Single-Channel Blind Separation for the Same Frequency and Modulation Signals Based on Multi-domain Residual Block Convolutional Network

With the growing scarcity of spectrum resources, mixed signals with the same frequency and modulation pose significant challenges for Single-Channel Blind Source Separation (SCBSS). To address the high computational complexity and poor separation performance of existing methods, we propose a multi-domain residual block convolutional neural network based on an encoder-separator-decoder framework. The encoder employs multi-scale 1-D convolution to extract mixed-signal features, while the separator enhances feature representation through a multi-domain fusion mechanism, which helps extract deep feature mask in the time and frequency domain. The decoder reconstructs source signals using masks and encoded features. Simulations demonstrate that the proposed method effectively separates mixed signals with minimal parameter differences under the same frequency and modulation conditions, achieving correlation coefficients above 0.95 at SNR ≥ 10dB.