You are on page 1of 15

Convolutional Neural Network Based Modulation

Classi cation over Multipath fading Channels


varun kumar reddy  (  varunkumarreddypg.phd@gmail.com )
Vel's College of Science: Vels Institute of Science Technology & Advanced Studies

Research Article

Keywords: Machine Learning, Support vector machine (SVM), Deep Learning, Modulation classi cation

Posted Date: June 29th, 2022

DOI: https://doi.org/10.21203/rs.3.rs-1567928/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License.  
Read Full License
Convolutional Neural Network Based Modulation
Classification over Multipath fading Channels
P G. Varna Kumar Reddy1, Dr.M.Meena 2
1
Research scholar, 2Associate Professor, Electronics & Communication Engineering,
Vels Institute of Science, Technology & Advanced Studies (VISTAS), Chennai, India

Email: varunkumarreddypg.phd@gmail.com

Abstract: The authors proposed a method to classify the required modulations among
the various kinds of distributed systems by using the Convolutional neural network
(CNN). A supervised Machine learning (ML) algorithm, Support vector machine
(SVM) is used to classify the required modulation among all the modulations due to
its advantages majorly low complexity. In this paper, the CNN and SVM are
explained in simple terms with relevant mathematical analysis. For better
understanding of the classification of modulated techniques, some analog modulation
techniques such as Binary phase shift keying (BPSK), Quadrature phase shift keying
(QPSK), 8-ary phase shift keying (8-PSK), 16-ary Quadrature amplitude modulation
(16-QAM), 64-ary Quadrature amplitude modulation (64-QAM), 4-ary pulse
amplitude modulation (PAM4), Gaussian frequency shift keying (GFSK), Continuous
phase frequency shift keying (CPFSK) and digital modulations such as Broadcast FM
(B-FM), Double sideband amplitude modulation (DSB-AM), Single sideband
amplitude modulation (SSB-AM) are considered. The time domain and the
spectrogram representation of different modulations are presented by transmitting
signals over Rayleigh and Rician channels. The impact of the noise over these signals
is also studied.

Keywords: Machine Learning, Support vector machine (SVM), Deep Learning,


Modulation classification

1. Introduction

The role of signal analysis and processing in modern wireless communication is


acknowledged [1], [2]. It is performed only when the modulation of a given signal is
recognized [3]. Due to the rapid development of wireless communication technology, the
communication environment has become more complex and diverse. This environment
has led to the automatic classification of various types of modulations and their
complexity with the help of software-defined radio (SDR) and this concept is called
Automatic modulation classification (AMC) [4, 5, 6]. The AMC is typically performed
before the detection of symbols. It can be used in various civilian and military
applications to improve the representation of weak and strong channels. The complexity
of taking care of the AMC issue has been restrictive, prompting research that applies
calculations from Machine learning (ML) to give computationally manageable
arrangements, as in [7][8].

Machine learning (ML) is an important component of artificial intelligence (AI). Its


various advantages include: the ability to learn without being programmed; the ability to
perform complex tasks; and the ability to recognize and classify complex data. ML is
categorized in to supervised learning, unsupervised learning and reinforcement learning.
In supervised learning, the agent has to train the machine by providing a lot of examples
with proper inputs and outputs, then the machine tries to predict the unknown input based
on the train data. In unsupervised learning, the agent doesn't need to store or label data.
Instead, it tries to find some structures within its input. While in reinforcement learning, it
constantly interacts with an environment in order to generate a good policy. Due to the
emergence of parallel processing units and the increasing number of data sources, deep
learning has gained a lot of advantages over the years. One of these advantages is that it
can overcome the limitations of traditional models and localization algorithms [9].

In this article, we will present the method for the classification of modulated signals
through Deep Learning (DL). It is carried out through the CNN algorithm [8], which is
commonly used for data processing. Images of constellation diagrams are also fed to the
network. The AlexNet [10] model is commonly used for network training and
classification, where as the classification of modulation is done based the Caffe
framework [11].

2. Convolutional Neural Network


There are two main parts of CNN such as convolutional layer and fully connected
layer as shown in Figure 1. A convolutional tool that combines the features of an image
for analysis. It identifies the features that are most useful for the process known as feature
extraction. It also predicts the class of image based on the extracted features.
Fig. 1. The CNN architecture with the constellation input and modulation type output.

