A prediction method for bonding breakout based on stacking multi-classifier fusion

Abstract

The present invention discloses a method for predicting steel breakout based on stacking multi-classifier fusion, belonging to the field of iron and steel metallurgy. The method comprises the following steps: constructing a sample library of true and false steel breakouts; extracting temperature characteristic data and time-series temperature rates of true and false steel breakouts; constructing feature vectors of true and false steel breakout samples; preprocessing the true and false steel breakout feature data; constructing a steel breakout prediction model based on stacking multi-classifier fusion using random forest, K-nearest neighbor classification, and support vector classification as primary classifiers, and logistic regression as a secondary classifier; and finally determining an optimal threshold point through a receiver operating characteristic curve. If the threshold is greater than the optimal threshold, the prediction is judged as steel breakout. The present invention combines the steel breakout temperature characteristic vector with a multi-classifier fusion recognition method to establish a strong classification and recognition model for steel breakouts. While ensuring a 100% steel breakout reporting rate, the method reduces the false alarm rate and improves the prediction accuracy.

Description

Technical Field

[0001] The invention belongs to the technical field of continuous casting in iron and steel metallurgy, and is a method for predicting bonding breakout based on Stacking multi-classifier fusion. Background Art

[0002] Steel leakage from bonding is a serious safety accident in continuous casting production. It not only greatly disrupts the normal production order, but also causes serious damage to the continuous casting equipment, resulting in huge economic losses to steel companies.

[0003] Bonding breakouts typically form near the meniscus, resulting from tearing of the shell due to poor lubrication and excessive friction. Currently, the mainstream bonding breakout prediction systems are based on logical judgment and artificial intelligence. Crystallizer breakout prediction is primarily based on thermocouple temperature measurement, with models primarily consisting of a logical judgment model and a machine learning-based bonding breakout prediction model. By identifying the crystallizer thermocouple temperature curve, the likelihood of bonding breakouts can be determined and predicted, thereby reducing and avoiding bonding breakout accidents. However, due to inherent defects and problems inherent in the single classifier model, overfitting is prone to occur when predicting and diagnosing bonding breakouts, affecting the accuracy of the prediction and leading to omissions and false alarms. Therefore, how to eradicate omissions and avoid false alarms is a core issue currently facing bonding breakout crystallizer process monitoring. Prediction methods for crystallizer breakouts need further improvement and development.

[0004] Patent document CN107096899 discloses a crystallizer leak prediction system based on logical judgment. Based on the theory of bonding leak, the patent smoothes and filters the temperature data, adopts a logical judgment method, and uses the temperature rise of a single thermocouple and the temperature change of surrounding thermocouples to check for light and heavy alarms for bonding leaks. It overcomes the shortcomings of the existing bonding leak prediction methods and devices in terms of prediction accuracy and alarm timeliness, and provides a crystallizer leak prediction system based on logical judgment. It achieves timely and accurate prediction of bonding leaks on the basis of ensuring no missed reports, ensuring smooth continuous casting production and improving the quality of cast billets. However, due to the diversity of the occurrence of bonding leaks, the logical judgment model must be debugged based on on-site equipment and process parameters. Therefore, changes in actual production conditions will cause the accuracy of this method to decrease.

[0005] Patent document CN108705058 discloses a method for predicting crystallizer leakage based on K-Means clustering. The patent uses thermocouples embedded in the crystallizer to measure and extract temperature, constructs a historical temperature data sample library, and uses the K-Means clustering method to divide it into two sample clusters. Finally, the judgment threshold is obtained by calculating the distance between the sample and the sample center in the cluster to determine whether it is bonding leakage. This method overcomes the shortcomings of existing bonding leakage prediction methods, has high stability, and can predict bonding leakage in real time and quickly. However, under stable casting conditions, large temperature fluctuations often occur. The prediction accuracy of rare category samples of this method is low, which is prone to false alarms. Summary of the Invention

[0006] The purpose of the present invention is to overcome the shortcomings of the existing bonding steel leakage prediction method and propose a bonding steel leakage prediction method based on Stacking multi-classifier fusion. The method combines the time-series temperature characteristics of bonding steel leakage with the multi-classifier fusion recognition method to establish a strong classification and recognition model for bonding steel leakage. While ensuring that the bonding steel leakage reporting rate is 100%, the false alarm rate is reduced, bonding steel leakage is accurately predicted, and the necessary conditions are provided for online detection of bonding steel leakage.

