Studies On Classification And Recognition Of Underwater Acoustic Targets Based On Deep Learning

In order to cope with the international changes and accelerate the construction of maritime power,China has continuously increased efforts in deep-sea observations and investigations every year.At present,traditional underwater acoustic recognition algorithms generally have the problems of low computational efficiency,dependence on manual adjustment,low reliability,and poor generalization.Contrary to the bottlenecks of traditional methods,the emerging deep learning(DL)methods still have broad development space in the field of underwater acoustic target recognition.On this basis,based on the signal analysis method and the deep learning method,this paper proposes several highly separable feature extraction algorithms of underwater acoustic target,and combines feature fusion and decision fusion strategies to propose a number of high-precision deep learning classification and recognition models,aiming to improve the performances of underwater acoustic target classification and recognition methods.This paper focus on two parts:(1)As for the measured underwater acoustic signals,the extraction of more robust and separable underwater acoustic features is studied based on signal analysis methods and DL methods,respectively;(2)For diffierent underwater acoustic features,construct high-precision deep learning models,several highly reliable deep learning model are flexibly constructed to improve the generalization ability and classification accuracy of models.The main contributions are listed as below:1.By analyzing the physical mechanism of underwater acoustic signals,from the perspective of separating signal and noise and improving feature separability,several underwater acoustic feature extraction algorithms with stronger detection performance and higher separability are studied and improved.To solve the problem of weak proprller shaft frequency detection ability of traditional spectrum analysis,an improved method for cyclic shaft frequency harmonics combined with wavelet decomposition and cyclic modulation spectrum is proposed,which realized the synchronous detection of shaft frequency and blade frequency.To overcome the spectrum leakage and non-Gaussian noise interference caused by the cyclic modulation spectrum,the cyclic bicoherence spectrum(CBS)extraction method based on cross-correlation and all-phase filtering is proposed,which effectively enhances the robustness of the CBS and the accuracy of harmonic line spectrum.To improve the auditory feature extraction method to meet the characteristics of underwater acoustic signal reception,CBS and principal component analysis are introduced into the processing of Mel cepstral to propose the Cyclic Bispectrum Compressd Mel-frequency cepstral coefficients feature,which greatly reduces the marine ambient noise interference and increases the feature separability.In addition,based on the studies on the deep-sea sound field interference characteristic,the simulation and extraction methods of the interference fringe features are introduced.2.Drawn on the theories of deep learning,several more generalized deep learning models for feature extraction are proposed for Mel-frequency cepstral coefficient,spectrum and sound field interference fringes,aims to make up for the deficiencies of signal analysis methods.In order to improve the generalization of Mel-frequency cepstral feature,a stack sparse autoencoder is proposed to compress the sparse Mel-frequency cepstral coefficients in batches,which reduces data redundancy and saves computational efficiency.When stable denoising features cannot be obtained directly,a correlation noise reduction feature learning framework is proposed based on the adaptive correlation of target signal and background noise,which solves the issue of dynamic deviation in different application scenarios of traditional denoising methods.In view of optimizing the fuzzy interference fringes,a deep belief network is built to generate clear and separable fringe features based on the simulated fringe images.On this basis,an adversarial generation network is constructed to realize the extension of the reception range of interference fringe features in the same scene.3.Based on supervised and unsupervised learning,respectively,the classification and recognition frameworks are studied and proposed for underwater and surface targets.Based on the principle of supervised learning,the separability of signal analysis features and DL features are compared and analyzed.The feature filtering strategy is used to build a feature fusion framework to integrate multiple features,resulting in a feature fusion model with richer information and higher classification accuracy than a single feature.In addition,based on the parallel training and decision fusion strategy,a DL classification and a recognition model are separately proposed to improve the accuracy and reliability of algorithms.Based on the theory of unsupervised clustering,the classification performances of underwater acoustic targets are compared via Gaussian mixture model and fuzzy C-means method,respectively.The experiment imples that the clustering performance of compressed Mel-frequency cepstral coefficient is superior to other features and it is more generalized for clustering underwater acoustic targets.4.To classify the datasets containing both labeled and unlabeled samples,the semi-supervised the binary classification based on interference fringe features is studied,and two semi-supervised classification models are proposed to achieve underwater and surface targets classification.To exploit the sensitive characteristic of interference fringe feature on target depth,the feasibility of the interference fringe feature in the semi-supervised binary classification task is studied.By improving the adversarial autoencoder,a convolutional adversarial autoencoders and adversarial autoencoding cooperative random forest models are proposed respectively,which aim to make full use of a small number of labeled samples to achieve effective classification of underwater targets without matching labels.It fills the blank in the studies on underwater acoustic target classification in the field of semi-supervised learning.Overall,this paper carried out algorithm innovations from the aspects of: improving the feature extraction methods of underwater acoustic signals,building the supervised deep neural network models and information fusion frameworks,and designing the semi-supervised classification models of underwater acoustic targets,resulting in the more separable fusion feature of underwater acoustic targets,and the classification and recognition models of underwater acoustic targets with higher classification performance.