Ship radiated noise separation based on auditory scene analysis and deep learning

A method for separating ship-radiated noise from mixed signals has been developed, combining computational auditory scene analysis (CASA) with deep learning. This method follows the general framework of CASA and divides the separation process into two stages: auditory segmentation and auditory reorganization. In the auditory segmentation stage, the mixed signal is divided into time-frequency frames to construct auditory segments. A Dense-UNet is then employed to extract data features and construct separation masks. The Dense-UNet integrates the encoder-decoder structure of the traditional UNet with the dense connections of DenseNet, enabling efficient extraction of multi-scale features in the encoder and effective recovery of fine-grained signal structures in the decoder through skip connections and dense connections. In the auditory reorganization stage, the separated frame-level signals are re-adjusted and paired based on the correlation analysis of adjacent frames, thereby achieving the reorganization of the separated signals. Experiments conducted on actual ship-radiated noise dataset demonstrate that the proposed method achieves superior separation performance and stability compared to baseline networks, even with a reduced network scale.