Continuous casting microalloy peritectic steel component optimization method based on limit gradient lifting tree and particle swarm optimization

Optimizing the composition of microalloy clad steel through the ultimate gradient enhancement tree and particle swarm algorithm, solving the problems of low component optimization efficiency and insufficient adaptability in the existing technology, significantly reducing the incidence of longitudinal cracks on the surface of continuous casting billets, and improving casting Blank quality and casting machine output.

CN118918997APending Publication Date: 2024-11-08DALIAN UNIV OF TECH
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411001771.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The prior art has problems such as high trial and error costs, low efficiency and poor adaptability of optimization methods when optimizing the composition of microalloy clad steel, which makes it difficult to effectively reduce the risk of longitudinal cracks on the surface of continuous casting billets.

Method used

Using the method based on the ultimate gradient enhancement tree and particle swarm algorithm, the steel grade composition data sample library is constructed, and the ultimate gradient enhancement tree classification model is established. Combined with Bayesian optimization and particle swarm algorithm, the steel grade composition is optimized to reduce the tendency of longitudinal cracking. .

Benefits of technology

The efficient optimization of the continuous cast microalloy crystal-clad steel composition is achieved, which significantly reduces the incidence of longitudinal cracks and improves the surface quality of the casting billet and the stability of the casting machine output.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918997A_ABST
    Figure CN118918997A_ABST
Patent Text Reader

Abstract

The invention discloses a continuous casting microalloy peritectic steel component optimization method based on a limit gradient lifting tree and a particle swarm algorithm, and belongs to the technical field of ferrous metallurgy continuous casting. The method comprises the following steps: firstly, detecting and recording the specific content of alloy elements in microalloy peritectic steel on line, and constructing a sample database containing normal working conditions and longitudinal crack working conditions in combination with a casting blank production quality report; secondly, performing parameter optimization on the limit gradient boosting tree model to obtain a limit gradient boosting tree classification model under an optimal hyper-parameter combination, and realizing accurate prediction of the surface longitudinal crack; then, a corresponding longitudinal crack nonlinear mapping function is extracted, and the occurrence tendency of surface longitudinal cracks is evaluated; and finally, the optimal solution of the nonlinear mapping function is searched, and the purpose of reducing the longitudinal crack occurrence tendency by optimizing steel grade components is achieved. The surface quality of the casting blank can be improved, and the longitudinal crack occurrence tendency can be reduced; the advantages of a machine learning algorithm are fully played, high efficiency and good adaptability are shown, and a technical means is provided for efficient continuous casting production and intelligent optimization of casting blank quality.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of continuous casting of iron and steel metallurgy, and relates to a composition optimization method for continuous casting microalloy peritectic steel based on extreme gradient boosting tree and particle swarm algorithm. Background Art

[0002] As a common surface defect, longitudinal cracks in continuous casting billets seriously affect the quality of billets and the output of casting machines. Longitudinal cracks are usually formed near the meniscus of the crystallizer. Due to the premature appearance of air gaps between the solidified billet shell and the copper plate of the crystallizer, the heat flow distribution is uneven and the growth of the solidified billet shell is uneven. When the stress at the weak position of the billet shell exceeds the strength limit it can withstand, microcracks will initiate on the surface of the billet shell and eventually expand into longitudinal cracks along the casting direction. Due to the peritectic transformation, the solidification shrinkage of peritectic steel is large, and it is more likely to produce surface cracks than other continuous casting steel grades. Although avoiding the carbon content range with severe peritectic reaction as much as possible can significantly reduce the risk of longitudinal cracks, due to the actual production needs, the composition of microalloyed peritectic steel cannot completely avoid this range. Therefore, how to reasonably control and optimize the chemical composition of microalloyed peritectic steel, improve the high-temperature mechanical properties of steel grades, and reduce the risk of surface longitudinal cracks is of great significance to improving the quality of billets and the output of casting machines.

[0003] Authorized patent CN115747616A proposes a smelting method for marine storage tank steel, which adopts peritectic steel composition design to effectively ensure the toughness of the product and improve the strength after heat treatment. Authorized patent CN112813355A proposes a method for reducing corner cracks in titanium-containing peritectic steel continuous casting slabs by adjusting the alloy composition. Authorized patent CN107630172A proposes a method for preventing surface cracks in low-carbon boron-containing steel. This method effectively reduces the incidence of surface cracks in low-carbon boron-containing steel after hot rolling by controlling factors such as steel grade composition, protective slag basicity, and process parameters. However, the above-mentioned optimization methods for steel grade composition are all derived from production practice experience, and there are shortcomings such as high trial and error costs, low efficiency, and poor adaptability of optimization methods. Summary of the invention

[0004] The purpose of the present invention is to reduce the tendency of longitudinal cracks on the surface of continuous casting billets by optimizing the steel composition. A composition optimization method for continuous casting micro-alloy peritectic steel based on extreme gradient boosting tree and particle swarm algorithm is proposed, which provides an effective intelligent optimization method for longitudinal cracks for continuous casting billet quality control.

[0005] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0006] A method for optimizing the composition of continuous casting microalloyed peritectic steel based on extreme gradient boosting tree and particle swarm algorithm. The optimization method first detects and records the specific content of alloying elements for the main steel grade composition of microalloyed peritectic steel online, and builds a sample database containing normal working conditions and longitudinal cracking conditions in combination with the ingot production quality report; secondly, based on the sample library, the Bayesian algorithm is used to optimize the parameters of the extreme gradient boosting tree model to obtain the extreme gradient boosting tree classification model under the optimal hyperparameter combination, so as to achieve accurate prediction of surface longitudinal cracks; then, according to the optimal extreme gradient boosting tree classification model, the corresponding longitudinal crack nonlinear mapping function is extracted to evaluate the occurrence tendency of surface longitudinal cracks; finally, the particle swarm algorithm is used to search for the optimal solution of the nonlinear mapping function, so as to achieve the purpose of reducing the occurrence tendency of longitudinal cracks by optimizing the steel grade composition. Specifically, the following steps are included:

[0007] Step 1: Build a sample database of steel composition data

[0008] For the ingots when the casting process is stable, a specific microalloy peritectic steel continuous casting grade with a high incidence of longitudinal cracking is selected. Within the same protective slag type and specified carbon content range, the historical data of the steel plant is traced back, and c-1 chemical elements to be optimized that have a significant impact on the solidification characteristics and defects of the steel grade are selected to construct a sample database containing normal working conditions and longitudinal cracking conditions. Each sample represents a piece of ingot, and each feature of the sample is the actual chemical composition of the casting of this piece of ingot. Each sample is marked and recorded in combination with the production report, including two types of normal samples and longitudinal cracking samples.

