Underwater Acoustic Target Recognition Method Based On Deep Learning

Intelligent underwater acoustic target recognition is an important research in the field of underwater acoustic,which has important economic and military value for marine resources development and marine security.The researches on underwater acoustic recognition technology mainly contains two aspects,one is feature extraction,and the other is recognition model construction.The traditional feature extraction algorithm mostly use a single feature,which cannot accurately describes the target characteristics.In addition,the traditional recognition methods are mostly based on the expert experience,which can not keep up with the increasing data furthermore,and it is difficult to extract the deep features of the target.In recent years,deep learning has made remarkable progress in the fields of image recognition and speech recognition.This thesis proposes an underwater target recognition method,which takes multidomain feature combination as input and uses convolutional neural network,long short-term memory network and residual block as backbone.The main contents and innovations of this thesis are as follows:(1)In order to expand the amount of available data,a time-domain transform data enhancement strategy is proposed.Due to the complexity of the marine environment,it is difficult to collect real ship noise data,leading to that there are few public data sets.This thesis considers the underwater audio waveform has the following characteristics: tone,loudness,quality,in the data enhancement,the data enhancement method based on time domain transformation is used to enhance the underwater acoustic data by time stretching,treble correction and Gaussian noise processing of the original signal.The training data set is increased to make the data set as diverse as possible,so that the training model has strong generalization ability;(2)A fusion feature extraction method based on Mel-Frequency Cepstral Coefficient and third-order Mel-spectrogram is proposed.Due to the high background noise intensity in the marine environment,the signal-to-noise ratio of actual underwater acoustic signals is often low,which greatly affects the ability of feature expression data.Inspired by the timbre perception mechanism of the auditory system,this thesis proposes a multi-feature fusion mechanism for underwater acoustic target radiated noise based on Mel-Frequency Cepstral Coefficient,Melspectrogram and its first-order difference and second-order difference,and constructs highdimensional data features,which contain rich target information;(3)A composite neural network model is proposed to improve the performance of deep network in intelligent recognition.The ship noise signal is generated by mechanical vibration,which signal is a time series signal.Considering the advantages of one-dimensional convolutional neural network and long short-term memory network in processing time-series signals,and the important role of residual blocks in deeper network models,this thesis proposes a composite neural network as a classification model based on the improved method of traditional recognition network model.Intelligent target recognition of underwater acoustic signals based on composite networks is carried out to better describe the internal information of data.The experimental results show that the data enhancement,multi-feature fusion and composite neural network proposed in this thesis can extract the feature information of underwater acoustic data well,the evaluation indexes obtained by network model training and testing are excellent.In the current main public datasets,excellent recognition accuracy can be obtained.