Anda di halaman 1dari 10


A Novel Adaptive Control System for Raw Material Blending

By Cs. Bányász, L. Keviczky, and István Vajk

ement is a fundamental commodity in any industrial These materials are purchased or quarried and then
society. As a key ingredient in concrete, it is the transported to the cement factory by road, railway, or con-
world’s most widely used building material. With ap- veyor. The raw materials are crushed, usually in the quarry
plications ranging from residential and commercial buildings but sometimes in the factory, to a size that the machinery
to key infrastructure elements such as roadways and bridges, can handle or grind.
it is hard to imagine life without it. Worldwide cement produc- The prepared raw materials are stored in silos, ready for
tion totals some 1.5 billion tons annually. Such large volumes the next production step (i.e., the material grinding and
place it among the top ten power consumers in many coun- blending operation). In this step, which can be either a wet
tries and, along with the precision required for raw materials or a dry process, ball mills are used to crush and blend the
blending, make it a prime candidate for automation. materials in varying proportions according to the required
The techniques of cement production have been known quality of the final product. The materials are proportioned
for centuries and are discussed extensively in the literature. using a control system to ensure the desired chemical com-
The main problems with automation of cement production position of the final ground mix.
are also well known, as discussed in several survey papers Adding water to the raw materials makes them easier to
(e.g., [1]-[3]), and thus it is difficult to introduce new ideas. grind and mix (grinding can also be done separately), but a
We believe, however, that recent advances in computational major disadvantage of the wet process from an energy
power have considerably increased the possibility of apply- standpoint is that the water later evaporates. Thus, we will
ing certain modern control theory methods and algorithms

Quarrying and Preparation

to the automation of cement production.
In this article, we discuss a new generic optimal control-
ler structure for raw material blending in the cement indus- Crushing
try. We focus on an important phase of the proportioning → Storage
burning → grinding operation triplet that essentially deter- Purchasing Crushing
mines the cement quality; namely, the composition control
of the raw mill system. We begin by describing the applica-
tion of the generic optimal controller structure introduced
in [1] to this important phase, followed by a discussion of

Raw Material
Raw Mill
the control engineering background and the design of the re-

alized controller. Finally, the control algorithm is discussed
in a technology-independent manner.

Description of the Technology

The main elements of cement production are shown in Fig- Preheater

ure 1. Some of the blocks shown may be missing in a given


Rotary Kiln
factory, but those denoted by thick lines are found in most
cases. Cooler
The raw materials for cement production are selected
based on their chemical composition. The primary compo-

Cement Mill
nent is the calcium oxide (CaO) source, which is usually
limestone, but chalk, marl, dolomite, and oyster shells can
also be used. The next two components are silica (SiO2) and

alumina (Al2O3), obtainable from clay or shale. Quartz and
bauxite or other minerals can also be used. The fourth, and
last, component is ferric oxide (Fe2O3), which can be ob-
tained from iron ore, pyrite, or blast furnace slag. Figure 1. Typical basic technological operations for a cement
Keviczky ( and Bányász are with the Computer and Automation Research Institute, Hungarian Academy of Sciences,
H-1111 Budapest, Kende u 13-17, Hungary. Vajk is with the Department of Automation and Applied Informatics, Budapest University of
Technology and Economics, H-1111 Budapest, Hungary.

February 2003 IEEE Control Systems Magazine 87
concentrate on the dry process. In this process, cement is • silica modulus
made by baking (at about 1450 °C) and grinding the homoge-
neous raw materials mixture with a prescribed oxide con-
centration. The continuous process is divided into serially S
MS =
connected subprocesses by using large material storage ar- A+ F
eas and applying several parallel elements to increase pro- (3)
duction capacity. Together, these large material storage
areas and parallel working elements ensure the safe and eco- and their possible linear combinations. (Here, the oxide
nomic operation of the overall cement plant. contents are given in percentages.) The most difficult prob-
Control of the chemical composition of the raw meal is lem of composition control is the measurement of oxide
necessary because the relative amounts of dicalciumsili- compositions having a considerable time delay in the closed
cate (C2S), tricalciumsilicate (C3S), tricalciumaluminate control loop. Its value is so large (30 minutes to one hour)
(C3A), and tetracalcium-aluminoferrite (C4AF) formed in the that many experts expect greater improvements in the qual-
kiln strongly depend on the oxide composition of the ground ity of control due to decreasing the delay than to applying
mix. (For simplicity, the four most important oxides, CaO, the most sophisticated algorithms. Earlier, the chemical
SiO2, Al2O3, and Fe2O3, are denoted by C, S, A, and F, respec- analysis was made by laboratory tools using material sam-
tively.) The purpose of the control is to maintain the relative pled by hand. Recently, automatic sampling has been intro-
proportions of these oxides. The relative proportions can duced, and an X-ray fluorescence analyzer (RFA) typically
be expressed by their so-called modulus values. The most provides the oxide compositions.
widely used ones are: The complexity of the control problem of raw material
• lime standard (or modulus) proportioning depends on the number of materials neces-
sary to blend the desired chemical composition. In many ce-
100C ment factories, the most important three to five basic
ML =
2.8 S + 11
. A + 0.8 F (1) materials are used; however, there are cases where eight to
ten different materials are to be mixed. The achievable qual-
• aluminum modulus ity specifications of this control depend on changes in the
basic material compositions. In the best case, there is no
A need for control. In many cases, however, the makeup of the
MA =
F (2) raw materials is so variable that only computer control can
achieve the desired composition. It is no accident that the
Feeder Tanks first computer control appeared at this point in the process.
The simplified scheme of a raw material blending sys-
Component nof
Component 1

