With the advancement of artificial intelligence (AI) technology and the increasing demand for deeper digital-physical integration, the traditional L1-L5 five-layer system in the steel industry can no longer meet the requirements of digital transformation. This paper begins by analyzing the connotation and characteristics of digital transformation in group-type steel enterprises.It elaborates on the digital system architecture of such enterprises driven by new-generation information technologies, as well as the three foundational pillars essential for digital transformation and smart factory construction: the industrial internet platform, big data center, and steel AI engine platform. Practical applications in smart governance, smart manufacturing, and smart services are illustrated through specific scenarios. Finally, using Baowu Steel Group, a typical group-type steel enterprise, as an example, the paper explains the evolution path of digital transformation—based on the integration of digital and realworld processes—in response to changes in steel manufacturing models and advancements in AI technology. This path progresses from specialized management at the individual enterprise level to crossdomain integrated operations, from AI-steel integration enabling fullprocess intelligence, and from singleentity management to ecosystem collaboration across the entire value chain. The digital transformation path and system architecture proposed in this paper provide significant reference value for promoting the digital transformation of group-type steel enterprises in China.

In automation control for metallurgical unmanned overhead cranes,a critical challenge lies in balancing operational speed and anti-swing measures to enhance crane efficiency while ensuring safe operation. A metallurgical unmanned crane anti-swing full-range speed control method was pro- posed based on controller parameter switching derived from human expertise under multi-run states in this article. The method first plans the full-range speed control for the overhead crane linked by a large and small vehicle under anti-swing conditions. Then anti-swing control is optimized by switching PID parameters based on human crane operation experience under different running states,forming an integrated feedforward-feedback hybrid control for the unmanned overhead crane. Simulation experi- ments demonstrate that,compared to the commonly used manual operation of large and small vehicle linkage control,this method reduces overall operation time by 24. 9% . Field experiments show thatthe maximum swing angle is controlled within the range of 0. 45°-0. 85 °,and positioning accuracy is maintained within 10 mm. The relevant indicators are in a leading position in similar domestic scenar- ios. This method has been successfully applied and validated in the steel industry in China.

At present , a neWWaVe of technological reVolution is emerging globally , represented by in- dustrial technologies such as artificial intelligence , big data and industry 4 . 0 .  The traditional chemical plant control model can no longer meet the  needs  of digital  manufacturing.   Typical  problems  such  as complex production processes and difficulty in collaboratiVe operations in the steel industry haVe gradu- ally emerged.  Enterprises also haVe uneVen information leVels  and serious information island phenome- na.  In this regard , the factory and production line based on emerging technologies such as digital tWin technology , internet of things , machine learning and Video intelligent recognition , accesses Video ima- ges , sensor monitoring data , system  control  data , external  draWing  information  and  management  sys- tems , and the status of production line  equipment and materials  in real time Were  established , Visual linkage of  data  information  Within  the  production  line  Was  displayed ,  and  information  islands  Were eliminated.  At the same time , artificial intelligence algorithms Were used to merge and splice the com- plex and scattered Video images , and splice the production line monitoring images.  For large-scale key areas , real-time monitoring Was achieVed in the form of“ one picture ”.

With the vigorous development of computer technology, the application of AI and large models in the steel industry has become a key force in promoting the intelligent transformation of the industry. This research mainly focuses on the application of large-scale models in the steel industry. Firstly, it summarizes the construction methods and typical application areas of industrial large-scale models; Secondly, elaborate on the characteristics of the steel industry and summarize the relevant technologies of the steel industry's large-scale model; Finally, a discussion will be conducted on the application scenarios of the large model, highlighting typical application scenarios of the large model in the steel production process, such as breakthroughs in perception and cognitive tasks. In the future, the steel industry is expected to deeply apply big models to the development of new products and systems, as well as provide comprehensive decision support, realizing the application of energy, raw material scheduling, and full process monitoring of steel enterprises, promoting the high-end, intelligent, and green development of the steel industry, and providing new ideas for digital development.

In response to the systemic challenges faced in the “multi-variety，small-batch” production model for special seamless steel pipes—including complex quality inspection dimensions，the absence of full-process single-unit traceability，heavy reliance on manual experience for production scheduling，and inadequate closed-loop quality control—this paper proposes and implements a “four-in-one” digital and intelligent factory solution encompassing “inspection-tracking-scheduling-control.” This solution establishes a quality inspection system based on deep learning and multi-modal visual fusion，achieving high-precision online measurement of pipe defects and dimensions. It pioneers a single-pipe tracking method that integrates “video AI+event logic”, overcoming the industry-wide challenge of accurately mapping data to individual pipes throughout the entire process. A dynamic scheduling optimization model，based on a hybrid genetic and simulated annealing algorithm，was developed，significantly enhancing scheduling efficiency and resource utilization in complex order environments. Furthermore，a collaborative quality management system based on the plan-do-check-act (PDCA) cycle was established，enabling systematic and proactive quality management. Application results demonstrate that the system achieves a defect detection rate of 99.95%，a material tracking accuracy of 9998%，an increase in scheduling efficiency of over 60%，and a 50% reduction in product non-conformance rates. This work provides a valuable and exemplary model for the digital transformation of China’s special steel industry.

