7164 Ind. Eng. Chem. Res.

2005, 44, 7164-7174

Development and Application of a Cement Raw Meal Controller

Dimitris C. Tsamatsoulis*
Halyps Cement, Italcementi Group, 19300 Aspropyrgos, Attica, Greece

The development and implementation of a cement raw meal controller are illustrated. The study
has established that the controller parameters shall be adjusted and optimized under the existing
field of application. The optimization was performed by the utilization of actual production data
and simulation studies. The control action follow-up was carried out using the control chart
technique. The application of the controller not only improves the raw meal stability but also
facilitates the inquisition on probable deviation sources. By utilizing the controller and the
developed simulation, a series of further improvements can be considered.

1. Introduction silos, where it was homogenized before being fed to the

Cement quality is primarily characterized by its
stability concerning the compressive strength in mortar
and concrete. There are three main factors that influ-
ence the variability of the cement strength:1 (a) clinker
activity variability, (b) composition variability, and (c)
cement fineness variability.
The clinker activity depends on the minerals composi-
tion, free lime content, and conditions of clinker forma-
tion. For a stable kiln operation and fuel quality, the
above characteristics depend on the fineness and chemi-
cal composition of the feed to the kiln raw meal. The
first parameter is regulated during the grinding of the
raw meal. The variation of the second parameter is
related to the deviations of the raw meal quality at the
mill outlet and the mixing ratio of the homogenizing silo.
Figure 1 depicts a typical flowchart of raw meal produc-
tion. In the existing closed-circuit raw meal production
process, the feeding of the raw materials is performed
via three weight feeders, first feeding a crusher. The
crusher outlet goes to the recycle elevator and from
there to a dynamic separator, the speed and gas flow of
which controls the product fineness. The fine exit stream
of the separator is the main part of the final product.
The coarse separator return is directed to the mill,
where it is ground, and from there it feeds the separator
via the recycle elevator. The material in the mill and
dynamic classifier are dried and dedusted by hot gas
flow. The gases are directed to the dedusting filter, the
product of which is a minor part of the final product. It
should be pointed out that the regulation of the raw
meal composition has a strong effect on efficient kiln
operation as well as on all performance indices. Because
of theimportance of this regulation, different auto-
mated systems are now provided from several manu-
facturers for the sampling and analyzing of the raw
material as well as for the adjustment of the mill weight
feeders. The regulation is primarily performed via PID
(proportional-integral-derivative) and adaptive con-
trollers. Keviczky et al.2 presented in 1978 a self-tuning
adaptive controller, which was tested and compared
against a traditional PI controller in a raw meal cir-
cuit for 90 h. The mill outlet raw meal fed a system of
* To whom correspondence should be addressed. Tel.:
0030 210 5518310. Fax: 0030 210 5572070. E-mail:
d.tsamatsoulis@halyps.gr.
10.1021/ie049284c CCC: $30.25 © 2005 American Chemical Society
Published on Web 08/06/2005
kiln. Swain3 used the singular value decomposition to
calculate the raw meal proportioning and compared
the derived controller with the well-known QCX pro-
gram produced by F. L. Smidth. In the paper appears a
5 h comparison between the two systems. In a recent
paper, Banyasz et al. presented a novel adaptive control
system for raw meal blending.4 Their purpose is to
proportion an entire silo content with the desired
average composition and to reach minimal variances
around the target values. The silo content has to be
homogenized before going to the kiln. As the authors
reference, if the raw material composition allows, e.g.,
if it is stable enough and the control operates correctly,
then a continuous silo can be applied instead of the
batch one. The controller was tested in actual conditions
for 200 h and compared with a conventional PI regula-
tor. As the authors declare, the large number of pub-
lished papers indicates that no unique or best method
has been found. The optimum solution depends on
different factors. This is a firm belief of the author of
this paper.
1.1. Objective and Field of Application. The
objective of this study is to prove that the deviations of
the kiln feed composition not only depend on the raw
materials variations and the mixing capacity of the silos
but also strongly depend on the effectiveness of the
regulating action. Consequently, any PID controller
used to regulate the kiln feed composition should be
adjusted in the existing field of application, e.g., at the
specific conditions of an existing plant, to achieve
optimum performance. In the Halyps cement plant, a
PD controller for the regulation of the raw meal
composition has been developed and applied. Further-
more, an analysis of the method was conducted in order
to achieve its optimum performance. Because of the
general characteristics of the method, it can also be
applied in a variety of other cases of raw meal automatic
control systems.
1.2. Proportioning Moduli and Clinker Minerals.
The proportioning moduli are used to indicate the
quality and activity of the raw materials, raw meal, and
clinker in order to regulate them properly. For the
main oxides, the following abbreviations are commonly
used in the cement industry: C ) CaO, S ) SiO2, A )
Al2O3, and F ) Fe2O3. The main moduli characterizing
the raw meal and the corresponding clinker are as
follows:5

