Abstract
Single-channel blind source separation (SCBSS) aims to recover multiple source signals from a single-channel mixture of signals. Traditional maximum likelihood methods incur prohibitively computational complexity, while existing deep learning approaches often overlook domain-specific knowledge of communication, which limits their separation accuracy. To address this gap, we propose a complex-valued convolutional encoder-separator-decoder network (C2ESDNet) for efficient high-precision SCBSS of co-channel communication signals. Unlike real-valued approximations, our complex-valued network inherently models in-phase and quadrature components without information loss. The encoder-separator-decoder architecture conforms to the weighted superposition principle of SCBSS. Furthermore, stacked dilated convolution blocks are employed to capture multi-scale waveform features. More importantly, we analyze and prove that the conventional scale-invariant source-to-noise (SISNR) ratio loss fails to preserve amplitude-phase characteristics essential for demodulation. Therefore, we propose a demodulation-aware loss, where the mean squared error focuses on point-wise accuracy and SISNR enhances global phase fidelity. Additionally, we further extend C2ESDNet to symbol-aware C2ESDNet-S, which integrates waveform separation with downstream demodulation. It features an extended decoder that mirrors the synchronization-and-demodulation pipeline of communication receivers in a data-driven manner. Numerical results show that our methods achieve a lower bit error rate compared to state-of-the-art (SOTA) baselines. The visualization result shows superior waveform fidelity in the separated signals of C2ESDNet compared to the SOTA baseline. Ablation studies confirm the critical contribution of the demodulation-aware loss to the performance gains.

This publication has 42 references indexed in Scilit: