International Journal of Electrical and Computer Engineering (IJECE)
Vol. 9, No. 1, February 2019, pp. 323~331
ISSN: 2088-8708, DOI: 10.11591/ijece.v911.pp323-331 o 323

Enhanced signal detection algorithm using trained neural
network for cognitive radio receiver

Sheetal D. Borde!, Kalyani R. Joshi?
Department Electronics and Telecommunications Engineering, College of Engineering, India
*Department Electronics and Telecommunications Engineering, PES’s Modern College of Engineering, India

Article Info ABSTRACT

Article history: Over the past few years, Cognitive Radio has become an important
Received Apr 15, 2018 research area in the field of wireless communications. It can play an
; important role in dynamic spectrum management and interference
Revised Aug 1, 2018 identification. Th 2 hni d

Accepted Aug 29, 2018 identification. There are many spectrum sensing techniques propose
? in literature for cognitive radio, but all those techniques detect only

presence or absence of the primary user in the designated band and
Keywords: do not give any information about the used modulation scheme. In
certain applications, in cognitive radio receiver, it is necessary to
identify the modulation type of the signal so that the receiver
parameters can be adjusted accordingly. Most of the modulated
signals exhibit the property of Cyclostationarity that can be used for
the purpose of correct detection of primary user and the modulation
type. In this paper, we have proposed an enhanced signal detection
algorithm for cognitive radio receiver which makes use of
cyclostationarity property of the modulated signal to exactly detect,
the modulation type of the received signal using a trained neural
network. The algorithm gives better accuracy of signal detection even
in low SNR conditions. The use of a trained neural network makes it
more flexible and extendible for future applications

Artificial neural network
Cognitive radio
Cyclostationarity
Enhanced signal detection
Spectrum sensing

Copyright © 2019 Institute of Advanced Engineering and Science.
All rights reserved.

Corresponding Author:

Sheetal D. Borde,

Department of Electronics and Communication Engineering,
College of Engineering, Pune-5, Pune University,

Pune, Maharashtra- 411005, India.

Email: sdbpune@gmail.com

1. INTRODUCTION

The existing wireless communication systems and networks are operating based on fixed spectrum
assignment policy laid down by regulatory authorities. The exclusive radio frequency spectrum assignment,
which is licensed by government regulatory bodies, is an efficient way for interference mitigation. However,
the fixed spectrum assignment policies lead to inefficient use of spectrum since most of the channels actively
transmit the information only for short duration while the certain portion of the spectrum remains idle for
most of the time when the licensed primary users are not transmitting any data [1]. This indicates that the
inefficient radio spectrum usage has been a serious bottleneck for deployment of upcoming wireless systems.
From this discussion it can be noted that the scarcity of RF spectrum is not a result of lack of spectrum but a
result of static spectrum allocations and thereby inefficient usage of the available spectrum. In order to
overcome the spectrum scarcity problem, recently Cognitive Radio (CR) system has been proposed to
facilitate the spectrum sharing so that the upcoming wireless radio systems can be accommodated without
assigning a new spectrum band. The popularity of Software Defined Radio and Cognitive Radio technologies

Journal homepage: http:/iaescore.com/journals/index.php/IJECE

324 o ISSN: 2088-8708

is increasing over past few years. The tremendous growth in the wireless communications has led to a huge
demand on the implementation of new wireless services in both licensed and unlicensed frequency spectrum.
Recent studies show that the fixed spectrum assignment policy results in poor spectrum utilization [1], [2].
Cognitive Radio (CR) is a promising technology to enable access to unoccupied spectrum bands, called white
spaces or spectrum holes, thereby achieving efficient spectrum utilization.

