Performance Analysis of Convolutional Neural Networks with Different Window Functions for Automatic Modulation Classification

Abstract

Automatic modulation classification is one of the main tasks for developing information security and radio signal surveillance systems. It is also becoming increasingly significant for spectrum monitoring, management, and secure communications. Recently, deep learning models have been widely applied in many fields due to their outstanding feature extraction and classification accuracy. In this paper, the automatic modulation classification performance of several deep convolutional neural networks (CNN) is analyzed on the Fast Fourier Transform (FFT) based signal spectrum and the Short Time Fourier Transform (STFT) based signal spectrogram. By using ResNet18 and MobileNet cross-combined with FFT and STFT input data, the simulation shows that the STFT data provides a higher AMC accuracy than that of FFT data. On the same STFT data, the ResNet18 model outperforms three other models (SqueezeNet, GoogleNet, and MobileNet) in classifying 26 modulation types under the influence of five levels of fading noise with SNRs from −20 dB to +18 dB. Besides, the impact of different window functions in STFT is also investigated. Numerical results indicate that the considered window functions cause an insignificant difference in the AMC accuracy.