Automatic Modulation Classification Based on High-order Cumulants Using Complex-valued Neural Networks

A novel automatic modulation classification (AMC) method is proposed based on high-order cumulants (HOC) via complex-valued neural networks. Benefited from the insensitiveness of HOC to noises, various orders of HOC are attached to the baseline signal, and leveraged as incipient features for deep features mining. In AMC, the baseline signal and HOC are all complex-valued and reserve plenty of phase information, which is useful for modulation, especially phase modulation recognition. However, conventional deep learning based AMC methods always treat a complex-valued signal as two columns of real-valued signals so that they cannot utilize phase information directly. To this end, complex-valued convolutional neural networks (CVNN) are applied finally. In addition, the HOC features, which alleviate the phase information contamination by noises, ensure the robustness of CVNN-based AMC. Experiments based on RadioML 2018.01A dataset demonstrates the effectiveness of the proposed method.