Single Channel Blind Source Separation Scheme with RFFI-assisted for Terrestrial-Satellite Links

Single-channel blind source separation (SCBSS) is essential for radio map construction by separating mixed signals in satellite communication systems. However, current SCBSS methods encounter significant challenges in satellite-to-ground link scenarios, where signals are severely attenuated due to long-distance transmission and atmospheric effects, resulting in low signal-to-noise ratio (SNR) and inadequate separation performance. To address these issues, we propose a novel radio frequency fingerprint identification (RFFI)-assisted SCBSS scheme, which utilizes RFFI to perform output-target assignment outside the neural network, thereby effectively enhancing the separation performance of deep learning-based SCBSS. In addition, by employing RFFI-assisted separation for multi-task training and integrating Transformer networks with signals transformed to the frequency domain via short-time Fourier transform (STFT), our approach enables comprehensive feature extraction even under low SNR conditions. The proposed method significantly reduces the impact of noise on separation performance. Experimental results show that, compared to existing methods, our approach achieves an 11.9% reduction in mean square error (MSE) at -10 dB SNR and a 29.7% reduction in bit error rate (BER), along with faster inference speed.