ABSTRACT
A major approach for quantifying the failure probability of a highly reliable industrial component, for which no failure data is available, is to use computer experiments reproducing the physical phenomena of degradation affecting its reliability. Often mathematically described as deterministic functions with random physical inputs, these computer codes represent expert approaches to the problem. Their outputs are thus typically random, and statistical tools can be used to estimate the failure probability. However, such code runs are usually time-consuming, therefore a large panel of methods has been developed to account for a moderate computational budget. The form constraints on the code output remain however little exploited, especially the properties of monotonicity that can be frequently encountered in such reliability studies. This article aims at presenting the benefits of accounting for this property: The probability can be deterministically bounded and its estimation can be considerably accelerated in terms of computational budget.
