AI solutions for the steel industry

Intelligent sintering technology based on big data

Sintering is one of the main processes in the blast furnace ironmaking process, and sintering ore blending is an important part of the blast furnace ironmaking process. Fluctuations in the sintering process and sinter quality have a great impact on the stability of the blast furnace. During the sintering process, complex physical and chemical reactions occur simultaneously under the influence of external random factors. Gas, solid and liquid multi-phase substances coexist, involving hundreds of raw materials, operations, states and finished product quality parameters. The number of parameters is large and the frequency of data is complex. It has the characteristics of multi-variable, strong coupling, non-linearity and hysteresis. To accurately describe the sintering process, various modern detection methods are applied to sintering production, which reduces the difficulty of monitoring the sintering process to a certain extent. However, the sources of sintering raw materials are complex, the production and smelting cycle is short, the reaction speed is fast, and the inside of the material layer during the sintering process is a black box that cannot be directly observed and lacks real-time and accurate monitoring information. Whether it is on-site operators or a small number of expert systems based on experience and knowledge, it is difficult to accurately grasp and predict the quality of the sintering process in real-time.

Advanced steel companies have been equipped with complete data collection and storage systems for sintering and other iron-making processes, forming an enterprise iron-making data platform. The large amount of sintering production process data accumulated in the platform contains sintering production rules and experience information, but related data has not yet been fully mined. With the development of big data intelligent technology, intelligent technologies such as machine learning and deep learning can be used to quickly and efficiently find the correlation between sintering process parameters, and deeply integrate big data intelligent technology with metallurgical theory and sintering production experience. Establishing an intelligent model of the sintering process with high accuracy and fast response speed, forming a positive interaction between the results of the intelligent model of the sintering process and actual sintering production, is expected to solve the constraint distortion and optimization difficulties of the sintering ore blending model, and the accuracy of the sintering process parameter prediction model traditional problems include low production line adaptability, difficulty in characterizing the quality of the sintering state, etc.

This article reviews the intelligent sintering technology in recent years and summarizes and analyzes the representative methods and important results of intelligent sintering from the aspects of ironmaking data platform construction, intelligent sintering ore blending, sintering process parameter prediction, and sintering process comprehensive evaluation and optimization. The future development trends of research and application of intelligent sintering technology are discussed.


Robot, coking equipment, Sintering trolley, sintering technology


1. Construction of an ironmaking data platform for iron and steel enterprises

Because intelligent technology based on big data has unique advantages in improving product quality and reducing labour intensity, steel companies such as Lianyang Iron and Steel, Shaogang, Shagang, Zhongtian Iron and Steel, Daye Special Steel, Nanjing Iron and Steel, Baosteel, and Maanshan Iron and Steel are all There are reports on the construction of an ironmaking data platform. The existing ironmaking data platform is based on the data layer-computing layer-application layer and adopts the Internet of Things-big data-intelligent manufacturing-cloud technology-mobile Internet technology to conduct large-scale collection of sintering, pelletizing, coking, and blast furnace process data. Control, realize real-time and efficient collection of ironmaking process data, massive data storage, unstructured data processing, multi-source data interconnection, automatic report generation and other functions, and conduct integrated intelligent management and control of iron area data. As production process data continues to accumulate, the data in the ironmaking data platform has become an intangible asset that supports enterprise intelligence. It can efficiently and accurately provide the required production data for subsequent intelligent sintering technology, and accelerate the on-site application and development of intelligent sintering technology. development process.

2. Intelligent sintering ore blending

The sintering ore blending process is to mix several kinds of iron ore powder, flux, fuel, returned ore, miscellaneous materials and other sintering raw materials under certain constraints to form a homogeneous material for sintering. The physical and chemical properties of the sintering mix are the key to affecting the quality of the sinter. At the same time, the sintering ore blending process is also one of the processes with the greatest cost reduction potential in blast furnace ironmaking. Therefore, sinter ore blending has always attracted the attention of steel companies. The researchers developed a mechanism-driven sintering intelligent ore blending model. When the mechanism model was difficult to accurately meet on-site needs, the mechanism-driven model and sintering process data were integrated, and data analysis was conducted on the composition, structure and performance of the sintering raw materials. , with the assistance of intelligent algorithms, a data-driven approach is used to further optimize the sintering intelligent ore blending model, accurately control the sintering raw material ratio, and improve the applicability of the model on-site.

