Documente Academic
Documente Profesional
Documente Cultură
n
s1
j=0
w
s1
ij
y
s1
j
(1a)
and
y
s
i
= f (x
s
i
); (1b)
where y
s
i
is the output of the neu(s; i); n
s
is the
total number of neurons in the sth layer, w
s
j;i
is the
weight from neu(s; i) to neu(s 1; j), and f () is a
nonlinear activation function.
With BP algorithm, it performs the input to output
mapping by minimizing the following cost function for
each given pattern pair (U
k
; Y
k
):
E
k
=
1
2
n
M
i=0
(y
i;k
^ y
i;k
)
2
; (2a)
where ^ y; y are the computed output and the real output,
respectively,
E =
N
k=1
E
k
; (2b)
where N is the number of training samples. The amount
of change for connection weight w
s
i; j
at iteration time t
can be determined by the following equation:
Dw
s
ij
(t) = m
@E
@w
s
ij
aDw
s
ij
(t 1); (3)
where m is the learning-rate coecient and a is a
momentum coecient, m; a > 0
@E
@w
s
ij
= d
s
i
y
s1
j
;
d
s
i
=
(y
i
^ y
i
)f
/
(x
s
i
) if s = M;
n
s1
k=1
d
s1
k
w
s
ki
_ _
f
/
(x
s
i
) if 2 _ s _ M 1:
_
The standard BP algorithm, however, frequently causes
the local minima problem and/or very slow learning
owing to the typical handicaps of all steepest descent
approaches. There are many improved learning al-
gorithms that have been proposed, such as using
heuristic rules to nd optimal parameters, adopting
conjugate gradients, Newtons method and quasi-New-
ton techniques, etc. (Chen and Xu, 1998; Ergezinger and
Thomsen, 1995). From Eq. (3), we can nd that the
convergence rate lowers dramatically when the deriva-
tive of the activation function f
/
(x) is close to zero.
Suppose that the activation function f (x) is a sigmoidal
function, that is,
f (x) =
1
1 exp(x)
(4)
and its derivative takes the form
f
/
(x) =
exp(x)
(1 exp(x))
2
= f (x)(1 f (x)): (5)
When f (x) ~ 0 or 1, then, f
/
(x) ~ 0, and the standard
BP algorithm may converge to a local minimum. In this
case, to escape from a local minimum, it is important to
keep f
/
(x) ,= 0. According to the above analysis, we try
to promote f
/
(x)to a special value b at some points, that
is, we dene a new activation function f
1
(x) and its
derivative takes the form
f
/
1
(x) = f
/
(x) b; (6)
where b is a small number and b > 0 (e.g. 0.0050.025).
From Eq. (6), the value of f
/
1
(x) will never be zero. If b
is equal to zero, it is obviously that the improved
algorithm will degenerate to the standard BP algorithm.
If b is small enough, it is believed that the improved
algorithm may have the properties of the standard one.
The training procedure is the same as the standard BP
algorithm. The performance comparison between the
improved algorithm and the standard BP algorithm is
described in Chen and Xu (1998).
In general, there are some shortcomings with ANNs,
such as (1) ANNs models are dicult to interpret; (2)
ANNs contain more parameters to estimate than do
most statistical models, which may result in overtting.
Fig. 2. Illustration of an ANN.
J. Chen / Engineering Applications of Articial Intelligence 14 (2001) 7785 79
To improve the performance of ANNs models, in this
paper, we integrate an ANNs model with qualitative
analysis in a predictive system and expect to increase its
predictive accuracy.
4. Qualitative reasoning
Qualitative reasoning is a relatively new eld studied
originally from AI research and focuses on using
incomplete knowledge. It appears to be an appropriate
or even a necessary approach for complex systems (e.g.,
a blast furnace) where complete numerical information
for problem under study is not available at the time of
analysis. The basic procedure of qualitative reasoning is
to obtain system structure, i.e. components and con-
nections among them that are described by qualitative
equations or by causal constraints, then to solve these
equations or analyze the constraints, and all possible
future behaviors will nally be explored through reason-
ing. They provide greater expressive power for incom-
plete knowledge than quantitative methods. Over two
decades qualitative reasoning has developed and
achieved much both in theory and practice. Three main
techniques of qualitative reasoning have been proposed:
De Kleer and Browns ENVISION (De Kleer and
Brown, 1984), Forbus qualitative process theory (For-
bus, 1984) and Kuipers QSIM (Kuipers, 1986).
