Research On Underwater Acoustic Signal Processing And Target Recognition For Underwater Attack And Defense System

Since the underwater attack and defense system(UADS)is supposed to work automatically long hours in the ocean,it should have the abilities to detect and classify sonar targets.Due to the complicated oceanic environment,acoustic signals usually possess the characteristics of low signal to noise ratio(SNR).Thus,the signal detecting and processing system has to extract the target features effectively and recognize these targets correctly in order to ensure the normal work of the UADS.According to the actual demand of UADS,this dissertation focuses on the technologies including underwater acoustic signals de-noising,underwater acoustic signals classification and sonar image recognition.The main works in this dissertation are outlined as follows:1.Firstly,a novel de-noising algorithm for ship-radiated signals is provided.We propose a threshold coefficient calculation algorithm for wavelet hard threshold de-noising,which uses a genetic adaptive threshold method to search for the optimum threshold in the sense of the minimum MSE.The proposed method not only runs more than four times faster,but also solves the problem that the threshold value is affected by the signal length.Secondly,for the underwater echo signal de-noising,we propose a wavelet de-noising algorithm based on adaptive threshold function,which could preserve the signal singularity features.The proposed algorithm suppresses the wavelet coefficients by using a continuous and differentiable threshold function.The function could be adjusted adaptively according to the signal wavelet coefficients.The corresponding simulation results demonstrate that the proposed model performs better performance in preserving the signal singularity features than the traditional threshold filtering method.Finally,we focus on the sonar image de-noising and present an image de-noising algorithm based on adaptive threshold function for the wavelet coefficients.The corresponding simulation results show that the proposed model has advantages of preserving the image edge features and overcoming the vision distortion.2.Aimed at the classification for the underwater echoes signal,we introduce the all-pole model and energy feature extraction method firstly.Then we propose a method to find the optimal wavelet basis based on the Shannon entropy and an algorithm to find the most reasonable wavelet transform layer based on the distinguish entropy.The simulation results of the underwater echoes classification in the pool demonstrate the excellent performance of the proposed method.Furthermore,considering the technique of passive sonar target classification,we introduce the feature extraction of power spectrum and the linear spectrum in low frequency,and then propose a novel adaptive genetic-back propagation algorithm for training neural network target classifier.The classification experiment conducted on the ship-radiated noise shows that the designed classification system has higher classification accuracy.3.The dissertation studies the sonar image recognition algorithm.Due to the fact that the sonar image is often covered by other things,the recognition accuracy would decrease seriously.Accordingly,we adopt the sparse representation to identify the sonar image,which shows strong robustness in solving the above mentioned problems.In order to extract the effective features from the sparse representation of the sonar image,a novel method is proposed to optimize the projection matrix on Compressed Sensing(CS)algorithm.The proposed method shrinks the off-diagonal entries of the Gram matrix corresponding to the mutual coherence between the projection matrix and sparse dictionary,and adopts a gradient descent approach based on the Wolf's-conditions to solve the optimization projection matrix.The vectors extracted by using the optimization projection matrix have smaller number of dimension but contain larger feature information.The simulations results demonstrate that the proposed algorithm based on sparse representation could improve the performance of classification accuracy.