2.1 Mechanism-driven sintering intelligent ore blending

The mechanism-driven sintering intelligent ore blending model sets the sintering ore blending constraint conditions based on metallurgical theory and sintering production experience. Taking the cost of sinter ore, the cost of molten iron or the energy consumption of the sintering process as the goal, it solves and optimizes the sintering ore blending plan through mathematical methods. The earliest researched and applied mechanism-driven sintering ore blending model is based on material balance and field experience to set the constraints of normal temperature indicators such as raw material composition, raw material dosage, raw material price, sinter composition, sinter alkalinity, and sintering return amount, etc., based on the sinter ore The linear programming model is solved by mathematical methods such as simplex method and first-order gradient method with the lowest cost as the goal. This type of model has been applied in steel companies such as Xinyu Steel and Tianjin United Special Steel. The earlier mechanism-driven sintering ore blending model has fewer constraints and runs faster, but its constraints only include some normal temperature indicators. In actual sintering production, other normal temperature characteristics and high-temperature characteristics of raw materials will also largely affect the sintered mineral products. Quality, therefore there is a certain deviation between the model and the actual production situation, and the accuracy of the model needs to be improved. To improve the accuracy of the model, the researchers carried out theoretical analysis through micro-sintering tests, sintering cup tests, on-site production data, etc., to add more constraints in line with the sintering production conditions into the model and improve the adaptability of the model. Brazil’s Vale has proposed new normal temperature index constraints, dividing the hydration degree, grain size and Al2O3 content of iron ore powder into three levels. The combination of the three indicators at different levels can optimize different sintering indicators, and based on the iron ore powder index Adjust the ore blending plan about the sintering index. Research by Wu Shengli and others found that metallurgical performance indicators such as sinter drum, low-temperature reduction powdering, reducibility, and droplet properties are closely related to the high-temperature characteristics such as assimilation temperature, liquid phase fluidity, and bonding phase strength of iron ore powder for sintering. , therefore when setting the sintering ore blending model constraints, considering the influence of the high temperature basic characteristics of iron ore powder can better generate a sintering ore blending plan that meets the sinter quality requirements.

2.2 Data-driven sintering intelligent ore blending

As the constraints of the mechanism-driven sintering intelligent ore blending model increase, the model becomes more complex and it becomes difficult to solve the model. Therefore, model constraints can be adjusted through data analysis, and intelligent algorithms can be used to solve the model to optimize the model-solving speed. Chang Jian et al. used principal component analysis to reduce the dimensionality of five iron ore powder high-temperature characteristic constraints, including assimilation, liquid fluidity, SFCA (needle-shaped composite calcium ferrite) generation characteristics, bonding phase strength, and continuous crystal consolidation characteristics, to 2 A principal component constraint reduces the constraints of the ore blending model while retaining variable information as much as possible. In terms of model solution methods, intelligent algorithms are more effective than mathematical methods when there are many constraints. Genetic algorithms and particle swarm algorithms can be used to achieve rapid optimization of sintering intelligent ore blending, and the ant colony optimization particle swarm algorithm and genetic algorithms based on linear programming solution sets are used further to improve the accuracy of the sintering ore blending model.

Existing research on intelligent ore blending in sintering has made great progress in mechanism analysis. The constraint settings of the ore blending model are highly consistent with on-site conditions, and the optimization speed and accuracy of the ore blending model have been accelerated through data analysis and intelligent algorithms. However, existing research relies on mechanism analysis in terms of ore blending model constraints and does not conduct in-depth mining of information on ore blending-related variable data. There is still room for further optimization of the ore blending model constraint settings. One of the core goals of sintering intelligent ore blending is to produce sinter with quality that meets the requirements of blast furnaces. However, in existing research, there is a lack of verification of the blast furnace use effect of the optimal solution of the sintering intelligent ore blending model corresponding to the sinter. Therefore, researchers still need to further dig into the field historical data and laboratory data information from the two aspects of mechanism analysis and data analysis, optimize the constraint setting of the sintering ore blending model, and develop intelligent sintering blending that deeply integrates the mechanism data and adapts to the production line conditions. ore model, and deeply optimize the sintering intelligent ore blending model based on the blast furnace usage effect.

