Data-driven Process Modelling in Stamping and Bending Technology

• Description, explanation and prediction of complex interactions in the stamping and bending process
• Data-driven modelling of the stamping and bending process with a combination of physical models and robust neural networks
• First formal specification and formal verification of forming processes with robust neural networks based on reachability analyses

The aim of the research project is to describe, explain and predict the complex interactions in stamping and bending processes using data-driven modelling approaches. In addition to real process data, classically motivated simulation models and implicit domain knowledge are to be used for the data basis. Influence modules are to be used to introduce specific disturbances that enable a comprehensive view of the stamping and bending process. Of particular interest are the defined quality criteria, which can be both geometric and physical in nature. As can be seen in Figure 1, each process step of a selected stamping and bending process is to be modelled using a combination of a physical model and a learned neural network. Robust neural networks trained with a novel set-based training method will be used for this purpose. The linking of the models of the individual process steps forms a model of the complete process chain, which is to be formally verified against the defined quality criteria on the basis of achievability analyses. Furthermore, a new type of quantity-based conformity is to form the basis for the formal verification of the process model, the investigation of process robustness and the transfer of implicit specifications into a formal description. Finally, the selected process is to be optimised on the basis of the model by identifying the most robust process sequence and the most robust active and effective surfaces.

The expected results of the project include the realisation of an industry-oriented and real production process for stamped-bent components. Furthermore, concepts will be developed and implemented that enable the introduction and measurement of disturbance variables in the stamping and bending process. Novel quantity-based training methods for robust neural networks are also being developed. In addition, a new type of set-based conformity should enable the transferability of implicit specifications of the model to the real process