10:40am (US Central) 9:40am (US Mountain) 8:40am (US Pacific) · Receiver Operating Characteristic OBE e t 0.90 0.00 0.20 0.60 0.80 1.00 b 0.5602 0.5608 Area Receiver Operating Characteristic

1

up

Up next 11:40am (US Eastern) 10:40am (US Central)

9:40am (US Mountain)8:40am (US Pacific)

Transform your data into strategic business value with predictive analyticsModerator: Lucian Fogoros, IIoT WorldSpeaker: Serg Posadas, Clockwork Solutions

Ask a question!Use the chat tool or

tweet using #iiotvirtualconf

2

3

You’ve never had more data on your strategic assets

• historical data on operations, maintenance, and inspections

• real-time and sensor data

• digital and virtual asset models

4

But data ALONEprovides NOstrategic VALUE

5

gathering data+ predictive modeling+ actionable insights= Strategic Value

Transforming data into timely insights and relevant actions

66

“million dollar” questions

• How can I improve total effective equipment performance?

• How do I get the most value out of my assets?

• How do I avoid costs when operating and maintaining my assets?

• How do I best manage spare parts to keep my assets running?

• How do I best design my operation to employ my assets?

• How do I maintain optimal uptime and asset-generated revenue?

…So you can answer these

7

Rear-facing BI is not accurate

Historical business “intelligence”is looking back at historical data

in attempt to react.

What Just Happened?

8

…buthow do you get the right answers?

traditional analytics predictive analytics

• rear facing

• helpful after the fact

• reactive

• forward looking

• provides warning and actionable insight

• supports well-developed strategies

• quantifies risk

traditional forecasting

• tied to historical data

• doesn’t account for operation changes

• can’t anticipate dynamic environment, aging, etc.

Stra

tegi

c va

lue

? •

What just happened?What will happen?

What should we do?Most solution providers

Next gen predictive analytics

9

Traditional forecastinganalytics

• Limited to historical view

• Can’t include future events, policy changes, evolution of business environment

• Does not measure risk

• Very likely to incur greater costs or more down time

• Lacks ability to support strategic planning

Traditional forecasting trades accuracy for ease of implementation

?

10

Predictiveanalytics

• Insights well beyond rear facing analytics

• Historical data only defines starting point

• Models future events for each asset and its components

• Simulates operations hour by hour, including failures, repairs, shipments, part buys, refurbishment, retirement, obsolescence, etc.

• Provides a holistic view of complex scenarios

Predictive analytics provides a higher degree of accuracy

11

Predictive analytics is determiningindustry leaders Within the next five years, advanced implementation of

Industry 4.0 will become a ‘qualifier to compete’ and is also likely to be seen by investors as a ‘qualifier for funding’.

Industry 4.0: Building the digital enterprise, PwC 2016 Global Industrial Survey, April 2016

