TransAoA: Transformer-Based Angle of Arrival Estimation for BLE Indoor Localization

Abstract

Bluetooth low-energy (BLE) technology, characterized by its low-energy consumption, cost-effectiveness, and scalability, has gained prominence as a viable solution for indoor localization within industrial contexts. However, the dynamic nature of industrial environments poses considerable challenges to the accuracy of BLE-based indoor positioning systems (IPSs), particularly those dependent on signal strength for localization. Accordingly, this article proposes a novel method framework TransAoA that leverages the Transformer deep learning architecture to enhance angle of arrival (AoA) estimation for BLE indoor positioning. First, a data filtering method is designed to eliminate low-quality in-phase and quadrature (I/Q) samples affected by noise. Second, a specialized feature extraction method is developed to distill multiple informative features from I/Q samples prior to the prediction model to enable rapid convergence and improve accuracy. Third, the Transformer-based AoA estimation model is constructed to establish a mapping relationship between angles (azimuth and elevation) and the combined I/Q samples and features. Fourth, several BLE anchors collaborate to localize targets using a least squares (LSs) approach, and a self-adjusting calibration mechanism is devised to bolster the long-term robustness and stability of the IPS. Finally, experiments are conducted in a lab that simulates industrial conditions to verify the effectiveness of the framework. By comparison, the TransAoA shows superiority over existing benchmark methods, achieving estimation errors within 5° for 98.85% of azimuth and 99.97% of elevation measurements.