With the rapid development of digital transformation in China′s steel industry, steel enterprises have accumulated massive knowledge and data assets. How to efficiently unearth the value of knowledge and data assets and gradually transform from digitalization to intelligentization has become a challenging problem. Large model has entered the stage of large-scale application, and the industry large-scale model is the key to its deep penetration into vertical fields. As a typical process industry, the steel industry has abundant scene resources and data assets, and urgently needs to be empowered by industry big models to create a new business model driven by the integration of knowledge, data, and intelligence, in order to achieve intelligent upgrading and high-quality development. This article first proposes the architecture design concept of the steel industry′s large model, and studies the data architecture, platform architecture, and application architecture; Then, the application of knowledge engines, intelligent agents for deep knowledge insight reports, metallographic detection models, and embodied intelligence models were introduced, exploring the application modes of natural language models, visual models, and multimodal models; Finally, the future development of the steel big model was discussed from the aspects of industry data space, collaborative system of large and small models, and application security protection.

Slab number is a critical identifier in steel manufacturing for process tracking and intelligent logistics. However, poor print quality and harsh industrial environments often damage image data, seriously reducing the accuracy of slab number recognition. To improve the recognition capability of damaged images, a novel algorithm—Feature Recognition Inference Network (FRI-Net)—which combines damaged region detection, contextual feature reasoning, and attention mechanism to achieve high-quality restoration and accurate recognition of degraded slab number images. FRI-Net adopts a modular architecture, introduces a feature feedback optimization mechanism and Knowledge Consistent Attention (KCA), and significantly enhances the restoration capability for complex defective regions. Experimental results on multiple public and industrial datasets demonstrate that FRI-Net outperforms existing mainstream methods in recognition accuracy and fault tolerance, effectively enhancing the stability and intelligent level of slab tracking systems.

To address the issue of low thickness control accuracy during non-steady-state rolling processes, such as specification switching and roll change in hot continuous rolling, this paper proposes a data-driven model based on Stacking ensemble learning for thickness prediction during specification switching. The optimal hyperparameter configurations of both the base learners and the meta-learner are determined using a Bayesian optimization algorithm. The performance of the Stacking model is systematically compared with that of other mainstream models. Additionally, the SHAP (SHapley Additive exPlanations) method is introduced to interpret the Stacking model and evaluate the importance of input features. The results demonstrate that the proposed Stacking model effectively integrates the predictions of base learners, significantly enhancing the accuracy of thickness prediction. The performance evaluation metrics of the developed Stacking model are as follows: root mean squared error (RMSE) of 0.019 9, mean absolute error (MAE) of 0.014 9, and coefficient of determination (R2) of 0.999 9. The probability that the error between the predicted and actual thickness is within ±35 μm reaches 94.8%, while the probability within ±50 μm is 97.8%, achieving high-precision thickness control during the specification switching process in hot continuous rolling.

The rolling mill is a key piece of equipment in steel production, and the health of its rotor directly affects equipment safety and production efficiency. However, in practice, fault data is scarce and class imbalance exists, limiting the performance of traditional diagnostic models. To address this, this paper proposes a data augmentation method based on SN-WGAN-MMD to improve the effectiveness of rotor fault diagnosis. The method combines Wasserstein loss with Maximum Mean Discrepancy (MMD) loss to generate realistic samples from both global distribution and high-dimensional feature perspectives, while spectral normalization (SN) is used to enhance training stability. The generated samples are added to the original dataset to alleviate class imbalance. Experiments conducted on two datasets show that the proposed method outperforms GAN, WGAN, and WGAN-GP in terms of sample similarity, class balance, and diagnostic accuracy. The highest accuracy reaches 97.3%, representing a 15% improvement over the original dataset, demonstrating its advantages in global alignment, local feature representation. This provides an efficient and novel data augmentation solution for rotor fault detection.