tem is shown in Figure 2. This is the case of the three most

... often applied basic materials. The feeder silos contain raw
materials of different composition. The weigh feeders are
controlled by a computer. A conveyor belt feeds the rubble
... Weighs into the raw mill. Next, the ground mix of raw materials
(Feeders) (raw meal) enters the silo. Before the meal enters the silo, a
sample is analyzed by an RFA that provides the oxide com-
positions for the computer. The computer controls this au-
Ground Meal tomatic sampling, conveyance, preparation, and analysis
Raw Mill
process and computes the new set points (scale factors)
Sample for the weigh feeders using the modulus values as refer-
ences. Hereinafter, only the composition control itself will
be treated.

Control Engineering Model of the Process
Computer A simplified control engineering discrete-time model of the
mill-silo system is shown in Figure 3 (see the details in [5]),
where r ( t ), v( t ), and q ( t ) are the input feed, the mill outlet
Reference Values (silo inlet), and the silo content, respectively. The vectors ox
and M denote the oxides and modulus values; w stands for
Figure 2. Simplified technological scheme of a raw material the scale factors of the weighs (feeders), and C is the compo-
blending system. sition matrix. The linear noninteracting (diagonal) dynamics

88 IEEE Control Systems Magazine February 2003

 gi z − di  methods and algorithms published in the past decades. The
Gm ( z −1 ) = diag K , −1
,Kz − d
 1 + (1 − g i )z  large number of papers indicates that no unique or best
≈ diag{K , z − d i ,K}
method has been found. The solution depends on the
model, criteria, uncertainties, disturbances, etc. (some-
≈ z −dI times even on the designer’s preference).
A control system is internally stable if the bounded sig-
for the partial material flows is generally acceptable for the nals injected at any point of the closed loop generate
mill, and the first-order unity gain time lag approximation bounded responses at any point. Thus, a linear time-invari-
with delay has proved to be adequate in most practical ant, discrete-time control system is internally stable if the
cases. Because of the relatively large sampling interval transfer functions between any two points of the closed
( h = 20 - 30 min ) necessary for automatic material sam- loop are stable (i.e., have all poles within the open unit disc).
pling, conveying, and the RFA and the typically applied aver- The Youla- (Y or Q) parametrization is a classical method for
aging measurement technique, different time delays characterizing all realizable stabilizing (ARS) regulators by
without any lag effects can also be used, and the simplest ap-
proximation is a common time delay block. Here, distinct R=
channel time delays z − d i will be used with additional dynam- 1 − QS (6)
ics caused by the noninteger discrete delays (modified
z-transform). for the open-loop stable plant S ∈ S, where S is the closed
The input and output vector variables rinp ′ of this
′ and rout set of all stable proper real-rational systems having all poles
noninteracting internal multivariable linear segment are the within the closed unit disc. The “parameter”
partial material flows (lime, clay, pyrite). Models for the ox-
ides are generally coupled because C is not diagonal. The Q= ; Q ∈S
dynamics of a batch silo can be easily derived and is given 1 + RS (7)
by a simple integrator
ranges over all proper (Q(ω = ∞ ) is finite), stable transfer
functions [6].
z −1
Gs ( z −1 ) = I, Using the K-B-parametrization [7], shown in Figure 4, the au-
1 − z −1 (5) thors introduced a generic two-degree-of-freedom (G2DF) con-
trol system, provided that it virtually opens the closed loop
where I is the unit matrix [5]. The vector rsilo′ represents the
integrated (accumulated) partial material flows. Gs is some- y = Qr Sy r − (1 − QS )n = y t + y d (8)
times called silo integration. The blocks denoted by NL refer
to the nonlinear computation of the modulus values. where the tracking properties y t = Qr Sy r can be designed in-
Based on the above model, it can be established that a dependent of the regulating behavior y d = (1 − QS )n by Qr .
multivariable coupled system has to be controlled. If the The G2DF control system is shown in Figure 5, where
problem can be formulated for the oxide variables, then we y r , u , y, and n are the reference, process input, output, and
have a linear system; if it can be done only for the modulus
values, then the system is nonlinear.
The purpose of the control is to proportion an entire silo Mill Silo
content with desired average compositions and to reach r(t) Dynamics Dynamics
minimal variances around these reference values. The task w(t) r′inp(t) r′out(t) r′silo(t) q(t)
× Gm(z−1) Gs(z−1) Σ
is to decrease the composition inhomogeneities of a silo
batch by mechanically homogenizing the silo content be-
C(t) Σ
fore it gets to the kiln.
If the changes in the raw material compositions allow, v(t)
and the control operates accurately, a continuous silo can ÷
also be applied instead of the batch one. In such cases, the ÷
silo can be abandoned, affording considerable savings in oxout(t) oxsilo(t)
capital investment. Note that only the variables ox out are
measured and the values ox silo , M out , and M silo are computed.
Mout(t) Msilo(t)
A Generic Optimal Controller Structure
The need to design high-performance control systems has
not diminished in importance despite the thousands of Figure 3. Model of the raw mill-silo system.

