EuroCon 2013 • 1-4 July 2013 • Zagreb, Croatia

Identification of Spectrum Holes using ANN Model

for Cognitive Radio Applications
Sandhya Pattanayak 1, R .Nandi2
1
Dept. of Electronics and Communication, Narula Institute of Technology, JIS Group
Kolkata 700109, India
[email protected]
2
Dept. of Electronics and Communication, Jadavpur University
Kolkata 700032, India
2
[email protected]

Abstract— With the burgeoning popularity & necessity of all (SU) Opportunistic Spectrum Access. The spectra is broadly
things wireless, the radio spectrum has become a scarce utility; divided into three categories: the black spaces, grey spaces and
cognitive radio (CR) is now the relevant technology under white spaces[2]. Recently soft computing techniques like
development that enables one to utilize the spectrum more ANN, genetic algorithm , Ant colony optimization are applied
efficiently. Here we propose a new artficial neural network
in the field of engineering for speech processing[6],
(ANN) model that predicts the channel capacity of the received
signal. This information is analyzed theoretically which is telecommunications [7]. Now ANN is used for the design of
subsequently verified by a suitable simulation scheme for the AI core of the cognitive radio[8].
identifying possible white space in a given band. In this paper we propose a ANN based spectrum hole
Keywords: ANN, Cognitive radio, spectrum sensing, Cognitiv engine, detector with conceptual background based on analysis of the
Channel capacity, mutual Information. SNR of channel, channel capacity, bandwidth efficiency and
mutual information. The SNR over the channel is given as
I. INTRODUCTION input to the ANN model and the channel capacity is predicted.
The radio spectrum is a scarce resource, but recent studies This output is analyzed to identify a channel as unoccupied.
reveal that substantive portion of the bands allotted to a A scenario having cognitive radio based scanning system,
licensed primary user (PU) remains underutilised. The key replicated as the secondary user (SU), with artificial
advantage of the CR technique is based on dynamic spectrum intelligence placed within the coverage area of the primary
access that enables one to identify and subsequently utilize the base station at a distance ‘d’ is taken for our simulation. The
spectrum holes for another user opportunistically. The study scanning system is intended to identify the white spaces based
also shows that in the current fixed spectrum allocation on the SNR over the channel. This is performed by predicting
scheme some frequencies in the spectrum are not occupied or the channel capacity using an ANN mode. The analysis of the
partially occupied and others are heavily used[1]. These output is done by calculating the bandwidth efficiency and
unoccupied frequencies have been termed as spectrum holes mutual information. The simulation suggests that the channel
[1]. A spectrum hole is defined as a band of frequencies having low channel capacity has low spectral efficiency and
assigned to a PU, but, at a particular time and specific mutual information. The channel having spectral efficiency
geographic location, the band is not being utilized by that and mutual information nearly equal to zero are considered as
user[1]. not transmitting and hence unoccupied. Some simulation
The cognitive radio has been suggested as an intelligent results on mutual information and bandwidth efficiency with
respect to varying SNR and channel capacity values are
device that can help in the efficient use of spectrum. In[2] the
presented graphically.
OOPDAL Loop with a cognitive cycle having the steps :
observe, orient, plan, learn, decide and act is suggested. The In section II a brief literature survey of spectrum sensing
radio senses the outer world and the internal parameters of the techniques is given: in III ANN is introduced briefly; in IV
system so as to make a choice of RF band, air interface, we discuss the proposed ANN model for spectrum sensing and
protocol etc [3].The cognitive tasks are categorized as : Radio finally in section V we present a comprehensive conclusion
scene analysis which is done to estimate the interference indicating some future research directions.
temperature of the radio environment and detect spectrum
holes, channel identification transmit power control and II. LITERATURE SURVEY
dynamic spectrum management [4]. The knowledge base The detection of the primary signal is based on the famous
stores the experience or the prior knowledge of the cognitive hypothesis [7].
system and is further used to take decisions [5] – a new
feature to the OOPDAL loop. H0 :  = 0, primary signal absent
The spectrum sensing is the capability of the cognitive H1 :  = 1, primary signal present
radio to detect unused band and allot to a secondary user- Where  is 0 or 1 which determines the presence or
absence of the primary signal respectively

