© 2016 DataRPM – Proprietary and Confidential11
Cog n i t i ve 	Pre d i c t i ve 	
M ai ntena nc e	( CPd M ) 	Pl atfor m
F o r 	 I n d u s t r i a l 	 I o T 	 ( I I o T )
Predictive Maintenance for Automotive Industry
© 2016 DataRPM – Proprietary and Confidential22
Predictive Maintenance - Automotive
2
Quality Control
• Assembly Line Uptime
• Production Quality / Warranty
• Early Detection & Prevention
• Paint Shop Quality
Service & Maintenance
• Maintenance & Fault Detection
Performance of Assets
• Service SLAs
Increase Productivity
• Increase uptime of assets
• Increase Jobs per Hour
• Lower Downstream costs of lost
productivity
• Spare Parts Availability
• Lead time indicator for expensive
parts
Optimize Inventory Customer Satisfaction
• Enhanced Reliability
• On time delivery
• Reduced maintenance problems
Reduce Warranty Costs
• Equipment Warranty
• Reduced Malfunctions
• Decreased Recalls
© 2016 DataRPM – Proprietary and Confidential33
Business Value We Help Unlock From PdM
For Automotive Industry Ecosystem
3
Minimize
Insurance
Risks
Prevent Car
Breakdowns
& Part
Failures
Prevent
Quality On
Assembly
Line & Paint
Shop
Minimize
Warranty
Claims
Minimize Car
Maintenance
Costs
Predicting
Potential
IssuesWith
Assets Ahead
OfTime Optimize
Parts
Inventory and
Field
Resources
Predictive maintenance will help
companies save $630 billion by 2025
McKinsey
© 2016 DataRPM – Proprietary and Confidential44
PdM Is Not New! But What Has Changed Now For $$$?
4
Sensors Everywhere And
Now ConnectedTo Internet
Big Data PlatformsTo Collect, Store And
Process Data At Scale
Price of sensors are rapidly
dropping close to $1.
The number of sensors
shipped has increased more
than five times in 2 years
from 4.2 billion in 2012 to
23.6 billion in 2014 and
growing rapidly!
Advancement In Data ScienceTechnologies
Meta Learning
Our Customers Our Partners
Us
© 2016 DataRPM – Proprietary and Confidential55
And… Data Science Is Hard For Machine Data
5
Weak SignalsTo Noise
The true signals hidden in the data
scattered over millions of data points
coming from various different
sensors at different time windows
Obsolete Models
The machine data patterns change
too often. Models get obsolete by
the time they are productionized
No LabeledTraining Data
No labeled data for training, lack
of knowledge about machine
signals and monitoring each
sensor in isolation doesn’t work
Not Human Scale
Timely and accurate insights is
impossible to get by manually
analyzing data samples
Besides there is a 200k - 1M Shortfall of Data Scientists
© 2016 DataRPM – Proprietary and Confidential66
Traditional Approach Of Analysis Don’t Work Anymore
Sensors individually monitored for Spikes Manually generated alert rules
Large Workforces required to
filter signals from noisy alerts
Manual rule-based monitoring solutions don’t deliver PdM ROI &TCO is high!
Data Science & Machine Learning driven approach is the only way to efficient PdM
All sensors analyzed continuously in
combination to learn machine states
Automatically Identify only the
critical signals
Recommend prescriptive
actions & learn from results
© 2016 DataRPM – Proprietary and Confidential77
The Only Solution: Teaching Machines to do Machine Learning
7
MLML Meta-Learning
on Machine-Learning
DataRPM is one of the first Enterprise-
grade applications of Meta-Learning.
Massive Economic Value is thus
delivered through Cognitive Predictive
Maintenance (CPdM) for Industrial IoT &
Manufacturing applications.
How Machines learns to do ML:
“Algorithmic Survival-of-Fittest”
7
1
Run many live automated ML
experiments on datasets in parallel
2
Extract meta-data from every
experiment based on
3
Train an Ensemble of Models on
this meta-data repository
4 Apply models to predict the best
algorithms & hyper-parameters
5 Build machine-generated and
human verified ML models for PdM
Dataset Characteristics
Selected Features
Selected Algorithm
Selected Hyper-Parameters
ResultantValue Of Objective Function
© 2016 DataRPM – Proprietary and Confidential88
We deliver a Cognitive Predictive Maintenance [CPdM]
software platform for the Industrial IoT [IIoT] that automates
Data Science with Meta-Learning on Machine Learning [MLML]
to solve large-scale problems
Automotive
Power | Energy | Utilities
Manufacturing
Healthcare
Oil | Gas
Transportation | Travel
SectorsResults
Results In
as quickly as
1/30th
the Time
with up to
30%
in Cost
Savings
at least a
300 %
increase in
Prediction
Power
© 2016 DataRPM – Proprietary and Confidential99
What Makes Us Different From Other PdM Solutions?
