Research on Signal Enhancement of Visible Optical Audio Communication Based on Deep Learning in Complex Noise Environment at Sea

Aiming at the problem of serious degradation of visible light communication signal in complex noise environment at sea, this paper proposes a dynamic frequency domain separation enhancement method based on deep learning. By constructing a dual-branch network architecture, Short-time Fourier Transform (STFT) and Convolutional Neural Network (CNN) were combined to extract the time-frequency features of noise, and Long Short-Term Memory network (LSTM) was used to model the timing characteristics of optical signal to realize the dynamic separation of noise and signal. The physical propagation model of optical signal was innovatively introduced to constrain the network training, and the adaptive ability of the model to multi-period (day and night) noise characteristics was enhanced. Experimental results show that the proposed model improves the Signal-to-Noise Ratio (SNR) by 8.2 dB compared with the traditional optical filtering method, and reduces the Bit Error Rate (BER) to $1 \times 10^{-5}$, which is significantly better than the existing deep learning baseline models. This study provides theoretical support and practical solution for Marine dynamic optical noise suppression.