You are on page 1of 5

2014 International Conference on Signal Processing and Integrated Networks (SPIN)

Effect of Dynamic Threshold & Noise Uncertainty


in Energy Detection Spectrum Sensing Technique
for Cognitive Radio Systems
Kriti Chabbra1 P Banerjee3
Garima Mahendru2 Senior Member IEEE
Amity School of Engineering and Technology Amity School of Engineering and Technology
Amity University- Noida, India Amity University- Noida, India
1 3
kriti_chhabra2006@yahoo.com pbanerjee@amity.edu
2
gmahendru@amity.edu

Abstract—Cognitive Radio has arrested the attention of detector is the energy of the input signal over a certain time
researchers for the simple reason of optimum utilization of interval [6, 7].
available spectrum of wireless communication in the recent
scenario. The spectrum sensing by energy detection is one of the The presence or absence of a primary user is usually
techniques for detecting spectrum holes. This paper revisits the formulated by the probability statistics of detection.
existing energy detection technique. The influence of dynamic
Let us assume that Ns is total number of samples during
threshold and noise uncertainty factor on performance
parameters of a communication system namely: probability of
sensing time. The average power of the (input) received signal
detection and probability of false alarm is studied and an attempt R(n) is known and noise is White Gaussian in nature with
has been made to deduce significance of their ratio on the sensing variance as ın2 . The test statistic of the energy detector is
technique. An empirical relationship between the sampling considered as:-
number and SNR is also proposed based on the computations 1
¦ s R(n)R(n)
N −1
D(R) = (1)
and tabulation using simulated results through MATLAB.
N s n =0
Keywords— Cognitive radio, spectrum sensing, energy
detection, probability of detection, probability of false alarm, where D(R) refers to decision variable and Ȗ is decision
sampling number, dynamic threshold variable threshold [7]. The decision variable or the test
statistic is compared to a threshold to detect the presence or
I. INTRODUCTION absence of the primary signal. The performance of all
Detection of available frequency spectrum or “spectrum spectrum sensing techniques can be analysed and indicated by
sensing” is of utmost importance and supposedly the first step two probabilities i.e. the probability of detection (PD) and the
towards Cognitive Radio Implementation [1-3]. There are probability of false alarm (PFA).
number of spectrum sensing techniques that examine the
availability of the frequency spectrum. Among them, energy Here, Probability of Detection (PD) is defined as the probability
detection technique involves estimation of energy of the that the sensing algorithm has successfully detected the
received signal at Cognitive Radio Receiver. The estimated presence of the primary signal and Probability of False Alarm
sensed energy is then compared with a set threshold to signify (PFA) is the probability that the sensing algorithm asserts to
the presence or absence of primary signal [3, 5]. detect the presence of the primary signal even when the
This paper reviews the background work of the existing primary signal is absent but falsely driven by noise.
energy detection techniques for spectrum sensing. The effect of
dynamic threshold and noise uncertainty on the performance of The expressions for the performance analysis parameters
detection technique [6] has been investigated through from central limit theorems are given [6] as:-
MATLAB simulation. Further, an attempt has been made to
formulate the empirical relationship between sampling number § ·
¨ ¸
and SNR for energy detection technique. ¨ Ȗ − ı 2n ¸  (2)
PFA =Q¨ ¸
¨ 2 ı2 ¸
¨ N n ¸
II. SYSTEM MODEL © S ¹
Normally in most of communication systems, the noise is
assumed to be Gaussian with zero mean, and has band limited
power spectrum density [6, 7]. The output of the energy

