A Prediction System of Burn through Point Based on Gradient Boosting Decision Tree and Decision Rules

Abstract

According to the characteristics of sintering process, a sintering end-point prediction system based on gradient boosting decision tree (GBDT) algorithm and decision rules is proposed in this paper. The on-line parameters of the sintering machine, which can characterize the change of the properties of the sintered raw materials in real time, were selected as the input of the model. The soft measurement results of the burn-through point position and temperature were selected as the output. The problem of establishing a system model based on the data collected in the sintering process to dynamically predict the state of burn through point (BTP) was solved. With the combination of process knowledge and several feature selection methods, the important characteristic variables related to the BTP were screened out. the algorithm of GBDT was used to establish the prediction model of BTP and burn through temperature (BTT). The parameters of the ensemble algorithm were optimized by using the methods of grid search and cross-validation, and the system model based on training data was established. On this basis, the corresponding decision model was added to the output of the prediction model, and the prediction accuracy of the system was improved. The establishment process of system model is introduced in detail. The operation results show that the system has better performance.

1. Introduction

In sintering production, the sintering end point is an important technological parameter that affects the quality, output and cost of sinter ore. Suitable and stable sintering end point is the key to ensure the high quality technical indexes of blast furnace, When the sintering end point is advanced, the production capacity of the sintering machine cannot be fully utilized, resulting in a decrease in the yield of the sintered ore; When the sintering end point is delayed, the mixture is transported to the tail unloaded before it is completely burned out, resulting in a decrease in the qualification rate of the sintered ore and an increase in cost.^1,2,3,4,5) however, due to the sintering production with large lag and dynamic time-varying, there is no instrument to detect the sintering end point directly at present.⁶⁾ Whether it’s through the methods of the exhaust gas temperature,^7,8) the negative pressure⁹⁾ and the exhaust gas composition to estimate the BTP,¹⁰⁾ or observing the red layer at the tail¹¹⁾ to qualitatively judge the status of the end point, which are lag in time. It will inevitably affect the stability control of sintering end point and even the whole sintering process.

In order to avoid the adverse effects mentioned above, Japan Kawasaki Steel Company’s Mizushima Plant developed a diagnostic sintering operation control expert system in the early 1990s,¹²⁾ which realized long-term and short term prediction of the end point based on raw material permeability and bellows temperature of temperature rise point, respectively; Vescovi M. R. et al.¹³⁾ used the pallet speed and the middle bellows pressure as inputs to achieve the prediction of the sintering end point and the sintering end temperature. Wu Xiaofeng et al.¹⁴⁾ based on the analysis of the thermal state of sintering process, a prediction model of sintering end-point state was established based on the exhaust gas temperature of the bellows. In the above research, the prediction model based on local variables (such as temperature rise point, raw material permeability or bellows pressure etc.) has the defect of prediction error because the influence of process parameters is not considered, and the models are more suitable for the prediction under steady state. The practical application has a greater limitation.

With the improvement of computer processing ability, some artificial intelligence methods have been applied to sintering process, such as support vector machine (SVM), artificial neural network (Ann) and so on.^15,16,17) This kind of algorithm has powerful nonlinear approximation ability. It is commonly used to establish a model of the relationship between observed data and state parameters.^18,19,20) Huang Xiaoxian²¹⁾ used fuzzy clustering algorithm to analyze sintering production data, and built a prediction model of bed permeability state by support vector machine. Wu Min et al.³⁾ derived the prediction model of BTP by combining the grey theory and BP neural network, and realized the hybrid fuzzy predictive control of the end point. The above models played a guiding role in practical production. However, because of the small amount of data used in the modeling process, even though the prediction accuracy of the training model is very high, there still exists the problem of poor generalization ability. In addition, most models were built using a single algorithm. With the rapid development of integrated algorithms such as bagging and boosting, it turned out that the prediction accuracy of the ensemble algorithm is higher than that of the single algorithm.

Therefore, based on a large number of historical data in sintering production, the relevant parameters of sintering terminal state were comprehensively screened by combining process experience with various feature selection methods. The gradient boosting decision tree (GBDT) integrated algorithm and the inference engine method were used to establish an integrated forecasting system for the BTP and BTT, which realized the advance and accurate prediction of the end point state, it played an important role in guiding the sintering production.