Three persistent challenges in steelmaking-continuous casting process optimiZation were  ad- dressed : inefficient process modeling , imprecise parameter control , and inadequate  multi-process co- ordination.  Three core technological innovations were developed : intelligent ladle furnace  (LF) metal- lurgy , precision-controlled continuous casting solidification , and dynamic multi-process collaboration.Firstly , an intelligent LFcontrol system through the  integration of metallurgical principles with inter- pretable machine  learning  was  established.    This  system  enables  accurate  regulation  of molten  steel  temperature , alloy composition  ( si/Mn content) , and  argon  stirring parameters , substantially reduc - ing both material consumption and energy usage while eliminating traditional reliance on empirical ad- justments and repeated sampling.  secondly , for continuous casting optimiZation , a solidification cool- ing strategy based on steel phase-transformation characteristics was proposed.  Through noZZle arrange- ment optimiZation and secondary-phase  precipitation control , this  approach effectively mitigates  crack formation and segregation defects in microalloyed steels .  A predictive model combining principal com- ponent analysis with deep neural networks further enhances process control by guiding real-time param- eter adjustments .   To  resolve  production  coordination  challenges  in  complex  steelmaking  workshops , dynamic collaborative operation technologies rooted in metallurgical process engineering theory was de- veloped.  The implementation of quantitative  coordination metrics  ensures  efficient  material flow  man- agement , significantly improving operational synchroniZation across multiple processes .

In order to achieve more accurate calculation of alloy addition amount,a comprehensive al- gorithm based on genetic algorithm ( GA) optimization for BP neural network was adopted. During the training process of the BP neural network,the refining starting composition of the molten steel was used as input parameters for the BP neural network model. Then,the fitness function of the GA was used to optimize and adjust the weight and threshold of BP neural network,predicting the refining endpoint steel composition of the LF furnace. By comparing the prediction results of BP neural net- work algorithm and GA-BP neural network algorithm, it was found that the GA-BP algorithm has smaller mean absolute error (MAE) and mean square error (MSE) ,and the prediction results aremore accurate and basically consistent with the actual steel composition,indicating that this model can be used in production. Based on the GA-BP neural network model,the alloy addition amount was determined according to the composition of the molten steel at the beginning of LF refining and con- trol composition requirements. After deploying a predictive model in the 140 ton ladle LF refining sys- tem of a steel plant and tracking 188 furnace data,the difference between the actual amount of alloy added and the predicted amount was within ±30 kg,then the prediction accuracy of high manganese alloy was 91. 3% ,high chromium alloy was 90. 4% ,ferrosilicon alloy was 90. 2% ,and the predic- tion accuracy of carburant was 91% in furnaces,from which can guide the determination of alloy ad- dition amount in the actual refining process. 

The  current  status  of  steel  scrap  utilization ,  classification ,  and  classification  standards  in china were analyzed , the current situation of steel scrap utilization and classification sorting were  ex- plored , and  the  existing  problems  were  put  forward.   On  this  basis ,  development  and application of steel scrap intelligent identification system and rapid detection technology were analyzed.  It is consid- ered that there is still a big gap between the development level of steel scrap classification and sorting technology and the actual demand of steel enterprises .  In the future , the coupling of the two technolo- gies must be  achieved  in  order  to  realize the  intelligent  smelting  of  electric  arc  furnace.  The  article points out that the classification and sorting of steel scrap in china is still relatively primary and exten- sive , it is not yet possible to achieve rapid and effective identification of the  composition of steel scrap.under the background of“ dual-carbon”, as china , s steel industry shifts from the stage of large-scale and high-speed development to the  new  stage  of green  low-carbon  and high-quality  development , the importance of steel scrap  quality  is  becoming  more  and  more  important .   The  utilization  level  of  steel scrap  should be improved as soon as possible , not only to improve the scrap ratio ofsteelmaking , more importantly , it is necessary to improve the technical level of steel scrap classification and sorting , so as to lay a foundation for the realization of intelligent smelting .

The prediction of the endpoint carbon content and molten steel temperature in converter steelmaking is of great significance for the precise control of molten steel composition and temperature and the improvement of product quality. Due to the high dimension, high noise and strong nonlinearity of the converter steelmaking process data, it is difficult to obtain high-quality data. Direct modeling not only has a low hit rate but also is prone to overfitting. To address this problem, this paper proposes a random forest prediction model based on adaptive SMOTE data enhancement technology was proposed. Firstly, the feature selection of the endpoint carbon content and molten steel temperature iswas performed by recursive elimination method. Secondly, the adaptive SMOTE algorithm iswas used to enhance the original data. Finally, random forest iswas used to predict the endpoint carbon content and molten steel temperature respectively. The actual industrial data shows that the prediction hit rate of the end-point carbon content within the target error value ±0.02% is 88.9%, and the prediction hit rate of the end-point molten steel temperature within the target error value ±20°C is 92.3%, which is significantly improved, and provides a reference for end point prediction and control of converter steelmaking.

Crane scheduling in steelmaking and continuous casting section is a typical multi-machine and multi-task constraint problem,which is of great significance to the connection and forward flow of each process and the control of the production rhythm of the whole steelmaking plant. In order to im- prove the efficiency of crane operation and ensure the stability of production,a crane scheduling mod- el based on spatio-temporal heuristic algorithm was established by analyzing the process flow and the constraints of crane job scheduling in steelmaking and continuous casting section. Partition rules,task allocation rules,collision avoidance rules and state update rules are designed to characterize the oper- ation process of the crane. The model is solved by a heuristic method,which can avoid the high com- putational complexity of traditional theoretical methods in large-scale problem solving. The results show that the model can provide a reasonable scheduling scheme and effectively avoid the spatio-tem- poral conflict during the operation of the crane. Compared with the actual production scheduling method,the performance of the crane under the heuristic algorithm model has been improved,and the operation efficiency of the crane has been improved.

