AC-26,NO. 4,
AUGUST
885
1981
Short Papers
A Digital Quality Control System for an Industrial
Dry Process Rotary Cement Kiln
TAP10 WESTERLUND, MEMBER, IEEE
A bstract-A multivariate autoregressive moving average (ARiMA) model
for an industrial dry process rotary cement lciln is identified, from real
process data, using the maximum likelihood method.The model obtained is
then used in computing a controller for quality control of clinker production. Itis shown that it is relevant to compute a minimum variance
controller subject to restrictions both in the controller structure and the
variances of the control signals. The resulting controller is finally implemented on the cement kiln process, together with a target adaptive controller for automatic adjustment of the clinker quality setpoint, in order to
save energy.
11 I
t
I
I. INTRODUCTION
In the last few years methods of saving energy have become ever more
important. In the cement industry one of the most important steps in this
direction has been the change from the so-called wet method to the dry
method. It has therefore been considered worthwhile to investigate different control strategies for dry process cement kilns [8], [IO]. [13], [18]-[20].
The present paper describes how a digital control system has been
implemented on a dry process cement kiln. The cement kiln has a capacity
of 1000 ton clinker/day and belongs to the company Oy Lohja Ab
located in Virkby, Finland. The actual cement kiln was previously manuaUy controlled by two operators, although the kiln process also contained
some efficient conventional control loops. Today the cement kiln has a
digital control system, realized on an Intel SBC 80/20 microcomputer,
which has been in continuous on-line operation since the summer of 1980.
The digital control system mainly consists of a multivariable "minimum
variance" controller and a target adaptive controller for automatic adjustment of the free lime content setpoint. Improvements both in production
time and fuel economy have been established without deterioration in
product quality. Some of these results are presented in the present paper.
The paper is organized as follows. In Section I1 a brief description of
the cement kiln process is given. In Section 111 the process model and the
identification method are discussed. Results from the maximum hkelihood identification are presented in Section IV. In Section V the controller design is discussed. Finally, in Section VI practical results from the
implemented control system are presented.
11. T H E CEMENT
KILN
.-
--
(I.,,
I
- _ - _ I
___..____.._
At-?*
Il.:,ccOm,iL,
Ylltl
J2111
Fig. I .
order to start the endo- and exothermic chemical reactions taking place in
the raw meal. The product is called clinker and is a complex chemical
compound of mainly dicalciumsilicate, tricalciumsilicate, tricalciumaluminate, and tetradciumaluminofenite. Finally, the clinker is c o o l e d in
a planetary cooler before it leaves the cement kiln process. The quality of
the clinker is specified by the free lime content in the clinker. The
disturbances that occur in cement kilns are mainly of a stochastic nature.
It therefore seems natural to use a stochastic approach in modeling the
process. This has also been discussed, for example, by Otomo er al. [ 151in
modeling a wet process cement kiln.
MODELAND IDENTIFICATION
METHOD
111. PROCESS
In this section we give a brief description of a general linear stochastic
process model and a method for estimating the model parameters from
real data
Consider a linear discrete-time multivariable stochastic process model
described by
PROCESS
+C,e(r)+ ...+C,ce(t--n,+l)+e(t+l)
Committee.
Tbe author is with the Department of Chemical Engineering at
(1)
(2)
Manuscript received December 19, 1979; revised December 9, 1980. Paper recommended
by D. G . Fisher, Past Chairman of the Applications, Systems Evaluation,and Components
:,,n*cr
0018-9286/81/0800-0885300.75
. . .+A,"Z""
B ( z ) = B O + B l z + . . . +B",Z.B
C ( z ) = l + C , z + . . . +cncznc.
A(z)=I+A,z+
0 198 1 IEEE
(3)
886
A(q-)y(r+I)=B(q-)u(r)+C(q-)e(r+l).