Although other interesting approaches have been
proposed, our work is based on Kuipers QSIM method,
in which the relationships between system variables are
represented by equation-like constraints. It allows us to
represent profound knowledge very easily and provides
the tree of all possible behaviors of the system studied.
The structure of a system is described by a set of
parameters linked by constraints. For each instant t, the
qualitative state of a parameter QS( f ; t) is dened by
the pair (qval, qdir), where qval stands for the quali-
tative value and qdir for the direction of change,
therefore the whole system qualitative states is a set of
the qualitative states of all the parameters. In the
simplest case, we can determine qval and qdir as follows:
qval =
if f (t) > 0;
0 if f (t) = 0;
if f (t)50;
_
_
_
qdir =
inc if f
/
(t) > 0;
sdt if f
/
(t) = 0;
dec if f
/
(t)50;
_
_
_
where inc, std and dec represent increasing, steady and
decreasing, respectively, and f
/
(t) denotes the time
derivative of f .
The concept of general qualitative algebra has been
introduced as the theoretical fundamentals. In the
simplest case, we can dene U = (; ; 0; ?) as the
universe of qualitative calculus, which has a lattice
structure with respect to set inclusion.
In this case, however, qualitative reasoning ap-
proaches work well only in a highly abstracted level.
When the size or complexity of the problem increases,
qualitative reasoning usually renders uninteresting
results, i.e. the results are too general to draw any
useful conclusion. In order to solve that, a more rened
qualitative universe should be represented by landmark
set. A landmark point is (1) a critical point that the
objective function or constraint equation reaches, (2)
stage segmentation point that the problem solver are
interested in. With rened landmark points, let us dene
it as a set L = L
j
; j= 1; . . . ; m, we can extend the
denition of qualitative value to
qval =
L
j
if f (t) = L
j
;
(L
j
; L
j1
) if L
j
5f (t)5L
j1
:
_
QSIM has the ability of asserting new landmarks during
simulation [11]. Generally, seven basic types of
constraints are used in QSIM: Add (addition), Minus,
Mult (multiplication), M
; M
n(m2)
_
1, where s; n and m are the number of
hidden, input and output units (in our model, m = 1).
The blast furnace is a dynamic system. To adapt to
the change of the process, the sample set will be
updated when new data are available. The number of
samples in the sample set for training the model is 50
and the oldest sample of the sample set will be
eliminated when a new one is available and it will be
added to the sample set. Then, an improved BP
algorithm mentioned in session 3 is used to train the
model. When training error E (dened by Eq. (2b)) is
less than a given value e = 0:005, the training process
stops. The iteration number for training the model is
between 3500 and 5000 depending on how many
variables are selected as inputs.
*
Predicting silicon content of next heat: After the
model is trained, it is used to predict the silicon
content of next heat with proper input data.
*
Evaluating performance: The predictive value will be
compared with the observed one when it is available.
Consequently, according to the comparison, some
renements will be made for both qualitative model
and the ANNs model. For the qualitative model,
some landmarks or hypotheses may be adjusted if the
bias is signicant. And the ANNs model will be
retrained with the new data. With the feedback from
the comparison, the system will be adapted to the
change of the process. Fig. 3. Overview of system structure.
J. Chen / Engineering Applications of Articial Intelligence 14 (2001) 7785 81
With the integration of qualitative analysis and
quantitative methods, we could expect to make full use
of the information/knowledge (both qualitative and
quantitative) and furthermore to improve the predictive
accuracy.
