| While multivariate methods have proven useful in modeling and monitoring chemical processes, difficulties arise in that sensors are inherently prone to malfunction and processes are susceptible to unmodeled upsets. It is the goal of this dissertation to address the deficiencies of current monitoring procedures and develop a unified methodology for accurately characterizing chemical systems in the presence of process and sensor variations. This entails the automatic detection, identification and diagnosis of disturbances and selection of the optimal error correction procedure.; A novel, automated method based on principal component analysis (PCA) is presented for the detection and identification of disturbed sensors during a process monitoring application. As opposed to previous approaches which are capable of identifying a disturbance in only a single sensor, the backward elimination sensor identification (BESI) algorithm can identify disturbances in multiple sensors.; A method integrating wavelet processing and techniques from multivariate statistical process control (MSPC) is presented, providing a means for the simultaneous localization, detection and identification of disturbances in process spectral data. Unlike the traditional MSPC approach where disturbance detection is carried out in the original wavelength domain using a single PCA model, detection employing wavelet transform processing results in the generation of multiple models within the wavelength-scale domain.; A voting system procedure, based on probabilistic and empirical rules, is described for distinguishing between process upsets and sensor malfunctions. In contrast to traditional voting techniques which require the strict duplication of sensor elements, the redundant sensor voting system (RSVS) presented is based on the concept of state redundancy. As an example, RSVS is successfully applied to process data from a liquid-fed ceramic melter.; The utility of sample weighting in updating regression models is explained theoretically and a leverage-based criterion for selecting weights for the new calibration samples is presented. In essence, weighting new samples more heavily improves the experimental design of the model, which is of particular importance when the new source of variance is represented by only a few calibration samples. Employing simulations and process data, a close correspondence is demonstrated between weights selected using prediction error and leverage criteria. |
