3.Processstatemonitoring



Similar documents

South East of Process Main Building / 1F. North East of Process Main Building / 1F. At 14:05 April 16, Sample not collected

Monitoring of Complex Industrial Processes based on Self-Organizing Maps and Watershed Transformations

RS-232 COMMUNICATIONS

The Parts of a Flower

Visualization of large data sets using MDS combined with LVQ.

Visualization of Breast Cancer Data by SOM Component Planes

Data Mining for Customer Service Support. Senioritis Seminar Presentation Megan Boice Jay Carter Nick Linke KC Tobin

Enrollment Data Undergraduate Programs by Race/ethnicity and Gender (Fall 2008) Summary Data Undergraduate Programs by Race/ethnicity

THE KENYA UNIVERSITIES AND COLLEGES CENTRAL PLACEMENT SERVICE KUCCPS. Procedure for Online Application for Placement to

Factor Models for Gender Prediction Based on E-commerce Data

Why can't I make or receive telephone calls (cordless phones)?

LONDON BOROUGH OF LAMBETH APPLICATION FOR DISCRETIONARY RATE RELIEF

Why can't I make or receive telephone calls (cordless phones)?

2014 Only Influencers Marketing Salary Guide

Violence against women: key statistics

Excel Charts & Graphs

Nottinghamshire County Council. Customer Service Standards

System Behavior Analysis by Machine Learning

Applying Data Analysis to Big Data Benchmarks. Jazmine Olinger

Best Practice in SAS programs validation. A Case Study

TURKISH REPUBLIC KARABÜK UNIVERSITY

Art and Design Teacher Education- Application Pack

FÉDÉRATION INTERNATIONALE DE GYMNASTIQUE. Artistic Gymnastics

GAZE TRACKING METHOD IN MARINE EDUCATION FOR SATISFACTION ANALYSIS

Computerized Micro Jet Engine Test Facility

AudioJoG (TM) Pro 8 Connector PIN LABEL LED Connector PIN LABEL LED. Operations Manual 3.5mm & 6.35mm Mono/Stereo Jacks

37 JIC Flare (Nuts) 304C (070112) 306 (070118) 318 (070110) Pg Bulkhead Locknut (Use with #2700) Pg. 23 Tube Nut (Use with #319)

Location: Clovis North High School 2770 E International Ave Fresno, CA 93730

4 on 4 Intramural Volleyball Rules

VIDEO SCRIPT: Data Management

Counting the Ways to Count in SAS. Imelda C. Go, South Carolina Department of Education, Columbia, SC

Using NeuralTools to Move Yellow Pages Dollars to Digital Solutions

Lecture 1: Review and Exploratory Data Analysis (EDA)

Chapter 4 Displaying and Describing Categorical Data

MCT-7 Multiple Cable Test System

APPLICATION FORM FOREIGN STUDENTS (PLEASE PRINT)

Segmentation of stock trading customers according to potential value

Chapter 2 Introduction to SPSS

Methodology for Emulating Self Organizing Maps for Visualization of Large Datasets

Automatic Detection of PCB Defects

BLUE RIBBON CORP. BC001 Birdcage Installation Manual

FULL-TIME APPLICATION FORM

Guido s Guide to PROC FREQ A Tutorial for Beginners Using the SAS System Joseph J. Guido, University of Rochester Medical Center, Rochester, NY

Finding Supporters. Political Predictive Analytics Using Logistic Regression. Multivariate Solutions

The UCC-21 cognitive skills that are listed above will be met via the following objectives.

Chapter 5 Analysis of variance SPSS Analysis of variance

CONNECTING PHONES, FAXS & DEVICES TO TALKSWITCH

Release: 1. ICPPRN493 Set up and monitor in-line printing operations

BA (Hons) Fashion Design

22 Annual Report Vaibhav Gems Limited

SECTION TESTING OF FIBER OPTIC CABLES

Financial Responsibility. Costs of Owning a Vehicle Trip Planning

Texas A&M University at Qatar Electrical Engineering

Malawi Data profile 2012

Independent t- Test (Comparing Two Means)

4. Do you have a Round Rock Public Library borrower s card? response: percent count Yes 82.4% 324 No 15.8% 62 Do not know 1.8% 7. page A2.4.

