The aluminum industry is at a critical juncture of transitioning from experience-driven to cognitive intelligence, and it is confronted with common technical challenges such as the unpredictability of ″black boxes″ under extreme working conditions, strong coupling of multiple physical fields, and time-varying raw materials. This article reviews the technological iteration and development paths of artificial intelligence technology in the three core processes of alumina production, electrolytic aluminum smelting and casting processing. Studies indicate that in the alumina process, artificial intelligence-related technologies have made a leap from optimization regression to mechanism-constrained soft measurement and edge-cloud collaborative architecture, effectively alleviating the physical inconsistency and control lag problems of pure data models. In the electrolytic aluminum process, in response to challenges such as uneven spatial distribution of large pre-baked cells and cell control black boxes, the diagnostic paradigm has evolved from single-time series analysis to high-dimensional spatio-temporal topology perception. The related technologies have achieved transparent mapping of unmeasurable states such as the shape inside the furnace. In the aluminum casting processing stage, the machine vision based on the YOLO algorithm and the physically guided AI model have effectively improved the efficiency of surface quality inspection and process optimization. Finally, this paper looks forward to the new era of cognitive intelligence centered on industrial large models, and analyzes deep-seated challenges such as ″semantic islands″, ″physical consistency″ of models, and the crisis of interpretability, providing theoretical references and technical guidance for building a new green, low-carbon, and intelligent aluminum industry ecosystem.

For a long time, the harsh operating conditions of industrial aluminum reduction cells—such as high temperature, strong magnetic fields, and severe corrosion—have limited real-time monitoring to cell voltage alone, which has become a bottleneck for the digital and intelligent transformation of the electrolysis process. This paper describes the principles and applications of traditional methods for measuring anode current based on voltage drop and magnetic field intensity. It focuses on the schemes and application progress achieved by the project team in recent years in measuring anode current, especially zonal current, based on the magneto-optic effect. The advantages of zonal current in predicting and pre-controlling abnormal events such as anode effects, voltage fluctuations, and voltage-drops, are demonstrated. A control scheme for zonal feeding to achieve uniform alumina concentration in large-scale aluminum reduction cells is proposed. The unique advantages and development potential of optical fiber current sensing technology in the digitalization of aluminum reduction cells are presented.

Nonferrous smelting is a strategic raw-material industry that provides fundamental support for national economic development and defense-related industries. At present, the nonferrous smelting industry is confronted with increasingly complex raw materials, increasingly stringent environmental requirements, and an urgent need to reduce costs and improve efficiency. Therefore, it is imperative to promote the development of smelting plants toward high-end, intelligent, and green manufacturing. Focusing on the practical needs and challenges of intelligent factory construction in the nonferrous smelting industry, this paper designs a system and functional architecture based on an industrial Internet platform, proposes the key technologies for intelligent workshops featuring perception-decision-control-execution linkage as well as the key technologies for intelligent factories characterized by factor interconnection, centralized control, and business collaboration, clarifies an implementation pathway of progressing from points to lines, from lines to planes, and from planes to an integrated whole, and further puts forward a priority analysis strategy for implementation content by considering business dependency, technological maturity, economic feasibility, and urgency of demand. Finally, the paper discusses the expected outcomes and future development of intelligent factory construction in nonferrous smelting, with the aim of providing a reference for the planning and development of intelligent factories in this industry.