February 2003 IEEE Control Systems Magazine 89

disturbance signals, respectively. The optimal ARS regula- optimal computation of the Gr and Gn filters. If Gr and Gn are
tor of the G2DF scheme [4], [7], [8] is given by optimally selected, then Ropt denotes the optimal ARS regu-
lator in (9). It is interesting to see what the transfer charac-
Pn K n Qo PnGn S −+1 teristics of this system look like:
Ropt = Ro = = = ,
1 − Pn K n S 1 − Qo S 1 − PnGn S − z − d (9)
y = Pr K r S y r − (1 − Pn K n S ) n
where = Pr Gr S − z − d y r − (1 − PnGn S − z − d ) n
= yt + yd .
Qo = Qn = Pn K n = PnGn S −1
+ (10) (13)

is the associated Y-parameter [8]; furthermore, Here, Pr and Pn are stable and proper transfer functions,
which are partly capable of placing the desired poles in
Qr = Pr K r = Pr Gr S −+1 , K w = Gw S −+1 , K r = Gr S −+1 , (11) the servo and the regulatory transfer functions; further-
more, they are usually referred to as reference signal and
assuming that the process is factorable as output disturbance predictors. They can even be called
reference models, so reasonably the gains Pr (ω = 0 ) = 1 and
S = S + S− = S + S− z − d , (12) Pn (ω = 0 ) = 1 are selected.
The G2DF control system has a specific form, shown in
where S + means the inverse stable (IS) and S − the inverse un- Figure 6, representing the so-called invariant model part
stable (IU) factors, respectively. z − d corresponds to the dis- S − z − d of the process, which cannot be canceled or elimi-
crete time delay, which is the integer multiple of the nated by any method. Only the influence of S − can be attenu-
sampling time. ated by the optimally selected embedded filters Gr and Gn .
It has been shown [8] that the optimization of the G2DF When Gr = 1and Gn = 1(which are optimal selections only for
scheme can be performed in H2 and H∞ norm spaces by an IS plant), the regulator is
proper selection of the serial K r and embedded K n filters
(compensators). These optimizations will be reduced to the Pn S −+1 Pn
R= −d
= R S −+1 ; R = ,
1 − Pn S − z 1 − Pn S − z − d
R n
yr + Q + + y
Qr S S which results in very simple closed-loop characteristics [8]
1−QS + +

Internal Linear Plant
Model y = Pr S − z − d y r − (1 − Pn S − z − d ) n = y t + y d . (15)

Figure 4. The general form of the K-B-parametrization. In control engineering practice, another interpretation of
the plant factorization (12) can be used when S − does not
contain only the IU factors but includes any part of the pro-
n cess model that is not desirable to cancel. Mostly poorly
+ +
yr + PnKn y damped zeros are considered here that could result in un-
Pr Kr S 1−PnKnS + S + wanted intersampling ripples.

Ro It is widely accepted that “all 2DF configurations have ba-
sically the same properties and potentials” [9]. However, it
Figure 5. The G2DF control system. has been shown that the above scheme is primus inter pares,
and therefore the term generic applies to more than just the
equivalent feedforward/feedback closed-loop control sys-
True n tems. The generalization of this scheme for general multi-
Process ple-input, multiple-output (MIMO) systems is not very
yr Pn G n y difficult, but it is a time-consuming and not obvious task (con-
PrGr S− z−d + S−1 S+S− z−d
1−PnGnS− z−d + + sider the noncommutativity of matrix products). In such spe-

Invariant Inverse S cial cases (see the following sections), when the decoupling
Model Part Model
of the MIMO system can be solved with some special (not al-
Controller ways linear) algebraic algorithm, the above generic scheme is
applicable for the decoupled “channels” and is the simplest
Figure 6. The G2DF control system for the IS plant. structure for reaching optimal control performance.