Grenada. Enterprise Survey Country Bulletin. The average firm in Grenada

TEACHERS SERVICE COMMISSION TEACHERS PERFORMANCE APPRAISAL REPORT

5 Point Choice ( 五 分 選 擇 題 ): Allow a single rating of between 1 and 5 for the question at hand. Date ( 日 期 ): Enter a date Eg: What is your birthdate

Validation of a Computerized Color Vision Test

Tech Bulletin. Hose / Cord Replacement and Required Wraps. Series RT

Computer-System Architecture

20th. Annual Bridge. Building Contest. Open to Students from Grade March is...

Name: Date: Use the following to answer questions 2-3:

MBA Data Mining & Knowledge Discovery

APPLICATION FORM FOR EXCHANGE STUDENTS

Network Intrusion Detection Systems

Minimally Invasive Mitral Valve Surgery

Diaphragm Seal with flush diaphragm and flange acc.to SMS. Completed with pressure gauge in stainless steel, liquid filled or pressure transmitter

Figure 1.1 Percentage of persons without health insurance coverage: all ages, United States,

Advanced visualization with VisNow platform Case study #3 Vector data visualization

Answer: Quantity A is greater. Quantity A: Quantity B:

A quick overview of geographic information systems (GIS) Uwe Deichmann, DECRG

MOTION COORDINATOR MC206X Quick Connection Guide

LVQ Plug-In Algorithm for SQL Server

Creating a Custom Class in Xcode

ICRS implantation with the Femto LDV laser in stabilized KC patients: 6 months results

GLENSOUND ELECTRONICS LTD

Self-Organizing g Maps (SOM) COMP61021 Modelling and Visualization of High Dimensional Data

SPSS Resources. 1. See website (readings) for SPSS tutorial & Stats handout

RapidIO Network Management and Diagnostics

AMERICAN NATIONAL STANDARD

Campus of Performing Arts (PTY) Ltd.

CAN/ULC-S Installation of Fire Alarm Systems Amendment 1. Canadian Fire Alarm Association (CFAA) D. GOODYEAR FIRE CONSULTING

CYB Credit for youth in business (Under the auspices of National Youth Council of Namibia)

AHA Instructor Renewal

First Nation Membership Database. Sample Screens

Transcription:

Chapter14 Processmonitoringandvisualisation O.SimulaandJ.Kangas HelsinkiUniversityofTechnology,LaboratoryofComputerandInformation usingself-organizingmaps Science,Rakentajanaukio2C,02150Espoo,Finland,Fax:358(0)4513277, E{mail:Olli.Simula@hut.,Jari.Kangas@hut. 1.Overview algorithmcanbeusedtoinvestigatecomplexdependenciesbetweenvariousprocess possibletopredictallpossibleerrortypesinadvance.theself-organizingmap problemarea.incomplicatedsystems,asinchemicalprocesses,itisnot systemstomonitorcomplicated,dynamicalprocessesandtovisualisetheprocess development. parametersaswellasinputandoutputvariables.itcanalsobeutilisedtocreate Analysisandcontrolofcomplexnonlinearprocessesconstitutesadicult algorithmisthetopologicalnatureofthemapping;similarsignalpatternsare inputpatternstomapunits.thekeypointintheapplicabilityofthesom sequenceanalysisbyndingthemappinglocationsofsubsequentpatternsand 2.Introduction mappedtonearbylocationsonthemap.suchamappingcanbeappliedtopattern observingthetrajectory,thatisthecurveofthelocationsintime.thepattern applications. twodimensionaltrajectorywhichcanbeused,forexample,inprocessmonitoring sequenceinmulti-dimensionalinputspacecaninthatwaybetransformedintoa Theself-organizingmap(SOM)algorithm[8][9]createsamappingfrom monitoringisdescribed.someapplicationsinvisualisationoftheprocessoperation Inthefollowing,theuseoftheself-organizingmapinprocessstate

