臺灣博碩士論文加值系統

最小平方法在參數鑑別的應用上有下列下列限制:(1)無法處理非線性系
統;(2)受限於資料相關性與有限的觀測數目.近年來蓬勃發展的類神經網
路雖然可以處理非線性系統,但其效能仍受到前述第二項限制的影響.有鑑
於此,一些學者提出混合式類神經網路的模式架構,其原理乃結合類神經網
路及由守恆原則導出的數學模式,使得此一架構除了保有類神經網路對非
線性系統的處理,更因為結合了物理數學模式,讓混合式類神經網路能在更
短的時間內達到學習收斂目標,並對訓練集合外的狀態作出更正確的估測
值. 在本論文中,我們探討各種不同的反應系統,利用混合式類神經網路
進行系統的狀態估測.其中未知的程序參數交由類神經網路元件去鑑別,再
將鑑別結果輸入守恆數學模式,以計算系統下一個時刻的狀態變數.同時,
為了適應實際反應中未知程序參數可能不只一個的情形,我們將混合式類
神經網路擴充至包含兩個以上的類神經網路元件,每個神經網路元件模擬
一項未知的程序參數,使得混合式神經網路的適用範圍更為廣泛.另外,由
於混合式神經網路乃透過系統狀態對未知參數的靈敏度,間接引用模式輸
出值來學習,我們針對此項特點導出一組簡易的學習通式,俾使後來的研究
者能很容易的將此架構套用於大部份的反應系統.

The application of the least-squares method in parameter
estimation hasthe following shortcomings: (1) It cannot be
applied to nonlinear Systems;(2) It is influenced by corrected
input data and limited observations. Neural nets developed
recently can deal with nonlinear systems. However, their
performance is straitened by the second limitation mentioned
above.To overcome these, some researchers proposed the theory of
hybrid neural networks which combine neural nets with a
mathematical model derivedfrom the first principle. The hybrid
network inherits the capability ofneural nets in dealing with
nonlinear problems. Moreover, it converges inmuch shorter
training time and obtain more accurate estimates for a
widerrange of operrating conditions by virtue of the
mathematical model. In this thesis, we study the applications
of hybrid neural networks in state identification for various
reactions systems. The neural net unit isresponsible for
estimation of the process parameter of reaction with
unknownkinetics. The estimation is then applied to the first
principle unit to calculate the next states. For reactor systems
with more than one processparameter of reaction, more neural net
units can be added to simulate their kinetics. In addition,
since the output of the neural net unit is not measurable , the
training of a hybrid neural network must resquire the
sensitivity of statevariables towards unknown parameters. We
then derive a simple general form ofthe sensitivity function
such that the hybrid neural network can deal with awide variety
of reactor systems more easily.