Subregion-Based Machine Learning for Wideband Amplitude-Only Direction-Finding Systems

In recent years, access to improved computing resources has contributed to the increased interest in the practical deployment of machine learning (ML) based systems. One area where ML has been considered is direction finding (DF) [1] . Typically, ML approaches for DF utilize preprocessing, or feature extraction. This produces good results, but it requires additional computations which increase the DF estimation time. Conventional DF approaches typically use the correlation method, where the received signal is correlated with pre-saved steering vectors in lookup tables. Some ML-based methods, such as the neural network presented in this work, can maintain a much smaller footprint than the baseline correlation approach, and omit the need for feature extraction. Often, these systems deploy multiple, parallel estimators which individually cover a fraction of the entire field-of-view (FOV), but collectively estimate over the system’s entire region of interest. This work presents a case study to quantify the impact of restricting the field-of-view (FOV) over which each subregion estimator operates.