Sintered ore alkalinity intelligent control method based on mixed information

Abstract

The invention discloses an intelligent sinter alkalinity control method based on mixed information, and belongs to the technical field of intelligent sinter production. The method comprises the following steps: acquiring test data of raw materials of iron, silicon, aluminum, magnesium and calcium and online detection data of a mixture in real time, and constructing an input data set after standardization treatment; a particle swarm optimization (PSO) algorithm is adopted to optimize hyper-parameters of a support vector regression (SVR) model, a weighted multi-fusion data prediction system is constructed, and a mixture component enhancement prediction model (MCEP) is formed; calculating a short-period and long-period mean value sequence through time grouping, optimizing prediction precision in combination with dynamic weight adjustment and clustering analysis, and updating the model in real time based on production conditions; and finally, realizing closed-loop control through component fluctuation mean value analysis and early warning. The method has remarkable benefits in the aspects of improving production process efficiency, reducing cost, optimizing resource utilization, improving product quality and promoting production intelligence.

Description

Technical Field

[0001] The present invention relates to the field of intelligent sintered ore production, and in particular to a mixed information-based intelligent control method for sintered ore basicity. Background Art

[0002] Sintered ore is one of the most important raw materials currently used in blast furnace ironmaking. The stability of its basicity (R) is closely linked to key indicators such as sinter yield and drum strength. These indicators directly impact the stable and smooth operation of the blast furnace and the optimization of technical and economic indicators. Therefore, controlling the chemical composition of sintered ore must focus on basicity, as changes in sintered ore basicity can cause fluctuations in its chemical composition. Furthermore, the structure of the blast furnace charge and the characteristics of the blast furnace smelting slag are directly based on this fundamental factor. Fluctuations in sintered ore basicity not only affect the composition and quality of the charge but can also impact the blast furnace's smelting performance, leading to unstable blast furnace production and even affecting the company's economic benefits.

[0003] For example, at a certain steel plant, the total amount of sintered ore fluctuated significantly. The quality control of the sintered ore produced by the No. 3 sintering machine was relatively good, but the pass rate for sintered ore basicity R±0.08 averaged 94.62% from January to October 2020, a significant gap compared to the advanced domestic and international standards of 98%-99%. This fluctuation had a serious impact on the long-term stable operation of the blast furnace and the improvement of its performance indicators. Therefore, to further improve the quality control of sintered ore and promote the in-depth application of intelligent sintering production, it is urgent to adopt advanced online composition detection technology, combined with intelligent control systems, to comprehensively improve the chemical composition stability and basicity control level of the sintered ore.

[0004] In recent years, online mixture analysis equipment has garnered widespread attention and initial application. A steel plant's No. 3 sintering machine has introduced online mixture analysis equipment for online composition testing of the sinter mixture. This real-time detection method enables high-frequency and high-precision monitoring of the chemical composition of the sintered ore. However, relying solely on mixture analysis equipment for composition monitoring is difficult to achieve closed-loop control of the sintered ore composition, necessitating the development of a supporting intelligent composition control system. This system embeds the knowledge of sintering experts into the control software, enabling closed-loop intelligent control of the sintered ore composition. This approach not only improves the stability of the sintered ore basicity but also provides a more stable feedstock for the blast furnace, significantly improving blast furnace output.

[0005] During the sintering production process, most composition fluctuations often go undetected and uncontrolled. The introduction of online mixture composition analysis equipment will significantly enhance the ability to visualize and control the entire composition process. Furthermore, intelligent composition control systems effectively combine online analyzer data with expert knowledge to improve the stability of sinter quality. This not only helps blast furnaces maintain feedstock stability but also reduces coke consumption, lowers production costs, significantly reduces the number of blast furnace overhauls, and extends blast furnace life. Currently, relevant patented technologies provide solutions for online detection and intelligent control of sinter composition. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for intelligently controlling the basicity of sintered ore based on mixed information to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solution: a method for intelligently controlling the basicity of sintered ore based on mixed information, characterized in that it comprises the following steps:

[0008] S1: Real-time acquisition of raw material test data of iron, silicon, aluminum, magnesium, and calcium in the sintering batching system and online detection data of the mixture to form dynamic and continuous time series data;

[0009] S2: Combine the two types of data and standardize them to construct a high-quality input dataset;

[0010] S3: Particle swarm optimization (PSO) is used to optimize the key hyperparameters of the support vector regression (SVR) model, including kernel function parameters, penalty coefficients, and error tolerance.

