Artificial Intelligence & Steelmaking Summit November 2021 by Quartz Business Media

2021

Technologies to accompany our steelmaking clients in their digital transformation journey

Virtual operator training

Eagle Eye CoatingTM A unique and intelligent coating technology

www.steeltimesint.com. S U M M I T C A TA LO G U E - N O V E M B E R 2 0 21

Digitalization

Explore the Nature of Your Supply Chain Smart Plans for Sustainable Metals Production

www.psimetals.com/planning

Software Excellence for Steel & Aluminium Producers

CONTENTS

Cover: John Cockerill Intelligent zinc coating technology, interactive process control or virtual operator training software. These are only three examples of John Cockerill’s technologies efficiently accompanying steelmakers on their digital transformation journey.

Editorial

Editor / Programme Director Matthew Moggridge +44 1737 855151 matthewmoggridge@quartzltd.com

Production Editor Annie Baker

Advertisement Production Martin Lawrence

Sales

International Sales Manager Paul Rossage +44 1737 855116 paulrossage@quartzltd.com

Sales Director

Ken Clark +44 1737 855117 kenclark@quartzltd.com

Corporate

Managing Director Tony Crinion

Published by: Quartz Business Media Ltd Quart House, 20 Clarendon Road Redhill, Surry RH1 1QX, UK +44 1737 855000 www.steeltimesint.com © Quartz Business Media, 2021

Steel Times International

Contents 3

Introduction by Matthew Moggridge, programme director.

Danieli: Every bar counts.

52 SST and PSI: Three AI-based applications for steelmakers.

Meet the sponsors. 6

Summit programme.

Everguard: Using AI to reach ESG initiatives in steel.

8 Speakers’ biographies.

Presentation abstracts.

Five reasons why steel manufacturers need machine learning.

Severstal: Looking ahead to a digital future.

IBA AG: High-resolution data acquisition is key.

PSI: Breaking the black-box nature of predictive models.

Falkonry: AI-based operational excellence.

Ternium: Managing the risks of ransomware.

Affine.ai: Machine learning technique to predict coke quality.

Posco: AI-Human ensemble operation in steel manufacturing processes.

SAP: From proof-of-concept to real use.

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

THE LEADER FOR AI IN STEEL NOW IN

AMERICA

4+-,'5&;24/<=/2=</>?</?1= 678#/(5 -+@A+B,CB9)-,$B,++#$,+&'6D#DE"+BF&D9 9'25 (((FB9)-,$B,++#$,+&'6D#DE"+BF&D9

!"#$%&!%''(&)'*+,-(-./'0&1"23 !"##"$%&'()*+$%,-).+/0 01234/5+-#"67/8+-9)6:

Welcome!

Matthew Moggridge, Programme Director, AI Summit

The Artificial Intelligence & Steelmaking Summit is our second online event of 2021 and I hope you will agree that it is top of the pile in terms of the calibre of those presenting papers, some of whom have written in-depth articles for this digital show guide. ArcelorMittal, Ovako from Sweden, Ternium from Latin America, Severstal from Russia and POSCO from South Korea are all taking part in this important summit. Specialists in the field of artificial intelligence – including the USbased Everguard.ai and Affine.ai of India – are writing and presenting papers as, indeed, are our most of our eight sponsors, who are themselves AI experts. In fact, a big thank you to SMS digital, PSI Metals, Danieli Automation, Fero Labs, Smart Steel Technologies, Tebulo Industrial Robotics, iba AG and Falkonry. Artificial intelligence is playing an increasingly important role in every aspect of steel manufacturing as operators from around the world resort to ‘high tech’ to reduce emissions, improve production efficiencies, worker safety and, of course, the quality of their end products. Steel Times International will continue its focus on steelmaking and artificial intelligence at the Future Steel Forum, which takes place live at the Grandior Hotel, Prague, 8-9 June 2022. We are currently calling for abstracts so please get in touch if you wish to speak. Matthew Moggridge, Programme Director

Steel Times International

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

MEET THE SPONSORS

DANIELI AUTOMATION

FALKONRY

Danieli Automation is a company, within the Danieli Group, responsible for the transfer of technological know-how from other Danieli technological divisions to end users, supplying the interface between plant process and operator. Our mission is to provide process automation and control systems for the metals industry covering a wide spectrum of Danieli Technology, ranging from iron ore to long and flat products.

Falkonry is a leading enabler of AI-based operational excellence for manufacturing and defense organizations looking to achieve significant improvement in production uptime, quality and yield. Leading steel manufacturers have deployed Falkonry’s operational AI in their production environment, reducing equipment downtimes and quality rejects. By analyzing process and operational data, Falkonry’s patented AI equips plant engineers with improved real time operational visibility so that they can stop the events that adversely impact operations. The initial results surface within as little as a few weeks, providing an opportunity of reducing unplanned equipment downtimes by 30-50% year-overyear. Falkonry’s products easily scale across the enterprise and can be deployed onsite, at the edge, or in the cloud, optimized for Azure and AWS IIoT platforms.

DIGI&MET

DIGI&MET is the cross-functional business unit Danieli Group has created to develop and implement new plant design concepts, based on digital innovation, and also new business models based on servitization and outcome economy principles, to ensure consistency in quality, plant utilization, OpEx and faster deliveries.

FERO LABS

Fero makes steel manufacturing more efficient through machine learning. With Fero’s easy-to-use dashboard, plant personnel get real-time recommendations and tools to prevent quality issues—no data science degree required. We’ve helped some of the world’s biggest steelmakers reduce raw material costs by more than $4/ton, improve yield and quality, and hit critical sustainability goals. For further information, log on to www.ferolabs.com/steel

For further details, log on to www.falkonry.com/metals/

For further information, log on to www.dca.it

iba AG It is our mission to bring transparency to the world of industrial production, power generation and energy distribution plants. By means of an iba system, the user can understand and master the growing technological complexity of automated processes and mechatronic systems. As with a flight recorder, all essential system and process data from various signal sources, field buses and automation

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

systems are recorded continuously and synchronously. For analyzing these data, we have developed powerful analyzing tools which comfortably support interactive work as well as automatic information generation. For further information, log on to www.iba-ag.com

Steel Times International

MEET THE SPONSORS

PSI AG

SMS DIGITAL

PSI AG develops its own software products for optimizing the flow of energy and materials for utilities (energy grids, energy trading, public transport) and throughout several industries (metals production, automotive, mechanical engineering, logistics). As a subsidiary of PSI Software AG, PSI Metals is the world’s leading provider of production management in the metals industry and has more than 50 years of experience in this field. PSI was founded in 1969 and employs more than 2,000 people worldwide.

SMS digital is a leading digital solution provider in the metals industry and develops cutting-edge digital technology to achieve operational excellence with optimized processes. As the digital unit of SMS group, SMS digital identifies and creates innovative products for the metals industry, building on the most advanced development techniques, in-depth metallurgical process know-how and technological expert knowledge. With digital applications using artificial intelligence and machine learning, SMS digital supports customers in implementing IIoT-solutions while disclosing the efficiency potentials of their learning steel plants.

For further information, log on to www.psi.de

For further information, log on to www.sms-group.com

SMART STEEL TECHNOLOGIES

Smart Steel Technologies supports the steel industry in its transformation towards intelligent AI-supported optimized production. The company delivers AI software products that improve quality, optimize energy demand and ensure accurate management of CO2 efficiency. SST excels in a portfolio of professional AI-powered optimization packages. They lead to a permanent performance increase of 5-10% per process step. For further information, log on to www.smart-steel-technologies.com

TEBULO INDUSTRIAL ROBOTICS Tebulo offers innovative robotics with the flexibility to integrate with your operations. Our systems fit yours – not the other way around. We’ll bring your business to a leading position with automation that significantly increases productivity, accuracy and safety. The world of manufacturing industries has gone global, and it’s powered by robots. It’s time to move forward. Productivity opens up when production line speeds run faster, longer and more efficiently using AI. Accuracy increases and the costly mistakes and downtime through human error are eliminated. Safety is maximized because robots do the most dangerous, Steel Times International

dull and dirty tasks. Freeing up employees for more rewarding and safer positions, all while helping companies to become more competitive. Our mission here at Tebulo is to bring the steel manufacturing industry up to a front-running position through the use of state-of-the-art robotics. Backed by our diverse industry experience and proven performance, you’ll be moving in the right direction. It’s time to move fast. It’s time to move forward. Move fast. Move forward. For further information, log on to www.tebulorobotics.com ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

CONFERENCE PROGRAMME

09:00hrs

Introduction by Matthew Moggridge, programme director

09:10hrs

Opening Keynote by Göran Nyström, executive vice president, Ovako.

09:40hrs

Boosting production with data science solutions at Severstal, by Boris Voskresenskii, chief digital officer, Severstal; and head of Severstal Digital.

10:10hrs

Industrial cybersecurity – Managing the risks of ransomware by Carlos Russell, business conduct compliance and cybersecurity director, Ternium.

10:40hrs

Scrap yards management. How digitalisation can help the new steelmaking transformation challenges by Dr. Asier Vicente, head of primary process (EAF & Scrap) department, ArcelorMittal Basque Country Research Centre, co-ordinator of ArcelorMittal Global R&D EAF programme.

11:10hrs

Coffee Break

11:40hrs

AI-human ensemble operation in steel manufacturing processes by Dr. Kisoo Kim, head, Process and Engineering Research Lab, POSCO.

12:10hrs

How artificial intelligence can simplify steelmaking complexity by Luca Cestari, manager of process control systems, Danieli Automation and DIGI&MET.

12:40hrs

Breaking the black-box nature of predictive models by Luc Van Nerom, deputy managing director, PSI Metals, managing director, PSI Metals Belgium.

13:10hrs

Lunch Break

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

CONFERENCE PROGRAMME

14:30hrs

Major advancement with new AI-based software solutions to increase production efficiency and quality – Smart Steel Technologies and PSI Metals launch three AI-based applications for the steel industry, by Dr. Falk-Florian Henrich, founder and CEO, Smart Steel Technologies; and Jörg Hackmann, managing director of PSI Metals.

15:00hrs

Turn data into value – how to improve operational efficiency in the metals industry by utilising the full potential of digitalisation, by Bernhard Steenken, CEO, SMS digital.

15:30hrs

Extended Tea Break

16:30hrs

A proactive approach: Using AI to reach ESG initiatives in steel by Sandeep Pandya, chief executive officer, Everguard.ai.

17:00hrs

AI in steelmaking – opportunities ahead by Arindam Sarkar, consulting partner, energy and resources business unit, Tata Consultancy Services.

17:30hrs

Optimizing heat chemistry in real time using explainable machine learning by Berk Birand, CEO, Fero Labs.

18:00hrs

Optimising the use of surplus gases at SSAB’s steel plant in Lulea, Sweden using artificial intelligence, by Tomas Lundberg, Department of Metallurgy, Swerim AB.

18:30hrs

Measurement data-based deep learning for steel applications – a method to predict saw blade wear by Dr. Tobias Seitz, product manager for analysis and automation software products at iba AG.

19:00hrs

AI-based Operational Excellence in steel manufacturing by Crick Waters, senior vice president, enterprise, Falkonry.

19:30hrs

Coal blend optimization and simulation using machine learning to predict coke quality by Satish Agarwal, principal Affine.ai and Ravi Teja, Affine Analytics.

20:00hrs

Closing remarks

20:10hrs

Conference closes

Steel Times International

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

SPEAKERS

GÖRAN NYSTRÖM

EXECUTIVE VICE PRESIDENT, OVAKO Göran is responsible for marketing and technology as part of Ovako Group’s management team. Ovako is a leading European producer of long special steel products for the heavy vehicle, automotive and engineering industries in the form of bars, tubes, rings and pre-components. The company has production at nine sites and sales companies globally. Göran’s past experience includes VP roles within Sandvik, managing the end-to-end supply chain (production, purchasing, logistics) of machining tools and mining products, and before that the role of VP of sales and marketing for the steel products business area. Göran was educated at Uppsala University in Sweden, and holds a Master of Science in Engineering Physics and Business Administration from the University of Western Ontario in Canada.

BORIS VOSKRESENSKII

CDO OF SEVERSTAL, HEAD OF SEVERSTAL DIGITAL Boris was previously responsible for risk management at the National Factoring Company and worked as a data scientist in the Centre of Application Data at Sberbank. In 2017 he joined Severstal as head of data science and in 2020 became head of Severstal Digital. Since 2018, he is also a Kaggle Competition Master.

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

SPEAKERS

ASIER VICENTE RESPONSIBLE FOR PRIMARY PROCESS (EAF AND SCRAP) DEPARTMENT OF ARCELORMITTAL’S BASQUE COUNTRY RESEARCH CENTRE AND CO-ORDINATOR OF ARCELORMITTAL GLOBAL R&D’S EAF PROGRAMME Dr. Asier Vicente received a BEng degree on civil engineering (2002) and in 2007 received an MEng degree on material engineering with honours from the School of Engineering of Bilbao, and a MSc degree in industrial technologies by UNED (2016) and PhD (2020). In 2002 he started at ArcelorMittal Sestao as steel mill shift manager and in 2010 moved to ArcelorMittal Basque Country Research Centre. He currently works as a principal investigator in the electric arc furnace and raw materials area, leading numerous projects. Currently, Asier is co-ordinating the EAF R&D programme worldwide, as well as the ferrous scrap and raw materials and molten phases alongside specialists within ArcelorMittal Global R&D.

CARLOS RUSSELL BUSINESS CONDUCT COMPLIANCE AND CYBERSECURITY DIRECTOR, TERNIUM Carlos Russell is business conduct compliance and cybersecurity director for Ternium. Over the last 10 years he has built Ternium’s cybersecurity programme, including industrial cybersecurity, privacy and technology risk management concerns. Prior to joining Ternium, Carlos spent 20 years dealing with complex business and technology transformation programmes in various industries, working for global management consulting companies such as PwC, Capgemini Consulting and EY in the UK. He holds a BSc in computing and information systems from Goldsmiths College (University of London) and is certified in cybersecurity (CISSP, CISM), IT Audit (CISA) and IT Governance (CGEIT).

Steel Times International

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

SPEAKERS

DR. KISOO KIM HEAD OF PROCESS AND ENGINEERING RESEARCH LAB, POSCO

Dr. Kisoo Kim is the head of the process and engineering research Lab at POSCO. He is currently working on Industry 4.0 technology such as big data and artificial intelligence, and is leading innovative research on steel manufacturing facilities and processes. He has served as a researcher at the POSCO technical research labs for 31 years, and has developed high-quality steel products. He graduated from the Department of Metallurgical Engineering at Seoul National University and obtained a master’s degree in materials science and engineering from POSTECH. He received his PhD in mechanical engineering from the University of Sheffield, UK.

LUCA CESTARI DIGITAL TRANSFORMATION MANAGER, DANIELI AUTOMATION/ DIGI&MET Luca Cestari holds a degree in electronic engineering from the University of Padua and a masters degree in metallurgical engineering. He started his professional career in Concast and later in SMS Group working on the development of Level 2 systems and numerical models for meltshops and continuous casters. He was then employed in Danieli as continuous casting process engineer. With the start-up of Digi&Met he embraced the new digitalization challenge to bring to the steel market the concepts and paradigms of Industry 4.0 and is now responsible for the development of digital solutions in quality management and product owner of an Industrial IoT platform for metals.

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

SPEAKERS

LUC VAN NEROM DEPUTY MANAGING DIRECTOR PSI METALS, MANAGING DIRECTOR PSI METALS BELGIUM

Luc Van Nerom studied mathematics and computer science and started as a researcher at the AI Lab of the Vrije Universiteit Brussel. In 1986 he created a spin-off company ‘Artificial Intelligence Systems’ bringing AI and optimization technology to the metals industry. After a merger, the products of this company have been embedded in the production management software of PSI Metals. Today, Luc is focusing on product innovation management and industrial intelligence at PSI and PSI Metals.

