Cross-domain structural damage identification based on updated FEM and WGAN-GP: Bridging the gap between FEM simulations and actual structural responses

Structural damage identification often faces challenges due to discrepancies between finite element model (FEM) simulations and actual structural responses. These domain discrepancies are inevitable and can significantly hinder the practical application of model-driven damage identification. To address these challenges, this paper proposes a two-stage cross-domain damage identification framework that integrates FEM updating with cross-domain deep learning. In the first stage, a set of FEMs with parameter variations is constructed, and a Convolutional Neural Network (CNN) is employed to select the one that best approximates the real structure, serving as the updated FEM. In the second stage, a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) is utilized to generate synthetic responses from FEM simulations, aiming to align with the feature distribution of actual structural responses. This approach enables effective feature alignment and large-scale sample augmentation, thereby improving the robustness and accuracy of damage identification. Finally, a CNN is adopted to extract essential damage features from the synthetic responses and construct a mapping to the damage states, thereby achieving accurate cross-domain damage identification. The effectiveness of the proposed method is validated through the damage identification of a steel truss model. The results demonstrate a clear progression in damage identification performance: starting from 33.3 % accuracy with the baseline FEM, increasing to 46.2 % with updated FEM, then substantially improving to 82.4 % with WGAN, and ultimately reaching 96.2 % with WGAN-GP, which offers enhanced training stability and feature alignment capability.