Abstract:
Abstract
This work presents the development of a data-driven digital twin at sensor level for structural health monitoring of a metallic suspension bridge in Portugal, with a focus on local validation of strain sensors using long-term monitoring data acquired between 2014 and 2017. The central challenge lies in the automated distinction between normal environmental variations and the onset of critical sensor faults. Using the Prophet library,
the proposed model effectively captured the dominant 12-hour and 24-hour cycles governing bridge behavior. The results demonstrate that integrating external temperature and solar position as predictive variables results in a substantial improvement compared to models based solely on solar radiation, reducing the Mean Absolute Percentage Error to 2.5%. The anomaly detection system, based on a Binomial Hypothesis Test, exhibited
high sensitivity, successfully identifying debonding faults as small as 5%. It is concluded that the reliability of the digital twin depends on empirically calibrated control limits, as the theoretical uncertainty intervals provided by the software proved unreliable for this application. Overall, this work establishes a robust methodology for long-term sensor self- validation in critical infrastructure.
