Training robust neural networks for uncertainty prediction in stamping technology

Stamping and bending technology uses a sequence of stamping and bending steps to manufacture complex metal parts. The execution order and design of each step are crucial for the quality of a produced part. However, simulating different process orders is hard due to uncertainties in the production process. We propose a data-driven model using neural networks to optimize the execution order to reduce the effects of uncertainties on the deviation in product quality. The process steps are modeled individually by neural networks, which are concatenated to model different process orders. To ensure accurate predictions, we use set-based training to make the neural networks robust against input uncertainties. We demonstrate the usefulness of our model by numerical experiments for the production process of a busbar.