JÖRG HACKMANN MANAGING DIRECTOR PSI METALS Mr. Jörg Hackmann is Managing Director of PSI Metals. He has 30 years experience in the metals industry as solution architect and project manager for many MES and SCM projects. Since the beginning of 2020 Mr. Hackmann has been managing director of PSI Metals and responsible for products and technology, business consulting and global support. He holds a masters degree in applied mathematics.

Steel Times International

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

SPEAKERS

DR. FALK-FLORIAN HENRICH

FOUNDER AND CEO, SMART STEEL TECHNOLOGIES Dr. Henrich holds a PhD in mathematics and contributed substantial research to the theory of loop spaces of Riemannian manifolds and artificial intelligence.

BERNHARD STEENKEN

CEO, SMS DIGITAL After Bernhard Steenken completed his master’s degree in metallurgical engineering at RWTH Aachen University, he joined the SMS group in Düsseldorf, Germany, in corporate development until he took over the position as CEO of SMS digital in 2018.

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

SPEAKERS

SANDEEP PANDYA

CHIEF EXECUTIVE OFFICER, EVERGUARD.AI Sandeep is CEO of Everguard, an AI joint venture, backed by Boston Consulting Group Digital Ventures and SeAH Global Inc, a global steel conglomerate. Everguard’s mission is to make the world’s industrial environments safer and sustainable with machine learning/computer vision solutions, and to drive a paradigm shift in safety and other ESG initiatives from reactive to proactive. Prior to Everguard, he served as president of Netradyne, a start-up focused on commercializing machine learning/computer vision innovations in the areas of autonomous vehicles, HD mapping, and fleet telematics. Prior to Netradyne, he served as vice president of product management at Qualcomm, Inc. During his time at Qualcomm, he was instrumental in driving the introduction of more than two dozen of their leading Snapdragon smartphone and 3G/4G IOT baseband chipsets. Sandeep believes that the world is experiencing a technological renaissance and that the nexus of artificial intelligence, computer vision, ubiquitous connectivity, and access to the world’s knowledge base through the internet will fundamentally reshape and improve the way people live, work and play globally. He is determined to help lead that change.

ARINDAM SARKAR CONSULTING PARTNER, ENERGY AND RESOURCES BUSINESS UNIT, TATA CONSULTANCY SERVICES Arindam is a metallurgical engineer with 25 years of overall experience, out of which the first 15 years were in the metals industry in multiple leadership roles followed by 10 years in Tata Consultancy Services (TCS) as a subject matter expert in metals manufacturing, closely associated with the global metals customers of TCS. While in the industry, he has gained invaluable experience in flat rolling operations in business functions like production, planning, R&D and technology. In TCS, he has worked closely with the customers in their digital adoption journeys from strategy, planning and to execution.

Steel Times International

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

SPEAKERS

BERK BIRAND

CEO, FERO LABS Berk is the CEO of Fero Labs, an industrial process optimization software company based in New York. He is passionate about helping large industrial companies advance their digital transformation goals using explainable machine learning. Birand holds a Ph.D. in electrical engineering and computer science from Columbia University. His academic research focused on optimizing wireless and optical networks with efficient cross-layer algorithms. He developed scheduling algorithms for optimizing cellular base stations in 5G networks and has several patents in IoT systems for resilient fibre-optic networks.

TOMAS LUNDBERG RESEARCH ENGINEER, DEPT OF METALLURGY – RESOURCE EFFICIENCY & ENVIRONMENT, SWERIM AB Tomas Lundberg holds a PhD in theoretical physics from Linköping University, Sweden. Prior to joining Swerim he worked for Ericsson Research for 19 years, mainly with objective quality models for speech, audio, and video, or other modelling works related to user behaviour. From around 2014 this modelling work naturally led to an interest in applying machine learning and big data technologies in the models. Since joining Swerim in 2018, he has been involved in several projects related to big data, data management, and advanced analytics using machine learning for the steel industry.

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

SPEAKERS

TOBIAS SEITZ

PRODUCT MANAGER, IBA AG • • •

Product manager for analysis and automation software products at iba AG Education in mathematics, engineering and computer science. PhD in numerical mathematics with focus on computational fluid dynamics and optimization

CRICK WATERS

SENIOR VICE PRESIDENT, (ENTERPRISE) FALKONRY Crick Waters leads customer success for Falkonry and in that role he engages both customers and partners. Crick is an expert at connecting Falkonry’s cognitive capabilities to industrial yield improvement objectives due to his experience on both sides. He has successfully founded or co-founded several companies, one of which, Ribbit, provided a $100M+ exit upon its acquisition by British Telecom. He has created game-changing physical and digital products in diverse industries such as telecoms, paper, and non-wovens, and has been awarded three patents. Crick is also an InfoWeek Top-Ten Start-up award winner. His career spans nuclear engineering, industrial engineering, telecommunications, and software. He studied electrical engineering at Duke University followed by nuclear engineering at the Bettis Reactor with associated work in the US Navy. He also has an MBA from Duke.

Steel Times International

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

SPEAKERS

SATISH AGARWAL PRINCIPAL MANUFACTURING COE AFFINE.AI Eleven years of experience in Analytics and Data Science Solution development in steel and healthcare industries and across geographies. Current designation – Principal Manufacturing CoE @ Affine.ai LinkedIn – https://www.linkedin.com/in/satishagarwal15/

RAVI TEJA LEADING MANUFACTURING COE AFFINE.AI Fourteen years of experience in consulting, analytics and data science solution development in multiple industries and across geographies. Current designation - leading manufacturing CoE @ Affine.ai LinkedIn - https://www.linkedin.com/in/tejaravi3/

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

IT‘S MORE THAN JUST A MACHINE.

KOCKS rolled ®

focused

R ELIABILITY

unique

S PECIALISTS

ingenious

B ELIEF

Do you want more?

WE MAKE YOUR PRODUCT GOLD

ABSTRACTS

0910hrs Opening keynote by Göran Nyström, executive vice president, Ovako. 0940hrs Boosting production with data science solutions at Severstal, by Boris Voskresenskii, chief digital officer, Severstal; and head of Severstal Digital. Severstal Digital develops and implements machine learning models to boost productivity and efficiency of equipment at Severstal, one of the leading steel producers in Russia. The company produced 11.3Mt of steel in 2020. Last year, the implementation of data science solutions at Severstal translated into the production of more than150kt of additional finished products and a financial gain of $18.9m, compared to $9.5m in 2019 (an increase of 98.9%). There are two projeсts that are worth discussing in this field: an AI project at continuous pickling line #3 at Cherepovets Steel Mill and several machine learning models at the Kolpino production site of Mill 5000. The performance of continuous pickling line #3 has improved by more than 6.5% and the productivity of the Mill 5000 line by 5.2%. 1010hrs Industrial cybersecurity – Managing the risks of ransomware by Carlos Russell, business conduct compliance and cybersecurity director, Ternium. During the last few years international threat actors have been consistently successful in attacking global and regional industrial companies’ technology infrastructure to get unauthorized access into their networks, compromising their valuable data assets and, therefore, resulting in reputational damage, personal data protection fines, data theft and/or paying ransom for recovering their data. This session provides an overview of the risk landscape, the challenges faced by organizations and proposes insights and recommendations as to where to focus investment and resources into managing these risks in an industrial context” 1040hrs Scrap yards management. How digitalisation can help the new steelmaking transformation challenges by Dr. Asier Vicente, head of primary process (EAF & Scrap) department, ArcelorMittal Basque Country Research Centre, co-ordinator of ArcelorMittal Global R&D EAF programme. Analyzing steelmaking trends of recent years, and taking into account the sustainability and environmental aspects, as well as the new challenges imposed by the aggressive global economic situation, which has developed since 2008, the steel industry has been pushed towards continuous development of new grades of steel with higher technical performance and a lower environmental footprint. The use of ferrous scrap (considered in the past as waste) has been consolidated as one of the main raw materials for producing steel, imposing additional pressures on the steelmaking sector for two main reasons: • In the medium-long term, increased steel scrap consumption can be expected, thus conserving raw materials and energy and reducing CO2 emissions. • The available ferrous material is becoming more complex (combination of steel with plastics and fibres, more complex joints, technical coatings). This new paradigm will drastically modify the characteristics of the obsolete scrap available in the near future, worsening the quality of the material that is expected to be available in the future.

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

ABSTRACTS

Considering that around 85% of the total costs incurred by EAF plants are associated with steel shop activities, and that ferrous scrap can represent up to 60% of total production costs, to work on deploying new solutions for a better understanding scrap materials for optimizing operational decisions offers many opportunities for steelmaking plants. Additionally, the multiplication of complex artificial intelligence algorithms in the last decade, coupled with a drastic increase in computational capabilities, and the development of vision-based analytical sensors, open up new opportunities in this field. This presentation will present the strategy followed by one ArcelorMittal’s EAF plants for transforming the scrap yard into a smart facility, showing several cases of success. 1140hrs AI-human ensemble operation in steel manufacturing processes by Dr. Kisoo Kim, head, process and engineering research lab, POSCO. POSCO has applied AI algorithms in more than 90 sites in steel works since 2016. It has found that AI can enable autonomous operations in steel-making processes, but the combination of AI and existing physical models ensures more stable quality control and cost reduction; the Al-Human Ensemble Model defines this concept. The level of completion and accuracy of the Ensemble Model varies depending upon the level of collaboration between academia, SMEs, start-ups and experienced operators, which is key to a smart factory equipped with IoT technologies and AI models in the steel industry. 1210hrs How artificial intelligence can simplify steelmaking complexity by Luca Cestari, manager of process control systems, Danieli Automation and DIGI&MET. Simplifying metals complexity and improving plant efficiency, reducing time losses, production stops and eliminating the encumbrances for in-line installations: these are the main challenges of the steel market. Danieli Automation has developed many different solutions based on AI to improve the performance of a steel plant, from machine learning systems to digital twins, from production scheduling optimization to quality analysis based on artificial vision. The presentation will focus on a success story based on accurately counting the bars in a rolling mill with the use of artificial vision systems. 1240hrs Breaking the black-box nature of predictive models by Luc Van Nerom, deputy managing director PSI Metals, managing director PSI Metals Belgium Machine learning is rapidly being adopted across several industrial sectors, including the metals industry. The adoption of machine learning techniques in a real live industrial environment requires an insight into the results produced by predictive models. The maturity of the model should be confirmed by thorough inspection employing machine learning interpretability techniques. On one hand, it facilitates a deep understanding of predictive model behaviour under a variety of circumstances. In connection with the domain knowledge, it brings an efficient tool for root cause analysis. In this presentation we aim to show how to apply the aforementioned techniques to the problem of defect detection in the metals industry. In our use case the process data gathered during coil production was used for the creation of machine learning models, where the predictive target was the occurrence of a defect. These models, subsequently, were exposed for the machine learning interpretability techniques. We will show the application of these techniques, in particular how to extract business value from certain aspects of machine learning interpretations. A special emphasis will be placed on the prediction breakdown, which is a decomposition of a single prediction into the contributions from all involved predictors. This is a measure of their importance and provides precise information about the impact of a given data/process feature in the context of a particular prediction. All mentioned techniques are applied to understand the decisions proposed by the models. This allows for building confidence and trust that the predictions are fair and based on clear presumptions.

Steel Times International

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

ABSTRACTS

1430hrs Major advancement with new AI-based software solutions to increase production efficiency and quality – Smart Steel Technologies and PSI Metals launch three AI-based applications for the steel industry, by Dr. Falk-Florian Henrich, founder and CEO, Smart Steel Technologies; and Jörg Hackmann, managing director of PSI Metals. Smart Steel Technologies and PSI Metals present their three jointly developed AI-based software solutions that will help steel companies improve production efficiency and quality of advanced steel products to meet the highest market demands. PSI’s extensive metals industry experience and process knowledge is further elevated with SST’s profound AI and machine learning expertise in the steel sector. The joint solutions are achieved by combining the implementation of AI-based software with production management-related software. The new solutions digitize applications at every single stage of the steelmaking process – from the liquid phase to strip finishing. The jointly developed solutions will be applied in the areas of product-to-order-reallocation, slab and coil classification and temperature and liquid steel control. The solution for temperature and liquid steel control will lead to a more efficient use of resources, assisting the steel industry in its approach to produce green steel. 1500hrs Turn data into value – how to improve operational efficiency in the metals industry by utilising the full potential of digitalisation, by Bernhard Steenken, CEO, SMS digital. Increasing operational efficiency in complex and volatile markets requires a new way of thinking in combination with the application of disruptive technologies. Digitalization has a significant effect on the transformation of the global steel industry. By applying state-of-the-art digital solutions fueled by artificial intelligence and customer-focused services, enriched with metallurgical domain expertise, steel producers improve plant performance in terms of availability, yield, product quality and energy efficiency. By that the profitability and sustainability of the metals industry can be pushed to the next level and thus scrap, downgrading and energy waste will be reduced. SMS digital has developed a cutting-edge digitalization framework that interlinks various areas of optimization such as asset and process condition and product quality, production planning and energy management creating synergies that go far beyond. 1630hrs A proactive approach: Using AI to reach ESG initiatives in steel by Sandeep Pandya, chief executive officer, Everguard.ai Now, more than ever, the world is not just interested in what steel you make but how you make it. We hear cries for the steel industry to reduce carbon emissions and carry out net zero promises that have been made. But these environmental initiatives are just one of the three pillars that comprise the move to using environmental, social, and corporate governance (ESG) factors as part of the buying and investing process. Within those social criteria, the safety of those making the steel is a major concern for buyers and investors. Artificial Intelligence (AI) has the power to revolutionize safety in steel and help steelmakers meet and exceed ESG criteria. In this presentation, we will cover: 1. The role of ESG factors in buying and investing in the steel industry. 2. How industrial safety can move from a reactive to a proactive process using AI-powered by sensor fusion, and computer vision. 3. The most critical use cases where AI can impact safety in steelmaking. 1700hrs AI in steelmaking – opportunities ahead by Arindam Sarkar, consulting partner, energy and resources business unit, Tata Consultancy Services. Over the past two decades, the steel industry has gained a lot of momentum in adopting digital technologies to solve operational challenges across the value chain. To facilitate this journey to scale up the digital maturity curve, artificial intelligence and machine learning are playing a key role by opening up new horizons of possibilities. This presentation analyzes some of the possibilities

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

ABSTRACTS

that lie ahead for the steel industry and depicts the contribution made by Tata Consultancy Services to create value for their customers in the steel industry. 1730hrs Optimizing heat chemistry in real time using explainable machine learning by Berk Birand, CEO, Fero Labs. Changes to product chemistry boundaries are made only a few times a year. However, during the period between chemistry modifications, there are still opportunities to optimize chemistry limits within the boundaries for cost and product performance measures. This includes accounting for process and raw material cost fluctuations, which have a direct impact on operating margins. Machine learning software optimizes alloy additions during the refining stage in the melt shop, ensuring that mills maximize cost savings while continuing to meet mechanical property specifications. This presentation will discuss the details of this novel technology and how it’s been deployed to five mills in North America in real-time optimization mode. In addition, we’ll share the savings achieved as a result of integrating ML into the chemistry optimization process. 1800hrs Optimising the use of surplus gases at SSAB’s steel plant in Lulea, Sweden, using artificial intelligence, by Tomas Lundberg, Department of Metallurgy, Swerim AB. In steel production, a number of energy-rich gases are formed as by-products. Most of the gases are collected and stored in gas holders and used internally at various process units inside the steel plant. When excess process gases exist, they are usually used externally as energy carriers. In general, the energy system at an integrated steel plant is very complex and non-linear as the process gases are generated from various process units, either continuously or discontinuously with varying energetic content, flow rate and chemical analysis. Therefore, it is quite difficult to model and optimize the energy system dynamically. At SSAB’s integrated steel plant in Luleå, the excess process gas is delivered to a combined heat and power (CHP) plant to produce district heating for the surrounding community and electricity. However, due to the limited capacity of the gas holders and the lack of dynamic interaction of related process units, from time to time some process gases need to be flared. Due to this, external fuels such as oil or LPG might then be required at the CHP plant to compensate for the lack of process gases. An improvement of the regulation of gases could lead to an increase in the utilization of the surplus gases, which in turn would reduce oil consumption at the CHP. The benefit here is thus twofold: instead of flaring gas at SSAB, it is used to reduce oil consumption at the CHP plant; consequently it will reduce both fossil CO2 emissions and production cost. The purpose of this work is to optimize and visualize the complex energy system by using an artificial intelligence approach. A system for energy visualization with AI, ‘EnVisA’, will be createda real-time view of the current energy performance and the availability of gases. In addition, EnVisA is will optimize and predict the supply of surplus gases for the CHP plant, thus contributing to a more climate-friendly district heating system by joint efforts from involved stakeholders, ie, the steel plant, CHP plant, and the community. 1830hrs Measurement data-based deep learning for steel applications – a method to predict saw blade wear by Dr. Tobias Seitz, product manager for analysis and automation software products at iba AG. We will introduce an architecture as well as our approach of using AI-based assistance systems that combine expert knowhow with AI algorithms. We will show how, by providing meaningful information to the operators during production, a quick return on investment can be achieved.