90 IEEE Control Systems Magazine February 2003

The Control Problem From a technological standpoint the moduli and oxides
The raw meal composition is controlled by adjusting the of the silo content are of interest, since these parameters de-
feed ratios of the raw material components. In the literature, termine the quality of the next technological step, the burn-
the applied multivariable controller is generally a fixed-pa- ing process.
rameter decoupled regulator utilizing the classical I, PI, or The complexity of the problem is increased by the fact
PID regulator structure. The advanced versions of these that the silo is not able to homogenize its entire contents,
controller classes use optimization techniques (e.g., as in even if a mixing silo is used. To take this effect into consider-
[10]). A practical control problem arises because the com- ation, a pseudosilo is introduced: beyond a given silo con-
position of raw materials varies from time to time and can- tent, continuous homogenization is calculated instead of a
not be measured directly. In this case, a proper adaptive batch one:
controller should be used. Several early attempts (see [11]
and [12]) are well known for having used adaptive tech-
q ( t ) = q ( t − 1) = q max (18)
niques for composition control. These applications con-
sider the plant as approximated by a black-box model and
implement multivariable self-tuning regulators.
Our present approach attempts to use the structural and
ox silo ( t − 1)(q max − v( t )) + ox out
( t )v( t )
ox silo = ,
parametric a priori information, so only the unknown plant q( t )
parameters are learned (adapted) continuously. Our ap- j = 1,K , nox ,
proach is based on the adaptive adjustment of the decoup- (19)
ling network by estimation of the composition matrix, using
flexibility in modulus definition and in assigning the moduli where q max is the limit size of the pseudosilo.
to be controlled, elimination of the silo effect and that of the As mentioned earlier, the objectives of the production are
feeder errors, the capability to use a priori information, and expressed by ratios: by moduli and oxide concentrations. It is
separation of silo and sample control. The controller struc- worthwhile to handle these ratios in a common way because
ture is based on the multivariable version of the special the ratios applied for control change from factory to factory.
G2DF scheme introduced by the authors. Also, this is the way to make the results applicable to mixing
In several cement plants, the sampling is based on the problems exhibited by different technologies. The general-
30-minute or one-hour (depending on the analysis process) ized moduli can be defined by the following expressions:
average of the mill product. The sample transport, the sam-
ple preparation, and the analysis require about 20 minutes. nox

The time delay and the time lag of the feeder-mill system may ∑α
kj ox j
be different for the different raw material components. The mk = nox
time delays are relatively constant and can be known a priori. γ k + ∑ β kjox j
These values are determined by the mechanical construc-
tion, the transportation method, the mill type, the capacity of
the mill, and the grinding property of the components. Since
whereα kj ,β kj , and γ k are constants characterizing the modu-
the time lag is relatively small compared to the total time de-
lus mk . Thus, the control objective can be expressed by a
lay, the mass flow of the components can be modeled by pure
unified formula.
time delay blocks. The content of the silo and its oxide con-
The ultimate control problem is as follows: determine the
centration cannot be measured. They are calculated from the
flow rates of the raw materials to fill the silos with the appro-
mill product. By applying the mass and concentration bal-
priate meal, taking into consideration that the oxide concen-
ances, the dynamics of the silo (see (5)) can be obtained:
tration of the raw material changes, the sampling is rela-
tively rare, the time delay is relatively large, the definition of
q ( t ) = q ( t − 1) + v( t ) (16) the moduli is given by nonlinear functions, the homogeniza-
tion of the silo is far from ideal, and the technological con-
straints must be satisfied.
q ( t − 1)ox silo
( t − 1) + v( t )ox out
(t ) Only quality control of the mill system was discussed
ox silo (t ) = , j = 1,K , nox, above. In this regard, the energy consumption for grinding is
q( t )
(17) very high and the efficiency of the process is very low, so
finding the maximum mill production that could result in the
where ox out ( t ) is the jth measured oxide concentration of lowest relative energy consumption is highly recom-
the mill product, ox silo ( t ) is the jth oxide concentration of mended. The usable set point for quantity control is deter-
the silo content, and nox is the number of oxides used. mined by the technological parameters of the mill system.

