Abstract
Recently, the application of machine learning on Computer-Aided Design
(CAD) models has emerged. However, there is a lack of robust methods for the
conversion of boundary representation (B-Rep) CAD models from engineering
software to appropriate input representations for a machine learning algorithm.
Those that do exist break the link with the B-Rep, meaning the ability to use
machine learning to support future engineering operations on the B-Rep are
challenging. This paper presents a method for the creation and labelling of point
clouds from B-Rep CAD models for machine learning techniques, while maintaining
a link between the two representations. This method allows for the creation of a
dataset with additional input features determined from the CAD model such as BRep face labels, that could increase the accuracy of certain problems when machine
learning is utilized. First, an open-source software called CloudCompare is used for
point cloud creation. Fast interrogation of the CAD model using face bounding
boxes are utilized to link points to their corresponding faces. This link allows for
easy traversal of the CAD model topology to gain other geometric features if
needed. A deficiency of the approach is that some B-Rep faces result in being
under sampled. For these insufficiently sampled faces, a method is presented to resample them. Experiments on the approach are outlined to illustrate the efficiency
of the proposed method with the approach taking approximately 10 seconds per
CAD model within the tested dataset.
Original language | English |
---|---|
Pages (from-to) | 760-771 |
Number of pages | 12 |
Journal | Computer-Aided Design and Applications |
Volume | 18 |
Issue number | 4 |
Early online date | Nov 2020 |
DOIs | |
Publication status | Early online date - Nov 2020 |