Method and system for predicting rolling force of last pass

Abstract

The invention provides a last-pass rolling force prediction method and system, and relates to the technical field of rolling technologies. The method comprises the steps that a data set is obtained; screening features associated with the rolling force of the last pass in the data set by using a feature engineering technology to obtain a training set; constructing a quantile deep learning network model; training the quantile deep learning network model by using the training set to obtain a target quantile deep learning network model; the target quantile deep learning network model is configured with a plurality of quantile models of different quantiles; inputting process parameter data of a to-be-measured slab into the target quantile deep learning network model, and obtaining a final-pass rolling force predicted value of the to-be-measured slab according to the final-pass rolling reduction; the invention aims to solve the problems that an accurate and reliable prediction result cannot be provided and the high-precision requirement of actual production is difficult to meet when the last-pass rolling force is predicted due to the fact that the current last-pass rolling force prediction method lacks attention and prediction capability on an extreme value.

Description

Technical Field

[0001] The present application relates to the field of rolling process technology, and in particular to a method and system for predicting the final rolling force. Background Art

[0002] In the complex process of steel production, rolling plays a central role in shaping raw materials into key products such as plates and profiles. Accurately predicting the rolling force during the final pass is crucial for ensuring the final thickness accuracy of the plates and improving product surface quality.

[0003] In practice, due to factors such as site conditions, equipment performance, and production safety, the final rolling pass is subject to multiple constraints, including force limits and strict regulation of reduction distribution. These factors combine to lead to an imbalanced data distribution, meaning that some key data (especially rolling force data under extreme conditions) is relatively scarce, posing a significant challenge to accurate prediction. To address this data imbalance, data balancing techniques are currently being used to balance rolling force data, such as the synthetic minority oversampling technique (SMOTE), the random oversampling and undersampling method (ROSE), and sample augmentation strategies.

[0004] While the aforementioned methods are somewhat effective in addressing general data imbalance issues, they lack the ability to focus on and predict extreme values when faced with data imbalance in the highly complex and specific working conditions of steel rolling. In actual rolling, the extreme values of the final-pass rolling force often play a decisive role in determining the quality limit of the plate. However, existing prediction technologies mostly focus on predicting average values and are unable to predict extreme values in rolling force data. This results in an inability to provide accurate and reliable prediction results when processing final-pass rolling force predictions, making it difficult to meet the high-precision requirements of actual production. Summary of the Invention

[0005] The present application provides a method and system for predicting the last-pass rolling force, which makes up for the current inability to predict the last-pass rolling force separately. By integrating the advantages of quantile regression and deep learning, combined with intelligent optimization algorithms, a method for constructing a target quantile deep learning network model is provided to solve the technical problem that the current full-pass rolling force model cannot take into account the extreme and heteroscedastic characteristics in the last-pass rolling force data when predicting the last-pass rolling force, making it difficult for the last-pass rolling force prediction results to meet the high-precision requirements of actual production.

[0006] In a first aspect, the present application provides a method for predicting the final rolling force, comprising:

[0007] Acquire a data set; the data set includes: chemical composition of a plurality of slabs, rolling process parameters of each pass, and rolling force;

[0008] Using feature engineering technology to screen features associated with the final-pass rolling force in the data set to obtain a training set;

[0009] Build a quantile deep learning network model;

[0010] The quantile deep learning network model is trained using the training set to obtain a target quantile deep learning network model; the target quantile deep learning network model is configured with quantile models of several different quantiles; the target quantile deep learning network model is configured to determine the quantile corresponding to the final pass reduction according to the final pass reduction; and the quantile model corresponding to the quantile is determined according to the quantile;

[0011] The process parameter data of the slab to be tested is input into the target quantile deep learning network model, and the predicted value of the final rolling force of the slab to be tested is obtained according to the final rolling reduction; the process parameter data includes: the chemical composition of the slab to be tested and the final rolling process parameters.

[0012] In some embodiments, after the step of obtaining the data set, the following steps are included:

[0013] Based on the data set, the upper limit and lower limit of each indicator in the data set are obtained using statistical methods.

[0014] Determine the setting range of process parameters;

[0015] Based on the upper limit and lower limit of each indicator in the data set, the process parameter setting range is used to obtain abnormal values, and the abnormal values are cleared and filled.

[0016] In some embodiments, the step of using feature engineering technology to screen the features associated with the final rolling force in the data set to obtain a training set includes:

[0017] According to the data set, a first feature is obtained; the first feature is a feature in the data set associated with the final rolling force;

[0018] Based on the first feature, using feature engineering technology to perform reduction and expansion reconstruction operations on the first feature to obtain a second feature;

[0019] Based on the second feature, a training set is obtained using a standardization method.

[0020] In some embodiments, the step of training the quantile deep learning network model using the training set to obtain a target quantile deep learning network model includes:

[0021] Based on the quantile deep learning network model, using an intelligent optimization algorithm, determining the hyperparameters of the quantile deep learning network model;

[0022] The quantile deep learning network model is trained using the training set to minimize the quantile regression loss function and obtain a target quantile deep learning network model.