2. Model Structure and Modeling Method

On-line real-time monitoring of sintering end point is a necessary means to realize optimal control of sintering end point. Through the analysis of sintering process mechanism, the relationship between the exhaust gas temperature and the position of the bellows is approximately quadratic near the sintering end point, and the location of the bellows at the highest point of exhaust gas temperature is the sintering end point. A soft measurement of the BTP can be achieved by using quadratic curve fitting. The temperature of three adjacent thermocouples centered on the highest exhaust gas temperature is used to fit the quadratic curve, and the general formula of the curve is as follows:²²⁾   

$$T=A{X}^2+BX+C$$(1)

In the equation, T is the exhaust gas temperature of the bellows (°C) and X is the number of bellows. The adjacent three point coordinates ((X₁,T), (X₂,T), (X₃,T)) with the highest exhaust gas temperature as the center are added to the formula (1) for solving the coefficients A, B and C. The position of the bellows corresponding to the maximum point X_max=−B/(2A) of the curve is the BTP. In order to accurately predict the BTP, the distance of the BTP is obtained by the length of bellows and bellows corresponding to the end point.

We used a variety of feature selection methods: Recursive Feature Elimination (RFE),²³⁾ Stability Selection,²⁴⁾ and Random Forests²⁵⁾ and screened the relevant parameters of the sintering end-point state to determine the optimal feature data set. The main idea of RFE is to iteratively construct extra tree classifiers (or extra tree regression for regression), recursively deleting features that assign the least weights in those classifiers to select the most important features. The stability selection combines quadratic sampling with lasso regression algorithm, and the feature selection results are finally summarized by repeatedly running the feature selection algorithm on different data subsets and feature subsets. The random forest uses multiple decision trees to construct, and its final results are outputted by voting. For classification problems and regression problems, the algorithm uses Gini impurity and MSE to measure the correlation between features and target values, respectively.

Supervised learning is divided into classification and regression. Classification is the problem of predicting the output of discrete class labels, while regression is the problem of predicting continuous quantity output. GBDT is a kind of ensemble learning algorithm, which has been widely mentioned in recent years. It can be used for both classification and regression. The algorithm implements common decision by iterating multiple regression trees. The base learner uses the regression tree whose goal is to minimize the square error, each regression tree learns the conclusions and residuals of all previous trees, and finally the whole decision tree is obtained by accumulating. The algorithm has the advantages of flexible processing of various types of data, high prediction accuracy and strong robustness to abnormal values.^26,27,28)

The sintering process has complex reactions, time lags and parameters uncertainties. A single model can hardly meet the various problems encountered in a changing production environment. Therefore, this study comprehensively screens the model input variables by comparing various feature selection methods, and uses the GBDT algorithm to establish the BTP prediction model and the BTT prediction model respectively, and then, the corresponding decision model is added to the output of the forecast model. According to the model prediction results and the corresponding expert rules, the final decision is realized.

3. Modeling

3.1. Feature Extraction

In the sintering production process, there are many parameters affecting the state of BTP. It mainly includes raw material parameters (such as amount of each raw material, composition of mixed ore, etc.), operating parameters (such as sintering machine speed, air volume, ignition temperature, etc.) and state parameters (such as negative pressure, exhaust gas temperature, etc.). Among them, the mixed mineral component parameters are detected every 4 hours, and the operating parameters and state parameters are collected in real time. Using pearson correlation coefficient method,²⁰⁾ the effects of mixture ratio(the amount of vanadium-containing fine powder, ordinary mineral powder and miscellaneous materials), the amount of flux (calcium ash and magnesium ash) and the amount of fuel (coke powder and white coal) on BTP were analyzed. As shown in Fig. 1.

Fig. 1.

Relationship between the amount of raw material used and the BTP. (Online version in color.)

The properties of sintered mixture are an important parameter that affect the sintering process. It can be seen from the calculation results in the figure that the correlation between these parameters and the burning through point is relatively weak, all of which are less than 0.3, but this cannot deny the significance of the influence of the sintering mixture on the burning through point. Because the composition, particle size and moisture of the sintered mixture are difficult to be obtained in real time and accurately, only the dynamic amount of sintering mixture, flux and fuel and the state of sintering end point are selected for single factor analysis. The above results can not comprehensively characterize the relationship between the properties of sintering materials and the state of sintering end point. Sintering production is a continuous dynamic process, when the properties of sintered raw materials change, the on-line parameters of the sintering machine (including operating parameters and state parameters) will fluctuate in real time. Therefore, the relationship between the on-line parameters of sintering and the state of sintering end point is studied in this paper. The on-line parameters of sintering machine were taken as inputs, and then the variables were further selected in combination with the feature selection method. Sintering machine online parameters have different units or dimensions, which will affect the results of data analysis. To eliminate the dimensional effects between the indicators, all features are processed using the Z-score normalization method. It is obtained that the data of each attribute are clustered around 0 and the variance is 1 to ensure that all the features are scaled in the same way.

