DOA Estimation of Underwater Acoustic Signals Based on Deep Learning

Direction of arrival (DOA) estimation is an essential part of array signal processing and also one of the main tasks in the field of sonar arrays. The most commonly method among DOA estimation problems is to perform subspace decomposition of the covariance matrix. Since traditional neural networks can't handle real and imaginary numbers at the same time, the subspace decomposition method is not suitable for neural networks. Inspired by the extensive application of ResNet in the field of computer vision, this paper proposes a method of using the covariance matrix as an image processing, which uses a dual-channel matrix image containing the imaginary covariance matrix and the real covariance matrix as the input of the ResNet to estimate the DOA of the underwater acoustic array. It provides a new perspective for DOA estimation to solve the acoustic field problem. The ResNet algorithm is compared with the KNN algorithm and traditional MUSIC algorithm in terms of estimation accuracy and time. Simulation experiments prove that the ResNet algorithm has greater accuracy and shorter prediction time in a low signal-to-noise ratio environment.