Regular Article
Analysis and Prediction of Sticker Breakout Based on XGBoost Forward Iterative Model
2024 Volume 64 Issue 8 Pages 1272-1278
Details
2024 Volume 64 Issue 8 Pages 1272-1278
All 61 sticker breakouts and 183 false sticker breakouts were obtained based on the on-line mould monitoring system during the conventional slab continuous casting. The 16-dimensional temperature characteristics and temperature velocity characteristics of the sticker breakout were extracted. The sticker breakout recognition based on the XGBoost forward iterative model was developed and optimized by the mean square error algorithm. The results show that the prediction probability of the sticker breakout after optimization is in the range of 0.72–1.00. The smallest output value is 0.5 higher than that before optimization. When the threshold is set to 0.65, the optimized XGBoost model can correctly predict all sticker breakouts and has a 99.5% accuracy rate. The XGBoost model has a stronger generalization ability and higher prediction accuracy, which promotes the intelligent production of continuous casting.
The technology of continuous casting simplifies the production process of the slab and makes the caster’s capacity release. In recent years, the casting speed has been increased with the development of “high-efficiency continuous casting”, which requires more stable heat transfer and lubrication in order to avoid some accidents in mould.1,2,3) As a serious safety accident, sticker breakout not only reduces production efficiency, but also causes some potential safety hazards.4,5)
Sticker breakout usually forms near the meniscus, which is influenced by lubrication conditions,6,7) molten steel composition8,9) and mould powder.10) On the other hand, external factors, such as manual operation and the change of casting speed, are also important triggers.11,12,13) In recent years, the thermocouple temperature measurement method has been the main method to predict sticker breakout.14,15,16) Meanwhile, the computer vision method was also used to detect the hot and cold regions of sticker breakout and predict breakout.17) Some logical judgment and intelligent models were also established to predict breakout.18,19,20,21,22,23) Although the intelligent models have strong adaptability and can deal with nonlinear problems of complex temperature characteristics, some false alarms occurred with large temperature fluctuations due to the difference of processing ability between different models. In order to reduce false alarms, some ensemble models were proposed, which combined the advantages of different models.24,25) However, the ensemble model has not been investigated in the prediction of sticker breakout. Therefore, the research of ensemble model for sticker breakout is of great significance to improve the prediction accuracy.
In this paper, a total of 61 samples of true sticker breakouts and 183 samples of false sticker breakouts were collected, and their temperature characteristics and temperature velocities characteristics were extracted. The XGBoost forward iterative model was developed to predict sticker breakout. The mean square error algorithm was used to optimize the model. Finally, a high accuracy and low false alarm rate was obtained.
The experiment was based on the arc continuous caster. The arc radius was 5 m. The mould length was 1.0 m. The width and thickness of conventional slabs were 0.9–2.0 and 0.18 m. The casting speed was in the range of 0.3–1.9 m·min−1. The main parameters of caster are included in Table 1.
| Item | Parameter |
|---|---|
| Type | Arc continuous caster |
| Strand | 2 |
| Slab thickness | 0.18 m |
| Slab width | 0.9–2.0 m |
| Drive | Hydraulic drive |
| SEN | Bilateral pore |
| Mould length | 1.0 m |
| Caster arc radius | 5.0 m |
| Caster length | 22.4 m |
| Casting speed | 0.3–1.9 m·min−1 |
In order to capture the characteristics of sticker breakout, the thermocouple temperature method was adopted. Figure 1 shows the schematic diagram of thermocouple arrangement. Three rows of thermocouples were arranged in mould copper plates. The distances from the top of the mould were 0.21, 0.33 and 0.45 m, respectively. The wide face had 7 columns of thermocouples, and the narrow face had one column of thermocouples. There were 48 thermocouples in total. The spacing between the adjacent thermocouples was 0.3 m except the first and second column of thermocouples, which was 0.15 m.
