PiGLAD · Physics-informed Generative Learning for Anomaly Detection in Transport Infrastructures under Moving Loads

Transport infrastructures are crucial for the mobility of people and goods. Beam-like transport infrastructures like railways and bridges, subjected to moving loads, face increasing safety risks due to growing traffic demands. Timely detection of structural anomalies is therefore essential for risk mitigation and condition-based maintenance. This project aims to develop and experimentally validate an innovative input-output structural anomaly detection framework that integrates moving loads and acceleration responses. In contrast to extensively studied approaches that only use structural responses, the input-output framework, free from external load assumptions, can potentially transform anomaly detection from an ill-posed inverse problem into a manageable forward problem. Despite its promise, such a framework faces two common challenges in structural anomaly detection: data discrepancies between numerical and real-world domains and the scarcity of data in damaged states. Domain adaptation (DA) and physics-informed machine learning (PIML) have demonstrated great promise in addressing these challenges. However, current PIML studies are not directly applicable to anomaly detection when multiple damage states are involved, while existing DA approaches fall short in transferring physical knowledge across numerical and real-world domains. To overcome this, we propose a physics-informed generative learning model for the input-output anomaly detection framework. This model will generate synthetic structural responses in multiple damage states, facilitating damage detection through comparison with actual measurements. The framework will be validated from laboratory to in-service railway bridges using state-of-the-art V-Track and CTO Measurement Train. In addition, a comprehensive reliability analysis and uncertainty quantification will be performed to identify the key factors that impact its various performance metrics, forming recommendations for the framework design.