Real-time monitoring of chemical processes based on variation information of principal component analysis model

In industrial processes, the change of operating condition can obviously affect the relations among process data, which in turn indicate the corresponding operating conditions. Considering that the loadings and eigenvalues, generated from the principal component analysis (PCA) model, contain primary data information and can reflect the characteristics of data, this article proposes novel monitoring statistics which quantitatively evaluate the variation of these two matrices, collected from real-time updated PCA model for process monitoring. Given that abnormal data may be submerged by normal data, a combined moving window which selects both real-time data and normal data is employed to collect data for model construction. By comparing with other PCA-based and non-PCA-based methods through a simple numerical simulation and the Tennessee Eastman process, the proposed data-driven method is demonstrated to be effective and feasible. Additionally, some other PCA-based methods are utilized for comparison.