Figure 2 shows the temperature variation trend during sticker breakout. In Fig. 2(a), the temperature fluctuation of the upper and lower thermocouples is small before the appearance of sticker breakout. The shell at the lower thermocouple is thicker than that at the upper thermocouple, which causes the temperature of the first thermocouple to be usually higher than that of the second thermocouple. In Fig. 2(b), when the deterioration of lubrication conditions is poor, the shell and mould will contact directly under function of the static pressure, and a sticker breakout occurs. When the sticker point passes through the first row of thermocouples, the breakout of molten steel causes the temperature rise in the local area, which makes the thermocouple temperature rise rapidly. In Fig. 2(c), with the mould’s vibration and the slab’s downward movement, the sticker point moves down. The temperature of the second row of thermocouples increases, while the temperature of the first row of thermocouples decreases, which is usually called “temperature lag”. Meanwhile, the temperature of the first row of thermocouples is usually lower than that of the second row of thermocouples, which is called “temperature inversion”. In Fig. 2(d), the sticker point passes through the first and second rows of thermocouples, and the temperature of the first and second rows of thermocouples decreases. Until the sticker point out of the mould, the molten steel flows out and causes accident. The accurate capture of thermocouple temperature is very important to predict the sticker breakout.
According to temperature fluctuation during the process of sticker breakout occurrence, the temperature rise of the corresponding thermocouples is a key characteristic for sticker breakout. Therefore, the temperature rise amplitude is calculated by Eq. (1) to express the rising characteristics of temperature.
| (1) |
where, Trise is the temperature rise amplitude, °C; Tmax is the maximum temperature for 30 seconds, °C; Tave is the average temperature for 30 seconds, °C.
Figure 3 is the temperature rise amplitude of sticker breakout and stable casting. In Fig. 3(a), the sticker point passes through the first row of thermocouples at the 15th s. The temperature rise amplitude increases from 0.04°C to 6.02°C, which shows an obvious temperature rise. In Fig. 3(b), when there is no sticker point at the stable casting, the temperature of the first row of thermocouples only rises from −0.13°C to 1.02°C. It can be seen that the temperature rise is very small and even negative.
When the sticker breakout occurs in mould, it also has a large fluctuation of thermocouple temperature. Therefore, the standard deviations of the first and second row thermocouple temperatures are extracted to characterize the temperature fluctuation of sticker breakout. The standard deviation can be calculated by Eq. (2).
| (2) |
where, DT is the standard deviation of temperature for 10 seconds, °C; Ti is the temperature at present, °C; Tav is the average temperature for 30 seconds, °C.
Figure 4 shows the standard deviation and time series temperature curve of sticker breakout. The standard deviation at the stable casting is lower than that of sticker breakout, which fluctuates around 0.15°C. With the occurrence of sticker breakout, the standard deviation of thermocouple temperature rises. At the 10th s and 15th s, the standard deviation of the first row and the second row thermocouples increased to 4.41 and 4.39, respectively.
The sticker breakout is a dynamic evolution process, and the temperature usually fluctuates with an “up-down” trend. However, the different casting processes affect the temperature fluctuation interval in mould, which is not conducive to analysis. The temperature velocity has a quick reaction and the same interval, which is conducive to model training and testing. Figure 5 shows the time series temperature velocities during the process of sticker breakout. The maximum temperature velocity of the first and second rows of thermocouples was set as the starting point. Two temperature velocities were extracted forward and backward with 2 s as the time interval. A total of 5 temperature velocities were extracted for each row of thermocouples.
In order to obtain the relationship between the temperature velocity of the first and second rows of thermocouples, the temperature velocity difference is calculated, which can be calculated by Eq. (3).
| (3) |
where, Vminus is the temperature velocity difference, °C·s−1; V1 is the temperature velocity of the first row, °C·s−1; V2 is the temperature velocity of the second row, °C·s−1.
Figure 6 shows the time series curve of the temperature velocity difference. The temperature velocity difference also has an “up-down” trend. At the 12th s, the sticker point passes through the first row of thermocouples, and the temperature velocity difference increases to 1.02°C·s−1. At this time, the temperature velocity of the first row of thermocouples is higher than that of the second row of thermocouples. At the 16th s, the temperature velocity inversion occurs, and the temperature velocity difference decreases to 0°C·s−1. The minimum value of temperature velocity difference is −1.34°C·s−1 occurred at the 21st s. The temperature velocity difference reflects the relative change of the temperature velocity between the first and second row of thermocouples and eliminates the abnormal situation of the temperature velocity rising or falling at the same time. Therefore, the maximum value and extreme difference of the temperature velocity difference are extracted as the temperature velocity characteristics of the sticker breakout as shown in figure. The extreme difference can be calculated by Eq. (4).
| (4) |
where, ED is the extreme difference of temperature velocity difference, °C·s−1; Vdmax is the maximum value of temperature velocity difference, °C·s−1; Vdmin is the minimum value of temperature velocity difference, °C·s−1.