Flotation is a crucial step in non-ferrous metal beneficiation, and fluctuations in concentrate grade can significantly impact subsequent smelting energy consumption and production organization. Accurately predicting the copper grade of rougher concentrate is challenging due to the multivariate, strongly coupled, time-lag, and significant cumulative effects of flotation data, as well as its strong instantaneous fluctuations and noise errors. To address this issue, this paper uses the TsMixer time-series prediction model with a fully multilayer perceptron architecture as a baseline and introduces the low-rank adaptation concept to structurally modify the time mixing mechanism. Four TsMixer time mixing variants based on low-rank adaptation (TsMixer-LoRA1～LoRA3, TsMixer-LR) are designed, and the impact of the training/freezing strategy of the base weights and the low-rank order on model performance is investigated. Experiments are conducted using real flotation data, incorporating multivariate inputs including flotation process variables and foam visual features. Historical 60 min data is used to predict the grade for the next 15 min, and comparisons are made with models such as LSTM, GRU, TimesNet, FEDformer, PatchTST, and iTransformer. The results show that applying a low-rank constraint to the temporal mixing layer effectively improves the model′s generalization performance; TsMixer-LR achieves the best overall performance and exhibits a clear optimal rank interval. Further parameter scale analysis reveals that TsMixer-LR not only achieves the best prediction results at medium to low rank orders, but its total parameter count remains lower than or close to the benchmark TsMixer, demonstrating a better balance between parameter scale and prediction accuracy. Mechanistic analysis indicates that the low-rank bottleneck prompts the model to preferentially learn the dominant temporal structure, suppressing overfitting to instantaneous fluctuations and noise perturbations, thereby enhancing prediction stability and robustness.

In the gas suspension roasting process of alumina, the main furnace temperature is the core variable reflecting the operating conditions. Accurate multi-step ahead prediction of this variable can provide a prerequisite for the implementation of fault warning and model predictive control, thereby stabilizing the phase transformation rate of products and reducing production risks. However, the inherent characteristics of the roasting process, including strong multivariable coupling and nonlinear dynamic evolution, increase the difficulty of prediction. Existing data-driven models are prone to autoregressive error accumulation and often lose high-frequency dynamic details, resulting in overly smooth prediction curves that fail to provide effective guidance for actual production. To address the above issues, this paper proposes a data-driven multi-step prediction method based on the Time-Frequency dual-domain Large Foundation Model (TF-LFM). With feature decoupling and time-frequency dual-domain optimization as the core, this method uses the Seasonal-Trend Decomposition Procedure Based on Loess (STL) to decouple complex process time-series signals into trend components and periodic components. Subsequently, the model leverages a large language model to extract long-range trends for temporal reasoning, and introduces a large speech model to reconstruct the frequency-domain features of high-frequency periodic fluctuations, so as to compensate for the defect of dynamic detail loss in traditional models. Under the constraint of the time-frequency joint loss function, the dynamic synergy and restoration of the dual-domain prediction results are completed. The verification results based on actual production data from a large aluminum plant show that the proposed method overcomes the problems of error accumulation and dynamic detail loss. Compared with the 3 time-series models for comparison, the prediction error of the proposed method in 32-step long horizon prediction is reduced by about 60.5% compared with the optimal comparison model, which provides reliable model support for fault warning and refined control of the roasting process.

To improve the accuracy of refractory lining life prediction in copper smelting rotary anode furnaces, this paper proposes a prediction method integrating physical mechanisms and deep learning. Aiming at the problem of sparse and unevenly distributed brick thickness measurement data in on-site high-temperature and high-dust environments, the model establishes a nonlinear exponential relationship between temperature and erosion rate based on the analysis of metallurgical reaction kinetics, constructs a data augmentation model under physical constraints, and introduces Monte Carlo noise simulation to expand discrete and sparse measurement records into a high-frequency time-series dataset. On this basis, a multilayer perceptron (MLP) model is constructed to realize the online evolution prediction of refractory brick thickness. Experimental results show that this method effectively solves the modeling difficulties under small-sample conditions.The model achieved a coefficient of determination (R²) greater than 0.99 and a root mean square error (RMSE) of approximately 0.1 mm on the test set, verifying the effectiveness of the mechanism and data dual-driven mode in the service life prediction of industrial thermal equipment.

