Predictive Maintenance with

MATLAB

Amit Doshi,
Senior Application Engineer – Data Analytics
MathWorks India
[email protected] © 2015 The MathWorks, Inc.
1
Agenda:

1. What is Predictive Maintenance? Who is benefiting by doing it?

2. How can you develop a predictive maintenance algorithm using MATLAB?

3. How can you get started quickly?

3
 Reactive maintenance

Types of maintenance Preventive maintenance

Predictive maintenance

4
 Reactive maintenance
Machine
health

x
Failure
Time

5
 Reactive maintenance
Machine
health

x
Failure
Time

Aircraft
engine

6
 Reactive maintenance
Machine
health

x
Failure
Time

Preventive maintenance
Machine
health
?
?
Time

7
 Reactive maintenance
Machine
health

x
Failure
Time

Preventive maintenance
Machine
health

x Still usable
condition

Time

8
 Reactive maintenance
Machine
health

x
Failure
Time

Preventive maintenance
Machine
health

x Still usable
condition

Time

Predictive maintenance
Machine
health
Optimum time to
now do maintenance
x Predicted
failure
Time 9
A Predictive Maintenance Algorithm Answers
These Questions
For example:
Is my machine
operating
Anomaly
Detection
I need help.
normally?

Why is my
machine behaving
Condition One of my cylinders is blocked.
Monitoring
abnormally?

How much longer Remaining

can I operate my Useful Life I will shut down your line in 15 hours.
machine ? Estimation

10
 Predictive Maintenance Success Stories

Pump Health Monitoring System Online engine health monitoring Production machinery failure warning
▪ Spectral analysis and filtering on binary sensor data ▪ Real-time analytics integrated with enterprise ▪ Reduce waste and machine downtime
and neural network model prediction service systems
▪ MATLAB based HMI warns operators of
▪ More than $10 million projected savings ▪ Predict sub-system performance (oil, fuel, potential failures
liftoff, mechanical health, controls
▪ > 200,000 € savings per year

11
Link to user story

12
 Transocean uses MATLAB tools to transition from preventative
maintenance to CPM* for a critical deep sea drilling component
*Condition and Performance Monitoring

Challenges
• Minimize unplanned downtime of the component
• Use as-close-to-real-time data for CPM
• Deploy CPM solution to components in other locations

Solution
▪ Monitoring data from the drill is collected in a PI archive. MATLAB is used
Drilling data is stored in an OSISoft PI Archive. MATLAB
to create an app that takes in the data, preprocesses the data, and Production Server is used in the application specific
generates a quality indicator. processing

▪ The app is deployed onto MATLAB Production Server for real-time use;
Interfacing directly with data in PI gives you increased
results are stored back on the PI archive. performance at the cost of having to write/deal with
low-level code (“getting under the hood”). It was noted
Results that using MATLAB parallel tools (which entails a
higher-level interface) was not only easier but also it
Able to move to CPM for their component(s) with improved decision-making
provided “hands-down” faster performance than
capabilities through faster access to data and quicker analytics deployment. interfacing with PI directly.

Link to Expo Talk 13

Predictive Maintenance Toolbox for Developing Algorithms

Is my machine
Anomaly
operating
Detection
normally?

Why is my
Condition
machine behaving
Monitoring
abnormally?

How much longer Remaining

can I operate my Useful Life
machine ? Estimation

14
Workflow for Developing a Predictive Maintenance Algorithm

Machine Learning

Acquire Preprocess Identify Train Deploy &

Data Data Features Model Integrate

15
Agenda:

1. What is Predictive Maintenance? Who is benefiting by doing it?

2. How can you develop a predictive maintenance algorithm using MATLAB?

3. How can you get started quickly?

16
 Develop Predictive Maintenance Algorithm: Use cases

Is my machine
Anomaly
operating
Detection
normally?

Why is my
Condition
machine behaving
Monitoring
abnormally?