[0011] S4: Build a weighted multi-fusion data prediction system based on the optimized SVR model and determine the weight distribution through cross-validation;

[0012] S5: Use the optimized “Mixture Composition Enhanced Prediction Model (MCEP)” to predict the key components of the mixture in real time and calculate the binary alkalinity R;

[0013] S6: Calculate the mean series of short-term and long-term periods by time grouping, perform multi-level data fusion, and balance short-term fluctuations with long-term trends;

[0014] S7: Dynamically adjust model training weights to optimize the prediction accuracy of key components and capture the fluctuation patterns of chemical composition through cluster analysis methods;

[0015] S8: Update the MCEP model in real time based on actual production conditions, improve the model's generalization ability through the accumulation of historical data, and ensure long-term stability of alkalinity;

[0016] S9: Generate component fluctuation means of different time periods and implement closed-loop control in combination with early warning information.

[0017] Preferably, in step S1, the raw material analysis and the mixed material coverage in the test meet the following requirements: the FeO mass fraction is 6.59% to 10.74%; the mass fraction of sintered ore SiO2 is 4.53% to 6.09%; the CaO mass fraction is 8.59% to 13.74%; the Al2O3 mass fraction is 1.51% to 3.17%; the MgO mass fraction is 1.02% to 2.65%; the sintered ore R is controlled between 1.72 and 2.35; wherein the chemical components of the sintered ore are mainly: TFe, SiO2, FeO, Al2O3, MgO and sintered ore basicity R.

[0018] Preferably, in step S2, the raw material test data and the mixture online test data are combined to clean and standardize the data to reduce the impact of dimensional differences on model training. The data cleaning includes outlier detection and elimination, and abnormal data that deviates significantly from the mean is eliminated by setting upper and lower limits; the standardization process uses a normalization method to map the value of each variable to the [0,1] interval.

[0019]

[0020] Among them, X norm is the normalized variable value, X is the original variable value, and X min 、X max They are the minimum and maximum values ​​of the variables, respectively. After standardization, a high-quality input data set is generated to ensure that data from different sources are modeled and analyzed at the same scale, providing consistent data input for subsequent model training and optimization.

[0021] Preferably, in step S3, a particle swarm optimization algorithm (PSO) is used to optimize the key hyperparameters of the support vector regression model (SVR), wherein the hyperparameters include the kernel function parameter C penalty coefficient, γ, and error tolerance ∈. The PSO algorithm initializes the position and velocity of the particle swarm and iteratively updates the position of the particles according to the fitness function to find the optimal solution. The fitness function is the optimization objective function of the SVR and is defined as follows:

[0022]

[0023] Where n is the number of samples, y i is the mixture test value, is the SVR predicted value, λ‖w‖ 2 It is a regularization term used to prevent overfitting. The PSO algorithm updates the position and velocity of particles using the following formula:

[0024]

[0025] in, is the particle velocity, is the particle position, w is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, and p i is the optimal position of the individual particle, g is the global optimal position, and the hyperparameters optimized by PSO are applied to the SVR model, which significantly improves the prediction accuracy and generalization ability of the model.

[0026] Preferably, in step S4, after particle swarm optimization, a prediction function of a support vector regression (SVR) model is constructed to process the nonlinear relationship of the input data, and the prediction function is:

[0027]

[0028] Among them, K(x,x i ) is the kernel function, usually the radial basis kernel function (RBF):

[0029] K(x,x i )=exp(-γ||xx i || 2 )

[0030] α i is the weight of the support vector, x is the input feature, and b is the bias term, which is determined through the training process. The optimized SVR model is used to model the fused raw material test data and the mixed material online detection data to achieve accurate prediction of the sintered ore composition. Furthermore, by combining the SVR output results, a weighted multi-fusion data prediction system is constructed. The prediction formula of this system is:

[0031]

[0032] in and Respectively represent the raw material test and mixture test results, weight w i Through random grid search cross validation optimization, the weights are optimized by random grid search and cross validation methods to achieve weighted fusion of prediction results.

[0033] Preferably, in step S5, the optimized "Mixture Composition Enhanced Prediction Model (MCEP)" is used to predict the real-time content of key components (silicon and calcium) of the mixture to obtain the binary alkalinity

[0034] R predicted =Ca predict / Si predict .

