A Sequential Test Based Cooperative Spectrum Sensing Scheme for Cognitive Radios

Yeelin Shei and Yu T. Su

Department of Communications Engineering
National Chiao Tung University
1001 Dar-Shei Rd., Hsinchu, 30056, TAIWAN, +886-3-573-1820
E-mail: ytsu@mail.nctu.edu.tw
Abstract-Fast and accurate spectrum sensing is crucial in
realizing a reliable cognitive network. Cooperative spectrum
sensing can help reducing the mean detection time and increasing
the agility of the sensing process. However, when the number
of cognitive users is large, the bandwidth need for the control
channel that are used to report the secondary user nodes' results
to the fusion center may become excessively large. this paper,
we apply the sequential probability ratio test (SPRT) control
the average number of the reporting bits. It is shown that the
proposed technique not only reduces the mean detection time and
bandwidth but also outperforms its non-sequential counterpart.
We derive the relationships amongst the global performance, miss
probabllity and false alarm probability and show how to control
the average number of reports by thresholding the distributed
cognitive users.
I. INTRODUCTION
Cognitive radio (CR) technique has been proposed to exploit
the spectrum holes-the frequency bands which are not used
at some time or space-for license-exempt usages [3]. The
spectrum below 3 GHz has become increasing crowded but
report [1] has shown that the utilization of licensed spectrum
ranges from 15% to 85% only. Inspired by the CR concept
and the fact that the some TV channels are unused in many
rural areas, IEEE has approved the establishment of a working
group to develop a CR-based wireless standard utilizing the
spectrum between 54 MHz and 862 MHz [2].
The realization of a CR-based wireless network depends,
among other things, on the assumptions that network users are
able to accurately sense the existence of spectrum holes and a
proper coordination protocol among the unlicensed users is in
place. The sensing result is used to indicate the absence
o
)
or presence of a primary user in the band of concern.
It is desired that the sensing method gives high detection
probability, that is, the probability that the sensing output is
when the spectrum is in use, which is a measure on how well
the primary user are protected. On the other hand, the false
alarm probability, i.e., the probability that the sensing result
is HI when the spectrum is not used, must be low enough
to ensure efficient usage of the spectrum for a false alarm
will prevent a secondary user (SU) from using the licensed
band even though the spectrum is available. Another critical
concern about the sensing method used is the average time
needed to make a spectrum decision. As the availability of a
given band is non-deterministic, it is important for a SU to
seize the transmission opportunity as soon as possible.
It has been shown [4,5] that cooperative spectrum sensing
978-1-4244-2644-7/08/$25.00 2008 IEEE
improves the detection and false alarm probability perfor-
mance [6] and enhances the agility [7]. A cooperative sensing
scheme is usually conducted in two successive-sensing and
reporting-stages. In the first stage, every cognitive user per-
forms sensing independently using some detection method that
requires a fixed observation interval to make a local decision
which is then sent to the common receiver called fusion center
(FC) in the second stage through a control channel. The fusion
center then make a final decision as to if 0) or ( is
accepted.
To reduce the control channel's bandwidth requirement, a
cognitive user needs to quantize its decision before sending
it to the FC. Quantization of local observation in distributed
detection has attracted much research interest [8-12]. Although
quantization error and signal-to-noise ratio (SNR) loss are
introduced [9], three-bit quantization is enough to recover most
of the performance loss [10]. [11], it is shown that a decision
rule based on one-bit quantization can be asymptotically
optimal as the number of cooperative users and hence that
of the reporting bits approaches infinity. In general, the more
reporting bits the FC collects, the more reliable the decision
is. Similarly, the reliability of the sensor-to-center report is an
increasing function of the sensor's observation duration.
Conservation of the reporting (control) channel bandwidth
and reduction of the average observation time can be accom-
plished if a sequential test instead of a fixed sample size
test is used for the former can make a sensing decision as
soon as it collects sufficient evidence (observations). This
paper proposes a cooperative sensing scheme that employs
sequential tests in both sensor nodes and the FC. We restrict
our investigation to the case of one bit quantization reporting.
The sequential test we used is the so-called probability ratio
test (SPRT) [13]. SPRT is an optimal test that minimizes the
average required sample size among all tests which achieve
the same detection and false-alarm probabilities performance,
if the samples (observations) are independent. [12] suggests
a non-sequential approach which requires that each cognitive
user reports to the FC only if its decision is reliable. Although
the required control channel bandwidth can be reduced, it is
difficult to determine the reliability thresholds that meet the
performance requirement. We present a systematic method to
derive the detector parameter values that are guaranteed to
yield the designed performance.
The rest of the paper is organized as follows. Section
II, the system model is introduced and the major properties
of the SPRT is briefly reviewed. In Section III we present
key equations that relate various parameters of the proposed
sequential distributed sensing method. Simulation results are
provided in Section IV, followed by conclusion in Section V.
II. SYSTEM DESCRIPTION AND BASIC SEQUENTIAL TESTS
The setup of this system is based on the IEEE 802.22
WRAN scenario. The system model we consideration is shown
in Fig. 1, which includes a primary user, an FC, and cognitive-
premiss equipments (CPEs) as SUs. The SUs are randomly
distributed within the coverage radius of the FC.
(3)
(4)
p(Yk!H IT p(YkIHI)
p(YkIHo) k=1 p(YkIHo)
p(YkIHl) If p(YkIHl)
p(YkIHo) k=1 p(YkIHo)
A(Yk)A(Yk- 1)
types of detection errors, namely, the false alarm probability,
and the miss probability, Let Yk be the observation
at time k and Yk [Yl, Y2, ... , Yk]T be the column vector
consisting of i.i.d. observations. Then the likelihood ratio
(LR) of the kth observation is
A(Yk) p(YkIHI)
p(YkIHo)
and that for Yk is
30
20
10
The decision rule for the SPRT with thresholds "lo and "l1,
denoted by T(170, T}I) is given by
-10
-20
-30
A(Yk) 2:: T}I
A(Yk) :::; "lo
"l1 A(Yk) 2:: "lo
accept
accept
o
taking another observation (5)
Fig. 1. A CR network that consists of a primary user terminal and several
SU terminals with one as the fusion center (common receiver).
The received power Pi at the ith SU terminal and the
corresponding SNR ri are respectively given by
Pi 1,,M
z
(8)
(7)
log + (3) log
E[L(YIH
1-{3 {3
"l1 "lo
1-
Theorem 2: The average sample size of 171) is

