Unwind While Unwinding - Solving Complex Automation Problems with Machine Learning
Niklas Körwer, Martin Bischoff, Rik W. De Doncker · 2024
In this paper we want to propose a new form of control for winding and unwinding applications based on artificial neural networks that are trained through reinforcement learning. In many machines and plants of today’s industry, winding and unwinding applications are present. From steel mills to the paper industry up to production machines for everyday goods. Each of theses tasks poses different challenges and for each, specific unwinding and winding concepts have been developed. Our approach aims at creating a method by which these algorithms can be developed by using machine learning instead of conventional programming. This has the potential to quicken development time for these machines, as the reinforcement learning process is faster than developing algorithms through conventional programming. Our method also differs from most other approaches, where reinforcement learning is only used to enhance the conventional control or to tune specific parameters of the conventional control whereas our method aims at creating the entire control algorithm. The presented approach is tested on a real machine with promising results.