A Hybrid Cooperative Spectrum Sensing Technique

for Cognitive Radio Networks Using Linear

Classifiers
Mohamed Gaafar, Yasmine Hassan and Tallal Elshabrawy
Communications Department, German University in Cairo
Cairo, Egypt
Email: muhamed.gaafar@gmail.com

Abstract-In this paper, a hybrid cooperative spec

trum sensing technique

HCSST

between energy detection

feature detection

(CFD)

( ED)

is proposed merging
and cyclostationary

such that both low compu

tational complexity as well as the superior detection

performance are achieved. In HCSST, individual Cog

nitive Radio CR

nodes decide independently on using

ED or CFD as their spectrum sensing technique based

on the received signal-to-noise ratio at each cognitive
end. Cooperative decision about spectrum availability
is made using a trained linear classifier
fusion center

( FC ) .

( LC )

at the

Results have shown that HCSST

provides superior performance over ED and a slightly

degraded performance compared to CFD. Further, the
computational complexity is reduced noticeably at high
SNR regimes.

Index Terms-Cognitive Radio, Cooperative Spec

trum Sensing, E nergy Detection, Cyclostationary fea
ture detection, Linear Classifier

I. INTRODUCTION
Cognitive Radio (CR) is a promising technology to solve
the spectrum scarcity and the inefficiency in spectrum
usage problems. The key characteristic of CR systems is
sensing the electromagnetic environment to adapt their
operational parameters to the dynamic radio environment.
CR networks improve the spectrum utilization by allowing
unlicensed secondary users (SUs) to opportunistically use
frequency bands not utilized by licensed primary users
(PUs). SUs in CR networks are restrained by causing no
interference to PUs. Hence, they need to employ effective
and efficient spectrum sensing techniques that ensure the
Quality of Service (QoS) for PUs and exploit all dynamic
spectrum opportunities [1].
Spectrum sensing is one of the main challenges that
researchers face in the dynamic opportunistic spectrum
access. It is responsible for providing efficient and fair
spectrum access among licensed and unlicensed users
and mitigating interference with PUs. Spectrum sens
ing techniques can be divided into two main categories:
parametric and non-parametric. In parametric spectrum
sensing, a prior information about the transmitted signal
should be known to SUs such as operating frequency,
bandwidth, modulation type, etc. On the other hand,
in non-parametric methods, a prior knowledge about the

978-1-4799-5807-8/14/$31.00 @2014

transmitted signal does not have to be available at the

CR node. The performance of the parametric algorithms
is usually better than of the non-parametric techniques.
However, if non-accurate information is used in parametric
algorithms, the detection performance degrades signifi
cantly. In this paper, a sensing scheme merging between a
non-parametric technique which is Energy Detection (ED)
and a parametric one which is Cyclostationary Feature
Detection (CFD) is proposed. The overall objective of
any spectrum sensing scheme is to minimize the false
alarm probability Pj (no PU is transmitting but the CR
decides that the spectrum is not idle) while maximizing
the detection probability Pd (there is a PU and the CRs
decide that the spectrum is not available).
ED is the most common spectrum sensing technique due
to its simplicity and not requiring any prior knowledge
about the transmitted signal to be detected. Further, it
has low complexity [2]. However, the performance of ED
is degraded greatly in low signal-to-noise (SNR) ratio.
On the other hand, CFD has a robustness detection in
low SNR conditions. However, its implementation and
computational complexity is very high compared to ED
[3].
Real world constraints may affect the detection perfor
mance of the spectrum sensing algorithms such as noise
uncertainty, multipath fading and shadowing effects [4][5].
The key solution to all these problems is cooperative
spectrum sensing. It takes advantage of the spatial diver
sity, i.e. different channel conditions, in which CR nodes
cooperate together to make the global decision. The global
decision can be made in a fusion center (FC) according to
hard decision rules, such as logical OR, AND, N-out-of-M
rules, or soft decision rules, such as likelihood ratio test,
soft linear combining, polynomial classifiers [1][6].
Cooperation of cognitive users is performed utilizing
a soft decision rule, in which a linear combination of
received energies is used to make a decision in [7]. The
advantage of linear combining is that users with better
channel conditions are given higher weights in making
the decision. Hence, more reliable decision is obtained.
However, the problem of finding the optimal combing
weights becomes more challenging when multiple sensing
techniques are utilized in the proposed scheme. In [6],