[0009] The data sample library obtained by backtracking is defined as the matrix D n×c , where n is the number of rows, indicating the number of samples, including normal samples and longitudinal fissure samples; c is the number of columns, and the 1st to c-1 columns represent the c-1 features of the sample, that is, c-1 elemental chemical compositions, and the cth column is the label corresponding to the sample, label 1 represents the longitudinal fissure sample, and label 0 represents the normal sample.

[0010] Step 2: Construct the optimal extreme gradient boosting tree classification prediction model

[0011] The specific implementation steps of the optimal extreme gradient boosting tree classification prediction model are as follows:

[0012] (2.1) Dividing the model data set

[0013] First, a model dataset K is constructed, which contains N samples, each of which comes from the steel composition data sample library D established in the first step. n×c Specific: According to the data sample library D n×c The label records of the cth column in (1 for longitudinal fissure samples, 0 for normal samples) are selected from the data sample library D n×c All longitudinal fissure samples in the dataset and randomly select the data sample library D n×cThe normal samples in the dataset K are used to form the dataset K, ensuring that the ratio between the number of longitudinal fissure samples and the number of normal samples in the dataset K is 1:3. The total number N of normal samples and longitudinal fissure samples in the dataset K is not less than 150.

[0014] Secondly, the selected data set K is randomly divided into training set K train and the test set K test , set the ratio of the number of training set P and the number of test set samples Q to 4:1; at the same time, keep the ratio of longitudinal fissure samples and normal samples in the training set and test set to 1:3. train With the test set K test The expression is as follows:

[0015] K train ={(X 1 ,0),(X 2 ,0),(X 3 ,0),...,(X P-2 ,1),(X P-1 ,1),(X P ,1)} (1)

[0016] K test ={(X 1 ,0),(X 2 ,0),(X 3 ,0),...,(X Q-2 ,1),(X Q-1 ,1),(X Q ,1)} (2)

[0017] Among them, X m represents a sample, and the subscript m represents the mth sample.

[0018] (2.2) Constructing the Extreme Gradient Boosting Tree Classification Prediction Model

[0019] The extreme gradient boosting tree model gradually reduces the prediction error by iteratively building multiple decision trees. Each decision tree improves the model performance by fitting its residuals based on the previous model. The structure expression of the extreme gradient boosting tree model is as follows:

[0020]

[0021] In the formula, F(X) represents the prediction result of the model, f k (X) represents the prediction result of the kth decision tree, and K represents the total number of decision trees.

[0022] (2.2.1) Define the model optimization objective function

[0023] The extreme gradient boosting tree model optimizes model parameters by minimizing the loss function. For classification problems, the logarithmic loss function is used as the loss function l of the extreme gradient boosting tree model, and its expression is as follows:

[0024] l(y i ,p i )=-y i log(p i )-(1-y i )log(1-p i ) (4)

[0025] In the formula, y i represents the true label of the i-th sample, p i It represents the probability that the i-th sample is predicted to be a positive class.

[0026] For a training set K containing P samples train , consisting of K decision tree models f k (X i The extreme gradient boosting tree model composed of )(k=1,2,...,K), the corresponding optimization objective function expression L(θ) is as follows:

[0027]

[0028] In the formula, θ refers to all parameters of the model; P represents the number of samples in the training set; f k (X i ) represents the prediction function added in the kth training; Ω[f k (X i )] represents the sum of the model complexity of the first k decision trees.

[0029] (2.2.2) Objective function solution

[0030] The incremental superposition training method is used to solve the objective function (5). For the i-th sample (X i ,y i ), and its superposition training expression is as follows:

[0031]

[0032] In the formula, Represents the k-th predicted value of the model; Represents the k-1th predicted value of the model;

[0033] In formula (6), Substituting into formula (5), we can get the objective function of the k-th training, which is expressed as follows:

[0034]

[0035] The purpose of the kth iteration is to find the most suitable f k (X i ) minimizes the objective function. In the process of optimizing the objective function, the second-order Taylor formula is used to expand equation (7) to obtain:

[0036]

[0037] In the formula, g i is the first-order gradient of the loss function l, as shown in formula (9); h i is the second-order gradient of the loss function l, as shown in formula (10).

[0038]

[0039]

[0040] To prevent the model from overfitting, a regularization term is added to the objective function of the extreme gradient boosting tree, and the complexity Ω[f k (X i )]. The decision tree model is further expressed as:

[0041]

[0042] In the formula, ω represents the weight function of the leaf node; q represents the structure function of the tree, which maps the sample to the leaf node index; Represents the tree for sample X i The predicted value of ; d represents the sample feature dimension.

[0043] For a given decision tree f k (X i ), and its complexity Ω[f k (X i )] is defined as follows:

[0044]

[0045] In the formula, γ represents the regularization parameter that controls the number of leaf nodes; T represents the number of leaf nodes in the tree; λ represents the regularization parameter that controls the weight of the leaf nodes; ω j Represents the weight of the jth leaf node.

[0046] Assume that the training sample X i is divided into leaf node j, and the sample set on the corresponding node is I j ={i|q(X i )=j}, then substitute equation (11) and equation (12) into equation (8), and the optimization objective function can be organized as:

[0047]

[0048] In the formula, This formula is an independent quadratic function of T.

