Condition-Based Maintenance Optimization of a Large-Scale System with a POMDP Formulation: Evaluation of a Heursitic Policy

This study investigates the question of the evaluation of a heuristic condition-based maintenance policy applied to a distributed multi-unit system. In particular, the system is composed of many units which function and degrade independently. We call it distributed since the system's total output is the sum of the individual output of each unit, where the failure of one unit has no impact on the functioning state of other units (i.e., no series/parallel structure). Taking into account the large-scale nature of the problem, which consists of a large number of units, is crucial because a good maintenance policy should coordinate the decisions at the scale of the system. Said differently, maintenance decisions cannot be taken independently unit-by-unit for the following reasons. First, the maintenance resource is limited and should be wisely allocated across the system. Second, as deploying on-site a maintenance crew is expensive, a good maintenance policy should also try to limit the number of deployments and group maintenance interventions. A condition-based maintenance policy relies on condition monitoring information, which we assume to be imperfect. We model this imperfection by assuming that remote sensors inaccurately estimate the true degradation state of the units. The decision-maker should then choose, at each time step, whether a maintenance operation or an inspection must be performed based on the information one has collected. We formulate the problem as a partially observable Markov decision process (POMDP). However, due to the curse of dimensionality, it cannot be solved via well-known approximate dynamic techniques. We then propose a heuristic algorithm based on a decomposition of the problem. The contribution of this work is to propose a framework to evaluate and validate this algorithm. We first validate the approach on a realistic-sized instance and show that the obtained policy has the expected properties (in terms of structure or value of information for different qualities of condition monitoring). Second, we validate the design of our procedure by showing that in a variety of scenarios, our heuristic performs better than its simpler (and more naive) alternatives.

Keywords: Condition-based maintenance, Partially Observable Markov Decision Process (POMDP), Large-scale multiunit system, Heuristic policy validation.

Download PDF