3. Prediction of sintering process parameters

The sintering process includes hundreds of high-frequency status parameters and dozens of inspection frequency quality parameters. Among them, the state parameters of common concern include the end position, material layer permeability, end temperature, sintering machine air leakage rate, etc. The quality parameters include sinter composition, alkalinity, drum index, high-temperature metallurgical properties, etc. Sintering site operators pay attention to and predict certain state quality parameters based on experience, hoping to stabilize each sintering state quality parameter within the ideal range by adjusting the sintering machine operating parameters, improving the quality of sintered minerals, and extending the life of sintering equipment. Therefore, it is the unanimous demand of sintering operators to accurately predict in advance state parameters such as the sintering end position and end temperature, as well as quality parameters such as sinter alkalinity and sinter drum index, and give on-site operators sufficient time for operation adjustment.

3.1 Prediction of sintering state parameters

The sintering end position is one of the core state parameters of the sintering process. It is usually controlled at the penultimate bellows position according to process requirements. If the sintering end position is advanced or delayed, resulting in over- or under-firing of the sintering process, it will cause the sintering utilization coefficient to decrease. Problems such as the decline in sinter quality. Calculating the wind box exhaust gas temperature can determine the sintering end position. Therefore, a wind box exhaust gas temperature prediction model can be established through genetic programming, Elman neural network or ridge regression, and then the corresponding sintering end position can be calculated based on the wind box exhaust gas temperature prediction results to achieve advanced prediction of the sintering end position. The calculation of the sintering end point usually requires the exhaust gas temperature data of three or more wind boxes, so the above method needs to predict the exhaust gas temperatures of multiple wind boxes at the same time. Each wind box exhaust gas temperature has a corresponding prediction error, so this method has the risk of error superposition.

Therefore, researchers’ related work focuses more on using historical data analysis of the sintering process and intelligent algorithms to directly establish an intelligent prediction model for the sintering endpoint. After collecting historical sintering data for some time, stepwise regression, principal component analysis, mechanism analysis and other methods are used to select sintering process parameters that are highly correlated with the sintering end point position as model input. Through multiple regression, ANFIS neural network (Adaptive Fuzzy Neural Network) ), LS-SVM (least squares support vector machine) and other algorithms to establish a sintering end position prediction model. To further improve the accuracy and field applicability of the model, the prediction of the sintering end position was optimized from the aspects of model data volume and model algorithm. When the data sources are the same, thousands or tens of thousands of sets of high-frequency data contain more information than hundreds of sets of data, and the end-point position prediction model established is also more suitable for on-site sintering production; BSO ( Particle Group) optimization LS-SVM, DSFA-PLS-LSTM (deep slow feature analysis-partial least squares- long short term memory, slow feature analysis-partial least squares long short term memory neural network), PSO-RBF (particle swarm optimization-radial basis function, particle swarm optimization radial neural network), genetic-particle swarm hybrid optimization neural network, SRP-ERVFLNs (sparse representation pruning-random weight neural network) and other improved algorithms have improved the accuracy of endpoint position prediction. The accuracy of the existing sintering end point position prediction model is shown in Table 1. Under different evaluation criteria, the accuracy of the existing sintering end position prediction model performs well on the test set.

In addition to the sintering end position, the air permeability of the sintering material layer, the sintering end temperature, and the sintering air leakage rate are also important indicators that characterize the state of the sintering process. Among them, the air permeability of the sintered material layer affects the quality of sintered minerals by affecting the vertical sintering speed; the sintering end temperature is a key parameter that reflects the thermal state of sintering, and the sintering air leakage rate affects the air volume of the sintered material layer, which in turn affects the quality of sintered minerals. The researchers screened the parameters of the sintering process through different model feature selection methods such as Pearson correlation analysis, RFE (recursive feature elimination), random forest, and mechanism analysis, and used neural networks, GBDT (gradient boosting decision tree), and PSO-BP neural networks (particle Algorithms such as swarm optimization-back propagation neural network (particle swarm neural network) predict the values or categories of sintering target state parameters in advance, and optimize through sintering process working condition classification, algorithm prediction effect comparison and selection, model hyperparameter grid search cross-validation and other methods Prediction model of sintering state parameters such as material layer air permeability. The accuracy of the existing prediction model for sintering state parameters such as the air permeability of the sintered material layer, endpoint temperature, and air leakage rate is shown in Table 2.