9
§ We are designed specifically to handle the challenges of doing PdM for IIoT
§ We cognitively automate the Data Science process at mass-scale
§ We utilize Meta-Machine-Learning [MLML] to teach machines to teach themselves
§ We Operationalize the best Ensembles & continually modify in-line & real-time
§ We partner with our customers to solve real problems and deliver ROI quickly
© 2016 DataRPM – Proprietary and Confidential1010
High-Level Automated Workflow for CPdM with MLML
10
Sensors (BatchTime Series Data)
Temperature
Pressure
Accelerometer
Noise
Feature Engineering
Features
Engg
Meta-
Learning
ALIGN
RESAMPLE
IMPUTE
MISSING
VALUES
ROLL-UP
MEAN /
STDEV
MIN /
MAX
CHANGE
RATE
FFT /
DWT
Anomaly Detection
Clustering
Meta-
Learning
LabeledTraining Data
USER VALIDATION
FREQUENT
PATTERNS SEEN IN
PRIOR FAILURES
Classifier /
Regression
Meta-
Learning
Model
Gen
Cross-
Validate
Is Quality
Good?
Tune
Params NO
Test
Correctly
Classified
?
YES
Add To New
Training Set
NO
Prediction Modeling
Sensor (Streaming Data)
Model Ensembles
Visualize on
DataRPM
Visualize on
Tableau etc.
Integrate via
APIs into
ServiceCloud
Feedback
Feedback
Production
CLASS BALANCE
Up-sample Minority Classes
Down-sample Majority Classes
CONNECTORS FEATURE ENGG RECIPE
SEGMENTATION RECIPE
INFLUENCING FACTORS
RECIPE
PREDICTION RECIPE
API FRAMEWORK + SCORING RECIPE +
RECOMMENDATION RECIPE + DASHBOARD
MLML
MLML
MLML
Prior
Maintenance/
ServiceAction
Records
© 2016 DataRPM – Proprietary and Confidential1111
Use Case | CPdM for Connected Cars
MANUAL
DATA ANALYSIS
CHALLENGE
Resulted in poor
prediction with lots
of false positives
Sensors
were
monitored
individually
Manual
rules were
written to
raise alerts
Impossible
to capture
all
failure scenarios
ALL
sensors
in parallel
Months
of sensor data
used to train Data
Models accurately
< 2 Weeks
Highly Accurate
Prediction Model
Automated building of
thousands of models
in parallel to deliver
the optimal model
Predictions of
Breakdowns &
Recommend
Maintenance
USE CASE
Identify the
indicators of
malfunction for
Connected Cars
011101
11001
0101
0110
110
DATA OVERLOAD
300%
Increase in
Prediction
Accuracy
Results
Results
Delivered
30 X
Faster
Each sensor
records multiple
data points in
millisec range
Unique Sensor
Recordings in
HDP Platform
75%
Reduction in
Breakdowns
Identified
Accuracy SavingsSpeed
AUTOMATED
DATA SCIENCE
SOLUTION w/
MLML
A Luxury
Automotive
Company
© 2016 DataRPM – Proprietary and Confidential1212
Use Case | CPdM For Assembly Line Quality
MANUAL
DATA ANALYSIS
CHALLENGE
AUTOMATED
DATA SCIENCE
SOLUTION
Resulted in poor
prediction with lots
of false positives
Sensors
were
analyzed
individually
Manual
rules were
written to
raise alerts
Impossible
to capture
all
failure scenarios
ALL
sensors
in parallel
Months
of sensor data
used to train Data
Models accurately
< 4 Weeks
Highly Accurate
Prediction Model
Automated building of
thousands of models
in parallel to deliver
the optimal model
Prevent defects
and reduce recalls
and warranty
claims
USE CASE
Identify the indicators
of poor quality in
production assembly
line
Each sensor
records multiple
data points in
millisec range
011101
11001
0101
0110
110
Unique Sensor
Recordings in
HDP Platform
DATA OVERLOAD
Results
30 X
Faster
328%
Increase in
Prediction
Accuracy
Results
Significant
Reduction in
Bad Parts
A Luxury
Automotive
Company
© 2016 DataRPM – Proprietary and Confidential1313
CPdM Platform For Connected Cars
13
Prediction Timeframe
Predicted Stage (Criticality) For
All Connected Car (Vin numbers)
Current Sensor Readings
For Selected Car
Available Info For
Selected Car
Components Requiring Predicted
Maintenance For Selected Car
Maintenance Alerts
With Recommended Time To
Maintenance For Selected Car
Predicted Sensor Readings
For Selected Car
© 2016 DataRPM – Proprietary and Confidential1414
14
THANK YOUF o r 	 M o r e 	 I n f o r m a t i o n
E m a i l 	 m a r k e t i n g @ d a t a r p m . c o m
V i s i t 	 	 w w w . d a t a r p m . c o m

Cognitive Predictive Maintenance for Automotive

  • 1.
