ABSTRACT
Bayesian updating has been widely used in civil engineering for many different purposes such as parameter identification, model updating, reliability analysis. Bayesian updating usually requires high computation cost, and some methodologies are available to reduce the computational costs One of the methodologies that has been commonly used for system identification/control in time-series problem includes unscented transformation (UT). In UT, the prior and posterior probability density functions (PDFs) are approximately estimated with 2n + 1 data points (n is the number of parameters of interest): it means that only 2n + 1 times simulations are required for Bayesian inference. However, UT assumes that prior and posterior PDFs follow Gaussian PDFs and has limitations for some practical applications. This paper investigates the performance and limitations of UT in Bayesian updating in terms of estimation accuracy of posterior PDFs through synthetic examples. Comparison of UT with Monte Carlo filter (MCF) is also made.