To ensure maximum information and noise minimization into the model variables, targeted selection of variables is performed. Firstly, the correlation coefficient between each input variable and the output variable is calculated by the above feature selection method. Then, combined with the prediction accuracy of the GBDT model, the forward selection method (That is, according to the importance ranking of the characteristic variables, the candidate independent variables are introduced into the regression equation one by one.) is used to filter the variables to complete the selection of the optimal feature set. For the three feature selection methods (RFE, stability selection, and random forest), the number of features considered is incremented according to the order of features obtained from each selection method, and then the cut-off points of a set of optimal features are determined by comparing the values of the cross validation scores of the three methods.

3.2. Modeling Based on GBDT Algorithm

3.2.1. GBDT Algorithm

The GBDT algorithm can be viewed as an additive model composed of k trees:   

$${\hat{\mathrm{y}}}_i=\sum _{k=1}^K{f}_k(x_i),f_k\in F$$(2)

Where F is a functional space composed of all trees, and the parameters of the model are: Θ={f₁, f₂, ···, f_k}. Unlike general machine learning algorithms, the addition model does not learn the weights in the d-dimensional space, but directly learns the function(decision trees) set. The objective functions of the above addition model are: $Obj={\sum }_{i=1}^{n}l(y_i,{\hat{y}}_i)+{\sum }_{k=1}^K\mathrm{\Omega }(f_k)$, where Ω represents the complexity of the decision tree. Forward distribution algorithm is used to solve the addition model, which is guided by minimization of objective function. In step t, the model predicts x_i as follows: ${\hat{y}}_i^t={\hat{y}}_i^{t-1}+f_t(x_i)$, where f_t (x_i) is the function (decision tree) we want to learn, and the objective function at this time can be written as:   

$$Ob{j}^{(t)}=\sum _{i=1}^{n}l\left(y_i,{\hat{y}}_i^{t-1}+f_t(x_i)\right)+\mathrm{\Omega }(f_t)+constant$$(3)

According to the Taylor formula, the function f(x+Δx) is expanded at the second order of the point x, and the following equation is obtained:   

$$f(x+\mathrm{\Delta }x)\approx f(x)+f^{\prime }(x)\mathrm{\Delta }x+\frac{1}{2}{f}^{″}(x)\mathrm{\Delta }{x}^2$$(4)

It can be known from Eq. (3) that the objective function is a function of the variable ${\hat{y}}_i^{t-1}+f_t(x_i)$. If the variable ${\hat{y}}_i^{t-1}$ is regarded as x in the Eq. (4) and the variable f_t(x_i) is regarded as Δx in the Eq. (3), the Eq. (3) can be converted into:   

$$\begin{array}{l}Ob{j}^{(t)}=\sum _{i=1}^n\left[l\left(y_i,{\hat{y}}_i^{t-1}\right)+g_i{f}_t(x_i)+{\displaystyle \frac{1}{2}}{h}_i{f}_t^2(x_i)\right]+\mathrm{\Omega }(f_t)\\ \hspace{1em}\hspace{1em}\hspace{1em}+constant\end{array}$$(5)

Where g_i is defined as the first derivative of the loss function, h_i is defined as the second derivative of the loss function, and when the loss function is square loss, $g_i=2({\hat{y}}_i^{t-1}-y_i)$, h_i=2, it is taken into Eq. (5) to obtain the objective function (6):   

$$Ob{j}^{(t)}=\sum _{i=1}^n\left[2({\hat{y}}_i^{t-1}-y_i)f_t(x_i)+f_t{(x_i)}^2\right]+\mathrm{\Omega }(f_t)+constant$$(6)

