Está en la página 1de 4

IEEE SIGNAL PROCESSING LETTERS, VOL. 17, NO.

10, OCTOBER 2010

867

Improved Energy Detectors for Cognitive Radios


With Randomly Arriving or Departing Primary Users
Norman C. Beaulieu, Fellow, IEEE, and Yunfei Chen, Senior Member, IEEE

AbstractNew and improved energy detectors for cognitive radios are derived by considering the effect of the primary user traffic
on spectrum sensing. The new energy detectors are designed based
on the assumption that the primary user randomly arrives or departs during the sensing period. Numerical results show that the
new energy detector outperforms the conventional energy detector
in all the cases examined. The performance gain depends on the
operating signal-to-noise ratio as well as the sample size used.
Index TermsCognitive radio, energy detector, primary user.

I. INTRODUCTION

NE of the most important elements of cognitive radio is


the detection of vacant licensed bandwidth in the radio
environment. This is accomplished by spectrum sensing. Energy
detection is one of the simplest methods [1][3]. Although the
energy detector doesnt perform as well as the matched filtering
detector or the feature-based detector [4], its simple structure allows a quick sensing decision to be made within a short sensing
period. Thus, the energy detector is widely used in cognitive
radio systems. In the conventional energy detector [1][3], the
primary user is assumed to be either present or absent for the
whole sensing period. This is a reasonable model when the primary user has low traffic. However, for medium to high traffic,
the primary user could randomly appear or disappear during the
sensing period. In this case, it was found in [5] and [6] that the
sensing performance of the conventional energy detector degrades. Instead of evaluating the performance of the conventional energy detector when the primary user randomly appears
or disappears in the sensing period, as was done in [5] and [6],
one can also optimize the energy detector based on the primary
user traffic employing the averaged likelihood ratio method in
[7].
In this letter, the energy detector is improved for primary
users having random arrival or departure time during sensing.
In the derivation the arrival time, or the departure time, of the
primary user is assumed to be exponentially distributed over the
sensing period. The performance of the new detector is compared with that of the conventional detector. Numerical examples show that the new energy detector outperforms the conven-

tional energy detector in all the cases considered. The performance gain depends on the number of samples used as well as
on the average signal-to-noise ratio (SNR).
II. DERIVATION
as the hypothesis that the primary user is absent
Denote
as the hypothesis that the primary user is present. Specand
trum sensing determines which hypothesis is most likely by processing the received signal from the interested frequency band.
We assume that the primary user either arrives or departs only
once during the sensing period.
, if the primary user departs randomly during the
Under
sensing period, the received signal from the band of interest can
be expressed as

(1)
where indexes the received samples,
represents the unknown signal from the primary user,
represents independent Gaussian noise samples each with mean zero and variance
represents the total number of samples available in the
sensing period and the random departure time of the primary
and
. Note that
user occurs between samples
yields the case when the primary user is absent during the whole
sensing period. Note also that the primary user has to leave by
.
the end of the sensing period. Thus,
Under
, if the primary user arrives randomly during the
sensing period, the received signal from the frequency band of
interest can be written as

(2)
where the random arrival time of the primary user occurs between samples and
and all other symbols are defined
as before. When
, the primary user is present during the
whole sensing period. Also, the primary user has to arrive by the
end of the sensing period such that
.
Using (1), the likelihood function under
can be derived as

(3)
Manuscript received May 17, 2010; revised July 29, 2010; accepted July 29,
2010. Date of publication August 09, 2010; date of current version August 19,
2010. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Teng Joon Lim.
N. C. Beaulieu is with the Department of Electrical & Computer Engineering,
University of Alberta, Edmonton, AB T6G 2V4 Canada (e-mail: beaulieu@ece.
ualberta.ca).
Y. Chen is with the School of Engineering, University of Warwick, Coventry,
U.K. CV4 7AL (e-mail: Yunfei.Chen@warwick.ac.uk).
Digital Object Identifier 10.1109/LSP.2010.2064768

where
.
hood function under

1070-9908/$26.00 2010 IEEE

Similarly, using
can be derived as

and
(2), the

likeli-

(4)

