AWEA Cognitive Analytics for Predictive Futures

PredictiveFuturesCognitiveAnalytics

TodaysSpeakers

StuartGillenDirector,

BusinessDevelopment

MaggiePakulaManager,

PerformanceAnalysisEngineering

JohnHensleyManager,

IndustryData&Analysis

“”

GlobalDataPowerestimatesthemaintenanceexpenditureonwindturbinesvitaltoproductivityisexpectedtorisefrom$9.25Bin2014to$17Bin2020

http://www.edie.net/news/6/Win-turbine-maintenance-costs-to-nearly-doubl/

“”

Itisestimatedthatin2011,nearly$40billionworthofwindequipmentintheU.S.willbeoutofwarranty,thrustingthefinancialriskontheownertoprovidecost-effectiveoperationandmaintenance.

http://www.renewableenergyfocus.com/view/26582/wind-getting-o-m-under-control/

CostofGearboxFailures

u Romax Studyestimatedcostofplanetarybearing failures>350k[1]

u In2014Siemenswrotedown€223Mtoreplacebearings infleet<2yrs.old[2].

u Controllingwindturbineswithdata-drivensoftwarecould,modelsshow,increaseenergyproductionbyatleast10%andgainsof14-16%arepossible[3]

u Theaveragegearboxfailurerateover10yearsisestimatedat5%[4].

1 0.75 0.5 0.25 0 2 4 6 8

ElectricalSystems

ElectronicControl

Sensors

HydraulicSystems

YawSystem

RotorBrake

MechanicalBrake

RotorHub

Gearbox

Generator

SupportingStructure/Housing

DriveTrain

UseCase

Aboutu Develops,Owns,andOperatesPowerGenerationandEnergyStorageUnitsinUSandEurope

u NorthAmerica’slargestindependentwindpowergenerationcompany

u Currentlyoperatingover4MWofwind

Headquarters

RegionalOffice

WindProject

NaturalGas

SolarProject

Storage

GearboxMonitoringApplicationTrial

u DesiredResults

u Predictgearboxfailureswith30-60dayadvancednotice

u Zeroorminimal falsepositives

u “DummyLight”output

u DataProvided

u 4yearsofhistoricaldatafromsiteof~100turbines

u 27datavariablesat10minuteresolution,novibrationvariablescollected

u Majorcomponentfailurelogs

GeneratedPredictionSignaturesforallCatastrophicGearboxFailures

RiskIndexforGearboxHealthTrialOutcomes

• Impendingfailure(redalert)predictionforcatastrophicfailure>1month

• Advanceddegradationwarning(amberalert)forfailuresis>2months

• Wehadzerofalsepositives,thatisnoalertswereraisedwhichdidnothaveafailurefollow

• Wehadzerofalsenegatives,thatisnofailuresweremissed

For100Turbines

67352

0

40

60DaysofWarning

500

1000

67Days

35Days

OutputOptions

OverallFleetHealth DetailedAssetView

OR

Connectoutputtoexistingsystems

GMSSCADA

CustomerUserInterface

Somesecondaryobservationsu Otherfailures(likelyblade),haveveryshortfailuresignatures

u Failurepredictionmadeindayscomparedtolonger signaturesforgearboxfailures

u SeasonalityofGearboxfailuresu Catastrophicgearboxfailuresshowhighcorrelationtoseasons

u Mostfailuresinsecondhalfoftheyear:Q1– 1,Q2– 1,Q3– 9,Q4–7

u Otherfailuresrelativelyindependentofseasonsu Catastrophicfailuresdistributedmoreuniformly:Q1– 3,Q2– 4,Q3– 5,Q4– 5

NextSteps

u Expandto5sites

u Pendingresultsofexpandedsites,committedtoenterprisewideroll-out

u Explorepredictivemodelsforothermajorcomponents

UsingMachineLearningandCognitiveFingerprinting™toDriveResults

Category KeyFeatures

Business Intelligence(BI)

• Centralizedanalysis• Uniformdatacollection• Averagevisualizations

RulesBasedModeling

• Fixedrulesmustaccountforalltypesoftransactionsinalltypesofconditions;leadtoruleproliferationandmanagementchallenges

• Maybegoodmeasuresforsomesimplesituations,butaverage(orevensub-par)measuresforothers

StatisticalAnalysis

• Identifiesdeviationsfrom“normal”• Moreaplatformformodelbuildinganddatascientists

thananalertgeneratingsolution• Notautomatedtoaccountforchangingconditions

PhysicsBasedModeling

• Asset-type specific• ModelbuildingIs averyhands-onprocessinvolving

laboratoryexperiments• Domainexpertsapplythesephysicalmodelsuniversallyto

assets

CommonApproaches

Enablesmachinestopenetratethecomplexityofdatatoidentifyassociations

Presentspowerfultechniquestohandleunstructured data

Continuously learnsnotonly frompreviousinsights,butalsofornewdataenteringthesystem

ProvidesNLPsupport toenablehumantomachineandmachinetomachinecommunication

Doesnot requirerules, insteadreliesonhypothesisgenerationusingmultipledatasetswhichmightnotalwaysappearconnectedorrelevant

BenefitsofCognitiveAnalytics

NLP: Natural Language Processing

CognitiveAnalyticsisinspiredbythewaythehumanbrainoperates:

ProcessesInformation

DrawsConclusions

Codifies Instincts&ExperienceintoLearning

BasicsofMachineLearning

Howdoyoulabelthese?

