A method for judging whether the head component of a continuously cast mixed billet is qualified based on SVM

Abstract

The application relates to a SVM-based method for judging whether the head component of a continuously-cast mixed-cast blank is qualified or not, which comprises the following steps: collecting mixed-cast process data and detection results at a molten steel impact time point of a subsequent furnace in each mixed-cast production and establishing a historical data set; using the data with consistent head component detection results and performance detection results, determining optimal SVM model superparameters through a gridding parameter optimization method; and according to the optimal model superparameters, training the model and using the trained model to judge whether the head component of the mixed-cast blank is qualified or not. The application has high adaptability to the data characteristics of the mixed-cast blank head component, i.e. small sampling data volume and a large number of abnormal values, and simultaneously uses the artificial intelligence and parameter optimization methods to improve the model judgment precision.

Description

Technical Field

[0001] This patent application belongs to the technical field of continuous casting methods in the metallurgical industry. More specifically, it relates to a method for determining whether the composition of the head of a continuously cast mixed billet is qualified based on SVM. Background Technology

[0002] Faced with increasingly fierce competition in the steel market, steel companies' customer contracts are characterized by multiple varieties, specifications, and small batches, leading to increased overlap in production processes. Continuous casting of dissimilar steel grades can effectively meet these needs, ensuring production continuity and stability while significantly improving production efficiency and reducing costs, making it an important choice for steel companies. However, due to differences in steel composition between successive furnaces, mixed-cast billets are unavoidable. If the mixed-cast billets cannot be accurately identified, it will significantly impact the performance of subsequent rolled products. Because the previous furnace's molten steel still has a liquid core within the crystallizer when the next furnace is cast, the subsequent furnace's molten steel impacts the unsolidified liquid core of the previous furnace's billet during the casting process, mixing with it. This process is influenced by multiple factors, including material diffusion, liquid flow field, and the gradual solidification of the billet. Therefore, in actual production, verifying the composition in this area requires extensive sampling and analysis of the billet's center. However, sampling the billet's core is not only extremely difficult compared to the edges, but destructive sampling also results in billet loss, making it difficult to accumulate a large amount of effective data in actual production.

[0003] Because the composition of the transition billet caused by the unsteady process of casting impact is highly variable, actual production requires consideration of billet length requirements to ensure subsequent performance meets standards. For safety, excessive cutting is often performed in the direction of the previous heat's composition. Consequently, most samples taken from the center of the billet contain only the composition of the previous heat. This only indicates that the billet is acceptable at that location, but it does not mean that this location is precisely the boundary between acceptable and unacceptable. Under the constraint of billet length requirements, accurately finding the interface between the previous and subsequent heat compositions and cutting accordingly is extremely difficult. Data generated under these circumstances will exhibit significant biases when using conventional methods for statistical analysis or mathematical regression modeling. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for determining whether the composition of the head of a continuously cast mixed billet is qualified based on SVM. By adopting the support vector method in SVM, it has high adaptability to the characteristics of data with small sample size of mixed billet composition and a large number of outliers. At the same time, the use of artificial intelligence and parameter optimization methods improves the accuracy of model judgment. It can effectively solve the problems of insufficient effective data accumulated in actual production and insufficient representativeness of a large number of sampled data in the background technology.

[0005] To solve the above problems, the technical solution adopted by the present invention is as follows: A method for determining whether the composition of the head of a continuously cast mixed billet is qualified based on SVM includes the following steps: S1. Calculate the impact time of the next batch of molten steel in each mixing process using the following formula, t 冲击 =t 开浇 +W / (V×L×H×23.4), where t 冲击 The time point for the impact of the next batch of molten steel; t 开浇 The time point at which the next heat of molten steel is poured; W is the weight of the tundish at the start of the next heat of molten steel pouring, in tons; V is the casting speed at the start of the continuous casting machine pouring, in m / min; L is the width of the billet, in meters; H is the thickness of the billet, in meters. S2. Collect historical datasets for each batch of mixed casting production, including the casting speed at the impact time of the next heat of molten steel, the width and thickness of the produced billet, the composition difference ratio between the previous and subsequent heats, the cutting length of the mixed casting head, the composition determination result of the head of the produced mixed casting billet, and the performance determination result of the steel plate subsequently processed from the last good billet of the previous heat. The composition difference ratio between the previous and subsequent heats is the maximum value among the larger and smaller values ​​of any alloy component in the molten steel from the previous and subsequent heats; when the calculated result is greater than 5, it is set to 5. The cutting length of the mixed casting head is the distance between the cutting position of the mixed casting billet head and the meniscus position of the billet when the next heat of molten steel begins casting. A qualified composition determination result for the head of the produced mixed casting billet is recorded as 1, and an unqualified result as 0. A qualified performance determination result for the steel plate subsequently processed from the last good billet of the previous heat is recorded as 1, and an unqualified result as 0.