2.1 How CNN works?

Let's say we want the computer to recognize the handwritten digit 9. The way
computer looks at this is as a grid of numbers here we are using -1 and 1 (shown in Figure
2). In reality it will use RGB numbers from 0 to 255. The issue with this presentation is
that this is too much hard-coded if we have a little shift in digit 9. For example, here the
digit 9 is in the middle but if it is in the left and the representation of numbers just changes
it doesn't match with our original number grid and computer will not be able to recognize
that this is number 9. There could be a variation since it is a handwritten digit which will
change the two-dimensional representation of numbers and again we will not be able to
match it with the original grid.

Therefore, we use Artificial Neural Network (ANN) [12] for this kind of case to
handle the variety in the deep learning. In that we created a one-dimensional array by
flattening the two-dimensional representation of our hand written digit number and then
we build a neural network with one hidden layer and output layer; and this dense neural
network will work okay for a simple image like handwritten digit but may not work out
for bigger images (shown in Figure 3). In addition, ANN has some more disadvantages
such as one-too much computation, two-treating the local pixels same as pixels far apart
and three-high sensitive to location of an object in an image.
Let's see a koala (an animal) the image size is 1920 x 1080, we have three RGB
channel such as red, green and blue. In this case, the first layer neuron itself will be six
million if we have let's say hidden layer with 4 million neurons we're talking about 24
million weights to be calculated just between the input and hidden layer. However, deep
neural networks have many hidden layers so this can go easily into like 500 million or 1
billion of weights that we have to compute and that's too much computation for our
computer.

There are little edges (shown in Figure 2) which come together and form a loopy circle
pattern (can be assumed as eyes of Koala in Figure 3) which is kind of like a head of digit
nine at the top, a vertical line in the middle, a diagonal line at the bottom [13]. Sometimes
we don't have diagonal line at all but we know that whenever there is a loopy circle
pattern at the top, vertical line in the middle, diagonal line in the end that means digit
nine. But, how can we make computers recognize these tiny features? we use the concept
of filter, in case of digit 9 we have three filters; the first one is the head which is a loopy
circle pattern ,in the middle you have vertical line, in the end you have diagonal filter.

The working of filters is shown in Figure 4, the way convolution operation works is
we take 3 x 3 grid from the original image and multiply individual numbers with this
filter. In the end we get a result and then we find the average which is divided by 9
because there are total 9 numbers and whatever number we get we put it in other box
which is called a feature map. We repeat the process for each and every element, in the
end we get a complete feature map, the benefit here is wherever we see number 1 or a
number that is close to 1 it means we have a loopy circle pattern. In the case of koala eyes,
nose, ears are the features obtained by applying loopy pattern detector.

Fig. 2. The classification of filters (Loopy pattern, Vertical and Diagonal filters) on digit
9.
Fig. 3. The features extraction and convolution of an image ‘Koala’.

Fig. 4. The process of applying the loopy filter.

As shown in Figure 5, if our number is shifted it is difficult to extract the features


with the filtering alone. Therefore max pulling along with convolution helps us with
position invariant feature detection, then it doesn't matter where eyes or ears are in the
image still it will detect that feature. The other benefits of pooling number are reduction in
the dimension and computation; the second benefit is reduce overfitting because there are
less parameters and the third one is model is tolerant towards variation and distortion
because if there is a distortion and if we're picking just a maximum number, we are
capturing the main feature and we are filtering all the noise. So the complete
convolutional neural network (shown in Figure 1) has a convolution and value layer then
we will have pulling then there will be another convolution value pulling there could be n
number of layers for convolution and pulling and in the end we will have fully connected
dense neural network [14]. The same concept is applied to detect the symbols in
constellation mapping for various modulation schemes in our contribution.

Fig. 5. The application of Max pooling on shifted digit 9.

In this paper, the CNN is a multi-convolutional system comprises of six layers. The
first four layers are followed by batch normalization (for regularization purpose), rectified
linear unit activation, and max pooling; and the last layer is the output layer which has
softmax activation [15]- [19].

2.2 Support Vector Machine

A supervised machine learning algorithm, Support Vector Machine (SVM) is


commonly used for classification problems. It plots the data items as points in the n-
dimensional space, and it performs the function of the particular coordinate, where n is the
number of features we have. The classification on data points is done based on the hyper-
plane as shown in Figure 6 (a). A, B, C are three hyper-planes, where as hyper-plane ‘B’
is separating the data points perfectly compared to ‘A’ and ‘C’ hyper-planes. So in this
case, ‘B’ is a perfect classifier. If we observe Figure 6 (b), both ‘B’ and ‘D’ hyper-planes
are classifying data points perfectly, but how can we say which is the best hyper-plane
between these two? The thumb rule to identify the right hyper-plane is to maximize the
marigin which is a distance between a hyperplane and the nearest data point.