Table 1 Accuracy of the existing model for predicting BTP

Serial Number Prediction Method Root Mean Square Error Tolerance Scope Hit Rate
1 Multiple Regression 0.5 wind boxes Over 90%
2 ANFIS 0.02m 97%
3 LS-SVM 0.13 wind boxes 100%
4 Time series adaptation 0.2 trolleys 86.94%
6 GBDT 0.5 wind boxes Over 99%
7 PSO-RBF 0.56
8 SRP-ERVFLNs 0.08


Table 2 Accuracy of an existing prediction model for other sintering state parameters

Serial Number Sintering State Parameters Prediction Method Average Relative Error/% Root Mean Square Error Tolerance Scope Hit Rate/%
1 Material layer breathability Neural Networks 2.73
2 Material layer breathability Support Vector Machines 87.5
3 Material layer breathability Deep Neural Network 0.10
4 End temperature Grid Search Dynamic Subspace 10℃ 100
5 End temperature GBDT 18
6 Air leakage rate PSO-BP neural network 0.0086


3.2 Prediction of sintering quality parameters

The composition of the sinter is one of the most basic and critical requirements for sinter quality in blast furnaces. It directly affects the quality of blast furnace hot metal and slag composition, and also has a certain impact on the metallurgical properties of sinter. Therefore, researchers have carried out a lot of work on the prediction of sinter composition. Changes in the TFe content of the sinter directly affect the blast furnace fuel ratio and output. However, due to the small fluctuations in the TFe content of sinter in on-site production, in addition to the multi-timescale TFe prediction model established by CHEN not much.

In addition to the TFe content of the sinter, the alkalinity of the sinter is also one of the important components of the sinter composition index. Researchers used a grey neural network, genetic programming, support vector machine, BPNN (back propagation neural network) and other algorithms to establish the sinter Alkalinity prediction model. The FeO content of the sinter has a strong impact on the metallurgical properties of the sinter and the economic and technical indicators of the blast furnace. Therefore, it is also the sinter composition that on-site sintering engineers focus on. Pearson correlation analysis, Xgboost, MIC (maximal information coefficient), MIV (mean decrease impurity variables) and other algorithms can be used to screen the sintering process parameters that affect the FeO content of the sinter from the aspect of data analysis, and then use RBF neural network to improve SVR (Support vector regression) algorithm, RNN (recurrent neural network), genetic algorithm optimized BP neural network, etc. to establish a sinter FeO content prediction model. In addition to raw material composition and sintering process parameters, the infrared thermal imaging image of the sintering machine tail is also an important basis for judging the FeO content of the sinter. Obtaining the infrared image of the sintering machine tail and establishing a convolutional neural network FeO prediction model is one of the ways to predict the FeO content of sinter ore.

The metallurgical properties of sinter have a great influence on the stability and smoothness of the blast furnace. Therefore, the accurate prediction of normal temperature metallurgical properties such as sinter drum index and high-temperature metallurgical properties of sinter such as low-temperature reduction and powdering, reducibility, and reflow dripping properties has always been intelligent. The focus of sintering technology. Researchers have established an intelligent prediction model for the sinter drum index by using methods such as multivariate linear regression, BP neural network, El-man neural network, and gradient boosting regression. The model input is composed of mechanism analysis, support vector machine recursive feature elimination, and Granger causality. Analysis and other screening methods.

3.3 Forecast of sintering economic and technical indicators

The existing sintering economic and technical indicator predictions include the prediction of the amount of sinter return ore and the prediction of energy consumption in the sintering process. The amount of sinter return represents the quality of the sinter, and the energy consumption in the sintering process represents the cost of the sintering process. Both of them also represent the level of the sintering operation. In terms of predicting the return amount of sinter ore, the BP neural network and random forest algorithm are used to establish a prediction model for the return amount of sinter ore. The prediction model for the return amount of sinter ore can be further improved by integrating different prediction models and data set classification predictions through optimal weighting coefficients. Accuracy. In terms of energy consumption prediction in the sintering process, a sintering energy consumption prediction model is established through BP neural network, RBF neural network and other methods, and a genetic algorithm is used to optimize the BP neural network, the AdaBoost algorithm is improved and combined with the ELM network, and the ELmanRNN prediction is combined with the extreme learning network prediction. The sintering energy consumption prediction model is optimized using methods such as MMEC (max-min ensemble classifier) algorithm working condition identification.