12

IIoT predictive analytics

Challenges

Data

• Multiple sensors

• High volume & velocity

• Complex distribution of sources

Obstacles

• Simple data but requires advanced techniques

• Combine asset health monitoring with maintenance &

operations data

• Need automation

13

CBMchallenge

Asset health monitoring for predictive maintenance analytics

Benefits• Leverage advances in predictive health maintenance

• Reduced unplanned downtime

• Control costs

Challenges• Data quality

• Data structures

• Volume & velocity of real-time and historical data sets

• Prediction accuracy

• False positives

14

Data aggregation

• Automate data aggregation & transformation

• Leverage latest techniques for data conditioning

• Combine data silos sources for complete, accurate prognostics

15

MachineLearning& CBM

16

1. Separate data into training and test sets.

2. Describe the data attributes in the training data set

Normal Pre-Failure FailureNormal NormalNormal

MachineLearning& CBM Training

17

Normal Pre-Failure Normal Pre-Failure Failure

MachineLearning& CBM

1. Separate data into training and test sets.

2. Describe the data attributes in the training data set.

3. Apply a predictive technique

4. Evaluate predictor with test data and measure error.

5. If not satisfied, try another predictor. Repeat while minimizing error.

Test Predictions

18

MachineLearning& CBM

1. Separate data into training and test sets.

2. Describe the data attributes in the training data set.

3. Apply a predictive technique

4. Evaluate predictor with test data and measure error.

5. If not satisfied, try another predictor. Repeat while minimizing error.

6. Select best prediction algorithm. Predict based on new values.

7. May have to re-train as conditions evolve.

NormalPre-Failure Pre-Failure Failure Normal Normal

New Predictions

19

CBMreality

Sensor data is noisy & inconclusive

sensor 1

sensor 2

sensor 3

sensor 4

sensor 5

not so fast…

20

CBMconditionindicators

traintest

False Positive

Rate 10%

Anomaly Detection

Rate 65%

21

Complete predictive view

Sufficient data set

0

2

4

6

8

10

0 100 200 300 400

Minor maintenance

actionsOperational profile

change Location of

operations

Operating Hours

Operating Hours

CI Value

Sufficient data set Time series behavior of each CI

Ideal data setAdd maintenance and operational events

• Maintenance Actions

• Operational Profile Changes

• Operation Locations

Goal: Develop accurate prognostic

Requirement: Study Condition Indicator (CI) across lifetime of a component.

CI Value

22

Super condition indicator

• Individual CIs often may not have sufficient prognostic power

• Leverage Super CI to increase predictive resolution

Se

nsitiv

ity

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.00 0.20 0.40 0.60 0.80 1.00

1-Specificity

a

b

class

0.7209

0.7209

Area

Receiver Operating Characteristic

OBE

Se

nsitiv

ity

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.00 0.20 0.40 0.60 0.80 1.00

1-Specificity

a

b

class

0.5602

0.5608

Area

Receiver Operating Characteristic

Se

nsitiv

ity

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.00 0.20 0.40 0.60 0.80 1.00

1-Specificity

a

b

class

0.5218

0.5217

Area

Receiver Operating Characteristic

MDR IDA3

Se

nsitiv

ity

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.00 0.20 0.40 0.60 0.80 1.00

1-Specificity

a

b

class

0.5747

0.5747

Area

Receiver Operating Characteristic

ODA1

Se

nsitiv

ity

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.00 0.20 0.40 0.60 0.80 1.00

1-Specificity

a

b

class

0.5094

0.5099

Area

Receiver Operating Characteristic

OFM0

Se

nsitiv

ity

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.00 0.20 0.40 0.60 0.80 1.00

1-Specificity

a

b

class

0.6650

0.6649

Area

Receiver Operating Characteristic

IFM0

Individual CI’s

COMBINED SUPER CI

23

Determine the best fit

24

• Evaluate many prognostic algorithms to determine best fit

COMBINED SUPER CI

0

0.2

0.4

0.6

0.8

1

1.2

1.4

0100200300400500

CI V

alu

e

Hours to Removal

Degradation

Hours to Removal 200 175 150 125 100 75 50

LDA 35% 37% 57% 59% 71% 93% 100%

Gaussian Naïve Bayes 39% 42% 46% 53% 68% 89% 100%

K Neighbors 65% 57% 57% 58% 68% 82% 100%

QDA 37% 39% 43% 51% 65% 84% 100%

Linear Support Vectors 38% 38% 42% 47% 60% 77% 80%

Non-linear Support Vectors 29% 30% 34% 43% 55% 77% 88%

Logistic Regression 37% 37% 39% 43% 55% 70% 56%

Stochastic Gradient Descent 21% 15% 24% 34% 35% 59% 4%

Ridge Classifier 35% 34% 37% 43% 53% 73% 68%

Determine the best fit

25

Anomaly detection

• Identify distinct probability distributions for Super CI indications

• Produce failure lead time

COMBINED SUPER CI

Healthy

Component

Anomalous

Component

Anomalous

Component with

>100 hours

to causal removal

26

AbatingFalse positives • Analysis select wait time required to distinguish

between false positive and true anomaly

• Reduces negative impact on maintenance and supply

• Account for spikes and dips in the data

• Manage data quality with cleansing and transforation

• Determine the optimal time for maintenance

27

CBMresults

• Provide repair lead time– Reduce wait times for maintenance & parts– Optimize labor