UnsupervisedLearning

UnsupervisedLearning

SM

MD

LG

SupervisedLearning

WH

GR

BL

Unsupervisedvs.SupervisedLearning

Unsupervised Supervised

Index Date Time AssetID Value2 5-Apr-10 7:01 750 8993 22-Mar-13 8:19 904 7927 20-Oct-14 8:26 545 745 10-Jul-12 7:38 552 8668 15-Sep-11 8:13 942 7429 1-Jun-11 8:44 900 7291 20-Jul-11 7:14 587 5054 12-Jul-10 7:36 765 9520 5-Sep-14 8:25 813 3944 30-Jun-11 7:07 983 71100 5-Oct-12 7:35 802 3466 12-Mar-10 7:39 726 4745 6-May-11 7:30 973 9884 10-Dec-12 7:17 504 6843 9-Jul-14 8:07 567 74

ActionTaken ComponentRepair Blade

Unknown BladeRepair Gearbox

Replaced GearboxReplaced Gearbox

NTF GeneratorGood GeneratorNTF BladeRepair GeneratorNTF GearboxNTF BladeRepair Gearbox

Unknown GearboxRepair BladeRepair Gearbox

TheSparkCognitionMethodology-CognitiveFingerprinting™

OurAlgorithms

Artemis• Proprietary regularization tool for

feature selection

• Automated class balancing

• Automated model selection

• Automated checks on overfitting

• Turn-key solutions with health index for industrial use cases

Iris• Proprietary clustering algorithm

• Optimal clustering of data leading to state generation

• Semantic indexing of states

• Classification from indexed states

• Turn-key solutions with health index for industrial use cases

Pythia• Proprietary regularization tool

for feature selection

• Genetic algorithms for optimizing neural networks

CognitiveAlgorithms-SparkArtemis™

Overall Vibration

MAXTemp

MinTemp

TensileForce

ShearStrength

FirstOrderFeatures SecondOrderFeatures

Wavelets

Enveloping

JointTimeFrequency

DoubleIntegration

CreatedFeaturen

ThirdOrderFeatures

CrestFactor

Integration

RunningAverage

Cauchy StressTensor

Created Feature1

CognitiveAlgorithms-SparkPythia™Artemis

ArtemisFeaturesTakeArtemisfeatures

Capturesthestateofandevolutiontofailure/eventincludingsubtleinfluencers

StartNeuralnetgeneticcomp

PredictBasedonaFunction

Significantly advancedcomparedtoexistingalgorithms

FeatureSelection

Automatically findsignificantdata

Adaptive&Self-learning

Identifymultiple topperformers

DefineRelationships

CognitiveAlgorithms-SparkPythia™

CognitiveAlgorithms-SparkIris™

WhichisBetter?

…Model1 Model1 Model1 Model1 Modeln

CognitiveAlgorithms-SparkIris™

Answer:Neither

Model1

Model2

Model3

Objectives Monitor Critical Assetsduringstartupsandcoast-downs

Predict RemainingUsefulLife

Analyzefailures, alertonimpendingfailures, optimizedesign

Client

Asset

BigUtility

Turbine Generator

BigUtility

WindTurbine

On-shoredriller

Electrical Submersible Pump

• Datacollectionfrommultipleassets

• Detectsfailures,graduatingtopredictions

• Self-learningsystemwithaccesstoin-contextadvisorypoweredbyIBMWatson

• RUL(RemainingUsefulLife)predictionandanomalydetection

• Automatedmodelbuilding,selection&management

• Insightsthroughdeeper-orderanalyses

• Failureidentificationandclassification

• Automatedfailurealerting• Criticalvariableidentification• Designandprocessoptimization

toreducespecificfailures

SolutionFeature

BusinessImpact

• Estimatedincreaseinproductivityof25%–30%

• 50XROI

• Estimatedsavingsof~40%inO&Mbudgets

• ~$2MMperyearfor100MWpowergenerationplant(wind),40XROI

• 3XincreaseinlifeofESPthroughpropermonitoringanddesign

• Savingsofupto$150,000perassetperyear,50XROI

OtherEnergySectorApplications

UseCase-ImproveSafetyandReduceRemediationCostThroughIntelligentPrognosticsu SparkCognition hasdeveloped anIBMWatson

“Advisory” application forAssetMaintenance

u SparkCognition’s poweredbyIBMWatsonwillallow

Directors ofMaintenance andtechnicians to:

§ Conductmachinetohumandialoguetotroubleshoot

faultcodes

§ Predictimpendingfailuresandfaults

§ Identifytherightfaultcodesandtroubleshootingtips

usingnaturallanguagequeries

§ Findsolutionstoproblemsandadvisetechnicians

§ Optimizeworkflowanddeliverrelevant

documentationforafasterturnaround

MachineLearning&CognitiveAnalyticscandeliverseveralbenefits

ExternalFactorsCanincorporateexternalfactors(e.g.environmentalissues suchasbirds&bats)

ScalabilityAutomatedmodel building capabilitydoesnotrequiremanualmodel buildingofeveryasset/component

In-contextRemediationAdvisor thatunderstandsnaturallanguagetohelptechnicalteams

SecurityOut-of-band, symptom-sensitive approachbeyond ITsecurity

AdaptabilityAdaptstonewandchangingconditionsautomatically

HigherAccuracyAutomatedfeatureenrichmentandextractionthatcandeliver betterinsightsandhigheraccuracy

Questions