Blast furnace burden distribution is one of the crucial aspects of blast furnace ironmaking.  Due to the airtight nature of the top loading equipment , it is not possible to intuitiVely and accurately obserVe the actual distribution of furnace burden in the furnace throat , such as the distribution of bur- den layers and ore-to-coke ratio.  During production , blast furnace operators typically rely on the distri- bution of cO2  , coal gas temperature , or gas floWrate in the coal gas at the furnace throat for upper ad- justments.  With the adVancement of detection technology , operators can utilize infrared cameras , ther- mal imaging , laser technology , and radar technology to perform more precise upper adjustments.  HoW- eVer , these technologies are only capable of capturing information pertaining to the  surface of the fur- nace material , and they fail to proVide a detailed understanding of the distribution information regard- ing the material layer and ore-to-coke ratio.  Based on the mathematical model of blast furnace burden distribution , the research comprehensiVely applies cutting-edge technologies such as numerical simula- tion , supercomputing , and artificial intelligence.  Through big data cloud platforms and intelligent neu- ral netWorks , an intelligent monitoring system for the distribution of the entire blast furnace burden lay- er Was constructed.  This system can achieVe online tracking and simulation of the material distribution process , assisting blast furnace operators in understanding the  distribution of furnace materials.   After the application of the system , Various economic and technical indicators of the blast furnace haVe been significantly improVed.   compared  With  the  past ,  the  monthly  output  has  increased by  an  aVerage  of 4. 7% , the  fuel  ratio  has  decreased  by 2. 37  kg/tFe ,  and  the  Vanadium  recoVery  rate  is  all  aboVe 8o% .

The  prediction  of  the  endpoint  carbon  content  and  molten  steel  temperature  in  conVerter steelmaking is of great significance for the precise control of molten steel composition and temperature and the improVement of product quality .  Due to the high dimension , high noise and strong nonlinearity of the conVerter steelmaking  process  data ,  it  is  difficult  to  obtain  high-quality  data.  Direct  modeling not only has a loWhit rate  but  also  is  prone  to  oVerfitting.   To  address this problem ,  a random forest prediction model based on adaptiVe SMOTEdata  enhancement technology  Was  proposed.   Firstly ,  the feature  selection of the  endpoint carbon content and molten steel temperature Was performed by recur- siVe elimination method.   Secondly ,  the  adaptiVe  SMOTEalgorithm  Was  used  to  enhance  the  original data.   Finally , random forest Was used to predict the endpoint carbon content and molten steel tempera- ture respectiVely .  The actual industrial  data  shoW that  the  prediction  hit  rate  of the  end-point  carbon content Within the target error Value   ±0 . 02%  is 88 . 9% , and the prediction hit rate  of the end-point molten steel temperature Within  the  target  error  Value   ± 20  ℃   is  92 . 3%  ,  Which  is significantly im- proVed , and proVides a reference for end point prediction and control of conVerter steelmaking.

With the rapid development of industry 4 . 0 and intelligent manufacturing , the  steelmaking industry is facing great opportunities  and challenges of transformation and upgrading .   The  application of steelmaking  intelligent  technology  not  only  improves  production  efficiency  and  optimiZes  product quality , but  also  significantly  reduces  energy  consumption  and  environmental  pollution ,  which  pro- motes the development of the steel industry in the direction of green , intelligent and sustainable.  The current status and future development trend ofsteelmaking intelligence was reviewed , and the integrat- ed application of artificial intelligence , big data , internet of things and other advanced technologies in the electric arc furnace , converter and refining process was focused on.  The key technologies of steel- making intelligence , including intelligent control system , data acquisition and monitoring technology , machine learning  algorithms , refining  process  optimiZation ,  etc.  ,  were  detailed.   It  also  summariZes the application cases of various types of intelligent technologies in improving the control of molten steel composition , smelting process precision , and energy utiliZation efficiency .  In addition , the challenges and bottlenecks ofsteelmaking intelligence in practical applications were discussed , such as data qual-ity , system integration , real-time and adaptability issues .  Finally , the future development direction of intelligent steelmaking technology was looked into , especially the potential in the fields of digital twin , automation control , intelligent prediction and optimiZation , aiming to  provide  reference  and guidance for the technological innovation and intelligent transformation of the steel industry .

 

Against the backdrop of intensifying global market competition, fluctuations in energy and raw material prices, and the implementation of carbon tariff policies, Xingcheng Special Steel has addressed issues such as insufficient enterprise customization capabilities, a lack of systematic product quality evaluation, high energy consumption due to “black-box” operations in blast furnace ironmaking, poor production coordination in the steel rolling process, and the absence of systematic energy management. Leveraging digital technologies such as the industrial internet, artificial intelligence, big data, and simulation, and drawing on the concept of lighthouse factory promoted by the world Economic Forum, Xingcheng Special Steel has deployed over 40 use cases of the fourth industrial revolution. This has led to remarkable improvements, including a 35.3% increase in customized orders, a 47.3% reduction in nonconforming product rates, and a 10.5% decrease in energy consumption per ton of steel. Through typical use case practices such as manufacturing process customization design supported by big data analysis, transparency in blast furnace black-box operations through multimodal simulation, steel quality enhancement via intelligent closed-loop control, efficient steel rolling process enabled by AI, and optimization of energy, water, and carbon resources driven by advanced analytics, the company has achieved significant success in improving production efficiency, reducing costs, innovating product development, enhancing product quality, and boosting customer satisfaction. These efforts provide a valuable reference for the digital transformation of the steel industry.

Under the background of green and intelligent construction in the steel industry , traditional single-process quality control systems become insufficient to meet the demands of smart steel manufac - turing.  They need to eVolVe into full-process integrated process control systems.  These systems need to transition from  mere  model  calculations  to  collaboratiVe  Whole-process  process  model  manufacturing , production data  mining ,  and  intelligent  system  decision-making ,  self-adaptation ,  and  self-learning.  with this objectiVe in mind , the system architecture , research and deVelopment content , and anticipa- ted Vision for the deVelopment and application of intelligent process control systems and process models across  four  dimensions : intelligent  egaipment , horizontal  process  floW, Vertical  information  transfer , and time Were explored.