Among them, (${\hat{y}}_i^{t-1}-y_i$) is called residual. Each step of the GBDT algorithm only needs to fit the residual of the previous models when generating the decision tree. Since the constants in the function do not work during the minimization of the function, we can remove the constant from Eq. (6) and get:   

$$Ob{j}^{(t)}=\sum _{i=1}^n\left[g_i{f}_t(x_i)+\frac{1}{2}{h}_i{f}_t^2(x_i)\right]+\mathrm{\Omega }(f_t)$$(7)

Suppose a generated decision tree has the number of leaf nodes T, The decision tree is composed of a vector ω∈R^T consisting of values corresponding to all leaf nodes, and a function q: R^d→{1,2, ···, T} that maps the feature vector to leaf node index. Therefore, the decision tree can be defined as f_t(x)=ω_q(x), and the complexity of the decision tree can be defined by the regular term $\mathrm{\Omega }(f_t)=\gamma \mathrm{T}+\frac{1}{2}\lambda {\sum }_{j=1}^T{\omega }_j^2$, that is, the complexity of the decision tree model is determined by the number of leaf nodes and the L2 norm of the corresponding vector value.

The set I_j={i|q(x_i)=j} is defined as a collection of all training samples that are divided into leaf nodes j. Equation (7) can be reorganized into the sum of T independent quadratic functions according to the leaf nodes of the tree:   

$$\begin{array}{l}Ob{j}^{(t)}=\sum _{i=1}^n\left[g_i{f}_t(x_i)+{\displaystyle \frac{1}{2}}{h}_i{f}_t^2(x_i)\right]+\mathrm{\Omega }(f_t)\\ \hspace{1em}\hspace{1em}\hspace{0.5em}=\sum _{j=1}^T\left[G_j{\omega }_j^2+{\displaystyle \frac{1}{2}}(H_i+\lambda ){\omega }_j^2\right]+\gamma T\end{array}$$(8)

Among them, $G_j={\sum }_{i\in I_j}{g}_i$, $H_j={\sum }_{i\in I_j}{h}_i$. Assuming that the structure of the tree is fixed, that is, the function q(x) is determined. Let the first derivative of the function Obj^(t) be equal to 0, and the corresponding value of the leaf node j can be obtained:   

$${\omega }_j^{*}=-\frac{G_j}{H_j+\lambda }$$(9)

At this point, the value of the objective function is as follows:   

$$Obj=-\frac{1}{2}\sum _{j=1}^T\frac{G_j^2}{H_j+\lambda }+\gamma T$$(10)

In summary, the learning process of the GBDT algorithm can be roughly described as follows: 1. The algorithm generates a new decision tree every iteration; 2. Before the beginning of each iteration, the first derivative g_i and second derivative h_i of the loss function at each training sample point is calculated. 3. Generate a new decision tree through the greedy strategy, calculate the predicted value corresponding to each leaf node by Eq. (9); 4. Add the newly generated decision tree f_t(x) to the model: ${\hat{y}}_i^t={\hat{y}}_i^{t-1}+\epsilon f_t(x_i)$, Where ε is called the learning rate. According to the addition model Eq. (2), the model to be finally learned can be obtained.

3.2.2. Model Construction

There are three main steps of the construction process of GBDT model: data set dividing, model construction and parameter optimization, model evaluation.

1) Data set dividing. By using random sampling, the samples in the data set were divided into training set and testing set at the ratio of 9:1. To facilitate management, two tables were created to manage input and output data, and they had one-to-one correspondence.

2) Model construction and parameter optimization. In this study, the GBDT machine learning model of Python library was used to establish the prediction model of BTP and BTT, respectively. The grid search and cross-validation method were combined to optimize the parameters of estimator, max_depth, min_samples_split, min_samples_leaf, learning_rate in GBDT model. Taking the prediction model of BTP as an example, the optimization process is as follows: (1) Starting with step size and iteration number, a larger step size and a smaller iteration number are selected. The n_estimator is too small, which is easy to under-fitting, the n_estimator is too large, which is easy to over-fitting, so a moderate value is generally chosen. Because the training set used in this study was large, the number of iterations for the model was increased. Step size was set to 0.1, iterations ranged from 100 to 200, and grid search was performed. (2) After finding the most suitable number of iterations (n_estimator), a grid search was performed on the max_depth and min_samples_split of the decision tree, and the max_depth range was set to 8–16, and the min_samples_split range was 50–100. (3) Since min_samples_split is related to the min_samples_leaf, when the maximum depth was obtained, the two parameters were adjusted together. (4) After finding the best min_samples_split and min_samples_leaf, then the absolute number was used to perform the optimal search for max_features. (5) Now that we have basically got all the parameters, we try to further improve the generalization ability of the model by reducing the step size and increasing the number of iterations.

