Converter end point carbon content prediction method based on wavelet neural network

Abstract

The invention relates to a converter end point carbon content prediction method based on a wavelet neural network, which comprises the following steps: firstly, carrying out data preprocessing on original data, sequentially finishing blank value elimination, abnormal value screening and normalization processing, then screening out weak correlation process parameters through Pearson's correlation coefficient analysis in combination with a metallurgy mechanism, and finally, carrying out prediction on the end point carbon content of a converter. Subsequently, dimensionality reduction is carried out through principal component analysis, and finally an optimized input parameter set is formed to replace original full-amount parameters. After the optimization parameter set is substituted into the wavelet neural network model, the model complexity is greatly reduced, and the training efficiency is improved. And finally, inputting production data into the wavelet neural network prediction model to obtain an end point carbon content prediction result. According to the method, modeling is carried out based on the wavelet neural network, deep learning driving replaces traditional mechanism driving, complex reactions and nonlinear relations in automatic learning data can be achieved, and the prediction precision of the converter end point carbon content is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a converter endpoint carbon content prediction method based on wavelet neural network, belonging to the technical field of converter steelmaking, especially to the intelligent prediction method of converter steelmaking endpoint carbon content. BACKGROUND

[0002] Converter steelmaking is the core link of steel production, and the endpoint carbon content as a key quality index of steelmaking process, it plays a key role in ensuring the quality of molten steel, improving production efficiency, reducing production cost and ensuring safety production. Precise prediction of endpoint carbon content can effectively reduce the number of supplementary blowing, reduce oxygen and auxiliary material consumption, shorten the smelting cycle, and avoid the risk of molten steel scrap due to excessive carbon content, which plays an irreplaceable role in achieving high quality, high efficiency, low cost and green safety goals of steel production.

[0003] However, there are a large number of complex chemical reactions and dynamic conditions in the process of converter steelmaking, such as decarburization, dephosphorization, temperature rise and other multi-coupling physical and chemical reactions in the furnace, and the reaction rate is dynamically affected by many factors such as molten iron composition, scrap ratio, oxygen blowing intensity, slag making system, etc., so it is difficult to accurately predict the endpoint carbon content.

[0004] The current converter endpoint carbon content prediction method of small and medium-sized steel enterprises mainly establishes a mechanism model by studying the material balance and heat balance in the converter process. For example, CN202411695219.8 discloses a converter steelmaking endpoint carbon temperature prediction method based on adaptive data enhancement. The technical method preprocesses the data by using adaptive SMOTE data enhancement technology, uses random forest to build a steelmaking endpoint carbon temperature prediction model, and obtains the converter steelmaking endpoint carbon temperature prediction result. CN202411110990.4 discloses a converter steelmaking endpoint carbon temperature real-time online prediction method based on instant learning. The technical method locally models by using similar historical heats as a local sample set, uses a multiple linear regression algorithm to solve in reverse, and obtains the endpoint carbon content prediction result after substituting into the prediction model. Careful analysis of the above technology shows that these process methods have the following shortcomings: (1) the mechanism model cannot accurately describe the complex reaction process and interaction in the converter, such as the complex mass transfer and heat transfer process in the furnace, which is difficult to fully and accurately quantify the indicators, and only a large number of assumptions can be made; some parameters are difficult to accurately measure in time, such as the time lag problem in the detection of molten steel composition information, which cannot capture the dynamic changes of the composition in the blowing process in time, resulting in large errors in the model input data and affecting the prediction accuracy of the endpoint carbon content. (2) The accuracy of the process parameters is extremely high. If the input process parameters (such as molten iron composition, oxygen blowing flow, etc.) are not accurate, the prediction error will be large; when the process conditions change during production, the model is difficult to automatically adapt, further reducing the prediction accuracy of the model, and professional maintenance personnel need to re-adjust and optimize the model parameters.

[0005] Therefore, developing a converter endpoint carbon content prediction method that can balance prediction accuracy, generalization ability, and process adaptability, and is suitable for steel plants of different production scales, has become a technical problem to be solved in the process of intelligent upgrading of the steel industry. SUMMARY

[0006] The present application provides a converter endpoint carbon content prediction method based on a wavelet neural network. Based on wavelet neural network modeling, deep learning is used to replace traditional mechanism driving, which can automatically learn the complex reactions and nonlinear relationships in the data, and improve the prediction accuracy of the converter endpoint carbon content.

