predictive maintenance advanced solution

PDM

1. © 2 0 1 9 S P L U N K I N C .
© 2 0 1 9 S P L U N K I N C .
Implementing
Predictive
Maintenance
Young Cho, Sr IoT Practitioner
Sept 8, 2019 | Ver 1.0

2. During the course of this presentation, we may makeforward‐lookingstatements regarding
future events or plans of the company. We caution you that such statements reflect our
current expectations and estimates based on factors currently known to us and that actual
events or results may differ materially.Theforward-lookingstatements made in the this
presentation are being made as of the time and date of its live presentation. If reviewed after
its live presentation, it may not contain current or accurate information. We do not assume
any obligation to update anyforward‐lookingstatements made herein.
In addition, any information about our roadmap outlines our general product direction and is
subject to change at any time without notice. It is for informational purposes only, and shall
not be incorporated into any contract or other commitment.Splunk undertakes no obligation
either to develop the features or functionalities described or to include any such feature or
functionality in a future release.
Splunk, Splunk>, Turn Data Into Doing, The Engine for Machine Data, Splunk Cloud, Splunk
Light and SPL are trademarks and registered trademarks of Splunk Inc. in the United States
and other countries. All other brand names, product names, or trademarks belong to their
respective owners. © 2019 Splunk Inc. All rights reserved.
Forward-
Looking
Statements
© 2 0 1 9 S P L U N K I N C .

3. © 2 0 1 9 S P L U N K I N C .
Sr. IoT Practitioner | Splunk Inc
Young Cho
▶Firm believer of human creativity with data will
conquer all problems in the world.
▶Long computer and data career journey
−Unix system admin during BBS time
−PERL programmer –primarily for data parsing
−Data ware house design and development for Telcos
−Big data consultant and architect
−Splunk solutions architect, product marketing, IoT
practitioner
▶Amateur “Go” player
−Favorite documentary, Netflix AlphaGo –Story of Deepmind
challenging human “Go” champion

4. © 2 0 1 9 S P L U N K I N C .
Welcome to “Implementing Predictive Maintenance”
Why do we care about
Predictive Maintenance?
Intro to the solutions that
will make a big impact in
your org.
Get you pragmatic
analytics skills for
Predictive Maintenance.
Why? What? How?
Give you foundational knowledge to make you a PdMpractitioner.

5. © 2 0 1 9 S P L U N K I N C .
1. Introduction to
Predictive Maintenance
1

6. © 2 0 1 9 S P L U N K I N C .
1)MIT Sloan Review, “GE’s Big Bet on Data and Analytics”
2)ThomasNet, “Downtime Costs Auto Industry $22k/Minute”
3)TechTarget, “Predictive maintenance software points to machinery problems”
“Predictive Maintenance is the Holy Grail of Industrial IoT”
-Heather Ashton, manufacturing industry analyst at IDC
3
$25M / Day
Liquefied Natural Gas Platform
$7M / Day
Offshore Oil Platform
$1.3M / Hour
Auto Manufacturing
1 1 2
Equipment Downtime Costs Millions of $

7. © 2 0 1 9 S P L U N K I N C .
ML
Predictive
Maintenance
Condition-based
Maintenance
Preventative Maintenance
Reactive Maintenance
Scheduled and planned maintenance
based on usage and time
Fix when broken
The global process industry loses $20 billion annually from unplanned downtime*
*ARC Advisory Group
O
E
E
Proactive,
Strategic
Operations
Real-time analytics and sensing
insights to predict machine reliability
Rules-based logic for sensor data
Advanced machine learning and AI
driven maintenance
80% of Industrial Operations Are Reactive

8. © 2 0 1 9 S P L U N K I N C .
Current Maintenance Strategy / Methods
How the world is maintained.
Failure
Reactive Maintenance Preventive Maintenance
Scheduled
Predictive Maintenance
Optimal time
Per asset
Predicted Failure
•Maintaining when the asset
fails
•Example : Light bulb
•Applies to assets with minimal
failure impact
•Majority of operational asset
can’t not do this. Costly
downtime.
•Maintaining at the regular
schedule regardless of condition
•Example : Engine Oil Change
•Applies to majority of assets in
operation.
•High maintenance expense –time,
availability, materials
•Maintaining at the optimaltime
based on data and prediction
•Example : Impact YOU can make!
•Should be applied to all assets in
operation.
•High availability, cost savings,
organizational efficiency