Steel Times International

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

ABSTRACTS

1900hrs AI-based Operational Excellence in steel manufacturing by Crick Waters, senior vice president, enterprise, Falkonry. Thanks to the rapid adoption of digital technologies in steelmaking, a typical modern steel plant is capturing operational data at a volume and velocity like never before. Yet there are numerous performance gaps leading to operational blindspots – unplanned downtimes still happen, process variances go undetected and quality defects lead to scrap. Just like having a wearable fitness tracker doesn’t make one a doctor, operational data without any intelligence cannot provide the required preventive or curative insights at the right time. What is required is real time operational visibility and actionable insights that can prevent unwanted events before they happen. Join this presentation to learn about Falkonry’s experience in scaling AI in production and the impact it has on business outcomes by minimizing unplanned downtimes by up to 50% for its customers. The presentation concludes with some lessons learned about integrating the operational AI solution to existing decision making systems and what it takes to make the AI on the ground more trustworthy and adaptable. 1930hrs Coal blend optimization and simulation using machine learning to predict coke quality by Satish Agarwal, principal Affine.ai and Ravi Teja, Affine Analytics. Coal is a raw material that plays a very important role in the coke making process. Four to five coals are blended to produce coke. The coal blend optimization model, developed by Affine Analytics, is based on coal properties and process parameters and is developed based on linear programming. The model takes optimized coal properties as input alongside the process parameters and predicts the coke CSR on a daily basis, based on historical data. Programme up to date at the time of going to press.

CONNECTING THE INTERNATIONAL STEEL INDUSTRY Published eight times a year, Steel Times International examines every aspect of iron and steel manufacturing processes, as well as bringing you the very latest market and company developments

CHOOSE YOUR SUBSCRIPTIONS PACKAGE... PRINT + DIGITAL SUBSCRIPTION 8 print copies a year posted to your chosen address Our annual Steel Times International Directory Digital issue of every magazine and two special digital only supplements sent directly to inbox Regular premium content including podcasts, video interviews and specialist reports

Access to our digital archive of past issues An annual yearplanner The weekly email newsletter Exclusive subscriber only content Priority places on industry webinars

DIGITAL ONLY SUBSCRIPTION With a digital subscription you receive all the benefits of the Print + Digital option, excluding print copies!

Discount to attend the Future Steel Forum

PLUS...access to all digital issues of our sister publication Furnaces International

Email subscriptions@quartzltd.com to take advantage of our special subscription packages

WWW.STEELTIMESINT.COM/SUBSCRIPTIONS

@SteelTimesInt BUSINESS MEDIA

Our goal, help our clients produce sustainably For us this means helping our clients to focus on quality and performance, as well as improved operatability, while minimizing the use of resources, energy and emissions throughout the lifecycle of the metal processing equipment & lines that we supply.

The first continuous CGL in China to produce heavy-gauge HR steel coils with high corrosion protection ZM Coatings, & the ingenious zinc coating system Eagle Eye Coating™ providing unequaled steel strip flatness & coating weight control, are prominent examples of our desire to provide efficient responses to the needs of our clients.

JOHN COCKERILL INDUSTRY METALS Reheating Furnaces Pickling Acid Regeneration Cold Rolling Strip Processing Automation

Design | Engineering | Erection & Commissioning I Technical Assistance & Training | After Sales

johncockerill.com/industry

20210511_JC Industry Metals_Help Clients Produce Sustainably_Final.indd 1

31/05/2021 11:19:10

ARTIFICIAL INTELLIGENCE SUMMIT

Looking ahead to a digital future How a Russian steel major applies digital solutions to boost productivity, improve product quality and reduce emissions. By Boris Voskresenskii* As technology advances, the global steel industry is constantly innovating to improve productivity and efficiency, reduce emissions and improve safety – and Russian producers are ahead of the curve in applying digital solutions. By implementing data science solutions in its business, in 2020 Severstal, a major vertically-integrated steel maker, produced over 150kt of additional finished products and delivered $18.9 million in additional profits, up from $9.5 million in 2019 (a 98.9% increase). In this article, Boris Voskresenskii describes how Severstal is using artificial

intelligence, machine learning and advanced computer vision to strengthen a competitive advantage, decrease production time and improve its product offerings. Boosting the productivity of a hot rolling mill with machine learning Severstal produces large diameter pipes (LDP) for the energy industry from metal processed at Mill 5000, at the company’s Kolpino production site in St Petersburg. To improve the mill’s productivity, the company has developed a set of machine learning (ML) models that together control

three elements of the production line – the heating furnaces that are responsible for the heating on the slabs, the transportation line that delivers the slabs, and the four-high stand that controls the rolling process. The rolling speed is controlled by a complex of ML models – gradient boosting to predict the maximum motor currents that drive the rolls; the nearest neighbour algorithm to estimate the approximate speed based on historical data, and a number of linear regression models that use the casual inference approach to determine the influence of speed on the maximum current.

* CDO of Severstal, CEO of Severstal Digital.

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

The maximum rolling speed increased from 3.2 m/s set by the operator, to 4.5 m/s set by the model. The control functions for the rolling mill have been automated, with the rolling speed maximized to match the energy capabilities of the units. To synchronize the operation of the furnaces with the mill, the company has developed models based on linear regression that predict the rolling time of the current group of slabs. As a result, the average pause downtime between rolling groups has been reduced by over a third to 40 seconds, with minor deviations; previously the average downtime was 62 seconds. Overall, this combination of solutions and their synergies have resulted in a 5.2% increase in productivity at the mill, resulting in a higher number of rolled slabs produced. In 2021 Severstal added a model that controls the heating intensity of the furnaces based on the analysis of the current conditions, and the conditions at which the furnaces are expected to be most efficient based on historical data. When the situation changes, the model automatically adjusts the heating intensity to quickly achieve uniform heating – reducing the time it takes to reach the target temperature by up to one hour.

Steel Times International

With the furnace model the company expects to boost productivity by an additional 1.5 %. Detecting quality defects with advanced computer vision The timely and accurate detection of steel surface defects is essential to maintain high standards of product quality control. In order to improve product quality and increase its competitive advantage, Severstal has implemented a disruptive deep learning technology that recognises steel surface defects using advanced computer vision. The technology is operating at Severstal’s Mill 2000 in Cherepovets, which produces 65% of the asset’s hot rolled coils. To implement the technology, the company adapted modern computer vision (CV) models, and introduced algorithms to detect anomalies, as well as the human-inthe-loop (HITL) concept, a model requiring human interaction. The operator trains the model by ‘liking’ or ‘disliking’ its predictions, which provides feedback for the algorithm – enabling the model to attain a human-level performance of detection for certain defect types and determine whether coils meet client quality requirements. The solution can be scaled to all hot and cold rolling lines

using transfer learning, and two systems are currently in production, with a third now under development. Using complex state-of-the-art algorithms Severstal’s in-house developed solution is achieving impressive results – the system detected 244% more real defects and 11% fewer false defects than the previous proprietary software. The company stores information about the defects in its Data Lake, the largest data storage system in the Russian industry, for use in future big data projects. Reducing energy consumption of air compressors with machine learning Air compressors at Severstal’s Cherepovets steel mill consume more than 50,000 kWh per year. The variable consumption of compressed air by the mill’s units results in constant changes to the compressors’ operating modes, with the redistribution of loads impacting the total consumption of electric power, and the compressor’s efficiency dependent upon its design, technical state and environmental conditions. Traditionally, operators determine how to distribute the load between compressors based on their experience, taking into consideration many factors defining the

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 25

ARTIFICIAL INTELLIGENCE SUMMIT

optimal distribution for the units. To solve this issue, Severstal has introduced an optimization model based on machine learning to control the individual load of air compressors operating in a group, and thereby reducing the total electric power consumption. Several machine learning models were developed to predict the performance of individual compressors. The models consider the operation of each compressor individually, in addition to as a group, allowing them to adjust to changes in external conditions, equipment wear and other factors that might impact performance. As a result, the electric power consumption of the air compression is minimized while maintaining the volume of the air at the intake, with the model recommending optimal distribution among the compressors for adjustment. Within the first month of the pilot, the compressors’ specific energy consumption was reduced by 1%. After a year, with further improvements to the model efficiency, the total specific energy consumption of the air compressors was reduced by 2.5%, corresponding to energy savings of 15.3 GWh per year and an annual decrease of 8,400 tons of CO2 emissions. The model has been applied at 70% of all of Severstal’s air compressors.

Steel Times International

Improving efficiency with artificial intelligence and reinforcement learning The Cherepovets steel mill’s continuous pickling line #3 (NTA-3) processes over 1Mt/yr – around 50% of all metal for cold rolling. While the new pickling line was under construction, the increase of the productivity of the pickling shop became crucial. Previously, the speed of the unit was set for each strip based on table values, with the unit’s operator adjusting the speed manually. This process is now automated, with the speed controlled by an AI model and adjusted automatically every second – this enables the model to react to unexpected situations immediately, improving the flexibility and safety levels in the unit. To implement this, the company developed a mathematical model, named Adelina. Adelina monitors over 100 parameters at all times, which the model uses to control the speed of the unit. More recently, the company brought in an AI module that uses reinforcement learning (RL) – named Ruban. Ruban operates in collaboration with Adelina. It was trained by exploring a digital twin environment of NTA3, learning how to choose the optimal speed. The training system used by Ruban is based on a system of rewards and penalties, and

the model experimented to find a solution to gain maximal reward. In real-time, Adelina and Ruban work together, with Adelina calculating the initial speed, and Ruban adjusting it to produce the best result. This approach is innovative for the steel industry, and since the first pilot productivity has improved by more than 6.5% – resulting in 100kt of additional steel produced in 2020. The solution has now been modified and implemented on one of Severstal’s cold rolling mills. Looking ahead: a digital future With digital solutions becoming increasingly integrated into the everyday workings of the company, taking on complex decisions and high-frequency control, improving safety standards and providing real-time quality assessment, Severstal is continuing to invest in the latest technologies across its business. In 2021, the company allocated $124 million towards investments in IT and digital projects, including automating production, supply chain management and digital tools. This ties in with Severstal’s ambitions to expand digital solutions further, providing increased productivity and flexibility in the product mix, as well as applying digital solutions in new areas. �

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 27

ARTIFICIAL INTELLIGENCE SUMMIT

Breaking the black-box nat predictive models

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

ture of

Machine learning (ML) has been rapidly adopted across several industries. The adoption of ML techniques in real live industrial environments requires an insight into the results produced by predictive models. The maturity of the model should be confirmed by thorough inspection employing ML interpretability techniques, which provide a deep understanding of predictive model behaviour under a variety of circumstances. What is more – and in connection with the domain knowledge – it brings an efficient tool for root cause analysis. In this article we aim to show how to apply the aforementioned techniques to the problem of defect detection in the metals industry. By Paulina Wawdysz*, Grzegorz Miebs*, Luc Van Nerom*, and Rafał A. Bachorz* In our use case, the process data gathered during coil production was used for the creation of ML models where the predictive target was the occurrence of a defect. These models, subsequently, were exposed to ML interpretability techniques. In this article, therefore, we will show the application of these techniques, in particular how to extract business value from certain aspects of ML interpretations. A special emphasis will be placed on the prediction breakdown, which is a decomposition of a single prediction into the contributions from all involved predictors. This is a measure of their importance and provides precise information about the impact of a given data/process feature in the context of a particular prediction. All mentioned techniques are applied to understand the decisions proposed by the models. This brings the confidence and trust that the predictions are fair and based on clear presumptions. Defect detection Defect detection in the metals industry is a very important step to ensure product quality. Defects and defect types can be identified via manual visual inspection (investigated by human), machine vision (investigated by machines), various types of machine learning or other semi-manual procedures [1]. Flaws may be originated and detected at any stage of coil production (Fig 1). Applying machine learning to detect the defects in different stages of the coil

production could increase the efficiency and reliability of the defect inspection, lower the costs, allow to detect defect patterns, and contribute to an overall increase in productivity. Proper management of defective coils can bring significant benefits. Depending on the nature and severity of the defect, the coil can be directed into a different production line or, in the extreme case, turned back into the very beginning of its life cycle. Machine learning approaches ML techniques are capable of embedding an overall past experience into an object called the predictive model. Just by having the data and relevant statistical model architecture we can create a tool that can reliably predict how probable the defect is. One can potentially take into account all the available information: • the chemical composition of the heat or the slab • the metallurgical data • the production process sensors data (time series) • the equipment data • the wear of tundishes and rolls • the maintenance history • the defects data – origin line, detection line • the product and production order data, the description of the final customer, etc. All gathered data describing the

* PSI Metals Steel Times International

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 29

ARTIFICIAL INTELLIGENCE SUMMIT

Fig 1. Stages of coil production

Fig 2. Model exploration stack by DALEX [4]

coils produced in the past as well as the discovered defects was turned into a matrix containing coils in rows and features in columns. The matrix might include the variety of cases, ie the defect-free coils, the coils with oxidation defects, dimensional defects, and so on. After pre-processing, a data frame containing all information describing the coil (called the set of predictors) and occurring defects (the targets) is formed. Within the use-case discussed here we have chosen one of the available machine learning techniques in order to transfer the knowledge from the data into the statistical model. The knowledge transfer is usually carried out

through a learning process. In particular, one can distinguish the supervised learning, unsupervised learning, semi-supervised learning and reinforcement learning [2]. For structured and labeled dataset, supervised learning is the most appropriate approach. Within this approach, the classification and regression problems can be addressed. The binary classification task involves a class that is considered ‘normal’, and the other class that reﬂects the anomaly. In the context of the problem presented in this paper, the normal class is related to the defect free coil (labeled as ‘0’) and the anomaly class – coils with defects (labeled as ‘1’). In order to solve the

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

classification problem, several algorithms were considered[3], in particular: the support vector machine (SVM), random forest (RF), gradient boosting machine (GBM), and eXtreme gradient boosting (XGB). The details about these methods can be found elsewhere [4], [5], [6] . For the presented research the Extreme Gradient Boosting approach was chosen as a final method. Machine learning interpretability A machine learning model can be considered as interpretable if we can easily determine the reasons for decisions made by the model in the prediction process. The

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

Language for Exploration and eXplanation) [7] . This tool helps analyze the predictive model’s behaviour and offers an entire stack of methodologies. Some of them can be applied at the general data set-level, the other locally, for particular prediction (Fig 2). The predictive model created here was carefully inspected by machine learning interpretability techniques and both instancelevel and dataset-level exploration was performed. The results provide the user with an in-depth insight into the nature of the predictions. The domain expert experience can be confronted with the decisions made by the statistical model.