E[KIHoJ = E[L(yIH
o
)]
The following properties of the SPRT are well-known [13].
Theorem 1: Let Pp a, (3 be the false-alarm and
detection probabilities associated with the SPRT T("lo, "l1),
then the two thresholds 770, "l1 satisfy
where K is the stopping time and
1-{3 {3
"l1 :::; --, "lo
a -a
If at the stopping time (i.e., the time when a decision to
accept or
o
is made), the LR is exactly equal to the
corresponding threshold, which happens if we have continuous
observation and the LR is a continuous process, the above two
inequalities become equalities.
L(YKIH
i
) log [A(YKIH
i
)], L(yIH
i
) log [A(yIH
Again, if the log-likelihood ratio (LLR) is a continuous pro-
cess, we have
For discrete observations, these two equations are only approx-
imations but in many cases they are excellent approximations.
(2)
Pi
"Yi 10 log 2' 1", ,M
(j
-10
and
where is the transmit power of the primary user,
i
is the
distance between the primary user and the ith SU, O'.L is the
path loss factor, is a scaling factor, is the noise power
and is the total number of SUs.
Approaches to solve a binary hypothesis test are generally
categorized into fixed sample size tests and sequential or vari-
able sample size tests. For the former class, one of two possible
actions is taken-accept or reject the null hypothesis Ho-after a
fixed number of samples are observed. In a sequential test, the
number of samples needed to make a decision is not predeter-
mined but depends on the values of the received samples. As it
can stop testing whenever the actual received samples provide
sufficient evidence for accepting or rejecting a hypothesis, the
observation time needed to make a decision is random. The
SPRT is a special class of sequential tests that offer an optimal
property when the samples are identical and independently
distributed (i.i.d.). It is a Neyman-Pearson type test whose
thresholds are functions of the required performance. More
specifically, the two thresholds, "lo, "l1, are determined by two
Fig. 2. Block diagram of the detector used a secondary user.
i
) == { log = 7702, = 0 (14)
log (11;= ) 7712, i 1
The FC computes its LLR by summing the LLR of each d
i
transmitted from the SUs and then compares with the two
thresholds determined by the designed global false-alarm and
miss probabilities and This fusion method is similar
(15)
(16)
p log
p
(1 - p ) log 1 _
p
where is the LR of the FC at the stopping time, and
is the corresponding number of sensing bits received from the
SUs. Invoking the approximation that the the FC's LR is equal
to one of the thresholds at the stopping time, we obtain
to the combination of the Chair-Varshney method [15] and the
SPRT.
Substituting (14) into (10) and (11), we obtain
log (1 + log (1 )18)
As and where == i ==
I} are set by the FC , the FC obtain and from
(15) and (16) ,and assign them to SUs. In particular, if ==
and == the thresholds for the FC's SPRT are
== while those for the SUs' SPRT are "702 == -1}12.
The overall distributed sensing method is shown in Fig. 3.
Each cooperative SU obtains its observation (sample), com-
putes the LR or LLR and compares it with the predetermined
thresholds. If the LR value exceeds one of the threshold, the
corresponding decision is reported to the FC, otherwise, no
sensing bit is transmitted. The process continues until the FC
notifies the SUs to stop. The FC collects sensing bits from the
cooperative SUs. The sensing process stops when the LR or
LLR computed by the FC exceeds one of the corresponding
thresholds.
Recall that our design parameter (thresholds) values are
derived based on the assumption that the stopping LR or LLR
values at both the FC and the SU terminals are equal to one
of the thresholds. In reality, these values would most likely
exceed one of the thresholds. If we define the excess as the
difference between the stopping LR value and the threshold,
then the excess at the FC will result in larger K 's and that
at the SU terminals leads to smaller K
>'71
or
<'70
Predetetion
filter
inpu W Hz
III. SPRT-BASED DISTRIBUTED SENSING
We use the SPRT at both the distributed sensor nodes and
FC. Each sensor node uses a energy detector with pre-detection
bandwidth Hz (see Fig. 2) to generate observations at t ==
== 1,2, ..., and to compute the resulting LRs.
A. SPRT in Cooperative Users
For the sensor (SU) nodes, the two thresholds used in the
test are determined by substituting the required false alarm
and miss probabilities, == and == {3, of the SUs into
(7).
Since the pdf of the energy detector output, y, is [14]
f(

