DOA Estimation Of Vector Hydrophone Based On Deep Learning

Compared with scalar hydrophone,vector hydrophone array can often obtain more sound field information.Vector hydrophone has sound pressure receiving channel of scalar hydrophone and vibration velocity receiving channel.For DOA estimation,the more information represents the higher estimation accuracy.Traditional DOA estimation algorithms often face the problem of poor robustness and poor effect in underwater low signal-to-noise ratio environment.At the same time,the complexity of traditional algorithms is often too high to meet the timeliness requirements of azimuth estimation.With the rapid development of deep learning,it has brought a new direction for target positioning.In this thesis,the vector hydrophone uniform linear array is taken as the research object,and the depth learning knowledge is introduced as a new underwater azimuth estimation method.Its specific research includes:(1)In the face of underwater low signal-to-noise ratio environment,the traditional array azimuth estimation algorithm has poor robustness and poor effect,and the traditional algorithm is often too complex.The depth neural network is introduced as the network model as a new underwater azimuth estimation method.The covariance matrix received by the vector array is decomposed into real and imaginary matrices,which are regarded as two parallel real matrices for input.The simulation results show that the accuracy of underwater azimuth estimation algorithm based on depth neural network is slightly better than that of traditional azimuth estimation algorithm under the condition of low signal-to-noise ratio.(2)For the deep neural network,with the increase of the number of full connection layers,the network model parameters increase sharply,resulting in the decline of model training efficiency,which is easy to cause the problem of over fitting or under fitting.The convolutional neural network,which is the fastest developing at present,is used as the training model of azimuth estimation.The simulation results show that the convolutional neural network model has a more robust estimation performance under the condition of low signal-to-noise ratio,and its accuracy is about 10% higher than that of the traditional algorithm.(3)Aiming at the problem that the scalar input-output mode of traditional neural network leads to insufficient data feature extraction or even loss of some features,a capsule network is introduced to realize the input-output of data vector through dynamic routing.The underwater vector DOA estimation network model based on deep learning is optimized and improved.The experimental results show that the accuracy and stability of the capsule network azimuth estimation model are greatly improved compared with other algorithms in the environment of low signal-to-noise ratio,and meet the rapidity requirements of DOA estimation.