[0049] The goal of solving the above objective function is to find the optimal decision tree, that is, to solve the optimal leaf node weight. j ={i|q(X i )=j}), by j Derivation obtains the optimal weight value of the leaf node:

[0050]

[0051] Substituting formula (14) into the objective function shown in formula (13), we can obtain the optimal value of the objective function:

[0052]

[0053] In the extreme gradient boosting tree model, a split point search algorithm is usually used, that is, assuming that each leaf node adds a split point, when a leaf node is decomposed into two leaf nodes, the score gain obtained is:

[0054]

[0055] In the formula, G L is the sample g that is assigned to the left leaf node i The sum of G R is the sample g that is assigned to the right leaf node i The sum of L is the sample h that is assigned to the left leaf node i The sum of R is the sample h that is assigned to the right leaf node i sum.

[0056] The main process of this node splitting principle is as follows: select a feature, scan from left to right according to a certain step size, and calculate the gain of each possible split point. Find the point with the largest gain and use it as the split point. However, if the calculated gain is less than the preset threshold γ, the split is not performed.

[0057] By solving the objective function as above, the extreme gradient boosting tree model will continuously optimize itself according to the residual in the iterative process, thereby improving the accuracy and performance of the model, and finally obtaining the expression of the extreme gradient boosting tree classification model, as shown in formula (3).

[0058] (2.3) Using Bayesian optimization to optimize model parameters

[0059] In the extreme gradient boosting tree model algorithm, the key hyperparameters that significantly affect the model performance include learning rate, maximum depth, number of iterations, subsampling ratio, minimum child node weight, and minimum loss value. In order to fully utilize the potential of the extreme gradient boosting tree model, the Bayesian optimization method is used to tune the hyperparameters of the model. The Bayesian optimization method constructs a probability model of the objective function and iteratively selects the best hyperparameter combination. The main optimization steps are as follows:

[0060] First, the accuracy of the validation set is defined as the objective function of Bayesian optimization, and the key hyperparameters and their search ranges of the extreme gradient boosting tree model are defined; secondly, a Gaussian process is constructed as a surrogate model to approximate the objective function and predict the effects of different hyperparameter combinations; then, the hyperparameter combination to be evaluated is selected by defining an acquisition function, balancing the exploration of new areas and the use of known optimal areas; then, cross-validation is used to evaluate the performance of the selected hyperparameter combination on the training data, the surrogate model is updated and the acquisition function is used to select a new hyperparameter combination for evaluation, and the above process is repeated until the preset number of iterations is reached or the objective function converges; finally, Bayesian optimization will find the hyperparameter combination that performs best on the objective function and complete the parameter optimization process of the extreme gradient boosting tree model.

[0061] (2.4) Measuring model performance

[0062] The test set K test Enter the extreme gradient boosting tree classification model under the optimal hyperparameter combination to test the performance of the extreme gradient boosting tree model. Accuracy, precision, recall and F1 score are used to comprehensively measure the model prediction performance indicators. Accuracy A refers to the proportion of correctly classified samples to the total number of samples; precision P refers to the proportion of true positive classes in all results predicted as positive classes; recall R refers to the proportion of true positive examples to all actual positive examples; F1 score is the harmonic mean of precision and recall, which is a comprehensive indicator used to evaluate the performance of the classification model. The calculation formulas for accuracy A, precision P, recall R and F1 score are as follows:

[0063]

[0064]

[0065] In the formula, TP represents the number of longitudinal fissure samples correctly predicted by the model as longitudinal fissure; TN represents the number of normal samples correctly predicted by the model as normal; FP represents the number of normal samples incorrectly predicted by the model as longitudinal fissure; FN represents the number of longitudinal fissure samples incorrectly predicted by the model as normal.

[0066] Step 3: Extract the nonlinear mapping function of longitudinal fissure

[0067] In the extreme gradient boosting tree classification model, the decision value of each sample represents the model's original prediction score for the sample, which is accumulated by the output results of each decision tree. The decision value of the extreme gradient boosting tree classification model has positive and negative properties. The predicted category of the sample can be determined by the sign of the model decision value, where negative values ​​represent normal working conditions and positive values ​​represent longitudinal cracking conditions. The absolute value of the model decision value indicates the model's prediction strength for the sample. The larger the absolute value, the more confident the model is in the prediction result, and the more obvious the tendency of the corresponding category. Therefore, the decision function of the extreme gradient boosting tree classification model under the optimal hyperparameter combination can be regarded as a nonlinear mapping function, which is used to evaluate the tendency of longitudinal cracks. The larger the absolute value of the positive decision value, the greater the tendency of longitudinal cracks, and the larger the absolute value of the negative decision value, the greater the tendency of normal working conditions.

[0068] The independent variable of the nonlinear mapping function is the steel composition data sample X, the dependent variable is the decision value of the optimal limit gradient boosting tree model, and the longitudinal crack nonlinear mapping function F e The specific expression of (x) is as follows:

[0069]

[0070] In the formula, F e (X) represents the prediction result of the extreme gradient boosting tree model under the optimal hyperparameter combination obtained in the second step, f u (X) represents the prediction result of the u-th decision tree, and U represents the total number of decision trees in the current model.

[0071] Step 4: Use particle swarm algorithm to optimize steel composition

[0072] The specific steps of optimizing the composition of peritectic steel using particle swarm optimization algorithm are as follows:

[0073] (4.1) Initialize particle population

[0074] The particle swarm algorithm iteratively searches for the optimal solution to the problem through the collaborative work of each particle. M particles are randomly initialized in the search space as the initial particle population P(0). Each particle represents a possible solution and has two attributes: position and velocity. The current steel composition data sample has c-1 input features, that is, c-1 chemical element components, that is, the target problem is a c-1-dimensional optimization problem. The position of the i-th particle in the population in the c-1-dimensional solution space can be expressed as x i =(x i,1 ,x i,2 ,...,x i,c-1 ), the velocity of the ith particle in the c-1 dimensional solution space can be expressed as v i =(v i,1 ,v i,2 ,...,vi,c-1 ).