Aiming at the core pain points in the aluminum plate, strip and foil processing industry such as low timeliness of data collection, prominent information silos, low collaborative efficiency and absence of real-time analysis and online process optimization capacity between production and quality, the 2 ms-level high-frequency data collection technology and the parsing method for multi-source heterogeneous system protocols were studied. According to the nonlinear and coupling characteristics of the aluminum plate, strip and foil production process, a product quality judgment model based on random forest was proposed, and an equipment health diagnosis model based on long short-term memory network was established for the deep abstract fault features of equipment health status, and a production and quality analysis platform suitable for aluminum plate, strip and foil enterprises was developed. The application verification of the platform has achieved the following results: the collection delay of key process data is reduced from an average of 2 h to less than 50 ms, the time required for real-time data analysis has been compressed from over 30 min to within 2 s, and the response time for online optimization of process parameters has been shortened from more than 2 h to within 10 s. The time consumed for manual data processing was cut by 84%, and the one-time qualification rate of finished products was increased from 92.3% to 96.5%. The results show that the platform effectively improves the management level of production and quality, and provides a replicable technical solution and practical demonstration for the efficient data collection, development and utilization of aluminum plate, strip and foil enterprises.

Predicting the properties of steel materials is crucial for optimizing production processes and enhancing product quality. In recent years, Machine Learning (ML) has demonstrated significant advantages in this field due to its powerful capabilities in data mining and pattern recognition. This paper provides a systematic review of the current applications and challenges of machine learning in predicting steel material properties, with a focus on analyzing research progress and industrial application cases of algorithms such as Backpropagation Neural Networks (BPNN), Support Vector Machines (SVM), Random Forests (RF), and deep learning. Key challenges including data quality, model generalizability, and interpretability are discussed. Furthermore, future innovative directions are outlined, such as the application of Large Language Models (LLM), multi-modal data fusion, and mechanism model deep integration, aiming to offer guidance and technical references for the intelligent transformation of the steel industry.

Few-shot classification of surface defects in cold-rolled steel strips is a significant challenge in industrial quality inspection, with the core difficulties being sample scarcity and class imbalance. To systematically investigate the efficacy of different solution strategies, this paper constructs a large-scale dataset comprising 53 categories and 14 499 samples, and systematically compares the impact of traditional transformation-based augmentation and generative augmentation methods－such as Projected GAN and Diffusion－on the performance of classification models. Our findings reveal that although both Projected GAN and Diffusion are capable of generating high-quality images, with Diffusion demonstrating strong cross-domain migration capability, their augmentation effects are limited. Generative methods only slightly improve accuracy at the cost of a decline in recall and F1-score, while traditional image transformation-based augmentation even leads to negative effects. Surprisingly, a simple balanced sampling strategy achieves the best performance. This study confirms that directly addressing class imbalance is crucial for improving few-shot defect classification performance. Furthermore, it provides important empirical evidence for selecting technical solutions in industrial inspection scenarios: compared to complex generative augmentation, lightweight balanced sampling may represent a more efficient and practical choice.

To address the low efficiency in tracing defects of traditional steel products and the poor adaptability of existing knowledge extraction methods, this paper proposes a knowledge extraction approach for multi-source steel product quality defect tracing texts. First, a multi-source integrated text dataset for steel product quality defect tracing was constructed; second, a knowledge extraction model for steel product quality defect tracing based on a pipeline architecture was designed to accurately extract triple-structured knowledge from domain-specific texts; finally, named entity recognition experiments and knowledge extraction model experiments were carried out to systematically validate the effectiveness of the model, and the extracted structured knowledge was presented in the form of a knowledge graph. The results indicate that the proposed knowledge extraction method can provide technical support for accurate and efficient defect tracing of steel products.

To address the requirement for raw material and finished product transport trains in the steel metallurgy industry to precisely align within the designated weighing zone during static weighbridge operations, this paper designs a visual positioning system based on the multimodal large model DeepSeek-VL2-tiny. The system captures images via high-definition industrial cameras deployed at the static weighbridge edges. It employs the DeepSeek-VL2-tiny model, fine-tuned using low-rank adaptive adjustment, to detect critical components in real time, including wheels, couplers, and weighbridge edge markers. By establishing a mapping relationship from pixel coordinates to physical space, it calculates the actual offset distance of the railcar relative to the standard weighing zone, thereby enabling intelligent guidance for parking positions. Experimental results in real industrial settings demonstrate the system's stable recognition and localization of critical components, validating the reliability of multimodal large models in practical industrial applications.