[0006] S3. The casting speed, width, thickness, compositional difference ratio between the previous and next furnaces, and cutting length of the mixed casting head in the historical dataset at the time of impact of the next furnace of molten steel are used as independent variables of the model, and the compositional determination result of the mixed casting head is used as the dependent variable of the model. S4. Select all data in the historical dataset whose head component determination results are consistent with the performance determination results, randomly shuffle their order, and select 80% of them as the model training set and the remaining 20% ​​as the model test set. S5. In the model training set and model test set, each data point with unqualified head component judgment result and performance judgment result is copied into multiple identical data points (e.g., 5 to 7 data points, taking 5 data points as an example) and incorporated into the original model training set or model test set as the final dataset. S6. Train multiple (e.g., 49) SVM models on the training set data, and test the accuracy of each SVM model on the test set data. The hyperparameters of the SVM model with the highest accuracy are determined as the optimal hyperparameters. The kernel functions of the 49 SVM models are all radial basis functions, and the hyperparameters are permutations and combinations of relaxation factors C={0.125, 0.25, 0.5, 1, 2, 4, 8} and kernel function parameters gamma={0.025, 0.05, 0.1, 0.2, 0.4, 0.8, 1.6}.

[0007] S7. Reconstruct the SVM model using the optimal hyperparameters and train it using the final dataset. The final dataset is organized as follows: for each data point in the historical dataset described in S2 where both the head component determination result and the performance determination result are unqualified, copy it into five identical data points and incorporate them into the original dataset as the final dataset.

[0008] S8. The final SVM model after training is used to determine whether the composition of the head of the mixed-cast billet to be predicted is qualified.

[0009] Due to the adoption of the above technical solution, the beneficial effects achieved by this invention are: This method is simple, feasible, and highly accurate. It adopts the support vector method in SVM (Support Vector Machine), which is highly adaptable to the characteristics of data with small sample size of mixed casting components and a large number of outliers. At the same time, the use of artificial intelligence and parameter optimization methods improves the accuracy of model judgment and has extremely high value for promotion and application. Attached Figure Description

[0010] Figure 1 This is a flowchart of the method. Detailed Implementation

[0011] The present invention will be further described in detail below with reference to the embodiments.

[0012] A method for determining the quality of the head composition of continuously cast mixed billets based on SVM, such as... Figure 1 For example, in the implementation example: The specific steps are as follows: Collect the process data and test results of the mixing process at the impact time point of the last heat of molten steel in each mixing production and establish a historical dataset. Using data where the head composition test results and performance test results are consistent, determine the optimal hyperparameters of the SVM model through a gridded parameter optimization method. Train the model based on the optimal model hyperparameters and use the trained model to determine whether the head composition of the mixed billet is qualified.

[0013] Step 1: Calculate the impact time of the next heat of molten steel in each batch of casting production using the following formula: t_impact = t_start_casting + W / (V × L × H × 23.4), where t_impact is the impact time of the next heat of molten steel; t_start_casting is the start_casting time of the next heat of molten steel; W is the weight of the tundish at the start of the next heat of molten steel, in tons; V is the casting speed at the start of the continuous casting machine, in m / min; L is the width of the billet, in meters; and H is the thickness of the billet, in meters. Taking a specific mixed casting production as an example, the start time for the next heat of molten steel is 03:12:30. At the start of casting, the ladle weight is 8.9t, the casting speed is 0.65m / min, and the produced billet width is 2.2m and the thickness is 0.3m. Therefore, the calculated impact time of the molten steel is t_impact = 03:12:30 + 8.9 / (0.65 × 2.2 × 0.3 × 23.4) = 03:13:23. Because the next heat of molten steel must pass through the tundish and mix with the previous heat before reaching the crystallizer to impact and mix with the liquid core of the previous heat, this empirical formula can be used to find the accurate impact time.

