A method and system for optimizing continuous casting process parameters based on shap values
By using a continuous casting process parameter optimization method based on SHAP values, a billet defect prediction model was constructed and SHAP values were calculated. This solved the problems of experience dependence and poor interpretability in continuous casting process parameter optimization, achieving efficient and interpretable parameter optimization and improving production adaptability and product quality.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- AUTOMATION RES & DESIGN INST OF METALLURGICAL IND
- Filing Date
- 2026-02-11
- Publication Date
- 2026-06-05
AI Technical Summary
Existing methods for optimizing continuous casting process parameters suffer from problems such as strong reliance on experience, poor model interpretability, low computational efficiency, and insufficient adaptability, making it difficult to adapt to dynamic production environments and steel grade switching.
A continuous casting process parameter optimization method based on SHAP values is adopted. By constructing a billet defect prediction model, training the dataset using the XGBoost algorithm, calculating the SHAP values, and conducting quality control analysis, the recommended range of process parameter values is determined.
It significantly reduced experimental costs and time, improved the interpretability and adaptability of parameter optimization, met the needs of online applications, and enhanced product quality stability and production efficiency.
Smart Images
Figure CN122153299A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of steel continuous casting technology, and in particular to a method and system for optimizing continuous casting process parameters based on shap values. Background Technology
[0002] Continuous casting is a crucial step in steel production, and its process parameters directly impact billet quality and production efficiency. Common methods for optimizing continuous casting process parameters include: expert system methods based on empirical rules, which establish an expert knowledge base and match historical experience parameters based on steel grade, cross-sectional specifications, and other conditions; experimental design methods based on statistical analysis, which design experimental schemes, establish statistical models of process parameters and quality indicators, and obtain parameter combinations through model optimization; and optimization methods based on traditional machine learning, which establish mapping relationships between process parameters and quality indicators through neural networks and search for optimal parameters using optimization algorithms.
[0003] However, existing technologies have the following problems: First, expert methods are highly dependent on experience, making it difficult to discover better parameter combinations that surpass experience; second, statistical analysis methods require a large amount of experimental data, resulting in long experimental cycles and high costs; third, traditional models have poor interpretability, making it difficult for engineers to understand and trust the optimization results, and they are computationally inefficient, with multi-objective optimization algorithms being computationally complex and unable to meet the needs of online optimization; at the same time, existing methods are difficult to adapt to dynamic production environments such as changes in working conditions and steel grade switching, resulting in poor applicability and stability. Summary of the Invention
[0004] Based on the above analysis, the embodiments of the present invention aim to provide a method and system for optimizing continuous casting process parameters based on shap values, in order to solve the problems of strong reliance on experience, poor model interpretability, low computational efficiency and insufficient adaptability in the existing technology.
[0005] On one hand, embodiments of the present invention provide a method for optimizing continuous casting process parameters based on shap values, including:
[0006] Historical process parameters and historical quality detection data in steel continuous casting production were collected, and a historical training dataset was obtained. Based on the historical training dataset, the historical process parameters are used as input, and the defect identifiers in the historical quality detection data are used as output to construct a billet defect prediction model. Real-time process parameters of the target billet are collected and preprocessed to form a real-time process parameter dataset. Based on the real-time process parameter dataset and the trained billet defect prediction model, the SHAP value corresponding to each process parameter of the target billet is calculated. Quality control analysis is performed on each SHAP value of the target billet, and the recommended range of process parameters for the target billet is determined.
[0007] Furthermore, the obtained historical training dataset includes: Based on the product material number, the historical process parameters and the historical quality detection data are associated to form a one-to-one corresponding data group, and a historical training dataset is obtained.
[0008] Furthermore, the historical training dataset is represented in the following form: ; in, Represents the historical training dataset, , ... This represents a sample of historical process parameters, and each sample... All are N-dimensional feature vectors, with each dimension representing a specific historical process parameter; , ... This represents each historical quality detection data sample, with each sample... Both are scalars, taking values of 0 or 1; V represents the number of samples in the historical training dataset.
[0009] Furthermore, the billet defect prediction model is implemented based on the XGBoost algorithm, and the model training process includes: Precision, recall, and F1 score are used as evaluation metrics to evaluate the prediction results of the billet defect prediction model, wherein the F1 score is based on the harmonic mean of precision and recall. Based on the model evaluation results, the model was validated using the K-fold cross-validation method, and the hyperparameters of the model were optimized using the grid search method to obtain the trained billet defect prediction model.
