forward handover). The soft handover tries
moving without restrictions. The size of such an area to eliminate the disadvantage of the hard handover,
is not limited, it could take the size of the entire consisting of the interruption of the data stream during
network. the connection switchover, by establishing and
ACS
activating a second radio connection to the target
RAP. The areas covered by the Mobile ATM
handover techniques so far are presented in Fig. 2.
with losses
(WATM CS1)
Backward
ATM Switch
lo s s le s s
Hard
with losses
MAS MAS MAS MAS
(WATM CS1)
Forward
lo s s le s s
Handover
Backward
RAP
Zone 1
Zone 3 Zone 4
Zone 2 Intra-zone Soft
handover Inter-area handover
Inter-zone handover
Fig. 1: Architecture of the Wireless ATM network Forward
(impossible)
It is not mandatory that all the switches should be Fig. 2: Mobile ATM handover schemes
able of supporting end-system mobility, therefore we
introduce an hierarchically superior instance, called Until now, no WATM handover mechanisms which
Area Communication Server (ACS), providing allow forward handover and, at the same time,
mobility control for a specific area. The ACS maintain cell loss QoS guarantees, have been
represents a mobility supporting ATM switch in proposed. Nevertheless, future Mobile ATM systems
charge of processing the protocoll requests in case of a will demand the flexibility and robustness in handover
inter-zone handover. It also serves as anchor point control as well as an increase in QoS.
(AP) for the active connections of the terminals inside Soft handover is considered to be the handover
this area. By using the ACS, the impact of the end- sollution of future wireless network systems, the so
system mobility on the network can be significantly called 4G or Next Generation Mobile Systems. During
reduced, because there is no need anymore for a soft handover process, the MT is able to
mobility specific functionality outside the ACS area. communicate simultaneously with both RAPs.
The disadvantage consists in the fact that connectivity Therefore, each connection posesses two active
cannot be guaranteed for terminals leaving this area. mobile segments between the MT and the COS. There
A consequence of the high mobility of the are several known methods, developed for the fixed
terminals is the requirement of a permanent ATM networks, able to establish cell synchronicity
reestablishment of the virtual connection, in order to between two different paths [9],[10],[17]. Cell
reach their current point of access to the network. This synchronicity is mandatory for a adaptive and
implies, beyond signaling and handover control, a latency-free switching between the two paths,
process of rerouting of the connection in the ATM achieving by this a quality gain (macrodiversity). The
network. QoS control based on requirements coming periodical in-slot signalling procedure upon which the
from the connection itself has to be provided in order lossless handover scheme proposed in this article is
to ensure the lossless and in-sequence delivery of the based, is also suited to synchronise the two paths, as it
ATM cells during the handover process. has been described in the Alignment Server Method
[9],[10].
III. HANDOVER FUNCTIONS OF A MOBILE
ATM NETWORK IV. A LOSSLESS HANDOVER SCHEME
Several different handover protocols have been During the handover of virtual connection there
described in the literature [1][2][3][5][7][8][15][16] are some situations when errors occur [17]. These
[17]. Based on the number of simultaneously active errors cause on the downlink the loss of in transit cells
radio connections, one can distinguish between two due to forced handover decisions and they can also
main streams in the handover techniqes: the hard cause on the uplink disruption of ATM cell sequence
handover and the soft handover. In the case of the hard at the handover Cross Over Switch (COS) due to a
handover, it always exists only one active radio transfer delay mismatch between the old and the new
connection and the handover control flow can be path (Fig.3).
directed either across the current RAPs air interface The result of the cell loss is a degradation of the
QoS of a virtual circuit connection by affecting the
416
data stream integrity. Several methods, based on an in- confirmed cell stream segments in the COS, there is
band signalling approach with two-way handshake, for no need to perform an in-slot signalling handshake
preventing cell loss and disruption have been before the handover to enforce zero cell loss.
described in the literature [15],[18],[19]. Their
MT RAP old COS RAP new MT
advantage consists in their relative simplicity which
easies the implementation. Their main disadvantage
comes up in the case of a single signaling cell loss or
in emergency handover situations, when the mobile
Handover Initiation Signaling
looses the connection to the old cell before the
signalling cell reaches him.
MAS
417
segment basis. The in-slot tag completing the selected reception of the receipt tag, which makes the fill-up
cell stream segment is then looped back on both level of the buffer at least equal to twice the product
uplink diversity paths. This enables the COS to also bandwith-delay, to which the number of tags,
activate its dynamic path selection algorithm. Once generated during this time interval, has to be added.
the radio conditions are sufficiently stable to guarantee The maximum level can be calculated with the
reliable communication via the new RAP, dualcasting formula
can be terminated and the handover completed. Soft 2 t oldd
handover is highly resource intensive in terms radio C RB = RC (2 t oldd + T ) + (1)
frequency spectrum and cell transport bandwidth. On
T
the other hand, STH based soft handover does not rely in which RC represents the average cell rate of the
on cell segment retransmission and therefore does not connection. The result represents at the same time, in
require any retransmission buffering space in the case of a handover, the maximum number of cells
COS. The only buffering space required is a small which have to be retransmitted to the new base station
path synchronization buffer in MT and COS. RAPnew. The optimal tag-interval can be calculated
Moreover, soft handover provides the best QoS out of the equation
performance of all handover alternatives. The old and
C RB 2 t oldd
the new path are synchronized on the ATM cell level, = RC =0 (2)
facilitating path selection and therefore handover T ( T )
switching/connection rerouting in realtime without out of which we obtain the best tag-interval as being
intro-ducing any delay or disruption into the cell
2 t oldd
stream. It is therefore the most attractive handover Topt = (3)
alternative for real time services which require an RC
outstanding cell loss performance during handover. During the retransmission of the cells stored already in
the buffer, the new ones, coming from the fixed
V. BUFFER ADMINISTRATION segment of the connection have to be queued first, in
order to be sent segmentwise afterwards on the new
The central element of the STH scheme is the buffer mobile path. The buffer can be emptied only if the
storing copies of the ATM cells sent in downlink reading rate is higher than the average cell rate of the
direction. As soon as the in slot signalling function
has been activated for a connection, a first tag is connection R EB > RC . In this case, the time interval
introduced in the cell stream and, at the same time, the in which the content of the buffer is retransmitted is
first copy is stored in the buffer. The procedure is C
continued until the moment in time when the second t EB = RB (4)
R EB
tag sent is received from the MT, confirming the
reception of the first segment of the cell stream. Buffer
max
MT C RB
COS RAP new
old t DCOS
t Dd
tD M T
new
t Dd
tD H O
C RB
R dT +1
C
418
result, that all the cells forwarded from the buffer did down. The user data cells contained in the
not yet reach the MT. They are now retransmitted with retransmission buffer are sent to the new RAP
the increased rate R EB > RC towards MT and followed by the normal cell stream received from the
fixed segment. Handover completion is signalled by a
segmentwise confirmed. The buffer is emptied step by
HCI OAM cell which has also the role of triggering
step and the system regains a normal, stable state. The
the queueing of user data cells following this signal,
maximum fill-up level of the buffer can be calculated
until the post handover resynchronisation is initialized
out of the formula below:
by the MT at the new RAP. The buffer in the new
RC 2 t new RAP is activated upon receipt of HCI and starts
max
C RB = C RB (1 + ) + 2 RC t new + d (5)
T
d
R EB storing cells. The MT itself starts the transmission on
the new link, after completing the pysical part of the
VI. OAM IMPLEMENTATION OF THE STH handover, by sending also a HCI signal containing the
SCHEME numbering tag of the last SAI correctly received on
the old downlink together with the value of the
Obviously, the proposed algorithm relies on the synchronisation counter. In a similar way, the RAP
standard ATM OAM functionality [7],[14]. Basically, discards the appropriate number of cells from the
the in-band signaling mechanism used in order to buffer and starts regular transmission in both up- and
protect data corresponds directly to the OAM downlink direction. A final HCI signal sent on the
principle, in which connection specific management uplink by the MT is flagging successful handover
and operations information are transmitted inside the completion to the COS. The described mechanism
user-data stream. The delay measurement mechanism protects connections with high requirements regarding
used in the STH method for dynamically adapting the QoS from cell loss or cell sequence mismatch. The
retransmission buffer size is an adaptation of the VCC signalling flow used by the STH handover scheme is
OAM cell loopback. The loopback capability of the similar to the OAM F5 flow. In addition to the
ATM-OAM enables the dynamic insertion of both the described handover procedure, new signaling
intermediated connection points and endpoints and to functions are needed for proper handover triggering
be transmitted back by a third, remote point. and setup or shutdown of the mobile segment.