(5)
1
-
7.5
Parameter Estimation
Consider some input-output pairs { u( t ) . y ( t ) , t = 1,2, ...,N } generated
by the real process. The identification problem can then be understood as
a problem of finding the best, in respect to some specified criterion,
parameters of the model (I). One of the most commonly used identification methods is the maximum likelihood method. Using the model
structure ( I ) it can be shown that maximization of the likelihood function
is equivalent to minimization of the loss function [3], [6]:
V=det-
I
N-1
2N t ? ( ~ ) B ( t ) ~
(6)
r=l
AIC=(W-l)Iog(V)+(I-n,)(N-I)log(N-l)
f n , ( N - 1)(1 + l o g ( 2 r ) ) + 2 n p
(7)
IDENTIFICATION EXPERIMENT
The cement kiln was manually controlled before the actual investigation, although there were also Some efficient conventional control loops,
The purpose of the identification experiment was therefore not to obtain a
of the totally uncontro~edkiln but the partly controlled kiln as it
behaved for the operator.
Two of the primary manipulated variables in the rotary kiln are the fuel
rate and the total exhaust gas rate, each of which has an important effect
on the heat transfer, and, indirectly, the reaction characteristics [9]. In the
actual kiln where coal is used as fuel, the oxygen concentration in the
exhaust gas is controlled by the fuel rate using an efficient conventional
control loop. The manipulated variable is thus the kiln exhaust fan speed,
which then corresponds to the energy flow into the kiln. This variable was
taken as the first control variable U l ( t )in the model to be identified.
The conventional manipulated variables in rotary kiln operation also
include the kiln rotation speed, which controls the material residence
time, and the feed rate of the raw material to the kiln, which controls the
production rate [9]. In the actual kiln the kiln speed is synchronized with
the raw material feed rate in order to make it possible to manipulate the
feed rate of the raw material without sigmficant changes in the internal
loading. This variable was taken as the second control variable U 2 ( t ) .
Although
- there also exist other possible maniuulated variables in the kiln
operation, the above mentioned variables are the most important ones in
the actual kiln. It was therefore decided to restrict the identification
experiment and the model to these inputs.
The identification experiment was made in June 1979. It was planned to
use PRBS in exciting the process. For practical reasons it was. however,
not possible to fulfill the PRBS experiment without any manual adjustments during the excitation procedure. The excitation of the process was
carried out during an 8 h long period (see Fig. 2) and instruments were
read every fifth minute. Samples of the clinker were also taken every fifth
minute and then ground and analyed by wet chemical methods in order
to get the sampled free lime content in the clinker during the experiment.
Ideal kiln operation will minimize fuel consumption while achieving the
desired production rate of quality clinker [9]. I n order to achieve the
20
40
60
80
100
t
Fig. 2. Input sequences during the identificationprocedure.
log(Y3(t+20min))=aY2(t)+b+e(t)
(8)
could adequately describe the relationship betrveen the kiln drive power
Y2(r) and the free lime content in the clinker Y3(r).It was therefore found
relevant to use the kiln drive power as one of the outputs in the model to
be identified, as it is easy to measure.
The numerical values of all the variables otherthan the free lime
content in the clinker (weight percent) and the exhaust gas temperature
(percent of the full scale O-lOOOC) are in the following given in the
absence of any units, primarily because their physical interpretation have
been changed, as described previously in this section.
o d l Structure
When it was decided which variables should be the inputs and the
outputs of the model to be identified. the next problem was to estimate
the model parameters, It was decided to use the
St~cture
A(q-)Y(r+l)=B(q-)U(r)+c(q-)e(t+l)+d
(9)
887
-1
'2
' [ I " ,
'
'
1 1 ,
-1
10
15
20
At
Fig. 4. Sample autocorrelation functions for the prediction errors illustrated in Fig. 3.
-4.0
L
tion calculated with the model (IO). Thirdly, the deterministic output from
the model is illustrated. The deterministic output was calculated from the
model
.
0
Y(t+I),=A,Y(t),+B,U(t)+d.
' "
20
"
'
"
60
40
80
' 100
' 1
Fig. 3.
(IO)
where
A,=
( 12)
[ -0.126
0.914
0.0800]
0.917
Bo=
[ -0.211
2.091
-0.0156
R="
'
(Y(f)-Y(f))(Y(r)-f(t)y
N-1
(11)
*=I
[0000257
~ 1 = ~ n ~ ( ( ~ ~ ( ~ ) - ~ ~ . ~ p ) * + ( l o g ( ~ ~ ( ~ ) (13)
)-~og(~3.s
A= O Y 4 4
V. CONTROLLER
DESIGN
0.000257]
0.0214
.