In this example, two parallel lines are drawn with respect to ‘B’ and ‘D’ by
maintaining nearest data points, respectively. G1 and G2 are the distance (Marigin) from
‘B’ and ‘D’ hyper-planes, respectively. Where G1 > G2, so it is concluded that B is the
best hyper-plane compared to ‘D’ hyper-plane.
2.2.1 Advantages of SVM
 SVM is a good fit for large data sets with many features. For instance, it can work
seamlessly with large numbers of words.
 The SVM transforms the various variables into new dimensions, which can be used to
represent the classification of the data.
 A linear classifier known as SVMs is an easy to understand algorithm that can be used
in data analysis.

Fig. 6. The classification of datasets.

3. System Model

Figure 7 shows the system model that uses a CNN to predict the type of modulation
that will occur in each frame. It uses a random generator to generate the known data
frames from the allocated samples. To achieve the ideal frequency response, the filter's
transfer function must be based on the square root that corresponds to the desired
response. For these systems, the square-root raised cosine filter is used to achieve the
optimum signal to noise ratio. Then, the signal is transmitted over multipath fading
channels such as Rayleigh and Rician channels. In wireless systems, fading may be
caused by multipath propagation or by the weather phenomenon known as shadow fading.
It is usually triggered by obstacles that can affect the wave propagation. After that, the
phase/ frequency offset is added intentionally to mitigate the interference with other
transmitters. Timing drift is added to maintain the synchronization among the symbols
and then passed through a white noise. Then the modulation frames will be generated and
the constellation data is given to SVM classifier to predict the particular modulation type.
The mathematical model analysis for the received signal y(t) can be expressed as

y (t )  x(t )  c  n(t ) (1)


Where x(t) is a time series signal that has a varied frequency, phase, amplitude, and
trajectory. c is a path loss or constant gain term that appears on the signal and n(t)
describes the process of reflecting thermal noise. This expression is often used in the
development of decision statistics and expert features, however the detailed expression is
represented by [14],

 0

 x( n
j  n Lo ( t )
y (t )  e Clk (t   )h( )  n Add (t )
 0 (2)

This paper considers a number of non-trivial world effects which are presented as
non-trivial to model. Some of these effects include random walk sampling (nLo(t), nClk(t)),
convolution, and noise (nAdd(t)),which may not be white. Simulations of harsh propagation
conditions are non-trivial and often forces us to simplify assumptions. This paper aims to
analyze the performance of simulated harsh propagation environments. The phase-
mapping for the case of QPSK this phase-mapping is defined as

2 ci 1
j 2f c t 
x(ti )  e 4
, ci  0, 1, 2, 3
(3)

Fig. 7. System model

4. Results and Analysis


Figure 8 (a) and (b) show the classification scores of various modulation schemes
which are predicted by the CNN for a modulation order of 4 and 8, respectively. It can be
observed that due to the change in order of modulation, the overlapping of 64QAM is
there on one of the modulation schemes. The impact of SNR on prediction is presented in
Table 1. It is observed that higher SNR gives the better prediction compared to low SNR,
but higher SNR requires high power amplifiers which will cause high complexity and
cost. The impact of SNR on various modulation schemes is shown in Figure 9. It is
observed that if there is a low SNR, it is very difficult to identify the required symbol and
vice versa. Therefore, higher SNR is required for the better capturing of required
modulations, but higher SNR requires high power amplifiers which are cost effective.
Figure 10 (a) and (b) show the time domain representation of obtained modulations over
Rician and Rayleigh channels, respectively. Rayleigh and Rician fade are terms used for
the fast fading of signals in wireless communication systems. When the multipath signals
contain uniformly distributed phase, the resulting signal is called Rayleigh scattering and
this occurs when there is no dominant signal path in any given system. When a signal has
a higher amplitude than its peers, it is called Rician fading and its origin is usually an LOS
path.

Figures 11(a) and (b) show the spectrogram of modulation schemes over Rician and
Rayleigh channels, respectively. It is observed that the Rayleigh channel shows the high
impact of noise compared to Rician due to the non-line of sight behavior. In future, we
may develop an algorithm to reduce the noise of Rayleigh channel too.

Table 1. Obtained modulations for respective SNR values.

SNR
10 dB 30 dB 60 dB
BPSK PAM PAM
BPSK PAM PAM
B-FM 64QAM PAM
B-FM PAM PAM
PAM PAM PAM
QPSK PAM PAM
QPSK GFSK PAM
Fig. 8. Classification of modulation schemes for a order of (a) 4 and (b) 8.

Fig. 9. The impact of SNR on signal constellation.


