Investigación sobre la inspección de rendimiento de puentes atirantados de gran luz bajo la guía de información de...

The temporal degradation of mechanical performance in large-span bridges necessitates the precise updating of Finite Element model parameters to guarantee accurate safety assessments and service life predictions. However, existing deep learning-based updating methodologies predominantly rely on single-physical-field inputs and assume homogeneous data topologies, thereby failing to capture complex, high-order mechanical interactions across heterogeneous physical domains. To overcome these limitations, this study proposes the Integrating Multi-Physical-Field Encoding Heterogeneous Graph Neural Network (IMPFE-HGNN). This novel architecture explicitly models the heterogeneous topology among strain, deflection, temperature, cable force, and acceleration sensors via meta-path subgraphs and relationship-aware encodings, enabling the extraction of high-order multi-physics semantics inaccessible to traditional architectures. Validated through a case study on a long-span cable-stayed bridge, the IMPFE-HGNN demonstrates substantial efficacy in parameter identification, yielding maximum correction rates of 43.33% for Poisson's ratio and 10.20% for elastic modulus. Consequently, the predictive fidelity of the updated FE model is significantly enhanced: strain prediction error is reduced by a median of 61.4% (peaking at 77%), while deflection prediction accuracy improves by a median of 72.8% (peaking at 87%). Ablation studies substantiate the critical contributions of meta-path subgraphs and relationship encoding mechanisms, while sensitivity analyses determine optimal hyperparameters, identifying a meta-path length of 5 and a feature-mapping dimension of 128. Overall, this study presents a physically interpretable heterogeneous GNN paradigm for multi-source data fusion, offering a robust and precise framework for the structural performance assessment of long-span cable-stayed bridges.