February 2003 IEEE Control Systems Magazine 91

The Control Strategy • recalculation of controller state.
The realized adaptive controller for composition control The calculation of the feedback signal is as shown in Fig-
can be regarded as composed of two parts: an estimator and ure 9. The controller system actually performs sample con-
a feedback regulator. The structure of the controller is trol for a modified mill product value instead of control of
shown in Figure 7. The identification part estimates the ox- the silo content. The reason is that the silo control error may
ide concentration of the components from the input and drastically modify the mill product and move the state of the
output measurements of the mill. The controller system at- controllers to a point that is not optimal for the next control
tempts to eliminate the disturbances and to guide the pro- step. So the controlled variables are generated from the
cess (the feeder → mill → silo system) in such a way that the measured oxide concentration by elimination of the silo
mill → silo output would be close to the required reference control effect. The correcting signal depends on the opera-
values of the moduli. The feedback controller uses the mea- tion of the feeder servo system as well. Assume that for
sured values of the feeders, the measured quantity and ox- some technological reason one of the feeders is temporarily
ide concentration of the mill product, and those of the silo stopped for a few minutes while the others are still in opera-
content. tion. Due to the feeder error, there will be an error in the mill
The feedback controller system consists of four major product as well. This error can be measured some sampling
blocks (Figure 8): periods later. Because of this error, a controller without
• calculation of feedback signals these modifications would increase or decrease the set
• controller points of the feeders, but these adjustments would be incor-
• decoupling system rect, as the technological problem has already been solved.
The above structure guarantees that the silo control does
not disturb the mill control and eliminates the disturbances
caused by the feeder system.
As mentioned earlier, the feeder → mill → silo system is a
multivariable nonlinear process where the process inputs
Controller Feeder are the set values for the feeders and the outputs are the
Mill Silo
System System measured oxide concentrations of the mill product. To elimi-
nate the nonlinearity introduced by the modulus definition,
its inverse is used in the nonlinear decoupling network. This
network guarantees that the transfer function from the
Figure 7. Adaptive composition control. moduli-type input of the decoupling system to the moduli of
the silo are pure time delay blocks if the technology permits
adjustment of the feeder reference values based on the de-
coupling network values. The decoupling requires the solu-
tion of the following programming problem:
Controller State • determining the reference values of the (weigh) feed-
C q oxsilo ersw rk ( k = 1,K , nof), where nof is the number of feed-
ers used for control
wr • minimizing the cost of the mill production
Mr System r′inp oxout r
Mreg nof

Calculation Mout ∑cw ,

Feedback Signal
where ci is the cost of the ith component
• the sum of the feeder reference values w r is given by
Figure 8. Structure of the controller system.
wr = ∑w
Mreg r r′inp oxout
− Mout
Feeder + Mill Modulus • technological constraints for the feeders must be sat-
Model Model Calculation isfied:
Silo Effect

Figure 9. Calculation of the feedback signal. w ri, min ≤ w ri ≤ w ri, max i = 1,K , nof (23)

92 IEEE Control Systems Magazine February 2003

• moduli of the mill production are limited: G1i = S −i z −2 ; G1 = diag{K , ai + (1 − ai )z −1 ,K}z −2 (27)

 nof 
M min ≤ M k  ∑ C jiw ri , j = 1,K , nox ≤ M max
 i =1  G2i = ,
k = 1,K , nom, 1 − [ai + (1 − ai )z −1] z −2
 1 
G2 = diag K , −1 −2
where nom is the number of moduli to be controlled  1 − [ai + (1 − ai ) z ] z  (28)
and M k (K ) denotes the calculation rule of the kth
moduli from oxides
• the moduli of the mill production corrected by the silo G3i = 1 ; G3 = I. (29)
effect in a given prediction horizon must be equal or
as close as possible to the output of the controller Here, the reference models Pr = Pn = 1 were selected,
Mreg : which correspond to a deadbeat-type regulator with no
intersample ripples. (If the delay times for the different com-

∑ d |M |
ponents are significantly different, the exact decoupling and
reg − M silo
k =1 (25) deadbeat effect can be reached only by the real MIMO ver-
sion of the optimal controller given by (15)). It is easy to ver-
Here, M silo depends on the oxide quantity in the silo, the ify that the resulting channel regulators Ri = G2i are integrat-
previous commands, the reference values of the moduli, the ing ones having a pole at z = 1.
prediction horizon to control, the estimated oxide concen- In the controller → decoupling → feeder → mill → sam-
tration of the components, the model parameters of the mill pling analysis → modulus calculation process, the order of
system, and the feeder ratios to be calculated (see Figure the linear and the nonlinear elements cannot be changed,
10). dk is a properly selected weighting factor. but we would like to eliminate a special fail-operation. This
The above problem can be solved using a two-step linear is why the control system recalculates the controller state
programming algorithm. in every control step. The effect of the estimation and the
The task of the controller is to eliminate the disturbances control can be separated, and the disturbances caused by
that cannot be predicted from the a priori information and the parameter adjustment due to the estimation can be
the a posteriori estimation. As the process can be decoup- eliminated. At the same time, the integrator wind-up can be
led, parallel working single-loop controllers can be applied neglected and the limit values of the feeders can be taken
for control. The block scheme of the “semi-MIMO” control- into consideration.
ler used in our application is shown in Figure 10. It is easy to
Identification Strategy
see a one-to-one correspondence with the G2DF scheme in-
In every adaptive control system, two major compromises
troduced in Figure 6. The saturation of the controller limits
must be made: one is related to the time horizon, and the
the error signal and reduces the fail-effect of the occasion-
other is related to the information available for estimation. It
ally incorrect analysis.
is usually assumed that the parameters of the system are
To demonstrate the controller adjustment, assume that
slowly varying and the measurements are noisy. The parame-
the feeder → mill → sampling analysis system can be mod-
ter tracking can be realized by forgetting old data. The speed
eled by the following transfer function:
of tracking and the precision of the parameter estimation im-