[0010] Furthermore, the real-time process parameters of the target billet are collected and preprocessed to form a real-time process parameter dataset, including: The real-time process parameters are cleaned and feature-engineered to form the real-time process parameter dataset. The real-time process parameter dataset includes multiple real-time process parameter samples, and each real-time process parameter sample includes multiple real-time process parameters processed by feature engineering.
[0011] Furthermore, the real-time process parameter dataset is represented in the following form: ; in, This represents a real-time process parameter dataset. , ... This represents a sample of each real-time process parameter, and each sample... All are N-dimensional feature vectors, with each dimension representing a specific real-time process parameter; M represents the number of samples in the real-time process parameter dataset.
[0012] Furthermore, the formula for calculating the SHAP value includes: ; in, This indicates that for model f and samples The SHAP value; This indicates the completed training of the billet defect prediction model. This represents a sample in the real-time process parameter dataset; Represents a subset of features; Indicates that features are not included. any subset of; Indicates the size of the feature; Indicates the portfolio weights; Indicates use In feature subset The process parameters in the data and the current number are added. The average value of the prediction results obtained by model f for each process parameter; Indicates use In feature subset The process parameters are the average values of the prediction results obtained through model f.
[0013] Furthermore, the quality control analysis includes at least: global analysis, local analysis, and dependency analysis regarding the SHAP value; The global analysis includes: determining the global importance of each real-time process parameter based on the SHAP value corresponding to each process parameter; The local analysis includes: interpreting the prediction results of a single real-time process parameter based on the SHAP value corresponding to each process parameter, and quantifying the individual contribution of each real-time process parameter to the prediction results; The dependency analysis includes analyzing the interaction between each of the real-time process parameters and the corresponding SHAP value.
[0014] Furthermore, the step of performing quality control analysis on each SHAP value of the target billet and determining the recommended range of process parameters for the target billet includes: Based on the combined results of the global analysis, local analysis, and dependency analysis, the parameter optimization analysis results of the target billet are obtained, and the recommended range of process parameters for the target billet is determined.
[0015] On the other hand, embodiments of the present invention provide a continuous casting process parameter optimization device based on shap values, comprising: The training data acquisition module is used to collect historical process parameters and historical quality detection data in steel continuous casting production, and to obtain historical training datasets. The prediction model building module is used to build a billet defect prediction model based on the historical training dataset, with the historical process parameters as input and the defect identifiers in the historical quality detection data as output. The SHAP value calculation module is used to collect real-time process parameters of the target billet, and form a real-time process parameter dataset after preprocessing; based on the real-time process parameter dataset and the trained billet defect prediction model, the SHAP value corresponding to each process parameter of the target billet is calculated. The SHAP value analysis module is used to perform quality control analysis on each SHAP value of the target billet and determine the recommended range of process parameters for the target billet.
[0016] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: First, unlike related technologies that rely excessively on human experience, have high experimental costs, and poor adaptability, this invention constructs a billet defect prediction model through a data-driven approach. This model can automatically learn complex and nonlinear relationships from massive production data, significantly reducing experimental costs and shortening the cycle. At the same time, the model can adaptively learn quality laws under different complex working conditions, effectively solving the problem of insufficient adaptability when working conditions change and steel grades are switched.
[0017] Secondly, unlike related technologies where the optimization of process parameters is not high and the interpretability of models is poor, this invention improves the interpretability of parameters by introducing a SHAP value quality control analysis mechanism. It clarifies the degree and direction of the influence of each process parameter on quality indicators, giving the optimization results a clear physical meaning. This enables rapid parameter optimization to meet the needs of online applications, and also improves the ability of multi-objective collaborative optimization and self-adaptation. In turn, it provides precise guidance for the dynamic adjustment of process parameters, resulting in a significant improvement in product quality stability.
[0018] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description
[0019] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.
[0020] Figure 1 This is a flowchart of the continuous casting process parameter optimization method based on shap value according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the SHAP values of the steel billet process parameters in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the dependency analysis between the outbound Ti value and the SHAP value in an embodiment of the present invention; Figure 4 This is a schematic diagram of the main modules of the continuous casting process parameter optimization device based on shap value according to an embodiment of the present invention. Detailed Implementation
[0021] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.