IEEE

2
the performance of pattern recognition models such as
linear and polynomial classifiers has been studied as a
soft decision combining rule for energy detection. Results
indicated that both classifiers have comparable perfor
mance in terms of the detection probability. Hence, in this
paper, a linear classifier ( LC ) will be utilized for global
decision making where the weights are obtained through a
supervised pattern recognition model. LCs are chosen here
due to their reduced complexity and similar performance
to polynomial classifiers.
The rest of the paper is organized as follows: In Section
II, we present the problem formulation and system model.
Section III illustrates the ED and CFD techniques. The
proposed technique is fully investigated in Section IV.
Linear classifier based cooperative Soft decision rule is
explained in Section V. Simulation results are given in
Section VI. Section VII concludes the paper.
II. PROBLEM FORMULATION AND SYSTEM MODEL
In this paper, we consider spectrum sensing in a CR
network consisting of !vI CR nodes with a central node ( i.e.
FC ) that decides on the channel availability. PU network
and CR network are assumed to be present within the
same geographical area. CR nodes must not cause harmful
interference to PUs.
In the proposed system, SUs are constantly sensing
the spectrum band for primary signal detection. \Vithin
a SU receiver, discriminative features are extracted from
the observed signal based on one of the spectrum sens
ing techniques discussed in Section III. The features are
transmitted to the FC through a relatively low data rate
control channel. A global decision is made by CR base
station using LC that is discussed in Section V. The
binary hypothesis test for spectrum sensing is formulated
as follows:

(1)
where i = 1,2, ..... !vI, Yd n] is the received signal by
the ithreceiver at the nthtime instant, x ( n ) is the PU's
signal to be detected. Further, Ho represents the null
hypothesis meaning that there is no primary signal and
only A\VGN noise, ni(n) with variance 17; , exists, while
HI describes the existence of a PU's signal in addition to
AWGN noise, ni ( n ) . The primary signal is passed through
a wireless channel with channel gain equivalent to hi which
is modeled as a slow fiat rayleigh fading channel, i.e. the
coherence time of the channel Tc is set to be much larger
than the primary signal's symbol duration Ts.
Different CR nodes are assumed to have independent
and identically distributed ( i.i.d. ) channel coefficients. The
ith CR is assumed to receive a signal with SN Ri that
differs according to its position from the PU. The SN Ri
( in dB ) follows a normal distribution with a variance 172
and mean equivalent to SN Ravg to account for the large
scale path loss.

III. PRIOR SPECTRUM SENSING TECHNIQUES

Over the past few years different techniques were pro
posed for spectrum sensing based on energy detection
( ED ) and cyclostationary feature detection (CFD ) .
A.

Energy Detection (ED)

ED is one of the most commonly used methods in

spectrum sensing as it is simple to implement and does not
require any prior information about the transmitted signal
to be detected. Furthermore, it has reduced execution
time, low computational and implementation complexity.
The detection statistics Ti is defined as follows:

(2)
where i = 1,2, ..., !vI. N is the number of samples during
one observation period and Yd n] is the received signal by
the ithreceiver. The decision rule at each CR node is given
by:
(3)
where Ai is the decision threshold. The selection of the
threshold in ED depends on the noise power; hence, it is
highly affected by noise uncertainty. Furthermore, in low
signal to noise ratio ( SNR ) regimes, reliable detection of
the transmitted signal is cumbersome.
B.