Fig 3. The histogram plot and the box plot for Feature_431

Table 1. The predictive model quality measures

models employing the techniques such as linear regression, logistic regression and decision trees are potentially interpretable, the application of appropriate ML interpretability tolling make them fairly easy to understand for humans. Interpretability in machine learning allows for better understanding of the decisions made by the created model. This understanding builds the trust and, to some extent at least, alleviates the unwanted black-

box nature of the predictive model. Moreover, understanding the prediction opens the doors to finding remedies for problems in the process, in particular the features that inﬂuence most defect occurrences. This knowledge can be coupled with the domain perspective later on and used for the defect Root Cause Analysis. In order to explore and explain the predictions served by the model we employed the DALEX library (moDel Agnostic

Fig 4. Precision recall curve

Steel Times International

The data The presented results are based on anonymized data from real production plant. The time series reﬂecting the process signals were turned into the set of static descriptors (mean, median, standard deviation, kurtosis, skewness, etc.), and this time-independent data frame was a subject of severe anonymization. The final dataframe contained 500 observations, and each was defined by 680 features including the physical characteristics of coils, the process data turned into a set of statistical features mentioned above, equipment data and defects data – origin line, detection line. An example of the distribution and the box plot for selected variable (Feature_431) is shown on Fig 3. ML modeling – techniques The data described above was utilized in order to create an efficient classifier capable of distinguishing between defectfree and defected coils. As mentioned before, for this purpose the eXtreme Gradient Boosting method was used [8] in the frame of a standard supervised machine learning loop. According to the predictive modeling workﬂow, the initial raw data was acquired from existing systems and appropriately processed. Then, after the anonymization, the data was the subject of data cleaning and ultimately became the final data set which can be used in the training and inference phases. The next step was the actual training and hyper-parameters optimization of the chosen XGBoost model. According to good data science practices, the dataset was divided into training and test sets. The former

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 31

ARTIFICIAL INTELLIGENCE SUMMIT

Fig 6. Partial dependence plots for selected features

Fig 5. Variable importance plot

Fig 7. SHAP plot for “Feature_206”

was applied exclusively within the training process. The latter, being the part of data that was not exposed to the model during the training, was used for estimating the quality of the resulting model. Results As already mentioned, the quality of the obtained predictive model was checked against the test data set. The results are shown in table 2.

Each binary prediction can be a true positive, true negative, false positive or false negative. The precision reﬂects the ability of the classifier not to label as positive the sample that is, in fact, negative. Recall reﬂects the ability of the classifier to actually detect the positive samples. The F1 score is simply the harmonic mean of precision and recall. The precision-recall curve presents

Fig 8. SHAP plot for “Feature_431”

the relation between those two metrics at a variety of thresholds – probabilities level separating the false and true classes (Fig 4). The ROC is the so-called receiver operating characteristic (ROC) and it reﬂects the relationship between the true positive and false positive rate. The ROC AUC and the AP AUC are the integrals (areas) under the respective curves. Accuracy, on the other hand, is an overall measure of all correctly identified observations (true positives and true negatives). All presented quality measures were obtained within the 4-fold cross-validation procedure which makes these quantities as objective as possible. Clearly, the obtained model poses noticeable predictive strength which is proven by relatively high values of ROC AUC and AP AUC. At the same time, one can easily notice that this is not a perfect model and for sure there is room for improvement.

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Feature importance The variable importance plot shows the relative importance of the most important features (Fig 5) calculated as a drop in the model performance after a permutation of a given feature. The information is available for all features, but for obvious reasons, only the top part matters here. It is noticeable that variables ‘Feature_431’ and ‘Feature_192’ have the most significant impact on the decisions made by the predictive model. These are clearly the most decisive features, and the significant fraction of the prediction strength is related to them. One can also notice that the strength of the features decays quickly here, which means that the predictions can probably be explained by just a few of them. Partial dependence plots In order to better understand the relation between the feature and the model output,

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

Fig 9.

Breakdown plot for defect #1

Fig 11. Breakdown plot for non-defect #1

the partial dependence plots (PDP) were created (see Fig 6). They give a graphical depiction of the marginal effect of a variable on the response, and the effect is calculated in terms of the change in the mean response of the model. Thus, on the y-axis there is a defect probability as a function of the feature value. One can easily find the threshold values above which the defect probability increases rapidly, which can be somehow related to the technological process. The four most important features were selected for presentation, namely: ‘Feature_431’, ‘Feature_192’, ‘Feature_206’, ‘Feature_394’ (Fig 6). Shapley values Shapley additive explanations (SHAP) plots for the whole dataset show the contribution of the features for each instance (or row of data). Within a slightly simplified picture, the SHAP values measure the importance of the feature by comparing model efficiency with Steel Times International

Fig 10 Breakdown plot for defect #2

Fig 12. Breakdown plot for non-defect #2

and without a particular feature[9]. It is worth noting that the importance of the feature is now provided for different values of this feature, clearly giving an additional insight into the model, compared with the feature importance information discussed earlier (see Section 3.3.2.1). Feature importance for single prediction In order to check feature importance distribution for particular observation, model predictions were visualized with the prediction breakdown plots (Figs 9, 10, 11, 12) These plots show the contribution of each variable into the total defect prediction. Two coils with defects and two without defects were taken into consideration. For both presented defects (Figs 9, 10), the most significant impact on model decision is related to variables ‘Feature_431’ and ‘Feature_192’ which agrees well with the feature importance picture. The prediction

made for defect #1 depends strongly on ‘Feature_438’, which is not the case of defect #2. For both non-defects cases (Figs 11, 12) the contribution from features is weaker and there are no clearly dominating features. This is, in a way, an expected behaviour. Since the expected probability should be at low level, there should not by any strong contributions. Ceteris paribus for single prediction In order to investigate the inﬂuence of a particular feature with the assumption that the values of all other variables do not change, the Ceteris paribus method was examined. ‘Ceteris paribus’ means ‘things held constant’. As previously, the Ceteris paribus plots were created for two defect cases and two non-defect observations. The four most important features were chosen as illustrative examples. Figs 14, 15 show how changes in the values of chosen variables affect the model’s prediction. Dots indicate the values of the feature on the horizontal axis and the

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 33

ARTIFICIAL INTELLIGENCE SUMMIT

Fig 13. Ceteris Paribus for defect 1

Fig 14. Ceteris Paribus for defect 2

Fig 15. Ceteris Paribus for non-defect 1

respective probability on the vertical axis. For ‘Feature_431’ and ‘Feature_192’ there is a cut-off value above which the model’s prediction changes significantly. In the case of ‘Feature_206’ and ‘Feature_394’ changing the feature value does not affect the predictive response of the model. In the case of both non-defect observations, for all considered variables one can find the critical values, for which there is a probability change. The magnitude of this effect varies and clearly depends on the feature. This again provides a deeper insight into the predictive model. One can estimate the prediction variability as a function of a certain feature, from the perspective of the given observation. Root cause analysis considerations At a very high level of consideration, one can consider the predictive model as an entity that embeds the industrial reality into a

Fig 16. Ceteris Paribus for non-defect 2

single object. This object has a well-defined interface, ie for a given set of predictors, it produces well-defined output, for example, the defect probability. The predictive model was trained on a certain portion of historical data according to time; moreover, the data provided was most likely to be just a subset of all the data available in terms of features/process variables. This means that the predictive model perspective is rather limited as the model was equipped with the necessary information in the context of the predictive task, ie the defect detection. Therefore it is unjustified to claim that the analysis of the predictive model behaviour can bring real root cause analysis. However, as mentioned above, the predictive model encapsulates a fraction of the industrial reality landscape. By means of ML Interpretability techniques (see Section 3.3) one can get an insight into the prediction structure. The feature space can be carefully inspected

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

and one can point out the areas which are more susceptible to defect occurrence. In particular, the prediction breakdowns presented in Figs 8, 9 show significant contributions from variables ‘Feature_431’ and ‘Feature_192’. This means that these features play an important role in inferring information about the defect. The importance of these features is also confirmed at the general model level, ie by the feature importance plot (see Section 3.3.2.1). But one should keep in mind that correlation does not imply causation. Within the limited landscape of the model, it is impossible to resolve the reason for this correlation (ie what is the causation). It is impossible to judge whether the defect is caused by this particular value of ‘Feature_431’, the presence of the defect causes this value, or neither of them. One can only state that there is a relationship between defect occurrence and the fact that a certain feature is within a given value range. From a

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

practical point of view, however, this can be of noticeable business value. If the domain expert is able to understand the particular behaviour of the feature, for example that some process value tends to behave differently for certain defect types, then perhaps the same domain expert can remove the real reason and thus remove the defects in the future. Summary In the presented study we have shown the application of ML interpretability for a particular use case, i.e. the defect detection on coils. The data used for this research came from a real production plant, but was the subject of anonymization. The obtained predictive model clearly has noticeable predictive strength, which was proven by the values of considered quality metrics. On top of this model, the ML interpretability techniques were applied in order to shed light on the prediction structure. These techniques allow for deep inspection of the predictive model behaviour, both at the general data/ model level and at the level of particular

Steel Times International

prediction. Actually, all the explanations showed the importance of two variables, namely ‘Feature_431’ and ‘Feature_192’. This should be considered as a clear indication that these variables have a clear relationship with the defect predictions, and thus they are natural candidates for further inspection by the domain experts. Finally, we provide our perspective of the connection between ML interpretability and root cause analysis. We emphasize clear limitations of the predictive analytics and, at the same time, the necessity of the incorporation of domain knowledge. As a final remark, it should be noted that the application of ML interpretability can significantly support the understanding of the industrial reality. It also makes the predictive model behaviour clear and explainable, which really makes them white-box, rather than black-box. �

References [1] W. Zhao, F. Chen, H. Huang, D. Li, W. Cheng, “A New Steel Defect Detection

Algorithm Based on Deep Learning“, 202103-22. [2] J. Brownlee , “14 Different Types of Learning in Machine Learning“, 2019-11-11 [3] T. Ghosh, “Machine Learning Algorithms: A comparison of different algorithms and when to use them“, 2018-05-26. [4] A. Liaw, M. Wiener, “Classification and regression by randomforest”, R News s, 2(3):18–22. [5] M.N. Wright, A. Ziegler, “A fast implementation of random forests for high dimensional data in C++ and R”, Journal of Statistical Software, 77(1): 1–17. [6] C. Corted, V. Vapnik, “Support-vector networks”, Machine Learning, 20(3):273– 297. [7] P. Biecek, T. Burzykowski, “Explanatory Model Analysis. Explore, Explain, and Examine Predictive Models. With examples in R and Python.“, https://ema.drwhy.ai/, last access 2020-12-12. [8] XGBoost Documentation, https://xgboost. readthedocs.io/en/latest/, 2021. [9] SHAP documentation, https://shap. readthedocs.io/en/latest/, 2018.

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 35

ARTIFICIAL INTELLIGENCE SUMMIT

Managing the risks of ransomware During the last few years international threat actors have been consistently successful in attacking global and regional industrial companies’ technology infrastructure to get unauthorized access into their networks, compromising their valuable data assets and therefore, resulting in reputational damage, personal data protection fines, data theft and/ or paying ransom for recovering their data. This article provides an overview of the risk landscape, as well as the challenges faced by organizations, and proposes insights and recommendations as to where to focus investment and resources into managing these risks in an industrial context. By Carlos Russell*

* Business conduct, compliance and cybersecurity director, Ternium.

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

If you happen to look into a dictionary definition, for example the good old Oxford English Dictionary, you will find that ‘ransomware’ is succinctly defined as ‘a type of malicious software designed to block access to a computer system until a sum of money is paid’. This means that someone creates code to ensure you cannot get access to your computers, servers, databases, or critical and confidential files until you pay a ransom. In spite of what you may have heard in the press over the last few years, ransomware is hardly a ‘new threat’. If we turn our clocks back to 1989 (yes, THAT far back), people in the US healthcare sector were struck by what was called the ‘AIDS Trojan’, a trojan virus that was distributed in 20,000 infected floppy disks (millennials can Google what they were). Once someone used an infected floppy disk, the virus encrypted the hard drives and posted a message to send $189

Steel Times International

to a P.O. Box in Panama to restore access to their computers. Fast forward to 2021, a successful ransomware attack on US Colonial Pipeline disrupted fuel supply to the east coast of the US for several days. The CEO decided to pay the ransom (circa $4m) and the security forces managed to trace and recover half of that money. Whenever you read press coverage about a ransomware attack you will also find they mention that there is an obscure gang – the threat actors – that claims responsibility, noticeably with funny names (i.e REvil, Avvadon, Conti, Babuk, DarkSide, etc). These full-time organised criminal groups are operating usually under friendly jurisdictions that will neither prosecute nor extradite them. They are either mercenaries, governmentsponsored or completely rogue, with the sole purpose to gain ransom money. According to a recent report from Palo Alto Networks,

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 37

ARTIFICIAL INTELLIGENCE SUMMIT

the average ransom payment rocketed this year to $570,000. https://www. paloaltonetworks.com/blog/2021/08/ ransomware-crisis/ . Ransomware gangs are very successful as they usually target people first, mostly employees, either by baiting them with a phishing e-mail or directly compromising their passwords to connect to their employers’ network. During 2020 and 2021, the pandemic sent millions of employees to work from home, and more often than not, their home networks were not protected to the same standard as a corporate network. They became easy prey. Once they have a malicious foothold on a corporate computer, they will attempt to disseminate the malware to the nearest network server, or as many of those as they can. Once installed the malware ‘calls home’ and retrieves the special encryption codes for this particular attack and the gang will set a date and time for a co-ordinated simultaneous encryption, maybe stealing some data beforehand. Once encrypted, the ransom demand is set – the victims receive the single or double extortion demand –‘pay up or you won’t recover your data’, and maybe ‘pay up or we will publicly disclose your data’. Meanwhile, the gang could also be selling the victim’s data to other criminal third parties over the

Dark Web – you can quickly understand why these threat actors are very successful and why a ransomware attack quickly becomes a costly, reputationally damaging, careerlimiting, serious business risk. Two specific attacks in the thousands recorded during 2021 can shine a light on the potential impacts of cyberattacks: 1) Water treatment plant in Florida, US – an intruder boosted the level of sodium hydroxide in the water supply to 100 times higher than normal. 2) Ireland’s Health Service Executive (HSE) – government organisation that runs all public health services in Ireland shut down IT systems in the wake of a significant ransomware attack. These two had potentially devastating impacts on the population, by either poisoning running water or disrupting health services during the peak of the pandemic. If somebody is prepared to put people’s lives at risk, would they care too much about putting businesses at risk? It turns out that industry technology analyst Gartner recently published a paper where they predicted that by 2025 cyber attackers will have weaponised Operational Technology (OT) environments to successfully harm or kill people:

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

“Gartner also predicts that the financial impact of Cyber-Physical attacks resulting in fatal casualties will reach over $50 billion by 2023. Even without taking the value of human life into account, the costs for organizations in terms of compensation, litigation, insurance, regulatory fines and reputation loss will be significant. Most CEOs will be personally liable for such incidents.” https:// www.gartner.com/en/newsroom/pressreleases/2021-07-21-gartner-predicts-by2025-cyber-attackers-will-have-we Risk management Steel manufacturing businesses are not exempt from these risks. Companies have invested in automation, productivity and performance improvement, digitally transforming the production shop floor, increasing personnel health and safety measures and incorporating environmentally friendly targets and initiatives to the programme portfolios. All of these meant a convergence of technology, process and capabilities that were treated and maintained separately for decades – Operational Technology (OT) versus Information Technology (IT). The aforementioned business risk context means that companies had to embed cybersecurity risks within their traditional

