Robust Fault Diagnosis Of Nonlinear Systems

Fault detection and diagnosis (FDD) is receiving more and more attention due tothe increasing demands for higher safety and reliability of dynamic systems, where themodel-based FDD strategy is a hot research area. Due to the universal existence ofnonlinearities and model uncertainties in practice, robust fault diagnosis of nonlinearsystems is of great significance in theory and practice, but few results were presentedbefore. In this dissertation two robust FDD strategies for nonlinear systems with uncer-tainties are mainly studied, i.e. unknown disturbance decoupling and adaptive learning.Firstly, two new nonlinear filtering algorithms with disturbance decoupling areproposed, which are then successfully applied in the nonlinear robust fault detectionand isolation. The first presented algorithm is based on the well-known extendedKalman filter (EKF), and the convergence is also proved using linear matrix inequality(LMI) technique. However, the EKF has deficiency in applications due to the lin-earization procedure. Therefore a disturbance decoupling particle filtering is proposed,which has much wider application scope. Based on it, the fault diagnosis of nonlinearstochastic systems are achieved with help of multi-model validation method based onlog-likelihood ratio.The adaptive learning strategy is then deeply studied. Firstly, the fault detectionand estimation problem is considered, where the detection speed is increased by in-troducing the silde mode item in the observer design, which is proved in theory andillustrated by simulations. The modification of adaptive law ensures the boundednessof estimation errors. The fault isolation problem is then considered, where an adap-tive separated estimation algorithm is presented which have many advantages over thenormal adaptive observer. Adaptive thresholds for fault detection and isolation are alsoappropriately designed, and the fault detectability and isolability conditions are sys-tematically analyzed by the defined concept of fault response space. Since the systemdescriptions in the two works above are limited, an adaptive observer for the nonlinear Lipschitz system, which is a more general case, is proposed and the stability is proveddue to small gain theorem. Then it is applied in the fault diagnosis under the similarframework of previous works.It is found that the two fault diagnosis strategies above could be unified in theframework of state and parameter estimation of nonlinear systems. Therefore, twonew hybrid intelligent optimization algorithms are proposed to estimate the states andparameters directly based on the moving horizon estimation (MHE) theory in order todiagnose fault.Last, some future directions are pointed out, such as optimal performance design,residual evaluation, and FDD for networked control systems.