[0023] In some embodiments, before the step of inputting the process parameter data of the slab to be tested into the target quantile deep learning network model, the step includes:

[0024] Acquire first process parameter sub-data; the first process parameter sub-data includes: the chemical composition of the slab to be tested and the final rolling process parameters;

[0025] Based on the first process parameter sub-data, using a statistical method to obtain abnormal values, and performing clearing and filling processing on the abnormal values;

[0026] Based on the first process parameter sub-data, using feature engineering technology to perform reduction and expansion reconstruction operations on the first process parameter sub-data to obtain second process parameter sub-data;

[0027] Based on the second process parameter sub-data, process parameter data is obtained using a standardization method.

[0028] In some embodiments, the step of inputting the process parameter data of the slab to be tested into the target quantile deep learning network model and obtaining the predicted value of the final rolling force of the slab to be tested according to the final rolling reduction comprises:

[0029] Determining a target quantile model according to the final pass reduction;

[0030] Inputting process parameter data of the slab to be tested into the target quantile model, and determining a target data path in the target quantile model that matches the process parameter data;

[0031] Based on the target data path, a predicted value of the final rolling force of the slab to be tested is obtained.

[0032] In some embodiments, the step of determining a target quantile model based on the final pass reduction includes:

[0033] Determining a quantile corresponding to the final pass reduction according to the final pass reduction;

[0034] Determining a target quantile based on the quantile;

[0035] According to the target quantile, a target quantile model corresponding to the target quantile is determined.

[0036] In some embodiments, the chemical components include: carbon, silicon, manganese, sulfur, and phosphorus; the rolling process parameters include: temperature, inlet and outlet widths, inlet and outlet thicknesses, and reduction.

[0037] A second aspect of the present application provides a final-pass rolling force prediction system, comprising:

[0038] An acquisition module is configured to acquire a data set; the data set includes: chemical composition of a plurality of slabs, rolling process parameters of each pass, and rolling force;

[0039] a screening module configured to screen features associated with the final-pass rolling force in the data set using feature engineering technology to obtain a training set;

[0040] A building module, configured to build a quantile deep learning network model;

[0041] a training module configured to train the quantile deep learning network model using the training set to obtain a target quantile deep learning network model; the target quantile deep learning network model is configured with quantile models of several different quantiles; the target quantile deep learning network model is configured to determine the quantile corresponding to the final pass reduction according to the final pass reduction; and determine the quantile model corresponding to the quantile according to the quantile;

[0042] The output module is configured to input the process parameter data of the slab to be tested into the target quantile deep learning network model, and obtain the predicted value of the final rolling force of the slab to be tested based on the final rolling reduction; the process parameter data includes: the chemical composition of the slab to be tested and the final rolling process parameters.

[0043] The present application provides a method and system for predicting the final rolling force, the method comprising: obtaining a data set; the data set comprising: the chemical composition of a plurality of slabs, rolling process parameters of each pass, and rolling force; using feature engineering technology to screen features associated with the final rolling force in the data set to obtain a training set; constructing a quantile deep learning network model; using the training set to train the quantile deep learning network model to obtain a target quantile deep learning network model; the target quantile deep learning network model is configured with quantile models of a plurality of different quantiles; the target quantile deep learning network model is configured as According to the final reduction, the quantile corresponding to the final reduction is determined; according to the quantile, the quantile point model corresponding to the quantile is determined; the process parameter data of the slab to be tested is input into the target quantile deep learning network model, and according to the final reduction, the final rolling force prediction value of the slab to be tested is obtained; the process parameter data include: the chemical composition of the slab to be tested and the final rolling process parameters, so that the rolling force prediction method can predict the extreme value of the final rolling force, thereby providing accurate and reliable prediction results when predicting the final rolling force, so as to meet the high precision requirements of actual production. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solution of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0045] Figure 1 This is a flow chart of the method for predicting the rolling force of the last pass in this application;

[0046] Figure 2 Flowchart for predicting the final-pass rolling force using the target quantile deep learning network model in this application;

[0047] Figure 3 This is the architecture diagram of the target neural network model in this application;

[0048] Figure 4 This is a comparison chart of the hit rate within 1000KN of the predicted deviation of the final rolling force of each slab in this application;

[0049] Figure 5 This is a histogram comparison chart of the predicted deviation of the final rolling force of each slab within 1000KN in this application. DETAILED DESCRIPTION

[0050] In order to enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0051] Because some technologies lack the ability to focus on and predict extreme values in rolling force prediction methods, neural network models are unable to provide accurate and reliable prediction results when predicting the final-pass rolling force, making it difficult to meet the high-precision requirements of actual production. To address this technical problem, the present application provides a final-pass rolling force prediction method and system. The final-pass rolling force prediction method and system are described below:

[0052] For example, in the complex process system of steel production, the rolling process plays a core role in shaping raw materials into key products such as plates and profiles. In this process, accurate prediction of the rolling force of the last pass is of great significance to ensuring the final thickness accuracy of the plate and improving the surface quality of the product. However, the current research focus of rolling force prediction technology is mostly concentrated on the comprehensive consideration of all passes in the entire rolling process, ignoring the challenges and limitations faced by the last pass in the actual production environment. In actual operation, due to factors such as site conditions, equipment performance and production safety, the last pass of rolling is often subject to multiple constraints such as force size limitation and strict control of reduction distribution. These factors work together to lead to an imbalance in data distribution, that is, some key data (especially rolling force data under extreme conditions) are relatively scarce, which brings huge challenges to accurate prediction.