[0007] To achieve the above object, the present invention adopts a technical solution: a method for predicting bonding breakout based on Stacking multi-classifier fusion, which mainly includes the following steps:

[0008] 1) Construct a sample library of genuine and fake bonded steel leaks

[0009] ①Use the crystallizer online monitoring system to obtain data on steel type, pouring temperature, casting speed, liquid level and crystallizer temperature;

[0010] ②According to the pouring records and mold temperature data, we obtain samples of bonding breakout. At the same time, we extract the mold temperature data under normal pouring, pouring start, and water change processes as false bonding breakout samples to build a library of true and false bonding breakout samples.

[0011] 2) Extracting temperature characteristic data and time series temperature rate of true and false bonding leakage

[0012] ① Obtain the first row temperature standard deviation DT of the thermocouples in the same row. (1) ;

[0013] ② Obtain the second row temperature standard deviation DT of the thermocouples in the same row. (2) ;

[0014] ③ Obtain the difference between the maximum and average values ​​of the first row temperature of the thermocouples in the same row. (1)max-ave ;

[0015] ④ Obtain the difference T between the maximum and average values ​​of the second row temperature of the thermocouples in the same row. (2)max-ave ;

[0016] ⑤ According to formula (1), calculate the temperature rate V of the true and false bonding leakage steel, and obtain the 9s temperature rate V including the temperature rate rise-fall process of the first row of thermocouples (1)i-4 , V (1)i-2 , V (1)i , V (1)i+2 , V (1)i+4 , where V (1)i The temperature rate when the temperature rate of the first row of thermocouples reaches the maximum value;

[0017]

[0018] Where V represents the temperature rate, °C / s; T i represents the temperature of the thermocouple at the i-th time point, °C;

[0019] ⑥According to formula (1), obtain the 9s temperature rate V including the temperature rise-fall process of the second row of thermocouples (2)i-4 , V (2)i-2 , V (2)i , V (2)i+2 , V (2)i+4 , where V (2)i The temperature rate when the temperature rate of the second row of thermocouples reaches the maximum value;

[0020] ⑦ According to formula (2), calculate the temperature rate difference V between true and false bonding leakage minus ;

[0021] V minus =V (1) -V (2) (2)

[0022] Where, V minus Represents the temperature rate difference, ℃ / s; V (1) Represents the temperature rate of the first row of thermocouples, ℃ / s; V (2) represents the temperature rate of the second row of thermocouples, °C / s;

[0023] ⑧ Get the maximum value V of the temperature rate difference minus(max) With the minimum value V minus(min) ;

[0024] ⑨ Obtain the difference between the maximum and minimum temperature rate difference V minus(max-min) ;

[0025] 3) Constructing the feature vector of true and false bonding and leakage steel samples

[0026] ①Construct the characteristic vector of bonding leakage sample (DT (1) , DT (2) , T (1)max-ave , T (2)max-ave , V (1)i-4 , V (1)i-2 , V (1)i , V (1)i+2 , V (1)i+4 , V (2)i-4 , V (2)i-2 , V (2)i , V (2)i+2 , V (2)i+4 , V minus(max) , V minus(max-min) );

[0027] ②Construct the characteristic vector of pseudo-bonding steel leakage sample (DT (1) , DT (2) , T (1)max-ave , T (2)max-ave , V (1)i-4 , V (1)i-2 , V (1)i , V (1)i+2 , V (1)i+4 , V (2)i-4 , V (2)i-2 , V (2)i , V (2)i+2 , V (2)i+4 , V minus(max) , V minus(max-min) );

[0028] 4) Preprocessing of characteristic data of true and false bonding and steel leakage

[0029] According to formula (3), the feature data is normalized, and the label of the bonding steel leakage sample is set to 1, and the label of the pseudo-bonding steel leakage sample is set to 0;

[0030]

[0031] Where X is the normalized feature data; min is the minimum value of the feature data; max is the maximum value of the feature data;

[0032] 5) Constructing a stacking multi-classifier fusion bonding leak prediction model

[0033] ① Divide the processed true and false bonding leakage sample feature vectors into training set and test set in a ratio of 7:3

[0034] ② Set random forest, K-nearest neighbor classification, and support vector classification as primary classifiers, pass the training set to the primary classifier, use 5-fold cross-validation for training, and use the output results as the new training set. At the same time, the test set follows the primary classifier for prediction, and the output results are averaged and used as the new test set.

