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 .
Although scrap bundles have many advantages , the diversity of scrap types has an impact on smelting . considering the complexity of the steel plant environment and the complexity of the current scrap type recognition technology , the use of mobile devices to realiZe the accurate recognition of scrap bundles in complex scenes is of vital significance to improve the accuracy and productivity of smelting models . The dataset was enriched by adding new pictures of scrap bundle under complex lighting scenes in the original dataset , and the improved hybrid network model was applied to study the scrap bundle recognition algorithm. The results of the study show that the improved Edge Next hybrid model has a better performance in the recognition scenarios . on the experimental dataset , its test accuracy is improved by 2. 81% compared to Mobilenetv3 ; one round of training time consumed is reduced by 16 seconds compared to the viT model ; and the model shows better convergence speed and oscillation amplitude during the training process . In summary , the improved Edge Next model provides solid theo-retical support for improving the intelligent recognition of scrap bundles .
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 .
The semi-quantitative detection of metal coating thickness on scrap steel surfaces using laser- induced breakdown spectroscopy (LIBs) was investigated , based on the characteristic regular changes in spectral intensity and ablation crater morphology during the laser ablation process . Firstly , by stud- ying the morphology of laser ablation craters , a mathematical model was established to correlate the depth of the ablation crater with the number of laser pulses . subsequently , by combining the judgment of the critical point of laser penetration through the coating , the thickness of the metal coating on the scrap steel surface is quantified. Anovel method for the rapid assessment of coating thickness on scrap steel surfaces was provided.
Real-time identification of the thickness of the scrap steel material line is of great signifi- cance for its transportation and the electric arc furnace steelmaking process . Currently , there is no al- gorithm research based on deep learning methods for this scenario. Aiming at the above problems , the current mainstream segmentation model networks were compared and PP-Liteseg for the segmentation of the edge contours of scrap steel was used. Furthermore , optimiZation strate gies such as Dice loss and LovasZsoftmax loss were introduced to improve the original cross-entropy loss function of PP-Lite- seg , achieving the best balance between global supervision and local optimiZation of local details . The mean intersection over union (mIoU) of segmentation reaches 81 . 11% . Finally , based on the opti- miZed model , a calculation method for the height of the scrap steel material line was designed , which extracts key reference lines by using the segmentation results , enabling real-time monitoring and quan- titative evaluation of the maximum and average stacking heights of scrap steel. The experimental re- sults show that this method has the ability of high-precision segmentation under complex working condi- tions and excellent generaliZation performance for different positions , providing reliable technical sup-port for the accurate identification of the thickness of the scrap steel material line in the electric arc furnace steelmaking process .
The laser-induced breakdown spectroscopy ( LIBS) technology to analyZe the evolution of spectral characteristics of scrap steel with and without coatings as a function of laser pulse count was utiliZed. The research reveals that for coated scrap steel , the spectral line intensity of coating elements initially increases and then decreases with the increase in pulse count , while the spectral line intensity of Fe element gradually increases . For uncoated scrap steel , the spectral line intensity of Fe element rapidly increases after the initial pulses and then stabiliZes . Based on these patterns , using the stand- ard deviation threshold of the normaliZed intensity of Fe element ( set at 0 . 02) to identify the presence of coatings on scrap steel surfaces was proposed. Furthermore , by analyZing the changes in the normal- iZed intensity of coating elements , a method based on the cumulative value of normaliZed spectral line intensities to determine the type of coating element was introduced , where the element corresponding to the maximum cumulative value was identified as the coating element . An effective technical means for the rapid identification and classification of coatings on scrap steel surfaces was provided.
Intelligent scrap steel grading can optimiZe the scrap steel grading process of steel enterpri- ses and scrap steel bases in the process of purchasing scrap steel , improve the efficiency of scrap steel grading , and reduce the emotional quality of scrap steel grading . For the intelligent scrap steel grad- ing , based on the original target detection algorithm , a method based on the target tracking algorithm was proposed to optimiZe the automatic termination of the grading function in the process of intelligent scrap steel grading. The target detection algorithm was used to first detect and identify the grading vehicles appea- ring in the gun-type camera ( gun camera) screen. There may be incomplete vehicles , multiple grad- ing vehicles , etc . , especially in these special cases , it is necessary to design a target tracking algo- rithm for vehicle identification to improve the accuracy of identifying scrap steel vehicles to be graded.The target tracking algorithm to realiZe the automatic termination function of the scrap steel intelligent grading system was adopted. compared with the original single target detection algorithm , the kalman filter and the Hungarian algorithm were used to solve the state prediction and trajectory matching asso- ciation of the targets detected in the image , so as to improve the robustness of the target tracking algo- rithm in complex scenes . The accuracy of the automatic resolution function of the scrap steel intelligent grading system has been increased from 86% to 93% . The multi-object tracking by associating every detection box (ByteTrack) algorithm explored in this article has improved the multiple object tracking accuracy(MOTA) evaluation index by 17. s% and 1s. 3% respectively compared with the simple on- line and realtime tracking(SORT) and deep learning for simple online and realtime tracking (Deep- SORT) algorithms , and the IDF1 evaluation index has improved by 16. 1% and 9. 9% respectively . The automatic termination function based on the target tracking algorithm can improve its accuracy , especially in complex scenarios , and more accurately determine whether the grading vehicle is moving , so as to trigger whether the scrap steel grading of the current vehicle needs to be terminated.