How much longer Remaining

can I operate my Useful Life
machine ? Estimation

17
Develop Predictive Maintenance Algorithm for -
Use case 1: Fault Classification
Why is my
Condition
machine behaving
Monitoring
Condition indicators abnormally?

Identify
fault type
Predictive
maintenance
algorithm
Determine what
needs to be fixed

18
Develop Predictive Maintenance Algorithm for -

Use case 2: Remaining useful life

How much longer Remaining
can I operate my Useful Life
machine ? Estimation

Condition indicators
Current
condition
Machine x
health
Estimate
Predictive remaining useful life
maintenance x
algorithm x x

Time

Remaining useful life

19
20
 Predictive Maintenance Algorithm
Acquire Data

Sensor Data Identify

Preprocess Condition Train Deploy &
Data Indicators Model Integrate
Generated
Data

Sensor
healthy Collect data using sensors data
+
faulty
Temperature
Flow Time
Triplex pump
Pressure
*For simplification purposes, healthy and faulty
operation are represented by single
measurements. In a realistic scenario, there
may be hundreds of measurements for both
types of operations.

21
 Predictive Maintenance Algorithm
Identify
Acquire Preprocess Condition Train Deploy &
Data Data Indicators Model Integrate

Sensor
healthy Collect data using sensors data
+
faulty
Temperature
Flow Time
Triplex pump
Pressure

*Quality data->Robust Algorithms

22
 Predictive Maintenance Algorithm
Identify
Acquire Preprocess Condition Train Deploy &
Data Data Indicators Model Integrate

Sensor
healthy Collect data using sensors data
+
faulty
Temperature
Flow Time
Triplex pump
Pressure

23
 Predictive Maintenance Algorithm
Identify
Acquire Preprocess Condition Train Deploy &
Data Data Indicators Model Integrate

Operating temperature: 30°F

Fluid viscosity: Low

Sensor
healthy Collect data using sensors data
+
faulty
Temperature
Flow Time
Triplex pump
Pressure Operating temperature: 90°F
Fluid viscosity: High

24
 Predictive Maintenance Algorithm
Identify
Acquire Preprocess Condition Train Deploy &
Data Data Indicators Model Integrate

Sensor
data

Time
Triplex pump
What if real failure data
is not available?

25
 Predictive Maintenance Algorithm
Identify
Acquire Preprocess Condition Train Deploy &
Data Data Indicators Model Integrate

Sensor
data

Time
Triplex pump
Sensor data

Refine model Synthetic data

Time
1
3
1Sealleakage Inject faults 2
2Blocked inlet
3Worn bearing Mathematical model Time
of the triplex pump
(Digital Twin)
26
 Health condition (fault code)

• Healthy (0)
• Blocked inlet (010)
• Worn bearing (001)
• Seal Leakage (100)
Flow Pressure • Blocked inlet, worn bearing (011)
• Seal leakage, worn bearing (101)
• Seal leakage, blocked inlet (110)
• Seal leakage, blocked inlet, worn be

240 measurements of
flow and pressure with a
duration of 1.2 seconds

Learn further: Data Ensembles 27

 Predictive Maintenance Algorithm

Acquire Data

Sensor Data Identify

Preprocess Condition Train Deploy &
Data Indicators Model Integrate
Generated
Data

Challenges

▪ Data clean up
– Poorly formatted files
– Irregularly sampled data
– Redundant data, outliers, missing data etc.

▪ Data specific processing

– Signals: Smoothing, resampling, denoising, Wavelet transforms, etc.
– Images: Image registration, morphological filtering, deblurring, etc.

28
 Spikes to sensor’s
maximum value
Raw data
10

9.5
Pressure (bar)

8.5

7.5

7 Offset in time
6.5
0.5 1 1.5 2 2.5 3 3.5 4

Time (s) 29
30
 Preprocessed data
Raw data 7.35
10

7.3
9.5

7.25
Pressure (bar)

Pressure (bar)
9

7.2
8.5
7.15
8
7.1
7.5
7.05

7
7

6.5
0.5 1 1.5 2 2.5 3 3.5 4 6.95
0 0.2 0.4 0.6 0.8 1 1.2
Time (s) Time (s)

31
 Predictive Maintenance Algorithm

Acquire Data

Sensor Data Identify

Preprocess Condition Train Deploy &
Data Indicators Model Integrate
Generated
Data

• What are Condition Indicators?