[0035] Preferably, in step S6, the chemical composition data of the mixture is obtained in real time by an online mixture detection device, and the detection data is grouped and counted according to the time dimension, and the mean sequence of the short period (such as 15 minutes, 30 minutes) and the long period (such as 1 hour, 4 hours) is calculated respectively. The mean sequence of the short period data is used to characterize the short-term fluctuation characteristics of the chemical composition of the mixture, and the mean sequence of the long period data is used to describe the long-term change trend of the composition. The mean sequences of the short period and the long period are calculated by the following formula:

[0036]

[0037] Among them, X short represents the short-term mean series, X long represents the long-period mean sequence, n and m are the time window lengths of the short-period and long-period respectively, is the chemical composition detection value at time point t, and the short-period and long-period mean sequences are multi-level fused to construct the multi-level feature vector X of the time series fused , the formula is as follows:

[0038] X fused =w short X short +w long X long

[0039] where w short and w long are weight parameters for short-term and long-term data respectively. The optimal weight configuration is determined by cross-validation method to achieve a balance between short-term fluctuations and long-term trends. Finally, based on the multi-level feature vector X fused Towards modeling and predicting the fluctuating nature of chemical composition.

[0040] Preferably, in step S7, the model training weights are dynamically adjusted based on the time series features to assign higher weights to the detection and prediction of key components (silicon and calcium) in the mixture, thereby enhancing the model's ability to capture the fluctuation patterns of core components. The captured features are used as input data to train the "Mixture Composition Enhanced Prediction Model (MCEP)". The dynamically adjusted weight optimization formula is:

[0041]

[0042] Among them, y i is the fusion data of mixture and raw materials, f(X fused ) is the predicted value of the model, λ‖w‖ 2 is a regularization term. The final optimized model has the comprehensive prediction ability of short-term fluctuations and long-term trends. The specific methods include:

[0043] Feature extraction and clustering analysis: Based on short-term and long-term data, analyze the fluctuation characteristics of the chemical composition of the mixture, combine the time series characteristics of key ingredients such as silicon and calcium, and use clustering analysis method to optimize the input data of the model, making the alkalinity prediction more accurate.

[0044] Dynamic weight adjustment: According to the importance of different ingredients, by increasing the weight of key ingredients (silicon, calcium) in the training process, to ensure that the model learns the features that have the greatest impact on alkalinity prediction first;

[0045] Model optimization and generalization ability improvement: Through dynamic learning and historical data accumulation, improve the comprehensive prediction ability of the model for short-term fluctuations and long-term trends, ensure the robustness and stability of alkalinity prediction.

[0046] Preferably, in step S8, the mixture composition enhanced prediction model (MCEP) is updated in real time according to the actual production conditions (such as changes in raw material properties, adjustment of batching ratio), through dynamic accumulation and learning mechanism of historical data, continuously optimize the generalization ability of the model, ensure the long-term stable control of the mixture alkalinity, including the following steps:

[0047] Real-time data update and dynamic training: According to the raw material test data and mixture real-time data detected in actual production, dynamically adjust the training data set D t of the model, and combine historical data D h to form a comprehensive training set D combined

[0048] D combined = αD t + (1-α)D h

[0049] Where, α is the weight parameter, used to balance the influence of current real-time data and historical data, to ensure the adaptability and stability of the model;

[0050] Model parameter self-adaptive adjustment: In the production process, real-time monitoring of the change trend of key ingredients (such as silicon, calcium), based on dynamic error feedback mechanism to adjust the key parameters (such as learning rate and regularization parameter) of MCEP model, the adjustment target is to minimize the following loss function:

[0051]

[0052] Where, y i is the target alkalinity value, is the predicted value, λ‖θ‖ 2 is the regularization term, θ represents the model parameters, and λ is the regularization coefficient;

[0053] Cumulative learning and long-term optimization: Utilizing the cumulative learning mechanism of historical data, the model is regularly retrained to enhance generalization capabilities. The weights of historical data features are dynamically adjusted through a time decay function:

[0054] w(t)=e -βt

[0055] Where t is the step size and β is the decay coefficient, which is used to control the contribution of historical data to model training. Forecast performance evaluation and adaptive iteration: Regularly evaluate the forecast performance of the model and measure the model stability by the weighted average of the real-time error and the historical error:

[0056] E=αE t +(1-α)E h

[0057] Among them, E t is the current error, E h is the historical error, α is the weight of the current error, and when E exceeds the preset threshold, the model update mechanism is triggered. Ultimately, through the above dynamic update and cumulative learning mechanism, the MCEP model can continuously maintain high-precision prediction and stable control of the mixture alkalinity while responding to changes in production conditions, thereby supporting the intelligent adjustment and optimization of sintering production.

