Convolutional neural network based blind automatic modulation classification robust to phase and frequency offsets

Abstract

In this study, the authors propose robust automatic modulation classifiers using convolutional neural networks (CNNs). CNNs can identify features directly from a received signal when appropriately trained. They trained the classifier by taking the magnitude, phase, real, and imaginary parts of the received signal for CNN-based automatic modulation classification, which is insensitive to phase and frequency deviation. Also, in order to make the proposed classifier more robust, they generated signal-to-noise ratio, phase offset, and frequency offset in the training signal set. They compared the proposed methods with conventional automatic modulation classification methods using higher-order cumulants, order statistics, and a CNN-based method. The simulation results show that the proposed methods outperform the conventional methods not only in the additive white Gaussian noise channel but also in the channels having severe phase or frequency offsets.