Rebooting Operational Excellence in Automotive Paint Shops Using Analytics

1

Rebooting Operational Excellence In the Automotive Paint Shop

© 2016 DataRPM – Proprietary and Confidential 2

ABHISHEK TANDONBusiness Insights Manager

Why are we discussing this today

v The paint shop corresponds to nearly third of the time taken in the automotive build process.

v Automation and Manual Processes: The process of paint application involves a large number of steps with a high degree of automation along with manual processes and inspections.

v As per research, 40% of the cars that exit a paint shop are likely to undergo some sort of rework (either on a part or in total).

v Environmental Impact: Paint is one of the highest waste generating process with a major environmental impact.

v Capital Intensive: Setting up and changes to the paint shop is a capital intensive process (almost a third of the total cost).

v Automation is generating large amount of data which is unused for asset analysis3

Henry Ford Once Famously Said

4

“Any customer can have a car painted any colour that he wants so long as it is black.”

Customer Preferences Have Changed Over Time

Car Colors have always reflected the mindset of the era and hence have a huge impact on the customer purchase mindset

Customer tastes have transformed over time and car manufacturers have to match up to their requirements with no tolerance in drop of quality

5

All Blacks 1920’s

Wacky 1960’s

Apple Effect of 2000’s

Paint Application Process Has Adopted to Change

6

Paint Shop Process


Observations

• Entire Paint Shop Process is a 7+ hour multi-step process with a number of manual and automated steps

• Defects in the paint process can occur due to:• ExternalInfluencersliketemperature, dirt, particles, paint quality, mix etc• ProductionLineInfluencerslikefault in the robotic arms• ManualInfluencerslikelindt particles, hair, PVC particles fromworkers

• There are multiple inspections done on the quality of the paint of the car both manually and through automated techniques to diagnose defects

• In case a defect is observed, it is recorded in the system and the car is sent back for rework

• On average, each manufacturer could spend >£76,000 in re-work costs for every 10,000 cars produced.

***Based on 63 million manufactured cars produced globally in 2012***8

Issues Are Many But Solutions Are Few


Thenumberofobserveddefectsandassociatedfactorsinthepaintshopareextremelyhigh.

Mostoftheanalysisisdoneafter thedefectisobserved.Thiscausesline disruptions and rework.

Mostofthefactorsareattributedtosomeexternal influence butthetrue root cause isnotdeterminedwithcertainty.

Changesinmultiplefactorswhichmayhavehappenedtogetherarenot takenintoconsiderationduetothemanualapproachofanalysis.

Applying Preventive Maintenance

10

Man Machine Method Material

IdentifythePossibleCauseAcrosstheEntireProcess

ConductanAssetLevelAnalysisforeachRobotic

Arm

PredictWorkingStatesofallthe

assetsindependently


Change in Scale requires a Change in Approach

AnalyzeRootCauseoftheErrors

DetectAnomaliesinmachinedata

AnalyzeRootCauseofWithinMachineError

CreateaPredictionModelforEveryMachine RecommendActionables

ApplyFeatureEngineeringtoenhancemachinedata

SegregateMachineandMethod

Transform

Enhance

TimetoFrequencyDomain

Mean,Skewness,Kurtosis

UnsupervisedLearning

IntrinsicFactors

ExtrinsicFactors

Predictpossiblefailuresinadvance

ReduceDowntimeandimprovequality

ImproveSystemSettings

AdjustTemperatureRange

IncreasePaintDensity

Process Level Analysis - Root Cause Analysis

12

ApplyDataScienceTechniqueswhichworkonlargeamountofdata,usemachinelearningtolearnandcorrelatemultipledatapatternsandfinallycreateassetlevelpredictivemodelsaretheneedofthehour.

AnalyzeRootCauseoftheErrors

• SimultaneouslyobservethechangesintheVitalX’sacrosstheentireproductionlinetodifferentiate betweennormalandabnormalworkingconditions

• UsingAssociativeMiningRules,determinecorrelationsbetweendefectsandobservedbehaviorinordertodeterminecandidates of causality

• Thismaybelinkedtooccurrence ofacertainoperator,aparticulartimeofday,linetemperature,numberofunitsprocessedbeforethedefect,aparticularpartofthecar,roboticarmsettingsetc.

• Thefactorsneedtobeobservedindependently and in combination todeterminetheeffectofinteractionacrossthecandidateX’s

Asset Level Analysis – Engineering Features on Sensor Data

13

• Fromthepreviousanalysis,therearelikelytobecertain machines and processes whichwillgethighlightedasVitalXcandidates.

• Owingtohighdegreeofautomation,themachinearmsarefittedwithmachine sensor data.Thesedatapointsneedtobecapturedandanalyzedforpotentialdeviations.

• BuildingaPredictiveMaintenancePipelinewillallowyoutoensurethatdeviationsinthosemachineandprocessescanbecapturedearly in the cycle anddonotleadtoqualitydisruptions.

ApplyFeatureEngineeringtoenhancemachinedata

TimetoFrequencyDomain

Mean,Skewness,Kurtosis

• Feature Engineering isessentialtounderstandtheeventsthatprecededthesensorvalueatagiventimeaswellasdeviationsinthereadings.

• Thesefeaturearecriticaltodifferentiatenormalworkingconditionstoanomalies

Choiceoffeatureswillvary fromassettoasset.Hence,itisimperativethatmultiplefeaturesarecreatedandthealgorithmsdeterminewhichfeaturesneedtobeselectedforthefinalmodel.

