Sintering low-temperature reduction degradation index prediction stepwise regression method based on factor analysis whole process

Abstract

The invention provides a sintering low-temperature reduction degradation index prediction stepwise regression method based on a factor analysis whole process. The method comprises the following steps: S1, data acquisition and preprocessing; s2, factor analysis dimension reduction and comprehensive factor extraction; s3, constructing a prediction model based on optimal subset regression; s4, a model verification and dynamic updating mechanism; and S5, carrying out model deployment and online prediction. According to the method, advanced accurate prediction and intelligent regulation and control of the sintering production process quality can be realized, the finally established prediction model is integrated into a sintering process control system, real-time online prediction of the RDI index is realized, advanced and accurate data support is provided for operators to adjust key parameters such as alkalinity and fuel ratio, and the service life of the operators is prolonged. Finally, the purposes of stabilizing the sinter quality, reducing the blast furnace fuel ratio and improving the economic benefits of enterprises are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steel metallurgical sintering process, and particularly relates to a sintering low-temperature reduction pulverization index prediction stepwise regression method based on factor analysis full process. BACKGROUND

[0002] Iron ore sinter is a metallurgical raw material produced by sintering process from raw materials such as iron ore. It is commonly used in ironmaking and other smelting processes, which can improve production efficiency and reduce costs. The quality and characteristics of sinter have an important influence on the smelting process and the performance of the finished product.

[0003] The low-temperature reduction pulverization of sintered ore occurs during the low-temperature reduction process, and the low-temperature reduction pulverization index (RDI) is commonly used to represent it. At present, sinter is the main raw material for blast furnaces, with a ratio of more than 70-80%, and the high or low of the low-temperature reduction pulverization index of sinter has a great influence on the permeability of the material column in the blast furnace, which is related to the smooth operation of the blast furnace, and also directly affects the technical and economic indicators of blast furnace smelting. Studies have shown that for every 5% increase in the low-temperature reduction pulverization rate of sinter, the blast furnace output decreases by 1.5%, which is not conducive to the stable operation of the blast furnace. Therefore, improving the low-temperature reduction pulverization index of sinter in the blast furnace is of great significance to the blast furnace smelting. With the gradual depletion of high-quality iron ore resources, diversified economic materials make the low-temperature reduction pulverization of sinter prominent, which restricts the improvement of the technical and economic indicators of blast furnace smelting.

[0004] Ironmaking experts have done a lot of research on the low-temperature reduction pulverization index of sinter, but these studies are all biased towards certain influencing factors, such as the papers “Quantitative analysis of factors affecting the low-temperature reduction pulverization index of sinter”, “Research on factors affecting the low-temperature reduction pulverization index of sinter”, “Effect of alkalinity on the low-temperature reduction pulverization performance of low-silicon sinter in Nanjing Steel”, etc., which cannot quantitatively analyze the change trend and degree of low-temperature pulverization index.

[0005] In the sintering production process, due to the complex source of raw materials and large fluctuation of working conditions, there is strong coupling and multiple collinearity between the key process parameters affecting RDI (such as alkalinity, MgO content, sintering machine speed, layer thickness, etc.). This complexity makes the prediction model constructed directly based on the original parameters fall into the curse of dimensionality, the model coefficient estimation is distorted, the adaptability to production fluctuations is poor, and it cannot be effectively applied to online early warning and optimization control in actual production. The existing technology lacks a preprocessing method that can effectively analyze and simplify the relationship between such complex process parameters. SUMMARY

[0006] The present application is proposed to solve the above-mentioned deficiencies, and aims to provide a sintering low-temperature reduction disintegration index prediction stepwise regression method based on factor analysis whole process, which is based on the principle of optimal subset regression, combined with a stepwise search strategy, automatically screens the most critical comprehensive factor combination for RDI prediction, thereby establishing a linear regression model with high prediction accuracy and strong generalization ability, and realizing advanced and accurate prediction and intelligent control of sintering production process quality.

