Advanced neural network controllers and classifiers based on sliding mode training algorithms

Publication Type:
Thesis
Issue Date:
2006
Full metadata record
Files in This Item:
Filename Description Size
Thumbnail01front.pdf1.06 MB
Adobe PDF
Thumbnail02whole.pdf6.29 MB
Adobe PDF
This thesis presents the research undertaken to develop some novel learning algorithms based on the sliding-mode control techniques for the neural network classifiers and controllers. Although the feedforward neural network with the backpropagation learning algorithm is the most widely used approach for classification and control applications, the slow convergence rate. the local minima problem, the difficulties in system identification and the lack of robustness are the issues existing for these neural networkbased systems. The combination of the sliding-mode control techniques and the backpropagation algorithm, as described in this thesis, leads to three novel learning algorithms, which offer effective solutions for these problems. The first learning algorithm, derived from the integration between the chattering-free sliding-mode control technique and the backpropagation algorithm, can obtain fast and global convergence with less computation. Experiment results relating to the headmovement neural classifier for wheelchair control show that the proposed approach considerably improved the convergence speed, global convergence capability and even the generalisation performance of the neural network classifier, in comparison with various popular learning algorithms. The second learning algorithm, also derived from the integration between the chattering-free sliding-mode control technique and the backpropagation algorithm, can guarantee the stability and robustness of the neural control system with parameter uncertainties. Based on this stable neural controller, a neural control design methodology is developed for a class of uncertain nonlinear systems with transportation lag, wherein a new training procedure is proposed to avoid the difficult choice of the training inputs always associated with the conventional neural network identifier. The implementation results with a real-time Static VAR Compensator system indicate the effectiveness of the proposed method. The third on-line learning algorithm, developed from the reaching law method combined with the backpropagation algorithm, offers a robust adaptation approach for the neural control systems with parameter uncertainties and disturbances. The neural control approach is further developed to design a novel decentralised neural controller for a class of uncertain large-scale systems with bounds of interconnections and disturbances. The stability and robustness of the neural control system are guaranteed based on the Lyapunov synthesis. Real-time implementation results for a Coupled Electric Drives CE8 system show the effectiveness and feasibility of the proposed approach.
Please use this identifier to cite or link to this item: