A data-driven recurrent event model for system degradation with imperfect maintenance actions

Although a large number of degradation models for industrial systems have been proposed by researchers over the past few decades, the modeling of impacts of maintenance actions has been mostly limited to single-component systems. Among multi-component models, past work either ignores the general impact of maintenance, or is limited to studying failure interactions. In this article, we propose a multivariate imperfect maintenance model that models impacts of maintenance actions across sub-systems while considering continual operation of the unit. Another feature of the proposed model is that the maintenance actions can have any degree of impact on the sub-systems. In other words, we propose a multivariate recurrent event model with stochastic dependence, and for this model we present a two-stage approach which makes estimation scalable, thus practical for large-scale industrial applications. We also derive expressions for the Fisher information so as to conduct asymptotic statistical tests for the maintenance impact parameters. We demonstrate the scalability through numerical studies, and derive insights by applying the model on real-world maintenance records obtained from oil rigs. In the online supplemental material, we provide the following: (i) sketch of proof for likelihood, (ii) convergence analysis, (iii) contamination analysis, and (iv) a set of R codes to implement the current method.