Understand and evaluate geometries of 3D CAD data for the Digital MockUp

Overview and questions

The project is located in the field of digital prototype construction in vehicle development. In the so-called Digital MockUp, the components of a vehicle, given as 3D CAD data, are secured with regard to their functionality. This is not only done individually for each component, but also for the interaction of the parts planned by many different designers. An important aspect here is to check that components do not conflict with their neighbours, i.e. do not occupy the same installation space. The collisions between components that occur frequently in practice sometimes show the engineers relevant errors that require a design modification of the components. However, the majority are irrelevant to the engineers. A frequent example are collisions involving small parts such as screws, bolts or clips. Their task is to attach components to each other and the collisions thus created are intentional. The evaluation of collisions in critical or uncritical is currently still done by experts. In GeoCADUp, AI-supported procedures for an automated classification and initial evaluation of collisions are developed.

Procedure

For the classification of 3D geometry data we train neural networks with self-rendered images of the objects and rely on proven, pre-trained models from image recognition. Our models recognize small parts with high reliability and on a variety of data sets.

In the field of image synthesis, we are researching alternative methods such as (rotationally invariant) cylindrical projections to improve the information content of the input images and to take into account the rotational position of the objects.

3D geometry data can be represented not only by images but also by point clouds. The neural network LocALNet developed in GeoCADUp achieves one of the world's best results for the classification of 3D CAD data based on point clouds in the academic competition on the MoldelNet40 dataset of the University of Princeton.

Neural networks based on point clouds can also be used for segmentation. LocALNet is trained on industrial data and used to segment motor connection points.

3D geometry data can be represented not only by images but also by point clouds . The neural network LocALNet developed in GeoCADUp achieves one of the best results worldwide in the academic competition on the MoldelNet40 dataset of Princeton University for the classification of 3D CAD data based on point clouds.

Desired results

With the results of GeoCADUp a higher degree of automation in virtual product development is achieved. The expert will receive specialized intelligent tools, which allow a safer and faster evaluation of collisions.

ManagementProf. Dr. Nicola Wolpert
Partner (external)invenio Virtual Technologies GmbH, Johannes Gutenberg University Mainz
Project e-mail addressnicola.wolpert@hft-stuttgart.de
FundingFederal Ministry of Education and Research (BMBF)
Call for proposal"Qualifizierung von Ingenieurnachwuchs an Fachhochschulen" ("Young engineers")
Duration01.05.2017 – 30.04.2021 extend until 30.06.2022

Team

Name and position Field Email and phone Room
Academic employee Research +49 711 8926 2633 2/583
Professor Mathematics +49 711 8926 2697 2/368
Akademischer Mitarbeiter Forschungsprojekt GeoCADUp +49 711 8926 2633 2/583