Comparative Analysis of Machine Learning Algorithms for BLE-Based Indoor Localization System

Indoor positioning systems (IPS) play a vital role in applications such as navigation and asset tracking. This study presents a comprehensive evaluation of machine learning algorithms on a newly introduced Bluetooth Low Energy (BLE) Received Signal Strength Indicator (RSSI) dataset, specifically developed for localization tasks. The dataset undergoes systematic robustness testing, including RSSI masking, added noise, and missing anchor nodes, to assess model resilience in challenging conditions. Five machine learning algorithms—K-nearest neighbors (KNN), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Fully Connected Neural Network (FCNN)—are examined to benchmark their performance. The models are evaluated based on localization error (LE) and success rate (SR) across different accuracy thresholds, utilizing a custom dataset collected in diverse room environments. Key contributions include an extensive robustness analysis under realistic BLE signal conditions—such as RSSI noise, partial signal masking, and missing anchor nodes—as well as the use of Kalman filtering as an effective preprocessing step to enhance model performance. Results demonstrate that XGBoost and RF achieve superior accuracy and scalability compared to other methods. RF delivers strong localization accuracy with low computational overhead, while XGBoost balances high accuracy with robustness to environmental noise and partial data loss. However, challenges such as performance inconsistencies in rooms with poor signal coverage and susceptibility to signal fluctuations are noted. These findings establish a benchmark for BLE RSSI-based localization research, offering key insights into the strengths and limitations of different machine learning models for various real-world applications. Key contributions include an extensive robustness analysis under realistic BLE signal conditions—such as RSSI noise, partial signal masking, and missing anchor nodes—as well as the use of Kalman filtering as an effective preprocessing step to enhance model performance.