Digital twin-driven cross-domain damage identification for experimental cable-stayed bridge models under sparse sensing

Reliable damage identification is paramount to structural integrity and operational safety of infrastructure. However, conventional structural health monitoring (SHM) methods often encounter implementation bottlenecks, primarily constrained by data labeling and the inherent discrepancy between numerical and physical responses. This study presents a novel digital twin-driven framework for structural damage identification using cross-domain feature adaptation. A residual network was trained on extensive simulated damage states, and the transfer learning was utilized to extract domain-invariant features across both numerical and experimental domains. This hybrid strategy integrates data-driven modeling and adaptive transfer learning to effectively bridge the gap in cross-domain generalization. An experimental study on a scaled experimental cable-stayed bridge was conducted to demonstrate the efficacy of the proposed framework in mitigating domain-shift challenges. Notably, it achieves 90.9% accuracy with sparse training sets, while simultaneously yielding a significant decrease in computational overhead. By contrast, the target only training strategy and the partial transfer strategy without domain alignment achieved accuracies of 83.1% and 73.4%, respectively. In addition, comparisons with GoogLeNet and AlexNet networks further validate the superior performance and robustness of the proposed framework in cross-domain damage identification tasks.