Exploration and optimization of underwater acoustic multibeam signal processing algorithm via deep learning and optical communication techniques

With the advancement of optical communication technologies, such as fiber-optic hydrophones and photonic crystal fibers, the high bandwidth and low noise characteristics of these systems provide new avenues for preprocessing underwater acoustic signals. These technologies enhance the quality of signal acquisition and offer potential for improving underwater acoustic communication systems. As the demand for exploration of marine resources continues to rise, related technologies in the field of underwater acoustic communication are increasingly gaining the attention of researchers from various sectors. Among them, technology for the classification and recognition of underwater acoustic multibeam signals based on deep learning combines the hot topic of artificial intelligence with traditional signal processing techniques and represents an important branch of research in the field of underwater acoustic communication. In traditional signal processing, time-frequency analysis is one of the effective methods for analyzing time-varying signals, which can accurately and efficiently estimate the instantaneous frequency of the signal at various moments. However, according to the uncertainty principle, time-frequency analysis based on traditional methods has an upper limit on time-frequency resolution, which often cannot meet the requirements of precision in fields such as military and exploration that demand high accuracy. This paper proposes a lightweight deep neural network for the classification of underwater acoustic multibeam signals. Drawing inspiration from the basic ideas of AlexNet and GoogLeNet, this deep learning model is implemented in a modular fashion. The network mainly consists of an improved multiscale feature extraction module innovatively assembled through the concept of contextual encoding. Compared to traditional deep neural network structures, the proposed model has better classification performance and fewer network parameters, thus reducing the consumption of computational resources in terms of both space and time.