• Avoid catastrophic failures

• Reduce logistics response time

• Control impact of failures on operations

• Extend asset life

• Minimize unplanned downtime

28

But What About…

• Parts with no sensors

• Long term strategies

• Impact on costs

• Inventory Optimization

• Future Performance metrics

29

But What About…

• Parts with no sensors

• Long term strategies

• Impact on costs

• Inventory Optimization

• Future Performance metrics

Asset Life Cycle Management

30

Life Cycle Management(LCM)

Benefits• Strategic approach to long-term asset planning

• Accurately managing future costs and expenses

• Maximizing uptime and revenue

Uses• Managing components and assets with and without sensors

• Accounting for changes in operations, upgrades, …

• Anticipating dynamic conditions & evolving environment

• Evaluating alternate future scenarios

Asset

Operations

Maintenance

Supply

Logistics

Sustainment

31

CBM & LCMbenefits

• Evaluate impact of full asset BOM

• Design maintenance strategies

• Integrate operational changes with maintenance planning

• Optimize future enterprise inventory• Control Costs

• Attain future business goals– Maximize up time & readiness– Control risk

• Budget• Operations

– Maximize Revenue

32

About us

33

Quick facts• Unique focus on capital

intensive physical assets

• Experience across multiple industries

• Software built out of service focus

• Over 30 years of experience

• Based in Austin, Texas; deployed worldwide

Predictive health management

Condition-based maintenance

Life Cycle Management

Performance based logistics

Data management

34

…and trusted by our clients(including but far from limited to…)

35

Turning data into information supporting decisions, processes and automation

Raw Data Aggregation

Data Cleansing & Consolidation

Predictive Model Design

Operationalizing Predictive Output

Automated Decision Making

Optimization

36

PlatformStudio Backbone

Life Cycle Management

Sensor Prognostics Design

Data Acquisition

Database access

Flat file

Big Data

Streaming Data

Data Quality & Transformation

ProductizedModels

Analytics Libraries

Machine Learning

Neural Networks

NLP and more

Data Visualization

Viz Libraries

Reports

Dashboards

Discrete Event Simulation

Simulation Engine

Asset models

Metric Aggregation

Analytics Software Platform

3737

Studio: data mgt and analytics

Supports data scientists and analysts with powerful tools for managing, analyzing and visualizing data

3838

LCM: modeling asset lifecycles

Represents

• Aging of assets

• Changes in operations

• Retirements & Acquisitions

• Repair degradation

• Risk and Uncertainty

Includes

• Deep asset indentures

• Asset age & condition initialized from raw data

• Detailed baseline and alternative cases

• Simulation output data that would otherwise not exist

• Future supply, sustainment and maintenance changes

• Best performance at least cost

Asset

Operations

Maintenance

Supply

Logistics

Sustainment

39

Business challenges

• Equipment maintenance is an extensive expense

• Multiple process control sensors with potential to detect impending issues before failure

• Long maintenance lead times lowering customer satisfaction

Solution

• Smoothing algorithm to the raw sensor data and maintenance history data

• Baseline anomaly detection prognostic by combining five sensor readings

• One month historical data to train algorithm

• Advance warning on impending failures– 5 months for heater wire– 2 months gerotor

ROI

• Predictive Heath Maintenance (PHM) prognostics

• Advance scheduling of maintenance before issues arise

case study: Industrial Packaging Machines

39

Global provider of high volume packing equipment

for shipping

40

Questions

41

up

Up next 12:30pm (US Eastern) 11:30am (US Central) 10:30am (US Mountain)

9:30am (US Pacific)

Panel discussionHost: Lucian Fogoros, IIoT WorldPanel: Benson Chan, Strategy of Things

Aaron Allsbrook, ClearBladeSerg Posadas, Clockwork Solutions

Ask a question!Use the chat tool or

tweet using #iiotvirtualconf

42

Ask a question!Use the chat tool or

tweet using #iiotvirtualconf

[email protected] [email protected] [email protected] [email protected]