Method and device for predicting mechanical property of hot-rolled product

Abstract

The invention provides a method and a device for predicting the mechanical property of a hot-rolled product, and aims to efficiently and accurately predict the mechanical property of the hot-rolled product, reduce the production cost of an enterprise, improve the yield and shorten the delivery cycle of the product. According to the method, historical production data, including production process parameter data and mechanical property data, of a hot-rolled product is collected from a production and marketing system, the historical production data is divided into a training set and a test set after being cleaned, and the training set and the test set are preprocessed; thirdly, taking the mechanical property index as a prediction target, and training the XGBoost model, so that the XGBoost model can predict the value of the mechanical property index according to the production process parameters; the model hyper-parameters are adjusted and optimized through a Bayesian optimization algorithm, features are sorted and screened based on feature importance, features which do not meet conditions in a training set and a test set are deleted, and the model is retrained to obtain an optimized prediction model. And finally, inputting production actual performance data of the to-be-tested hot-rolled product into the prediction model to obtain a predicted value of the mechanical property index.

Description

Technical Field

[0001] The present application relates to the interdisciplinary technical field of big data and artificial intelligence modeling in steel industry production, and in particular to a method and device for predicting the mechanical properties of hot-rolled products. Background Art

[0002] In existing steel production, the evaluation of the mechanical properties of hot-rolled products primarily relies on laboratory testing of sampled coils. This process is time-consuming and costly, making this traditional evaluation method particularly uneconomical in the modern industrial production environment, which strives for high efficiency and cost control. Laboratory testing typically requires destructive testing of sampled coils representing a batch to determine key mechanical properties such as yield strength, tensile strength, elongation after fracture, and impact energy. This process inevitably reduces yield, increases production costs, and prolongs the product cycle from production to shipment.

[0003] At present, most prediction technologies for the mechanical properties of hot-rolled products use a modeling method based on a combination of physical metallurgical principles and semi-empirical methods. Statistical analysis is used to determine the key parameters of the model, thereby constructing a mechanical property prediction model. Although the principles of such models are clear, their implementation is highly dependent on the accumulation of experimental data, resulting in high modeling costs and limited popularization and application in actual industrial production. In recent years, with the rapid development of computer technology, industrial Internet, big data, and artificial intelligence technology, multivariate regression models based on industrial production big data have begun to be applied to the prediction of the mechanical properties of hot-rolled products, attempting to reduce dependence on experimental data and improve prediction efficiency through data-driven methods. However, such models still have some limitations in practical applications, such as large prediction deviations, which have led to their lack of widespread application and recognition.

[0004] Therefore, the current hot-rolled product mechanical properties prediction technology still has obvious deficiencies in terms of efficiency, cost and prediction accuracy. There is an urgent need for a more efficient, accurate and low-cost method for predicting the mechanical properties of hot-rolled products. Summary of the Invention

[0005] The present application provides a method and device for predicting the mechanical properties of hot-rolled products, which can efficiently and accurately predict the mechanical properties of hot-rolled products, and can reduce production costs, improve yield rates and shorten product delivery cycles for enterprises.

[0006] In a first aspect, a method for predicting mechanical properties of hot-rolled products is provided, comprising:

[0007] Collecting historical production data of hot-rolled products from a production and marketing system, including production process parameter data and mechanical property data, wherein the mechanical property data includes values of mechanical property indicators, and the production process parameter data includes the value of at least one production process parameter;

[0008] Perform cleaning processing on historical production data to remove redundant and abnormal data in historical production data;

[0009] Divide the cleaned historical production data into a training set and a test set;

[0010] Preprocess the training set and test set;

[0011] Based on the preprocessed training set and test set, the mechanical performance index is used as the prediction target to train the XGBoost model, so that the XGBoost model can predict the value of the mechanical performance index based on the production process parameter data;

[0012] Tune the hyperparameters of the trained XGBoost model using the Bayesian optimization algorithm;

[0013] Based on the feature importance ranking, the features input to the tuned XGBoost model are screened, and the features that do not meet the importance conditions in the training set and test set are deleted. The tuned XGBoost model is trained again based on the latest training set and test set to obtain an optimized prediction model, whose features are production process parameters;

[0014] The actual production performance data of the hot-rolled product to be tested is input into the prediction model to obtain the predicted values of the mechanical performance indicators output by the prediction model.

[0015] In a feasible design, before cleaning the historical production data, the method includes:

[0016] Conduct correlation analysis on historical production data to generate correlation heat maps corresponding to mechanical performance indicators;

[0017] Before retraining the tuned XGBoost model, the method includes:

[0018] According to the correlation heat map corresponding to the mechanical performance indicators, determine whether the mechanical performance indicators correspond to feature groups with multicollinearity;

[0019] If the mechanical performance index corresponds to a feature group with multicollinearity, determine the redundant features in the feature group;

[0020] Remove redundant features from the training and test sets.

[0021] In a feasible design, before cleaning the historical production data, the method includes:

[0022] Conduct descriptive analysis on historical production data to generate distribution histograms corresponding to mechanical performance indicators;

[0023] If the absolute value of the skewness of the distribution histogram is greater than a first preset threshold, a logarithmic transformation or power function adjustment is performed on the value corresponding to the mechanical performance index in the mechanical performance data to obtain new mechanical performance data.

[0024] In a feasible design, the hyperparameters of the trained XGBoost model are tuned using a Bayesian optimization algorithm, including:

[0025] Define the hyperparameter search space and set the value range for each hyperparameter;

[0026] Determine the Bayesian optimization tool library;

[0027] Initialize the Bayesian optimization cross-validation algorithm by setting the number of iterations, cross-validation folds, and model performance evaluation indicators through the Bayesian optimization tool library;

[0028] Run the Bayesian optimization toolkit to iteratively select candidate hyperparameter combinations from the hyperparameter search space.