[0007] In order to achieve the above-mentioned purposes, the present application provides a sintering low-temperature reduction disintegration index prediction stepwise regression method based on factor analysis whole process, characterized in that it comprises the following steps: S1: data acquisition and preprocessing: collecting historical sample data in the sintering production process, wherein the historical sample data includes multiple original variables affecting the low-temperature reduction disintegration index RDI and laboratory measured RDI values corresponding to the original variables; and the multiple original variables are subjected to standardization processing; S2: factor analysis dimension reduction and comprehensive factor extraction: taking the multiple original variables subjected to standardization processing in S1 as input, performing factor analysis, extracting multiple public factors that are not correlated with each other, and establishing a linear combination relationship between the public factors and the multiple original variables; S3: constructing a prediction model based on optimal subset regression: taking the multiple public factors extracted in S2 as candidate independent variables and the laboratory measured RDI values in S1 as dependent variables, performing optimal subset regression analysis, screening an optimal public factor subset for predicting the laboratory measured RDI values, and establishing an RDI prediction model equation; S4: model verification and dynamic updating mechanism: establishing a model verification and updating mechanism combining periodic and triggered types; when the verification finds that the prediction deviation exceeds the preset range or the production process is significantly changed, triggering model updating, automatically adding the latest production data to the historical sample data, re-executing steps S2 and S3 to generate a new RDI prediction model equation, and performing safe replacement; S5: model deployment and online prediction: integrating the RDI prediction model equation established in S3 or generated in S4 into a control system; real-time collecting real-time process parameters corresponding to the multiple original variables in S1, and automatically calculating the scores of the multiple public factors according to the linear combination relationship determined in S2, substituting the scores into the RDI prediction model equation, and real-time calculating and outputting RDI prediction values.

[0008] Further, the multiple original variables in S1 include the chemical composition of sinter, sintering process operation parameters and sinter quality indexes.

[0009] Further, the factor analysis in S2 extracts common factors by principal component method, and rotates the factors by maximum variance method, so that the cumulative variance contribution rate exceeds a threshold, realizing dimension reduction of common factors.

[0010] Further, the principle of extracting the plurality of common factors in S2 is that the eigenvalue is greater than 1.

[0011] Further, the optimal subset regression analysis in S3 includes: traversing all possible combinations from 1 to the maximum number of the plurality of common factors, and fitting a multiple linear regression model for each independent variable subset.

[0012] Further, the optimal common factor subset is screened out in S3 by comparing all models, and the adjusted R 2 The maximum model is selected.

[0013] Further, the model verification and updating mechanism in S4 includes: Periodic verification, the system automatically summarizes the predicted value and the corresponding laboratory measured RDI true value, calculates the average deviation size and hit rate; Deviation judgment, set the allowed deviation range, when the verification result continuously exceeds the allowed deviation range for many times, judge that the model prediction is inaccurate, trigger model updating.

[0014] Further, the trigger type combined mechanism in S4 also includes: when the production process is changed significantly, manually trigger model updating.

[0015] Further, the safety replacement in S4 includes: after the new RDI prediction model equation is generated, first test it with historical data, confirm that the new model predicts more accurately than the old model, and then formally replace the old RDI prediction model equation for online prediction.

[0016] Further, the S5 automatically calculates the scores of the plurality of common factors according to the linear combination relationship determined in S2, uses the factor loading matrix obtained by the factor analysis in S2, substitutes the real-time process parameters collected in real time, and calculates the scores of the plurality of common factors.

[0017] Compared with the prior art, the present application has the following beneficial effects: Firstly, the method is based on factor dimension reduction and feature reconstruction of process mechanism: for the high correlation among the 18 original parameters affecting RDI, the method innovatively uses factor analysis technology. The innovation is not simply reducing the number of variables, but deeply mining the potential process driving factors behind the parameters (such as "production process index factor", "sinter quality factor", "sintering ore-forming mechanism factor", etc.). Through analyzing the factor loading matrix, the process rules hidden behind the massive operation data are clearly revealed, the feature extraction and reconstruction of the complex process system are realized, and pure and efficient explanatory variables are provided for subsequent modeling.

[0018] Secondly, the method is based on the robust model construction strategy of best subset search: on the basis of dimension reduction, the method discards the traditional "trial and error" or "human intervention" variable selection method, and innovatively uses the BestSubsetRegression algorithm to systematically traverse all possible factor combinations. Through adjusting the adjusted R 2 2, Mallows Cp criterion and other statistical quantities as objective evaluation criteria, the optimal prediction variable combination and the corresponding regression equation are automatically and efficiently determined. This method ensures that the model is the globally optimal linear model under the given data conditions, significantly improving the prediction accuracy and reliability of the model.

