A Novel Off-Grid Deep Learning framework for DOA Estimation
Yunye Su, Xianpeng Wang, Liangliang Li · 2024
In this paper, a novel off-grid deep learning-based Direction of Arrival (DOA) estimation algorithm is proposed, specifically designed for accurately determining the direction of reflected signals. Accurate DOA estimation is crucial in a wide range of applications including radar, sonar, communications, and healthcare. Despite the prevalence of advanced deep learning algorithms for DOA estimation, nearly all suffer from errors of grid. Consequently, a neural network framework with a corresponding loss function was developed to eliminate off-grid errors. The proposed model does not require pre-determining the number of targets, unlike traditional off-grid algorithms, and it delivers more precise DOA estimations. Viewing DOA as a regression issue, this model utilizes the signal’s covariance matrix as input. After processing through a neural network module to obtain pseudo-spectrum, another neural network module refines these into coordinate error values, ultimately yielding precise DOA estimations. The simulation results demonstrate significant advantages.