[0053] For example, current data balancing technologies, such as the synthetic minority oversampling technique (SMOTE), the random oversampling and undersampling method (ROSE), and sample augmentation strategies, while effective in addressing general data imbalance, are increasingly limited when faced with data imbalance in the highly complex and specific working conditions of steel rolling. These technologies often struggle to fully capture and reflect the subtle changes and extreme working conditions of the on-site rolling process, particularly the critical information at the tail of the rolling force distribution. This limits the accuracy and generalization capabilities of the prediction models, resulting in insufficiently accurate and reliable prediction results for the final-pass rolling force prediction, making it difficult to meet the high-precision requirements of actual production.

[0054] Finally, current rolling force prediction technologies lack the ability to address and predict extreme values. In actual rolling, extreme values of the final-pass rolling force often play a decisive role in determining the quality limit of the plate. However, existing prediction technologies mostly focus on predicting average values and are unable to predict extreme rolling force variations. This leads to significant deviations between predicted results and actual requirements during the final-pass rolling process, which requires high-precision control.

[0055] like Figure 1 As shown, it is the flow chart of the method for predicting the rolling force of the last pass in this application.

[0056] In response to the above problems, the first aspect of the present application provides a method for predicting the final rolling force, comprising the following steps:

[0057] S100: Acquire a data set; the data set includes: chemical composition of several slabs, rolling process parameters of each pass, and rolling force; the chemical composition includes: carbon, silicon, manganese, sulfur, phosphorus, etc.; the rolling process parameters include: temperature, inlet and outlet widths, inlet and outlet thicknesses, reduction, etc.

[0058] For example, the chemical composition of each slab and its rolling process parameters in each pass are systematically collected from the entire process of steel production. The data includes physical properties such as temperature, width, and thickness, as well as detailed information on key alloying elements (such as carbon C, silicon Si, manganese Mn, sulfur S, phosphorus P, etc.). Through comprehensive data collection, it is ensured that the model can be trained based on the most detailed information, providing a solid foundation for subsequent analysis. Among them, by paying attention to the data correlation at different stages, various factors that may affect the quality of the final product in the production process are obtained, thereby continuously improving the product information in the data set to reduce the deviation between the prediction results and actual demand.

[0059] After the step of obtaining the data set, the following steps are included:

[0060] S110: Based on the data set, using statistical methods to obtain the upper and lower limits of each indicator in the data set; using the data set, using statistical methods to obtain the upper and lower limits of each indicator from the data distribution level by adding percentile lines.

[0061] S120: Determine the process parameter setting range; the staff sets the process parameter setting range based on professional knowledge in this field, and the process parameter setting range is used as an indicator for screening abnormal values.

[0062] S130: Based on the upper limit and lower limit of each indicator in the data set, using the process parameter setting range, obtaining abnormal values, and clearing and filling the abnormal values.

[0063] Exemplarily, the upper and lower limits of each indicator in the data set constitute an indicator interval, and the values in the indicator interval that are outside the set range of the process parameters are outliers. Statistical methods (such as mean square error, box plot or cluster analysis) are used to perform data cleaning operations on the outliers, and the outliers in the data set are cleared and filled, thereby improving the quality of the data set and ensuring the consistency and integrity of the data set.

[0064] For example, taking the box plot method as an example, by drawing a box plot, you can visually observe the distribution of data in the data set, and thus obtain which data points are beyond the normal range. Using the mean square error and the box plot method combined with business experience to determine the upper and lower limits of the abnormal points, first calculate the mean value u and standard deviation s of each indicator. d , the third quarter point Q3 and the fourth quarter point Q1, then the upper limit of the abnormal point H max , lower limit H min They are:

[0065] H max =[Q3+1.5×(Q3-Q1),u+3×s d , upper limit of business threshold] max ;

[0066] H min =[Q3-1.5×(Q3-Q1),u-3×s d , business threshold lower limit] min ;

[0067] Among them, H max =Q3+1.5×(Q3-Q1), u+3×s d and the upper limit of the business threshold; H min Q3-1.5×(Q3-Q1), u-3×s d The minimum value among the three values of H and the lower limit of the business threshold; the upper and lower limits of the business threshold are determined based on business experience. max 、H min Data points with outliers are considered outliers. Outliers can be removed by directly deleting the rows or columns containing them. By removing outliers that adversely affect data analysis or modeling, the accuracy and stability of the target neural network model generated subsequently are improved.

