A new unbiased metamodel method for efficient reliability analysis

Abstract

Metamodel method is widely used in structural reliability analysis. A main limitation of this method is that it is difficult or even impossible to quantify the model uncertainty caused by the metamodel approximation. This paper develops an improved metamodel method which is unbiased and highly efficient. The new method formulates a probability of failure as a product of a metamodel-based probability of failure and a correction term, which accounts for the approximation error due to metamodel approximation. The correction term is constructed and estimated using the Markov chain simulation. An iterative scheme is further developed to adaptively improve the accuracy of the metamodel and the associated correction term. The accuracy and efficiency of the new metamodel method is illustrated and compared with the classical Kriging metamodel and high dimensional model representation methods using a number of numerical and structural examples.