Using external information for statistical process control

Abstract

<p>This research presents practical solutions for several problems in the case of multivariate statistical process control (MSPC) for process monitoring and fault diagnosis of industrial processes. Various types of external information are used with multivariate statistical models to deal with fault detection and isolation (FDI) problems. The first part of the research deals with the fault isolation problem. It is based on the fact that MSPC approaches have been found to be very useful for fault detection, but less powerful for fault isolation because of the non-causal nature of the data. To improve fault isolation, an approach is proposed that uses additional data on past faults to supplement existing contribution plot methods. This approach extracts steady-state fault signatures of faults in both the correlation model space and the residual space. The use of transient fault trajectory or initial fault signature would minimize detection delay and false isolation of a fault. An indirect usage of process dynamic information to deal with FDI problems is proposed in the second part of the work. Since process signals represent the cumulative effects of many underlying process phenomena, multiresolution analysis via wavelet transformations is used to decompose signals. The third part of this thesis examines both the fundamental and the practical differences between the causal and statistical model-based approaches to FDI. The causal-model-based approach is based on causal state variable or parity relation models developed from theory or identified from plant test data. Faults are then detected and isolated with structured or directional residuals from these models. MSPC approaches are based on non-causal models built with multivariate latent variable methods using historical process data. Faults are then detected by referencing future data against these covariance models, and isolation is attempted through examining contributions to the breakdown of the covariance structure. Most processes are subject to change of the operating conditions such as feed rate and composition, product grade, controller status, and so on. Sometimes these large common-cause variations disguise or distort the relevant information to faults. This difficulty can be minimized if one builds a correlation model that includes only the process variations relevant to FDI. By incorporating various types of prior knowledge into the empirical model building process, one can estimate a hybrid correlation model that includes both raw data and prior knowledge. Application of the hybrid correlation model for process monitoring and fault diagnosis is proposed and used for analyzing a real industrial dataset. (Abstract shortened by UMI.)</p>