Lightweight Neural Network Denoising Model for Long-Range Underwater Acoustic Signals based on Frequency-Domain Features

Long-range underwater acoustic signal transmission faces great challenges of low signal-to-noise ratio caused by complex marine environments. Traditional denoising methods rely heavily on prior knowledge and experience severe performance degradation in non-Gaussian noise environments. In this study, we propose a lightweight neural network denoising model based on depthwise separable convolution in the frequency domain. To address the problem of scarce training samples, we transmit hyperbolic frequency modulation signals through a parabolic equation model simulated channel and superimpose alpha-stable distribution noise to build the dataset. We design a frequency-domain lightweight denoising network with noise estimation, magnitude processing, and phase correction modules based on the depthwise separable convolution architecture. Simulation experiments demonstrate that the proposed scheme achieves better noise reduction performance than traditional methods and classic deep learning approaches.