Most major companies invest heavily in advanced measurement and data collection technology for manufacturing quality assurance. One of the most promising measurement technologies that is broadly applicable in discrete parts manufacturing is noncontact dimensional metrology using laser and/or vision systems. Such optical coordinate measuring machines (OCMMs) produce large volumes of profile, point cloud, and high resolution image data that represent parametric and nonparametric surface geometry characteristics. OCMM technology is well developed, and accuracy and throughput are now at levels that render it suitable for quality control of parts with reasonably high precision. There is a large body of existing work on analyzing OCMM data that pertains to fitting specific geometric features for individual parts (e.g., fitting a circle to the perimeter of a drilled hole) for the purpose of verifying the dimensional integrity of the individual parts. However, although there is also a wealth of information on the nature of part-to-part variation buried in the rich and complex structure of the spatially dense OCMM data, there is currently no comprehensive and generic approach for uncovering this information. To fill this void, this research will develop a paradigm for identifying and visualizing complex part-to-part variation patterns in high-dimensional, spatially dense OCMM data, which will provide a powerful tool to facilitate the discovery and elimination of major root causes of manufacturing variation. Intellectual Merit: The emphasis on part-to-part variation here is fundamentally different than the emphasis of current OCMM data analysis software, which fits geometric features separately to individual parts. As such, the proposed methodology for discovering information from raw OCMM data represents a critical advancement in data analysis to keep pace with the tremendous advancements in noncontact dimensional metrology and data collection that have occurred in recent years. The approach is intended to be broadly applicable for discovering generic variation patterns in any type of OCMM data, requiring no prior knowledge nor premodeling of the types of anticipated patterns. To accomplish this objective, we propose a novel manifold learning framework for identifying and visualizing the variation patterns. Within this framework, the proposed research will address challenges that include handling measurement noise structure that differs from what is typically assumed in manifold learning, transforming OCMM data to reduce the extent of nonlinearity in the patterns, and handling heterogeneous data that represent different types (e.g., laser and image and/or different process stages) through transfer learning and sequential leaning methods. Broader Impact: Many discrete parts manufacturing industries invest heavily in OCMM equipment, but lack knowledge discovery tools for fully utilizing the technology to understand part-to-part variation. By creating a methodology for more effectively analyzing high-dimensional, spatially dense OCMM data, the proposed research fills a critical void. First, it will provide more sophisticated, badly needed six-sigma tools suitable for manufacturing variation reduction programs that employ modern OCMM technology, which will ultimately increase the competitiveness of US manufacturers. Second, it will allow for a greater return on investment in OCMM technology, which will increase its demand and spark further technological advances. The results will be integrated into six-sigma and data mining courses that the PIs teach to undergraduate and Ph.D. students and to an ethnically diverse spectrum of engineers and technical managers in professional masters' courses; this includes heavily enrolled Web conferencing courses. Infrastructure and dissemination will be enhanced by a Web service, through which collaborative teams can access the technology on a server to interactively visualize and study variation sources in their processes. This will provide a transformative research and education platform for variation reduction in discrete parts manufacturing, which will promote its use to large numbers of science, engineering, and technology students and professionals. The PIs have recruited and advised students from underrepresented groups in previous projects and plan to continue this here.
|Effective start/end date||8/15/13 → 7/31/16|
- National Science Foundation (NSF): $167,400.00