[0019] Thirdly, the method can solve the problem of poor model stability and low prediction accuracy caused by the large number of independent variables and serious multicollinearity when directly predicting the sinter low-temperature reduction pulverization rate (RDI +3.15 ) based on multiple linear regression in the prior art. An advanced data preprocessing and feature extraction technology is provided: through factor analysis method, the 18 original variables representing the chemical composition of sinter, sintering process operation parameters and sinter quality indicators are converted into fewer (13) comprehensive factors that are not correlated and have clear process significance, which fundamentally eliminates the multicollinearity problem and lays a foundation for building a high-precision prediction model.

[0020] Fourthly, the method establishes an efficient and robust model construction method, based on the BestSubsetRegression principle and combined with the step-by-step search strategy, automatically selects the most critical comprehensive factor combination for RDI prediction, and establishes a linear regression model with high prediction accuracy and strong generalization ability, overcoming the limitations of traditional stepwise regression or subjective variable selection method.

[0021] Fifthly, the present application realizes the advanced accurate prediction and intelligent regulation of the sintering production process quality, integrates the finally established prediction model into the sintering process control system, realizes the real-time online prediction of the RDI index, provides the advanced and accurate data support for the operators to adjust the key parameters such as the basicity and fuel ratio, and finally achieves the purposes of stabilizing the sinter quality, reducing the blast furnace fuel ratio and improving the economic benefits of enterprises. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 A process schematic diagram of the sintering low-temperature reduction pulverization index prediction stepwise regression method based on the factor analysis whole process of the embodiments of the present application. DETAILED DESCRIPTION

[0023] The implementation of the present application will be described in detail below in combination with implementation cases, but they do not constitute the limitation of the present application, but only serve as examples. Meanwhile, the advantages of the present application will become more clear and easy to understand through the description.

[0024] A sintering low-temperature reduction pulverization index prediction stepwise regression method based on the factor analysis whole process of the present application, comprising the following steps: S1: data acquisition and preprocessing: collecting the historical sample data in the sintering production process, wherein the historical sample data comprises a plurality of original variables affecting the low-temperature reduction pulverization index RDI and the laboratory measured RDI values corresponding to the original variables; and the plurality of original variables are subjected to standardization processing; S2: factor analysis dimension reduction and comprehensive factor extraction: taking the plurality of original variables subjected to the standardization processing in S1 as the input, performing factor analysis, extracting a plurality of public factors irrelevant to each other, and establishing the linear combination relationship between the public factors and the plurality of original variables; S3: constructing a prediction model based on the best subset regression: taking the plurality of public factors extracted in S2 as the candidate independent variables, and taking the laboratory measured RDI values in S1 as the dependent variables, performing the best subset regression analysis, screening out the optimal public factor subset for predicting the laboratory measured RDI values, and establishing the RDI prediction model equation; S4: model verification and dynamic updating mechanism: establishing a model verification and updating mechanism combining the periodic and triggered types; when the verification finds that the prediction deviation exceeds the preset range or the production process is significantly changed, triggering the model updating, automatically adding the latest production data into the historical sample data, re-executing steps S2 and S3 to generate a new RDI prediction model equation, and performing the safe replacement; S5: Model deployment and online prediction: integrate the RDI prediction model equation established in S3 or generated in S4 into the control system; collect real-time process parameters corresponding to the multiple original variables in S1 in real time, and automatically calculate the scores of the multiple common factors according to the linear combination relationship determined in S2, and then substitute the scores of the multiple common factors into the RDI prediction model equation, to calculate and output the RDI prediction value in real time.

[0025] In some embodiments, the multiple original variables in S1 include chemical composition of sintered ore, sintering process operation parameters, and sintered ore quality indicators.

[0026] In some embodiments, the factor analysis in S2 extracts common factors by principal component method, and performs factor rotation by maximum variance method.

[0027] Specifically, the principle of extracting the multiple common factors in S2 is that the eigenvalue is greater than 1.