3) Model evaluation. For the prediction of the BTP and the BTT, both the fitting effect of the model and the error loss value between the predicted value and the real value are paid attention to. Therefore, the model was evaluated by both the goodness of fit (R²) and mean square error (MSE), as shown by Eq. (11). In order to reflect the performance of the model more intuitively, we also calculated the hit ratio of the prediction model under the specific error.   

$$MSE={\displaystyle \frac{1}{m}}\sum _{i=1}^m{({\mathrm{y}}_i-{\hat{y}}_i)}^2,R^2=1-{\displaystyle \frac{\sum _{i=1}^m{({\mathrm{y}}_i-{\hat{y}}_i)}^2}{\sum _{i=1}^m({\hat{y}}_i-y_i{)}^2}}$$(11)

3.3. Decision Model Based on Multi-orientation Information

The end-point location prediction model can realize the quantization and digitization of the prediction target. However, when some input parameters appear noise signal or fluctuate greatly, the prediction accuracy of the model will be reduced and the output result will be abnormal. By combining the actual production experience, the influence law of the parameters in each process and the results of the sintering end-point prediction model, the decision model of sintering BTP was established based on the inference machine method. The big data prediction model can reveal the rule that is difficult to find by traditional experience, while process experience can guide the model operation and judge the accuracy of model prediction results. Therefore, the effective integration of prediction model and decision model is an important means for the field personnel to obtain stable, accurate and interpretable production plan. The structure of the decision model is shown in Fig. 2.

Fig. 2.

Structure diagram of decision model.

4. Results and Discussion

4.1. Feature Extraction

Figure 3 shows the cross-validation scores of each feature selection method (RFE, stability selection and random forest), which use feature numbers in the dataset corresponding to the BTP regression model (Fig. 3(a)) and BTT regression model (Fig. 3(b)).

Fig. 3.

Relationship between characteristic variables and model prediction accuracy. (a) BTP and variables; (b) BTT and variables. (Online version in color.)

In Fig. 3(a), the average R² score of 3 cross-validations was used to evaluate the fitting performance. The R² scores of RFE, stability selection and random forest increase rapidly with the addition of features, and converge to around 0.61 after 15 features. Compared with the other two methods, the R² score of the random forest increases more slowly with the number of features, and the R² score of the extreme point is relatively low. Among the three feature selection methods used in this paper, the cross-validation R² score of the RFE method is the highest. The RFE uses a gradient boosting regression tree model that can handle various types of data flexibly with high prediction accuracy. In addition, RFE implements backward feature elimination, which removes irrelevant features from the beginning of the feature selection process, limiting the negative effects of irrelevant features on cross-validation scores. Therefore, this helps it produce the best cross-validation R² score. When the optimal feature number was selected, it was found that the RFE cross-validation R² score hardly increased after about 12 features and converged to the maximum value of 0.62. As a result, we decided to use only the first 12 features to fit the BTP regression model. Using as few features as possible without causing severe loss of precision is the best chance to avoid over-fitting and constructing the simplest and most accurate predictive model. In Fig. 3(b), the above analysis method was also used. When the first nine features were selected, the RFE method had the highest cross-validation R² score of 0.74. We used this feature set to fit the BTT regression model. The input parameters and output parameters of the end position prediction model and the end temperature prediction model are shown in Fig. 4.

Fig. 4.

Relationship between input and output parameters of prediction model. (Online version in color.)

