[PPT] - Deep Learning for Predictive Maintenance Pawel Morkisz GTC 2017 PowerPoint Presentation

SLIDE 1

Deep Learning for Predictive Maintenance

Pawel Morkisz GTC 2017

SLIDE 2

www.relia-sol.pl

Agenda

Problem
Introduction and notion of deep neural networks
Convolutional layers
Residual networks (ResNet)
One dimensional convolutional networks

in failure prediction

Approach
Results and the best architecture

SLIDE 3

www.relia-sol.pl

The problem

Inefficient operations

unexpected downtimes, repairs
lower productivity and safety

Delayed timeline

costly delays,
missing critical deadlines,
damaged customer relationships

PdM aMarket

$4.9B by 2021, at CAGR of 28.4%

Substantial cost and safety hazards caused by machinery failure

SLIDE 4

www.relia-sol.pl

Problem setting

Sensor data collected in predefined time intervals
Well specified failure records

Data collected through thousands of sensors Pattern recognition indicate oncoming failure, malfunction or anomalies Clear insights related to

perations, services,

logistics, design

SLIDE 5

www.relia-sol.pl

Data format

Timestamp Sensor 1 Sensor 2 Sensor 3 Sensor 4 Sensor 5 Sensor 6 …. Failure? 01.02.2011 00:03 999,7 5,300 5,547 0,087 3491,7

0,942

... 01.02.2011 00:04 744,6 6,053 20,665 0,178 1436,9

0,820

... 01.02.2011 00:05 4,7 9,111 3,116 0,226 6151,9

0,410

... 01.02.2011 00:06 840,9 4,413 7,863 0,059 7759,8

0,065

... 01.02.2011 00:07 756,7 0,606 22,314 0,131 4474,9

0,429

... 01.02.2011 00:08 750,9 6,303 4,633 0,092 3664,1

0,318

... 01.02.2011 00:09 639,8 3,826 5,382 0,206 3999,1

0,271

... 01.02.2011 00:10 274,2 9,073 16,963 0,066 2834,0

0,514

... 01.02.2011 00:11 551,6 4,383 16,822 0,183 1808,3

0,334

... 01.02.2011 00:12 983,7 3,497 22,169 0,087 9260,7

0,632

... 01.02.2011 00:13 742,7 3,012 23,503 0,042 7537,9

0,481

... 01.02.2011 00:14 24,7 1,394 2,590 0,085 163,9

0,048

... 01.02.2011 00:15 568,9 5,846 4,161 0,133 8403,1

0,909

... 1 01.02.2011 00:16 329,2 8,313 7,152 0,006 5390,7

0,456

... 1 01.02.2011 00:17 269,1 9,835 3,013 0,098 2576,4

0,908

... 1

SLIDE 6

www.relia-sol.pl

Problem setting

…. Failure? ... ... ... ... ... ... ... ... 1 ... 1 ... 1 ... 1 ... 1 ... 1 ... 1 ... 1

𝑼

Determine the time horizon 𝑼

for failure prediction

Observations during pre-failure period

marked as the distinguished class

Failure records itself can be removed or not,

depending on how much they differ from the rest of the set

Binary classification - evaluation of probability

that observation precedes failure

SLIDE 7

Predictive maintenance - interpretation

www.relia-sol.pl

1 2 3 4 5

Time Sensors

SLIDE 8

Predictive maintenance - interpretation

www.relia-sol.pl

1 2 3 4 5

Time Sensors

SLIDE 9

www.relia-sol.pl

Evaluation

Data set is divided into three parts
model learning,
validation of hiper-parameters
final evaluation
Model search criterion in selected

class is the quality on the second set

The quality criterion between

classes is quality on the third set

Chronological division
prevents ‘prediction of the past

using future’

Timestamp Sensor 1 …. Failure? 01.02.2011 00:03 999,7 ... 01.02.2011 00:04 744,6 ... 01.02.2011 00:05 4,7 ... 01.02.2011 00:06 840,9 ... 01.02.2011 00:07 756,7 ... 01.02.2011 00:08 750,9 ... 01.02.2011 00:09 639,8 ... 01.02.2011 00:10 274,2 ... 01.02.2011 00:11 551,6 ... 01.02.2011 00:12 983,7 ... 01.02.2011 00:13 742,7 ... 01.02.2011 00:14 24,7 ... 01.02.2011 00:15 568,9 ... 1 01.02.2011 00:16 329,2 ... 1 01.02.2011 00:17 269,1 ... 1

Learning Validation Test

SLIDE 10

www.relia-sol.pl

Evaluation

𝐶 – number of not predicted failures,

𝐷 – number of false alarms

industrial problem

significantly different costs of 𝑪, 𝑫

Class cost coefficient 𝑦

(included in model training)

Real False True Classified False

A 𝑪

True

𝑫 D 𝐹𝑠𝑠 = 𝑪 𝑦 1 + 𝑦 + 𝑫 1 1 + 𝑦

SLIDE 11

www.relia-sol.pl

Independence of the observations

Actual Series Seasonal Component Breakdown Trend Component Remainder remainder trend seasonal data time

Sensor data
Collected cyclically,
Multidimensional time series
Dependent!
Many machine learning methods

require independence

Data transformations
Decomposition (trend, periodicity,

etc.)

