Informative Score-Loading Plots For Multi-Response Process Monitoring (pp. 13-19)
Authors: Ergon, R. (Telemark University College, Porsgrunn, Norway)
Abstract: An ordinary principal component regression (PCR) latent variables model, using any number of components, can be reduced to a model with as many components as the number of response variables. This is done by projection onto a subspace spanned by the PCR estimators for the responses. By defining new orthogonal loading vectors in that subspace, a new latent variables model with exact correspondence between loadings and scores is obtained. For the special case with two response variables, the result of this can be shown in a single score-loading biplot, and in this plot the contributions to the score from the different regressor variables can be shown by use of contribution vectors. A confidence ellipse defining upper control limits for score movements may also be included. The confidence ellipse and the contribution vectors give both fault detection and fault diagnosis.