There are many spectrum sensing algorithms proposed for spectrum sensing in cognitive radio.
In [2] and [3], various spectrum sensing techniques are discussed. The basic three spectrum sensing methods
that are usually used are Match filter detection (MFD), Energy Detection (ED) and Cyclostationary Feature
Detection (CFD). But most of these methods mainly help to detect the presence or absence of Primary User
(PU) by detecting the energy of the sensed signal or by detecting cyclic feature. They do not give any
information about the modulation scheme used. Sometimes for a cognitive radio receiver it is not only
sufficient to just detect the presence of the signal but also to detect its modulation type and then demodulate
that signal to get back the original information signal. In CR receiver, identification of correct modulation
scheme of Primary User (PU) is often a critical issue when there is no prior knowledge about signal power,
carrier frequency etc. [4].

This work presents an enhanced spectrum sensing algorithm to addres the above discussed problem.
The main purpose of the proposed Enhanced Signal Detection (ESD) algorithm for a cognitive radio receiver
is not only to identify the existence of a signal in a given frequency band, at a given location, at a given time
but also to correctly identify the modulation scheme used to modulate that signal. As we know, modulation is
the process of varying a periodic waveform in order to use that signal to convey a message, the algorithm
makes use of this periodic signal to confirm the presensce primary user signal. This is because, noise signal
do not contains any kind of periodicity, and it is totally random in nature. So, detection of periodicity or
cyclostationary feature almost confirms the presence of a modulated carrier which might be carrying some
information signal of the primary user. After detection of periodicity, the algorithms further facilitates for the
detection of the modulation scheme of the detected signal. Thus, the proposed ESD algorithm using Trained
Nueral Network (TNN) provides an intermediate step between signal detection and demodulation. It plays an
important role in dynamic spectrum management and interference identification for civilian, commercial and
military applications.

2. LITERATURE REVIEW

A large number of automatic modulation classification methods have been developed as per the
research of at least two decades. Some modulation classification algorithms are based on the use of signal
cyclostationarity that falls under the category of feature based methods. In [1], the author has introduced a
spectral correlation theory for cyclostationary time series.Second order periodicity of the signal is exploited
using cyclostationary properties. A brief survey of the literature on cyclostationarity and the applications of
cyclostationarity in communication, signal processing and many other research areas are described in [4].
The authors in [5] have used the Spectral Correlation Function (SCF) and Cyclic Domain Profile (CDP) for
signal detection and hidden Markov model for signal classification to process extracted signal features due to
its strong pattern-matching capability. The definitions of SCF and CDP are also mathematically explained in
[5]. In [6] and [7], the authors have presented signal classification using cyclic spectral analysis and the
simulation results of different modulation types under various noise conditions. In [8], the author has
discussed how the cyclostationary features in the signals modulated using different schemes can be detected
correctly. In [9] authors have suggested the Likelihood-Based Approach to Modulation Classification. In [10]
and [11], the authors have discussed the basics of neural networks, their classification and applications in
pattern recognition. The objective of paper [12] and [13] is to summarize the methods used in various stages
of a pattern recognition system and advantages of Back Propagation Algorithm over other techniques to
recognize the complex patterns along with other real-life applications.

The authors have given the overview of Artificial Neural Network, its working principle and
training in [14] and [15]. In [16], the authors have provided comprehensive fundamentals in detailed manner
with examples and computer-oriented experiments. Moreover, the concepts associated with neural networks
are explained. The mathematical treatment of layered neural networks is given in detail. In [17], the author
has introduced supervised learning algorithm named Scaled Conjugate Gradient (SCG) and its performance
is compared to that of standard Back-Propagation Algorithm. According to [18], the spectral correlation
analysis can be preferably applied for signals that are produced by some periodic modification or modulation
of stationary random noise. In [19], the author investigates the spectral features of several basic analog and
digital modulated signals from point view of modulation recognition and parameter estimation. In [20] and
[21], the spectral correlation function that is a generalization of the power spectral density function is
described. Explicit formulae for the spectral correlation function for various types of analog and digital

Int J Elec & Comp Eng, Vol. 9, No. 1, February 2019 : 323 - 331

Int J Elec & Comp Eng ISSN: 2088-8708 Oo 325

modulated signals are also derived. In [22], authors have provided an experimental performance evaluation of
the Energy Detector based sensing using NI USRP-2930, which is a Software Defined Radio (SDR)
transceiver. In [23], problem of future wireless network has been examined and new learning scheme using
artificial neural network is proposed.

