Research On Underwater Acoustic Target Recognition Method Based On Automatic Feature Extraction

With the development of information technology,the perception and confrontation technology of marine information is valued by military powers.Sonar technology is the main means of underwater target information detection.Underwater acoustic target recognition is one of the key technologies of underwater information countermeasure.At present,underwater acoustic target recognition methods face the challenges of complex marine environment and few labeled samples.This paper comprehensively studies the theory of ship radiated noise,underwater acoustic target recognition and automatic feature extraction,and proposes an underwater acoustic target recognition method based on automatic feature extraction.The main contents of this paper are as follows:1.The generation mechanism and characteristics of ship radiated noise are studied,and the characteristics of ship radiated noise samples are characterized by power spectrum and demodulation spectrum.This paper studies manual feature extraction methods such as GFCC,EMD and KL transform,as well as two pattern recognition algorithms: Gaussian mixture model(GMM)and BP neural network.Finally,two classifiers are used to test the recognition effect of three manual features.2.Aiming at the problem of insufficient generalization ability of traditional manual features in complex marine environment,an automatic feature extraction method based on RBM auto-encoder is proposed.In this method,the RBM auto-encoder is constructed by stacking single-layer models,and the RBM auto encoder optimized globally is used for automatic feature extraction.Different from the traditional manual feature filtering information according to a specific model,RBM auto-encoder fits the probability distribution characteristics of the sample set layer by layer to realize the automatic feature extraction driven by data.This method makes full use of unlabeled data to improve the generalization ability of features,so as to effectively improve the recognition performance of target noise in complex marine environment.3.Due to the high acquisition cost of ship radiated noise,the sample size of ship radiated noise signal is often difficult to meet the needs of neural network training.This paper designs a data expansion method based on RBM reconstructor.In this method,the trained RBM automatic encoder is connected with a decoder with symmetrical structure and the same parameters to form an RBM reconstructor.Through the data expansion of the original samples by RBM reconstructor,we can obtain new samples that meet the overall probability characteristics of the sample set,increase the diversity of the sample set,and improve the performance of the recognition system.4.An underwater acoustic target recognition system based on automatic feature extraction is built.The sample set based on the measured ship radiated noise is constructed,the performance of feature extraction of RBM auto-encoder is evaluated,the improvement of system recognition performance by expanding data samples with RBM reconstructor is evaluated,and the recognition effects of power spectrum input,demodulation spectrum input and fusion spectrum input are compared to verify the effectiveness of this work.