Dynamic identification and thermo-modal response of a steel-glass hall under long-term monitoring

Long-term vibration-based monitoring of heritage structures is often hindered by environmental variability, with temperature-induced frequency shifts masking damage-sensitive changes. This study presents a digital twin-oriented framework for the thermo-modal analysis and dynamic identification of a heritage steel–glass hall under continuous monitoring. The proposed approach integrates probabilistic inference, physics-informed learning, and decision-oriented analysis within a unified pipeline. In the first stage, a Bayesian linear regression model is employed to describe the dependence of modal frequencies on the steel–glass temperature differential, providing uncertainty-aware predictive envelopes. In the second stage, a physics-informed neural network is trained on the residuals to capture nonlinear thermo-mechanical interactions while enforcing smoothness constraints consistent with physical behavior. A damage-sensitive indicator is constructed from thermo-normalized residuals using an epistemically weighted statistical distance metric, enabling identification of anomalous deviations from baseline structural behavior. The framework is applied to the Marcus Aurelius Exedra Hall at the Capitoline Museums in Rome, monitored through a continuous IoT-based sensing system. Results show that the hybrid model improves prediction accuracy and reduces temperature-induced variability. Moreover,the residual-based indicator provides a normalized and uncertainty-aware measure of deviation from the expected structural behaviour, improving the interpretability and robustness of anomaly detection compared to the use of raw residuals alone. The study highlights the potential of integrating Bayesian inference and physics-informed learning within digital twin architectures for robust, uncertainty-aware structural health monitoring of complex heritage structures.