Maintenance optimization for multi-component systems using Markov decision processes

Modern engineering systems are often comprised of multiple components that deteriorate with use. In a civil structure, this deterioration will, over time, lead to an unacceptable risk of failure. In manufacturing and production systems, component failures caused by deterioration lead to unforeseen downtime. Since more manufacturing processes are being automated and the requirements for safety become stricter, the maintenance function accounts for an increasingly larger fraction of the total operational costs for such systems. An effective maintenance policy is, therefore, necessary to compensate for this trend. Both industry and academia are now mostly focused on a maintenance paradigm called Condition-Based Maintenance (CBM), where maintenance activities are carried out based on the monitored deterioration state of the system. In a more traditional approach, Time-Based Maintenance (TBM), the next maintenance activity is scheduled based on the elapsed time since the previous maintenance activity without consideration to the condition of the system at the scheduled time. By monitoring the condition of the system, the CBM approach can potentially reduce the number of redundant maintenance activities and unforeseen failures. In this dissertation, we consider the problem of maintenance optimization in multicomponent systems, both for the TBM and the CBM approach. The dissertation is divided into six chapters, three of which are academic paper manuscripts. Many heuristic maintenance policies have been developed for multi-component systems. We focus on optimal maintenance policies identified using Markov Decision Process (MDP) models and dynamic programming optimization algorithms. Identifying the optimal policy in a general MDP model is computationally demanding if the state space has multiple dimensions. In paper A, we use one MDP state dimension for each component in a unifying model framework for the TBM and CBM approach in a multi-component system. We then perform numerical experiments to determine the practical computational size limit of the MDP, that is, the largest number of components in the system, for which we can obtain an optimal maintenance policy. In Paper B, we consider a CBM system with continuously deteriorating components and investigate the effects of discretization, which is a necessary step for dynamic programming optimization. We compare different methods for discretization and demonstrate that a relatively coarse discretization still results in a near-optimal maintenance policy. Even though the discretization is primarily a technical matter pertaining to the optimization procedure, we can also draw a connection between the results from one of the tested discretization methods and a common practice of classifying the system condition on a discrete scale. Many companies are currently seeking to improve their maintenance practices by implementing CBM. The rationale is that CBM is more cost-efficient than TBM, because maintenance can be performed just in time before a component fails in CBM. In Paper C, we quantify this benefit by comparing the performance of optimal TBM and CBM policies for a multi-component system. Specifically, we show how changing the number of components and the degree of stochastic and economic dependence between components affect the difference between the performance of the TBM and CBM policies.