Unsupervised and supervised learning for the reliability analysis of complex systems

Abstract

In this paper, a strategy to deal with high-dimensional reliability systems with multiple correlated components is proposed. The goal is to construct a state function that enables the classification of the states of components in one of two categories, that is, failure and operative, in case of dealing with a large number of units in the system. To this end, it is proposed a new algorithm that combines a factor analysis algorithm (unsupervised learning) with local-logistic and isotonic regression (supervised learning). The reliability function is estimated and system failures are predicted in terms of the variables in the original state space. The dimensions in the latent state space are defined by blocks of units with a certain dependence structure. The flexibility of the model allows quantifying locally the effect that a particular unit has on the system performance and a ranking of components can be obtained under the philosophy of the Birnbaum importance measure. The good performance of the proposal is assessed by means of a simulation study. Also a real data case is considered to illustrate the method.