[0022] A specific embodiment of the present invention discloses a method for optimizing continuous casting process parameters based on shap values, such as... Figure 1 As shown, the steps S1 to S4 are as follows: Step S1: Collect historical process parameters and historical quality detection data in steel continuous casting production, and obtain historical training dataset.
[0023] During implementation, process parameters are collected from the primary and secondary equipment of the production system, while quality inspection data is automatically determined by the plotting instrument and further assisted by manual judgment. Historical process parameters collected need to be filtered; based on operational experience in steel billet casting, process parameter data irrelevant to subsequent defect prediction tasks, such as ladle number, shift number, and furnace number, are directly removed, while process parameter data from the continuous casting production process are retained.
[0024] The historical process parameters that need to be retained include at least: final temperature, final carbon, final oxygen level, carbon-oxygen product, final slag TFE (total iron content in the final slag), final slag R (basicity in the final slag), modifier addition amount, number of reblowing cycles, inlet oxygen level PPM (oxygen content when molten steel arrives at the refining station), final deoxidation level PPM (oxygen content after final deoxidation operation), oxygen blowing rate, settling time, aluminum consumption, total aluminum consumption, outlet Als (aluminum content dissolved in acid when molten steel leaves the station after refining), outlet Ti (titanium content when molten steel leaves the station after refining), Al / Ti (aluminum-titanium ratio), RH oxygen (oxygen content during RH vacuum refining), RH slag TFe (total iron content in the slag at the RH refining station), Ca / Al (… The parameters include: calcium-aluminum ratio, slag basicity (R), superheat, casting cycle, finished product Als (acid-soluble aluminum content), finished product Ti (titanium content), Als burn-off (aluminum burn-off), Ti burn-off (titanium burn-off), tundish oxygen, tundish nitrogen, nitrogen addition, molten steel time in the ladle, and overall temperature drop. It can be understood that these 32 process parameters correspond to the N-dimensional feature vector of each historical process parameter sample in the historical training dataset, i.e., N≥32; it can also be understood that once the process parameters are determined, N is a fixed value, for example, N=32.
[0025] For the historical quality inspection data collected, it is necessary to mark whether the corresponding steel billet has defects according to the quality inspection results. For example, if the quality inspection data of a steel billet indicates that it has one of the following: slag inclusions or peeling, then the steel billet is marked as having quality defects; otherwise, it is marked as having no quality defects.
[0026] Preferably, the collected historical process parameters and historical quality detection data need to undergo data preprocessing operations such as data cleaning and feature engineering. Data cleaning includes at least outlier removal and missing value imputation. For example, outlier removal includes: filtering and deleting outliers from each historical billet data based on the 3σ principle; missing value imputation includes: filling missing values using the mode of the historical data. Feature engineering includes steps such as feature transformation, feature construction, feature selection, and feature encoding; for example, feature transformation converts unevenly distributed historical process parameters into a standard normal distribution; feature construction performs calculations on some original parameters and generates new features; feature selection is performed using the chi-square test, and process parameter dimensions irrelevant to billet defect prediction are filtered out; feature encoding discretizes continuous historical quality detection data, marking samples with quality defects as 1 and samples without quality defects as 0; thus, a historical training dataset suitable for model training is constructed after the preprocessing operations.
[0027] The process of constructing the dataset specifically includes: based on the product part number, associating the historical process parameters and the historical quality detection data to form a one-to-one corresponding data group, such as... And obtain the historical training dataset. The final historical training dataset is represented in the following form: ; in, Represents the historical training dataset, , ... This represents a sample of historical process parameters, and each sample... All are N-dimensional feature vectors, where N represents the number of process parameters; each dimension represents a specific historical process parameter. , ... This represents each historical quality detection data sample, with each sample... Both are scalars, taking values of 0 or 1; V represents the number of samples in the historical training dataset.
[0028] Step S2: Based on the historical training dataset, using the historical process parameters as input and the defect identifiers in the historical quality detection data as output, construct a billet defect prediction model.
[0029] In implementation, the billet defect prediction model can be based on the XGBoost algorithm. The model training process includes: evaluating the prediction results of the billet defect prediction model using accuracy, recall, and F1 score as evaluation indicators; verifying the model based on the evaluation results using the K-fold cross-validation method; and optimizing the model hyperparameters using the grid search method to obtain the trained billet defect prediction model.
