Inverse uncertainty quantification using large-scale database from system level tests

Abstract

Reliability-based design optimization (RBDO) utilizing computer simulations can lead to a highly reliable optimum design. However, conventional RBDO methods require full statistical information of input variables to estimate reliabilities of engineering systems or components, which is not easy to obtain in most engineering applications. Application engineers suffer from lack of information to assume well-known distributions or to trust supplier’s document using few test sample. This is the main reason why many engineering companies have to validate their designs through physical test several times before mass production. In this study, an uncertainty quantification method with mean-correlated simulations is proposed to estimate the statistical information of input variables from corresponding system response distributions obtained using large-scale test database. The proposed approach employs an error-lumped inverse method and kernel density estimation (KDE) for the uncertainty quantification. All possible errors such as measurement error, simulation error, and error by input variable difference are lumped into one to minimize residual errors of responses. Because quantified uncertainties using the error-lumped inversed method could be too scattered, distribution correction is proposed to reduce effective range of input variables while maintaining response distributions. Numerical and engineering examples show that the proposed approach can well estimate uncertainties of input variables using mean-correlated simulation and system response distributions obtained from test results.