Underwater Acoustic Signal Classification And Recognition Method Based On Improved Deep Learning Network

With the increasing demand for the exploration of Marine resources,the related technologies in the underwater acoustic communication are paid more and more attention by researchers from all circles.Among them,underwater acoustic signal classification and recognition technology based on deep learning combines the current hot Artificial Intelligence(AI)and traditional signal processing technology,and it is an important branch of underwater acoustic communication research.In the traditional signal processing technology,Time-Frequency Analysis(TFA)is one of the effective methods for analyzing time-varying signals,which is Instantaneous Frequency(IF)of the signal to be analyzed at every moment.It has been a hot topic in the field of signal processing for years.With the development of this technology,more and more advanced time-frequency analysis methods have been proposed,and the results of time-frequency analysis have become more and more refined.However,according to the uncertainty principle,the time-frequency analysis based on the traditional method has the upper limit of time-frequency resolution,which often cannot meet the task requirements in the military,exploration and other fields that require high accuracy of the results.Along with the development direction of the current time-frequency analysis technology,thesis designs a novel time-frequency analysis method and combines with the improved deep learning network to conduct intelligent classification of underwater acoustic signals.The specific work is as follows:Firstly,a novel time-frequency analysis method,Multi-Angle Optimal Chirplet Transform(MAOCT),is proposed,which can solve the problems of low resolution,low accuracy and low anti-interference performance of the analysis results in the time-frequency analysis.Based on the time-frequency rotation characteristics of the chirp transform,the MAOCT algorithm matches the signal to be analyzed as best as possible with the optimal frequency resolution of the current moment in each piecewise time window,which improves the time-frequency resolution of the final analysis results.Theoretical analysis and simulation results show that the proposed method can describe the time-frequency characteristics of nonlinear analog signal and measured underwater acoustic signal well,and is superior to the traditional time-frequency analysis method in various performance indexes.Moreover,it still has good performance in the environment with severe Gaussian white noise interference,and has high adaptability to the processing of underwater acoustic signal.Secondly,a lightweight deep n Eural Network for classification,Light-Weight n Eural network(LWENet)is proposed.Based on the basic ideas of AlexNet and GoogLeNet,the LWENet model is implemented in a modular way in thesis.The network model is composed of an improved multi-scale feature extraction module,which is innovatively combined with the idea of context coding.Compared with the traditional deep neural network structure,LWENet has better classification effect and fewer network parameters,which reduces the loss of computing resources in space and time.Thirdly,based on the proposed time-frequency analysis method and light weight deep neural network,a classification and recognition system of underwater acoustic signals based on LWENet is constructed.Through the establishment of training set and test set on the Watkins Marine mammal biological data set,MAOCT method was used for time-frequency analysis,and then the results were sent to LWENet for classification and recognition.Through the comparison of different analysis methods and different classification networks,the experimental results show that: Compared with the existing conventional time-frequency analysis methods and deep learning classification networks,MAOCT method is used to analyze the data set,while LWENet-based classification recognition network is used to classify and recognize underwater acoustic signals,which has higher training efficiency and classification recognition accuracy,and is not easy to overfit,gradient disappearance and other problems.The research results of thesis can provide theoretical support for the passive classification and recognition of Marine acoustic signals,and is expected to be further popularized and applied in Marine resource detection,passive sonar and other fields.