| dc.description.abstract | In Bayesian model updating, probability density functions of model parameters are updated accounting both for the information contained in the data and for uncertainties present in the measurements and model predictions, requiring a probabilistic model for the error between predictions and observations. Most often, a zero-mean uncorrelated Gaussian prediction error is assumed, although in many engineering applications prediction errors will show non-negligible spatial and/or temporal correlation (e.g. when densely populated sensor grids are used). In this paper, the effect of prediction error correlation on the results of the Bayesian model updating scheme is studied, and it is investigated how the challenging task of selecting a suitable prediction error correlation structure can be addressed appropriately. In two illustrative applications, it is demonstrated that Bayesian model class selection can be effectively applied to this end, ensuring more realistic modeling and corresponding Bayesian model updating results. (C) 2013 Elsevier Ltd. All rights reserved. | en |
