tackling the last efficiency frontier in steel
TRANSCRIPT
![Page 1: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/1.jpg)
TACKLING THE LAST EFFICIENCY FRONTIER IN
STEEL MANUFACTURING THROUGH AI
Smart Steel Technologies GmbH
Am Treptower Park 75
12435 Berlin, Germany
Phone: +49 30 403 673 720
E-Mail: [email protected]
Web: www.smart-steel-technologies.com
Dr. Jan Daldrop
Team Lead
Machine Learning
![Page 2: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/2.jpg)
2
Smart Steel Technologies Company Profile
TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL MANUFACTURING THROUGH ARTIFICIAL INTELLIGENCE
Blending AI and metallurgical expertise in one team
▪ Interdisciplinary integrated team of 20 associates with …▪ … world-class AI and steel know-how …▪ … PhDs in metallurgy, math, physics.▪ Languages: English, German, RussianTEAM
REFERENCES
EXPERIENCE
BUSINESS
▪ ArcelorMittal (3 production sites)▪ Buderus Edelstahl▪ Ternium, British Steel
▪ Commissioning and optimizing steel production lines: 20 years▪ Data transformation, industrial AI applications: 10 years▪ Building up steel focus: 4 years
▪ Improving the performance of steel manufacturing operationsthrough the deployment of ready-to-use AI software
▪ Enabling steelmakers to achieve superior quality levels▪ Achieving superior energy consumption, CO2 production standards
![Page 3: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/3.jpg)
E.g., 920 continuous casting lines (1.9 Mt each) worldwide 3
▪ 3500 steel grades, from low carbon up to grain oriented electrical steel
▪ 4 major processing steps up to casting
▪ 4 to 6 thermo-mechanic processing steps after casting
Steel Manufacturing: More Than 10 Shops / Site
10 AND MORE SHOPS PER PRODUCTION SITE
![Page 4: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/4.jpg)
4
Secondary Metallurgy and Casting
Basic Oxygen Furnace• hot metal to steel• decarburization• de-Si, de-P
Ladle Furnace• temperature• chemical
composition
Ruhrstahl Heraeus• vacuum degassing• H, O
Continuous Casting• solidification• shaping
A SOPHISTICATED PROCESS CHAIN FOR HIGH QUALITY PRODUCTS
![Page 5: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/5.jpg)
5
Rolling and Galvanizing
Hot Rolling• thickness reduction• above recrystallization
temperature• mechanical properties
Surface Inspection• automated visual
inspection
Pickling• acid bath• remove oxides,
scale
Cold Rolling• thickness reduction• mechanical
properties• surface finishing
Galvanizing• zinc coating to
prevent corrosion• liquid zinc bath
A SOPHISTICATED PROCESS CHAIN FOR HIGH QUALITY PRODUCTS
![Page 6: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/6.jpg)
Steel Plant Impressions
A SOPHISTICATED PROCESS CHAIN FOR HIGH QUALITY PRODUCTS
Basic Oxygen Furnace Hot Strip Mill Galvanized Coils
![Page 7: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/7.jpg)
Building Blocks
1 CNN Surface Defect Classification
2 Training Data Optimization
3 SST Image Search
4 Online Integration
7
SST Surface Inspection AI Project
STATE-OF-THE-ART DEEP-LEARNING TECHNOLOGY FOR RELIABLE SURFACE INSPECTION
![Page 8: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/8.jpg)
CNN Surface Defect classification
DEEP-LEARNING-BASED SST CLASSIFIERS SIGNIFICANTLY OUTPERFORM EXISTING CONVENTIONAL SYSTEMS
• Existing ASIS systems typically based on classical image features
• SST uses advanced CNN ensembles optimized for steel surface defect classification
• Steel-specific image augmentation, problem-specific class weights + losses
• Transfer learning + Semi-supervised learning
![Page 9: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/9.jpg)
GOOD TRAINING DATA IS THE KEY FACTOR FOR MOST AI APPLICATIONS
![Page 10: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/10.jpg)
10
SST Defect Image Search for QUICK Classifier Tuning
QUICKLY SCAN THROUGH 500,000,000 DEFECT IMAGES
Search through 100,000,000 defect images within 100 ms
![Page 11: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/11.jpg)
SST Image Search Technology
FAST SEARCHES BY DEEP-LEARNING-BASED DIMENSIONALITY REDUCTION AND FAST ANN SEARCH
Figure: Malkov, Y.A. and Yashunin, D.A., IEEE transactions on pattern analysis and machine intelligence (2018)
Image features / embedding
• Images are indexed by CNN features(cosine distance in Euclidean space)
• Imagenet pre-training
• Select CNN architecture by class-based image retrieval on a test set (MAP, truncated VGG16)
• Dimensionality reduction: Transfer learning (of cosine distance) to a smaller CNN with less outputs + PCA
• Image augmentation during transfer learningimproves MAP
Fast approximate nearest neighbor search
• Fast image retrieval with Hierarchical Navigable Small World graphs
11
![Page 12: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/12.jpg)
UMAP Projection of Feature Vectors
CNN FEATURES SEPARATE DEFECT CLASSES WELL
Colors: different defect types
![Page 13: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/13.jpg)
Building Blocks
1 Centralized Coil Map