As a key equipment in the hot-rolled Wide and thick plate production line , it is a complex system With multidimensional factors coupled With each other.  Especially during operation , due to the strong coupling and nonlinear characteristics  of Various  influencing  factors , the  rolling  mill  operation process is in a“ black box ”state , Which makes it difficult to accurately grasp the rolling mill operation status in real time and formulate preVentiVe measures.  This has become  a bottleneck problem that re- stricts the high-precision and stable rolling of the hot-rolled Wide  and thick plate  production line.   To solVe this industry problem , data-driVen monitoring and analysis technology for the status of hot-rolled Wide  and thick plate rolling mills Was introduced.  Based on deep perception of Various data in the roll- ing process , big data algorithms are applied for data correlation analysis , and a rolling process dynam- ic model and rolling mill status online monitoring system platform With multi-source data-driVen fusion mechanism are constructed.  After being applied in releVant production lines , the accuracy of the sickle bending offset model has exceeded 95% , and the Vibration speed of the precision rolling mill has been reduced by 64 . 8%  ~72 . 8% .   It can effectiVely  identify 90%  of the rolling mill state data , effectiVely improVing the real-time accurate perception of the rolling mill state and the ability to analyze and pre- dict abnormal states.

LFrefining is a metallurgical process inVolVing multi-process , multi-steel type and complex enVironment , a single-process  control  model  cannot  meet  the  demand  for  full-process  automation  and intelligence in smelting.   Based on data and experience , by adopting  self-learning algorithms , mecha- nistic models , expert system , image  recognition , mathematical model to  create  models  for alloy , slag formation , temperature control , argon bloWing , and Wire feeding , and using a Variety of detection de- Vices  and multi-functional robots to  achieVe the  intelligent temperature  sampling , and  closely integra- ting multi-process model  and  L1  automated  control  system ,  an  intelligent  control system for  the full- process  of LFrefining Was realized.  The system adopts a modularized and multi-threaded control meth- od , completing the  smelting steps by a single-step incremental Way through time sequence analysis.  It is  applied in 150t double-station LF, and the aVerage operation time is shortened about 4 min , the hit rate  of molten steel prediction temperature error Within  ± 8  ℃  is more than 93%  , and the success rate of temperature measurement and sampling is 97% .

In order to ensure the manufacturing quality and yield of high-end products, a data-driven quality insight analysis system based on HBIS digital industrial internet platform was constructed. The system integrates machine learning algorithms and image processing technology through multi-source data collection and intelligent processing, achieving quality data tracking of the entire iron production process. The system has established a linkage mechanism of online monitoring-intelligent diagnosis-closed-loop control, which optimizes process parameters, traces defects, and predicts abnormal working conditions for key processes such as steelmaking, hot rolling, and cold rolling, forming a digital control loop covering quality index prediction and process diagnosis. Since the application of the system, through the analysis of quality data throughout the entire process, the product qualification rate has been improved, the narrow window control level of indicators has been enhanced, and the quality assurance needs of high-end manufacturing have been effectively supported.

In the iron and steel industry, traditional PID control of blast furnace hot stoves faces challenges such as nonlinear hysteresis and poor multi-condition adaptability, which is difficult to meet the dual requirements of energy efficiency optimization and environmental compliance. A fusion control method based on AI agents was proposed, and a three-in-one intelligent control system of “prediction-knowledge-optimization” was constructed. Through the double-layer coupling of the LSTM algorithm and the knowledge embedding technology, the mean absolute error (MAE) of the dome temperature prediction is achieved at 5.3 ℃, which is 58% higher than that of the traditional ARIMA model. By establishing a four-dimensional knowledge representation system and combining it with the Rete rule engine, a response of 9.8 s for abnormal working conditions is achieved. By developing the DeepSeek lightweight decision-making engine, multi-objective optimization was achieved with 3.2 iterations of convergence, reducing the unit consumption of hot air per ton of iron by 14.8% and achieving a quarterly energy-saving benefit of 9 million yuan. Industrial validation demonstrates that this system significantly enhances thermal efficiency and environmental performance in 3200 m3blast furnace applications, providing theoretical and engineering references for the intelligent upgrading of industrial furnaces.

The connection and scheduling of the casting-rolling interface is an important factor affecting the production efficiency and economic benefits of iron and steel enterprises.  For the continuous cast- ing and hot rolling production site of a steel enterprise , a multi-agent simulation model of the casting- rolling interface Was established , and the  complete  production  and  transportation  process  on  the  pro- duction site Were modeled by designing the roles and tasks of Various agents and the cooperatiVe inter- action mechanism betWeen  agents.   A Variety  of  different  production  rhythms  Was  simulated  and  ana- lyzed by multi-agent model , and the precise matching betWeen different production rhythms of continu- ous  casting and different production rhythms of hot rolling on the production line Was studied.  The sim- ulation results shoWthat the model can accurately reflect the actual production process , Which Verifies the effectiVeness of the simulation model.  According to the simulation model , as the interVal time be- tWeen the different streams of the casting machine increases , the conVeying time of the hot slab in the process of traVersing trolley and the total conVeying time of the hot slab shoWed a trend of first decrea- sing and then rising , and reached their loWest When the interVal time betWeen the  different streams of the casting  machine  Was  2  min.    when  only  producing  hot  billets ,  With  the  increase  in  the  casting speed , the aVerage total conVeying time increases , the total production time  decreases , and the  aVer- age reheating time increases first and then decreases  after the  casting  speed  aboVe  1 . 9  m/min.   As  a result of the optimization , the aVerage reheating time Was reduced to 95 . 7%  and the aVerage total con- Veying time Was reduced to 68 . 2%  of the pre-optimization leVel.