9. © 2 0 1 9 S P L U N K I N C .
So, what is Predictive Maintenance?
Every asset has different life expectancy –Different operating environments
–Weather conditions, Types of workload, Different operating schedules & Frequency
Data is the key to insights –contains similarity in how things will behave
Key consideration element of Predictive Maintenance
Time / Age
Optimal time
Per asset
Predicted Failure Predicted Failure Predicted Failure
Condition Indicator
100
Cycles
65
Cycles
145
Cycles

10. © 2 0 1 9 S P L U N K I N C .
Intro to the Predictive Maintenance
business case
Splunk App –Designed for you to learn and experiment predictive maintenance
•Includes Jet Engines dataset from Nasa
•Key analytics techniques on how to understand and design your analysis model
•Step-by-step guides on using Splunk for predictive maintenance analysis + algorithm creation.
Jet engine use-case : Splunk Essentials for Predictive Maintenance App

11. © 2 0 1 9 S P L U N K I N C .
Predictive Maintenance And
Business Operations
Introduces challenges and opportunities

12. © 2 0 1 9 S P L U N K I N C .
Critical Business Implication
It’s not as simples as as we may think it could be
A Safety Threshold
•Fine balance between
efficiency and safety
•Each day of downtime could
cost up to $1 million USD for a
jumbo jetliner.

13. © 2 0 1 9 S P L U N K I N C .
Let’s look at what the data look like?
unit_cyclesname_Bleed_Enthalp
y
sname_Bypass_R
atio
sname_Corr_Cor
e_Speed
sname_Corr_Fan
_Speed
sname_Fuel_Flo
w_Ratio
sname_HPC_Outl
et_Temp
sname_HPT_Cool
ant_Bleed
sname_LPC_Outl
et_Temp
sname_LPT_Outl
et_Temp
sname_Phys_Cor
e_Speed
sname_Phys_Fan
_Speed
sname_Static_HP
C_Outlet_Pres
sname_Total_HP
C_Outlet_Pres
1 392 8.4195 8138.62 2388.02 521.66 1589.7 39.06 641.82 1400.6 9046.19 2388.06 47.47 554.36
2 392 8.4318 8131.49 2388.07 522.28 1591.82 39 642.15 1403.14 9044.07 2388.04 47.49 553.75
3 390 8.4178 8133.23 2388.03 522.42 1587.99 38.95 642.35 1404.2 9052.94 2388.08 47.27 554.26
4 392 8.3682 8133.83 2388.08 522.86 1582.79 38.88 642.35 1401.87 9049.48 2388.11 47.13 554.45
5 393 8.4294 8133.8 2388.04 522.19 1582.85 38.9 642.37 1406.22 9055.15 2388.06 47.28 554
6 391 8.4108 8132.85 2388.03 521.68 1584.47 38.98 642.1 1398.37 9049.68 2388.02 47.16 554.67
7 392 8.3974 8132.32 2388.03 522.32 1592.32 39.1 642.48 1397.77 9059.13 2388.02 47.36 554.34
8 391 8.4076 8131.07 2388.03 522.47 1582.96 38.97 642.56 1400.97 9040.8 2388 47.24 553.85
9 392 8.3728 8125.69 2388.05 521.79 1590.98 39.05 642.12 1394.8 9046.46 2388.05 47.29 553.69
10 393 8.4286 8129.38 2388.06 521.79 1591.24 38.95 641.71 1400.46 9051.7 2388.05 47.03 553.59
11 392 8.434 8140.58 2388.01 521.4 1581.75 38.94 642.28 1400.64 9049.61 2388.05 47.15 554.54
12 391 8.3938 8134.25 2388.02 521.8 1583.41 39.06 642.06 1400.15 9049.37 2388.09 47.18 554.52
13 393 8.4152 8128.1 2388.08 521.85 1582.19 38.93 643.07 1400.83 9046.82 2388.12 47.38 553.44
14 393 8.3964 8134.43 2388 521.67 1592.95 39.18 642.35 1399.16 9047.37 2388.09 47.44 554.48
15 391 8.4199 8127.56 2388.08 522.5 1583.82 38.99 642.43 1402.13 9052.22 2388.11 47.3 553.64
16 392 8.3936 8136.11 2388.07 521.49 1587.98 38.97 642.13 1404.5 9049.34 2388.05 47.24 553.94
17 392 8.4542 8137.27 2388.04 521.89 1584.96 38.81 642.58 1399.95 9054.92 2388.06 47.12 553.8
18 392 8.4028 8132.73 2388.07 521.76 1591.04 38.89 642.62 1396.12 9049.55 2388.05 47.21 554.2

