DATA SCIENCEPOP UP

AUSTIN

Applied Machine Learning for IoT

Hank RoarkData Scientist & Hacker, H20.ai

hankroark

DATA SCIENCEPOP UP

AUSTIN

#datapopupaustin

April 13, 2016Galvanize, Austin Campus

http://www.dominodatalab.com

Appl iedMachineLearningfor IoT

HankRoark

13-April-2016DataSciencePop-upAustin

WHOAMI I

Lead,CustomerDataScience@H2O.ai

JohnDeere:Research,SoftwareProductDevelopment,HighTechVenturesLotsoftimedealingwithdataoffofmachines,equipment,satellites,weather,radar,handsampled,andon.Geospatial,temporal/timeseriesdataalmostallfromsensors.

SystemsDesignandManagement:MITPhysics:GeorgiaTech

hank@h2o.ai @hankroarkhttps://www.linkedin.com/in/hankroark

WHYIOT?

Because it’s trending (so it must be promising)!

Internet of Things

Deep Learning

IF YOUAREINTODATA,THEIOTHASIT

LOTSOFDATA

LOTSOFDATA

GETREADYFORBRONTOBYTES!!

WOW,HOWBIGIS ABRONTOBYTE?

Image courtesy http://www.telecom-cloud.net/wp-content/uploads/2015/05/Screen-Shot-2015-05-27-at-3.51.47-PM.png

I OT ARCH I T EC TURE I S S E TUP FOR RE IN FORC ING , E XPONENT IA L GROWTH

MYPARTICULARCURIOSITY

EXAMPLEFROMTHEIOTDomain:PrognosticsandHealthManagementMachine:TurbofanJetEnginesDataSet:A.SaxenaandK.Goebel(2008)."TurbofanEngineDegradationSimulationDataSet",NASAAmesPrognosticsDataRepository

PredictRemainingUsefulLifefromPartialLifeRuns

Sixoperatingmodes,twofailuremodes,manufacturingvariability

Training:249jetenginesruntofailureTest:248jetengines

SENSORSSHOWTRENDSOVERTIME

Time

We also know remaining useful lifedecreases by at least one (1) cycle each operation.

How can one take advantage of this knowledge?

INCORPORATINGPRIORSTATE

Disentangling the dynamic core: a research program for a neurodynamics at the large-scale MICHEL LE VAN QUYEN Biol. Res. v.36 n.1 Santiago 2003 http://dx.doi.org/10.4067/S0716-97602003000100006

One option: Phase Space Embedding

Drawbacks: Incorporates knowledge from onlysmall number of prior states

Curse of dimensionality

Another parameter (tau)

KALMANFILTER

FINALPIPELINE

Featureengineering• Signalprocessing,featurecreation,featureselection

Regression Models• Supervised

MachineLearning(H2O)

PostProcessing• Kalmanfilter(pykalman,inthiscase)

Not really that much different than typical machine learning pipeline

BEFOREANDAFTERPOST-PROCESSING

CYCLE NUMBER CYCLE NUMBER

PR

ED

ICTE

D R

UL

UNIT #2UNIT #2

This presentation, data, and associated Jupyter notebooks (look in 2016_04_13_AppliedMachineLearningForIoT):https://github.com/h2oai/h2o-meetups/

WHYH2O

FLEETTELEMATICS:PREVENTIVEMAINTENANCE

PROBLEM

• H2Osupportforcustomer’sKerberosauthenticationmechanismforHadoop

• SupportforMapReduce,YARN,R,PythonandSparkinHadoop

• In-memory,distributedarchitecture • RapiddeploymenttoproductionwithPOJO • QuickprototypingwithH2OFlow

• Fleettelematics—analyzemaintenancerecordsandvehicleperformancetomakepredictionsonwhentodopreventivemaintenance

• Couldn’tscalebysamplingdata • Tookdaystocreatemodels

IMPACT

“AnnualSavingsare$7M” –Anonymized,MemberTechnicalStaff

• Whenyoulookatthecostoftowingastrandedvehicle,technicianlossofproductivity,andthecustomerlifetimevalue,theannualsavingsis$7M.”–Anonymized,LeadMemberTechnicalStaff

Leading Mobile Telecom Operator

Telecommunications

TAKE-AWAYS

• Knowyour“physics”o ”Physics”iswillbelikeafishfinderinthisseaofBrontobytes

o Lineardynamicsystems,Queues,Signals,etc.• BigdataandBigmodelsandSmallmodels• Keepingandeyeonpromisingnewmethods(e.g.,ConvolutionNeuralNetsandLSTM-RNN)

H2ORESOURCES

• Downloadandgo:http://www.h2o.ai/download• Documentation:http://docs.h2o.ai/• Booklets,Datasheet:http://www.h2o.ai/resources/• Github:http://github.com/h2oai/• Training:http://learn.h2o.ai/

THANKYOU

DATA SCIENCEPOP UP

AUSTIN

@datapopup #datapopupaustin