Deep State Space Transformer for underwater acoustic recognition.

Underwater acoustic target recognition is crucial in marine engineering and information technology. However, the challenging nature of this task arises from the low radiated noise level of targets compared to ocean ambient noise and the physical complexities such as Doppler and multipath effects, leading to the issue of “indistinctness.” The emergence of deep learning has revolutionized underwater acoustic target recognition, drawing significant attention to extracting more effective hidden features from underwater acoustic signals using advanced learning methods. This paper proposes a novel Deep State Space Transformer (DSS-Transformer) model, which addresses the long-duration correlations present in underwater acoustic target signals. Leveraging the state space mapping method and bilinear approach in the Transformer, the DSS-Transformer directly extracts practical features from underwater acoustic signals and benefits from the multi-scale feature extraction capabilities of the Transformer model. Through experimentation on the Shipsear dataset, the DSS-Transformer demonstrates promising recognition results while substantially reducing the computational complexity associated with underwater acoustic target recognition.