[0007] The technical solution adopted by the present application to solve its technical problems is: A converter endpoint carbon content prediction method based on a wavelet neural network, specifically comprising the following steps: Step S1, collect data, collect process parameter information of historical heats; Step S2, the process parameters in step S1 are pretreated, first, the blank values in the process parameters are removed, then the abnormal values are removed through normal distribution, the process parameters beyond the fluctuation range are removed by combining the preset production range, and finally the process parameters are normalized; Step S3, Pearson correlation coefficient analysis is performed on the pretreated process parameters to determine the correlation size of different process parameters to the converter endpoint carbon content; Step S4, the correlation size of different process parameters obtained by step S3 Pearson correlation coefficient analysis is sorted, and the significant correlation critical value of the correlation is obtained considering the number of collected process parameters; Step S5, the process parameters are screened, the characteristic data with the highest correlation to the endpoint carbon content are retained, and the process parameters with low correlation are subjected to principal component analysis to extract the process parameter characteristic values with small correlation factors to obtain a new process parameter data set; Wherein, the screening condition of low correlation is that the absolute value of Pearson correlation coefficient in step S3 is less than the significant correlation critical value in step S4; Step S6, a wavelet neural network model is established, the new process parameter data set in step S5 is divided into training set and test set according to the proportion, and is substituted into the wavelet neural network model for optimization training to obtain the optimal wavelet neural network model; Step S7, the optimized wavelet neural network model in step S6 is installed in the converter main control room, the process parameter information generated in the converter smelting process is read in real time, and the prediction result of the converter endpoint carbon content is obtained through the wavelet neural network model; Step S8, the prediction result obtained in step S7 is combined with the phosphorus content in the first sample molten steel to determine whether the molten steel meets the preset tapping condition, if the preset tapping condition is met, the molten steel is directly tapped without waiting for the second sample molten steel detection time; Further, in step S1, the collected process parameters include the composition of molten iron, molten iron temperature, molten iron weight, scrap steel weight, oxygen blowing amount, oxygen blowing time, auxiliary material addition amount, and converter endpoint carbon content; Wherein, the composition of molten iron includes the carbon content, phosphorus content, sulfur content, silicon content and manganese content of molten iron after pretreatment by KR desulfurization station, the molten iron temperature includes the molten iron temperature after desulfurization, and the auxiliary material addition amount includes the addition amount of lime, pellets, magnesium ball, iron oxide ball and light burned dolomite; Further, the specific steps of step S2 for pretreating the process parameters in step S1 are as follows: Step S21, remove the blank values in the process parameters; Step S22, check whether the process parameters after removing the blank values obey normal distribution, if yes, go to step S23 directly, if not, determine that this process parameter is an abnormal value, remove it and go to step S23; Step S23, in combination with the actual working condition, the production range of the preset process parameters is determined as an abnormal value when the fluctuation range exceeds the production range, and the abnormal value is removed; wherein, the preset process parameter production range includes: molten iron temperature range: 1320 ~ 1420℃, molten iron weight range: 155 ~170t, oxygen blowing time: 12 ~17min, lime, light-burned dolomite, magnesium ball addition amount: not 0, converter endpoint carbon content range: 0.02 ~ 0.1%; Step S24, the process parameters screened in step S23 are normalized, and the processing formula is: (1) In formula (1), is the normalized data corresponding to the original process parameter, is the original process parameter, is the maximum value of the normalized original process parameter, is the minimum value of the normalized original process parameter, is the maximum value of the original process parameter, is the minimum value of the data of the original process parameter; Further, in step S3, the formula used for Pearson correlation coefficient analysis on the preprocessed process parameters is: (2) In formula (2), r is the Pearson correlation coefficient, the value range is [-1, 1]; is the normalized process parameter; is the converter endpoint carbon content; n is the number of normalized process parameters; is the mean value of the normalized process parameter, is the mean value of the converter endpoint carbon content; Further, in step S4, the correlation significant association critical value calculation formula is: (3) In formula (3), q is the significant association critical value; n is the number of normalized process parameters; t is the t distribution obtained by inputting the normalized process parameter; α is 0.05; Further, in step S5, the specific steps of obtaining the new process parameter data set are: Step S51, based on screening of process parameters, process parameters with low correlation are obtained; Step S52, the process parameters screened in step S51 are centralized, and the processing formula is: (4) (5) In formula (4) and formula (5), is the mean value of the i th process parameter; j x ij is the value in the i th row and the j th column; i j is the number of normalized process parameters; n j is the number of process parameters; X cent is the centralized data; is a matrix composed of the mean value of each normalized process parameter; X is a matrix of n × p ; Step S53, the covariance matrix is calculated, and the calculation formula is: (6) In formula (6), S is a symmetric matrix of p × p ; Step S54, the eigenvalue and eigenvector of the covariance matrix are calculated, and the calculation formula is: (7) In formula (7), is the eigenvalue after the eigenvalue decomposition of the covariance matrix S; is the eigenvector after the eigenvalue decomposition of the covariance matrix S; Step S55, the data is reduced in dimension, and the processing formula is: (8) In formula (8), Y is the reduced process parameter; V is a matrix composed of the eigenvector corresponding to the maximum eigenvalue; Further, in step S6, the specific steps of obtaining the optimal wavelet neural network model are: Step S61, the input layer, the hidden layer and the output layer of the model are constructed, wherein the number of input process parameters determines the input layer, the output layer is the final carbon content of the converter, and the number of hidden layers is obtained by multiple actual simulation results The optimal value is calculated by the formula (9) In formula (9), h is the number of hidden layer nodes;​​​a is a constant between [0, 10]; m is the number of input layer neurons; n is the number of output layer neurons; Step S62, the new process parameter data in step S5 is proportionally divided into a training set and a test set, which is substituted into the wavelet neural network model constructed in step S61 to perform optimization training, forming optimal model parameters to obtain an optimal wavelet neural network model.