[0058] Preferably, in step S9, the predicted value and historical data are divided into different time periods (such as 15 minutes, 30 minutes, 1 hour, etc.), and the mean and standard deviation of the component fluctuations are calculated to form the time series characteristics of the component fluctuations. When the fluctuation exceeds the preset control range, an early warning information is triggered, and finally closed-loop control is achieved through automatic adjustment.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] This invention significantly improves the real-time prediction accuracy of key sinter ore components (such as silicon and calcium) by incorporating online mixture composition detection technology, combined with an optimized particle swarm optimization (PSO) algorithm and support vector regression (SVR) model. Real-time prediction of binary basicity (R) accurately guides ingredient adjustments, reduces the impact of composition fluctuations on production stability and product quality, and significantly improves production efficiency.

[0061] Through multi-level data fusion technology, the short-term and long-term component fluctuation characteristics are combined to form time series features. Through optimized weight distribution, comprehensive control of short-term fluctuations and long-term trends is achieved. This method can respond to raw material changes and process fluctuations in real time, better adapting to the dynamic changes in the actual production environment and improving the stability of the sintering production process.

[0062] Based on historical data accumulation and dynamic weight adjustment, the proposed MCEP model can adaptively update according to changes in the chemical composition and proportions of raw materials. Cluster analysis is also used to optimize the characteristics of the model input data, giving the model strong generalization and long-term stability. This feature ensures high prediction reliability in response to raw material fluctuations and process adjustments.

[0063] In summary, the present invention has significant benefits in improving production process efficiency, reducing costs, optimizing resource utilization, improving product quality, and promoting intelligent production, and provides a new solution for the technological upgrading and industrial application of sintered ore production processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic diagram of the present invention;

[0065] Figure 2-6 The average laboratory error is calculated based on the deviation of the results of 78 groups of samples (i.e. 156 parallel samples), and a data chart comparing the mixture test results with the laboratory component analysis results;

[0066] Figure 7 Comparison of Ca composition results predicted by the final model;

[0067] Figure 8 Comparison of Si composition results predicted by the final model.

[0068] In the figure: 1. Upper pipe stop surface; 2. Lower connecting seat; 3. Spring; 4. Hinge pin; 5. Nut; 6. Guide pin; 7. Lug seat; 8. Coupling flange; 9. Cast pipe; 10. Rotating spindle; 11. Coupling; 12. Bearing seat; 13. Pipe stop frame; 14. Cylinder frame; 15. Cylinder. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the implementation regulations described are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0070] A method for intelligently controlling sinter basicity based on mixed information comprises the following steps:

[0071] S1: Real-time acquisition of raw material test data and online detection data of the mixture of iron, silicon, aluminum, magnesium and calcium in the sintering batching system to form dynamic and continuous time series data; the raw material test and mixture coverage in the experiment meet the following requirements: FeO mass fraction is 6.59% to 10.74%; the mass fraction of SiO2 in sintered ore is 4.53% to 6.09%; the mass fraction of CaO is 8.59% to 13.74%; the mass fraction of Al2O3 is 1.51% to 3.17%; the mass fraction of MgO is 1.02% to 2.65%; the R of sintered ore is controlled between 1.72 and 2.35.

[0072] S2: Combine the two types of data and standardize them to construct a high-quality input data set; combine the raw material test data and the mixed material online detection data, clean and standardize the data to reduce the impact of dimensional differences on model training. The data cleaning includes outlier detection and elimination, and abnormal data that deviates significantly from the mean is eliminated by setting upper and lower limits; the standardization process uses the normalization method to map the value of each variable to the [0,1] interval.

[0073]

[0074] Among them, X norm is the normalized variable value, X is the original variable value, and X min 、X max They are the minimum and maximum values ​​of the variables, respectively. After standardization, a high-quality input data set is generated to ensure that data from different sources are modeled and analyzed at the same scale, providing consistent data input for subsequent model training and optimization.

