Multivariate Process Analysis With Lattice Data

Abstract

Modern dimensional inspection techniques often produce more measurements per part than parts measured within reasonable timeframes. This poses a problem for multivariate process monitoring and capability analysis: The sample covariance matrix is not positive definite, hence not full rank and not invertible. When the measurement sites form a multidimensional lattice, spatially stationary covariance models provide positive definite estimates regardless of the number of measurements per part. I show that these estimates may be used in place of the sample covariance matrix to extend, and in some cases improve, standard multivariate methods. I describe a general class of lattices for which positive definite estimates are obtained via simple averaging or a closed-form EM algorithm. The proposed estimation and analysis procedures are illustrated in three case studies.

KEY WORDS:

• EM algorithm
• Exploratory data analysis
• Patterned covariance matrix
• Positive definite estimation
• Process capability analysis
• Spatial correlation
• Statistical process control