A Method for Predicting the Final Carbon Temperature in Converter Steelmaking Based on CJS-SLLE Dimensionality Reduction

By introducing carbon temperature labeling information into converter steelmaking using the CJS-SLLE dimensionality reduction method, and combining it with local linear embedding and adaptive projection algorithms, the accuracy problem of carbon temperature detection at the end point of converter steelmaking is solved, and efficient carbon temperature prediction for small and medium-sized converter steelmaking is realized.

CN115600486BActive Publication Date: 2026-11-17KUNMING UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211064606.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-08-31
Publication Date
2026-11-17
Estimated Expiration
2042-08-31

AI Technical Summary

Technical Problem

Existing methods for detecting the final carbon temperature in converter steelmaking are limited by the professionalism of operators and the precision of equipment, making it difficult to achieve accurate real-time detection. This is especially true in small and medium-sized converter steelmaking production, where the complexity of the equipment and the high dimensionality and nonlinearity of the data lead to insufficient prediction accuracy.

Method used

A CJS-SLLE-based dimensionality reduction method is adopted. By introducing carbon temperature label information to construct a supervised measurement method, and combining local linear embedding and adaptive local linear projection algorithms, low-dimensional features of the converter steelmaking process are extracted to establish a carbon temperature prediction model.

Benefits of technology

It enables real-time and accurate prediction of the final carbon temperature in converter steelmaking, improves prediction accuracy and stability, reduces smelting costs, and is suitable for small and medium-sized converter steelmaking production.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115600486B_ABST
    Figure CN115600486B_ABST
Patent Text Reader

Abstract

This invention discloses a method for predicting the final carbon temperature in converter steelmaking based on CJS-SLLE dimensionality reduction. Its key features include: constructing a supervised measurement method by introducing quantized carbon temperature label information into the distance metric to adjust intra- and inter-class variance; then introducing inter-data orientation information based on the labeled information, thus achieving a novel CJS similarity measurement strategy that fuses sample label, orientation, and distance information, which is applied to local linear embedding to obtain the low-dimensional coordinates of high-dimensional training samples; secondly, proposing an adaptive local linear projection strategy for unlabeled test samples to ensure that their low-dimensional coordinates also contain label information; and finally, using an instant learning algorithm to select a subset of samples to establish a partial least squares local regression model to predict the final carbon temperature.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of complex industrial process modeling and soft measurement technology, specifically involving a method for predicting the carbon temperature at the end point of converter steelmaking based on CJS-SLLE dimensionality reduction. Background Technology

[0002] Converter steelmaking is my country's primary steelmaking method, especially driven by the increasing demand for steel due to China's booming economy, which has further propelled the development of the steel industry. A key aspect of steelmaking is the precise measurement of the carbon temperature at the end of the converter process, ensuring the carbon content in the molten steel reaches the controlled range for the desired steel grade, and guaranteeing the tapping temperature is sufficient for smooth refining and casting. Therefore, accurate real-time forecasting of the carbon temperature at the end of converter steelmaking can effectively control tapping time, reduce the number of re-blowing operations, and is of great significance for improving smelting efficiency, reducing smelting costs, and promoting green development.

[0003] Currently, the main methods for detecting carbon content and temperature at the end of converter steelmaking include contact-based methods such as empirical steelmaking, secondary lance detection, and furnace gas analysis. With the development of computer technology and artificial intelligence, non-contact methods for carbon temperature detection have emerged, such as optical imaging, furnace mouth flame image processing, and data-driven soft measurement modeling. Among these methods, the experience-based steelmaking method relies mainly on operators to determine the endpoint of converter steelmaking by combining the materials added during smelting, the blowing conditions, and observing the dynamic characteristics of the furnace flame and changes in sound. This method is highly influenced by human factors and has low efficiency. The auxiliary lance detection method involves attaching a disposable probe to an auxiliary lance and inserting it into the molten pool for measurement. This method has a complex equipment system, strict requirements on plant height, and can only be used in large converters, not in small and medium-sized converter steelmaking. The furnace gas analysis method involves collecting furnace gas from pipelines and sending it to specialized detection equipment to analyze the furnace gas composition and predict the carbon temperature of molten steel. This method has high requirements for the gas supply system in steelmaking, and needs to avoid the impact of oxygen pressure fluctuations on blowing, as well as the impact of the lag time in furnace gas sampling and testing. It is difficult to accurately detect the carbon temperature at the endpoint of converter steelmaking blowing.

