Selective Maintenance Optimization Under Limited Maintenance Capacities: A Machine Learning-Enhanced Approximate Dynamic Programming

Selective maintenance, as a prevalent maintenance policy for engineered systems under limited maintenance resources, has been widely adopted in industrial and military settings. Most existing works on selective maintenance only considered limited consumable resources, such as time and budget. In many real-world applications, the implementation of maintenance activities, however, has to be supported by several maintenance capacities, including repairpersons and repair facilities, which are occupied during the execution of a maintenance action and released upon its completion. Moreover, due to the stochasticity of maintenance actions’ durations, the available number of maintenance capacities in each period is also uncertain. In this article, a novel dynamic selective maintenance model is introduced by taking account of the limited maintenance capacities and stochastic action durations. The dynamic optimization problem is formulated as a Markov decision process where maintenance actions are dynamically selected in accordance with components’ states, remaining consumable resources, and available maintenance capacities. An approximate dynamic programming algorithm based on the rollout policy is put forth to estimate the optimal selective maintenance policy, and an importance-based heuristic is proposed to enhance its performance. To overcome the inefficiency of the solution algorithm in coping with large-scale instances, a machine learning-enhanced framework that enables neural networks trained by small-scale instances to solve large-scale instances is developed. A manufacturing system and an aircraft fleet are exemplified to demonstrate the effectiveness of the proposed approach.