With the rapid development of the iron and steel industry, there is an increasingly urgent demand for automation and intellectualization in slag skimming operations. Aiming at the current situation where slag skimming operations mainly rely on manual work, which suffers from low efficiency, poor accuracy and high safety risks, this paper proposes an automatic molten iron slag skimming system based on machine vision and path planning algorithm. On the basis of fully investigating the on-site working logic of molten iron desulfurization and slag skimming, the system integrates deep learning convolutional neural network and intelligent visual perception methods. It can accurately segment the elliptical area of the molten iron ladle mouth, and complete the distinction between molten iron and light/dark slag through image preprocessing, gray feature analysis and Otsu multi-threshold segmentation, thereby realizing the functions of intelligent determination of slag amount on the molten iron surface and automatic path planning of the slag skimming arm. This paper elaborates on the design principle, implementation method of the system and its effect in practical application. Through the operation verification in the intelligent control center of a steel plant, the recognition error of molten iron area is no greater than 5%, and the determination qualification rate based on 500 ladles of molten iron is no less than 96%. These results significantly demonstrate its advantages in improving the efficiency and accuracy of slag skimming, providing new ideas and practical cases for the intelligent transformation of the iron and steel industry.

To improve the slab yield rate in the production of medium and heavy plates in iron and steel enterprises, and address issues such as excessive unplanned material feeding and over-reliance on the experience of planners, this paper, based on the actual production condition of medium and heavy plates at Xiangtan Iron and Steel Co., Ltd., analyzes the whole-process mathematical model for the slab design of medium and heavy plates, and studies and constructs a set of medium and heavy plate slab design methods and systems oriented to yield rate control and optimization. Based on the idea of hierarchical modularization, this method establishes a standardized slab design rule model, introduces additional large-slab design rules to supplement the large-slab design and arrangement system, optimizes the selection logic of rolling directions, the margin reservation strategy for slab feeding, and the dimension correction method, and constructs an intelligent slab design algorithm. This algorithm, which incorporates mutually independent slab design constraints, is developed on the basis of mixed integer programming (MIP) and constraint programming (CP) theories, and yields results oriented toward multi-objective optimization. This system has been put into operation, and operational results indicate that it has enhanced the flexible adaptability between the enterprise′s slab design methods and business operations, reduced the generation of residual materials, improved the slab yield, while the yield control strategy has increased the plan fulfillment rate of slab production, and the intelligent slab design algorithm has reduced employees′ workload and improved the quality of production plans, thereby demonstrating favorable application value and promotion potential overall.

In order to reduce the adverse impacts of abnormal conditions that occur during the converter smelting process on the stability of steel production, efficiency, and cost control. This paper, through a deep analysis of the changing relationship of the volume fraction curves of CO, CO₂ and O₂ in the flue gas during actual production, has developed and integrated a prediction model for abnormal states in the smelting process by combining Python language and PLC programming. The model mainly encompasses the following components: a prediction model for failure to ignite at the start of blowing, a prediction model for the molten pool temperature in the early stage of smelting, a prediction model for re-drying and splashing during the blowing process, and a prediction model for the double-slag operation during the smelting process. After the implementation of the model, it provides strong guidance for stabilizing the process operation. It effectively eliminates abnormal interruptions during the blowing process. Specifically, the proportion of re-drying in the process is reduced by 10%, the splashing rate is controlled to less than 0.5%, the amount of iron entrained in the slag during the double-slag operation is decreased by 50%, and the dephosphorization rate in the early stage is stabilized at approximately 80%. The stabilization of the smelting process has significantly enhanced technical and economic indicators. The converter smelting cycle is maintained at around 28 min. The qualified rate of the final carbon content for low-alloy steel has increased to around 88%. The proportion of first-time carbon pulling and steel tapping reaches over 95%. The mass fraction of oxygen at the end of the low-alloy steel smelting is stably maintained at 370×10^-6-380×10^-6, and the consumption of steel and iron materials in the converter is stabilized within the range of 1 030-1 033 kg/t.