A Dynamic Updating-Based RSSI Fingerprint Localization Algorithm Using TabNet and Transfer Learning

The widespread deployment of Internet of Things (IoT) applications relies on indoor positioning systems that are both accurate and sustainable. However, conventional RSSI fingerprinting methods suffer from signal fluctuations in dynamic environments and require costly manual updates to maintain long-term performance. To address these challenges, this paper proposes a dynamic indoor positioning framework that integrates TabNet and transfer learning (TL). A small number of reference points are first deployed, and an offline fingerprint database is efficiently constructed using pedestrian dead reckoning (PDR) and Kriging interpolation, significantly reducing manual effort. A TabNet-based localization model is then trained to learn the nonlinear RSSI–location mapping while mitigating signal instability. Moreover, a reliability evaluation mechanism is introduced to identify high-confidence online position estimates, which are incorporated via TL to continuously update the localization model without additional data collection. Experimental results demonstrate that the proposed method outperforms state-of-the-art localization approaches and maintains stable performance over time, with the root mean square error increase remaining below 0.2 m after 20 days.