[ ]
S i = ai + (1 − ai )z −1 z −2
pose contradictory requirements. The effect of the noise can
be reduced by averaging, but the speed of parameter tracking
S i
= [ a + (1 − a )z ] z −1 −2
= S −i z − d , d =2 can be increased by rapid forgetting of old data or old mea-
i i (26)
surements. One has to distinguish between the effects of
for the ith channel, which corresponds to an IU time delay
process, where the delay is common for all channels. We as-
sumed that the time lag effect is not significant and the sys-
tem can be modeled by pure time delay. The reasons for two S− z−d Compensator Inverse Model
+ +
parameters in the transfer function are that the delay time is
G1 Limiter G2 G3
not a multiple of the sampling period and the sampling device Mr − + Mreg
works continuously. This transfer function does not differ Invariant Part R
of the Model Mout
from a first-order lag, but it is not worthwhile to compensate
this effect by a lead block. The application of the optimal
G2DF controller results in the following transfer functions: Figure 10. Structure of the controller.

February 2003 IEEE Control Systems Magazine 93

noise and the effects of changes in parameters. The other and C ji ( t ) is the estimated jth oxide concentration of the ith
compromise is related to the information available for esti- component and ox out ( t ) is the jth measured oxide of the mill
mation by the excitation of the process. The requirement for product. There are components whose concentrations are
process input is different for control action and for parameter relatively stable and known, but some oxide concentrations
estimation. The identification requires significant variation of of certain materials are unknown and slowly varying. The es-
the process input. Effective control action reduces the con- timation process should determine these C ji values. Here,
trol error and does not ensure a good input signal for the esti- the identification algorithm for one oxide is presented, but
mation. In a mill control system, an additional excitation similar algorithms may be constructed for the other oxides
signal cannot be used. The identification must be realized on used as well.
the basis of the normal control operation. The parameter update of the often applied identification
′ ) relative to the
The quantity of the raw materials ( r ′ = rout methods leads to the matrix expression
actual measured oxide concentration of the raw meal can be
calculated based on the mill model using the measured val-
ues of the material flow. The identification process uses these
mass flow values and the measured oxide concentration to [
c j ( t + 1) = c j ( t ) + P( t ) r ( t ) ox out
( t ) − r T ( t )c j ( t ) , (32)

estimate the oxide concentration of the various components.

where the ith element of c j ( t ) isC ji ( t ), the ith element of r( t )
For the jth oxide, we can apply the following model:
is ri ( t ), and P( t ) is a weighting matrix. Here, c j is the trans-
pose of the jth row of C (i.e., a column vector). The identifica-
ox out ( t ) ≈ ∑ C ji ( t ) ri ( t ), (30)
tion processes differ in the choice of the weighting matrix
P( t ). This determines the convergence, the precision, and
where the partial material flow ratios can be calculated by
the learning capability of the estimation strategy.
r ′( t ) ri′( t ) It is well known that a least-squares-type choice of P( t ) may
ri ( t ) = nof
= and r = [r1 ,K , rnof]T guarantee good noise reduction, but it does not allow parame-
v( t )
∑ r ′( t )
ter tracking. For the estimation, however, the input signal must
excite the process sufficiently. The old data (old measure-
ments) are required to account for parameter tracking. Several
forgetting strategies can be found in the literature. A few at-
105 tempt to use variable forgetting factor algorithms that scale
Adaptive Control Switched on at t = 45
the covariance matrix and leave the search direction unal-
100 tered. In the search direction, scaling may result in smooth pa-
rameter tracking, but numeric problems may arise with ill
conditioning of the covariance matrix [13]. The directional for-
getting attempts to eliminate the persistent excitation prob-
lem [14], forgetting the collected information only in the
90 direction of the actual situation vector with constant factor.
ML Reference Value = 94 The forgetting strategy applied in adaptive composition
85 control combines directional forgetting with a variable for-
50 100 150 200 getting strategy. The application of a constant weighting ma-
Number of Samples [1 h = 3 Samples] trix P means that, in every step, as much of the collected
information is forgotten as is added by the new step. To in-
crease the robustness of the identification, the step size of
the parameter update is limited to the normal projection
90 step. A critical point is the choice of weight matrix. As the
size of the oxide concentration of various components can
50 100 150 200
be expected to fluctuate, the diagonal elements of the
Silo Filling Time = 45 weighting matrix can be chosen inversely proportional to
Average the expected variation. For example, if the oxide concentra-
tion of a component is known and does not change, the diag-
onal element of the weighting matrix corresponding to this
90 oxide can be chosen to be zero. In our experiment, the above
mixed-solution compromise provided the necessary adap-
50 100 150 200
tive estimation robustness with an acceptable learning
Figure 11. Recorded operating data of the lime standard. speed for abrupt changes.

