You are on page 1of 5

CLUSTERING METHODS FOR DISTRIBUTED SPECTRUM SENSING IN

COGNITIVE RADIO SYSTEMS


Amy C. Malady and Claudio R. C. M. da Silva
Bradley Department of Electrical and Computer Engineering
Virginia Polytechnic Institute and State University
Blacksburg, VA USA 24061
email: {amalady, cdasilva}@vt.edu
ABSTRACT

50

100

It is known that the reliability of spectrum sensing in


cognitive radio systems has the potential to improve when
this critical process is performed distributively by multiple
radios. In situations in which a large number of cognitive
radios are present in a given geographical area, these
radios will likely have to be divided into multiple sensing
clusters in order to keep bandwidth and power requirements manageable. In this paper, we propose different
clustering methods for distributed spectrum sensing and
evaluate their performance in terms of sensing reliability
and resultant spectrum efficiency. The use of position
information by clustering methods is also considered.
INTRODUCTION
In recent years, the concept of cognitive radios (CRs)
introduced a new approach to communication systems
that promotes more efficient use of the spectrum [1], [2].
As a result of their distinctive learning and adaptation
capabilities, CRs are able to find frequencies unused by a
primary user (PU) at a given location and time and dynamically access this spectrum. However, dynamic spectrum
access demands that CRs perform, either independently
or distributively, reliable and efficient spectrum sensing. If
the spectrum sensing estimates are not reliable, spectrum
access decisions will be based upon noisy information,
which can lead to harmful interference.
Given that spectrum utilization is a spatial phenomenon, it
is only natural to perform spectrum sensing in a distributed
manner. Among other advantages, distributed spectrum
sensing has the potential to reduce sensitivity requirements of individual CRs and to minimize the probability
of interference due to the hidden node problem [3]. In
distributed sensing, each CR obtains relevant information
of the spectrum, processes this information, and then shares
a summary of its sensing with other CRs. We define a
set of CRs that perform distributed spectrum sensing as a
sensing cluster. Using the summaries, the sensing cluster
makes a decision on whether a given frequency band is
being used by a PU. Based on this decision, either all CRs
This work was supported in part by a gift from Texas Instruments.

c 2008 IEEE
978-1-4244-2677-5/08/$25.00 

150

200

250

300

350

400
200

400

600

800

1000

1200

Fig. 1.
Illustration of a cognitive radio network (represented by
handheld radios) and a primary user (antenna tower). The grey area
represents the region of potential interference to the primary user.

in the sensing cluster use (no PU detected) or do not use


(PU detected) the band of interest.
A number of recent papers (including ours) have proposed
different approaches to distributed spectrum sensing, and
addressed some of its design and implementation issues.
See, among others, [4]-[8]. However, the clustering of
CRs in the context of distributed spectrum sensing is an
important problem that remains mostly unaddressed. Previous work on clustering and spectrum sensing include [9][11]. Our contribution in this paper is to develop and
evaluate different clustering methods that, by exploiting
any position information available, result in more reliable
and/or efficient spectrum sensing. In addition, we propose
and analyze new performance metrics for the evaluation of
clustering methods for distributed spectrum sensing.
PROBLEM DEFINITION AND
PERFORMANCE METRICS
When properly designed, the reliability of distributed spectrum sensing is expected to increase with the number of
CRs in a sensing cluster [4]-[10]. However, as the number
of CRs increases, more bandwidth and power is spent
during the exchange of spectrum sensing data, and fusing
the sensing summaries from multiple CRs becomes more
complex. Also, for the case in which either all or none of
the CRs in a sensing cluster access the spectrum (based on
the results of their distributed sensing), this increase in the
number of CRs might also lead to a larger number of false
spectrum access denials, which we define next. Therefore,
it might be advantageous to divide the CRs in a given area
into multiple sensing clusters [9]-[11].
Consider the scenario shown in Fig. 1. The grey area in