This remote point is in the case of STH the MT, VC-Segment for handover
which sends back on the mobile segment the
information generated by the COS. A separate counter COS
is used to determin the number of user data cells MT
RAP
received on the corresponding VC after the last
SAI
received numbering tag. This information is then used
to resynchronize the data stream after the handover. HCI
HCI
the buffer belonging to the corresponding connection,
in other words, the cells stored in it prior to the OAM flow of the F5 handover segment
transmission of the SAI signal. This assures that the
buffer contains only copies of user data cells either Fig. 7: OAM signals of the STH method
not yet received or not yet acknowledged by the MT.
The SAI process is continuous, in a loop, and is Due to the very small cell size, handover are
reinitialized after every successful handover attempt. occuring more and more frequently and they should be
Its main purpose is the management and the content- seen as a normal process of continuous improvement
refresh of the STH retransmission buffer. After the of the radio link rather than an emergency situation.
validation of a handover attempt and the establishment The aim of every handover scheme should be
of the connection between the COS and the new RAP, therefore to reduce the impact to the connection in
the buffer content is transmitted to the new RAP. terms of QoS to a minimum. To achieve this goal, the
Once the MT received the permission for OAM functionality for handover should be in charge
handover, it has to finish the transmission of user data of in-time recognition of a handover situation and
cells on the old connection, this being done with the fast switching of the active connections to the new
help of the MDI signal. This OAM cell marks the last acces point RAP, if possible before radio link failure.
user data cell transmitted in the uplink direction via This is different than standard ATM fault situations,
the old RAP. Upon receiving the MDI cell, the COS where a certain delay responding to alarm indications
ceases transmitting on the old downlink and updates is provided in the ATM layer waiting for lower layer
its routing tables to switch over the connection to the protection mechanism to be activated and at the same
new route via the new RAP and connect the time a short disruption of service in the order of
connections fixed segment to the new mobile several hundred milliseconds is regarded as acceptable
segment. After that, the old mobile segment is shut due to the singular nature of ATM VPC/VCC faults.
419
Therefore, the standard OAM fault management is not administration of the buffer storing copies of the ATM
suited to handle handover situations, a distinct cells sent in downlink direction on the mobile
handover management being necessary. The segment. Finally, the OAM implementation of the
implementation of this can still be realized together procedure is presented, underlying the advantages of
with the ATM OAM functions because, as seen this scheme in terms of maintaining QoS and easy
before, the special requirements are mainly in the field implementation in existing ATM OAM.
of the performance of the handover OAM functions.
REFERENCES
Table 1: OAM signals for STH handover management
Signal Use for [1] Acharya, A.; Rajagopalan, B.; Raychaudhuri, D.: Mobility
First cell sent to the new RAP Management in Wireless ATM Networks, IEEE Communications
Handover Complete Magazine, vol. 35, no. 11, November 1997, pp. 100-109.
for resynchronisation after
Indication (HCI) [2] Acampora, A. S.; Naghshineh, M.: Control and Quality-of-
successful physical handover
Service Provisioning in High-Speed Microcellular Networks, IEEE
Mobile Detach Indication Last cell sent to the old RAP
Personal Communications Magazine, vol. 1, no. 2, 2nd Quarter
(MDI) before handover 1994.
Synchronisation and Returned OAM signal [3] Akyol, Bora A.; Cox, Donald C.: Rerouting for Handoff in a
Acknowledgement containing the cell numbering Wireless ATM Network, IEEE Personal Communications
Information (SAI) information tag Magazine, vol. 3, no. 5, October 1996, pp. 26-33.
[4] ATM-Forum: ATM User-Network Interface Specification,
The ATM Forum, Version 4.1, November 2002.
In the particular case of STH handover, an OAM [5] ATM-Forum: Domain-based rerouting for active point-to-
cell stream per mobile VCC is required, meaning that point-calls, The ATM Forum, Version 1.0, August 2001.
the handover management cells are belonging to an F5 [6] ATM-Forum: Wireless Mobile Terminal/Network Anchor
type of OAM flows. This in-band signalling handover Switch Handover Model, The ATM Forum Technical Committee
Wireless ATM Working Group, Contribution ATMF 97-0265.
management stream is at the same time a segment [7] Chen, T.S.; Liu, S.S.: Management and Control Functions in
oriented stream, having the endpoints at the mobile ATM Switching Systems, IEEE Network, July/August 1994, pp.
terminal MT, access point RAP and respectively, 27-40.
COS. The OAM signals are shown in Tab.1 and the [8] Dellaverson, L.: Reaching for the new frontier, 53 Bytes
The ATM Forum Newsletter, vol. 4, no. 3, October 1996.
message flow for the acknowledgement of received [9] Edmaier; Fischer; Eberspcher; Klug: Alignment server for
numbering information and the synchronisation hitless path switching in ATM networks, Proc. of the International
process during handover is displayed in Fig. 7. Switching Symposium ISS, vol. 2, Berlin, 23.-28. April 1995, pp.
The signal having the most frequent occurrence is 403-407.
[10] Edmaier, B.: Pfad-Ersatzschalteverfahren mit verteilter
SAI, due to the fact that it is transporting a numbering Steuerung fr ATM-Netze, Ph.D. Thesis, Technische Universitt
label in order to exchange synchronisation points in Mnchen, 1996.
the user data stream between COS and MT. The SAI [11] Eng, K.Y.; Karol, M.J.; Veeraraghavan, M.; Ayanoglu, E.;
cells are the one that are inserted periodically into the Woodworth, C.B.; Pancha, P.; Valenzuela, R.A.: BAHAMA: A
broadband ad-hoc wireless ATM local-area network, ICC 95,
user data stream of a VC by the COS. Those cells Conference Proceedings, IEEE, 1995, S. 1216-1223.
contain the sequentially numbered tag defined in the [12] Iselt, A.: Ausfallsicherheit und unterbrechungsfreies
STH handover scheme. The user data cells arriving Ersatzschalten in Kommunikationsnetzen mit Redundanzdomnen,
after SAI are copied by the OAM module to the STH Ph.D. Thesis, Technische Universitt Mnchen, 1999.
[13] Marsan, M.A.; Chiasserini, C.-F.; Lo Cigno, R.; Munaf, M.:
retransmission buffer. The SAI cells have to be Local and Global Handovers for Mobility Management in Wireless
returned to the COS by the endpoint of the mobile ATM Networks, IEEE Personal Communications Magazine, vol.
segment of the VC which is the mobile terminal MT. 4, no. 5, October 1997, pp. 16-24.
When receiving an SAI signal, MT would reset the [14] Minoli, D.; Golway, T.: Planning & Managing ATM
Networks. Greenwich: Manning, 1996.
synchronisation counter, save the numbering tag [15] Mitts, H.; Hansn, H.; Immonen, J.; Veikkolainen, S.:
received with the SAI signal and send back the SAI Lossless handover for wireless ATM, MONET Mobile Networks
cell to the originating point. and Applications Volume 1 (1996), no. 3, pp. 299-312.
[16] Toh, C.-K.: A hybrid handover protocol for local area
wireless ATM networks, MONET Mobile Networks and
CONCLUSIONS Applications Volume 1 (1996), no. 3, pp. 313-334.
[17] Vgel, H.-J.: Handover switching in mobile ATM networks,
Based on a mobile ATM reference architecture, Conference Proceedings EPMCC97, Bonn, 30. Sept. 2. Okt.
consisting of a wireless access system paired with the 1997. ITG-Fachbericht 145, Berlin, Offenbach: VDE-Verlag, 1997,
S. 375-381.
support of mobile end-systems within the ATM [18] Vgel, H.-J.: Robust and Soft: handover design for high-tier
network, and after detailing the handover functions Mobile ATM systems, Wireless99, Conference Proceedings,
related to it, a new handover procedure called Sync Munich, October 6-8, 1999. B. Walke (Hrsg.), ITG-Fachbericht
Tag Handover STH is being proposed and discussed 157, Berlin: VDE-Verlag, 1999, pp. 333-338.
[19] Walke, B; Petras, D.; Plassmann, D.: Wireless ATM: Air
in this paper. To justify the need of a lossless Interface and Network Protocols of the Mobile Broadband System,
handover scheme, the potential error scenarios are IEEE Personal Communications Magazine, August 1996.
presented and also the way in which the new
procedure overcomes them. The tag insertion
mechanism is explained in detail together with the
420
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Department of Information Technology, University of Cooperative Education Loerrach,
Hangstrasse 46-50, D79539 Loerrach, Germany, e-mail: sikora@ba-loerrach.de
421
However, there are caveats connected with was postponed several times, giving room to the
standardization. argument that standards impede short time to
market.
