Research On Underwater Acoustic Target Recognition Based On Convolutional Neural Network

Underwater acoustic target recognition is critical for ocean exploration and has significant applications in both military and civilian domains.However,the complex and dynamic ocean environment makes it a challenging task,further compounded by the scarcity of underwater acoustic data.This thesis focuses on classifier designing and feature extraction methods for underwater acoustic target recognition using deep learning and multi-feature fusion.Additionally,the problem of recognizing multiple targets is explored in underwater environments here.This thesis examines the generation mechanism of ship-radiated noise,investigates real-world ship-radiated noise,so as to provide a theoretical basis for the research on ship radiation noise identification.Furthermore,it is discovered that machine learning methods based on statistical learning are inadequate for underwater acoustic target recognition in realistic scenes,leading to the application of deep learning techniques.To address the issue of data scarcity,waveform augmentation and spectrum augmentation methods are employed to simulate acoustic signal propagation in the ocean,and the dataset are enriched in this way.Experimental results demonstrate that the data augmentation techniques utilized in this study effectively increase the quantity and quality of the data.To overcome issues with traditional convolutional neural networks,including network degradation,insufficient multi-scale representation ability,and easy overfitting,improvements are made to the standard network.These include adding residual structures to capture more original information,introducing a multi-scale structure to enhance multi-scale representation ability,implementing an attention mechanism to focus on important information in the frequency domain,and using online label smoothing to improve anti-overfitting ability.Combining the above improvements,the ECAPA-OLS(Emphasized Channel Attention,Propagation,and Aggregation-Online Label Smoothing)model is designed,and the experimental results verify the effectiveness of the model.To address the problem of incomplete characterization of signal characteristics by a single feature,the performance of different features on underwater acoustic target recognition is studied.Static and dynamic features are extracted and fused into the AMRG-Delta(AMS,MFCC,RASTAPLP,GFCC-Delta)feature,which significantly improves characterization capabilities.Experimental results show that the fusion features used in this study improve both intra-class aggregation and inter-class separation performance,leading to better performance in underwater acoustic target recognition.Considering the situations where multiple ships appear simultaneously in same areas of water,this thesis regenerates the underwater acoustic multi-target recognition dataset and designs the ECAPA-ASL(Emphasized Channel Attention,Propagation,and Aggregation-Asymmetric Loss)model,which incorporates the fusion features to complete the multi-target recognition task.Experimental results demonstrate that the combination of ECAPA-ASL+AMRG-Delta effectively recognizes multiple targets.