( 14)
ontrol
888
AC-26,NO. 4,
1981
AUGUST
L
Y
Fig. 5.
5 XC.0
5 XCaO
The frequency function for the free Lime content in the clinker during the
manually controlled period.
1.5
1.0
0.5
0
-1
1 log (IC.0)
-1
1 lag (XCaO)
Fig. 6. n e frequency function for the logarithmic free lime content in the clinker during
the manually controlled period.
Fig. 7. The frequency functions for the free Lime content in the clinker during the period
when the quality control system was in on-line operation.
U(f)=
[ -0.490
-1.49
2.491 y ( t ) ,
71.1
EuI(f)=0.00185
(15)
E~,(t)=0.0644
Ey2(r)2=0.0214
(16)
E u 2 ( t ) 2 = 108.
(17)
= mintr( Q, c
subject to
control signals for the system considered. In this case the variances of the
output signals are, however, close to that of the open-loop system. The
variances of the outputs of the open-loop system, as in (IO), are
,Ey,(t)=0.375
cy2(
t)=0.476.
(19)
J=-(tr(Q,c,)+tr(Q,P,))
y,(t)*=0.347
cy2(
t)=0.426.
(20)
(25)
where the weighting matrix Qu has been given the nice interpretation
889
Q,=
Eul(f)=0.004
2 hisi
Eu2(t )2 = 1.5.
i= I
(35)
it thus
EXPERIMENTAL RESULTS
a,=[
890
results the successful maximum likelihood identification and the controller design, using constrained optimization. should be mentioned.
REFERENCES
H Akaike. A new look at statistical model identification. I E E E Tram Auromat.
Contr.. uol. AC-19. pp. 716-722. 1974.
K J. htram. Introductron to Srochasrrc Conrrol Thheoq. lieu York:Academic. 1970.
K. J. Astram and P. Eykhoff. System identification. A surne!. Auronroncu. vol. 7.
pp. 123-162. 1971.
G. P. Box and G . M. Jenkins. T m e Serres Anohsrs. Forecasrrrrg and Conrrol. San
Francisco:Holden Day. 1970.
D. F. Church. Current and projected pulp and paper indust? problems In process
control and process modeling. Chenural Process Conrrol: AIChE Srmp Senes. vol.
72.
P. Eykhoff. s y s m n Idenrrficafion. New York:Wiley. 1974.
W . M. Gale and C. A. Stapelton. Experimental verification of the correlation
between the calcining rate and other kiln variables, 5th IFAC World Congress.
Paris. France.
J. D. Higham. Dynamic Computer regulation of a dry-process cement kiln. Pror.
I E E . vol. 118. pp. 609-617, 1971.
V. A. Kaiser. Computer control in the cement industry, Proc. IEEE. vol. 58. pp.
70-77. 1970.
L. Kevicky. J. Hetthessy. M. Hilger. and J. Kolostori. Self-tuning adaptive control
of cement raw material blending. Auromatica. vol. 14, pp. 525-532. 1978.
L. Lpng. On Positive real transfer functions and the Convergence of some recursive
schemes. I E E E Truns. A u r o M . Contr., vol. AC-22. pp. 539-551. 1977.
L. Llung. Convergence analysis of parametric identification methods. I E E E Tram.
Automat Conrr.. rol. AC-23. pp. 770-783, 1978.
A. Lundan and 0. Mattila. A system for the control of the homogenization of the
cement raw meal, 4th IFAC Symp. Digital Computer Applications Process Control,
Zurich, Switzerland. 1974.
P. M W i Optimal structure constrained controllers, Process Control Laboratory.
i\bo Akademi. Finland. Rep. 79-1 1.
T. Otomo, T. Nakaga*-a. and H. Akaike, Statistical approach to computer control
of cement rotary kilns. Autoornorm. vol. 8. pp. 35-48, 1972.
L. K. Rastogi. A flat die extmder control system using distributed ~crocomputers.
in Proc. 1978. Joint Automat. Conrr. Conf., 1978.
H. Toivonen T. Westerlund. and P. M M & On the applicatjon of stochastic control
theory for process control, Praffs Control Laboratory, Abo Akademi. Finland.
Rep. 80-10.