[0029] The model performance of each combination was evaluated using a cross-validation algorithm;

[0030] Extract optimized target hyperparameter combinations and corresponding performance indicators through the Bayesian optimization tool library;

[0031] Apply the target hyperparameter combination to the trained XGBoost model.

[0032] In a feasible design, the historical production data is cleansed, including:

[0033] Identify abnormal data in mechanical performance data based on the distribution histogram corresponding to the mechanical performance index;

[0034] Delete abnormal data in mechanical properties data.

[0035] In one feasible design, redundant features in a feature group are determined, including:

[0036] Obtaining a first correlation coefficient between a candidate feature in the feature group and a mechanical performance index;

[0037] The candidate feature whose first correlation coefficient in the feature group is not the highest is determined as a redundant feature.

[0038] In one feasible design, the historical production data includes multiple samples, and the method includes:

[0039] If the number of features with missing data in the sample exceeds a second preset threshold, the sample is deleted.

[0040] In one possible design, the method includes:

[0041] If the number of features with missing data in the sample does not exceed the second preset threshold, and the number of features with missing data is not 0, the average value corresponding to the features with missing data is used to fill the sample.

[0042] In one feasible design, the importance condition is that the feature is in the top 80% quantile.

[0043] In a second aspect, a device for predicting mechanical properties of hot-rolled products is provided, comprising:

[0044] A production data acquisition module is used to collect historical production data of hot-rolled products from the production and marketing system, including production process parameter data and mechanical property data, wherein the mechanical property data includes the values of mechanical property indicators, and the production process parameter data includes the value of at least one production process parameter;

[0045] The production data processing module is used to clean the historical production data to remove redundant and abnormal data.

[0046] The production data processing module is also used to divide the cleaned historical production data into a training set and a test set;

[0047] The production data processing module is also used to preprocess the training and test sets;

[0048] The model training module is used to train the XGBoost model based on the preprocessed training set and test set, taking the mechanical performance indicators as prediction targets, so that the XGBoost model can predict the values of mechanical performance indicators based on production process parameter data;

[0049] The model optimization module is used to tune the hyperparameters of the trained XGBoost model using the Bayesian optimization algorithm;

[0050] The model optimization module is also used to screen the features input to the tuned XGBoost model based on the feature importance ranking, delete the features that do not meet the importance conditions in the training set and test set, and retrain the tuned XGBoost model based on the latest training set and test set to obtain an optimized prediction model, the features of which are production process parameters;

[0051] The prediction module is used to input the production performance data of the hot-rolled product to be tested into the prediction model to obtain the predicted value of the mechanical performance index output by the prediction model.

[0052] In summary, in the embodiments of the present application, a prediction model for the mechanical properties of hot-rolled products was established by improving the XGBoost algorithm. Specifically, the data collected from the production and marketing system were cleaned and preprocessed to ensure the quality of the data set and improve the accuracy and completeness of the data. The XGBoost model parameters were adjusted by Bayesian optimization to improve the model performance and achieve improvements to the XGBoost model. Reasonable selection and elimination of unimportant features reduced the amount of calculation and improved the efficiency of model training. Ultimately, a prediction model with high prediction accuracy and good results was obtained. It can also reduce production costs, increase yield rates, and reduce product delivery cycles for enterprises, and can also provide guidance for inspection and testing of hot-rolled products. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] 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.

[0054] Figure 1 This is a schematic flow chart of a method for predicting mechanical properties of hot-rolled products provided by an exemplary embodiment of the present application;

[0055] Figure 2 This is a distribution histogram drawn by performing a descriptive analysis on yield strength data provided by an exemplary embodiment of the present application;

[0056] Figure 3 This is a correlation heat map drawn by performing correlation analysis on yield strength data provided by an exemplary embodiment of the present application;

[0057] Figure 4 is a diagram of an ongoing Bayesian optimization process provided by an exemplary embodiment of the present application;

[0058] Figure 5 This is a feature importance ranking diagram drawn after model training provided by an exemplary embodiment of the present application;

[0059] Figure 6 This is a scatter plot comparing predicted values and actual values of a yield strength prediction model based on Q355B steel test data provided by an exemplary embodiment of the present application;

[0060] Figure 7 This is a scatter plot comparing predicted values and actual values of a tensile strength prediction model based on Q355B steel test data provided by an exemplary embodiment of the present application;

[0061] Figure 8This is a scatter plot comparing predicted values and actual values of an elongation prediction model based on Q355B steel test data provided by an exemplary embodiment of the present application;

[0062] Figure 9 This is a scatter plot comparing predicted values and actual values of an impact energy prediction model based on Q355B steel test data provided by an exemplary embodiment of the present application;

[0063] Figure 10 This is a line chart comparing the predicted values and actual values of a yield strength prediction model based on two groups of sampling test data of Q355B steel provided by an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described 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 are within the scope of protection of this application.

[0065] In order to solve the problem that the current hot-rolled product mechanical property prediction technology still has obvious deficiencies in efficiency, cost and prediction accuracy, this application processes the hot-rolled product production data collected in the production and marketing system, introduces Bayesian optimization to improve the Extreme Gradient Boosting (XGBoost) algorithm, and then ranks the importance of each feature after model training and reasonably selects features. After retraining, the optimal hot-rolled product mechanical property prediction model is obtained. This model has high prediction accuracy, which is beneficial for enterprises to reduce production costs, improve yield rate and shorten product delivery cycle, and can also provide guidance for hot-rolled product inspection and testing. Among them, XGBoost is a machine learning algorithm based on Gradient Boosting Decision Tree (GBDT).

[0066] Figure 1 This is a schematic flow chart of a method for predicting the mechanical properties of hot-rolled products provided by an exemplary embodiment of the present application. Figure 1 As shown, the method includes:

[0067] S110, collecting historical production data of hot-rolled products from the production and marketing system.

