Maintenance Planning with Deterioration by a Reinforcement Learning Approach - A Semiconductor Simulation Study

Manufacturing companies are often faced with deteriorating production systems, which can greatly impact their overall performance. Scheduling the preventive maintenance activities optimally is vital for these companies. This study utilizes historical data on product quality deterioration in the last machine of a real-world serial, buffered production line to refine maintenance decisions. Currently, maintenance intervals for the last machine are fixed. The objective is to improve this policy by increasing the production system's long-term throughput. A simplified two-machine-one-buffer (2M1B) system is modeled to devise and assess policies. Various optimization techniques are applied, with the average reward Reinforcement Learning (RL) technique showing the most promising results in numerical experiments. The RL-optimized policy exhibits significant potential by considering product quality deterioration, buffer levels ahead of the last machine, and the machine states of the last two machines.