Predictive Maintenance

From Development to IoT Deployment

Phil Rottier
MathWorks UK Consulting

© 2015 The MathWorks, Inc.

1
Outline

 Introduction
 Examples
– Machine Learning
– Digital Twin
 Predictive Maintenance Toolbox
– Features
 Deployment
– Options
 Summary

2
 Predictive Maintenance
for polymer-based production machines
Sensor Data (~1 minute)
10s-100s sensors/machine
Quality State (~40 minutes)

Classification using Statistics, Machine

Learning, and Neural Networks

3
Deployment – a MATLAB App
used by machine operators

OK

4
 Deployment – a MATLAB App
used by machine operators

“Our financial control department

determined that we are saving more
than 50,000 euros per year by using
MATLAB for predictive maintenance,”
says Dr. Kohlert. “That total is based
on just eight machines. We expect
that to increase at least fourfold as
we analyze the data from more of
our machines.”
NOT OK

5
Predictive Maintenance interests and objectives

 Maintenance cares about day-to-day operations

– Need actionable information, readily available
– Dashboards, Hand held devices, Alarms
– Reduced downtime, reduced cost of service delivery
 Operations & IT look at the bigger picture
– Integration with IT & OT systems
– Improved operating efficiency
 Engineering get product feedback
– Enhanced design management through life
 Management wants to drive growth
– Automated reports
– New revenue streams
6
Outline

 Introduction
 Examples
– Machine Learning
– Digital Twin
 Predictive Maintenance Toolbox
– Features
 Deployment
– Options
 Summary

7
Machine learning example: typical workflow

 Get data into MATLAB

– Clean / sanitise / verify as appropriate
 Explore options with the “Learner” apps
– Designed to be quick, easy and intuitive to use
 Write code to formalise (and document) chosen approach
– Programmatic approach offers more options, in a more formal manner
 Deploy for operational use
– Package your choices in a MATLAB App
– Compile to stand-alone executable
– Compile to Add-in for Excel
– Deploy to web
– Deploy to hardware
8
Machine learning example: engine failure data

 100 failure events

 Time-axes normalised so failures occur at T = 0
 Can we predict whether any single point measurement is within, say, 100
days of failure ?

10
Machine learning example: classification

 How to get started ?

 >> classificationLearner

11
Machine learning example: prediction

12
Outline

 Introduction
 Examples
– Machine Learning
– Digital Twin
 Predictive Maintenance Toolbox
– Features
 Deployment
– Options
 Summary

13
Digital twin example: Predictive Maintenance workflow
Build Digital
Twin
 Data isn‘t always readily available
Sensor Data
– Measurements difficult to acquire Obtain and
– Time consuming or costly or impracticable Explore Data

– Need to cater for multiple failure modes

Preprocess
Solution: Build digital twin and generate Training Data
sensor data using simulation
Develop
 Developing algorithm is complex Predictive Model

– Requires complex concepts and analysis

Solution: Use MATLAB to simplify process of Deploy
Algorithm Control
developing and deploying algorithm Simulation Unit

14
Digital twin example: Physical modelling using Simscape

Power Systems

Electronics

MultiBody

Driveline
Fluids
 Enables physical modeling
of multi-domain physical systems
 Eases process of
Simscape
modeling physical systems
MATLAB, Simulink
– Build models that reflect V+
structure of physical system
– Leverage MATLAB to create
reusable models V-

 Build models
with the same structure
as the physical system

15
Digital twin example: Simscape model of a Triplex Pump

 Crankshaft drives three plungers Failure Algorithm

– Each 120 degrees out of phase Diagnosis
Pressure
Sensor
– One chamber always discharging Component
– Smoother flow than single Failure

or duplex piston pumps

Outlet

Crankshaft

Inlet

16
Video for Simulink fault data

17
Digital twin example: Simscape model of a Triplex Pump

 Ability to simulate plant, with fault conditions, is essential

 3-D visualisation is nice to have, not essential

18
Digital twin example: model validation

 e.g. by tuning to test data where possible

19
Digital twin example: simulate plant performance

 Create dataset under wide range of

– Normal operations
– Fault conditions
 Multiple failure modes
 Levels of degradation

20
Digital twin example: characterise plant performance

 Data reduction: extract features of interest

 Create labelled dataset, for training

21
Video for feature extraction and code gen

22
Digital twin example: train diagnostic model

 Using the machine learning techniques

23
Digital twin example: Deploy diagnostic model

24
Outline

 Introduction
 Examples
– Machine Learning
– Digital Twin
 Predictive Maintenance Toolbox
– Features
 Deployment
– Options
 Case studies

25
Predictive Maintenance Toolbox

 New product in Release 2018a

26
Toolbox contents

 Data management
– Customised datastores for managing “simulation ensembles”
 Pre-processing functions
 Condition indicators
 Fault detection & prediction
– Remaining useful Life (RUL) models
 Deployment

28
Example – Remaining Useful Life

29
Remaining Useful Life

 Set up model and train

30
Video for RUL estimation & Compiler

31
Outline

 Introduction
 Examples
– Machine Learning
– Digital Twin
 Predictive Maintenance Toolbox
– Features
 Deployment
– Options
 Summary

32
 Deployment Options

Asset w. smart Edge Devices OT Systems IT Systems

sensors

Algorithm, Analysis, Reports

Models
Data

Engineer
33
Deployment options

 Cloud
– If data is being stored on the cloud
– Removing the need to transfer data between the cloud and local machines that are
running the prognostics and health monitoring algorithm makes the maintenance
process more effective
– Results calculated on the cloud can be made available through tweets, email
notifications, web apps, and dashboards
– For cloud implementations, you can use ThingSpeak or MATLAB Production Server.

34
Deployment options

 Edge
– Algorithm can run on embedded devices close to the actual equipment.
– The amount of information sent is reduced as data is transmitted only when needed,
– Updates and notifications about equipment health are immediately available.
– For embedded implementations, you can use MATLAB Compiler, MATLAB Coder, or
Simulink Coder to generate code that runs on a local machine.

35
Deployment options

 Edge + Cloud
– In systems that run continuously and generate huge amounts of data, storing all the
data on board or transmitting it is not always viable because of cellular bandwidth and
cost limitations.
 Pre-processing and feature run on embedded devices
 Predictive model can run on the cloud
 Generate notifications as needed
– Using an algorithm that operates on streaming data or on batches of data lets you store
and send data only when needed..

36
Outline

 Introduction
 Examples
– Machine Learning
– Digital Twin
 Predictive Maintenance Toolbox
– Features
 Deployment
– Options
 Summary

39
Summary: Predictive Maintenance workflow

 Reduce downtime, reduce cost, improve efficiency

40
Summary: products

 Analytics
– MATLAB
– Statistics and machine learning
toolbox
– Deep learning toolbox
– Predictive Maintenance toolbox
 Simulation
– Simulink
– Simscape
– Simulink Design Optimisation
 Deployment
– Compiler, Coder

41
Summary: why use MathWorks’ tools ?

 Single environment for

– Data acquisition
– Data pre-processing / reduction
– Data characterization
– Machine learning / algorithm design & development
– Modelling / simulation
– Deployment – wide range of options available
 MATLAB provides
– Easy access - get started quickly
– Scalability to real-world problems
– New product in 2018a: Predictive Maintenance Toolbox

42