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 .
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 .
The process of steelmaking and continuous casting ( Scc) was accompanied by complex chemical reactions and physical changes , and has high uncertainty . The production scheduling of the Scc process has always been one of the difficult problems in iron and steel enterprises . In theory , this kind of problem can be reduced to the dynamic flexible flow shop scheduling problem , which is one of the hot issues in the academic research. Genetic programming (GP) has been successfully applied to the automatic design of heuristic rules for dynamic scheduling problems . However , in traditional GP, the selection of the parent solution through the tournament mode will slow down the convergence speed of the algorithm. A two-stage semantic selection mechanism was proposed to accelerate GPconver-gence by providing more convergent parent solutions for GP. In addition , in order to retain the valid information in the parent individual more efficiently , a correlation based variation strate gy was de- signed , which replaces the negative correlation subtrees in the individual tree with randomly generated subtrees . The performance of the proposed algorithm was compared and analyZed in 8 scenarios with different configurations , and the results show that the proposed algorithm has a good competitiveness compared with the existing algorithms .
The mold is a critical component of continuous casting , playing a vital role in maintaining the safety and efficiency of the casting process . Due to the complex heat transfer mechanisms within the mold , direct investigation of its thermal behavior is quite challenging . Therefore ,the interior-point method to solve the inverse heat flux distribution problem along the broadside of the mold and performs longitudinal fitting of the results was employed. The results indicate that the relative root mean square error between the calculated and measured temperatures is only 0. 94% , demonstrating the accuracy and effectiveness of the proposed inverse problem model. The solution to the inverse problem reveals a consistent relationship between the non-uniform heat flux distribution at the mold-casting interface and the uneven temperature distribution of the copper plate. Further longitudinal fitting of the results shows that the heat flux increases , and its fluctuations along the broadside become more pronounced as the distance to the meniscus decreases . Conversely , both the heat flux and its fluctuations decrease as the distance from the meniscus increases .
In the blast furnace ironmaking process , the silicon content in hot metal is a crucial indica- tor for evaluating production stability and iron quality . However , the prediction accuracy of silicon content in hot metal is affected , due to the complex mechanism of blast furnace ironmaking , multi - field and multi-phase coupling , data loss and noise. with the development of industrial intelligence ,industrial foundation models have shown application potential. Based on the concept of industrial foun- dation models , a silicon content prediction method that integrates probabilistic distributions and graph convolutional networks was proposed. The proposed method consists of a multi-channel univariate fea- ture extractor , a graph neural network , and a predictor. The multi-channel univariate feature extractor employs a variational autoencoder to extract probabilistic representations of variables . The graph neural network captures variable interactions and compensates for missing data. Real-world data collected from a blast furnace digital twin system was utiliZed for validation. Comparative experiments demon- strate that the proposed method improves the prediction accuracy of silicon content by 4. 31% com- pared to existing approaches . The novel practical reference for the application of industrial foundation models in blast furnace ironmaking quality prediction was provided.
In the hot rolling production process , the imbalance in industrial data distribution increases the difficulty of diagnosing strip crown abnormalities , which severely affects product quality . To ad- dress this issue , a method combining the Extra Trees algorithm with the resampling technique SMOTE- Tomek link was proposed. The approach effectively solves the class imbalance problem in multi-class data. On this basis , an improved chaotic genetic algorithm (ICGA) was used to optimiZe the model/s hyperparameters and determine the optimal hyperparameter combination. Model interpretability analy- sis was conducted using the local interpretable model-agnostic explanations ( LIME) method , which further reveals the key process parameters affecting the strip crown and their contributions . Experimen-tal results show that the proposed method achieves accuracy , recall , precision , and F1 Score of0.9907,0.9908,0.9910 , and 0. 9908 , respectively , significantly outperforming traditional mod- els , which verifies the effectiveness and advantages of the method in strip crown diagnosis and provides support for fault diagnosis in the hot rolling production process .
In the process of hot strip rolling , the temperature rise curve of billet in reheating furnace has a significant impact on the quality of products . Due to the bad conditions in the furnace , the actual temperature of the billet is difficult to be directly monitored , so it is necessary to establish a real-time temperature prediction model for the billet heating process . The traditional mechanism model based on partial differential equation (PDE) is usually difficult to meet the needs of real-time prediction due to its high computational complexity ; The neural network model has the characteristics of low accuracy and relying on a large number of training labels based on the actual temperature of billet , so it has not been well applied in practice. Aprediction model of hot rolling billet temperature based on generaliZed physics-informed (GPINN) was proposed. Firstly , the position of the billet in the reheating furnace was tracked according to the action signal of the walking beam , and the temperature of the whole fur- nace was predicted by cubic spline interpolation ; secondly , the PDE describing billet heating was solved by PINN, and combined with the“ branch net-trunk net ”structure of deep operator network (DeepoNe t ) , the different initial temperatures and specific time and space positions of billet were co- ded respectively , which effectively realiZes the real-time prediction of billet temperature in the heating process with different initial temperatures ; Finally , the effectiveness of the method was verified by the case study of a real steel plant . Compared with the traditional neural network method and mechanism - based method , GPINN integrates the advantages of physical information and neural network , better captures the heat conduction characteristics in the process of billet heating and improves the interpret- ability and prediction accuracy of the model.
