Sparse Coding Based Spectrum Sensing in Presence

of Multiple Frequency Hopping Primary Users

Kukil Khanikar, Rohit Sinha, Ratnajit Bhattacharjee
Department of Electronics and Electrical Engineering
Indian Institute of Technology Guwahati
Guwahati-781039, India
Email: {kukil, rsinha, ratnajit}@iitg.ernet.in

AbstractA novel sparse coding based wideband spectrum additional challenges. Since FHSS PUs remain in a particular
sensing technique for multiple frequency hopping spread spec- sub-band for very short period of time, the detection need to
trum primary users has been proposed. The proposed method be done within the short dwell time with sufficient time being
relies on the sparsity of the received target signal on a learned
exemplar dictionary. The learned exemplar dictionary is derived left for transmission by a SU. Detection of a single FHSS PU
from training data comprising of undistorted primary user has been dealt with in [15] and [16]. In [15], an FFT based
transmissions through all the channels in the band of interest algorithm referred to as FFT based averaging ratio (FAR) [17]
using K-SVD algorithm. The high sampling requirement for is adopted. In [16], a sparse representation classification based
the wideband sensing is addressed by using a set of lower rate approach was adopted for detection of single FHSS PU. In
samplers, each sampling the received baseband signal at different
time offsets. The proposed method maintains constant false alarm contrast to these works, this work deals with the detection of
rate under various SNR conditions and when the number of active multiple unknown number of FHSS PUs present in the BOI.
primary users in the band is varying. On comparing with an The effect of transmission channel was not accounted for in
existing FFT based wideband sensor, the detection performances [16] which is also addressed in this work.
of the proposed method for varying SNR and multiple active
The rest of this paper is organized as follows. Section II
primary users are noted to be promising.
describes the scenario under which spectrum sensing is carried
I. I NTRODUCTION out and the assumed signal model. Section III elaborates on
The apparent scarcity of radio spectrum [1] has led to the sparse representation based detection technique proposed
development of techniques for opportunistic spectrum access in this work to deal with presence of multiple PUs. Section IV
(OSA) for wireless communication. Spectrum sensing enables discusses the simulation setup and the results. Finally, Sec-
secondary user (SU) or cognitive radio [2] to utilize the radio tion V concludes the paper.
spectrum allocated to licensed or primary user (PU). Several
spectrum sensing techniques based on energy [3][6], cyclo- II. O PERATING S CENARIO AND S IGNAL M ODEL
stationary property [7], [8], eigenvalue decomposition [9]
and compressive sensing [10] have been proposed till date. First we describe the broad setting under which the spectrum
Wideband spectrum sensing offers possibility of enhancing sensing task is carried out by an SU. We consider a non-
the spectrum utilization. However, there are typical challenges cooperative scenario in which the SU employs a spectrum
in achieving this goal. Wideband sensing can be performed sensor to assist in opportunistic transmission through a radio
using multiple narrowband radio frequency (RF) front-ends frequency band licensed to a set of FHSS PUs.
to sense individual sub-bands or a single such front-end to The following assumptions are made regarding the spectrum
perform sequential sensing. Alternatively, the entire wideband sensing scenario.
or the band of interest (BOI) can be scanned using a single A1) There are number of non-overlapping frequency chan-
wideband RF front-end. The latter is advantageous from the nels in the BOI.
point that it can simultaneously scan the entire BOI using A2) A maximum of number of frequency hopping PUs
lesser RF components. The challenge however comes in the operate in the BOI and .
form of requirement of very high sampling rates. A3) The baseband bit stream transmitted by PUs are mod-
In [11], a bank of bandpass filters has been used to achieve ulated using a digital modulation scheme such as am-
the spectrum sensing task in multiple sub-bands. For joint plitude shift keying (ASK), phase shift keying (PSK) or
wideband sensing, FFT and wavelet transform have been used frequency shift keying (FSK).
in [12] and [13], respectively. Spectrum sensing proposed in A4) The active PUs employ orthogonal hopping sequences
[14] demonstrates the application of compressive sampling for and their hopping instants are synchronized so that there
wideband spectrum sensing. is no interference among themselves.
The detection of frequency hopping spread spectrum A5) The hop periods of the PUs are equal and constant. The
(FHSS) signals for OSA is a wideband sensing problem with SU has the complete knowledge about the hop period of
978-1-4799-6619-6/15/$31.00
c 2015 IEEE the PUs.
A6) The number of active PU transmitters change with time frequency sinusoid with frequency between the start and end
in a random manner. limits of the BOI. Training signals comprise of associated I
A7) The information regarding the number of FHSS PUs and Q waveforms corresponding to each channel such that on
active at any instant of time and the frequency channels modulating with the same sinusoid used in demodulation of the
they occupy are not available to the SU. received signal, their spectrum would fall in the corresponding
A8) The channel information is not available to the SU. RF frequency channel.
The signals transmitted by the PUs suffer multipath fading Each sub-dictionary is learned using training signals
and are corrupted by additive white Gaussian noise (AWGN) from the channel and K-SVD [18] algorithm. Each sub-
at the receiver end. The signal received at the SU detector can dictionary has atoms. Additional details of dictionary
be expressed as learning are provided in Section IV. The learned exemplar
dictionary is denoted by and is formed by horizontal
concatenation of sub-dictionaries s corresponding to each
() = () () + () (1) of the channels. The dictionary has a total of
atoms and is defined as
where,

