Underwater Acoustic Target Recognition Based On Multi-Dimensional Feature And Attention Neural Network

Underwater acoustic target recognition technology is a research priority in the field of underwater acoustics in various countries,and is also an important technology in areas such as ocean detection,underwater navigation,and marine ecological environment monitoring.However,due to the complex marine environment,underwater acoustic data is difficult to obtain,and the development of ship acoustic stealth technology makes it difficult to achieve the expected accuracy in target recognition.Therefore,accurately and rapidly identifying underwater acoustic targets remains a challenge in underwater acoustic target recognition technology.Additionally,the complex and diverse underwater environment often makes it difficult for a single feature to reflect the various characteristics of the target.To address these issues,this paper introduces deep learning methods into underwater acoustic target recognition and designs a model based on multi-dimensional feature fusion and attention mechanisms.Specifically,this paper studies the characteristics of ship radiation noise and the principles of convolutional neural networks and residual networks,extracts LOFAR spectrum features,MFCC features,and self-encoding features of measured ship radiation noise signals as input to the network model,and realizes the recognition of underwater acoustic targets.The paper also compares and analyzes the recognition performance of models with multi-dimensional feature fusion and single-type feature.The experimental results show that fused features have better performance.Furthermore,considering the temporal nature of underwater acoustic target signals,an attention mechanism is introduced into the underwater acoustic target recognition task.The principles of attention mechanisms and self-attention weight calculation are studied,and an attention feature fusion module is constructed.By introducing attention during the multidimensional feature fusion stage,the fused features with stronger representation information are obtained.The effectiveness of the attention model is verified by comparing the recognition results on residual network models and attention models.The feasibility of multi-dimensional feature fusion is also demonstrated by comparing the recognition results of single-type features and fused features.