High-Order Modulation Recognition based on Complex-valued Residual Network

With the rapid development of artificial intelligence technologies, deep learning methods are widely used in the field of modulation recognition of communication signals due to their superior performance. However, most real-valued deep neural networks ignore the phase-related features in the complex signals, resulting in poor performance in higher-order modulation recognition. To address the above problems, we propose a higher-order modulation recognition method based on complex-valued residual network (CV-ResNet). The network can directly extract the amplitude-related features and phase-related features in the complex signals, so it can better recognize the higher-order modulations which are more difficult to distinguish in the time domain. The experimental results on a public signal dataset containing 12 higher-order modulations show that, the complex-valued residual network achieves the best recognition accuracy for the higher-order modulations compared to the complex-valued convolutional neural network and the real-valued residual network. Moreover, the recognition accuracy of the complex-valued residual network continuously improves with an increase of the signal-to-noise ratios (SNRs) and tends to be stable when the SNRs exceed 18 dB.