[0030] Predicting whether a steel billet has quality defects is understandably a binary classification task; that is, the model divides each data sample (corresponding to one billet) into two categories: "defective" or "defect-free." Therefore, to evaluate the model's classification performance, its predictions need to be compared with known actual quality conditions, and all samples need to be divided into the following four cases: True positive (TP): This indicates that the actual situation is defective, and the model also correctly predicts that it is defective, representing that the model correctly identifies the defective billet.
[0031] False positive (FP): This means that the actual situation is defect-free, but the model incorrectly predicts that there is a defect. It represents a false alarm by the model, misclassifying a qualified product as a defective product.
[0032] True negative (TN): This indicates that the actual situation is defect-free, and the model also correctly predicts that it is defect-free, representing a qualified billet correctly identified by the model.
[0033] False negative (FN): This indicates that the actual situation is defective, but the model incorrectly predicts that it is defect-free. This means that the model missed the report and failed to identify the real defective billet.
[0034] By statistically analyzing the number of samples corresponding to the four scenarios mentioned above, evaluation metrics such as accuracy, recall, and F1 score can be calculated, thereby providing a comprehensive quantitative evaluation of the model's performance. For example, let TP, FP, TN, and FN represent the number of samples corresponding to these four scenarios, then TP + FP + TN + FN = total number of samples. Accuracy P is defined as: Recall rate R is defined as: The F1 value is defined as: In other words, the F1 score is based on the harmonic mean of accuracy P and recall R, and the value is between 0 and 1. The larger the value, the better the training effect of the model.
[0035] Furthermore, overfitting can be prevented through K-fold cross-validation, specifically by first randomly dividing the historical training dataset... Divide into K disjoint subsets of the same size; then, use K... The model is trained using one subset of data and tested using the remaining subset; finally, the model with the largest average F1 score is selected by combining the evaluation results of K iterations.
[0036] Grid search is an exhaustive search method that traverses all possible combinations of hyperparameters in a model to find the optimal hyperparameters. Specifically, the grid search method involves: assigning a set of candidate values for each hyperparameter of the model, generating a Cartesian product of these candidate values, and forming a grid of hyperparameter combinations. Then, model training and performance evaluation are performed on each parameter combination in the grid, and the best-performing hyperparameter combination is found.
[0037] Therefore, by constructing and training a billet defect prediction model, it is possible to automatically learn complex and nonlinear relationships from massive production data, discover patterns that are difficult to summarize manually, and better adapt to complex and new working conditions.
[0038] The billet defect prediction model can also be implemented based on the random forest algorithm or deep learning algorithm (such as the DNN model); details about the construction and training process of the billet defect prediction model can be found in existing technologies, which will not be elaborated here.
[0039] Step S3: Collect real-time process parameters of the target billet, and form a real-time process parameter dataset after preprocessing; calculate the SHAP value corresponding to each process parameter of the target billet based on the real-time process parameter dataset and the trained billet defect prediction model.
[0040] The formation of a real-time process parameter dataset specifically includes: performing data cleaning on the real-time process parameters, and performing feature engineering on the cleaned real-time process parameters to form the real-time process parameter dataset; the data cleaning and feature engineering of the real-time process parameters are similar to the data preprocessing operations of the historical process parameters mentioned above, and will not be described in detail here.
[0041] After data preprocessing, the real-time process parameter dataset includes multiple real-time process parameter samples, and each real-time process parameter sample includes multiple real-time process parameters processed by feature engineering. The real-time process parameter dataset is represented in the following form: ; in, This represents a real-time process parameter dataset. , ... This represents a sample of each real-time process parameter, and each sample... All are N-dimensional feature vectors, with each dimension representing a specific real-time process parameter; M represents the number of samples in the real-time process parameter dataset. Furthermore, the N-dimensional feature vectors of the real-time process parameter samples here have the same dimension as the aforementioned historical process parameter samples, i.e., N≥32. Once the process parameters are determined, N is a fixed value, for example, N=32. However, the number of samples in the real-time process parameter dataset may differ from the number of samples in the historical training dataset, which will not be elaborated upon here.
[0042] The technical principle behind calculating the SHAP value is that it assesses the change in model predictions caused by adding a feature to any subset of features. In other words, the SHAP value characterizes the contribution of a single feature relative to the baseline prediction (the average prediction when all features are missing). A positive value indicates that the feature improves the prediction result, while a negative value indicates that it reduces the prediction result. Simultaneously, the SHAP value aims to explain the model's prediction of the input data, i.e., the probability of whether the steel billet has quality defects. Therefore, in calculating the SHAP value, this invention does not rely on the final defect classification label output by the model. Instead, it is based on a trained billet defect prediction model and its real-time process parameter inputs. The contribution of each parameter—the SHAP value—is derived by analyzing the model's internal prediction results for different parameter subsets.
