Diagnosing manufacturing variation using second-order and fourth-order statistics

This article discusses a method that can aid in diagnosing root causes of product and process variability in complex manufacturing processes, when large amounts of multivariate in-process measurement data are available. A linear structured model, similar to the standard factor analysis model, is used to generically represent the variation patterns that result from the root causes. Blind source separation techniques form the basis for identifying the precise characteristics of each individual variation pattern in order to facilitate the identification of their root causes. The second-order and fourth-order statistics that are used in various blind separation algorithms are combined in an optimal manner to form a more effective and black-box method with wider applicability.