o
Y 1 HI
where is the received SNR, is the time-bandwidth product.
denotes a central chi-square distribution with 2u degrees
of freedom (d.f.) and a non-central chi-square distri-
bution with 2u d.f. and non-centrality parameter
Moreover, from equation (12),with a sampling interval,
== 1 and the LLR of the kth sample received by the ith
SU is given by
log
log exp
0
exp (-
where
0
(.) is the zero-th order modified Bessel function of
the first kind. Upon receiving a new sample, a sensor terminal
makes a decision based on the SPRT rule defined by (5). It
sends' to the FC when HI is accepted, if H
o
is accepted,
and continues by waiting for the next sample without sending
any reporting bit if no threshold is crossed.
B. SPRT in the FC
Let be the random variable representing binary decision
of the ith SU (sensor node) received by the FC. Using the
approximation assumption that, at the stopping time, the LR
of
i
is equal to one of the thresholds, we obtain the LLR of
d
i
as
V. CONCLUSION
An SPRT-based cooperative spectrum sensing scheme for
use in CR networks has been proposed and analyzed. The
where r 1is smallest integer greater than or equal to and
the excess of the LLR is given by
~ ~ ; ~ ~ ; ) r/02, if
0
is accepted (23)
( ~ ~ ; 1- ~ ~ ; ) 7]12, if o is rejected
(22)
if
o
is accepted
if
o
is rejected
"'Of "1
== "102 02,
"112 7]12
IV. SIMULATION RESULTS
Simulation results of the proposed SPRT-based sensing
scheme in the CR network shown in Fig. 1 are reported in
this section. The parameters values used are as follows.
SUs are randomly distributed within the 5 km radius of the
FC. The secondary BS is 59.7 km away from the primary
user. During the sensing period, each SU keeps sampling its
energy detector output at a rate of W samples per second and
reporting its LR value, if necessary, to the FC until it is told
by the FC to stop sensing. The path loss exponent factor Q,
in (1) is set to be 3.5, and are set at a value such that
the primary user's SNR at the secondary BS is -2 dB. The
SUs' thresholds are determined by == == 0.01, and
== It is also assumed that each SU can estimate its
received SNR perfectly.
Fig. 4 shows the relation between the theoretical and
K derived by analysis and those estimated by simulation
with 20 SUs. The discrepancy is due to our zero-excess (over
two thresholds) assumption. Fig. 5 plots the total sensing time
as a function or Under the same environment, the
average sensing time for a distributed detection strategy with
fix-sample-size energy detection at the SU sites and the "AND"
fusion [16] at FC is == 94.448. If the "OR" fusion is used
instead, the sensing time becomes == 67.794. Obviously,
our approach achieves significant improvement on the mean
sensing time performance. However, these curves do exhibit
some discontinuities. (21) shows that the average sensing
time is a function of and
But it is that results in the discon-
tinuities. This is because the FC may receive more than
one reporting sensing bits at the stopping time and if the
LLR exceeds one of the thresholds, since == and
== by design, we have == and 7]02 == -1712,
Fig. 6 plots and for different and
_ "'02 "'02 "'02
K 'So It is found that there is a discontinuity in the excess
whenever the integer part of changes. Fig. 7 plots the four
"'02
terms in the denominator of with 20 SUs. We find that
remains constant, is small and insensitive
to the threshold while
i
] is a decreasing function
of the threshold but as a function of is not
continuous but exhibit a saw-like shape.
ith CR user node
==
+ + +
(21)
Fig. 3. Flow chart of the proposed SPRT-based distributed spectrum sensing
method
C. Sensing Time Analysis
The complete distributed sensing process consists of two
SPRT stages. If there are SUs in this system joining the
cooperative sensing process, M samples are available at each
sampling epoch. The conditional average LR of one M -sample
block observation is then given by the sum
== L
where
j
] is the conditional average LR of one
sample associated with the ith SU's observation From
(8), we have
== ~ (20)
where is the total sensing time and is the stopping
LR under includes four parts:
the LR threshold used by the FC.
the excess LR at the FC.
the excess LR at a SU node.
L non-excess stopping LR value at a SU node.
According to equation (8) and (20), it can be proved, after
some algebraic manipulations, that
, ,0 110111
...... 110111
110111
02
1-110111
02
20
18
16
14
o

