PHM frontiers in Korean manufacturing – success episodes

Copyright © 2018 Laboratory for System Health and Risk Management in Seoul National University2018-04-13 - 1 -

YOUN, Byeng DongProfessor & CEO

System Health and Risk Management (SHRM) Laboratory (shrm.snu.ac.kr)Department of Mechanical and Aerospace Engineering

Seoul National University

OnePredict Inc. (onepredict.com)

NIST Workshop

PHM frontiers in Korean manufacturing –

success episodes and issues


Part Ⅰ: Smart Factory Activity in Korea01

Contents

PartⅡ: PHM Discipline and Activity02

PartⅢ: Successful PHM Episodes03

PartⅣ: Closing Remarks04

* PHM: Prognostics and Health Management


4. Shunt Reactor

Part ⅠSmart Factory in Korea


Manufacturing Power

100

99.5

93.9

80.4

76.7

75.8

72.9

69.5

68.7

68.4

China

United States

Germany

Japan

South Korea

United Kingdom

Taiwan

Mexico

Canada

Singapore

(100: High, 10: Low)

Ref: Delotte Touche Limited and US Council on Competitiveness, “2016 Global Manufacturing Competitiveness Index”


Some Figures of Smart Factory Effort in Korea

Ref: ’18.04 MOTIE Smart Factory Korea Foundation “2017 Successful Episodes of Smart Factory”

• Number: 18 times from ’14 to ’17• Level: Elementary (76.4%)

277

‘14 ‘15 ‘16 ‘17 ‘20 ‘22

12402800

5003

12000

20000

•Elementary76.4%

Intermediate I21.5%

Intermediate II2.1%

<Number of smart factoriessupported by the KOST Program>

<The level of smart factories in domestic companies>

• Elementary: ERP with data acqn.• Intermediate I: RT equipment data acqn.• Intermediate II: RT decision making and

control

Korea Smart Factory Foundation (KOSF) Newly Launched in 2014


Outcomes out of Smart Factory Effort in Korea

Productivity30% ↑

Manufacturing Cost

15% ↓Quality

45% ↑

Improvement of Overall Manufacturing Capability(2800 companies with smart factory, Dec 2017)

Downtime Cost16% ↓

Process Automation

DeviceMonitoring

Power Control

Risk Management

Life CycleManagement

Big DataAnalytics

Condition Monitoring

IoT

IoT

Product Monitoring

Industry 4.0 in Manufacturing

Ref: S&T Market Report Vol. 56 2018.02 Smart Factory Industry and Market Trend


4. Shunt Reactor

Part ⅡPHM* Discipline and Activity

PHMOverview

SensingSolution

ReasoningSolution

DiagnosticsSolution

PrognosticsSolution

PHMArchitecture *PHM – Prognostics and Health Management


Global Recognition of SNU & OnePredict

“Five times winner of Global PHM Data Challengesover Various Industrial Sectors”


PHM Standard Architecture (Collab. w/ UNIST, KAU, and Kookmin Univ.)

Domain Knowledge

Data-driven

Rule-based Physics-based

Data-driven + Physics-based

Dat

a Si

ze

Health Features

Raw Data

Feature Engineering

Preprocessing

Diagnostics/Prognostics

Health Prediction

Power & Energy

Hydro-system

Electrical &Electronics

Power Trans.

Machining

Driving

Cleaning Filtering

EnvelopingScalingStatistical Modeling

Organization

Domain Knowledge

Data-driven



Dat

a Si

zeDomain Knowledge

Data-driven



(Hybrid)

Dat

a Si

ze


PHM Standard Architecture

http://onepredict.com/blog/newsView.do

*M: MatureD: Developing

P: Promising

http://shrm.snu.ac.kr/index.php?idx=511&hCode=BOARD&bo_idx=5&sfl=&stx=&page=view&pg=1


4. Shunt Reactor

Part IIISuccessful PHM Episodes

Case Study 1Industry Robots

Case Study 2Industrial Bearing

Case Study 3Overhead HoistTransport

Case Study 4Deep Learning -Steam/Gas Turbine


1. Steam Turbine2. Power Generator

3. Power Transformer4. Shunt Reactor

Industrial Bearing: Rolling-Element Bearing

Case Study 2Industrial Bearing


REB Failure Prognostics (Schaeffler & Samsung Heavy Industry)

2. Analysis 3. Diagnosis & Prognosis

1. Sensing

ISO 10816-1 (Vibration measured on Non-Rotating Parts)

Waveform analysis Spectrum analysisEnvelope spectrum

& Feature extraction

0 100 200 300 400 500 600 700

time(hr)