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

business risk frameworks, the potential impact skyrocketed, the probability of occurrence is almost certain and, therefore, it requires mature risk management processes to address these risks accordingly. There are a few cultural constraints to get out of the way first. From an awareness perspective, from the CEO to the industrial COO they need to understand that everybody can and will be attacked at some point, and that most likely an attack of a certain level of sophistication is occurring as you are reading this paragraph. Recognising the basic level of threats that every business is exposed to is a good start for a reasonable risk management process. Now we can focus on modeling risks: • How long can a steel manufacturing company survive without access to their Industrial Control Systems (ICS)? (if you really think about it, not very long). • What are the potential impacts of such a cyberattack, in monetary, reputational, regulatory, market, etc? (I’d assume those are quite hefty impacts). • How quickly can you attain full recovery from a wide scale ransomware attack? (probably not as quickly as you originally thought). Mitigating these risks usually requires the thorough implementation of an industrial cybersecurity programme, one that not only protects industrial assets and networks, but also provides the foundation blueprint for the manufacturing facilities’ ‘cyber secure’ operating model. Some companies may also decide to complement such a programme with a risk transfer to an insurance policy – all options are on the table in love, war and risk management. Recommendations for an industrial cybersecurity programme Once an organization decides to move into protecting its OT networks and assets, it faces the traditional challenges of a cultural and digital transformation programme. An industrial cybersecurity programme requires: • Right governance – both senior industrial and technology executives leading a joint industrial cybersecurity transformation

Steel Times International

programme including floor shop operations /maintenance, IT, cybersecurity experts with clear roles and responsibilities as well as accountability. • Right people – you will probably need some outside help from consulting/OT vendor or even outsourced managed service providers, and at the same time, you will need to build your own in-house skills and expertise. There will be a strong requirement for specialist certification and training as well as the deployment of industrial cybersecurity awareness training for OT operators. • Right technology tools – if you have not done so, you will definitely need to upgrade your traditional anti-virus model to an endpoint detection and response (EDR) technology to ensure you have timely and deep protection on your industrial endpoints. You will need to incorporate OT-specific threat management technology that will be able to identify the most insecure components of your OT estate and help you prioritise a remedial plan. Please, do check your online/ offline back-up capability and ensure that you have periodically tested your successful ability to recover. Finally, you will need to engage with your OT vendors to understand their own cybersecurity roadmap and align your technology upgrade plans accordingly to incorporate more secure OT assets in the future. • Right processes – you will need to ensure a varying degree of network segregation between your OT and other networks, which require careful planning and deployment, usually involving a strong change management process. As you move into industrial cybersecurity, you will need to create a cybersecurity incident management process tailored for OT. Envisioning cybersecurity challenges Implementing an industrial cybersecurity programme requires organisations to continue monitoring disruptive change as we move into a more complex industrial operating model, one with innovative and more advanced equipment and digital transformation technologies embedded within Industry 4.0 trends or even beyond. Here are a few examples of trends and innovation which industrial cybersecurity programmes should be following very

closely: • Convergence of IT/OT + IIoT /IoT – I’d like to point out two simultaneous trends: • The convergence between traditional information technology (IT) and operational technology (OT) to simplify and improve the joint management of these complexities, resulting in a joint IT/OT cybersecurity programme. • And at the same time, the proliferation of Internet of Things (IoT) sensors around corporate wireless networks as well as within the shop floor networks – Industrial Internet of Things (IIoT) – both technologies tend to be very insecure ‘out of the box’, hence constantly providing new threats to be exploited by attackers and the constant need to keep them secure. • Autonomous systems (and agents) within the shop floor – the roadmap from manual to automation and further into autonomy is well documented. Autonomous cars, autonomous agents and systems that collectively can make decisions on different parts of the production cycle. From an industrial cybersecurity perspective, it means a renewed monitoring on ensuring data flow integrity and certifying data provenance, encapsulating access levels and additional network zoning/segregation. • Use of digital transformation technologies (RPA + AI/ML + Big Data) for improving preventive controls – by embedding cybersecurity right on the design of the application of digital transformation technologies, the industrial cybersecurity programme can achieve both an ‘insideout’ and ‘outside-in’ capability of protection within the industrial operating model and architecture. These are data-driven technologies, where data flows need to be certified and data (and metadata) segregated and protected accordingly. As we have discussed before, taking on all these layers of complexity within an industrial cybersecurity programme has several challenges. Applying a structured approach to risk management helps to justify and realise the business benefits of an industrial cybersecurity programme. As Confucius said, “It does not matter how slowly you go as long as you don’t stop”. �

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 39

TURN YOUR DATA INTO INFORMATION AND INFORMATION INTO VALUE

Also visit us online: www.sms-digital.com/solutions/the-learning-steel-plant

Building the Learning [Steel] Plant SMS digital develops innovative solutions to boost your business. Benefiting from cutting-edge development methods, our solutions for plant and process condition, product quality,production planning, and energy management contribute in streamlining your maintenance efforts,decrease quality deviations and optimize plant utilization, even down toa short-term rescheduling. The digital future has already begun

Learning Steel Plant Ad.indd 13-14

21.10.21 18:06

TURN YOUR DATA INTO INFORMATION AND INFORMATION INTO VALUE

Also visit us online: www.sms-digital.com/solutions/the-learning-steel-plant

Learning Steel Plant Ad.indd 13-14

21.10.21 18:06

ARTIFICIAL INTELLIGENCE SUMMIT

AI-Human ensemble operation in steel manufacturing processes By Dohun Kim1, Hyunsuk Kang2, and Kisoo Kim3 Steel has been one of the key materials in human history for about 3,000 years and it is vital to economic growth in modern times. Today, the industry produces 1.8 billion tons of steel from over 70 nations. This has been possible thanks to historic breakthrough technologies for steel production such as blast furnaces, Basic Oxygen Furnaces (BOF) and continuous casting. Since joining the global steel industry in 1968, POSCO has benefited from these innovative technologies and was an early adopter of automation, digital sensors and control systems. However, the industry now faces technological challenges and environmental regulations. Various efforts are being made to discover new ways of overcoming them and POSCO has shifted its focus from traditional hardware-driven improvements to smart technology to find innovative solutions, including artificial intelligence (AI). Not only does AI overcome the limitations of existing technological development, but it also learns the operational know-how of experienced

experts. This enables new operators to control main production facilities with the help of AI. Currently, POSCO has made a plan to establish an intelligent steel mill where humans and AI collaborate together. AI-human ensemble operation is highlighted as a practical solution for steel manufacturing processes. The concept and some of the examples are presented in this article. POSCO currently has the world’s sixthlargest crude steel production capacity and has been selected as the most competitive steelmaker by World Steel Dynamics for 11 consecutive years. However, it has faced several challenges recently. First, quantitative growth can no longer be expected. Korea recorded the highest steel consumption per capita, that is, more than 1 ton. Crude steel production in Korea was about 71Mt in 2019, but the growth rate is gradually slowing down as the Korean steel industry turns its focus towards competitive, high-end products from mass production. Second, there is an issue of environmental regulation.

The global steel industry is under pressure to reduce carbon dioxide emissions in line with the Paris Agreement of 2015. This will lead to higher steel manufacturing costs due to carbon taxes and the use of hydrogen in ironmaking in the future. Workplace safety is another challenging issue. Despite great efforts to improve steel mill safety, the number of casualties has not been significantly decreased. Fourth, it is crucial that the knowhows possessed by high-level operators from the baby boom generation – who are retiring in the near future – is maintained and it is believed that smart technology is the key to overcome these challenges. In 2016, POSCO was determined to develop smart technology by converging domain knowledge and ICT technologies towards the development of intelligent steelworks. This has been triggered by the AlphaGo event in 2016, and it has believed that the possibility of applying AI technology to manufacturing processes has become critical. Further, POSCO has recognized that

1. Senior researcher, POSCO, hunkim@posco.com 2. Senior researcher, POSCO, kangpoh@posco.com 3. Head of process and engineering research lab., POSCO, kisookim@posco.com

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

Fig 1. POSCO’s smart factory strategy

Steel Times International

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 43

ARTIFICIAL INTELLIGENCE SUMMIT

Fig. 2 The concept of Smart Blast Furnace

smart technology could be a new growth engine creating new values for the steel industry. POSCO’s ‘smartization’ strategy can be divided into three categories (Fig 1). First, it focuses on developing smart sensing technology. By developing a smart sensor, it is possible to convert massive unstructured data obtained from video, audio and sensor data to structured data that can be analyzed. Next, POSCO is developing smart analytics technology that can analyze operations in real-time by utilizing both smart

sensing data and operational data. Lastly, the right AI algorithms are chosen to optimize operational conditions with a level of highly skilled operators, or to predict facility failures based on real-time analysis. With a strategy of leading by example, POSCO started with four pilot projects, which gradually expanded to other processes. Currently, more than 100 AI models are in operation covering various areas from marketing to production, including logistics and safety. Realizing that it is difficult to

automate the entire process with AI alone in the steel industry, POSCO combines Al models with 4D (Dull, Dirty, Dangerous, Difficult) tasks. Some case studies follow. Autonomous AI control of blast furnace (smart blast furnace) A smart blast furnace equipped with AI algorithms has been developed (Fig 2). Traditionally, the operation of the blast furnaces depends entirely on the operators’ experience and their intuition. Five AI models

Fig. 3 Smart CGL

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

Fig 4. POSCO’s digital twin technology

Fig 5. PIMS: POSCO Intelligent Maintenance System

Steel Times International

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 45

ARTIFICIAL INTELLIGENCE SUMMIT

have been developed for automating blast furnace operation. High quality data for these models are obtained by installing new IoT sensors to quantify iron ore quality, air volume, and coke usage ratio, etc. These developed AI models can predict the process conditions of the blast furnace and guide operators on how to optimize the set-up for operation. This makes it possible to keep the blast furnace in a normal steady condition. The models increase the productivity of a blast furnace by 5% and the cost is also reduced by 1%. It is noted that the CO2 has been reduced by 1%. After finishing the Smart Blast Furnace project for a medium-sized one, new AI models are being developed for other bigger blast furnaces. AI control of air-knife (Smart CGL) Smart CGL (Continuous Galvanizing Lines) have been chosen for downstream processes (Fig 3). Hot-dip galvanized steel sheets are products in which Zn is coated with a thickness of 6 to 25 micrometres to prevent the corrosion of base iron. The coating thickness is controlled using an air-knife device positioned near the molten Zn pot. The operators control the air pressure and the gap of the air knife to wipe out molten Zn in order to produce steel strips with a desired uniform coating thickness. However, the time delay between measurement of Zn-coating thickness and the operator’s judgement of operating conditions cause the non-uniform thickness of Zn-coating layers. The developed AI model predicts the Zn-coating thickness with a high accuracy instead of human operators. The deviation of the coating thickness and the amount of the over-coating are reduced by 60% and 50%, respectively as a result of the better performance of AI. Digital Twins for VCS & VTS Digital twins as digital replicas of manufacturing facilities have developed for better maintenance and smart engineering, including 3D-based layout design, VCS (Virtual Commissioning System), and VTS (Virtual Training System) (Fig 4). These are the basis for creating a virtual factory synchronized with a physical factory. A virtual factory based on 3D drawings can optimize the plant layout. The VCS helps the development of control logics by performing

various pre-tests such as sequence and tracking control and logic tests for abnormal situations. VTS allows operators to have theoretical and technical training to cope with abnormal situations in a physical world. Operators can learn about the new plants more efficiently using these digital twin components, shortening engineering time and reducing total engineering costs. POSCO Intelligent Maintenance System (PIMS) PIMS (POSCO Intelligent Maintenance System) has been developed to prevent downtime and unexpected failures of facilities (Fig 5). The main role of this system is to predict the failure of physical assets using AI algorithms built on data from various types of sensors. Before having this system, maintenance engineers had to periodically check and determine when to replace parts. This could not prevent unexpected disruption or maximize the lifespan of a physical asset. PIMS automatically predicts when each part needs to be replaced and preventive maintenance can be performed before failures occur. A failure prediction model based on IoT sensors such as sound, video, thermal image, vibration, motor current, laser and radar has been developed. Smart safety solution The smart safety solution is one of the key tools for achieving zero accidents in the workplace. For example, the Smart Safety Ball has been developed and commercialized (Fig 6). Before entering the confined area, they can throw or roll the ball into the hazardous space to detect oxygen, carbon monoxide, and hydrogen sulfide. This tennis ball size smart gas detector can be used for two years without battery replacement. Using Smart Safety Ball, it is now possible to detect gas before entering the enclosed space by checking gas concentration displayed on smartphones. Lessons learned The experience of over six years and hundreds of industrial AI show that collaboration is key to obtaining successful implementation. Building a successful team is required between AI experts and domain experts including sensor developers. The

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

team includes universities that are in charge of developing AI technology, and SMEs that are responsible for developing the right sensor for the right data. The manufacturer provides the domain knowledge to universities and SMEs. The team can make decisions for their projects by understanding each part. It is found that the role of asset owners is crucial because they will use or co-work with Al algorithms and they will have to transfer their experiences to the algorithms. Applying AI models to the works reveals that it is still difficult to lead to full factory automation. The concept of ensemble AI, therefore, is developed in a practical way in terms of AI model implementation. The concept of ensemble AI includes collaboration between AI and human, as well as a combination of physical/metallurgical and AI models (Fig 7). Ensemble AI might be a practical solution for steel manufacturing at the moment. POSCO was recognized for its achievements and has continued to focus on driving smart manufacturing technology. As a result of these challenges, it was selected as a Lighthouse Factory in steel manufacturing by the World Economic Forum in July 2019. POSCO aims to continually offer best practice in the field of AI, and wants to be a beacon for smart manufacturing solutions that can lead the 4th industrial revolution in the steel industry.

Summary Since AI models were first introduced, more than 100 algorithms have been in operation at POSCO. During the development of AI models, the company has learned that the realization of full automation using standalone AI is impractical for the steel manufacturing processes because data-based models are interpolating, not extrapolating, which means AI cannot cope with unexpected changing conditions due to a lack of data or data availability. Therefore, a human-machine ensemble operation with rule-based and a physical database AI ensemble model is a realistic way towards implementing a smart factory in a steelworks in order to reduce the risks. Again, the collaboration between programmers and operators is the key success factor for building smart factories. �

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

Fig 6. Smart Safety Ball

Fig 7. POSCO Ensemble AITM

Steel Times International

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 47

ARTIFICIAL INTELLIGENCE SUMMIT

Every bar counts Danieli has employed artificial intelligence for its Q-VID DAN COUNTER, a next generation bar counter capable of accurate measurements of hot and cold products. Danieli Automation entered the intelligent sensors market for the steel industry with a new family of smart and flexible devices based on the use of off-the-shelf highresolution cameras and image processing algorithms. Joining together more than 20-years of experience in the use of these types of technologies directly on steel production lines, with the wide capabilities now offered by industrial cameras and processing units, Danieli Automation has developed a modular hardware and software architecture called Q-VID that makes it possible to take a novel approach with process control, observing either the classic or more innovative requirements. The Q-VID range is capable of accurate measurements of hot and cold products by groundbreaking and robust computer vision techniques, specifically designed to fit accuracy and real-time constraints.

Nowadays, vision-based applications are globally recognised as a new trend in the steelmaking industry. They are used in a wide range of situations thanks to their advantages, which include contact-less operation mode, quickness of response and quality of measurement, coupled with the reduced need for human intervention and, thus, a reduction in the potential for error. Vision systems are appealing for a range of applications, from simple area monitoring to more demanding tasks such as profile detection, surface inspection or robot guidance. However, these systems often come as standalone machines with strict space requirements or a specific, fixed layout, which forces them to reside in ground plant areas and precludes incorporating others. This is one of the most interesting aspects of the Q-VID family of products, which boasts

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

the ability to be mounted in a variety of configurations from ground level to below the roof in a scattered fashion or grouped together for maximum compactness. Such flexibility opens up a broader spectrum of possibilities and allows for the creation of a range of optical applications that work in inaccessible areas of the plant, like, for instance, tracking the position of very long ladles or semi-finished products. In the Q-VID family, there are several applications that suit the most diverse requirements, but with an underlying theme of accuracy without being invasive. Digitalization and artificial intelligence can support an easier and simpler metals industry. A perfect example is the Q-VID DAN COUNTER, a new counting bars unit that adopts artificial vision technology to execute an essential task: counting the bars in a Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

each single phase of counting and separation process, in the fastest and most efficient way. Today we can simply look at the bars and count them very simply and with high accuracy based solely upon the image from a camera and artificial intelligence technology. This solution was developed by Danieli Automation with the support of its Long Products division and the Danieli Research Centre, after a benchmark of the best AI solutions available. It is an off-the-shelf product for the needs of the bar mill, all-in-one and easy-toassemble, dust, water and high temperature resistant. Q-VID DANCOUNTER was designed to avoid an invasive installation, reducing downtimes and improving the efficiency of the rolling mill. In fact, the solution consists of a simple camera, which can be installed in line with no impact on civil works and minimum room requirement. It is also natively integrated with the automation systems level 1 and 2, for setting the number of bars per bundle. The real-time counting system gives an accuracy at the higher frame rate available, powered by NVIDIA Technologies and supported by our partner Beantech. Integrated separation equipment is activated depending on the objective number of bars counted. With Q-VID DAN COUNTER the achievable counting precision is above 99.99%, adopting also a self-diagnosis system and enabling the user in the pulpit to see exactly what the camera sees.