[0043] Specifically, in this embodiment of the invention, the real-time process parameter dataset is input into the trained billet defect prediction model, and the SHAP value of each process parameter can be output by analyzing the response inside the model.
[0044] Furthermore, the formula for calculating the SHAP value includes: ; in, This indicates that for model f and samples The SHAP value; This indicates the completed training of the billet defect prediction model. This represents a sample in the real-time process parameter dataset; Represents a subset of features; Indicates that features are not included. any subset of; Indicates the size of the feature; Indicates the portfolio weights; Indicates use In feature subset The process parameters in the data and the current number are added. The average value of the prediction results obtained by model f for each process parameter; Indicates use In feature subset The process parameters are the average values of the prediction results obtained through model f.
[0045] Therefore, based on the SHAP values of each process parameter of the target billet, the influence of process parameters of different process links on the quality defects of steel billets can be determined, clearly presenting the parameter distribution and variation law, improving interpretability, and thus providing a clear direction for process parameter adjustment.
[0046] Step S4: Perform quality control analysis on each SHAP value of the target billet and determine the recommended range of process parameters for the target billet.
[0047] The quality control analysis includes at least: global analysis, local analysis, and dependency analysis of the SHAP value.
[0048] The global analysis includes: determining the global importance of each real-time process parameter based on the SHAP value corresponding to each process parameter; the global analysis is mainly used to evaluate the overall impact of each feature in the model on the prediction results, and quantifies the contribution of each feature through the SHAP value.
[0049] For example, for real-time process parameter datasets Each sample in the dataset can be individually analyzed for its first... SHAP value of each process parameter We take the average of the absolute values to get the first... Global importance of each process parameter The calculation formula is as follows:
[0050] Where M represents the dataset The number of steel billet samples; Indicates sample The Middle SHAP values of each process parameter; The larger the value, the higher the number of . The greater the correlation between process parameters and the presence of quality defects in steel billets, the better. By determining the value of the parameter, we can filter out the process parameters that need to be optimized.
[0051] The local analysis includes: interpreting the prediction results of a single real-time process parameter based on the SHAP value corresponding to each process parameter, and quantifying the individual contribution of each real-time process parameter to the prediction results; the local analysis helps to understand the model decision-making process by quantifying the contribution of each feature to the prediction.
[0052] Assuming that the global analysis of the aforementioned SHAP values has identified the key areas requiring optimization... Given a process parameter, then for the dataset... Specific samples , its first The SHAP value of each process parameter is expressed as follows: ,if If the value of is positive, then it means that the th Several process parameters can contribute to quality defects in steel billets; if If the value of is negative, then it means that the th Several process parameters play a role in suppressing quality defects in steel billets.
[0053] The dependency analysis includes analyzing the interaction between each of the real-time process parameters and their corresponding SHAP values. Through dependency analysis, a visualized SHAP dependency graph can be constructed, which is used to analyze the relationship between the feature values and SHAP values (feature contribution) of each sample in the model, thereby revealing the nonlinear impact and potential interactions of the process parameters in the samples on the model prediction.
[0054] For example, based on each sample and its corresponding SHAP value, multiple data points can be constructed as follows:
[0055] in, Represents sample points The Middle The values of each process parameter; Indicates sample The Middle The SHAP value of each process parameter; M represents the number of steel billet samples in dataset D2. By constructing the above multiple data points, a SHAP dependency graph can be plotted, and the SHAP dependency graph can be analyzed. The positive and negative correlation between process parameters and whether steel billets produce quality defects.
[0056] Furthermore, by integrating the analysis results of global analysis, local analysis, and dependency analysis, the parameter optimization analysis results of the target billet are obtained, and the recommended value range of the process parameters of the target billet is determined.
[0057] The parameter optimization analysis results are a comprehensive interpretation based on global, local, and dependency analyses of the SHAP value. Specifically, global and local analyses identify process parameters that significantly influence the prediction of casting defects, thus determining the set of parameters to be prioritized for optimization. Dependency analysis further reveals the dynamic relationship between these parameter values and the SHAP value. By analyzing the SHAP dependency graph, if the SHAP value is consistently negative, indicating that the interval corresponding to the process parameter has a suppressive effect on defects, then this interval can be used as the recommended interval.
