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

Area of Science:

• Marine acoustics
• Signal processing
• Machine learning for underwater applications

Background:

• Underwater acoustic (UWA) hyperbolic frequency modulation (HFM) signals face low signal-to-noise ratio (SNR) in complex environments.
• Traditional denoising methods struggle with non-Gaussian noise and require prior knowledge.
• Existing neural networks for denoising are computationally intensive for UWA applications.

Purpose of the Study:

• To propose a lightweight time-frequency gated fusion network (LTFG-Net) for effective long-range UWA signal denoising.
• To address the challenge of limited data in UWA environments.
• To develop a computationally efficient denoising solution for marine acoustics.

Main Methods:

• A dataset was created using simulated channels and alpha-stable noise to overcome data scarcity.
• A lightweight dual-branch network processing signals in the time-frequency domain was designed.
• A two-stage training strategy combining pre-training and transfer learning was employed for environmental adaptation.

Main Results:

• LTFG-Net significantly improved Generalized SNR (GNR) from 0 dB to 43.13 dB in simulations.
• The correlation coefficient increased from 0.1928 to 0.8954 in simulations.
• In sea trials, GNR improved from 4.56 dB to 38.65 dB, with correlation rising from 0.3542 to 0.9122, using only 0.23M parameters.

Conclusions:

• The proposed LTFG-Net demonstrates superior performance for UWA signal denoising compared to traditional methods.
• The lightweight design and efficient training strategy make LTFG-Net suitable for practical underwater acoustic applications.
• The study validates the effectiveness of LTFG-Net through both simulations and real-world sea experiments.