To address the challenges of infrequent detection and prolonged measurement cycles for FeO content in iron ore sinter, this study proposes a soft-sensor model for online estimation of FeO content using routinely monitored parameters from the sintering process as input variables. The model development first employs the Random Forest (RF) algorithm to identify high-impact variables, subsequently constructs a Multilayer Perceptron (MLP) to capture the intricate nonlinear relationships between input variables and FeO content, and finally integrates the Wild Horse Optimizer (WHO) algorithm to optimize MLP hyperparameters, thereby enhancing both model fitting accuracy and generalization capabilities. Validation using real-world production data demonstrates that the proposed model achieves an accuracy of 91.26% within ±0.5% error margin. This framework provides actionable insights for industrial applications, effectively mitigating FeO content fluctuations and advancing operational precision in sintering production, with significant implications for optimizing process stability and elevating metallurgical manufacturing standards.

Longitudinal cracks are common and serious quality defects in slab production. Accurate prediction of longitudinal cracks is of great significance to improving slab quality and production efficiency. However, in actual production, the number of longitudinal fissure samples is far less than that of normal samples, resulting in extremely unbalanced data distribution, which affects the effect of model training and brings huge challenges to model construction. How to improve the generalization ability and prediction accuracy of the model has become a key problem. To this end, firstly based on a large amount of measured temperature data of thermocouples on the copper plate of the mold, the sliding window technology was applied to extract temperature samples of longitudinal cracks′ and other conditions. Then, a slab surface depression-type longitudinal crack prediction method based on grid search optimized Convolutional Neural Network (CNN) and Bidirectional long Short-Term Memory (BiLSTM) network was proposed. The samples were input into the CNN-BiLSTM network, CNN was used to obtain the local features of the time series, and BiLSTM was used to obtain the long-term dependency features. Finally, the slab longitudinal crack prediction output was performed through the fully connected layer. Experimental results show that the proposed grid search optimized CNN-BiLSTM model performs significantly better than other models on the test set, with a prediction hit rate of 99.29% for longitudinal crack temperature waveforms, a false alarm rate of 0.71%, and a Matthews Correlation Coefficient (MCC) as high as 0.96, and the training and prediction speeds of the model are relatively fast. The research results provide a reliable theoretical basis for the identification of longitudinal cracks on the slab surface.

To address the issue of shape prediction in the roller quenching process, this paper proposes a novel method based on clustering analysis and an improved Bidirectional Gated Recurrent Unit (BiGRU) network. Key process parameters influencing plate shape during quenching are first analyzed, and K-means clustering is applied to categorize sample data, thereby identifying the distribution characteristics of different steel grades and specifications. For continuously produced steel plate batches, a BiGRU model is employed to extract high-dimensional temporal features, capturing the sequential dependencies between preceding and succeeding shape variations. By integrating Convolutional Neural Networks with bidirectional gated units, an enhanced BiGRU model is constructed for plate shape prediction. Experimental results demonstrate that the K-means-guided improved BiGRU model achieves accurate shape predictions within acceptable process error margins. This method provides a reliable foundation for plate shape control in steel quenching processes and contributes to intelligent control in advanced manufacturing.

In hot strip rolling mills, loopers are installed between finishing stands to ensure the balance of metal mass flow and the stability of strip tension during the rolling process. Therefore, the control accuracy and dynamic response characteristics of the looper system directly determine the quality of the final strip product. To address the control challenges caused by the multivariable strong coupling, nonlinear and time-varying characteristics of the hydraulic looper system, this paper designs a composite intelligent control strategy that combines an Improved Particle Swarm Optimization (IPSO) algorithm with Back Propagation Neural Networks (BPNN) Proportional-Integral-Derivative (PID). A feedforward compensation decoupling control loop is constructed to weaken the coupling effects in the system. An improved PSO algorithm with dynamic adjustment mechanisms for inertia weights and learning factors is developed to solve problems in traditional BP neural network PID, such as strong randomness in weight initialization, tendency to fall into local optima, and slow convergence speed. Simulation experiments based on the MATLAB/Simulink platform demonstrate that the proposed IPSO-BP-PID algorithm achieves better control performance compared to both BP neural network PID and traditional PID under unit step input signals, verifying the effectiveness of the IPSO-BP-PID algorithm.

