Angle Feature Extraction And Recognition Of Underwater Target Echo Based On Depth Neural Network

In actual underwater detection,when an active sonar is used to identify the attitude and orientation of an underwater artificial quiet small target,the incident angle of the sound wave to the target is unknown,it is influenced by underwater environmental noise or other factors,the target echo exists interference and distortion,it is difficult to extract features that can characterize target echo at different incident angles.With the continuous development of underwater acoustic technology,the observation data has also increased rapidly.Using existing underwater acoustic observation data to realize the attitude recognition of underwater quiet small target is an urgent problem.When the pulse width of the transmitted signal is wider than the time delay between two adjacent scattered echo components,each scattered echo component of the target produces serious aliasing in time domain and frequency domain,the time domain waveform and spectral structure are difficult to identify different incident angles as echo features.According to the formation mechanism and signal characteristics of the scattered echo components,analyze the structure of the acoustic scattering components in the target echo,and use two time-frequency analysis methods of the Wigner-Ville distribution and the fractional Fourier transform to extract the WVD time-frequency image characteristics of the target echo and the time-sequence structure features in the best fractional order domain?There are serious cross-term interferences in WVD time-frequency image features,and it is difficult to calculate the effective features such as the number of bright points and the intensity of the bright points in the echo based on the image information for the traditional pattern recognition classifier to judge.Existing cross-term removal methods usually cause a decrease in time-frequency resolution,leading to the loss of feature information.Using two types of deep convolutional neural network models(Alexnet and Resnet),through the iteration of the multilayer network,constructing a deep network structure to learn the mapping relationship between the incident angle and the echo WVD time-frequency image,which overcomes the previous pattern recognition algorithms in high-dimensional data Insufficient characterization capabilities.The results show that compared with the WVD-Alexnet model,the WVD-Resnet model achieves a higher recognition rate,and effectively avoids the occurrence of overfitting.Time-sequence structure features in the optimal fractional order domain have the problem of high dimensions.The traditional dimensionality reduction can also lead to the loss of feature information.There is a lack of quantitative criteria for the study of the time-delay between elastic scattering echoes in the time-sequence structure at different angles.Using long short-term memory neural networks(LSTMs)in recurrent neural networks,features between time series information can be stored in the hidden state of the network,complex function mapping relationships can be learned,and dynamic modeling of sequence signals can be completed.The structure composition and training process of the two types of deep neural networks are analyzed and deduced,the parameters of the model and the selection of the optimization algorithm are established.The underwater target echo angle recognition model is constructed to realize the classification and recognition of echoes at different angles.Compared with the model using time-domain waveform and spectral structure as input features,the two deep neural network models proposed in this paper,WVD-Resnet and FRFT-LSTMs,have significant advantages in various evaluation indicators,which verify the effectiveness of the proposed method.