| More and more large dynamic processes with a high degree of automation demand that an on-line fault diagnosis function is incorporated into their supervision and monitoring system. In this dissertation, research literature on fault diagnosis since the 1970s are reviewed, and, to contribute to this area, new techniques have been invented. A new state and fault observer is developed for dynamic systems with discrete time stochastic models. This observer is able to give unbiased state estimates even in the presence of target faults. The fuzzy decision making is developed to remove the hard boundaries of hypothesis tests. It also provides a way to take advantage of experts' knowledge about the process and fault. Some improvements to state-of-the-art techniques are made and included in the fault diagnosis system to increase its effectiveness. An improvement to the parity space approach is proposed for redundant sensor management systems, and it is more robust than the original parity space algorithm to noises in the sensors. For general dynamic systems, a Kalman filter and Rauch-Tung-Striebel smoother combination is proposed to extract the fault information from the measurements and control signals. All of the above fault detection and isolation methods are applied to a dynamic fault diagnosis system designed for a power plant boiler model. Biased sensor, noisy sensor, biased control actuators, and waterwall leakage faults are simulated, and successfully detected and isolated. |
