AI-Based Motion Control with System Dynamics Flexibility

Abstract

In today’s production machines, algorithms are getting more complex every day to keep up with market demands for increased productivity and flexibility. Precise motion control tasks tailored to a specific machine type often require months of development. This paper proposes a method to create these algorithms through reinforcement learning and artificial neural networks to reduce development time and increase transferability. The reinforcement learning agent is trained on a digital twin of the machine, allowing for fast learning times and robust solutions by training on multiple simulated machines simultaneously. The artificial neural network is integrated into the control loop of an electric motor that consists of a programmable logic controller, a frequency converter and the motor itself. In addition, we propose a concept that allows the neural network to adjust the given setpoints to the underlying system dynamics of the used frequency converter and electric motor. With this concept, the neural network is able to adapt the setpoints in such a way that the following errors between the setpoints and actual values are reduced and an overall smoother control is achieved. This not only enables an optimal control but also provides machine builders with flexibility in their design of the machines since the same neural network can be used for various different machine configurations.

Authors 4

1. Niklas Körwer

   Siemens (Germany)

   Affiliation as printed

   Technology Siemens AG,Munich,Germany

2. Martin Bischoff

   Siemens (Germany)

   Affiliation as printed

   Technology Siemens AG,Munich,Germany

3. Michael Laumen Aachen ISEA Institute

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,ISEA Institute,Aachen,Germany

4. Rik W. De Doncker Aachen ISEA Institute

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,ISEA Institute,Aachen,Germany