Deep Learning for Bluetooth Low Energy Indoor Positioning: A Comparative Analysis of Fingerprinting Accuracy and Generalization to Unseen Locations

Indoor positioning remains an important challenge in ubiquitous computing, particularly when balancing cost, accuracy, and adaptability. This work evaluates deterministic, probabilistic, classification, and regression models for indoor localization using a low-cost Bluetooth Low Energy (BLE) mesh platform, specifically examining their comparative resilience to spatial and temporal domain shifts. We compare Log-distance Least Squares, Log-normal Maximum Likelihood Estimation, Bidirectional Long Short-Term Memory (BiLSTM), Convolutional Neural Network (CNN)-Attention-BiLSTM, Extreme Gradient Boosting (XGBoost), and a hybrid CNN-XGBoost architecture across multiple scenarios designed to assess both peak accuracy and generalization to previously unseen locations. Our findings reveal a critical trade-off: the classification-based BiLSTM model achieves near-perfect precision in fully mapped environments but suffers a significant performance drop when tested on unseen positions. In contrast, the hybrid CNN-XGBoost regressor demonstrates better generalization, obtaining the lowest Mean Distance Error (MDE) of 0.89 m on unseen locations and maintaining resilience as the number of anchors is reduced. While the classification models perform best for environments with exhaustive site surveys, the hybrid regression offers better adaptability in constrained or dynamic deployments, highlighting that the best model is application-dependent.