As a key index to evaluate the shape quality of medium and heavy plate,the plate crown plays an important role in determining the market competitiveness of steel companies. With the devel- opment of computer technology,intelligent control has become the most focused subject in steel in- dustry. However,artificial intelligence models of predicting plate crown still lacks interpretation and are difficult to effectively guide the practice. Based on problems above,a medium and heavy plate crown prediction model based on data mining and multi-model fusion was proposed. Firstly,rolling data was projected onto spaces which have less dimensions,which solved the multi-source heteroge- neity problem of data. Secondly, samples were expanded by using the improved synthetic minority oversampling technique ( SMOTE) method based on Bootstrap,which solved the problem of low mod- el learning performance and accuracy caused by imbalance data. Then,an integrated model was pro- posed,which used predicted value of Ridge regression model as the main value and prediction of BPneural network model as the deviation. The experimental results show that mean square error (MSE) of this model is 0. 004 mm 2 ,and absolute error within 100 of model is more than 95% . On one hand, the range of prediction error is further reduced and the prediction accuracy of the medium and heavy plate crown is improved through multi-model fusion. On the other hand,the changes of feature varia- bles affect the plate crown is quantitatively obtained by reversely deducing plate crown prediction for- mula,which not only made data reflect and effectively guide the actual production,but also improved interpretability of plate crown prediction model as well. This paper provided a strong reference for control of plate crown and useful new ideas for future development and research of related fields.

With the advancement of technology and economic development, the steel industry is facing new challenges and opportunities. Intelligent factories are a new type of industrial life form that deeply integrates digital technology and manufacturing industry. They have become the core force driving the upgrading of manufacturing industry during the critical period of industrial intelligence transformation and upgrading. This article takes Hebei Iron and Steel Group Shijiazhuang Iron and Steel Co., Ltd. (hereinafter referred to as “Shisteel”) as an example, focusing on the construction and operation of a life form intelligent factory, and systematically studying the transformation and upgrading path of special steel enterprises driven by the deep integration of digital technology and process innovation. Through the integration of industrial Internet, big data, artificial intelligence and other cutting-edge technologies, the cluster deployment of intelligent equipment and the implementation of intelligent manufacturing in the whole process, the intelligent high-quality development of production lines will be pushed to a new height, that giving life a strong physique and intelligent thinking. The fact shows that the construction of intelligent factories in the form of living organisms has achieved significant results in production efficiency, quality and cost control, and green development for Shigang, providing a new model for the high-quality development of the steel industry that can be referenced.

In order to realize the rapid construction of the knowledge graph for the convertible steelmaking process scene while balancing its accuracy and practicality, this paper proposes a knowledge graph construction mode that integrates top-down and bottom-up approaches. On the one hand, guided by the principle of event evolution and under the direction of expert experience, this method completes the definition and classification of ontology classes and constructs the schema layer of the knowledge graph for converter steelmaking process scenarios using a top-down approach. Concurrently, it leverages a knowledge extraction tool based on LLM Graph Transformer to achieve automatic extraction of entities, relations, and attributes, thereby constructing the data layer of the knowledge graph for converter steelmaking process scenarios through a bottom-up approach. By integrating the schema layer and data layer, this method efficiently constructs the knowledge graph for converter steelmaking process scenarios.

The quality  disputes  during  inspection  of  scrap  steel  have  always  been  an important issue plaguing major steel enterprises .   In order to  solve the problems  such as the  great influence  of subjec - tive factors , the  difficulty  in  tracing  the  grading  process ,  and  quality  disputes  existing  in the  scrap steel quality  inspection  process ,  the  scrap  steel  intelligent  grading  system  based on  artificial  intelli- gence technology has emerged as the times require and has received great attention in the steel indus- try .  As a new thing , many enterprises have  doubts , incomplete  and unscientific  understandings , and even misunderstandings about the scrap steel intelligent grading system.  Based on the technical princi- ples  and  the  functions  of  the  technical  architecture ,  the  key  technological  breakthroughs  in  aspects such as the automatic collection of pictures , standard unification , the intelligent grading ofspecial ma- terial types like  briquettes , and  intelligent  deduction  of  impurities  were  expounded  on.   At  the  same time , it points  out  the  engineering  challenges  faced  in  aspects  such  as  the  identification  of  chemical compositions , the recognition of small material types , and the internal quality inspection of briquettes .Finally , it puts forward the development process of the application of the scrap steel intelligent grading system in enterprises and its future trends .

In the sintering process , accurate prediction of the sintering burn-through point is essential to control the operation of the  sintering machine , as  it determines  the  quality  of the  sintered  material and the efficiency of energy use.  HoWeVer , the current common sintering burn-through point prediction models mainly focus on single-step prediction , and they tend to ignore the spatio-temporal characteris- tics of the sintering process data.  In VieWof the  multi-step prediction task requirements of the  sintering burn-through point and the  complex  spatio-temporal  characteristics  of process  data , a  multi-step  pre- diction model of the sintering burn-through point based on encoder-decoder architecture Was proposed.  The spatio-temporal encoder-task decoder architecture  effectiVely  extracts  the  spatio-temporal  features of sintering process data by stacking temporal conVolutional netWorks and spatial attention mechanism , and adopts a task-specific guidance decoder to correspond to each step of prediction in the form of in- dependent units.  The effectiVeness of the proposed method Was Verified by an actual sintering industrial process case.  This study can assist operators to obtain information about the future state of the sintering process in adVance , so as to adjust the operating parameters more accurately , reduce the quality fluc - tuation and energy Waste caused by the lag , and improVe the quality stability of the sinter.