Figure 1. Flowchart of raw meal production.
Lime Saturation Factor:
100C
LSF ) (1)
2.8S + 1.18A + 0.65F
S
Silica Modulus: SM ) (2)
A+F
A
Alumina Modulus: AM ) (3)
F
Under equilibrium conditions, the above oxides give
the following mineral phases that constitute the poten-
tial clinker composition:6
Alite: C3S ) 4.07(C - CaOf) - 7.60S -
6.72A - 1.43F - 2.85SO3 (4)
Belite: C2S ) 2.87S - 0.754C3S (5)
Aluminate: C3A ) 2.65A - 1.69F (6)
Ferrite: C4AF ) 3.04F (7)
The regulation of some or all of the indices (eqs 1-3)
contributes significantly to the achievement of a con-
stant clinker composition.
2. Controller Derivation
2.1. Equation Formulation. The number of the
chemical indices (eqs 1-3) that can be adjusted depends
on the number of the existing weight feeders. For p
weight feeders, p - 1 indicators can be regulated.7 The
existing installation shown in Figure 1 includes three
independent weight feeders. As a result, two indi-
cators can be used as control variables. In the current
process, LSF (lime saturation factor) and SM (silica
modulus) were selected. As independent variables, the
limestone and iron ore percentages in the mill feed were
used. The controller responses act on the above inputs.
The controller accepts as inputs the raw mill exit
analysis results in order to decrease the raw meal
standard deviations not only in the mill exit but also
primarily in the kiln feed. The above is achieved by
taking into account the transient functions of the mill
and of the silo by calculating the corresponding transfer
functions.8
Ind. Eng. Chem. Res., Vol. 44, No. 18, 2005 7165
Figure 2. Block diagram of the regulation process.
To formulate the controller action, the following
equations are derived:
%limnext ) %limcurr - kp((LSFcurr - LSFtarget) +
kd(LSFcurr - LSFprev)) (8)
%addnext ) %addcurr + kA(SMcurr - SMtarget) (9)
2.2. Block Diagram. Limestone and clay are fed to
the mill via two silos: the first silo contains limestone,
while the second one is loaded with a mixture of lime-
stone and clay. The third silo contains either the
corrective material of high iron oxide or high alumina
content or both of them. The ratio of limestone/clay in
the mixture is 1:2. The controller regulates the percent-
age of the limestone and the corrective material in the
mixture in order to absorb the disturbances of the LSF
and SM resulting from the variation of the composition
of the clay contained in the mixture. The block diagram
of the regulation process is clearly illustrated in Figure
2. Regarding the Y[I] ZZ (where ZZ ) mix, mill, etc.)
and Gc[I] variables, the integer I takes two values. If I
) 1, the variable represents the corresponding LSF
variable. Otherwise if I ) 2, the SM variables are
considered. As also concerns the X[I] variables, I ) 1 or
2. If I ) 1, then X[1] ) % lim. On the other hand, X[2]
) % add.
The block diagram clearly depicts that the construc-
tion of the regulator requires the knowledge of the
transfer functions of the grinding process and the
homogenizing process as well. The average delay time
for transport and measurement of the samples should
also be known.
2.3. Mill Dynamic Response. To estimate the
response and the corresponding transfer function of the
7166 Ind. Eng. Chem. Res., Vol. 44, No. 18, 2005

