Use of reinforcement learning for the automated optimisation of an injection moulding process during production

• Applicant: Professor Dr.-Ing. Reinhard Schiffers
• Subject Area: Plastics Engineering
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 548155694

The motivation for this research project arises from the fact that although the injection moulding process is considered to be very reproducible, the quality of the injection moulded products is not constant due to internal and external disturbance variables such as batch fluctuations of the raw material, temperature fluctuations in the vicinity of the machines, etc. The adjustment of the machine settings is carried out manually by trial-and-error or based on experience by the machine operator. Readjustment of the machine settings is done manually by trial-and-error or experience-based by the machine operator. Besides the necessary use of experienced skilled personnel, this can also mean an intervention of the machine operator in the process after bad part production. Current research is therefore investigating the use of data-based applications such as machine learning to predict part quality. The focus so far has been on the use of unsupervised and supervised learning methods. An alternative to these two methods is reinforcement learning. Reinforcement learning is an independent category of machine learning in which an agent independently develops strategies through interaction with its environment. In doing so, the agent pursues the goal of obtaining the best possible reward for its problem-solving behaviour. The data sets are generated dynamically and the learning procedure independently develops an action strategy that leads to the solution of a problem within a previously defined environment. The overarching goal of the research project is to independently control the injection moulding process with the help of a reinforcement learning approach in such a way that quality-relevant process parameters are automatically recognised and optimised. In this research project, quality is defined by the component weight, as this correlates with many quality parameters of moulded parts and can be recorded precisely and robustly. Partial requirements for the research project are the improvement of the component quality with the lowest possible data generation and low resource input. In order to achieve the overall goal, two sub-goals can be defined. Firstly, the methodological development of a reinforcement learning approach for use in the injection moulding process. This process takes place offline, i.e. without real component production on the injection moulding machine. The second step is the online phase and involves the technical implementation and optimisation of the developed reinforcement learning approach on the real injection moulding machine. The data generated in this research project using the reinforcement learning approach will be used for inline quality control of an injection moulding process.

• DFG Programme: Research Grants