Research On Underwater Acoustic Taget Recogntion Based On Attention Mechanism

As a hot research content in the field of hydroacoustics,underwater acoustic target recognition(UATR)is of great significance to marine exploration and national defense security.UATR mainly takes ship radiated noise as the research object.However,traditional classificaiton schemes often fail to meet the needs of actual scientific research,production and application in terms of model inference speed and accuracy.This thesis proposes two attention-based models that can improve the accuracy of underwater target identification while also enhancing model computational efficiency.The main research contents are as follows:1.The characteristics of ship radiated noise have been analyzed.The source and characteristics of ship radiated noise are introduced,a general feature extraction scheme for ship radiated noise is introduced,and three different cepstrum features are extracted while using modern signal processing methods and a research framework is established.2.To verify the accuracy of ship radiated noise information,this thesis extracts frequency bands of different categories of ship radiated noise from a public dataset.Five features are then selected and combined as input to the analysis,then uses uniform manifold approximation and projection dimensionality reduction techniques to eliminate redundant features,and finally adopts extreme learning machine algorithm for classification.The experiment proves that this scheme can effectively improve the accuracy of ship radiation noise.Compared with the dimensionality reduction algorithms such as principal component analysis and t-SNE,the uniform manifold approximation and projection dimensionality reduction technique achieves 86.11% accuracy,which is better than other two algorithms.3.A method for underwater acoustic target recognition based on convolutional attention recurrent neural network model is proposed.This model combines the advantages of convolutional neural networks,attention mechanisms,and recurrent neural networks,enabling better processing of sequential data.In this model,convolutional neural networks are used to extract local features of spectral features,while attention mechanisms are used to capture important information in spectral features.At the same time,recurrent neural networks can model long-term dependencies in feature vectors,thereby better processing time series data.By dynamically fusing the convolution with the attention module and the bidirectional LSTM,the model can learn more scale time series features,thereby enhancing the prediction accuracy of the model.4.A method for underwater acoustic target recognition based on attention multi branch network model is proposed.This network can effectively utilize multi-scale features of different depths in the network,including global and local feature clues of shallow finegrained features and deep features,and provide richer descriptions of spectral features.In addition,this chapter utilizes the spatial attention module and channel attention module in the convolutional attention module to couple them into a multi branch network,enabling the network to focus on key features,enhancing the network’s learning ability,and thereby improving the accuracy of target noise recognition.