Table 1. LSF Dynamic Response

t (min) LSF X
0 104.4 0
7 127.3 0.1338
20 202.9 0.5761
35 233.6 0.7557
50 243.0 0.8108
65 262.6 0.9253
87.5 267.3 0.9531
110 275.3 1.000

mill where the raw meal is prepared, the technique of

the step change of a characteristic property of the raw
meal was selected. The selection of the above property
was based on the following criteria: (a) The property
should be intrinsic and representative for the material
ground. (b) Its dynamic behavior has to be equivalent
Figure 3. Residence time distribution of the grinding circuit.
to the response of the raw mix ground in the given
circuit. (c) Its sudden change should not produce a
variation in the steady state operation of the grinding 3. The pulse response of the grinding system, derived
circuit larger than the natural variance caused by the from the step response, is also depicted in the same
grindability or humidity of the raw materials in small figure. The large departure of the pulse response from
time intervals. The lime saturation factor conforms to the normal Gauss curve and the long tail are due to the
all the above criteria, and it was chosen as the previ- circulating load of the system.
ously mentioned system variable. To achieve the step Using the trapezoidal rule, the average residence time
change of the LSF in the mill inlet, two steps were of the material in the circuit can be calculated according
followed: to the following formula:
(1) Initially, the mill operation was stabilized for
around 2 h, in relation to the following: (a) the total
feed rate, (b) the circulating load, and (c) the flow and
temperature of the hot gases inserted to dry the raw
τmill ) ∫0
1
t dXmill )
N

∑
i)1
( 2
)
ti-1 + ti
∆Xmill (13)

materials.
(2) A step change was performed in the raw mix
composition fed in the mill via the weight feeders. An
initial feed consisting of 40% limestone and 60% mixture
was changed to a feed with 80% limestone and 20%
mixture. Samples from the final product stream were
taken immediately after the step change, the chemical
analysis was performed, and the LSF values were
calculated. To estimate the average residence time of
the material in the grinding circuit, the control variable
was also calculated in dimensionless form.
LSF - LSFmin
X(t) ) (10)
LSFmax - LSFmin
The values of LSF and X are shown in Table 1.
A second-order Laplace transform of the transfer
function is considered:
LSFOUT(s) 1 The raw meal extracted from this silo is fed directly to
) (11)
LSFIN(s) (1 + τ1s) (1 + τ2s)
From the reverse Laplace transformation, the following
transfer function was obtained:
The numerical integration of the data produces τmill )
27.1 min. This result, of course, does not represent the
residence time of the material in the mill, which is
shorter enough. It corresponds to the total grinding
circuit residence time. The value of τmill, as well as the
size of the tail of the pulse response, depends on the
amount of the coarse material recirculated via the
separator. We should point out that the test was
performed under normal operating conditions with a
circulating ratio around 3 (t/h of return to t/h of fresh
feed) and a fresh feed of ∼150 t/h.
2.4. Silo Dynamic Response. The raw meal pro-
duced in the mill is homogenized in two silos connected
in series. The raw meal is first alimented in one
relatively small silo of 600 t capacity. In this silo, it is
fully mixed and continuously extracted by overflow.
Afterward, it is fed to the second silo, which is consid-
ered a stock silo, with a maximum capacity of 6000 t.
the kiln. An appropriate number of compressors providea
the necessary air flow to homogenize the raw meal. The
two silos can be considered as two reactors connected
in series. A transfer function of first order is used for
each of them. The resulting function is the product of

( ( ) ( ))
τ1τ2 1 the two first-order functions, and consequently, it is of
t 1 t
Xmill(t) ) 1 - exp - - exp - second order and describes the behavior of a silo of total
τ1 - τ2 τ2 τ1 τ1 τ2 capacity 6600 t. To obtain the residence time distribu-
(12) tion in such a large capacitor is not easy, because the
experimental procedure followed in the case of the
The experimental data were fitted to eq 12 in order to grinding circuit is not applicable, without causing an
estimate the time constants τ1 and τ2. Using a nonlinear enormous and costly disturbance of the kiln’s normal
regression technique, the following solutions were operation. Instead of that technique, the actual analysis
found: τ1 ) 2.7 min and τ2 ) 23.4 min. The experimen- of the material at the inlet and outlet of the silo was
tal and calculated results of Xmill as a function of time, performed, by applying the convolution theorem. The
representing the residence time distribution of the Laplace transform of the LSF in the kiln feed is given
material in the grinding circuit, are illustrated in Figure by the following formula:
 Ind. Eng. Chem. Res., Vol. 44, No. 18, 2005 7167

LSFOUT(s) ) Gsilo(s)LSFIN(s) (14)

