Smart manufacturing is the inevitable trend of technological development for joint-stock companies under the new situation, and it is also an important guarantee for winning the initiative and remaining invincible in the fierce market competition. The construction of 1580 hot-rolling smart workshop is Baosteel’s first demonstration smart project piloted by the Ministry of Industry and Information Technology. By introducing and upgrading existing technology and equipment, Baosteel has taken steps to upgrade the smart steel industry.
This article will introduce in detail the project practice of how Baosteel applies the intelligent algorithm based on JMP Pro to predict the natural width of 1580 hot rolling and finishing rolling. It is also a supplement to the speech delivered by Mr. Zhu Jianqin, Chief Engineer of Digital Modeling of Baosteel Co., Ltd. Equipment Department at the 2020 JMP Discovery Summit China. .
one Project background — why needed JMP Pro ?
In the Baosteel 1580 hot rolling smart workshop construction project, the finishing rolling dynamic width control (DWC) system was developed by Primetals Technologies Germany GmbH (PT) using its finishing rolling natural width prediction and self-adjusting neural network model. Embedded in the L2 system, through the input of L2 process parameters, the predicted value of the natural width of the finishing rolling is calculated, which is used to determine the rough rolling target width and post-rolling self-adjustment, and is transmitted to L1 as a reference value, and L1 realizes the finishing rolling width. Dynamic control, thus significantly improving the control accuracy of the exit width of 1580 hot rolling finish rolling. The average deviation between the model prediction value and the actual value is only 1.38mm (January to September 2019), see Figure 1.
figure 1
However, the fine rolling natural width prediction and self-adjusting neural network model provided by Pratt in Germany is a black box model. Its original program code is encrypted and cannot be read or modified. There is no description of the model structure, so it cannot be carried out. Parameter adjustment, model tuning and other improvement work.
Therefore, the team hopes to develop an intelligent model by applying other machine learning, deep learning and other artificial intelligence modeling tools of the JMP statistical analysis product family to predict the natural width of finish rolling at Baosteel Co., Ltd. 1580 hot rolling mill. It replaces the black box model of the German PT company and makes the deviation between the predicted value and the actual value of the self-developed model reach the monthly average value of ≤1.7mm, thus basically reaching the control level of the German black box model.
Based on this, the team began to make further attempts and explorations with the help of predictive modeling tools in JMP Pro.
two Establishment of intelligent prediction models
1 Data collection and cleaning
The team collected 1,580 hot-rolled finishing rolling production data for a total of nine months from 2019.01 to 2019.09, and compiled a SAS program to clean and integrate the data from January to August 2019. A total of 85,423 pieces of valid data were extracted and summarized Form an Excel data file.
2 Data filtering
01 Use smart algorithms to screen key factors
The Bootstrap forest algorithm in the JMP Pro statistical analysis software was used to perform fitting training on this data set (January to August 2019, 85423 records, 206 influencing factors), and then based on the training results based on the actual natural width For the contribution (based on the regression sum of squares), influence factors with a variance sum of squares > 3000 were selected, a total of 31 factors, see Figure 2.
figure 2
02 Pratt ( PT ) Company experience inheritance
The input/output interface of the natural widening neural network model analysis and diagnosis system provided by Pratt (PT) of Germany. It is not difficult to see that the neural network black box model requires 33 parameters such as chemical composition, finishing rolling entrance width, and exit width. Input variables, see Figure 3.
image 3
03 Data synthesis
The team compared the key factors selected by JMP Pro's Bootstrap forest algorithm with the input variables required by Pratt (PT)'s neural network model. It was found that 4 influencing factors were the same, and they were pre-calculated and accurate. Rolling target width and chemical composition silicon (Si), niobium (Nb), etc.
At the same time, based on production quality analysis experience, it is not difficult to see that there is an autocorrelation between factors such as the actual rough rolling width, the actual second half width of rough rolling, and the rough rolling target width, and these variables are consistent with the target actual natural width that the team needs to fit. Spread (spread_ACT) also has a linear correlation. Applying JMP Pro's multivariate linear correlation analysis tool to analyze these eight influencing factors and the actual values of natural width, it can be seen that there is a strong correlation between the eight influencing factors (correlation coefficient r>0.99), and it is closely related to the natural width. The correlation of the actual value of the spread is also relatively strong (r>55%). See Figure 4.
Figure 4
In addition, there are more than 1,000 tapping marks (alloyCode). Although they have an impact on the natural width and contribute the most, they cannot be fitted because the dimensions are too large. A table method can be used for self-adjustment learning.