analysistaskstoprovidethevisualisationofcertainsignalcharacteristics. 372 arepresented.inadditiontovisualmonitoring,thesomalgorithmcanbeapplied infaultdetectionandanalysis.examplesoffaultdiagnosisofdevicesandprocesses ApplicationswheretheSOMalgorithmhasbeenusedinvisualisationofspeech signalsandspeechsignalvariationsintimearealsodescribed.finally,other applicationsofthesomalgorithminmonitoringandsignalanalysisareshortly covered.moredetailedreviewoftheuseandapplicationofthesomalgorithm willbefoundin[6]. Theself-organizingmapalgorithmcanalsobeappliedtoothersignal 3.Processstatemonitoring Inpracticalapplications,alsotheonlineprocesscontrolbasedonthestateanalysis anecientandunderstandableway.theonlineprocessstatusmeasurements visualisethetypicallycomplexrelationsbetweenvarioussystemparametersin shouldbeconvertedtosomesimpledisplaysthat,despitethedimensionality processmonitoringisoftencarriedoutinordertoestimatethefuturebehaviour. Reliablepredictionofthefuturebehaviourcouldresultinecientfaultdiagnosis. reduction,preservetherelationshipsbetweenstates.simultaneously,itshould bepossiblefortheusertofollowtheprocessstatedevelopment.furthermore, Inprocessstatemonitoringapplications,theproblemistoanalyseand spacetotwodimensionalsurfaceoftheprocessingunits.themappingis, furthermore,doneinsuchawaythatthetopologicalrelationsbetweentheinput isanimportantaspect. featurevectorsarepreserved.inmonitoringapplications,theself-organizingmap processparameters. algorithmcaneasilybeusedtoanalysethecomplexrelationsbetweenthevarious Theself-organizingmapmethodmakesamappingfromamultidimensional preservingpropertyofthemap,similarfeaturescorrespondingtosimilarstatesof theprocessaremappedclosetoeachotherresultinginclustersonthemap. variablesordata,(2)theprocessparameters,and(3)outputsoftheprocess.all theseparameterscanbeconcatenatedtoformafeaturevectorwhichisusedas aninputtotheself-organizingmap,asshowninfigure1.duetothetopology threetypesofvariablesorprocessparametersthatmustbeconsidered:(1)input Themappingiscreatedinanunsupervisedwayfromthemeasureddata Incomplexprocesses,e.g.inchemicalengineering,thereareusually theprocess.thisisdepictedinthelowerpartoffigure1.onlyalimitednumber eachclusteronthelabeledmapcorrespondstoacertainstateoroperationpointof thislearningphase.thephysicalinterpretationofthemapcanbeobtainedby labelingthenodesofthemapaccordingtotheknownprocessbehaviour.the nodescorrespondingtosimilarfeaturesaremergedinthelabelingprocess.thus, andparameters,i.e.noknowledgeoftheprocessbehaviourisrequiredduring

Measurement vector (Feature vector) 373 Input measurements Output measurements Process Material Material flow in flow out ofpreclassiedsamplesarerequiredinthelabelingphase.usingthelabeledmap Process parameters inmonitoring,wecannowidentifythestateoftheprocesscorrespondingtoa certainoperationpoint. Figure1.Featurevectorobtainedfromprocessparametersanddata (above).physicalinterpretationoftheself-organizingmapbylabeling A C 3.1.Visualization theclustersofsimilarfeaturestocorrespondingprocessstates(below). E B Map training D and labeling behaviour[7].theparameterofinterestcanbeextractedfromthefeaturevector Theself-organizingmapisanecienttoolforvisualisingtheprocess Self-Organizing Map infigure2. bydisplayingitsvalueasagraylevelonthemap.followingthetrajectoryofthe operationpoint,wecaneasilymonitortheparametervalue.anexampleisshown oneofthetemperatures.darkgraytonescorrespondtolowandlighttonestohigh analysedduringaperiodof24hours.thetrajectoryisdrawnonthemapdisplaying Thetrajectoriesofthesuccessivedayswereverysimilarduringnormaloperation. temperatures,respectively.itcanbeseenthattheoperationpointhasmovedfrom darktolightandbackagaincorrespondingtothedaytimeoperationofthesystem. Inthisexample,measurementsoftheprocessparametershavebeen

parameterdependenciesandtheireectonprocessbehaviourcaneasilybe 374Simulateddataandprocessparameterscanalsobeusedinthelearning investigated.forinstance,theeectofcertainprocessparametersonprocess phase.theself-organizingmapcan,thus,beusedinprocessmodeling.various stateoroutputcanbevisualisedonthemap. InFigure3,eightdierentparametersofinteresthavebeenextractedfromthe featurevector.theparametervaluesandtheirdependenciescanbeanalyzed parametervaluescorrespondingtothestateoftheprocess. directlybycomparingthegraylevelsofcorrespondingmapunits.thetrajectory oftheoperatingpointcanalsobedisplayedoneachseparatemapgivingthusthe Figure2.Exampletrajectoryofaprocessstateduring24hours. Severalparameterscanbevisualisedsimultaneouslybyusingasetofmaps. vectoritispossible,forinstance,tomakeoverallcomparisonsofdierentprocesses. areoftencalculated.usingtheseguresofmeritasparametersinthefeature Theeectofvariousparameterstotheprocessbehaviourcanbeanalyzedandto someextenteven\optimised".asanexample,thesomalgorithmiscurrently beingappliedintheanalysisofapulpprocess,whereespeciallytheairandwater emissionsoftheprocessareofinterest. Inevaluatingthequalityofcomplexprocesses,variouscharacteristicgures 3.2.Faultdiagnosis toidentifythefault.inpracticalapplications,wecandistinguishbetweentwo diagnosis.themapcanbeusedintwoways:(1)todetectthefaultand(2) Anotherimportantapplicationoftheself-organizingmapisinfault

