Online Maintenance Prioritization via Monte Carlo Tree Search and Case-based Reasoning in Complex Manufacturing Systems
Open Access
- Author:
- Hoffman, Michael L
- Graduate Program:
- Industrial Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- June 07, 2021
- Committee Members:
- Soundar Kumara, Co-Chair, Major Member & Dissertation Advisor
Russell Barton, Outside Unit & Field Member
Daniel Finke, Major Field Member
Michael Brundage, Special Member
Steven Landry, Program Head/Chair
Eunhye Song, Co-Chair & Dissertation Advisor - Keywords:
- maintenance optimization
condition-based maintenance
case-based reasoning
optimization via simulation
discrete event simulation
predictive maintenance
monte carlo tree search
artificial intelligence - Abstract:
- Maintenance serves a critical role in manufacturing systems by ensuring that machines and other assets remain in a productive working condition. The primary objective of maintenance optimization is to determine when to conduct maintenance and which machines should be maintained. Recent advances in industrial maintenance have sought to use online information obtained from sources such as machine sensors and manufacturing execution system software to provide real-time decision support. Such predictive maintenance strategies combine abundant online manufacturing data with techniques in simulation, planning, and artificial intelligence to make effective maintenance decisions and support the overall performance of the system. In this dissertation we examine several challenges associated with adopting real-time maintenance decision support in complex manufacturing systems. One such challenge is that of modeling complex machine configurations that are often found in modern manufacturing systems. It can be difficult or impossible to model the behavior of these systems analytically without imposing unrealistic simplifying assumptions. One of the goals of this work is therefore to propose a method of maintenance optimization and planning that is generalizable to arbitrarily configured systems. We also introduce a discrete-event simulation package that has been developed as a part of this work and is capable of modeling these systems of interest. Additionally, real-world manufacturing systems are typically subject to constraints on available maintenance resources which limits the number of maintenance jobs that may be conducted simultaneously. In these settings, the maintenance planner must determine how to prioritize competing maintenance activities and allocate these limited resources throughout the system. This work addresses these challenges by proposing a simulation-based maintenance optimization and planning approach to seek an optimal maintenance policy and prioritize maintenance in complex systems. We formulate condition-based maintenance policy optimization as a discrete optimization via simulation problem and seek a solution using the Gaussian Markov Improvement Algorithm and a genetic algorithm. The result is a degradation threshold for each machine in the system that determines when a machine should request maintenance. To overcome the problem of capacity-constrained maintenance resources, we first propose a dynamic priority scheduling heuristic that aims to minimize throughput disruption due to downtime for maintenance. We then improve upon this scheduling heuristic by employing a reinforcement learning approach to seek the best maintenance action in each state of the system. We use Monte Carlo tree search to progressively build a search tree within the system state space and evaluate alternative sequences of actions in order to find that which maximizes the expected reward. We demonstrate that our proposed method of online prioritization results in improved system-level performance when compared to commonly used maintenance prioritization methods. Furthermore, we apply a case-based reasoning framework to retain and reuse relevant experience that improves the decision-making efficiency over time. In addition to improved system productivity, the proposed approach results in reduced time needed to identify optimal maintenance actions which is particularly important when critical maintenance decisions must be made quickly.