Statistical Process Monitoring with Multivariate Exponentially Weighted Moving Average and Independent Component Analysis.

Abstract

The ever increasing number of variables measured in chemical and biological plants has led to increased emphasis on monitoring performance and fault detection in process system engineering. However, conventional T2 and squared prediction error (SPE) charts based on principal component analysis (PCA) and partial least squares (PLS) are ill-suited to detecting small disturbances resulting from process faults because these monitoring techniques only use information from the most recent samples. In this paper, a new statistical process monitoring algorithm is proposed for detecting process changes resulting from small shifts in process variables. This new algorithm is based on the multivariate exponentially weighted moving average (MEWMA) monitoring concept combined with independent component analysis (ICA) and kernel density estimation. ICA is a recently developed statistical technique for revealing hidden, statistically independent factors that underlie sets of measurements. In this research, three monitoring charts (I2, Ie2 and SPE) obtained using a combination of ICA and MEWMA are developed to better monitor processes undergoing small mean shifts with autocorrelation, where the control limits for these statistics are obtained by kernel density estimation. The proposed monitoring method is applied to fault detection in both a simple multivariate process and the simulation benchmark of the biological wastewater treatment process (WWTP). For a small shift in these processes, the simulation results illustrated the monitoring power of MEWMA-ICA and ICA-MEWMA versus various existing methods (conventional PCA, ICA, MEWMA-PCA and PCA-MEWMA monitoring).