[0068] The historical production data includes multiple samples, each of which includes production process parameter data and mechanical property data. The mechanical property data includes the value of a mechanical property index, and the production process parameter data includes the value of at least one production process parameter. In other words, each piece of data in the production process parameter data of this application includes the name and value of a production process parameter, and each piece of data in the mechanical property data includes the name and value of a mechanical property index.

[0069] Exemplarily, the production process parameter data include one or more of the following parameters: steel coil number, heating time in the soaking section of the heating furnace, temperature of the soaking section out of the heating furnace, total heating time of the heating furnace, inlet temperature of the initial rolling mill, final rolling temperature of the initial rolling mill, inlet temperature of the finishing mill, final rolling temperature of the finishing mill, water cooling temperature, water cooling speed, coiling temperature, coiling speed, coil thickness, coil width, coil length, actual thickness average value, intermediate billet thickness after rough rolling, slab thickness, carbon mass fraction, silicon mass fraction, manganese mass fraction, phosphorus mass fraction, sulfur mass fraction, niobium mass fraction, vanadium mass fraction, and titanium mass fraction.

[0070] Exemplarily, the mechanical performance index is yield strength, tensile strength, elongation or impact energy.

[0071] For example, the hot-rolled product mechanical property prediction method of the present embodiment was tested on Q355B hot-rolled coils. A total of 11,433 production samples were obtained from the production and marketing system over a six-month period. Each sample contained production process parameter data including: coil number data, heating time in the soaking section of the heating furnace, temperature data at the exit of the soaking section of the heating furnace, total heating time in the heating furnace, blooming mill inlet temperature data, blooming mill final rolling temperature data, finishing mill inlet temperature data, finishing mill final rolling temperature data, water cooling temperature data, water cooling speed data, coiling temperature data, coiling speed data, coil thickness data, coil width data, coil length data, actual thickness average data, intermediate billet thickness data after roughing, slab thickness data, carbon mass fraction data, silicon mass fraction data, manganese mass fraction data, phosphorus mass fraction data, sulfur mass fraction data, niobium mass fraction data, vanadium mass fraction data, and titanium mass fraction data. Each sample also contained yield strength data, tensile strength data, elongation data, and impact energy data. In the process of training the model, the mechanical property data used in this application is the data corresponding to the mechanical property index used as the prediction target. For example, the yield strength prediction model is trained using yield strength data as the mechanical property data, and so on.

[0072] In a feasible design, before cleaning the historical production data, the method includes:

[0073] Perform correlation analysis on historical production data to generate correlation heat maps corresponding to mechanical performance indicators.

[0074] This application takes the correlation analysis between the yield strength of Q355B samples and various production process parameters as an example, and generates the following Figure 2 From the correlation heat map shown, it can be found that the yield strength is positively correlated with the coil thickness, the average actual thickness, the thickness of the intermediate billet after rough rolling, and the mass fraction of manganese and niobium elements, and negatively correlated with the coil length and the final rolling temperature of the finishing mill. Figure 2 The correlation strength between each production process parameter and yield strength is clearly demonstrated, making it easier to screen out features that have a significant impact on the prediction, thereby optimizing the model input and improving the prediction accuracy.

[0075] In the above example, correlation analysis of historical production data helps understand data distribution, providing important guidance for subsequent feature selection. Correlation analysis results can also reveal potential production process optimization opportunities. For example, adjusting coil thickness and manganese and niobium content could further increase yield strength, thereby improving product quality and stability.

[0076] It should be noted that the features described in this application refer to the various production process parameters input into the model. In other words, the features are the names of the various production process parameters.

[0077] In a feasible design, before cleaning the historical production data, the method includes:

[0078] Perform descriptive analysis on historical production data and generate distribution histograms corresponding to mechanical performance indicators.

[0079] This application takes the descriptive analysis of the yield strength data of each sample in Q355B as an example and generates the following Figure 3 The distribution histogram shown in Figure 2 shows that the yield strength data is normally distributed, the data distribution is relatively reasonable, and the data stability is relatively strong.

[0080] Exemplarily, the method further includes: if the absolute value of the skewness of the distribution histogram is greater than a first preset threshold, performing a logarithmic transformation or a power function adjustment on the value corresponding to the mechanical performance index in the mechanical performance data to obtain new mechanical performance data.

[0081] Accordingly, if the absolute value of the skewness of the distribution histogram is less than or equal to the first preset threshold, no data transformation is required and the original data can be used directly for subsequent analysis. This step helps ensure the normality and stability of the data.

[0082] The first preset threshold is set according to actual needs. For example, the first preset threshold is set to 0.5 to ensure that the skewness of the data distribution is within an acceptable range.

[0083] The above example effectively identifies the mechanical performance data that needs to be adjusted by comparing the absolute value of the skewness of the distribution histogram with the first preset threshold value, thereby ensuring the reliability of the analysis results. If the distribution of the values corresponding to the mechanical performance indicators is slightly skewed, the new mechanical performance data distribution obtained by performing a logarithmic transformation or a power function adjustment can be closer to a normal distribution, which can improve data stability and facilitate subsequent statistical analysis and model training. This data processing method not only improves the accuracy of data analysis, but also lays a solid foundation for the subsequent construction of mechanical performance prediction models. By optimizing the distribution form of the data, the potential laws in the data can be more effectively mined, thereby improving the performance and reliability of the prediction model. Furthermore, the optimized data distribution helps to reduce the deviation in model training, enhance the generalization ability of the model, and make the prediction results closer to actual production conditions.

[0084] S120, cleaning the historical production data.

[0085] Among them, cleaning processing is used to remove redundant data and abnormal data in historical production data.

[0086] Redundant data refers to repeated records or values of characteristics or mechanical performance indicators that have no practical significance.

[0087] For example, when processing data, there may be a situation where the value corresponding to a certain feature or a certain mechanical performance indicator is empty, that is, there is a situation where data of a certain feature or a certain mechanical performance indicator is missing.