    © 2016 DataRPM– Proprietary and Confidential11 Cog n i t i ve Pre d i c t i ve M ai ntena nc e ( CPd M ) Pl atfor m F o r I n d u s t r i a l I o T ( I I o T ) Predictive Maintenance for Automotive Industry
  • 2.
    © 2016 DataRPM– Proprietary and Confidential22 Predictive Maintenance - Automotive 2 Quality Control • Assembly Line Uptime • Production Quality / Warranty • Early Detection & Prevention • Paint Shop Quality Service & Maintenance • Maintenance & Fault Detection Performance of Assets • Service SLAs Increase Productivity • Increase uptime of assets • Increase Jobs per Hour • Lower Downstream costs of lost productivity • Spare Parts Availability • Lead time indicator for expensive parts Optimize Inventory Customer Satisfaction • Enhanced Reliability • On time delivery • Reduced maintenance problems Reduce Warranty Costs • Equipment Warranty • Reduced Malfunctions • Decreased Recalls
  • 3.
    © 2016 DataRPM– Proprietary and Confidential33 Business Value We Help Unlock From PdM For Automotive Industry Ecosystem 3 Minimize Insurance Risks Prevent Car Breakdowns & Part Failures Prevent Quality On Assembly Line & Paint Shop Minimize Warranty Claims Minimize Car Maintenance Costs Predicting Potential IssuesWith Assets Ahead OfTime Optimize Parts Inventory and Field Resources Predictive maintenance will help companies save $630 billion by 2025 McKinsey
  • 4.
    © 2016 DataRPM– Proprietary and Confidential44 PdM Is Not New! But What Has Changed Now For $$$? 4 Sensors Everywhere And Now ConnectedTo Internet Big Data PlatformsTo Collect, Store And Process Data At Scale Price of sensors are rapidly dropping close to $1. The number of sensors shipped has increased more than five times in 2 years from 4.2 billion in 2012 to 23.6 billion in 2014 and growing rapidly! Advancement In Data ScienceTechnologies Meta Learning Our Customers Our Partners Us
  • 5.
    © 2016 DataRPM– Proprietary and Confidential55 And… Data Science Is Hard For Machine Data 5 Weak SignalsTo Noise The true signals hidden in the data scattered over millions of data points coming from various different sensors at different time windows Obsolete Models The machine data patterns change too often. Models get obsolete by the time they are productionized No LabeledTraining Data No labeled data for training, lack of knowledge about machine signals and monitoring each sensor in isolation doesn’t work Not Human Scale Timely and accurate insights is impossible to get by manually analyzing data samples Besides there is a 200k - 1M Shortfall of Data Scientists
  • 6.
    © 2016 DataRPM– Proprietary and Confidential66 Traditional Approach Of Analysis Don’t Work Anymore Sensors individually monitored for Spikes Manually generated alert rules Large Workforces required to filter signals from noisy alerts Manual rule-based monitoring solutions don’t deliver PdM ROI &TCO is high! Data Science & Machine Learning driven approach is the only way to efficient PdM All sensors analyzed continuously in combination to learn machine states Automatically Identify only the critical signals Recommend prescriptive actions & learn from results
  • 7.
    © 2016 DataRPM– Proprietary and Confidential77 The Only Solution: Teaching Machines to do Machine Learning 7 MLML Meta-Learning on Machine-Learning DataRPM is one of the first Enterprise- grade applications of Meta-Learning. Massive Economic Value is thus delivered through Cognitive Predictive Maintenance (CPdM) for Industrial IoT & Manufacturing applications. How Machines learns to do ML: “Algorithmic Survival-of-Fittest” 7 1 Run many live automated ML experiments on datasets in parallel 2 Extract meta-data from every experiment based on 3 Train an Ensemble of Models on this meta-data repository 4 Apply models to predict the best algorithms & hyper-parameters 5 Build machine-generated and human verified ML models for PdM Dataset Characteristics Selected Features Selected Algorithm Selected Hyper-Parameters ResultantValue Of Objective Function
  • 8.
    © 2016 DataRPM– Proprietary and Confidential88 We deliver a Cognitive Predictive Maintenance [CPdM] software platform for the Industrial IoT [IIoT] that automates Data Science with Meta-Learning on Machine Learning [MLML] to solve large-scale problems Automotive Power | Energy | Utilities Manufacturing Healthcare Oil | Gas Transportation | Travel SectorsResults Results In as quickly as 1/30th the Time with up to 30% in Cost Savings at least a 300 % increase in Prediction Power
  • 9.
