即時適應性類神經控制在非線性多輸入多輸出系統上之追蹤

Abstract

本篇論文提出一種應用遞迴神經網路模型為學習基礎之非線性多輸入多輸出系統之即時適應性類神經滑動增益追蹤控制理論。使用非線性狀態回授來設計控制輸入項，其中包含由對角遞迴神經網路即時估測系統的多個未知參數來建立模型。這個使用滑動更新模式設計導出的適應性學習法則，可以在短時間內調整遞迴神經網路的權值到最佳模型狀態。模型系統的誤差會漸近收斂至一個小的區間 ，且滑動模式控制補償了此類神經近似的誤差。只要所有的滑動面達到一個小的範圍，則反饋式滑動增益的調整就會停止，此舉也可以避免不必要的參數漂移現象發生。本篇論文也證明了導出的閉迴路系統是穩定的，而多輸入多輸出非線性系統軌跡的追蹤也可完成。最後，也將此即時適應性遞迴神經網路控制器與其他適應性修正型倒傳遞神經網路控制器作比較，可得到較好的結果。

A approach for on-line adaptive neural control of MIMO nonlinear systems is explored in this thesis with sliding gain using diagonal recurrent neural network (DRNN) learning model. We use the nonlinear state feedback to design the control input of a nonlinear system containing several unknown parameters of nonlinear systems. These parameters are estimated by diagonal recurrent neural networks with on-line modeling. The resulting adaptive learning law, which is designed by sliding mode update, can adjust the weights of DRNN to the optimal value for modeling in short time. System modeling errors are asympotatically converged to the small region , and a sliding-mode control which compensates for the neural approximation errors is proposed. The adaptation of the feedback sliding gain will stop as soon as all sliding surface have reached a small range. As a result, the undesirable parameter drift phenomenon can be avoided. It is proved that the resulting close-loop system is stable and the trajectory tracking of MIMO nonlinear systems is achieved. Some simulation results are also provided to evaluate the design. Finally, simulation results show that on-line adaptive DRNN controller provide better performance than the other adaptive modify back-propagation neural controller.