[0028] In some embodiments, the optimal subset regression analysis in S3 includes traversing all possible combinations from 1 to the maximum number of the multiple common factors, and fitting a multiple linear regression model for each subset of independent variables.

[0029] Specifically, the optimal common factor subset is screened out in S3 by comparing all models, and the model with the smallest adjusted R 2 maximum or Cp value closest to p+1.

[0030] In some embodiments, the model verification and updating mechanism in S4 includes: Periodic verification: the system automatically summarizes the prediction value and the corresponding laboratory measured RDI true value, calculates the average deviation size and hit rate; Deviation judgment: set the allowed deviation range, when the verification result continuously exceeds the allowed deviation range for multiple times, judge that the model prediction is inaccurate, and trigger model updating.

[0031] Specifically, the trigger combination mechanism in S4 further includes: when the production process is significantly changed, manually trigger model updating.

[0032] Specifically, the safe replacement in S4 includes: after the new RDI prediction model equation is generated, first test it with historical data, confirm that the new model predicts more accurately than the old model, and then formally replace the old RDI prediction model equation for online prediction.

[0033] In some embodiments, the scores of the plurality of common factors are automatically calculated in S5 according to the linear combination relationship determined in S2, the factor loading matrix obtained by factor analysis in S2 is used, and the real-time process parameters collected in real time are substituted to calculate the scores of the plurality of common factors.

[0034] Embodiments: The sinter low-temperature reduction pulverization index prediction stepwise regression method based on the whole process of factor analysis in the embodiment includes the following steps: S1: data acquisition and preprocessing Sample data in a historical time period in the sintering production process are collected, each sample including: a) chemical composition of sinter (such as TFe, SiO2, CaO, MgO, Al2O3, basicity R, etc.); b) sintering process operation parameters (such as bed depth, etc.); c) sinter quality indexes (such as drum strength, sinter particle size, etc.); d) laboratory measured RDI value of the sinter batch corresponding to the sample. The collected original data are subjected to cleaning, removal of abnormal values, and standardization processing.

[0035] S2: dimensionality reduction and comprehensive factor extraction by factor analysis The above 18 kinds of influencing factors after preprocessing are taken as initial variables, factor analysis (common factors are extracted by principal component method, and factor rotation is performed by maximum variance method) is performed, and 13 main common factors (F1, F2,..., F13) are extracted according to the principle that the characteristic value is greater than 1. Each common factor is a linear combination of original variables (Fi=Wi0+Wi1*X1+Wi2*X2+...+Wi18*X18), wherein X1, X2, X3...X18 are original variables; Wi0 is a constant, and Wi1, Wi2, Wi3...Wi18 are variable coefficients of original variables X1, X2, X3...X18, the factor loading matrix thereof reveals the contribution degree of each original variable to the common factor, thereby giving each common factor a clear process meaning.

[0036] S3: constructing a prediction model based on the best subset regression The 13 common factors extracted are taken as candidate independent variables, and RDI is taken as the dependent variable to perform best subset regression analysis.

[0037] All possible combinations of independent variables from 1 to 13 are traversed.

[0038] For each independent variable subset, a multiple linear regression model is fitted, and evaluation indexes such as adjusted R 2 , Cp value, etc. are calculated.

[0039] All models are compared, and the adjusted R 2The model with the largest (or Cp value closest to p+1) is the optimal model. Assuming the optimal model contains k factors (k≤13), the final RDI prediction model equation is: RDI +3.15_pred = β0+ β1*F1+ β2*F2+... + βk*Fk where RDI +3.15_pred is the mass percentage of particles larger than 3.15 mm after a specific standard simulation reduction experiment; β0 is a constant, F1, F2, F3...Fk are candidate independent variables; β1, β2, β3...βk are variable coefficients of candidate independent variables F1, F2, F3...Fk.

[0040] S4: Model verification and dynamic updating mechanism Establish a model verification and updating mechanism combining regular and triggered methods to ensure the long-term effectiveness of the prediction model.

[0041] Regular verification: The system automatically aggregates all prediction values and their corresponding RDI true values verified by the laboratory within a period of time every week / month. The average deviation between the prediction values and the true values within the period is calculated, as well as the proportion of prediction values within the allowed error range (hit rate).