12
"iii
10
;C
W 8
10 12 14 16 18 20
E[Kfil
-
....

:::M.,

....


o
2 8 10 12 14 16 18 20
Fig. 4. R used in simulation.
Fig. 6. Behaviors of 1and 1- for different
26
20 16 8 10 12 14
12 :.
.2 10

o 8

6
"T
....J 4

2 K., .
",
2 4
8 10 12 14 16 18 20
.."ea
o
..
.......',.......... lit ..
2
12
10
[20
Q)
E
18
C>
c:
16 .
14

Fig. 5. Normalized sensing time as a function of

Fig. 7. The average LR of the four terms in (28) for different
scheme has the advantage of minimizing the control channel
bandwidth and the total sensing time while rendering little or
no compromise in performance. In fact, the proposed scheme
is capable of controlling the average number of sensing bits
sent to the FC with any given false alarm and miss probabilities
performance requirement.
ACKNOWLEDGEMENT
This work is supported in part by the NCTUIITRI Joint
Research Center under Contract G1-97006.
REFERENCES
[1] Federal Communications Commission, "Spectrum Policy Task Force,"
Rep. ET Docket no. 02-135, Nov. 2002.
[2] C. Cordeiro, et al. "IEEE 802.22: An introduction to the first worldwide
wireless standard based on cognitive radios," Commun., vol.l, pp. 38-
47, Apr. 2006 .
[3] S. Haykin, "Cognitive radio: brain-empowered wireless communications,"
IEEE Select. Areas Commun., vol. 23, pp. 201-220, Feb. 2005.
[4] A. Ghasemi and E. S. Sousa, "Collaborative spectrum sensing for
opportunistic access in fading environments," in Proc. 1st IEEE Symp.
New Frontiers in Dynamic Spectrum Access Networks, Baltimore, USA,
Nov. 8-11, 2005, pp. 131-136.
[5] G. Ganesan and Y. Li, "Cooperative spectrum sensing in cognitive radio
networks," in Proc. DySPAN, Nov. 2005, pp. 137-143.
[6] G. Ganesan and Y. G. Li, "Agility improvement through cooperation
diversity in cognitive radio," in Proc. IEEE GlobeCom, St. Louis, USA,
Nov. 28-Dec. 2, 2005, pp. 2505-2509.
[7] E. Peh and Y. C. Liang, "Optimization for cooperative sensing in cognitive
radio networks," in Proc. IEEE WCNC, Hong Kong, Mar. 11-15, 2007,
pp.27-32.
[8] A. Sahai, et ale "Some fundamental limits on cognitive radio," in Proc.
Allerton Conf., Monticello, USA, Oct. 2004.
[9] W. A. Hashlamoun and P. K. Varshney, "Near-optimum quantization for
signal detection," IEEE Trans. Commun., vol. 44, pp. 294-297, Mar. 1996.
[10] R. S. Blum, "Distributed detection for diversity reception of fading
signals in noise," IEEE Trans. Inform. Theory, pp. 158-164, Jan. 1999.
[11] J. F. Chamberland and V. V. Veeravalli, "Decentralized detection in
sensor networks," IEEE Trans. Signal Proces., Vol. 51, Issue 2,pp. 407-
416, Feb. 2003.
[12] C. Sun, W. Zhang, and K. B. Letaief, "Cooperative spectrum sensing for
cognitive radios under bandwidth constraints," in Proc. IEEE Int. Wireless
Commun. Networking Conf., Hong Kong, Mar. 11-15,2007, pp. 1-5.
[13] A. Wald, "Sequential tests of statistical hypothesis," Annals Math.
Statistics, vol. 16, no. 2, pp. 117-186, Jun. 1945.
[14] F. F. Digham, et al., "On the energy detection of unknown signals over
fading channels," in Proc. ICC, Anchorage, USA, May 11-15,2003, pp.
3575-3579.
[15] Z. Chair and P.K. Varshney, "Optimal data fusion in multiple sensor
detection systems," IEEE Trans. Aerospace, Elect. Syst., pp. 98-101, Jan.
1986.
[16] P. K. Varshney, Distributed detection and data fusion, Springer-Verlag,
New York, 1997.