You are on page 1of 5

ISIJ International, Vol.

56 (2016),
ISIJ International,
No. 12 Vol. 56 (2016), No. 12, pp. 2113–2117

Bed Permeability State Prediction Model of Sintering Process


Based on Data Mining Technology

Xiaoxian HUANG, Xiaohui FAN,* Xuling CHEN, Guiming YANG and Min GAN

School of Minerals Processing and Bioengineering, Central South University, Changsha, Hunan, 410083 China.
(Received on April 1, 2016; accepted on August 4, 2016)

The bed permeability state prediction model of sintering process based on data mining technology was
proposed in this study. Firstly, the sintering production data were analyzed by fuzzy clustering algorithm,
to make a comprehensive evaluation of the bed permeability state. Then the prediction model of bed
permeability state was established via support vector machine, based on the sample data that obtained
in the cluster analysis. The bed permeability prediction model has a good learning and generalization abil-
ity, its prediction hit rate reached 87.5%. The practical application showed that: the sintering process could
be stabilize effectively, since the operation parameters was adjusted according to the prediction results of
bed permeability state; the standard deviation of burn through temperature and burn through point was
decreased by 47% and 34% respectively.

KEY WORDS: iron ores sintering; bed permeability; prediction model; data mining.

original material bed or granulation of single iron ore.


1. Introduction
Data mining is a technology to discovery the relationship
Iron ore sintering is a complex industrial process, which between the pattern, knowledge and data. At present, data
involves a large number of physical parameters such as mining is the research and application focus of telecom-
temperature, pressure, velocity and flow. It also referred munications, banking, transportation, retail and other com-
to many issues like the gas distribution in the material bed mercial areas, and it also has the technology advantage in
and the temperature field distribution.1) The sintering pro- biology, astronomy and other scientific research fields.9–12)
cess state is the major factor affecting the sinter quality and In sintering production, engineers or operators usually adjust
yield indices, it mainly includes the thermal state and bed the production parameters according to the bed permeability
permeability state. Since the temperature data of sintering state, which is evaluated through current operating param-
process are relatively easy to obtain, thermal state has been eters and thermal state. Therefore the large amount of sin-
extensively studied. The typical evaluation methods of bed tering production data contained the relevant information of
permeability like Ergun formula and Voice formula are lim- bed permeability. In this study, the data mining technology
ited in the practical application,2,3) because the online mea- was used to analysis the sintering production data, and the
surements of their variables are difficulty and inaccuracy. In clustering and classification models were established. It will
the Operation Guidance System that developed by Japan’s play an important role in guiding the sintering production.
Kawasaki Steel Corporation in 1980s,4) the bed permeability
state was evaluated with a three-dimensional matrix which
2. Model Structure and Modeling Method
including the air flow of main exhaust fan, the maximum
temperature of wind box and the burn through point (BTP). The bed permeability state of sintering process is dif-
Bo Jiang et al.5,6) established a comprehensive evaluation ficult to measure directly, and the Ergun formula or Voice
model of bed permeability state via fuzzy logic method, formula cannot be used in practical application due to the
in which the bed permeability was evaluated through the inaccurate measurement of related variables. However, the
parameters of raw materials, operating parameters and permeability state determines the vertical sintering speed of
process state parameters. But this model almost considered material bed, then effect the highest temperature of waste
all factors related with bed permeability, it had the problem gas and its position in the moving direction of sintering
that some variables cannot be detected in continuous in the pallet, which calls the burn through temperature and burn
practical application. In addition, some researchers7,8) estab- through point. Meanwhile, the permeability state determines
lished the bed permeability prediction model that considered the adjustment of moving speed of sintering pallet or air
the parameters such as particle size distribution, air velocity quantity below the material bed. In other words, the burn
and chemical composition, but it is only suitable for the through temperature, burn through point and pallet speed or
air quantity can reflect the bed permeability state. Since the
* Corresponding author: E-mail: csufanxiaohui@126.com bed permeability state is mainly influenced by the physical
DOI: http://dx.doi.org/10.2355/isijinternational.ISIJINT-2016-193 and chemical properties of raw materials, the moisture con-