1 of 5

this figure represents what we define here as the region of


potential interference (RPI) of the PU. CRs inside the RPI
have the potential to interfere with the PU as a result of
a missed detection, but would be allowed to dynamically
access the spectrum if it were not being used by a PU1 .
If the six CRs in Fig. 1 form a sensing cluster, it is likely
that their distributed sensing will indicate the presence of
the PU (since four of the CRs are inside the RPI, where the
PUs signal is the strongest). If this is the case, the six CRs
will not access the licensed spectrum. However, note that
the two CRs outside of the RPI normally would be able to
access the spectrum even when the PU is communicating.
They will not in this case due to the imposed restriction
that all CRs of a sensing cluster must reach a consensus
on spectrum sensing and access2 . We define this event as
a false spectrum access denial.
It is worth noting that because the two CRs outside of the
RPI in Fig. 1 are far from the PU, the individual sensing
of these two CRs will likely detect very low levels of PU
signals; a possible indication that it is safe to use licensed
spectrum. Thus, false spectrum access denial would likely
not have occurred if the two CRs outside of the RPI
formed a sensing cluster by themselves. In this scenario,
the number of false spectrum access denials would be
minimized by forming a sensing cluster with the two CRs
outside the RPI, and a second sensing cluster with the
four CRs located inside the RPI. However, this new cluster
formation has a lower spectrum sensing reliability, as each
sensing cluster has a smaller number of CRs. Therefore,
there is a trade-off between spectrum sensing reliability and
likelihood of false spectrum access denials when dividing
CRs in a given geographical area into sensing clusters.
In order to quantify this trade-off, we propose the following
performance metrics for the distributed sensing stage:
Probability of detection (PD ) Average probability
that a sensing cluster will detect the presence of a
PU, given that there is a PU transmitting in the area
of interest.
Probability of false alarm (PF A ) Average probability that a sensing cluster will indicate the presence of
a PU, given that there is not a PU transmitting in the
1
In the same way, CRs not inside any RPIs could access licensed
spectrum with no spectrum sensing. However, this paper is mostly
concerned with the situation in which CRs do not have information
about the PUs position (and corresponding RPIs). In this case, the CRs
must make spectrum access decisions based only on spectrum sensing.
2

The distributed sensing stage might have indicated the presence of


PU signals because some of the CRs in the sensing cluster are inside
a RPI (and the corresponding PU is transmitting see Fig. 1) or, if all
CRs in the sensing cluster are not in any RPI, they can still measure
relatively high levels of signals from nearby transmitting PUs.

area of interest.
Probability of false spectrum access denial (PF SAD )
Average probability that a CR outside a RPI will not
access the spectrum because its distributed spectrum
sensing indicated the presence of PU transmissions,
given that there are PU transmissions in the area of
interest.
Probability of potential interference (PP I ) Average
probability that a CR inside a RPI will access the
spectrum because distributed spectrum sensing did not
detect the presence of PU transmissions, given that
there are PU transmissions in the area of interest.
CLUSTERING METHODS

The main objective of this paper is to investigate different


methods for the division of CRs into a given number
of sensing clusters and characterize their performance in
terms of PD , PF A , PF SAD , and PP I . The methods we consider have different assumptions regarding the information
available about the position of the CRs and/or PUs.
In the case in which the position of both the CRs and
the PUs is unknown, we consider the following clustering
method:
Random clustering The CRs are divided into a
given number of sensing clusters randomly, with the
same (or approximately the same) number of CRs per
cluster.
We consider two methods for the case in which the position
of the CRs is known and of the PUs is unknown. The
position information of the CRs could be obtained through
the use of GPS or by using one of the many different
algorithms available for the localization of the nodes of a
wireless network [12], for example. These methods are:
Reference-based clustering In this method, the CRs
are divided into clusters according to their positions
with respect to a given reference. For example, CRs
in Manhattan could be divided into three clusters depending on whether they are in downtown, midtown,
or uptown; or into two clusters depending whether
they are on the east or west side of Fifth Avenue.
Statistical clustering The CRs are divided into
clusters using a statistical method that takes the relative position of all CRs in the system into account.
That is, CRs are clustered according to their relative
proximities.
Clustering has been extensively studied in areas such as
data compression and pattern classification [13], [14], and,
more recently, in the context of ad-hoc and sensor networks
(for example, as an efficient approach for organizing the