The author of [24] has discussed a modulation classification method capable of classifying MFSK
digital signals without a priori information using modified covariance method. In reference [25]
cyclostationary based spectrum sensing method and a modified intrusion elimination (AIC) algorithm to
mitigate the interference between the primary and cognitive users had been proposed. Thus from this
literature review, it is clear that cognitive radio is apromising technology to overcome the problem of
spectrum scarcity and spectrum sensing is an very important aspec t of cognitive radio receiver. There are
various method of spectrum sensing and very few help in recognising parameters of primary signal which is
of great importance in certain applications. The proposed algorithm in this paper helps to overcome the
problem of correct detection of the primary user presence as well as modulation type. Thus it is not merely a
spectrum sensing algorithm but going beyond that to give additional information about primary user, hence
the named as enhanced spectrum sensing algorithm.

3. CYCLOSTATIONARITY, SCF AND CDP

A cyclostationary process has statistical properties that vary periodically over time. Cyclostationary
feature detection method deals with the inherent cyclostationary properties or features of the signal. Such
features have a periodic statistics and spectral correlation that cannot be found in any interference signal or
stationary noise. This periodicity in the received primary signal is exploited to identifythe presence of
primary users, and that is why the cyclostationary feature detection method possesses higher noise immunity
than any other spectrum sensing method. In this method, the cyclic spectral correlation function is the
parameter that is used for detecting the primary user signals.

Cyclostationary spectrum sensing method performs better than other detection schemes such as
energy detection method in low SNR regions, because of its noise rejection capability. This occurs because
noise is totally random and does not exhibit any periodic form of behaviour. When we have no prior
knowledge about primary user’s waveform, which is the scenario in real life, then best technique to be
adopted is cyclostationary feature detection. As an advantage, the cyclostationary spectrum sensing method
can be used to find out the type of modulation scheme used by the primary user signal.

Spectral correlation function (SCF) is a powerfultool to study cyclostationary signals. It can be used
to analyze and extract information from modulated signals in noise according to their cyclostationarity [19].
The expression for SC can be derived as follows:

Signal x(t) satisfies second order periodicity with cyclic frequency a + 0 if and only if

T/2

sS : 1 j
RE = lim > Lx et x(t = Je i2mat dt (1)

exists and is not identically zero as a function of t. Above equation is known as the cyclic autocorrelation
(also called limit cyclic autocorrelation). When a=0, it reduces to conventional autocorrelation.

We know that, the Power Spectral Density (PSD) is obtained by taking the Fourier transform of the
autocorrelation function. Mathematically,

S.(f) = fo Ry (te"2™ de (2)

Similarly, the Spectral Correlation Density (SCD) is the Fourier transform of the cyclicautocorrelation
function and is given by,

SSCP) = J R&(t)e 2dr (3)

When a=0, (3) reduces to (2).
For practical implementation, Cyclic Periodogram can be defined as [5],

1 *
sO = =Xr (tf +5) x7 (tf 5) (4)
Where Xr is the time-variant Fourier transform defined as follows:

TEDES re xe"? du (5)

Enhanced signal detection algorithm using trained neural network for cognitive... (Sheetal D. Borde)

326 Oo ISSN: 2088-8708

The estimated SCF obtained by frequency smoothing of theCyclic Periodogram in (4) is,

a 1 cftAf/2
Syr(t, far = Af f-Af/2 Syr(t, v)dv (6)

The SCF can be obtained by increasing observation lengthT and reducing the size of the smoothing window
Af [5],