868

IEEE SIGNAL PROCESSING LETTERS, VOL. 17, NO. 10, OCTOBER 2010

where
. In (3) and (4), the primary users signal
is unknown to the cognitive radio and
therefore, it needs to be removed from the likelihood function.
This can be done by using the generalized likelihood ratio test
[8] together with the maximum likelihood estimate of
to replace
in (3) and (4). This gives
(5)
where
and the relationship
has been used. The values of
and
in (5)
are also unknown and random. They need to be averaged
before the final likelihood ratio test. In [7], the averaging is
performed directly over the likelihood ratio in (5) that contains
the exponential functions. For the system considered in [7],
this method gives a tractable decision variable as a sum of
lognormal random variables in [[7, eq. (4)]]. If the same method
is used for the system considered in this letter, one has the
decision variable as either
or

,
where
and
are the probabilities that the primary
user disappears or appears at
and
, respectively. The
distributions of these decision variables are unknown, and
are difficult to obtain. Consequently, the detection threshold
cannot be determined. To overcome this difficulty, one has
to simplify the method in [7] by first taking the logarithm of
both sides of (5) to give
(6)

where is the detection threshold to be determined. Then, the


values of
and
are averaged in (6) over their distributions. This method is suboptimal but tractable, compared with
the method in [7].
In this letter, assume that the arrival or the departure of the
primary user follows a Poisson process. Denote
as the arrival rate and
as the departure rate of the primary user. The
probability that the primary user doesnt arrive or depart during
and
, respecthe sample interval is given by
tively, and the probability that the primary user arrives or departs during the sample interval is given by
and
, respectively. Then, the probability that the primary
user departs or arrives in the -th sample or the -th sample
is given by
and
, respectively. Three different cases are
discussed.
1) Conventional: When
and
such that the
low traffic of the primary user doesnt affect sensing, one has
(7)
giving the conventional energy detector derived in [1][3],
where
is the detection threshold to be determined. In this
case, the primary user is either present or absent for the whole
sensing period. The random variable
follows a
gamma distribution. Using the Neyman-Pearson criterion, the

detection threshold for a predetermined probability of false


alarm can be derived from (7)
(8)
where

is

the

inverse

function of
representing the cumulative distribution function (CDF) of a gamma
random variable with shape parameter and scale parameter .
One may also use a Gaussian approximation to the probability
density function of
by matching its mean and
variance with those of a Gaussian random variable to give an
approximate detection threshold as
(9)
where

is the inverse function of


representing the Gaussian error function and tabulated values of
can be found in many
statistical handbooks [9].
, the primary user poten2) Random Arrival: When
tially starts transmission in the sensing period due to its medium
to high traffic, and the likelihood ratio in (6) depends on the
is exponentially distributed, one has
random variable . If
the averaged likelihood ratio as

(10)
which is the improved energy detector when the arrival time of
the primary user follows an exponential distribution during the
is the detection threshold for the imsensing period. In (10),
proved detector with random arrival to be determined. One sees
that the received samples in (10) are exponentially weighted by
their arrival time, as expected, as the late-arrival samples are
more reliable in determining the presence of the primary user.
is difficult to
The exact distribution of
obtain. Using the gamma approximation,
can be approximated as gamma distributed by matching its mean
and variance with those of a gamma random variable. In this
case, the detection threshold for the NeymanPearson criterion
can be determined from (10) as
(11)
where
tion of

the

gamma

is
the
inverse
CDF defined as

funcbefore,
and

.
Alternatively,
can be approximated as Gaussian
distributed by matching its mean and variance with those of a
Gaussian random variable. In this case, the detection threshold
can be estimated from (10) as
(12)
, the primary user po3) Random Departure: When
tentially terminates transmission during the sensing period, and
the likelihood ratio in (6) depends on the random variable .