The relationship between the input and output parameters of the model is calculated by RFE feature selection method. When the input parameters of the BTP prediction model are sintering machine speed, nine roller velocity, exhaust gas temperature of main pipe, air volume of main exhaust fan, the valve opening of main pipe, the exhaust gas temperature of 1, 9, 11, 13 bellows and the negative pressure of 5, 9, 11 bellows, R² score converges to a maximum of 0.62. When the input parameters of the BTT prediction model are sintering machine speed, nine roller velocity, exhaust gas temperature of main pipe, air volume of main exhaust fan, the valve opening of main pipe, the exhaust gas temperature of 1, 11, 13 bellows and the negative pressure of 5 bellows, R² score converges to the maximum value of 0.74. Therefore, the above parameter set is selected as the input parameter of the model. As shown in Fig. 4, the input parameters of the BTP prediction model are sintering machine speed, nine roller velocity, exhaust gas temperature of main pipe, air volume of main exhaust fan, the valve opening of main pipe, the exhaust gas temperature of 1, 9, 11, 13 bellows and negative pressure of 5, 9, 11 bellows. The input parameters of the BTT prediction model are those labeled in blue. Among them, the sintering machine speed is the running speed of sintering pallet; The nine roller velocity is an important parameter of the sintering charge distributor, and this parameter can reflect the change of the pallet mixture surface to a certain extent; the air volume of the main exhaust fan is an important parameter to reflect the permeability of sinter layer; the valve opening of main pipe is a method to adjust the air volume of the main exhaust fan in the sintering process; the exhaust gas temperature of the main pipe and bellows refer to the temperature detected in the exhaust gas of the main pipe and the bellows; the negative pressure of bellows are the value of the pressure detected in the bellows. The output parameter of the end position prediction model is the position distance of the BTP, which unit is centimeter. The output parameter of the BTT prediction model are continuously distributed exhaust gas temperature values, which unit is degree Celsius.

4.2. Evaluation and Verification of the Model

The actual production data of a 360 m² sintering machine in a steel plant for nearly three years was selected. Combined with the statistical results of massive data and the experience of field experts, the interval definition method was used to formulate screening criteria for all process parameters related to the sintering end point, and the abnormal and repeated data clearing was completed. The variables filtering criteria are shown in Table 1.

Table 1. Filtration criteria for sintering process parameters.

Variable name	Minimum	Maximum	Incremental range
Round roll speed (r/h)	180	715	±1
Nine roller velocity (r/h)	370	820	±1
Sintering machine speed (m/min)	1.40	2.50	±0.01
Coal gas pressure (kpa)	2.5	8.6	±0.1
Coal gas flow (m3/h)	1000	3000	±1
Combustion air pressure (kpa)	4	8.5	±0.1
Combustion air flow (m3/h)	6500	13850	±1
Ignition temperature (°C)	1100	1200	±1
Trolley material thickness (mm)	600	700	±1
Air volume of main exhaust fan (m3/h)	3500	9000	±1
Valve opening of main pipe (%)	46.00	92.00	±0.01
Exhaust gas temperature of bellows (°C)	0	550.00	±0.01
Negative pressure of bellows (kpa)	6.00	16.00	±0.01
Temperature of main pipe (°C)	0	250.00	±0.01
Negative pressure of main pipe (kpa)	8.00	16.00	±0.01

In order to eliminate the influence of the dimensions and numerical values of each parameter, the collected data was standardized using the z-score method, namely:   

$${\text{y}}_i=\frac{x_i-\overline{x}}{s}$$(12)

where, $\overline{x}=\frac{1}{n}{\sum }_1^n{x}_i$, $s=\sqrt{\frac{1}{n-1}{\sum }_{i=1}^n{(x_i-\overline{x})}^2}$. X_i, $\overline{x}$ and s represent the original value, mean and standard deviation, respectively.

After data preprocessing by the above method, a total of 15292 sets of modeling data were obtained, 90% of the data were randomly selected from the data set as the training set (13763 sets), and the remaining 10% data was used as the test set (1529 sets). The end position and end temperature prediction models were trained and tested. The above prediction models of the sintering BTP and the BTT were trained and tested. The optimal parameters of the prediction model in the training set are shown in Table 2. The prediction accuracy of the model in the test set is shown in Table 3. The prediction results of the BTP and the BTT are shown in Figs. 5 and 6, respectively.

Table 2. Optimal parameters of forecasting model.

Model	n_estimators	max_depth	min_split	min_leaf	learning_rate
BTP	150	13	80	6	0.1
BTT	460	8	90	40	0.1

Table 3. Test results of BTP prediction model and BTT prediction model.

Model	BTP		BTT
Index	R2	MSE	R2	MSE
Result	0.671	20967.22	0.809	325.41

Fig. 5.

Result of the BTP prediction model in the test set. (Online version in color.)

Fig. 6.

Results of the BTT prediction model in the test set. (Online version in color.)