2113 © 2016 ISIJ


ISIJ International, Vol. 56 (2016), No. 12

tent and coke breeze ratio of mixture, the way of granulation the sintering production, xk(k = 1,2,…,n) is a 3-dimensional
and distributing. It means that the permeability state can be vector, each feature value represents the burn through point,
predicted by the variables such as mixture size, moisture burn through temperature and air quantity respectively.
of mixture, coke breeze ratio, mixture temperature, bed The data set X is divided into 5 clusters, V = {v1,v2,…, v5},
height, negative pressure of ignition furnace and No.1 wind and vi(i = 1,2,…,5) is the center of each cluster. The fuzzy
box. The relationship between bed permeability state and membership function µik ∈( 0,1) is used to describe the
other sintering parameters is shown in Fig. 1. Where, the membership degree between sample data xk and cluster cen-
mixture size is the average particle size of mixture, and it ter vi. Fuzzy C-means clustering algorithm (FCM) is using
is calculation results according to the size composition and iterative optimization method that based on the least square
proportion of raw materials; the moisture of mixture refers principle to divide the data set. The expression of objective
to the water content of mixture after mixing and granulating; function Jm is as follows:
the coke breeze ratio refers to the proportion of coke breeze 5 n
in the mixture; the mixture temperature is the temperature of J m (U ,V ) = ∑ ∑ µikm dik2 ....................... (1)
mixture after steam preheating; the bed height is the height i =1 k =1

of mixture bed on the sintering pallet before ignition; the Where, dik = xk − vi is represent the Euclidean distance of
negative pressure of ignition furnace and No.1 wind box sample data xk and cluster center vi; m(m > 1) is a weighted
refers to the pressure values that detected in the ignition index, which is used to adjust the degree of membership
furnace and No.1 wind box. degree in different fuzzy cluster. The sum of all the mem-
Clustering and classification are both commonly used bership degree of any one sample data for every cluster must
data mining methods. Clustering is to divide a data set into be 100%, as follows
different classes or clusters according to a particular stan- 5

dard, and fuzzy clustering is to determine the degree of each ∑ µik = 1, ∀k = 1, 2, n ....................... (2)
i =1
data belonging to each cluster via membership function. Using Lagrange multiplication to solve the optimization
Classification is a supervised learning method, which is to problem of minimizing the objective function Jm under
find the classification rules by training the data sets that has above constrains. The fuzzy membership function μik and
known classes, and to predict the class of new data. Support cluster center vi can be written as
vector machine (SVM) is a classification algorithm, and a 1
new kind of machine learning method which is based on the µik = 2
, ∀i = 1, 2,, 5; ∀k = 1, 2,, n
5
 dik  m −1 .... (3)
structure risk minimization principle of statistical learning
theory. SVM can solve the problem of high dimensional and ∑  d jk 
j =1
local optimal, and has good learning ability and generaliza-
n
tion ability. In this study, the fuzzy clustering was used to
analyze the variables that can reflect the bed permeability ∑ µikm xk
state in sintering process, to realize the comprehensive vi = k =1
n
, ∀i = 1, 2,, 5 .................... (4)
evaluation of permeability state, and the SVM was used ∑µ m
ik
k =1
to establish the prediction model of bed permeability state.
The optimal fuzzy classification matrix and cluster
centers are obtained through iterative algorithm according
3. Modeling
to formula (3) and formula (4), and finally complete the
3.1. Evaluation of Bed Permeability State fuzzy clustering division. In addition, considering the FCM
For the sample data set X = {x1, x2,…, xn} that extracted from algorithm is a local search algorithm essentially. It is very

Fig. 1. The relationship between bed permeability state and other sintering parameters.

© 2016 ISIJ 2114


ISIJ International, Vol. 56 (2016), No. 12

sensitive to the initial value of cluster center, and easy to fall onto the 2-dimensional plane, and the different clusters were
into local minimum point. Therefore, the genetic algorithm, marked in different colors and shapes. It can be seen that
which has the strong global search ability, was introduced the fuzzy clustering algorithm can divide the sintering data
to search the whole state space in this study, which made set into several cluster effectively. Fuzzy clustering model is
the individual get rid of the local optimal solution. The an effective method to evaluate the bed permeability state,
algorithm flow as illustrated in Fig. 2. In this algorithm, the and it can provide important data for establish the predic-
genetic algorithm is used to optimize the cluster center. So tion model.
in the population initialization process, the cluster centers
are randomly initialized according to the population size. 3.2. Prediction of Bed Permeability State
The algorithm evaluates the individual by the objective The sample data set {(xi,yi), i = 1,2,…,N} can be set up
function of fuzzy C-means clustering, that the fitness func- by combining the fuzzy clustering results of bed perme-
tion can be expressed by formula (1). The iterative process ability and sintering production data, where xi is the input
is to do the genetic operations included selection, crossover parameters, and yi is the output parameter. xi is a 7-dimen-
and mutation, to find the best cluster center and optimized sional vector, each feature value represents the mixture
the objective function of fuzzy C-means clustering. size, moisture of mixture, coke breeze ratio, bed height,
The sample data of clustering were selected from the his- mixture temperature, negative pressure of ignition furnace
torical data of sintering production of Steel Corp. in China, and No.1 wind box; yi represents the bed permeability state;
where the air quantity is represented by the frequency of N represents the total number of samples. For the nonlinear
exhaust fan. Accordingly, three feature parameters that
reflect the bed permeability state are frequency of exhaust
fan, burn through point and burn through temperature. For Table 1. The clustering centers of bed permeability state.
the fuzzy C-means clustering, the weighted index m in Bed permeability Frequency of Burn through Burn through
formula (1) is 3, and the number of iterations is 20. For state exhaust fan (Hz) point (m) temperature (K)
the genetic algorithm, the population size, which means Very good 42.84 39.26 625.15
the group number of cluster centers, is 20, the maximum
Good 43.17 39.78 597.87
generation, which means the highest number of iterations,
is 100, the crossover probability is 0.7, and the mutation Normal 43.71 38.87 647.64
probability is 0.05. After clustering analysis via above algo- Bad 44.96 39.02 626.09
rithm, the objective function Jm = 89.43, and the results of Very bad 45.77 39.45 591.24
multiple runs were very close, which means the algorithm
can converge to the global optimal point. Table 1 shows
the clustering centers, and the corresponding relationship
between each cluster and bed permeability state can be con-
firmed according to expert experience. The clustering results
as shown in Fig. 3, the 3-dimensional data set was projected