BEAULIEU AND CHEN: IMPROVED ENERGY DETECTORS FOR COGNITIVE RADIOS WITH RANDOMLY ARRIVING OR DEPARTING PRIMARY USERS

Fig. 1. ROC curves for the improved energy detector and the conventional
(cross line),
energy detector using the gamma approximation for
(plus line) and
(circle line) when the primary user (a)
dB.
departs or (b) arrives randomly during the sensing period at

N = 200

N = 400

N = 100
SNR = 05

Again, if
follows an exponential distribution, the averaged
likelihood ratio is

(13)
which is the improved energy detector when the departure time
of the primary user is exponentially distributed over the sensing
period. In (13),
is the detection threshold for the improved
detector with random departure to be determined. One sees that
(13) has the same form as (10). Since one cannot tell if the primary user is randomly arriving or departing in practice,
has
to be set equal to
in (11) or
in (12), depending on
which approximation is used. In the following, we compare the
performance of the conventional detector in (7) with the performance of the new detector in (10).
III. NUMERICAL RESULTS AND DISCUSSION
In this section, the performance of the improved energy detector in (10) is compared with that of the conventional energy detector in (7). In the gamma approximation, the comparison is made by deriving the receiver operating characteristic (ROC) curve for the detector in (7) with
in (8) and
the ROC curve for the detector in (10) with
in (11). In
the Gaussian approximation, the comparison is made by deriving the ROC curve for the detector in (7) with
in (9)
and the ROC curve for the detector in (10) with
in (12).
Denote
as the probability of detection and
as the probability of false alarm. Define the average signal-to-noise ratio
(SNR) as
. Also,
for
. This effectively assumes a rectangular pulse
for the primary user signal. Other pulse shapes, such as the

869

Fig. 2. ROC curves for the improved energy detector and the conventional energy detector using the gamma approximation for
dB (cross line),
dB (plus line) and
dB (circle line) when the primary
user (a) departs or (b) arrives randomly during the sensing period for
.

SNR = 05

SNR = 010

SNR = 0

N = 200

raised-cosine pulse, can be examined in a similar way. In a practical system, the primary user traffic has different intensities.
The primary user signal is filtered at the cognitive radio by a
bandpass filter with bandwidth , and the filtered signal is sampled at an interval of , giving
samples for energy detection. The measurement bandwidth is normally set equal to
the bandwidth of the licensed channel. The sensing performance
will change with
and the type of bandpass filter used. In the
simulation, the values of and are generated using random
number generators to examine the case when the primary network has medium to high traffic, as and would be fixed to
zero for low primary user traffic. However, the simulation does
not generate a primary network with medium to high traffic directly. Also, the simulation generates the samples after filtering
and sampling and therefore, the filtering and sampling methods
arent specified. The values of SNR examined are
dB,
dB and
dB, as the received primary
user signal at the cognitive radio is often weak. In Figs. 14,
, while in Fig. 5, they are specified otherwise.
Fig. 1 compares the improved energy detector and the conventional energy detector using the gamma approximation for
when the primary user randomly arrives
different values of
or departs. Several observations can be made. First, the probability of detection increases as the probability of false alarm
increases. Thus, one can achieve a larger probability of detection for larger values of probability of false alarm, as expected.
Second, the new energy detector outperforms the conventional
energy detector in all the cases considered. The performance
gain decreases when increases. The reason for this becomes
clear by comparing (10) with (7). As the number of samples
increases, the term
approaches 1 and an increasing
number (proportion) of the samples in (10) have weight close
to 1 which is the weight for all the samples in (7). So, the two detectors become more similar as increases. Third, comparing
Fig. 1(a) with Fig. 1(b), one sees that the performance gain
doesnt depend strongly on whether the primary user is randomly departing or arriving.