[0088] Therefore, for each feature or each mechanical performance index, if there is missing data, the corresponding average value of the corresponding feature or mechanical performance index in the historical production data is used to fill in the missing data to ensure the integrity of the data.

[0089] Furthermore, if the number of features with missing data in the sample exceeds a second preset threshold, the sample is deleted.

[0090] In this example, if the number of features with missing data in a sample exceeds a second preset threshold, the sample is considered an invalid sample and is removed to ensure the overall quality of the data set.

[0091] Furthermore, if the number of features with missing data in the sample does not exceed a second preset threshold, and the number of features with missing data is not 0, the sample is filled with the average value corresponding to the features with missing data.

[0092] In this example, if the number of features with missing data in the sample does not exceed the second preset threshold, the feature's average value in the historical data is used to fill in the gaps, ensuring data integrity while avoiding analytical bias caused by missing data. This approach effectively improves the usability of the dataset and the accuracy of the analysis results.

[0093] The third threshold can be set according to actual needs.

[0094] For example, for each feature or each mechanical performance index, if there are abnormal values that deviate significantly from the normal range, the abnormal values are modified to the average value of the corresponding feature or mechanical performance index to avoid these abnormal values from having an adverse effect on data analysis and results.

[0095] The above examples use rigorous data cleaning and missing value processing to ensure a solid and reliable data foundation for subsequent analysis, further improving model training results and prediction accuracy.

[0096] In a feasible design, the historical production data is cleansed, including:

[0097] Identify abnormal data in mechanical performance data based on the distribution histogram corresponding to the mechanical performance index;

[0098] Delete abnormal data in mechanical properties data.

[0099] For example, when the value of the mechanical performance index falls within the abnormal interval of the distribution histogram, the data is considered to be abnormal data.

[0100] By removing these outliers, the above example further cleans the dataset and reduces the impact of noise on subsequent analysis. This process ensures that the data used for model training is more accurate and reliable, helping to improve the model's predictive accuracy and generalization capabilities.

[0101] S130, dividing the cleaned historical production data into a training set and a test set.

[0102] For example, the historical production data after cleaning is divided into a training set and a test set in a ratio of 8:2. The divided training set is used for model training, and the test set is used for evaluating model performance.

[0103] It should be understood that for each mechanical property indicator, the production process parameter data in the training and test sets are identical, and the mechanical property data only includes data corresponding to the mechanical property indicator being predicted. For example, for four mechanical property indicators, four sets of training and test sets are required, with each set of mechanical property data including data corresponding to the mechanical property indicator being predicted.

[0104] S140, preprocessing the training set and the test set.

[0105] Exemplarily, the preprocessing includes normalization and / or standardization.

[0106] Specifically, the normalization process is shown in the following formula (1):

[0107]

[0108] Among them, X is the original data of the type to be normalized, X min and X max are the minimum and maximum values of this type of data in the data set, respectively. norm It is the normalized data of this type. The type is divided according to different production process parameters and different mechanical performance indicators.

[0109] The normalization process is shown in the following formula (2):

[0110]

[0111] Among them, X' is the original data of the type to be standardized, μ is the mean of the data of this type in the data set, σ is the standard deviation of the original data of this type in the data set, X std This is the data after standardization of this type.

[0112] The above example improves the accuracy and stability of the model's prediction of the mechanical properties of hot-rolled products by normalizing and standardizing the training and test sets. Normalization scales the data to a specific range, making different features numerically comparable and thus enhancing the model's generalization capabilities. Furthermore, by applying the same normalization rules to the training and test sets, the model ensures consistent data distribution during training and testing, further improving the reliability of the prediction results. Standardization converts the data to a distribution with a mean of 0 and a standard deviation of 1, eliminating the impact of abnormal data on model training. In the converted data, abnormal data is mapped to a more distant location, reducing its impact on the model. At the same time, standardization makes the data distribution more concentrated, which helps accelerate model convergence and improves training efficiency. In the prediction of the mechanical properties of hot-rolled products, combining normalization and standardization can fully leverage their respective strengths and further enhance the performance of the prediction model.

[0113] For example, according to the descriptive analysis results and / or correlation analysis results of different prediction targets (such as yield strength, tensile strength, elongation after fracture, impact energy), the corresponding data preprocessing method is determined to improve the model prediction accuracy.

[0114] For example, a histogram analysis of the tensile strength distribution revealed significant differences in the range of values between coiling temperature and carbon mass fraction (e.g., a temperature span of 250°C and a carbon content span of 0.08%). A heat map analysis of the tensile strength correlation revealed a strong positive correlation between tensile strength, coiling temperature, and finishing temperature.

[0115] Because the input features have large dimensional differences (temperature parameters have a wide span, while chemical composition parameters have a narrow span), direct modeling can easily lead to a bias in favoring features with larger values. Normalization can be used to scale all features to a uniform range ([0,1]).

[0116] For example, through the correlation heat map analysis of impact energy, it was found that it was weakly correlated with various features. By generating interaction terms (such as "sulfur mass fraction × cooling rate") through feature engineering and combining the various features of the input with standardized processing, the model's ability to capture nonlinear relationships was enhanced.

[0117] S150, based on the preprocessed training set and test set, the mechanical performance index is used as the prediction target to train the XGBoost model, so that the XGBoost model can predict the value of the mechanical performance index according to the production process parameter data.

[0118] For example, different preprocessing combinations are used to process the training and test sets, and the model performance (e.g., mean squared error, accuracy) on the validation set is evaluated to determine the appropriate preprocessing method. Preprocessing combinations include normalization only, standardization only, and both normalization and standardization. This allows for dynamic optimization of data processing strategies based on model training results, further improving model performance and generalization capabilities.

[0119] S160, tune the hyperparameters of the trained XGBoost model using the Bayesian optimization algorithm.

