PROCEEDINGS OF THE 4th INTERNATIONAL CONFERENCE ON CROWNCOM 2009

Implementation of Cyclostationary Feature Detector

for Cognitive Radios
Vesa Turunen∗ , Marko Kosunen∗ , Anu Huttunen† , Sami Kallioinen‡ ,
Petri Ikonen‡ , Aarno Pärssinen‡ and Jussi Ryynänen∗
∗ HelsinkiUniversity of Technology
Department of Micro- and Nanosciences,
SMARAD-2, P.0 Box 3000, FIN-02015, Finland
† Nokia Research Center, Helsinki, Finland
‡ Nokia Research Center, Otaniemi Lablet, Espoo, Finland

Abstract— Spectrum sensing is needed in cognitive radios to
provide information about the surrounding radio spectrum. This
enables cognitive radio system to communicate among existing
radio systems without interfering them. This paper describes
an FPGA implementation of a cyclostationary feature detector,
which has an improved detection performance achieved by
decimation of the cyclic spectrum. Decimation also provides a
simple way to control detection time and, thus, allows trading the
detection time to better probability of detection and vice versa.
Measured detection performance is presented and detection of a
802.11g WLAN signal from air is demonstrated.
I. I NTRODUCTION
As wireless communications systems evolve, demand for
spectral resources is continuously growing. However, tradi-
tional frequency allocation policy, practiced all over the world,
has resulted in situation where unallocated spectrum bands are
running short, while measurements have shown that spectrum
utilization in already allocated bands is usually low.
Cognitive radios (CR) [1] [2] provide a solution for taking
advantage of the underutilized spectral resources and have
been a popular research topic for several years. Cognitive
radio’s capability to recognize the surrounding radio environ-
ment and operate accordingly (i.e. change operation frequency,
modulation etc.) permits operation among existing commu-
nication systems without interfering the primary users. This
enables major increase in spectrum utilization.
To produce awareness of the surrounding radio spectrum,
cognitive radio device needs to incorporate a spectrum sensing
unit, which is able to sense spectral opportunities reliably and
at very low signal-to-noise ratios (SNR). Also, appearance
of the primary user must be detected in reasonable time to
minimize interference produced by the secondary network to
the primary system.
Spectrum sensing can be implemented for example with
energy detectors or feature detectors such as cyclostationary
based detectors. Energy detectors are very simple to imple-
ment, but their performance degrades when noise levels are
unknown and they are also incapable to differentiate between
signals from different systems [3]. Therefore, energy detectors
are best suited for fast and coarse scanning of the spectrum.
Feature detectors, in general, can operate reliably at very low
978-1-4244-3424-4/09/$25.00 ©2009 IEEE
SNR levels and can also differentiate certain signal types from
others.
Much of the recent work is concentrated on IEEE802.22 [4]
[5], which is the first standard based on cognitive radio
technology. It defines a radio interface for Wireless Regional
Access Network (WRAN) that operates at the frequency bands
currently mainly occupied by digital TV broadcast services. In
IEEE802.22 networks, sensing task is somewhat simplified due
to two facts. Firstly, primary signals that must be detected in-
clude only DTV broadcasts and Part 74 (wireless microphones
etc.) transmissions and, thus, purpose-built sensing algorithms
can be used. Secondly, locations of both base stations and
customer equipment are fixed, which relaxes implementation
constraints. Approaches for spectrum sensing in IEEE802.22
are presented in [6] [7] [8] [9].
A more general approach is taken in this work. The
detection algorithm, presented in Section II, is based on
cyclostationarity that the received signal inherits for example
from modulation, cyclic prefixes or spreading codes. Since the
parameters are system dependent, cyclostationarity can be used
to identify the signal. Detectors based on cyclostationarity has
been presented in [10] [11] [12]. In the early stage of the
research, effort is put on identifying OFDM-based systems
(e.g. WLAN, DVB-T, LTE) only, but the concept is readily
extendable to other signal types as well [13].
In general, detector performance is characterized by two
metrics: probability of detection and false alarm rate. They
are both equally important, since low probability of detection
increases the amount of interference inflicted on the primary
users, whereas high false alarm rate increases the amount of
missed spectral opportunities in the secondary network. Other
important parameters are the detection time resolution and
bandwidth, and, of course, power consumption and area of
the implementation.
This paper is organized as follows: Section II briefly
presents the implemented detection algorithm and shows
how the detector performance can be further improved by
decimation. Section III explains the FPGA implementation
and discusses implementation related issues. Simulation and
measurement results are presented in Section IV. Finally, a
conclusion is given in Section V.
PROCEEDINGS OF THE 4th INTERNATIONAL CONFERENCE ON CROWNCOM 2009