The transfer function of the silo is the response of the

silo if the Dirac function δ(t) is applied to the inlet. The
inverse Laplace transformation of eq 14 gives

LSFOUT(t) ) ∫0tLSFIN(τ)Gsilo(t-τ) dτ )
∫0tLSFIN(τ)Gsilo(t-τ) d(t - τ) ) ∫0tLSFIN(τ) dXsilo(t-τ)
(15)
In eq 15, the silo pulse response Gsilo(t) is substituted
by the second-order step response Xsilo, given by eq 12
for a second-order system.
N
Figure 4. Raw meal LSF in the silo outlet (experimental and
∫0
t
LSFIN(τ) dXsilo(t-τ) ) ∑ LSFIN(τi) dXsilo(t-τi) calculated values).
i)0
t
τi ) i∆t ∆t ) (16)
N
By taking into account the above formulas and using a
nonlinear regression technique, it is possible to calculate
the time constants, τ1 and τ2, of eq 16. The composition
data of the raw meal fed to and extracted from the silo
were taken into consideration for a period of four
continuous months. The variability of the LSF at the
mill outlet results from the previously applied regula-
tion of the mill weight feeders. Each sample of raw meal
collected by an automatic sampler at the silo entrance
is a 1 h average sample. On the other hand, silo outlet
sampling is instantaneous. By applying eqs 12 and 16,
the LSF values of the raw meal in the kiln feed were Figure 5. Step (Xsilo) and pulse (δsilo) response curves of the silo.
derived. These were compared with the actual values
and the standard error between them was minimized 17-21): The time parameters in the mill function were
by the regression algorithm. As a result, the following
time factors were estimated: τ1 ) 4 h and τ2 ) 11 h. GC[LSF](s) ) kp(1 + kds) (17)
The value of τ1 corresponds exactly to the ratio of the
mass of the material in the small silo to the average GC[SM](s) ) kA (18)
inlet mass flow to the system. Because the small silo
operates in overflow condition, it always remains full
with material. The actual and calculated LSF values of GD(s) ) exp(-tds) (19)
the kiln feed are shown in Figure 4. The LSF actual
standard deviation is calculated as sact ) 2.357, while 1
Gmill(s) ) (20)
the estimated one from the model is smodel ) 2.230. This (1 + 2.7s)(1 + 23.4s)
means that the residual variance is found to be at a low
level, near the analysis reproducibility. The fit of the 1
Gsilo(s) ) (21)
actual data to the calculated ones is good. The step and (1 + 4s)(1 + 11s)
pulse response curves of the silo, Xsilo and δsilo, respec-
tively, using the above time constants are shown in defined in minutes, while those in the silo function were
Figure 5. The above technique of estimation of the defined in hours.
dynamic response using only routine analysis data was In the installed sampling system of the raw meal at
very useful, since the application of a step change in the mill outlet, an automatic sampler continuously
the input certainlyhas an excessively harmful influ- collects material for a certain time period. By the end
ence on the kiln operation. The above-described pro- of this time, the sample is fully homogenized in one
cedure proves that, in many cases, current industrial stirrer and automatically extracted from the sampler.
data, which usually include a lot of noises, can give a The sampling period was selected to be 1 h. Afterward,
perfect description of the process if the correct math- the sample is transferred manually in the laboratory
ematical treatment is applied and can be utilized to and the preparation procedure follows. The procedure
adjust and calibrate the appropriate mathematical consists of a fused bead preparation using a flux to
models. decrease the fusing temperature. The analysis of the
2.5. System Transfer Functions. On the basis of bead is performed in an XRF analyzer. The delay time
the mathematical treatment presented in Sections 2.3 td under the existing conditions of sampling, transfer,
and 2.4 as well as in eqs 8 and 9 describing the regu- preparation, and measurement of the sample is 30 min.
lating action of the controller, all the transfer func- The sampling at the silo outlet is instantaneous. The
tions of the system can be derived. These functions in results of the sample analysis provide important infor-
Laplace form are given in the following equations (eqs mation to the kiln operator for the regulation of the kiln