It leads to additional overhead in the systems, if
functionality has to be implemented just for III. LEVELS OF STANDARD
conformitys sake. CONFORMANCE
Applications / Profiles
Application Application Application (1) Distinguishes the type of device from an end-user perspective
Object 1 Object 2 Object 3 Specified in Profiles
(2) Distinguishes the Physical Device Types deployed in a specific
ZigBee Applications Framework API ZigBee network
ZigBee Application Support Sub-Layer (3) Distinguishes type of ZigBee hardware - Based on 802.15.4 RFD
and FFD definitions
ZigBee General
ZigBee Alliance
422
Level of
Characteristics Rule 1: Lightweight devices shall not disturb
Standardization standard devices.
Devices use the cheap and low-power RF
Layer 1:
chips with proprietary L2-protocols Rule 2: Standard devices shall understand
Devices use the IEEE802.15.4-library for
messages from lightweight devices.
Layer 2: medium access, but use proprietary L3-
protocols
Devices use the ZigBee network Rule 3: Lightweight devices shall ignore messages
Layer 3: funktionality with own application not included in the lightweigt standard and shall
protocols not be obstracted.
Layer 7: Devices use ZigBee application profiles
Rule 4: All routines in the lightweight devices
Table 2: Characteristics of standardized devices with vertical covering parts of the full standard shall comply
modularity
with the format and the behavior. This is essential
for a smooth migration path to future
B. Necessity of compromises enhancements.
423
synchronized with the beacons that define REFERENCES
contention access periods (CAP) and contention
[1] http://www.chipcon.com/files/CC2420_Data_Sheet_1_2.pdf
free periods (CFP). However, if the guaranteed
[2] Road vehicles -- Controller area network (CAN) ISO
time slots (GTS) in the CFP are not kept free, as Standard 11898; http://www.iso.org
some light devices do not run a time-based access [3] http://www.ieee802.org/15/pub/SG4a.html
scheme, this has a destructing impact on the [4] http://webs.cs.berkeley.edu/
[5] http://www.zigbee.org.
quality of service in the slotted network.
Therefore, it seems to be necessary, that a device
not supporting slotted access detects a beacon
with GTS definition, should leave this channel by
selecting another channel (rule 4).
IV. CONCLUSIONS
424
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
1
Facultatea de Electronic i Telecomunicaii, Departamentul
Comunicaii Bd. V. Prvan Nr. 2, 300223 Timioara, e-mail marius.oltean@etc.utt.ro
2
Facultatea de Electronic i Telecomunicaii, Departamentul
Comunicaii Bd. V. Prvan Nr. 2, 300223 Timioara, e-mail andy.vesa@etc.utt.ro
3
Facultatea de Electronic i Telecomunicaii, Departamentul
Comunicaii Bd. V. Prvan Nr. 2, 300223 Timioara, e-mail eugen.marza@etc.utt.ro
425
generates difficulties regarding the RF power [Xk]
Signal [xk ] CP [xcp k ]
amplifiers that must be used in a practical IFFT Channel
mapping insertion
implementation. The transmitter power amplifier
requires an increased power back-off for obtaining a [ycp k ]
[Xk ] Channel [Yk ] [yk] CP
wide range of linearity in order to faithfully reproduce invert
FFT
extraction
all the peaks of the signal envelope, which will
significantly increase the cost of this component.
Detection
SC-FDE is an approach that counteracts this
inconvenience, keeping a low complexity of the
whole equalization process, similar to that of OFDM. Fig.1: Block diagram of an OFDM system
The data are transmitted using a single carrier with a
classical modulation scheme, which will eliminate the The data are first grouped in blocks of M bits.
problem of an expensive peak-to-average power ratio. Each m bits are then mapped to one of 2m complex
The whole complexity of the equalization is valued symbols Xk, k=0,1,,N-1 using a digital
distributed at the receiver and is still based on FFT modulation scheme.
processing. The basic mathematics of this approach The N time domain signal samples forming
are practically the same as in OFDM, as we will se in an OFDM symbol (x0, x1, ..., xN-1) are obtained
the next section. A frequency domain - linear or through IFFT processing according to (1):
adaptive - equalization must be performed in order to
counteract the inherent time-dispersive nature of the 2
N 1 jk n
radio channel. x[n] = X[k] e N , n = 0,1,..., N 1 (1)
k =0
In the next section of the paper we will study the
equalization concepts proposed by the two
approaches, which moreover could efficiently be used Afterwards, L-1 cyclic prefix samples are
simultaneously in a dual-mode. The BER added in front of the signal sequence, that becomes:
performance of both methods is investigated in the [xcp]=(xN-L+1, xN-L+2,...,x0,..., xN-1). This signal is
section III by means of computer simulation. The analog converted and modulates a RF carrier, before
relevant conclusions regarding the two methods are being transmitted in the channel. If we consider the
outlined in the final section. equivalent baseband discrete model of the channel as
a FIR filter of order L, then the Z-domain channel
II. OFDM AND SC-FDE: BASIC CONCEPTS response is given by:
L 1
A. OFDM
H (z) = h[n ] z n (2)
n =0
OFDM represents an optimized version of
the multicarrier modulation techniques. The finding of Since we must assume our channel to be time
this approach was to replace a single-carrier serial variant, its impulse response will depend on the time
transmission at a high data rate with a number of at which the impulse is applied. We shall assume,
slower parallel data streams that will simultaneously however, that the channel impulse response is static
modulate orthogonal carriers. By creating N parallel (will not change) for the duration of an IFFT frame.
substreams, the bandwidth of the modulation symbol The equivalent baseband signal at the
is reduced by the same factor, or, equivalently, the channel output can be obtained by the well known
duration of the modulation symbol will be N times operation of convolution:
higher. The lengthening of the transmitted symbol
will significantly reduce its sensitivity to ISI. The N
y cp [n ] = x cp [n ] * h[n ] (3)
parallel transmission channels do not interfere each
other, since their correspondent subcarriers are
orthogonals. This is in fact the basic idea that lies Discarding the L-1 CP samples from the
behind OFDM. The generation of the multiple carriers received sequence, the remaining (useful) signal can
is done by performing Inverse Fast Fourier Transform be expressed as:
(IFFT) processing at the transmitter. To the receiver,
data are recovered using FFT processing, which y[n]=x[n]h[n] (4)
extracts the subcarriers. In addition, a cyclic prefix is
inserted in front of each symbol, in order to prevent where denotes the circular convolution operator.
two consecutive blocks to interfere because of the The relation above is in fact the main reason that lies
time-dispersive channel character and, furthermore, to behind the use a cyclic prefix. Its insertion will
facilitate the equalization process to the receiver. In transform the convolution between the data sequence
the figure 1, the block diagram of an OFDM system is and the channel impulse response into a circular
shown. convolution [4], which preserves the temporal support
of the signal, thus avoiding the interference of two
successive OFDM symbols due to the time dispersive
426
nature of the channel. If the cyclic prefix duration spread of the channel, the interference introduced by
spans more than the multipath delay spread of the temporal dispersion of the previous transmitted block
channel, the interference from the previous is totally absorbed by the circular extension, which is
transmitted blocks is totally eliminated through this discarded to the receiver. Moreover, since the
operation of CP insertion/extraction. Furthermore, the equalization process is still based on FFT processing,
equalization process is facilitated at the receiver, a the appearance of periodicity that cyclic prefix
simple channel inversion allowing theoretically a confers to the signal will facilitate this process. The
perfect data recovering, as we will see next. mathematics of this method are essentially the same
Since x[n]=IFFT{X[k]} and taking into account as for OFDM with the difference that in SC-FDE, all
the FFT demodulator, the received symbols Y[k] can the equalization complexity is allocated to the
be expressed as: receiver. The data block [yk] arrived to the receiver is
Y[k]=FFT{IFF T{X[k]}h[n]} (5) first FFT-processed, then the influence of the
frequency-selective channel impulse response is
eliminated by an simple channel inversion operation.
