Multivariate statistical techniques for modelling, monitoring, and controlling continuous processes

Statistical process control (SPC) and statistical quality control (SQC) techniques are popular in efforts to improve profitability, competitiveness, and safety of manufacturing processes. Standard SPC methods may not be satisfactory for multivariable continuous processes. Correlation in these processes can lead to an increase in false alarms and missed alarms. Quantitative means for isolation of likely disturbances is difficult to handle in a multivariate situation.A new method using the multivariate exponentially weighted moving average (EWMA) with PCA has been demonstrated to provide marginal improvements for detection of out-of-control conditions compared to existing techniques. Additional novel quantitative tools for diagnosis of disturbances in operation using PCA have been demonstrated.The use of angular information with PCA has been introduced to diagnose upsets, to compare models for different modes of operation, and to explore diagnosis of multiple faults. Applications of quantitative measures of similarity between PCA models have been related to success in diagnosis of single and multiple process faults.Forming an integrated, modular framework of quantitative process monitoring and disturbance diagnosis tools, the techniques have been applied to detect contamination of fermentation samples and to analyze performance of a plant-wide chemical system.Statistical modelling of these processes has been investigated to relate quality and process measurements using Kalman filtering and partial-least squares (PLS) regression. Data from normal process operation is used with PLS and principal component analysis (PCA) techniques to derive empirical linear models with statistical confidence limits to detect outliers. Besides providing baseline models for monitoring, PCA models for operation with a variety of process upsets have been built.