[0008] Through the above technical solution, compared with the prior art, the present application has the following beneficial effects: 1. The converter endpoint carbon content prediction method based on the wavelet neural network provided by the present application performs data preprocessing on the original data, sequentially completes blank value elimination, abnormal value screening (in combination with a metallurgical process preset range) and normalization processing, and ensures the effectiveness and consistency of the input data; then the Pearson correlation coefficient analysis is combined with the metallurgical mechanism to screen out weakly correlated process parameters, which are subsequently reduced in dimension through principal component analysis, and finally an optimized input parameter set is formed to replace the original full parameter set, effectively increasing the robustness of the data, reducing data deviation and noise interference, laying a solid foundation for subsequent use of the wavelet neural network model, not only greatly reducing the model complexity and improving the training efficiency, but also fully exerting the core advantage of the wavelet neural network that it can approximate any continuous function with arbitrary precision; 2. The converter endpoint carbon content prediction method based on the wavelet neural network provided by the present application fully utilizes the strong self-learning ability of the wavelet neural network, and when facing complex and variable data, automatically adjusts the weight and threshold of the network according to the characteristics of the input data and the training target, and improves the fitting and prediction ability of the data. BRIEF DESCRIPTION OF DRAWINGS

[0009] The present application will be further described below in combination with the drawings and examples.

[0010] Figure 1 is a wavelet neural network model structure diagram in the preferred embodiment provided by the present application. DETAILED DESCRIPTION

[0011] The present application will be further described below in combination with the drawings and examples. The specific dimensions used in the description of the present application are only for the purpose of illustrating the technical solutions and do not limit the protection scope of the present application.

[0012] As set forth in the background, the prediction techniques in the prior art about the converter endpoint carbon content are mostly predicted by solving complex equation sets, but these methods based on traditional machine learning have limited fitting ability for the non-linear and strongly coupled smelting process, and cannot meet the precise control requirements of complex working conditions. Some updated techniques are to filter similar historical heats to construct a local sample set, and then use multivariate linear regression algorithm to solve the prediction model inversely, which avoids the limitations of offline models, but when the converter process changes greatly, the reference value of the historical similar sample will be greatly reduced, resulting in that the model cannot achieve effective prediction.