Figure 10. Time domain representation of modulations over (a) Rician channel (b)
Rayleigh channel

Figure 11 Spectrogram representation of modulations over (a) Rician channel (b)


Rayleigh channel

5. Conclusions
A detailed analysis on CNN with SVM is presented for the classification of analog and
digital modulations. The whole process involved in deep learning is presented by taking
an input digit 9 and koala. From the obtained results, it is concluded that higher SNR
gives the better prediction compared to low SNR, but higher SNR requires high power
amplifiers which will cause high complexity and cost. It is also observed that the
Rayleigh channel shows the high impact of noise compared to Rician due to the non-
line of sight behavior.

References:
[1] G. Q. Yang, “Modulation classification based on extensible neural networks,”
Mathematical Problems in Engineering, vol. 9, p. 6416019, 2017.
[2] F. Wang, S. Huang, H. Wang, and C. Yang, “Automatic modulation classification
exploiting hybrid machine learning network,” Mathematical Problems in Engineering, vol.
2018, pp. 1–14, 2018.

[3] Y. Wang, H. Zhang, Z. Sang, L. Xu, C. Cao, and T. A. Gulliver, “Modulation classification
of underwater communication with deep learning network,” Computational Intelligence and
Neuroscience, vol. 2019, pp. 1–12, 2019.
[4] B. Kim, J. Kim, H. Chae, D. Yoon, and J. W. Choi, “Deep neural network-based automatic
modulation classification technique,” in Proceedings of the International Conference on
Information and Communication Technology Convergence(ICTC), Jeju Island, South Korea,
October 2016.
[5] Reddy, Bathula Siva Kumar, Kiran Mannem, and K. Jamal. "Software defined radio based
non-orthogonal multiple access (NOMA) systems." Wireless Personal Communications 119.2
(2021): 1251-1273.

[6] Wu, Peng, et al. "Automatic modulation classification based on deep learning for software-
defined radio." Mathematical Problems in Engineering 2020 (2020).

[7] L. Han, F. Gao, Z. Li, and O. A. Dobre, “Low complexity automatic modulation
classification based on order-statistics,” IEEE Trans. on Wireless Commun., vol. 16, no. 1,
pp. 400–411, Jan 2017.

[8] Abu-Romoh, Mohannad, Ahmed Aboutaleb, and Zouheir Rezki. "Automatic modulation
classification using moments and likelihood maximization." IEEE Communications
Letters 22.5 (2018): 938-941.

[9] Sun, Y., Peng, M., Zhou, Y., Huang, Y., & Mao, S. (2019). Application of machine learning
in wireless networks: Key techniques and open issues. IEEE Communications Surveys &
Tutorials, 21(4), 3072-3108.

[10] Reddy, Bathula Siva Kumar. "Experimental validation of non-orthogonal multiple access
(NOMA) technique using software defined radio." Wireless Personal
Communications 116.4 (2021): 3599-3612.

[11] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, et al., ”Caffe:


Convolutional Architecture for Fast Feature Embedding,” arXiv:1408.5093, 2014.

[12] Fan, Feng-Lei, et al. "On interpretability of artificial neural networks: A survey." IEEE
Transactions on Radiation and Plasma Medical Sciences (2021).
[13] Galety, Mohammad, et al. "Deep Neural Network Concepts for Classification using
Convolutional Neural Network: A Systematic Review and Evaluation." (2021).

[14] O'Shea, T. J., J. Corgan, and T. C. Clancy. "Convolutional Radio Modulation Recognition
Networks." Preprint, submitted June 10, 2016. https://arxiv.org/abs/1602.04105

[15] O'Shea, T. J., T. Roy, and T. C. Clancy. "Over-the-Air Deep Learning Based Radio Signal
Classification." IEEE Journal of Selected Topics in Signal Processing. Vol. 12, Number 1,
2018, pp. 168–179.

[16] Liu, X., D. Yang, and A. E. Gamal. "Deep Neural Network Architectures for Modulation
Classification." Preprint, submitted January 5, 2018. https://arxiv.org/abs/1712.00443v3

[17] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by
reducing internal covariate shift,” in International Conference on Machine Learning, 2015,
pp. 448–456.
DECLARATION

Funding:

The authors declare that no funds, grants, or other support were received
during the preparation of this manuscript.

Competing Interests:

The authors have no relevant financial or non-financial interests to


disclose.

Author Contributions:

All authors contributed to the study conception and design. Material


preparation, data collection and analysis were performed by P G. Varna
Kumar Reddy and Dr.M.Meena. The first draft of the manuscript was
written by P G. Varna Kumar Reddy and all authors commented on
previous versions of the manuscript. All authors read and approved the
final manuscript.

Data Availability:

No datasets are used in this paper.

You might also like