Complementary database generation for machine learning for quality prognosis using the example of ring rolling

• Applicants: Professor Dr.-Ing. Alexander Brosius; Professor Dr.-Ing. Bernd Kuhlenkötter
• Subject Area: Primary Shaping and Reshaping Technology, Additive Manufacturing
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 499350001

The proposed research project deals with the central research question of the usability of machine learning (ML) methods by means of synthetically generated training data sets in the field of radial-axial ring rolling (RARR). For the use of ML in production areas such as RARR, data sets of good and reject parts have to be recorded. Here, balanced data sets are advantageous with regard to the ratio of good and reject parts. However, this is not the case for industrially acquired data sets. Therefore, the approach of employing data enhancement by synthetic data generated by simulations is used within this research project. For RARR processes, however, fast analytical simulations do not exist by which a sufficiently large number of synthetically produced data sets with "rolling defects in geometry or process" can be generated. For this reason, the research question is investigated to what extent a similar process can be used for the transfer to RARR. Here, the cold rolling of rings is a suitable process. The advantage is the reduction of complexity due to the exclusively radial forming instead of the parallel forming in the radial and axial direction in the RARR as well as the reduced temperature influence during cold ring rolling. A central advantage of cold rolling for the research project is the semi-analytical model of the cold rolling process developed by the Chair of Forming and Machining Processes (FF). This semi-analytical method is used to simulate the cold rolling process quickly and precisely in order to generate process data. This approach will be used in a first project phase to determine if synthetically produced data sets can be used for the training of ML. These results are generalised for RARR and defined in form of an experimental environment for ML in ring rolling. The learning algorithm shall be supported by the integration of domain knowledge to accelerate the process prediction or to impose limits according to process specific limits. Another central question is the required data quality of the real and synthetic data and how differences in data quality are reflected in the process prediction. In the second project phase, the findings will be transferred to the hot rolling process. Through the cooperation of the Chair of Forming and Machining Processes (FF) and the Chair of Production Systems (LPS) new research questions at the intersection of classic simulation methods and ML and new possibilities for research transfer can be addressed. The combination of the semi-analytic model of cold ring rolling and its extension to RARR (FF) with the research and evaluation of shape defects by means of data driven methods in the field of RARR (LPS) forms the approach for a new method of machine learning.

• DFG Programme: Research Grants