[0013] To avoid the above problems, the present application provides a converter endpoint carbon content prediction method based on wavelet neural network. As the biggest innovation point of the present application, the prediction of the converter endpoint carbon content adopts a wavelet neural network, which can automatically adjust the weights and thresholds of the network according to the characteristics of the input data and the training target. When facing complex and variable data, it can adaptively optimize its structure and parameters to improve the fitting and prediction ability of the data.

[0014] In order to more accurately fit the complex function relationship and improve the hit rate of the model, the data input to the wavelet neural network needs to be optimized, so another innovation point of the present application is to preprocess the production data to form an optimized input parameter set to replace the original full parameter set.

[0015] Next, the prediction method is described, which specifically includes the following steps: Step S1, collecting data, collecting process parameter information of historical heats; traditional partial prediction methods have one-sidedness in data collection, such as focusing only on a few parameters such as molten iron weight and oxygen blowing amount, while ignoring the indirect influence of auxiliary material addition amount, scrap steel weight, etc. on the temperature and reaction in the furnace. Therefore, in this step, the collected process parameters include the composition of the molten iron, the temperature of the molten iron, the weight of the molten iron, the weight of the scrap steel, the oxygen blowing amount, the oxygen blowing time, the auxiliary material addition amount, and the converter endpoint carbon content; wherein the composition of the molten iron includes the carbon content, phosphorus content, sulfur content, silicon content and manganese content of the molten iron after pretreatment by the KR desulfurization station, the temperature of the molten iron after desulfurization, and the addition amount of lime, pellets, magnesium balls, iron oxide balls and light-burned dolomite.

[0016] Collecting multi-dimensional key parameters avoids prediction deviation caused by data missing, and provides a solid foundation for the wavelet neural network to learn the carbon content variation law under different process conditions more comprehensively.

[0017] Step S2, preprocessing the process parameters of step S1, the specific steps are: Step S21, eliminating blank values in the process parameters; Step S22, check whether the process parameters of blank value elimination obey normal distribution, if obey directly into step S23, if not obeyed is determined that this process parameter is an abnormal value, after elimination into step S23; Step S23, combining the actual working condition preset process parameter production range, the fluctuation range beyond the production range of process parameters is determined as an abnormal value, continue to remove; wherein, the preset process parameter production range includes: hot metal temperature range: 1320 ~ 1420℃, hot metal weight range: 155 ~170t, oxygen blowing time: 12 ~17min, lime, light calcined dolomite, magnesium ball addition: not 0, converter endpoint carbon content range: 0.02 ~ 0.1%; Step S24, the process parameters screened in step S23 are normalized, and the processing formula is: (1) In formula (1), is the normalized data corresponding to the original process parameter, is the original process parameter, is the maximum value of the normalized original process parameter, is the minimum value of the normalized original process parameter, is the maximum value of the original process parameter, is the minimum value of the original process parameter.

[0018] Step S2 first eliminates the blank value, then removes the abnormal value through normal distribution, then removes the abnormal value from the collected data combined with the actual production data, and finally normalizes the data. By removing the abnormal value through normal distribution, and then removing the abnormal value from the collected data combined with the actual production data, the invalid data interference caused by equipment failure and manual recording error can be avoided. As for normalization, different process parameters often have different dimensions and dimension units, such as the range of hot metal temperature is 1320-1420℃, but the range of Si content in hot metal is only 0.15-0.6%, if the original data is directly brought into the prediction model, it will directly affect the prediction result of the model; In order to eliminate the influence of the dimension between the indicators, it is necessary to normalize the data, so that the process parameter data is in the same order of magnitude, and then the data is substituted into the prediction model. The normalization processing is placed in the later period of pre-processing, because the normalization processing depends on the extreme value (maximum value, minimum value) or statistical characteristics (mean value, standard deviation) of the data, and the abnormal value will greatly deviate these parameters. For example, if the abnormal value of hot metal temperature is 1000℃, the normal 1320-1420℃ data will be compressed to a very small range after normalization, losing the original distribution difference, and the subsequent cannot reflect the real process fluctuation, so the abnormal value needs to be removed first, and then the remaining data is normalized.