Cyclostationary Feature Detection (CFD)

As most communication signals introduce built-in pe

riodicity in their mean and autocorrelation, they can
be modeled as cyclostationary random processes. This
underlying periodicity is produced as a result of coupling
stationary message signals with sinusoidal carriers, cyclic
prefixes, sampling, etc. The spectral correlation density
( SCD ) of a cyclostationary signal is the Fourier transform
of its cyclic autocorrelation function ( CAF ) as follows [8]:

1
lim
"
T--'tCXJ t--'tCXJ T ut

S (j) = lim

6.t

6.t

-2

YT (t , j + ) Y;' (t , J - ) dt
2

2
(4)

where

1-t+ y (m )e-j27rmldm
T

(5)
t
is the complex envelope of the spectral component of
y (t), a is the second order cycle frequency, J = j, is the
frequency resolution and t is the averaging time over
which the SCD is estimated. It is obvious from (4) that
the SCD represents the temporal cross correlation between
shifted versions of the signal y (t) in the frequency domain
by . The hypothesis model of the CFD is as follows:

YT(t , l) =

"2

(6)

Rx Signal

where Si;, represents the SCD of the A\VGN at a certain

cyclic frequency a = ao when no signal is present, h is the
channel gain due to shadowing and multipath fading and
S;:,o [k] is the SCD of the primary signal. The decision rule
at each CR node is given by:

HI

'Y
T
t < It

(7)

SNRavg

>=

SNRth

Ho

where Ii is the decision threshold. The CFD has a

robustness detection in low SNR regime. However, its
computational complexity is very high compared to the
ED method.
ED

IV. PROPOSED SYSTEM

As illustrated in Section III, CFD has a better detection
performance than ED for the same average received SNR.
ED has many advantages such as short execution time,
simplicity and not requiring any prior knowledge of the
PU's signal. Thus, it is considered as the optimum spec
trum sensing technique at good SNR regimes. On the other
hand, CFD has a better performance even under bad SNR
conditions. However, its high complexity, processing time
and necessity of prior knowledge of the PU's signal give
birth to the idea of merging the two techniques together
in order to get a good performance comparable with CFD
detection performance with much less complexity and
processing time. This method is called hybrid cooperative
spectrum sensing technique (HCSST).
A.

Principle of HCSST

In HCSST, ED can be utilized by CR nodes till a certain

SNR threshold SN Rtf!) satisfying a certain requirement
on Pd. In other words, ED can be used if the average
received primary signal's SNR is above this SNR thresh
old; otherwise, CFD will be deployed. Hence, when the
average received SNR is above the SNR threshold, the
required detection performance can be achieved with less
complexity and execution time due to adopting ED. On
the other hand, when the average received SNR is below
the SNR threshold, CFD is adopted and the detection
performance is way better than the case if ED is deployed.
The block diagram of the proposed method is depicted in
Figure 1.
After each CR node selects the sensing technique based
on the average received SNR, it transmits its calculated
features, energy or peaks of the SCD, to the FC which
makes the global decision using LC. The detailed expla
nation of LC is shown in Section V. Overall, HCSST
makes use of the short processing time of ED at good
SNR regimes and the better performance of CFD at low
SNR regimes. Hence, the overall complexity and sensing
time is greatly reduced compared to the CFD time and the

eFD

Figure 1. Block diagram of the proposed method adaptation

detection performance is much better than the detection

performance of ED.
B.

Selection of the SNR Threshold

The detection performance of ED degrades severely at

low SNR regimes. Hence, the SNR threshold (SNRth) at
which the CR network must convert to CFD, must be
determined based on ED detection performance. In this
work, SNRth is defined as the average received SNR at
which Pd is smaller than 90% at Pf of 10% in order
to fulfill the requirements of the IEEE 802.22 standard
[9]. SNRth can be derived as follows: When AWGN only
presents in the spectrum with variance o'f, ED detection
statistics Ti will be the sum of the square of N zero mean
Gaussian random variables. Hence the detection statistics
will have a central Chi-square distribution with N degrees
of freedom. \Vhereas, when the primary signal is present
the detection statistics will have a non-central Chi-square
distribution with N degrees of freedom as follows [7]:
(8)
where
(9)
is the local SNR at each CR node. Es is the signal energy
of PU and hi is the channel gain. The average SNR at the
E Ihl2
output of the local energy detector is equal to ----:-:-+-.
:v OJ i
For a large number of samples N, Ti can be consIdered
asymptotically normally distributed with mean:

(10)

Pr(ED) = Pr(SNR

and variance
_

(11)

1 !=
-(J"y'27T

SNRt h

) 2 ) d'
-21 ( Y-SNRavg
Y
(J"

exp -

Pr(ED)

(12)

( SNRth SNRavg )

therefore:

-1

vkr Ix

ex)

exp( -

)du.

