东南大学学报(英文版)
東南大學學報(英文版)
동남대학학보(영문판)
JOURNAL OF SOUTHEAST UNIVERSITY
2004年
1期
65-69
,共5页
黄凯%罗正鸿%陈丰秋%吕德伟
黃凱%囉正鴻%陳豐鞦%呂德偉
황개%라정홍%진봉추%려덕위
偏最小二乘法%催化剂%甲烷氧化偶联%神经网络%建模
偏最小二乘法%催化劑%甲烷氧化偶聯%神經網絡%建模
편최소이승법%최화제%갑완양화우련%신경망락%건모
partial least square%catalyst%oxidative coupling of methane%neural network%modeling
神经网络偏最小二乘法(NNPLS)被应用于一种甲烷氧化偶联多组分催化剂的鲁棒反应模型的建立. 重点研究了内层神经网络学习算法、激活函数、网络结构(包括隐含节点数、隐含层)、网络权值初始化及主元的选取原则等. 研究表明, 内层神经网络分别采用1-10-5-1,1-8-4-1,1-8-5-1, 1-7-4-1, 1-8-4-1, 1-8-6-1的拓扑结构是合适的; Levenberg-Marquardt方法被用于网络的学习算法可以加快学习速度; 同时采用了sigmoid函数为激活函数. 计算结果显示, 四主元可以满足建模的需要. 与单纯的神经网络催化剂模型相比, NNPLS方法压缩分解了变量, 减少了计算量, 同时使模型的推广能力得到提高, 有效地改善了直接神经网络建模过程中催化剂模型泛化能力较差的缺点.
神經網絡偏最小二乘法(NNPLS)被應用于一種甲烷氧化偶聯多組分催化劑的魯棒反應模型的建立. 重點研究瞭內層神經網絡學習算法、激活函數、網絡結構(包括隱含節點數、隱含層)、網絡權值初始化及主元的選取原則等. 研究錶明, 內層神經網絡分彆採用1-10-5-1,1-8-4-1,1-8-5-1, 1-7-4-1, 1-8-4-1, 1-8-6-1的拓撲結構是閤適的; Levenberg-Marquardt方法被用于網絡的學習算法可以加快學習速度; 同時採用瞭sigmoid函數為激活函數. 計算結果顯示, 四主元可以滿足建模的需要. 與單純的神經網絡催化劑模型相比, NNPLS方法壓縮分解瞭變量, 減少瞭計算量, 同時使模型的推廣能力得到提高, 有效地改善瞭直接神經網絡建模過程中催化劑模型汎化能力較差的缺點.
신경망락편최소이승법(NNPLS)피응용우일충갑완양화우련다조분최화제적로봉반응모형적건립. 중점연구료내층신경망락학습산법、격활함수、망락결구(포괄은함절점수、은함층)、망락권치초시화급주원적선취원칙등. 연구표명, 내층신경망락분별채용1-10-5-1,1-8-4-1,1-8-5-1, 1-7-4-1, 1-8-4-1, 1-8-6-1적탁복결구시합괄적; Levenberg-Marquardt방법피용우망락적학습산법가이가쾌학습속도; 동시채용료sigmoid함수위격활함수. 계산결과현시, 사주원가이만족건모적수요. 여단순적신경망락최화제모형상비, NNPLS방법압축분해료변량, 감소료계산량, 동시사모형적추엄능력득도제고, 유효지개선료직접신경망락건모과정중최화제모형범화능력교차적결점.
In this paper neural network partial least square (NNPLS) was used to establish a robust reaction model for a multi-component catalyst of methane oxidative coupling. The details, including the learning algorithm, the number of hidden units of the inner network, activation function, initialization of the network weights and the principal components, are discussed. The results show that the structural organizations of inner neural network are 1-10-5-1, 1-8-4-1, 1-8-5-1, 1-7-4-1, 1-8-4-1, 1-8-6-1, respectively. The Levenberg-Marquardt method was used in the learning algorithm, and the central sigmoidal function is the activation function. Calculation results show that four principal components are convenient in the use of the multi-component catalyst modeling of methane oxidative coupling. Therefore a robust reaction model expressed by NNPLS succeeds in correlating the relations between elements in catalyst and catalytic reaction results. Compared with the direct network modeling, NNPLS model can be adjusted by experimental data conveniently and the calculation of the model is simpler and faster than that of the direct network model.