Underwater Acoustic Target Classification Based on Multi-Feature Fusion

The marine environment contains not only undulating seabeds and inhomogeneous water layer structures but also abundant natural and artificial noise. The low signal-to-noise ratio (SNR) environment caused by noise is the greatest obstacle to underwater acoustic target recognition, significantly increasing the difficulty of target classification in the acoustic field. In recent years, with the development of deep learning and artificial intelligence, new intelligent target classification methods have gradually emerged in the field of underwater acoustic target recognition. This paper introduces an attention mechanism and proposes a lightweight convolutional network model for underwater acoustic target classification. This method comprehensively utilizes two key features of ship noise: LOFAR spectrum and Mel spectrum. It demonstrates the inter-class separability of different acoustic features through analysis, and subsequently verifies the effectiveness of the proposed method by comparing and analyzing the classification results of various acoustic features based on convolutional neural networks. Multiple experiments using the Monte Carlo method show that the average classification accuracy of the model reaches 93.24%, with the minimum accuracy being 85.8%, indicating its promising application prospects in the field of underwater acoustic target classification.