• Why are they important?
• How to identify significant ones?

32
A condition indicator can be any feature that is useful -
• for distinguishing normal from faulty operation or
• for predicting remaining useful life

Raw data Condition indicator

Healthy Healthy
operation operation
Faulty operation Faulty operation

Time Time

Identify condition indicators

33
 Signal-Based Condition Indicators

Time-domain Frequency-domain Time-frequency

features features domain features

Mean
Standard deviation
Skewness
Root-mean square
Kurtosis
...

34
 Healthy Seal Leakage Seal Leakage, Blocked Inlet
Faulty 7.3 7.3

Pressure (bar)
Pressure (bar)
Increasing 7.25 7.25

fault severity 7.2 7.2

7.15 7.15

7.1 7.1

7.05 7.05

7 7
0.62 0.63 0.64 0.65 0.66 0.67 0.68 0.69 0.7 0.71 0.72 0.62 0.63 0.64 0.65 0.66 0.67 0.68 0.69 0.7 0.71 0.72

Blocked Inlet Seal Leakage, Worn Bearing

7.3 7.3
Pressure (bar)

Pressure (bar)
7.25 7.25

7.2 7.2

7.15 7.15

7.1 7.1

7.05 7.05

7 7
0.62 0.63 0.64 0.65 0.66 0.67 0.68 0.69 0.7 0.71 0.72 0.62 0.63 0.64 0.65 0.66 0.67 0.68 0.69 0.7 0.71 0.72

Worn Bearing Blocked Inlet, Worn Bearing

7.3 7.3
Pressure (bar)

Pressure (bar)
7.25 7.25

7.2 7.2

7.15 7.15

7.1 7.1

7.05 7.05

7 7
0.62 0.63 0.64 0.65 0.66 0.67 0.68 0.69 0.7 0.71 0.72 0.62 0.63 0.64 0.65 0.66 0.67 0.68 0.69 0.7 0.71 0.72

Time (s) Time (s)

35
 MEAN

Blocked Inlet 7.25

7.3
Pressure (bar)

7.25
7.245

7.2 No overlap
7.24
Minimum
7.15

7.1

7.05 7.235 25th percentile

7
0.62 0.63 0.64 0.65 0.66 0.67 0.68 0.69 0.7 0.71 0.72 Median
7.23

75th percentile
7.225

Maximum
Blocked Inlet Healthy

36
 MEAN

7.25

7.245

Healthy
7.24

Blocked Inlet

Seal Leakage
7.235

Worn Bearing
7.23

Overlap Blocked Inlet &

Worn Bearing
7.225

Seal Leakage
&
7.22 Worn Bearing
Seal Leakage
&
7.215 Blocked Inlet

7.21

Fault type
37
 MEAN 10-3 VARIANCE
7.25
6 Outlier
7.245
5
7.24

7.235 4

7.23
3 Healthy
7.225

7.22 2 Blocked Inlet

7.215
1
Seal
7.21 Leakage
Fault type Fault type Worn Bearing

Blocked Inlet &

SKEWNESS KURTOSIS Worn Bearing
0
Outlier
-0.2 7 Seal Leakage
-0.4 &
-0.6 Worn Bearing
6
Seal Leakage
-0.8 &
-1 5 Blocked Inlet
-1.2
-1.4 4

-1.6
3
-1.8

-2
2
Fault type Fault type 38
 Healthy

Blocked Inlet
Variance

Seal Leakage
Worn Bearing

Blocked Inlet, Worn Bearing

Seal Leakage, Worn Bearing
Seal Leakage, Blocked Inlet

Mean

39
 Signal-Based Condition Indicators

Time-domain Frequency-domain Time-frequency

features features domain features

Mean
Standard deviation
Skewness
Root-mean square
Kurtosis
...

