Abstract—
Modulation identification shows great significance
for any receiver that has little knowledge of the modulation
scheme of the received signal. In this paper, we compare the
performance of a deep autoencoder network and three shallow
algorithms including SVM, Naïve Bayes and BP neural network
in the field of communication signal modulation recognition.
Firstly, cyclic spectrum is used to pre-process the simulation
communication signals, which are at various SNR (from -10dB
to 10dB). Then, a deep autoencoder network is established to
approximate the internal properties from great amount of data.
A softmax regression model is used as a classifier to identify the
five typical communication signals, which are FSK, PSK, ASK,
MSK, QAM. The results for the experiment illustrate the
excellent classification performance of the networks. At last, we
discuss the comparison of these methods and three traditional
shallow machine learning models.
Keywords-deep learning; Autoencoders; Cyclic spectrum;
softmax
Modulation scheme identification plays an essential role in
a receiver system which has limited information about
received communication signals, and it has well application
for military and civil activity. The basic task of the modulation
recognition of communication signals is to configure the
modulation scheme and signal parameters of the received
signals under multi-signal environment and the noise
interference conditions, in order to provide a basis for further
analysis and signal processing. The classic recognition
methods are statistical pattern recognition [1-2] and statistics
judgment [3-4]. Statistical pattern recognition is widely used
in real systems since it needs to select proper features and
decision criteria of the received signals, which is easier to
implement.
Statistical pattern recognition usually includes three parts:
pre-processing, feature extraction and classification. Many
machine learning algorithms have been used as shallow
classifications. However, these classifications need manual
feature extraction, which couldn’t perform well when it comes
to high dimension and large volume of data. Deep learning can
help to learn representations of data with multiple levels of
abstraction according to the mathematical model which is
composed of multiple processing layers [5]. It has been
successfully applied in many aspects such as speech
recognition [6] and visual object recognition [7] since Hinton
published papers [8] in Science in 2006.
In this paper, a stacked sparse autoencoder and softmax
regression based deep learning network is introduced to
identify communication signals of different modulation
schemes, which are FSK, PSK, ASK, QAM and MSK. In the
simulation experiment, the performance of this method is
compared with that of three traditional automatic modulation
recognition systems, which use shallow classifiers.
Digital modulation signal, which is called cyclostationary
signal, is a kind of special non-stationary random signal, and
the statistical property of the signal is periodic. Both the first
order statistic property and the second order property of the
generalized cyclostationary signals are periodic [9].
The first-order and second-order statistical features of the
signal, which are known as expectation and autocorrelation
function can be expressed as follows.
)
2
,
2
()
2
,
2
();( nTtnTtRttRtR
xxx
can be expanded as
the form of Fourier series.
2
2
( ; ) ( ) ( )
j kt
j t
T
x x x
m m
R t R e R e
are constant, and the factor of its
Fourier series can be expressed as follows:
2/
2/
2
);(
1
)(
T
T
tj
xx
dtetR
T
R
is cyclic frequency of the signal,
and
represents the cyclic autocorrelation function.
Then the cyclic autocorrelation function can be utilized to get
the spectral density function of the signal via Fourier
transform.
dteRfS
ftj
xx
2
)()(
are cyclic frequency and spectrum
frequency respectively.
Research on Modulation Identification of Digital Signals Based on
Deep Learning
Jiachen Li, Lin Qi, and Yun Lin*
College of Information and Communication Engineering,
Harbin Engineering University, Heilongjiang, China, 150001
Email: lijiachen@hrbeu.edu.cn