S%(Ð = Jim Jim S$r(t, Dar (7)

The SCF is a cross-correlation function between frequency components separated by (f+a/2) and
(f-a/2). If x(t) contains no spectral components at f=+o/2, then the SCF is actually the covariance of the two
spectral components. We can define the correlation coefficient for the SCF, which is known commonly as
spectral coherence (SC). Mathematically, the spectral coherence is defined as [2], [5], [6]

Å sto)

Cx(f) = [S(f+a/2)S(f-a/2)]1/2 (8)

The magnitude of the SC ranges from 0 to 1. It allows us to measure thestrength of second-order
periodicity contained within a timeseries in a simplified way. When SCF is plotted, the occupancy status of
the spectrum can be found out. If a primary user signal is present in the operating frequency range, the SCF
gives a peak at its centre. The peak will not be present in the case when there is no primary user signal
present in the concerned frequency range [7].

However the use of SC requires large amounts of data and hence one of the solutionsis to use only
the peak values in the SC. These peak values in the SC are called as cyclic frequency domain profile or
simply cyclic domain profile (CDP) or a-profile [5]. The cyclic frequency is inherently a discrete parameter.
It is convenient to assess the cycle-frequency parameter by evaluating the maximum over spectral
frequency f.

I(a) 4 max| cg (P (9)

Neural networks trained using the cyclic domain profiles (CDP) can be used for signal classification
due to their patternmatching capabilities. Training the neural network with different realizationsof the signal
allows it to extract features such as carrierand keying-rate features of the signal. When the neuralnetwork is
trained with a variety of signal realizations withdifferent SNRs, the network performs exceptionally, evenat
low SNR levels.

4. THE PRPOSED ALGORITHM

The proposed Enhanced Signal Detection (ESD) algorithm primarily aims at estimation of
modulation scheme of the received PU signal. Most modulated signals exhibit the property of
Cyclostationarity that can be exploited for the purpose of classification. A feature-based method called
Cyclostationary Feature Detection is able to identify different modulation schemes. The Spectral Correlation
Function obtained from the sensed signal is used as a cyclic feature. The Cyclic frequency Domain Profile
(CDP) derived from Spectral Correlation Function is used as a discriminator in the classification process
since several modulation schemes have unique cyclic frequency domain profiles. The neural network
approach based on the learning mechanism is employed for pattern matching. It is used for classification of
data patterns and distinguishing them into predefined set of classes. The two layered neural network is
trained using the Back Propagation Algorithm.

Enhanced Signal Detection (ESD) Algorithm using a Trained Neural Network (TNN)

Step 1: Detect the noisy modulated signal.

Step 2: Estimate its Cyclic Periodogram.

Step 3: Perform Time Smoothing.

Step 4: Perform Frequency Smoothing.

Step 5: Calculate Spectral Coherence.

Step 6: Calculate Cyclic Domain Profile.

Step 7: Train the Neural network with various Cyclic Domain Profile patterns.
Step 8: Obtain Classification Confusion Matrix.

Step 9: Verify the results for different modulation schemes.

Int J Elec & Comp Eng, Vol. 9, No. 1, February 2019 : 323 - 331

Int J Elec & Comp Eng ISSN: 2088-8708 Oo 327

The main Design Considerations for the Enhanced Signal Detection (ESD) Algorithm using a
Trained Neural Network (TNN) are as follows:
a. Channel is assumed to be noisy with Additive White Gaussian Noise (AWGN) with varying values of
standard deviation.
Number of data symbols are 2000 and symbol rate is 3 GHz.
Sampling frequency selected satisfied Nyquist criterion.
FFT is used for time smoothing.
Raised Cosine filter is used for frequency smoothing.
400 CDPs of different modulation schemes are input to the neural network.
Number of hidden neurons is 10.

mmoaas

5. RESULTS AND DISCUSSIONS

The LabVIEW software from National Instruments is a powerful simulation tool which makes use
of graphical programming language. The proposed algorithm is simulated with combined use of LabVIEW
and MATLAB. Different types of synthetic signal realisaions are generated to prepare a database to train the
ANN. For each of the case SCF and CDP signals are studied. After analysis of the CDP plot for the different
synthetic signal realization, the modulation type of the received signal can be recognized. CDP plot is
actually a two dimentinal view of the SCF plot which can represent the important characteristics of the signal
of our interest with less data samples.