By taking
, squaring of both sides and defining a
number 2 as follows:
2

vVe

A; - 2N(J"; [Q-l

(Pd(i)) r -2AiO";N +N2(J";

Pr(CFD) =

C. Complexity of HCSST
The main aim of the HCSST method is to reduce the
computational complexity with respect to CFD. Here, we
need to note that we are considering Software defined
Radios (SDRs), hence there is no hardware complexity is
considered. The main complexity comes due to the number
of operations needed to be performed. Table I shows the
computational complexity of ED and CFD for single user
where L is the smoothing window size 10]. It is clearly
shown that the computational complexity of CFD is very
large, approximately log2N times, comparable with ED.
The complexity of HCSST can be formulated as follows:

O(HCS S T ) = O(ED) *Pr(ED) +O(CFD) *Pr(CFD)

(16)

where O(x) is the complexity of the method (x) and

Pr(x) is the probability that a CR node selects the method
(x). As mentioned in Section II, SNRi follows a normal
distribution with mean SN Ravg and variance (J"2, hence
its probability density function (PDF) is as follows:

(J"y'27T

))

1 y-SNRavg 2

(J"

SNRavg

(21)

Table

METHOD

( SNRth

(20)

ROUGH COMPUTATIONAL COMPLEXITY OF THE ED AND CFD

Technique

exp -"2

l-Q

(13)

(J";T); + (2N(J"; -2Ai(J"; -4(J";[Q-l (Pd(i) ) ]2) T)i +2 0

(14)
By solving (14), one can get the value of T)i and calculate
SN R;; as follows:
(i) T)i
(15)
SNRth
N
In this case, each CR node selects SN Rth to achieve Pd
of 90% at Pf of 10% at each node.

(19)

Hence, the complexity of HCSST can be calculated

using (16).

can derive the following equation:

fSNRi (Y) =

(18)

By solving the integral in (19), we can formulate

Pr(ED) as follows:

The SN Rth can be derived as follows:

where the Q-function is defined as

SNRth)

>

(17)

CFD
ED

Real Multiply
2N log2 N

Real Add

+ 5L

3N log2 N

4N

+ 3L

3N

V. LINEAR CLASSIFIER BASED COOPERATIVE SOFT

DECISION RULE

Spectrum sensing in CR networks has been introduced

as a pattern recognition problem 6]. Generally speaking,
the main aim of pattern recognition is assigning a signal
to one of a number of known categories based on features
derived to emphasize commonalities between those signals.
Typically, pattern recognition is used to address problems
including speech and face recognition. Input signals that
need to be classified are usually referred to as patterns. In
most cases, patterns are not useful for classification, and
they need to be pre-processed in order to acquire more
useful input to the classifier. This processed information
is called features and the process of acquiring them is
called feature extraction. Pattern recognition is used at
the cognitive network FC to detect the available spectrum
holes so that detection probability is maximized for a
desired false alarm rate 6]. Cognitive users monitor the
spectrum through the appropriate sensing scheme accord
ing to the received SNR. The sensed information by each
user provides discriminative features to the FC. At the
FC, the received features are applied to the designated
classifier model where a global decision is made about the
presence of primary transmission.
As shown in Figure 2, the steps of the global decision
making can be summarized as follows: The first step,
sensing, involves collecting the signal received by antennas
at different CR receivers. The following stage is the feature
extraction, in which each CR computes more useful data
i.e. energy or peaks of cyclic spectrum, after deciding
on the appropriate sensing scheme to be used. Thereafter,

ti,

the extracted features are applied to the designed classifier

at the FC to obtain a global decision variable. Features
extracted by any of the above sensing schemes will follow
a certain pattern when the spectrum is occupied by a
primary user. The pattern extracted would be different
when only noise is present in the spectrum. The difference
between these two patterns will be exploited as discrim
inative input data to the classifier for decision making.
In this paper, the classifier model to be implemented
for spectrum sensing in CR network is LC. The prob
lem of spectrum sensing can be formulated as a multi
input single-output (MISO) polynomial classifier. Features
extracted by different receivers, t [tl ... tM], form an !vI
dimensional input vector to the classifier. The classifier is
required to provide a single output score Yd, representing
the decision of whether or not a primary signal is present.
The output score fi is obtained at the output block after
linearly combining the expansion terms t as follows:

