Statistical dynamic modeling and monitoring methods for multivariable continuous processes

Industrial continuous processes have a large number of process variables and usually are operated under closed loop control. On-line statistical process monitoring (SPM) ensures that their operation safety, product quality, and productivity meet the expectations. New methods for statistical modeling and SPM techniques are proposed and developed. The techniques are based on multivariate statistics and system theory. The methods are especially appropriate for continuous chemical processes operated under closed-loop control. The parameter change detection (PCD) method is proposed for tightly monitoring vital process variables that are highly positively correlated. The canonical variate (CV) stochastic realization method is proposed for monitoring large numbers of process variables which yield measurements that are collinear, crosscorrelated and autocorrelated. Simulation studies and experimental work at the National Center for Food Safety and Technology, Chicago, Illinois and the National Institute of Standards and Technology, Gaithersburg, Maryland have been used to refine and validate the techniques developed.;The performance of the PCD method is compared to the performance of residuals based SPM techniques via Monte Carlo simulations and experimental trials. The results indicate that PCD performs better than the classical residuals based SPM tools in monitoring univariate measurements that are highly positively autocorrelated. The CV realization is used to obtain vector autoregressive moving average (VARMA) type models which explain the autocorrelations and crosscorrelations between multiple process measurements in such a way that residuals (innovations) become serially independent. CV state space models were successfully built for processes with 6, 22, and 5 measurement variables. CV residuals are used to audit process sensors by resolving multiple sensor faults. A single statistic and pair wise trajectory plots based on the orthogonal CV states are used for monitoring the total variability of an experimental process with six measurements.