Training the neural networks with different realizations of the signal allowed them to extract and pick out
the carrier and cyclostationary features of the signal from the background noise. Time constraints limited the
amount of data we were able to generate to test and train the ANN, however considering its sufficient performance
with a considerable observation window (on the order of 256 samples) we believe it is possible for this type of
network to classify signals with low SNRs and fewer observed samples while preserving the levels of performance
and computational complexity.

As of now, we have tested here some specific modulation types using the proposed algorithm, but
other advanced modulation schemes can also be detected in future. Based on the some typical nature of SCF
and CDP plots, different modulations schemes are discussed in following cases.

Case 1: Noisy Signal: If the received signal consists of only noise (assumed AWGN for simulation),
then no signal peaks are observed in the SCF as well as CDP plots. This indicates the absence of signal in the
sensed channel i.e. the primary user is absent. This is because noise signal do not contain any form of
periodicity, it is totally random in nature.

Case 2: Binary Phase Shift Keying (BPSK) Signal: When a BPSK modulated signal is sensed, then
three distinct peaks along a axis are observed, as shown in Figure 1. The pick signal at a perticular frequency
represents the maximum signal content at that frequency. As there are only two frequencies, one is
information signal and other is career signal in BPSK, the first peak corresponds to the symbol rate f;, the
second one corresponds to the carrier frequency f and the last one corresponds to f,+f-.

Spectral Coherence Function
Cyclic Domain Profile

Spectra Coherence Function

Figure 1. SCF and CDP for BPSK modulated signal showing presence of three peaks

Case 3: Quadrature Phase Shift Keying (QPSK) Signal: When the sensed signal is QPSK
modulated, the CDP plot is very simplified with presence of only one peak along a axis, which corresponds
to the symbol rate f, as shown in Figure 2. The extra two peaks in BPSK are missing in QPSK, because,

Enhanced signal detection algorithm using trained neural network for cognitive... (Sheetal D. Borde)

328 Oo ISSN: 2088-8708

QPSK is a balanced modulation (has both, in phase and quadrature phase components) and hence peaks
corresponding to the carrier frequency get cancelled out.

Spectral Coherence Function
Cyclic Domain Profile

0.2

0.18 -

Spectral Coherence Function
Amplitude
ke
ji 1

Figure 2. SCF and CDP for QPSK modulated signal showing presence of one peak

Case 4: Frequency Shift Keying (FSK) Signal: If the sensed signal is FSK modulated, there are two
impulses of equal amplitude in the CDP as shown in Figure 3, which corresponds to two carrier frequencies
involved, i.e. ‘mark’ frequency and ‘space’ frequency. Also, the two peaks that overlap the two impulses are
observed. These two peaks along with the third one are analogous to the one observed in CDP of BPSK.

Spectral Coherence Function
Cyclic Domain Profile
02

0.18 -

Spectral Coherence Function

7 7 T T T 7 7 7 T 1
oo o1 02 03 04 05 06 07 08 09 10
Alpha

Figure 3. SCF and CDP for FSK modulated signal showing presence of two main peaks

Case 5: Minimum Shift Keying (MSK) Signal: When the received signal is MSK modulated, then
the SCF and CDP shows only two peaks as shown in Figure 4. These peaks correspond to the two quadrature
carrier Sin2r and Cos2z. This enables easy identification of the MSK signal.