[0075] S3: Particle swarm optimization (PSO) is used to optimize the key hyperparameters of the support vector regression (SVR) model, including the kernel function parameter C, the penalty coefficient γ, and the error tolerance ∈. The PSO algorithm initializes the position and velocity of the particle swarm and iteratively updates the particle positions according to the fitness function to find the optimal solution. The fitness function is the optimization objective function of SVR and is defined as follows:

[0076]

[0077] Where n is the number of samples, y i is the mixture test value, is the SVR predicted value, λ‖w‖ 2 It is a regularization term used to prevent overfitting. The PSO algorithm updates the position and velocity of particles using the following formula:

[0078]

[0079] in, is the particle velocity, is the particle position, w is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, and p i is the optimal position of the individual particle, g is the global optimal position, and the hyperparameters optimized by PSO are applied to the SVR model, which significantly improves the prediction accuracy and generalization ability of the model.

[0080] S4: Construct a weighted multi-fusion data prediction system based on the optimized SVR model and determine the weight distribution through cross-validation; after particle swarm optimization, construct a prediction function of the support vector regression (SVR) model to process the nonlinear relationship of the input data. The prediction function is:

[0081]

[0082] Among them, K(x,x i ) is the kernel function, usually the radial basis kernel function (RBF):

[0083] K(x,x i )=exp(-γ||xx i || 2 )

[0084] α i is the weight of the support vector, x is the input feature, and b is the bias term, which is determined through the training process. The optimized SVR model is used to model the fused raw material test data and the mixed material online detection data to achieve accurate prediction of the sintered ore composition. Furthermore, by combining the SVR output results, a weighted multi-fusion data prediction system is constructed. The prediction formula of this system is:

[0085]

[0086] in and Respectively represent the raw material test and mixture test results, weight w i Through random grid search cross validation optimization, the weights are optimized by random grid search and cross validation methods to achieve weighted fusion of prediction results.

[0087] S5: Use the optimized "Mixture Composition Enhanced Prediction Model (MCEP)" to predict the key components of the mixture in real time and calculate the binary alkalinity R; Use the optimized "Mixture Composition Enhanced Prediction Model (MCEP)" to predict the real-time content of the key components of the mixture (silicon, calcium) to obtain the binary alkalinity

[0088] R predicted =Ca predict / Si predict .

[0089] S6: Calculate the short-term and long-term mean series by time grouping, perform multi-level data fusion, and balance short-term fluctuations with long-term trends; obtain the chemical composition data of the mixture in real time through the online mixture detection device, and group and count the detection data according to the time dimension, and calculate the short-term (such as 15 minutes, 30 minutes) and long-term (such as 1 hour, 4 hours) mean series respectively. The mean series of the short-term data is used to characterize the short-term fluctuation characteristics of the chemical composition of the mixture, and the mean series of the long-term data is used to describe the long-term change trend of the composition. The short-term and long-term mean series are calculated using the following formula:

[0090]

[0091] Among them, X short represents the short-term mean series, X long represents the long-period mean sequence, n and m are the time window lengths of the short-period and long-period respectively, is the chemical composition detection value at time point t, and the short-period and long-period mean sequences are multi-level fused to construct the multi-level feature vector X of the time series fused , the formula is as follows:

[0092] X fused =w short X short +w long X long

[0093] where w short and w long are weight parameters for short-term and long-term data respectively. The optimal weight configuration is determined by cross-validation method to achieve a balance between short-term fluctuations and long-term trends. Finally, based on the multi-level feature vector X fused Towards modeling and predicting the fluctuating nature of chemical composition.

[0094] S7: Dynamically adjust the model training weights to optimize the prediction accuracy of key components and capture the fluctuation patterns of chemical components through cluster analysis methods; dynamically adjust the model training weights based on time series features to give higher weights to the detection and prediction of key components (silicon and calcium) in the mixture, enhance the model's ability to capture the fluctuation patterns of core components, and use the captured features as input data to train the "Mixture Composition Enhanced Prediction Model (MCEP)". The dynamically adjusted weight optimization formula is:

[0095]

[0096] Among them, y i is the fusion data of mixture and raw materials, f(X fused ) is the predicted value of the model, λ‖w‖ 2is a regularization term. The final optimized model has the comprehensive prediction ability of short-term fluctuations and long-term trends. The specific methods include:

[0097] Feature extraction and cluster analysis: Based on short-term and long-term data, the fluctuation characteristics of the chemical composition of the mixture are analyzed. Combined with the time series characteristics of key components such as silicon and calcium, cluster analysis methods are used to optimize model input data to make alkalinity prediction more accurate;

[0098] Dynamic weight adjustment: Based on the importance of different components, the weight of key components (silicon, calcium) is increased during training to ensure that the model prioritizes learning the features that have the greatest impact on alkalinity prediction;

[0099] Model optimization and generalization capability improvement: Through dynamic learning and historical data accumulation, the model's comprehensive prediction capabilities for short-term fluctuations and long-term trends are improved, ensuring the robustness and stability of alkalinity prediction.