14. © 2 0 1 9 S P L U N K I N C .
Let’s look at what the data look like?
Different assets with different degradation cycles
A full life-cycle of a engine
lasted 500+ cycles
A full life-cycle of another engine
Only lasted 135cycles

15. © 2 0 1 9 S P L U N K I N C .
What are we trying to accomplish?
Big difference in length of life

16. © 2 0 1 9 S P L U N K I N C .
Predictive Maintenance
Solution
Demonstration
1

17. © 2 0 1 9 S P L U N K I N C .
Demo

18. © 2 0 1 9 S P L U N K I N C .
Insert your own screenshot here.
For best results, use an image sized at 1450 x 850

19. © 2 0 1 9 S P L U N K I N C .
Insert your own screenshot here.
For best results, use an image sized at 1450 x 850

20. © 2 0 1 9 S P L U N K I N C .

21. © 2 0 1 9 S P L U N K I N C .
Predictive Maintenance
Development Process
0

22. © 2 0 1 9 S P L U N K I N C .
Predictive Maintenance Analytics Process
STAGE 1 : Data collection and ingestion focuses on how you can easily use Splunk software to collect, store, and structure
asset metrics (dataset from airplane jet engines included).
STAGE 2 : Data exploration covers key methods for pre-processing and exploring data to help you understand the type of
data in use and the characteristics of the dataset, which is crucial for getting the desired outcome.
Predictive maintenance analytics methodology:Essential predictive analytics knowledge
STAGE 3 : Analysis teaches you the 3
key analysis options (Anomaly Detection,
Unsupervised Learning, Supervised
Learning) when doing predictive
maintenance analysis.
STAGE 4 : Operationalization teaches
you how to apply the analytics model to a
broader implementation, and how to
create reports and alerts for operational
actions.

23. © 2 0 1 9 S P L U N K I N C .
1. Getting Data In

24. © 2 0 1 9 S P L U N K I N C .
Splunk for All of Enterprise Asset Data
Universal data platform for events and metrics

25. © 2 0 1 9 S P L U N K I N C .
How your industrial data is collected
Different protocols and collection methods supported

26. © 2 0 1 9 S P L U N K I N C .
Insert your own screenshot here.
For best results, use an image sized at 1450 x 850

27. © 2 0 1 9 S P L U N K I N C .
2. Explore Data /
Feature Engineering

28. © 2 0 1 9 S P L U N K I N C .
Understand how the asset operate
Turbo Jet Engines has 3 major components
Speed values
Temperature values
Flow / Bypass ratios
Bleed values
Pressure values

29. © 2 0 1 9 S P L U N K I N C .
Data Normalization Method
Prepares the data to create analytics model

30. © 2 0 1 9 S P L U N K I N C .
Event Windowing
Captures a full comparable operational cycle of an asset

31. © 2 0 1 9 S P L U N K I N C .
Condition Feature Selection
Understanding what information correlates with condition

32. © 2 0 1 9 S P L U N K I N C .
Feature Profile for an Asset
Visual inspection and understand of asset characteristics
Speed values
Temperature Values
Flow / Bypass ratios
Bleed values

33. © 2 0 1 9 S P L U N K I N C .
Feature Engineering
Different types of
condition indicators
•Time based features
•Frequency based
features
Creating new features from statistics –Best condition indicators

34. © 2 0 1 9 S P L U N K I N C .
Insert your own screenshot here.
For best results, use an image sized at 1450 x 850

35. © 2 0 1 9 S P L U N K I N C .
Insert your own screenshot here.
For best results, use an image sized at 1450 x 850

36. © 2 0 1 9 S P L U N K I N C .
3. Analysis

37. © 2 0 1 9 S P L U N K I N C .
Predictive Maintenance Analysis Domain
Two main approaches to Predictive Maintenance analysis
•Anomaly Detection
•Remaining Useful Life
•Unsupervised
Learning
•Supervised Learning
Statistical Analysis
Approach
Machine Learning
Approach

