Shortwave Signal Recognition Based on the AP-TCN Model

The research on shortwave signal modulation recognition is of significant importance for battlefield situational awareness, intelligence acquisition, and analysis in military operations. It also holds substantial research value in civilian emergency communications and rescue operations. Due to the strong time-varying nature, low signal-to-noise ratio (SNR), and severe interference caused by the electromagnetic environment during the propagation of shortwave signals, existing recognition models insufficiently extract deep feature information of shortwave signals. This paper proposes an Attention Mechanism Adaptive average Pooling-Temporal Convolutional Networks (AP-TCN). This model utilizes a causal dilated convolution structure with progressively expanding feature maps to capture deep spatiotemporal features of the signal. It introduces average pooling to reduce model complexity and incorporates channel attention to further significantly enhance model performance. In the classification module, the adaptive average pooling layer replaces the fully connected layer to handle variable-length data. Experimental results show that this model achieves a recognition rate of 97.5<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> for shortwave signals at an SNR of 8dB, which is about 3% higher than models with higher parameter counts, such as LeNet and ResNet. Furthermore, it maintains a high recognition accuracy of 89.1% even at a low SNR of -4dB, indicating that the proposed model can accurately recognize complex signals.