[0100] S8: Update the MCEP model in real time based on actual production conditions, improve the model's generalization ability through the accumulation of historical data, and ensure long-term stability of alkalinity; update the mixture composition enhancement prediction model (MCEP) in real time according to actual production conditions (such as changes in raw material properties and adjustments to ingredient ratios), and continuously optimize the model's generalization ability through the dynamic accumulation of historical data and learning mechanisms to ensure long-term stable control of mixture alkalinity. Specifically, the following steps are included:

[0101] Real-time data update and dynamic training: Dynamically adjust the model’s training data set D based on the raw material test data and real-time mixed material data detected in actual production t Combined with historical data D h Form a comprehensive training set D combined

[0102] D combined =αD t +(1-α)D h

[0103] Among them, α is a weight parameter used to balance the influence of current real-time data and historical data to ensure the adaptability and stability of the model;

[0104] Adaptive adjustment of model parameters: During the production process, the changing trends of key components (such as silicon and calcium) are monitored in real time. Based on the dynamic error feedback mechanism, the key parameters of the MCEP model (such as the learning rate and regularization parameter) are adjusted. The adjustment goal is to minimize the following loss function:

[0105]

[0106] Among them, y i is the target alkalinity value, is the predicted value, λ‖θ‖2 is the regularization term, θ represents the model parameter, and λ is the regularization coefficient;

[0107] Cumulative learning and long-term optimization: Utilizing the cumulative learning mechanism of historical data, the model is regularly retrained to enhance generalization capabilities. The weights of historical data features are dynamically adjusted through a time decay function:

[0108] w(t)=e -βt

[0109] Where t is the step size and β is the decay coefficient, which is used to control the contribution of historical data to model training. Forecast performance evaluation and adaptive iteration: Regularly evaluate the forecast performance of the model and measure the model stability by the weighted average of the real-time error and the historical error:

[0110] E=αE t +(1-α)E h

[0111] Among them, E t is the current error, E h is the historical error, α is the weight of the current error, and when E exceeds the preset threshold, the model update mechanism is triggered. Ultimately, through the above dynamic update and cumulative learning mechanism, the MCEP model can continuously maintain high-precision prediction and stable control of the mixture alkalinity while responding to changes in production conditions, thereby supporting the intelligent adjustment and optimization of sintering production.

[0112] S9: Generate the mean value of component fluctuations in different time periods and combine it with early warning information to achieve closed-loop control; by dividing the predicted value and historical data into different time periods (such as 15 minutes, 30 minutes, 1 hour, etc.), calculate the mean and standard deviation of the component fluctuations to form the time series characteristics of the component fluctuations. When the fluctuation exceeds the preset control range, the early warning information is triggered, and finally closed-loop control is achieved through automatic adjustment.

[0113] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent control of sintered ore basicity based on mixed information, characterized by: The following steps are involved: S1: Real-time acquisition of raw material test data of iron, silicon, aluminum, magnesium, and calcium in the sintering batching system and online detection data of the mixture to form dynamic and continuous time series data; S2: Combine the two types of data and standardize them to construct a high-quality input dataset; S3: Particle swarm optimization (PSO) is used to optimize the key hyperparameters of the support vector regression (SVR) model, including kernel function parameters, penalty coefficients, and error tolerance. S4: Build a weighted multi-fusion data prediction system based on the optimized SVR model and determine the weight distribution through cross-validation; S5: Use the optimized "Mixture Composition Enhanced Prediction Model (MCEP)" to predict the key components of the mixture in real time and calculate the binary alkalinity R; S6: Calculate the mean series of short-term and long-term periods by time grouping, perform multi-level data fusion, and balance short-term fluctuations with long-term trends; S7: Dynamically adjust model training weights to optimize the prediction accuracy of key components and capture the fluctuation patterns of chemical composition through cluster analysis methods; S8: Update the MCEP model in real time based on actual production conditions, improve the model's generalization ability through the accumulation of historical data, and ensure long-term stability of alkalinity; S9: Generate component fluctuation means of different time periods and implement closed-loop control in combination with early warning information.