({1, 2, . . . , }) and (.) denotes a power set, () = [ 1 2 . . . ] (2)

is the transmitted signal by the PU, () is the multipath
channel between the PU and the SU, () is AWGN, and
It is to be noted that dictionary learning is an offline process
denotes the convolution operator.
so the complexity associated with K-SVD algorithm has no
III. D ETECTION OF PU TRANSMISSIONS USING S PARSE affect during sensing. An additional fact to be noted here is
C ODING that during the learning phase only the characteristics of the
primary user signals are captured rather than the channel char-
In this paper, we propose a detection technique enabled by acteristics. The learned characteristics (in the form of atoms of
sparsity of the received signal. The BOI is assumed to be the dictionary) are then employed during the detection phase.
sparsely occupied by the active PUs. This assumption follows As a result the detector is blind to the channel characteristics.
from the fact that the spectrum remains under-utilized and the
communication opportunities remain hidden in the concerned
RF band. In our earlier work [16], we have proposed a learned B. Detection of PU Transmissions
dictionary based approach for spectrum sensing using sparse
In contrast to the work in [16], in this work multiple PUs are
coding. Motivated by that, the same approach is also followed
considered in the BOI. As a result of this the training signals
in this work. The columns of the dictionary matrix are called
are required to be sampled at a much higher rate. In order to
atoms and a group of atoms are associated with a class which
avoid use of faster analog-to-digital converters (ADCs) at the
in this case refer to a channel. Thus there are classes
detector front-end we adopt a sub-sampling approach while
in the dictionary corresponding to the number of channels.
sampling the received target signal.
The detection task involves recovering a sparse representation
The sub-sampling factor is given as
of the received target signal over such a dictionary which
gives an indication about the channels that are free, i.e., not
occupied by the PUs. The description of how such a dictionary = (3)