978-1-4673-2232-4/13/$31.00 ©2013 IEEE 133

 EuroCon 2013 • 1-4 July 2013 • Zagreb, Croatia

H o : y (t ) = n (t ) This cr based scanning system is placed within the coverage

area of the primary base station(BS) at a distance d from it as
H1 : y (t ) = hx(t ) + n(t )
shown in fig 1. The scanning system scans the channels
where x(t) is the primary user’s signal to be detected licensed for the primary user PU and the snr over the channel
n(t) is the additive white Gaussian noise, is given as input to the ANN model to predict the channel
h is the channel gain capacity. The channel capacity predicted is an index to
Ho is a null hypothesis indicating white space and H1 identify white spaces and the unoccupied channels are allotted
channel is occupied with a primary signal. to the secondary user
We know that the power received Pr at a receiver placed at
The various spectrum sensing methods are discussed in [9].
a distance ‘d’ from the transmitter base station is dependent on
To detect whether the frequency band is vacant or not the radio
the pathloss exponent ‘n’ and transmit power Pt, which is
signal is filtered, if it contains only noise then frequency band
assumed to be 1W [19]. So the power received at the receiver
is not occupied, otherwise if there is a unknown signal and
at a distance d from the transmitter is given as
noise then it is occupied[10]. Matched filter detection
Pr(d) = Pt( d0 / d )n for do = 10m ……………….(1)
technique [11], Energy detection techniques [12][13] and
Noise power spectral density No is assumed as 10-9 W/Hz and
cyclostationary feature detection technique[14][15] can be used
B is 5KHz bandwidth and d0 is the reference distance. For Pt
for spectrum sensing. The cyclostationary feature detection
equal to 1W, n = 3 and distance d = 200m SNR will be
technique can easily differentiate noise from the primary user
because noise is a widesense stationary signal with no
SNR = Pr(d)/(N0B) = 13.9dB ……… (2)
correlation while modulated signals are cyclostationary with
spectral correlation due to redundancy of the signal
periodicities [16].But this technique cannot distinguish
between the deep fade signal and unused band [17].
.

III. ARTIFICIAL NEURAL NETWORKS

Artificial Neural Networks are neuroscience inspired
computational tools that are trained using input-output data to
generate a desired mapping between input stimulus to the
targeted output [18]. The artificial neuron is the unit model of
the ANN structure which gets input from all neighboring
neuron and gives an output depending on its synaptic weight
and activation functions. The neurons are arranged in three
layers the input layer, hidden layer and the output layer.
Learning is accomplished by adjusting the weights till the
difference between the desired output and the output of the
network is negligible. The ANN has been used to make the
cognitive radio learn from the environment and take
decision[8]
Fig. 1. Proposed scenario
IV. PROBLEM FORMULATION
In this paper we have tried to put up a new technique to So the channel capacity over a channel with bandwidth B and
identify spectrum holes. A neural network model which can SNR is given as [11]
predict some channel state information is designed for C= B log2(1+SNR) ……………… (3)
spectrum sensing. The channel information here is the channel For this scenario channel capacity over a transmitting or
capacity which is predicted from the SNR of the channel occupied channel for SNR equal to 13.9 dB is calculated as
scanned. Spectrum sensing implies the detection of white C= 23.502 Kbps. At this location if the channel capacity is
spaces. Here the proposed ANN model predicts channel found less than this value then it is a clear indication that the
capacity and this output is analyzed to find whether the channel is unutilized. So the ANN model which is placed at
channel is occupied or not, based on information-theoretic 200m from the transmitting base station predicts the channel
analysis. capacity and analysis is done to find whether the channel is
occupied or not. The SNR at a particular location is
We consider a QPSK transmission and reception system dependent on the distance between the primary base station
over an AWGN channel. The bandwidth of the channel is and the geographical location at which the CR based scanning
taken as 5KHz. The concept of cognitive radio is inculcated system is placed. We know that the SNR will reduce as the
by allotting the vacant band to the SU. So a cognitive radio distance d between the BS and CR based scanning system
based scanning system with artificial intelligence is proposed. increases. Hence the channel capacity will reduce even though

