Data-Based Global Control of the Part’s Geometry During Free-Form Bending

Freeform bending is a kinematically-controlled bending process that can be used to achieve complex 3D geometries. Geometry control of the bent part is particularly challenging because a geometric error that has already occurred, for example due to material fluctuations, can only be detected after a certain amount of time and after it has left the bending unit. In order to still meet various quality criteria on the part geometry, the error can be compensated for at a later point in the bending process and at a different location in the part by adjusting the amplitudes, thus minimizing a global quality measure. However, this requires a global consideration of the amplitudes and their effect on the resulting geometry. The aim of this paper is to train a neural network based controller that is able to improve the global geometry of the bent component for different optimization criteria with respect to material variations. This is done by training a data-based surrogate model that is fast enough to perform inverse optimisation of the global part geometry. This surrogate model is trained with simulation data and provides the training data to train a neural network that makes the control decisions in the process.