Automatic Modulation Recognition based on SE Module and CNN_ResNet

In response to the problem that most current automatic modulation recognition models cannot achieve high recognition rates while maintaining a certain training speed, this study proposes a neural network model based on depthwise separable convolution, residual connection, and channel attention mechanism. By using residual structures and Swish activation to alleviate gradient problems to support deep network training, dynamically optimizing feature channels using SE modules, and significantly reducing computational costs through depthwise separable convolution and global pooling, the model not only achieves lightweighting and ensures a certain training speed, but also has certain feature extraction capabilities, which can achieve high recognition rates. This study conducted comparative experiments to train different models in five SNR environments on the DeepSig RadioML 2018.01A dataset. The CNN_ResNet model can achieve a recognition rate of 92.67% in a 10dB environment, which is 20.07% higher than the basic CNN model. The results of the experiments demonstrate that the improved model exhibits a significantly higher recognition rate compared to other models, while maintaining a certain training speed.