Dynamic Environmental Risk Identification for Historic Buildings Based on Multi-Source Sensing Digital Twins

Buildings with heritage suffer from various disturbances like hygrothermal disturbances, moisture, vibration, and crack problems which cannot be easily identified from sensor readings independently. In this research, a multi-source sensing-based digital twin is developed for identifying environmental risks dynamically in a monitored two-story building made of bricks and timber structure. Variables such as temperature, relative humidity, wall moisture, vibration, crack displacement, and weather data outside are synchronized every five minutes and mapped into IFC-based Heritage Building Information Model in Unity. With the Temporal Fusion Transformer, a multivariate history with 24 time steps is taken into account, and normal, hygrothermal instability, moisture infiltration, vibration abnormalities, and crack displacement risk predictions are done. The architecture includes class weights, adaptive temporal consistency, and spatial regularization of components. With temporal testing, macro-F1 of 0.891±0.012, AUROC of 0.962±0.006, and mean warning delay of 6.7±1.1 min were obtained which is better than LSTM and XGBoost models with p<0.001. Removal of moisture sensors yielded maximum drop of macro-F1 to 0.847±0.014. Digital twin localization error was 0.43±0.17 m.