[0004] The accuracy of the above-mentioned contact-based carbon temperature detection methods is mainly limited by the operator's expertise and the precision of the equipment. However, with the development of intelligent detection technology, a carbon temperature endpoint prediction model can be established using spectral information from the steelmaking endpoint, furnace flame images, and production process data. Among these, the spectral image method suffers from the difficulty of effectively acquiring flame spectral images due to the complex environment and flue gas interference within the actual molten pool, thus affecting the identification of spectral information; the flame image processing method suffers from the difficulty of extracting complex and variable texture features, thus impacting prediction accuracy.

[0005] Therefore, a method is needed to reduce external influences and accurately measure the final carbon temperature of converter steelmaking. This has led to the development of data-driven soft measurement modeling methods to address the difficulty in measuring key variables in actual industrial processes. Due to the high dimensionality and nonlinearity of production process data, directly measuring the similarity of high-dimensional data is challenging; therefore, dimensionality reduction is necessary for feature extraction to address the problem of excessively high data dimensionality. Existing technologies improve model prediction performance by extracting important low-dimensional features through dimensionality reduction; however, these methods have limited performance improvement because they do not consider the supervisory role of output label information in dimensionality reduction. Summary of the Invention

[0006] To address the technical challenges of high dimensionality, nonlinearity, and large fluctuations in converter steelmaking process data, this invention provides a CJS-SLLE-based method for predicting the carbon temperature at the converter steelmaking endpoint. This method constructs a supervised measurement mechanism by introducing quantized carbon temperature label information into the distance metric, thereby adjusting intra- and inter-class variance. Then, based on the labeled information, directional information between data points is introduced, resulting in a novel CJS similarity measurement strategy that fuses sample label, direction, and distance information. This strategy is applied to local linear embedding to obtain low-dimensional coordinates of high-dimensional training samples.

[0007] To achieve the above-mentioned technical effects, the present invention is implemented through the following technical solution: a method for predicting the final carbon temperature of converter steelmaking based on CJS-SLLE dimensionality reduction, characterized by the following steps: S1. Collect actual converter steelmaking production process data from steel plants; S2. Normalize the data collected in S1 above; S3. Divide the original high-dimensional process data into training samples. and test samples ; S4. Use the CJS-SLLE supervised dimensionality reduction algorithm to extract low-dimensional features from high-dimensional training samples and test samples respectively; S4.1 Construct a CJS similarity measurement strategy that integrates fusion distance, carbon temperature label information, and data orientation information; S4.2. Supervised dimensionality reduction of high-dimensional training samples based on CJS similarity measurement strategy to extract low-dimensional features; S4.3. Using the adaptive local linear projection algorithm, extract the low-dimensional features containing label information of the test sample based on the high-dimensional training samples and their low-dimensional features. S5. Based on the real-time learning algorithm, the subset of similar samples with the strongest correlation to the test sample is selected from the low-dimensional training samples using grey relational analysis as local model training samples. S6. Input the current sample to be tested and the corresponding similar sample subset into the partial least squares local regression prediction model to predict the endpoint carbon content and temperature; S7. Repeat steps 5 and 6 for the next sample to be tested.

[0008] Furthermore, S4.1 specifically includes the following steps: S1. The JS divergence, which is used to measure the difference (or similarity) between the probability distributions of two random variables, is used as a distance metric in similarity measurement to measure the degree of proximity between samples.

[0009]

[0010] S2. Extract the final carbon content or temperature label value of converter steelmaking to B = [b1, b2, ..., b n The carbon content and temperature label values ​​are uniformly quantified into label information k. i ;

[0011] Among them, b i For sample x i The corresponding carbon temperature label value, c represents the number of decimal places in the sample label value. k i This refers to the quantized label information corresponding to the sample; S3. Introduce carbon temperature label information into the JS divergence distance metric to reconstruct the metric function;

[0012] S4. Utilizing cosine similarity and the JS divergence similarity criterion that incorporates sample carbon temperature label information, a CJS similarity measurement strategy integrating distance, label, and orientation information is constructed. The algorithm is as follows:

[0013]

[0014] D CJS The higher the (P,Q) similarity value, the higher the similarity between the samples.