Spectral Coherence Function Cyclic Domain Profile
0.2

Spedra Coherence Fundion

Figure 4. SCF and CDP for MSK modulated signal showing presence of only two peak

Int J Elec & Comp Eng, Vol. 9, No. 1, February 2019 : 323 - 331

Int J Elec & Comp Eng ISSN: 2088-8708 o 329

The generated CDPs from the synthetic modulated signals are then given to an artificial neural
network. The use of ANN makes it simpler to classify the CDPs and then identify correctly the modulation
scheme after training phase. Figure 5 shows the block diagram of the artificial neural network. It is trained
with 256 point CDPs of various synthetic modulated signals. There is one hidden layer and one output layer.

Hidden Output

Figure 5. Neural network trained using 256 point CDPs

Figure 6 shows a plot of mean square error versus number of training iterations (epochs) according
to which the best validation performance is 0.00026501 at epoch 36. It can be concluded from the plot that
the mean square error decreases with increase in number of epochs and attains a stable value at some instant.

Figure 7 is called Classification Confusion Matrix for training, validation, testing as well as overall
classification result. It shows the various types of errors that occurred for the final trained network. In other
words, it represents the accuracy of classification.The diagonal cells show the number of cases that were
correctly classified (in green) and the off-diagonal cells show the misclassified cases (in red). The blue cell in
the bottom right shows the total percentage of correctly classified cases and the total percent of misclassified
cases. For training, the accuracy achieved is 98.9%, for training, it is 98.3% for testing and 100% for
validation. The results show very good recognition since overall accuracy of all confusion matrix is 99%.
Thus, the network performs exceptionally well, even at low SNR levels.

The proposed work over here is much different than [4] and [6], as they make use of Hidden
Markov Model (HMM) and MAXNET based classifier. It is much different from the work proposed in [23]
which focuses mainly on the spectrum sensing of predicted free channels using ANN. Instead, in this work
we have proposed an enhanced algorithm for signal detection and classification using ANN which is
implemented using LabVIEW and Matlab softwares. This approach firstly helps to fully automise the process
of signal classification and secondly, provides flexibility for future enhancements for detection of advanced
modulation schemes. The accuracy of classification achieved with this algorithm is almost 100% which was
not achieved with any of the past research work mentioned in the references. These are the some noteworthy
contributions of this work and will be proved useful for development of efficient cognitive radio systems in
future.

Best Validation Performance is 0.00026501 at epoch 36

F — Train
= Validation
— Test

-zv Best

Mean Squared Error (mse)

20 25
42 Epochs

Figure 6. Training, validation and test performance plot

Enhanced signal detection algorithm using trained neural network for cognitive... (Sheetal D. Borde)

330 o ISSN: 2088-8708

Training Confusion Matrix Validation Confusion Matrix

=

1

N

2

3

3

>
Output Class

o
a
=
o
2
Fi
a
=
5
(0)

1 2 3 4 1 2 3 4
Target Class Target Class
Test Confusion Matrix All Confusion Matrix

Output Class
Output Class

1 2 3 4 1 2 3 4
Target Class Target Class

Figure 7. Classification confusion matrix

6. CUNCLUSION

The proposed Enhanced Signal Detection algorithm using Trained Neural Network not only helps to
detect the signal but classifies it correctly and identifies the modulation scheme. Because of use of
cyclostationarity feature of the PU signal, the algorithm can work well even in low SNR conditions and still
gives more accurate results as compared to energy detection scheme whose performaaance degrades in noisy
environment. The calculated SCF and then CDP of the received signal helps to correctly identify the type of
the modulation scheme with the help of neural network. The results show very good recognition rate since
overall accuracy of all confusion matrix is 99%. The classification accuracy of the algorithm is also 98.3%
which is the highest among the other classification schemes discussed in literarture review. The use of
LabVIEW software for implementation of ANN provides automization of signal classification process and
gives the flexibility for future adaptations. It also reduces computational complexity. The proposed algorithm
not only gives the assurance of the presence of signal by detecting its scheme of modulation but also helps to
demodulate the signal by knowing its type. In this paper, the performance of the ESD algorithm is tested by
making use of limited number of modulation schemes, but in future more modulation schemes can be
considered. The neural network gives the flexibility for training with advanced modulation schemes such as
OFDM, DVB using fading channel environment with noise uncertainty which can further extend the scope of
this algorithm in future.