2 of 5

network topology for the purpose of balancing the load and


prolonging the network lifetime [15]). In our application,
as we assume the CRs positions and desired number of
clusters to be known, we use the classic k-means clustering
method as the statistical clustering method. In the k-means
method, beginning with an initial guess for the centroid
of the clusters, the following two steps are performed
iteratively until a convergence condition is satisfied: 1) the
points (CR positions) are assigned to the cluster with the
closest centroid, then 2) compute the centroids for the
new clusters [13], [14].
Lastly, we consider the following clustering method for the
case in which the positions of both the CRs and the PUs
are known:
Distance-based clustering Out of the K CRs in a
sensing cluster, only the k (k < K) closest CRs to
the PU perform distributed spectrum sensing.
The PU position information could be obtained from
databases of regulatory agencies or estimated as part of the
spectrum sensing process, for example. (Note that spectrum
sensing is necessary even if the position of the PU is known
because the PU may or may not be communicating in a
given instant of time.)
SYSTEM DEFINITION
The IEEE 802.22 standard, the CR-based standard for
dynamic spectrum access and sharing of TV bands, divides
spectrum sensing into two stages [16], [17]. In the first
stage, usually referred to as fast (or coarse) sensing, a
relatively simple sensing technique (e.g., energy detection)
is used to detect the presence of signals in a given
frequency band over a short period of time. In the second
stage, usually referred to as fine sensing, a more detailed
sensing (e.g., feature detection [8]) is performed in order
to decrease the false alarm rate and detect the presence of
weak signals. We focus our analysis on the fast sensing
stage. However, the qualitative results obtained here (for
clustering in the context of spectrum sensing) are believed
to also be valid for the fine sensing stage.
Consider a scenario in which at most one PU transmits at
a given location and time, and denote its transmitted signal
by s(t). When the PU transmits, assuming an additive
white Gaussian noise (AWGN) channel, the received signal
ri (t) at the ith CR is written as

ri (t) = ai s(t i ) + Ni ni (t),
(1)
where ai and i are the channel gain and time delay, respectively, and Ni is a normalization constant. The thermal
noise ni (t) is modeled as a zero-mean Gaussian process
with power spectral density N0 /2 and bandwidth W Hz.

Let us first assume that the CRs in a sensing cluster


transmit their complete observation ri (t) to a fusion
center (which can either be a dedicated entity responsible
for fusing sensing data or one of the CRs), and that this
fusion center uses a general linear combination of the form
r(t) =

NS


ci ri (t),

(2)

i=1

where NS is the number of CRs in the sensing cluster. For


the case in which the noise processes ni (t) are uncorrelated
and the time delays i in (1) can be tracked (such that
the PU signal components in ri (t), for 1 i NS , are
essentially identical except in amplitude), it is well known
that among all systems of the form (2), the one that yields
maximum signal-to-noise ratio (SNR) of the output signal
r(t) is the maximal-ratio combiner (MRC) [18]. Denoting
the SNR at the ith CR as i , the MRC realizes
=

NS


i .

(3)

i=1

Based on this result, we assume that each CR estimates


the SNR in a given frequency band, sends this value (the
summary of its observation) to a fusion center, and then
the fusion center uses (3). If is greater than a given
threshold , the decision is made that there are PUs using
the band. Otherwise, the decision is made that there are
no PUs using the band. Note that if the estimation process
was perfect, = 0 (i.e., dB) for the case in which no
PU transmissions are present.
A good review (and comparative performance study) of
SNR estimation techniques for the AWGN channel is found
in [19]. In this paper, we consider the maximum-likelihood
(ML) estimator for SNR derived in [20]. Assuming that the
signal s(t) in (1) is known and has bandwidth W Hz, and
taking i to be zero (recall the assumption that i can be
tracked), the ML estimator for the SNR is given by [20]

2
1 2W T
r
s
i,j
j
j=1
2W T
i = 
 
2 , (4)

2W T 2
1
1 2W T

r
r
s
i,j j
j=1
j=1
i,j
2W T
2W T
where ri,j = ri (jTs ) and sj = s(jTs ) are time samples of
ri (t) and s(t) spaced Ts = 1/2W s apart, respectively. In
the derivation of (4), it was assumed that s(t) and ni (t) are
normalized to unit power over the observation period T so
that a2i and Ni are the signal and noise powers, respectively.
NUMERICAL ANALYSIS
To evaluate the performance of the proposed clustering
methods, we consider a scenario in which NCR CRs and

3 of 5

10

All
Probability of False Spectrum Access Denial

Probability of Detection

All

Individual
1

10

10

10

10

Individual

10

10

10

10

10

10

10

10

10

Probability of False Alarm

10

10

10

Probability of Potential Interference

Fig. 2. PF A vs. PD for scenario 1. Random (solid), statistical (dashed),


and reference-based (dash-dot) clustering methods.

Fig. 4. PP I vs. PF SAD for scenario 1. Random (solid), statistical


(dashed), and reference-based (dash-dot) clustering.

