On the generalization of Deep Learning models for AoA estimation in Bluetooth indoor scenarios

Abstract

Indoor positioning remains as one of the major challenges of the Internet of Things (IoT), where assets are usually deployed over environments where the Global Navigation Satellite Systems (GNSS) are denied. Angle-of-arrival (AoA) estimation techniques are among the most prominent candidates to address this challenge and multiple techniques based on conventional signal processing, such as the Multiple Signal Classification (MUSIC) algorithm, have been extensively studied. However, they require a deep knowledge of the elements involved in the AoA system, such as the characterization of the receivers’ antennas. On the other hand, Deep Learning (DL) models are considered as promising solutions, as they can provide AoA estimations without resorting to a deep knowledge of the physical system. Instead, they only need to be trained with labeled AoA signal measurements. Nevertheless, in the training process, non-desired effects like overfitting appear, yielding models with poor generalization capabilities. Although these models have already been considered in the literature, the problem of DL models generalization has not been addressed in depth. Therefore, this manuscript first compares the performance of DL models to the MUSIC algorithm as a baseline, to then analyze their generalization to different situations. These include changing the position of the receivers within the same area, considering measurements at different time instants and deployments in new scenarios, specifically in locations different from the training one. The assessment showed that the generalization of the DL models is weak compared to the MUSIC algorithm, especially when applied to environments not previously observed.