Manufacturing process identification from 3D point cloud models using semantic segmentation
Xiaofang Liu, Zhichao Wang, Shreyes N. Melkote, David W. Rosen
Abstract
When a part requires more than one manufacturing process to produce it, traditional manufacturing process identification approaches do not work well. In contrast, we present a deep learning approach that utilizes semantic segmentation of the 3D CAD model of a part to automatically identify the manufacturing processes from a combination of casting, turning, and milling. To enable the approach, a dataset consisting of 3D CAD models of eight categories of common industrial parts manufactured using a combination of processes such as casting, turning, and milling is created. We utilize this dataset to develop two novel deep learning-based semantic segmentation methods for manufacturing process identification. The first method, MRIConv++ (Modified Rotation Invariant Convolutions), is purely geometry-based, utilizing a 3D point cloud representation of the part. To enhance its suitability for real-world applications, we develop the PRIConv++ (Perturbed Rotation Invariant Convolutions) method, which uses both geometric features and surface finish information as inputs. The PRIConv++ architecture processes perturbed 3D point clouds that model surface roughness to enable the identification of suitable manufacturing processes. We show that PRIConv++ provides a slight improvement in segmentation accuracy over MRIConv++, delivering state-of-the-art performance across all three manufacturing processes. While MRIConv++ is optimized for applications requiring only geometric features, PRIConv++ is better suited for scenarios involving complex surface finish characteristics, making it ideal for detailed semantic segmentation in real-world manufacturing environments. • Mapping CAD model geometric features to manufacturing processes enables accurate process prediction for each part feature. • Dataset of 8 part categories and 3 processes (casting, turning, milling) supports multi-process segmentation. • Achieves state-of-the-art 3D CAD segmentation for manufacturing process selection using MRIConv++ and PRIConv++. • Enhancing segmentation accuracy with point sampling and boundary post processing techniques.