Figures 5 and 6 are the predicted results of the BTP and the BTT in the test set respectively, and diagrams (a1) and (b1) are the partial amplification of the test result. Among them, the abscissa represents the number of test samples, and the ordinate represents the value of the predicted result. Each sample has an actual value and a predicted value, represented by different colors and shapes in the figure. When the predicted value matches the actual value, the predicted result is correct, otherwise it is incorrect. It can be seen from the graphs and Table 3 that the predicted value of the BTT is basically consistent with the trend of the actual value, and the fitting degree R² reaches 0.809. However, there is a problem that the end position prediction has a large prediction error. Therefore, the decision model is used to judge the output of the prediction model, and the result of the end point prediction model with decision rules is shown in Fig. 7.

Fig. 7.

Results of BTP prediction model with decision rule added. (Online version in color.)

From the comparison of Figs. 5 and 7, it can be seen that the prediction performance of the BTP prediction model with added decision rules is significantly improved, especially when the end position is distributed between 70 m and 76 m, the model can accurately predict the position and change trend of the BTP. In order to evaluate the model more intuitively, we calculated the hit rate of the BTP under different errors, as shown in Table 4. With the decrease of error precision, the hit rate of the end position prediction model continues to increase. When the error is controlled within the range of 1.25 m, the prediction hit ratio of the model with decision rule can reach 85.6%. Compared with the traditional method based on the exhaust gas temperature to determine the BTP corresponding to the bellows number, the error accuracy is reduced by 3 times. Therefore, the end-point location prediction model with decision rules can meet the requirements of the site for end point judgment.

Table 4. Hit ratio of the BTP prediction model with different error accuracy.

No.	accuracy error/m	No decision rules/%	Add decision rules/%
1	0.5	46.9	50.3
2	0.75	59.7	66.9
3	1	70.7	79.6
4	1.25	77.4	85.6
5	1.5	83.1	89.4
6	1.75	86.2	91.6
7	2	88.4	93.3

4.3. Application of the Model

Based on the above prediction model, the sintering terminal state prediction system was developed by using Python and JavaScript language, and applied to a domestic sintering plant. The main function of the system is to predict the position and temperature of the current BTP about 15 minutes in advance and realize the on-line visualization of the sintering end-point state. The end-point state forecasting system is based on the combination of the integrated learning algorithm and the decision-making model. When the end-point state fluctuates, the operation suggestions of air volume of main exhaust fan and the thickness of pallet mixture can be given to assist the operator in the sintering site to adjust the production parameters in advance.

After the system was put into operation on a 360 m² sintering machine in a domestic steel plant, the stability of the BTP was improved, and the drum index and the screening index of the sinter ore were increased by 1–1.5%. According to the ironmaking process theory, the powder content of sinter ore is changed by 1%, which affects the coke ratio of the blast furnace by 0.5%. The annual production capacity of this sintering machine is 3.7 million tons/year. Suppose an average grade of 58% for sinter ore, a coke unit price of 1500 yuan/ton, a coke water content of 5%, a blast furnace iron loss of 0.05%, and a coke ratio of 380 kg/t, the forecasting system can save 42889.854 tons of coke per year for the blast furnace process, and save costs of about 6434800 yuan.

5. Conclusion

According to the characteristics of the sintering process and the difficulty of measuring the BTP, a sintering end-point prediction system based on the GBDT algorithm and decision rules was proposed. The problem of establishing a system model based on-line acquisition data of sintering process to dynamically predict the state of BTP was solved.

(1) The on-line parameters of the sintering machine, which can characterize the change of the properties of the sintered raw materials in real time, were selected as the input of the model. The soft measurement results of the burn-through point position and temperature were selected as the output. Based on the combination of process experience, the three methods of RFE, stability selection and random forest selection were compared, and the important characteristic variables related to the end state of sintering were determined.

(2) The BTP prediction model and the BTT prediction model were established by using the integrated learning algorithm (GBDT algorithm). The corresponding decision rules were added to the output of the forecast model, which is helpful to improve the prediction accuracy. The method proposed in this paper is suitable for the advance and accurate judgment of the state of BTP.

(3) The practical application in sintering production shows that the prediction system can accurately predict the state of the sintering end point, enhance the stability of the end point, improve the metallurgical properties of the sinter.

Acknowledgement

Thanks are give to the financial supports form the key Program of National Nature Science Foundation of China (U1360205),the Hebei Province Graduate Innovation Project (CXZZBS2017120), the Scientific Research Program Project of Hebei Education Department (QN2019200).