Learn more about Condition Indicators

41
Machine with rotating
components

Amplitude
Disc • In time-domain, we observe
Bearing the combined effect of
different sources of vibration.
Time

Motor
shaft

Three different Bearing Motor Disc

Amplitude
vibration sources: shaft • Using frequency-domain
- Bearing Nominal
analysis, we can
- Motor shaft Faulty distinguish different
- Disc Frequency sources of vibration.

42
 Seal Leakage Seal Leakage, Blocked Inlet
0.02 0.02

Magnitude
0.015 0.015

0.01 0.01

0.005 0.005

0 0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000

Blocked Inlet Seal Leakage, Worn Bearing

0.02 0.02
Frequency-domain
features:
Magnitude

0.015 0.015

0.01 0.01
• Peaks
0.005 0.005 • Peak frequencies
0 0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000

Worn Bearing Blocked Inlet, Worn Bearing

0.02 0.02
Magnitude

0.015 0.015

0.01 0.01

0.005 0.005

0 0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000

Frequency (Cycles/min) Frequency (Cycles/min)

43
Time-domain Frequency-domain

44
 Signal-Based Condition Indicators

Time-domain Frequency-domain Time-frequency

features features domain features

Mean Power bandwidth Spectral entropy

Standard deviation Mean frequency Spectral kurtosis
Skewness Peak values

...
Root-mean square Peak frequencies
Kurtosis Harmonics

...
...

Learn more about Condition Indicators

46
Diagnostic Feature Designer App
Predictive Maintenance Toolbox R2019a

▪ Extract, visualize, and rank

features from sensor data

▪ Use both statistical and

dynamic modeling methods

▪ Work with out-of-memory data

▪ Explore and discover

techniques without writing
MATLAB code

47
48
 Predictive Maintenance Algorithm

Acquire Data

Sensor Data Identify

Preprocess Condition Train Deploy &
Data Indicators Model Integrate
Generated
Data

49
Machine Learning Workflow

Machine learning uses data and produces a program to perform a task

Train: Iterate till you find the best model using historical data

HISTORICAL PREPROCESS SUPERVISED MODEL

DATA DATA LEARNING

FILTERS PCA CLASSIFICATION

SUMMARY CLUSTER REGRESSION

STATISTICS ANALYSIS

Predict: Integrate trained models into applications

NEW PREDICTION
DATA

50
Machine Learning: Types

Type of Learning Categories of Algorithms

Regression
Supervised
Learning
Classification
Develop predictive
Condition Machine
model based on both
Indicator Learning input and output data

Unsupervised
Learning

Group and interpret

data based only
on input data
51
52
Video showing App in action

53
Summary: Develop Predictive Maintenance Algorithm:

Use case 1: Fault Classification

Diagnostic Feature
Designer

Classification Learner App 54

Develop Predictive Maintenance Algorithm: Use case 2

How much longer Remaining

can I operate my Useful Life
machine ? Estimation

Condition indicators
Current
condition
Machine x
health
Estimate
Predictive remaining useful life
maintenance x
algorithm x x

Time

Remaining useful life

55
 Machine deterioration
profile
Condition indicator
Current condition
Failure condition

What is RUL?