[0120] In one feasible design, the hyperparameters of the trained XGBoost model are tuned using the Bayesian optimization algorithm in the following way:

[0121] Define the hyperparameter search space and set the value range for each hyperparameter;

[0122] Determine the Bayesian optimization tool library;

[0123] Initialize the Bayesian optimization cross-validation algorithm by setting the number of iterations, cross-validation folds, and model performance evaluation indicators through the Bayesian optimization tool library;

[0124] Run the Bayesian optimization toolkit to iteratively select candidate hyperparameter combinations from the hyperparameter search space.

[0125] The model performance of each combination was evaluated using a cross-validation algorithm;

[0126] Extract optimized target hyperparameter combinations and corresponding performance indicators through the Bayesian optimization tool library;

[0127] Apply the target hyperparameter combination to the trained XGBoost model.

[0128] For example, the hyperparameters include one or more of the following: sampling feature ratio, learning rate, maximum tree depth, and number of trees. A value range can be set for each hyperparameter according to actual needs.

[0129] For example, the Bayesian optimization tool library uses scikit-optimize (skopt).

[0130] For example, the number of iterations and cross-validation folds are determined based on the complexity of the problem (such as the dimension of the features, that is, whether the relationship between the dependent variable and the independent variable is complex), the amount of data, and actual computing resources.

[0131] For example, the mean squared error (MSE) is used as an evaluation indicator of model performance. Given that MSE is sensitive to outliers and is particularly suitable for processing continuous variables, it is very suitable for evaluating the impact of each input feature on the model in this application.

[0132] For example, different initial hyperparameter values are set for different prediction objectives, and corresponding parameter ranges are defined. This can reduce the time required for Bayesian optimization (e.g., by narrowing the search range to speed up model convergence). Different initial hyperparameter values can be set for different prediction objectives based on historical experimental data.

[0133] In the above example, this application used the Bayesian Optimization toolkit to fine-tune and optimize the hyperparameters of the XGBoost model. This approach aims to maximize the model's accuracy in predicting the mechanical properties of hot-rolled products. Specifically, the Bayesian Optimization toolkit constructed a probabilistic model that iteratively selected the most promising hyperparameter combinations, effectively narrowing the search range and significantly improving the efficiency of the optimization process.

[0134] The implementation of this step ensures that the XGBoost model can demonstrate efficient and stable performance when processing complex data sets. By adjusting hyperparameters, this application can not only improve the computational efficiency of the model when processing large-scale data sets, but also significantly enhance the accuracy and reliability of the prediction results. Therefore, in practical applications, the model can more accurately capture the key characteristics of the mechanical properties of hot-rolled products, providing strong technical support for related production and quality control.

[0135] Exemplarily, the method further includes:

[0136] Evaluate the performance of the tuned XGBoost model on the test set.

[0137] By evaluating the performance of the tuned model on the test set, the generalization ability of the model can be further verified, thereby ensuring the reliability of the prediction results in practical applications.

[0138] S170, based on the feature importance ranking, the various features input into the tuned XGBoost model are screened, the features that do not meet the importance conditions in the training set and the test set are deleted, and the tuned XGBoost model is trained again based on the latest training set and the test set to obtain the optimized prediction model.

[0139] Among them, the characteristics are production process parameters.

[0140] Specifically, XGBoost can rank the importance of features by calculating the average information gain brought by the features in all trees.

[0141] Before retraining the tuned XGBoost model, you can also optimize the number of features based on the correlation analysis results to speed up model training. The methods include:

[0142] According to the correlation heat map corresponding to the mechanical performance indicators, determine whether the mechanical performance indicators correspond to feature groups with multicollinearity;

[0143] If the mechanical performance index corresponds to a feature group with multicollinearity, determine the redundant features in the feature group;

[0144] Remove redundant features from the training and test sets.

[0145] A feature group with multicollinearity refers to a combination of certain production process parameters (i.e., features) in a predictive model where there is a high linear correlation between them (e.g., the correlation coefficient between two production process parameters is greater than 0.8). This correlation can lead to decreased model stability or biased feature importance assessments. It should be understood that the number of features in a feature group is greater than or equal to 2, and the number of redundant features in a feature group is greater than or equal to 1.

[0146] The above example can reduce the dimensionality of the input data, reduce the complexity of the model calculation, and significantly shorten the training time by identifying and removing redundant features with multicollinearity (such as features in the feature group that have a weak correlation with mechanical performance indicators).

[0147] In addition, eliminating redundant features can effectively suppress the risk of overfitting caused by the high correlation between features, making the model's prediction results on the test set more stable and reliable.

[0148] In a feasible design, redundant features in a feature group are determined by:

[0149] Obtaining a first correlation coefficient between a candidate feature in the feature group and a mechanical performance index;

[0150] The candidate feature whose first correlation coefficient in the feature group is not the highest is determined as a redundant feature.

[0151] The following is the definition of redundant features:

[0152] In the above example, the candidate feature with the highest first correlation coefficient with the mechanical performance index is assumed to be the most critical feature. Other features that have a high linear correlation with this most critical feature are considered redundant. This is because the information provided by these redundant features can largely be replaced by the most critical feature. Their presence in the model not only increases computational complexity but may also interfere with the model's capture of key information, thereby affecting prediction accuracy.

[0153] The above example calculates the correlation between each candidate feature and the mechanical property index to measure their contribution to the prediction results. By identifying and eliminating these redundant features, the prediction model can be further optimized, improving its efficiency and accuracy.

[0154] In one feasible design, the importance condition is that the feature is in the top 80% quantile.

[0155] The above example sets the importance condition to only feature values within the top 80% quantile, effectively filtering out features with significant impact on mechanical property prediction. Further reducing the number of features and model complexity can also preserve key features that significantly influence prediction results. In practice, this importance condition can be flexibly adjusted to achieve optimal prediction results based on specific prediction requirements and dataset characteristics.

