Incorporating second order statistics in process monitoring

Abstract

One of the most applicable approaches in data driven process monitoring techniques is Principal Component Analysis (PCA). This approach assumes existence of uncorrelated stationary observations. Restriction of PCA-based process monitoring approaches on distribution function of observations has turned attentions to the use of Independent Component Analysis (ICA) algorithms. ICA is based on the assumption that at most one of the sources is Gaussian. Therefore, recent process monitoring approaches are based on FastICA algorithm which maximizes non-Gaussianity. As process variables can have any form of distribution function, implementing a method that has the ability to face all of the situations improves the monitoring quality. While both PCA-based and ICA-based monitoring approaches are restricted methods, this paper proposes extracting both Gaussian and non- Gaussian sources through Second Order Blind Identification for process monitoring. Besides, a new criterion for sorting sources is introduced. The applicability of the proposed method will be investigated through Tennessee Eastman Process.