Research On Fault Tolerance Of Neural Network Controller Based On Error Estimation

With the fast development of pattern recognition and automatic control in theories andapplications, a new control field which is called intelligent control is now widely studiedusing intelligent techniques, for example, fuzzy, neural network and specialist systems.Intelligent control is proved superior to traditional control methods when tackling problemsinvolving high nonlinearity and serious model uncertainty. Therefore, intelligent control hasbeen paid much attention. Recently, many mature models and theories have developed inpattern recognition. Artificial intelligence techniques, such as genetic algorithms, supportvector machine, etc, have been successfully applied on image processing, data mining andinformation indexing, etc. Merging pattern recognition and automatic control will gain a hugebenefit.A control system may get interrupted by any kinds of disturbances, which can be outsideor inside the system itself. The controller will be downgrade or even crash by disturbances.How to tackle these disturbances and keep a control system away from their negative effect isa critical issue when designing controllers. In traditional control, scholars have introducedrobust control, fault tolerant control and disturbance rejection.In this thesis, the fault tolerant control with intelligent control methods is tackled bycombining the concept of the Localized Generalization Error Model (L-GEM) in patternrecognition research field. In pattern recognition, L-GEM is used to select certain parametersof a classifier and thus to improve its generalization ability. In our proposed learning method,some virtual points around the point represented the current status are generated to simulatethe current status that is affected by some disturbances. The controller is updated not onlybased on the error between actual output and desired output but also the error of the virtualpoints. In addition, the concept of multiple classifier system, or called ensemble, utilized inpattern recognition to eliminate internal limits of a single classifier, is applied to controllers. Aneural network controller ensemble with base neural network controllers is proposed toincrease the performance of control.The simulation results show that better fault tolerance is achieved by the neural networkcontroller with L-GEM when compared with traditional neural network controller. Moreover,when our neural network controller ensemble takes charge as the main control unit, it couldmake even better fault tolerant performance.