[0156] S180, inputting actual production performance data of the hot-rolled product to be tested into the prediction model to obtain predicted values of the mechanical property indicators output by the prediction model.

[0157] Specifically, after the production performance data is cleaned and preprocessed, it is input into the prediction model to obtain the predicted values of the mechanical performance indicators output by the prediction model.

[0158] For example, when constructing training samples, the production data of multiple brands of hot-rolled products are selected, and independent model training and analysis are performed on different brands of hot-rolled products. This can effectively improve the pertinence of the constructed model and improve the prediction accuracy.

[0159] The following is an example of the hot-rolled coil with grade Q355B and the prediction target being yield strength to illustrate the embodiment of the hot-rolled product mechanical property prediction method of the present application.

[0160] Step 1: Obtain a production sample of Q355B hot-rolled coil from the production and marketing system. The sample information is as described in the above example and will not be repeated here.

[0161] Step 2: Perform data cleaning, missing value processing, and outlier modification on each sample.

[0162] Step 3: Draw a correlation heat map and a distribution histogram for each sample processed in step 2, and determine the correlation analysis results and descriptive analysis results based on the correlation heat map and the distribution histogram.

[0163] by Figure 2 The correlation heatmap shown and Figure 3 Taking the distribution histogram shown as an example, correlation analysis results show that yield strength is positively correlated with coil thickness, average actual thickness, intermediate bar thickness after roughing, and the mass fractions of manganese and niobium, and negatively correlated with coil length and finishing mill temperature. Descriptive analysis results indicate that yield strength exhibits a generally normal distribution, indicating a reasonable data distribution.

[0164] Step 4: Based on the descriptive analysis results, determine whether the distribution of the values corresponding to the mechanical property indicators is slightly skewed. If so, perform a log transformation or power function adjustment on the data corresponding to the mechanical property indicators to ensure a normal distribution of the new mechanical property data. Furthermore, identify abnormal data within the mechanical property data based on the descriptive analysis results.

[0165] Step 5: Divide the cleaned historical production data into a training set and a test set in a ratio of 8:2.

[0166] Step 6: Determine the corresponding data preprocessing method based on the descriptive analysis results and / or correlation analysis results of the yield strength, and preprocess the training set and the test set respectively.

[0167] Step 7: Train the XGBoost model based on the preprocessed training set and test set.

[0168] Step 8: Tune the hyperparameters of the trained XGBoost model using the Bayesian optimization algorithm.

[0169] Specifically, the optimization process is as follows Figure 4After setting the hyperparameter search space, we fine-tune key parameters such as colsample_bytree (feature sampling ratio), learning_rate (learning rate), max_depth (maximum tree depth), and n_estimators (number of trees) through a continuous iterative process to determine the optimal combination. Based on this, we construct an XgBoost regression model and evaluate its performance on the test set.

[0170] Step 9: Sort the features fed into the tuned XGBoost model based on their importance. Remove features that do not meet the importance criteria in both the training and test sets, retaining only those in the top 80% quantile. This reduces model computational overhead and speeds up training.

[0171] For the prediction target of yield strength, the XGBoost model ranks the importance of each feature as follows: Figure 5 As shown in the figure, considering that temperature control during hot rolling production is relatively stable and contributes little to model accuracy, temperature-related parameters can be deleted. After repeated testing and training, for yield strength prediction, removing the production process parameter data corresponding to six features from the training and test sets: heating time in the heating furnace soaking section, finishing mill temperature, coiling temperature, finishing mill temperature, finishing mill inlet temperature, and finishing mill inlet temperature significantly improved model accuracy, thus determining the optimal prediction model.

[0172] Step 10: Use tensile strength, elongation, and impact energy as prediction targets to train the optimal XGBoost prediction model. The method is similar to steps 1 to 9. Figures 6 to 9 The scatter plots comparing the predicted values and actual values based on the test set data for the yield strength prediction model, tensile strength prediction model, elongation prediction model, and impact energy prediction model respectively show that the differences between the predicted values and the actual values are almost all within the ±10% control line, indicating that the model prediction effect is good. In addition, taking the yield strength as an example, Figure 10 A line chart comparing the predicted and actual yield strength values of two sets of Q355B steel sample test data is shown. The horizontal axis represents the sample number, and the vertical axis represents the yield strength value. It can be seen that the predicted and actual values are highly consistent, with the errors remaining within a controllable range.

[0173] Step 11: Obtain actual production performance data for the hot-rolled product to be tested. This data includes data corresponding to 29 production process parameters. This data is replicated to generate four sets of production performance data to be tested. For each set of production performance data, redundant features corresponding to the corresponding mechanical performance indicators are rationally removed. Substituting the four sets of de-redundant production performance data into the optimal prediction models for yield strength, tensile strength, elongation, and impact energy, the predicted values for the yield strength, tensile strength, elongation, and impact energy of the current steel coil are obtained.

[0174] In summary, in the embodiments of the present application, a prediction model for the mechanical properties of hot-rolled products was established by improving the XGBoost algorithm. Specifically, the data collected from the production and marketing system was cleaned and preprocessed to ensure the quality of the data set and improve the accuracy and integrity of the data. The XGBoost model parameters were adjusted by Bayesian optimization to improve model performance and achieve improvements to the XGBoost model. Rational selection and elimination of unimportant features reduced the amount of computation and improved model training efficiency. Ultimately, a prediction model with high prediction accuracy and good results was obtained.

[0175] Furthermore, this application solution also has the following effects:

[0176] Reduce production costs for enterprises:

[0177] Traditional methods require destructive laboratory testing of each batch of hot-rolled products to obtain mechanical property data, resulting in waste of raw materials. This solution replaces some laboratory testing with a high-precision prediction model, reducing the number of sampling times and directly saving material and labor costs. Furthermore, the model's prediction results can guide companies to focus on adjusting key process parameters, avoiding blind trial and error and reducing ineffective energy and resource consumption.

