Research On Imbalanced Underwater Acoustic Target Detection Based On Deep Reinforcement Learning

Underwater acoustic target detection,as one of the research hotspots in the field of underwater acoustics,has great civil and military values.However,due to the influence of various factors such as the complex marine environment,short tracking time of targets,beamforming,and so on,the sample distribution in the obtained underwater acoustic data set is frequently imbalanced,resulting in a reduction in the performance of traditional underwater acoustic target detection methods.This thesis analyzes the current state of imbalanced underwater acoustic target detection and summarizes the method of combining underwater acoustic target detection and imbalanced learning to realize imbalanced underwater acoustic target detection in response to the actual needs of imbalanced underwater acoustic target detection.The main work and research of this thesis are as follows:Firstly,this thesis explains the theoretical basis of the study,including necessary signal processing technology of underwater acoustic target detection and basic theory of deep reinforcement learning chosen for imbalanced learning in this thesis.The part of underwater acoustic signal processing technology mainly introduces the theories of beamforming,reverberation suppression and feature processing,and the part of deep reinforcement learning theory mainly introduces Markov decision process,Q-learning and deep Q-network.Secondly,this thesis proposes an audio feature acquisition strategy for imbalanced underwater acoustic dataset.After the imbalanced underwater acoustic dataset is bagged,a general audio feature set is used for feature extraction,and then a feature selection method is used to filter the appropriate number of features with high overlap and high weight as the feature set of the imbalanced dataset.The imbalanced underwater acoustic feature dataset obtained by this strategy can obtain better feature representation than the convolution based on the spectrograms,which can help to simplify the subsequent network structure and speed up training and predicting of the model.Then,this thesis transforms imbalanced underwater acoustic target detection into Markov decision process of sequential decision,and proposes an imbalanced classification model based on deep Q-network named DQNiDTC.Through the design of imbalanced reward function and dual termination condition,the influence of positive and negative samples on the model can be balanced in the process of gradient descent,thereby achieving efficient imbalanced two-class and imbalanced multi-class underwater acoustic target detection.Finally,experiments of imbalanced two-class and imbalanced multi-class underwater acoustic target detection are carried out on the ocean dataset,and the experimental results verify the performance of the proposed algorithm.This thesis also implements DQNiDTC model deployment on Windows platform,which can meet the quasi-real-time requirement of system through multi-process acceleration processing.