[0015] Furthermore, S4.2 specifically includes the following steps: S1. Find the k nearest neighbors: Use the CJS similarity measurement strategy to calculate the similarity between sample points and other sample points in the original training data, and select the k sample points with the highest similarity to the current sample as the nearest neighbors of the sample point. S2. Linear Reconstruction: Representing sample point x linearly using k-nearest neighbors with weights. i Then, the local reconstruction weight coefficients are calculated by minimizing the error function;

[0016] The constraints are Q i For x i The set of k-nearest neighbors; S3. Calculate the low-dimensional embedding coordinates: The weight coefficients in the local linear embedding algorithm remain unchanged in both high-dimensional and low-dimensional spaces. Therefore, minimizing the error function in the low-dimensional space yields the low-dimensional feature coordinates of the training samples. ; Will Construct a weight matrix W Define a sparse, symmetric positive semidefinite matrix M. (IW) yields the objective function as follows:

[0017] Right now The constraint is YY T =I. Calculated using the Lagrange multiplier method, we get... In other words, obtaining the eigenvectors of M is equivalent to obtaining the low-dimensional embedding coordinates Y. The matrix formed by the eigenvectors corresponding to the first d smallest non-zero eigenvalues ​​of matrix M after eigenvalue decomposition is the low-dimensional spatial coordinate matrix Y of X.

[0018] Furthermore, S4.3 specifically includes the following steps: S1. Calculate the similarity of the test samples using a CJS similarity measurement strategy that does not include label information. Compared with all samples in the original high-dimensional training samples similarity ,right Sort the samples from largest to smallest, and determine the sample points with the closest similarity to the current test sample by setting a custom threshold. This yields the high-dimensional and low-dimensional coordinates of the nearest neighbor training samples of the test sample. , , Select the top n samples with the highest similarity to the original high-dimensional training samples. The mean of the samples is used as the threshold for selecting one similar sample, and the number of similar samples is determined, n=30;

[0019] S2. Find the mapping matrix between the high-dimensional and low-dimensional data of the corresponding nearest neighbors of the sample to be tested. Make it satisfy the mapping relationship ;

[0020]

[0021]

[0022] in , represent The One element, It is a regularization parameter. express The identity matrix, composed of noise levels Sure value, represent The non-zero eigenvalues ​​sorted from smallest to largest; S3. Based on the specific mapping matrix of the sample to be tested. Calculate the test sample Corresponding low-dimensional embedding coordinates ; .

[0023] The beneficial effects of this invention are: The original high-dimensional data is used to extract features by introducing sample carbon temperature labels and orientation information into the distance metric using the CJS-SLLE dimensionality reduction algorithm, which transforms the unsupervised LLE into a supervised dimensionality reduction process. Meanwhile, the adaptive local linear projection method adaptively searches for sample points and their corresponding low-dimensional coordinates that can locally linearly represent the current test sample in the original training samples, and finds the corresponding mapping matrix between the two, thus solving the problem of lack of label information in supervised dimensionality reduction of the test sample.

[0024] 2. For converter steelmaking production process data, the proposed CJS-SLLE algorithm is used to achieve supervised dimensionality reduction of training samples and test samples respectively. Then, the supervised dimensionality reduction method is combined with the real-time learning strategy to construct a local model and establish a soft measurement model of carbon temperature at the end of converter steelmaking, realizing real-time and accurate prediction of carbon content and temperature.

[0025] 3. By comparing with existing technologies, using prediction accuracy (PA), root mean square error (RMSE), and mean absolute percentage error (MAPE) as evaluation metrics, based on... Figure 1 and Figure 2 Simulation results show that the method presented in this paper has better prediction performance than existing technologies, compared with other methods.