870

IEEE SIGNAL PROCESSING LETTERS, VOL. 17, NO. 10, OCTOBER 2010

Fig. 3. ROC curves for the improved energy detector and the conventional en(cross line),
ergy detector using the Gaussian approximation for
(plus line) and
(circle line) when the primary user (a)
dB.
departs or (b) arrives randomly during the sensing period at

N = 200

N = 400

N = 100
SNR = 05

Fig. 5. ROC curves for the improved energy detector and the conventional energy detector using the gamma approximation for
dB and
or (b) arwhen the primary user (a) departs with different  T and  T
randomly during the sensing period.
rives with different  T and  T

SNR = 05

=1

N = 100
=1

Fig. 4 examines the effect of noise uncertainty on the conventional energy detector and the new energy detector. As expected,
, apthe probability of missed detection, which equals
proaches an upper limit when the SNR becomes very small.
From this figure, since the probability of missed detection for
the new detector approaches the upper limit at almost the same
rate as the probability of missed detection for the conventional
detector, one sees that the new detector improves the detection performance but it does not improve the detection performance under noise uncertainty. Fig. 5 compares the conventional energy detector with the improved energy detector for difor
. One sees that the improved energy
ferent values of
detector outperforms the conventional energy detector for all
values of
and
examined, including when
.
REFERENCES
Fig. 4. Probability of missed detection for the improved energy detector and
the conventional energy detector using the gamma approximation for noise uncertainty 1 dB (cross line) and noise uncertainty 3 dB (circle line) when the
primary user (a) departs or (b) arrives randomly during the sensing period for
.

N = 100

Fig. 2 compares the two detectors using the gamma approximation for different SNR. Similar observations can be made.
Again, the probility of detection increases when the the probability of false alarm increases, and the performance gain of the
improved energy detector for randomly departing users and randomly arriving primary users is similar. Moreover, the performance gain increases when the value of SNR decreases.
Fig. 3 compares the improved energy detector and the conventional energy detector using the Gaussian approximation for
different values of . Similar observations to those drawn from
Fig. 1 can be made for Fig. 3. One also sees that the detectors using thresholds from the Gaussian approximation outperform those using thresholds from the gamma approximation.

[1] H. Urkowitz, Energy detection of unknown deterministic signals,


Proc. IEEE, vol. 55, pp. 523531, Apr. 1967.
[2] V. I. Kostylev, Energy detection of a signal with random amplitude,
in Proc. ICC 2002, New York, May 2002, pp. 16061610.
[3] F. F. Digham, M.-S. Alouini, and M. K. Simon, On the energy detection of unknown signals over fading channels, IEEE Trans. Commun.,
vol. 55, pp. 2124, Jan. 2007.
[4] D. Cabric, S. M. Mishra, and R. W. Brodersen, Implementation issues
in spectrum sensing for cognitive radios, in Proc. IEEE Asilomar Conf.
Signals, Systems and Computers 2004, Pacific Grove, CA, Nov. 2004,
pp. 772776.
[5] T. S. Shehata and M. El-Tanany, A novel adaptive structure of the energy detector applied to cognitive radio networks, in Proc. Can. Workshop on Information Theory 2009, May 2009, pp. 9598.
[6] T. Wang, Y. Chen, E. L. Hines, and B. Zhao, Analysis of effect of
primary user traffic on spectrum sensing performance, in Proc. Chinacom 2009, Aug. 2009, pp. 15.
[7] N. C. Beaulieu, W. L. Hopkins, and P. J. McLane, Interception of
frequency-hopped spread-spectrum signals, IEEE J. Selec. Areas in
Commun., vol. 8, pp. 853870, June 1990.
[8] S. M. Kay, Fundamentals of Statistical Signal Processing: Detection
Theory. Upper Saddle River, NJ: Prentice-Hall, 1998.
[9] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables. New York:
Dover, 1972.

También podría gustarte