6. Qualitative structure of a blast furnace
In order to reason the behaviors of complex systems
qualitatively, it is necessary to build a qualitative model
for the system. The causal approach is widely used to
model systems qualitatively. Through causality, we can
focus on any part or component of a system, without
having to deal with the model as a whole. This is useful
when trying to understand and explain the behaviors of
a system. The causal inuences (or causal explanations)
between variables are usually represented as an oriented
graph, where nodes in the graph represent variables or
physical parameters, arcs, oriented from an input or a
causal variable to an eect variable, are representative
of the relationship between variables. The silicon
content of the pig iron is inuenced greatly by the
operating situations, which are strongly related with
various conditions including in-furnace heat balance,
burden descending, gas permeability and the likes. In
actual operations, operators of a blast furnace obtain
information relating to the blast furnace conditions
from many sensors. By combining all these pieces of
information (absolute values, relative values, amount of
changes and so on), they judge the type of the blast
furnace conditions at any given time.
After making a thorough investigation and study on a
blast furnace, the qualitative structure of the operating
process in a blast furnace is worked out in the form of a
causal graph. Fig. 4 describes a part qualitative structure
of #9 blast furnace in Anshan Iron and Steel Company
(AISC) that is the largest Iron and Steel Company in
China. According to the causal graph, a qualitative model
is constructed with qualitative equation-like constraints.
The initial landmarks (important values) of each
variable are determined based on the specic parameters
of the blast furnace by discussing with experts.
7. Empirical results
To empirically implement the system described above,
610 heats from #9 blast furnace in AISC are checked.
According to the operating conditions of #9 blast
furnace, the following variables are considered to be
the input variables of the neural network for predicting
silicon content in pig iron:
V
1
(t): the quantity of blast
V
2
(t): the temperature of blast
V
3
(t): the pressure of blast
V
4
(t): the quantity of coal powder
V
5
(t): the index of ventilating performance
V
6
(t): the pressure of top gas
V
7
(t): the temperature of top gas
V
8
(t): the quantity of raw material
V
9
(t): the coke ratio
V
10
(t): the additional quantity of coke
Here, V
1
(t); V
2
(t); V
3
(t); V
4
(t); V
8
(t); V
9
(t) and V
10
(t)
are control variables and V
5
(t); V
6
(t) and V
7
(t) are
state variables. Generally, V
1
(t); V
2
(t) and V
4
(t) are
most frequently used to adjust the operating process,
V
8
(t) is constant for some weeks. However, some
additional quantity of coke V
10
(t) is fed some times if
necessary. In #9 blast furnace of AISC, the time
constant (T) of the iron-smelting process is about 6 h
and 10 heats are produced every day, that is, about
144 min on average for producing a heat. According to
engineering experience (Wang and Fang, 1992), the
sampling period t can be determined as t =
1
6
~
1
10
_ _
T,
that is, the sampling period for #9 blast furnace of AISC
could be 3660 min. Considering the average time
interval between two heats is 144 min, we set the
sampling period t of 48 min. In this case, the interval
between two heats is exactly 3t. Our objective is to
predict the silicon content in pig iron of next heat (heat
t 1) based on the information of heat t, that means we
need predict 3 sampling steps ahead (k = 3), which
could help the operators to run the blast furnace with a
more ecient control strategy. With experiments and
analysis, the time delay indices of inputs are determined
as d
1
= 1; d
2
= 1; d
3
= 1; d
4
= 2; d
5
= 1; d
6
= 3; d
7
=
2; d
8
= 6; d
9
= 6; d
10
= 6. For example, under station-
ary condition, V
1
; V
2
and V
3
are selected as input
variables. Corresponding to y(t), the inputs are
V
1
(t 1); V
2
(t 1) and V
3
(t 1).
Once the qualitative states of the process in
blast furnace are derived through qualitative
analysis, the relevant input variables of the fore-
casting model can be determined, and the model will
be built automatically. With the model, silicon content
in pig iron will be predicted. Fig. 5 shows the
comparison between the predicted values and observed
data.