[0068] For example, missing values in the dataset are filled using statistical filling methods. For missing or abnormal data points, advanced filling strategies can be used, such as time series trend filling or median filling of adjacent samples, to maintain the overall continuity of the data. Furthermore, data consistency checks can be performed to ensure that the logical relationships between all relevant variables are correct, thereby providing high-quality data support for the subsequent establishment of the target neural network model.

[0069] For example, taking the median filling method as an example, by calculating the median of the non-missing values in the data set, using the median to fill all missing values, and using statistics (such as mean, median, etc.) to fill missing values, data deviation can be reduced to a certain extent.

[0070] S200: Utilize feature engineering technology to screen the features associated with the last-pass rolling force in the data set to obtain a training set; the training set includes: the features associated with the last-pass rolling force in the data set; and through in-depth analysis of the data set, key indicators that have a significant impact on the non-last-pass rolling force are mined. In addition to the chemical composition and process parameters used directly, mechanism knowledge is introduced as an auxiliary feature, and feature engineering technology is used to reduce and expand the reconstruction to generate more representative new features. Subsequently, an appropriate normalization method (such as min-max normalization, z-score normalization) is selected according to the current data distribution to ensure that data of different dimensions can be compared at the same scale and optimize the learning effect of the model. The training set is the best input prepared for model training, thereby enhancing the predictive ability of the target neural network model generated subsequently.

[0071] The step of using feature engineering technology to screen the features associated with the final-pass rolling force in the data set to obtain a training set includes the following sub-steps:

[0072] S210: Obtain a first feature based on the data set; the first feature is a feature associated with the final rolling force in the data set; data on the final rolling force, such as the temperature, width, thickness and other data of the final slab, are searched from the data set; wherein, through mechanism knowledge, an indicator that can reflect the features in the data set can be designed, and the indicator can expand the feature quantity of the data set, thereby reducing the deviation between the prediction result and the actual demand.

[0073] S220: Based on the first feature, perform reduction and expansion reconstruction operations on the first feature using feature engineering technology to obtain a second feature. The second feature includes a collection of some features in the first feature.

[0074] For example, feature reduction aims to reduce the number of features while preserving as much key information as possible from the original data, thereby reducing model complexity, reducing computational costs, and improving model generalization. Feature expansion and reconstruction aims to enhance the expressiveness of data by constructing new features or combining existing features, thereby revealing hidden patterns and correlations in the data and improving the model's predictive performance.

[0075] S230: Based on the second feature, a training set is obtained using a standardization method.

[0076] For example, an appropriate normalization method, such as the Z-Score normalization method, is used to normalize the second feature, converting data of different magnitudes into a score of a unified scale, thereby making the data features of different metrics comparable. Using the training set obtained by the normalization method to train the model to be trained can improve model performance, accelerate model training, enhance model generalization capabilities, and facilitate data comparison and analysis.

[0077] S300: Construct a quantile deep learning network model. The quantile deep learning network model is a model that combines deep neural networks and quantile regression, and is used for the subsequent prediction of the final rolling force. The quantile deep learning network model can perform interval prediction: compared with the point prediction method, the quantile deep learning network model can predict the predicted values at different quantiles, thereby constructing a prediction interval to more comprehensively reflect the uncertainty of the prediction; nonlinear feature capture: the deep neural network part can extract complex features in time series data, including nonlinear features; heteroscedasticity processing: the quantile regression part can capture the heteroscedasticity in the data, that is, the uncertainty of the predicted value may vary with the independent variable.

[0078] like Figure 2 As shown, this is a flow chart of the target quantile deep learning network model used in this application to predict the final rolling force.

[0079] S400: The quantile deep learning network model is trained using the training set to obtain a target quantile deep learning network model; the target quantile deep learning network model is configured with quantile models of several different quantiles; the target quantile deep learning network model is configured to determine the quantile corresponding to the final pass reduction according to the final pass reduction; determine the quantile model corresponding to the quantile according to the quantile; perform model training at different quantiles according to the reduction to obtain a target quantile deep learning network model and comprehensively estimate the rolling force distribution. By replacing the traditional mean loss function with the quantile regression loss function, the model can better capture the information of the tail of the rolling force distribution, especially showing higher accuracy in the prediction of extreme values. This method helps to more accurately understand the variation law of non-final pass rolling force.

[0080] The step of using the training set to train the quantile deep learning network model to obtain a target quantile deep learning network model includes the following sub-steps:

[0081] S410: Based on the quantile deep learning network model, the hyperparameters of the quantile deep learning network model are determined using an intelligent optimization algorithm; in order to determine the optimal hyperparameters of the quantile deep learning network model, an intelligent optimization algorithm (such as genetic algorithm, particle swarm optimization, whale optimization algorithm, etc.) is used to traverse the possible parameter space. Intelligent optimization algorithms can be used in combination to search for the optimal solution space and enhance the robustness of the model. The hyperparameters include: the number of layers, the number of nodes, the learning rate, the activation function, etc. of the neural network. Figure 3 As shown, for example, by using the whale optimization algorithm, the number of hidden layers, the number of nodes, the learning rate, and the activation function of the model to be trained can be determined.