10

10

Probability of False Spectrum Access Denial

All

Probability of Detection

All

10

Individual

10

10

10

Individual

10

10

10

10

10

10

10

10

Probability of False Alarm

10

10

10

10

Probability of Potential Interference

Fig. 3. PF A vs. PD for scenario 2. Random (solid), statistical (dashed),


and reference-based (dash-dot) clustering methods.

Fig. 5. PP I vs. PF SAD for scenario 2. Random (solid), statistical


(dashed), and reference-based (dash-dot) clustering.

one PU are uniformly distributed in a rectangular region of


size L Lm. Two sets of values for these parameters are
considered, NCR = 10, L = 100m, and 2 sensing clusters
(scenario 1), and NCR = 20, L = 200m, and 4 sensing
clusters (scenario 2). For a given scenario realization, the
probabilities PD , PF A , PF SAD , and PP I are computed
for each clustering method by simulation from a number
of SNR estimates obtained for each CR using (4). These
probabilities are then averaged over multiple scenario realizations (that is, positions of CRs and PU). The decision
threshold is taken to be in the range (26dB, 10dB).
Assuming a free-space model, the path loss between the
PU and a CR at distance dm is given by

for the four sensing clusters case. In Figs. 2-6, the cases in
which the CRs perform spectrum sensing independently
(labeled individual) and when the CRs form a single
sensing cluster (all) are shown for reference. For Figs.
4 and 5, the RPI is defined to be a circular area of radius
20m around the PU.

PL dB = 10 log10 (Gl c)2 / (4df )2 ,

(5)

where f is the signal frequency, Gl is the product of the


transmit and receive antenna field radiation patterns in the
LOS direction, and c is the speed of light [21]. We take f
and Gl to be equal to 900 MHz and 1, respectively. Also,
we take the SNR at the transmitter to be 50 dB.
For the reference-based method, CRs are divided according
to whether they are on the left or right sides of the rectangular region for the 2 sensing clusters case, or on the four
quadrants (upper left, upper right, lower left, lower right)

The receiver operating characteristic (PF A vs. PD ) of the


proposed clustering methods are presented in Figs. 2 and
3 for scenarios 1 and 2, respectively. It is seen for both
scenarios that the single cluster case provides the best
PD for a given PF A , followed by the three clustering
methods (with 2 and 4 clusters, respectively), and lastly
all CRs acting individually has the worst performance,
as expected. It is interesting to note in these figures
that the random clustering method outperforms both the
statistical and reference-based clustering. Note that the
CRs in clusters formed by the statistical and referencebased methods are relatively closer to each other than those
within the random clusters. Thus, the worst case scenario
for distributed detection (namely, PU distant from all CRs
in the cluster) is more likely to happen for the statistical
and reference-based methods than for random clustering.
The superior performance of the random clustering method
is also due to the fact that while the clusters in this method

4 of 5

smaller for methods that use CR position data. As expected,


our results also indicate that more reliable sensing can be
achieved if the position of PUs is known.

0.9
0.8

Probability of Detection

0.7
0.6

R EFERENCES

0.5

All
0.4

0.3

10

10

10

10

10

Probability of False Alarm

Fig. 6. PF A vs. PD for distance-based clustering. Closest 1 (dotted),


2 (dash-dot), 4 (dashed), and 6 (solid) CRs out of 10. L = 200m.

always have the same number of CRs, the clusters in the


other two methods may have different sizes. For these two
methods, while one of the clusters will likely have a better
performance (more CRs), the other one will likely have a
worse performance (less CRs). On average, this results in
an overall performance loss.
While it is advantageous to have sensing clusters with
more CRs in order to increase the reliability of the sensing
stage, Figs. 4 and 5 depict that large clusters lead to a
larger number of false spectrum access denials for a given
PP I . These figures indicate that in order to better utilize
the spectrum, each CR should make its spectrum access
decisions based only on its own sensing. The problem is
that the required circuitry and processing capability for a
single CR to provide a given sensing reliability might be
too costly (hence the need for distributed sensing).
The performance of the proposed clustering method for the
case in which the position of the PU is known (distancebased clustering), is presented in Fig. 6. It is seen in this
figure that the detection performance can be improved
by combining only the observations of the CRs that are
closer to the PU (and, consequently, that have higher SNR).
However, there is an optimum number of CRs that should
be combined. For example, for the scenario considered in
Fig. 6, the 4 out of 10 CRs outperforms the 6 out of 10
CRs case. Incorporating low SNR CRs (which also have
higher SNR estimation errors [20]) can lead to an overall
decrease in detection performance.
CONCLUSIONS
This paper analyzed the clustering of CRs in the context of
distributed spectrum sensing. We observed that clustering
allows for a good compromise between detection reliability
and spectrum use efficiency. Our results show that the random clustering method outperforms (higher PD for a given
PF A ) the two proposed methods that exploit CR position
information. However, the number of false spectrum access
denials for a given probability of potential interference is