T. Westerlund, Application of stochastic control theory to the cement manufacturing process. Process Control Laboratory, k o Akademi. Finland, Rep. 80-5.
T, Westerlund, Oy Lohja Ab, Virkby Finland. Internal Rep. 79-2.
T. Westerlund, H. Toivonen, and K-E Nyman, Stochastic modeling and self-tuning
control of a continuous cement raw material mixing system, in Proc. 18th I E E E
Canf. Decrsian Contr.. Fort Lauderdale. FL, 1979.
T. Westerlund. Identification and controlof an industrial d q process cement kiln,
1980 Joint Automat Contr. Cod., San Francisco. C A
AC-26, NO.
4 AUGUST 1981
Remark: Process noise effects are not discussed here because the
mapping decouples from the process noise inclusion. Bierman [2] and
Thornton and Bierman [7] quite adequately handle the effects of process
noise using rank-one matrix modification.
One direct approach that has been used to construct
and
is to
compute Z= W D W 7 (which could take advantage of possible W structure) and then to factor Z using a Cholesky type reduction [2]. Unfortunately, this procedure is both costly in operation counts (compared to
the solution to be presented) and, while the Cholesky reduction is
numerically stable, the X computation can contribute an information loss
due to round off. In problems where high accuracy is crucial. or where
there is poor observability. such an information loss is unacceptable. To
illustrate this point. let c be positive and small enough so that ( I +z2)= = 1
and consider
W=(
I ti:D=Diag(l.l).
);
Then
m.
having eigenvalues of 1 2
If we were to compute the U-D factors
of Z, the way that the Cholesky algorithm computes them, one obtains
the results indicated in the third column of Table I, and it is clear from
G. J. BIERMAN, SENIOR MEMBER, IEEE
the negative computed Dl that a negative eigenvalue has been introduced.
Abstract-Time propagation of the Kalman filter covariance matrix As is illustrated by the Table I results, information lost in constructing
invokes an operation of the form CPPB. For many applications CP has a cannot be recovered by accurately computing the Cholesky factors of Z,.
Canear: Factorization filter algorithms transform (time or measureblock triangular structure. When the filter implementation employs U-D
covariance factorization (i.e., recursions for U and D are used, where ment update) matrix factors, and good algorithms never explicitly
P= UDU, with U unit upper triangular and D diagonal) this matrix construct covariance matrices as intermediate results. The point to be
structure can be used to reduce the computation cost of time propagation. aware of is that even with D strictly positive one can. in ill-conditioned
The contribution of this paper is a streamlined Gram-Schmidt orthogonali- problems. compute a Z with negative eigenvalues. as in (3). Incidentally,
zation algorithm that can dramatidy reduce U-D time update computa- (LrDL) diagonals always majorize the corresponding D entries: one is
thus assured that, at least. the computed variances must be positive.
tion costs.
In this paper Bre discuss computation of 6and 5 directly from W and
D. One reliable way to do this is to apply the Gram-Schmidt procedure
I. INTRODUCTION
described in [2], [7]. and [8]. When @ and consequently W have arbitrary
Time propagation of the Kalman filter covariance matrix involves an structure in regard to the number and placement of nonzeroqntjes, then
operation of the form Z=OPCP7, where P is the filter covariance matrix as noted in [7] and [8], the computation of @Pa7and U-D via the
and CP, the state transition matrix, often has a block triangular structure. Gram-Schmidt process requires on the order of 1.7n3 and 1.5n3 flops.*
In Kalman filter applications involving relatively few measurements per respectively. Unless n is large, the difference in computation times is
time step, the computational cost of the transition propagation is of hardly discernible.
In many Kalman filter applications. particularly aerospace applications.
central importance [l]. It is, therefore, important to take advantage of
transition and dynamics structure in problems where the filter dimension problems that have large dimensions have sparse transition matrices.
Partitioning and using sparse arithmetic code dramatically reduces the
is large or when there is a (near) real time processing constraint.
Manuscript received April 14. 1980 revised February 5. 1981. Paper recommended by
A. J. Laub. Chairman of the Computational Methods and Discrete Systems Committee.
The author is with Factorized Estimation Applications. Inc.. Canoga Park, CA 91307.
0018-9286/81/0800-0890$00.75
a1981 IEEE
____