[0042] Deviation judgment: Set an allowed deviation range (e.g., the proportion of prediction values deviating from the true values by more than 5% cannot exceed 5%). If the verification result exceeds this range continuously for multiple times, it is judged that the model prediction is inaccurate and needs to be updated.

[0043] Triggered update: Once it is found that the model prediction is inaccurate, the system automatically alarms and prompts that the model needs to be updated. At the same time, when there are major changes in production process (such as significant adjustment of raw material structure, equipment overhaul, etc.), manual triggering of update can also be performed.

[0044] Model update: When updating, the system automatically adds a large amount of new production data in the recent period (e.g., three months) to the training set, re-performs factor analysis and regression calculation, and generates a new prediction model. This process is equivalent to allowing the model to "re-learn" the latest production rules.

[0045] Safe replacement: After the new model is generated, it is first tested with past data to confirm that the new model is more accurate than the old model before replacing the old model for online prediction, which can prevent the model from deteriorating and ensure production safety.

[0046] S5: Model deployment and online prediction The finalized regression model equation is integrated into the intelligent ironmaking platform. The system collects the original process parameters described in S1 in real time, automatically calculates the scores of 13 comprehensive factors according to the weight coefficients determined in S2, and then substitutes the factor scores into the regression equation obtained in S3 to calculate and output the predicted RDI value under the current production status in real time.

[0047] Maanshan Iron and Steel 300-380m 2 Taking the 2024 production data of the sintering machine as an example, the implementation method of the present invention will be explained in detail.

[0048] 1. Data preparation: More than 300 sets of valid sample data were collected, each set containing 18 process parameters and 1 measured RDI value.

[0049] 2. Factor Analysis: Factor analysis was performed using SPSSPRO software. The KMO test value was 0.78 > 0.6, and the Bartlett's test of sphericity showed a significance level of p < 0.001, indicating that factor analysis was suitable. Thirteen common factors were extracted, with a cumulative variance contribution rate of 85.7%. For example, factor F1 had high loadings on CaO and alkalinity and could be named the "Alkalinity Characteristic Factor"; factor F2 had high loadings on ignition temperature and could be named the "Hot State Intensity Factor".

[0050] 3. Model Construction: Optimal subset regression was performed using the `leaps` package in R. Analysis results show that when 11 common factors are included, the adjusted R... 2 The highest value was 0.71, and the lowest Cp value was obtained. The final regression equation is as follows: y = -18.199 + 5.413 * FeO + 4.937 * SiO2 + 28.045 * R - 18.92 * MgO / Al2O3 + 39.88 * Al2O3 / SiO2 - 147.344 * V2O5 - 69.721 * P - 91.709 * (K2O + Na2O) - 0.405 * Drum coefficient + 0.342 * LD_5 - 10(%) - 24.928 * S(%) 4. Model Validation and Update: The system automatically checks the previous week's forecasts weekly. For example, in May 2024, due to a change in the main iron ore powder type, resulting in raw material changes, the system found that the predicted values ​​for two consecutive weeks deviated significantly from the laboratory measured values, with 10 out of 15 batches exceeding our set allowable range of 5%. The system immediately triggered an alarm and automatically started the update process. The system collected new data from the past three months, recalculated, "learned" the new raw material patterns, and generated a new model. After testing, the new model's prediction accuracy returned to a high level, and it was officially replaced by the old model, ensuring the continued accuracy of the predictions.

[0051] 5. Application deployment: write the final equation into the company's intelligent ironmaking system. The system automatically reads real-time process data every 10 minutes, calculates the factor score and completes the RDI prediction, and displays the prediction results on the sintering control room operation picture to guide production.

[0052] In addition, through the actual application verification on the sintering machine of a certain company, the present application has achieved remarkable technical and economic effects: High prediction accuracy. The average absolute error (MAE) of the prediction value of the prediction model for RDI and the actual test value is stable within ±5%, the prediction hit rate (absolute error ≤5%) is more than 95%, and it fully meets the requirements of industrial production for prediction accuracy.

[0053] Fast response. The present application realizes the hour-level or even minute-level advanced prediction of the RDI index, compared with the current situation that the traditional laboratory test results lag behind for 8-12 hours, which wins valuable time for production adjustment.