II. A LGORITHM 1

A process x(t) is second-order cyclostationary if its mean

0.8
and autocorrelation are periodic in time [14]. Thus, for a cyclo-
stationary process, the cyclic autocorrelation function (CAF)
is nonzero for a set of cyclic frequencies α = 0. Here, we
0.6

concentrate on signals that exhibit conjugate cyclostationarity 0.4

such as OFDM signals. The conjugate cyclic autocorrelation
function at cyclic frequency α can be estimated as 0.2


N −1
1
R̂xα = x(n)x∗ (n − τ )e−j2παn/N = Rxα + (α), (1) 0
N
0 2 4 6 8 10 12 14 16 18
Cyclic frequency (MHz)
n=0

in which (α) is the estimation error. Here, τ is lag parameter Fig. 1. Example cyclic spectrum of OFDM signal sampled at 20MHz.
Number of subcarriers is 52 and length of cyclic prefix is 12.
in the autocorrelation. In practice, values of the CAF are
seldom exactly zero and decision has to be made whether the
value presents a zero or not.
If the cyclic autocorrelation does not exist, Rxα = 0 and 1

R̂x = (α), which is asymptotically normal zero mean

α
0.8
complex random variable
R̂xα = (α) = X(α) + jY (α). (2) 0.6

X(α) and Y (α) are normal distributed zero mean random 0.4

variables. For vector of zero mean random variables, an

estimate of the covariance matrix can be computed as 0.2

 
E[X 2 ] E[XY ]
Σ̂2c = , (3) 0
E[XY ] E[Y 2 ]
0 0.5 1 1.5 2
Cyclic frequency (MHz)

