Copyright © SAS Institute Inc. All rights reserved.
SAS Visual Defect Detection System
with SAS Deep Learning and ESP
By Steel Dragon Team
SAS China
Jan 2020
Copyright © SAS Institute Inc. All rights reserved.
Agenda
Background
Demo
Solution Value
Copyright © SAS Institute Inc. All rights reserved.
Background
Copyright © SAS Institute Inc. All rights reserved.
BaoSteel Co.
Mission: Becoming a top steel product, technology and service provider in the world.
• About the customer
• Largest steel company in China
• RMB 305.2 billion (US 44 billion) business
revenue in 2018
• RMB 27.82 billion ( US 4 billion) profit
in 2018
• No.1 Silicon steel sale in the world
• No.2 Crude Steel output in the world.
• No.3 Automobile plate sale in the world.
• Enhancing Strategy
• Technical leadership
• Digital transformation
• Service foremost
• Environment managementSource: http://www.baosteel.com/
Copyright © SAS Institute Inc. All rights reserved.
Business Challenge
Efficiency
Improve the efficiency of
inspector’s manual work
by automating the
process
Accuracy
Improve the predictive
model accuracy by adopt
Computer Vision
capability
Optimization
Continuously embed
expert’s domain knowledge
into the model and apply it
into production seamlessly.
Around 8.6 – 21.6 million
pictures generated everyday,
around 2 million defects
pictures needs to be
analyzed.
The accuracy of the result
from current defect
detection system (Parsytec)
was just ~60%.
How to support experienced
QC staff embed their rich
knowledge into the production
line to achieve continuous
optimization?
People
Technology
Process
Copyright © SAS Institute Inc. All rights reserved.
Quality Control – Defect Detection and Classification
1. Surface Sensor: Take pictures
2. Junction Box: Image analysis
3. Inspection Server: Store defects
4. Inspection Terminal: Monitor defects
Parsytec
Main Defects are important.
ACC—Accuracy
FNR—False Negative Rate
Production
Line
Main
Defects
Fake
Defects
Main
Defects
ACC
Main
Defects
FNR
2050 23 14 63% 4.2%
1580 27 13 64% 3.7%
1880 21 13 59% 6.9%
Current Situation
Copyright © SAS Institute Inc. All rights reserved.
Defect Examples
Black line:
Internal crack of slab caused during rolling.
Scratch
Mechanical damage to the surface during rolling.
Black Stripe:
Stripe-like pseudo-defect due to different light
and shadow conditions.
Red iron:
Scale consists of thin layers of iron oxide crystals.
Copyright © SAS Institute Inc. All rights reserved.
Demo
Product
Line
Baoshan
Black Stripe Black Line Red Iron Scratch
Factory Hot Rolled Steel Line 2050
Product Type Product Line Analytics Beyond Vision
342 defects of 4 defect types has been
detected, released 57 labor-hours from
Factory Baoshan, Hot Rolled Steel, Product
Line 2050.
SAS Visual Defect Detection System – Real-time Monitor
Black_Stripe
51 151 27 33
Total defects
Stripe-like pseudo-defect due to
different light and shadow conditions.
Internal crack of slab caused
during rolling.
Scale consists of thin layers
of iron oxide crystals.
Mechanical damage to the
surface during rolling.
SAS Visual Defect Detection System – Continuous Optimization
Product
Line
Baoshan
Factory A 1
Workshop Product Line 2019/11/05 12
DateTime 20
Condition This interface was
designed to help
operator, to modify the
classification result of
image and then re-train
the CV model for better
performance.
