Abstract: To address the challenges of low model accuracy and unquantified uncertainties in the health monitoring of long-serving hydraulic structures,this study developed a Bayesian theory-driven approach for constructing digital twin models.The method integrates data-driven and model-based strategies by combining finite element model (FEM) updating with a Bayesian probabilistic framework,establishing a dynamically self-calibrating model through a two-step updating process.First,a multi-zone FEM was constructed,followed by modal identification and parameter sensitivity analysis to identify sensitive parameters for structural model updating.A two-step updating methodology was then applied:deterministic updating optimized the initial parameters,while Bayesian uncertainty updating quantified parameter uncertainties using the Metropolis-Hastings (MH) sampling algorithm to enhance computational efficiency.The proposed method was validated via field tests and numerical simulations on a 50-year-old rectangular aqueduct in Shaanxi Province,China.Results show that the two-step updating method improves convergence efficiency by approximately 25%.The frequency errors after updating range from 1.23% to 6.42%,with the exception of the second-order mode error at 18.17%.The zoned FEM also improves damage identification,with a 5.2% error in elastic modulus reduction in damaged zones.This automated updating approach,implemented through MATLAB-ABAQUS integration,offers a high-precision and dynamic solution for lifecycle safety assessment of hydraulic structures.