Time
Remaining useful life (RUL)

[ Number of days ]
[ Miles ]
[ Cycles ]
…

56
 RUL Estimator Models

Similarity model Degradation model

Safety
threshold

Healthy state Failure

Survival model
Healthy state Failure

Healthy state Failure

57
RUL Methods and when to use them
Requirement: Need to know what constitutes failure data

58
59
 Predictive Maintenance Algorithm
Identify
Acquire Preprocess Condition Train Deploy &
Data Data Indicators Model Integrate

Cloud
Edge device

60
Deploy & Integrate analytics using MATLAB:

Embedded Hardware Enterprise Systems

Standalone Excel Hadoop/ MATLAB

C, C++ HDL PLC Application C/C++ Java ++ Python .NET Production
Add-in Spark Server

MATLAB
Runtime
61
Feature Extraction Algorithm at the Edge

Pump flow sensor 1 sec ~ 1000 samples ~16kB

▪ 1 day ~ 1.3 GB

▪ 20 sensors/pump ~26 GB/day

▪ 3 pumps ~ 78 GB/day
Asset w. smart
sensors

Challenge:
Data transmission cost is pretty high

Solution:
Extract only relevant information and send it to predictive model

62
Video showing Codegen with MATLAB Coder

63
Video showing Codegen with MATLAB Coder

64
Code Deployment for Machine Learning
Deploy trained models as standalone C/C++ code

▪ SVM Classification
▪ Linear Classification
▪ Linear Regression
▪ Generalized Linear Regr.
▪ Decision trees
▪ Ensembles for Class.
▪ Ensembles for Regr.
▪ SVM Regression
▪ KNN Classification
▪ Gaussian Process Regr.
▪ Discriminant Analysis

▪ Non-tree Ensembles
▪ KNN with kd-tree

65
What do your end users want?

Flexible Deployment AWS

Kinesis
▪ Maintenance needs simple, quick Azure
IoT Hub
information
– Hand held devices, Alarms Azure Azure Streaming Data
Blob SQL

▪ Operations needs a birds-eye view

– Integration with IT & OT systems
Data Sources Analytics Platforms
Dashboards
Fleet & &
Inventory Analysis
▪ Customers expect easy to digest Hand held Devices
information
– Automated reports

66
Agenda:

1. What is Predictive Maintenance? Who is benefiting by doing it?

2. How can you develop a predictive maintenance algorithm using MATLAB?

3. How can you get started quickly?

67
MathWorks can help you get started TODAY

▪ Examples

▪ Documentation

▪ Tutorials & Workshops

▪ Consulting

▪ Tech Talk Series

68
Training Services
Exploit the full potential of MathWorks products

Flexible delivery options:

▪ Public training available in several cities
▪ Onsite training with standard or
customized courses
▪ Web-based training with live, interactive
instructor-led courses

More than 48 course offerings:

▪ Introductory and intermediate training on MATLAB, Simulink,
Stateflow, code generation, and Polyspace products
▪ Specialized courses in control design, signal processing, parallel computing,
code generation, communications, financial analysis,
and other areas

www.mathworks.in/training 69
Machine Learning with MATLAB

▪ This two-day course focuses on data analytics and machine learning techniques in MATLAB. The
course demonstrates the use of unsupervised learning to discover features in large data sets and
supervised learning to build predictive models. Topics include:
▪ Organizing and preprocessing data
▪ Clustering data
▪ Creating classification and regression models
▪ Interpreting and evaluating models
▪ Simplifying data sets
▪ Using ensembles to improve model performance

70
Statistical Methods in MATLAB

After this 2-day course you will be able to:

Import, visualize, explore, and model data

▪ Fit probability distributions to data, and perform

hypothesis tests

▪ Develop and fit regression models to data

▪ Generate random numbers

and perform simulations

71
Summary: Why MATLAB for Predictive Maintenance?

▪ Dedicated toolbox for data preprocessing and feature extraction and developing
predictive models

▪ Apps to make the task simple

▪ Support for taking these algorithm to edge and enterprise

▪ Get started quickly…examples, training and consulting

72
Please provide feedback for this block of sessions

▪ Scan this QR Code or log onto link below

(link also sent to your phone and email)

▪ http://bit.ly/expo19-feedback

▪ Enter the registration id number displayed

on your badge
▪ Provide feedback for this session

73
Speaker Details:
Email: [email protected]
LinkedIn: https://www.linkedin.com/in/amit-doshi/

© 2015 The MathWorks, Inc.

74