Aiming at the problems of low production efficiency and poor information collaboration in the traditional steel rolling industry, this study carries out the construction practice of a digital rolling plant based on the industrial Internet platform. Through a series of measures, including constructing a centralized control center, upgrading digital production lines, promoting full-process digitalization of production operations, realizing intelligent decision-making for production control, establishing a visual performance dialogue mechanism, strengthening precise energy consumption control for processes, implementing digital management of the entire equipment life cycle, and improving the digital monitoring system for safety and environmental protection, the factory’s intelligent system is comprehensively improved. During the practice, the data barriers in all production links are effectively broken, and the deep sharing and integration of full-process data flow and information flow are achieved. This successfully promotes the transformation of the enterprise from the traditional heavy industry model to a new innovative development model with significantly improved automation, informatization, and scientific management levels. Specifically, product energy consumption is reduced by 12.7%, the yield rate and qualification rate are increased by 0.1% and 0.7% respectively, and production efficiency is improved by 18%. This study provides practical experience and theoretical basis for the digital upgrading of the steel rolling industry.

In modern industrial production, precise and efficient steel grade matching plays a pivotal role in material selection, scheduling, and process optimization. However, confronted with the escalating demands for complex multi-dimensional, fuzzy matching, similarity ranking, and weighted queries, traditional relational database-based query methods exhibit inherent limitations, resulting in low efficiency and insufficient flexibility. To address these challenges, this paper proposes an innovative framework for steel grade matching based on vector embedding and high-performance vector databases. The method first standardizes diverse and heterogeneous attributes of steel grades, such as chemical composition, mechanical properties, and heat treatment status, transforming them into unified-dimensional vector representations (embeddings). Subsequently, these vectors are stored in a high-performance vector database employing a hierarchical navigable small world (HNSW) graph index, enabling rapid and accurate steel grade querying and matching through Approximate Nearest Neighbor (ANN) search techniques. This paper elaborates on how various practical matching scenarios—including target value matching, range matching, multi-dimensional weighted matching, and similarity-based lookup from existing steel grades—are unified and implemented as retrieval problems within the vector space. Experimental results clearly demonstrate that when handling steel grade data at the scale of hundreds of thousands, the proposed vector-based method achieves millisecond-level query responses, showcasing significant efficiency gains and a breakthrough in capabilities for fuzzy and complex query scenarios. This research significantly enhances the efficiency, accuracy, and scenario compatibility of steel grade matching, offering a novel technical paradigm and implementation path for data-driven intelligent material management in industrial applications.

A method of unifying the coordinate system of multi-line laser cameras and the coordinate system of rotating axis based on the light cylindrical shaped target was proposed,which aims to realize the high precision dimensional measurement of the shaped workpiece. By combining the data collect- ed by three linear laser cameras and the dimensional prior information of cylindrical shaped targets,a series of nonlinear equations with the attitude of linear laser cameras as variables are established. By using RANSAC and nonlinear least squares algorithm to solve the equations,the attitude of the line laser camera is obtained and verified by practical measurement. The results show that the calibration accuracy of this method can reach 0. 043 mm,and the accuracy of measuring the train wheel with a diameter of 760 mm can reach 0. 27 mm,and the percentage accuracy can reach 0. 036% . This cali-bration method provides a new idea in the field of rotation measurement of shaped workpiece. 

In response to challenges such as overcapacity in the steel industry, the “dual carbon” goals, the upgrading of customer customized demands, and the technological transformation of Industry 4.0, Xichang Vanadium and Titanium Steel has initiated the construction of a smart factory for high-quality vanadium-titanium steel. The aim is to achieve cost reduction, efficiency improvement, flexible production, and sustainable development through digital transformation. Addressing the problems of data isolation horizontally and attenuation vertically in the traditional pyramid architecture of information systems, Xichang Vanadium and Titanium Steel has carried out top-level architecture design, adjusted business, technical, data, and application architectures, and explored the migration of existing applications to a cloud-edge collaborative industrial Internet platform to achieve deep integration of systems and data. Focusing on scenarios such as “digital design of processes, dynamic optimization of processes, intelligent online detection, online operation monitoring, intelligent warehousing, real-time monitoring and emergency response of safety risks, environmental pollution monitoring and control, and supply chain optimization”, data empowerment has been carried out. A domestic largest-scale integrated control center for the iron-making area of vanadium-titanium magnetite smelting has been built, and a three-in-one digital system for consistent quality management has been constructed. Centering on the customer, the entire supply chain data and collaborative optimization from production to sales, distribution, and transportation have been integrated, significantly enhancing product quality and enterprise competitiveness.

With the rapid advancement of general-purpose language models such as DeepSeek, research in the industrial domain is increasingly shifting from foundational model exploration to their specialized application in vertical, domain-specific scenarios. Taking converter steelmaking as a representative application context and the technological evolution of industrial-scale foundation models was systematically reviewed, including language, vision, and temporal models. Based on this analysis, it proposes a preliminary design for a multimodal large-scale model architecture tailored to the converter steelmaking process. The proposed architecture leverages heterogeneous data sources—such as scientific literature, flame video streams, flue gas composition, and molten iron chemistry—to address key technological challenges, including complex furnace condition recognition, dynamic endpoint prediction, and real-time process control. The objective is to overcome the theoretical limitations of conventional approaches in modeling high-temperature multiphase reactions, multivariate coupling behavior, and real-time decision-making under complex operating conditions. Furthermore, this study analyzes the core challenges encountered in the deployment of large-scale models in metallurgical applications—such as data heterogeneity, domain adaptation, and model interpretability—and discusses corresponding mitigation strategies. Finally, future development directions are explored to provide a theoretical foundation and methodological reference for the intelligent transformation of the steelmaking industry.

