A general methodology for reliability modeling of component failures and model discrimination using cross validation is developed. First, the requirements for collection of failure, maintenance, and operation data are outlined, including left and right censored data. Cross validation is then used as a probabilistic measure of predictive performance for selection of the optimal model from a set of reliability model candidates. In addition, cross validation is used to determine the classification or hierarchical decomposition of systems into component-classes which provides the best overall set of predictive models. As a measure of predictive performance for model selection, we demonstrate that cross validation is superior to likelihood function maximization and modeling error minimization, since both have bias for over-parameterized models and may not be generally applicable to reliability models with wear and shock variables, and with as-good-as-old maintenance. Case studies are used to demonstrate these points and the overall methodology.