[0058] like Figure 2 As shown, Figure 2 The horizontal axis represents the SHAP value of each process parameter, and the vertical axis represents the process parameters that need to be optimized. A color gradient visualization design is used to intuitively represent the numerical variation of the process parameters. The color spectrum transitions continuously from blue to red, representing the numerical trend of the process parameter values from small to large. Taking Ti (titanium content at the time of leaving the station) as an example, when the value of Ti at the time of leaving the station is small, corresponding to the blue value point in the graph, its SHAP value is less than 0, indicating that Ti at the time of leaving the station has an inhibitory effect on the quality defects of the steel billet; when the value of Ti at the time of leaving the station is large, corresponding to the red value point in the graph, its SHAP value is greater than 0, indicating that Ti at the time of leaving the station has a promoting effect on the quality defects of the steel billet.
[0059] Furthermore, such as Figure 3 As shown, Figure 3 The horizontal axis represents the value of Ti at the outlet of the process parameter (such as content), and the vertical axis represents the SHAP value of Ti at the outlet of the process parameter. It can be seen that when the value of Ti at the outlet is greater than 0.065, the SHAP value of Ti at the outlet increases with the increase of the value of Ti at the outlet, indicating that the probability of quality defects in steel billets increases. When the value of Ti at the outlet is between 0.056 and 0.065, the SHAP value of Ti at the outlet is less than 0, indicating that Ti at the outlet has an inhibitory effect on the quality defects in steel billets. Therefore, it can be used as a recommended value range.
[0060] By analyzing SHAP values, the degree and direction of the influence of each process parameter on quality indicators are clarified, giving the optimization results physical meaning. Production operators can adjust relevant process parameters according to the recommended value range of each process parameter, effectively optimizing the continuous casting production process conditions, minimizing the probability of billet defects, and improving product quality and production efficiency.
[0061] It is understood that the above embodiments are only for ease of understanding and simplification of description, and should not be construed as limiting the present invention. The present invention does not impose specific limitations on the selection of process parameters, model construction, calculation methods, etc.
[0062] Therefore, the embodiments of the present invention, on the one hand, construct a billet defect prediction model through a data-driven approach, which can automatically learn complex and nonlinear relationships from massive production data, significantly reducing experimental costs and shortening the cycle; at the same time, the model can adaptively learn quality laws under different complex working conditions, effectively solving the problem of insufficient adaptability when working conditions change and steel grades are switched. On the other hand, by introducing a SHAP value quality control analysis mechanism, the interpretability of parameters is improved, giving the optimization results clear physical meaning. Compared with traditional genetic algorithms, the convergence speed is increased by 3-5 times, meeting the needs of online applications, improving optimization efficiency, and enhancing multi-objective collaborative optimization capabilities and adaptability, thereby providing precise guidance for the dynamic adjustment of process parameters. Experimental data shows that after the implementation of the present invention, the incidence of billet cracks is reduced by 45%, and the pass rate of central porosity is increased by 30%, resulting in a significant improvement in product quality stability.
[0063] In another embodiment of the present invention, a continuous casting process parameter optimization device based on shap value is proposed, such as... Figure 4 As shown, it specifically includes the following modules: The training data acquisition module is used to collect historical process parameters and historical quality detection data in steel continuous casting production, and to obtain historical training datasets. The prediction model building module is used to build a billet defect prediction model based on the historical training dataset, with the historical process parameters as input and the defect identifiers in the historical quality detection data as output. The SHAP value calculation module is used to collect real-time process parameters of the target billet, and form a real-time process parameter dataset after preprocessing; based on the real-time process parameter dataset and the trained billet defect prediction model, the SHAP value corresponding to each process parameter of the target billet is calculated. The SHAP value analysis module is used to perform quality control analysis on each SHAP value of the target billet and determine the recommended range of process parameters for the target billet.
[0064] The above-described method and apparatus embodiments are based on the same principle, and their related aspects can be referenced from each other to achieve the same technical effect. For specific implementation processes, please refer to the foregoing embodiments, which will not be repeated here.
[0065] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.
