Automatic modulation classification based deep learning with mixed feature

Abstract

The automatic modulation classification (AMC) plays an important and necessary role in the truncated wireless signal, which is used in modern communications. The proposed convolution neural network (CNN) for AMC is based on a method of feature expansion by integrating I/Q (time form) with r/Ɵ (polar form) in order to take advantage of two things: first, feature expansion helps to increase features; the second is that converting to polar form helps to increase classification accuracy for higher order modulation due to diversity in polar form. CNN consists of six blocks. Each block contains symmetric and asymmetric filters, as well as max and average pooling filters. This paper uses DeepSig: RadioML which is a dataset of 24 modulation classes. The proposed network has outperformed many recent papers in terms of classification accuracy for 24 modulation types, with a classification accuracy of up to 96.06 at an SNR=20 dB.