BLE-based indoor localization with temporal convolutional network

Bluetooth Low Energy (BLE) is a promising technology for indoor localization due to its strong signal penetration and low power consumption, especially in environments where Wi-Fi signals are unreliable. BLE beacons are compact and battery-powered, enabling flexible deployment without the need for wired infrastructure. In this study, we propose a low-cost indoor localization approach based on a Temporal Convolutional Network (TCN), and perform a comparative analysis using two signal filtering techniques: a median filter (MF) and a Recursive Moving Average Filter (RMAF). The method was evaluated in a cluttered indoor environment measuring 10.7 m × 6.3 m using only three BLE beacons. Its performance was compared with state-of-the-art machine learning methods, including weighted k-nearest neighbors, support vector machine, and bagged trees. Experimental results demonstrate that the proposed TCN consistently achieved high accuracy, with at least 98% accuracy across all input conditions (raw RSSI, MF, and RMAF), showcasing robust baseline reliability. However, the F1-score, which provides a more balanced measure of precision and recall, revealed the most significant improvements in performance. Compared to the second-best machine learning method in each condition, the TCN achieved F1-score improvements of 27.1% with raw input, 22.9% with MF, and 9.7% with RMAF. The best overall result was achieved with the TCN + RMAF configuration, which reached an outstanding accuracy of 99% and an F1-score of 84.8%. These results emphasize that while all configurations of the proposed method perform reliably in terms of accuracy, the use of filtering-particularly RMAF further enhances the model’s ability to correctly detect user positions, making it a robust and practical solution for BLE-based indoor localization.