machine learning and predictive analyticssesam-world.com/_pdf/sesam-132/06-velux.pdf · we are...

41
MACHINE LEARNING AND PREDICTIVE ANALYTICS Sigurd Lazic Villumsen

Upload: phunganh

Post on 16-May-2019

226 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

MACHINE LEARNING AND PREDICTIVE ANALYTICSSigurd Lazic Villumsen

Page 2: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

2

Page 3: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

WE ARE WORLD CHAMPIONS OF ROOF WINDOWS

3

AW

AR

EN

ES

S

END-USER — AWARENESS / PREFERENCE 2016

PREFERENCE

VELUX GROUP

COMPETITOR X COMPETITOR Y

100

80

60

40

20

00 20 40 60 80 100

No matter how good you are, you can always do it better faster or cheaper

Page 4: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

Light grey, use

eyedropper to

THE MAIN POINT

We need these technologies to unlock uninvestigated business potential.

• Increase productivity

• We want to grow (acquisitions and increased sales of core products)

• Increase quality even further

• Mitigate manufacturing challenges• Increasing people turnover

• Increasing product program complexity

• Aging workforce

4

Page 5: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

5 VELUX GROUP PRESENTATION

CULTIVATE ⁄ A SIMPLE IDEA

2017 numbers (EUR)

VKR Holdingrevenue

VKR Holdingnet profit

1740 10,200

2.5bn

340m

Company facts

The VELUX Group is owned by VKR Holding A/S, a limited company wholly owned by the foundations and family.

The VELUX Group’s financial results are incorporated into VKR Holding’s consolidated accounts.

sales companies around the world

production sites in nine countries

employeesglobally

Page 6: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

UNCHARTED TERRITORY

6

Possible applications

We don’t know where we are so need to explore before we prioritize.

But we need data to be able to do this

Page 7: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

POWERBI DASHBOARD FOR MACHIENS

7

We are using PowerBI to visualise the data we are collecting from machines in our production.

Benefits

We can easily share data visualizations in our organization.

Very good at making first level analytics.

VELUX COMPANY

Page 8: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

IS MACHINE LEARING ALWAYS NEEDED?

8

Just by making data more accessible, we gain a lot, and also by using fx the insight function in PowerBI we can learn new stuff about our machines.

VELUX COMPANY

Page 9: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

SMART DATA, NOT ONLY BIG DATA

9

In VELUX we have our own PackML interface made for data collection from machines.

This gives us the same data foundation for all machines, making it easy to make standardised dashboards fx OEE calculation, Alarm statistics ect.

VELUX COMPANY

Page 10: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

UNCHARTED TERRITORY

10

Possible applications

Process optimization

Page 11: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

IMPROVING PRODUCTION PROCESSES

11

What we think

ML can be used as a way of improving production processes.

Why is it important

We need to be in control of our processes if we wish to stay in control of quality, delivery cost.

Why ML

We have a big toolbox for optimizing processes, ML could be an additional tool for that.

Status:

Not very far in these areas

Cases:

1. Aluminium saw

2. PU painting line.

Page 12: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

CASE 1: IMPROVING ALUMINIUM SAW

12

Why

Varying length is a problem when cutting aluminum on flying saw.

Goal: Remove the varying length by using machine learning

Result.

?

Page 13: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

CASE 1: IMPROVING ALUMINIUM SAW

13

Why

Varying length is a problem when cutting aluminum on flying saw.

Goal: Remove the varying length by using machine learning

Result.

Machine learning was never necessary as linear correlation was found just by looking at the data.

Machine learning not always needed

Page 14: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

CASE 2: IMPROVING PU PAINTING

14

Goal:

• Improve quality of painted components

• Reduce the used amount of paint.

• Predict maintenance e.g. change of nozzle

• Data available

Page 15: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

Visual inspection

UNCHARTED TERRITORY

15

Possible applications

Process optimization

Page 16: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

VISUAL INSPECTION

16

What we think

We think that we can push the boundaries for visual inspection by using new technology.

Why is it important

Manual inspection is a cost heavy process.

We need to ensure uniform and global quality references

Why ML

We are working with natural products e.g. wood. So far traditional vision solutions has not been good enough

Cases:

3. Wood grading

4. Paint inspection

5. Pane crack recognition

Page 17: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

CASE 3: USING DEEP LEARNING FOR AUTOMATIC WOOD GRADING

LUKE WHITE

MARCH 4, 2019

Page 18: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

Light grey, use

eyedropper to

USING DEEP LEARNING FOR AUTOMATIC WOOD GRADING

18

Goal: To develop a machine vision system to find defects in wooden components as well as or better than current human inspection.

