Research On Sound Source Localization Method Based On Deep Learning

Underwater sound source target location technology is indispensable in the exploration of marine resources and the location of hostile submarine targets in the naval battlefield.Its importance in the field of underwater sound and national defense is self-evident.At present,the traditional method of underwater sound source localization is to match the sound source signal with the established sound field model,but there are a series of problems such as unknown ocean parameters and environmental mismatch.With the excellent performance of deep learning in various fields,according to its characteristics of being unaffected by environmental factors and its potential ability in the direction of acoustic signal processing,this thesis proposes an underwater sound source localization method based on deep learning.The main work of this thesis is as follows:First of all,in view of the multi-dimensional information of the target contained in the underwater sound source signal,the complex and changeable features are difficult to extract and the subtle features are easily ignored.Through the study of convolution operations in deep learning and the application of time convolutional networks in time series tasks,A feature extraction method of underwater acoustic signal based on multi-dimensional perception is proposed.This method is a network model with a multi-scale deformable convolution filter bank structure.Aiming at the time domain and frequency domain characteristics of acoustic signals,the model uses convolution kernels of multiple scales and samples in each convolution kernel.An offset value is added to the point area to improve the range and ability of feature perception.Aiming at the problems of many parameters and high computational time complexity in the model,the network model is light-weighted and improved,and different convolution kernels are used for different input channels to convolve,thereby improving the feature extraction ability and efficiency of the model in this chapter.Then regression analysis proposed in this thesis to deal with the localization problem.Aiming at the impact of predicting abnormal points,the current deep learning method to achieve sound source localization has the problem of easily deviating from the true value.By analyzing the characteristics of the sound source movement and reducing the impact of abnormal points,a loss function and evaluation standard based on error penalty weights are proposed.This method uses different error adjustment methods for different error values of the sound source position depth and distance,and The network parameters are adjusted by back-propagating the positive and negative errors of the predicted value to improve the effect of the regression model.Finally,the feature extraction effect of this model and the sound source localization regression experiment effect are verified experimentally on the underwater acoustic signal data set.First,the effectiveness of this model on the underwater acoustic signal feature extraction is verified on the data set,and it is compared with Lenet5 and TCN.The comparison of the models verifies that the model has a certain degree of improvement in the feature extraction of underwater acoustic signals.Secondly,for the regression experiment of sound source localization,the feasibility of the regression model is verified on the data set and the model has the effect of improving the accuracy of the sound source location prediction in depth and distance.