Process fault detection and isolation: Integrating statistical techniques and analytical redundancy

The detection and isolation of faults in engineering systems have been of great practical importance. Early detection of the occurrence of faults is critical in avoiding product deterioration, performance degradation, major damage to equipment and harm to human health, or even loss of life. The goal of this dissertation is to investigate integrated fault detection and isolation methods which can be applied to complex industrial processes.; Principal Component Analysis (PCA) and other multivariate statistical methods have recently become the tools of choice in the monitoring of complex chemical processes. However, they provide little support for fault isolation. It has been shown recently that a close equivalence exists between PCA and parity relations, which belong to the analytical redundancy methods (Gertler and McAvoy, 1997). The analytical redundancy methods have well developed fault isolation capabilities. However, one needs a process model to use such methods to isolate faults. By virtue of the equivalence, the fault isolation capabilities in parity relations may be transferred into the PCA framework, resulting in methods which combine the convenience of statistical-type modeling with the powerful isolation properties of analytical redundancy.; The first approach presented is Isolation Enhanced PCA (IEPCA). Instead of requiring a model, the IEPCA approach generates the structured residuals algebraically, relying on a single PCA model. These residuals have the same isolation properties as analytical redundancy residuals. Fault isolation for dynamic systems is also studied. Three concepts are introduced to handle the time dependency in dynamic systems: time-lagged data, pseudo-variables and time-shifted relations. Application results showed the capability of the IEPCA method in both static and dynamic cases.; IEPCA is a linear method and the extension of IEPCA to nonlinear systems is very difficult because it relies on an algebraic transformation. Through an alternative to IEPCA, Partial PCA, the extension to nonlinear systems can be achieved. In the partial PCA framework, structured residuals are obtained by performing separate eigenstructure analyses for subsets of the variables, chosen in accordance with the incidence matrix. Based on two nonlinear extensions of PCA and the cluster analysis technique, three nonlinear extensions of the partial PCA approach are proposed in this dissertation: “Generalized” Partial PCA, Nonlinear Partial PCA and Clustered Partial PCA. Application results show that these nonlinear approaches give better representations of the systems and higher accuracy in fault isolation than the linear approaches.