Several iterations of training utilizing both “un-supervised” and “supervised” methods where tested.

Page 19: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

SUPERVISED VS. UNSUPERVISED

19

Unsupervised Training Example

Supervised Training Example

Page 20: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

AUTOMATIC WOOD GRADING

20

Reclassified good/bad parts and re-trained system. New results of 98%+ certainty.

Determined unsupervised training not allowing different quality standards for different wood surfaces (A, B, C, etc.)

Supervised Training

Allowed different surfaces to have different quality standards.

Resulted in certainties above 99% on defined defects.

Current work:

Testing fine crack/lamination seam testing using current lighting

Page 21: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

CASE 4: WHITE PAINTED WOOD COMPONENTS

21

Page 22: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

DETECTION OF PAINT DEFECTS

22 19 DECEMBER 2018 MACHINE LEARNING

Goal: Classify defects in white painted wood components

Simple POC

Result: Good preliminary results. We can detect contaminations and non uniform paint

Page 23: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

RESULTS FROM PRE-STUDY

23 19 DECEMBER 2018 MACHINE LEARNING

Page 24: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

CASE 5: DETECTION OF CRACKED PANES

24

Page 25: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

MOTIVATION

Damaged glass is a source of claims and internal scrap for VELUX

25

Manual inspection Hand programmed scanner Machine learning

Currently glass is subject to manual inspection.

A scanner will be able to take high quality and consistent

images of glass. Performance is untested on edge damages,

but images can be fed to a machine learning model

Machine learning will ensure a high detection rate (95%+), while keeping maintenance low, as the models are self

learning.

Page 26: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

DEEP LEARNING FOR QUALITY INSPECTION• Using a deep learning, we teach a computer model to recognize glass with damaged edges and scratches through four individual steps

• By introducing the computer model to a large number of images of damaged and undamaged glass. You can say that the computer model is “trained”

• The computer model then predicts whether or not it thinks the glass is damaged.

• We then compare these predictions to the actual state of the glass.

• The computer model then updates itself based how right it predicted the state of the glass.

Window glass Deep Learning model

Layer 1 Layer 2 Layer 3 Layer 4

Predictions

Evaluation

Actual state

Model learns based on error

Page 27: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

DATSET• The final dataset consisted of 1200 image tiles. However, only 55 of the images have an actual damage on them, which in this case is the important

information. As such, we have a very skewed dataset, which will have to be taken into consideration when training the models and validating the results.

• Further image preprocessing is performed as part of the data pipeline, this being:

• Random image augmentation is added to increase image diversity. Augmentations include random 90-interval degrees rotation, color scaling, adding random noise, tilting and shifting images by small amounts.

Random image augmentation

Page 28: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

INITIAL RESULTS – RANDOM VALIDATION• To solve the issue of the unbalanced dataset (55 damaged image images when subtracting the validation data, vs 1145 undamaged images), we have

to manually balance the dataset. If we leave it unchanged, the model will have a very hard time distinguishing between damaged and undamaged glass.

• Balancing the dataset simply means copying the damaged images until we reach a balanced dataset (1125 damaged and undamaged images). This will make it very easy to overfit the data, which can be alleviated by applying enough data augmentation.

• A validation set can be chosen in various ways, usually it is done by randomly holding back 20% of the data. In this study, we experiment with various small validation sizes, due to the small sample size of data (number of images). We train the model on validation samples of 4, 8 and 12 images of each class for balanced interpretable validation – A total of 8, 16 and 24 images. To validate the results, we do this sampling the validation data 50 times, training a model for each sample and reporting the average validation loss and accuracy across the 50 models.

• Accuracies should be compared to the accuracy of just guessing the majority class – which as we have balanced the dataset is 50%.

• We utilize a model type pre-trained on another very large dataset, such that the model has already learned useful representations. Due to the short time, we utilize an older model architecture (VGG16 with convolutional top layers) that is known to converge well during training – Higher performance is expected when using a more recent model architecture.

Overview of results

Training data1176-1192 images

Validation data4-12 images sampled from 55 damaged

images and 4-12 images sampled from 1145 undamaged images. Each sampled 50 times.