where elements of the matrix are computed as Fig. 2. Simulation in Fig. 1 modified by applying decimation by ratio 8.
N −1
1 
E[X 2 ] = {R̂xαk }2 (4)
N
k=0
−1
where Fχ22 is the cumulative distribution function of χ22 -
1 
N
distribution and p is the false alarm rate. The test can be
E[Y 2 ] = {R̂xαk }2 (5)
N modified to include multiple lag values [10] or cyclic frequen-
k=0
−1 cies [11]. An alternative approach to estimate the covariance
1 
N
E[XY ] = {R̂xαk }{R̂xαk }. (6) matrix is presented in [10].
N
k=0 In the derivation of the algorithm, samples of the process
The error introduced to expectation by cyclic frequency com- x(t) are assumed to be well separated in time and thus
ponent Rxα , if it exists, is not significant in critical cases (low approximately independent. In practice, however, presence of
SNR), and converges to zero with large N . any narrowband interferer (relative to detection bandwidth) vi-
In order to find out if there exists cyclic components in olates this assumption and can degrade detection performance.
R̂xα , a hypothesis test is developed by following the guidelines Often the case is that the information in the cyclic spectrum
presented in [10]. Hypotheses are resides on low cyclic frequencies. Therefore, the high end of
the cyclic spectrum is of no interest and the autocorrelation
H0 : ∀α ∈ A → R̂xα = (α) (7)
product can be resampled at lower frequency before calcu-
H1 : f or some α ∈ A → R̂xα = Rxα + (α), (8) lating the DFT. This is illustrated in Fig. 1, where a cyclic
where set A contains all cyclic frequencies for a fixed value spectrum of an OFDM signal is simulated without decimation.
of τ , which are assumed to be known a priori. Under null In Fig. 2 the simulation is repeated and decimation by 8 is
hypothesis, test statistic applied.
Decimation increases the detection time by a factor cor-
T = R̂xα Σ̂−1
2c (R̂x )
α T
(9) responding the decimation ratio, and therefore improves the
is χ22 -distributed and the following constant false alarm rate probability of detection. This is because a longer signal can be
test for presence of cyclostationarity is derived: processed with a fixed-length FFT. Maximum decimation ratio
depends on cyclic frequencies of the signal under detection,
Fχ22 (T ) > 1 − p, (10) or could also be limited by detection time constraints.
PROCEEDINGS OF THE 4th INTERNATIONAL CONFERENCE ON CROWNCOM 2009
III. I MPLEMENTATION
The cyclostationary feature detection algorithm was imple-
mented based on FFT. This approach was selected to support
later incorporation of an energy detector, which requires the
FFT and, thus, can be implemented with small additional
hardware. It is obvious, however, that implementing just the
feature detector would require less hardware if the FFT was
omitted and estimation of the covariance matrix was done in
the time-domain.
The implementation is shown in Fig.3. First, complex input
signal, sampled at 20 MHz, is multiplied with delayed version
of itself. This requires four multipliers and a dual-port random
access memory (RAM) to implement the lag. Size of the
memory is determined by expected maximum lag value and
was left as an open synthesis parameter.
The autocorrelation product is then resampled using variable
rate decimator. The decimator consists of a CIC-stage [15] and
three decimate-by-2 stages that were implemented using FIR
filters. This combination supports decimation ratios M=1,2,4,8
and 16.
Following the decimator is the FFT unit. 2048-point radix-
22 DIF FFT algorithm [16] was selected because of it’s
relatively low complexity, small area and power consumption.
It is an pipelined architecture that is suitable for real-time
high-speed applications.
Finally, elements of the covariance matrix are calculated
by three multiply-and-accumulate blocks. Normally this type
of FFT implementation requires a reordering memory block,
since the pipeline output samples appear in bit-wise reversed