In summary, a total of 16-dimensional sticker breakout characteristics are extracted based on the above research. As included in Table 2, the temperature characteristics have 4 characteristics, such as temp_max_ave_1, temp_max_ave_2, temp_dt_1 and temp_dt_2. The temperature velocities include 12 characteristics, such as 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 and v_minus_max-min.
| Item | Name | Meaning |
|---|---|---|
| Temperature characteristic | temp_max_ave_1 | The temperature rise amplitude of the first row |
| temp_max_ave_2 | The temperature rise amplitude of the second row | |
| temp_dt_1 | The temperature standard deviation of the first row | |
| temp_dt_2 | The temperature standard deviation of the second row | |
| Temperature velocity | v(1)i−4 | The velocity before its maximum value 4 s of the first row |
| v(1)i−2 | The velocity before its maximum value 2 s of the first row | |
| v(1)i | The maximum velocity of the first row | |
| v(1)i+2 | The velocity after its maximum value 2 s of the first row | |
| v(1)i+4 | The velocity after its maximum value 4 s of the first row | |
| v(2)i−4 | The velocity before its maximum value 4 s of the second row | |
| v(2)i−2 | The velocity before its maximum value 2 s of the second row | |
| v(2)i | The maximum velocity of the second row | |
| v(2)i+2 | The velocity after its maximum value 2 s of the second row | |
| v(2)i+4 | The velocity after its maximum value 4 s of the second row | |
| v_minus_max | The maximum value of temperature velocity difference | |
| v_minus_max-min | The extreme difference of temperature velocity difference |
The method of extreme gradient boosting (XGBoost) is an integrated learning model widely used in medical, financial and other fields.26,27,28) XGBoost has higher predictive performance than common classifier models.29,30) It trains the weak model into a strong model by residual fitting. The optimal solution of the objective function ensures the prediction accuracy of the XGBoost model and reduces the influence of overfitting. The objective function can be calculated by Eq. (5).
| (5) |
where, yi is the real value;
Figure 7 shows the flow chart of the XGBoost forward iterative model. A total of 61 true sticker breakouts were collected as positive samples. Due to the complexity of continuous casting process, false sticker breakouts were selected from a total of five process conditions as negative samples, such as SEM change process, casting start-up, temperature with large fluctuation, temperature with small fluctuation and normal casting condition. The five process conditions included the most conditions during continuous casting. It was designed to simulate the actual production processes as best as possible. The 16-dimensional characteristics were also extracted from temperature characteristics and temperature velocity. And the 70% samples are randomly selected as training samples, and the others are used as testing samples. So the data of training and testing samples will include the true sticker breakouts and false sticker breakouts together. And all data has been used for training and testing. The CART tree is established for obtaining residuals, and the next tree fits the residuals. The strong model is built by iteration, which is decided by n CART trees. The final output is the prediction probability.
Figure 8 shows the prediction results of XGBoost model. The prediction result of the sticker breakout is in the range of 0.22–1.00. There is a lower output value (0.22) for sticker breakout, which is lower than 0.5. It is mainly due to the unstable temperature fluctuation caused by manual operation according to the record from steel plant. When the threshold is 0.5, there are two false sticker breakouts judged to be true sticker breakouts by this original XGBoost model. The whole prediction accuracy is 98.7%.
The mean square error can measure the relationship between the true value and the predicted value of the sample, which is used to reflect the stability and accuracy of the model performance. The mean square error can be calculated by Eq. (6).
| (6) |
where, n is the number of samples; Yi is the true value of the sample;
Figure 9 shows the mean square error curve of the training set and the testing set. It can be seen that the mean square error will decrease with the iteration number increase. The mean square error of the testing set is higher than that of the training set. In Fig. 9(a), the mean square error of the training set before optimization reaches stable at the 200th iteration (0.02). However, the mean square error of the testing set reaches the lowest point (0.22) at the 20th iteration, and then the mean square error begins to increase with the iteration number. The mean square error reaches 0.28 at the 200th iteration, which is higher than the training set of 0.26. From the mean square error curve, it can be seen that the XGBoost model has an over-fitting phenomenon before optimization. In Fig. 9(b), the mean square error of the training set and the testing set at the 200th iteration is 0.04 and 0.24, respectively. The difference is 0.20, which has 0.06 lower than that before optimization.