But, the FFT of a circular convolution of two In adaptive SC-FDE, the adaptation of FDE transfer
discrete time signals yields a spectral multiplication: function can be done using least mean square (LMS),
root least square (RLS), or least-square minimization
Y[k ] = FFT{IFFT{X[ k ]}} FFT{h[n ]} = X[k ] H[ k ], methods. An inverse FFT returns the equalized signal
in the time domain prior to the detection of data
k = 0,1,..., N 1 (6) symbols. In terms of complexity, for channels with
severe delay spread, SC-FDE is simpler than time
where H[k] represents the sampled frequency domain equalization, because equalization is
response of the equivalent discrete channel, performed on a data block at a time, using a
corresponding to the frequencies k=k(2/N). The computationally efficient FFT algorithm.
crucial consequence of the relation above is that the Furthermore, as proposed in [2], OFDM and SC-
modulation symbol X[k] could be recovered at the FDE could simultaneously be used with high
receiver by a simple pointwise division operation, efficiency in a transmission modem. It is easy to
commonly referred to as a one-tap frequency domain notice from the presented block-diagrams that the two
equalizer. Thus, the CP theoretically eliminates both types of systems mainly differ in the placement of
IBI (each block preserves its temporal support) and IFFT operation. A system in which a radio modem
inter-carrier interference (ICI) (each serial symbol can be configured to work in both OFDM and SC-
received on the k-th carrier will depend only on the FDE mode could be simply implemented by
corresponding k-th carrier transmitted symbol, not switching the IFFT block between transmitter and
being affected by the adjacent carriers). receiver. As observed in [2] since SC-FDE
concentrates all the complexity to the receiver, such a
B. SC-FDE system could be appropriated for an uplink, thus the
most complex part of the communication issues being
SC-FDE is an alternative equalization approach, solved by the base station. Using OFDM in downlink
which eliminates some of the OFDM disadvantages will reduce the complexity of the processing that must
(especially the high peak-to-average power ratio), be done by the mobile station. Such an arrangement
while keeping approximately the same low has two obvious advantages: concentrates the main
complexity. The block-diagram of a SC-FDE system amount of processing in the base station and reduces
is illustrated in the figure 2. the power consumption of the mobile station that uses
a single-carrier mode for transmission and an OFDM
Signal [xk ] CP [xcp k ] mode for reception.
Channel
mapping insertion
III. EXPERIMENTAL RESULTS
[Xk ] Channel [Yk ] [yk] CP [ycp k ]
IFFT FFT
invert extraction BER performances of both OFDM and SC-FDE
systems were studied by means of computer
[ x k ] Detection simulation. Both Rice (Line-of-Sight) and two-ray
Rayleigh (Non-Line-of-Sight) conditions with perfect
channel knowledge were taken into account in order
Fig.2: Block diagram of a SC-FDE system to simulate the signal propagation in the radio
channel. As a parameter of the simulation, the relative
The data are still transmitted in blocks of N power of the two multipath components denoted by P1
samples, but this time the transmission is a classical and P2 in a Rayleigh fading channel was considered.
serial single-carrier transmission. The data are The Rice factor K, defined as the ratio between the
encoded using a digital modulation scheme, obtaining power of line-of-sight deterministic signal and the
the sequence [xk], where k=0,,N-1. A cyclic prefix power of the multipath components is also modified
is added in front of each data block. If the cyclic during the simulations. The fading was modeled as
prefix duration is longer than the multipath delay
427
quasi-static that is it remains unchanged during the almost identically in respect to these parameters. BER
transmission of a data block. For the simulated coded performance slightly improves when a LOS
transmissions, a BCH encoder with rate 1/2 was component is introduced, or when the relative power
implemented. This block-code can correct 6 bits in a of the second multipath decreases. The effectiveness
block of length N=64, as was taken in all the of the last parameter becomes clear especially for high
simulations. The binary information is mapped using SNR values, when the spread of the results is within
two classical constant-envelope modulation schemes: about 1dB (for SC-FDE) and 4dB (for OFDM). Thus,
QPSK and DBPSK respectively. the OFDM system proved to be slightly more
In the figure 3, a comparison of BER sensitive to the considered parameters than the other
performances of OFDM and SC-FDE methods is system.
illustrated. One can observe that for low values of
SNR, OFDM performs generally better than SC-FDE,
especially for coded transmission. When SNR
exceeds the critical range of 12dB to 15dB, SC-FDE
becomes more reliable, mainly when uncoded data
blocks were transmitted. This confirms that, while
desirable, the coding is not strictly imposed in single
carrier method, unlike in OFDM where the coding is
mandatory in order to combat the high amount of
errors on the carriers attenuated by the frequency-
selective channel response.
428
to a cyclic prefix that is insufficient in order account the modulation method used. As expected,
counteract the time-dispersive nature of the channel. DBPSK generally performs better than QPSK.
Though, one can notice that for high values of SNR,
SC-FDE overcomes OFDM.
All simulations were repeated using a differential
modulation scheme, namely DBPSK. In the figure 7,
BER performance of both systems with and without
encoding is illustrated.
429
SC-FDE seems to not improve in the same manner its
performances. Thus, the coding gain of SC-FDE REFERENCES
system with QPSK modulation is significantly inferior
to the same system used with DBPSK modulation. [1] D. Matic, OFDM as a possible modu lation technique
for multimedia app lica tions in the range o f mm waves ,
Furthermore, the coded SC-FDE systems, as available on-line at:
previously seen, performs generally worse than coded w ww .u bicom. tudelf t.nl/M MC/D ocs/introOFDM .pdf
OFDM, observation confirmed by the performance [2] M. Huemer, A. Koppler, L. Reindl, R. Weigel, A Review of
curves in the figure 12. Thus, the coding gain in this Cyclically Extended Single Carrier Transmission with Frequency
Domain Equalization for Broadband Wireless Transmission,
case covers a range of values between approximately
European Transactions on Communications (ETT), Vol. 14, No. 4,
4dB and 8dB. pp. 329-341, July/August 2003
[3] D.Falconer, S.L. Ariyavisitakul, A. Benyamin-Seeyar, B.
Eidson, Frequency Domain Equalization for Single-Carrier
Broadband Wireless Systems, IEEE Communications Magazine,
pp. 58-66, April 2002
[4] Chini, A., Analysis and Simulation of Multicarrier Modulation
in Frequency Selective Fading Channels ,Ph. D. Thesis, 1994,
Chapter 3
[5] Werner Henkel, Georg Taubck, Per lding, The cyclic prefix
of OFDM/DMT an analysis, 2002 International Zrich Seminar
on Broadband Communications, February 19-21 Zrich,
Switzerland
IV. CONCLUSIONS
430
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
I. INTRODUCTION
Fig. 1. Adjacent channel interference
The term battlefield spectrum management refers to
managing electromagnetic spectrum resources in
Interference power caused by the first upper and
order to support telecommunications (including
lower interfering signals is designated as ACI-1 and
weapon systems) and electronic warfare (EW)
the one caused by second signals is designated as
requirements [1]. This type of management includes
ACI-2.
allocating and assigning generated frequency
It is important to know the values of the spacing
resources and the distribution of the variables for
between channel (W) and bandwidth of the receiver
frequency hopping (FH) radio systems. The items of
filter. In several system specifications the receiver
the management include frequencies, TSK
filter is assumed to be a brick wall filter [2].
(Transmission Security Keys) variables, net
The integrated (total) ACI power can be expressed as
identifiers, COMSEC (Communications Security)
[2]:
variables, and time. However, equipment parameters
S y ( f ) H ( f w)
2
impose some constraints on the distribution schemes df
for TSK, net identifiers, and frequency allocation ACI = A (WTb ) = (1)
S y ( f ) H ( f ) df
2
schemes of the hop sets.
Radio frequency interference is inherent in all
wireless systems and is one of the most significant where: Sy(f) is the power signal density (PSD) of the
design parameters of cellular and other mobile radio signal; H(f) is the receive filter transfer function; w =
systems. WTb is the normalised carrier spacing between
This paper investigates the power efficiency adjacent channel.
performance of frequency hopping radio systems The method used to assess the ACI levels is a
operating in an adjacent channel interference (ACI) practical one and is based on measuring this
environment. parameter in two representative operating modes: the
Adjacent Channel Interference (ACI) is caused by fixed secure frequency mode and frequency hopping
modulation, filter and radio design imperfections. mode.
Transmitted signal is not band-limited to a brick The instrument we used is a R&S FSH3 Spectral
Analyzer [3]. Comparing the results we can provide
1
coala de Aplicaie pentru Transmisiuni, Informatic i Rzboi Electronic,
Departamentul de Cercetare, Bd. Vasile Milea 2-3, Sibiu, e-mail Paul.Bechet@personal.ro
2
Academia Forelor Terestre, Str. Revoluiei 3-5, Sibiu, e-mail demeter@actrus.ro
3
Centrul 196 Rzboi Electronic, Buzu
4
Academia Forelor Terestre, Str. Revoluiei 3-5, Sibiu
431
an efficient allocation of the resources for frequency hopping mode. Using a different frequency position in
hopping radio systems. the narrow hopping band, ACI level values can be
observed for a variable operation frequency of the
II. RESULTS AND DISCUSSION equipment. Due to the limited spectral analyzer
capability we are focused only to the results between
The results are representative for a simple tactical two frequencies of the successive hopps. The values
network where the power of the frequency hopping of these frequencies are 35 MHz and 36 MHz. For a
systems is uniform distributed in the coverage area. hop speed of 100 hops/s, it is necessary to set the
Fixed secure operation mode is one specific for acquisition time of the analyzer to 20 ms. The ACI
frequency hopping systems due to the spectral levels are presented in figures 3, 4 and figure 5. More
performance in the critical cases of propagation like details are shown in table 2.
urban or mountain areas. The levels of ACI are
presented in figure 2 and more details are shown in
table 1.
