Maintenance policy analysis for multi-unit systems with dynamic programming and reinforcement learning

Abstract

This thesis investigates the engineering systems consisting of multiple units subject to gradual performance degradation. This degradation eventually results in catastrophic system failure. Conventionally, system engineers try to reduce system failure by performing periodic maintenance. However, this time-periodic maintenance policy often considers states of the units only not the inter-relationship among the states of different units. Meanwhile, due to the advancement of the technologies, real-time based state identification is possible. In this thesis, we analyze the state-dependent maintenance policy considering the state of the units as well as the inter-relationship among the states of the units. The maintenance decision process is formulated with a Markov decision process (MDP) and stochastic dynamic programming (DP) is used to find optimal maintenance policies. With the insights gained from the MDP with the DP approach, a model-free methodology using reinforcement learning (RL) is proposed the validity of the model is verified.