is constructed is given in the following subsection.
where, is the sampling frequency of the training signals,
A. Dictionary Learning
and is the rate at which the ADCs sample the received
As explained in [16], first sub-dictionaries are learned using signal.
K-SVD [18] algorithm on the training signals for each of
the channels. These channel-specific learned dictionaries Instead of employing a single pair of uniform samplers at
are then concatenated to form a learned exemplar dictionary the I-Q sub-branches, we have employed ( ) pairs of
(LED) which is used for the sparse coding. In this work we uniform sampling ADCs to obtain the sampled versions of
consider that the baseband signalling scheme of all FHSS PUs the I-Q sub-branches at different offsets. (.) denotes the
is identical, i.e., the baseband digital modulation scheme, baud ceiling function. Each ADC pair installed in a branch samples
rate and pulse shaping filter adopted by all PUs are same at ( 1 ) rate at which the training signals are sampled and
and follow some common standard. Training data consists of all ADCs are time synchronized. In the discrete domain, the
distortionless signals, i.e., in absence of AWGN and multipath signal received at each pair of ADC branch may be expressed
fading. More details about how these training signals are as the projection of the higher sampled vector on a sub-
generated are discussed in Section IV. At the SU receiver, the sampling matrix. For instance, if a real signal of 1 sec duration
received wideband RF signal encompassing the entire BOI is originally sampled at 10 Hz is sub-sampled by factor of 2, the
I-Q demodulated and translated to the baseband using a high corresponding sub-sampling matrices would be
 C. Implementation Complexity of the Proposed Detector
1 0 0 0 0 0 0 0 0 0
The proposed sparse coding based detector implements the
0 0 1 0 0 0 0 0 0 0
sensing operation in two stages:
0 0 0 0 1 0 0 0 0 0
0 0 0 0 0 0 1 0 0 0 a) Computation of sparse coefficients of the received signal
0 0 0 0 0 0 0 0 1 0 at each of the -branches using OMP algorithm.
b) Combination of decisions from all branches as described
and by Equation 8 to determine channel occupancy status.
0 1 0 0 0 0 0 0 0 0 The time complexity of computing the sparse coefficients
0 0 0 1 0 0 0 0 0 0
at each of the parallel branches using OMP is approximately
0 0 0 0 0 1 0 0 0 0
2 + 3 2 . Here, denotes number of samples
0 0 0 0 0 0 0 1 0 0
obtained at the output of an ADC pair combined during a
0 0 0 0 0 0 0 0 0 1 sensing period and denotes the chosen sparsity constraint
Let denote the sub-sampling matrix corresponding to the value. This includes
pair of ADC sub-branch. If represents the received signal i) (2 1) number of floating point operations (flops)
vector sampled at rate and represents the 2 normalized involved in the matrix-vector multiplication operation re-
version of given as quired for determining highest correlated dictionary atom
at each iteration,
=
(4) ii) 4 flops required for computing least squares solution
2
and 2 flops required for residual update at the
then the signal vector appearing at the ADC output is given iteration.
as The second stage is common to any detector which may
work in a fixed frequency-slotted spectrum sensing scenario

= (5) considered for this work. This stage involves ( 1) flops
for the branch is derived as follows associated with taking binary decisions for channels using
The dictionary
atoms per channel and sensors.
= ( )
(6) The time complexity of the proposed algorithm is linear
with dictionary size while quadratic with sparsity constraint
where, () denotes the function normalizing the value. It is therefore, necessary to keep the sparsity constraint
columns of the matrix . value to a minimum, while maintaining the detection perfor-
The branch signal is sparse coded over dictionary mance.
using OMP [19] to obtain the sparse vector such that
IV. S IMULATION , R ESULTS AND D ISCUSSION

= (7) For simulation, we have modulated the baseband binary
bit stream using Gaussian frequency shift keying (GFSK)
The PU transmission is declared to be present in the
modulation scheme. The signal closely resembles Bluetooth
channel if,
signal transmitting at basic rate [20]. Baud rate is set at 1 Mega
(y 1 ) > (8) symbols per second or equivalently the symbol period is
set to 1s. The time-bandwidth product of the of the Gaussian
where, filter is 0.5. The modulation index is kept as 0.32. The
bandwidth of each channel is 1 MHz. We have considered 16
1 channels for simulation. Thus, the total bandwidth of interest
y = ( ) (9)
is 16 MHz.
is a 1 size vector containing the original Euclidean The LED is created using training signals from each of 16
lengths of the columns of ( ) before transformation into channels. As mentioned in Section III-A, pure and undistorted
normalized version using Equation 6, (.) is an signals fragments generated using GFSK modulation and spec-
operator that shapes a given vector into a diagonal matrix, ifications stated previously are used for the training. Equally
and denotes the element-wise vector multiplication operator, sliced portions at random instants from a long baseband
denotes a 1 indicator vector selecting the indices transmission are extracted as training specimens. Signals are
corresponding to channel such that sampled at rate and the time length of training signals is
set equal to the desired sensing time of the detector. The
() = 1, for = ( 1) + 1, ..., (10) signals so generated are modulated into 16 different carrier
= 0, otherwise frequencies spaced 1 MHz apart to simulate their presence
in different channels. This operation is done in the baseband
and is a predefined threshold to detect the presence of the instead of the actual RF frequency band. For instance, in order
signal in the observed data. to translate a signal to an RF channel in the 2.4 GHz band, it
 TABLE I
P ERCENTAGE OF DETECTION ( ) AT VARYING SNR AND DIFFERENT NUMBER OF ACTIVE PU S FOR FAR AND LEARNED - EXEMPLAR DICTIONARY
(LED) BASED METHODS WITH SENSING TIME EQUAL TO 5 SYMBOL PERIODS . T HE NUMBER OF ATOMS CHOSEN FOR LED IS 30 16 . F OR SIMULATION ,
THE RECEIVED SIGNAL IS SAMPLED AT 128 MH Z . F OR THE DICTIONARY BASED METHOD A SUB - SAMPLING FACTOR , OF 4 IS APPLIED REDUCING THE
EFFECTIVE SAMPLING RATE TO 32 MH Z .

