Time-Dependent Reliability Using Metamodels With Transformed Random Inputs

We have recently proposed a method for time-dependent reliability based on metamodels with random inputs. In that method, we employed multiple sets of inputs sampled from the input distribution to construct a new metamodel as a mixture of classical metamodels. Because the sampled inputs may cluster around a mode of the input distribution, they may result in a metamodel of reduced quality. We address this issue in this paper by using a transformation to de-cluster the sample inputs and then use our previously developed metamodel with random inputs. We first obtain the output of the computer model for a limited number of transformed input draws which do not cluster in high probability regions of the input space. Then, conditioned on these transformed sampled inputs, we construct a classical Kriging surrogate and obtain the distribution of the new surrogate as the marginal of the joint distribution between the classical surrogate and the transformed sampled inputs. The proposed method is illustrated with a corroding beam example. A more accurate time-dependent reliability estimation is obtained compared with our previously developed metamodel method.