(22)
where Wi is the model of class i. Hence, each feature

ti received from a certain CR is multiplied by a weight

according to the received SNR. Therefore, users with
higher reliability are given higher weight in making the
global decision on spectrum availability. In supervised
pattern recognition, a set of training data is assumed to
be available and the classifier model weights are obtained
by exploiting this a prior known information. Once the
model parameters are estimated, the model can be used
to classify new novel data. Assuming a set of training data
M, which is a Q x !vI matrix where Q is the number of
feature vectors used in the training process and lvI is the
dimensionality of the input feature vectors (provided by
lvI CR users). Our goal is to solve for the best model
parameters {wd that minimizes the Euclidean distance
between Ii in (22) and the desired ideal output {zd
for class i. The ideal output Zi is a column vector of
length Q consisting of elements equal to ones for the
indices corresponding to primary signal present and zeros
otherwise. LC is trained to find an optimum set of weights,
w that minimizes the Euclidean distance between the ideal
target vector Zi and the training matrix M using mean
squared error as objective criterion, by solving the problem
in (23).
opt

wi

arg minllMw-zil12
w

(23)

The problem of (23) can be solved using the method of

normal equation:

(24)
which is used to compute class models {Wi }, i = 1, 2.
After training, the estimated class models {Wi} are used
for classification of novel data sets. At the end, the output
score Ii is compared to a threshold A to decide on Hi, for
i = {O, I}, corresponding to the binary hypothesis in (1).

r---,-------,

..:...I

CR,

CR",

<E

t,

Qassification

Ii

rt
FC

<EHi
Figure

2.

Block diagram of Linear Classifiers (LCs)

The threshold A can be calculated iteratively to achieve a

certain a desired constant Pf .
Calculating the threshold, in addition to the training
process, is performed offline. Thus, no additional complex
ity is added to the overall complexity calculated in Section
IV except for the linear combining when calculating the
output score. However, this linear combining complexity
can be neglected because it is a function of !vI, the number
of CR nodes.
VI. SIMULATION RESULTS
In this section, Monte Carlo simulation results are
presented to show the detection performance for HCSST
compared to ED and CFD. The network consists of !vI = 3
CR nodes who contribute in making the global decision at
the FC. The selection of SN Rth is based on the fulfillment
of the requirements, Pd = 90% at Pf = 10% for ED at each

CR node according to (15). After calculating SN R , each

CR node selects its sensing technique, ED or CFD, based
on the received SNR and extracts the features. Features
are send to the FC to be classified into exactly two classes;
channel is idle or channel is busy. The primary signal is
assumed to follow a linearly modulated signal model. For
simplicity, the received signal is modeled as all Is if the
channel is busy modulated by a carrier of Ie = 100kHz.
A slow fiat Rayleigh fading channel is considered for the
channel model. The observation window size N = 800
samples. The path loss variance 0'2 = 4 dB and Pf = O.l.
The LC network is trained using a very large training data
samples to obtain the optimum model weights resulting in
maximum Pd with a constant Pf . A novel data is used in
order to evaluate the model weights in order to caculate
the output score Yd.
The results presented in Figure 3 show the detection
probability of ED, CFD and HCSST as the received
average SNR varies. It can be shown that ED detection
performance degrades significantly at low SNR values;
hence, it will be used to determine SN Rth. It is clear
that the detection performance of the proposed technique

is better than the performance of the ED method and

it is very close to the CFD method. For example, at
SNRavg = -18 dB, Pd is around 0.2, 0.76 and 0. 84 for
ED, HCSST and CFD respectively. For Pd of 90%, the
required SNRavg is -16 dB, -14 dB and -7.5 dB for
CFD, HCSST and ED respectively resulting in a gain of
6.5 dB for HCSST over ED. Moreover, CFD is better than
HCSST by around 2 dB only on average.
Figure 4 shows the complexity comparison between
ED, CFD and HCSST. The results clearly show that the
complexity of HCSST is significantly reduced compared
to the CFD. At low SNR values, SNRavg < -10 dB, the
complexity of HCSST and CFD are almost equal and for
SN Ravg 2': -10 dB, the complexity of HCSST starts to
decrease rapidly and the detection performance is very
close to CFD. For example, at SNRavg = -2 dB, the
complexity is 1.7 x 104, 6 x 104,12 x 104operations for ED,
HCSST and CFD respectively.

