A Novel Neural Network Approach for Coherent Source DOA Estimation

Abstract

Coherent sources often exist due to various factors such as multipath effects and electronic interference. How to estimate the parameters of coherent sources is a significant part of spatial spectrum estimation. The traditional algorithm for coherent signals has the defect of losing the effective aperture of the array, which affects the accuracy and resolution of the estimation. To solve the problem, this paper models coherent DOA estimation as multi-label classification based on neural network. Sparse autoencoder, spatial filter, and multiple parallel DNN classifiers are employed to complete the multi-label classification task. The whole framework can also adapt to close DOA scenario, and simulation results have demonstrated the superiority of the method. Moreover, this paper discussed the reason of DOA estimation failure and a staggered grid method is utilized to improve the classification accuracy.