Rational Method for FE-Model Updating: Comparison Between Conventional and AI Methods: A case-study of Tvärbanan, stage Solvalla

Abstract

The aim of this thesis is to evaluate and develop alternative and rational methods for model updating. Ordinary and already established optimisation methods (Nelder- Mead simplex, pattern search, genetic algorithm and particle swarm) were compared against the contemporary and emerging principles of neural networks and AI. Their performance was compared using three different beam test cases, with increased complexity, with intention of solving unknown variables based on the corresponding “true” deflection. Through the test cases, each method was tested and evaluated based on accuracy, computational demand and scalability against complexity. The results indicated that the traditional and already established methods performed best, yielding high accuracy for low computational demand. Therefore, the developed optimisation methods based on neural networks and use of the ChatGPT AI could not compete. However, findings indicate that neural networks methods and AI have future potential in other areas then initially assumed. The conclusion of this thesis is that neural network methods could be useful, particularly when implemented as a tool rather than a replacement for model updating. Neural networks with the capacity to predict deflections based on input variables could be simultaneously trained based on information gained trough a traditional model updating routine. Once trained, re-evaluation of model behaviour could be produced without a FEM-model, requiring just a few seconds, providing improved abilities in engineering judgements.