Asset Level Analysis – Determining the Anomaly State

14

DetectAnomaliesinMachineData

AnalyzeRootCauseofWithinMachineError

UnsupervisedLearning

IntrinsicFactors

ExtrinsicFactors

• BasedontheFeaturescreatedinthepreviousstep,weneedtodifferentiatebetweenthenormalandabnormalworkingconditionsofthemachine.

• TherightoperatingcriteriacanbeinfluencedbymanualrulesbuttheunsupervisedlearningalgorithmwilldeterminethevariousoperatingstateswithintheMachine

• AssociatedConditionsinthesestateswillhighlightvariousintrinsicandextrinsicfactorswhicharecausingthemachinetobeinthatstate

• Thiswillfurtherhelpdeterminetherootcauseofdeteriorationofthemachinehealth

Asset Level Analysis - Preparing for the Future

15

CreateaPredictionModelforEveryMachine

Predictpossiblefailuresinadvance

ReduceDowntimeandimprovequality

• Allthisanalysisculminateswithanassetlevelpredictionmodelwhichdeterminesthelikelymachinestateinthenearfuture.

• Themodelisregularlytweakedbyafeedbackloopwhichtunesitbasedonthechangingworkingconditionsofthemachine.

• Ratherthanpost-mortemmonitoring,thefloormanagerscanbebetterpreparedtotweakprocessesiftheyobserveanysignificantdeviationinthemachine.

RecommendActionables

ImproveSystemSettings

AdjustTemperatureRange

IncreasePaintDensity • Further,basisthestrengthofthemodelvariables,variousrecommendationscanbemadetothelinemanageraroundimprovingtheprocess.

• Thesetooaredrivenbymachineobservedpatternsandbasistheassetconditions

The Solution: Meta-Learning on Machine-LearningTeaching Machines to automate Machine Learning


DataRPM is the first Enterprise-Grade application of Meta-Learning for Machine Learning & Data Science Automation.

Massive Economic Value is thus delivered via our Cognitive Predictive Maintenance (CPdM)software platform for Industrial IoT & Manufacturing applications.

MLML

Meta-Learning on Machine-Learning

Automating how Machines learn to do Meta Learning:“Algorithmic Survival-of-Fittest”


1 3

4

5

2

Run many live automated ML Experiments in parallelfor each Asset

Extract Meta-Data from every Experiment:

• Dataset Characteristics• Selected Features• Selected Algorithm• Selected Hyper-Parameters• Resultant Value of Objective Function

Train an Ensembleof models via our

Meta-Data Repository

Apply Models to Predict the best Algorithms & Hyper-Parameters

for each asset

Build Machine-Generated& Human-Verified Models

for each & every Asset

DataRPM End-to-End Workflow


Consumption

APIs

Micro Apps Framework

Security

DataManagement

Data Sync

Data Lake

Metadata

Machine Learning& Analytics

Spark Engine

Workflow Builder

Data Science Recipes

Meta Learning

Natural Language

Visualization

IIoT Sensors Data

Enterprise Asset Management Systems

RDBMSDatasources

Hadoop

Insights App

Discovery App

Admin App

PdM Apps

DataRPM Key Differentiators


Natural LanguageSearch

DistributedComputing

API Driven+ + + + +Digital

TwinMeta

LearningOpen-box

SolutionCognitiveAutomation+

$250M+ in Annual Cost Savings identified

20

Average

> 80%+increase in Prediction Power

Go Live in

Days - WeeksAverage

30%Cost Savings

DataRPM End-to-End Full Tech Stack

21© 2016 DataRPM – Proprietary and Confidential

Pluggable App Container

Citizen Data

Scientists

Nat

ural

Lan

guag

e (N

LQA

) Sea

rch

& D

isco

very

Eng

ine

OA

uth

Bas

ed A

PI A

uthe

ntic

atio

nHadoop Data Lake

Spark (ML &

MLLIB)

Adaptive Indexing Cheetah (Powered

by ES)

Meta Data Store (MongoDB)

SAP HANA, Teradata, Hadoop/Hive, HBase,

MongoDB, Oracle, MySQL, PostgreSQL, SQL Server,

DB2, Redshift, CSV, Salesforce, Google

Spreadsheet…

Business Analysts

Data Scientists

Data Engineers /

BI Architects

App

licat

ion

Inte

grat

ion

RES

T C

onsu

mpt

ion

INSIGHTS App

DISCOVERY App

ETL App

API

Lay

er

Func

tiona

l & D

ata

Secu

rity

Meta Learning Engine

Data Science Recipes

Recipe Builder

Workflow Builder

Visualization Library

Recipe Building Framework

Data

RPM

Anal

ytic

s En

gine

REP

L En

gine

(Zep

pelin

)

EYWA DATA CONNECTOR &

LOADING MANGER

TASK ENGINE

EVENT LOADER

UNIOSGI BUNDLE

MICRO WORKFLOW EXECUTION

ENGINE

Dat

aRPM

Uni

fied

Dat

a A

cces

s La

yer

DataRPM ProprietaryDataRPM Secret Sauce Open Source Multitude of Data Sources & Connectors

Rebooting Operational Excellence in Automotive Paint Shops Using Analytics

Data & Analytics