[0066] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for optimizing continuous casting process parameters based on shap values, characterized in that, include: Historical process parameters and historical quality detection data in steel continuous casting production were collected, and a historical training dataset was obtained. Based on the historical training dataset, the historical process parameters are used as input, and the defect identifiers in the historical quality detection data are used as output to construct a billet defect prediction model. Real-time process parameters of the target billet are collected and preprocessed to form a real-time process parameter dataset. Based on the real-time process parameter dataset and the trained billet defect prediction model, calculate the SHAP value corresponding to each process parameter of the target billet. Quality control analysis is performed on each SHAP value of the target billet, and the recommended range of process parameters for the target billet is determined.
2. The optimization method according to claim 1, characterized in that, The obtained historical training dataset includes: Based on the product material number, the historical process parameters and the historical quality detection data are associated to form a one-to-one corresponding data group, and a historical training dataset is obtained.
3. The optimization method according to claim 2, characterized in that, The historical training dataset is represented in the following form: ； in, Represents the historical training dataset, , ... This represents a sample of historical process parameters, and each sample... All are N-dimensional feature vectors, with each dimension representing a specific historical process parameter; , ... This represents each historical quality detection data sample, with each sample... Both are scalars, taking values of 0 or 1; V represents the number of samples in the historical training dataset.
4. The optimization method according to claim 1, characterized in that, The billet defect prediction model is implemented based on the XGBoost algorithm, and the model training process includes: Precision, recall, and F1 score are used as evaluation metrics to evaluate the prediction results of the billet defect prediction model, wherein the F1 score is based on the harmonic mean of precision and recall. Based on the model evaluation results, the model was validated using the K-fold cross-validation method, and the hyperparameters of the model were optimized using the grid search method to obtain the trained billet defect prediction model.
5. The optimization method according to claim 4, characterized in that, The real-time process parameters of the target billet are collected and preprocessed to form a real-time process parameter dataset, including: The real-time process parameters are cleaned and feature-engineered to form the real-time process parameter dataset. The real-time process parameter dataset includes multiple real-time process parameter samples, and each real-time process parameter sample includes multiple real-time process parameters processed by feature engineering.
6. The optimization method according to claim 5, characterized in that, The real-time process parameter dataset is represented in the following form: ； in, This represents a real-time process parameter dataset. , ... This represents a sample of each real-time process parameter, and each sample... All are N-dimensional feature vectors, with each dimension representing a specific real-time process parameter; M represents the number of samples in the real-time process parameter dataset.
7. The optimization method according to claim 6, characterized in that, The formula for calculating the SHAP value includes: ； in, This indicates that for model f and samples The SHAP value; This indicates the completed training of the billet defect prediction model. This represents a sample in the real-time process parameter dataset; Represents a subset of features; Indicates that features are not included. any subset of; Indicates the size of the feature; Indicates the portfolio weights; Indicates use In feature subset The process parameters in the data and the current number are added. The average value of the prediction results obtained by model f for each process parameter; Indicates use In feature subset The process parameters are the average values of the prediction results obtained through model f.
8. The optimization method according to claim 1, characterized in that, The quality control analysis includes at least: global analysis, local analysis, and dependency analysis of the SHAP value; The global analysis includes: determining the global importance of each real-time process parameter based on the SHAP value corresponding to each process parameter; The local analysis includes: interpreting the prediction results of a single real-time process parameter based on the SHAP value corresponding to each process parameter, and quantifying the individual contribution of each real-time process parameter to the prediction results; The dependency analysis includes analyzing the interaction between each of the real-time process parameters and the corresponding SHAP value.
9. The optimization method according to claim 8, characterized in that, The quality control analysis of each SHAP value of the target billet and the determination of the recommended range of process parameters for the target billet include: Based on the combined results of the global analysis, local analysis, and dependency analysis, the parameter optimization analysis results of the target billet are obtained, and the recommended range of process parameters for the target billet is determined.
10. A device for optimizing continuous casting process parameters based on shap values, characterized in that, include: The training data acquisition module is used to collect historical process parameters and historical quality detection data in steel continuous casting production, and to obtain historical training datasets. The prediction model building module is used to build a billet defect prediction model based on the historical training dataset, with the historical process parameters as input and the defect identifiers in the historical quality detection data as output. The SHAP value calculation module is used to collect real-time process parameters of the target billet and form a real-time process parameter dataset after preprocessing. Based on the real-time process parameter dataset and the trained billet defect prediction model, calculate the SHAP value corresponding to each process parameter of the target billet. The SHAP value analysis module is used to perform quality control analysis on each SHAP value of the target billet and determine the recommended range of process parameters for the target billet.