VII. CONCLUSIONS
In this paper, a novel hybrid cooperative spectrum sens
ing technique ( HCSST ) is proposed. HCSST combines ED
and CFD according to a certain SNR threshold (SNRth).
Cooperative decision about existence of PU's signal is
made using a trained linear classifier ( LC ) at the fusion
center (FC ) . The mathematical expression of SN Rth is
derived based on the detection performance of ED. The
overall complexity of HCSST is calculated mathematically.
Simulation results show that the overall computational
complexity, sensing time of HCSST are significantly re
duced compared to CFD and the detection performance is
much better than the performance of ED and comparable
with CFD.
REFERENCES
[1]

T. Yucek and H. Arslan, "A survey of spectrum sensing al

gorithms for cognitive radio applications,"

Communications

Surveys Tutorials, IEEE, vol. 11, no. 1, pp. 116 -130, quarter
2009.
[2]

H. Urkowitz, "Energy detection of unknown deterministic sig

nals," Proceedings of the IEEE, vol. 55, no. 4, pp. 523 - 531,
april 1967.

[3]

S. Xu, Z. Zhao, and J. Shang, "Spectrum sensing based on

cyclostationarity," in Power Electronics and Intelligent Trans
portation System, 2008. PElTS '08. Workshop on, aug. 2008,

"O
:

pp.171-174.

:0
ro
.0

[4]

R. Tandra and A. Sahai, "Fundamental limits on detection

in low snr under noise uncertainty," in

c.
c
o

Communications and Mobile

Computing,

Wireless Networks,
2005 International

Conference on, vol.1, june 2005, pp.464- 469 vol.l.

U
Q)
1ii

[5]

S. Atapattu, C. Tellambura, and H. Jiang, "Performance of an

energy detector over channels with both multipath fading and

shadowing," Wireless Communications, IEEE Transactions on,

vol.9, no.12, pp.3662-3670, december 2010.
[6]

Y. Hassan, M. El-Tarhuni, and

K. Assaleh, "Comparison of

linear and polynomial classifiers for co-operative cognitive radio

-20

-15

-10
-5
Average received SNR (dB)

networks," in Personal Indoor and Mobile Radio Communica

tions (PIMRCj, 2010 IEEE 21st International Symposium on,

sept.2010, pp.797-802.
[7]

Z.Quan, S.Cui, and A.Sayed, "Optimal linear cooperation for

spectrum sensing in cognitive radio networks," Selected Topics
in Signal Processing, IEEE Journal of, vol.2, no.1, pp.28-40,

Figure 3. T he detection performance of ED, CFD and HCSST

feb.2008.

[8]

W. Gardner, "Exploitation of spectral redundancy in cyclo

stationary signals," Signal Processing Magazine, IEEE, vol. 8,

no.2, pp.14-36, april 1991.

[9]

8"

Q)

...J

"

7
6

c;;

>.
'lOi

Q)
Q.
E
o
u

[10]

Ikuma,

"Autocorrelation-based spectrum

S. Da, G. Xiaoying, C. Hsiao-Hwa, Q. Liang, and X. Miao,

"Significant cycle frequency based feature detection for cognitive
radio systems," in Cognitive Radio Oriented Wireless Networks
and Communications, 2009.

S;!

T.

nology, vol. 59, pp. 718-733, February, 2010.

(f)

10

Naraghi and

sensing for cognitive radios, " IEEE Trans. On Vehicular Tech

11

M.

CROWNCOM '09. 4th Interna

tional Conference on, june 2009, pp.1-4.

5
4
3
2

Average Received SNR (dB)

Figure 4. T he computational complexity of ED, CFD and HCSST