[0075] (4.2) Determine the fitness function

[0076] In order to achieve the purpose of reducing the tendency of longitudinal cracking by optimizing the steel composition, the nonlinear mapping function obtained in the third step can be used as the fitness function of particle swarm optimization, as shown in formula (21). The main goal of steel composition optimization is to reduce the tendency of longitudinal cracking of the ingot as much as possible, that is, to minimize the fitness function value.

[0077] (4.3) Update individual optimal position and global optimal position

[0078] The fitness function determined by (4.2) is used to calculate the fitness values ​​of the M particles in the initial particle population P(0). According to the optimization requirement of minimizing the fitness value, the i-th particle is recorded at the current optimal position p i By comparing the optimal positions of the particles, the global optimal position g of the current P(0) particle population is recorded to ensure and guide the particles to search for the optimal solution.

[0079] (4.4) Update speed and position

[0080] In particle swarm optimization, the speed and position of each particle are updated in each generation. The speed update formula of the i-th particle in the t+1 generation particle population P(t+1) is as follows:

[0081] v i (t+1)=wv i (t)+c 1 r 1 [p i (t)-x i (t)]+c 2 r 2 [g(t)-x i (t)](22)

[0082] In the formula, v i (t+1) is the velocity vector of the ith particle in the t+1th generation; v i (t) is the velocity vector of the ith particle in the tth generation; p i (t) is the optimal position vector of the ith particle in the tth generation; x i (t) is the position vector of the ith particle in the tth generation; g(t) is the global optimal position vector of the tth generation; w is the inertia weight, which controls the inertia of the particle; c 1 and c 2 is the acceleration constant, which controls the acceleration of the particle to the personal optimal position and the global optimal position; r 1 and r 2A random number between 0 and 1 that increases the randomness of the algorithm.

[0083] The position update formula of the i-th particle in the t+1 generation particle population P(t+1) is as follows:

[0084] x i (t+1)=x i (t)+v i (t+1)(23)

[0085] In the formula, x i (t+1) is the position vector of the ith particle in the t+1th generation.

[0086] When the updated position exceeds the upper or lower limit of the position, the following out-of-bounds processing is required:

[0087]

[0088] In the formula, x max is the upper limit of the position; x min is the lower limit of the position.

[0089] (4.5) Iterative Optimization

[0090] Repeat the above steps (4.2) to (4.4), and continuously update and optimize each generation of particle swarm until the maximum number of iterations T set by the particle swarm algorithm is reached, thus ending the particle swarm iterative optimization process.

[0091] (4.6) Extraction of steel grade composition optimization results

[0092] Through the particle swarm update optimization process of steps (4.1) to (4.5), the optimal population P(T) is determined. The optimal population at this time is defined as the matrix P M×c , where M is the number of rows, representing M particles in the population; c is the number of columns, and the 1st to c-1 columns represent the c-1 characteristics of the particles, that is, the components corresponding to c-1 chemical elements, and the cth column represents the fitness function value corresponding to the individual.

[0093] Extract the optimal population database P M×c The value range corresponding to the c-1 property is obtained, and the value range corresponding to the composition of each chemical element is obtained, and finally the optimized composition range of each element in microalloy peritectic steel is obtained.

[0094] Furthermore, the above method of improving the longitudinal cracks on the surface of the ingot by optimizing the chemical element composition of the peritectic steel is applicable to ingot types such as slabs, square billets, round billets, and profiled billets.

[0095] The beneficial effects of the present invention are:

[0096] The present invention establishes a composition optimization model for continuous casting microalloyed peritectic steel based on extreme gradient boosting tree and particle swarm algorithm for the specific composition of microalloyed peritectic steel, and uses it in the production practice of continuous casting to improve the surface quality of the ingot and reduce the tendency of longitudinal cracking. The intelligent optimization method integrating extreme gradient boosting tree and particle swarm algorithm gives full play to the advantages of machine learning algorithm, shows high efficiency and good adaptability, and provides technical means for efficient continuous casting production and intelligent optimization of ingot quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 It is a flow chart of composition optimization of continuous casting microalloyed peritectic steel based on extreme gradient boosting tree and particle swarm algorithm;

[0098] Figure 2 This is a schematic diagram of the test results of the extreme gradient boosting tree classification model of Bayesian optimization;

[0099] Figure 3 It is the relationship between the number of particle swarm iterations and the optimal fitness. DETAILED DESCRIPTION

[0100] The present invention will be further described below through specific embodiments in conjunction with the accompanying drawings, but the present invention is not limited thereto.

[0101] A composition optimization method for continuous casting microalloyed peritectic steel based on extreme gradient boosting tree and particle swarm algorithm, comprising the following steps:

[0102] Step 1: Build a sample database of steel composition data

[0103] like Figure 1 The figure shows the composition optimization flow chart of continuous casting microalloyed peritectic steel based on extreme gradient boosting and particle swarm algorithm. Firstly, for the main components of microalloyed peritectic steel, the specific content of alloy elements is detected and recorded online, and a sample database containing normal working conditions and longitudinal cracking working conditions is constructed in combination with the billet production quality report; secondly, based on the sample library, the Bayesian algorithm is used to optimize the parameters of the extreme gradient boosting tree model to obtain the optimal extreme gradient boosting tree classification model; then, according to the optimal extreme gradient boosting tree classification model, the corresponding longitudinal crack nonlinear mapping function is extracted to evaluate the occurrence tendency of surface longitudinal cracks; finally, the particle swarm algorithm is used to search for the optimal solution of the nonlinear mapping function, so as to achieve the purpose of reducing the occurrence tendency of longitudinal cracking by optimizing the steel composition.

[0104] First, for the ingots when the casting process is stable (no process operations such as width adjustment and water inlet change are performed), the chemical element compositions of 11 microalloyed peritectic steels are selected within the same protective slag type and the specified carbon content range of 0.12% to 0.14% (the range with a higher longitudinal cracking rate), mainly including manganese (Mn), sulfur (S), phosphorus (P), copper (Cu), nitrogen (N), silicon (Si), nickel (Ni), molybdenum (Mo), chromium (Cr), titanium (Ti), aluminum (Al), etc. By tracing back the historical data of a steel plant when casting slabs, a data sample library containing the above 11 steel grades is obtained, including 419 normal slabs and 45 longitudinally cracked slabs. Each slab represents a sample, and the 11-dimensional feature is the specific content of the alloy elements in this slab. According to the production report records, the label of each sample is determined, and finally the data sample library D is obtained. 464×12 ,The distribution of each feature data is shown in Table 1.

