Reduction of rejects Research project wants to make massive forming processes more stable

Editor: Alexander Stark

Germany — At the Machine Tool Laboratory WZL of RWTH Aachen University, a new research project for process optimisation and quality enhancement in massive forming has been launched. The project examines whether machine learning algorithms can be used to automatically adapt to instabilities and thus reduced rejects in the forming process.

Related Companies

The new research project "Irlequm" for process optimisation and quality improvement in massive forming starts at the WZL of the RWTH Aachen.
The new research project "Irlequm" for process optimisation and quality improvement in massive forming starts at the WZL of the RWTH Aachen.
(Source: Mubea Tailor Rolled Blanks)

Instabilities due to external influencing variables, unknown interactions between process parameters or quality characteristics of products often lead to rejects in massive forming processes despite existing process controls. Current control concepts are based on implicit operator knowledge and rely on manual adjustment of the process parameters. It is often not possible to adjust the processes in time to maintain the manufacturing tolerances of the products. Quality control loops are a means of overarching compensation for quality deviation. In combination with machine learning approaches, such as reinforcement learning and transfer learning, they offer the potential to reduce rejects. This is done by automatically adapting the system parameters when instabilities occur.

The aim of the research project called “Irlequm” is therefore the development of a procedure based on reinforcement and transfer learning for the implementation of novel controllers in quality control loops of solid forming processes. In order to enable a reinforcement learning-based control, the necessary IT infrastructure is first defined and implemented. Such a control offers the advantages that, on the one hand, all quality-relevant information, such as process parameters, environmental conditions or raw material properties, can be included in the control. On the other hand, the implicit operator knowledge of the control can be made permanently usable.

In order to reduce the learning time of the reinforcement learning algorithm and to save resources, it is not trained directly on the real process, but on a stochastic process simulation. The knowledge gained from the simulation is then transferred to the quality control process of the control loop by means of transfer learning. The result of the research project will be a quality control system for solid forming processes that regulates processes automatically, comprehensively and in real time and optimises the quality of the processes. The increased process quality will in turn increase the quality of the products and reduce rejects.

The research project, with a project duration of three years, started on 1 June 2021 and is being carried out in cooperation with the Chair of Production Metrology and Quality Management, the Chair of Manufacturing Process Technology (both from the Machine Tool Laboratory WZL of RWTH Aachen University) and the companies Mubea Tailor Rolled Blanks GmbH (consortium leader), Eichsfelder Schraubenwerke, Iconpro, Schomäcker Federnwerk, Quality Automation and the associated partners Mawi and Schiller Pressen.

(ID:47487722)