Integrating climate projections with Gaussian processes for structural anomaly detection via cointegration

Ensuring the long-term safety and durability of civil infrastructure increasingly depends on data-driven structural health monitoring systems capable of objectively tracking changes in structural performance. However, their reliability is undermined by the evolving environmental conditions associated with climate change. Variations in temperature and other environmental factors can modify the dynamic response of a target structure, producing apparent changes in damage-sensitive features that may be misinterpreted as degradation. Therefore, distinguishing environmental influences from genuine structural deterioration is essential for a reliable assessment. This study presents an integrated framework for resilient structural monitoring that separates climate-induced variability from damage effects. The methodology combines automated operational modal analysis with Gaussian process regression to model the nonlinear dependence of natural frequencies on temperature and quantify uncertainty. Climate projections from the CH2018 model under Representative Concentration Pathway scenarios are incorporated to forecast the long-term evolution of modal properties. The approach is demonstrated on the case study of the Chillon Viaduct in Switzerland, where continuous monitoring revealed a consistent reduction in natural frequencies with increasing temperature, reflecting temperature-dependent material, revealing a consistent reduction in natural frequencies with increasing temperature, reflecting a combination of thermally driven mechanisms, including material property changes and structural boundary variations. Sparse Gaussian process models enabled efficient long-term forecasting of frequency variations while maintaining predictive accuracy. Cointegration analysis was used to remove temperature-driven trends, isolating residuals sensitive to structural damage. Simulated progressive-damage scenarios confirmed the method’s capacity to detect incipient deterioration under varying environmental conditions. Overall, this work establishes a robust and generalizable framework that integrates modal analysis, probabilistic modeling, and climate projections to improve the resilience, interpretability, and long-term applicability of structural monitoring systems.