order. In this application reordering is not necessary, since
the following summation can be done in arbitrary order and
the control logic can be designed to pick up the right cyclic
bin from the FFT output. Since calculation of the final test
statistic includes a matrix inversion, rest of the operations are
performed with an external microprocessor.
The FFT is the major area and power consuming unit in
this application and also dictates the minimum detection time.
Detection time, that is the time interval over which test statistic
is calculated from input samples, can be expressed as
M ∗ NF F T
Td = , (11)
fs,in
where fs,in is the input sampling frequency, M is the deci-
mation ratio and NF F T is 2048. By adjusting the decimation
ratio before the FFT, better detection performance is obtained
at the cost of increased detection time. For example, when
input is sampled at 20 MHz, Td , ranges from 100 us (M = 1)
to 1.6 ms (M = 16). In addition, using higher decimation ratio
also decreases power consumption, since most of the logic is
then sampled at lower clock frequency and power consumption
of CMOS logic is approximately directly proportional to the
sampling frequency.
Detection algorithm was implemented on an FPGA that
was accompanied by a commercial WLAN RF receiver for
measurement purposes. FPGA was connected to a laptop via
USB, where Matlab was used to control the measurements
M=1,2,4,8
^ a}
or 16 Re{R x
Re{x} E[X2 ]*N
RAM
2048− E[XY]*N
FFT
RAM
Im{x} E[Y2 ]*N
−1 ^ a}
Im{R x
Fig. 3. Detector’s block diagram.
WLAN 10−bit Digital Feature
receiver ADC front−end detector
USB MATLAB
Control
PCB FPGA
Fig. 4. Measurement setup.
and analyze the results. The measurement setup is illustrated
in Fig. 4. The digital front-end consists of simple baseband
filters and a DC rejection filter.
IV. S IMULATIONS AND M EASUREMENTS
Probability of detection of the cyclostationary feature de-
tector was simulated in Matlab as a function of SNR for
various decimation ratios and results are shown in Fig. 5. The
signal under detection utilizes orthogonal frequency division
multiplex (OFDM) modulation with 52 subcarriers, the length
of the cyclic prefix is 12 samples and subcarrier modulation
employed is 16-QAM. Thus, cyclostationarity is observed
at cyclic frequency α = fs /64 when lag is 52 samples.
Probability of false alarm is set to 0.05 and noise is AWGN.
Increasing the decimation ratio clearly improves the detection
performance without having to increase size of the FFT.
Implementation performance was measured using the same
OFDM signal. First, a vector signal generator was used
to modulate the baseband signal to WLAN channel 6
(2.437 GHz). Then, probability of detection was measured
from 20000 consecutive detections as a function of input
power to the RF receiver for various decimation ratios. The
measured curves (Fig. 6) match well with the simulation
results, except when the signal is no longer detected, the
probability of detection tends to zero instead of the theoretical
false alarm rate. This is explained by the distribution of the
noise, which is not AWGN in practice. Especially, at very low
RF input power levels, the noise is dominated by interference
that is coupled to the analog baseband.
Finally, detection of 802.11g WLAN signal from air was
demonstrated. In this case, WLAN traffic was generated with
a laptop, which was located in the same room with the
detection equipment. 802.11g OFDM-signal has FFT/IFFT
period (TF F T ) of 3.2 μs, which corresponds to 64 samples
at 20 MHz sampling frequency. Thus, peaks at the cyclic
spectrum are found by setting the lag parameter to 64. The
signal exhibits cyclostationary with cycle frequency α =
(TF F T + TCP )−1 = 0.25 M Hz, where TCP = 0.8 μs is
length of the cyclic prefix. This can be observed from the
PROCEEDINGS OF THE 4th INTERNATIONAL CONFERENCE ON CROWNCOM 2009
100 1
M=1
90 M=2
M=4
80 M=8
M=16
Probability of Detection (%)
70 0
0 50 100 150 200 250 300 350 400
time (ms)
60