Input and output data are divided into two sets:
the initial training set which includes 50 samples, and
the test set that includes 610 samples. The initial
training set is used for learning the initial parameters
(weights) of the ANN model. The test set is used to
assess the performance of the system. The model will be
retrained when a new sample from test set is available
and the oldest sample of the training set will be
eliminated at the same time to keep the size of training
set. To evaluate the performance of the prediction,
some important criteria used in practice are considered
J. Chen / Engineering Applications of Articial Intelligence 14 (2001) 7785 82
as follows:
J
1
=
1
610
610
t=1
N
t
_ _
100%;
where
N
t
=
1 where y(t) ^ y(t=t k) [ [50:1
0 else;
_
y(t) is the observed value at time t and ^ y(t=t k) the
predicted value k steps ahead at time t k
J
2
=
1
610
610
t=1
(y(t) ^ y(t=t k))
2
;
J
3
=
1
610
610
t=1
(y(t) y(t=t k))
y(t=t k)
100%:
Generally, we say that the prediction hits the target in
the case [y(t) ^ y(t=t k)[50:1 and the percentage of
hits (J
1
) is the most important criterion for operating a
blast furnace. Table 1 gives the comparison of dierent
methods for predicting silicon content in pig iron.
It is always believed that to predict the tendency of the
process (i.e. the direction of change of the process) is the
most important guide for taking correct actions.
Consider those heats where prediction does not hit the
target with the criterion J
1
, we also counter the
prediction hits the target if the predicting tendency is
consistent with real one, that is, the sign of
(^ y(t=t k) ^ y(t 1=t k 1)) is the same to the sign
of (y(t) y(t 1)). With this criterion, the percentages
of hits for IS and KAP will increase to 93.6 and 89.7%,
respectively. In practice, sometimes the criterion J
1
is
changed as follows to meet higher requirement of
operation:
J
/
1
=
1
610
610
t=1
N
/
t
_ _
100%;
where
N
/
t
=
1 where y(t) ^ y(t=t k) [ [50:075
0 else
_
The percentages of hits for IS and KAP are 81.5 and
68.1%, respectively. From the results of comparison
with other methods, we can nd that the performance of
Fig. 4. Illustration of a part qualitative structure for a blast furnace.
J. Chen / Engineering Applications of Articial Intelligence 14 (2001) 7785 83
Fig. 5. Comparison between the predicted values and observed values.
Table 1
Comparison of dierent methods
Integrating
system (IS)
Combining knowledge-base
with an adaptive predictor
(KAP) (Chen, 1993)
Adaptive predictor
a
(Chen, 1993)
Experienced operator
(Chen, 1993)
J
1
(%) 88.7 81.3 68.9 62.0
J
2
0.0062 0.0082 0.0117 0.0214
J
3
(%) 14.3 16.3 19.3 24.9
a
For the detailed description of the method, please see Keyser and Canwenberghe (1981).
J. Chen / Engineering Applications of Articial Intelligence 14 (2001) 7785 84
the system presented in this paper is much better than
those of others.
8. Conclusion
In this paper, we present a predictive system for blast
furnaces to predict silicon content in pig iron. A blast
furnace is a very complex system, in which gas, liquid
and solids coexist. For a complex system, information is
very likely to be highly qualitative and it involves both
qualitative and quantitative aspects. Unlike traditional
methods, in this paper, we have designed a predictive
system by integrating a neural network with qualitative
analysis and expect to make use of both quantitative and
qualitative information. Six hundred and ten heats of
pig iron are used for verifying the performance of the
system. From the empirical results, it is evident that
excellent performance has been obtained, at least for the
test samples. It is very important that the predictive
system only makes a suggestion to help the operators to
make the right decision, that is, they can use it as a guide
or can ignore it and follow their own intuition. In this
sense, the more accurate the system performs, the more
helpful the system will be. We believe that to synthesize
qualitative reasoning with quantitative methods is a way
for treating complex problems. Further application of
this system to expand areas of operations is highly
expected. To improve the performance of the system
furthermore, we need to acquire more deep knowledge
about the process of blast furnaces, and explore ecient
methods for representing and processing qualitative
knowledge as well. On the other hand, eorts should
also be made toward a closed loop operation of blast
furnace heat control beyond the stage of guidance levels.