0,1629
0,3729
0,1024
0,3617
0 0,1 0,2 0,3 0,4
Black Line
Black Stripe
Red Iron
Scratch
Probability of Defects
CHANGE
Black Lines Black Stripe Sratch Red Iron
Train
Black Stripe
Architecture
JavaScript
User
Interface
ESPJS
connect
Source Calculate
(Image Resizing)
Score
(Classification)
Model Reader
(ASTORE
Loading)
ESP Server
Viya VDMML
SAS Server
Retrain
Analytic
store
Hardware
& Software
EEC 171_V35
• Viya 3.5
• ESP 6.2
• CentOS 7.6
Hardware
• Memory:
128G
• Storage:
530G
• CPU:
12,
10 *2.27
Ghz Intel Xeon
E7560
Machine Vision Life cycle
Source
Calculate
(Image Resizing)
Score
(Classification)
Model Reader
(ASTORE Loading)
Viya VDMML
Training Once
Use everywhere
Database
Real Time Scoring
in ESP Edge
Nvidia TX2
Process Flow
Video: 2 Pics/s
Source
Calculate
(Image Resizing)
Score
(Classification)
Model Reader
(ASTORE Loading)
ESP Server
Black_line
Process Flow
Video: 2 Pics/s
Source
Calculate
(Image Resizing)
Score
(Classification)
Model Reader
(ASTORE Loading)
ESP Server
31
32
Black_line
+1
Copyright © SAS Institute Inc. All rights reserved.
Solution Value
Copyright © SAS Institute Inc. All rights reserved.
Quality optimization
IMPACT
Inefficiency of defect
analysis
CONSEQUENCE
CURRENT STATE
Inaccuracy of defect
classification
Gap between quality inspector
and quality analyst
Hard to continuously improve
the quality by embed expert
knowledge into the predictive
model in interactive manner
Longer product quality
problem solving time
More manual work
Limited understanding of what
drives product quality
Unforeseen down time is
driving down margins
Root cause analysis and quality
issues is costing too much time
and effort
Reduced Brand equity
 Increased
maintenance cost
 Increased rework
cost and call back size
 Reduced Production
capacity resulting in a
potential hit to
revenue
 Reduced yields
 Increased
down time
 Increased
waste/scrap
Product quality problem
repeating
 Increased
labor cost
 Customer
satisfaction
Copyright © SAS Institute Inc. All rights reserved.
Quality Optimization
Continuous Optimization
• Visual , interactive
interface for quality
analyst
• Embed expert
knowledge to achieve
optimization
Process
Seamlessly, integrated and
continuous quality
optimization process enabled
 Reduced labor cost
 increased yields
SAS® SOLUTION BUSINESS CHANGE ENABLED
 Reduced rework and scrap
BENEFITS
Product Quality
Reduce Yield variations
Automatic , real-time defect
analysis minimize product
quality failures
Real-time monitor
• Visual real-time
• interactive dashboards
• Support mobile/edge
device
Aftersales
Minimize the impact of
aftersales defects
 Increased brand equity
 Quicker root cause
determination
Production Efficiency
Reduce rework efforts and call
back sizes – early intervention
Minimize efforts and
accelerate the process to
determine potential quality
issues
Increased quality – optimize
and tune the process
 Prioritizes problems based
on business impact
 Collaborative environment
enable more innovation
 Increased product quality
 Increased Customer
satisfaction
Copyright © SAS Institute Inc. All rights reserved.
Hot-rolled Steel Production Saving
Revenue Saved= Qualified Products Rev * FNR Improved (due to missed defects)
+
Defect Products Rev * Acc Increased (due to miss-classification)
SAS Visual Defect Detection Value Calculator (Million USD)
Revenue Accuracy Before FNR Before Hired Inspectors
Total
Labor Cost
11,272 62.00% 4.20% 280 4.86
Defects Rate Accuracy After FNR After
Labor-Cost
(Man/Year in USD)
Missed
Defects Rev
5.00% 90.00% 1.00% $17,340 340.87
Substandard
Rate
Accuracy
Improved
FNR
Decreased
Misclassified
Defects Rev
0.50% 28.00% 3.20% 173.59
Customer Data Input Total Rev
Saved
SAS Input Acc: Model Accuracy Rate
FNR: False Negative Rate
Totals 519.31
Copyright © SAS Institute Inc. All rights reserved.
