Abstract - Gertler 9714891 Implicit modeling approaches to fault detection are data driven and are based on multivariate statistical methods. Principal component analysis (PCA) has been applied to model normal chemical process behavior, unusual events are then detected by referencing the observed process variable against this model. While these approaches are effective at detecting process faults, they do not specifically address the issue of fault isolation. Analytical redundancy methods have been created by the electrical engineering control community, which rely on explicit models of the monitored plant. Based on the model, consistency (parity) relations are derived which return residuals indicative of faults. These equations may be viewed as generalizations of material and energy balance calculations. The parity relations may be so designed ("structured") that their residuals follow clear patterns in response to various faults, thus providing a convenient means of fault isolation. The fundamental results of parity relations have been developed for linear systems, they may be applicable to nonlinear systems (the category for most chemical manufacturing facilities). The PI's are planning a combination of the above two methods, multivariate statistical techniques and model-based monitoring methods, to isolate faults. They plan to build a structured set of partial PCA models using the following steps: 1. A full PCA will be performed to determine the number of relations among the variables; 2. A structure matrix will be designed utilizing results from the theory of parity relations; and 3. A number of partial PCA models will be created, according to the selected structure, by simply omitting variables from the full PCA model.
