Sensitivity-driven model updating for bridge structures using operational modal analysis and enhanced batched Thompson sampling-based Bayesian optimization

This paper proposes a robust and effective finite element model updating (FEMU) framework for inverse parameter identification of bridge structures by integrating operational modal analysis (OMA) and variance-based global sensitivity analysis (GSA) with an enhanced batched Thompson sampling-based Bayesian optimization (ETS-BO) method. Specifically, a Gaussian process (GP) model is constructed to approximate the nonlinear relationship between uncertain structural parameters and the corresponding modal responses. Leveraging the trained GP model, the variance-based GSA identifies the most influential parameters governing the dynamic behavior of the bridge, thereby significantly reducing the dimensionality of the FEMU problem and improving computational efficiency. Building upon the reduced parameter space, the ETS-BO method is developed by integrating an adaptive trust-region selection strategy with an enhanced batched Thompson sampling (ETS) method to efficiently solve the FEMU problem. The adaptive trust-region selection evaluates the improvement potential at the regional level, guiding the search toward promising regions of the parameter space while avoiding over-exploration of less promising ones. The trust region method is then applied to perform local optimization, whereby each trust region is adaptively expanded or contracted based on its observed improvement history, and local optimization runs exhibiting greater potential for further improvement are selectively continued. Within the selected trust regions, the ETS method is employed to determine the optimal updating parameters by iteratively selecting batches of candidate solutions sampled from the posterior distributions of the corresponding independent GP models. The effectiveness of the proposed ETS-BO is validated through a real-world case study of a bridge structure equipped with wireless accelerometers for vibration monitoring. Ambient vibration responses recorded under normal traffic conditions are analyzed using OMA to identify modal parameters, which subsequently serve as targets for updating the finite element (FE) model developed to represent the structural behavior of the bridge. The results demonstrate that the proposed ETS-BO requires substantially fewer FE evaluations than conventional optimization methods while achieving rapid convergence toward the optimal solution. Furthermore, the updated FE model closely reproduces the experimentally identified modal parameters, including natural frequencies and mode shape correlations, thereby providing a reliable numerical representation of the in-service dynamic behavior of the bridge and demonstrating the potential of the proposed framework to support long-term structural health monitoring of large-scale civil infrastructure systems.