Att-DSConv: Separation of I/Q Signals Based on Self-Attention and Depthwise Separable Convolution

To solve the problem of low signal separation accuracy in communication scenes, an attentional Feature extraction network (Att-DSConv) separation method based on deep learning is proposed. First, the encoded feature representation is extracted from the mixed signal. The masks of sound source separation are then trained by stacked depth-separable convolution, and the attention layer is embedded to learn to separate the most relevant parts of the data. Finally, the processed features are mapped back to the original dimensions by a decoder. The problem of aliasing interference of the same frequency signal during signal transmission is effectively solved. The signal separation accuracy and robustness of this method are obviously superior to the other four methods used in the experiment.