Acknowledgements
The author thanks the anonymous referees and the
editor for their valuable comments and suggestions. This
work is supported partly by the National Science
Foundation of China, Natural Science Foundation of
State Education Commission and National Defense
Foundation.
References
Chen, J., 1993. Predicting system based on combining an adaptive
predictor and a knowledge base as applied to a blast furnace.
Journal of Forecasting 12 (2), 93102.
Chen, J., Xu, D., 1998. An improved back propagation algorithm.
In: Gu, J. (Ed.), Proceedings of Systems Science and Systems
Engineering. Scientic and Technical Documents Publishing House,
Beijing, pp. 408411.
De Kleer, J., Brown, J.S., 1984. A qualitative physics based on
conuence. Articial Intelligence 24, 783.
Ergezinger, S., Thomsen, E., 1995. An accelerated learning algorithm
for multilayer perceptrons: optimization layer by layer. IEEE
Transactions on Neural Networks 6, 3141.
Fielden, C.J., Wood, B.I., 1968. A dynamic digital simulation of a
blast furnace. Journal of Iron and Steel Inst. 206, 650658.
Forbus, K.D., 1984. Qualitative process theory. Articial Intelligence
24, 85168.
Fouche, P., Kuipers, B.J., 1992. Reasoning about energy in qualitative
simulation. IEEE Transactions on Systems, Man and Cybernetics 22
(1), 4763.
Iida, O. et al., 1989. Blast furnace control by articial intelligence.
In: Rodd, M.G. Suski, G.J. (Eds.), Articial Intelligence in Real-
Time Control. Swansea, UK, pp. 7378.
Jin, G., 1986. A mathematic model for predicting the Si-content and
temperature of pig iron. Iron Smelting 5 (2), 816.
Keyser, R., Van Canwenberghe, A., 1981. A self-tuning multistep
predictor application. Automatica 17, 167174.
Kuipers, B., 1986. Qualitative simulation. Articial Intelligence 29,
289338.
Li, T., et al., 1984. Self-learning algorithm of fuzzy semantic inference.
Acta Automatica Sinica 10, 329336.
Liu, Y., 1985. The inuence of the relation between the temperature
and Si-content of pig iron by technique and melting condition. Iron
Smelting 4 (4), 5557.
Nakajima, R., 1987. Operation control system of blast furnace by
articial intelligence. AIME 46th Iron-Making Conference Proceed-
ings, pp. 155157.
Rumelhart, D.E., et al., 1986. Learning internal representations
by error propagation. In: Rumelhart, D.E., McClelland, J.L.
(Eds.), Parallel Distributed Processing: Explorations in the Micro-
structures of Cognition, Foundations, Vol. I. MIT Press, Cam-
bridge, MA.
Say, A.C.C., 1998. LHopitals lter for QSIM. IEEE Transactions on
Pattern Analysis and Machine Intelligence 20 (1), 18.
Unbehauen, H., Diekman, K., 1982. Application of MIMO identica-
tion to a blast furnace. Sixth IFAC Symposium on Identication
and System Parameter Estimation, pp. 235240.
Wang, Y., Chen, J., 1995. A framework of decision analysis integrating
qualitative reasoning and quantitative models. Proceedings of 1995
IEEE International Conference on Systems, Man and Cybernetics,
Vancouver, pp. 346450.
Wang, J., Fang, C., 1992. In: Computer Systems for Process Control.
Tsinghua University Press, Beijing, p. 37.
Jian Chen received the B.Sc. degree (1983) in Electrical Engineering,
M.Sc. degree (1986), and the Ph.D. degree (1989) both in Systems
Engineering from Tsinghua University, China. He is now a Professor
and Chairman of Management Science Department, Tsinghua
University. He serves as a member of the Administrative Committee
of IEEE systems, man and cybernetics society, a member of the
Standing Committee of Systems Engineering Society of China. He has
over 80 technical publications and has been a principal investigator for
about 20 grants or research contracts. His main research interests
include modeling and control for complex systems, forecast and
optimization techniques.
J. Chen / Engineering Applications of Articial Intelligence 14 (2001) 7785 85