375 (a) (b) (c) (d) dierentsituations;eitherwehavenopriormeasurementsofthefaultysituations Figure3.Asetofmapscorrespondingtodierentparameters. Therefore,onlythesituationsincludedintothetrainingdatacanberecognisedby thelabeledmap. orwehavebeenabletorecordalsofaults. situations,theoperationspaceonthemapcoversonlynormalsituations.fault processthatarecoveredbythemeasurements.thus,thestatespacewillbedivided intotwoparts:(1)thepossibleoperationspaceand(2)itscomplementaryspace. Incasethetrainingdatahascontainednomeasurementsfromfaulty Inthelearningphase,themapistrainedtorecogniseonlythosestatesofthe (e) (f) (g) (h) faultysituation.thisconclusionisbasedontheassumptionthatduetothelarge distanceofthefeaturevectorfromallthemapnodestheoperationpointmust mapunits.ifthedierenceexceedsapredeterminedthresholdtheprocessisina correspondingtothemeasurementiscomparedtotheweightvectorsofallthe belongtothecomplementaryspace,notcoveredbythetrainingdata.therefore, detectioncannowbebasedonthesocalledquantisationerror.thefeaturevector thesituationhasnotoccurredbeforeandsomethingispossiblygoingwrong. situations.inthiscase,clusterscorrespondingtocertainfaultsarecreatedon themapandtheseclusterscanbeconsideredas\forbidden"areas.thefaultcan nowbeeasilyidentiedbyfollowingthetrajectoryoftheoperatingpoint.ifthe mapincludingoneforbiddenareaisshowninfigure4.inthisparticularexample, trajectorymovestoaforbiddenareathefaultwillbeidentied.anexampleofthe Infaultidentication,thetrainingdatamustcontainsamplesoffaulty

376 thetrajectorypassestheforbiddenareaandthefaultcorrespondingtotoohigh temperatureincertainpartoftheprocessisidentied. correspondingtovariousfaultscanbecreatedonthemap.thisispossibleif canbeproduced.thisisespeciallyimportantinsituationswherefaultsarerare faultysituationsandtheirreasonsareknownwellenoughsothatsimulateddata andtruemeasurementsarethusnotavailable. Figure4.Forbiddenareaonthemapcorrespondingtoafaultysituation. 3.3.Faultdetectionandidenticationsystem Infaultanalysis,simulateddatacanalsobeused.Forbiddenareas anaesthesiasystemisdescribed.theanaesthesiasystemcomprisestheanaesthesia machine,thepatient,andtheanaesthesiapersonnel.thepurposeofmonitoringis tominimisetherisksofanaestheisiabydetectingandidentifyingthefaultsbefore theycauseinjurytothepatient[22],[23]. Inthissection,anexampleofthefaultdetectionandidenticationinan reasonwillbeidentied.thefaultdetectionandidenticationsystemisdepicted inanalysingfaultsoralarms,asshowninthelowerpartoffigure5.ontherst usingthesocalledfault-detectionmap.onlyafterthedetectionofthefault,its indetailinfigure5. notpossibletoexactlydene\normal"situations,twolevelsofmapscanbeused thefaultisdetectedbasedontheincreaseinthequantizationerror.thisisdone Thedetectionandidenticationoffaultscanbedoneintwostages.First, Indynamicalsystemswheretheoperationpointisnotstableorwhereitis