978-1-4799-2866-8/14/$31.00 ©2014 IEEE 377


2014 International Conference on Signal Processing and Integrated Networks (SPIN)

§ ·
¨ ¸
(
¨ Ȗ − P + ı 2n ¸
PD = Q ¨
) (3) 
C. Noise Uncertainty& Dynamic Threshold
¸
¨ 2 (P + ı 2 ) ¸
¨ N n ¸
© S ¹ In above cases, we have taken noise uncertainty and dynamic
where Q (.) is the Standard Gaussian Complementary threshold factor separately. Let us combine both these factors
Cumulative Distribution Function (CDF) and P is the average for finding probability of detection. The performance
signal power. Computing Ȗ threshold in terms of PFA & PD parameters PD & PFA are computed by modifying equation (2)
from (2), (3) we get, & (3). In the expression for PD of (3), the factor (P+ ın2) is
replaced by (P+ ın2/ ȡ) & Ȗ as Ȗ/ȡ’. For PFA of (4), ın2 is
replaced as ȡın2 & Ȗ as ȡ’Ȗ. The sampling number or the
2 2
Ȗ = Q −1 (PFA ) ı 2n + ı 2n = Q −1 (PD )(P + ı 2n ) + ( P + ı 2n ) sensing duration is now expressed in terms of ȡ’ (dynamic
Ns Ns threshold) & ȡ (noise uncertainty) as:
(4)
Ns = 2 [(ȡ/ ȡ') Q-1 (PFA) – ȡ'(1/ ȡ + SNR) Q-1 (PD)] 2 [ȡ'SNR +
-1
where Q (.) is the inverse standard Gaussian
ȡ'/ ȡ – ȡ/ ȡ')]-2 (8)
Complementary CDF. Ns is expressed in terms of PFA, PD &
SNR, Probability of detection (PD) and probability of false alarm
(PFA) are related as given by
[
N S = 2 Q −1 (PFA ) − Q −1 (PD )(1 + SNR) / SNR 2 ]
2
(5) ª ρ −1
« Q (PFA ) -
Ns ρ' ρ º
(ȡ' SNR + − ) »
ρ ' 2 ȡ ρ' »
PD = Q «
SNR = P/ın2 is the signal-to-noise ratio. It may note that the « 1 »
« ρ ' (SNR + ) »
expression of Ns in (5) is free from the variable Ȗ (decision ¬ ȡ ¼ (9)
threshold).
A. Noise Uncertainty It is observed that (8) & (7) incorporate the ratio of dynamic
threshold factor (ȡ') to noise uncertainty factor (ȡ) also.

Assuming that noise uncertainty does not exists, poses a


limitation on the performance of the given system. Thus, we III. SIMULATION RESULTS
need to incorporate noise uncertainty factor in the system for
reliable performance [6]. Let us introduce in the noise model a Let us investigate carefully above formulations through
factor ȡ as the noise uncertainty coefficient such that ȡ •1 (ȡ=1 MATLAB simulations to have deeper illustrative insight into
when there is no noise uncertainty implying no variation of different aspects of spectrum sensing technique.
noise power). So limits of noise variance will lie between
[ 2qƒ2 /ƒandƒ 2
qƒƒ`1ƒFor obvious reason the ın
2
in (2) and A. Without Noise uncertainty
(3) would be replaced by these limiting values respectively. The probability of false alarm PFA ‫( א‬0, 0.9) has been set to a
The average signal-to-noise ratio defined as SNR=P/ın2. Thus minimum value of 0.01 which may be accepted as negligible.
(5) gets modified to:
[
N = 2 ρQ
s
−1
( PFA ) − (1 / ρ + SNR )Q
−1
]2
( PD ) ( SNR − ( ρ − 1) / ρ ))
−2

(6)
Energy detection algorithm which involves fixed threshold
offers degraded performance with noise uncertainty. This
indicates that the dynamic threshold would yield better
performance in presence of noise uncertainty [8, 9].

B. Dynamic Threshold
The dynamic threshold factor ȡ’ (ȡ ‘ • 1) has been introduced
in such a way that threshold Ȗ lie in the interval [Ȗ/ȡ’, ȡ’Ȗ],
instead of remaining constant [6]. In a way similar to the case
of noise uncertainty showed in the previous section, Ns may
expressed as Fig. 1 Variation of Probability of Detection (PD) with SNR for different values
of Number of samples
[ ][ 2
N s = 2 Q −1 ( PFA ) − ρ '2 (1 + SNR )Q −1 ( PD ) ρ '2 SNR + ( ρ '2 − 1) ]−2

(7)

378
2014 International Conference on Signal Processing and Integrated Networks (SPIN)

