Gaussian and non-Gaussian Double Subspace Statistical Process Monitoring Based on Principal Component Analysis and Independent Component Analysis

Abstract

This study proposes a new statistical process monitoring method based on variable distribution characteristic (VDSPM) with consideration that variables submit to different distributions in chemical processes and that principal component analysis (PCA) and independent component analysis (ICA) are, respectively, suitable for processing data with Gaussian and non-Gaussian distribution. In VDSPM, D-test is first employed to identify the normality of process variables. The process variables under Gaussian distribution are classified into Gaussian subspace and the others belong to non-Gaussian subspace. PCA and ICA models are respectively built for fault detection in Gaussian and non-Gaussian subspaces. Bayesian inference is used to combine the monitoring results of the two subspaces to create a final statistic. The proposed method is applied to a numerical system and to the Tennessee Eastman benchmark process. Results proved that the proposed system outperformed the PCA and ICA methods.