A Complex-valued hybrid deep learning models for automatic modulation recognition

Modulation recognition methods with real-valued deep network models fail to fully leverage the complex nature of in-phase/quadrature(I/Q) signals. Automatic modulation recognition based on complex-valued operations has attracted the attention of researchers. A Complex-valued hybrid deep learning models for automatic modulation recognition is proposed, which is named CCTL-Net. The CCTL-Net employs a complex-valued Convolutional Neural Network (CNN) to extract local features from I/Q signals, thereby effectively preserving the complex-related information inherent to I/Q signals. To enhance the correlation of different channels with CNN, a complex-valued Squeeze-and-Excitation module is designed. The complex-valued Transformer and complex-valued LSTM are employed to extract global features from both the attention and temporal relation perspectives. To reduce mode complexity, a light model with smaller channels and complex-valued 1D convolutional networks was designed. CCTL-Net achieved average accuracy rates of 70.71% and 62.97% on the RML22 and RML2016.10a datasets, respectively. Compared with real-valued hybrid model and complex-valued convolutional networks, the experimental results show that CCTL-Net can improve recognition accuracy under lower complexity conditions.