[0019] Step S2 is to standardize the collected data, solve the problem of traditional data processing roughness, and improve the stability of the whole prediction process. After preliminary preprocessing, the correlation of the quantitative parameters needs to be quantified to quickly locate the significant core parameters that affect the carbon content. The traditional prediction technology relies on the metallurgical experience of engineers to screen parameters, and this subjective screening is easy to cause data omission. Therefore, step S3 performs Pearson correlation coefficient analysis on the preprocessed process parameters to determine the correlation size of different process parameters to the converter endpoint carbon content; through objective quantitative analysis, the importance of parameters is determined more accurately than experience, and a scientific screening standard is provided for the subsequent dimension reduction step, reducing the interference of invalid parameters on the model.

[0020] The formula used for Pearson correlation coefficient analysis is: (2) In formula (2), r is the Pearson correlation coefficient, whose value range is [-1, 1]; is the normalized process parameter; is the converter endpoint carbon content; n is the number of normalized process parameters; is the mean value of the normalized process parameter, is the mean value of the converter endpoint carbon content.

[0021] After calculating the Pearson correlation coefficient, step S4 sorts the correlation size of different process parameters obtained by step S3 Pearson correlation coefficient analysis, considers the number of collected process parameters, and obtains a significant correlation critical value; step S5, the process parameters are screened, and the characteristic data with the highest correlation to the endpoint carbon content is retained. The process parameters with low correlation are subjected to principal component analysis, and the characteristic values of the process parameters with small correlation factors are extracted to obtain a new process parameter data set; wherein, the screening condition of low correlation is that the absolute value of the Pearson correlation coefficient in step S3 is < the significant correlation critical value in step S4.

[0022] In step S4, the significant correlation critical value calculation formula is: (3) In formula (3), q is the significant correlation critical value; n is the number of normalized process parameters; t is the distribution obtained by inputting the normalized process parameters; t α is 0.05.

[0023] After calculating the significant correlation critical value of the correlation coefficient through statistical significance, step S5 is entered, and the steps are as follows: step S51, based on The process parameters were screened to obtain those with low correlation. Step S52: The process parameters with low correlation after screening in step S51 are centered to eliminate the interference of the mean on the data analysis and ensure the accuracy of covariance calculation and principal component orientation. The centered formula is as follows: (4) (5) In formulas (4) and (5), For the first j The average value of each process parameter; x ij For the first i Line 1 j The value of the column; n The number of process parameters after normalization; j The number of process parameters; X cent For centralized data; A matrix consisting of the mean values ​​of each normalized process parameter; X for n × p Matrix; Step S53, calculate the covariance matrix. The calculation formula is as follows: (6) In formula (6), S is p × p symmetric matrix; Step S54: Calculate the eigenvalues ​​and eigenvectors of the covariance matrix using the following formula: (7) In formula (7), These are the eigenvalues ​​of the covariance matrix S after eigendecomposition; These are the eigenvectors of the covariance matrix S after eigendecomposition. Step S55: Perform dimensionality reduction on the data using the following formula: (8) In formula (8), Y represents the process parameters after dimensionality reduction; V It is a matrix consisting of the eigenvectors corresponding to the largest eigenvalue.

[0024] Step S5, the process parameters with a Pearson correlation coefficient absolute value less than the threshold value (i.e. no statistically significant linear correlation) are screened out in combination with the metallurgical mechanism to obtain process parameters with low correlation, the process parameters are reduced in dimension, the data before dimension reduction is replaced, and finally an optimized input parameter set is formed to replace the original full-quantity parameters. Both the model training speed and the scientificity of prediction are improved, and the balance between dimension reduction and information retention is achieved.

[0025] Step S6, a wavelet neural network model is established, first, an input layer, a hidden layer and an output layer of the model are constructed, the input layer is determined by the number of input process parameters, the output layer is the converter endpoint carbon content, the number of hidden layers is obtained by multiple actual simulation results, and the calculation formula is (9) In formula (9), h is the number of hidden layer nodes; a is a constant between 0 and 10; m is the number of input layer neurons; n is the number of output layer neurons; Secondly, the new process parameter data in step S5 is divided into a training set and a test set according to a proportion, is substituted into the wavelet neural network model for optimization training, and the optimal model parameters are formed to obtain the optimal wavelet neural network model. Compared with the traditional neural network, the wavelet neural network has stronger fitting ability and higher calculation precision in function approximation, and finally the prediction precision of the converter endpoint carbon content is effectively improved.

