ESTMST-ST: An End-to-End Soft Threshold and Multiloss Self-Distillation Based Swin Transformer for Underwater Acoustic Signal Recognition
Fan Wu, Haiyang Yao, Zhao Zhongda, Xiaobo Zhao, Yuzhang Zang, Haiyan Wang · IEEE Transactions on Geoscience and Remote Sensing · 2024
Underwater acoustic signal recognition (UASR) is significant for marine life and ecological environment protection. However, 2-D fixed-parameter inputs are inadequate for adapting to the variable underwater acoustic environment, and learnable-parameter inputs with inductive bias priors in transformers lead to difficulties in model convergence. Additionally, the differing optimization objectives between noise reduction and recognition methods can cause signal distortion, hindering recognition accuracy. To address these issues, this article proposes an innovative end-to-end soft threshold Swin Transformer model based on a multiloss self-distillation training strategy (ESTMST-ST) for robust recognition of weak underwater acoustic targets. Building on the previously proposed time-frequency Swin Transformer (TFST), we design a learnable dual filter module (LDFM) that decomposes underwater acoustic signals in the frequency direction, with parameters obtained through model training. To improve the model’s antinoise performance, we incorporate a soft threshold strategy within TFST to reduce nonstationary interference in underwater acoustic signals. For enhanced robustness and training efficiency, we introduce a self-distillation training strategy with four specific loss functions in selected stage in TFST. Using publicly available datasets, ShipsEar and DeepShip, we conduct three experiments: fixed signal-to-noise ratio (SNR) UASR, multi-SNR UASR, and model generalization ability tests. The experimental results demonstrate that ESTMST-ST achieves superior performance (at least a 1.6 improvement in${F}1$scores and a 2.2 improvement in kappa coefficients) compared to five state-of-the-art methods across two open-source datasets.