| With the depletion of terrestrial resources due to the daily consumption of human activities,the ocean will inevitably become one of the most important sources of resources for mankind in the future,and the corresponding hydroacoustic target recognition technology will become an integral part of mankind’s exploration of marine resources.However,the complex and changing ocean environment,the prevalence of various ocean noises and the increasingly obscure characteristics of underwater targets make the identification of underwater target recognition a challenging problem.Based on hydroacoustic signal preprocessing,this paper explores and studies the classification and recognition of classical hydroacoustic target-ship signals from the non-linear,non-smooth and non-Gaussian characteristics of hydroacoustic signals.Firstly,Aiming at the problem that the actual ship signals are full of natural background noises,this paper denoises the ship signal.Considering the problems of empirical mode decomposition and ensemble empirical mode decomposition with modal aliasing and weak robustness,variational mode decomposition is used to decompose the ship signal,and the common denoising method combing the correlation coefficient on the basis of this decomposition algorithm to divide the signal into two parts: noise components and signal components,and directly discarding the noise components will often cause the loss of some effective signals.So on this basis,permutation entropy is introduced,and the noise components are further divided into pure noise components and noise dominant components.Based on wavelet soft threshold denoising,the dominant noise component is processed and reconstructed with the signal components.The newly proposed denoising method achieves good results in the denoising of simulated signals and actual ship signals,which not only greatly improves the signal-to-noise ratio but also reduces its mean squared error.Secondly,considering that the traditional feature extraction process is complex and the efficiency is low,and the resolution of the low-band signal of the ship target is not high.The spectral features of the ship signal are reconstructed,the mapping relationship between the pixel and the signal spectrum of the ship is determined,and the F-C mapping spectrogram is constructed.Compared with the traditional method,the new feature map not only simplifies the feature extraction process but also improves the low-frequency resolution of the ship signal,so as to prepare data for subsequent classification and recognition tasks.Furthermore,to improve the adaptability of target recognition network to sample ship data,and to shorten the training time of the network and improve the classification recognition efficiency,we can alternate structures of convolution and pooling layers are used to achieve a simplified structure of the hidden layer of the target classification recognition network,and the size of the convolution kernel is uniform.This method obtains better classification recognition results in relatively few network layers.The network is used to classify and identify three types of feature maps before and after denoising,and the recognition accuracy of the F-C mapping spectrogram after denoising reaches 96%,which further verifies the superiority of the proposed denoising algorithm and the new feature extraction method.Finally,to improve the practicality of hydroacoustic target identification method,a GUI system for hydroacoustic signal preprocessing is designed by combining signal denoising and feature extraction algorithms.Under this preprocessing system,the task of hydroacoustic signal denoising and feature extraction can be realized.At the same time,the actual collected ship signals are passed through the preprocessing GUI system to get the corresponding feature maps,and the experimental verification of classification recognition based on convolutional neural network method is carried out,which further notes the effectiveness of the method in this paper. |
PDF Full Text Request