软测量技术若干问题的研究及工业应用

《软测量技术若干问题的研究及工业应用》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、浙江大学博士学位论文软测量技术若干问题的研究及工业应用姓名：刘瑞兰申请学位级别：博士专业：控制理论与控制工程指导教师：褚健;苏宏业20040401浙江大学博士学位论文明，该方法在线建立的软测量模型精度高，很适合慢时变对象、且训练样本分布不均匀情况下的软测量建模。5．主导变量与过程变量之间的时序匹配是软测量技术不可缺少的组成部分。时序匹配实际上是确定主导变量相对于每个过程变量的滞后时间。提出了两种无需人为干涉而是直接使用现场采集到的数据来确定滞后时间的方法：最大相关系数法和模糊曲线法。最大相关系数法一般应用于线性和弱非线性对象；模糊曲线法既可以用于线性对象也可以应用于非线性对象。仿真和工业应用表

2、明这两种方法都是适用的。II浙江大学博士学位论文ABSTRACTSoftsensortechnologyjsoneofthemostimportantresearchdirectionsintheareaofprocesscontr01．Inthisdissertation，severalissuesandthecorrespondingsolutionsaboutsoftsensortechnologyarediscussedbasedontherealindustrialprocessandthemaincontributionsaredescribedasfollows．1．Asofts

3、ensormodelingalgorithmbasedonimprovedfuzzyneuralnetworkispresented．Thenormalizedaverageoutputmembershipfunctionsaredefinedasfuzzybasisfunctionsfurdefuzzificationcalculation．Inordertoimprovethepropertyofconvergence，someparametersofthefuzzyneuralnetworkaretrainedbyLevenberg—Marquardtalgorithm，andtheot

4、hersaretrainedbygradientdescentalgorithm．Finally,asoftsensormodelofmeltindexinpolymerreactionbasedontheproposedmethodisestablished，andthesimulationresultsshowthatincontrasttothetraditionalfuzzyneuralnetworktheproposedmethodisnotsensitivetoinitialparametersandpossessesgoodconvergencecapabilityandpred

5、ictionprecision．2AnewhybridlearningalgorithmisproposedtotrainthefuzzyneuralnetworkbasedonTSKfuzzymodel．Firstly,fuzzyc-meansalgorithmisappliedtoinitializetheparametersofthefuzzyneuralnetwork．Secondly,theparametersofthepremisepartofthefuzzyrulearelearnedbythegradientdescentalgorithm．Finally,theparamet

6、ersofconsequentpartarelearnedbythepartialleastsquaresalgorithm．TheproposedhybridmethodCanautomaticallygiveappropriateinitialparametersofthefuzzyneuralnetworkandpreventthefuzzyrulenumberfromincreasingforhigh—dimensionalsystems．Theresultsofsimulationandindustrialapplicationshowthatthehybridlearningalg

7、orithmhaspropertiesoffastconvergenceandhighaccuracy．3Asoftsensormodelingmethodbasedonhybridmodelcombiningthesimplifiedfirstprinciplemodelanddata-drivenmodelisproposed．Severalsimplifiedfirstprinciplemo