1

1.1

1.2

1.3

1.4

1.5

prog

nost

ic in

dex

prognostics

Prognostic index & RUL Prediction

thresholdRUL

Learning Prediction

Industrial Bearings (custom order)

Wind Turbine Motor

Spindle Pump & Compressor

Robots


1 Real-time Monitoring of Health Condition

2 Color Image of Health Condition

3 Trend Monitoring of Health Condition

4 Remaining LifeTrend Monitoring

“First-ever Commercial Solution for Bearing RUL Prediction”

REB Failure Prognostics (Schaeffler & Samsung Heavy Industry)




Case Study 4Steam/Gas Turbine

Steam/Gas Turbine w/ Journal Bearing(Data-driven, Deep Learning)


Autonomous machine learning algorithms to extract data features

through abstractionof massive data sets

Deep LearningHiddenLayer 1

𝑥𝑥2

𝑥𝑥1

…

𝑥𝑥𝑁𝑁

… …

…………

……

Inputlayer

…

……

RandomImages

HiddenLayer N

2012 2014 2016 2018

Year

0

20

40

60

80

100

Rel

ativ

e In

dex

World

Korea

Search Trends in Google

Popularity of Deep Learning- Improved computing power- Enlarged data size- Advanced DL techniques

Vision Recognition

Applications of Deep Learning

Voice Recognition

Healthcare PHM

Deep Learning

*R. Zhao et al. (2016), IEEE Tran. Neural Networks and Learning Systems (Submitted); K. Fragkiadaki (2012) Computer Vision-ECCV; Q. V. Le (2012) ICML


DataAcquisition

FeatureExtraction

FeatureSelection

Feature Learning

Domain Knowledge Based Deep Learning Based

Domain Knowledge Required Not Required

System Dependency Dependent Independent

Feature Representation Shallow Deep

Big Data Learning Over-fitted Model Well-fitted Model

DataAcquisition Autonomous Feature Engineering by DL

DL Type Input Data Type Input Data Labels Others

DBN n/a Unsupervised Break through in deep learning (2006)

CNN Vision Data (Images) Supervised Parameter sharing, local connectivity

RNN Sequential Data (Speech) Supervised Storing sequential information

AE n/a Unsupervised Representation learning, dimension reduction

Deep Learning in PHM

*H. Oh et al. (2017), IEEE Tran. Industrial Electronics** R. Zhao et al. (2016), IEEE Tran. Neural Networks and Learning Systems (Submitted)


Deep Belief Network

Convolutional Neural Network

Recurrent Neural Network

Auto Encoder

RBM 1 RBM 3RBM 2 RBM 4

2nd layer shared weights

1st layer shared weights Encoder Decoder

Code

Input Output

OutputInput Output Input

Input Output

RNN Cell

Deep Learning in PHM

*THE ASIMOV INSTITUTE (http://www.asimovinstitute.org)


Deep Learning for Steam Turbine

2. Analysis

3. Scale-free Turbine Prognostics (Fully validated, 92% prediction accuracy)

1. Sensing

Gap Sensor

(Gap)HousingShaft

Vibration Data Acquisition

Vibration Image

Vib-Imaging Technique

Fault Log

# Norm. Rubb. Misalign. Oil Whirl

#1 0.975 0.010 0.014 0.006

#2 0.991 0.005 0.004 0.000

#3 0.482 0.288 0.186 0.042

#4 0.828 0.135 0.034 0.003

#5 0.037 0.958 0.005 0.001

# Status

#1 Normal

#2 Normal

#3 Normal

#4 Normal

#5 Rubbing

Deep Learning Result

vs.

Omni-directional Regeneration (ODR)

HP IP LP Gen. Ext.

Brg. #1 #2 #3 #4 #5


4. Shunt Reactor

Part ⅣClosing Remarks


Industrial Information Prognosis (IIP)

“Industrial Information is the one with industrial value, which includes availability, quality, productivity, energy efficiency, safety, etc.”

Conventional Business: Raw Material Product

RawMaterial Product

IndustrialRaw Data

IndustrialInformation

New Business: Industrial Data Information


Issues in PHM

01

03

02

04

?

Data QualityLack of Labeled Data

Cyber Security & Data OwnershipLack of Resources

Missing, Unsynchronized, Noisy Data

Class Imbalance, Labeling Quality

PHM Experts, Budgets

Protection of Cloud Network Information




THANK YOUFOR LISTENING

ANY QUESTION?

Documents

PHM frontiers in Korean manufacturing – success episodes