94 IEEE Control Systems Magazine February 2003

Since the moving average of the observation vector is far 2
from zero, the adaptation speed can be increased by using Adaptive Control Switched on at t = 45
the continuously updated working point. So it is worthwhile 1.8
to apply P′ instead of P, where P′ is calculated by

[ ]
−1 −1
P′ ( t ) = P ( t ) + r ( t ) r ( t )
(33) 1.4

Here, r( t ) is the filtered values of the ratios of the raw ma- 1.2
terial flows by means of which the moving average of the ob- MA Reference Value = 1.5
servation vector is estimated. 1
50 100 150 200
Operational Records Number of Samples [1 h = 3 Samples]
The normal operational records given in Figures 11, 12, and
13 show the time curves of the three adaptively controlled
(lime, aluminum, and silica) modulus variables for filling five 1.5
homogenization silos. While filling the first silo, a conven-
tional (PI-like) regulator was operating, and at the start of 1
50 100 150 200
filling the second silo (at t = 45), the adaptive control system
was switched on. The sampling interval was 20 min. The ap- Silo Filling Time = 45
plied reference values are MLr = 94, MA r = 1.5, and MS r = 2.1 Average
index. The continuously updated average (pseudosilo) val-
ues are also presented. These results came from actual ex- 1.5
periments with the real blending system.
It is obvious that the variances of the three standards 1
50 100 150 200
with the PI regulator are considerably higher than with the
adaptive regulator. Because the adaptive controller started Figure 12. Recorded operating data of the aluminum modulus.
to learn during the normal operation, no additional long ad-
aptation period is visible. 3
Adaptive Control Switched on at t = 45
The physical explanation of the better performance of
the introduced adaptive controller is simple. The applied 2.5
strategy decomposed the original multivariable nonlinear
control paradigm to subproblems, which are much easier to
handle using the special decoupling approach presented. 2
Thus, the adaptation relates only to the parameter estima-
tion of the oxide models, which are basically static relation- 1.5
ships. The dynamics of the process is not continuously MS Reference Value = 2.1
adapted because it is not time varying and can be identified
to a good approximation a priori. At the same time, the ap- 50 100 150 200
plied G2DF controller scheme provides the best achievable Number of Samples [1 h = 3 Samples]
performance given the process dynamics. In this low-order
case, the final controller is a minimum-variance-type con-
troller ensuring H2 optimality.

Conclusion 1
50 100 150 200
This article described a simplified scheme for an adaptive
composition control algorithm. The applied control system Silo Filling Time = 45
is based on a new generic controller structure and has the Average
following main features:
• estimation of the raw material composition using an a 2
priori dynamic model of the mill
• adaptive adjustment of the decoupling network based 1
50 100 150 200
on the identified composition matrix
• flexibility in modulus definition Figure 13. Recorded operating data of the silica modulus.