[0178] Improve the yield rate:

[0179] By analyzing the relationship between historical production data (such as temperature and chemical composition) and mechanical properties, the model predicts product performance in real time, helping companies optimize rolling parameters (such as final rolling temperature and cooling rate), reduce defective products or scrap caused by process deviations, and achieve precise process control. Furthermore, the predictive model provides real-time quality feedback during the production process, enabling operators to adjust processes in a timely manner and avoid overall downgrade of an entire batch of products due to local defects.

[0180] Reduce product delivery cycle:

[0181] Traditional laboratory testing takes hours to days, while model predictions can be completed in seconds, significantly shortening the waiting time from production to quality inspection.

[0182] Provide guidance for hot-rolled product inspection and testing:

[0183] The model prediction results can identify high-risk batches (such as steel coils with predicted values close to the lower quality limit), guide laboratories to prioritize the testing of these batches, and improve testing efficiency.

[0184] The present application also provides a hot-rolled product mechanical property prediction device, comprising:

[0185] A production data acquisition module is used to collect historical production data of hot-rolled products from the production and marketing system, including production process parameter data and mechanical property data, wherein the mechanical property data includes the values of mechanical property indicators, and the production process parameter data includes the value of at least one production process parameter;

[0186] The production data processing module is used to clean the historical production data to remove redundant and abnormal data.

[0187] The production data processing module is also used to divide the cleaned historical production data into a training set and a test set;

[0188] The production data processing module is also used to preprocess the training and test sets;

[0189] The model training module is used to train the XGBoost model based on the preprocessed training set and test set, taking the mechanical performance indicators as prediction targets, so that the XGBoost model can predict the values of mechanical performance indicators based on production process parameter data;

[0190] The model optimization module is used to tune the hyperparameters of the trained XGBoost model using the Bayesian optimization algorithm;

[0191] The model optimization module is also used to screen the features input to the tuned XGBoost model based on the feature importance ranking, delete the features that do not meet the importance conditions in the training set and test set, and retrain the tuned XGBoost model based on the latest training set and test set to obtain an optimized prediction model, the features of which are production process parameters;

[0192] The prediction module is used to input the production performance data of the hot-rolled product to be tested into the prediction model to obtain the predicted value of the mechanical performance index output by the prediction model.

[0193] In a feasible design, the device further includes a correlation analysis module for performing a correlation analysis on the historical production data before cleaning the historical production data to generate a correlation heat map corresponding to the mechanical performance index;

[0194] Among them, the model optimization module is used to determine whether the mechanical performance indicators correspond to a feature group with multicollinearity based on the correlation heat map corresponding to the mechanical performance indicators before re-training the tuned XGBoost model; if the mechanical performance indicators correspond to a feature group with multicollinearity, determine the redundant features in the feature group; and delete the redundant features in the training set and test set.

[0195] In one feasible design, the device further includes a descriptive analysis module for performing descriptive analysis on the historical production data before cleaning the data to generate a distribution histogram corresponding to the mechanical performance index. If the absolute value of the skewness of the distribution histogram is greater than a first preset threshold, a logarithmic transformation or power function adjustment is performed on the value corresponding to the mechanical performance index in the mechanical performance data to obtain new mechanical performance data. The first preset threshold can be set according to actual needs.

[0196] In one feasible design, the model optimization module is implemented in the following way, using the Bayesian optimization algorithm to tune the hyperparameters of the trained XGBoost model:

[0197] Define the hyperparameter search space and set the value range for each hyperparameter;

[0198] Determine the Bayesian optimization tool library;

[0199] Initialize the Bayesian optimization cross-validation algorithm by setting the number of iterations, cross-validation folds, and model performance evaluation indicators through the Bayesian optimization tool library;

[0200] Run the Bayesian optimization toolkit to iteratively select candidate hyperparameter combinations from the hyperparameter search space.

[0201] The model performance of each combination was evaluated using a cross-validation algorithm;

[0202] Extract optimized target hyperparameter combinations and corresponding performance indicators through the Bayesian optimization tool library;

[0203] Apply the target hyperparameter combination to the trained XGBoost model.

[0204] In a feasible design, the production data processing module is implemented in the following way to clean the historical production data:

[0205] Identify abnormal data in mechanical performance data based on the distribution histogram corresponding to the mechanical performance index;

[0206] Delete abnormal data in mechanical properties data.

[0207] In one feasible design, the model optimization module is implemented to determine redundant features in a feature group by:

[0208] Obtaining a first correlation coefficient between a candidate feature in the feature group and a mechanical performance index;

[0209] The candidate feature whose first correlation coefficient in the feature group is not the highest is determined as a redundant feature.

[0210] In a feasible design, the historical production data includes multiple samples, and the production data processing module is further used to:

[0211] If the number of features with missing data in the sample exceeds a second preset threshold, delete the sample;

[0212] In a feasible design, the production data processing module is further used to:

[0213] If the number of features with missing data in the sample does not exceed the second preset threshold, and the number of features with missing data is not 0, the average value corresponding to the features with missing data is used to fill the sample.

[0214] In one feasible design, the importance condition is that the feature is in the top 80% quantile.

[0215] For other implementations and effects of the above-mentioned device, please refer to the description in the embodiment of the method for predicting mechanical properties of hot-rolled products, which will not be repeated here.

[0216] The basic principles of the present application have been described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, and effects mentioned in this application are merely illustrative and not restrictive, and it should not be assumed that these advantages, strengths, and effects are required of each embodiment of this application. In addition, the specific details disclosed above are merely illustrative and facilitating understanding, and are not restrictive. The above details do not limit this application to necessarily being implemented using the above specific details.

