Wasserstein generative adversarial networks for form defects modeling

Geometric deviations of mechanical products are specified by tolerancing in the design stage for a functional purpose. In order to verify the impact of geometric deviations on functional surfaces while considering the manufacturing process, form defects have been considered in tolerance analysis in recent years. As a digital representation of geometrical defects in mechanical parts and assemblies, Skin Model Shapes enables the rapid and comprehensive generation of non-ideal shapes from either measurement or via data augmentation using simulation approaches. This paper presents a novel method for form defects modeling using Generative Adversarial Networks (GAN). The form defects of cylindrical surfaces considering machining process are represented and used for training a Wasserstein GAN. The pre-trained network is able to generate realistic form defects for cylindrical Skin Model Shapes rapidly and automatically without explicitly formulated representations. Manufacturing errors in turning process are considered in this approach and the generated samples from WGAN can be re-used for generating new cylindrical surfaces with a mapping strategy considering specification. A case study of a cylindricity specification is used in the paper to illustrate the effectiveness of the proposed method.