[0082] Exemplarily, the steps of determining the hyperparameters of the model to be trained using an intelligent optimization algorithm are as follows: first, according to the selected intelligent optimization algorithm, a population or set containing multiple candidate solutions (model configurations) is initialized; the fitness of each candidate solution is evaluated, that is, its performance on the training set is calculated using a loss function; selection, crossover, mutation and other operations are performed on the population or set until the set number of iterations or convergence conditions are reached; finally, the individual with the highest fitness is selected from the final population / set as the optimal solution, that is, the model hyperparameters and initial weights are output as the results.

[0083] S420: Using the training set to train the quantile deep learning network model, the quantile regression loss function is minimized to obtain a target quantile deep learning network model. After optimizing the hyperparameters of the quantile deep learning network model using an intelligent optimization algorithm, the quantile deep learning network model is trained using the training set, and the model weights are adjusted to minimize the quantile regression loss function, ensuring that the model can capture complex nonlinear relationships and improve prediction accuracy.

[0084] Exemplarily, the specific steps of training the quantile deep learning network model using the training set are as follows: first, define the quantile regression loss function: according to the principle of quantile regression, define the quantile loss function; train the model to be trained: use the training set and the defined quantile regression loss function, and iteratively update the weights of the model to be trained through an optimization algorithm (such as gradient descent, quantum gradient descent, etc.); finally, in each iteration, calculate the loss function and update the weights to minimize the value of the loss function.

[0085] Exemplarily, the quantile deep learning network model is trained using the training set, and the resulting target quantile deep learning network model has the following advantages: by minimizing the quantile regression loss function, the target quantile deep learning network model can more accurately predict different quantiles of the target variable, thereby providing more comprehensive prediction information; the quantile regression loss function imposes different penalties on situations where the predicted values are too high or too low, thereby helping the model to maintain robustness when dealing with outliers or noisy data; the target quantile deep learning network model can provide detailed information about the distribution of the target variable, which helps to solve complex decision-making processes.

[0086] It is worth noting that the intelligent optimization algorithm is not limited in this embodiment and can be used as one or in combination of multiple algorithms.

[0087] like Figure 3 As shown, the architecture diagram of the target neural network model in this application.

[0088] Exemplarily, the parameter configuration obtained by the intelligent optimization algorithm is obtained through iterative search, which may involve the coordinated optimization of multiple parameters. Due to the randomness and exploratory nature of the intelligent optimization algorithm, the parameter configuration obtained may not be unique and may be affected by the initial conditions and algorithm parameters. The parameter configuration obtained by introducing the regularization term is obtained by adding the regularization term to the loss function through an optimization algorithm (such as gradient descent). The existence of the regularization term constrains the model parameters during the training process, thereby tending to select simpler and more generalized models.

[0089] Before the step of inputting the process parameter data of the slab to be tested into the target quantile deep learning network model, the following steps are included:

[0090] S430: Obtain the first process parameter sub-data; the first process parameter sub-data includes: the chemical composition of the slab to be tested and the last rolling process parameters; S440: Based on the first process parameter sub-data, obtain abnormal values using statistical methods, and clear and fill the abnormal values; S450: Based on the first process parameter sub-data, fill the missing values in the first process parameter sub-data using statistical filling method; S460: Based on the first process parameter sub-data, use feature engineering technology to reduce and expand the first process parameter sub-data to obtain the second process parameter sub-data; S470: Based on the second process parameter sub-data, obtain process parameter data using standardization method.

[0091] It is worth noting that the effects of the above-mentioned embodiment process of obtaining process parameter data can be referred to the effects of the above-mentioned embodiment process of obtaining the training set, which will not be repeated here.

[0092] S500: Inputting the process parameter data of the slab to be tested into the target quantile deep learning network model, and obtaining a predicted value of the final-pass rolling force of the slab to be tested based on the final-pass rolling reduction. The process parameter data includes the chemical composition of the slab to be tested and the final-pass rolling process parameters. The rolling process parameters include temperature, inlet and outlet widths, inlet and outlet thicknesses, and reduction. Based on the final-pass reduction, an appropriate quantile (i.e., the output node of the target neural network model) is dynamically selected for prediction. For example, if the final reduction is large, a higher quantile (e.g., the 90th percentile) can be selected to reflect a larger rolling force estimate; if the final reduction is small, a lower quantile (e.g., the 10th percentile) can be selected to reflect a smaller rolling force estimate. Through this approach, the target neural network model can flexibly adjust its prediction strategy based on actual conditions, ensuring the accuracy and reliability of the final-pass rolling force prediction, thereby providing strong support for precise control of the thickness of the slab to be processed.

