Physics-Informed Machine Learning for Hybrid Digital Twin–Enhanced Damage Detection and Localization

In structural health monitoring (SHM), structural damage detection and localization have made significant advances with deep learning–based methods. While finite element model (FEMs) have been employed to simulate a variety of damage scenarios for network training, mismatches between FEMs and actual structures persist due to uncertainties in materials, boundary conditions, construction processes, and aging. As a result, data-driven approaches trained on FEMs struggle to accurately capture the true dynamics of physical, limiting their reliability in representing digital twins of civil infrastructure. To bridge this gap, we propose a hierarchical physics-informed domain adaptation (HierPhyDA) approach that integrates physics-based and data-driven models to construct a hybrid digital twin of the physical structure. The proposed solution employs an initial phase of unsupervised anomaly detection using a deep autoencoder approach, followed by a novel physics-informed domain adaptation method. More specifically, a convolutional neural network (CNN) is pretrained to predict frequency response functions (FRFs) of the structure, and the weights of the pretrained model are utilized to initialize a discriminator-free adversarial learning network (DALN). This modal-informed weight initialization imposes physical constraints on DALN and enables it to learn modal-related features, which are conducive to damage localization. This approach is rigorously evaluated using a numerical ASCE benchmark structure and an experimental geodesic dome structure. The effects of model updating, modeling errors, structural uncertainties, and impact locations on the performance of the HierPhyDA approach are systematically analyzed. The results demonstrate that the proposed approach outperforms existing state-of-the-art methods in damage detection and localization tasks, even when damage cases from the actual structure are excluded from network training and are mutually exclusive from those generated by the FEM used for training. The introduced physical constraints significantly enhance the model’s generalization capability. The capabilities and limitations of the proposed approach are discussed.