Detecting The Number Of Operational Modes In Baseline Multivariate SPC Data

When building a multivariate statistical process control model, it is commonly assumed that there is only one operational mode in the baseline data. However, multiple operational modes may exist due, for example, to several suppliers of raw materials or seasonal changes. It is important to know the number of modes in the data in order to construct an effective process control system. Each operational mode appears as a cluster in the baseline data. This paper proposes a new method to identify whether there is one cluster, the most common case, or more than one cluster. If there is more than one, the proposed method identifies the correct number. Unlike the many existing clustering methods, the proposed method has the following three features that are critical when analyzing baseline data. The proposed method can identify if there is only one cluster, a capability that some existing clustering methods do not have. The proposed method can identify clusters that are convex or non-convex. And finally, the proposed method is insensitive to user-specified parameters. The paper includes a comparison of the existing and proposed methods on four datasets, which include data from a simulated multiple-zone industrial oven under PID control and data from a real system. The proposed method consistently gives the correct number of clusters while the existing methods do not.