2. The intelligent control method for sintered ore basicity based on hybrid information according to claim 1, characterized in that: In step S1, the raw material analysis and the mixed material coverage in the experiment meet the following requirements: the FeO mass fraction is 6.59% to 10.74%; the mass fraction of SiO2 in the sintered ore is 4.53% to 6.09%; the CaO mass fraction is 8.59% to 13.74%; the Al2O3 mass fraction is 1.51% to 3.17%; the MgO mass fraction is 1.02% to 2.65%; the sintered ore R is controlled between 1.72 and 2.35; and the chemical components of the sintered ore mainly include: TFe, SiO2, FeO, Al2O3, MgO and the sintered ore basicity R.

3. The intelligent control method for sintered ore basicity based on hybrid information according to claim 1, characterized in that: In step S2, the raw material test data and the mixture online test data are combined to clean and standardize the data to reduce the impact of dimensional differences on model training. The data cleaning includes outlier detection and elimination, and abnormal data that deviates significantly from the mean is eliminated by setting upper and lower limits. The standardization process uses a normalization method to map the value of each variable to the [0,1] interval. Among them, X norm is the normalized variable value, X is the original variable value, and X min 、X max They are the minimum and maximum values ​​of the variables, respectively. After standardization, a high-quality input data set is generated to ensure that data from different sources are modeled and analyzed at the same scale, providing consistent data input for subsequent model training and optimization.

4. The intelligent control method for sintered ore basicity based on hybrid information according to claim 1, characterized in that: In step S3, the particle swarm optimization algorithm (PSO) is used to optimize the key hyperparameters of the support vector regression model (SVR). The hyperparameters include the kernel function parameter C penalty coefficient, γ, and error tolerance ∈. The PSO algorithm initializes the position and velocity of the particle swarm and iteratively updates the position of the particles according to the fitness function to find the optimal solution. The fitness function is the optimization objective function of SVR and is defined as follows: Where n is the number of samples, y i is the mixture test value, is the SVR predicted value, λ‖w‖ 2 It is a regularization term used to prevent overfitting. The PSO algorithm updates the position and velocity of particles using the following formula: in, is the particle velocity, is the particle position, w is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, and p i is the optimal position of the individual particle, g is the global optimal position, and the hyperparameters optimized by PSO are applied to the SVR model, which significantly improves the prediction accuracy and generalization ability of the model.

5. The intelligent control method for sintered ore basicity based on mixed information according to any one of claim 1, characterized in that: In step S4, after particle swarm optimization, a prediction function of a support vector regression (SVR) model is constructed to process the nonlinear relationship of the input data. The prediction function is: Among them, K(x,x i ) is the kernel function, usually the radial basis kernel function (RBF): K(x,x i )=exp(-γ||x-x i || 2 ) α i is the weight of the support vector, x is the input feature, and b is the bias term, which is determined through the training process. The optimized SVR model is used to model the fused raw material test data and the mixed material online detection data to achieve accurate prediction of the sintered ore composition. Furthermore, by combining the SVR output results, a weighted multi-fusion data prediction system is constructed. The prediction formula of this system is: in and Respectively represent the raw material test and mixture test results, weight w i Through random grid search cross validation optimization, the weights are optimized by random grid search and cross validation methods to achieve weighted fusion of prediction results.

6. The intelligent control method for sintered ore basicity based on hybrid information according to claim 1, characterized in that: In step S5, the optimized "Mixture Composition Enhanced Prediction Model (MCEP)" is used to predict the real-time content of key components (silicon and calcium) of the mixture to obtain the binary alkalinity R predicted =As predict / And predict 。 7. The intelligent control method for sintered ore basicity based on hybrid information according to claim 1, characterized in that: In step S6, the chemical composition data of the mixture is obtained in real time through the online mixture detection device, and the detection data is grouped and counted according to the time dimension, and the mean series of short periods (such as 15 minutes, 30 minutes) and long periods (such as 1 hour, 4 hours) are calculated respectively. The mean series of the short-period data is used to characterize the short-term fluctuation characteristics of the chemical composition of the mixture, and the mean series of the long-period data is used to describe the long-term change trend of the composition. The short-period and long-period mean series are calculated by the following formula: Among them, X short represents the short-term mean series, X long represents the long-period mean sequence, n and m are the time window lengths of the short-period and long-period respectively, is the chemical composition detection value at time point t, and the short-period and long-period mean sequences are multi-level fused to construct the multi-level feature vector X of the time series fused , the formula is as follows: X fused =w short X short +w long X long where w short and w long are weight parameters for short-term and long-term data respectively. The optimal weight configuration is determined by cross-validation method to achieve a balance between short-term fluctuations and long-term trends. Finally, based on the multi-level feature vector X fused Towards modeling and predicting the fluctuating nature of chemical composition.

