Phase-preserving gray box model for robust sonar signal separation

Passive sonar systems monitor underwater acoustic signals from diverse sources, including vessels, submarines. However, separating these signals from noise and overlapping sources remains a significant challenge. Traditional spectrogram-based methods often lose critical phase information, reducing separation performance, while existing deep learning models adapted from speech signal separation lack interpretability, limiting practical application in sonar systems. This study introduces Cycle Conv-Transformer TasNet, a novel gray box model for underwater signal separation that addresses these limitations. By leveraging time series representations of signals and preserving phase information, the model achieves more robust separation. Built upon ConvTasNet, it incorporates Transformer-based self-attention to effectively capture temporal correlations. Furthermore, the model is designed as an Explanation-Producing System, providing interpretable visualizations of time-frequency regions influencing its decisions and enhancing explainability alignment with XAI principles. Additionally, cycle consistency enhances signal reconstruction reliability even with limited training data. The proposed model processes signals using a convolutional encoder, Transformer-based source separator, and decoder, enabling efficient and accurate signal separation. Experimental results demonstrate a 1.2× improvement in processing speed compared to DPRNN and a 3.9 dB gain in SI-SNRi, highlighting its superior performance in both speed and separation quality.