Thus, neglecting the effect of noise uncertainty (i.e. ȡ = 1) four Let us now try to derive an empirical relation between SNR
different sample points with values Ns =10, Ns =100, Ns =2000 and Ns, through a linear regression technique. The equation is
and Ns =5000 over SNR range of -20 to 5 dB are considered. of the form f(x) = p1x + p2. Table 1 shows the polynomial
With above values, the graph of probability of detection vs. coefficients for PD = 0.5, 0.7 and 0.9 in descending powers.
SNR has been plotted through MATLAB. As discussed in
Section II (A), the energy detector model without noise TABLE I. LIEAR FIT COOEFFICIENTS FOR PROBABILITY OF DETECTION
uncertainty yields a performance as depicted in Fig. 1, the (PD) = 0.5, 0.7, 0.9
simulated graph of SNR values on X-axis and probability of
Coefficients of Linear fit
detection on Y-axis for different number of sample points Ns PD
(Ns=10, Ns=100, Ns=2000 and Ns=5000). p1 p2
It is observed in Fig. 1 that for a fixed value of SNR, the 0.5 -0.501 5.18
probability of detection is improved by increasing the number
0.7 -0.5165 6.749
of samples. It is also, observed in Fig. 1 that i.) At very low
values of Ns (small sampling number), the PD does not 0.9 -0.5249 8.201
increases significantly even with increasing SNR. ii.) As Ns
increases, there is a significant increase in PD even at low
values of SNR. Thus, the optimum detection at low SNR
values would require higher sampling number or sensing Figs. 3 shows typical the linear fit and the corresponding
duration for a case without noise uncertainty factor. original data points for PD = 0.5 with dots while the solid line
represents the computed data regression line.
Now, from Fig.1 we attempt to find a correlation between
the SNR and number of samples. For this purpose, we generate
many more number of graphs (not shown in the Fig.1 for
brevity) corresponding to different sampling number ranging
from 100 to 20000 with varying interval. So, that one gets
more data corresponding to a fixed value of PD.

With the help of the resulting graph, different values of SNR


for all sample numbers are noted at a particular value of PD say
0.5. Similar procedures are carried out for PD = 0.7 and 0.9
also. In this way, we tabulated different SNR values for
Fig.3 Polynomial fit of the graph shown in Fig. 3 for the value of PD = 0.5
different number of samples (Ns) at a given PD. We first started
with a reasonable value of probability of detection as 0.5, then
By careful inspection of the coefficients of the Table 1 the
0.7 and finally the maximum value of probability of detection
following empirical relation may be approximated
is considered i.e. 0.9 which is fairly a good value for confident
detection
1
log (SNR) = − log N s + PD (10)
2
Thus, (5) may be for all practical purposes may be replaced by
more simple relation of (10). Therefore, we can easily decide
the number of samples that are required to achieve a desirable
PD at some SNR values.

Fig. 4 shows the curves which provides the information


about PD vs. PFA for different SNR values and Ns=1000. From
the graph it is inferred that i.) For SNR values at -20dB, the
probability of detection becomes almost equal to probability
of false alarm which depicts that energy detector becomes
Fig. 2 Variation of SNR with Number of samples for different values of non-robust under low SNR. ii.) The performance becomes
Probability of Detection (PD) = 0.5, 0.7 and 0.9 better at higher SNR values as the PD increases.

B. With Noise Uncertainty


By using the obtained values of SNR and Ns, Fig. 2 is
generated showing the variation of SNR values on Y-axis and The model of energy detetcor in Section II (B) considers the
number of samples (Ns) on X-axis. SNR and Ns are in dB unit. noise uncertainty factor ȡ. Fig. 5 illustrates the corresponding

379
2014 International Conference on Signal Processing and Integrated Networks (SPIN)

performance. Here, Ns=1000 at varying SNR values and C. With Dynamic Threshold
ȡ=1.12. The simulated results clearly depicts that the
performance of sensing technique is very sensitive to noise
Enhancement of the performance of a energy detector
uncertainty.
technique by introduction of concept of dynamic threshold (ȡ')
has been illustrated in Fig. 7 for Ns=1000. It is observed that
i.) when ȡ' =1, (fixed threshold), PD is low and increases only
at higher SNR values. ii.) The probability of detection is
greatly improved as the value of ȡ' is increased even at low
SNR. Introduction of dynamic threshold helps in improvement
of performance.
.