The hyperparameters optimized by the mean square error curve were set into the XGBoost model. The small learning rate is helpful to increase the number of weak learners which is used to fit residuals. So it is set to 0.05. The reg_lambda is helpful to reduce the complexity of the model which is used to avoid overfitting. So it is raised to 3. The reg_alpha is helpful to adapt to complex data composition which is used to handle outliers. So it is set to 0.5. The optimized hyperparameters and their meanings are included in Table 3.
| Item | Meaning | Parameter |
|---|---|---|
| n_estimators | Number of weak classifiers | 200 |
| subsample | Sampling ratio of sample | 1 |
| scale_pos_weight | Ratio of positive and negative samples | 2 |
| max_depth | Maximum depth of weak classifiers | 6 |
| learning_rate | Weight of residual fitting | 0.05 |
| reg_lambda | Weight of L2 regular term | 3 |
| reg_alpha | Weight of L1 regular term | 0.5 |
The threshold is an important parameter to predicate the sticker breakout. A high threshold can reduce the risk of false alarms, but it can also lead to some missed alarms. Therefore, in order to ensure that the model has a better generalization performance, the threshold should be increased on the basis that all sticker breakouts can be predicted truly. Figure 10 shows the recall rate of XGBoost in different thresholds. When the threshold is 0.5 or 0.6, the recall rate of the XGBoost model before and after optimization is 98.3% and 100%, respectively. When the threshold is 0.7, the recall rate of the original XGBoost model is 96.7%, while the optimized model is still 100%. It can be seen that this threshold is higher than the initial threshold 0.5, and the generalization performance of the model is improved.
Figure 11 shows the prediction results of the optimized XGBoost model. The prediction result of the sticker breakout is in the range of 0.72–1.00. The smallest output value is changed from 0.22 to 0.72, which has an increase of 0.5 than that before optimization. When the threshold is set to 0.65, the XGBoost model ensures that all sticker breakouts are true predicted. It can be seen that the optimized XGBoost model is better than the initial one.
Table 4 includes the comparison between XGBoost and other models for breakout prediction. It can be seen that only the recall rate of XGBoost model is 100%, and the recall rates of the other four models are 96.7%, 98.3%, 90.1% and 72.1%, respectively. The XGBoost model with optimal hyperparameters has the highest accuracy rate (99.5%). It not only predicts all true sticker breakouts correctly, but also has very few false predictions. Therefore, it has a better applicability and generalization ability.
| Models | Recall rate/% | Accuracy rate/% |
|---|---|---|
| XGBoost | 100 | 99.5 |
| Decision Tree | 96.7 | 98.3 |
| Support Vector Classification | 98.3 | 92.6 |
| K-Nearest Neighbor | 90.1 | 94.2 |
| Logistic Regression | 72.1 | 89.3 |
(1) A total of 16-dimensional sticker breakout characteristics were extracted, including temperature rise amplitude, standard deviation, temperature velocities, temperature velocity difference. Compared with steady casting, there is an obvious temperature rise trend between the upper and lower rows of thermocouples of sticker breakout, accompanied by “temperature lag” and “temperature inversion” phenomena, which proves the feasibility of characteristics selection.
(2) A prediction model of sticker breakout based on XGBoost forward iterative model was developed. The mean square error algorithm was used to optimize this model. The optimized mean square error of the training set and the testing set at the 200th iteration are 0.04 and 0.24, respectively. The difference is 0.20, which has 0.06 lower than that before optimization. The threshold optimized by the mean square error algorithm is increased from 0.5 to 0.65, which brings a stronger generalization ability and higher prediction accuracy.
(3) A total of 61 samples of true sticker breakouts and 183 samples of false sticker breakouts were used for training and testing. The prediction probability of the sticker breakout is in the range of 0.72–1.00. When the threshold is set to 0.65, the XGBoost model can correctly predict all sticker breakouts and has a higher accuracy rate 99.5%. It is better than decision tree, support vector classification, k-nearest neighbor and logistic regression models. The XGBoost forward iterative model not only improves the accuracy rate of sticker breakout prediction, but also promotes the intelligent production of continuous casting.
All authors contributed to the research concept and design, the data was provided by Xudong Wang and Man Yao, the data analysis were performed by Yali Gao and Miao Yu and Zhiqiang Xu, the XGBoost forward iterative model was developed by Zhixin Ma, while the first draft of the manuscript and constructive discussions were written by Yu Liu and Zhixin Ma.
“This work was supported by the Science and Technology Development of Jilin Province (20230101335JC).”
The authors declare no potential conflict of interest relevant to this article’s content.
The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.