[0026] Comparison of experimental results between this application and existing technologies Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart of the converter steelmaking endpoint carbon temperature prediction method based on CJS-SLLE dimensionality reduction of the present invention; Figure 2 This is a diagram of the CJSSLLE supervised dimensionality reduction algorithm in step 4 of this invention; Figure 3 This is a comparison of the carbon content prediction accuracy of the present invention with that of various methods. Figure 4 This is a comparison of the temperature prediction accuracy of the present invention with that of various methods. Detailed Implementation

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Example 1: As Figure 1-2 As shown, this paper presents a method for predicting the final carbon temperature in converter steelmaking based on CJS-SLLE dimensionality reduction. By introducing quantized carbon temperature label information into the distance metric, a supervised measurement method is achieved. Then, considering the directional factors between data, a novel CJS similarity measurement strategy that conforms to the characteristics of process data is constructed by introducing directional information between data on top of the supervised measurement. This strategy is applied to local linear embedding to achieve supervised dimensionality reduction of training samples. Furthermore, an adaptive local linear projection algorithm is used to quickly obtain the low-dimensional coordinates containing supervised information of newly added test samples. Finally, based on the JITL algorithm, the subset of samples with the strongest correlation to the test samples is selected from the low-dimensional training samples using grey relational analysis and input into the partial least squares regression local model to achieve the final carbon temperature prediction. The specific steps are as follows: Step 1: Collect actual converter steelmaking process data from the steel plant; Step 2: By applying the maximum-minimum normalization principle, the process data features are unified within the range of [0,1] to eliminate the influence of differences in unit scale between different features; Step 3: Divide the original high-dimensional process data into training samples X = [x1, x2, ..., x...]. n ],X∈R D×n and test sample X new =[x new1 ,x new2 ,... ,x newm ], X new ∈R D×m ; Step 4: Use the invented CJS-SLLE supervised dimensionality reduction algorithm to extract low-dimensional features from the high-dimensional training samples and test samples respectively; Step 5: Based on the real-time learning algorithm, select the subset of similar samples with the strongest correlation to the test sample from the low-dimensional training samples using grey relational analysis as local model training samples. Step 6: Input the current sample to be tested and the corresponding similar sample subset into the partial least squares local regression prediction model to predict the endpoint carbon content and temperature; Step 7: Repeat steps 5 and 6 for the next sample to be tested.

[0031] The supervised local linear embedding dimensionality reduction method described in step 4 has the following specific steps: Step 4.1: Construct a novel CJS similarity measurement strategy that integrates distance, carbon temperature label information, and orientation information between data points to accurately reflect the local neighborhood information of sample points; Step 4.1.1: Use the JS divergence, which measures the difference (or similarity) between the probability distributions of two random variables, as the distance metric in similarity measurement to measure the degree of proximity between samples.

[0032]

[0033] Step 4.1.2: Extract the final carbon content or temperature label value of converter steelmaking to B = [b1, b2, ..., b n The carbon content and temperature label values ​​are uniformly quantified into label information k. i ;

[0034] Among them, b i For sample x i The corresponding carbon temperature label value, where c is the number of decimal places in the sample label value, and k iThis represents the quantized label information corresponding to the sample. When c = 0, it indicates a temperature label. To avoid the label information being 0, the smallest unit of the label is added first, and the difference between this and the smallest value in the original label is used as the quantized label information. When c < 0, it indicates a carbon content label. To find the minimum unit value for carbon content, first take the difference between adding the minimum unit value to all labels and the minimum value in the original label, then multiply by 10. c The label information is post-quantized to the smallest positive integer.

[0035] Step 4.1.3: Introduce carbon temperature label information into the JS divergence distance metric to reconstruct the metric function, making the metric a supervised process, further bringing samples with similar labels closer together, while further increasing the similarity of samples with large differences in labels;

[0036] Step 4.1.4: Using the cosine similarity criterion for measuring directional similarity between samples and the JS divergence similarity criterion that incorporates sample carbon temperature label information, a novel CJS similarity measurement strategy that integrates distance, label, and directional information is constructed.

