High-Definition Metrology Based Spatial Variation Pattern Analysis for Machining Process Monitoring and Diagnosis

Abstract

Non-contact high-definition measurement technology, such as laser holographic interferometry, makes it feasible to quickly inspect dimensional variation at micron level, providing up to 2 million data points over a surface area of up to 300×300 mm2. Such high-definition metrology (HDM) data contain rich spatial variation information but it is challenging to utilize this information for process monitoring and control. The spatial distribution of the data is in high-dimensional form and may show nonlinear patterns. Conventional statistical process monitoring and diagnostic schemes based on simple test statistics and linear statistical process models are incapable of capturing the complex surface characteristics as reflected by large amounts of spatial data. This paper develops a framework for efficient monitoring of spatial variation in HDM data using principal curves and quality control charts. Since large scale surface variation patterns (caused by fixturing and part bending) may camouflage those in the smaller scale (generally associated with tooling conditions), it is essential to separate the patterns in these scales and monitor them individually. At each scale, process monitoring is implemented in a sequential manner by monitoring the overall spatial features followed by localized variation identification if an out-of-control condition is detected. To examine the overall spatial characteristics, a principal-component-analysis (PCA) filtered principal curve regression is proposed in conjunction with multivariate control charts whereby nonlinear patterns of spatial data are extracted and monitored. When the overall monitoring indicates a problem, the identification of a surface variation change can be achieved through localized monitoring over each surface region based on variogram pattern analysis and control charts. The location of surface region change provides clues for variation source diagnosis. The proposed method is illustrated using simulated HDM data.