Fig. 3. The clustering results. (a) is the relationship between fre-


quency of exhaust fan and burn through point; (b) is the
Fig. 2. The flow chart of fuzzy clustering algorithm that improved relationship between frequency of exhaust fan and burn
by genetic algorithm. through temperature.

2115 © 2016 ISIJ


ISIJ International, Vol. 56 (2016), No. 12

separable cases, the input vectors are mapped to a high Only small parts of the solution ai of formula (9) are not
dimensional feature space using the nonlinear function ϕ, zero, and the corresponding samples xi are the support vec-
and then one optimal hyper-plane is constructed in the high tor. Thus, the optimal classification function is as follows:
dimensional feature space:
 l 
f ( x ) = sgn  ∑ ai yi K ( xi , x ) + b  .............. (10)
ω T ⋅ φ ( x ) + b = 0 ............................ (5)  i =1 
Taking into account that some samples cannot be cor- Where sgn is the symbolic function, l is the number of
rectly separated in the case of linear non-separable, the support vectors, and b is the classification threshold. The
slack variables ξi ≥ 0, i = 1,2,…, N were introduced. Finally, model structure is shown in Fig. 4.
the classification problem of original sample space can be The prediction model of bed permeability state was
expressed as: established based on the above SVM classification algo-
rithm. According to the sintering production data and the
yi [ω T ⋅ φ ( x ) + b] ≥ 1 − ξi , i = 1, 2,, N ............. (6) clustering result of bed permeability state, 700 sets of data
According to the principle of construction risk minimiza- samples were constructed, among which 500 sets were used
tion, the minimum risk bounds of the classification problem as training sample, and the other 200 sets were used as test-
can be obtained from the following optimization problem: ing sample. The kernel function of support vector machine
N is RBF kernel, and the model parameters are optimized by
1
min J (ω , ξ ) = ω T ω + c ∑ ξi .................. (7) cross validation method. The prediction results of the model
2 i =1 are shown in Fig. 5. Where, the abscissa represents sample
Through a series of optimization methods, such as intro- number, which was optionally given, and the ordinate rep-
ducing the Lagrange function, the optimization problem is resents bed permeability state. Each sample has an actual
transformed into solving the maximum value of following state and a predict state, expressed in different colors and
function, under the constraint conditions of 0 ≤ ai ≤ c and shapes in this figure. When the predict state and actual state
n
matches, it indicates that the prediction result is correct,
∑ ai yi = 0 . Where ai is a Lagrange multiplier. otherwise it is not correct.
i =1

1 N N
Q a, φ ( xi ) = − ∑ yi y j φ ( xi ) φ ( x j ) ai a j + ∑ ai ...... (8)
T

2 i , j =1 i =1

The above optimization problem is a quadratic program-


ming problem under the inequality constraints. In order to
reduce the computational complexity, the kernel function
K(xi,xj) is introduced to replace the dot product of high
dimensional feature space. The quartic programming prob-
lem can be expressed as follows:
 1 N N

 max Q ( a , K ( xi , x j ) ) = − ∑
2 i , j =1
yi y j K ( xi , x j ) ai a j + ∑ ai
 i =1

 N
s.t. ∑ ai yi = 0 , i = 1, 2,, N
  i =1
 0 ≤ ai ≤ c
.......................................... (9)
Fig. 5. The prediction results of bed permeability state.

Fig. 4. The structure of the prediction model of bed permeability state.

