Research On Underwater Acoustic Target Recognition Based On Deep Learning

Underwater acoustic target recognition in complex marine environment is a key technology and difficult problem in the field of underwater acoustic signal processing.It has important theoretical and practical significance in the development of marine resources and the protection of maritime rights and interests.This thesis studied the underwater acoustic target recognition method based on deep learning,and proposed several underwater acoustic target recognition model inspired by the biological neural mechanisms.These models effectively improved the accuracy and generalization of underwater acoustic target recognition system.The key contents and contributions of this thesis are as follows:(1)According to the generation mechanism of ship radiated noise and the basis of ship type classification,the characteristics of ship radiated noise of different ship types are analyzed profoundly.These analyses provide a priori knowledge of the main factors that restrict the intraclass compactness and inter-class separability of underwater acoustic targets.The framework of underwater acoustic target recognition system based on deep learning is proposed.The superiority of deep learning for underwater acoustic target recognition is expounded by comparing with the traditional underwater acoustic target recognition system.(2)For underwater acoustic target recognition under small sample condition,inspired by the competition mechanism of biological neurons,a competitive deep belief underwater acoustic target recognition model is proposed.The model mainly includes three learning mechanisms:unsupervised pre-training,competitive learning and feature selection.Firstly,unsupervised pretraining could use unlabeled underwater acoustic dataset to optimize the model parameters to a relatively reasonable space.Secondly,by integrating competitive learning into the training process of deep belief network,the feature learning of the deep belief network could improve the intra-class compactness and inter-class separability of underwater acoustic targets.Finally,feature selection based on mutual information is deployed to remove the redundant features and further optimize the structure of the network.Experiments based on measured ship radiated noise verified the effectiveness of the model for underwater acoustic target recognition tasks under small sample conditions.The proposed model could not only retain the learning ability of deep belief network,but also have a concise structure.The model have the ability to solve the problem that the deep belief network is prone to overfitting a small set of underwater acoustic dataset.(3)Extracting the information of the inter-class difference from the complex frequency structure of underwater acoustic signal is an effective way to improve the recognition performance of underwater acoustic target.Inspired by the frequency selectivity of the human cochlea,a deep auditory convolutional underwater acoustic target recognition model is proposed.Firstly,the frequency distribution information of underwater acoustic signal is obtained by convolution of auditory filter and underwater acoustic signal,and the auditory spectrum features of underwater acoustic target are obtained.Then,the multi-layer convolution operations are carried out in time-frequency space of the auditory spectrum to discover spectrotemporal patterns of the underwater acoustic signal,so as to extract the deep auditory feature.The model integrates the auditory filter learning and the task of underwater acoustic target recognition.The multi-layer convolution operation simulates the multi-level signal processing of auditory system,so as to extract the information from frequency structure of underwater acoustic signal.The underwater acoustic target recognition experiments based on measured data indicated the effectiveness of the proposed model.The analyze of convolution kernel confirmed the effect of the model on auditory filter optimization.(4)A large amount of underwater acoustic data over a long-time span has characteristics of low information value density and large distribution difference.According to the deep structure of human auditory system and a variety of auditory neural mechanisms,an ultra-deep auditory convolutional underwater acoustic target recognition model is proposed.The model is suitable for underwater acoustic target recognition under the condition of a large number of underwater acoustic data with complex distribution.Inspired by the frequency selectivity mechanism of cochlear,auditory filter and dilated convolution are used to perform convolution operation on underwater acoustic signal,then auditory features of the underwater acoustic target are obtained.The dilated convolution makes the model parameters more sparse,while preserving characteristics of the auditory filter.The proposed supervised feature calibration module is introduced in the model.The module could not only increase the optimization gradient of the auditory filter,but also calibrate the significance of extracted auditory feature.In the proposed model,multiscale convolution kernels and residual connections make it possible for increasing the number of layers and units to form the multistage of auditory system.These layers have the ability to learn spectro-temporal patterns of different scales.These layers could also preserve locality and reduce spectral variations from a large number of underwater acoustic data.A large number of measured underwater acoustic data were used for underwater acoustic target recognition experiments.The experiments confirmed the excellent recognition performance of the proposed model.Besides,the experimental results showed that the model has good robustness to navigation conditions of ships.(5)The applicability of competitive deep belief network,deep auditory convolutional neural network and ultra-deep auditory convolutional neural network in underwater acoustic target recognition is discussed.The structure and learning mechanisms of competitive deep belief network make it more suitable for underwater acoustic target recognition with small samples.The deep auditory convolutional neural network is suitable to extract the spectrum distribution information of underwater acoustic signals through the fusion of auditory filters and convolutional neural network.A variety of modules increase the depth and structural diversity of ultra-deep auditory convolutional neural network,which is more suitable for modeling large amounts of underwater acoustic data with large distribution differences and accomplish underwater acoustic target recognition.The underwater acoustic target recognition experiments were carried out with measured underwater acoustic data.The recognition performance of the proposed underwater acoustic target recognition model is compared.Based on the proposed ultra-deep auditory convolution underwater acoustic target recognition model,the transfer learning is studied and a number of indicators are proposed to evaluate the effectiveness of transfer learning.Transfer learning successfully expands the model’s generalization ability to different underwater acoustic target recognition tasks.