[0035] ③ Pass the new training set and the new test set to the logistic regression as the secondary classifier training output;

[0036] ④ In support vector classification, the penalty coefficient C is set to 3 and the kernel function coefficient g is set to 0.3;

[0037] ⑤ The maximum depth of the tree in the random forest is set to 1, the optimal attribute feature is set to 1, the number of decision trees is set to 39, and the random number seed is set to 123.

[0038] ⑥ In K-nearest neighbor classification, the number of neighbors K is set to 2;

[0039] ⑦ The regularization strength C in logistic regression is set to 2;

[0040] ⑧ Set the output of the constructed Stacking model as probability, and use the receiver operating characteristic curve (ROC curve) to find the optimal threshold point T. If the Stacking output is greater than T, it is judged as bonding leakage; if the output is less than T, it is judged as pseudo-bonding leakage.

[0041] Among them, T is set to 0.55.

[0042] The beneficial effects of the present invention's method for predicting bonding breakout based on Stacking multi-classifier fusion are as follows: the temperature change characteristics of the first and second rows of thermocouples in the bonding breakout occurrence area and their time-series temperature characteristics are respectively extracted, the temperature change characteristics and time-series temperature characteristics of bonding breakout are combined with multi-classifier fusion recognition, the sample library is trained and tested using the Stacking multi-classifier fusion model, the optimal parameters are found by the grid search method, and the optimal Stacking classification model finally obtained has a higher bonding breakout accuracy rate than the single classifier model, reduces the false alarm rate, and is of great significance to improving the quality of the casting. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of the bonding breakout prediction method based on Stacking multi-classifier fusion;

[0044] Figure 2 This is a flow chart of the training of the bonding breakout prediction model based on Stacking multi-classifier fusion;

[0045] Figure 3 is the receiver operating characteristic curve;

[0046] Figure 4 This is the bonding breakout prediction result diagram based on Stacking multi-classifier fusion;

[0047] Figure 5 This is a comparison chart of the accuracy of single classifier and Stacking multi-classifier; DETAILED DESCRIPTION

[0048] The present invention is further described below with reference to the following examples, but the present invention is not limited thereto.

[0049] like Figure 1 As shown, a method for predicting bonding breakout based on Stacking multi-classifier fusion includes the following steps:

[0050] Step 1: Build a sample library of genuine and fake bonded steel leaks

[0051] ①Use the crystallizer online monitoring system to obtain data on steel type, pouring temperature, casting speed, liquid level and crystallizer temperature;

[0052] ②According to the pouring records and mold temperature data, we obtain samples of bonding breakout. At the same time, we extract the mold temperature data under normal pouring, pouring start, and water change processes as false bonding breakout samples to build a library of true and false bonding breakout samples.

[0053] Step 2: Extract the temperature characteristic data and time series temperature rate of true and false bonding leakage

[0054] ① Obtain the first row temperature standard deviation DT of the thermocouples in the same row. (1) ;

[0055] ② Obtain the second row temperature standard deviation DT of the thermocouples in the same row. (2) ;

[0056] ③ Obtain the difference between the maximum and average values ​​of the first row temperature of the thermocouples in the same row. (1)max-ave ;

[0057] ④ Obtain the difference T between the maximum and average values ​​of the second row temperature of the thermocouples in the same row. (2)max-ave ;

[0058] ⑤ According to formula (1), calculate the temperature rate V of the true and false bonding leakage steel, and obtain the 9s temperature rate V including the temperature rate rise-fall process of the first row of thermocouples (1)i-4 , V (1)i-2 , V (1)i , V (1)i+2 , V (1)i+4 , where V (1)i The temperature rate when the temperature rate of the first row of thermocouples reaches the maximum value;

[0059]

[0060] Where V represents the temperature rate, °C / s; T i represents the temperature of the thermocouple at the i-th time point, °C;

[0061] ⑥According to formula (1), obtain the 9s temperature rate V including the temperature rise-fall process of the second row of thermocouples (2)i-4 , V (2)i-2 , V (2)i , V (2)i+2 , V (2)i+4 , where V (2)i The temperature rate when the temperature rate of the second row of thermocouples reaches the maximum value;

[0062] ⑦ According to formula (2), calculate the temperature rate difference V between true and false bonding leakage minus ;

[0063] V minus =V (1) -V (2) (2)

[0064] Where, V minus Represents the temperature rate difference, ℃ / s; V (1) Represents the temperature rate of the first row of thermocouples, ℃ / s; V (2) represents the temperature rate of the second row of thermocouples, °C / s;

[0065] ⑧ Get the maximum value V of the temperature rate difference minus(max) With the minimum value V minus(min) ;

