Deep learning for communication signal classification

Abstract

The ability to differentiate between different radio signals is important when using communication signals. This is achieved with modulation classification. In this study, the use of two deep learning approaches, the convolutional neural network and the gated recurrent unit for modulation classification of five different signal modulations schemes, Gaussian Frequency Shift Keying, Binary Phase Shift Keying, Broadcast Frequency Modulation, Double Sideband Amplitude Modulation and Single Sideband Amplitude Modulation are explored. The network performances are compared based on the classification accuracy of various test sets. Each network variation was trained with a range of signal sample sizes and the signal modulation classification accuracy evaluated for each network and sample variation. The overall test results for both networks indicate that the convolutional neural network outperforms the gated recurrent unit albeit by a small margin of 4% to 8%. The outcome of the study shows the exceptional potential for deep learning approaches in communication signal classification as seen from the promising results displayed by both networks evaluated in this study