[0026] Step S7, the wavelet neural network model optimized in step S6 is installed in the converter main control room, real-time process parameter information generated in the converter smelting process is read, and the prediction result of the converter endpoint carbon content is obtained through the wavelet neural network model.

[0027] Step S8, the prediction result obtained in step S7 is combined with the phosphorus content in the first sample molten steel (TSC) to determine whether the molten steel reaches the preset tapping condition, if the preset tapping condition is reached, the molten steel is directly tapped, and the second sample molten steel (TSO) detection time is not needed, and the production efficiency is improved. Embodiment

[0028] The application further provides a specific case of a converter endpoint carbon content prediction method based on a wavelet neural network, which comprises the following steps: Step S1, 8315 furnace data of a converter workshop of a certain plant from January to February 2025 are collected, including: furnace number, steel grade, hot metal composition (carbon, silicon, manganese, phosphorus and sulfur contents after hot metal pretreatment), hot metal temperature, hot metal weight, scrap steel weight, oxygen supply time, oxygen supply amount, lime, magnesium ball, pellet, light-burned dolomite addition amount and converter endpoint carbon content data.

[0029] Step S2, based on the selected silicon steel, 1750 furnace data after rejecting each parameter blank value; then the data that does not meet the normal distribution is rejected. The remaining data is normalized, and after data preprocessing, the silicon steel category can use data for 755 furnaces.

[0030] Step S3, import the data into the Pearson correlation coefficient analysis, and obtain the correlation between the process parameters and the end point carbon content, as shown in Table 1.

[0031] Table 1 Pearson correlation characteristic value

[0032] Analyzing the data in Table 1, the larger the value, the higher the correlation with the end point carbon content, and the positive and negative signs represent positive and negative correlations, respectively. The carbon content of the hot metal is the main source of the carbon content of the molten steel, and the main reason for its low correlation is that the end point carbon content reaches a certain range, which allows tapping, so the end point carbon content has a low correlation with the carbon content of the hot metal. The physical heat brought by the hot metal can affect the reaction rhythm in the furnace, including carbon oxidation, reduction and other reactions, which in turn affect the end point carbon content, so although the correlation value of the hot metal temperature is low, it is still one of the main factors affecting the end point carbon content. The selected 750 heats of scrap steel are almost consistent, between 18% and 19%, and the data fluctuation range is small, so the hot metal weight and scrap steel weight have little effect on the end point carbon content.

[0033] Step S4, calculate the statistical significance critical value of the Pearson correlation coefficient, wherein the sample number is 755, 1.96, and the statistical significance critical value obtained is 0.07.

[0034] Step S5, screen the process parameters with an absolute value of the Pearson correlation coefficient less than 0.07, and perform principal component analysis on the hot metal sulfur content, hot metal carbon content, magnesium ball, hot metal weight, scrap steel weight, lime, and pellet parameter data. The highest correlation data with the end point carbon content is recorded as Y , which is one of the input process parameters.

[0035] Step S6, as shown in Figure 1 , a wavelet neural network model is constructed, wherein the number of input layer neurons is 8, the number of hidden layer neurons is 17, and the number of output layer neurons is 1. The data is divided into a training set and a test set in a ratio of 9:1, with 680 furnaces in the training set and 75 furnaces in the test set. The data collected above is substituted into the wavelet neural network model, and the learning rate, learning step, and error parameters are continuously optimized to obtain the optimal model parameters.

[0036] Step S7, substitute the input data into the wavelet neural network prediction model to obtain the end point carbon content prediction result.

[0037] Step S8, the furnace length determines whether the direct tapping condition is reached according to the end-point carbon content obtained from the prediction model and the phosphorus content in the TSC sample taken by the sublance. If the tapping condition is reached, the direct tapping is performed.