Cold-rolled Steel Production Saving
SAS Visual Defect Detection Value Calculator (Million USD)
Revenue Accuracy Before FNR Before Hired Inspectors
Total
Labor Cost
13,588 62.00% 4.20% 330 5.72
Defects Rate Accuracy After FNR After
Labor-Cost
(Man/Year in USD)
Missed
Defects Rev
5.00% 90.00% 1.00% $17,340 410.89
Substandard
Rate
Accuracy
Improved
FNR
Decreased
Misclassified
Defects Rev
0.50% 28.00% 3.20% 209.25
Customer Data Input Total Rev
Saved
SAS Input Acc: Model Accuracy Rate
FNR: False Negative Rate
Totals 625.87
Revenue Saved= Qualified Products Rev * FNR Improved (due to missed defects)
+
Defect Products Rev * Acc Increased (due to miss-classification)
Copyright © SAS Institute Inc. All rights reserved.
Total-rolled Steel Production Saving
SAS Visual Defect Detection Value Calculator (Million USD)
Revenue Accuracy Before FNR Before Hired Inspectors
Total
Labor Cost
24,860 62.00% 4.20% 610 10.58
Defects Rate Accuracy After FNR After
Labor-Cost
(Man/Year in USD)
Missed
Defects Rev
5.00% 90.00% 1.00% $17,340 751.77
Substandard
Rate
Accuracy
Improved
FNR
Decreased
Misclassified
Defects Rev
0.50% 28.00% 3.20% 382.84
Customer Data Input Total Rev
Saved
SAS Input Acc: Model Accuracy Rate
FNR: False Negative Rate
Totals 1145.19
Revenue Saved= Qualified Products Rev * FNR Improved (due to missed defects)
+
Defect Products Rev * Acc Increased (due to miss-classification)
Copyright © SAS Institute Inc. All rights reserved.
S p e c i a l T h a n k s
Xiangqian Hu SAS Advanced Analytics R&D Cary
Wenyu Shi SAS Advanced Analytics R&D Cary
S t e e l D r a g o n
Olivia Wang Customer Advisory SAS China
Jackie Hu Customer Advisory SAS China
Tianlun Gu Customer Advisory SAS China
Copyright © SAS Institute Inc. All rights reserved.
Thanks

SAS Visual Defect Detection System VP Steel Dragons.pdf

  • 1.
    Copyright © SASInstitute Inc. All rights reserved. SAS Visual Defect Detection System with SAS Deep Learning and ESP By Steel Dragon Team SAS China Jan 2020
  • 2.
    Copyright © SASInstitute Inc. All rights reserved. Agenda Background Demo Solution Value
  • 3.
    Copyright © SASInstitute Inc. All rights reserved. Background
  • 4.
    Copyright © SASInstitute Inc. All rights reserved. BaoSteel Co. Mission: Becoming a top steel product, technology and service provider in the world. • About the customer • Largest steel company in China • RMB 305.2 billion (US 44 billion) business revenue in 2018 • RMB 27.82 billion ( US 4 billion) profit in 2018 • No.1 Silicon steel sale in the world • No.2 Crude Steel output in the world. • No.3 Automobile plate sale in the world. • Enhancing Strategy • Technical leadership • Digital transformation • Service foremost • Environment managementSource: http://www.baosteel.com/
  • 5.
    Copyright © SASInstitute Inc. All rights reserved. Business Challenge Efficiency Improve the efficiency of inspector’s manual work by automating the process Accuracy Improve the predictive model accuracy by adopt Computer Vision capability Optimization Continuously embed expert’s domain knowledge into the model and apply it into production seamlessly. Around 8.6 – 21.6 million pictures generated everyday, around 2 million defects pictures needs to be analyzed. The accuracy of the result from current defect detection system (Parsytec) was just ~60%. How to support experienced QC staff embed their rich knowledge into the production line to achieve continuous optimization? People Technology Process
  • 6.