[0093] The step of inputting the process parameter data of the slab to be tested into the target quantile deep learning network model and obtaining the predicted value of the final rolling force of the slab to be tested according to the final rolling reduction comprises the following steps:

[0094] S510: Determine a target quantile model according to the final pass reduction.

[0095] The step of determining the target quantile model according to the final pass reduction comprises the following sub-steps:

[0096] S511: Determine the quantile corresponding to the last-pass reduction based on the last-pass reduction; the quantile corresponding to the last-pass reduction, if the last-pass reductions are arranged from large to small, and the last-pass reduction to be measured is at the 10% quantile, that is, the last-pass reduction corresponding to the 10% quantile, then corresponds to the value at the 10% position in the last-pass reduction sample arranged from large to small.

[0097] S512: Determine a target quantile based on the quantile. The target quantile can be determined as a point through the quantile. For example, the target quantile model may include a 10% quantile, a 20% quantile, etc. The quantile corresponding to the final reduction is the 13% quantile. If there is no quantile equal to it, the quantile closest to it, such as the 10% quantile, is selected as the target quantile.

[0098] S513: Determine the target quantile model corresponding to the target quantile according to the target quantile. According to the target quantile, the target quantile model can be obtained from a plurality of quantile models.

[0099] S520: Inputting the process parameter data of the slab to be tested into the target quantile model, and determining a target data path in the target quantile model that matches the process parameter data. It is understood that the target quantile model stores a plurality of process parameter data and corresponding final-pass rolling force prediction values. By inputting the process parameter data of the slab to be tested into the target quantile model, a target data path in the target quantile model that matches the process parameter data is determined.

[0100] S530: Based on the target data path, a predicted value of the rolling force of the last pass of the slab to be tested is obtained. When the target data path is confirmed, the rolling force value corresponding to the target output node is output as an output result, thereby obtaining a predicted value of the rolling force of the last pass of the slab to be tested.

[0101] Exemplarily, when the target neural network model obtains the final pass reduction, it first obtains the corresponding quantile based on the final pass reduction value to determine the target quantile model; then, the process parameter data of the slab to be tested is input into the target quantile model, and matched with the process parameter data stored in the output node of the target data path. After the matching is completed, the corresponding final pass rolling force prediction value is output.

[0102] It is worth noting that the data of the slabs produced subsequently can be input into the model to be trained as a new data set after comparison and correction by the operator, so that the target neural network model can be continuously improved to improve the accuracy of the prediction results and the comprehensiveness of the data.

[0103] The present application provides a method for predicting the rolling force of the last pass, which is based on a deep learning network of quantile regression and aims to solve the limitations of existing technologies in dealing with data imbalance and extreme value prediction. First, the chemical composition and rolling process data of each slab in each pass are comprehensively collected from the steel production process. These data cover all relevant physical factors, such as C, Mn, Si, P, S, temperature, width, thickness, etc., and these raw data are cleaned to ensure the quality and consistency of the data. Then, in the process of model construction, the traditional mean loss function is replaced by the quantile regression loss function to better capture the tail information of the rolling force distribution and improve the accuracy of extreme value prediction. That is, an intelligent optimization algorithm is used to intelligently select hyperparameters such as the number of hidden layers, the number of nodes and the activation function of the quantile regression deep learning to construct the optimal model structure. Under this model structure, deep learning model training is performed at different quantiles to obtain the final rolling force prediction model. Then, when there is a new rolling task, the chemical composition and rolling process parameters of the new steel grade are input in real time for cleaning and standardization. Then, the deep learning model at the appropriate quantile is selected according to the current reduction to predict the final rolling force value.

[0104] This application proposes a method for predicting the rolling force of the last pass, which is designed to specifically solve the problem of data imbalance in the prediction of the rolling force of the last pass. By introducing the concept of quantile regression, this method can deeply explore and effectively utilize the information at the tail of the rolling force data distribution, enhance the model's prediction sensitivity under extreme rolling conditions, and combine the powerful feature extraction and pattern recognition capabilities of deep learning. This application can not only predict the rolling force of the last pass more accurately, but also adapt to dynamic changes under different production conditions, providing strong technical support for the intelligent and refined control of steel production. By capturing the information at the tail of the rolling force distribution, a more accurate prediction of the rolling force of the last pass can be achieved to better meet actual production needs.