Danieli Automation DIGI&MET HQ (Italy)

Danieli Q-VID DAN COUNTER

bundle. This task improves plant efficiency, reduces time losses and production stops and eliminates the inefficiencies of in-line installations. Q-VID DAN COUNTER Prior to Q-VID technology, solutions for counting bars adopted specific mechanical Steel Times International

methods to guarantee the required precision. Following high demand from customers who sell their bundles, based on quantity of bars, Danieli has completely revisited its existing counter unit, starting from the last technology devices available on the market, and applying its own knowledge, with the final scope to precisely guide via software,

Device operation – working description Q-VID DAN COUNTER is equipped with one camera capable of detecting, identifying and counting the head of the bars from the incoming layer. In a very simple way, the bars, carried by the existing chain transfers, pass in front of the camera, which starts to count them, till reaching the preset number, assigned at the beginning of the campaign. When the number of bars, established to keep in the final bundle, is achieved, then the chain transfer stops and the new smart preseparation arm starts the seeking cycle. The scope of this phase is to separate the head of

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 49

ARTIFICIAL INTELLIGENCE SUMMIT

Q-VID DAN BAR COUNTER hardware setup: industrial colour camera with 4 LED light modules

the last counted bar from the following head of uncounted products, and create the useful space required, from the existing separation unit, to start raising the arms, to finally get the division for the whole layer length. When the separation cycle is over, everything restarts as the normal operation sequences. Q-VID DAN COUNTER MAIN HIGHLIGHTS 1. Increased accuracy in counting and separation of bars: highest rate of bundles with right number of bars. 2. Wear free technology: no mechanical parts in contact with the bars to count, thanks to the new contact less solution. 3. Set-up free: whole range of bars counted without adjustment/fine tuning, specific for each size of bars. 4. Minimal level of maintenance required. 5. Perfect integration with the existing equipment: impact on installation negligible, thanks to special kit provided for matching the existing equipment. 6. Quick installation and commissioning. 7. No impact on existing productivity of bundling area: same separation time kept from Q-VID, giving the chance to optimize the performance (each case will be specifically studied). Conclusions Developed to aid the operators in common and repetitive monitoring tasks, artificial vision systems such as Q-VID DAN COUNTER are

Video user interface during counting operation

progressing in the marketplace as versatile, adaptable and autonomous applications dedicated to counting bars even in harsh environments. Q-VID DAN COUNTER components are thought to be self-containing, resistant and durable blocks of a distributed vision system; each one is comprised of interchangeable

parts that can be accessed easily by the operators to replace them. Danieli Automation carefully designed Q-VID DAN COUNTER to operate as individual microsystems that provide reliable and accurate data feedback to the workforce in a minimum time to market, with 99.9% counting precision. �

Q-VID BAR COUNTER installation

DANIELI INTELLIGENT PLANT A NEW CONCEPT FOR PLANT AND PROCESS SUPERVISION Data-driven approach, AI and machine learning for continuous improvement of plant performances, simplifying metals complexity Via Bonaldo Stringher, 4 33042 Buttrio (UD) Italy Phone +39 0432 518 111 www.digi-met.com www.dca.it

ARTIFICIAL INTELLIGENCE SUMMIT

“Our partnership with PSI Metals was established to

maximize efﬁciency gains and associated cost savings in steel companies.

’’

Dr. Falk-Florian Henrich, founder and CEO, Smart Steel Technologies.

Three AI-based applications Two leading suppliers of AI-based systems and production management technology have joined forces to help steel mills optimize production and improve efficiency and quality. Smart Steel Technologies (SST), a leading provider of artificial intelligence (AI) for the global steel industry, and PSI Metals, a specialist supplier of production management solutions for the metals industry, have announced a partnership to further optimize production in steel mills. “Our partnership with PSI Metals was established to maximize efficiency gains and associated cost savings in steel companies. It will help the steel industry to work even more effectively and with higher quality. The combination of PSI’s extensive production management know-how and SST’s profound AI expertise lifts evident synergies for everyone involved”, says Dr. Falk-Florian

Henrich – founder and CEO of Smart Steel Technologies. SST and PSI Metals have launched three AI-based software solutions in the areas of product-to-order reallocation, slab and coil classification and liquid steel quality optimization, and in particular temperature control. It is claimed that these new solutions will help steel companies to further improve the production efficiency and quality of advanced steel products to meet the highest market demands. This co-operative effort, says SST, will digitize applications at every stage of the steelmaking process – from the liquid phase to strip finishing. “In order to implement comprehensive

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

digital solutions and, therefore, increase the added value for our steel customers, it is crucial to join forces with a strong AI partner with excellent AI and machine learning capabilities,” says Jörg Hackmann – managing director of PSI Metals. “We have been able to convince ourselves of these capabilities over the last few months. Therefore, we are sure that the co-operation will bring high added value to the steel industry.” Three AI-based applications for steelmakers Smart (product-to-order) reallocation is one software solution that focuses on

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

For further information, log on to www.smart-steel-technologies.com and www.psimetals.de

“It is crucial to join forces with a strong AI partner

’’

with excellent AI and machine learning capabilities.

Jörg Hackmann - managing director of PSI Metals.

s for steelmakers smart product-to-order reallocation. SST will predict the expected surface quality after hot rolling, pickling and galvanizing at the time when the slab is cut off the strand. The defect classification is based on deep learning to reach much better classification accuracies. SST has developed three distinct proprietary, deep convolutional neural network architectures for hot rolled, pickled and galvanized strip images that will serve as an input signal for production management. PSI will provide a reactive cutting length schedule service that uses the surface quality data from SST as input. Based on both real time Level 2 system measurements and Level 3 system information on order book and caster sequences, this solution allows a close to real time re-optimization of casting cutting plans in case of production deviations. In combination, a dynamic slab cutting

Steel Times International

programme can be initialized, in order to maximize the length of prime quality material products. Grading of slabs and coils The second software solution is aimed at the automated grading of slabs and coils. The precise surface defect classification using SST’s proprietary deep learning technology is combined with precise quality decisions from PSI Metals. Data transfer takes place to surrounding software systems, to automatically take good slabs or coils into direct use and to sort out bad ones. This enables the steel companies to lift existing rule-based grading information to state-of-the-art AI-based grading of their products. Temperature and liquid steel control In the process of liquid steel production

optimization and temperature control, PSI Metals’ precise online heat scheduling is supplemented by Smart Steel Technologies’ SST Temperature AI and SST Steel AI. An improved treatment procedure (to optimize steel quality and temperature) may lead to changes in the production schedule by adapting the treatment times. SST provides recommendations on target temperatures and treatment times. PSI incorporates this information into its scheduling system and the result of the changed schedule is fed back into the SST optimization software. Hence, the joint product lifts the current static approach to a highly dynamic, closed loop AI-based planning system. As a result, the steel company permanently saves energy costs, optimizes process stability from BOF/EAF steelmaking to continuous casting to the maximum, and hence reduces CO2 emissions. �

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 53

ARTIFICIAL INTELLIGENCE SUMMIT

Using AI to reach ESG initiatives in steel In recent years, the demand for transparency on sustainable and socially responsible business practices from investors, employees, and most importantly, customers, has seen a significant rise. Now more than ever, the world is no longer only interested in just what you make, but in how you make it. By Sandeep Pandya* As environmental, social and governance (ESG) factors continue to impact buying and investing decisions, much of the emphasis has been on the environmental component. This is largely due to the climate crisis and the harsh environmental impact heavy industry has had on our planet – specifically

industries like manufacturing and construction. Understandably, the race to eliminate carbon emissions and reduce environmental risks has been the focus of many companies and investors. While environmental impact is certainly vital to the future of our planet, for these initiatives to be truly transformative,

employers must ensure the safety, health, and welfare of workers. And top among those responsibilities is worker safety. Even the most environmentally responsible company can’t truly say they are sustainable if workers are being injured, or killed, on the job. And while we’ve made great strides in

* CEO, Everguard.ai

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

environmental health and safety over the past decades, we still have a long way to go. Worldwide, 380,000 workers die annually due to work-related accidents. That’s over 1,000 fathers, mothers, sons, daughters and friends who don’t make it home at the end of the working day. When you consider how many non-fatal but serious debilitating injuries occur, the incident count jumps into the millions. We must do better. And we can by utilizing artificial intelligence (AI) powered by sensor fusion. Understanding sensor fusion We use our five senses – sight, touch, taste, smell, hearing – every day to understand our environment and assess any risks or dangers. One sense is valuable, but when all five are used together, our ability to prevent an accident before it happens increases dramatically. Think about an afternoon bike ride, enjoying the beautiful weather as you pedal along. Before you even see a vehicle traveling towards you, you hear it and your other senses begin gathering input. Steel Times International

By the time the vehicle travels into your line of vision, you’re already incorporating that input. Without realizing it, you may use your sense of touch to grip the handle bars a bit tighter to veer out of the way if necessary. And is that exhaust you smell? You’re innately processing input from each of your senses and instantaneously deciding if the situation poses any risks. If there is a risk, you naturally move away from traffic. If no risk is indicated, you continue biking and enjoying your day without ever consciously realizing all the data your body just processed. Now imagine walking through a steel mill, fabricator, or processor. Instead of that data coming from your senses, a technology platform is gathering input and data from multiple sources to continuously inform you, your workplace and your team members, proactively protecting you and your coworkers from incidents and accidents. The information it gathers, much like the inputs your senses gathered on your bike ride, is processed by artificial intelligence (AI) algorithms that can make the same type of

split-second decisions you made. Sensor fusion-based AI makes this concept a reality. Sensor fusion at its core Sensor fusion is the process of combining data from multiple components to generate an output or action that is more accurate and reliable. Much like our five senses, sensors are often used individually, but a combination of these inputs has been proven to be more valuable. Sensor fusion is how AI in health and safety understands an environment and decides if there are any definable risks or dangers present. Cameras, computer vision, radar, GPS, lidar and wearables are all devices that can be incorporated into a sensor fusion AI health and safety platform. Inputs gathered from these devices and potentially machine state status via PLC integration are sent to a central system for holistic analysis and processing by AI algorithms that create an almost immediate output. The more data input the system receives, the more it ‘learns’. While output

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

ARTIFICIAL INTELLIGENCE SUMMIT

results are more accurate with multiple sensor inputs, sensors are not dependent on each other. Should one sensor fail, the remaining sensors continue to actively gather data. Using multiple state-of-the-art sensors allows for constant gathering of consistent data even in the harshest conditions like a steel mill or construction site. A deep dive into AI The hardware sensors that feed data to the deep learning algorithms provide information used to contextualize the environment, specifically around worker activity. Examples include: • High-definition cameras and thermal imaging cameras act as inputs for computer vision (CV) which allows the system to ‘see’ and understand digital images and videos. Computer vision is the ability for systems to ‘see or sense a specific stimulus, contextualize what is being seen, and extract and communicate that data in such a way that it can be used to educate and inform the actions taken by AI algorithms. • Radar sensors detect moving objects, like a worker or mobile equipment, using radio waves. • Light Detection and Ranging (Lidar) uses a laser to project a beam on to a target and then measure the reflection to provide a measure of range. • Real-time location systems (RTLS),

much like indoor GPS, provides centimetreaccurate location detection. Workers, machinery, and mobile equipment equipped with location tags are tracked in real-time using RTLS. RTLS anchors create a mesh that keeps in constant contact with the sensors, creating a zone of protection. • Wearables allow for two-way communications with workers, providing inputs as well as acting as notification devices using haptic responses. Data is continuously gathered and passed wirelessly to edge servers situated at the ‘edge’ of the safety zone. True to its name, the edge server must be at the edge of the zone to process the sensory data in real-time and access the deep-learning models. These high-powered computers are powered by NVIDIA Titan RTX graphics processing units (GPU) capable of processing 130 trillion calculations per second, per card. Each edge server can be outfitted with a number of cards depending on the environment and the number of sensors a specific safety zone requires. Supervised learning Turning those trillions of calculations into input that educates the AI algorithms or ‘brain’ of the system requires a process called supervised learning. This is where human interaction from engineers and Ph.D. level

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

technologists becomes imperative to train the models. The process requires sensor fusion inputs to be translated into numeric data sets that create large matrix tables. The data within those tables is labeled by humans looking for positive and negative scenarios to train the AI system on what specifically to look for within the model. This process of feeding positive and negative inputs trains the AI algorithms to detect the underlying patterns and relationships within the data which represent real-world activities, be it the movement of a worker or the trajectory of a crane. The more data the system receives, the more generalized its learning can become as it looks for patterns within the data. This pattern recognition is how the system is able to contextualize the inputs even when presented with never-before-seen data. Once trained, this AI process occurs within the edge servers to trigger an immediate response from the system when it recognizes a positive pattern within the data it has been trained to process. Such action can involve an alert to a worker who has violated a safety protocol or the slowing of an overhead

Steel Times International

crane that is moving dangerously close to humans or machinery. Once sufficient data has been labeled and processed by the system, the models generate predictions with over 90% accuracy which mirrors or exceeds the accuracy of humans to detect potential hazards. However, technology doesn’t suffer from distraction, fatigue, or the need for a bio break.

found nearly everywhere in hospitals. The technology in the form of wearables are placed on patients to monitor heart rates, oxygen levels, blood pressure, body temperatures, and even moods. The combination of the data and inputs from these devices provides a much more accurate picture of a patient’s health than one sensor could alone.