FAR [17] LED (480 atoms)

SNR (dB)
Number of active PU Transmitters Number of active PU Transmitters
5 4 3 2 1 5 4 3 2 1
0 57.05 63.88 75.99 87.18 98.34 88.92 93.38 95.08 96.58 99.09
-5 56.28 65.05 73.14 84.92 97.90 86.86 89.85 91.24 95.43 97.42
-10 52.60 60.81 70.11 79.33 90.40 79.83 82.30 84.87 87.63 92.94
-15 43.40 49.97 57.56 66.85 75.44 60.50 65.87 67.70 70.19 73.64

TABLE II
P ERCENTAGE OF DETECTION ( ) AT VARYING SNR AND DIFFERENT NUMBER OF ACTIVE PU S FOR FAR AND LED BASED METHODS WITH SENSING
TIME EQUAL TO 2 SYMBOL PERIODS . OTHER PARAMETERS ARE SET SAME AS IN TABLE I.

FAR [17] LED (480 atoms)

SNR (dB)
Number of active PU Transmitters Number of active PU Transmitters
5 4 3 2 1 5 4 3 2 1
0 28.48 39.78 57.99 77.80 97.89 84.17 89.75 91.72 95.58 97.52
-5 26.97 39.41 54.90 71.58 90.58 83.02 83.59 87.05 88.85 90.46
-10 24.08 34.07 47.90 63.39 79.97 69.89 73.90 74.79 75.56 75.77
-15 18.74 25.24 34.54 43.64 52.28 47.59 52.02 53.60 53.71 53.97

would actually require additional modulation using a suitable algorithms based on cyclostationary or eigen analysis based
carrier of high frequency to do so. In our simulation, the features, the FFT based detectors are relatively less complex
carrier frequency that would have been required to translate due to minimal number of computations required and can
to that RF frequency is equal to the start band frequency easily be applied to wideband sensing scenario like the one
(2.4 GHz for instance) added to half the the total bandwidth under consideration. FFT based sensors are basically energy
of interest (8 MHz). This is done in order to bring down detectors and require less sensing time than cyclostationary
the sampling requirements. The converse process in order or eigen value based detectors to arrive at a decision under
to deal with the real RF signal would require the detector moderate SNR conditions. Since fast sensing is necessary
signal to bring down the received signal to baseband using for the scenario involving FHSS primary user, we chose to
I-Q demodulation and the associated carrier of frequency . compare the proposed method with FAR. In contrast to the
All signals in this work are simulated in the baseband. proposed method, FAR utilizes Fourier bases to decompose
The resultant training signals are complex. The real and the received signal into a non-sparse representation. The
imaginary components are stacked in the form of column innovation in FAR algorithm is in its ability to maintain a
vectors such that all the real components lie on the top half constant false alarm rate (CFAR) under varying and unknown
of the vectors followed by the imaginary components. The SNR scenarios. Similar CFAR property is also noted in the
training set for a particular channel consists of such vectors proposed method. The detection threshold need not be altered
from associated training signals stacked in the form of a to account for the changing SNR conditions while retaining
matrix. This training set is used to obtain the sub-dictionary the same noise rejection capabilities. The decision thresholds
for the channel using the K-SVD algorithm. The of both the methods also remains fairly invariant to the number
number of atoms in is which is kept constant for all of active transmitters that are present in the band of interest.
channel sub-dictionaries. The LED is then obtained using Figures 1 and 2 plot the sensitivity of the optimal detection
Equation 2. threshold to varying SNR conditions and number of PU trans-
The received signals are generated using the signal model mitters for the proposed dictionary based and FAR methods
described in Equation 1. ITU Indoor Office-A [21] channel respectively.
model is adopted for the simulations. The signal received at For the LED based method, similar to the training signals,
the detector front-end is an aggregate of transmissions from the received signals are separated into real and imaginary
variable number of PUs relayed through multipath channels components and stacked as column vectors of real values.
and corrupted by AWGN. Varying numbers of active PUs The received signals are 2 normalized as in 4 and effectively
ranging from 0 to 5 are considered to simulate the spectrum down-sampled at each of the ADC branches mentioned
sensing scenario. in Section III-B using the sampling matrices . This is
We have compared our method with an FFT based algorithm equivalent to using slower ADCs at the I and Q branches of
called FAR [17]. Unlike most conventional spectrum sensing each of the sampling branches and obtain a signal sampled at
 effect of inter-channel interference on the performances.
On comparing Table I and Table II we make following
Deviation of Threshold from mean value

