Finite Element Model Updating for Cable-Stayed Bridges Using Adaptive Wind Driven Optimization with Gradient Assistance

Finite Element Model Updating (FEMU) is crucial for ensuring bridge safety within Structural Health Monitoring (SHM), yet traditional methods face significant challenges including high computational cost, susceptibility of surrogate models to local optima, and imbalanced exploration-exploitation in metaheuristic algorithms. To overcome these limitations, this paper proposes an innovative Adaptive Wind Driven Optimization with Gradient Assistance (AWDO-GA) framework to optimize Backpropagation (BP) neural networks for efficient and high-precision FEMU. The core innovation lies in integrating the physics-inspired exploration mechanism of Wind Driven Optimization (WDO) with the gradient information of BP networks: firstly, adaptive global-local balance is achieved by dynamically tuning the Coriolis force coefficient, friction coefficient, and velocity constraint (e.g., g = 1.5g_base during early exploration, g = 0.8g_base during late exploitation); secondly, a gradient-physical fusion velocity update rule (vcorrected = (1-β)vwdo + β(-η∇L(W))) accelerates convergence and helps mitigate measurement noise effects. Applied to the Guanhe long-span steel-concrete composite cable-stayed bridge, sample data were generated using D-optimal experimental design, and the AWDO framework optimized BP network weights and thresholds. AWDO was compared with traditional algorithms (GA, PSO) and modern metaheuristics (DE, CMA-ES). Results show that AWDO reduces optimization time by 44% (only 6.5 minutes), improves convergence, and provides more accurate parameter identification. Specifically, AWDO reduces lateral mode (H1) error by 79% (to 1.21%) and achieves vertical mode (V1-V3) errors of ≤1%, surpassing GA and PSO. DE and CMA-ES also showed competitive performance, with CMA-ES achieving similar optimization times. Parameter sensitivity analysis highlights the dominance of material properties (steel modulus E 2 ) and boundary conditions (bearing stiffness K 1 ) in modal sensitivity. Despite a 6.26% residual error in torsional mode due to simplified boundary modeling, this study provides a robust algorithm for creating high-fidelity digital twins of bridges.