[0066] ⑨ Obtain the difference between the maximum and minimum temperature rate difference V minus(max-min) ;

[0067] Step 3: Construct the feature vector of the true and false bonding leakage samples

[0068] ①Construct the characteristic vector of bonding leakage sample (DT (1) , DT (2) , T (1)max-ave , T (2)max-ave , V (1)i-4 , V (1)i-2 , V (1)i , V (1)i+2 , V (1)i+4 , V (2)i-4 , V (2)i-2 , V (2)i , V (2)i+2 , V (2)i+4 , V minus(max) , V minus(max-min) );

[0069] ②Construct the characteristic vector of pseudo-bonding steel leakage sample (DT (1), DT (2) , T (1)max-ave , T (2)max-ave , V (1)i-4 , V (1)i-2 , V (1)i , V (1)i+2 , V (1)i+4 , V (2)i-4 , V (2)i-2 , V (2)i , V (2)i+2 , V (2)i+4 , V minus(max) , V minus(max-min) );

[0070] Step 4: Preprocessing of the characteristic data of true and false bonding and steel leakage

[0071] The sample data is normalized according to formula (3), and the label of the bonding steel leakage sample is set to 1, and the label of the pseudo-bonding steel leakage sample is set to 0;

[0072]

[0073] Where X is the normalized feature data; min is the minimum value of the feature data; max is the maximum value of the feature data;

[0074] Step 5: Construct a stacking multi-classifier fusion bonding leak prediction model

[0075] ① Divide the processed true and false bonding leakage sample feature vectors into training set and test set in a ratio of 7:3

[0076] ② Set random forest, K-nearest neighbor classification, and support vector classification as primary classifiers, pass the training set to the primary classifier, use 5-fold cross-validation for training, and use the output results as the new training set. At the same time, the test set follows the primary classifier for prediction, and the output results are averaged and used as the new test set.

[0077] ③ Pass the new training set and the new test set to the logistic regression as the secondary classifier training output;

[0078] ④ In support vector classification, the penalty coefficient C is set to 3 and the kernel function coefficient g is set to 0.3;

[0079] ⑤ The maximum depth of the tree in the random forest is set to 1, the optimal attribute feature is set to 1, the number of decision trees is set to 39, and the random number seed is set to 123.

[0080] ⑥ In K-nearest neighbor classification, the number of neighbors K is set to 2;

[0081] ⑦ The regularization strength C in logistic regression is set to 2;

[0082] ⑥ Set the output of the constructed Stacking model as probability, and use the receiver operating characteristic curve (ROC curve) to find the optimal threshold point T as 0.55, such as Figure 3 As shown in the figure, if T is greater than 0.55, it is judged as bonding leakage; if T is less than 0.55, it is judged as pseudo bonding leakage. The optimal Stacking classification model is set at this threshold.

[0083] Finally, based on the casting data of a domestic steel plant, 57 cases of true and false bonding leakage were predicted. The results are as follows: Figure 4 The accuracy of the bond breakout prediction model based on Stacking multi-classifier fusion is 98.2%. Figure 5 This is a comparison chart of the accuracy of a single classifier and a stacking multi-classifier. Under the same conditions, the method proposed in this patent has a higher accuracy than the single classifier, indicating that this method can more accurately and efficiently identify bonded steel leaks.

[0084] The above-described embodiments merely express the implementation methods of the present invention, but they cannot be understood as a prediction of the scope of the patent of the present invention. It should be pointed out that for those skilled in the art, several variations and improvements can be made without departing from the concept of the present invention, which all fall within the scope of protection of the present invention.