[0217] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0218] The block diagrams of the devices, devices, equipment, and systems involved in this application are intended only as illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As will be appreciated by those skilled in the art, these devices, devices, equipment, and systems can be connected, arranged, or configured in any manner. Words such as "include," "comprise," "have," and the like are open-ended words, meaning "including but not limited to," and can be used interchangeably therewith. The words "or" and "and" used herein refer to the words "and / or" and can be used interchangeably therewith, unless the context clearly indicates otherwise. The word "such as" used herein refers to the phrase "such as but not limited to," and can be used interchangeably therewith.

[0219] It should also be noted that in the apparatus, device, and method of the present application, each component or each step can be decomposed and / or recombined, and such decomposition and / or recombination should be regarded as equivalent solutions of the present application.

[0220] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present application. Therefore, the present application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0221] The above description has been provided for the purpose of illustration and description. Furthermore, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although a number of example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for predicting the mechanical properties of hot-rolled products, characterized in that: include: Collecting historical production data of hot-rolled products from a production and marketing system, including production process parameter data and mechanical property data, wherein the mechanical property data includes values of mechanical property indicators, and the production process parameter data includes the value of at least one production process parameter; Performing a cleaning process on the historical production data, wherein the cleaning process is used to remove redundant data and abnormal data in the historical production data; Divide the cleaned historical production data into a training set and a test set; Preprocessing the training set and the test set; Based on the preprocessed training set and test set, the mechanical performance index is used as a prediction target to train the XGBoost model, so that the XGBoost model can predict the value of the mechanical performance index according to the production process parameter data; Tune the hyperparameters of the trained XGBoost model using the Bayesian optimization algorithm; Screening each feature input into the tuned XGBoost model based on feature importance ranking, deleting features that do not meet importance conditions in the training set and the test set, and retraining the tuned XGBoost model based on the latest training set and test set to obtain an optimized prediction model, wherein the features are the production process parameters; The actual production performance data of the hot-rolled product to be tested is input into the prediction model to obtain the predicted value of the mechanical property index output by the prediction model.

2. The method according to claim 1, characterized in that Before cleaning the historical production data, the method includes: Performing correlation analysis on historical production data to generate a correlation heat map corresponding to the mechanical performance indicators; Before retraining the tuned XGBoost model, the method includes: Determining, based on a correlation heat map corresponding to the mechanical performance indicators, whether the mechanical performance indicators correspond to a feature group with multicollinearity; If the mechanical performance index corresponds to a feature group with multicollinearity, determining redundant features in the feature group; Redundant features in the training set and the test set are deleted.

3. The method according to claim 1 or 2, characterized in that Before cleaning the historical production data, the method includes: Performing descriptive analysis on historical production data to generate a distribution histogram corresponding to the mechanical performance index; If the absolute value of the skewness of the distribution histogram is greater than a first preset threshold, a logarithmic transformation or a power function adjustment is performed on the value corresponding to the mechanical performance index in the mechanical performance data to obtain new mechanical performance data.

4. The method according to claim 1 or 2, characterized in that The hyperparameters of the trained XGBoost model are tuned using the Bayesian optimization algorithm, including: Define the hyperparameter search space and set the value range for each hyperparameter; Determine the Bayesian optimization tool library; Initialize the Bayesian optimization cross-validation algorithm by setting the number of iterations, cross-validation folds, and model performance evaluation indicators through the Bayesian optimization tool library; Running the Bayesian optimization tool library to iteratively select candidate hyperparameter combinations from a hyperparameter search space; Evaluate the model performance of each combination using the cross-validation algorithm; Extracting the optimized target hyperparameter combination and the corresponding performance indicators through the Bayesian optimization tool library; Apply the target hyperparameter combination to the trained XGBoost model.

5. The method according to claim 3, characterized in that The cleaning process of the historical production data includes: Identifying abnormal data in the mechanical performance data according to a distribution histogram corresponding to the mechanical performance index; Delete abnormal data in the mechanical properties data.

6. The method according to claim 2, characterized in that The determining of redundant features in the feature group includes: Obtaining a first correlation coefficient between a candidate feature in the feature group and the mechanical performance index; The candidate feature whose first correlation coefficient in the feature group is not the highest is determined as a redundant feature.

7. The method according to claim 1 or 2, characterized in that The historical production data includes a plurality of samples, and the method includes: If the number of features with missing data in the sample exceeds a second preset threshold, the sample is deleted.

8. The method according to claim 7, characterized in that The method comprises: If the number of features with missing data in the sample does not exceed the second preset threshold, and the number of features with missing data is not 0, the sample is filled with the average value corresponding to the features with missing data.

9. The method according to claim 1 or 2, characterized in that The importance condition is that the feature is in the top 80% quantile.

10. A device for predicting mechanical properties of hot-rolled products, characterized in that: include: A production data acquisition module is used to collect historical production data of hot-rolled products from the production and marketing system, including production process parameter data and mechanical property data, wherein the mechanical property data includes values of mechanical property indicators, and the production process parameter data includes the value of at least one production process parameter; A production data processing module, configured to perform cleaning processing on the historical production data, wherein the cleaning processing is configured to remove redundant data and abnormal data in the historical production data; The production data processing module is further used to divide the cleaned historical production data into a training set and a test set; The production data processing module is further used to preprocess the training set and the test set; A model training module is used to train the XGBoost model based on the preprocessed training set and test set, using the mechanical performance index as a prediction target, so that the XGBoost model can predict the value of the mechanical performance index based on the production process parameter data; The model optimization module is used to tune the hyperparameters of the trained XGBoost model using the Bayesian optimization algorithm; The model optimization module is further used to screen the various features input into the tuned XGBoost model based on feature importance ranking, delete the features that do not meet the importance conditions in the training set and the test set, and retrain the tuned XGBoost model based on the latest training set and test set to obtain an optimized prediction model, wherein the features are the production process parameters; The prediction module is used to input the production performance data of the hot-rolled product to be tested into the prediction model to obtain the predicted value of the mechanical performance index output by the prediction model.

Images