Sensor fusion usage Today, sensor fusion-powered AI is used widely across multiple industries, the most publicized being in the evolution of the autonomous vehicle. Sensor fusion combined with AI is what gives these vehicles capabilities for static object detection, moving object detection and tracking, occupancy grid mapping, navigation and much more. This is a great example of how a properlytrained model can instantaneously process continuous outputs and take actions based on that data even if it has never processed it before, be it swerving to miss a newly fallen branch or slowing down to avoid a new pothole on a city street. Smart cities are also utilizing this technology. Multi-sensor systems like cameras, motion sensors, and others, allow for cities to efficiently monitor pollution, improve traffic flows, and enhance public safety. And finally, sensor fusion can be

Sensor fusion and the steel industry When it comes to employee health and safety in the steel industry, sensor fusionbased AI gives mills, fabricators and processors the ability to shift from a reactive response to a proactive, behavioural training programme. This technology helps bring the industry one step closer to reaching the goal of an accident-free steel industry. Consider an overhead crane in the process of moving a coil or beam in a shipping bay. The crane carrying the load is moving south, but there is a worker with his/her back to the crane moving north, not realizing he/she is about to walk directly under a heavy load. In a bay equipped with sensor fusion AI, the mix of hardware, software, bio-metric wearable devices and other sensors feed data to the AI algorithms embedded in the technology platform to remind that worker of safety protocols, all in under half a second. Camera A attached to the crane tells

Steel Times International

the system the crane is moving and in what direction, Camera B views the worker moving towards the crane, and also relays that input to the system. That data is processed almost instantaneously to create an alert which is sent to the worker via a bio-metric wearable device before he/she reaches the danger area beneath the moving crane. An alert can also be sent to the crane operator to notify him/her to stop movement. In addition to the cameras, a sensor can be placed on the crane itself to surround it with a geofence. Once the worker ventures within inches of the geofence, that same wearable on his or her wrist buzzes with an alert. With multiple input devices, the system does not rely on one set of data but instead continuously assesses the entire environment to train the AI algorithms to take action as outlined above. In this scenario, sensor fusion AI is able to provide behavioural training via the alerts to encourage worker behaviour that lowers incident rates and helps avoid dangerous situations before they happen, including fall protection, worker-to-worker and worker-tovehicle detection, and many others. Emphasizing the ‘S’ in ESG We’ve made significant progress in safety throughout the last decade. But statistics remind us that we still have more work to do. To fulfill ESG initiatives, we must not forget the social pillar – and the safety and health of employees that fall under it. Sensor fusionbased AI has the potential to not only help companies fulfil their social responsibility to their employees, investors, and consumers, but to proactively propel steel to be an accidentfree industry. Sandeep Pandya is CEO of Everguard.ai, a joint venture backed by Boston Consulting Group Digital Ventures and SeAH Global Inc, a global steel conglomerate. The company’s mission is to make the world’s industrial environments safer using AI and sensor fusion. Every day, there are an estimated 1,000 global workplace fatalities. The team at Everguard is committed to driving a paradigm shift in workplace safety from reactive to proactive approaches. Learn more at www.everguard.ai. �

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

ARTIFICIAL INTELLIGENCE SUMMIT

Five reasons why steel manufacturers need machine learning Fero Labs’ Berk Birand*shares how explainable machine learning can help steel manufacturers solve core business and technical challenges. 1. With the ability to learn and improve in real time, machine learning solutions outperform Six Sigma and SPC. Most steel plants face the same business challenge: reducing raw material costs to stay competitive, without jeopardizing product

quality. Traditional process improvement tools aim to help them achieve this goal, yet they have several key limitations. Specifically, these tools tend to plan conservatively for the worst-case scenario, ie a particular composition of scrap.

However, the frequency of variation in steel production—especially at the scrap level— means that this scenario rarely happens. As a result, the tools don’t perform as well as they could. A machine learning solution, however,

* CEO, Fero Labs.

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

adapts continuously to changes on the factory floor. It’s like designing an experiment that covers your entire production line multiple times a day—a task that would be impossible with Six Sigma. Moreover, traditional tools don’t do a great job of combining data from across the plant in an end-to-end process. In contrast, machine learning can rapidly process complex correlations and help you make smarter business decisions—as long as you choose an explainable solution. Why is explainability key? Most applications of machine learning at technology companies like Netflix and Uber are based on black-box algorithms. They give you predictions (i.e., what movie you’ll enjoy), but don’t have to reveal how they came up with them, as this information has little benefit for the average user. Industrial machine learning, in contrast, requires explainability. When optimizing for emissions reduction, a black-box model that simply predicts results is far less useful than understanding the complex interplay between different factory settings and how changing them affects the volume of emissions produced.

2. Machine learning delivers measurable ROI, from lower raw material costs to improved quality. With any digitalization project, your first question should be: How big is the opportunity? How can this impressivesounding technology drive real business value? Likely you’ve seen many AI and ML technologies that don’t live up to the hype. That’s why it’s crucial to focus on results. And that’s where explainable machine learning—in contrast to its black-box counterparts – can really deliver. Steel plants generate major savings by deploying Fero’s explainable machine learning solution to reduce alloy consumption. Within six months of deployment, the average plant generates $1-$6 per ton in raw material savings. How does this work? With a quick glance at Fero’s live recommendations, engineers and operators know the precise amount of alloys needed for each heat, rather than relying on conservative estimates. This makes Steel Times International

their job simpler and easier, and because alloys are costly, it also leads to substantial savings. Scaled across five or more plants, the results are even more dramatic, highlighting the vast amounts steel manufacturers can potentially save by deploying Fero across a wider number of grades. Customers often start out deploying our solution in individual plants, allowing them to compute the potential savings that are possible in terms of raw material reduction. That way, they’re able to see how much they’ll save over the next year, if they deploy machine learning-powered predictions and recommendations over every heat and every product grade.

3. Machine learning lets you optimize your operations in real time. A live machine learning solution streams data directly from the production line and tells engineers and operators how to adjust settings based on this real-time data. Instead of applying static assumptions to a broad range of processes, you get live predictions and optimizations tailored to your unique raw materials, or to the process setting impacting your quality at the time of the prediction/ optimization. Skip unnecessary tests or add fewer raw materials, while remaining confident that your quality falls within specs. One example? Minimizing costs. Everyone wants to keep costs low. With live machine learning, you can see exactly how much raw

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 59

With thousands of sensors in your factory, shouldn’t you turn that data into your competitive advantage?

material you need to reach a given quality standard, keeping costs as low as possible. Here’s a look at how this works in a steel mill. Mistakes aren’t cheap in steel, where a scrapped heat can cost as much as $100,000 due to wasted time and materials. So accuracy is of vital importance. As part of the initial Fero set-up phase, engineers in the steel mill tell the software what yield and tensile strengths they need to achieve. Based on the specified quality requirements, Fero can analyze a sample of each heat and tell operators the minimum amount of each alloy they need to add, in real time, to ensure they meet those requirements. As a result, costs and raw materials are kept as low as possible. It’s crucial to note that the software’s recommendations don’t come out of a mysterious black box. Rather, they’re explainable. Confidence bands and buffer zones displayed prominently on the interface ensure that these recommendations fall within quality standards. If there’s a slight probability—even 5%—of the heat being scrapped given the current settings, the prediction tells the operator how to change the settings to mitigate such an outcome. What’s the ROI? Within a month of enabling Fero’s live optimization feature to provide engineers and operators with constant real-time recommendations, one steel company reduced alloy consumption

by 16%, saving as much as $4 per ton as a result. When you consider that your typical steel mill generates over a million tons per year, this adds up to millions in savings. How does this compare to Six Sigma? While machine learning has many advantages over Six Sigma, it doesn’t contradict its core principles. In the context of the traditional DMAIC cycle, you can think of machine learning as a vastly more powerful ‘Analyze,’ continuously learning and improving in the background. Predictions and root cause analysis are two key benefits of this real-time analysis, and it doesn’t end there. A machine learning solution can tell you what raw material additions and process parameter settings will minimize costs, while ensuring that you still meet quality targets. This has implications on the entire workflow, as quality and operations groups can confidently change production settings while the product is still being processed—improving the bottom line.

4. Machine learning reduces emissions Industrial engineers face a daunting balancing act, forced to balance financial goals with quality targets such as producing strong enough steel or chemicals with the appropriate viscosity. In this context, the idea of minimizing emissions often falls by the wayside—even though US factories account

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

for roughly a quarter of emissions, falling behind only transportation and electricity. But with machine learning, this cumbersome process can be automated. Today’s engineers don’t have to choose between competing objectives—rather, they can simply upload plant data and tell the software what their goals are. By analyzing this data through machine learning, the software can pinpoint the settings that must be adjusted in order to achieve all the goals, making emissions reduction no more burdensome than any other task. Machine learning also eliminates the need for emissions-causing tests. Traditionally, manufacturers must suspend production to experiment with new settings. In contrast, explainable machine learning lets engineers test hypothetical what-if’ scenarios without running a single real-life test. By creating a digital twin of the plant, they can simulate how past results might have changed with different settings, or tinker with future production settings to achieve optimal quality. This avoids (or minimizes) physical testing, preventing additional emissions. Finally, machine learning improves production yields and efficiency. Despite advances in manufacturing technology, engineers still can’t always explain inconsistent yields—in other words, why the same quantity of input materials produces a different quantity of outputs.

Steel Times International

No more expensive "digitalization consultants.” Put engineers and operators in the driver's seat with Fero — the only machine learning software developed with steel manufacturers in mind.

Minimize raw material costs while ensuring quality specs are always met.

Improve production volume by optimizing process bottlenecks in the melt shop and the rolling mill.

Forecast the risk of defects

at the caster, rolling mill, and process line... and intervene before they happen.

Increase yield by optimizing

for each heat’s unique raw material and EAF/Blast Furnace parameters in real time.

We’ve driven $4+/ton savings for some of the world’s largest steel companies. Are you next? Get our whitepaper and schedule your free demo at www.ferolabs.com/steel.

ARTIFICIAL INTELLIGENCE SUMMIT

Machine learning can reduce emissions

Using explainable machine learning, it’s finally possible to understand what factors drive these inconsistencies, and to optimize around them. Steel production, for instance, requires combining various scrap metals until the result reaches a certain quality. With machine learning, engineers can predict exactly the amount of raw materials they need. By improving yields and efficiency, they can

operate the steel plant for less time to produce the same result—tackling one of the biggest causes of emissions, which is burning fossil fuels.

5. Machine learning can give you a competitive advantage. From cost savings and yield improvements to increasingly valuable sustainability metrics

such as emissions reduction, basing decisionmaking on existing factory data rather than conservative estimates can be transformative, optimizing both the production process and your bottom line. With thousands of sensors in your factory, shouldn’t you turn that data into your competitive advantage? �

Machine learning gives you a competitive advantage

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

MOVE FAST

MOVE FORWARD

Marking & Labelling Coils & Slabs

Destrapping Including High Strength Steel Straps

Product Handling & Specials

Dross Removing

Coil Eye Welding

Sample Plate Handling

www.tebulorobotics.com

ARTIFICIAL INTELLIGENCE SUMMIT

High-resolution data acquisition is key In this article we present different approaches to apply machine learning to measurement data of different aggregation levels and show why a common and consistent data acquisition system is the key to apply such methods successfully. By Tobias Seitz*

While machine learning or AI methods in general are becoming more widely-used in different sectors of industry, the preparation work for the application of such approaches is often underestimated. Especially for fast processes, the acquisition, recording, and management of data is a challenging task. Additional work is required in order to get clean, complete, and reliable measurement data. Why edge analytics is so important While in the context of Industry 4.0 and digital transformation, measurement data are available for nearly every asset or machine, the highest flexibility can only be achieved by using high-resolution process, machine, vibration, energy and sensor data as a basis for all consecutive computations and aggregation layers. Especially for the partially very fast processes in steelmaking this means that sampling rates down to sub-milliseconds are required. Obviously, sending these data volumes to a higher-level system ‘as is’ is impracticable in terms of costs and volume in most cases. That’s were edge analytics comes into play. By applying analytical or datadriven methods directly where the data are recorded makes it possible to immediately reduce the amount of data. The process data acquisition system has to offer an edge analytics integration which makes it possible

to apply suitable methods to the raw or pre-aggregated data. Only the results or condensed data should be passed to higher level systems continuously. Additionally, short raw data snapshots can be transferred to the higher level periodically or triggered when an anomaly is detected. One of the main functions of edge analytics is to aggregate data and compute KPIs based on the raw data. Any anomalies or problems detected in trends and values on higher aggregation layers have their root cause in the raw level. With references from the higher layers down to the raw level a verification and analysis on to the original data must be provided. In the case of machine learning, not all required steps should be done on the edge. An important point is the distinction between training and evaluation. The training which is running most preferably on raw or pre-aggregated data is not designed to be executed on edge devices. But the edge device can be used to transfer the raw or preaggregated data to a central storage system or data lake and provide the training data for powerful computation systems. The evaluation (or analytics) part of the machine learning can be deployed to an edge device. Different data layers and their relation to machine learning A huge variety of methods can be referred to

as machine learning. However, ‘the method’ usually does not exist. The results for each method will be different depending on the available data, the desired output, and the available time for engineering. Based on different data aggregation layers, see Fig 1, we want to give some examples where AI methods can be applied in a fruitful way. Most important for the application of any machine learning method is an appropriate data management together with a data acquisition, recording and processing system, which fulfills the requirements to support such applications. Especially for fast processes, data can be available in a different format and resolution and we categorize them in the following way: (Fig 1) • Layer 1: Raw data These time-series data are usually recorded continuously with high-resolution from all available sensors, actors, etc. Triggers, ie a segmentation of the process, need to be available in this layer. • Layer 2: Aggregated data As a first data preparation step, data are usually down-sampled, pre-selected, and filtered based on triggers (which are part of the data enrichment layer), and redundant information can be removed. • Layer 3: Production KPIs (key

*Product manager for analysis and automation software products at iba AG.

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

Fig 1. Different data layers based on high-resolution time-series data. A suitable data enrichment (meta-data) needs to be provided on each layer

Fig 2. While data acquisition and storage are done close to the process, online-evaluation of measurement values needs to be available and the results can be provided to higher-level systems. Additional local storage is recommended to avoid data loss

performance indicators) With KPI we usually refer to any value which was derived from raw or aggregated data and refers to a single product, batch, or other segmentation (e.g. average weight or different quality measures). • Data enrichment While we refer to process data acquired directly at the production with the three categories above, in all stages it is possible to enrich data with additional information from higher-level systems. Next to the triggers already available in the raw data, other metadata should also be added like, for example, material-mix or order information. Steel Times International

Depending on the data layer, different analytical and data-driven methods can be applied and we want to give some examples with a focus on the requirements of a data acquisition system, which is used to manage the data. It is important to point out that for each data layer, the application of machine learning methods usually requires there to be enriched data available. Otherwise, problems might be detected but not categorized or assigned to the different situations. Analyze time-series data For data layers 1 and 2 typically an edge analytics approach should be preferable in

order to avoid sending or copying high data volumes to higher-level systems. The result of any algorithm applied on these layers can usually be characterized as Layer 3 data. With this in mind, combined approaches are also possible by combining different stages or a ‘chain’ of individual methods. For continuous time-series data available in Layer 1, various stream processing methods can be used. The processes should run locally to avoid sending high data volumes over the network. Ideally, the data acquisition system itself already offers different possibilities to do online calculations. However, a broad connectivity is an important prerequisite to use external tools. Further, by storing the data

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 65

ARTIFICIAL INTELLIGENCE SUMMIT

Fig 3. When high-resolution data are not required, further processing can be done asynchronously based on locally stored data. Results need to be provided to higher-level systems using different interfaces and protocols

locally and transferring them to a central storage facility asynchronously, it should be possible to not lose data for further processing. (Fig 2) A good example for using raw (Layer 1) data is a coil mix-up identification based on thickness recordings. By using thickness measurements for each coil, a unique fingerprint can be derived and compared at different process stages. For such a method to work, different steps (length-conversion, alignment, and feature extraction) need to be applied to the raw data otherwise too much information is lost and the comparison will fail. Also note that partial data (and not the complete coil) can already be used to identify a possible mix-up. When working on segmented or triggered data, only a semi-online processing is required, ie whenever a batch or product is finished, the corresponding data block can be processed, which requires algorithms

based on data arrays rather than streaming data. Ideally, the data management system offers possibilities to define start- and stoptriggers, and to automatically process the data between those triggers batch-wise. For the integration of machine learning, suitable interfaces should be available to request training data and to evaluate models semionline, based on the corresponding triggers. (Fig 3) For the application of batch-wise and stream processing, the measurement system should be able to store historical data and triggers in order to use the recorded data for training or further evaluations on higher aggregation layers. Open interfaces should be available to allow asynchronous data requests and enable integration into thirdparty tools for custom applications. A typical example is the monitoring of repetitive and uniform processes like the

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

sawing of steel rods. Each sawing has a characteristic shape in the power profile of the saw and by suitable resampling of the time-series they can be fed into a neural network which can be trained to find and characterize anomalies. Since the process works stably in most cases, a semi-unsupervised method can be used to learn the ‘good’ process state. Any process deviation (abrupt or creeping) is detected and alarms or maintenance recommendations can be derived. Long-term analysis based on KPIs While the previous approaches rely on high-resolution data from, for example, a single production line, a far more common application of machine learning tries to find defects and quality problems by monitoring the complete production chain or plant. By using a common data acquisition system for all available assets, it is possible

Steel Times International

Monitor repetitive processes in real-time With iba you won’t only get a full-featured data acquisition system but can also rely on various modules for visualization, trending, and real-time data analysis. Our components for repetitive process monitoring offer either simple limit alarms, advanced reference curve monitoring, or even the application of neural networks by seamlessly integrating third-party machine learning systems. The monitoring can be easily configured and is based on already available measurement data without the need for additional sensors. Data analytics can run on an edge device close to your machine Typical applications include anomaly and wear detection to prevent failures or damages. This avoids high maintenance costs, unplanned downtimes, and thus results in a fast ROI.