978-1-4673-2232-4/13/$31.00 ©2013 IEEE 134

 EuroCon 2013 • 1-4 July 2013 • Zagreb, Croatia

the channel is transmitting. So the threshold of channel As the channel capacity increases, entropy increases.
capacity at which the channel is declared as unoccupied is If maxp(x) I(X;Y)  0
dependent on the distance ‘d ’. then it means that the channel is not transmitting. Hence it is a
For distance d being 200m the channel capacity C for an decision making factor to declare a channel as occupied or
occupied channel is 23.502Kbps. At this distance for a unoccupied.
channel with channel capacity less than this value can be Channel capacity is also an index to bandwidth efficiency
declared as unoccupied. But for a scanning system placed at a as [11]
distance 400m from the transmitting base station the SNR C/B = log2 (1+SNR) ………….(5)
received will be 6.9dB. The channel capacity of the occupied
channel will be 12.7Kbps which is less than the channel Less bandwidth efficiency is due to no transmission over the
capacity predicted at 200m, but the channel is occupied. channel. Spectral efficiency and mutual information if nearly
Hence to identify spectrum hole at a particular geographical zero or below a certain threshold then the channel can be
location the distance of the SU from the primary base station considered as unoccupied. Hence the availability of the
and the SNR over the channel at this location has to be given channel can be decided. A decision making threshold λ has to
as input to the ANN model. The ANN model will predict the be set for channel capacity above which the channel is
channel capacity which is further processed to declare a considered as occupied and below it the channel is considered
channel occupied or unoccupied. as unoccupied.
The method we are proposing may be applied both for
WLAN in infrastructure mode as well as in ad-hoc mode. In V. PROPOSED ANN MODEL
case of infrastructure mode the access point must be equipped
with CR based scanning system. The distance between the The efficiency of the neural network model depends on the
access point and the primary BS is fixed. So the ANN model parameters used for training. Database comprising the channel
used here is for predicting the channel capacity for a fixed capacity of the QPSK signal for a particular SNR and distance
distance say 200m. Depending on the channel capacity is developed. The channel capacity is calculated for SNR
predicted the channel status is decided. If the channel is ranging from -5dB to 20dB and for a bandwidth of 5KHz.
unoccupied than the access point allots the channel to the SU
in its coverage area.
TABLE I
Incase of ad-hoc mode all the mobile nodes have to be TRAINING PARAMETERS
equipped with CR based scanning system. The input to ANN
model is the distance between the mobile node and the Parameters Value
primary base station and SNR over the scanned channel. The Number of input neurons 2
mobile node (SU) can scan for the unoccupied channel at
every location. The channel status is decided based on the Number of output neurons 1
channel capacity predicted for a particular SNR and the
Number of neuron in hidden layer 1 15
distance ‘d ’.
In this paper we have assumed that the SU is in a Number of neuron in hidden layer 2 20
infrastructure system with the access point placed at a distance
of 200m from the primary base station. Spectrum sensing is Learning rate (η) 0.01
performed for this system. We intend to enhance the work by
performing spectrum sensing for infrastructure system placed Training tolerance 1 × 10-5
at various distances or mobile node(SU) placed within the
coverage area of the primary transmission system. Training time 5 min.
The channel capacity C is a very important parameter to
analyze the channel status. The channel capacity here can be Activation function in hidden layer and Log sigmoid
assumed as the transmission rate over the channel. The input layer function
channel capacity reduces as SNR reduces. If the channel is Activation function in output layer Pure linear
unoccupied then the noise power is high and the capacity will Training Algorithm Resilient Back
reduce. A low channel capacity indicates that the channel is propagation
vacant.
As per Shannon, channel capacity is also a measure of The ANN model is designed and trained for different
mutual information I(X;Y) and entropy (H) over all possible number of hidden layer and neurons in it and the results are
inputs [11]. compared. ANN model 1 is designed with 1 hidden layer
having 5 neurons in it. ANN model 2 is designed with 2
C = maxp(x) I(X;Y) = [H(X)- H(YX)] ………….(4) hidden layers having 5 and 20 neurons in them. The ANN
model 3 has 2 hidden layers with 15 and 20 neurons in them.
where X is the input and Y is the output . All the ANN models are trained to a targeted accuracy of
0.00001. Training is performed by resilent backpropagation
algorithm. Log sigmoid function is used as transfer function in

978-1-4673-2232-4/13/$31.00 ©2013 IEEE 135

 EuroCon 2013 • 1-4 July 2013 • Zagreb, Croatia

the neurons. The purpose of taking the resilent Hence the neural network model can be used for predicting
backpropagation algorithm is to avoid the harmful effects of the channel capacity. This model can be used as the AI core of
the magnitudes of partial derivatives. The sign of the the cognitive radio where QPSK is the modulation technique
derivative decides whether the weight and bias has to be and predict the free channels of the basis of channel capacity
increased or decreased. The weight is increased when the and decide the status of the channel.
derivative of the performance function is same for two
successive iterations The update value of each weight is Comparative study of Channel capacity predicted
increased by a factor delta_inc taken as 1.2 . The weight and 35
bias is reduced when the derivative with respect to that weight
30
changes sign from the previous iteration, Here the weight is

