Automatic Modulation Classification Based on SNR Estimation via Two-Stage Convolutional Neural Networks

Abstract

Automatic modulation classification (AMC) based on constellation diagrams that own distinct modulation features performs well in Gaussian channels. However, proper constellation diagrams for AMC are difficult to obtain in non-Gaussian channels due to incorrect estimation of symbol timings, which puts bad effects on AMC greatly. This paper proposes a novel AMC method via two-stage convolutional neural networks (CNN). Specifically, constellation diagrams with respect to all timings of any signal consist of a group of multi-offset constellation diagrams (MCD). Then, the first-stage CNN estimates the signal-to-noise ratio (SNR) for each constellation diagram so that the constellation diagram owning the maximum SNR can be selected out from its MCD. Finally, the second-stage CNN can classify modulations based on the selected constellation diagrams. Comparing with seven conventional methods, experimental results demonstrate the remarkable improvement of the proposed method.