This paper focuses on the intelligent flexible production smart factory，exploring the practical paths and methods of building wire production line transformation towards intelligence and digitalization. By analyzing the production characteristics of the wire industry and the pain points in intelligent and digital transformation,this study integrates advanced technologies such as 5G,digital twins, and industrial internet to construct an architecture and implementation plan for a flexible production smart factory based on the industrial internet.It summarizes innovative models of smart factories in flexible production, efficiency improvement, and quality optimization,providing replicable experience for peers in the industry. The research shows that building a smart factory can significantly enhance the flexibility and intelligence of wire production,achieving multiple goals including quality improvement,efficiency optimization, and cost reduction.The value of this research lies in its implementation from research to production practice, all elements involved in the intelligent factory of excellence, such as intelligent decision-making digital twin, human-machine collaboration, production and operation management, supply chain coordination, R&D and design functions are observable and available, and equipment intelligence, industrial internet technology, management, etc. represent the latest achievements in the industry, and as one of the 19 steel enterprises selected as the first batch of 235 national pilot enterprises, is a model factory for the industry to learn from.

In steel production, full-process billet tracking is critical for ensuring production efficiency and product quality. Traditional technologies rely on multi-source sensors, which suffer from issues such as information errors and delays. Existing visual cross-domain algorithms are difficult to apply due to the limitations of steel plant scenarios. To address these issues, a cross-domain tracking algorithm based on Agent trajectory information association was proposed. Leveraging the fixed transportation paths and well-defined process nodes of billets, the algorithm integrates visual perception data with production rules through the collaboration of overhead crane and roller table Agent at network nodes, achieving precise cross-camera trajectory association without overlapping views or significant appearance features. Experimental results show that the algorithm achieves multi-object tracking accuracy (MOTA) of 95.8% and 82.1% in roller table and overhead crane cross-domain scenarios, respectively, outperforming comparative algorithms. This research breaks through the limitations of traditional tracking technologies and existing vision algorithms, providing an effective solution for intelligent tracking in metallurgical production. Future work will focus on enhancing performance through multi-modal data fusion and model optimization.

Based on the 1 580 mm hot rolling production line of a certain factory,in response to the problem that traditional thickness models cannot accurately reflect actual thickness,an improved stochastic configuration network ( SCN) based strip thickness prediction model was proposed. Firstly, from the perspective of rolling mechanism, the reasons for thickness fluctuations in hotrolled strip products were analyzed. Secondly, based on the original SCN, a stochastic configuration network based on hunting prey optimization algorithm ( HPO-SCN) and a stochastic configuration network based on hunting prey optimization algorithm and orthogonal triangular decomposition ( HPO-QR-SCN) were proposed. Then,using onsite measurement devices,parameters of three different thick ness specifications of strip steel products were collected to form a database of strip thickness. Parame ters related to export thickness were selected as input values for the model,and the Pauta rule was used to preprocess the original rolling data. SCN,HPO-SCN, and HPO-QR-SCN prediction models were established,and their prediction results were compared. The experimental results show that the proposed HPO-QRS-CN thickness prediction model has the shortest prediction time and the highest accuracy,with a model determination coefficient of 0. 963 8. At the same time,based on the model with the best predictive performance,the influence of rolling force and roll gap on the exit thickness of the strip steel was tested,and the results were in line with actual physical laws. The asymptotic be havior of the model was tested,with a root mean square error ERMS (RMSE) of 0. 059 8,indicating good approximation performance. 

Currently , many imported temper mill flatness closed-loop feedback control systems employ a sequential control strategy , Which proVes to be ineffectiVe in controlling compound WaVes and fourth- order  flatness  deViations.    MoreoVer ,  these   systems  struggle  to  handle  the   collaboratiVe  regulation among different flatness control mechanisms.  Focusing on a six-high CvCtemper mill at a steel plant , the  existing sequential control strategy Was optimized and a flatness closed-loop feedback control strate- gy based on the  RMsprop  gradient  descent  algorithm  Was  proposed.  The proposed strategy considers three regulation methods : Work roll bending , intermediate roll bending , and CvCshifting , aiming to quickly determine the combination of regulation quantities that minimizes flatness deViations.  This en- sures the full utilization of each control mechanism s(/) potential.  simulation and comparatiVe experiments demonstrate that the  flatness  control  strategy  based  on  the  RMsprop  algorithm  significantly  enhances the control of high-order and complex flatness defects , markedly improVing flatness quality compared to the original sequential control strategy .  Furthermore , the RMsprop-based flatness control strategy Was implemented Via  CFC programming ,  addressing  program  load  rate  issues  on-site  by  integrating  the strategy into the primary  control  system  in  parallel.   Results  confirm  that  the  RMsprop-based  strategy achieVes excellent application outcomes.

The end-point carbon content and temperature  of the  basic  oxygen furnace  (BOF)  are  the crucial factors for  ensuring  the  seamless  operation  of  steelmaking  production.   consequently ,  a noVel end-point prediction model based  on  the  L6Vy  flights  Whale  optimization  algorithm   ( LWOA)  and ε - tWin support Vector regression  ( εTSVR)  has  been  established.   Firstly ,  the  box  plot  Was  employed  to filter the data.  Subsequently , the key factors influencing the end-point carbon content and temperature Were identified through  metallurgical  mechanism  and  spearman  correlation  analysis.  Finally ,  the  pa- rameters in the εTSVRalgorithm Were automatically optimized by utilizing LWOA, Which possesses the characteristics of simple  adjustment  parameters  and  rapid  conVergence speed.  The simulation  results indicate that the proposed LWOA-εTSVRprediction model haVe hit rates of 89%  and 93%   at the  end- point carbon content and temperature satisfying the error tolerance of  ±0 . 005%   and  ± 10 ℃ , respec- tiVely .   MeanWhile , the  double  hit rate  reaches  83% .   compared  With  the  other  three  prediction  mod- els , the proposed LWOA-εTSVRmodel demonstrates superior adVantages.  Furthermore , by setting dif- ferent error interVals , the reliability of the performance of the proposed prediction model Was also Veri- fied.   MoreoVer ,  the  proposed  prediction  model has higher  prediction accuracy than the  actual steel plant process , proViding robust technical support for steel enterprises.