50
Fig. 8. Detections of 802.11g WLAN signal over 400ms period. Black
40 color indicates time instances when signal is detected. Decimation ratio is 8,
resulting in time resolution of 0.82 milliseconds.
30

20

10
of the detection. Furthermore, decimation introduces a simple
0
−22 −20 −18 −16 −14 −12 −10
SNR (dB)
−8 −6 −4 −2 0 way to dynamically adjust the detection time for proper
detection performance and power consumption. The algorithm
Fig. 5. Simulated probability of detection as a function of SNR for multiple has been implemented on a FPGA and detection of 802.11g
decimation ratios (M). False alarm rate is set to 5%. WLAN signal from air has been demonstrated.
100
M=1 R EFERENCES
90 M=2
M=4
M=8
[1] J. Mitola III, “Cognitive radio: An integrated agent architecture for soft-
80
M=16 ware defined radio,” Ph.D. dissertation, Royal Institute of Technology,
May 2000.
Probability of Detection (%)

70
[2] I. F. Akyildiz, W.-Y. Lee, M. C. Vuran, and S. Mohanty, “Next
60
generation/dynamic spectrum access/cognitive radio wireless networks:
50 a survey,” Computer Networks, vol. 50, no. 13, pp. 2127–2159, 2006.
40 [3] F. F. Digham, M.-S. Alouini, and M. K. Simon, “On the energy detection
of unknown signals over fading channels,” "IEEE Trans. Commun.",
30
vol. 55, no. 1, pp. 21–24, 2007.
20 [4] IEEE 802.22 Working Group on Wireless Regional Area Networks,
10
“http://www.ieee802.org/22/.”
[5] C. Stevenson, G. Chouinard, Z. Lei, W. Hu, S. Shellhammer, and
0
−110 −105 −100 −95 −90 −85 −80 W. Caldwell, “Ieee 802.22: The first cognitive radio wireless regional
Input power (dBm) area network standard,” "IEEE Commun. Mag.", vol. 47, no. 1, pp. 130–
138, 2009.
Fig. 6. Measured probability of detection as a function of input power to [6] A. Sahai, S. Mishra, R. Tandra, and K. Woyach, “Cognitive radios for
the RF receiver. Results are averaged over 20000 detections. spectrum sharing,” "IEEE Signal Processing Mag.", vol. 26, no. 1, pp.
140–145, 2009.
[7] H.-S. Chen, W. Gao, and D. Daut, “Signature based spectrum sensing
algorithms for IEEE 802.22 WRAN,” Communications, 2007. ICC ’07.
measured cyclic spectrum that is presented in Fig. 7. In the IEEE International Conference on, pp. 6487–6492, June 2007.
measurement, detector’s decimation ratio was set to 8, thus [8] Y. Karmalita and H. Leib, “Statistical detection of NTSC signals in
detection time is 0.82ms. Detections over 400ms period are an IEEE 802.22 environment for dynamic spectrum access,” Signals,
Systems and Electronics, 2007. ISSSE ’07. International Symposium on,
presented in Fig. 8, where black lines corresponds to time pp. 63–66, 30 2007-Aug. 2 2007.
instances where WLAN signal was detected. Fig. 8 clearly [9] G. Zheng, N. Han, X. Huang, S. H. Sohn, and J. M. Kim, “Enhanced
shows quiet periods of tens of milliseconds where the spectrum energy detector for IEEE 802.22 WRAN systems using maximal-to-
mean power ratio,” Wireless Communication Systems, 2007. ISWCS
is not occupied by the WLAN signal. 2007. 4th International Symposium on, pp. 370–374, Oct. 2007.
[10] A. Dandawate and G. Giannakis, “Statistical tests for presence of
V. C ONCLUSION cyclostationarity,” "IEEE Trans. Signal Processing", vol. 42, no. 9, pp.
In this paper, an implementation of a cyclostationary feature 2355–2369, 1994.
[11] J. Lunden, V. Koivunen, A. Huttunen, and H. V. Poor, “Spectrum
detector has been described. It has been shown how decimation sensing in cognitive radios based on multiple cyclic frequencies,” in
of the correlation product can be used to improve performance Proc. Int. Conf. on Cognitive Radio Oriented Wireless Networks and
Communications, 2007, pp. 37–43.
[12] A. Tkachenko, A. Cabric, and R. Brodersen, “Cyclostationary feature
0.5 detector experiments using reconfigurable bee2,” in New Frontiers in
0.45 Dynamic Spectrum Access Networks, 2007. DySPAN 2007. 2nd IEEE
0.4 International Symposium on, 2007, pp. 216–219.
0.35
[13] M. Öner and F. Jondral, “Air interface identification for software radio
0.3
systems,” AEU - International Journal of Electronics and Communica-
Amplitude

0.25
tions, vol. 61, no. 2, pp. 104–117, Feb. 2007.
[14] W. A. Gardner, A. Napolitano, and L. Paura, “Cyclostationarity: Half
0.2
a century of research,” Signal Processing, vol. 86, no. 4, pp. 639–697,
0.15
Apr. 2006.
0.1
[15] E. Hogenauer, “An economical class of digital filters for decimation
0.05 and interpolation,” "IEEE Trans. Acoust., Speech, Signal Processing",
0
0 0.5 1 1.5 2
vol. 29, no. 2, pp. 155–162, 1981.
Cyclic frequency (MHz) [16] S. He and M. Torkelson, “A new approach to pipeline fft processor,” in
Proc. Int. Symp. Parallel Processing, 1996, pp. 766–770.
Fig. 7. Cyclic spectrum of 802.11g WLAN signal measured from air.
Detector’s decimation ratio is 8.