[0037] In the novel CJS similarity calculation method, (1-cosθ) uniformly standardizes the cosine similarity measure to the range [0,2], and incorporates D containing supervised information. JSL The (P,Q) metric is combined with the cosine similarity of the orientation information, taking the overall value as e. -x A novel CJS method for calculating sample similarity is derived from the following formula:

[0038] D CJS The higher the (P,Q) similarity value, the higher the similarity between the samples.

[0039] Step 4.2: Perform supervised dimensionality reduction on high-dimensional training samples based on the CJS similarity measurement strategy to extract low-dimensional features; Step 4.2.1: Find the k nearest neighbors. A novel CJS similarity metric strategy is used to calculate the similarity between a sample point and other sample points in the original training data. The k sample points with the highest similarity to the current sample are selected as the nearest neighbors of that sample point. Step 4.2.2, Linear Reconstruction. The sample point x is linearly represented using its k nearest neighbors with weights. i Then, the local reconstruction weight coefficients are calculated by minimizing the error function;

[0040] The constraints are Q i For x i Given the set of k-nearest neighbors, it is easy to know x i With x j Given the training samples, the local covariance matrix C can be obtained. jk =(x i -x j ) T (x i -x k Using the Lagrange multiplier method, the optimal weight w is obtained. ij The solution is as follows:

[0041] Step 4.2.3: Calculate the low-dimensional embedding coordinates. The weight coefficients in the Local Linear Embedding algorithm remain unchanged in both high-dimensional and low-dimensional spaces. Therefore, minimizing the error function in the low-dimensional space yields the low-dimensional feature coordinates of the training samples, Y=[y1,y2,...,y...]. n ],Y∈R d×n ; w ij Construct a weight matrix W = [w ij ] n×n Define a sparse, symmetric positive semi-definite matrix M = (IW) T (IW), the objective function is obtained as follows:

[0042] Right now The constraint is YY T = I. Calculated using the Lagrange multiplier method, we get... That is, obtaining the eigenvectors of M is equivalent to obtaining the low-dimensional embedding coordinates Y. The matrix formed by the eigenvectors corresponding to the first d smallest non-zero eigenvalues ​​of matrix M after eigenvalue decomposition is the low-dimensional spatial coordinate matrix Y of X.

[0043] Step 4.3: Extract low-dimensional features containing label information from the test sample based on the high-dimensional training samples and their low-dimensional features using the adaptive local linear projection algorithm; Step 4.3.1: Calculate the test sample x using the CJS similarity measurement strategy that does not include label information. newi Compared with all samples x in the original high-dimensional training samples i similarity Si i , for Si i Sort the samples from largest to smallest, and determine the sample points with the closest similarity to the current test sample by setting a custom threshold. This yields the high-dimensional and low-dimensional coordinates X of the nearest neighbor training samples of the test sample. l=[x1,x2,...,x l ], X l ∈R D×l Y l =[y1,y2,...,y l ], Y l ∈R d×l Select the first n samples with the highest similarity Si to the original high-dimensional training samples. i The mean of the samples is used as the threshold for selecting l similar samples, thereby adaptively determining the number of similar samples. Based on experience, n = 30 is determined.

[0044] Step 4.3.2: Find the mapping matrix A = [a1, a2, ..., a1] between the high-dimensional and low-dimensional data of the corresponding nearest neighbors of the sample to be tested. d ] T Make it satisfy the mapping relationship Y l =AX l ;

[0045] a j =(XX) T +βI) -1 Xy i

[0046] Where 1≤j≤d, Represents y i The j-th element, β>0 is the regularization parameter, I represents the D×D identity matrix, and is determined by the noise level δ. 2 Determine the value of β, γ i Representing XX T The non-zero eigenvalues ​​sorted from smallest to largest.

[0047] Step 4.3.3: Calculate the value of the sample x based on the specific mapping matrix A of the sample to be tested. newi The corresponding low-dimensional embedding coordinates y newi ; Y new =AX new .

[0048] The specific steps for gray relational analysis in step 5 are as follows: Step 5.1: Given the low-dimensional training set Y and the test set Y new Determine the analysis sequence, where the current sample to be tested is y. new For the reference sequence, all training samples y i To compare sequences. Where: Reference sequence: ynew =y new (k) |k=1,2...n; Compare sequences: y i =y i (k) |k=1,2...n,i=1,2...m.

