Abstract

Maintenance is a pivotal aspect of manufacturing systems, particularly those operating on a large scale. With the advent of data-driven methods and machine learning technologies, new avenues have opened for optimizing maintenance policies. In light of this, this thesis introduces advanced methodologies in Reinforcement Learning (RL) and Deep Reinforcement Learning (DRL) specifically tailored for large-scale parallel manufacturing systems. We conducted two major studies to advance the field: In the first study, an RL-based algorithm is proposed, moving beyond the traditional focus on system degradation levels to instead concentrate on the count of failed or unhealthy units. This shift allows for a more dynamic and nuanced approach to maintenance. Through Q-learning, our algorithm demonstrated significant superiority over conventional methods such as value iteration, particularly when applied to a system with four parallel units. In the second study, we delve into Deep Reinforcement Learning, developing a framework designed for multi-unit
systems experiencing stochastic degradation and unforeseen failures. Unlike traditional methods, our DRL approach incorporates a more intricate reward function that considers a wide array of factors ranging from production costs to maintenance crew deployment. Notably, this study was rigorously tested on a system comprised of 30 parallel units, making it particularly relevant for real-world, large-scale applications. Our research significantly broadens the applicability of machine learning methodologies in maintenance scheduling, demonstrating both robustness and adaptability. These contributions not only validate the efficacy of data-driven approaches in real-world settings but also lay the groundwork for future research in this crucial domain.