ibm predictive analytics iot presentation

Post on 10-May-2015

772 Views

Category:

Software

5 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Predictive Analytics for IoT

Michael Adendorff Architect, STSM IBM

IBM Predictive Maintenance and Quality michael.adendorff@ca.ibm.com

Evidence , Clues > Failure Prediction

Predictive Analytics

Valuable Insight

Maintenance Insight

Maintenance Insight

Failure Risk: Under Maintained Equipment

Maintenance Insight

Wasted $$$$$: Over maintained equipment

Work Order

Urgent Inspection Required: High probability of failure

High risk of failure before next scheduled maintenance

Maintenance Schedule Update Request

Bring forward scheduled maintenance to Jul 5

Bring forward scheduled maintenance to Aug 7

Delay scheduled maintenance to Dec 15

Parts Requirements Forecast : Main Bearing

June: July: Aug:

10 3 22

9 20 12

7 21 14

12 15 17

Business Results : Predictive Maintenance

Downtime

Unplanned

Planned

Predictive Analytics

Valuable Insight

How does it work?

Simplistic Illustration

Historic Data

Failure Records

Vibration Levels

Correlation

Failu

re C

ou

nt

Vibration Level

More failures have been witnessed when vibration levels are high

Univariate Model

Failu

re C

ou

nt

Vibration Level

Vibration Level

P Failure Confidence

< 0.1 0.1 % 2%

0.-0.5 1% 3%

0.5 – 2 3% 5%

2 – 5 15% 10%

+5 98% 80%

p(fail)

Simple univariate models are generally not very accurate. This one looks better than it is. High vibration strongly correlated with failure as it is a lagging indicator. Need leading indicators to predict.

Multivariate model

p(fail)

More accurate than the univariate model, but raw input data never reveals the whole story.

Correlates failures with combinations between multiple input variables

Historic Data

Advanced Data Prep + Ensemble Models

More accurate than the univariate model, but raw input data never reveals the whole story.

Historic Data

p(fail)

E(fail date)

Advanced Data Prep + Ensemble Models

More accurate than the univariate model, but raw input data never reveals the whole story.

Historic Data

Cumulative Cycles = f(speed, operating hours)

p(fail)

E(fail date)

Advanced Data Prep + Ensemble Models

More accurate than the univariate model, but raw input data never reveals the whole story.

Historic Data

Cumulative Fatigue Load = f(Cycles, Speed)

p(fail)

E(fail date)

Advanced Data Prep + Ensemble Models

More accurate than the univariate model, but raw input data never reveals the whole story.

Historic Data

Wear Damage Forecast

p(fail)

E(fail date)

Advanced Data Prep + Ensemble Models

p(fail)

More accurate than the univariate model, but raw input data never reveals the whole story.

Historic Data

Wear Damage Forecast

E(fail date)

Wear Modeling

Advanced Data Prep + Ensemble Models

More accurate than the univariate model, but raw input data never reveals the whole story.

Historic Data

Fatigue Damage Forecast

p(fail)

E(fail date)

Advanced Data Prep + Ensemble Models

p(fail)

More accurate than the univariate model, but raw input data never reveals the whole story.

Historic Data

Wear Damage Forecast

E(fail date)

Fatigue Modeling

Advanced Data Prep + Ensemble Models

Building models like this requires brute force number crunching as well as skills and knowledge. Payoff comes from more accurate predictions – but – it doesn’t end here.

Historic Data

Time series forecast + Combination Model

p(fail)

E(fail date)

Advanced Data Prep + Ensemble Models

Historic Data

Expected failure date is more actionable than current probability of failure

Building models like this requires brute force number crunching as well as skills and knowledge. Payoff comes from more accurate predictions – but – it doesn’t end here.

p(fail)

E(fail date)

Advanced Data Prep + Ensemble Models

Historic Data

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Ut lacinia semper gravida. Morbi vel orci in leo malesuada malesuada in ac enim. Nam pulvinar nec enim in venenatis. In nibh turpis, sodales at fermentum in

Sensors don’t record every causal factor. Text analytics is used to fill in some of the blanks.

p(fail)

E(fail date)

Predictive Analytics

Valuable Insight

Building models is only half the fun. Next step – OPERATIONALIZE

Feed Data

APIs for: • Describing target data structures • Describing calculations and aggregations • Running analytics • Exposing analytic results

REST Historian DB

Web

Se

rvic

e

MQTT Other

Data flows into DB in realtime

Event

Master Data

Profile

KPI

Predictive Analytics done in realtime

Event

Master Data

Profile

KPI

p(fail)

E(fail date)

Predictive Analytics done in realtime

Event

Master Data

Profile

KPI

p(fail)

E(fail date)

Predictive Outputs fed back as new events

Deciding on Recommended Actions

Event

Profile Action

KPI

Taking Action

REST DB

Web

Ser

vice

FTP Other

Valuable Insight

Build Models 1) Assemble historic data 2) Attempt to correlate historical data with a

known target 3) Improve results by putting more thought

about preparing inputs and algorithm selection

Operationalize 1) Feed raw data 2) Describe calculation and aggregation 3) Perform analytics 4) Carry out decision logic 5) Feed results 6) Retrain models regularly

Questions?

Michael Adendorff Architect, STSM IBM

IBM Predictive Maintenance and Quality michael.adendorff@ca.ibm.com

top related