8. The intelligent control method for sintered ore basicity based on hybrid information according to claim 1, characterized in that: In step S7, the model training weights are dynamically adjusted based on the time series features to assign higher weights to the detection and prediction of key components (silicon and calcium) in the mixture, thereby enhancing the model's ability to capture the fluctuation patterns of core components. The captured features are then used as input data to train the Mixture Composition Enhanced Prediction Model (MCEP). The dynamically adjusted weight optimization formula is: Among them, y i is the fusion data of mixture and raw materials, f(X fused ) is the predicted value of the model, λ‖w‖ 2 is a regularization term. The final optimized model has the comprehensive prediction ability of short-term fluctuations and long-term trends. The specific methods include: Feature extraction and cluster analysis: Based on short-term and long-term data, the fluctuation characteristics of the chemical composition of the mixture are analyzed. Combined with the time series characteristics of key components such as silicon and calcium, cluster analysis methods are used to optimize model input data to make alkalinity prediction more accurate; Dynamic weight adjustment: Based on the importance of different components, the weight of key components (silicon, calcium) is increased during training to ensure that the model prioritizes learning the features that have the greatest impact on alkalinity prediction; Model optimization and generalization capability improvement: Through dynamic learning and historical data accumulation, the model's comprehensive prediction capabilities for short-term fluctuations and long-term trends are improved, ensuring the robustness and stability of alkalinity prediction.

9. The intelligent control method for sintered ore basicity based on hybrid information according to claim 1, characterized in that: In step S8, the mixture composition enhancement prediction model (MCEP) is updated in real time according to actual production conditions (such as changes in raw material properties and adjustments to ingredient ratios). Through the dynamic accumulation and learning mechanism of historical data, the generalization ability of the model is continuously optimized to ensure long-term stable control of mixture alkalinity. Specifically, the following steps are included: Real-time data update and dynamic training: Dynamically adjust the model’s training data set D based on the raw material test data and real-time mixed material data detected in actual production t Combined with historical data D h Form a comprehensive training set D combined D combined =αD t +(1-α)D h Among them, α is a weight parameter used to balance the influence of current real-time data and historical data to ensure the adaptability and stability of the model; Adaptive adjustment of model parameters: During the production process, the changing trends of key components (such as silicon and calcium) are monitored in real time. Based on the dynamic error feedback mechanism, the key parameters of the MCEP model (such as the learning rate and regularization parameter) are adjusted. The adjustment goal is to minimize the following loss function: Among them, y i is the target alkalinity value, is the predicted value, λ‖θ‖ 2 is the regularization term, θ represents the model parameter, and λ is the regularization coefficient; Cumulative learning and long-term optimization: Utilizing the cumulative learning mechanism of historical data, the model is regularly retrained to enhance generalization capabilities. The weights of historical data features are dynamically adjusted through a time decay function: w(t)=e -βt Where t is the step size and β is the decay coefficient, which is used to control the contribution of historical data to model training. Forecast performance evaluation and adaptive iteration: Regularly evaluate the forecast performance of the model and measure the model stability by the weighted average of the real-time error and the historical error: E=αE t +(1-a)E h Among them, E t is the current error, E h is the historical error, α is the weight of the current error, and when E exceeds the preset threshold, the model update mechanism is triggered. Ultimately, through the above dynamic update and cumulative learning mechanism, the MCEP model can continuously maintain high-precision prediction and stable control of the mixture alkalinity while responding to changes in production conditions, thereby supporting the intelligent adjustment and optimization of sintering production.

10. The intelligent control method for sintered ore basicity based on hybrid information according to claim 1, characterized in that: In step S9, the predicted value and historical data are divided into different time periods (such as 15 minutes, 30 minutes, 1 hour, etc.), and the mean and standard deviation of the component fluctuations are calculated to form the time series characteristics of the component fluctuations. When the fluctuation exceeds the preset control range, an early warning information is triggered, and finally closed-loop control is achieved through automatic adjustment.

Images