[0105] The present application provides a deep learning method for predicting the rolling force of the last pass based on intelligent optimization and quantile regression, which aims to solve the shortcomings of the existing technology in dealing with the special conditions of the last pass. Specifically, the present application effectively deals with the data imbalance caused by the specific restrictions imposed on the last pass by on-site rolling (such as the magnitude of the force and the distribution of the reduction) by capturing the information of the tail of the rolling force distribution. At the same time, it focuses on improving the accuracy of extreme value predictions to ensure that high-precision control requirements are met, because the extreme values of the rolling force of the last pass are crucial to the quality limit of the slab. In view of the limitations of traditional data balancing technologies (such as SMOTE, ROSE, etc.) in dealing with complex rolling conditions and the tail information of the rolling force distribution, the present application introduces quantile regression to overcome the above problems. To this end, the present application proposes a deep learning network structure of quantile regression, which uses an intelligent optimization algorithm to find the optimal deep learning hyperparameters and initial weights to ensure rapid convergence and strong generalization capabilities. This method can not only accurately predict the changes in the rolling force of the last pass, providing a scientific basis for the optimization of rolling process parameters, but also provide reliable data support for the rolling force of each slab in the last pass, thereby better controlling the shape and performance of the slab, reducing the risk of unqualified products, and improving product quality and production efficiency.

[0106] A second aspect of the present application provides a final-pass rolling force prediction system, comprising:

[0107] An acquisition module is configured to acquire a data set; the data set includes: chemical composition of a plurality of slabs, rolling process parameters of each pass, and rolling force;

[0108] a screening module configured to screen features associated with the final-pass rolling force in the data set using feature engineering technology to obtain a training set;

[0109] A building module, configured to build a quantile deep learning network model;

[0110] a training module configured to train the quantile deep learning network model using the training set to obtain a target quantile deep learning network model; the target quantile deep learning network model is configured with quantile models of several different quantiles; the target quantile deep learning network model is configured to determine the quantile corresponding to the final pass reduction according to the final pass reduction; and determine the quantile model corresponding to the quantile according to the quantile;

[0111] The output module is configured to input the process parameter data of the slab to be tested into the target quantile deep learning network model, and obtain the predicted value of the final rolling force of the slab to be tested based on the final rolling reduction; the process parameter data includes: the chemical composition of the slab to be tested and the final rolling process parameters.

[0112] It is worth noting that the effects of the above system embodiment during operation can be found in the effects of the above method embodiment, which will not be described in detail here.

[0113] like Figure 4 As shown in the figure, the hit rate comparison chart of the predicted deviation of the final rolling force of each slab within 1000KN in this application is shown.

[0114] For example, 192 newly produced AH36 steel slabs were selected for final rolling force prediction, and the prediction results were compared with the original rolling force model of the secondary system at the production site. According to the on-site business needs, the proportion of samples with a prediction deviation within 1000KN was counted, and the results are as follows: Figure 4 Specifically, under the requirement that the rolling force error be controlled within 1000 kN, the hit rate of the quantile deep learning rolling force prediction model reached 96.35%, while the hit rate of the production site model was only 64.58%. This means that the final-pass rolling force prediction method provided by this application has a hit rate that is approximately 32 percentage points higher than that of existing methods, significantly improving the control level of product thickness accuracy.

[0115] like Figure 5 As shown, it is a histogram comparison diagram of the predicted deviation of the final rolling force of each slab within 1000KN in this application.

[0116] For example, in order to more intuitively show the difference between the two models, the rolling force prediction deviations of the quantile deep learning model and the production site model are plotted into a histogram, as shown in Figure 2. Figure 5 As shown. Figure 5 It is clear that the production site model has large prediction deviations, some of which are as high as 3000 kN or more, which has a negative impact on product quality. In contrast, the prediction deviation of the final rolling force prediction method provided by this application is mainly concentrated within 500 kN, greatly improving plate quality and production consistency. This not only helps to reduce scrap rates, but also improves production efficiency and reduces energy consumption, bringing significant economic benefits to steel manufacturers.

[0117] This application provides a method and system for predicting the final rolling force, which has the following beneficial effects:

[0118] By adopting the last-pass rolling force prediction method provided by this application, the medium and thick plate production line of the steel enterprise has achieved remarkable results. It not only effectively solved the limitations of the existing technology in dealing with data imbalance and extreme value prediction, but also significantly improved the prediction accuracy of rolling force and product quality, bringing many economic and social benefits to the enterprise. By introducing the quantile regression loss function, the multi-layer quantile neural network model can more accurately capture the information of the tail of the rolling force distribution, especially showing higher accuracy in extreme value prediction. This enables the enterprise to more accurately control the thickness of the last pass of each slab during the production process, and ultimately improves the thickness accuracy by 2%. In summary, the prediction method provided by this application not only reduces the scrap rate, but also ensures the consistency and stability of product specifications, meeting the market's increasingly stringent requirements for high-quality steel.

[0119] Traditional rolling force control methods rely on the experience and intuition of on-site operators, which can easily lead to large deviations and affect the final quality of the product. However, the application of this application significantly reduces the need for manual operation. The intelligent prediction system can receive the chemical composition and rolling process parameters of the new slab in real time and dynamically adjust the prediction strategy based on historical data to provide precise guidance to operators. This not only improves operational efficiency but also reduces the uncertainty caused by human factors, ensuring the continuity and stability of production.

