Introduction to Multivariate Process Monitoring Procedures

The first step in using the MVP procedures is to build a principal component model by running PROC MVPMODEL on an initial sample of process data, which is labeled "Phase I Data" in Figure 10.1. PROC MVPMODEL produces tabular and graphical output to enable you to determine an appropriate number of principal components for the model. You can also use cross validation to have the procedure automatically select the number of principal components.

After you determine the number of components, you use PROC MVPMODEL to save the eigenvalues, principal component loadings, and other information that describes the model to an OUTLOADINGS= data set. At the same time, you can save the Phase I data and the corresponding principal component scores, residuals, and and SPE statistics to an OUT= data set.

Next, you use the MVPMONITOR procedure to determine whether the process is stable. PROC MVPMONITOR reads the OUT= data set that is produced by PROC MVPMODEL as a HISTORY= data set and creates control charts for the and SPE statistics that are computed from the Phase I data. A point that falls outside a chart’s control limits indicates unusual variation.

The purpose of using a control chart to signal unusual variation is to distinguish special cause variation from common cause variation. Special causes, also referred to as assignable causes, are local, sporadic, or transient causes of variation, whereas common causes are inherent in a process. You should investigate out-of-control points to determine whether they signal assignable causes of variation.

Contribution plots show how the original process variables contribute to variation displayed by the and SPE charts. PROC MVPMONITOR can produce contribution plots automatically for out-of-control points. You can also use the MVPDIAGNOSE procedure to produce contribution plots of observations from the HISTORY= data set.

Based on your Phase I analysis, you can proceed to monitor the process by using the existing principal component model or collect more Phase I data and build a new model. The goal is to have a model of a stable process before you move on to Phase II analysis.