0.01 observations:
0.005 In general, the detection performances for LED based
method are superior to that of FAR under the observed
0
SNRs. Except for single PU case, the proposed LED ap-
0.005 proach is outperformed by the FAR method in particularly
0.01 at lower SNR values.
10
5 The performance of both FAR and LED based methods
0
4
10 3 decrease when the sensing period is reduced. The change
SNR 2
20 1 No of Active PUs in FAR method is more drastic than LED based method.
On comparing the performances for the 5 PU case, we
Fig. 1. Figure showing the deviation of the optimal decision threshold
note that the LED method with 2 symbol period sensing
from the mean value for the proposed method. The noise rejection outperforms that of FAR method with 5 symbol period
ratio is maintained at 10% under varying SNR conditions and different sensing.
number of active PUs. The SNR and number of PUs range from -20 The detection performances of LED based method is noted
dB to 5 dB and 1 to 5, respectively. The standard deviation of the
optimal threshold values considering all possible combination of SNR to be significantly better than that of FAR for most of the
conditions and number of active PUs turns out to be 1.7 106 . cases. In our understanding the possible reason for the same
lie in:
a) the use of a data-dependent dictionary rather than the fixed
Fourier bases
Deviation of Threshold from mean value

b) finding the sparse representation of targets rather than

0.3
0.2
orthogonal projections
0.1 Since the dictionary used in LED is derived from the type
0 of signals to be detected using K-SVD, the atoms tend to
0.1 have higher correlation with the target to be detected. Unlike
0.2
generic Fourier basis vectors used in FAR, the atoms of the
10 LED provide a degree of matching gain for a time limited
5
0
4 segment of the target. Due to finite length of observation of
10 3
SNR 2
No of Active PUs
the target the Fourier basis leaks the energy into the adjacent
20 1
channels leading to degradation in performance.
Secondly, the sparse representation adopted in LED based
Fig. 2. Figure showing the deviation of the optimal decision threshold method helps suppress unwanted noise as long as sparsity
from the mean value for the FAR method. Noise rejection ratio, SNR
constraint is appropriately chosen. During sparse coding only
conditions and number of active PUs are same as for the dictionary
based method. The standard deviation of the optimal threshold values the highest coefficients are retained which happen to corre-
considering all possible combination of SNR conditions and number spond to the PU signal(s) when it(they) are actually present.
of active PUs turns out to be 5.6 103 . Since the atoms have significantly higher correlation with PU
signals than they have with noise, sparse coding using such
atoms would yield a limited set of higher valued coefficients
a lower rate than the training signal. It is worth noting that the at even low SNR but only when signal is present. This holds
training signals are generated offline and as such the constrain for some minimum SNR beyond which the noise rejection
on ADC speed does not relate to these signals. capabilities are lost depending on the difference between the
For the FAR detector the sampling rate adopted is same as value of sparsity constraint chosen and actual number of
that of the training signals for the LED based method which transmitters present in the band of interest. The mismatch
means higher speed ADCs are installed at the detector front- between chosen sparsity and number of PU present explains
end. Unlike LED based method, the received vector is not why the performances in the a single PU case drops compared
transformed into a real vector but retained as it is. to the FAR method at very low SNR. The average performance
Table I and Table II give the average detection performances though is noted to be significantly better than that of FAR
for FAR and LED based methods for sensing periods even at lower SNR considering variable number of active PUs
equal to 5 and 2 symbol periods, respectively. The FAR present in the band.
method uses 1024 point FFT and 4 time-frames for averaging. The degradation of detection performances for both the
For the LED based method, the values for sparsity constraint methods due to reduced sensing time is expected. A shorter
are set 10 and 16 for testing and training, respectively. The signal leaves lesser scope for averaging out the white noise
false alarm rate is fixed at 10%. For these set of simulations, thereby decreasing the performance. However, the larger drop
the active channels are placed far apart so that there is minimal in performance of the FAR method can again be attributed to
the phenomenon of frequency leakage. Shorter sensing period raise the detection performance by introducing diversity gain.
leads to greater leakage of energy of the signal of interest into The assumption that all PUs use identical baseband modulation
adjacent channels. The matching gain provided by dictionary scheme may also be relaxed. These tasks are left as future
atoms reduces such kind of leakage for the LED based method. extension of this work.