Fig. 4 Probability of Detection Vs Probability of False Alarm for different


values of SNR

It is interpreted from the graph that i.) No matter Ns is high or


low, there is a specific range of values of SNR beyond which
the detetcor is able to detetct the presence of a signal, termed
as SNR wall. One may also comprehend that the value of SNR Fig .6 Probability of Detection Vs Probability of false alarm for different
wall increase with the value of uncertainty factor ȡ. ii.) As Ns values of noise uncertainty ȡ
increases there is an increase in PD even for a small increase in
SNR. Thus, for optimum detetcion Ns and SNR both should be
It can be concluded from the simulated graph that even when
high.
Ns is too high, the performance is unacceptable for a small
For the same detector model, Fig. 6 provides information
change in ȡ. This indicates that energy detection is very
about PD vs. PFA for SNR value equals to -10dB for Ns= 5000,
susceptible to noise uncertainty.
for different values of noise uncertainty factor (ȡ).
It is interesting to note from the Fig. 6, that for ȡ upto 1.03, the
Similarly, Fig. 8 shows the curves which provide information
performance is acceptable. But when for ȡ = 1.05 or more the
about PD vs. PFA for SNR =-10dB and Ns=1000, & ȡ'=1
performance is not at all acceptable, PFA equaling or exceeding
corresponds to the case with fixed threshold. Thus, it is clearly
PD . This indicates that energy detection is very sensitive to
seen that the system performance is significantly improves
noise uncertainty.
with increasing values of dynamic threshold ȡ'.

D. With Noise Uncertainty & Dynamic Threshold

Let us look into effect of noise uncertainty factor and dynamic


threshold taken together into consideration for energy
detection. The simulated results are presented in Fig. 9.

It is observed that for small values of ȡ' and ȡ ࠵ (1.005, 1.09)


the relative difference between the curves is very less. But as
the value of ȡ' and ȡ further increases, we observe significant
changes. It is important to note that when the ratio of ȡ' and ȡ
ȡ' and ȡ is one, the absolute values the respective parameters
may not be important but for other ratios, the absolute values
of ȡ' and ȡ are significant.
Fig.5 Probability of Detection Vs SNR with noise uncertainty .

380
2014 International Conference on Signal Processing and Integrated Networks (SPIN)

IV. CONCLUSIONS

In this paper, the effect of two factors noise uncertainty &


dynamic threshold on the performance parameters of the
energy detection technique i.e. probability of detection and
probability of false alarm has been analyzed. It is concluded
that the absolute values of dynamic threshold and noise
uncertainty are significant and not only their ratio. A simple
useful empirical relation between SNR to number of samples
could be derived by a linear regression fit.

REFERENCES

[1] 802.22 Working Group, IEEE P802.22/D0.1 Draft


Fig. 7 Probability of Detection Vs SNR for different values of dynamic Standard for Wireless Regional Area Networks.
threshold ȡ’ http://grouper.ieee.org/groups/802/22/, May 2006.
[2] FCC, “Facilitating opportunities for flexible, efficient, and
reliable spectrum use employing cognitive radio technologies,
notice of proposed rleu making and order,” in FCC 03-322,
Dec.2003.
[3] T. Yucek and H. Arslan, “A Survey of Spectrum Sensing
Algorithm for Cognitive Radio Applications,” IEEE
Communication Survey & Tutorials, Vol. 11, No. 1, Spring
2009, pp. 116-130.
[4] Ying-Chang Liang; Kwang-Cheng Chen; Li, G.Y.;
Mahonen, P., "Cognitive Radio Networking and
Communications: An Overview," Vehicular Technology, IEEE
Transactions on , vol.60, no.7, pp.3386,3407, Sept. 2011
[5] Ekram Hossain, Long B. Le, Natasha Devroye, and Mai
Vu, “Cognitive radio: From theory to practical network
engineering,” invited chapter in Advances in Wireless
Communications, (Editors V. Tarokh and I. Blake), Springer,
Fig. 8 Probability of Detection Vs false alarm for different values of 2009.
dynamic threshold ȡ’ [6] Guicai Yu, Chengzhi Long and Mantian Xiang, “ A Novel
Spectrum Detection Scheme Based on Dynamic Threshold in
Cognitive Radio Systems”. Research Journal of Applied
Sciences, Engineering and Technology, 4(21): 4245-4251,
2012.
[7] Urkowitz, Harry, "Energy detection of unknown
deterministic signals," Proceedings of the IEEE , vol.55, no.4,
doi:10.1109/PROC.1967.5573
[8] S.M.Kay, Fundamentals of Statistical Signal Processing:
Detection Theory. Englewood Cliffs: Prentice-Hall, 1998, vol.
2.
[9] R. Tandra, A Sahai, “SNR Walls for Signal Detection”.
IEEE Journal of Selected Topics in Signal Processing, 2008,
Vol. 2, No. 1. pp: 4 – 17.

Fig. 9 Probability of Detection (PD) Vs Probability of False Alarm (PFA) with


different values of ȡ' and ȡ.

381

You might also like