Research On Underwater Acoustic Target Recognition Technology For Unbalanced Data Based On Optical Fiber

In order to realize target recognition by modulation and demodulation of fiber optic hydrophone system,the research on underwater acoustic target recognition is an important solution method,which has very important application value in military and civil affairs.Due to the special application background of optical fiber hydrophone in the ocean,there are two problems in underwater acoustic target recognition: Firstly,feature extraction and classifier design are often characterized by low stability and low generalization.Secondly,the classification of underwater acoustic data is often unbalanced,which leads to the decline of the recognition performance of the recognition system.Focusing on these two existing problems,this paper studies the key technologies of fiber-optic underwater acoustic target recognition based on imbalanced data.To solve these two problems,a fusion recognition method based on trigonometric cos x weighted cross entropy loss function and multi-scale residual convolutional neural network with attention mechanism(named MR-CNN-A network)is proposed.This method starts from two aspects: one is to design a more robust front-end network model framework MR-CN-A,and the other is to design a back-end loss function of the network framework to resist the influence of imbalanced data.Firstly,based on the characteristics of experimental data and three feature extraction methods,single-feature multi-scale feature fusion recognition based on attention mechanism,multi-domain feature fusion recognition based on attention mechanism and anti-imbalanced data recognition based on triangular cos x weighted cross entropy loss function are studied respectively.All three of them have optimized their parameters through a large number of experiments,deeply studied the target recognition technology based on this method,and verified the effectiveness of this method.The main innovations and research achievements of this paper are as follows:1.Aiming at the low robustness of current recognition methods,a multi-scale underwater acoustic target feature fusion algorithm with attention mechanism is proposed.Experiments are carried out from two perspectives of the algorithm.One is from single feature MFCC.,the multi-resolution analysis relationship is formed according to the multi-scale convolution kernel and the feature graph,and the dominant feature weights are extracted and fused through the attention mechanism,so as to improve the robustness of the model in extracting target noise features and classification recognition from underwater acoustic data sets.Experimental results show that the recognition performance of the proposed network is more than 98% under the original data set,and it is more suitable for low SNR environment than other methods.The other is by modifying input and output parameters,the three domain features of MFCC,DEMON and HHT are simultaneously input into the network.Moreover,the attention mechanism is used to extract the weight value of each feature,to carry out fusion recognition.Experiments show that the method can achieve more than 98% recognition performance under the condition that the original DEMON and HHT are not well represented.These two algorithms are different representation surfaces of the same model,which verifies the robustness and superiority of the model respectively.2.Aiming at the problem of unbalanced data,an algorithm for underwater acoustic imbalanced data processing based on weighted cross entropy loss function of trigonometric function is proposed.By designing the cos x weighted loss function,this method resist the feature learning bias to most classes caused by the imbalanced underwater acoustic data.The effectiveness of the whole method is verified by experiments of loss function recognition.