Underwater Acoustic Classification Based On Deep Learning

Underwater target classification and recognition based on sound signal is a reliable way for underwater detection and target recognition,which is an important research content of underwater acoustic signal processing.A high recognition rate,wide range of underwater target recognition system depends on the extraction of features with good characterization and accurate and efficient classifier design.By machine learning from the data from the characteristics of self-expression,changing a long time to rely on theoretical analysis and experience to summarize the characteristics of artificial extraction of ideas.To avoid the loss of information when the signal is time-frequency transformed,it is more likely to obtain the feature vector which fully characterizes the signal characteristics.Combined with a reasonable classifier for accurate classification and recognition,which can effectively improve the underwater acoustic signal recognition accuracy and efficiency.In this paper,the depth of learning method for the use of underwater acoustic signal data feature extraction and classification recognition.The neural network structure of recurrent neural network(CNN)and recurrent neural network(RNN)are studied in this paper.The convolution neural network and recursive neural network which are suitable for processing deep structure of sound signal are designed to extract data.(HHT),Meyer cepstrum coefficients(MFCC),spectrograms and so on.In the meantime,the SVM,including the support vector machine(SVM),limit learning machine(ELM),linear regression(SOFTMAX)for multi-class recognition to compare the experimental results.And then the characteristics and advantages of machine learning features are analyzed.The experimental results are expected to be applied to the underwater multi-target voice recognition of passive sonar.The paper includes the following aspects:1,Introduce the research background,analyze the particularity and complexity of underwater target signal.The development course of underwater target recognition is briefly described,especially the development path under the thinking of physical analysis of time-frequency transformation.Including the simplest time-domain features,Fourier transform-based spectrum analysis,wavelet analysis,and Hilbert-Huang transform based on instantaneous frequency.The development of neural network in sound signal processing is introduced.Especially,Deep Learning has made a breakthrough in the field of speech recognition,which leads to the main idea of ??underwater acoustic target recognition using depth learning.2,This paper introduces several mature methods of signal processing and feature extraction,including data preprocessing,theoretical analysis of Hilbert-Huang transform,two methods of signal processing and feature extraction,including data preprocessing,two-part system design(feature extraction and classifier design)And implementation;MFCC principle and implementation.This paper introduces several practical classifiers: the basic principles of support vector machine(SVM);the principle of limit learning machine(ELM).3,For the latter part of the realization and comparison of the theory of bedding.The characteristics and development results of convolutional neural network(CNN)are introduced.The composition of convolution network and the training method are introduced in detail.A deep convolution neural network suitable for one-dimensional sequence signal input is designed to test and analyze the target signal.An improved convolution neural network structure for time signal is proposed and tested.4,The characteristics and development of recurrent neural network(RNN)are introduced.The advantages and development of recurrent neural network are introduced in detail.Introduce the basic structure and function principle of LSTM,and introduce the training method applied to RNN.The LSTM network,which is suitable for the input of sound signal,is designed to test and analyze the target signal.Compare the experimental results of different depth learning networks under the same task,and give the analysis.5,Comprehensive analysis of the experimental structure under various features and a variety of classifiers,draw conclusions and prospects for the future.