[1] J. Mitola III and G. Q. Maguire Jr., Cognitive radio: Making


software radios more personal, IEEE Pers. Commun., vol. 6, pp.
13-18, Aug. 1999.
[2] S. Haykin, Cognitive radio: Brain-empowered wireless communications, IEEE J. Sel. Areas Commun., pp. 201-220, Feb. 2005.
[3] W. Krenik and A. Batra, Cognitive radio techniques for wide area
networks, in Proc. Design Automation Conf., Anaheim, CA, 2005,
pp. 409-412.
[4] D. Cabric, A. Tkachenko, and R. W. Broderesn, Experimental
study of spectrum sensing based on energy detection and network
cooperation, in Proc. ACM TAPAS, Boston, MA, 2006.
[5] A. Ghasemi and E. S. Sousa, Opportunistic spectrum access in
fading channels through collaborative sensing, J. Commun., vol. 2,
pp. 71-82, Mar. 2007.
[6] M. Gandetto and C. Regazzoni, Spectrum sensing: A distributed
approach for cognitive radio terminals, IEEE J. Sel. Areas Commun., vol. 25, pp. 546-557, April 2007.
[7] C. R. C. M. da Silva, W. C. Headley, J. D. Reed, and Y. Zhao, The
application of distributed spectrum sensing and available resources
maps to cognitive radio systems, in Proc. Information Theory and
Appl. Workshop, La Jolla, CA, 2008, pp. 53-57.
[8] W. C. Headley, J. D. Reed, and C. R. C. M. da Silva, Distributed
cyclic spectrum feature-based modulation classification, in Proc.
IEEE Wireless Comm. and Netw. Conf., Las Vegas, NV, 2008, pp.
1200-1204.
[9] C. Sun, W. Zhang, and K. B. Letaief, Cluster-based cooperative
spectrum sensing in cognitive radio systems, in Proc. IEEE Intl.
Conf. Commun., Glasgow, Scotland, 2007, pp. 2511-2515.
[10] C.-H. Lee and W. Wolf, Energy efficient techniques for cooperative spectrum sensing in cognitive radios, in Proc. IEEE Consumer
Commun and Netw. Conf., Las Vegas, NV, 2008, pp. 968-972.
[11] P. Pawelczak, C. Guo, R. V. Prasad, and R. Hekmat, Cluster-based
spectrum sensing architecture for opportunistic spectrum access
networks, International Research Centre for Telecommunications
and Radar, Tech. Rep. IRCTR-S-004-07, Feb. 2007.
[12] S. Gezici, A survey on wireless position estimation, Wireless
Pers. Commun., vol. 44, pp. 263-282, Feb. 2008.
[13] A. Gersho and R. M. Gray, Vector Quantization and Signal
Compression. Springer, 1991.
[14] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification.
Wiley-Interscience, 2000.
[15] O. Younis, M. Krunz, and S. Ramasubramanian, Node clustering
in wireless sensor networks: Recent developments and deployment
challenges, IEEE Network, vol. 20, pp. 20-25, May/June 2006.
[16] C. Cordeiro, K. Challapali, and M. Ghosh, Cognitive PHY and
MAC layers for dynamic spectrum access and sharing of TV bands,
in Proc. ACM TAPAS, Boston, MA, 2006.
[17] IEEE 802.22 WG LAN/MAN Standards Committee, Cognitive
wireless RAN MAC and PHY specifications: Policies and procedures for operation in the TV Bands, draft standard, Nov. 2007.
[18] D. G. Brennan, Linear diversity combining techniques, Proc.
IRE, vol. 47, pp. 1075-1102, June 1959.
[19] D. R. Pauluzzi and N. C. Beaulieu, A comparison of SNR estimation techniques for the AWGN channel, IEEE Trans. Commun.,
vol. 48, pp. 1681-1691, Oct. 2000.
[20] R. M. Gagliardi and C. M. Thomas, PCM data reliability monitoring through estimation of signal-to-noise ratio, IEEE Trans.
Commun. Technol., vol. 16, pp. 479-486, June 1968.
[21] A. Goldsmith, Wireless Communications. Cambridge Press, 2005.

5 of 5

You might also like