[0120] The above specific implementation methods further explain in detail the purpose, technical solutions and beneficial effects of the embodiments of the present application. It should be understood that the above are only specific implementation methods of the embodiments of the present application and are not intended to limit the scope of protection of the embodiments of the present application. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solutions of the embodiments of the present application should be included in the scope of protection of the embodiments of the present application.

Claims

1. A method for predicting the final rolling force, characterized in that: include: Get the dataset; The data set includes: chemical composition of several slabs, rolling process parameters of each pass, and rolling force; Using feature engineering technology to screen features associated with the final-pass rolling force in the data set to obtain a training set; Build a quantile deep learning network model; The quantile deep learning network model is trained using the training set to obtain a target quantile deep learning network model; the target quantile deep learning network model is configured with quantile models of several different quantiles; the target quantile deep learning network model is configured to determine the quantile corresponding to the final pass reduction according to the final pass reduction; and the quantile model corresponding to the quantile is determined according to the quantile; The process parameter data of the slab to be tested is input into the target quantile deep learning network model, and the predicted value of the final rolling force of the slab to be tested is obtained according to the final rolling reduction; the process parameter data includes: the chemical composition of the slab to be tested and the final rolling process parameters.

2. A method for predicting the final rolling force according to claim 1, characterized in that: After the step of obtaining the data set, the following steps are included: Based on the data set, the upper limit and lower limit of each indicator in the data set are obtained using statistical methods. Determine the setting range of process parameters; Based on the upper limit and lower limit of each indicator in the data set, the process parameter setting range is used to obtain abnormal values, and the abnormal values are cleared and filled.

3. The method for predicting the final rolling force according to claim 1, wherein: The step of using feature engineering technology to screen the features associated with the final-pass rolling force in the data set to obtain a training set includes: According to the data set, a first feature is obtained; the first feature is a feature in the data set associated with the final rolling force; Based on the first feature, using feature engineering technology to perform reduction and expansion reconstruction operations on the first feature to obtain a second feature; Based on the second feature, a training set is obtained using a standardization method.

4. The method for predicting the final rolling force according to claim 1, wherein: The step of using the training set to train the quantile deep learning network model to obtain a target quantile deep learning network model includes: Based on the quantile deep learning network model, using an intelligent optimization algorithm, determining the hyperparameters of the quantile deep learning network model; The quantile deep learning network model is trained using the training set to minimize the quantile regression loss function and obtain a target quantile deep learning network model.

5. The method for predicting the final rolling force according to claim 1, characterized in that: Before the step of inputting the process parameter data of the slab to be tested into the target quantile deep learning network model, the method includes: Acquire first process parameter sub-data; the first process parameter sub-data includes: the chemical composition of the slab to be tested and the final rolling process parameters; Based on the first process parameter sub-data, using a statistical method to obtain abnormal values, and performing clearing and filling processing on the abnormal values; Based on the first process parameter sub-data, using feature engineering technology to perform reduction and expansion reconstruction operations on the first process parameter sub-data to obtain second process parameter sub-data; Based on the second process parameter sub-data, process parameter data is obtained using a standardization method.

6. The method for predicting the final rolling force according to claim 1, characterized in that: The step of inputting the process parameter data of the slab to be tested into the target quantile deep learning network model and obtaining the predicted value of the final rolling force of the slab to be tested according to the final rolling reduction comprises: Determining a target quantile model according to the final pass reduction; Inputting process parameter data of the slab to be tested into the target quantile model, and determining a target data path in the target quantile model that matches the process parameter data; Based on the target data path, a predicted value of the final rolling force of the slab to be tested is obtained.

7. The method for predicting the final rolling force according to claim 1, characterized in that: The step of determining the target quantile model according to the final pass reduction comprises: Determining a quantile corresponding to the final pass reduction according to the final pass reduction; Determining a target quantile based on the quantile; According to the target quantile, a target quantile model corresponding to the target quantile is determined.

8. The method for predicting the final rolling force according to claim 1, characterized in that: The chemical components include: carbon, silicon, manganese, sulfur, and phosphorus; the rolling process parameters include: temperature, inlet and outlet widths, inlet and outlet thicknesses, and reduction.

9. A final-pass rolling force prediction system, characterized in that: include: An acquisition module is configured to acquire a data set; the data set includes: chemical composition of a plurality of slabs, rolling process parameters of each pass, and rolling force; a screening module configured to screen features associated with the final-pass rolling force in the data set using feature engineering technology to obtain a training set; A building module, configured to build a quantile deep learning network model; a training module configured to train the quantile deep learning network model using the training set to obtain a target quantile deep learning network model; the target quantile deep learning network model is configured with quantile models of several different quantiles; the target quantile deep learning network model is configured to determine the quantile corresponding to the final pass reduction according to the final pass reduction; and determine the quantile model corresponding to the quantile according to the quantile; The output module is configured to input the process parameter data of the slab to be tested into the target quantile deep learning network model, and obtain the predicted value of the final rolling force of the slab to be tested based on the final rolling reduction; the process parameter data includes: the chemical composition of the slab to be tested and the final rolling process parameters.

Images