Abstract:

To address the issues of high computational complexity and poor separation performance in separating co-frequency and co-modulation mixed signals under single-channel conditions, a multi-scale fusion recursive convolutional neural network (RCNN) based on time-domain convolution is proposed. The proposed architecture adopts an encoding-separation-decoding structure to achieve single-channel blind source separation. Specifically, the encoding module extracts the encoding feature of the mixed communication signal. Then, the separation module employs convolutional blocks of varying scale sizes to further extract feature information from the signal. Subsequently, it utilizes 1×1 convolutional blocks to capture both local and global information of the signal, estimating masks for each source signal. Finally, the decoding module uses these masks along with the encoding features of the mixed signal to reconstruct the waveforms of the source signals. Simulation results demonstrate that the proposed multi-scale fusion RCNN not only achieves separation of mixed communication signals with only minor parameter differences, but also reduces the number of parameters and computational complexity by approximately 62% and 41% respectively, compared to the U-Net. Moreover, the network exhibits strong generalization ability and can effectively handle the challenges posed by complex communication environments.

Key words: single channel blind source separation, deep learning, co-frequency and co-modulation signals separation, multi-scale fusion recursive convolutional neural network (RCNN), communication signal processing