Learning-based robust direction-of-arrival estimation with array imperfections

Estimating the direction-of-arrival (DOA) of signals is a fundamental problem in radar applications and source localization. With practical antenna arrays, the performance of the DOA estimation often degrades significantly in low-cost systems due to the presence of numerous imperfect factors. This paper presents novel deep learning-based approaches for DOA estimation in two-dimensional (2-D) scenarios with imperfect factors. As the selection between using in-phase and quadrature (I/Q) sequences or covariance vector methods is a well-known open issue in DOA estimation, we investigate the performance of deep learning-based networks using these two different inputs. Our study aims to provide insights into which approach is more effective in achieving high accuracy for DOA estimation under non-ideal array conditions. We first train a ResNet using I/Q sequences for DOA estimation and then propose a novel approach using a convolutional attention neural network (CANN) with a covariance vector as input, incorporating frequency information to enhance network robustness. Furthermore, we derive the Cramér-Rao Lower Bounds (CRLBs) of mean squared errors (MSEs) for parameter estimators in a single-source scenario. Simulations are conducted to evaluate the accuracy of the proposed DOA estimation approach, demonstrating its superior performance over existing techniques and its close approximation to the corresponding CRLBs.