Demonstrating Reinforcement Learning for Maintenance Scheduling in a Production Environment

Abstract

As the automation of production lines in modern manufacturing environments becomes ubiquitous, their flexibility and resilience become increasingly important. Consequently, the scheduling of maintenance activities is growing more complex and at the same time ever more crucial for ensuring adequate system availability. In this paper a digital model of a production environment is presented, using building blocks and restrictions that can be found in most modern production environments. Maintenance and repair activities in the model are scheduled by a reinforcement learning agent for different proof-of-concept scenarios, which can be optimised using measures such as maximizing production capacity and minimizing maintenance costs. The results of this paper provide the basis for further work to improve the working conditions of human maintenance planners by providing a reliable decision support system which facilitates the task of scheduling planned and unplanned maintenance activities.