Evidence-Based Identification of Weighting Factors in Bayesian Model Updating Using Modal Data

Abstract

In Bayesian model updating, parameter identification of structural systems using modal data can be based on the formulation of the likelihood function as a product of two probability density functions, one relating to modal frequencies and one to mode-shape components. The selection of the prior distribution of the prediction-error variances relating to these two types of data has to be performed carefully so that the relative contributions are weighted to give balanced results. A methodology is proposed in this paper to select these weights by performing Bayesian updating at the model class level, where the model classes differ by having different ratios of the two prediction-error variances. The most probable model class on the basis of the modal data then gives the best choice for this variance ratio. Two illustrative examples, one using simulated data and one using experimental data, point out the effect of the different relative contributions of the modal frequencies and mode-shape components to the total amount of information extracted from the modal data.

Additional details