This paper focuses on the innovative applications and development trends of digital twin technologies in the field of metallurgical engineering. As core components of Industry 4.0, digital twin demonstrates enormous application potential and broad prospects in the steel industry. Firstly, this paper reviews the definition, background, and core development technologies of digital twin technology, and expounds its significance in the metallurgical field. Secondly, based on the production process of the steel manufacturing system, it systematically combs the practical paths and application scenarios of digital twin technology in the metallurgical field. Finally, from a multi-dimensional perspective, it analyzes the technical bottlenecks in the industrial implementation process of steel digital twin, and based on the characteristics of the industry and technological development, prospects its future development direction. It is worth noticing that this article tracks the latest progress of cutting-edge technologies such as generative artificial intelligence algorithms, large language models, and intelligent agents, and analyzes the integration potential of such technologies with digital twin technology at different levels, providing a theoretical direction for the construction of the next-generation metallurgical digital twin system with intelligent cognition.

:Iron molten steel transportation serves as a critical link connecting ironmaking and steelmaking in steel enterprises, with its efficiency constrained by factors such as solution quality and computation speed of multi-locomotive path planning algorithms at the iron-steel interface. To address existing bottlenecks including low computational efficiency and susceptibility to local optima, this paper proposes a Whale-optimized Dynamic Weight Time A* algorithm (WODWT-A*) for multi-locomotive path planning. Firstly, the algorithm introduces a dynamic weight mechanism that linearly combines cost ratios with difference functions to enhance the heuristic function. This innovation significantly improves temporal efficiency and pathfinding accuracy in dynamic environments, overcoming the efficiency limitations of traditional A* algorithms in complex scenarios. Furthermore, to enhance global optimization capability, the method incorporates the Whale Optimization Algorithm (WOA). By simulating humpback whales′ group predation strategies, it achieves dynamic optimization of initial solutions and adaptive parameter adjustment, thereby strengthening multidimensional exploration in complex solution spaces and effectively avoiding local optima traps. This synergistic mechanism between global search and local optimization enables WODWT-A* to significantly improve planning stability and solution quality. Practical case studies demonstrate that WODWT-A* exhibits high adaptability and reliability in multi-task concurrency, path congestion, and dynamic environments, providing an optimized solution for coordinated scheduling of multiple locomotives at the iron-steel interface.

The slag grinding process is essential for manufacturing slag powder, which is a green raw material. The quality of its indicators directly affects the economy and safety of the entire process. However, this process involves multiple controlled variables and features strong nonlinearity, large feedback delays, and frequent fluctuations in operating conditions, making it difficult to control the indicators of quality, throughput, and temperature. To address these issues, this paper designs an intelligent control system for slag grinding based on model predictive control. The system first performs intelligent optimization of the control targets through an existing platform, and then controls the quality, throughput, and temperature indicators using the model predictive control method. Practical operating results show that the application of this system reduces the unit gas consumption and unit power consumption from 35.7 m3/t and 40.4 kWh/t to 34.1 m3/t and 38.7 kWh/t, respectively. These results indicate that the proposed method can significantly reduce energy consumption while maintaining stable product quality, thus demonstrating its practicality and effectiveness.

Aming at the problems of rolling force calculation and process optimization in the hot continuous rolling process of high-strength non-oriented silicon steel, the influence of deformation temperature and deformation rate on the hot rolling deformation process of high-strength non-oriented silicon steel was studied through thermal simulation experiments and analysis of true stress-strain curves. A corresponding deformation resistance model was constructed, and the regression coefficient of the deformation resistance curve was solved. Based on the rolling force model, the influence of different rolling speeds and deformation temperatures on the difference in rolling force changes was analyzed. Furthermore, with the goal of balancing the equipment capacity of each stand rolling mill, improving product performance quality and production line efficiency, the objective functions of balancing the remaining proportion of rolling mill equipment capacity and controlling the temperature and speed of hot rolling process were established. The optimization technology of high-strength non-oriented silicon steel rolling process was developed, and the reasonable setting of inlet temperature and rolling speed was achieved, and significant application effects were achieved.