Passive Ranging Method Of Underwater Acoustic Target Based On Residual Neural Network

Underwater acoustic passive ranging problem is always the key point of the underwater acoustic field,the hot and difficult,due to the complicated and changeable Marine environment,the underwater acoustic channel in time domain,frequency domain,the airspace has the propagation path,and the sound propagation characteristics is complicated,made the positioning method in traditional matching field environment mismatch arrays,leading to poor positioning performance problems such as mismatch.Is different from the dependence on the environment parameter matching of locating methods,this article obtains from the deep learning and deep learning good ability to autonomous learning from the original data hidden rules in data mining,network parameters is used to instruct the mapping relationship between input data and labels,to adapt to the sound source and underwater acoustic channel and hydrophone array composed of complex nonlinear system,so as to realize the prediction of the sound source location.To simple sound source acoustic ranging problem,this paper studies the traditional matching field orientation method,discusses the matching positioning technology,uses the real sound field sound pressure information to match with the copy field sound pressure information,discusses the difference of matching field location under different parameters,verifies and analyzes its feasibility and accuracy by using swellex96 maritime data.Aiming at the problems of environmental mismatch and unknown model parameters,the matching field location methods are compared,Based on convolution neural network,self defined fully connected neural network and residual neural network are proposed underwater target location under three neural network structures.Firstly,the model structure,selection and construction of model parameters of the three neural network structures are introduced,and the frequently used activation function,loss function and optimization function are introduced.In this paper,the underwater target location problem is transformed into a classification problem.The cross spectral density matrix of sound pressure field vector is used as the input of the model,and the output of the model is the probability Classification of possible types takes the maximum probability type as the predicted sound source location.Using Leakyrelu function as the activation function,the effects of data preprocessing,known sound source information and network structure on underwater target positioning performance are discussed.Compare narrowband and wideband sound sources,the feasibility and effectiveness of the algorithm are verified by the simulation data and swellex96 maritime data analysis.The neural network ranging method with known and unknown environmental parameters is discussed respectively.Among them,the neural network ranging method with known environmental parameters can achieve "off-line" training,and then make rapid prediction of actual data,which is more valuable in practical application.The neural network ranging method under unknown environment can reduce the dependence of ranging method on environmental factors.The results of simulation and actual data show that,compared with the matching field ranging algorithm which has strict requirements on environmental parameters and sound field modeling,the localization method based on neural network is more suitable for underwater target localization,and its localization performance is better than the matching field localization algorithm.Among them,the ranging performance of residual neural network(Res Net)is slightly better than that of convolutional neural network(CNN)than that of fully connected neural network(FC).