Claims

1. A method for predicting bonding breakout based on Stacking multi-classifier fusion, characterized in that: The following steps are involved: 1) Construct a sample library of genuine and fake bonded steel leaks ①Use the crystallizer online monitoring system to obtain data on steel type, pouring temperature, casting speed, liquid level and crystallizer temperature; ②According to the pouring records and mold temperature data, we obtain samples of bonding breakout. At the same time, we extract the mold temperature data under normal pouring, pouring start, and water change processes as false bonding breakout samples to build a library of true and false bonding breakout samples. 2) Extracting temperature characteristic data and time series temperature rate of true and false bonding leakage ① Obtain the first row temperature standard deviation DT of the thermocouples in the same row. (1) ; ② Obtain the second row temperature standard deviation DT of the thermocouples in the same row. (2) ; ③ Obtain the difference between the maximum and average values ​​of the first row temperature of the thermocouples in the same row. (1)max-ave ; ④ Obtain the difference T between the maximum and average values ​​of the second row temperature of the thermocouples in the same row. (2)max-ave ; ⑤ According to formula (1), calculate the temperature rate V of the true and false bonding leakage steel, and obtain the 9s temperature rate V including the temperature rate rise-fall process of the first row of thermocouples (1)i-4 , V (1)i-2 , V (1)i , V (1)i+2 , V (1)i+4 , where V (1)i The temperature rate when the temperature rate of the first row of thermocouples reaches the maximum value; Where V represents the temperature rate, °C / s; T i represents the temperature of the thermocouple at the i-th time point, °C; ⑥According to formula (1), obtain the 9s temperature rate V including the temperature rise-fall process of the second row of thermocouples (2)i-4 , V (2)i-2 , V (2)i , V (2)i+2 , V (2)i+4 , where V (2)i The temperature rate when the temperature rate of the second row of thermocouples reaches the maximum value; ⑦ According to formula (2), calculate the temperature rate difference V between true and false bonding leakage minus ; V minus =V (1) -V (2) (2) Where, V minus Represents the temperature rate difference, ℃ / s; V (1) Represents the temperature rate of the first row of thermocouples, ℃ / s; V (2) represents the temperature rate of the second row of thermocouples, °C / s; ⑧Get the maximum value V of the temperature rate difference minus(max) With the minimum value V minus(min) ; ⑨ Obtain the difference between the maximum and minimum temperature rate difference V minus(max-min) ; 3) Constructing the feature vector of true and false bonding and leakage steel samples ①Construct the characteristic vector of bonding leakage sample (DT (1) , DT (2) , T (1)max-ave , T (2)max-ave , V (1)i-4 , V (1)i-2 , V (1)i , V (1)i+2 , V (1)i+4 , V (2)i-4 , V (2)i-2 , V (2)i , V (2)i+2 , V (2)i+4 , V minus(max) , V minus(max-min) ); ②Construct the characteristic vector of pseudo-bonding steel leakage sample (DT (1) , DT (2) , T (1)max-ave , T (2)max-ave , V (1)i-4 , V (1)i-2 , V (1)i , V (1)i+2 , V (1)i+4 , V (2)i-4 , V (2)i-2 , V (2)i , V (2)i+2 , V (2)i+4 , V minus(max) , V minus(max-min) ); 4) Preprocessing of characteristic data of true and false bonding and steel leakage According to formula (3), the feature data is normalized, and the label of the bonding steel leakage sample is set to 1, and the label of the pseudo-bonding steel leakage sample is set to 0; Where X is the normalized feature data; min is the minimum value of the feature data; max is the maximum value of the feature data; 5) Constructing a stacking multi-classifier fusion bonding leak prediction model ① Divide the processed true and false bonding leakage sample feature vectors into training set and test set in a ratio of 7:3 ② Set random forest, K-nearest neighbor classification, and support vector classification as primary classifiers, pass the training set to the primary classifier, use 5-fold cross-validation for training, use the output results as the new training set, and at the same time, use the test set to predict with the primary classifier, and average the output results as the new test set; ③ Pass the new training set and the new test set to the logistic regression as the secondary classifier training output; ④ In support vector classification, the penalty coefficient C is set to 3 and the kernel function coefficient g is set to 0.3; ⑤ The maximum depth of the tree in the random forest is set to 1, the optimal attribute feature is set to 1, the number of decision trees is set to 39, and the random number seed is set to 123. ⑥ In K-nearest neighbor classification, the number of neighbors K is set to 2; ⑦ The regularization strength C in logistic regression is set to 2; ⑧ Set the output of the constructed Stacking model as probability, and use the receiver operating characteristic curve (ROC curve) to find the optimal threshold point T. If the Stacking output is greater than T, it is judged as bonding leakage; if the output is less than T, it is judged as pseudo-bonding leakage.

2. The method for predicting steel breakout based on stacking multi-classifier fusion according to claim 1 is characterized in that In step 2), the characteristics of the genuine and fake bonding leakage samples are composed of temperature characteristic data and time series temperature rate.

3. The method for predicting steel breakout based on stacking multi-classifier fusion according to claim 1 is characterized in that The parameters described in ② to ⑤ in step 5) are obtained by the grid search method.

4. The method for predicting steel breakout based on stacking multi-classifier fusion according to claim 1 is characterized in that In step 5), the receiver operating characteristic curve (ROC curve) is used to find the optimal threshold point T of 0.55.

Images