[0038] Here, as to whether the standard is reached, taking the DANIELI converter intelligent smelting model applied in the converter workshop of a certain plant as an example, 1203 actual production data in a month is collected, the calculated carbon content is obtained through TSC sublance detection, and then the TSC calculated carbon content is substituted into the model to predict the end-point carbon content, wherein the prediction hit rate of the end-point carbon content within ±0.005% is 59.35%, the prediction hit rate within ±0.01% is 77.06%, and the prediction hit rate within ±0.015% is 77.3%.

[0039] After the data is substituted into the wavelet neural network model, the obtained result is shown in Table 2.

[0040] Table 2: End-point carbon content prediction result

[0041] In Table 2, the prediction hit rate of the end-point carbon content within ±0.005% is 62.67%, the prediction hit rate within ±0.01% is 81.33%, and the prediction hit rate within ±0.015% is 89.33%. Compared with the DANIELI converter intelligent smelting model, it is obvious that the prediction result precision obtained by the wavelet neural network-based converter end-point carbon content prediction method provided in the present application is higher.

[0042] Those skilled in the art can understand that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with those in the context of the prior art, and should not be interpreted with idealized or overly formal meanings unless defined as such herein.

[0043] The meaning of “and / or” described in the present application means that each single existence or both existences are included.

[0044] The meaning of “connection” described in the present application can be a direct connection between components or an indirect connection between components through other components.

[0045] With the above ideal embodiments according to the present application as the inspiration, through the above description, relevant staff can make various changes and modifications without deviating from the technical idea of the present application. The technical scope of the present application is not limited to the content of the specification, and must be determined according to the scope of the claims.

Claims

1. A method for predicting the carbon content at the end of a converter based on a wavelet neural network, characterized in that: Specifically, the following steps are included: Step S1: Collect data, collect process parameter information for historical furnace batches; Step S2: Preprocess the process parameters from step S1. First, remove blank values ​​from the process parameters. Then, remove outliers using a normal distribution. Next, combine the preset production range to remove process parameters that exceed the fluctuation range. Finally, normalize the process parameters. Step S3: Perform Pearson correlation coefficient analysis on the pretreated process parameters to determine the correlation between different process parameters and the carbon content at the converter endpoint. Step S4: Sort the correlation magnitudes of different process parameters obtained from the Pearson correlation coefficient analysis in step S3, and take into account the number of process parameters collected to obtain the significant correlation threshold. Step S5: Screen the process parameters, retain the feature data with the highest correlation to the endpoint carbon content, perform principal component analysis on the process parameters with low correlation, and extract the feature values ​​of the process parameters with small correlation factors to obtain a new set of process parameter data. Among them, the screening condition for low correlation is that the absolute value of the Pearson correlation coefficient in step S3 is less than the critical value of significant association in step S4. Step S6: Establish a wavelet neural network model. Divide the new process parameter data set from step S5 into a training set and a test set according to the proportion. Substitute them into the wavelet neural network model for optimization training to obtain the optimal wavelet neural network model. Step S7: Install the optimized wavelet neural network model from step S6 into the converter main control room, read the process parameter information generated during the converter smelting process in real time, and obtain the predicted result of the carbon content at the converter endpoint through the wavelet neural network model. Step S8: Combine the prediction result obtained in step S7 with the phosphorus content of the first sample of molten steel to determine whether the molten steel meets the preset tapping conditions. If the preset tapping conditions are met, tap the steel directly without waiting for the second sample of molten steel to be tested.

2. The method for predicting the carbon content at the converter endpoint based on wavelet neural networks according to claim 1, characterized in that: In step S1, the collected process parameters include the composition of molten iron, temperature of molten iron, weight of molten iron, weight of scrap steel, oxygen blowing amount, oxygen blowing time, amount of auxiliary materials added, and carbon content at the converter endpoint for historical heats. The composition of the molten iron includes the carbon, phosphorus, sulfur, silicon, and manganese content of the molten iron after pretreatment at the KR desulfurization station. The temperature of the molten iron includes the temperature of the molten iron after desulfurization. The amount of auxiliary materials added includes the amount of lime, pellets, magnesium balls, iron oxide balls, and lightly calcined dolomite added.