    Copyright © SASInstitute Inc. All rights reserved. Quality Control – Defect Detection and Classification 1. Surface Sensor: Take pictures 2. Junction Box: Image analysis 3. Inspection Server: Store defects 4. Inspection Terminal: Monitor defects Parsytec Main Defects are important. ACC—Accuracy FNR—False Negative Rate Production Line Main Defects Fake Defects Main Defects ACC Main Defects FNR 2050 23 14 63% 4.2% 1580 27 13 64% 3.7% 1880 21 13 59% 6.9% Current Situation
  • 7.
    Copyright © SASInstitute Inc. All rights reserved. Defect Examples Black line: Internal crack of slab caused during rolling. Scratch Mechanical damage to the surface during rolling. Black Stripe: Stripe-like pseudo-defect due to different light and shadow conditions. Red iron: Scale consists of thin layers of iron oxide crystals.
  • 8.
    Copyright © SASInstitute Inc. All rights reserved. Demo
  • 9.
    Product Line Baoshan Black Stripe BlackLine Red Iron Scratch Factory Hot Rolled Steel Line 2050 Product Type Product Line Analytics Beyond Vision 342 defects of 4 defect types has been detected, released 57 labor-hours from Factory Baoshan, Hot Rolled Steel, Product Line 2050. SAS Visual Defect Detection System – Real-time Monitor Black_Stripe 51 151 27 33 Total defects Stripe-like pseudo-defect due to different light and shadow conditions. Internal crack of slab caused during rolling. Scale consists of thin layers of iron oxide crystals. Mechanical damage to the surface during rolling.
  • 10.
    SAS Visual DefectDetection System – Continuous Optimization Product Line Baoshan Factory A 1 Workshop Product Line 2019/11/05 12 DateTime 20 Condition This interface was designed to help operator, to modify the classification result of image and then re-train the CV model for better performance. 0,1629 0,3729 0,1024 0,3617 0 0,1 0,2 0,3 0,4 Black Line Black Stripe Red Iron Scratch Probability of Defects CHANGE Black Lines Black Stripe Sratch Red Iron Train Black Stripe
  • 11.
    Architecture JavaScript User Interface ESPJS connect Source Calculate (Image Resizing) Score (Classification) ModelReader (ASTORE Loading) ESP Server Viya VDMML SAS Server Retrain Analytic store Hardware & Software EEC 171_V35 • Viya 3.5 • ESP 6.2 • CentOS 7.6 Hardware • Memory: 128G • Storage: 530G • CPU: 12, 10 *2.27 Ghz Intel Xeon E7560
  • 12.
    Machine Vision Lifecycle Source Calculate (Image Resizing) Score (Classification) Model Reader (ASTORE Loading) Viya VDMML Training Once Use everywhere Database Real Time Scoring in ESP Edge Nvidia TX2
  • 13.
    Process Flow Video: 2Pics/s Source Calculate (Image Resizing) Score (Classification) Model Reader (ASTORE Loading) ESP Server Black_line
  • 14.
    Process Flow Video: 2Pics/s Source Calculate (Image Resizing) Score (Classification) Model Reader (ASTORE Loading) ESP Server 31 32 Black_line +1
  • 15.
    Copyright © SASInstitute Inc. All rights reserved. Solution Value
  • 16.
    Copyright © SASInstitute Inc. All rights reserved. Quality optimization IMPACT Inefficiency of defect analysis CONSEQUENCE CURRENT STATE Inaccuracy of defect classification Gap between quality inspector and quality analyst Hard to continuously improve the quality by embed expert knowledge into the predictive model in interactive manner Longer product quality problem solving time More manual work Limited understanding of what drives product quality Unforeseen down time is driving down margins Root cause analysis and quality issues is costing too much time and effort Reduced Brand equity  Increased maintenance cost  Increased rework cost and call back size  Reduced Production capacity resulting in a potential hit to revenue  Reduced yields  Increased down time  Increased waste/scrap Product quality problem repeating  Increased labor cost  Customer satisfaction
  • 17.
