Process Monitoring Using an Online Nonlinear Data Reduction Based Control Chart

Abstract

Recent advances of multivariate Statistical Process Control (SPC) show that the introduction of Principal Component Analysis (PCA) methods for reduction of process data is a promising area in system monitoring and fault diagnosis. The advantage of these techniques is to identify sets of variables which describe the key variations of the operating data and which allow process handling and control based on a reduced number of charts. However, because the basic PCA method stipulates that relationships between process characteristics are linear, the application of such techniques to nonlinear systems that undergo many changes has been limited in many real cases. In order to overcome this issue, some recent studies suggested the use of nonlinear adaptive PCA methods in order to track process variation and detect abnormal events at early stages. For this reason, this study develops and analyses an online Kernel PCA chart as a key technique to model nonlinear systems and to monitor the evolution of non-stationary processes. Results based on an analysis of a simulated process show that the control chart is robust and provides a reduced rate of false alarms with high fault detection abilities.