Table 1
Frequency Attenuation
distance (dB)
(kHz)
100 58.5
300 69
500 72.8
526 60.8
600 73.9
626 56.6
700 75.2 Fig. 4. The ACI levels in the Hopping Frequency Mode
780 56.3 at a frequency distance of 800kHz and 1.8MHz.
800 76.3
1040 56.4
1100 77.9
1400 78
1600 78.4
432
Bar frequency band
Table 2
Frequency Attenuation 2 MHz 2 MHz
distance (dB)
(kHz)
200 0.3
Fixed secure frequency Hopping frequency Fixed secure frequency
300 0.9 band
500 4.4
700 6.2 Fig. 6. An analysis of two radio networks, one
operating in the fixed secure mode and the second
800 8.9 operating in the frequency hopping mode
900 14.5
1100 18.1 List of frequency hopping network
1200 20.9
f1 f2 f3 f4 fn
1300 25.4
1400 29.3
1500 34 2 MHz 2 MHz
1600 37.9
1700 43
2000 50
Fixed secure frequency network
Comparing with fixed secure mode, one can observe
increased values of the ACI levels. An acceptable 40
Fig. 7. An analysis of two radio networks, one
dB ACI level requirement imposes a minimum of 1.7 operating in the fixed secure mode and second
MHz frequency distance. This means considerable operating in a list of frequencies hopping mode
constraints in the process of managing the radio
resources.
the case of inserting the frequency value of the fixed
III. CONCLUSION secure network.
The situation that we considered in this paper is valid
In this paper we investigate the ACI level for only in the case of uniform power of the frequency
frequency hopping systems in two representative hopping equipment.
operating modes: fixed secure frequency and
frequency hopping.
Fixed secure mode provides good spectral efficiency REFERENCES
but is easy to detect by the enemy.
Frequency hopping mode is less power efficient but [1] *** - FM 11-32 , Combat Radio Operations, Headquarters
provides ECCM (Electronic Counter Counter Department of the Army Washington, DC, 1999.
Measures) communications. [2] K. Feher, Wireless Digital Communications: Modulation &
Spread spectrum Applications, Prentice-Hall, New Jersey, 2000.
In the process of resources planning it is [3] *** - Operating Manual: Handheld Spectrum Analyzer R&S
recommended to combine the advantages of both FSH3, 2001.
operation modes, so we expect to have, in same area,
complex radio networks.
A simplified situation is presented in figure 6 where Acknowledgements:
the spectrum manager must allocate the frequency Present work was supported by a CNCSIS Grant / 2004 from the
resource between two radio networks, one operating Romanian Ministery of Education and Research to the Military
on a fixed secure frequency and second one operating Application School for Communications, Informatics and
Electronic War, in Sibiu.
in the frequency hopping mode.
A minimum 2 MHz frequency distance, upper or
lower, is requested between the frequency resources
of the radio networks, if a 40-dB ACI level
performance is considered.
Another situation is the one represented in figure 7,
where the frequency hopping radio network is a
particular one, operating with a list of frequencies. A
minimum of 4 MHz frequency distance is
recommended between two consecutive frequencies
of hopping list. This requirement is necessary only in
433
Buletinul tiinific al Universitii "Politehnica" din Timioara
Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
where:
I. INTRODUCTION
b
k = a k2 + bk2 et k = arctan k (3)
The identification of the digital modulation type of a ak
signal has found applications in many areas, including It is clear that QAM signals contain a phase as well as
electronic warfare, surveillance and threat analysis. an amplitude modulation. Compared to PSK signals,
Recognition of communication signals became an one distinction in QAM signals is that it does not have
independent discipline in the electronic warfare. The constant amplitude. This type of modulation is used in
goal is to intercept, analyze, classify and, eventually, high-speed modems.
understand the message carried out by the The FSK (Frequency Shift Keying) modulation can be
communication signals. considered as a non-linear frequency modulation. It
In this paper a potential method is proposed to can be represented as:
classify different types of numeric modulations. In
section II the characteristics of the PSK, QAM, FSK, m FSK (t ) = cos(2f c t + (t )) (4)
and OFDM modulations are described. In section III
the time-frequency representations based on where (t) depends on the integral of modulator
polynomial phase signal processing are presented. signal. Thus, the frequency varies with the message
The new method for recognition of the OFDM and the information is carried out by the instantaneous
modulations is illustrated in section IV. Furthermore, frequency.
in section V some simulation results are depicted. In comparison with the modulations described
Section VI will close this communication. previously, which are single-carrier modulation
techniques, the OFDM (Orthogonal Frequency
Division Multiplex) modulation is a multiple-carrier
II. DIGITAL MODULATIONS: PSK, QAM,
technique. OFDM is a method that allows to transmit
FSK, OFDM high data rates over extremely hostile channels at a
comparable low complexity.
The class of PSK (Phase Shift Keying) modulations is The OFDM spread spectrum technique distributes the
widely used in numeric satellite TV broadcasting, data over a large number of carriers that are spread
being very robust to perturbations. The general regularly over a frequency band. This spacing
equation of PSK is given by: provides the "orthogonality" in this technique
resulting in a high spectral efficiency. Thus, the
mPSK (t ) = (t kT )cos(2f t +
k
c k ) (1) complex signal is represented by:
M 1 N 1
mOFDM (t ) = X
l =0 k =0
k e j 2f k t e j 2f c t p(t lT ) (5)
1
Facultatea de Electronic i Telecomunicaii, Departamentul Comunicaii Bd. V. Prvan Nr. 2, 300223 Timioara,
e-mail marius.salagean@etc.utt.ro
2
ENSIETA, Departement E3I2, 2 rue Francois Verny, 29806, Brest - France, e-mail ioanaco@ensieta.fr
3
ENSIETA, Departement E3I2, 2 rue Francois Verny, 29806, Brest - France, e-mail quinquis@ensieta.fr
434
Using the ml-HIM concept (relation (6)), Barbarossa
where Xk represents the k-th information symbol,
and al [3] introduced the Product HAF: the ml-HAFs
fk=f0+kf stands for the k-th subcarrier and p(t)
computed, via relation (7), for different lag sets:
represents the pulse shaping function. N and M are the
total number of subcarriers and total number of
(l ) (l )
transmitted blocks respectively.
OFDM has become very popular as it is used in such
T = K 1 ;
K 1
= { i }i = 1, K 1 (12)
l = 1, L
major applications as digital subscriber lines (DSL)
and digital audio broadcasting (DAB). are multiplied, obtaining also a more robust method
and a cross-terms free representation:
III. HIGH-ORDER TIME-FREQUENCY
METHODS K 1 ( l )
i
L
( l)
PHAF ( ; T ) = mlHAFK s ; K 1
i =1
, (13)
It is well known that there is no transformation from K 1
i
l =1 (1)
the Cohen's class that can produce the complete
i =1
concentration along the instantaneous frequency (IF)
when this one is a nonlinear function of time. Still, the effect of error propagation remains a serious
Therefore, different high order distributions have been limitation of the PHAF when we try to estimate a
developed in order to better match the non-linear deeply non-linear IF laws (underwater transitory
time-frequency behavior of the analyzed signal. On signals, digital modulations, etc). Therefore, in [5] is
the other hand, the polynomial phase signal (PPS) proposed a new procedure for polynomial order
constitutes a good model in a variety of applications, compensation, the WarpCom method, based on
e.g. radar imagery, mobile communication systems, unitary transform phase reducing [4].