References

• 1)   P. R.  Dawson: Ironmaking Steelmaking, 20 (1993), 135.
• 2)   M.  Wu,  C. H.  Xu,  J. H.  She and  W. H.  Cao: J. Process Control, 22 (2012), 925.
• 3)   M.  Wu,  P.  Duan,  W. H.  Cao,  J. H.  She and  J.  Xiang: Expert Syst. Appl., 39 (2012), 5971.
• 4)   M.  Wu,  C. S.  Wang,  W. H.  Cao,  X. Z.  Lai and  X.  Chen: Control Eng. Pract., 20 (2012), 1065.
• 5)   W. H.  Cao,  Y. Y.  Zhang,  J. H.  She,  M.  Wu and  Y.  Cao: Inform. Sci., 466 (2018), 1.
• 6)   R.  Usamentiaga,  J.  Molleda,  D. F.  Garcia and  F. G.  Bulnes: Sensors, 13 (2013), 10287.
• 7)   X. H.  Fan: Mathematical Models and Expert Systems of Iron Ore Agglomeration, Science Press, Beijing, (2013), 8 (in Chinese).
• 8)   J. Q.  Hu and  E.  Rose: Control Eng. Pract., 5 (1997), 247.
• 9)   Q. M.  Feng,  T.  Li,  X. H.  Fan and  T.  Jiang: Trans. Nonferr. Met. Soc. China, 10 (2000), 804.
• 10)   N. K.  Nath and  K.  Mitra: Ironmaking Steelmaking, 31 (2004), 199.
• 11)   M. H.  Li and  Y. F.  Sun: J. Huazhong Univ. Sci. Technol. (Nat. Sci.), 32 (2004), 71 (in Chinese).
• 12)   T.  Fukagawa,  S.  Iyama and  M.  Okuyama: Kawasaki Steel Giho, 23 (1991), 203 (in Japanese).
• 13)   M. R.  Vescovi,  M. M.  Lamega and  A.  Farquhar: IEEE Expert, 12 (1997), 42.
• 14)   X. F.  Wu and  M. R.  Fei: J. Zhejiang Univ., 41 (2007), 1722 (in Chinese).
• 15)   C.  Cortes and  V.  Vapnik: Mach. Learn., 20 (1995), 273.
• 16)   P. J.  Laitinen and  H.  Saxén: Ironmaking Steelmaking, 34 (2013), 119.
• 17)   J. E.  Meng,  J.  Liao and  J. Y.  Lin: IEEE Trans. Fuzzy Syst., 8 (2000), 314.
• 18)   P.  Harrington: Machine Learning in Action, Manning Publications, Shelter Island, NY, (2012), 101.
• 19)   H.  Zhou,  J. P.  Zhao,  C. E.  Loo,  B. G.  Ellis and  K. F.  Cen: ISIJ Int., 52 (2012), 1550.
• 20)   M. J.  Er,  J.  Liao and  J. Y.  Lin: IEEE Trans. Fuzzy Syst., 8 (2000), 314.
• 21)   X. X.  Huang,  X. H.  Fan,  X. L.  Chen,  G. M.  Yang and  M.  Gan: ISIJ Int., 56 (2016), 2113.
• 22)   X. H.  Fan and  H. D.  Wang: Mathematical Model of Sintering Process and Artificial Intelligence, Central South University Press, Changsha, (2002), 179 (in Chinese).
• 23)   I.  Guyon, J. Weston, S. Barnhill and V. Vapnik: Mach. Learn., 46 (2002), 389.
• 24)   N.  Meinshausen and  P.  Bühlmann: J. R. Stat. Soc. Ser. B, 72 (2010), 417.
• 25)   T.  Hastie,  R.  Tibshirani and  J.  Friedman: The Elements of Statistical Learning, Springer, New York, (2009), 587.
• 26)   X. L.  Ma,  C.  Ding,  S.  Luan,  Y.  Wang and  Y. P.  Wang: IEEE Trans. Intell. Transp. Syst., 18 (2017), 2303.
• 27)   R. A.  Fisher and  J. H.  Bennett: Statistical Methods, Experimental Design and Scientific Inference, Oxford University Press, USA, (1990), 8.
• 28)   J. H.  Friedman: Ann. Stat., 29 (2001), 1189.