Learning Interpretable Phase Difference Mapping for Scalable DOA Estimation via Deep Learning

Interpretability is the major bottleneck regarding the deep learning-based direction of arrival estimation (DL-DOA). This paper provides novel insight into the mapping relationship (MR) for DL-DOA, which can be interpreted as a mapping from the received signal to the spatial phase difference (SPD) with prior information of array geometry (AG). Based on the learned MR, two scalable DL-DOA methods are developed. One method refers to the variability of AG, enabling the trained model to be utilized for varying AGs without the requirement of retraining network. The other involves scaled resolution (SR) by using an equivalent finer grid, which significantly mitigates the problem of grid mismatch for off-grid DOA with a coarse pre-defined grid. The efficacy of proposed methods is demonstrated through simulations with different network structures. The code for this work is available at: https://github.com/hslcode/DOA-Scalable.