Blind Source Separation Algorithm Based on Improved Wave-U-Net Skip Connection

This paper proposes a full-scale skip connection model based on Wave-U-Net to further improve the separation performance of existing blind source separation algorithms. First， to solve the problem of signal feature loss during the Wave-U-Net Down-sampling process， the model adds a convolution operation to the skip connection. By connecting the feature maps of different time scales， it effectively combines the shallow features of the signal and Deep features improve the separation performance of the model. Aiming at the problem of the optimal depth of Wave-U-Net and the excessive parameters of the full-scale skip connection model， this paper further proposes a multi-scale skip connection model. In the multi-scale skip connection model， the convolution operation in skip connection is replaced by embedding Wave-U-Net of different depths， which reduces the model parameters at the expense of part of the separation performance. The model shares the Down-sampling block to reduce the model training time and the influence of the optimal depth of the model. Simulation experiments show that compared with other baseline models， the two models proposed in this paper can significantly improve the signal separation performance， and the SDR， SIR， and SAR increase awards are nearly 3~4 dB.