DOA Estimation Using Deep Neural Network with Angular Sliding Window

Abstract

Deep neural network (DNN) has shown great potential in direction-of-arrival (DOA) estimation. In high dynamic signal-to-noise (SNR) scenarios, the estimation accuracy of the weaker sources may degrade significantly due to insufficient training samples. This paper proposes a deep neural network framework with sliding window operation. The whole field-of-view (FOV) is divided into a series of sub-regions via sliding windows. Each sub-region is assumed to contain one source at most. Thus, the single-source data can be used to train all the networks, alleviating the need for the training samples and the prior information on the number of sources. A detector network and an estimator network are followed for each sub-region, enabling high estimation accuracy and the number of sources. Simulation and real data experiment results show that the proposed method can achieve excellent DOA and source number estimation performance. Specifically, in the real data experiment, the results show that the RMSE of the proposed method reaches 0.071, which is at least 0.03 lower than FFT, MUSIC, ESPRIT, and a deep learning method namely deep convolutional network (DCN), cannot estimate the lower SNR source in high dynamic SNR scenarios.