Integrated scheduling in steel production faces core challenges such as process coupling, resource constraints, dynamic disturbances, and conflicting objectives across stages including iron and steel ladle management, steelmaking, reheating furnaces, slab yards, hot rolling matching, and crane coordination. Existing intelligent agents, reliant on predefined rules and decision algorithms, exhibit limitations in comprehending complex rules and adapting to dynamic changes. This research proposes and validates a hierarchical hybrid multi-agent architecture based on large language models (LLM). Its core advantage lies in the effective integration of LLM advanced reasoning capabilities with the execution efficiency of traditional intelligent agents, compensating for the latter's deficiencies in adaptability, planning, and learning. This significantly enhances the system's capacity to manage complex rules and dynamic environments. Through the application of LLM, the system is endowed with the ability to autonomously understand business requirements, automatically utilize the model context protocol (MCP) to invoke a toolset for generating scheduling decisions, and provide natural language explanations. This approach not only optimizes human-computer interaction but also improves the system's scenario adaptability and scalability. This research offers crucial technical insights for developing next-generation intelligent steel scheduling systems capable of understanding requirements and adapting to changes, potentially accelerating the implementation of agents in scheduling systems.

In the context of new industrialization, digital and intelligent transformation has become a core pathway for the steel industry to break through efficiency bottlenecks and achieve high-quality development. Meanwhile, large language models, as a rapidly advancing artificial intelligence technology in recent years, have demonstrated remarkable results across various sectors. Therefore, how to leverage general-purpose large models to accelerate the intelligent transformation of the steel industry has emerged as a significant research and application direction. This paper focuses on exploring three mainstream methods for constructing large models in the steel domain: Domain-continual pretraining, which enhances the model’s understanding of industry-specific knowledge through continuous pretraining on steel-related corpora; supervised fine-tuning on scenario-specific tasks, which improves model performance on domain-specific tasks using labeled data from real-world industrial applications; and knowledge distillation, which leverages a powerful teacher model to extract knowledge from steel industry corpora and transfer it to a smaller student model during training, thereby enhancing the model’s retention of domain-specific knowledge. Finally, the advantages and disadvantages of these three methods were evaluated using a constructed steel knowledge assessment dataset. Through systematic analysis and comprehensive comparison, this research provides practical guidance and methodological references for building specialized large language models in the steel industry and other vertical domains, thereby facilitating industrial intelligent upgrading.

The intelligent upgrading of metallurgical industry standard interpretation faces multi-dimensional technical challenges, primarily characterized by the multi-version iterative nature of standard texts and strong semantic coupling between provisions. To address the limitations of retrieval-augmented generation (RAG) technology in processing standards, specifically knowledge fragmentation and weakened semantic relevance, this study proposes a hierarchical semantic recursive RAG. Firstly, a multi-level semantic parsing model based on document structural features was established, which achieving hierarchical decomposition and structured semantic representation of standard texts. Secondly, a recursive semantic aggregation algorithm was designed to enhance contextual semantic coherence through bottom-up multi-granularity summarization technology. Finally, a dynamic correlation discovery module was developed, leveraging large language models’ named entity recognition and semantic reasoning capabilities for deep exploration of cross-level constraint relationships. Experimental evaluations demonstrate that compared with baseline RAG methods, HSR-RAG achieves a 14.46% improvement in average F1-score on BERTScore metrics for QA results in professional test sets. In multi-hop QA scenarios, the framework reaches an average ROUGE-L score of 43.69%. This research provides a novel technological pathway for intelligent interpretation of metallurgical industry standards, offering significant practical value for promoting digital transformation in the sector.

Implementing a full-process material tracking system for high-speed wire rod (HSWR) production can significantly enhance an enterprise’s control over the production process and product quality, facilitating its transition towards specialty steel manufacturing. Current material tracking in continuous casting, reheating, rolling, and finishing areas faces challenges including reliance on manual input, difficulty in achieving individual coil and per-billet tracking throughout the entire process, and insufficient accuracy and stability. The less than 100% accuracy of machine vision technologies used for identifying steel heat numbers and C-hook numbers contributes to the instability of HSWR material tracking in industrial practice. This paper proposes methods to improve environmental perception accuracy, such as constructing state machines, utilizing multi-system data comparison, and adding verification bits, thereby enhancing the stability of material tracking. For instance, the verification bit method applied to C-hook identification improved system stability by a factor of over 600. By establishing and linking the correspondences between “steel heat number→cast billet number→wire rod material number→C-hook number”, the system moves beyond single-process tracking to achieve a stable and reliable full-process material tracking solution. Practical applications confirm the effectiveness of the proposed methods and system, which effectively support quality traceability and cross-process coordination.

In the sintering process, accurate prediction of the sintering burn-through point is essential to control the operation of the sintering machine, as it determines the quality of the sintered material and the efficiency of energy use. However, the current common sintering burn-through point prediction models mainly focus on single-step prediction, and they tend to ignore the spatio-temporal characteristics of the sintering process data. In view of the multi-step prediction task requirements of the sintering burn-through point and the complex spatio-temporal characteristics of process data, a multi-step prediction model of the sintering burn-through point based on encoder-decoder architecture was proposed. The spatio-temporal encoder-task decoder architecture effectively extracts the spatio-temporal features of sintering process data by stacking temporal convolutional networks and spatial attention mechanism, and adopts a task-specific guidance decoder to correspond to each step of prediction in the form of independent units. The effectiveness of the proposed method was verified by an actual sintering industrial process case. This study can assist operators to obtain information about the future state of the sintering process in advance, so as to adjust the operating parameters more accurately, reduce the quality fluctuation and energy waste caused by the lag, and improve the quality stability of the sinter.