[0105] Table 1 Data distribution of each sample feature

[0106]

[0107] Step 2: Construct the optimal extreme gradient boosting tree classification prediction model

[0108] (2.1) Dividing the model data set

[0109] According to the longitudinal fissure label record, select the data sample library D established in the first step 464×12 We randomly selected 135 normal samples from the sample library and constructed the dataset K required for the subsequent model. Then, we randomly divided the selected dataset K into training set K and training set K. train and the test set K test , the total number of samples in the training set and the test set are set to 144 and 36 respectively, while the ratio of longitudinal fissure samples to normal samples in the training set and the test set is still maintained at 1:3. The specific situation of the training set and the test set after the division is as follows:

[0110]

[0111] (2.2) Constructing the Extreme Gradient Boosting Tree Classification Prediction Model

[0112] The extreme gradient boosting tree model gradually reduces the prediction error by iteratively constructing multiple decision trees. Each tree selects the best split point through a greedy algorithm to maximize information gain. The model prevents overfitting through regularization and approximates the objective function using a second-order Taylor expansion, further improving optimization efficiency and model performance. Finally, through the weighted combination of multiple trees, the extreme gradient boosting tree model achieves powerful classification capabilities, and the resulting expression is as follows:

[0113]

[0114] In the formula, F 1 (X) represents the prediction result of the model, f k (X) represents the prediction result of the kth decision tree, and K represents the total number of decision trees.

[0115] (2.3) Using Bayesian optimization to optimize model parameters

[0116] (a) Define the key parameter search range

[0117] In the extreme gradient boosting tree model algorithm, the key hyperparameters that significantly affect the model performance include learning rate, maximum depth, number of iterations, subsampling ratio, minimum child node weight, and minimum loss value. The parameter search interval for the learning rate is set to [0.01, 0.3], the parameter search interval for the maximum depth is set to [3, 10], the parameter search interval for the number of iterations is set to [300, 500], the parameter search interval for the subsampling ratio is set to [0.6, 1.0], the parameter search interval for the minimum child node weight is set to [1, 10], and the parameter search interval for the minimum loss value is set to [0.06, 0.13].

[0118] (b) Using Bayesian algorithm to optimize parameters

[0119] Use 5-fold cross validation to evaluate the performance of each selected hyperparameter combination on the training data until the preset number of iterations is reached or the objective function converges. Finally, Bayesian optimization finds the best performing hyperparameter combination, and the corresponding parameter combinations are as follows: learning rate is 0.067, maximum depth is 3, number of iterations is 300, subsampling ratio is 1.0, minimum child node weight is 3, and minimum loss value is 0.06. At this time, the accuracy of the 5-fold cross validation of the training set is between 84.3% and 92.4%, and the average accuracy of the 5 cross validations is 88.6%. The verification results show that the model has good generalization ability.

[0120] (2.4) Measuring model performance

[0121] The test set K test Enter the extreme gradient boosting tree classification model under the optimal hyperparameter combination for testing. The test results are as follows Figure 2 As shown in the figure, the accuracy, precision, recall and F1 score of the optimal extreme gradient boosting tree classification model are 0.92, 0.8, 0.89 and 0.84 respectively, showing good prediction performance.

[0122] Step 3: Extract the nonlinear mapping function of longitudinal fissure

[0123] The extreme gradient boosting tree classification model under the optimal hyperparameter combination obtained in the second step is regarded as a function that performs nonlinear mapping and is used to evaluate the occurrence tendency of longitudinal cracks. e 1 The specific expression of (X) is as follows:

[0124]

[0125] In the formula, X represents the steel composition data sample, It represents the prediction result of the extreme gradient boosting tree model under the optimal hyperparameter combination obtained in the second step. represents the optimal prediction result of the u-th decision tree, and U represents the total number of decision trees.

[0126] Step 4: Use particle swarm algorithm to optimize steel composition

[0127] (4.1) Initialize particle population

[0128] 500 particles are randomly initialized in the search space as the initial particle population P(0). The current steel composition data sample has 11 input features, that is, 11 chemical element components, indicating that the target problem is an 11-dimensional optimization problem. In the particle swarm update process, the position and velocity of the i-th particle in the population in the 11-dimensional solution space are defined as x i =(x i,1 ,x i,2 ,...,x i,11 ) and v i =(v i,1 ,v i,2 ,...,v i,11 ), and combined with the data distribution of each sample in Table 1, the upper and lower limits of each position are defined. Position upper limit x max It is composed of the maximum values ​​of each chemical element, with a lower limit x min It is composed of the minimum values ​​of each chemical element, the details are as follows:

[0129] x max =[1.3219,0.0154,0.0205,0.0447,0.0051,0.0912,0.0469,0.0224,0.048,0.0257,0.0548]

[0130] x min =[0.3186,0.0018,0.0074,0.0104,0.0017,0.0054,0.0056,0.0001,0.0123,0.0002,0.0231]

[0131] (4.2) Determine the fitness function

[0132] In order to achieve the purpose of reducing the tendency of longitudinal cracking by optimizing the steel composition, the nonlinear mapping function can be used as the fitness function of the particle swarm optimization, as shown in formula (26). The main goal of steel composition optimization is to minimize the tendency of longitudinal cracking of the billet, which means that the fitness function value should be minimized in the particle swarm algorithm, that is, F e 1 (X) min .

[0133] (4.3) Update individual optimal position and global optimal position

[0134] The fitness function determined by (4.2) is used to calculate the fitness values ​​of the 500 particles in the initial particle population P(0). According to the optimization requirement of minimizing the fitness value, the optimal position of each particle is updated, and the global optimal position of the current P(0) particle population is updated.

