Integrated neural networks – Particle Swarm Optimization for probabilistic finite element model updating: application to a prestressed concrete girder bridge

The implementation of Finite Element Analysis (FEA) has grown due to its strong theoretical foundation and accuracy in describing structural behaviour. However, developing Finite Element Models (FEMs) that can precisely replicate the behaviour of existing structures for structural health monitoring remains challenging, primarily due to variations in actual structural parameters. To address this, an integrated Neural Network–Particle Swarm Optimization Finite Element Model Updating framework (NN-PSO FEMU) is proposed in this study to probabilistically reduce the gap between FE-simulated and actual dynamic responses. The framework employs multi-restart PSO to identify optimal model parameter combinations, while NN models serve as surrogate models to bypass the computationally expensive FEA, enabling probabilistic FEMU. Kernel Density Estimation then builds joint probability distributions of optimal parameters to ensure all possibilities, including the exact solution, are embraced. Applying this framework to an aged prestressed concrete girder bridge resulted in low-variance joint probability distributions of simulated dynamic responses closely matching actual measurements. The updated FEM, validated through static deformation comparisons, demonstrated improved realism, even when only dynamic responses were used in FEMU. Validation through a Bayesian FEMU framework with No-U-Turn Sampler confirms NN-PSO's effectiveness in handling complex and high-dimensional search spaces and avoiding overconfidence in local optima.