FRF-based model updating using neural networks

Model updating methodologies allow the improvement of the accuracy of an initial finite element (FE) model to get the predicted and measured dynamic properties closer. In this way, the updated FE model is more representative of the actual dynamic behavior of the structure under investigation, thus giving the opportunity to use it for further investigations. In this paper a response-based updating methodology using neural networks is proposed. The neural networks used in this study have a feed-forward architecture, their inputs are the amplitude FRFs and their outputs are the physical properties of the structure under investigation, in terms of its elemental mass and stiffness that allow improving the correlation between the predicted and measured modal parameters. In order to evaluate the effectiveness of the proposed updating approach both numerical and experimental analyses are carried out on simple cantilever beams. Furthermore, the influence of some setting parameters on the accuracy of the solution is investigated.