[0135] (4.4) Iterative Optimization

[0136] According to the speed and position update formula, each generation of particle swarm is continuously updated and optimized to reach the maximum number of iterations set by the particle swarm algorithm, which is 200, thus ending the particle swarm iterative optimization process. Figure 3 The relationship between the number of particle swarm update iterations and the optimal fitness is shown. As the number of particle swarm update iterations increases, the optimal fitness continues to decrease. When the number of iterations reaches 200, the optimal fitness remains almost unchanged, and the global optimal position obtained at this time is as follows:

[0137] g e =[0.4888,0.0018,0.0132,0.0447,0.0029,0.0470,0.0105,0.0001,0.0415,0.0006,0.0324]

[0138] (4.5) Extraction of steel grade composition optimization results

[0139] Through the particle swarm update optimization process, the optimal population P(200) is determined. The optimal population at this time is defined as the matrix P 500×12 , where 500 represents 500 particles in the population; c is the number of columns, the 1st to 11th columns represent the 11 features of the particles, and the 12th column represents the fitness function value corresponding to the individual.

[0140] The fitness function values ​​of 500 particles in the optimal population correspond to the interval [-5.090, -7.032]. The value range corresponding to each chemical element composition in the current population database is extracted and used as the recommended interval for chemical element composition optimization. The specific optimization is shown in Table 2. Taking the S element content as an example, the variation range in production before composition optimization is 0.0018-0.0154%, and the recommended value range after optimization is 0.0018-0.0029%. For other element contents, the recommended value range can also be determined by referring to the same method. Before the optimization of steel composition, 464 ingots were cast, and 45 ingots were found to have longitudinal cracks during cleaning, with an incidence rate of 9.70%; after optimizing the steel composition according to this method, for microalloy peritectic steel sensitive to longitudinal cracks, 200 ingots were cast on site, and 4 ingots were found to have longitudinal cracks during cleaning, with an incidence rate of 2%. Before and after the optimization of chemical element composition, the incidence rate of longitudinal cracks decreased significantly, and the quality of continuous casting ingots was significantly improved.

[0141] Table 2 Comparison of chemical element composition before and after optimization

[0142]

[0143] The above-described embodiments merely express the implementation methods of the present invention, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A composition optimization method for continuous casting microalloyed peritectic steel based on extreme gradient boosting tree and particle swarm algorithm, characterized in that: The optimization method comprises the following steps: firstly, detecting and recording the specific content of alloying elements in microalloy peritectic steel online, and building a sample database including normal working conditions and longitudinal cracking working conditions in combination with the ingot production quality report; secondly, optimizing the parameters of the extreme gradient boosting tree model based on the sample database, and obtaining the extreme gradient boosting tree classification model under the optimal hyperparameter combination, so as to achieve accurate prediction of surface longitudinal cracks; then, according to the optimal extreme gradient boosting tree classification model, extracting the corresponding longitudinal crack nonlinear mapping function, and evaluating the occurrence tendency of surface longitudinal cracks; finally, searching for the optimal solution of the nonlinear mapping function, so as to achieve the purpose of reducing the occurrence tendency of longitudinal cracks by optimizing the steel grade composition.

2. The composition optimization method for continuous casting microalloyed peritectic steel based on extreme gradient boosting tree and particle swarm algorithm according to claim 1, characterized in that: The specific steps include: The first step is to build a sample library of steel composition data; For the ingots when the casting process is stable, we select microalloy peritectic steel with high longitudinal cracking incidence rate, and trace back the historical data of the steel plant within the same protective slag type and specified carbon content range, select c-1 chemical elements to be optimized that have a significant impact on the solidification characteristics and defects of the steel, and build a sample database containing normal working conditions and longitudinal cracking conditions; each sample represents a piece of ingot, and each feature of the sample is the actual chemical composition of the ingot cast; each sample is marked and recorded in combination with the production report, including normal samples and longitudinal cracking samples; The sample database obtained by backtracking is defined as the matrix D n×c , where n is the number of rows, indicating the number of samples, including normal samples and longitudinal fissure samples; c is the number of columns, the 1st to c-1 columns represent the c-1 features of the sample, i.e., c-1 elemental chemical compositions, and the cth column is the label corresponding to the sample, label 1 represents the longitudinal fissure sample, and label 0 represents the normal sample; The second step is to build the optimal extreme gradient boosting tree classification prediction model; (2.1) Dividing the model data set First, a model dataset K is constructed, which contains N samples, each of which comes from the steel composition data sample library D established in the first step. n×c ;Specifically: According to the data sample library D n×c The label records of the cth column in the data set are as follows: 1 is a longitudinal fissure sample and 0 is a normal sample. Select the data sample library D n×c All longitudinal fissure samples in the dataset and randomly select the data sample library D n×c The normal samples in constitute the data set K; Secondly, the selected data set K is randomly divided into training set K train and the test set K test , set the ratio of the number of training set samples P to the number of test set samples Q, and keep the ratio of longitudinal fissure samples and normal samples in the training set and test set unchanged; the training set K train With the test set K test The expression is as follows: K train ={(X1,0),(X2,0),(X3,0),...,(X P-2 ,1),(X P-1 ,1),(X P ,1)} (1) K test ={(X1,0),(X2,0),(X3,0),...,(X Q-2 ,1),(X Q-1 ,1),(X Q ,1)} (2) Among them, X m represents a sample, and the subscript m represents the mth sample; (2.2) Constructing the Extreme Gradient Boosting Tree Classification Prediction Model The extreme gradient boosting tree model gradually reduces the prediction error by iteratively constructing multiple decision trees; each decision tree improves the model performance by fitting its residuals based on the previous model; the extreme gradient boosting tree model structure expression is as follows: In the formula, F(X) represents the prediction result of the model, f k (X) represents the prediction result of the kth decision tree, and K represents the total number of decision trees; (2.3) Using Bayesian optimization to optimize model parameters The Bayesian optimization method is used to tune the hyperparameters of the model. The Bayesian optimization method constructs a probability model of the objective function and iteratively selects the best hyperparameter combination. (2.4) Measuring model performance The test set K test The extreme gradient boosting tree classification model under the optimal hyperparameter combination is input to test the performance of the extreme gradient boosting tree model; the accuracy, precision, recall and F1 score are used to comprehensively measure the model prediction performance indicators. Step 3: Extract the nonlinear mapping function of longitudinal fissure The decision function of the extreme gradient boosting tree classification model under the optimal hyperparameter combination is regarded as a nonlinear mapping function to evaluate the tendency of longitudinal cracks. The larger the absolute value of the positive decision value, the greater the tendency of longitudinal cracks, and the larger the absolute value of the negative decision value, the greater the tendency of normal working conditions. The independent variable of the nonlinear mapping function is the steel composition data sample X, the dependent variable is the decision value of the optimal limit gradient boosting tree model, and the longitudinal crack nonlinear mapping function F e The specific expression of (x) is as follows: In the formula, F e (X) represents the prediction result of the extreme gradient boosting tree model under the optimal hyperparameter combination obtained in the second step, f u (X) represents the prediction result of the u-th decision tree, and U represents the total number of decision trees in the current model; The fourth step is to use the particle swarm optimization algorithm to optimize the composition of peritectic steel, as follows: (4.1) Initialize the particle population; (4.2) Determine the fitness function; (4.3) Update individual optimal position and global optimal position; (4.4) Update speed and position; (4.5) Iterative optimization; (4.6) Extract the steel grade composition optimization results.