[0014] The second step involves collecting historical datasets for each batch of mixed casting production. These datasets include the casting speed at the impact time of the next heat of molten steel, the width and thickness of the produced billet, the compositional difference ratio between the previous and subsequent heats, the cutting length of the mixed casting head, the compositional determination result of the produced mixed casting billet head, and the performance determination result of the steel plate subsequently processed from the last good billet from the previous heat. The compositional difference ratio between the previous and subsequent heats is the maximum value among the larger and smaller values ​​of any alloy component in the molten steel from the previous and subsequent heats. When the calculated result is greater than 5, it is set to 5. The cutting length of the mixed casting head is the distance between the cutting position of the mixed casting billet head and the meniscus position of the billet when the next heat of molten steel begins casting. A passing compositional determination result for the produced mixed casting billet head is recorded as 1, and a failing result as 0. Similarly, a passing performance determination result for the steel plate subsequently processed from the last good billet from the previous heat is recorded as 1, and a failing result as 0.

[0015] The third step is to use the casting speed at the impact time of the next heat of molten steel, the width and thickness of the produced billet, the proportion of composition difference between the previous and next heats, and the cutting length of the mixed casting head in the historical dataset as independent variables of the model, and the composition determination result of the mixed casting head as the dependent variable of the model.

[0016] Step 4: Select all data from the historical dataset whose head component determination results are consistent with the performance determination results. Randomly shuffle these data and select them proportionally as the model training set and model test set. For example, select 80% of these data as the model training set and the remaining 20% ​​as the model test set. In actual production, unsteady-state shock processes cause drastic changes in head components, which are often unevenly distributed even at the same cross-section. Therefore, head component sampling is sometimes not representative, and the performance of the final product fluctuates greatly. Therefore, using data whose head component determination results are consistent with the performance determination results to determine the hyperparameters of the SVM model in later steps can make the model results more stable and reliable, less prone to overfitting, and have stronger generalization ability.

[0017] Step 5: In both the model training and testing sets, each data point with unsatisfactory head component and performance evaluation results is copied multiple times (5-7 times) and incorporated into its original training or testing set as the final dataset. In this example, 5 data points are copied. In actual production, the costs associated with unsatisfactory head components and performance evaluation are often significant. Therefore, unsatisfactory data should be given more attention during model training. This operation adjusts its weight to 5 times that of normal data.

[0018] Step 6: Train multiple SVM models on the training set data separately, and test the accuracy of each SVM model on the test set data. The hyperparameters of the SVM model with the highest accuracy are determined as the optimal hyperparameters. In this embodiment, 49 SVM models are selected to train on the training set data separately. The kernel function of all 49 SVM models is the radial basis function, and the hyperparameters are permutations and combinations of relaxation factor C={0.125, 0.25, 0.5, 1, 2, 4, 8} and kernel function parameter gamma={0.025, 0.05, 0.1, 0.2, 0.4, 0.8, 1.6}. In this embodiment, the accuracy of the SVM models after training under different combinations of hyperparameters is shown in the table below. Therefore, the hyperparameters with the highest accuracy, C=0.25 and gamma=0.1, are selected as the optimal hyperparameters.

[0019]

[0020] Step 7: Reconstruct the SVM model using the optimal hyperparameters C=0.25 and gamma=0.1, and train it using the final dataset. The final dataset is organized as follows: for each data point in the historical dataset described in S2 where both the head component judgment result and performance judgment result are unqualified, duplicate it into five identical data points and incorporate them into the original dataset as the final dataset. Training the final model after weighting all the data can effectively improve the accuracy of the model's judgments.

[0021] Step 8: The final SVM model after training is used to determine whether the composition of the head of the mixed-cast billet to be predicted is qualified.

[0022] This invention collects process data and test results from the impact time of the next heat of molten steel in each batch of mixed casting production and establishes a historical dataset. Using data consistent with the head component test results and performance test results, the optimal hyperparameters of the SVM model are determined through a gridded parameter optimization method. The model is then trained based on these optimal hyperparameters, and the trained model is used to determine whether the head component of the mixed casting billet is qualified. By employing the support vector method in SVM, this invention demonstrates high adaptability to data characteristics such as a small sample size of head component data and a large number of outliers. Furthermore, the use of artificial intelligence and parameter optimization methods improves the model's judgment accuracy. This effectively solves the problems in the background technology, such as insufficient effective data accumulated in actual production and insufficient representativeness of large sample sizes.