[0049] Step 5.2: Calculate the correlation coefficient. Calculate the difference between the reference series and the comparison series, and let... The grey relational coefficients are shown below:

[0050] In the formula, ρ is usually taken as 0.5. To compare the correlation coefficients between the sequence and the reference sequence at each k-point.

[0051] Step 5.3: Calculate the correlation value. The current reference sequence y is calculated using the following formula. new Compare the sequence y i The correlation value.

[0052]

[0053] Among them, D GRA,i To compare the correlation coefficient (i.e. correlation degree value) between the sequence and the reference sequence at each k point.

[0054] Step 5.4, Relevance Ranking. Sort the reference sequence y... new Compare with all sequences y i The calculated correlation scores are sorted, D GRA,i The larger the value, the more relevant it is to the reference sequence.

[0055] The parameter estimation methods for the reduced dimension d, nearest neighbor k, and sample subset S in the implementation plan are as follows: The dimension d is selected in the interval [5, 11] with a step size of 1; the nearest neighbor k is selected in the interval [10, 20] with a step size of 1; and the optimal number of sample subsets S is verified in the interval [5, 100] with a step size of 1. All possible combinations of the three parameters are verified, and the globally optimal parameters are finally estimated.

[0056] Experimental results show that the carbon content prediction accuracy of this invention reaches 90% within an error range of ±0.02%, and the temperature prediction accuracy reaches 87% within an error range of ±10℃. This invention can effectively extract low-dimensional key features of process data for prediction using local models in real time, effectively solving the problems of high dimensionality, nonlinearity, and large data fluctuations in production process data, and improving the prediction accuracy of final carbon temperature.

[0057] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example.

[0058] Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0059] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for predicting the final carbon temperature in converter steelmaking based on CJS-SLLE dimensionality reduction, characterized in that, Specifically, the following steps are included: S1. Collect actual converter steelmaking production process data from steel plants; S2. Normalize the data collected in S1 above; S3. Divide the original high-dimensional process data into training samples X = [x1, x2, ..., x...]. n ],X∈R D×n and test sample X new =[x new1 ,x new2 ,…,x newm ], X new ∈R D×m ; S4. Use the CJS-SLLE supervised dimensionality reduction algorithm to extract low-dimensional features from high-dimensional training samples and test samples respectively; S4.1 Construct a CJS similarity measurement strategy that integrates distance, carbon temperature label information, and orientation information between data; S4.1.

1. The JS divergence, which is used to measure the difference or similarity between the probability distributions of two random variables, is used as a distance metric in similarity measurement to measure the degree of proximity between samples. S4.1.2, Extract the final carbon content or temperature label value of converter steelmaking to B = [b1, b2, ..., b i The carbon content and temperature label values ​​are uniformly quantified into carbon-temperature label information k. i ; S4.1.

3. Introduce carbon temperature label information into the JS divergence distance metric to reconstruct the metric function; S4.1.

4. Utilize cosine similarity and the JS divergence similarity criterion that incorporates sample carbon temperature label information to construct a CJS similarity measurement strategy that integrates distance, carbon temperature label information, and inter-data orientation information; S4.

2. Supervised dimensionality reduction of high-dimensional training samples based on CJS similarity measurement strategy to extract low-dimensional features; S4.

3. Using the adaptive local linear projection algorithm, extract the low-dimensional features containing carbon temperature label information of the test sample based on the high-dimensional training samples and their low-dimensional features. S5. Based on the real-time learning algorithm, the subset of similar samples with the strongest correlation to the test sample is selected from the low-dimensional training samples using grey relational analysis as local model training samples. S6. Input the current sample to be tested and the corresponding similar sample subset into the partial least squares local regression prediction model to predict the endpoint carbon content and temperature; S7. Repeat steps 5 and 6 for the next sample to be tested.

2. The method for predicting the final carbon temperature of converter steelmaking based on CJS-SLLE dimensionality reduction according to claim 1, characterized in that, S4.1 specifically includes the following steps: S4.1.