REFERENCES

[1] W. A. Gardner, “The Spectral Correlation Theory of Cyclostationary Time-Series”, Signal Processing, vol.11, no.1,
pp. 13-36, 1986.

[2] T. Yucek and H. Arslan, “A Survey of Spectrum Sensing Algorithms for Cognitive Radio Applications,” IEEE
Comm. Survey Tuts., vol. 11, no. 1, pp. 116—130, 2009.

[3] S. D. Borde, K. R. Joshi, “Signal Processing Techniques for Spectrum Opportunities and Interference Detection in
Cognitive Radio Networks”, proceedings of IET International Communication Conference of Wireless Mobile &
Computing, Nov 2011.

[4] B. Ramkumar, “Automatic Modulation Classification for Cognitive Radios using Cyclic Feature Detection”, JEEE
Circuits and Systems Magazine, vol. 9, no. 2, pp. 27-45, 2009.

[5] W. A. Gardner et al., “Cyclostationarity: Half a century of research”, Signal Processing, vol. 86, pp. 639-697,
2006.

[6] K. Kim, I. A. Akbar, K. K. Bae, J.-S. Um, C. M. Spooner, and J. H. Reed, “Cyclostationary Approaches to Signal
Detection and Classification in Cognitive Radio,” in Proc. IEEE DySpan, pp. 212-215, 2007.

[7] A. Fehske, J. Geaddert, and J. H. Reed, “A New Approach to Signal Classification Using Spectral Correlation and
Neural Networks,” in Proc. IEEE Int. Symp. New Frontiers in DySpan, 2005.

Int J Elec & Comp Eng, Vol. 9, No. 1, February 2019 : 323 - 331

Int J Elec & Comp Eng ISSN: 2088-8708 Oo 331

[8] Aparna P.S., M. Jayasheela, “Cyclostationary Feature Detection in Cognitive Radio Using Different Modulation
Schemes ”, International Journal of Computer Applications, vol. 47, no.21, pp. 12-16, 2012.

[9] F. Hameed, O. A. Dobre, D. C. Popescu, “On the Likelihood-Based Approach to Modulation Classification”, IEEE
transactions on wireless communications, vol. 8, no. 12, pp. 5884-5892, 2009.

[10] Andrej Krenker et al., “Introduction to the Artificial Neural Networks”, Artificial Neural Networks -

Methodological Advances and Biomedical Applications, 2011.

[11] B. Yegnanarayana, “Artificial Neural Networks for Pattern Recognition”, Sadhana, Vol19, Part 2, April 1994

[12] Dr. Rama Kishore and TaranjitKaur, “Backpropagation Algorithm: An Artificial Neural Network Approach for

Pattern Recognition” International Journal of Scientific & Engineering Research, vol. 3, no. 6, June 2012.

[13] TaranjitKaur, “Implementation of Backpropagation Algorithm - A Neural Network Approach for Pattern

Recognition” International Journal of Engineering Research and Development, vol. 1, no. 5, pp. 30-37

https://pdfs.semanticscholar.org/b0a6/5 78f40c 4/581 74670c755a7a084fa8a0c471.pdf, June 2012.

[14] S. B. Maind et al., “Research Paper on Basic of Artificial Neural Network” JJRITCC, 2014.

[15] Vidushi Sharma et al., “A Comprehensive Study of Artificial Neural Networks”, LJARCSSE, 2012.

[16] Simon Haykin, “Neural Networks-A Comprehensive Foundation”, 2nd Edition, 1999.

