You are on page 1of 25

by

Introducing Predictive
Maintenance
Qt World Summit 2016
Predictive maintenance
The What
What is predictive maintenance?
Moving from devices to smart connected devices

Corrective maintenance
• Wait for something to go wrong (spoiler: it will !)
• Easiest, but no planning, bad perceived quality

Preventive maintenance
• Guess when it will go wrong
• Easy planning, extra cost, requires consistent behavior

Predictive maintenance
• Be alerted before it goes (too) wrong
• Easy planning, optimal interventions

3
Is it for me?
Requirements
? It depends!

Failures are acceptable (for operations & perceived quality)


• Corrective maintenance
No budget to work at it or no signs before a failure
• Preventive maintenance

Failures are easily predicted


• Condition based predictive maintenance
Failures are harder to predict
• Model based predictive maintenance

4
Predictive maintenance
Condition based

• For simple cases


• Use conditions to trigger an alert
• When motor’s current is above 1A
• When CPU temperature is above 80°C
• When vibrations occur

+ Easy to implement
- Limited

5
Predictive maintenance
Model based condition monitoring

• Fits the more complex cases


• Use a set of data to learn (predict) when a failure will occur
• Machine learning
• Supervised learning requires a learning data set
• Preferrably experienced engineer or data scientist (or find some books !)

+ Can cover more complex cases


- More work to implement and maintain

6
When to plan for it ?
Impact on design

Prototyping & Hardware design


• Identify signs occuring before a failure
• Integrate the appropriate sensors (luminosity, vibration, temperature, …)
System software architecture
• Monitor sensors, notify changes
• Create a model manually or train one w/ machine learning
• Integrate model and prediction (web API, library or complete solution)

7
Predictive maintenance
The How
The right tools
Existing and custom solutions

CMMS: big commercial solutions (IBM Maximo, MVP Plant, …)


• More or less easy to integrate
• Usually best for large scale, complex operations
• Less technical knowledge needed

Custom solutions from open tools and technologies (like Qt !)


• Tailored to your context and tools
• Requires technical skills
9
Basic Predictive maintenance
Condition based
Statistics driven

• Connected device reporting usage stats

• Statistics driven automated


maintenance: “If… then”

• Allows increased lifespan and uptime


• Fixing issues before seeing damages

• Why should we need the cloud ?


• Evolutivity
• Connectivity with other services Nb of cup served
Qty coffee grounded
Qty milk used
Usability
10
Basic Predictive maintenance
Connecting simple tools
HTTPS

Mobile and Desktop


Device Supervision
website
• Qt application

Web API
• ElasticStack (or NodeJS, PHP, …) DB Web API

• Email and/or ticket on event


Smart
HTTPS device
Supervision website
• Jira (or redmine, custom, …) Cloud infrastructure Local devices

11
On the device
(Qt) Application’s role

Monitor
void Device::pollSensors()
QTimer, QThread {
QFile file("/sys/class/mysensor/value");
QFileWatcher […]
int value = QString::number(file.readAll());
Serialize / Log QNetworkAccessManager manager;
QJson classes QJsonDocument jsonDoc; QJsonObject jsonObject;
jsonObject["mysensor"] = value;
QLoggingCategory, msg handler […]
qCDebug(sensorsLogCat) << jsonData;
manager.post(QUrl("http://monitor.domain.com"), jsonData);
Notify QTimer::singleshot(60*1000, pollSensor);
HTTPS, AMQP or MQTT }

12
On the Web server
Cloud business intelligence

"actions": {
Parse "send_email": {
"email": {
LogStash, NodeJS "to": "operator@customer.com",
"subject": "Please check me !",
"body": "You should probably check machine
Store {{ctx.payload.hits.0.fields.name}}, something seems wrong on
the espresso motor !",
ElasticSearch, MariaDB (MySQL)
"attachments": {
Essential to build a dataset "machine_report": {
"http": {
Alert
"content_type": "application/pdf" ,
"request": {"url": "http://localhost/report[...]}
Watcher, NodeJS }
}
Email, Jira, Redmine }

13
That’s it !
Wait …

Isn’t that just a bunch of « if » ?

14
Advanced Predictive maintenance
Model based Intelligence driven
ML

• Bring in Machine Learning

• Intelligence driven automated


maintenance

• Optimized maintenance costs


• Self improving solution, efficiency
increases with data consolidation

• How do you do that … ?


Nb of cup served
Qty coffee grounded
Qty milk used
Usability
15
The right tools
Machine Learning

Choose your Machine learning toolbox


• The « good old » way Dedicated tool

• Matlab, R
• Machine learning OpenSource frameworks Library

• Shark (C++), Encog (Java), scikit-learn (python)


• Machine learning cloud APIs Online

• Google prediction API, Seldon, MS Azure Machine Learning, BigML

16
Advanced Predictive maintenance
Architecture
HTTPS

Message Broker Mobile and Desktop

• AMQP: QAmqp for Qt Technical


backend
• RabbitMQ server
Machine
• Disconnection msg, queues Learning
API

Web API

Machine Learning DB

• MS Machine Learning
Message Smart
MQTT
broker device

Cloud infrastructure Local devices

17
Advanced Predictive maintenance
Learning and testing

In 4 steps
• Choose your output metric Failure probability
• Remaining useful life, failure probability or maintenance
needed
• Build a complete dataset of values and failures (hard part !) Excel

• Generate a model using Machine learning and test it


• Integrate the model in your system Call to a Web API

18
Machine learning
Dataset & Learning (MS Machine Learning Studio)

Dataset (failure probability) Model for prediction & learning


19
Machine learning
Web API and integration

Production model Testing the webservice 20


Going a bit further
Full system supervision

• Example with Kibana

• Visual overview
• Helps identify visually
trends & anomalies

21
Sum up: Added value
A real leverage for a better business

+ Equipment lifespan thanks to anticipation


+ Better uptime and user satisfaction
+ Optimized maintenance
+ Possible new services and commercial models

And …
• Know your users: Predict their preferences, actions
• Security: Alert potentially fraudulous actions, from unsual
behavior
22
Key points
• Plan to integrate sensors
• Define the machine learning output
• Make sure you can update the prediction

• Enjoy presenting the result to your customers !


• … Put a sensor in that fuel tank !
Witekio, System Software Integrator

Automobile &
Navigation

Handheld & Mobility

Industry & Energy

Medical & Healthcare

Smart Object
Integration
Technical software Embedded system
expert expert
System software
integrator

Witekio helps customers to develop and integrate all the software layers
from the hardware to the cloud
24
Thank you
Witekio France Witekio USA Witekio Germany Witekio UK/EmbeddedBits Witekio Asia
4, chemin du ruisseau 3150 Richards Roads Suite 210 Bellevue, Am Wartfeld- 61169 Hollywood Mansion, Hollywood Lane, C/O 14F-3, No. 57, Fuxing Nth Rd,
69134, Ecully France WA, 98005, USA Friedberg, Germany Bristol BS10 7TW, UK Songshan District, Taipei, 10595, Taiwan
Phone : + 33 4 49 26 25 39 Phone : + 1 425 749 4335 Phone : + 49 6031 693 7070 Phone : + 44(0) 117 369 0930 Phone : +886 2 2740 0394
sales.emea@witekio.com sales.amer@witekio.com sales.dach@witekio.com sales.uk@witekio.com sales.asia@witekio.com

©2016 Witekio & Subsidiaries. All Rights Reserved.


This document and the information it contains is confidential and remains the property of our company. It may not be copied or communicated to a third party or used for any purpose other than that for which it is supplied without the prior written consent of our company.

You might also like