類神經網路最佳學習參數值找尋之研究:以空調溫濕度異常診斷與處理為 例

A Study on Finding the Optimum Learning Parameters Values of Artificial Neural Networks: Sampling with the Fault Diagnosis and Treatment of Air-Conditional System''s Temperature and Humidity

卓高平(Gau-Pyng Jwo)

指導教授 ： 葉若春 ； 鄭春生

元智大學/工程學院/工業工程與管理學系/碩士(1995年)

摘要

近年來，電腦之執行運算速度，快速增進，致使類神經網路(
Artificial Neural Networks)廣泛地應用於各領域中，然而，要構建好
的類神經網路的確是很複雜的，因其含有多個學習參數，如隱藏層數及其
運算元數、學習速率、動量因素、起始權數、學習樣本數等等，必須對其
一一做最佳設定，方能使類神經網路之學習與回想能力為最好。 本研
究是以國內一家相紙製造加工廠之空調系統溫濕度控制異常診斷與處理為
例進行研究，採用類神經網路之多層倒傳遞模式(Multilayer Back-
Propagation Model)進行學習訓練，計有十個輸入特徵訊號，以五個輸出
運算元做為異常診斷與處理。先以田口技術(Taguchi Techniques)設計方
法分析及找尋較佳之各學習參數值，作為啟始佳解，接著以其較顯著之學
習參數做為反應曲面法(Response Surface Methodology :RSM)之設計參
數，經實驗設計分析而後得最佳之各學習參數值，使類神經網路之輸出誤
差為最小，空調溫濕度異常診斷與處理效果為最好。

關鍵字

類神經網路 ； 最佳學習參數值 ； 異常診斷 ； 田口方法 ； 反應曲面法

並列摘要

The speed of the computer operation has been increasing in
recent years so that Artificial Neural Networks could be applied
to various fields widely. However, to construct a good
Artificial Neural Networks is complicated because it involves
several Learning Parameters, such as Hidden layers, Hidden layer
units, Learning rate, Momentum, Initial weight, Learning
samples, etc.. Therefore, they have to be well-set-up
respectively to make the learning and recall performance of
Artificial Neural Networks greater. This paper is based on
the fault diagnosis and treatment of air-conditional system''s
temperature and humidity control in a domestic film process
plant. It utilizes Multilayer Back-Propagation Model of
Artificial Neural Networks to proceed learning, including ten
input feature signals which use five output processing elements
as fault diagnosis and treatment. First, to analyze and find the
better learning parameters values as the initial optimal
solution by Taguchi Methods. Then, to use the significant
learning parameters as the designed parameters of Response
Surface Methodology, after the design and analysis of the
experiment, we obtain the optimal learning parameters values to
make the minimum output error of Artificial Neural Networks and
to achieve the best effects of the fault diagnosis and treatment
of air-conditional system''s temperature and humidity control.

並列關鍵字

Artificial Neural Networks ； Optimum Learning Parameters Values ； Fault Diagnosis ； Taguchi Methods ； Response Surface Methodology