Probabilistic finite element model updating across multiple model classes and response domains using integrated NN - PSO framework: application to a PC girder bridge

Abstract

While Finite Element Analysis provides high accuracy in describing structural behaviour, discrepancies between simulations and reality exist. This study applies probabilistic Finite Element Model Updating (FEMU) using multi-restart Particle Swarm Optimisation (PSO) to a prestressed concrete girder bridge. Real-world measurement data, encompassing static and dynamic responses, collected from static truck load and free vibration tests, is used to improve FEM accuracy. Several cases of the FEMU, including single- and multi-response domain FEMU with cross-validation, are performed to assess the impact of the updating domains and ensure the result validity. Additionally, multi-class FEMs with improved details are presented to eliminate portions of modelling errors, enhance the result interpretability, and explore the challenges of ill-posedness and ill-conditioning nature of inverse engineering problems. This is accelerated by Neural Network static and dynamic metamodels, which predict static and dynamic responses, thereby enabling probabilistic FEMU. The metamodel development is discussed in detail, including the feature engineering that covers an effective mode-matching method for the dynamic metamodels, addressing the performance limitations observed in previous studies. Transfer Learning between two comparable FEM classes is also discussed. As a result, updated FEMs that yield much more realistic static and dynamic responses are obtained from the multi-response domain FEMU.