Amar Pradhan: 2013 Sandia National Laboratoies Wind Plant Reliability Workshop
- 1. Use of Data to
Reduce Wind
Energy Costs
Machine to Machine Learning
for the Wind Industry
- 3. We are a machine to machine learning company
focused on improving performance
© 2011 Fluitec International. All rights reserved.
Fluitec Offices as Viewed from Space
USA
Belgium
China
Singapore
- 4. Awarded $3.3M
through the New
Jersey Clean Energy
Manufacturing Fund to
accelerate adoption of
Technology to Reduce
the Cost of Wind
Energy. Matched by
leading European
Cleantech VCs.
- 5. An award winning
company and a stellar
team with unparalleled
expertise.
“Most Promising
Innovation”
“People’s Choice
Award”
- 6. Fluitec Wind currently monitors
>5,000 turbines: 8 GW
Vestas V82, GE 1.5, & Acciona 1.5
turbines are best represented
Fluitec Wind has the Largest Aggregated Database
of Diagnostic & Operational Data in the World
Acciona, 34%
Vestas, 39%
GE, 13%
Suzlon, 4%
Gamesa, 5%
Mitsubishi, 2%
Nordex, 3%
- 8. Wind turbines have high variability
and are expensive to access.
Therefore significant remote
monitoring capabilities exist.
- 10. >60% of Unscheduled Downtime is on Nacelle
Rotating Components.
>40% of total downtime is unscheduled
>80% of downtime on lubricated components is unscheduled
Problem
Component
TOTAL Downtime
Events
Total Downtime
(days)
Unscheduled
Downtime
(days)
Unscheduled
Proportion
Lubricated
Component
Downtime
(days)
Lubricated
Component
Unscheduled
Downtime
- - Yes No
Gearbox 2300 802 1498 1,750 1,500 86% 1,750 1,500
Generator 1965 1651 314 834 595 71% 834 595
Yaw 603 272 331 253 216 85% 253 216
Hydraulic System 37 3 34 6 0 0% 6 0
Pitch Blade 956 564 392 547 389 71% -
Rotor 306 288 18 80 71 89% -
Other/Electrical 23988 3780 20208 5,264 1,039 20% -
Grid 430 6 424 37 1 3% -
Anemometer 12 7 5 2 2 100% -
30,598 7,374 23,224 8,773 3,811 43% 2,844 2,311
Unscheduled
Events
- 12. M2M
Analytics
Work Orders
Oil Analysis Data
Operational Configuration
Weather Data
Existing
Existing Existing
Existing Existing
SCADA Alerts & Sensors
You have existing data with immediate predictive
value. We can utilize this without any capital expense
to identify and reduce risks.
- 14. Genetic:
Equipment
Permutation
Location
>20 Input Attributes
Operational:
SCADA
O&M Logs
Insurance Claims
>1000 Outputs
We aggregate your data to enhance signal to noise
detection. Allowing for noisy, small datasets to be
utilized. We can corroborate the reliability of any data
point and use. We also report this quality so you can
improve in the future.
Diagnostic:
Sensors
Production
Weather
Oil Analysis
Vibration
>500 Input Attributes
- 15. Oil Analysis is the deepest and
widest diagnostic data set and
perfect for such analyses if unlocked.
We utilize “fingerprinting” technology on this big
data. Further minimizing the effect of outliers or poor
data. Similar to how doctors use blood analysis, and
the police use criminal data.
- 16. We aggregate global data to have a map of good and
poor states of various turbine permutations. Fleets of
various ages and models become predictable.
Current database size
is 8 GW, with an avg
farm age of 5 years
- 17. Allows an accurate method
to identify turbines that are
following a specific failure
mode or pattern. Has a
proven 95% accuracy in
gearbox failure prediction.
Finally, we match “fingerprints” to poor states, not
simply identifying “abnormal” turbines. This drastically
increases the predictive value of data. Police catch
criminals by matching to known offenders.
NN
- 18. N
Three Algorithmic Steps to Creating a Predictive Map
1. Similarity
Define the similarity
between each point
in time to every other
point in time
2. Cluster
Cluster the points in
time based on their
similarity
3. Severity
Asses the severity of
each cluster
- 20. In the following we will focus on the limitations of
individual attribute alarms, and discuss how we
develop multi-attribute alarms. The discussion is
centered on gearbox oil analysis, as it provides a large
number of attributes to consider, and the individual
attribute approach is particularly flawed. However, the
fundamentals of our recommendations should be
exercised on all turbine level data: temperatures,
speeds, direction, etc.
Disclaimer
- 21. Copper Iron Silicon
50 100% 100 99% 60 99%
30 99% 50 96% 45 96%
1 25% 1 16% 30 85%
Visc40 Oil Age PQ Index
384 99% 1691 90% 15 91%
320 61% 715 50% 8 50%
256 20% 178 10% 1 21%
Individual Attribute Alarms Are Not Working &
Are Not Predictive
Within an analysis of 25,000
samples: ~99% of the time, the
individual values received are
below the critical limits.
Average Visc40 and Iron prior
to gearbox failure is 318 and
11, respectively.
Individual attribute alarms are
inherently less predictive.
Percent of Oil Sample Data Below
Adjacent Value
- 22. Include:
1. Genetic Attributes
2. Multiple Attributes to define an Alarm Band
3. Rate of Change Attributes
Ideally use all of the above
Three ways to make Better Alarm Levels
- 23. Use Genetic Attributes to define Bands
Simply looking at
attributes in the
context of the
equipment
permutation gives
a clear picture of
what are normal
levels. Especially
Ingression and
Wear elements
- 25. 1. In the vast majority of instances just prior to failure,
oil attributes were within the “acceptable range”
provided by OEMs.
2. The pronounced difference in the profile prior to
failure, versus in general can be seen via multi-
attribute bands.
3. Rate of Change thresholds are more effective in
highly variable environments.
Case Study Summary
- 27. Raw
Unstructured
Data: equipment
model/year,
SCADA alerts,
production data,
oil analysis
Usable
Data
6-10 weeks
How to Get Started: Send Us Raw Data & We Provide Deliverables 1-3
Data
Cleaned &
Structured:
Returned in
any format
Analytical
Reports
Expert Risk
Assessment
Provided as
Report
Web portal:
visualization,
analysis, and
dynamic work
order toolkit
Provide Raw Data: best results are if
sample set has 2,000 turbine-years
of data, high failure rates, and/or use
of popular equipment permutation
(Vestas V82, GE 1.5, AW1500)