Autocorrelation Convolution Networks Based on Deep Learning for Automatic Modulation Classification

Automatic modulation classification (AMC) is challenging but significant in the field of cognitive radio. Despite recent deep learning methods have dominated as the best performers for AMC, they are challenged by the practical problem in low signal-to-noise ratios (SNRs). In this paper, we propose novel autocorrelation convolution networks (ACNs) to capture periodic representation for communication signals. In ACNs, modulation modes are classified with the periodic local features under an autocorrelation convolution criterion. The experimental results demonstrate that ACNs achieve a great improvement that outperforms recent deep learning methods in low SNRs.