1. The JS divergence, which is used to measure the difference or similarity between the probability distributions of two random variables, is used as a distance metric in similarity measurement to measure the degree of proximity between samples. S4.1.2, Extract the final carbon content or temperature label value of converter steelmaking to B = [b1, b2, ..., b i The carbon content and temperature label values ​​are uniformly quantified into carbon-temperature label information k. i ; Among them, b i For sample x i The corresponding carbon temperature label value, c represents the number of decimal places in the sample label value. k i This refers to the quantized carbon temperature label information corresponding to the sample. S4.1.

3. Introduce carbon temperature label information into the JS divergence distance metric to reconstruct the metric function; S4.1.

4. Using cosine similarity and the JS divergence similarity criterion that incorporates sample carbon temperature label information, a CJS similarity measurement strategy that integrates distance, carbon temperature label information and data orientation information is constructed. The algorithm is as follows. D CJS The higher the (P,Q) similarity value, the higher the similarity between the samples.

3. The method for predicting the final carbon temperature of converter steelmaking based on CJS-SLLE dimensionality reduction according to claim 1, characterized in that, S4.2 specifically includes the following steps: S4.2.1 Finding k nearest neighbors: The CJS similarity measurement strategy is used to calculate the similarity between sample points and other sample points in the original training data, and the k sample points with the highest similarity to the current sample are selected as the nearest neighbors of the sample point. S4.2.2 Linear Reconstruction: The sample point xi is linearly represented by the k nearest neighbors in a weighted manner, and then the local reconstruction weight coefficients are calculated by minimizing the error function; The constraints are Q i For x i The set of k-nearest neighbors; S4.2.3 Calculating Low-Dimensional Embedding Coordinates: The weight coefficients in the local linear embedding algorithm remain unchanged in both high-dimensional and low-dimensional spaces. Therefore, minimizing the error function in the low-dimensional space yields the low-dimensional feature coordinates Y = [y1, y2, ..., y]. n ],Y∈R d×n ; w ij Construct a weight matrix W = [w ij ] n×n Define a sparse, symmetric positive semi-definite matrix M = (IW) T (IW), the objective function is obtained as follows: Right now The constraint is YY T =I, calculated using the Lagrange multiplier method, yields... That is, obtaining the eigenvectors of M is equivalent to obtaining the low-dimensional embedded coordinates Y; the matrix formed by the eigenvectors corresponding to the first d smallest non-zero eigenvalues ​​of matrix M after eigenvalue decomposition is the low-dimensional spatial coordinate matrix Y of X.

4. The method for predicting the final carbon temperature of converter steelmaking based on CJS-SLLE dimensionality reduction according to claim 1, characterized in that, S4.3 specifically includes the following steps: S4.3.1 Calculate the test sample x using the CJS similarity measurement strategy that does not include carbon temperature label information. newi Compared with all samples x in the original high-dimensional training samples i similarity Si i , for Si i Sort the samples from largest to smallest, and determine the sample points with the closest similarity to the current test sample by setting a custom threshold. This yields the high-dimensional and low-dimensional coordinates X of the nearest neighbor training samples of the test sample. s =[x1,x2,…,x s ], X s ∈R D×s Y s =[y1,y2,…,y s ], Y s ∈R d×s Select the first n samples with the highest similarity Si to the original high-dimensional training samples. k The mean of the samples is used as the threshold for selecting s similar samples to determine the number of similar samples, n = 30; S4.3.2 Find the mapping matrix A = [a1, a2, ..., a] between the high-dimensional and low-dimensional data of the corresponding nearest neighbors of the sample to be tested. d ] T Make it satisfy the mapping relationship Y s =AX s ; a j =(XX T +βI) -1 Xy i Where 1≤j≤d, Represents y i The j-th element, β>0 is the regularization parameter. I represents a D×D identity matrix, determined by the noise magnitude δ. 2 Determine the value of β. γ i Representing XX T The non-zero eigenvalues ​​sorted from smallest to largest; S4.3.3 Calculate the value of the sample x based on the specific mapping matrix A of the sample to be tested. newi The corresponding low-dimensional embedding coordinates y newi ; Y new =AX new 。