FedGeoSwap++:A Novel Context-Aware and Privacy-Preserving Sensor Substitution Framework for Fault Tolerance in Metaverse-Driven Smart Cities

The integration of the Metaverse with real-time physical data from Internet of Things (IoT) sensors introduces stringent requirements for system reliability and continuity. Sensor failures at the network edge can disrupt immersive experiences due to data loss or inconsistency. To address this challenge, we propose FedGeoSwap++, a novel fault-tolerant framework that combines spatio-temporal indexing, federated learning, transfer learning, and an intelligent Dynamic Resource Adaptation (DRA) agent based on Double Q-Learning to enable intelligent, privacy-preserving sensor substitution in Metaverse applications. FedGeoSwap++ leverages a cloud-based spatio-temporal database with R-tree indexing to identify the geographically closest and data-wise most similar sensor upon failure. A Double Q-Learning agent dynamically adjusts the trade-off between spatial proximity and temporal correlation based on network load and environmental dynamics. Furthermore, a hard correlation threshold (ρ ≥ 0.6) ensures semantic consistency by filtering out spatially close but data-wise dissimilar sensors. We evaluate the framework using a simulated smart city environment with 50 sensors and 100 timesteps over 300 simulation runs. Results show that FedGeoSwap++ achieves a Mean Absolute Error (MAE) of 0.615°C, outperforming Nearest (0.624°C), CorrOnly (0.955°C), MeanImpute (3.029°C), and LSTM-Predict (1.299°C), while maintaining low latency (0.16 ms). Paired t-test confirms the statistical significance of this improvement (p < 0.0001). This work advances fault tolerance in Metaverse systems by ensuring seamless continuity, high accuracy, and robustness under sensor failures.