Reinforcement Learning-Based Maintenance Scheduling for Resource Constrained Flow Line System

摘要

This paper considers the problem of condition-based maintenance for a two-machine flow line which consists of an upstream production machine, a downstream production machine and an immediate buffer with maintenance resource constraints. Due to the constrained maintenance resource, the maintenance actions are imperfect, and they are purposefully initiated when the machines are operating on a deteriorated quality states represented by multiple decreasing yield levels. A continuous-time semi-Markov decision processes model is formulated to describe the machine degradation processes. The distributed model-free average reward reinforcement learning algorithm, semi-Markov average reward technique algorithm is used to implement within each machine and its adjacent buffer and determine the optimal maintenance policy of the total system. The numerical results show that the approach can converge to the approximate optimal solution. Also, the effects of the variation of some parameters on the optimal policy and on the average cost rate are given.