A. Effects of Inter-channel Interference on Detection Perfor- R EFERENCES

mance [1] FCC, Spectrum Policy Task Force, Report of the Spectrum Efficiency
Working Group, FCC, Tech. Rep. ET Docket 02-135, Nov. 2002.
Tables I and II list the detection performances under cases [2] J. Mitola III and G. Q. Maguire Jr., Cognitive radio: making software
when the active channels are sufficiently far apart so as not radios more personal, Personal Communications, IEEE, vol. 6, no. 4,
cause significant inter-channel interference. The performances pp. 13 18, Aug 1999.
[3] P. Sofotasios, E. Rebeiz, L. Zhang, T. Tsiftsis, D. Cabric, and S. Freear,
under the condition when two PUs are very closely placed in Energy detection based spectrum sensing over and extreme fading
the band are however, expected to be different. Table III shows channels, IEEE Transactions on Vehicular Technology, vol. 62, no. 3,
the detection performances when two active PU transmitters pp. 10311040, 2013.
[4] F. Digham, M.-S. Alouini, and M. Simon, On the energy detection of
lie in different locations in the band relative to each other. unknown signals over fading channels, in IEEE International Confer-
ence on Communications, vol. 5, May 2003, pp. 3575 3579.
TABLE III [5] Y. Chen, Improved energy detector for random signals in gaussian
P ERCENTAGE OF DETECTION ( ) AT VARYING SNR WHEN TWO ACTIVE noise, IEEE Transactions on Wireless Communications, vol. 9, no. 2,
PU S ARE PRESENT ( I ) AT FARTHEST POSSIBLE CHANNELS (C H 1, C H 16), pp. 558 563, February 2010.
( II ) WITH A VACANT CHANNEL IN BETWEEN (C H 8, C H 10), ( III ) IN [6] V. Kostylev, Energy detection of a signal with random amplitude, in
ADJACENT CHANNELS (C H 8, C H 9). S ENSING TIME IS 5 SYMBOL PERIOD . IEEE International Conference on Communications, vol. 3, 2002, pp.
OTHER PARAMETERS ARE SET SAME AS IN TABLE I. 16061610.
[7] J. C. Shen and E. Alsusa, An efficient multiple lags selection method
FAR [17] LED (480 atoms) for cyclostationary feature based spectrum-sensing, Signal Processing
SNR (dB) Letters, IEEE, vol. 20, no. 2, pp. 133136, 2013.
Busy Channels Busy Channels
[8] A. Tkachenko, D. Cabric, and R. Brodersen, Cyclostationary feature de-
1, 16 8, 10 8, 9 1, 16 8, 10 8, 9
tector experiments using reconfigurable bee2, in 2nd IEEE International
0 87.18 87.15 99.64 96.58 97.40 97.62
Symposium on New Frontiers in Dynamic Spectrum Access Networks.,
-5 84.92 83.60 98.97 95.43 93.53 95.78
2007, pp. 216219.
-10 79.33 77.88 95.75 87.63 86.68 87.48
[9] Y. Zeng and Y.-C. Liang, Eigenvalue-based spectrum sensing algo-
-15 66.85 64.49 83.98 70.19 71.54 68.94
rithms for cognitive radio, IEEE Transactions on Communications,
vol. 57, no. 6, pp. 17841793, 2009.
[10] Z. Tian, Compressed wideband sensing in cooperative cognitive radio
It is observed that the FAR method has an advantage when networks, in IEEE Global Telecommunications Conference, 2008, pp.
15.
the active channels happen to be just adjacent to each other. [11] B. Farhang-Boroujeny, Filter bank spectrum sensing for cognitive
This is attributed to the fact that the frequency leakage from radios, IEEE Transactions on Signal Processing, vol. 56, no. 5, pp.