etc. Let consider a signal modeled by a Kth order PPS
As it was illustrated in [1], [2], the classical HAF (relation (8)). Using a modern version of the HAF
algorithms present some limitations, related to the (PHAF operator or the approach proposed in [3]), we
noise robustness, the cross-terms presence and the obtain an accurate estimate of the Kth order
effect of the error propagation. In order to solve the polynomial coefficient, denoted by a$ K . With this
first two aspects, the multi-lag HAF (ml-HAF) value, we construct the following time axis warping
concept has been initially proposed in [2]. In fact, the function:
ml-HAF is based on the generalization of the high t
1/ K
t w( K ) = w K (t ) = (14)
wK
order instantaneous moment HIM [2]: w K : t
a K
MK[x(t); K-1] = MK1[x(t +K1); K-2]MK* 1[x(t K1); K-2] (6)
The effect of the associated unitary operator U (14) on
where i = ( 1 , 2 ,..., i ) . Applying the FFT to the PPS is depicted by:
1/ K
K
[ ]
(1), we obtain the ml-HAF of the signal x(t) :
K 1
(K ) m
(UK y )(tw(K ) ) = A~ exp jaK t
exp j am tw
a K m =0
m lH A F K [ s ; , ] = H IM K s ( t ) ; e j t d t (7)
(15)
[ ]
Assuming a PPS model for the analyzed ~ K 1 m
a
= A e xp j am t w(K ) exp j K t
signal, i.e.: m = 0 a K
14 442444 3 14 4 244 3
K SPP oforder (K 1)th residual
s ( t ) = A exp j ( t ) = A exp j ak t k (8)
k =0
1 / K 1
the main property of HIM is that, the Kth order where ~ 1 t
(16). Since all the terms
A=A
HIM is reduced to a harmonic with amplitude K a K a K
~
A2 , frequency ~ and phase :
k 2
in (16) are known and non-random, the induced
amplitude modulation can be compensated, for
M k [ y (t ); ] = A 2
K 1
( ~
exp j ~ k t + k ) (9) example, trough an amplitude weighting using the
~ = k! K 1a (10). inverse of relation (16).
where k k Therefore, the result of the warping transform of a Kth
Based on these results, Porat [1] has proposed an order PPS consists in a (K-1)th order PPS for the new
algorithm, which estimates sequentially the temporal variable t w(K ) . The (K-1)th order PHAF of
coefficients {ak}. At each step, using a spectral
analysis method, we estimate the spectral peak and, this signal, with respect to t w(K ) , peaks to a frequency
using the HAF, we compute an estimation value ( a k ) location related, via relation (10), to the aK-1
of ak. With this value, the effect of the phase term of coefficient. Once aK-1 is estimated, we construct the
the higher order is removed: (K-1)th order unitary operator UK-1:
{ }
1 /( K 1)
t (K )
s (k 1) (t ) = s (k ) (t ) exp ja k t k (11) w K -1 : t w(K ) K
w -1
( )
t w(K 1) = w K -1 t w( K ) = w
a K 1
(17)
435
V. RESULTS
which removes the (K-1)th order component. The
process is iterated (see figure 1) until all polynomial At first to test the efficiency of the proposed method,
coefficients are estimated. the signal analyzed is of type OFDM obtained by the
sum of four QPSK (Quaternary Phase Shift Keying)
y(m)(t)
PHAF-based {am }m=0, K quasi-analytic signals. The spacing of the carrier
estimation method frequencies is constant and the normalized values are
Construction of axis spread in the interval (00.5): 0.2, 0.224, 0.248,
Phase term removing
warping operator 0.273. In this case, the frequency band of interest is in
1/ m
(m) Um (m 1) = t w(m)
( ) ( ) range of 0.2 to 0.3. The orthogonal filter-banks
y m 1 Uy m tw tw considered are Chebyshev Type II filters. These filter
a
m batteries are composed of 2k, k=0, 1,2 filters. Thus,
Fig. 1. Polynomial coefficient estimation based on unitary the decomposition tree will have two decompositions
transform phase reducing
levels: first, composed by the output of the battery
with 2 filters and the second composed by the output
IV. METHOD FOR RECOGNITION OF OFDM of the battery with 4 filters.
MODULATION In this case the number of carrier frequencies and
their spectral localization is known. Thus the design
For separation of modulation OFDM the method of the filter batteries is progressive : with 2 filters and
proposed is based on the filtering process with an finally with 4 filters. The second bank of filters is
orthogonal filter-bank and a high-order time- constructed so that each filter is centered on each
frequency representation. The time-frequency space carrier. Normally, taking into account the procedure
of an OFDM signal is very complex due to numerous described above in step 5, the carriers (the minima for
carrier frequencies included in the signal. The main each branch) in the second level of decomposition of
idea is to recover and to locate these carrier the tree are retrieved. In other words, the detection of
frequencies in the frequency band of interest. the carrier frequency is accurate. This situation is
Therefore, this technique is based as follows: illustrated in figure 2. The minima are represented
1. Determination of the frequency band of with encircled stars.
interest in the time-frequency plane of the
signal using a detection method based on
cumulates of order 4 [6],
2. Construction of an orthogonal filter-bank in
order to perform a sub-bands filtering
operation in the frequency band of interest.
These filter batteries are composed of 2k,
k=0,N filters considering that each bank of
filters introduces a level of decomposition,
3. Stationarization and Extraction of the signal
using the WarpCom time-frequency
representation [5] for each filtered signal
obtained at step 2, Fig. 2. Decomposition tree for an OFDM signal with 4 carriers: 0.2,
4. Calculation of the frequency marginal and 0.224, 0.248, 0.273, in normalized frequency
the standard deviation {i,k}, where i is the
decomposition level and k is the position of Because of the possible errors that could appear in the
the filter in the filter-bank, for the normal method proposed, its possible to detect not all the 4
distribution (Pdf - Probability density carriers in the last level of the tree: only 3, 2, 1 or
function) corresponding to each stationary none, in the worse case. They could appear in the
signal obtained in step 3, upper levels.
5. Construction of a decomposition tree and Consequently, we apply this algorithm taking into
location of the minima {i,k} for each branch account the simulation method of Monte Carlo for
of the decomposition tree. N=50 signals described above. The results obtained
are presented in table 1.
Steps 1 and 2 are considered to be a filtering process Table 1
and steps 3 and 4 a time-frequency characterization 4 carriers 3 carriers 2 carriers
process. The types of the filters implied are : detected detected detected
Chebyshev Type II, Morlet, Sinc. Other types of
35 times 14 times 1 times
filters could be taken into account.
The results show good performances: from 50 times,
the proposed method have detected all the carriers 35
times in the second level of decomposition (the
situation depicted in figure 2), 14 times the algorithm
436
have picked up 3 carriers and 1 time only 2 carriers.
The efficiency of the technique is very dependent on
the construction of the batteries of filters and the band
of interest covered by these batteries. The
performances could be improved with a battery of
filters more centered on the carrier frequencies as well
as worse results are obtained with a battery of filters
that are not precisely centered on the carriers.
The next signal analyzed is of type QPSK with a
normalized frequency of 0.248. The batteries used are
the same as before In this case, the localization of the
minima in the decomposition tree has to be different. Fig. 4. Decomposition tree for a real FSK signal
Normally, on the second level of decomposition it has
to detect only one carrier. We can remark that for the FSK signal, is not needed
The results using the simulation method of Monte many levels of decompositions : the carrier can be
Carlo for N=50 such signals are presented in table 2. detected in the upper levels as shown in figure 4.
Table 2
VI. CONCLUSIONS
4 carriers 3 carriers 2 carriers 1 carrier
detected detected detected detected
In this paper, a new method for the discrimination of
0 times 0 times 17 times 27 times the OFDM modulation was proposed. The principle is
to filter the carrier frequencies in order to obtain, via
From the results obtained is evident the different WarpCom time frequency-method, the stationarized
structure of the decomposition tree, being possible to version of the signal. Furthermore, the corresponding
discriminate the OFDM modulation from others types standard deviation of the frequency is involved for the
of numeric modulation. construction of the decomposition tree. A further
Next, the signals analyzed are real communication research direction will be needed to improve and
numeric signals : an OFDM signal issued from a study the factors that affect the effectiveness of the
multi-path channel and an FSK signal. The first signal method.
has the following caracteritistics: 250 kHz sampling
frequency, 32 carrier frequencies (base carrier located REFERENCES
at 12.5 kHz), signal-to-noise (SNR) 5 db, 3201
[1] B. Porat, Digital Processing of Random Signals, Pretince
samples and 9 symbols. The algorithm generates with
Hall, 1993.
the following parameters : band of interest (1041 [2] Y. Wang and G. Zhou, On the use of high order ambiguity
kHz), 5 levels of decomposition (with the batteries functions for multicomponent polynomial phase signals, IEEE
composed of 2k, k=0, 1,2,3,4,5 filters of type Trans. on Signal Processing, vol. 65, pp. 283-296, 1998.
[3] S. Barbarossa, A. Scaglione and G.B. Giannakis, Product
Chebyshev Type II). The decomposition tree resulted High-Order Ambiguity Function for Multicomonent Polynomial-
is presented in figure 3. Phase Signal Modeling, IEEE Transactions on Signal Processing,
vol. 46, No. 3, 1998.
