Surface Variation Modeling by Fusing Multiresolution Spatially Nonstationary Data Under a Transfer Learning Framework
Surface Variation Modeling by Fusing Multiresolution Spatially Nonstationary Data Under a Transfer Learning Framework
复制标题
DOI：
10.1115/1.4041425
复制
发表时间：
2018-10
期刊：
影响因子：
--
通讯作者：
Jie Ren;Hui Wang
中科院分区：
文献类型：
--
作者：
Jie Ren;Hui Wang
高清计量（HDM）由于可以详细揭示空间表面变化，因此在表面质量检测方面受到了极大关注。由于其成本和耐用性，此类 HDM 测量偶尔会实施。这一限制创造了一个新的研究机会，通过将质量检查期间从有限的 HDM 数据获得的见解与广泛可用的低分辨率表面数据融合起来，改善表面变化表征。使用 HDM 进行的最先进研究得出的一个有用见解是，表面高度与某些可测量协变量（例如材料去除率 (MRR)）之间所揭示的关系和正相关性。假设这种关系在空间上恒定，并与表面测量相结合，以改进表面质量建模。然而，当协变量与不同表面积上的表面高度具有非平稳关系时，即协变量与表面高度关系在空间上变化时，该方法会遇到挑战。此外，非平稳关系只能通过 HDM 捕获，当大多数训练数据以低分辨率测量时，这增加了表面建模的挑战。本文提出了一种迁移学习（TL）框架来应对这些挑战，通过该框架，将来自 HDM 测量表面空间模型的公共信息迁移到仅提供低分辨率数据的新表面。在此框架下，本文开发并比较了三种表面模型来表征非平稳关系，包括两种基于变系数的空间模型和一种基于推理规则的空间模型。进行了现实世界的案例研究，以证明所提出的改进表面建模的方法。
High-definition metrology (HDM) has gained significant attention for surface quality inspection since it can reveal spatial surface variations in detail. Due to its cost and durability, such HDM measurements are occasionally implemented. The limitation creates a new research opportunity to improve surface variation characterization by fusing the insights gained from limited HDM data with widely available low-resolution surface data during quality inspections. A useful insight from state-of-the-art research using HDM is the revealed relationship and positive correlation between surface height and certain measurable covariates, such as material removal rate (MRR). Such a relationship was assumed spatially constant and integrated with surface measurements to improve surface quality modeling. However, this method encounters challenges when the covariates have nonstationary relationships with the surface height over different surface areas, i.e., the covariate-surface height relationship is spatially varying. Additionally, the nonstationary relationship can only be captured by HDM, adding to the challenge of surface modeling when most training data are measured at low resolution. This paper proposes a transfer learning (TL) framework to deal with these challenges by which the common information from a spatial model of an HDM-measured surface is transferred to a new surface where only low-resolution data are available. Under this framework, the paper develops and compares three surface models to characterize the nonstationary relationship including two varying coefficient-based spatial models and an inference rule-based spatial model. Real-world case studies were conducted to demonstrate the proposed methods for improving surface modeling.