38. © 2 0 1 9 S P L U N K I N C .
Analysis Techniques & Approach
How to decide how to approach the problem

39. © 2 0 1 9 S P L U N K I N C .
1.Understand standard deviation
2.Value distribution analysis
3.Defining threshold based on standard
deviation
Statistical
Method
Anomaly
Detection

40. © 2 0 1 9 S P L U N K I N C .
Analysis –Anomaly Detection
Conditional value “Standard Deviation”

41. © 2 0 1 9 S P L U N K I N C .
Analysis –Anomaly Detection
Conditional values distribution

42. © 2 0 1 9 S P L U N K I N C .
Analysis –Anomaly Detection
Standard Deviation Thresholding

43. © 2 0 1 9 S P L U N K I N C .
Insert your own screenshot here.
For best results, use an image sized at 1450 x 850

44. © 2 0 1 9 S P L U N K I N C .
1.Asset degradation model
2.Calculating remaining useful life
3.Defining condition indicators
4.Optimizing remaining useful life using
similarity model
Statistical
Method
Remaining
Useful Life
Analysis

45. © 2 0 1 9 S P L U N K I N C .
Analysis –Remaining Useful Life
What is remaining Useful Life?Failure Condition
ASSET DEGRADATION MODEL
C
o
n
d
it
io
n

I
n
d
ic
a
t
o
r
Remaining Useful Life
Time / Age
Current Condition

46. © 2 0 1 9 S P L U N K I N C .
Analysis –Remaining Useful Life
How to calculate remaining useful life?
Failure Condition
Condition Indicator
Remaining Useful Life
Time / Age
Current Condition
ASSET DEGRADATION MODEL
•Known “End Of Life” distribution
•Calculate condition indicator

47. © 2 0 1 9 S P L U N K I N C .
Analysis –Remaining Useful Life
Calculating remaining useful life using similarity model
Failure Condition
Condition Indicator
Remaining Useful Life
Time / Age
Current Condition
ASSET DEGRADATION MODEL
•Known “End Of Life” distribution
•Calculate condition
indicator

48. © 2 0 1 9 S P L U N K I N C .
Data reduction and combination technique
Enhancing representation of condition indicators
Physical Fan Speed
LPC Outlet Temperature
Pre-processing Data
•Standard Scaling
•Average 2 features into 1
2
▶Condition Indicator Model3
Enhanced Condition Indicator
| eval condition_ind=
((SS_sname_LPC_Outlet_Temp+
SS_sname_Phys_Fan_Speed) / 2) * -1
▶Pick feature as a reduction1

49. © 2 0 1 9 S P L U N K I N C .
What is remaining useful life?
•Input as current cycle(Age) and
conditional indicator value
•Group similarhistorical models
(nearest)
•Statistics of known “End of Life” –
meanof known eol.
How to calculate the remaining life based on historical similarity model
Asset Degradation Model
Failure Condition
Condition Indicator
Remaining Useful Life
Time / Age
Current Condition
Expected End Of Life
200 270
End Of Life
Range
235

50. © 2 0 1 9 S P L U N K I N C .
Insert your own screenshot here.
For best results, use an image sized at 1450 x 850

51. © 2 0 1 9 S P L U N K I N C .
1.Principal Component Analysis
2.Clustering
3.Various visual data explorations
Machine Learning
Approach
Unsupervised
Machine
Learning

52. © 2 0 1 9 S P L U N K I N C .
Analysis –Unsupervised Machine Learning

53. © 2 0 1 9 S P L U N K I N C .
Analysis –Unsupervised Machine Learning

54. © 2 0 1 9 S P L U N K I N C .
Analysis –Unsupervised Machine Learning

55. © 2 0 1 9 S P L U N K I N C .
Insert your own screenshot here.
For best results, use an image sized at 1450 x 850

56. © 2 0 1 9 S P L U N K I N C .
1.Convey meaning and be inspirational with
your message when possible
2.Use powerful imagery to support
your point
3.Use animation to support your message,
not just to entertain the audience
Machine Learning
Approach
Supervised
Learning

57. © 2 0 1 9 S P L U N K I N C .
Analysis –Supervised Machine Learning

58. © 2 0 1 9 S P L U N K I N C .
Analysis –Supervised Machine Learning

59. © 2 0 1 9 S P L U N K I N C .
Analysis –Supervised Machine Learning
Supervised learning analytics process components

60. © 2 0 1 9 S P L U N K I N C .
Analysis –Supervised Machine Learning
Applying training the “Good” model approach
When used?:
•Training a “Good” model, when there are other variations (Often many) of patterns not well known.
•So, anything other than a very specific “good” pattern as bad. Pick out rest of non-good pattern as bad.