[17] M. F. Moller, “A Scaled Conjugate Gradient Algorithm for Fast Supervised Learning”, Neural Networks, vol. 6,

1993, pp. 525-533, 1991.

[18] J. Goerlich et al., “Signal Analysis Using Spectral Correlation Measurement”, IEEE Instrumentation and

Measurement Technology Conference, pp. 1313-1318, 1998.

[19] Hong Wang, “Spectral Correlation Function of Basic Analog and Digital Modulated Signals”, IEEE, School of

Electronic Engineering, University of Electronic Science and Technology of China, 2013.

[20] W.A. Gardner, “Spectral Correlation of Modulated Signals: PART I - Analog Modulation”, JEEE Trans. on

Communication, vol. 35, pp. 584-594, 1987.

[21] W.A. Gardner, “Spectral Correlation of Modulated Signals: PART II - Digital Modulation”, IEEE Trans. on

Communication, vol. 35, pp. 595-601, 1987.

[22] F. Z. El Bahi, H. Ghennioui, and M. Zouak, “Performance Evaluation of Energy Detector Based Spectrum Sensing

for Cognitive Radio using NI USRP-2930, International Journal of Electrical and Computer Engineering (IJECE),

vol. 7, no. 4, pp. 1934 — 1940, August 2017.

[23] Rita Mahajan, Deepak Bagai, “Improved Learning Scheme for Cognitive Radio using Artificial Neural Networks
International Journal of Electrical and Computer Engineering (IJECE), vol. 6, no. 1, pp. 257-267.
http://www. iaescore.com/journals/index. php/IJECE/article/view/103/119, , February 2016.

[24] H. M. Hamee, J. Wadi, “Automatic Modulation Recognition for MFSK Using Modified Covariance Method” ,
International Journal of Electrical and Computer Engineering (IJECE), vol 5, no. 3, pp 429-435.
http://www. iaescore.com/journals/index. php/IJECE/article/view/563 6/4656, , Jane 2015.

[25] J. Avila, K. Thenmozhi, “Multiband OFDM for Cognitive Radio — A Way for Cyclostationary Detection and
Interference Cancellation” International Journal of Electrical and Computer Engineering (IJECE), vol. 6, no. 4,
pp. 1702-1709, , August 2016.

BIOGRAPHIES OF AUTHORS

Sheetal D. Borde has completed her bachelor’s degree in Electronics and Telecommunications
from Government College of Engineering, Aurangabad, Maharashtra, India in year 2001. She
completed her Masters degree in Electronics Engineering from Marathwada University,
Maharashtra, India in year 2003. Currenlty she is pursuing her Ph. D. in Electronics and
Telecommunications engineering from College of Engineering, Pune under Pune University.
Having an experience of 14 years, she is currently working as a Assistant Professor in
Electronics and Telecommunication Department of PES’s Modern college of Engineering, Pune
Maharashtra, India. Her research interests include wireless communication, mobile
communication and cognitive radio.

Dr. Mrs. Kalyani R. Joshi has currently been working as Principal and Professor E&TC in PES’s
Modern College of Engineering, Pune, Maharashtra, India. She has completed her Bachelor’s
degree in Electronics Engineering from Shivaji University and Masters degree in E&TC from
College of Engineering, Pune under Pune University. She has got her Ph. D. degree from
College of Engineering, Pune in year 2008 with specialization in biomedical image processing.
She has been worked as Chair, Signal Processing Chapter, IEEE Pune section for three
consecutive years from 2013 to 2015 and awarded as senioer member of IEEE. She has Co-
authored a book on Image And Video Compression Fundamentals,CRC Press-2015. She has
more than 30 national and international publications in reputed journals and conferences. Her
areas of interests are Digital Signal and Image Processing and signal processing in
Communications.

Enhanced signal detection algorithm using trained neural network for cognitive... (Sheetal D. Borde)