[0054] Significant guidance effect. Based on the prediction results, the operator can adjust the key parameters such as alkalinity and fuel ratio in advance, so that the RDI qualified rate is improved by about 15%, which effectively promotes the stable and smooth operation of the blast furnace.

[0055] Obvious economic benefits. Due to the stable quality of sinter and the improvement of blast furnace indicators, it is expected to reduce the fuel cost by hundreds of millions of yuan per year, and the economic benefits are significant.

[0056] The above is only a specific embodiment of the present application. It should be noted that any changes or replacements that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered within the protection scope of the present application. The remaining prior art not described in detail is the prior art.

Claims

1. A stepwise regression method for predicting sintering low-temperature reduction pulverization indices based on a full-process factor analysis, characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing: Collect historical sample data during the sintering production process. The historical sample data includes various original variables that affect the Low Temperature Reduction Pulverization Index (RDI), as well as the laboratory measured RDI values ​​corresponding to the original variables; and standardize the various original variables. S2: Factor analysis dimensionality reduction and comprehensive factor extraction: Using the standardized original variables from S1 as input, factor analysis is performed to extract multiple uncorrelated common factors and establish a linear combination relationship between the common factors and the original variables. S3: Construct a prediction model based on optimal subset regression: Using the multiple common factors extracted in S2 as candidate independent variables and the laboratory measured RDI value in S1 as the dependent variable, perform optimal subset regression analysis, screen out the optimal common factor subset for predicting the laboratory measured RDI value, and establish the RDI prediction model equation. S4: Model Validation and Dynamic Update Mechanism: Establish a model validation and update mechanism that combines periodic and trigger-based methods; when the validation finds that the prediction deviation exceeds the preset range, or when there is a major change in the production process, the model is updated, the latest production data is automatically added to the historical sample data, steps S2 and S3 are re-executed, a new RDI prediction model equation is generated, and a safe replacement is performed. S5: Model Deployment and Online Prediction: Integrate the RDI prediction model equations established in S3 or generated in S4 into the control system; Real-time process parameters corresponding to the various original variables mentioned in S1 are collected in real time, and the scores of the multiple common factors are automatically calculated according to the linear combination relationship determined in S2. The scores are substituted into the RDI prediction model equation, and the RDI prediction value is calculated and output in real time.

2. The method according to claim 1, characterized in that, The various original variables in S1 include: the chemical composition of the sinter, the operating parameters of the sintering process, and the quality indicators of the sinter.

3. The method according to claim 1, characterized in that, The factor analysis in S2 uses principal component analysis to extract common factors and the maximum variance method to perform factor rotation.

4. The method according to claim 1 or 3, characterized in that, The principle for extracting the multiple common factors in S2 is based on the eigenvalue being greater than 1.

5. The method according to claim 1, characterized in that, The optimal subset regression analysis in S3 includes: traversing all possible combinations from 1 to the maximum number of common factors, and fitting a multiple linear regression model for each subset of independent variables.

6. The method according to claim 5, characterized in that, The optimal common factor subset is selected in S3, and the adjusted R is selected by comparing all models. 2 The largest model.

7. The method according to claim 1, characterized in that, The model verification and update mechanism in S4 includes: Regular verification is performed, and the system automatically summarizes the predicted values ​​and their corresponding laboratory measured RDI values, and calculates the average deviation and hit rate. Deviation judgment: Set an allowable deviation range. When the verification result exceeds the allowable deviation range multiple times in a row, the model prediction is judged to be inaccurate, and the model is triggered to update.

8. The method according to claim 1 or 7, characterized in that, The trigger-based combined mechanism in S4 also includes: manually triggering a model update when a major change occurs in the production process.

9. The method according to claim 1, characterized in that, The safe replacement in S4 includes: after the new RDI prediction model equation is generated, it is first tested with historical data to confirm that the new model predicts more accurately than the old model, and then the old RDI prediction model equation is officially replaced for online prediction.

10. The method according to claim 1, characterized in that, In step S5, the scores of the multiple common factors are automatically calculated according to the linear combination relationship determined in step S2. The factor loading matrix obtained from factor analysis in step S2 is used to substitute the real-time process parameters collected in real time to calculate the scores of the multiple common factors.

Images