measurements (feature vector) fault detection map visualization of quantization error 377 Figure5.Faultdetection(upperpart)andidentication(lowerpart)system. fault registration of ring buffer operation point visualization of operationspaceintodierentparts.theamountofdeviationcausedbyafaulty beusedtolocatefaultsbasedonthedeviationsintheparameterscomparedtothe operatingpoint. trajectory situationmaystronglydependontheoperationpointofthesystem.thus,each measurements.onthesecondlevel,amoredetailedmap(orasetofmaps)can level,theself-organizingmapisusedtoidentifytheoperatingpointbasedonthe 1st oftherstlevelmap.bystoringasequenceoffeaturevectors,thebehaviourof theprocessbeforeandduringtheoccurrenceofthefaultcanbeanalysedinmore maponthesecondlevelcorrespondstoarelativelysmallpartoftheoperationspace Thepurposeoftherstlevelmapistoincreaseaccuracybydividingthe 2nd maps detail.themonitoringsystemwasimplementedbycollectingtrainingdataina obstructionsinthetubesindierentpartsoftheanaestehesiasystem.inthe realoperatingroomenvironment.dierentfaultconditionsincludedleaksand experiments,therecognitionaccuracyofthesecondlevelmapwas87%onthe oftime.figure6(a)showsthequantizationerrorofthefaultdetectionmap.in samplesfromtruesituationsaswell.inthisexample,thepositionofthepatient considerablefromthoseofthetrainingset,therecognitionaccuracydecreasedto 70%ontheaverage. pointsofthetrainingset.whenthelocationoftheoperationpointdeviated waschangedandtheintubationtubewasaccidentallyobstructedforashortperiod average,iftheoperationpointofthetestsetwasrelativelyclosetotheoperating Figure6(b)itcanbeseenthatthetrajectoryoftheoperationpointmovesfrom Theperformanceofthefaultdetectionandidenticationwastestedusing

378 theareacorrespondingtonormalsituation(n)toobstructionarea(o5). Figure6.Quantizationerrorofthefaultdetectionmap(a)andthetrajectoryof mapalgorithminanalysisoftimeseriesdata.wehaveusedthespeechsignalasthe 4.Visualizationofspeechsignals Thefollowingexampledemonstratestheapplicationoftheself-organizing theoperationpointduringthefaultysituation(b). inputdata,butsimilardynamicphenomenacanbedetectedinmanyprocessesand theanalysismethodsarerathergeneric.studingthesimilaritiesanddierences alsotakesintoaccounttheprobabilitydensityfunctionoftheinputsamplesso preservesthesimilarityoftheacousticsignals(toacertaindegree).themapping soundsaremappedtonearbylocationsonthemap(plane).themappingthus ofanydynamicprocess. betweenspeechsamplesisanalogoustocorrespondingstudiesforoperatingstates thatthedistributionofinputstothemapunitsapproximatesthedensityfunction amappingfromthespeechsignalspacetoatwodimensionalplane.similarspeech directly.perceptuallymeaningfulfeaturesofspeechderivefromcomplexfeatures Inspeechanalysistheself-organizingmapalgorithmhasbeenusedtocreate thereforeeasilycomprehendablewhencomparedwith,forexample,spectrograms. Becausetheprobabilitydensityfunctionoftheinputsamplesisalsotakeninto locationsonatwodimensionalmap.visualizationofvoicequalitywithsomis ofmorecommonsounds.itispossibletodirectlyobservethesimilaritiesofthe inthespeechspectra.suchfeaturesarerepresentedbyself-organizingmapas account,thosesoundsthataremorecommonarerepresentedwithmoredetails thanlessfrequentlyoccurringsounds.thusmoreemphasisisputonthedeviations 400.0 Quantization error Intubation tube obstructed (O5) 0.0 t (a) (b)

soundsbyobservingthedistancebetweentherespectiverepresentationsonthe map. 379 from15men.themapareasforthevowels,/s/and/r/areshownon Figure7.Aself-organizedmapcomputedwithFinnishspeechsamples somedysphonicpatientsaredepictedbythetrajectorycurve.someexamplesof timespectraformatrajectoryonthemapandchangesintimecanbeobserved thechangingofspeechspectraintime.therepresentationsofconsecutiveshort- fromtherepresentations.theabruptchangesinthespeechsignalobservedin bythelinebeginningfrom/s/areaandendingin/i/area. themap.thetrajectoryproducedbyanutteranceof/sa:ri/isindicated speechtrajectoriesonself-organizedmapsareshowninfigures7and8. Theself-organizingmapmethodisespeciallyusefulforthevisualisationof Figure8.Aself-organizedmapcomputedwithFinnishspeechsamplesof18 ThemapsinFigures7and8wereusedtoanalyzethevoiceduringlong women. u i e u o o y i e a r ö x a y x ö ä r ä male map female map s s

