基于CSA策略和改进ResNet的通信信号调制识别方法OA

Due to factors such as channel noise and multipath fading,the quality of modulated signals is significantly degraded.Current network models have limitations in handling such modulation signal recognition tasks,and hybrid network models are usually characterized by complex structures and high computational costs.To address these problems,a new modulation recognition network model is constructed based on an improved Residual Neural Network(ResNet),incorporating a parallel-processing Channel-Space Joint Attention(CSA)mechanism.This model leverages the attention mechanism's ability to automatically focus on key channels and spatial features,thereby enhancing its representational capability of core features.Additionally,the two-level residual connection is introduced into the residual network to effectively reduce the risk of overfitting in deep models.Simulation verification on the public dataset RML2018.01a demonstrates that when the Signal to Noise Ratio(SNR)is higher than 2 dB,the average recognition accuracy of the model reaches 93%.When the SNR surpasses 12 dB,the recognition accuracy stabilizes above 97%.Compared to models like ResNet,ResNet50,and Long Short-Term Memory(LSTM),this model achieves significant improvements in both average recognition accuracy and accuracy under low-SNR scenarios.Compared to the Convolutional,Long Short-Term Memory,Fully Connected Deep Neural Networks(CLDNN)hybrid model,this model achieves comparable recognition accuracy while having less than one-tenth of the trainable parameters,with the relative performance loss controlled within a reasonable range.The experimental results indicate that this network model holds considerable potential in the field of Automatic Modulation Recognition(AMR).