One-bit DOA Estimation with Array Imperfections: A Learning-based Approach

Low-power and low-complexity Direction-of-Arrival (DOA) estimation is a critical challenge in various signal processing applications. This paper presents a data-driven approach using a neural network framework designed for single-snapshot, one-bit quantized measurements, enabling significant hardware simplification. We explore a Fully Connected (FC) Neural Network architecture (which is identical to a 1D convolutional architecture in the single-snapshot case) trained end-to-end on raw quantized data. The performance of this learned approach is rigorously compared against the well-established MUltiple SIgnal Classification (MUSIC) and Complex-valued Binary Iterative Hard Thresholding (CBIHT) algorithms. Our findings reveal that the proposed neural network achieves superior estimation performance compared to these, especially under low SNR values and in the presence of realistic array imperfections. We observe significant performance gains even without explicitly training the network on data with such imperfections, highlighting the generalization capabilities of the proposed method.