Quantitative approaches for fault detection and diagnosis in process industries

Fault Detection and Diagnosis (FDD) is an important task that should be carried out continuously to ensure the safe and reliable operation of plants in chemical industries. A significantly large amount of research in FDD assumes the availability of an accurate mathematical model of the system under observation. The absence of such a model makes a majority of the approaches unsuitable for chemical processes.; The focus of this thesis is on quantitative techniques for FDD. It covers methods which use data to identify models for a variety of applications including FDD. Models developed using these techniques have varying fault diagnostic capabilities, from simple process monitoring to more detailed fault isolation.; A powerful novel approach for the joint identification of steady-state models and the noise covariance matrix (Iterative Principal Components Analysis (IPCA)) is reviewed. This method has analytical advantages such as, accurate determination of the number of principal components. Industrial case-studies are used to show the applicability of this method. Online predictions of Bitumen Recovery in an Oil Sands Extraction Plant using this approach is presented.; Structured residual approaches for fault isolation are reviewed. The importance of optimal residual generation and choice of decision rule are illustrated. The use of a novel single-testing decision rule combined with the Structured Residual Vector (SRV) approach is shown to improve the sensitivity of the approach. The applicability of the SRV approach for sensor and actuator fault diagnosis under closed-loop conditions is demonstrated. A novel continuous-time model-based extension of the SRV approach for the diagnosis of additive process faults, which cannot be satisfactorily isolated using existing techniques, is presented.; A novel strategy for identification of chemical processes with irregularly sampled outputs using the Expectation Maximization (EM) algorithm is presented. The applicability of this method for multivariate state-space identification is demonstrated. This method also produces optimal state estimates and predictions via a Kalman predictor-corrector mechanism constructed during the identification. Online predictions of quality variables in the bleaching operation of a Bleached-Chemi Thermo-Mechanical Pulp (BCTMP) mill, are presented.; Illustrative case-studies are included to demonstrate the theoretical ideas in this thesis.