Underwater Acoustic Target Recognition Methods Based On Deep Learning Theory

Target recognition is an important issue in underwater target detection and acquisition and a difficult problem in underwater signal processing task.Currently,the mainstream approach for target classification and recognition is to extract features by signal processing techniques,and then distinguish the targets attributes with pattern recognition method.However,in practical applications,there have two main problems in traditional systems: the mismatch between training and testing condition,and the lack of label information.To solve aforementioned problems,this thesis focuses on the study of feature extraction,recognition methods and the integration of the whole recognition system.Feature learning and enhancement,target recognition and unsupervised clustering methods are presented based on deep learning,factor analysis and generative probabilistic methods.The presented methods are validated by the experiments on the three measured underwater target data-sets collected in real underwater circumstances.The work of this thesis includes:1)Spectrum multiplication method(SMM)and factor analysis are proposed to enhance the traditional target features.The features are easily effected by the variation of target working condition,leading to the mismatch between training and testing conditions.This thesis proposed a SMM in signal level to enhance target features in the real noisy environment.Besides,factor analysis technique is also applied to extract common information and improve the robustness of features among different time interval.The performances of two methods are validated with supervised recognition and unsupervised clustering systems respectively.2)Deep learning features are proposed to enhance the adaptability of features in different data-sets.Selection of the features that most suitable for current data-sets is an important problem in the underwater target recognition.Many feature extraction methods have been proposed for underwater signals,and the performance of those features vary from different data-sets.Deep learning techniques can learn the parameters of bottom network on the basis of the current data-sets and then extract the most expressive features to the observed data.The experiment indicates that the presented learning strategy can extract robust deep learning features in different data-sets.3)The stacked deep neural networks and their applications in underwater target recognition are studied.Deep neural networks can be continually trained by increasing new data.This thesis studied their applications and performances in the recognition of underwater acoustical signal.The experiment shows those methods not only inherit the properties of deep learning features,which are not weaker than the traditional method of combing the optimal features and classifiers,but also can improve the whole performance of the system on the basis of the original stacked method.4)A new convolutional neural network framework is established for underwater acoustic target recognition.A new convolutional neural network(CNN)framework is also proposed to directly map the underwater acoustical waveform signals to their labels.Different regularization strategies,like zeros-mean,L2 regularization,and orthographic restraint,are presented to restrain convolutional kernels and enhance the performance.Experiment shows that the presented CNN framework makes significant progress in the recognition performance.5)The unsupervised underwater target clustering method based on the traditional features and deep learning features are studied.To meet the challenges of label information lacking of underwater targets,clustering methods are studied.k-means,Gaussian mixture model(GMM)are introduced to analyse the data-sets without label information.This thesis also introduces a Dirichlet process based GMM(DP-GMM)to infer the amount of clusters.The combination of traditional features,deep learning features and different clustering methods are studied based on the underwater acoustical signal data-sets.The result is similar to the supervised target recognition that deep learning features are more robust than traditional features in the clustering task.6)The integration between feature extraction and recognition/clustering method are studied.In both supervise recognition and unsupervised clustering,the integrated method is proposed respectively.In this thesis,deep neural networks are firstly constructed in supervised target recognition system,in which the bottom network is applied to extract features from those signals automatically by different fully-collected layers or convolutional layers.Then softmax layer is added to output label information.Experimental results indicates that deep neural networks can improve the performance by the integration of the feature extraction and recognition phases.In unsupervised clustering,a deep generative clustering method is proposed based on the graphic model.Corresponding joint optimization algorithm is also proposed to integrate the feature extraction and clustering methods.This algorithm is implemented by Gibbs sampling and two subsystems are collaboratively functioned as a whole to cluster the features that extracted from the raw sources of the target signals.The performance of these methods are validated by comparing with other different clustering methods using different data-sets.