Predictive Modelling of bridge bearing displacements with environmental effect filtering: a Physics-Informed Machine Learning approach to structural health monitoring

In the field of bridge Structural Health Monitoring, the possibility of an efficient damage detection is affected by confounding influences due to operational and environmental variability. In static monitoring of a bridge, not only temperature, but also other factors such as traffic loads, wind and friction, can have a big influence on the observed data. It is therefore necessary to understand in first place what constitutes a normal response to environmental and operational loadings, allowing to filter those effects and correctly identify anomalous behaviour. The primary objective of this thesis is to develop a predictive model that can filter the normal structural responses to environmental loads, establishing a baseline that highlights only those displacements caused by structural anomalies. This study focuses on the development of a static monitoring technique based on the use of displacement and temperature sensors to evaluate the longitudinal displacements to which a bridge deck is subjected, building a predictive model enabling to detect anomalous behaviour and presence of damage. Since thermal variations are the primary cause of bridge bearing movements and are easy to monitor, temperature is selected as the sole independent variable in the regression model alongside time. In this way it is possible to capture the influence of both daily and seasonal movement cycles, which cause a strong non-linearity between temperature and displacements. To build an accurate predictive model, regression-based Machine Learning algorithms, namely Gaussian Process Regression and Support Vector Machine models, are employed to identify a generic model of response of the structure to thermal variation, thus obtaining an algorithm capable of accurately predicting the expected longitudinal displacements of the bridge deck. In addition to conventional Black-box models, the research explores a Physics-Informed Machine Learning (PIML) approach, commonly known as a Grey-box model, which incorporates engineering knowledge of bridge behaviour. This hybrid model not only enhances prediction accuracy, but is also more interpretable and explainable, making the Grey-box approach a more trustworthy tool for supporting informed decision-making in bridge maintenance and safety. To further validate the effectiveness of the proposed models, an Early Warning System is implemented using thresholds for displacement anomalies. By simulating abnormal displacement scenarios, such as potential damage to a bearing device, the thesis assesses each model’s capability to detect simulated abnormal behaviours and differentiate them from environmental influences. Comparative analysis between the Black-box and Grey-box models reveals the superior performance of the Grey-box approach in predicting more accurately and with greater robustness the bearings displacements. The results highlight the significant potential of PIML-based predictive modelling to enhance anomaly detection in bridge structures, facilitating the development of automated Early Warning Systems with high reliability, thus providing a powerful tool for ensuring bridge safety. To summarise, this study demonstrates the potential of combining data-driven and physics-informed techniques to support proactive maintenance strategies and enhance infrastructure resilience, thus extending the lifespan of critical structures.