Data features-based likelihood-informed Bayesian finite element model updating

Abstract

A new formulation for likelihood-informed Bayesian inference is proposed in this work based on probability models introduced for the features between the measurements and model predictions. The formulation applies to both linear and nonlinear dynamic models of structures. A relation between likelihood-informed and likelihood-free approximate Bayesian computation (ABC) is also established in this study, demonstrating that both formulations yield reasonable and consistent uncertainties for the model parameters. In particular, the uncertainties obtained with the new formulation account better for the fact that different sampling rates used in recording response time history measurements often yield measurements that contain the same information and so the sampling rate should not affect the uncertainty in the model parameters. The effectiveness of the proposed approach is demonstrated using an example from model updating of a linear model of a dynamical spring-mass chain system. © 2019 Proceedings of the 3rd International Conference on Uncertainty Quantification in Computational Sciences and Engineering, UNCECOMP 2019. All rights reserved.