Claims

1. A method for determining whether the composition of the head of a continuously cast mixed billet is qualified based on SVM, characterized in that... Includes the following steps: S1. Calculate the impact time of the next batch of molten steel in each batch of mixed casting production according to the following formula. t 冲击 =t 开浇 +W / (V x L x H x 23.4), in t 冲击 This refers to the timing of the impact of the next batch of molten steel. t 开浇 This refers to the time point for pouring the next batch of molten steel. W represents the weight of the tundish when the next batch of molten steel is poured, in tons. V is the casting speed at the start of continuous casting, in m / min; L is the width of the casting billet, in meters (m). H represents the thickness of the cast billet, in meters (m). S2. Collect the following data in each mixed casting production process: the casting speed at the impact time of the next heat of molten steel, the width and thickness of the produced billet, the proportion of compositional difference between the previous and subsequent heats, the cutting length of the mixed casting head, the compositional determination result of the head of the produced mixed casting billet, and the performance determination result of the steel plate subsequently processed from the last good billet of the previous heat. The cutting length of the mixed casting head is the distance between the cutting position of the head of the mixed casting billet and the position of the meniscus of the billet when the next heat of molten steel is poured. S3. The casting speed, width, thickness, compositional difference ratio between the previous and next furnaces, and head cutting length of the mixed casting in the historical dataset at the time of impact of the next furnace of molten steel are used as independent variables of the model, and the compositional determination result of the head of the mixed casting is used as the dependent variable of the model. S4. Select all data in the historical dataset whose head component determination results are consistent with the performance determination results, randomly shuffle their order, and select the model training set and model test set according to the proportion. S5. In both the model training set and the model test set, each data point whose head component determination result and performance determination result are unqualified is copied into multiple identical data points and incorporated into the original model training set or model test set as the final dataset. S6. Train multiple SVM models on the training set data respectively, and test the accuracy of each SVM model on the test set data. The hyperparameters of the SVM model with the highest accuracy are determined as the optimal hyperparameters. S7. Reconstruct the SVM model using the optimal hyperparameters and train it using the final dataset to form the final SVM model; S8. The final SVM model after training is used to determine whether the composition of the head of the mixed-cast billet to be predicted is qualified.

2. The method for determining whether the composition of the head of a continuously cast mixed billet based on SVM is qualified according to claim 1, characterized in that: In step S2, the ratio of the difference in composition between the front and rear furnaces is the maximum value of the ratio of the larger value to the smaller value of any alloy component in the molten steel in the front and rear furnaces. When the calculated result of the maximum value is greater than 5, it is set to 5.

3. The method for determining whether the composition of the head of a continuously cast mixed billet based on SVM is qualified according to claim 1, characterized in that: In step S2, the result of the head composition determination of the produced mixed billet is recorded as 1 if it is qualified and 0 if it is not qualified; the result of the performance determination of the steel plate processed from the last good billet of the previous furnace is recorded as 1 if it is qualified and 0 if it is not qualified.

4. The method for determining whether the composition of the head of a continuously cast mixed billet based on SVM is qualified according to claim 1, characterized in that: In step S4, 70% to 80% of the set is selected as the model training set, and the remaining 20% ​​to 30% is selected as the model test set.

5. The method for determining whether the composition of the head of a continuously cast mixed billet based on SVM is qualified according to claim 1, characterized in that: In step S5, each data point that is unqualified in both the head component determination result and the performance determination result is copied into 5 to 7 identical data points and incorporated into the original model training set or model test set.

6. A method for determining whether the composition of the head of a continuously cast mixed billet based on SVM is qualified, as described in any one of claims 1 to 5, characterized in that: In step S6, 49 SVM models are used to train the model training set data respectively, and the kernel function of the 49 SVM models is the radial basis function.

7. The method for determining whether the composition of the head of a continuously cast mixed billet based on SVM is qualified according to claim 6, characterized in that: In step S6, the hyperparameters of the 49 SVM models are permutations and combinations of relaxation factor C={0.125, 0.25, 0.5, 1, 2, 4, 8} and kernel function parameter gamma={0.025, 0.05, 0.1, 0.2, 0.4, 0.8, 1.6}.

8. The method for determining whether the composition of the head of a continuously cast mixed billet based on SVM is qualified according to claim 7, characterized in that: The optimal hyperparameters are C=0.25 and gamma=0.

1.

9. The method for determining whether the composition of the head of a continuously cast mixed billet based on SVM is qualified according to claim 6, characterized in that: Accuracy rate ≥ 94.9%.

Images