long-range underwater acoustic hyperbolic frequency modulation signal denoising based on lightweight neural networka).

Marine acoustic techniques are crucial for exploring and developing the ocean. However, the underwater acoustic (UWA) hyperbolic frequency modulation (HFM) signal suffers from a low signal-to-noise ratio (SNR) in complex environments. Traditional denoising methods rely heavily on prior knowledge and perform poorly in non-Gaussian noise. While neural networks offer an alternative, their computational cost limits underwater applications. In this study, we propose a lightweight time-frequency gated fusion network (LTFG-Net) for long-range UWA signal denoising. First, to overcome scarce data, a dataset is built by transmitting HFM signals through a simulated channel and adding α-stable noise. Second, we design a lightweight denoising network with dual-branch parallel processing in the time-frequency domain. Moreover, by combining pre-training and transfer learning, a two-stage training strategy is proposed to quickly adapt to specific marine environments with limited measured data. Finally, simulation and sea experiments demonstrate the superiority of the proposed scheme. Specifically, numerical results show that LTFG-Net improves the Generalized SNR (GNR) from 0 dB to an average of 43.13 dB, with the correlation coefficient increasing from 0.1928 to 0.8954. In long-range sea trials, LTFG-Net boosts the GNR from 4.56 dB to 38.65 dB and the correlation coefficient from 0.3542 to 0.9122, with only 0.23 M parameters.