Channel Capacity in Kbps

reduced by a factor delta_dec taken as 0.5. The performance 25

of each ANN model is compared in the Fig. 2. It is observed 20

that the ANN model 1 has accuracy of mean square error 15

0.000518, ANN model 2 has accuracy with a mean square 10

error of 0.0001015 and ANN model 3 has accuracy of mean 5

square error 0.0000688. Hence model 3 yields relatively more
0
accurate result as compared with ANN models 1 and 2. -5 -3 -1 0.5 2 4 6 7 9 11 13 14 16 18 20
SNR in dB

Theoretically calculated channel capacity Channel capacity predicted by ANN model 1

ANN model training Channel capacity predicted by ANN model 2 Channel capacity predicted by ANN model 3

Fig. 3. Comparision of different neural network model predicted channel

0.002 capacity and theoretical channel capacity
0.0018
0.0016
The channel status is analysed on the basis on Imax(X,Y)
mean square error

0.0014
0.0012 over the channel. If the mutual information Imax(X,Y) of the
0.001 channel is nearly equal to zero then the channel is considered
0.0008 as unoccupied. A graphically analysis of Imax(X,Y) vs SNR in
0.0006 Fig 4 shows that as SNR reduces the mutual information
0.0004
becomes near to zero. If the CR scan device placed at a
0.0002
0
distance of 200m from the primary transmitting BS, CR
device observes SNR less than 13.9dB; the corresponding
0

0
00

00

00

00

00

00

00

00

00

00
10

50

90

mutual information Imax(X,Y) is nearly equal to zero then the

13

17

21

25

29

33

37

41

45

49

no: of iterations
channel can be considered as unoccupied. Hence the channel
ANN model 1 ANN model 2 ANN model 3 with low channel capacity is considered as unoccupied.
Fig. 2. Performance comparison of different ANN models

From the graph in Fig. 3 and Table II it is observed that the

channel capacity predicted by the neural network model is
nearly same as the theoretical calculations.
TABLE II
COMPARATIVE STUDY OF ANN OUTPUT AND THEOREICAL OUTPUT

SNR in Theoretical Channel Error in % of

dB Channel capacity Kbps error
capacity in predicted
Kbps by ANN in
Kbps

-5 1.982 1.9595 0.0235 1.18

-1 4.2172 4.223 -0.0058 0.948

5 10.2869 10.2917 0.0048 0.04

Fig. 4. Mutual information vs SNR in dB predicted over a
10 17.2972 17.2731 0.0241 0.13 channel from a distance of 200m and 400m from
transmitting BS
15 25.139 25.186 -0.047 -0.18 The bandwidth efficiency reflects how efficiently the
allotted bandwidth is utilized. It is the throughput data rate per
20 33.2911 33.1929 0.0982 0.29
hertz in a given bandwidth [19].

