Source Depth Estimation Method Under Sound Speed Disturbance Based on Self-Coded Feature Selection

In this paper, we propose an algorithm for depth estimation of underwater sound sources, which is divided into four parts: calculating interference fringes of sound field, Variational Auto-Encoder (VAE), feature selection and one-dimensional deep residual network (1D-resnet). Firstly, we use the ray model to calculate the sound field interference fringe diagram corresponding to different sound source depth and different sound speed profile. The diagram takes a set frequency range and a set distance range as independent variables, and the sound field propagation loss as dependent variable. Then, the Variational Auto- Encoder is used for training to realize the dimensionality reduction of the interference fringe pattern of sound field, and the one-dimensional characteristics of the interference fringe pattern are obtained. Then the statistical characteristics of different characteristics are calculated to remove the characteristics which are greatly affected by the disturbance of sound speed. Finally, the selected features are sent into a one-dimensional deep residual network to realize the classification of sound source depth. This method can effectively suppress the influence of sound speed disturbance on the depth estimation of underwater sound source and the accuracy of source depth estimation is 97%.