Abstract

The rapid expansion of wireless communication systems has intensified the challenge of co-channel interference, which has significantly affected signal reliability and spectral efficiency. In dense communication environments, multiple transmitters often have shared the same frequency band, which has created overlapping signals at the receiver. Traditional blind source separation techniques, which have relied on statistical independence assumptions or matrix factorization strategies, have faced limitations when signals have exhibited complex temporal correlations and nonlinear distortions. These limitations have motivated the need for adaptive learning models that have captured deeper signal representations without extensive labeled datasets. This study has addressed the problem of separating mixed communication signals under severe co-channel interference conditions. Conventional supervised learning frameworks have required labeled mixtures and ground truth signals, which have remained difficult to obtain in real wireless deployments. As a result, a robust representation learning strategy has become essential for extracting meaningful signal structures from unlabeled observations. To overcome this limitation, the study has proposed a Self-Supervised Contrastive Representation Separation Network (SCRSN), which has utilized self-supervised representation learning for blind source separation. The proposed method has learned latent signal embeddings through a contrastive objective that has encouraged the model to distinguish between temporally consistent signal patterns and unrelated interference components. An encoder–decoder architecture has extracted hierarchical signal features, while a clustering-based separation module has reconstructed the independent source signals. The model has leveraged signal augmentation strategies that have generated positive and negative sample pairs without manual labeling, which has enabled efficient representation learning from raw signal mixtures. The experimental evaluation demonstrates that the proposed SCRSN framework achieves 94.1% signal separation accuracy at 20 dB SNR, which exceeds the performance of the ICA, NMF, and Deep Autoencoder BSS approaches. The method produces 23.4 dB Signal- to-Interference Ratio and 22.4 dB Signal-to-Distortion Ratio, which indicate strong interference suppression and signal reconstruction capability. The framework also reduces the reconstruction error to 0.021 Mean Squared Error, while maintaining an efficient computational time of 11.7 seconds for large signal inputs.

Authors

G.L. Krishna Shri
Vellore Institute of Technology, India

Keywords

Blind Source Separation, Co-Channel Interference, Self-Supervised Learning, Wireless Signal Processing, Representation Learning