Therefore, after discussion, it was finally decided to delete these 9 variables, plus the repeated 4 variables, a total of 13 influencing factors were deleted. In this way, 18 of the 31 influencing factors remain. Adding the 33 factors inherited from Pratt's experience, 51 influencing factors are finally determined.
3 data preparation
For the final 51 influencing factors that affect the natural width of finishing rolling, the team collected data for a total of three months from June to August 2019 (a total of 30,533 strips), used JMP Pro to read the data, and The actual value of the natural width of finishing rolling is used as the target value (Y, response variable), 51 influencing factors are used as fitting data (X, quotation), and 80% of the fitting data set is randomly set as the training set, and 20% is used as the verification set. set.
three Smart model training and offline fitting
The team tried to apply two smart modeling methods in JMP Pro analysis software: neural network and Bootstrap forest method to train and verify the above data set respectively, and obtained the following results:
1 Apply neural network modeling method
Figure 5 Figure 6
As can be seen from Figures 5 and 6, the model obtained after training and fitting using the neural network algorithm has a coefficient of determination R2 of 0.858, an RMSE of 1.506, and an average of absolute deviations of 1.16 after verification using verification data. Theoretically, it has been achieved Goals set by the team.
Figure 6 shows the scatter diagram of the predicted values of the three models (from left to right, the JMP model, the PT model and Baosteel's existing model) and the actual values. It can be seen that the model fitted by JMP has the smallest degree of dispersion.
2 application Bootstrap Forest method modeling method
Figure 7 Figure 8
As can be seen from Figures 7 and 8, the model obtained after training and fitting using the Bootstrap forest method has a determination coefficient R2 of 0.863 and an RMSE of 1.48 after verification using the validation data, achieving higher prediction accuracy.
Similarly, Figure 8 shows the scatter diagram of the predicted values of the three models (from left to right, the JMP model, the PT model and Baosteel's existing model) and the actual values. It can be seen that the JMP fitted model has the smallest degree of dispersion. of. )
It can be seen from the application of the above two smart model algorithms that Bootstrap The forest algorithm has a better fitting effect and higher prediction accuracy than the neural network. Both modeling methods have achieved the goals of the project team.
However, this is only the result of theoretical analysis of the data. The actual application effect still needs to be determined by the actual control data on site. How to apply the models established by these intelligent algorithms to actual control is the truly meaningful work.
The model export function of JMP Pro software builds a bridge from theory to practice for the team, bridges the gap from offline analysis to online application, and makes the natural chasm clear.
Four Online verification of smart models
1 Export of model
JMP Pro software was used to convert these two intelligent model algorithms into C language codes, thus taking a critical step for the practical application of intelligent control models in the field.
2 Online verification of models
The C language code converted from the two smart model algorithms was compiled and integrated into the preset process of the 1580 hot rolling rough rolling process computer offline analog machine, and then tested using the data in September 2019, that is, input factors It is the actual strip steel data in September. The smart model is used to predict the natural width of the finishing rolling. Then the same standard deviation calculation formula of Pratt PT Company is used to calculate the standard deviation of the Baosteel model, PT Company model and JMP smart model respectively. See below Figure 9:
Figure 9
As can be seen from Figure 9, From the actual prediction results, JMP The prediction results of the model fitted by the forest method have reached the prediction results of Baosteel's existing model. !
five Project summary and experience
By using the intelligent analysis and modeling methods in JMP Pro statistical analysis software, the team trained, verified and fitted the predicted values, actual values and related parameters of the natural width of 1580 hot-rolled finish rolling, and obtained the neural network and The theoretical prediction accuracy and verification accuracy of the two smart control models of the Bootstrap forest have fully reached the accuracy of the neural network model of the German Pratt company.
At the same time, with the help of the model publishing function of JMP Pro software, the team obtained its C language program code and conducted actual testing of the preset process on the analog machine controlled by the on-site program, and achieved very good results, thus providing the basis for the next step. One-step online production testing and verification laid a solid foundation.
In addition, during the analysis and modeling process, the team also summarized the following experiences for your reference:
- The time span of training materials should not be too long, preferably 3 to 4 months;
- The amount of data for training and verification should not be too large, with 30,000 to 40,000 records being appropriate;
- If equipped with self-adjusting learning function, the prediction accuracy can be higher.
It can be seen that with the proposal and advancement of the Made in China 2025 strategy, the manufacturing industry is gradually transforming and upgrading from manufacturing to "intelligent manufacturing", and it is imperative for smart manufacturing to empower future steel plants. The attempt and exploration of Baosteel's 1580 hot-rolled finish rolling natural widening project provides a good reference and basis for Baosteel's other product lines on the road to smart manufacturing. Scientific analysis tools will surely become indispensable for the manufacturing industry to go further and further. powerful engine.