Average of 50 models with 16 validation images

Loss measure

Loss

Training accuracy

-

Validation accuracy

0.909

Average of 50 models with 8 validation images

Loss measure

Loss

Training accuracy

-

Validation accuracy

0.922

Average of 50 models with 24 validation images

Loss measure

Loss

Training accuracy

-

Validation accuracy

0.910

Page 29: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

29

Rolls Rolls

Viprotronscanner

Camera

Lighting

Viproton PC & software /w VELUX software

Mounted share

Robot arm Discarded Pellet large

Discarded Pellet small

Azure Data Box Edge

Robot com interface

Relay

Azure data lake storage

Reporting on performance

Quality monitor

VELUX TC-SkjernHardware: VELUX IT

PRODUCTION OPERATIONSViprotron

Software: Data & Analytics

Page 30: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

UNCHARTED TERRITORY

30

Possible applications

Process optimization

Visual inspection

Optimizing business

processes

Page 31: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

CASE 6: IMPROVING SERVICE REGISTRATIONS.

31

Project goal:

Predict the cost of a service technician and what spareparts he might need based on user input.

Data:

User inputs, Many years of service registrations

Page 32: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

PREDICTIVE ANALYTICS PROCESS. Input → Model → Output

32

Input:SCO

Product TypeProduct Size

Product VariantComponentSymptom

Ect.

Model:Classification

We run a model that creates a decision tree on the input variable dependent on the

associated output variables.

Output:Spare part

Number of SparepartsStdTextKey

HoursExecution factor

Probability

{

"json_data": {

"sales_company": "V-F",

"product_type": "GPL",

"product_size": "MK04",

"product_variant": "3076F",

"component_group_code": "0004",

"component_code": "4003",

"symptom": "9",

"id": "1"

}

}

{

"json_data": {

"Material": " IPL MK04 0076F",

"Material_Quantity": "1",

"Stdtextkey": "GL_WH",

"Hours": "1",

}

}

Page 33: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

PREDICTION MODEL VS. MACHINE LEARNING

33

HOW DOES A PREDICTION MODEL WORK

Model

Data

Prediction

Customer request

Service order

Confirmed data

Feedbackloop

What is the difference between a prediction model and machine learning?Prediction model: Uses data and model to create a customer specific prediction Machine learning: Set up that alters the model based on the comparison of the prediction and confirmation data

Learning algorithm

NOT PART OF THE CURRENT PROJECT SCOPE!!!

Page 34: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

ARCHITECTURE FOR PREDICTIVE MODELLING IN VELUX

Hana

Data storage on

BW

Development server

ComputeLocal environment

Localfilesystem

CodeTrained model

Trainedmodel

NAS

ODBC

Productionserver

REST API(Plumber)

JSON

JSON

Audit Trail

WebpageVELUX Network Cloud

SAP PEP SRS 2

MAGIC 1 MAGIC 2

Page 36: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

NOTICE: ARCHITECTURE FOLLOWS APPLICATION

36

Not necessarily a one size fits all, but sharing of best practice is extremely important

The application requirements will control the architecture.

Pane crack recognition Service predictions

Page 37: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

UNCHARTED TERRITORY

37

Possible applications

Process optimization

Visual inspection

Optimizing business

processes

Page 38: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

MACHINE LEARNING: WHERE ARE WE?

38

2018 2019 2020 2021

White painted wooden

components

Pane cracks

Wood defect detection

Sup

ply

Oth

er

White painted wooden

componentsPane cracks

Wood defect detection

Quality prediction Painting line

Aluminum sawPredictive

maintenance ØB

… …

……

……

……

……

……

……

……

……

……

Web solutions … (secret) … (secret)Web solutions

… …

……

……

……

……

……

Page 39: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

UNCHARTED TERRITORY

39

Possible applications

Process optimization

Visual inspection

Optimizing business

processes

Predictive maintenance

Lead generation

Unknown applications

Mordor

Page 40: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

CONCLUDING REMARKS

• A good data foundation is an absolute requirement.

• We have taken the first steps towards using machine learning in VELUX.

• We are in many cases still novices in the field, but making progress.

• I expect we will see this as a common tool in 1-2 years

• Remember your old tools.

• Don’t limit yourself to one supplier

• Partner up

• Work with it

40

Page 41: MACHINE LEARNING AND PREDICTIVE ANALYTICSsesam-world.com/_pdf/sesam-132/06-VELUX.pdf · we are world champions of roof windows 3 s end-user —awareness / preference 2016 preference

CONTACT INFO

FIND US HERE

pinterest.com/VELUXGroup/

linkedin.com/company/VELUX

youtube.com/user/VELUX

facebook.com/VELUX

twitter.com/VELUX

Sigurd Lazic Villumsen

[email protected]

+45 30707273