3. The method for predicting the carbon content at the converter endpoint based on wavelet neural networks according to claim 1, characterized in that: The specific steps for preprocessing the process parameters in step S1 in step S2 are as follows: Step S21: Remove blank values ​​from the process parameters; Step S22: Check whether the process parameter for removing blank values ​​follows a normal distribution. If it does, proceed directly to step S23. If it does not, determine that the process parameter is an outlier and remove it before proceeding to step S23. Step S23: Based on the actual operating conditions and the preset production range of process parameters, process parameters whose fluctuation range exceeds the production range are identified as outliers and are subsequently eliminated. The preset production range of process parameters includes: molten iron temperature range: 1320~1420℃; molten iron weight range: 155~170t; oxygen blowing time: 12~17min; lime, lightly calcined dolomite, and magnesia ball addition: not 0; converter endpoint carbon content range: 0.02~0.1%. Step S24: Normalize the process parameters selected in step S23 using the following formula: （1） In formula (1), These are the normalized data corresponding to the original process parameters. These are the original process parameters. This represents the maximum value after normalization of the original process parameters. This represents the minimum value after normalization of the original process parameters. This represents the maximum value of the original process parameters. This represents the minimum value of the original process parameters.

4. The method for predicting the carbon content at the converter endpoint based on wavelet neural networks according to claim 1, characterized in that: In step S3, the formula used for Pearson correlation coefficient analysis of the pretreated process parameters is: （2） In formula (2), r The Pearson correlation coefficient has a value range of [-1, 1]. These are the normalized process parameters; The final carbon content of the converter; n The number of process parameters after normalization; This represents the mean of the normalized process parameters. This represents the average carbon content at the converter's final stage.

5. The method for predicting the carbon content at the converter endpoint based on wavelet neural networks according to claim 1, characterized in that: In step S4, the formula for calculating the significant association threshold is as follows: （3） In formula (3), q The critical value for significant correlation; n The number of process parameters after normalization; t The result is obtained after inputting normalized process parameters. t Distribution; α is 0.

05.

6. The method for predicting the carbon content at the converter endpoint based on wavelet neural networks according to claim 1, characterized in that: In step S5, the specific steps for obtaining the new process parameter data set are as follows: Step S51, based on The process parameters were screened to obtain those with low correlation. Step S52: The process parameters selected in step S51 are centralized using the following formula: （4） （5） In formulas (4) and (5), For the first j The average value of each process parameter; x ij For the first i Line 1 j The value of the column; n The number of process parameters after normalization; j The number of process parameters; X cent For centralized data; A matrix consisting of the mean values ​​of each normalized process parameter; X for n × p Matrix; Step S53, calculate the covariance matrix. The calculation formula is as follows: （6） In formula (6), S is p × p symmetric matrix; Step S54: Calculate the eigenvalues ​​and eigenvectors of the covariance matrix using the following formula: （7） In formula (7), These are the eigenvalues ​​of the covariance matrix S after eigendecomposition; These are the eigenvectors of the covariance matrix S after eigendecomposition. Step S55: Perform dimensionality reduction on the data using the following formula: （8） In formula (8), Y represents the process parameters after dimensionality reduction; V It is a matrix consisting of the eigenvectors corresponding to the largest eigenvalue.

7. The method for predicting the carbon content at the converter endpoint based on wavelet neural networks according to claim 1, characterized in that: In step S6, the specific steps to obtain the optimal wavelet neural network model are as follows: Step S61: Construct the input layer, hidden layer, and output layer of the model. The input layer is determined by the number of input process parameters, the output layer is the carbon content at the converter endpoint, and the number of hidden layers is the optimal value obtained from multiple actual simulations. The calculation formula is as follows: （9） In formula (9), h This represents the number of hidden layer nodes; a A constant between [0, 10]; m This refers to the number of neurons in the input layer. n This refers to the number of neurons in the output layer. In step S62, the new process parameter data from step S5 is divided into a training set and a test set according to the ratio, and then substituted into the wavelet neural network model constructed in step S61 for optimization training to form the optimal model parameters, so as to obtain the optimal wavelet neural network model.

Images