    © 2016 DataRPM– Proprietary and Confidential99 What Makes Us Different From Other PdM Solutions? 9 § We are designed specifically to handle the challenges of doing PdM for IIoT § We cognitively automate the Data Science process at mass-scale § We utilize Meta-Machine-Learning [MLML] to teach machines to teach themselves § We Operationalize the best Ensembles & continually modify in-line & real-time § We partner with our customers to solve real problems and deliver ROI quickly
  • 10.
    © 2016 DataRPM– Proprietary and Confidential1010 High-Level Automated Workflow for CPdM with MLML 10 Sensors (BatchTime Series Data) Temperature Pressure Accelerometer Noise Feature Engineering Features Engg Meta- Learning ALIGN RESAMPLE IMPUTE MISSING VALUES ROLL-UP MEAN / STDEV MIN / MAX CHANGE RATE FFT / DWT Anomaly Detection Clustering Meta- Learning LabeledTraining Data USER VALIDATION FREQUENT PATTERNS SEEN IN PRIOR FAILURES Classifier / Regression Meta- Learning Model Gen Cross- Validate Is Quality Good? Tune Params NO Test Correctly Classified ? YES Add To New Training Set NO Prediction Modeling Sensor (Streaming Data) Model Ensembles Visualize on DataRPM Visualize on Tableau etc. Integrate via APIs into ServiceCloud Feedback Feedback Production CLASS BALANCE Up-sample Minority Classes Down-sample Majority Classes CONNECTORS FEATURE ENGG RECIPE SEGMENTATION RECIPE INFLUENCING FACTORS RECIPE PREDICTION RECIPE API FRAMEWORK + SCORING RECIPE + RECOMMENDATION RECIPE + DASHBOARD MLML MLML MLML Prior Maintenance/ ServiceAction Records
  • 11.
    © 2016 DataRPM– Proprietary and Confidential1111 Use Case | CPdM for Connected Cars MANUAL DATA ANALYSIS CHALLENGE Resulted in poor prediction with lots of false positives Sensors were monitored individually Manual rules were written to raise alerts Impossible to capture all failure scenarios ALL sensors in parallel Months of sensor data used to train Data Models accurately < 2 Weeks Highly Accurate Prediction Model Automated building of thousands of models in parallel to deliver the optimal model Predictions of Breakdowns & Recommend Maintenance USE CASE Identify the indicators of malfunction for Connected Cars 011101 11001 0101 0110 110 DATA OVERLOAD 300% Increase in Prediction Accuracy Results Results Delivered 30 X Faster Each sensor records multiple data points in millisec range Unique Sensor Recordings in HDP Platform 75% Reduction in Breakdowns Identified Accuracy SavingsSpeed AUTOMATED DATA SCIENCE SOLUTION w/ MLML A Luxury Automotive Company
  • 12.
    © 2016 DataRPM– Proprietary and Confidential1212 Use Case | CPdM For Assembly Line Quality MANUAL DATA ANALYSIS CHALLENGE AUTOMATED DATA SCIENCE SOLUTION Resulted in poor prediction with lots of false positives Sensors were analyzed individually Manual rules were written to raise alerts Impossible to capture all failure scenarios ALL sensors in parallel Months of sensor data used to train Data Models accurately < 4 Weeks Highly Accurate Prediction Model Automated building of thousands of models in parallel to deliver the optimal model Prevent defects and reduce recalls and warranty claims USE CASE Identify the indicators of poor quality in production assembly line Each sensor records multiple data points in millisec range 011101 11001 0101 0110 110 Unique Sensor Recordings in HDP Platform DATA OVERLOAD Results 30 X Faster 328% Increase in Prediction Accuracy Results Significant Reduction in Bad Parts A Luxury Automotive Company
  • 13.
    © 2016 DataRPM– Proprietary and Confidential1313 CPdM Platform For Connected Cars 13 Prediction Timeframe Predicted Stage (Criticality) For All Connected Car (Vin numbers) Current Sensor Readings For Selected Car Available Info For Selected Car Components Requiring Predicted Maintenance For Selected Car Maintenance Alerts With Recommended Time To Maintenance For Selected Car Predicted Sensor Readings For Selected Car
  • 14.
    © 2016 DataRPM– Proprietary and Confidential1414 14 THANK YOUF o r M o r e I n f o r m a t i o n E m a i l m a r k e t i n g @ d a t a r p m . c o m V i s i t w w w . d a t a r p m . c o m