© 2016 ISIJ 2116


ISIJ International, Vol. 56 (2016), No. 12

Table 2. The statistical data of burn through state.


5. Conclusions
Average Maximum Minimum Standard
Items Date (1) According to the characteristics of sintering process
value value value deviation
05:00~10:00
and the difficulty of bed permeability measurement, the bed
Burn (Manual 575 641 510 53.65 permeability state prediction model based on data mining
through control) technology was proposed in this study. This model evaluates
temperature 10:00~15:00 the bed permeability state through the clustering analysis of
(K) (Intelligent 638 656 555 28.29 historical production data, and then establishes the classifi-
control)
cation prediction model by support vector machine.
05:00~10:00 (2) The comprehensive evaluation of bed permeability
(Manual 39.35 42.81 37.32 1.25
Burn control)
state was realized through the fuzzy clustering algorithm
through that improved by genetic algorithm. The clustering results
point (m) 10:00~15:00
(Intelligent 38.72 41.56 38.74 0.83
show that this algorithm can divide the sintering data set
control) into a number of clusters clearly. The prediction model of
bed permeability state was established based on support
vector machine, it has a good learning and generalization
As can be seen from Fig. 5, there are only 25 sets of sam- ability, and its prediction hit rate reached 87.5%.
ples that the predicted state were not match with the actual (3) The practical application in sintering production
state, which means that the prediction hit rate is 87.5%. The show that the prediction model can accurately predict the
classification model of bed permeability state based on SVM bed permeability state, and realize the stable control of the
has a good learning and generalization ability. sintering process state while combined with soft sensor
model of thermal state and intelligent control model. The
standard deviation of burn through temperature and burn
4. Model Applications
through point was decreased by 47% and 34% respectively.
The bed permeability prediction model based on data
mining technology was applied to the sintering production Acknowledgements
of a Steel Crop. in China. The intelligent control system of The authors would like to thank the National Natural
sintering process was established, which combined with the Science Foundation of China (No. 51474237) and the
prediction model of bed permeability state, the soft sensor Postdoctoral Science Foundation (No. 2013M540639) for
model of thermal state and the intelligent control model. The the support given to this research.
intelligent control system has been put into used in sinter-
ing production. When the bed permeability state fluctuates, REFERENCES
the frequency of main exhaust fan is adjusted preliminarily
1) X. H. Fan: Mathematical Models and Expert Systems of Iron Ore
according to the expert control rules, to stabilize the sinter- Agglomeration, Science Press, Beijing, (2013), 8.
ing process state; when the process state gradually stabi- 2) X. C. Qiu, R. Z. Cai and S. C. Fu: Sinter. Pellet., 11 (1986), 16.
3) D. Z. Wang, G. Z. Qiu and Y. H. Hu: Mineral Processing. Science
lized, the fuzzy controller figure out the further adjustment Press, Beijing, (2008), 408.
advice of exhaust fan frequency according to the thermal 4) K. Nakashima: Sinter. Pellet., 12 (1987), 41.
state, to optimize the sintering process state. The intelligent 5) X. H. Fan, H. D. Wan, T. Z. Huang, J. Chen, S. N. Yang, T. Z. Zheng
and Y. L. Wang: Sinter. Pellet., 23 (1998), 13.
control system was put into online guide in a sintering plant 6) B. Jiang: Master’s Thesis, Central South University of Technology,
of China, the statistical results of burn through state are (2000), http://d.wanfangdata.com.cn/Thesis/Y357990, (accessed 2011-
08-18).
shown in Table 2. 7) R. Venkataramana, S. S. Gupta and P. C. Kapur: Int. J. Miner. Pro-
From the table, it can be seen that after the system was cess., 57 (1999), 43.
put into used, the average value of burn through temperature 8) J. Khosa and J. Manuel: ISIJ Int., 47 (2007), 965.
9) X. L. Lu and B. C. Xie: Data Mining Technologies and Applications.
was rise and the too low phenomenon was eliminated; the China Renmin University Press, Beijing, (2008), 2.
burn through point moved ahead and avoided the too lag sit- 10) Y. Otsuka, M. Konishi, K. Hanaoka and T. Maki: ISIJ Int., 39 (1999),
1047.
uation. Compared with the manual control, the burn through 11) M. Dutta, A. Mukhopadhyay and S. Chakrabarti: ISIJ Int., 44 (2004),
temperature and burn through point were more stable, the 129.
standard deviation was reduced by 25.36 and 0.42, with a 12) J. L. Zhang, W. C. Cao, H. W. Guo, B. X Su and T. Zhang: Metall.
Ind. Automat., 36 (2012), 6.
reduction of 47% and 34% respectively.

2117 © 2016 ISIJ

You might also like