2 Defect Classification
3 Caster Data, Model Tuning
4Automatic Casting Parameter Optimization
5 Testing in Production
13
SST Casting Optimization AI Project
TROUBLESHOOTING INSTEAD OF SIMPLE PREDICTION
![Page 14: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/14.jpg)
14
Installing Deep CNN Classifiers and Centralized Coil Map
ALL SURFACE INSPECTION RESULTS MAPPED PRECISELY ONTO ONE CENTRAL COIL MAP
Matching of defect positions across all routes: HSM, PL, TCM, CGL, inspection lines
Red = HSM
Green = PL
Blue = CGL
Coil Selection With Auto
Completion
Deep CNN Classifiers for
all Lines
Path of Current Coil: HSM -> PL -> TCM -> CGL
![Page 15: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/15.jpg)
15
Merging Quality, Caster, Melt Shop Data, Model Tuning
MERGING QUALITY AND MELT SHOP DATA
Data Density Mold Level
Center Defects Edge Defects
Heats / Sequences
Defect Rate
Caster Signals
Mapping defect rates onto strand position
Merging caster and melt shop data
Explainable AI:
Inspecting arbitrary subspaces of caster settings
![Page 16: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/16.jpg)
16
SST Casting Optimization AI: Significantly Reduce Defects
SST AUTOMATICALLY FINDS OPTIMAL CONFIGURATIONS
Applicable to manufacturing of both, flat and long products
Permanently reduced rate of casting defects without installation of new equipment
▪ Steel analysis
▪ Speed, mould level
▪ Heat flux, water flow
▪ More casting parameters
▪ Defect images
▪ Defect positions on strip
▪ Defect types
▪ Normalized defect rates
QualityInspection
Caster Slabs Reheating Furnace Hot Rolling Mill RolledProduct
PROCESS DATA QUALITY INSPECTION DATA
AUTOMATED IDENTIFICATION OF OPTIMAL SETTINGS IN HIGH DIMENSIONAL CONFIGURATION SPACE
▪ Changes: speed / width / taper, more ▪ Ranges: heat flux, more ▪ Wear: copper plates, more
SMART STEEL TECHNOLOGIES IMPLEMENTS
OPTIMIZATION ACTIONS
Example 1:Optimized mould settings for best primary solidification
Example 2:Adapted strand cooling for best surface and internal quality
Example 3:Process chain optimized speed management
![Page 17: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/17.jpg)
Building Blocks
1 Live Data Transformation
2 Live Integration in Melt Shop OT
3 Model Tuning
4 Testing in Production
17
SST Temperature Optimization AI Project
TROUBLESHOOTING INSTEAD OF SIMPLE PREDICTION
![Page 18: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/18.jpg)
18
Conventional Temperature Optimization (This is ***NOT*** the SST Approach)
FURTHER OPTIMIZATION REQUIRES NEW METHODS
Figures: Excellent paper by Hüttenwerke Krupp Mannesmann, Peiner Träger, BFI, Research Report No. 2.31.001
![Page 19: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/19.jpg)
= Ladle 319
SST Uses 100 % of Production Data
MORE DATA LEADS TO HIGHER PRECISION AND ROBUSTNESS
H1 H2
H1
H1
H1
H1
H6
H6H2
H2
H2
H2
H7H3 H4 H5
H5H3 H4
H5H4H3
H4H3
H3
BOF/EAF
LF
T1
CCM
T2
t= Ladle 1 = Ladle 2 = Ladle 4 = Ladle 5
CU
RR
ENT H
EAT
INPUT VALUES
INP
UT V
ALU
ES
INPUT VALUES
INP
UT
VA
LUES
![Page 20: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/20.jpg)
20
SST Process Timeline View
SST MODELS COVER ALL MELT SHOP PROCESSES
SST covers all melt shop configurations from BOF / EAF to caster
![Page 21: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/21.jpg)
21
SST Temperature Guidance For BOF Steelmaking
SST MODELS COVER ALL MELT SHOP PROCESSES
Live Predictions:▪ Expected tapping temperature of current heat based on local BOF model▪ Optimal target tapping temperature of current heat based on global temperature model▪ Recommendation to operator (e.g., set-point for blow-end)▪ Chemical composition
![Page 22: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/22.jpg)
22
SST Temperature Optimization AI – Reduce Temperature Levels By 10 Kelvin
SAVE ENERGY, CO2, IMPROVE PROCESS STABILITY AND PRODUCTIVITY
SST covers all melt shop configurations and processes
No cloud and no supercomputing needed
▪ Tapping
temperature
▪ Steel analysis
▪ Slag and more
▪ Heating time
▪ Temperature
▪ Bubbling time
▪ Alloys and more
ContinuousCasting
LadleFurnace
BOF / EAF DATA LF DATA
VD / RHDegassing
▪ Vacuum time
▪ Temperature
▪ Steel analysis
▪ More values
VD / RH DATA
▪ Speed, time
▪ Mould level
▪ Cooling water
▪ More values
CASTING DATA
LIVE TEMPERATURE MODEL FOR ALL CIRCULATING LADLES
Based on ladle / tundish histories, all possible steel grades / treatments. Data visualization in web interface. API for HMI integration
Permanent
Temperature
Prediction
at BOF/EAF, LF,
RH, CasterBOF / EAF
Steelmaking
![Page 23: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/23.jpg)
23
SST Builds On Future Proof IT Architecture
PROVEN SOFTWARE COMPONENTS AND OPEN DATA FORMATS
Live integration of machine learning applications is a lot of work!
![Page 24: TACKLING THE LAST EFFICIENCY FRONTIER IN STEEL](https://reader034.vdocuments.mx/reader034/viewer/2022042417/625f5fbfa4821a78335c9405/html5/thumbnails/24.jpg)
Thank you for your attention.