ofthesamples,thefollowingparameterswerecalculated. computingdiagnosticguresfromthetrajectory.fortheevaluationofthestability 380 /a:/vowels.wemeasuredthesmoothnessandregularityofthespeechsoundby 2.themeanlengthofshifts,and 3.thenumberofcaseswhereconsecutivesampleshavethesamerepresentations 1.Thetotallengthofthetrajectoryduringanintervalof150ms, issmoothwithoutanyabrupt`jumps'initbutitislocatedona`wrong'areaon variations)canalsobeobservedfromthetrajectory.inthesecases,thetrajectory andthenumberofconsecutivesampleshavingthesamerepresentationsgivesan indicationofasmallscalestability. Thelengthoftrajectoryindicateslargescalechangesinconsecutivesamples, onthemap. themap.whennormalspeakersutteraknownphoneme,weexpectthesamplesto beprojectedintoaspecicarea.normalinterspeakerdierencesareseenassmall dierencesinlocationswithinthisarea.bydeterminingasampletrajectoryin referencetothis`normalarea',wecanthusdiagnosethe`normality'ofthespeech sound. Thepermanentchangesinthespeechspectra(asopposedtocycle-to-cycle 4.1.Experiments experiments. wasdescribed.inthispaper,theself-organizingmapwasusedtovisualisethe speechsignalandthespeechsignalvariationsintime,andfromthemapping createdbythesomprocessitwasfurtherpossibletoconstructsomequantitative Theaboveideasonspeechsignalvisualisationhavebeentriedinseveral speechsignalanalysisisdescribed.inthispapertheself-organizedmap analyzed. measuresofthevoicequality.inthestudiesthevoicequalityofvowelswas In[11]anapproachtodevelopanautomaticdeviceforvoicequalityanalysis wasusedtodistinguishbetweenthe/s/samplesperceptuallyclassiedas themapwasasuitabletoolfortheextractionandmeasurementofacousticfeatures acceptable/unacceptable,asjudgedby21speechpathologists.itwasshownthat thevowelfollowingaword-initial/s/clearlyaectsthespectratowardstheend correspondingtotheaudibledeviationsof/s/. In[15]anotherapplicationoftheself-organizingmapalgorithmfor changesinspectraduetothepreparatorymovementsoflipsandtongueforthe ofthe/s/sound.thusthefollowingvoweltypecouldberecognisedbasedonthe followingvowel. Fromacousticvoiceanalysisusingspectrogramsitcanbeobservedthat

organizingmapalgorithmcouldextractuseful,perceptuallymeaningfulfeaturesin the/s/sound.theobjectivewastoseeiftherepresentationsofthe/s/samples beforedierentvowelscouldconsistentlyberecognisedfromthetrajectories producedbyprojectingtheconsecutivespeechsamplesintoamapplane. In[12]and[13]thephenomenonwasfurtherstudiedtondoutiftheself- 381 5.OtherApplications processmonitoringwasdescribed.self-organizingmapwasusedtodetectabnormal statesinareal-timeprocessbyexaminingthequantisationerrorbetweenthebest thebenetsofthealgorithmasappliedtomonitoringapplicationswerepointed usingneuralnetworksinfaultdiagnosisofachemicalprocesswasexplained. out.theexampleapplicationwasadistillationprocess.in[17]anotherstudyof andcontrolwasexplained.thesommethodwasexplainedinsucientdetailand In[1]apreliminarystudyofthepotentialoftheSOMalgorithmfor In[18]theapplicabilityoftheSOMalgorithmtoprocessstatemonitoring thatwereencounteredduringthetraining.risingquantisationerrorobviously determinedusingaknownsetoferrorsamples.asimilarsystemwasdescribedin indicatedsomeprocessstatethathadseldomoccurredduringthetrainingperiod. Asimilarprincipleoferrordetectionwasusedalsoin[7]wheresomeerrorstates wereclassiedbyobservingtheprojectionofsamplesto`forbidden'areasintomap, matchingunitonthemapandtheinputvector.thedetectionwasbasedonthefact [4],wherethetrainingalgorithmswereimproved. thatthemapunitsweredistributedinthespaceoccupiedbythoseinputsamples powerplant,whereitisimportanttonoticepreviouslyunknownsituations.the ofanenginecondition.theresultswerepromising.theauthorsfoundthesom quantisationerrorgivenbythesomalgorithmisusedtogivesomeindicationof tobeespeciallyusefulinthevisualisationofdatapropertiesandhighlightingthe thenoveltyoftheinputsample. autoassociativeback-propagationmodel)wereappliedtoamonitoringapplication In[2]asimilarsystemasabovewasusedtodetecterrorsinanuclear deviantdatavalues.in[21]au-matrixmethodforthevisualisationoftheprocess propertiesthroughthemapwasexplained(theu-matrixmethodwasintroduced in[19]and[20]). In[3]twoneuralnetworkmodels(self-organizingmapandakindof anexampleonecantakethepaper[16],whereself-organizingmapwasusedfor fordierentkindsoffailuredetection.in[5]severalstudieswerereviewed.as fromagroupofspanishbanks.itwasshownhowdierentregionsonthemap ofsuccessiveyears. representedsolventandbankruptbanks,respectively.itwasalsopossibletofollow thetimeevolutionofthebanksfromthetrajectorycreatedbymappingthedata In[14]theself-organizingmapwasusedtoanalyzesomenancialdata Theself-organizingmaphasbeenusedtomonitorelectricpowersystems