February 2003 IEEE Control Systems Magazine 95

• direct modulus control [16] F. Mellies and I. Vajk, “Adaptive mischungsregelung, ein baustein für das
laborautomationssystem POLAB,” Zement-Kalk-Gips, 1991.
• handling the sampling, continuous, and batch homog-
[17] I. Vajk, L. Keviczky, J. Hetthéssy, F.J. Adrian, M. Hilger, and J. Kolostori,
enization “Adaptive composition control,” in Proc. 9th IFAC/IFORS Symp. Identification
• taking the silo into consideration as the control strat- and System Parameter Estimation, Budapest, Hungary, 1991, pp. 511-515.
egy forecasts the future average composition in the silo
• separation of silo and sample control Cs. Bányász received the M.S. degree in electrical engineer-
• elimination of problems caused by local position con- ing in 1969 from the Budapest University of Technology and
trollers, if necessary Economics (BUTE) and the Ph.D. in identification from the
• optimal satisfaction of the technological requirements Hungarian Academy of Sciences (HAS) in 1976. Since 1969
• effective noise reduction and adjustable operation she has been with the Computer and Automation Institute
speed. (CARI) of HAS, currently working with the Systems and Con-
The adaptive composition controller presented is in op- trols Laboratory as senior research scientist. She is a mem-
eration in several cement plants (e.g., in the Vác Cement ber of the Hungarian Academy of Engineering. She was the
Work in Hungary). editor and one of the main organizers of major international
Acknowledgment scientific events in Hungary (IFAC SYSID 1991, IFAC ACASP
This work was supported by the U.S. Army Research Office, 1995, IFAC ROCOND 1997, and the International Conference
the Hungarian National Science Foundation (Országos on Women Engineers and Scientists in 1996). She is an active
Tudományos Kutatási Alap: OTKA), and the Control Engi- member of the IEEE Women-in-Engineering Group. She has
neering Research Group of the Hungarian Academy of Sci- published more than 110 papers and has more than 120 cita-
ences (MTA TKI). tions. Her research interests are system identification, pa-
rameter estimation, and systems engineering applications
References in various industrial problems.
[1] H. Hammer, “Computer controlled raw meal production in cement indus-
try,” Regelungstechnik und Prozessdatenverarbeitung, vol. 5, pp. 190-198, 1972.
[2] H. Hoenig, “Component control in a cement plant using a process com-
L. Keviczky graduated from the Electrical Engineering Fac-
puter,” Zement-Kalk-Gips, vol. 1, pp. 31-36, 1972. ulty of the TUB in 1968 and received the first doctoral degree
[3] L. Keviczky, “Control in cement production,” in Proc. IFAC Symp. MMM’83, in 1970. Since 1985 he has been a member of the Hungarian
Helsinki, Finland, 1983, pp. 1-13.
[4] L. Keviczky, “Combined identification and control: Another way,” in Proc.
Academy of Sciences and is a member of the Hungarian
ACASP’95: IFAC Symp. Adaptive Control and Signal Processing, Budapest, Hun- Academy of Engineering. In 1991 he was appointed a Foreign
gary, 1995, pp. 13-30. Member of the Royal Swedish Academy of Engineering Sci-
[5] L. Keviczky, M. Hilger, and J. Kolostori, Mathematics and Control Engi-
neering of Grinding Technology. Ball Mill Grinding. Boston, MA: Kluwer Aca-
ences. He was the director of the CARI from 1986 to 1993 and
demic, 1989. then elected secretary general of the HAS. He has been
[6] J.M. Maciejowski, Multivariable Feedback Design. Reading, MA: Addi- vice-president of the HAS since 1999. Since 1994 he has been
son-Wesley, 1989.
full professor of the Department of Automation and Applied
[7] L. Keviczky and Cs. Bányász, “Iterative identification and control design
using K-B parametrization,” in Control of Complex Systems, K.J. Åström, M. Informatics at the BUTE. He has held several positions
Blanke, A. Isidori, W. Schaufelberger, and R. Sanz, Eds. New York: Springer, within IFAC (council member, Application, Awards and Elec-
2000, p. 482.
tion Committee chairman). In 1999 he was elected
[8] L. Keviczky and Cs. Bányász, “Optimality of two-degree of freedom con-
trollers in H2 - and H∞ -norm space, their robustness and minimal sensitivity,” vice-chairman of the Policy Committee for the present trien-
in Proc. 14th IFAC World Congress, Beijing, PRC, vol. F, 1999, pp. 331-336. nial. His special fields of interest are system identification
[9] I.M. Horowitz, Synthesis of Feedback Systems. New York: Academic, 1963.
and parameter estimation, adaptive optimal control of in-
[10] A.K. Swain, “Material mix control in cement plant automation,” IEEE
Contr. Syst. Mag., vol. 15, pp. 23-27, 1995. dustrial processes, computer-controlled systems, simula-
[11] L. Keviczky, J. Hetthéssy, M. Hilger, and J. Kolostori, “Self-tuning adaptive tion and modeling, and intelligent and expert controls.
control of cement raw material blending,” Automatica, vol. 14, pp. 525-532, 1978.
[12] L. Keviczky and J. Hetthéssy, “Cement industry: Adaptive control,” in Sys-
István Vajk received the M.Sc. degree in electrical engineer-
tems and Control Encyclopedia: Theory, Technology, Applications, M.G. Singh,
Ed. Elmsford, NY: Pergamon, 1987, pp. 561-564. ing in 1975 and the Ph.D. degree in automatic control in 1977,
[13] T. Hagglund, “Recursive estimation of slowly time-varying parameters,” both from the BUTE. He was awarded the candidate of sci-
in Proc. IFAC Symp. Identification and System Parameter Estimation, York, U.K.,
ences degree in 1989 from the HAS. Since 1976 he has been at
1985, pp. 1137-1142.
[14] R. Kulhávy, “Restricted exponential forgetting in real-time identifica- the DAAI, Faculty of Electrical Engineering and Informatics,
tion,” in Proc. IFAC Symp. Identification and System Parameter Estimation, York, BUTE, where he is a professor, and he has been head of the
U.K., 1985, pp. 1143-1148. department since 1994. His main research interests are in
[15] A. Lundan and O. Mattila, “A system for the control of the homogeniza-
tion of the cement raw meal,” in Proc. 4th IFAC Symp. Digital Control Applica- the fields of system identification, adaptive control, and
tions for Process Control, Zürich, Switzerland, 1974. software engineering.

96 IEEE Control Systems Magazine February 2003