Manufacturing process classification based on heat kernel signature and convolutional neural networks
Manufacturing process classification based on heat kernel signature and convolutional neural networks
复制标题
DOI：
10.1007/s10845-022-02009-9
复制
发表时间：
2022-08
影响因子：
8.3
通讯作者：
Zhichao Wang;David Rosen
中科院分区：
文献类型：
--
作者：
Zhichao Wang;David Rosen
制造工艺分类的问题是确定适合于给定零件设计的制造工艺。以往对制造过程自动分类的研究受到数据集小、准确率低的限制。为了解决第一个问题，构建了一个由12,463个样本和四个制造工艺组成的更大的数据集。为了提高分类精度，提出了一种基于深度学习的CAD模型分类方法。首先，根据零件的三角形网格表示计算热核特征(HKS)。为了处理不同模型中不同数量的顶点，即每个模型中HKS的大小不同，提出了两种替代方法。第一种方法是将所有顶点分类成恒定箱，再将其送入传统的CNN进行分类。另一种方法通过最远点抽样搜索最具代表性的局部HKS，然后将具有代表性的局部HKS送入逐点CNN进行分类。本白皮书的范围仅侧重于使用零件形状对制造工艺进行分类，同时承认尺寸比例、公差和材料等其他信息在制造工艺选择中起着重要作用。结果表明，仅需零件形状信息即可获得较好的工艺分类性能。
The problem of manufacturing process classification is to identify manufacturing processes that are suitable for a given part design. Previous research on automating manufacturing process classification was limited by small datasets and low accuracy. To solve the first problem, a larger dataset composed of 12,463 examples with four manufacturing processes was constructed. To improve classification accuracy, a deep learning-based method for CAD models was proposed. To begin with, the heat kernel signature (HKS) is computed from a triangle mesh representation of the part. To deal with different numbers of vertices within different models, i.e., different sizes for HKS within each model, two alternative methods are proposed. The first one applies binning to sort all vertices into constant bins, which is further fed into a conventional CNN for classification. The other method searches the most representative local HKS by farthest point sampling, and the representative local HKS are then sent into a pointwise CNN for classification. The scope of this paper focuses on the use of only part shapes for manufacturing process classification, while acknowledging that other information such as size scales, tolerances, and materials, play important roles in manufacturing process selection. Results demonstrate excellent process classification performance with only part shape information.