Reinforcement learning for opportunistic maintenance optimization

Abstract

Intelligent systems, that support the maintenance of production resources, offer real-time data-based approaches to optimize the maintenance effort and to reduce the usage of resources within production systems. However, unused potentials remain regarding maintenance schedules with minimal opportunity costs of the measures taken. This work provides a novel, machine-learning-based approach for the exploitation of these remaining optimization opportunities as an exemplary extension of the current state of the art. The determination of an optimal maintenance schedule for parallel working machines, is based on the data of a production system. The main result of this work is the performance of the implemented reinforcement learning algorithms, both in terms of downtime reduction, which increases the production output, and in terms of reducing maintenance costs compared to existing maintenance strategies. Hence, this work provides a holistic approach to the optimization of maintenance strategies and gives further evidence of a meaningful applicability of reinforcement learning algorithms in manufacturing processes.