PHYSICS-INFORMED MACHINE LEARNING-DRIVEN STRUCTURAL DIGITAL TWIN FOR DAMAGE IDENTIFICATION THROUGH ANOMALY DETECTION

Structural health monitoring (SHM) and timely damage identification play crucial roles in ensuring the safety and longevity of civil infrastructure. With the recent advancements in artificial intelligence (AI), data-driven techniques such as autoencoders (AEs) have been extensively utilized for damage identification through anomaly detection. However, AE based methods primarily learn identity mappings from healthy-state data, making them less effective in detecting subtle damage, highlighting the need for more robust and physics aware approaches to damage identification. This study presents a physics-informed, machine learning-driven structural digital twin (DT) framework for damage identification through anomaly detection. The proposed approach integrates physics-informed neural networks (PINNs) to first reduce discrepancies between finite element model (FEM) predictions and actual structural responses which can arise due to several reasons such as modeling uncer tainties. By leveraging sensor data from the real structure, the structural DT refines the predicted structural response. Anomaly detection is then performed by comparing these pre dictions against real sensor measurements. The performance of the proposed structural DT is evaluated against state-of-the-art AE-based and long short-term memory (LSTM)-AE-based anomaly detection methods. Validation of this approach is conducted using the ASCE bench mark structure, both numerically and experimentally. The ability of the proposed approach to detect different damage scenarios is studied. Results demonstrate consistently superior performance, particularly in detecting minor damage cases compared to existing AE-based methods. To demonstrate robustness, the framework is validated under both hammer impact and shaker excitations. Notably, an improvement of mean squared error (MSE) by at least 10-fold across all study cases was observed compared to traditional FEMs, indicating the potential of the proposed structural DT in generating refined response prediction. The dam age detection accuracy of 100% was observed across all the major damage scenarios across numerical and experimental study. Moreover, even in the smaller damage scenarios, the ac curacy of 100% was obtained for at least 20% stiffness reduction offering an improvement of accuracy value by up to 55% compared to standard AE and LSTM-AE. The findings of this study highlight the ability of physics-informed digital twins for structural damage identification. The study also forms a basis for exploration of the potential of the proposed structural DT approach beyond this application such as surrogate models for FEM updating, enabling real-time structural assessment, and dynamic testing applications