[4] R. Baraniuk, Unitary Equivalence: a new twist on signal
processing, IEEE Trans. on Signal Processing, vol. 43, no. 10,
October, 1995.
[5] C. Ioana, A. Quinquis, "High-Order Warped Based Ambiguity
Function", IEEE-EURASIP Workshop on Non-Linear Signal and
Image Processing, Grado, Italy, 2003.
[6] C. Ioana, A. Quinquis, "Transient signal detection using
Overcomplete Wavelet Transform and high order statistics",
International Conference on Acoustic, Speech and Signal
Processing -ICASSP 2003, Hong Kong, April, 2003.
437
438
Index of Authors
439
Cleju Nicolae, fasc. 2, p. 354;
Codreanu N.C., fasc. 1, p. 98;
Coma Ciprian, fasc. 1, p. 117, fasc. 2, p. 382, fasc. 2, p.385;
Coma Ciprian-Romeo, fasc. 1, p.178;
Constantinescu Mihai, fasc. 1, p. 399;
Copaciu Flavius, fasc. 1, p. 388;
Cosac Andreea, fasc. 2, p. 205;
Costin Mihaela, fasc. 1, p. 291, fasc. 1, p.328;
Cotae Paul, fasc. 2, p. 385;
Crciun Florin, fasc. 2, p. 403;
Crainic Monica Sabina, fasc. 2, p. 201;
Crciun Adrian Virgil, fasc. 1, p. 40, fasc. 1, p.131;
Cremene Marcel, fasc. 2, p. 323;
Cre Clin, fasc. 1, p. 20;
Criste Daniel, fasc. 2, p. 175;
Croitoru Victor, fasc. 1, p. 399;
Dvid Lszl, fasc. 1, p. 209;
Dnei Marllene, fasc. 2, p. 346;
De Sabata Aldo, fasc. 2, p. 317;
Demeter Stefan, fasc. 2, p. 431;
Dimitriu Bogdan, fasc. 1, p. 159, fasc. 1, p.173;
Dobrea Dan-Marius, fasc. 2, p. 354;
Dobrot Virgil, fasc. 1, p. 367, fasc. 1, p.383, fasc. 1, p. 388;
Dogaru Ulieru Cristina, fasc. 2, p. 221;
Dogaru Ulieru Valentin, fasc. 2, p. 221;
Dragomir Ioan-Virgil, fasc. 1, p. 367;
Dragomir Livia, fasc. 2, p. 209;
Dughir Ciprian, fasc. 2, p. 348;
Duma Ioan I. , fasc. 2, p. 398;
Duma Petrut, fasc. 2, p. 55, fasc. 2, p. 154, fasc. 2, p. 187;
Dumitrescu Ctlin I. , fasc. 2, p. 247;
Emil Teodoru, fasc. 2, p. 300;
Ene Daniel Victora, fasc. 2, p. 193;
Enescu Andrei Alexandru, fasc. 2, p. 69, fasc. 2, p. 165, fasc. 2, p. 360;
Faur Daniela, fasc. 2, p. 205;
Fazakas Albert, fasc. 1, p. 136, fasc. 1, p. 168;
Frca Cristian, fasc. 1, p. 20;
Fericean Gabriel G. , fasc. 2, p. 50;
Fetil Lelia, fasc. 1, p. 136, fasc. 1, p. 168;
Feteanu Gigi, fasc. 1, p. 249;
Florea M., fasc. 1, p. 10, fasc. 1, p. 15;
Fratu Octavian, fasc. 2, p. 19, fasc. 2, p. 403;
Gal Jnos, fasc. 1, p. 164, fasc. 2, p. 87;
Galatchi Dan, fasc. 1, p. 358, fasc. 1, p. 362;
Gasparel Aida-Virginia, fasc. 1, p. 367;
Gavrincea tefan, fasc. 1, p. 184;
Gheu Drago, fasc. 2, p. 285;
Ghisa Laura, fasc. 2, p. 294, fasc. 2, p. 297;
Giurgiu Luminia, fasc. 1, p. 249;
Gvan M. , fasc. 2, p. 341;
Gontean Aurel, fasc. 2, p. 180, fasc. 2, p. 184;
Goras Tecla., fasc. 1, p. 7, fasc. 1, p. 102;
Gora Liviu, fasc. 1, p. 102;
Grama Lcrimioara, fasc. 2, p. 96;
Grama Gheorghe, fasc. 1, p. 34;
Groza Robert, fasc. 1, p. 136, fasc. 1, p. 168;
440
Gui Vasile, fasc. 1, p. 321;
Halunga Simona, fasc. 2, p. 19, fasc. 2, p. 403;
Hintea Sorin, fasc. 1, p. 136;
Hurgoi Florin, fasc. 1, p. 20;
Husu Adela, fasc. 2, p. 221;
Ignea Alimpie, fasc. 2, p. 236, fasc. 2, p. 255;
Ilioaei Ciprian, fasc. 2, p. 138;
Ioana Cornel, fasc. 2, p. 434;
Ionacu Cristian, fasc. 1, p. 159, fasc. 1, p. 173, fasc. 1, p. 178;
Ionel Sabin C. , fasc. 2, p. 240, fasc. 2, p. 346;
Ionescu C., fasc. 1, p. 98;
Ionica Tiberiu, fasc. 2, p. 197;
Iosif Florin D., fasc. 1, p. 333;
Isar Alexandru, fasc. 2, p. 110;
Isar Dorina, fasc. 2, p. 110;
Ivanovici Traian, fasc. 2, p. 221;
Jibleanu Roxana, fasc. 2, p. 180, fasc. 2, p. 184;
Jipa Rzvan, fasc. 1, p. 147;
Jurca Lucian, fasc. 1, p. 193;
Keller Guenter, fasc. 1, p. 63, fasc. 1, p. 85;
Kovaci Maria, fasc. 2, p. 33, fasc. 2, p. 38;
Krneta Radojka, fasc. 1, p. 47;
Laitinen Jyrki, fasc. 1, p. 321;
Lascu Dan, fasc. 1, p. 63, fasc. 1, p. 85;
Lascu Mihaela, fasc. 2, p. 253;
Lazar Gabriel, fasc. 1, p. 383;
Lazr A., fasc. 1, p. 10, fasc. 1, p. 15;
Lazr Luminia-Camelia, fasc. 1, p. 10, fasc. 1, p. 15;
Lzrescu Dan, fasc. 1, p. 253;
Lzrescu Vasile, fasc. 1, p. 253;
Levinthal Adam, fasc. 1, p. 147;
Loghin Clin, fasc. 2, p. 323;
Lojewski George, fasc. 2, p. 285;
Luca Adrian, fasc. 2, p. 132;
Lucaciu Radu, fasc. 2, p. 394;
Lungu erban, fasc. 1, p. 73, fasc. 1, p. 153;
Lupu Eugen, fasc. 1, p. 275, fasc. 1, p. 279;
Mailat Adrian, fasc. 1, p. 131;
Malutan Raul E., fasc. 1, p. 337;
Marchegay Philippe, fasc. 1, p. 142;
Marcus Ionel Urdea, fasc. 2, p. 217;
Mrza Eugen, fasc. 2, p. 389, fasc. 2, p. 425;
Mateescu Cosmin, fasc. 2, p. 403;
Matekovits Ladislau, fasc. 2, p. 317;
Matin Mohammad Abdul, fasc. 2, p. 114;
Mrnescu Valentin, fasc. 1, p. 189, fasc. 1, p. 193;
Mereu erban, fasc. 2, p. 244;
Mic Daniel, fasc. 1, p. 184, fasc. 1, p. 236;
Miclaus Simona, fasc. 2, p. 431;
Miclu Nicolae, fasc. 1, p. 226, fasc. 1, p. 256;
Micul Emil, fasc. 1, p. 184;
Mihaescu Adrian, fasc. 2, p. 294, fasc. 2, p. 297;
Mihiu Adrian, fasc. 2, p. 255;
Milenkovic Sanja, fasc. 1, p. 47;
Militaru Nicolae, fasc. 2, p. 285;
Mischie Septimiu, fasc. 2, p. 230;
441
Mitran Radu, fasc. 2, p. 431;
Moca Vasile V., fasc. 1, p. 279;
Mocanu Irina G. , fasc. 2, p. 142, fasc. 2, p. 146;
Moise Marius, fasc. 2, p. 409, fasc. 2, p. 415;
Moldoveanu Alexandru, fasc. 1, p. 399;
Moraru Bogdan, fasc. 1, p. 388;
