線上學習神經控制器

An On-Line Learning Neuro-Controller

凃瑞澤(Jui-Rze Too) ； 黃瑞圖(Jui-Tu Huang) ； 陳標達(Biau-Dar Chen)

《大葉學報》 6卷1期 (1997/12) Pp. 61-71

https://doi.org/10.7119/JDYU.199712.0061

摘要

通常神經網路在程序控制的應用有三個主要的缺點：（1）需要長時間的訓練，（2）需要大量的訓練數據，（3）不易線上學習。為了解決上述的問題，本研究提出一種線上學習並能同時控制的神經控制器，此一學習控制式神經控制器已能突破傳統神經控制器的瓶頸並改善不易線上學習的限制。此控制器設計的基本概念為“學習即控制、控制即學習”，其特色在於神經網路不是去學習控制的模式，而是使程序的實際溫度去趨近於學習目標（溫度控制器的設定點）。並將此學習控制式神經控制器應用在CSTR的溫度控制，以便了解其性能。電腦模擬的結果顯示該神經控制器擁有良好的控制能力，對於設定點的改變及外在的擾動，皆能使系統迅速地達到穩定控制。

關鍵字

神經網路 ； 程序控制 ； 學習控制式神經控制器

並列摘要

In general, there are three major drawbacks when a neural network is applied to process control. A network needs a lot of training data and takes long time to train. In addition, it can not be trained on-line. To overcome these problems, we have proposed a learning control neuro-controller (LCNC) which can provide control action while it is in training on-line at the same time. The basic concept of this neuro-controller is ”training is control and control is training”. Its major characteristic is that the neural network does not learn control mode instead it tries to make the actual output temperature of the process to match with its target value (the set point of the temperature controller). We have applied this learning neuro-controller to temperature control of a CSTR. Computer simulation has been carried out to evaluate the performance of the neuro-controller. Simulation results demonstrate that the proposed neuro-controller is able to bring the system back to its set point regardless of whether the system encounters a set point change or an external disturbance.

並列關鍵字

neural network ； process control ； learning control neuro-controller