382 powersystemstaticsecurityassessment.itwasshownthatthemapcanbeused collectedfromtestcarduringtestruns.thesystemconsistedoftwolayersofsoms wheretherstlayerhandledstaticmeasurementsandthesecondlayercollected inmonitoringofpowersystems. In[24]theself-organizingmapwasusedtovisualisemeasurementsignals REFERENCES datafromlongerdurationsoftime.thevisualisationofthemeasurementswas donebycolourcodingthemapunits;similardrivingstateshavethereforesimilar qualityofpaperfromprocessmeasurements.aseparateself-organizingmapwas colourcodes.thedrivingstatesduringthetestrunwasthenillustratedasacolour codedtrajectoryonamapoftestlane. usedtomonitorthemovementoftheoperatingpointoftheprocessandtogivea 6.Discussion hintoftheestimationerroroftheprimarynetwork. In[10]anerrorback-propagationmodelwasusedtoestimatethenal systems.duetothetopologypreservingpropertyofthesomalgorithmithas monitoringcomplexprocesses.thenonlinearmappingfromahighdimensional showntobeanextremelypowerfulvisualisationtool. inputspacetoausuallytwodimensionalgridecientlycharacterisescomplex Theself-organizingmaphasbeenappliedtovariousapplicationsin insimulatingandestimatingthebehaviouroftheprocess.variousparameterscan ofanysystemwheretruemeasurementsofprocessparametersorsimulateddata investigatedinastraightforwardmanner.themethodisapplicabletotheanalysis areavailable.forinstance,intheanalysisofcomplexchemicalprocessesvarious isnotasignicantdrawbackwithtoday'scomputers. dependenciescaneasilybeexamined.thehighdimensionalityofthefeaturevector Inadditiontomonitoringandvisualisation,theSOMalgorithmcanbeused UsingSOMthedependenciesofthesystemparametersandvariablescanbe beoptimisedbyfollowingtheprocessbehaviouronthemap.eventhecontrolof References complexsystemsmaybepossiblebyusingfeedbackfromthemonitoringsystem. [1]J.T.Alander,M.Frisk,L.Holmstrom,A.Hamalainen,andJ.Tuominen, [2]Y.Bartal,J.Lin,andR.E.Uhrig,Nuclearpowerplanttransientdiagnostics usinglvqorsomenetworksdon'tknowthattheydon'tknow,inproc. ICNN'94,Int.Conf.onNeuralNetworks,pages3744{3749,IEEE,Piscataway, Processerrordetectionusingself-organizingfeaturemaps,InT.Kohonen, K.Makisara,O.Simula,andJ.Kangas,editors,ArticialNeuralNetworks, USA,1994. pagesii{1229{1232,north-holland,1991.