61. © 2 0 1 9 S P L U N K I N C .
Analysis –Supervised Machine Learning
Applying training the “Good” model approach

62. © 2 0 1 9 S P L U N K I N C .
Analysis –Supervised Machine Learning
Applying training the “Bad” model approach
When used? :
•Training a “Bad” model, when a bad pattern is certain and known, also limited “bad” possibility.
•So, pick specific bad pattern from the data where there maybe multiple “good” patterns.

63. © 2 0 1 9 S P L U N K I N C .
Analysis –Supervised Machine Learning
Applying training the “Bad” model approach

64. © 2 0 1 9 S P L U N K I N C .
When used? :
•Training a model with multiple conditions as different labels. (0,1,2,3)
•Define multiple different patterns based on severity or different symptom.
Analysis –Supervised Machine Learning
Examples :
•Based on condition
(Good →Warning →Bad)
•Based on different symptom
(Compressor failure vs
fan failure)

65. © 2 0 1 9 S P L U N K I N C .
Analysis –Supervised Machine Learning

66. © 2 0 1 9 S P L U N K I N C .
Analysis –Supervised Machine Learning
Define the range of time series data, then label “state=” with numeric value.
Training a multiple conditions example

67. © 2 0 1 9 S P L U N K I N C .
Analysis –Supervised Machine Learning
Training a multiple conditions pre-training model labeling

68. © 2 0 1 9 S P L U N K I N C .
Analysis –Supervised Machine Learning
Different machine learning algorithm for trial and error

69. © 2 0 1 9 S P L U N K I N C .
Analysis –Supervised Machine Learning
Analyzing accuracy and precision of trained machine learning model
Accuracy and precision verification
Confusion Matrix

70. © 2 0 1 9 S P L U N K I N C .

71. © 2 0 1 9 S P L U N K I N C .
4. Apply

72. © 2 0 1 9 S P L U N K I N C .
Apply –Dashboard & Reports
Immediately operationalize Splunk predictive maintenance
Creating Dashboard : Predictive
Maintenance Dashboard
•Use the analyzed prediction, create dashboard to
show the status of engines in different state.
•Use Splunk Dashboard Examples App to
customize visualization and various user inputs
and controls.
Creating Reports : Using Splunk
Enterprise
•Watch training video on "How to create reports in
Splunk"
•Splunk documentation on "How to create reports"

73. © 2 0 1 9 S P L U N K I N C .
Apply –Generate Alerts &
Connected Experience
Notify your maintenance crew in real-time
Creating Alerts in Splunk
•Watch training video on "How to create alerts in
Splunk"
•Splunk documentation on "How to define alerts”
Splunk Connected Experience :
•Launch Splunk Mobile to empower your field
teams –to access analytics and alerts in mobile
phones
•Augmented reality to make analytics available to
all field crew in a easy to see augmented reality on
top of your assets.

74. © 2 0 1 9 S P L U N K I N C .
Apply –Automate and Control
Automate and Orchestrate your operations•Process multiple conditions
•Apply logic to decide
•Even involve human
Confirmation
•Interact with any assets with APIs
•Fully automated
response to intergrate
Automation using Phantom
•Define full automation using Phantom and integrate with your legacy assets

75. © 2 0 1 9 S P L U N K I N C .
5. Wrap Up

76. © 2 0 1 9 S P L U N K I N C .
Let’s get you hands dirty!
Methodology Tools / Technology Passion
This is where the subtitle goes
Now you know key
analytics techniques to
accomplish predictive
maintenance
Splunk Essentials for
Predictive Maintenance
+ ML Toolkit
+ Basic Statistics
It’s time for a big impact
-reach out to all the
resource –Vertical Team
Specialists + Practitioners

77. RATE THIS SESSION
Go to the .conf19 mobile app to
© 2 0 1 9 S P L U N K I N C .
You!
Thank