Research On Robust Model-Based Fault Diagnosis Technology For Control Systems

The model-based approach has been widely used in condition monitoring and fault diagnosis of control systems. Unfortunately, the model errors or uncertainties, the shift of operating mode and the other disturbances can bring negative influences on diagnosis accuracy since they are inevitable. So it is very important to improve the robustness of the fault diagnosis system.The main topics of this dissertation are focused on the robust model-based fault diagnosis of control system. The main contents and results are as following:1. How the model uncertainties influence the fault diagnosis accuracy is analyzed deeply. Some basic methodology for improving the robustness of fault diagnosis is outlined.In view of main factors inducing uncertainties, the model uncertainties are formally described by stochastic model, set model, fuzzy model and the unknown input model respectively. Based on such models, the mechanism how the model uncertainties influence fault diagnosis accuracy is analyzed systematically. Some basic methodology is then obtained to reduce model uncertainties and improve fault diagnosis robustness2. For the linear uncertain system, the observer/filter-based robust residual generation method is applied to improve the robustness of fault diagnosis.(1)The OUIO (optimal unknown input observer) -based robust residual generation method is investigated. If some conditions hold true, it is optimal to reduce the influence of the model uncertainties.(2)A robust residual generation method is proposed based on multi-objective optimal program and satisfactory estimation of the uncertain systems. The influence of model uncertainties is limited without debasement of fault diagnosis sensitivity by optimizing H_∞norm , fault-free residual stable variance, and H_ norm simultaneously.3. Two adaptive threshold decision-making methods are proposed to improve the robustness for the linear uncertain system.(1)An adaptive threshold design method applicable to analytical model is presented based on the theory of probabilistic robust. The adaptive threshold is obtained based on the upper bound of fault-free residual which controls system inputs and reflects the model uncertainties. Therefore the FDR(Fault Detection Rate) and FAR(False Alarm Rate) can be guaranteed simultaneously. (2) In case of disturbances, whose bound is difficult to be acquired, another adaptive threshold design method is presented based on the fuzzy theory. The influence residual rule of model uncertainty is modeled and designed by applying fuzzy reasoning techniques. The threshold modified according to the rule can adapt itself to the influence of modeling uncertainties.4. For the nonlinear uncertain system, two robust residual generation methods are presented based on SVM (Support Vector Machine) to improve the robustness of fault diagnosis.(1) For the nonlinear uncertain system whose nonlinear function satisfies with the Lipschitz condition and whose model uncertainty can be described by limited boundary interfere, a robust fault diagnosis method is presented based on SVM observer. A SVM approximator is constructed to monitor the behavior of dynamic system according to in-suit learning of the nonlinear fault function. Theε-insensitivity loss function of SVM is used to restrict the influence of modeling uncertainty. The diagnosis robustness is enhanced.(2) For the strong nonlinear system which is difficult to build analytical model, a fault detection and diagnosis method is presented. based on LS-SVM (least squares support vector machines) The nonlinear relation between inputs and outputs is modeled by LS-SVM approximation. Because the model is adaptive to the change of dynamic process, the fault detection and diagnosis have better robustness.5. In order to evaluate the efficiency of the techniques proposed above, a BIT (Built-in Test)system of one mechatronical system in our Lab is implemented. Its two subsystems, the strap-down inertial navigation subsystem and the control system, are chosen to demonstrate. The experimental results show that BIT performance is improved and the false alarms are reduced after using the method presented in this paper.