REFERENCES [3]S.Cumming,Neuralnetworksformonitoringofengineconditiondata,Neural [4]F.Firenze,L.Ricciardiello,andS.Pagliano,Self-organizingnetworks:A andp.g.morasso,editors,proc.icann'94,int.conf.onarticialneural challengingapproachtofaultdiagnosisofindustrialprocesses,inm.marinaro Computing&Applications,1(1):96{102,1993. 383 [5]R.Fischl,Applicationofneuralnetworkstopowersystemsecurity: [6]J.Kangas,Ontheanalysisofpatternsequencesbyself-organizingmaps,PhD [7]M.Kasslin,J.Kangas,andO.Simula,Processstatemonitoringusingselforganizingmaps,InI.AleksanderandJ.Taylor,editors,ArticialNeural Networks,2,pagesII{1531{1534.North-Holland,1992. Networks,pagesII{1239{1242,Springer-Verlag,1994. Technologyandtrends,InProc.ICNN'94,Int.Conf.onNeuralNetworks, [8]T.Kohonen,Self-organizingformationoftopologicallycorrectfeaturemaps, pages3719{3723,ieee,piscataway,usa,1994. thesis,helsinkiuniversityoftechnology,espoo,finland,1994. [11]L.Leinonen,J.Kangas,K.Torkkola,andA.Juvas,Dysphoniadetectedby [10]J.LampinenandO.Taipale,Optimizationandsimulationofqualityproperties [9]T.Kohonen,Theself-organizingmap,Proc.ofIEEE,78:1464{1480,1990. Networks,pages3812{3815,IEEE,Piscataway,USA,1994. inpapermachinewithneuralnetworks,inproc.icnn'94,int.conf.onneural patternrecognitionofspectralcomposition,journalofspeechandhearing BiologicalCybernetics,43(1):59{69,1982. [13]L.Leinonen,R.Mujunen,J.Kangas,andK.Torkkola,Acousticpattern [12]L.Leinonen,T.Hiltunen,K.Torkkola,andJ.Kangas,Self-organizedacoustic recognitionoffricative-vowelcoarticulationbytheself-organizingmap,folia Phoniatrica,45:173{181,1993. featuremapindetectionoffricative-vowelcoarticulation,journalofthe AcousticSocietyofAmerica,93(6):3468{3474,1993. Research,35:287{295,1992. [14]B.Martn-del-BroandC.Serrano-Cinca,Self-organizingneuralnetworks [16]D.NieburandA.J.Germond,Unsupervisedneuralnetclassicationofpower [15]R.Mujunen,L.Leinonen,J.Kangas,andK.Torkkola,Acousticpattern fortheanalysisandrepresentationofdata:somenancialcases,neural Phoniatrica,45:135{144,1993. Systems,14(2-3):233{242,1992. recognitionof/s/misarticulationbytheself-organizingmap,folia systemstaticsecuritystates,int.journalofelectricalpowerandenergy ComputingandApplications,1(3):193{206,1993.

384 [17]T.SorsaandH.N.Koivo,Applicationofarticialneuralnetworksinprocess [18]V.TrybaandK.Goser,Self-OrganizingFeatureMapsforprocesscontrolin ArticialNeuralNetworks,pagesI{847{852,North-Holland,1991. faultdiagnosis,automatica,29(4):843{849,1993. chemistry,int.kohonen,k.makisara,o.simula,andj.kangas,editors, REFERENCES [19]A.UltschandH.P.Siemon,Exploratorydataanalysis:UsingKohonen [20]A.UltschandH.P.Siemon,Kohonen'sself-organizingfeaturemapsfor networksontransputers,technicalreport329,univ.ofdortmund, Dortmund,Germany,1989. [22]M.Vapola,O.Simula,T.Kohonen,andP.Merilainen,Monitoringofan [21]A.Ultsch,Self-organizedfeaturemapsformonitoringandknowledge ICANN'93,Int.Conf.onArticialNeuralNetworks,pages864{867,Springer- pages305{308,kluweracademicpublishers,1990. Verlag,1993. exploratorydataanalysis,inproc.innc'90,int.neuralnetworkconf., T.Reponen,editors,Proc.oftheConf.onArticialIntelligenceResearchin anaesthesiasystemusingself-organizingmaps,inc.carlsson,t.jarvi,and acquisitionofachemicalprocess,ins.gielenandb.kappen,editors,proc. [23]M.Vapola,O.Simula,T.Kohonen,andP.Merilainen,Representationand Finland,number12inConf.Proc.ofFinnishArticialIntelligenceSociety, [24]P.WeierichandM.vonRosenberg,Unsuperviseddetectionofdrivingstates pages55{58,finnisharticialintelligencesociety,1994. editors,proc.icann'94,int.conf.onarticialneuralnetworks,pagesi{ 246{249,Springer-Verlag,1994. withhierarchicalself-organizingmaps,inm.marinaroandp.g.morasso, Int.Conf.onArticialNeuralNetworks,pagesI{350{353,Springer-Verlag, identicationoffaultconditionsofananaesthesiasystembymeansoftheself- OrganizingMap,InM.MarinaroandP.G.Morasso,editors,Proc.ICANN'94,