Reinforcement learning in series production: Multivariable optimization of machine tools to increase energy efficiency

Increasing energy prices and ecological demands are moving companies towards more energy-efficient production. Machine tools are one of the main consumers in a production facility. Control configurations, such as adjustments to pump pressures, of machine tools are either neglected in terms of energy efficiency or performed manually based on extensive domain knowledge. Modern manufacturing facilities are characterized by a landscape of heterogeneous production systems, creating a need for a generic approach to perform overall optimizations from the operator's perspective. The high complexity of production systems is accompanied by an extensive amount of process data, which favors the use of AI-based methods. However, those methods are often not compatible with the requirements in terms of process safety, so that an integration into series production can have a conflicting effect with respect to the process goals. In this paper, an approach to process optimization using reinforcement learning is presented and evaluated in view of securing the overall equipment effectiveness of a machine tool operated in series. The implementation is carried out in the series production of a production plant for engines and components of Mercedes-Benz AG. Initial savings of up to 12% of the total electric power consumption have been successfully demonstrated. © 2023 The Authors.