Research On Active Fault-Tolerant Control Of Nonlinear Systems Based On Learning Approach

Since modern systems trend to be more and more complicated, it is crucial to ensure their higher safety and reliability. Therefore, Fault Tolerant Control (FTC) is receiving more and more attention. FTC means that it can still accomplish the control task with a desirable or degraded (but acceptable) performance when some components in the system are faulty. Many results have been achieved in the research field of FTC. Unfortunately, most of the results are for linear systems.Due to the imperfect of theories in the nonlinear control system, most existing results are only for special nonlinear systems, and the study on the fault tolerant control of nonlinear systems is extremely limited. So there is an urgent need of further research in this field.Fault tolerant control is generally classified into two types: passive FTC (PFTC) and active FTC (AFTC). The former does not need the fault diagnosis, therefore, its design may inevitable be conservative and its performance may not be optimal. Even worse not only the performance but also the stability of the system may not be guaranteed when the system suffered unkown faults. The latter can use various information of fault to react to the fault actively. So AFTC is the research focus. Now the AFTC of nonlinear systems studied much is the method based on artificial intelligent technology. It's mainly divided into three kinds: modeling, estimation and control using artificial intelligent technology. Although the method is especially suitable for complex systems its most undesirable weakness is difficult to analysis the systems stability theoretically.The methods presented in this dissertation overcome the difficulties analytically modeling the nonlinear system by using the artificial technology to estimate the system's model from input-output data directly. And to avoid designing the observers or filters for nonlinear systems the unkown faults are estimated using the artificial technology. This dissertation focus on several different nonlinear systems to develop several fault compensation laws derived with the Lyapunov stability theory so as to realize the bounded stability under faulty conditions.For affine systems, a control structure based on TLC is presented for nonlinear normal system.The fault model is established using artificial neural network. The fault accommodation laws are established using Lyapunov methods from parameter variety. Thus the stability of the fault-suffered system is ensured.For affine systems, based on study above, considering the fact which the procedure of TLC design can become impractical when the plant dynamics are complex due to the repeated applications of time-varying differentiations. A TLC controller based on neural network is used to control the normal system. The stability of the system is theoretically proved considering the neural network modeling error and uncertainty and faulty which are considered as regular perturbation to the nonlinear normal system.The stability region is dedued under certain assumptions.For the systems described by NARMAX model SVM is used to establish the normal system's regression model. The unkown faults are diagnosed by the residual error between the regression model output and the real-time output. Then the normal controller's parameter regulation laws are established using Lyapunov methods. So the bounded stability under faulty conditions is achieved.For the systems described by NARMAX model SVM is used to establish the system's positive model. And it is adapted on-line learning the post-fault dynamics. Then the inversion of the SVM model is used as a controller to maintain the system stability and control performance after fault occurrence. Based on the adaptive SVM model the extended Kalman filter algorithm is developed to form an iterative inversion to evaluate the optimal control variable.Finally, the proposed methods are applied into the jet engine compression system model and three-tank system model which both are the well-regarded benchmark problems. Feasibility and efficiency of the proposed methods are validated by simulation results. The fruits of this dissertation has enriched the design theory of the FTC and meanwhile expanded the application fields of learning approach. The results also provide guidance for improving the safety and reliability of the complex systems.