Transparent AI-Based Process Modelling in Die Forging

• Applicants: Dr.-Ing. Kai Brunotte; Professor Dr.-Ing. Marco Huber
• Subject Area: Primary Shaping and Reshaping Technology, Additive Manufacturing ; Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 520195047

The aim of the research project is to improve quality characteristics in die forging processes by advancing the understanding of the complex collective of interacting factors. Data-based models in combination with global explanatory methods will help to identify correlations and process deviations that have not been understood so far, so that approaches can be developed that allow process stabilization in future generations even with lower safety factors such as flash or oversizing The existing domain knowledge regarding the correlations between process and target variables in hot forging is compiled, structured and evaluated in the form of a correlation matrix. In order to generate process data on a sufficient scale, an existing series forging press will be digitized by means of a data processing and database infrastructure. Sensors are integrated into this system to enable the collection, processing and exchange of all process data. Since the component geometry is an essential optimization variable and it must also be predictable by an AI model, an automated image processing system will extract the relevant geometry data from 2D image data. Reference data sets for the analyzed process will be generated in serial forging tests supported by numerical process simulations. These are expanded with further test series under introduction and elimination of disturbances as well as with varied control variables. In addition, FE simulations of the individual processes are carried out and included in the data set collection in order to support the identification of interactions in general with noise-free data. Since various methods exist in ML that can be used for quality feature prediction, several learning algorithms are tested and compared to each other. The goal is to identify the algorithm that allows both a high prediction quality of different quality criteria and a later interpretability of the results. Based on this model, the latest algorithms for surrogate generation can be applied, resulting in a white-box model that is used to identify the input variables responsible for the prediction. Various explainability algorithms are compared with each other. The goal is to obtain an explainable white-box model that has maximum fidelity compared to the original model. Finally, an evaluation of the determined correlations with respect to their plausibility and novelty is carried out by comparing them with the compiled domain knowledge. Furthermore, the model is validated and the potential for process optimization is evaluated on the basis of forging tests with a derived new generation of tools with modified dies.

• DFG Programme: Priority Programmes
• Subproject of: SPP 2422: Data-driven process modelling in metal forming technology
• Co-Investigator: Professor Dr.-Ing. Bernd-Arno Behrens