Model-based and data driven fault diagnosis methods with applications to process monitoring

This thesis discusses statistical model-based as well as data driven fault diagnosis approaches with applications to process monitoring.; A model-based technique, the Multiple Model Extended Kalman Filter (MMEKF), is proposed for the fault diagnosis of nonlinear stochastic systems. The MMEKF system consists of a bank of EKFs, where each filter is tuned to a specific fault. The residual, which is the difference between the measurement and the predicted measurement from the EKF, is an indicator of how well the filter is performing. Analysis of the residual can be used to evaluate the match between the actual measurement and the corresponding filter output, and therefore detect and isolate a possible fault. Hence, this multiple model approach has an inherent fault detection and isolation capability. A novel residual evaluation scheme combining a multivariate statistical technique, Principal Component Analysis (PCA), and an efficient search algorithm, is proposed to improve the fault evaluation process.; Simulation results on three different kinds of faults, actuator, sensor and process fault, show that the MMEKF fault diagnosis system can successfully perform the fault detection and isolation task. The fault magnitude can also be estimated by PCA residual evaluation technique.; Process monitoring based on PCA has been a very popular topic in the context of data driven fault diagnosis techniques in recent years. However, two fundamental statistical assumptions have limited the performance of the conventional PCA approach for many industrial process control applications. Three variations of the conventional PCA, namely Adaptive PCA, Moving PCA (MPCA), and Multi-Scale PCA (MSPCA), are presented to increase the range of applicability of PCA. Each of these three extensions deals with a certain limitation of the conventional PCA. When the process is normal but undergoing a slow change which is still considered to be normal, Adaptive PCA is preferred. MPCA and MSPCA might be good approaches for detecting a slow process drift since they are more sensitive to small changes.; An industrial application for the Adaptive PCA and simulation studies for the MPCA and MSPCA show superior performance over the conventional PCA.; In addition, these approaches are also expected to improve the robustness of residual evaluation for the MMEKF approach when the effect of model uncertainty and unmodeled dynamics and disturbance cause the residuals to deviate from the assumed white noise characteristics of an optimal filter.