Existing sintering process parameter prediction research has used different methods to construct an intelligent prediction model that meets the accuracy requirements under certain test conditions for the generally concerned key parameters of sintering state quality, and has improved the intelligent algorithm used. The model has achieved better results on the test set. The prediction effect is good. However, the parameters involved in the existing sintering process parameter prediction are not comprehensive. In addition to the generally concerned key parameters of sintering, there is a lack of research on other key parameters of the sintering process, which cannot fully meet the needs of on-site operators. Moreover, the sintering and smelting cycle is basically 30 to 45 minutes, and the practical application effect of the sintering state parameter prediction model established with low-frequency data in the field is limited. At the same time, in the existing research, the selection method of model input is fixed, that is, it is believed that the correlation between parameters discovered in the sintering historical data has always been suitable for actual on-site sintering production. However, the actual production situation is complex and the correlation between sintering parameters will also change. Therefore, the existing sintering process parameter prediction model lacks a process of dynamically screening model features based on actual sintering process data, which affects the adaptive ability of the model in the sintering production line.

In addition, after the sintering process parameter prediction model is constructed and optimized, the prediction results are usually directly provided to on-site operators to assist decision-making, lacking in-depth utilization of the prediction model. Therefore, it is necessary to comprehensively select the target parameters for sintering process prediction based on the data characteristics of the production line and the needs of on-site operators, and collect reasonably frequent model learning data based on the frequency of target parameter data, the sintering smelting cycle and the needs of operators to construct target parameters for the sintering process. Dynamically related relationships with other parameters and establish a sintering process parameter prediction model. While displaying the prediction results on the human-computer interaction interface, a feedback operation plan set is formed, and then the sintering state quality parameters corresponding to the feedback operation plan set are predicted based on the prediction model. It also pushes feedback operation plans corresponding to the optimal prediction results, giving full play to the function of the sintering process parameter prediction model to assist on-site personnel in making quick decisions.

4. Comprehensive evaluation and optimization of the sintering process

The operating logic of the sintering site operators is to predict and evaluate the overall state quality of the sintering production line, and then adjust the control parameters to maintain the overall state quality of the sintering production line within the ideal range. Therefore, comprehensive scientific evaluation and optimization of sintering process state quality based on accurate prediction of sintering process parameters is the ultimate goal of sintering intelligent technology. In existing research, two state parameters or two quality parameters are usually selected as targets for evaluation and optimization, such as the comprehensive evaluation and optimization of the linear weighted satisfaction function of the end position-mixing trough level, and the end position-mixing trough level. Expert rules coordinated comprehensive evaluation and fuzzy mathematics coordinated optimization, carbon efficiency-end position fuzzy satisfaction comprehensive evaluation and optimization, sinter drum index-sieving index Kmeans cluster analysis comprehensive evaluation, sinter drum index-low temperature reduction powder The comprehensive evaluation and optimization of the sintering process have been explored and achieved certain results.

In current research on the comprehensive evaluation and optimization of the sintering process, there are few parameters involved in the evaluation. The evaluation parameter information is not enough to represent the overall quality level of the sintering state. Moreover, the selection of evaluation parameters relies on experience. There is currently no sintering state that integrates experience and data. The comprehensive quality evaluation parameter selection process, the final evaluation and optimization results, and the on-site application effect are not yet clear. Therefore, it is necessary to select sintering process characterization parameters and sintering process operating parameters based on sintering production process data information and expert experience, establish a comprehensive evaluation system for sintering state quality that integrates data and experience, and carry out real-time evaluation of sintering state quality and evaluation of sintering state quality prediction results. Based on the demand for the overall level of the sintering process in blast furnace smelting and the prediction of sintering process parameters, the sintering operating parameters are adjusted to achieve comprehensive evaluation and optimization of the sintering process.

Leave a Reply

Your email address will not be published. Required fields are marked *

With machine vision as the core, DBM can provide intelligent equipment R&D, manufacturing, and transformation services such as unmanned factories in the metallurgical industry and other industries, machine vision, and robot application design.

Product category

Customer testimonials

Tell us your needs and a technician will contact you immediately.

Friendship Link

We are committed to becoming a “leader in visual applications”, advocating the core values of “integrity, cooperation, innovation, expertise, and excellence”, taking the road of “technology shaping the brand, innovation leading the industry”, and attaching importance to the investment in scientific research and technology. The research and development of new products and the construction of the team make unremitting efforts for the goal of scientific and technological development.

We have been supplied to world’s top steel manufacturer Arcelormittal, TATA Steel, EZZ steel etc. We do OEM for Concast and Danieli for a long time

Metallurgical Intelligence
Change the way of manufacturing

Save labor, improve working environment, improve efficiency, reduce error rate and save cost!

Металлургический интеллект
Изменить способ производства

Сэкономьте труд, улучшите рабочую среду, повысьте эффективность, уменьшите количество ошибок и снизьте затраты!