Benefits Reduce costs for maintenance and spare parts

Increase your productivity

Protect your machines, your personnel and the environment

Measurement Systems for Industry and Energy www.iba-ag.com

ARTIFICIAL INTELLIGENCE SUMMIT

Fig 4. Results (KPIs) from local instances can be collected centrally, enabling an overall evaluation. By using pre-aggregation methods, data volumes are kept low, and condensed information is provided to the upper layer

to transfer KPIs which are computed on the edge to a higher-level system via defined protocols and fixed schemata. With this approach a central and light-weight (in terms of storage capacity) data pool can be generated and used for training and evaluation of advanced analytical models. Due to the reduced amount of data traffic the storage and analysis tools can be easily designed as cloud-based systems and also different production-sites can be compared and analyzed. (Fig 4) When working on these data we are now in the regime of classical data science, data mining and machine learning methods. Depending on the available values or KPIs, different methods and approaches exist which have been adapted to almost every possible situation. Typical approaches include SPC analysis or condition monitoring. The difference, when setting up the data

acquisition and using available methods as described above, is that in such a scenario any additional data preparation work (selecting, sorting, clustering, or cleaning the data) is kept to an absolute minimum. The suitable aggregation and enrichment with meta-data information is already done on the lower layers before doing the final aggregation or evaluation. This is also the key to enabling the drill-down to the raw data at any time. Best practice While we tried to focus on the application of machine learning for different situations, we should mention that classical approaches also exist which are usually not referred to by AI methods. Those methods are already able to provide good ROI in many cases. However, independent of the method or algorithm used, a thoroughly configured and consistent data

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

acquisition and thus a high-quality data-basis is a crucial ingredient for good results. From our experience, machine learning methods perform best if existing insights and understanding of the process are integrated into the algorithms. Any information available should be used to support the used method. To rely on the ability of the approach to gain such insight automatically is not recommended. Finally, the most important ingredient is a clearly defined goal or scope for the application. What do you want to achieve? Which specific part of your production do you want to improve? How do you measure an improvement? With the answers to these questions, it will be much easier to define any specific measures needed starting with data acquisition, aggregation, and further evaluation and processing. �

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

AI-based operational excellence

Steel casting and milling are complex processes involving extreme physical stresses at considerable temperatures. The equipment that performs these tasks require large investments to procure, install, operate and maintain. Hence, maximizing revenue generation from existing capital assets is essential for the modern steel mill to remain competitive. Digital technologies are unlocking unparalleled opportunities

in achieving newer heights of operational efficiency and reliability, thereby allowing steel companies to maximize production uptime and maintain quality. The automation of the equipment and processes happening over the past couple of decades has given way to accelerated adoption of Industry 4.0 and IIoT-based initiatives that focus on sensors and connectivity. Operational data is being

collected at a velocity and volume like never before. Insights are no longer limited to human inspection and domain knowledge but are now acquired using digital technologies and advanced analytics practices. Application of AI unearths multiple new goals of operational excellence. The primary role of AI is to provide complete visibility of every action of every machine and process. Second, find patterns in the

* Falkonry

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

At least one day per month of production time is lost to machine failures, costing steel and heavy-industrial companies US$225 billion annually. Such instances of machine failures or quality deviations in production stem from the operational blindspots that exist in plant operations. Artificial Intelligence can help detect and explain these blindspots in real time, providing operations and maintenance teams with actionable insights to address events before they adversely impact operations.By Shreebhooshan B* and Siddharth Parwatay*

data and provide explanations, ie, discover where the problem is and what to do about it. As a result, predictive maintenance and quality monitoring procedures directly benefit from AI and lead to increased equipment availability, production uptime and throughput. Finally − and most importantly – they enable continuous improvement not only in existing processes and work practices, but also in enhancing feedback into engineering, product design and sustainability projects.

Steel Times International

Current challenges and role of advanced analytics Modern steelmaking is heavily instrumented with several process parameters being monitored, yet there are limited operational insights available in real time. Take, for instance, the continuous casting process − a facility producing 150 tonnes per hour can generate over US$5 million per day in production revenue, assuming current steel prices. Conversely, a single day of lost production is equivalent to US$5 million worth of losses. It is alarming to note that about

US$20 billion is attributed to reliability losses in steel manufacturing annually. Therefore, a manufacturer can unlock tremendous value by eliminating these unscheduled production downtimes. Casting molten steel, unsurprisingly, is hard on heavy equipment. Components wear under harsh conditions leading to failures or adverse product quality. Early detection of such conditions could warn the maintenance and production managers to schedule repairs before failures occur. Applying advanced analytics to machine and process data can

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 71

ARTIFICIAL INTELLIGENCE SUMMIT

Fig 1

help in predicting such unwanted events. Data-science projects are often designed for specific use cases thereby limiting the scope and interoperability of the model. The approach faces challenges in terms of model sustenance in production and scalability across use-cases or plants. This could be attributed to several factors:

changes (part replacements, process changes, etc) would make the resultant AI-based solution quickly irrelevant. Our experience with AI, and particularly when it comes to scaling, reveals that adequate benefits from AI accrue not from depth of application, but from broad application across a large number of applications.

1. Using historical data to build future models: A dynamic production environment, where operational states change rapidly, ensures that failures are not precisely repeatable. Equipment and processes evolve and old models don’t detect new failures; in fact, they might see new normals as false positives. This leads to what is known as the model-churn phenomenon wherein the training and validation of models requires so much time that any completed data model is obsolete as it nears deployment.

3. Limited buy-in from the ops team: When a technology solution is implemented from the top down, there is a natural resistance to change which hampers adoption. For an AI system to be selfadopted by on-ground practitioners, it has to be easy to use and its usefulness should be evident from the outset. When an AI system is able to ingest and analyse real time data to surface insights quickly, operations personnel find value in it and advocate the scalability. Hence, there is a strong demand for advanced analytics to provide real-time operational insights that are not only actionable, but also explainable and easy to use. In our experience, the discovery of patterns or anomalous conditions is only a part of the puzzle. The most value comes from the explanations of those patterns – answering why at a particular point in time a particular asset or system is identified as needing attention. Automated discovery and

2. Apply AI to only a limited number of use cases: Many steel manufacturers are still in proof of concept and pilot purgatory stages, with an intention to create highreliability prediction systems within a defined use case in order to prove ROI. The difficulty with this approach is that getting to the point of high prediction reliability tends to take a long time. During this time, any systemic

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

explanations of production events help the maintenance and reliability teams to find ‘what’ the problems are; ‘where’ they are happening right when they are happening; and ‘why’ the event is happening with the details of contributing factors. This helps the maintenance team use their operational know-how to prioritize actions across the assets, lines or the entire plant. One of the key characteristics of the Falkonry AI and ML platform is that it provides explanations for the conditions that it discovers and detects. Operational AI Operational AI is Falkonry’s software platform to enhance operational excellence by avoiding operational events that disrupt manufacturing and production operations. As such a broad set of problems need to be solved for increasing operational excellence. These include reducing downtime, increasing throughput, enhancing capacity utilization, increasing yield, reducing and eliminating defects, and lowering changeover time. For such operational excellence solutions, Operational AI needs to offer a broad set of process and asset-agnostic capabilities. Operational AI functions by applying machine learning to operational data to discover and report behaviours that engage operational experts and solicit their

Steel Times International

ARTIFICIAL INTELLIGENCE SUMMIT

Fig 2

know-how through accessible interfaces. Operational AI is easy-to-use, meaning that front-line subject matter experts such as maintenance and reliability engineers, process engineers, maintenance managers, and those with similar experience, can use it themselves. Operational AI produces applications that predict undesirable system behaviour and can be evolved without assistance from data engineers or data scientists. Hence results are achieved more rapidly, the solution scales more quickly, and at a lower cost than traditional analytics. From our experience working with leading international steel manufacturers, these are the features and benefits of Operational AI that have enabled scaled outcomes of AI in production: • Instant insights surfaced from the automated and continuous discovery of production events in real time • Explanations help understand why the event is happening with details of all the contributing factors and develops trust in the AI system • Capturing the human tacit knowledge enables collaboration across the operational teams and prioritizes actionables across the plant • Plant-level operational visibility provides granular reporting and workflow

Steel Times International

adoption and helps precisely understand the antecedents and consequences of events • Flexible deployment be it on cloud, edge, on premise or air gapped; the same solution is available from everywhere. Application on the field Operational AI can be applied across the steelmaking processes from casting to finishing lines. Various components of casting operations from the molds and pinch rollers to the bending roll can be monitored for predicting and detecting early degradations in critical components. Fig 1 illustrates the continuous monitoring of all the components of a vertical caster. By feeding the signals from all the components into the Operational AI system, we get a unified real time view of the dynamic operating modes of various components. The input signals involve multimodal data consisting of quantitative signals such as current, torque, motor speed, hydraulic pressure, etc and categorical signals such as value status, alarm codes and so on. Automated unsupervised machine learning identifies essential patterns of behaviour that give the operational experts a better idea of different operating modes. The anomalous events such as part failure, misalignments, or high force events are

marked as warning periods and the system then performs semi-supervised learning to create models that can discover and distinguish between patterns across supplied parameters. The ML application, now having learned to detect and recognize the relevant operations patterns occurring before conditions of interest, can be used to predict the undesirable events of interest from the real-time data stream. This intelligence-first approach readily scales up across assets, processes and plants and provides the operations team with explanations along with contributing factors, in-the-moment guidance (decision support) and superior operational visibility. Fig 2 puts this into perspective: Over 700 asset twins are monitored for a range of use cases by analyzing around half a billion assessments. By leveraging a combination of unsupervised and semi-supervised learning, the models are developed and validated rapidly despite such large variety and volume of data. The ability to automatically understand the contributing parameters accelerates the troubleshooting and corrective action determination for the predicted failure conditions. Fig 3 illustrates the flow of the AI-based predictive maintenance solution.

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 73

ARTIFICIAL INTELLIGENCE SUMMIT

Fig 3

There are also opportunities for detecting quality defects in cooling tables, strip break classification and laser cutters. To illustrate how Operational AI can generate value, let’s look at one particular use case – cooling tables. In the casting process, the steel slab that comes out goes on to rollers, which number in their hundreds and are driven by motors. A key challenge here is that as the motor, the coupling, the gears and the roller

surface itself degrade, they start to drag on the surface of the steel, causing defects in the steel surface. So the quality is an outcome of the need for maintenance. With the help of Operational AI, the earliest possible indication of a poorly performing roller could be detected. The maintenance team could get explanations from the captured information and drill down to know that the faulty drive motor, for

instance, needs a corrective action. Another case involves strip breaks. One of the biggest challenges in this application is not identifying whether the strip break has happened, but whether the potential causes of that strip break can be identified. Operational AI can characterize all of the associated equipment to identify potential causes of any series of strip breaks, and then use that information to correct the process.

Fig4

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT

Steel Times International

Conclusion For the AI to scale up and succeed in the steel mills, it is important to think beyond the use cases and cast a plant-wide net to discover and detect anomalous behaviours in the equipment and the processes. This allows the operations team to have a unified view of the operational excellence goals spanning across reliability, quality and process efficiency. By putting AI directly in the hands of the operational experts, the insights are developed and validated at a faster rate, which is a prerequisite for scaled implementation of AI in production (Fig 4). Most importantly, this ensures that the AI adapts seamlessly to the changing business and operating conditions, building trust and buy-in from the operations team. With scaled adoption of Operational AI, leading steelmakers expect to achieve up to 50% year-over-year reductions in unplanned Steel Times International

downtime. Falkonry has invested millions of dollars to build an AI platform that unlocks the above mentioned opportunities in digital operations. With over eight years of research and development and over a dozen issued and pending patents, Falkonry is committed to making industrial AI accessible, trustable and accountable to the frontline operations team in steel manufacturing, so that companies can set new standards of achieving AI-enabled operational excellence. �

aistech-technical-paper “Guide to Operational AI” Falkonry Inc., [Online] Available: https://info.falkonry.com/ buyers-guide-2021 C. Lee, “The Intelligence-First Path to Predictive Operations,” Falkonry Inc., 6 July 2020. [Online]. Available: https://info.falkonry.com/theintelligence-first-path-to-predictive-operations/ “Blog: International Society of Automation” [Online] Available: https://blog.isa.org/worlds-largestmanufacturers-lose-1-trillion/year-to-machinefailure

References C. Waters, B. Klemme, R. Talla, P. Jain, N. Mehta, “Transforming Metal Production by Maximizing Revenue Generation With Operational AI” Falkonry Inc., 29 June 2021. [Online]. Available: https://info.falkonry.com/

“Prediction at scale” McKinsey, 2021 [Online] Available: https://www.mckinsey.com/businessfunctions/operations/our-insights/prediction-atscale-how-industry-can-get-more-value-out-ofmaintenance

ARTIFICIAL INTELLIGENCE & STEELMAKING SUMMIT 75

LIGHTENING THE IMPACT OF HEAVY INDUSTRY

SPEAKERS CONFIRMED INCLUDE:

Lord Adair Turner Chair Energy Transitions Commission

CAN YOUR COMPANY CONTRIBUTE TO A CLEANER INDUSTRIAL FUTURE?

Anthony Hobley Executive Director Mission Possible Partnership

Andrew Purvis

Pernelle Nunez

Dr. Max Åhman

Gökçe Mete PhD

Director Safety Deputy Secretary General Environment and / Director – Sustainability Technology International Aluminium World Steel Association Institute

Christina Sobfeldt Jahn

Head of PPA Origination & Execution Ørsted

Henning Bloech

Ilhan Savut

Jean-Marc Moulin

Lead Analyst - Circular Economy BloombergNEF

Director of Sustainability Extruded Solutions Norsk Hydro

Chris Bayliss

Anne-Claire Howard

Global Director Sustainable Solutions Mitsubishi Chemical Advanced Materials

Director of Standards Aluminium Stewardship Initiative

Geoff Matthews

Dilip Chandrasekaran

CEO ResponsibleSteel

Show the world how at SIM Europe 2022. Hosting leaders from the worlds of industry, innovation, science, government and investment, Sustainable Industrial Manufacturing (SIM) in Brussels will provide an opportunity for those at the frontier of cleaner industrial manufacturing to present sustainable solutions to some of the world’s largest industrial companies and manufacturers across five hard-to-abate sectors.

40+ years of experience delivering events for the manufacturing industry

Head of Secretariat, Leadership Group for Industry Transitions (LeadIT) and Research Fellow Stockholm Environment Institute

Cédric de Meeûs

Dolf Gielen

Vice-President, Group Public Affairs & Government Relations Holcim

WHAT MAKES SIM EUROPE UNIQUE? No other event is inviting decision makers from across hard-to-abate sectors

Associate Professor in Environmental and Energy Systems Studies Lund University

Three exhibition zones will deliver end-to-end solutions for visitors

Director IRENA Innovation and Technology Centre

Modulation Specialist EnergyFlex Pty Ltd

Dr Jörg Rothermel

Managing Director Energy, Climate Protection, Raw Materials German Chemical Industry Association (VCI)t

Head of R&D and Technology Kanthal

Philippe Bastien Regional President, Architectural Glass Division AGC Glass Europe

Stefan Grüll

CEO and Co-Founder S1Seven GmbH

Zakia Khattabi

Minister of the Climate, The Environment, Sustainable Development and Green Deal, Belgium