An image-based, unified feature extraction framework for experimental and synthetic data for data-driven modelling in stamping and bending

Abstract

Stamping describes cutting and forming processes in a combined, usually multi-stage process. The order of the process sequence, material, its composition, spring-back behavior and any pre-deformations of the material, influence the result of such a stamping process. In order to model interdependencies, data-driven process models are increasingly used, which among other things use synthetic data from finite element (FE) simulations for training purposes. One issue that arises in this context is the acquisition of synthetic training data, which in general does not take into account the acquisition of the corresponding experimental data. This is especially problematic against the backdrop of holistic process and uncertainty modeling for both data types. To overcome this challenge, we propose a unified feature extraction framework, which models the experimental data acquisition in a virtual setting, using the same post-processing steps for synthetic as well as experimental data. The mesh data of the FE simulation is transformed into an image representation, considering specifically the highly non-convex nature of the geometry of the component as well as the resolution achieved by the experimental component inspection. The following post-processing steps are applicable to both experimental as well as synthetic data. An uncertainty quantification of the implemented framework was derived based on the existing CAD model of the considered component, which served as the ground-truth. Particularly good results were obtained for the detection and extraction of the bending angles, which are of importance for quantifying the springback during bending operations.