978-1-4673-2232-4/13/$31.00 ©2013 IEEE 136

 EuroCon 2013 • 1-4 July 2013 • Zagreb, Croatia

 = R/B bps/Hz……………...……(6) that a channel, which is not transmitting, would have low
channel capacity and the corresponding bandwidth efficiency
where R is the data rate in bit/sec and B is the bandwidth and mutual information will be nearly equal to zero. Hence we
allotted for the signal. Here R is the channel capacity i.e. the can identify white spaces in a particular geographical location.
transmission rate over the channel. Our further plan is to enhance the work by performing
spectrum sensing for infrastructure system placed at various
From Fig. 5 it is observed that the channel capacity distances d or mobile node (SU) placed at any location within
predicted by the ANN model is an index of bandwidth the coverage area of the primary transmission system.
efficiency. When a channel is occupied than for a SNR of
14dB the channel capacity predicted will be 23.502Kbps. For REFERENCES
a channel bandwidth of 5KHz the bandwidth efficiency will [1] P. Kolodzy et al., “Next generation communications: Kickoff meeting,”
be 4.7bps/Hz. But if the spectral efficiency over a channel at Proc. DARPA, Oct. 2001.
this geographical location is found to be far below 5bps/Hz [2] J. Mitola III and G. Q. Maguire.Jr., “Cognitive radio: making software
radios more personal,” IEEE Personal Commun. Mag., vol. 6, pp.13–18,
then the channel can be assumed as unoccupied. As the Aug. 1999.
channel capacity is reducing the bandwidth efficiency is [3] J. Mitola III, “Cognitive Radio: An Integrated Agent Architecture for
nearly equal to zero. Hence a channel with less channel Software Defined Radio”, Ph.D. Dissertation Royal Instt.Tech., Sweden,
capacity has low bandwidth efficiency i.e. the primary signal May 2000.
is absent and the channel can be declared as unoccupied. [4] S. Haykin, “Cognitive radio: brain-empowered wireless
communications,” IEEE J. Sel. Areas Commun., vol. 23, pp. 201- 220,
Feb. 2005.
[5] X. Dong,Y. Li, C. Wu and Y. Cai; , "A learner based on neural network
for cognitive radio," Comm. Tech. 12th IEEE International Conference
on ICCT, pp.893-896, Nov. 2010.
[6] S.Dhanjal , "Artificial neural networks in speech processing: problems
and challenges," Comm, Compu. Sig Proc., IEEE Pacific Rim
Conference vol.2, pp.510-513 .2001
[7] P. Giribone., R. Revetria ,M. Antonetti and R.Tabolacci, "Use of
artificial neural networks as support for energy saving procedures in
telecommunications," Telecomm Energy Conf., INTELEC. pp.159-162,
2000
[8] S.Pattanayak, P.Venkateswaran and R.Nandi ,”Artificial neural
networkk for cognitive radio : A preliminery survey” WiCOM 2012, ,
Shanghai, 21-23 Sept 2012
[9] T.Yucek and H. Arslan., "A survey of spectrum sensing algorithms for
cognitive radio applications," IEEE Comm. Surveys & Tutorials, ,
vol.11, pp.116-130, 2009.
Fig. 5. Simulation results on bandwidth efficiency for d= 200m from
[10] J. G. Proakis, Digital -Hill Pub. , 2001 Communications, McGraw.
transmitting BS
M.Ghozzi, M.Dohler, F. Marx and J..Palicot,”Cognitive radio methods
for the detection of free bands,”Comptes Rendous Physique, vol 7, pp
794-804, Sept. 2006.
The CR based scanning system with ANN model predicts [11] J. G. Proakis, Digital Communications, McGraw-Hill Pub. , 2001.
the channel capacity and is studied to identify the white [12] H. Urkowitz, “Energy detection of unknown deterministic signals,”
spaces. The optimized neural network model once trained Proc. IEEE, vol. 55, pp. 523–531, 1967.
need not be further trained and the optimized weights and bias [13] F. F. Digham, M. S. Alouini, and M. K. Simon, “On the energy
can be extracted to design the CAD model. We further plan to detection of unknown signals over fading channels,” IEEE Trans
Comm, vol. 55, pp. 21–24, 2007.
consider factors affecting transmission of signal like
[14] K. Kim, I. A. Akbar, K. K. Bae, J.-S. Urn, C. M. Spooner, and J. H.
interference and design a CAD model. This particular Reed, “Cyclostationary approaches to signal detection and classification
technique is most suitable for spectrum sensing and allocation in cognitive radio,” Proc. IEEE Int. Symp. New Frontiers in Dynamic
in a particular coverage area of the BS of the primary user. Spectrum Access Networks, Ireland, , pp. 212–215, Apr. 2007
[15] S. Shankar, C.Cordeiro, and K.Challapali, “Spectrum agile radios:
VI. CONCLUSION utilization and sensing architectures,” Proc. IEEE Int. Symp. Ibid.,
Maryland, USA, pp. 160–169, Nov. 2005.
In this paper we propose an ANN model which can predict
[16] D Cabric and R W Brodersen, “Physical layer design issue unique to
channel capacity. This channel state information is analysed cognitive radio systems,” in Proc. IEEE PIMRC, vol.2, pp.759-763,
by determining the bandwidth efficiency and mutual Sept 2005.
information for identifying the spectrum holes. We observe [17] K. Hamdi and K.B. Letaief, "Cooperative communications for cognitive
that channel capacity predicted by the ANN model can be radio networks," Proc. 8th Postgrad. Symp. Converg. Telecom.,
Liverpool J. M.University, June 2007.
considered as a decision making parameter to declare the
[18] S. Haykin, Neural Networks: A Comprehensive Foundation., NJ:
channel occupancy status as it provides a measure of the Prentice-Hall, 1999.
bandwidth efficiency, entropy (H) and mutual information (I) [19] A.Goldsmith, Wireless Communication, Cambridge Press, 2005.
over the channel. The analysis made from the graphs show

978-1-4673-2232-4/13/$31.00 ©2013 IEEE 137