3. The composition optimization method for continuous casting microalloyed peritectic steel based on extreme gradient boosting tree and particle swarm algorithm according to claim 2, characterized in that: In the aforementioned (2.1), the ratio between the number of longitudinal fissure samples and the number of normal samples in the dataset K is 1:3; the total number N of normal samples and longitudinal fissure samples in the dataset K is not less than 150 cases.

4. The composition optimization method for continuous casting microalloyed peritectic steel based on extreme gradient boosting tree and particle swarm algorithm according to claim 2, characterized in that: In (2.2) described above, the ratio of the number of training set samples P to the number of test set samples Q is set to 4:1, and the ratio of longitudinal fissure samples to normal samples in the training set and the test set is still maintained at 1:

3.

5. The composition optimization method for continuous casting microalloyed peritectic steel based on extreme gradient boosting tree and particle swarm algorithm according to claim 2, characterized in that: The (2.2) mentioned above is specifically: (2.2.1) Define the model optimization objective function The extreme gradient boosting tree model optimizes the model parameters by minimizing the loss function. The logarithmic loss function is used as the loss function l of the extreme gradient boosting tree model, and its expression is as follows: l(and i ,p i )=-y i log(p i )-(1-and i )log(1-p i ) (4) In the formula, y i represents the true label of the i-th sample, p i Represents the probability that the i-th sample is predicted as a positive class; For a training set K containing P samples train , consisting of K decision tree models f k (X i The extreme gradient boosting tree model composed of )(k=1,2,...,K), the corresponding optimization objective function expression L(θ) is as follows: In the formula, θ refers to all parameters of the model; P represents the number of samples in the training set; f k (X i ) represents the prediction function added in the kth training; Ω[f k (X i )] represents the sum of the model complexity of the first k decision trees; (2.2.2) Objective function solution The incremental superposition training method is used to solve the objective function (5). For the i-th sample (X i ,y i ), and its superposition training expression is as follows: In the formula, p i k represents the k-th prediction value of the model; p i k-1 Represents the k-1th predicted value of the model; In formula (6), p i k Substituting into formula (5), we can get the objective function of the k-th training, which is expressed as follows: The purpose of the kth iteration is to find the most suitable f k (X i ) minimizes the objective function. In the process of optimizing the objective function, the second-order Taylor formula is used to expand equation (7) to obtain: In the formula, g i is the first-order gradient of the loss function l, as shown in formula (9); h i is the second-order gradient of the loss function l, as shown in formula (10); In order to prevent the model from overfitting, a regularization term is added to the objective function of the extreme gradient boosting tree, and the complexity Ω[f k (X i )]; the decision tree model is expressed as: In the formula, ω represents the weight function of the leaf node; q represents the structure function of the tree, which maps the sample to the leaf node index; Represents the tree for sample X i The predicted value of; d represents the sample feature dimension; For a given decision tree f k (X i ), and its complexity Ω[f k (X i )] is defined as follows: In the formula, γ represents the regularization parameter that controls the number of leaf nodes; T represents the number of leaf nodes in the tree; λ represents the regularization parameter that controls the weight of the leaf nodes; ω j Represents the weight of the jth leaf node; Assume that the training sample X i is divided into leaf node j, and the sample set on the corresponding node is I j ={i|q(X i )=j}, then substitute equation (11) and equation (12) into equation (8), and the optimization objective function is: In the formula, This formula is an independent quadratic function about T; The goal of solving the above objective function is to find the optimal decision tree, that is, to solve the optimal leaf node weight; for the tree structure q(x)(I j ={i|q(X i )=j}), by j Derivation obtains the optimal weight value of the leaf node: Substituting formula (14) into the objective function shown in formula (13), we can obtain the optimal value of the objective function: In the extreme gradient boosting tree model, a split point search algorithm is usually used, that is, assuming that each leaf node adds a split point, when a leaf node is decomposed into two leaf nodes, the score gain obtained is: In the formula, G L is the sample g that is assigned to the left leaf node i The sum of G R is the sample g that is assigned to the right leaf node i The sum of L is the sample h that is assigned to the left leaf node i The sum of R is the sample h that is assigned to the right leaf node i sum; By solving the objective function, the extreme gradient boosting tree model will continuously optimize itself according to the residual in the iterative process, thereby improving the accuracy and performance of the model, and finally obtaining the expression of the extreme gradient boosting tree classification model, as shown in formula (3).

