Automated Functional Basis Modeling Using Multi-Label Machine Learning on 3D Models

Abstract Recent progress in neural network-based object detection algorithms allows engineers to automatically detect and track objects using camera data or LiDAR data. These algorithms have been applied to many engineering applications to automatically detect obstacles, defects, or objects of interest. However, these algorithms are only trained to detect the class label of an object representing a physical entity, not the object’s function, a concept that is more nebulous than a single entity. Functional Basis Modeling is a technique developed in mechanical engineering to abstract an object’s functionality from the form of the object. This paper describes a methodology for the automation of Functional Basis Modeling using the PointNet algorithm. A set of point clouds extracted from 3D models of common products were manually labeled with multiple functional basis labels. The resulting dataset was used to train and validate a modified version of the PointNet convolutional neural network algorithm. PointNet is traditionally a multi-class classification algorithm, but the authors modified it to perform multi-label classification. The authors were able to train PointNet to predict several mechanical functions of an input object. The method achieves a single-label maximum F1 Score of 89.6%. This work paves the way for an AI text input CAD generator and more useful object detection methodologies.