Monitoring the Variation in Your Multivariate Process: An Introduction to the MVP Procedures

Complex processes in modern manufacturing and business environments can generate hundreds and even thousands of process measurements that vary over time. Early detection of process instability is critical for avoiding costly failures and minimizing risk. When the process measurements are correlated, multivariate statistical process monitoring methods are appropriate. Three new procedures in SAS/QC® 12.1, the MVPMODEL, MVPMONITOR, and MVPDIAGNOSE procedures, implement methods that are based on a principal components approach to process monitoring, which was developed in the field of chemometrics. They provide T 2 and SPE charts, which are multivariate summaries of process variation. An example from social media sentiment analysis illustrates how the procedures work together and demonstrates the power of the methods for discovering and diagnosing unusual variation. INTRODUCTION Multivariate process monitoring based on principal components is an effective approach for dealing with hundreds or thousands of correlated process measurements. This approach was introduced during the 1990s for applications in the chemical process industries (Kourti and MacGregor 1995, 1996). In manufacturing applications, the goal of statistical process monitoring—more commonly referred to as statistical process control (SPC)—is to create a stable, predictable process by identifying and removing special causes of variation. In a business environment on the other hand, this form of SPC is seldom feasible because it is often impossible to eliminate the special causes. Nevertheless, process monitoring can be used to better understand process variability and to detect problems early on, thereby minimizing risk and reducing costs. An example of this type is early detection of sentiment change in social media, which is used in this paper to illustrate the use of new SAS/QC procedures for multivariate process analysis. The new MVPMODEL, MVPMONITOR, and MVPDIAGNOSE procedures in SAS/QC software enable you to use multivariate measurements that are collected over time to monitor a manufacturing or business process for unusual variation. You use these procedures in the following order: 1. You use the MVPMODEL procedure to create a principal components model that uses a small number of components to characterize the variation in the data. It builds a principal components model and saves the loadings and scores in output data sets. 2. With this model, you use the loadings and scores data sets as inputs to the MVPMONITOR procedure to create multivariate control charts of the T 2 and SPE (squared prediction error) statistics to find unusual variation. 3. If you discover unusual variation, you can use the MVPDIAGNOSE procedure to help diagnose and interpret the variation. This paper refers to these three procedures collectively as the MVP procedures. Statistical process control often occurs in two phases: 1. In Phase I, it is not assumed that the process is stable. The goal is to identify (and remove) special causes of variation by constructing a control chart from an initial set of data. You use all three MVP procedures in a Phase I analysis. 2. In Phase II, the goal is to monitor a stable process by constructing a control chart for new data with control limits derived from a previously established model for stable variation. You use model loadings saved by the MVPMODEL procedure as input to the MVPMONITOR procedure. Figure 1 shows how the MVP procedures work together. The time of measurement is displayed on the control chart although it is not used in the creation of the model. Although other SAS® procedures can perform a principal components analysis, the MVPMODEL procedure provides convenient features for working with the MVPMONITOR and MVPDIAGNOSE procedures. The MVPMODEL procedure also supports cross validation for selecting the number of components. 1 Statistics and Data Analysis SAS Global Forum 2012