    Copyright © SASInstitute Inc. All rights reserved. Quality Optimization Continuous Optimization • Visual , interactive interface for quality analyst • Embed expert knowledge to achieve optimization Process Seamlessly, integrated and continuous quality optimization process enabled  Reduced labor cost  increased yields SAS® SOLUTION BUSINESS CHANGE ENABLED  Reduced rework and scrap BENEFITS Product Quality Reduce Yield variations Automatic , real-time defect analysis minimize product quality failures Real-time monitor • Visual real-time • interactive dashboards • Support mobile/edge device Aftersales Minimize the impact of aftersales defects  Increased brand equity  Quicker root cause determination Production Efficiency Reduce rework efforts and call back sizes – early intervention Minimize efforts and accelerate the process to determine potential quality issues Increased quality – optimize and tune the process  Prioritizes problems based on business impact  Collaborative environment enable more innovation  Increased product quality  Increased Customer satisfaction
  • 18.
    Copyright © SASInstitute Inc. All rights reserved. Hot-rolled Steel Production Saving Revenue Saved= Qualified Products Rev * FNR Improved (due to missed defects) + Defect Products Rev * Acc Increased (due to miss-classification) SAS Visual Defect Detection Value Calculator (Million USD) Revenue Accuracy Before FNR Before Hired Inspectors Total Labor Cost 11,272 62.00% 4.20% 280 4.86 Defects Rate Accuracy After FNR After Labor-Cost (Man/Year in USD) Missed Defects Rev 5.00% 90.00% 1.00% $17,340 340.87 Substandard Rate Accuracy Improved FNR Decreased Misclassified Defects Rev 0.50% 28.00% 3.20% 173.59 Customer Data Input Total Rev Saved SAS Input Acc: Model Accuracy Rate FNR: False Negative Rate Totals 519.31
  • 19.
    Copyright © SASInstitute Inc. All rights reserved. Cold-rolled Steel Production Saving SAS Visual Defect Detection Value Calculator (Million USD) Revenue Accuracy Before FNR Before Hired Inspectors Total Labor Cost 13,588 62.00% 4.20% 330 5.72 Defects Rate Accuracy After FNR After Labor-Cost (Man/Year in USD) Missed Defects Rev 5.00% 90.00% 1.00% $17,340 410.89 Substandard Rate Accuracy Improved FNR Decreased Misclassified Defects Rev 0.50% 28.00% 3.20% 209.25 Customer Data Input Total Rev Saved SAS Input Acc: Model Accuracy Rate FNR: False Negative Rate Totals 625.87 Revenue Saved= Qualified Products Rev * FNR Improved (due to missed defects) + Defect Products Rev * Acc Increased (due to miss-classification)
  • 20.
    Copyright © SASInstitute Inc. All rights reserved. Total-rolled Steel Production Saving SAS Visual Defect Detection Value Calculator (Million USD) Revenue Accuracy Before FNR Before Hired Inspectors Total Labor Cost 24,860 62.00% 4.20% 610 10.58 Defects Rate Accuracy After FNR After Labor-Cost (Man/Year in USD) Missed Defects Rev 5.00% 90.00% 1.00% $17,340 751.77 Substandard Rate Accuracy Improved FNR Decreased Misclassified Defects Rev 0.50% 28.00% 3.20% 382.84 Customer Data Input Total Rev Saved SAS Input Acc: Model Accuracy Rate FNR: False Negative Rate Totals 1145.19 Revenue Saved= Qualified Products Rev * FNR Improved (due to missed defects) + Defect Products Rev * Acc Increased (due to miss-classification)
  • 21.
    Copyright © SASInstitute Inc. All rights reserved. S p e c i a l T h a n k s Xiangqian Hu SAS Advanced Analytics R&D Cary Wenyu Shi SAS Advanced Analytics R&D Cary S t e e l D r a g o n Olivia Wang Customer Advisory SAS China Jackie Hu Customer Advisory SAS China Tianlun Gu Customer Advisory SAS China
  • 22.
    Copyright © SASInstitute Inc. All rights reserved. Thanks