On Fault Identification of MEWMA Control Charts Using Support Vector Machine Models

Multivariate exponentially weighted moving average (MEWMA) control charts are widely used for detecting small mean shifts in manufacturing processes. However, the MEWMA control chart can only give out-of-control signals but provide no information on which variable or subset of variables that leads to the out-of-control signals. We propose a SVM (Support Vector Machine) based MEWMA fault identification model to help understand the underlying cause of the out-of-control signals. For each process variable, we build a SVM model for each variable to classify the out-of-control data of each variable into three classes: no mean shifts, downward mean shifts and upward mean shifts. The classification results are combined into the fault identification results. We also examine the effects of SVM parameters on classification performance and provide a SVM parameter optimization method.