6. The composition optimization method for continuous casting microalloyed peritectic steel based on extreme gradient boosting tree and particle swarm algorithm according to claim 2, characterized in that: The main optimization steps of the Bayesian optimization method (2.3) are as follows: First, the accuracy of the validation set is defined as the objective function of Bayesian optimization, and the key hyperparameters and their search ranges of the extreme gradient boosting tree model are defined; secondly, a Gaussian process is constructed as a surrogate model to approximate the objective function and predict the effects of different hyperparameter combinations; then, the hyperparameter combination to be evaluated is selected by defining an acquisition function, balancing the exploration of new areas and the use of known optimal areas; then, cross-validation is used to evaluate the performance of the selected hyperparameter combination on the training data, the surrogate model is updated and the acquisition function is used to select a new hyperparameter combination for evaluation, and the above process is repeated until the preset number of iterations is reached or the objective function converges; finally, Bayesian optimization will find the hyperparameter combination that performs best on the objective function and complete the parameter optimization process of the extreme gradient boosting tree model.

7. The composition optimization method for continuous casting microalloyed peritectic steel based on extreme gradient boosting tree and particle swarm algorithm according to claim 2, characterized in that: In (2.4), accuracy A refers to the ratio of correctly classified samples to the total number of samples; precision P refers to the ratio of true positive samples to all predicted positive samples; recall R refers to the ratio of true positive samples to all actual positive samples; F1 score is the harmonic mean of precision and recall, and is a comprehensive indicator used to evaluate the performance of classification models; the calculation formulas for accuracy A, precision P, recall R and F1 score are as follows: In the formula, TP represents the number of longitudinal fissure samples correctly predicted by the model as longitudinal fissure; TN represents the number of normal samples correctly predicted by the model as normal; FP represents the number of normal samples incorrectly predicted by the model as longitudinal fissure; FN represents the number of longitudinal fissure samples incorrectly predicted by the model as normal.

8. The composition optimization method for continuous casting microalloyed peritectic steel based on extreme gradient boosting tree and particle swarm algorithm according to claim 2, characterized in that: The fourth step is to use the particle swarm optimization algorithm to optimize the composition of peritectic steel, as follows: (4.1) Initialize particle population The particle swarm algorithm iteratively searches for the optimal solution of the problem through the collaborative work of each particle; randomly initialize M particles in the search space as the initial particle population P(0); the current steel grade composition data sample has c-1 input features, that is, c-1 chemical element components, that is, the target problem is a c-1-dimensional optimization problem, then the position of the i-th particle in the population in the c-1-dimensional solution space can be expressed as x i =(x i,1 ,x i,2 ,...,x i,c-1 ), the velocity of the ith particle in the c-1 dimensional solution space can be expressed as v i =(v i,1 ,v i,2 ,...,v i,c-1 ); (4.2) Determine the fitness function The nonlinear mapping function obtained in the third step is used as the fitness function of particle swarm optimization, as shown in formula (21); the main goal of steel composition optimization is to minimize the tendency of longitudinal cracking of the ingot, that is, to minimize the fitness function value; (4.3) Update individual optimal position and global optimal position The fitness function determined by (4.2) is used to calculate the fitness values ​​of the M particles in the initial particle population P(0) respectively; according to the optimization requirement of minimizing the fitness value, the i-th particle is recorded at the current optimal position p i ; By comparing the optimal positions of particles, the global optimal position g of the current P(0) particle population is recorded to ensure and guide the particles to search for the optimal solution; (4.4) Update speed and position In particle swarm optimization, the speed and position of each particle are updated in each generation; the speed update formula of the i-th particle in the t+1 generation particle population P(t+1) is as follows: v i (t+1)=wv i (t)+c1r1[p i (t)-x i (t)]+c2r2[g(t)-x i (t)](22) In the formula, v i (t+1) is the velocity vector of the ith particle in the t+1th generation; v i (t) is the velocity vector of the ith particle in the tth generation; p i (t) is the optimal position vector of the ith particle in the tth generation; x i (t) is the position vector of the ith particle in the tth generation; g(t) is the global optimal position vector of the tth generation; w is the inertia weight, which controls the inertia of the particle; c1 and c2 are acceleration constants, which control the acceleration of the particle to the personal optimal position and the global optimal position; r1 and r2 are random numbers between 0 and 1, which increase the randomness of the algorithm; The position update formula of the i-th particle in the t+1 generation particle population P(t+1) is as follows: x i (t+1)=x i (t)+v i (t+1) (23) In the formula, x i (t+1) is the position vector of the ith particle in the t+1th generation; When the updated position exceeds the upper or lower limit of the position, the following out-of-bounds processing is performed: In the formula, x max is the upper limit of the position; x min is the lower limit of the position; (4.5) Iterative Optimization Repeat the above steps (4.2) to (4.4), and continuously update and optimize each generation of particle swarm until the maximum number of iterations T set by the particle swarm algorithm is reached, thus ending the particle swarm iterative optimization process; (4.6) Extraction of steel grade composition optimization results Through the particle swarm update optimization process of steps (4.1) to (4.5), the optimal population P(T) is determined; The optimal population at this time is defined as the matrix P M×c , where M is the number of rows, representing M particles in the population; c is the number of columns, the 1st to c-1 columns represent the c-1 features of the particles, that is, the components corresponding to c-1 chemical elements, and the cth column represents the fitness function value corresponding to the individual; Extract the optimal population database P M×c The value range corresponding to the c-1 property is obtained, and the value range corresponding to the composition of each chemical element is obtained, and finally the optimized composition range of each element in microalloy peritectic steel is obtained.

9. A method for optimizing the composition of continuous casting microalloyed peritectic steel based on extreme gradient boosting tree and particle swarm algorithm according to any one of claims 1 to 8, characterized in that: The method for improving the longitudinal cracks on the surface of the ingot by optimizing the chemical element composition of the peritectic steel is applicable to slabs, square billets, round billets, profile billets or other types of ingots.

Citation Information

Patent Citations

  • Method for preventing surface cracks on low-carbon boron-containing steel

    CN107630172A

  • Method for reducing corner cracks of titanium-containing peritectic steel continuous casting slab

    CN112813355A

  • Smelting method of P690QL2 marine storage tank steel

    CN115747616A

  • Continuous casting slab surface longitudinal crack process optimization method based on support vector machine and genetic algorithm

    CN118296476A