
A New Method of Automatic Modulation
Recognition Based on Dimension Reduction
Hui Wang, LiLi Guo
College of Information and Communication Engineering
Harbin Engineering University
Harbin, China
wanghui@hrbeu.edu.cn
Abstract—To improve the recognition rate of signal
modulation recognition methods under the low Signal-to-noise
ratio (SNR), a modulation recognition method is proposed. In
this paper, we study an automatic modulation recognition
through the Artificial Neural Network (ANN). Implement and
design 7 digital modulations are: 2FSK, 4FSK, 8FSK, BPSK,
QPSK, MSK and 2ASK. The cyclic spectrum after reducing
dimension via Principle Component Analysis (PCA) is chosen as
key feature for digital modulation recognizer based on the ANN.
We corrupted the signals by additive White Gaussian Noise
(AWGN) for testing the algorithm. The simulation results show
that the ANN could classify the signals in its current state of
development.
Keywords- automatic modulation recognition; Artificial Neural
Network; cyclic spectrum; Principle Component Analysis
I. INTRODUCTION
Since the spectrum is limited, in order to meet the needs of
various users, and utilize the frequency resource more
adequately, signals are modulated in different ways.
Modulation recognition is increasing in importance for a
number of years
[1]
. Extensive work has been carried out in
both military and commercial applications
[2]
, and it is also a
key technology in the intelligence signal analysis and
processing
[3]
.
In general, the modulation recognition methods are divided
into two main categories: artificial recognition method and
automatic modulation recognition method. The artificial
recognition method means to transform the signals from
high-frequency to mid-frequency, and use the modem to
demodulate the signal, then judge the demodulation results
using related instruments, such as headphones and spectrum
analyzer. However, the artificial recognition method needs
great experience and knowledge, and the recognition rate is not
accurate when the symbol rate is high. The automatic
modulation recognition method is divided into three main
processes: data pre-processing, feature extraction, and
classificatory decision. Data pre-processing is the estimation of
carrier and symbol rate after signal has been down-conversion.
The use of feature extraction is to transform original data to
extract some features which could be classified more easily.
Classificatory decision is to judge the modulation type
according to the features extracted.
Many features have been adopted for the automatic
modulation recognition, including wavelet coefficients, higher
order statistics (HOS), etc. Meanwhile, different methods are
also employed for classificatory decision, such as probability
density function (PDF) matching methods, unsupervised
clustering techniques, and support vector machine (SVM).
However, the aforementioned modulation recognition
techniques are either computationally cumbersome or lead to
unsatisfactory performances and hence new robust efficient
modulation recognition schemes are still in demand.
[4]
In this work, we proposed a method on automatic
recognition of signal modulation based on cyclic spectral
feature and artificial neural network. As the characteristics of
the signal, the cyclic spectrum is not sensitive to noise, which
is helpful for signal modulation recognition in low SNR
environment. However, the cyclic spectrum of the signal is a
large number of data, there are a lot of redundant information if
it is directly recognized as the feature. On the one hand, it
increases the complexity, and on the other hand, it may
interfere with the final recognition. Therefore, this paper uses
the PCA dimension reduction method to reduce the dimension
of the cyclic spectrum feature. For the classifier, we choose
artificial neural network as classifier. The neural network
classifier has a strong pattern recognition ability, which can
cope the complicated nonlinear problem well. Meanwhile, it
has better robustness, and it is generally applied in the
modulation recognition.
The remainder of this paper is organized as follows.
Section II is the system model in which we introduce the signal
expression and the environment of our study. Section III
describes the cyclic spectrum after dimensionality reduction as
features. Section IV introduces the classifier of the automatic
modulation classification based on the neural network, and
argues with some parameters of network. Experiments are
conducted in Section V, and finally Section VI concludes the
paper.
316
CPGPS 2017 Forum on Cooperative Positioning and Service
978-1-5090-5022-2/17/$31.00 ©2017 IEEE
Harbin, China
May 19-21, 2017