BSSepFormer: A Blind Source Separation Method for Underwater Scenarios
Yan Wang, Yuan An, Hao Zhang, Wei Kang Huang, Zhen Wang · IEEE Sensors Journal · 2025
Hydroacoustic communication and underwater acoustic target recognition constitute critical research areas in underwater acoustic signal processing. The complexity and uncertainty of the underwater acoustic environment pose significant challenges to underwater acoustic signal processing. Specifically, ocean ambient noise, reverberation as well as other interferences may induce communication instability and recognition performance degradation. Blind source separation (BSS) technique constitutes a fundamental approach to address these challenges. However, conventional statistical characteristic-based BSS methods exhibit limited effectiveness in fast time-varying and spatial-varying hydroacoustic channels. Deep learning-based BSS methods (e.g. TasNet) demonstrate enhanced capability in learning feature representations and nonlinear mapping relationships between mixed and source signals, thereby reducing the conventional methods’ reliance on statistical characteristics. Motivated by the exceptional performance of Transformer in speech separation tasks, we propose an early-split Transformer-based blind source separation (BSSepFormer) model for underwater scenarios. To mitigate performance degradation when processing mixed signals with analogous time-frequency characteristics, the proposed method employs the early-split strategy through a dedicated split module integrated between the encoder and decoder for separating the feature representations extracted from different sources. Furthermore, to optimize computational efficiency, the additive attention mechanism is introduced to substituting the dot-product self-attention mechanism in Transformer. To evaluate the effectiveness, the proposed BSSepFormer is validated on UAMix and ShipsearMix, achieving separation performance improvements of 1.1 dB and 0.9 dB respectively over existing state-of-the-art methods. Experimental results demonstrate that the proposed method not only achieves better separation performance in different scenarios, but also enables accurate source signal estimation.