nearby channels contribute to the increased energy of the 18011811, May 2008.
[12] Z. Quan, S. Cui, A. Sayed, and H. Poor, Optimal multiband joint
adjacent channel leading to an improved performance. Though detection for spectrum sensing in cognitive radio networks, IEEE
this may seem as an obvious advantage of the FAR method, in Transactions on Signal Processing, vol. 57, no. 3, pp. 11281140, March
reality, the frequency leakage causes the detection performance 2009.
[13] Z. Tian and G. Giannakis, A wavelet approach to wideband spectrum
to drop in other cases when the active channels are not adjacent sensing for cognitive radios, in 1st International Conference on Cogni-
to each other. Since, it is very unlikely to have active PUs tive Radio Oriented Wireless Networks and Communications, June 2006,
located at adjacent channels most of the time, this effect may pp. 15.
[14] , Compressed sensing for wideband cognitive radios, in IEEE
not provide an overall advantage to the FAR method. The International Conference on Acoustics, Speech and Signal Processing,
performance of LED based method is however, noted to be vol. 4, April 2007, pp. IV1357IV1360.
consistent throughout all the observed cases for a particular [15] R. Prasanna and B. Amrutur, Cognitive radio implementation for a
frequency hopping primary signal, in National Conference on Commu-
SNR. nications (NCC), 2013, pp. 15.
[16] K. Khanikar, R. Sinha, and R. Bhattacharjee, Sparse representation
V. C ONCLUSION based tracking of frequency hopping primary user for cognitive radio,
in International Conference on Signal Processing and Communications
Spectrum sensing for detection of multiple unknown number (SPCOM), July 2014, pp. 16.
of FHSS PU has been proposed. The proposed LED based [17] Z. Chen, N. Guo, and R. Qiu, Demonstration of real-time spectrum
sensing for cognitive radio, Communications Letters, IEEE, vol. 14,
technique is compared with an FFT based method which no. 10, pp. 915917, 2010.
utilizes sampled signals obtained at higher rate than the pro- [18] M. Aharon, M. Elad, and A. Bruckstein, K-SVD: an algorithm for
posed method. The detection performances for the LED based designing overcomplete dictionaries for sparse representation, IEEE
Transactions on Signal Processing, vol. 54, no. 11, pp. 43114322, 2006.
method are still found to be superior when the sensing time is [19] T. Cai and L. Wang, Orthogonal matching pursuit for sparse signal
kept very short. For achieving decent detection performances, recovery with noise, IEEE Transactions on Information Theory, vol. 57,
slower ADCs can be installed at the detector front-end for the no. 7, pp. 46804688, 2011.
[20] Specification of the Bluetooth System, Volume 1, Bluetooth SIG, Inc.,
proposed method. February 2001.
The performance of the LED based method at lower SNRs [21] Guidelines for evaluation of radio transmission technologies for IMT-
may be further improved by using correctly estimated value 2000, ITU-R Recommendation M.1225, 1997.
of the sparsity constraint. Also, cooperative sensing may help