Moraru Simona, fasc. 2, p. 205;
Moussaoui A., fasc. 1, p. 199;
Munteanu Doru P., fasc. 1, p. 301;
Munteanu Valeriu, fasc. 2, p. 106;
Naforni Corina, fasc. 1, p. 164, fasc. 2, p. 126;
Naforni Miranda, fasc. 1, p. 372, fasc. 1, p. 377, fasc. 2, p. 389;
Nstase Ana, fasc. 2, p. 13;
Neagoe Victor-Emil, fasc. 1, p. 343, fasc. 1, p. 348;
Neddermeyer Werner, fasc. 2, p. 236;
Nedelcu Liviu, fasc. 2, p. 285;
Negoiescu Dan, fasc. 1, p. 79;
Nicolaescu Ioan, fasc. 2, p. 150;
Nicula Dan, fasc. 1, p. 147;
Oancea Eugeniu, fasc. 1, p. 301;
Obreja erban Georgic, fasc. 2, p. 370, fasc. 2, p. 378;
Olah A., fasc. 2, p.341;
Oltean Gabriel, fasc. 1, p. 220, fasc. 2, p. 328;
Oltean Ioana, fasc. 2, p. 328;
Oltean Marius, fasc. 2, p. 389, fasc. 2, p. 425;
Oniga tefan, fasc. 1, p. 184, fasc. 1, p. 232, fasc. 1, p. 236;
Orza B., fasc. 1, p. 311, fasc. 2, p. 341;
Osborne Dominic, fasc. 2, p. 120;
Oteteanu Marius, fasc. 1, p. 393, fasc. 2, p. 175;
Palade Tudor, fasc. 1, p. 352;
Paleologu Constantin, fasc. 2, p. 165, fasc. 2, p. 69;
Panaitopol Dorin, fasc. 2, p. 170;
Pan Gheorghe, fasc. 1, p. 40, 131;
Pantelimon Brandusa, fasc. 2, p. 205, fasc. 2, p. 209, fasc. 2, p. 213;
Partheniu Cezar, fasc. 2, p. 75;
Paca Sorin, fasc. 1, p. 53;
Pun Adrian Florin, fasc. 2, p. 370, fasc. 2, p. 378;
Perioar Lucian Andrei, fasc. 2, p. 28, fasc. 2, p. 60;
Petrescu Teodor, fasc. 2, p. 259, fasc. 2, p.285;
Petreu Dorin, fasc. 1, p. 20;
Pirnog Ionu, fasc. 2, p. 193;
Pletea I.V., fasc. 1, p. 7, 23;
Polgar Zsolt, fasc. 2, p. 7, fasc. 2, p. 13;
Pop Ovidiu Aurel, fasc. 1, p. 73;
Pop Petre G., fasc. 1, p. 275, fasc. 1, p. 279;
Popa Cosmin, fasc. 1, p. 122, fasc. 1, p. 126;
Popa Gheorghe Daniel, fasc. 1, p. 393;
Popa Rzvan, fasc. 2, p. 25;
Popa Rustem, fasc. 2, p. 335;
Popescu Camelia, fasc. 1, p. 98, fasc. 2, p. 350;
Popescu Sorin, fasc. 2, p. 44;
Popescu Victor, fasc. 1, p. 153;
Popescu Viorel, fasc. 1, p. 53, fasc. 1, p. 59, fasc. 1, p. 69, fasc. 1, p. 85;
Popescu Vladimir, fasc. 1, p. 305;
Popovici Adrian, fasc. 1, p. 59, fasc. 1, p. 89;
Pradel Gilbert, fasc. 1, p. 241;
442
Preda Radu O., fasc. 1, p. 315;
Puschita Emanuel, fasc. 1, p. 352;
Quinquis Andr, fasc. 2, p. 110, fasc. 2, p. 434;
Rados Ivan, fasc. 2, p. 289;
Rangu Marius, fasc. 1, p. 94;
Rp Adrian, fasc. 1, p. 98, fasc. 2, p. 350;
Rdescu Radu, fasc. 2, p. 25, fasc. 2, p. 65;
Rogers Derek, fasc. 2, p. 120;
Romanca Mihai, fasc. 1, p. 131;
Ropot Armand-Dragos, fasc. 1, p. 343;
Rosu-Niculescu Ana, fasc. 2, p. 259;
Rusu Mircea Sorin, fasc. 2, p. 170;
Sadi Francois, fasc. 1, p. 241;
Salagean Marius, fasc. 2, p. 434;
Sandu Ion, fasc. 2, p. 209;
Schnell Michael, fasc. 2, p. 236;
Scripcariu Luminita, fasc. 2, p. 55, fasc. 2, p. 187;
Sechelea Andrei-Tudor, fasc. 2, p. 354;
Serafin Petru, fasc. 1, p. 407, fasc. 1, p. 412;
Shu Fangwu, fasc. 2, p. 236;
Sikora Axel, fasc. 2, p. 421;
Silion Daniel, fasc. 2, p. 170;
Simion Stefan, fasc. 2, p. 265;
Simion Viorica, fasc. 2, p. 213;
Simon Csaba, fasc. 1, p. 372, fasc. 1, p. 377;
Simu Dan, fasc. 1, p. 44;
Sipos Emilia, fasc. 2, p. 328;
Sirbu Gabriel, fasc. 2, p. 309, fasc. 2, p. 313;
Srbu A., fasc. 1, p. 7;
Sorell Matthew, fasc. 2, p. 120;
Stanciu Lucian, fasc. 2, p. 102;
Stefanescu Costin, fasc. 2, p. 213;
Stoian Rodica, fasc. 2, p. 28, fasc. 2, p. 60;
Stoica Adrian, fasc. 1, p. 333, fasc. 2, p. 150;
Stoiciu Dan, fasc. 2, p. 253, fasc. 2, p. 348;
Sucholotiuc Cristina, fasc. 2, p. 193;
Sunaric Tanja, fasc. 2, p. 289;
Svasta Paul, fasc. 1, p. 98, fasc. 2, p. 350;
Szkely Iuliu, fasc. 1, p. 214;
Szopos Erwin, fasc. 2, p. 81;
chiop Adrian, fasc. 1, p. 69;
elaru Clin, fasc. 1, p. 352;
erbanescu Alexandru, fasc. 2, p. 366;
tefan Demeter, fasc. 1, p. 249, fasc. 2, p. 300;
Trniceriu Daniela, fasc. 2, p. 106;
Telescu Mihai, fasc. 2, p. 294, fasc. 2, p. 297;
Teodorescu T.D., fasc. 1, p. 266;
Teodorescu Tiberiu, fasc. 1, p. 112;
Tigaeru Liviu, fasc. 1, p. 29;
Tiponut Sebastian V. , fasc. 2, p. 332;
Tiponu Virgil, fasc. 1, p. 236, fasc. 2, p. 305;
Toadere Florin, fasc. 2, p. 275, fasc. 2, p. 271, fasc. 2, p. 280;
Toderean Gavril, fasc. 1, p. 262;
Tofilescu Pompilian, fasc. 1, p. 34;
Toma Liviu Jr. , fasc. 2, p. 236;
Toma Norbert, fasc. 2, p. 81;
443
Tome Marin, fasc. 1, p. 53;
Tordai Botond, fasc. 1, p. 377;
Trip Nistor Daniel, fasc. 1, p. 69;
Tudos Ana Maria, fasc. 1, p. 328;
Tulbure Traian, fasc. 1, p. 147;
Turalija Pero, fasc. 2, p. 289;
igeru Liviu, fasc. 1, p. 205;
opa Marina, fasc. 2, p. 81;
Udrea Radu Mihnea, fasc. 1, p. 297, fasc. 1, p. 315, fasc. 2, p. 193;
Ungureanu Mihaela, fasc. 1, p. 253;
Ursaru Ovidiu, fasc. 1, p. 205, fasc. 1, p. 29;
Varga Mihaly, fasc. 2, p. 7;
Vesa Andy, fasc. 2, p. 425;
Vilda Pedro Gmez, fasc. 1, p. 337;
Vizireanu Drago Nicolae, fasc. 1, p. 297, fasc. 1, p. 315;
Vizitiu Constantin I. , fasc. 1, p. 333 , fasc. 2, p. 150;
Vlad A., fasc. 2, p. 341;
Vlad Adriana, fasc. 2, p. 132;
Vladescu Clin, fasc. 2, p. 366;
Vlaicu Aurel, fasc. 1, p. 311, fasc. 2, p. 341;
Vldeanu Clin, fasc. 2, p. 394;
Voina Catalin, fasc. 2, p. 205;
Zinca Daniel, fasc. 1, p. 367, fasc. 1, p. 383;
Zoican Roxana, fasc. 1, p. 358, fasc. 1, p. 362;
Zoican Sorin, fasc. 2, p. 160.
444
SPONSORS: