Maintenance optimization of multiple component systemsusing probabilistic graphical models

Maintenance optimization is a difficult task in today's manufacturing environment, especially when the system has multiple components. Thus, it is essentially critical to cope with the uncertainty and the complexity of the systems while deciding on the correct maintenance actions. Taking maintenance decisions in a planning horizon is one of the well-known stochastic sequential decision problems under uncertainty. Partially Observable Markov Decision Processes (POMDPs) are powerful tools for such problems under uncertainty in partially observable stochastic environments. However, since their state spaces can quickly explode with the increasing number of variables, POMDPs may not be preferable for addressing maintenance problems of multi-component systems. Factored representations are used for POMDPs by exploiting the inherent factored structure of the problem. This study aims to demonstrate how to formulate the maintenance problem of systems consisting of partially observable deteriorating components using factored POMDPs on two maintenance problems. The first one is an experimental model to perform in depth sensitivity analyses and to compare with some predefined policies proposed in the study. The second model belongs to a real-life implementation in thermal power plants. Sensitivity analyses are conducted under various scenarios with several settings. The results show that factored POMDPs are advantageous in modeling, solving and analyzing of maintenance problems with multi-components. Furthermore, the generated factored POMDP policies perform considerably better than the myopic policies.