A lot of additional variables

SLIDE 12

www.relia-sol.pl

Convolution approach

32 32 3

1 1 1 1 1x1 1x0 0x1 4 3 4 1x0 1x1 1x0 2 4 3 1x1 1xo 0x1 1 1

Image Convolved Feature

SLIDE 13

www.relia-sol.pl

Convolutional Neural Network (CNN)

source:

Weights sharing - less parameters
Better understanding of inherent data structure

SLIDE 14

www.relia-sol.pl

Residual networks

Deeper architectures because

the residual layers usually learn small, near zero values

The winning architecture in

many competitions

Great stability improvement
bserved

weight layer weight layer + x relu F(x) H(x)=F(x)+x relu Identify x

SLIDE 15

Sensor 1 Sensor 2 Sensor 3 …. 999,7 5,300 5,547 ... 744,6 6,053 20,665 ... 4,7 9,111 3,116 ... 840,9 4,413 7,863 ... 756,7 0,606 22,314 ... 750,9 6,303 4,633 ... 639,8 3,826 5,382 ... 274,2 9,073 16,963 ... 551,6 4,383 16,822 ... 983,7 3,497 22,169 ... 742,7 3,012 23,503 ... 24,7 1,394 2,590 ... 568,9 5,846 4,161 ... 329,2 8,313 7,152 ... 269,1 9,835 3,013 ...

www.relia-sol.pl

Convolutional neural networks in failure prediction

One dimensional filters
Applied only to columns,
i. e. on subsequent

measurements from one sensor

SLIDE 16

www.relia-sol.pl

Network architecture

Input layer 4 x Conv (3 x 1 filter) 8 x Conv (3 x 1 filter) 16 x Conv (3 x 1 filter) Dense layer (256 neurons) Output layer (2 neurons)

SLIDE 17

www.relia-sol.pl

Techniques used

Loss function taking into account

large disproportions in the number of classes

Batch normalization
L2 regularization
PReLU \ ReLU activation functions

SLIDE 18

www.relia-sol.pl

Results

Class\Real False True 𝐹𝑠𝑠 XGBoost False 50892 5 5.02 True 29 49 DNN ResNet Average False 47553 1.06 4.10 True 2893 52.94 DNN ResNet

Best

False 50391 0.06 True 55 54 𝑦 = 950 𝐹𝑠𝑠 = 𝑪 𝑦 1 + 𝑦 + 𝑫 1 1 + 𝑦

SLIDE 19

www.relia-sol.pl

Occupancy data set

Attempt to use the same

architecture

Setting hyper-parameters on

validation data

Only changes:
Weights of classes
L2 regularization coefficient

SLIDE 20

www.relia-sol.pl

Occupancy data set

SLIDE 21

www.relia-sol.pl

Our solutions

Cloud based IIoT solution
Fast and painless deployment and integration
Unlimited possibilities in:
Adding new machines (with hierarchy)
Quick generatingpredictive models
On-the-fly monitoring of assets
Identifying the causes of failures
Immediate access to information worldwide
Small plug & play predictive maintenance

device

Predictive model adjusted for the machine
Scalable
Onboard computations – no necessity of

constant Internet connection

Integrable

with majority

f

industrial transmission protocols

Low cost of purchase and deployment

The Mind Cloud, IIoT system The Eye Edge device

SLIDE 22

www.relia-sol.pl

Our solutions

SLIDE 23

THANK YOU FOR YOUR ATTENTION!

Reliability Solutions Sp. z o.o. Lublańska 34, 31-476 Kraków Head-office: +48 (12) 394-11-21 Sales : +48 (12) 394-11-23 ACC: +48 (12) 627-77-15 R&D: +48 (12) 394-11-31 IT: +48 (12) 394-11-29

ffice@relia-sol.pl

Deep Learning for Predictive Maintenance

Pawel Morkisz GTC 2017

Agenda

in failure prediction

The problem

Inefficient operations

Delayed timeline

PdM aMarket

Substantial cost and safety hazards caused by machinery failure

Problem setting

Data collected through thousands of sensors Pattern recognition indicate oncoming failure, malfunction or anomalies Clear insights related to

logistics, design

Data format

Problem setting

𝑼

for failure prediction

marked as the distinguished class

depending on how much they differ from the rest of the set

that observation precedes failure

Predictive maintenance - interpretation

1 2 3 4 5

Time Sensors

Predictive maintenance - interpretation

1 2 3 4 5

Time Sensors

Evaluation

class is the quality on the second set

classes is quality on the third set

using future’

Learning Validation Test

Evaluation

𝐷 – number of false alarms

significantly different costs of 𝑪, 𝑫

(included in model training)

Real False True Classified False

A 𝑪

True

𝑫 D 𝐹𝑠𝑠 = 𝑪 𝑦 1 + 𝑦 + 𝑫 1 1 + 𝑦

Independence of the observations

require independence

etc.)

Convolution approach

32 32 3

1 1 1 1 1x1 1x0 0x1 4 3 4 1x0 1x1 1x0 2 4 3 1x1 1xo 0x1 1 1

Image Convolved Feature

Convolutional Neural Network (CNN)

source:

Residual networks

the residual layers usually learn small, near zero values

many competitions

weight layer weight layer + x relu F(x) H(x)=F(x)+x relu Identify x

Convolutional neural networks in failure prediction

measurements from one sensor

Network architecture

Techniques used

large disproportions in the number of classes

Results

Class\Real False True 𝐹𝑠𝑠 XGBoost False 50892 5 5.02 True 29 49 DNN ResNet Average False 47553 1.06 4.10 True 2893 52.94 DNN ResNet

False 50391 0.06 True 55 54 𝑦 = 950 𝐹𝑠𝑠 = 𝑪 𝑦 1 + 𝑦 + 𝑫 1 1 + 𝑦

Occupancy data set

architecture

validation data

Occupancy data set

Our solutions

The Mind Cloud, IIoT system The Eye Edge device

Our solutions

THANK YOU FOR YOUR ATTENTION!

We invite you to our booth 1132!