BLE Channel Classification with Stability Encoding for Indoor Positioning

Bluetooth Low Energy (BLE) based indoor localization systems often face performance degradation when applied to new environments because signal patterns captured through the Received Signal Strength Indicator (RSSI) vary with obstacles, layouts, and device types. Models trained on one environment’s data frequently fail to maintain accuracy on unseen data, due to unstable and fluctuating signal characteristics. To overcome these issues, this paper presents an enhanced multi-source domain generalized BLE channel classification framework by combining statistical feature encoding, attention-based pooling, and a variance-guided loss for stability-aware learning across multiple source domains. The experimental evaluation with multi-environment (room and corridors) BLE RSSI datasets shows that the proposed hybrid framework using LSTM achieves 90.6% accuracy in training and 78.5% in an unseen target domain, outperforming the baseline DANN model by approximately 17%. Furthermore, the CNN-BiLSTM variant further improves accuracy in the unseen target domain by 6.6% over the LSTM-based hybrid framework.