Machine learning–enhanced structural health monitoring of the steel–glass exedra under year-round environmental loading

This paper presents a comprehensive approach to dynamic identification and environmental effect filtering for the Exedra Hall at the Capitoline Museums in Rome. Measurements collected continuously over 1 year were analyzed using the Stochastic Subspace Identification (SSI) method to extract the modal parameters. A k-means clustering algorithm was employed to identify and track the first three natural frequencies. To address temperature-induced frequency variations, a feedforward Artificial Neural Network (ANN) was trained on 8 months of frequency and temperature data, enabling the isolation and removal of temperature effects from the identified frequencies. The filtered frequencies were extrapolated and validated against the identified data for the entire observation period. To improve robustness, physical plausibility, and extrapolation capability under seasonal transitions, a physics-informed neural network (PINN) constrained by finite-element-derived thermomechanical priors was additionally introduced. Finally, a model updating of the Finite Element Model (FEM), has been carried out to describe the temperature effect. This work contributes to Structural Health Monitoring (SHM) practices by integrating advanced identification techniques, machine learning, and model-based simulations.