锅炉设计外文翻译燃煤锅炉的个案事故研究.docx

锅炉设计外文翻译燃煤锅炉的个案事故研究.docx

《锅炉设计外文翻译燃煤锅炉的个案事故研究.docx》由会员分享，可在线阅读，更多相关《锅炉设计外文翻译燃煤锅炉的个案事故研究.docx（9页珍藏版）》请在冰豆网上搜索。

锅炉设计外文翻译燃煤锅炉的个案事故研究

附录A

英文原文

A.KusiakandA.Burns,MiningTemporalData:

ACoal-FiredBoilerCaseStudy,Proceedingsofthe9thInternationalConference,KES2005,Melbourne,Australia,September14-16,2005,inR.Khosla,R.J.Howlett,L.C.Jain（Eds）,Knowledge-BasedIntelligentInformationandEngineeringSystems:

Vol.III,LNAI3683,Springer,Heidelberg,Germany,2005,pp.953-958.

MiningTemporalData:

ACoal-FiredBoilerCaseStudy

AndrewKusiakandAlexBurns

IntelligentSystemsLaboratory,IndustrialEngineering

3131SeamansCenter,TheUniversityofIowa

IowaCity,IA52242–1527,USA

andrew-kusiak@uiowa.edu

Abstract

Thispaperpresentsanapproachtocontrolpluggageofacoal-firedboiler.Theproposedapproachinvolvesstatistics,datapartitioning,parameterreduction,anddatamining.Theproposedapproachwastestedona750MWcommercialcoal-firedboileraffectedwithafoulingproblemthatleadstoboilerpluggagethatcausesunscheduledshutdowns.Therare-eventdetectionapproachpresentedinthepaperidentifiedseveralcriticaltime-baseddatasegmentsthatareindicativeoftheashpluggage.

1Introduction

Theabilitytopredictandavoidrareeventsintimeseriesdataisachallengethatcouldbeaddressedbydataminingapproaches.Difficultiesarisefromthefactthatoftenasignificantvolumeofdatadescribesnormalconditionsandonlyasmallamountofdatamaybeavailableforrareevents.Thisproblemisfurtherexacerbatedbythefactthattraditionaldataminingdoesnotaccountforthetimedependencyofthetemporaldata.Theapproachpresentedinthispaperovercomestheseconcernsbydefiningtime

windows.

Theapproachpresentedinthispaperisbasedonthetwomainconcepts.Thefirstisthatthedecision-treedata-miningalgorithmcapturesthesubtleparameterrelationshipsthatcausetherareeventtooccur[1].Thesecondconceptisthatpartitioningthedatausingtimewindowsprovidestheabilitytocaptureanddescribesequencesofeventsthatmaycausetherarefailure.

2EventDetectionProcedure

Inthecasestudydiscussedinthenextsectionrareeventscanbedetectedbyapplyingthefivestepprocedure.Thesefivestepsinclude:

Step1:

ParameterCategorization

Theparameterlistisdividedintotwocategories,responseparametersandimpactparameters.Responseparametersarethosethatchangevaluesduetoarareeventorafailure,e.g.,anairleakinapressurizedchamber.Impactparametersaredefinedasparametersthatareeitherdirectlyorindirectlycontrollableandmaycausetherareevent.Thesearetheparametersthatareofgreatestinterestforthedeterminationofrareevents.

Step2:

TimeSegmentation

Timesegmentationdealswithpartitioningandlabelingthedataintotimewindows（TWs）.Atimewidowisdefinedasasetofobservationsinchronologicalorderthatdescribeaspecifiedamountofcontinuousobservations.Thisstepallowsthedataminingalgorithmstoaccountforthetemporalnatureofthedata.Themosteffectivemethodtosegmentthedataisbydetermining/estimatingtheapproximatedateoffailureandsetthatasthelastobservationofthefinaltimewindow.

Step3:

StatisticalandVisualAnalysis

Thisstepinvolvesstatisticalanalysisofthedataineachtimeperiodthatwasdesignatedinthepreviousstep.Processshifts,changesinvariation,andmeanshiftsinparametersarehelpfulinindicatingthattheappropriatetimewindowsandparameterswereselected.

Step4:

KnowledgeExtraction

DataminingalgorithmsdiscoverrelationshipsamongparametersandanoutcomeintheformofIF…THENrulesandotherconstructs（e.g.,decisiontables）[1],[5].Dataminingisnaturalextensionofmoretraditionaltoolssuchasneuralnetworks,multivariablealgorithms,ortraditionalstatistics.Inthedetectionofrareevents,thedecision-treeandrule-inductionalgorithmsareexploredfortwosignificantreasons.First,thealgorithmsgenerateexplicitknowledgeintheformunderstandablebyauser.Theuserisabletounderstandtheextractedknowledge,assessitsusefulness,andlearnnewandinterestingconcepts.Secondly,thedataminingalgorithmshavebeenshowntoproducehighlyaccurateknowledgeinmanydomains.

Step5:

AnalysisofKnowledgeandValidation

Thisstepdealswithvalidationoftheknowledgegeneratedbythedataminingalgorithm.Ifavalidationdatasetisavailableitshouldbeusedtovalidatetheaccuracyoftherules.Ifnosimilardataisavailablethenunuseddatafromtheanalysisora10-foldcross-validationcanbeutilized[6].

