The role of finite element model updating in homogeneous transfer learning for damage classification in structural health monitoring

This paper investigates the potential of employing transfer learning (TL) for structural damage classification using synthetic data generated from un-updated finite element models (FEMs). While finite element model updating (FEMU) is commonly used to align FEM outputs with experimental data, it does not fully eliminate residual discrepancies between simulated and real measurements. These discrepancies can undermine the reliability of damage classification models trained on FEM-generated data, when used to classify experimental observations. Recent studies have addressed this issue by leveraging TL to align synthetic data from updated FEMs with real measurements. However, FEMU is time-consuming, prone to ill-conditioning, and its necessity in TL-based damage classification remains unclear. Motivated by these factors, this study investigates whether synthetic data from un-updated FEM can still enable effective damage classification through the Joint Domain Adaptation (JDA) method, a form of homogeneous TL where real and simulated domains share consistent labels and features. To this end, an investigation using both numerical and experimental case studies is conducted, along with a detailed discrepancy analysis to evaluate how domain mismatch affects classification performance. The results demonstrate that the TL-based approach achieves high classification accuracy and reliable damage characterization, even without an ad-hoc FEM calibration.