3PowerBoilerCaseStudy

Theapproachproposedinthisresearchwasappliedtopowerplantdata.Dataminingalgorithmsarewellsuitedforelectricpowerapplicationsthatproducehundredsofdatapointsatanytimeinstance.

Thiscasestudydealswithanashfoulingconditionthatcausesboilershutdownsseveraltimesayearonacommercial750MWtangentially-firedcoalboiler.Theashfoulingcausesabuildupofmaterialandpluggageinthereheatersectionoftheboiler.Oncethebuildupbecomessubstantialtheboilerperformanceisnegativelyaffected.Thisleadstothederatingandtheeventualshutdownoftheboiler.Thecleaningoftheboilerduringtheshutdownrequires1to3days.Thisproblemismademoredifficultbythefactthereisnomethodtodeterminethelevelofashbuildupwithoutshuttingdowntheboilertophysicallyinspectthearea.Furthermore,inanalysisallparameterswerewithinspecifications,sotherewasnoobvioussingleparameterthatiscausingthepluggage.Toinvestigatetheproblemconsideredinthispaper,datawascollectedon173differentboilerparameters.Thisincludedflows,pressures,temperatures,controls,demands,andsoon.Thedatawascollectedinone-minuteintervalsoverthecourseofthreemonths.Thedatacollectionbegandirectlyfollowingashutdownwherethereheatersectionoftheboilerhadnopluggage.Thecollectionperiodendedapproximatelythreemonthslaterwhentheboilerhadtobeshutdownforpluggageremoval.Thisdatasetcontainedover168,000observations.

Thelistof173parameters,whichincludedbothresponseandimpactparameters,wasanalyzed.Thelistwasreducedtoincludetwenty-siximpactparameters.Thisparametercategorizationandreductionwasaccomplishedwiththeassistanceofdomainexpertsaswellasstatisticalanalysissuchascorrelationandmultivariateanalysis.

Theinitialstepfortimesegmentingthedatawastodetermineanapproximatedateforthefailureevent.Inthisapplicationthefailureeventwasdefinedbythedatewhentheboilerwasderatedduetothepluggage.Thecauseoftheshutdownwasconfirmedthroughvisualinspectionoftheaffectedregion.Thisdatewasthensettobethelastdayofthefinaltimewindow（TW6）.

Thewindowsweresettobeapproximatelyoneweeklong.Aweekwaschosenforseveralreasons.First,theboilerwasinspectedapproximatelyonemonthpriortoitsderating.Duringtheinspectionthereheatersectionoftheboilerwascompletelyfreeofash.Thisinformationprovidedtheknowledgethatthepluggagerequiredlesstheonemonthtomanifestitselftothepointofshutdown.Itwashypothesizedthatthepluggagerequiresseveraldaystobuildup.Basedonthisinformationoneweekwasdeemedtobeanadequatetimewindow.Oneweekalsoprovidedasufficientnumberofobservations（over10,000perwindow）forthedataminingalgorithms.

Usingthederatedateandaone-week-longtimewindow,thedatawasdividedintosixtimewindowsshowninFigure1.Timewindow1（TW1）wasincludedtoensurethattherewasadequatedatatodescribenormaloperatingconditions.

Thereappearsbeaprocessshiftbetweentimewindows3and5inFigures1.Thewesttiltdemonstratesameanshiftduringwindowthreeandthehotreheatsteamtemperaturedisplaysameanshiftaswellasalargeincreaseinvariationstartingintimewindowfourandculminatinginwindowfive.Theresultsofthisanalysisleadtothehypothesisthattheeventsthatleadtotheeventualpluggageoccurbetweentimewindowsthreeandfive.Italsoconfirmstheselectionofparametersandwindowsize.

Thedataminingapproachwasthenappliedtothedatasettopredictthepredefinedtimewindows（decisionparameter）.Thealgorithmproducedasetofrulesthatdescribedtheparameterrelationshipsineachtimewindow.

Theknowledgeextractedbythealgorithmhadanoverall10-foldclassificationaccuracyof99.7%.Theconfusionmatrix（absoluteclassificationaccuracymatrix）isshowninFigure2.Thematrixdisplaystheactualvaluesandthevaluespredictedbytherulesduringthecross-validationprocess.

ItcanbeseenfromthedatainFigure2thattherearefewpredictedvaluesthatareoffbymorethanonetimewindowfromtheactualwindow.Theresultsprovidedintheconfusionmatrixprovideahighconfidenceintheproposedsolutionapproach.

Anothertestdatasetwasextractedfromtheweekfollowingtimewindow1andwaslabeledtimewindow2（TestTW2）.Thelastportionofthedata（TestTW3）wasobtainedfromtheweekafterthegeneratorwasderatedandtheoutcomewaslabeledtimewindow6（TW6）.Thetotaltestsetcontainedover30,000observations.

Therulesandknowledgethatwereextractedfromtheoriginaldatasetwerethentestedusingthetestdataset.Forpurposesofanalysistimewindows1–3wereconsiderednormalandtimewindow4–6wereconsideredfaulty.TheresultingconfusionmatrixisshowninFigure3.

Therulesaccuratelypredictedthenormalcases,buttheywerenotaseffectiveinpredictingthefaultcases.Thisismostlikelyexplainedbythefactthatthetestdatalabeled,timewindow6,wasextractedaftertheboilerhadbeenderated.Thederatingof