7168 Ind. Eng. Chem. Res., Vol. 44, No. 18, 2005
parameters. The analysis results are also the main
factors defining the produced clinker handling.
The LSF value of an instantaneous sample at the mill
outlet is the convolution between the LSF of the feed
and the Xmill function. The time interval where this
function takes nonzero values is 2 h, as illustrated in
Figure 3. Therefore,
LSFout(t) ) ∫0t LSFin(τ) dXmill(t-τ) )
∫t-120
t
LSFin(τ) dXmill(t-τ) (22)
A new time variable, u ) t - τ, is considered, defined
in the interval [0, 120]. The value of LSFout(t) is
influenced by three compositions in the weight feeders
because of the 30 min delay between sampling and
measurement. For each composition, an average LSFin
is assumed. On the basis of this assumption, the
following equation is derived from eq 22:
LSFout(t) ) LSFin(t-30) ∫030 dXmill(u) +
LSFin(t-90) ∫3090 dXmill(u) + LSFin(t-150)∫90120
dXmill(u) ) 0.6864LSFin(t-30) +
0.2895LSFin(t-90) + 0.0241LSFin(t-150) (23)
An important conclusion can be drawn from the above
correlation. The last composition contributes only 68.6%
to the composition of the instantaneous sample. Even
2 h before, the composition influences the LSFout by
2.4%. The sample extracted from the automatic sampler
represents the mean value of the instantaneous samples
over 1 h. The LSF of this sample is given by the
following formula:
LSFs(t) ) ∫
1 t
60 t-60
LSFout(u) du )
0.3432LSFin(t-30) + 0.4880LSFin(t-90) +
0.1568LSFin(t-150) + 0.0120LSFin(t-210) (24)
From the above equation, the large influence of the
previous feeders’ composition on the composition of the
sampled raw meal in the mill output becomes more
evident. Exactly the same formulas are valid for the
second control variable, SM. Equation 24 can be utilized
in simulation algorithms to optimize the controller
performance.
3. Feedback Control System Stability Criteria
The transfer functions of the feedback control system
including the controller, the grinding process, and the
delay time are the following:
Gc[LSF](s)Gmill(s)
Gs[LSF](s) ) (25)
1 + Gc[LSF](s)Gmill(s)Gd(s)
Gc[SM](s)Gmill(s)
Gs[SM](s) ) (26)
1 + Gc[SM](s)Gmill(s)Gd(s)
The total system also includes the transfer function of
the homogenizing silos, connected in series with the
grinding system. The resulting transfer functions de-
scribing the total system from the feeding of the mill
up to the feeding of the kiln are the following:
Gc[LSF](s)Gmill(s)Gsilo(s)
Gs,tot[LSF](s) ) (27)
1 + Gc[LSF](s)Gmill(s)Gd(s)
Gc[SM](s)Gmill(s)Gsilo(s)
Gs,tot[SM](s) ) (28)
1 + Gc[SM](s)Gmill(s)Gd(s)
The denominator of each transfer function represents
the characteristic equation of the feedback control
system. The denominators of eqs 25 and 27 and eqs 26
and 28 are respectively the same. The reason is that
there is not a feedback from the silo to the manipulating
variables. As a result, the stability criteria applied to
the functions of eqs 25 and 26 give exactly the same
results as the functions of eqs 27 and 28. The stability
of the system was investigated using the Nyquist
analysis. This criterion uses the frequency response of
the open loop GO(jω). For the existing system, GO(jω) )
Gc(jω)Gmill(jω)Cd(jω). The condition for the system to be
stable is the following: For phase φ ) -180°, it must
be |GO(jω)| < 1. This means that for Im(GO(jω)) ) 0,
then Re(GO(jω)) > -1. The transfer function of the delay
time Cd(jω) ) exp(-jωtd) ) cos(ωtd) - jsin(ωtd). A stable
LSF control system for a pair of [kp, kd] is shown in
Figure 6.
The area of [kp, kd] giving a stable LSF system is
presented in Figure 7. The delay time was considered
as an additional parameter. As can be observed from
Figure 7, the area of the controller parameters to ensure
a stable operation is sensible enough in the case that
the delay time changes. In the existing system of
sampling, transferring, and measurement, the average
delay time is 30 min. But for stability assurance
purposes, the selection of the controller coefficients
should be in the area below the 40 min curve. The
stability of the SM control system corresponds to the
vertical segment where kd ) 0.
4. Optimization of the Controller Parameters
4.1. Controller Simulator. To estimate the optimum
range of the controller parameters, a system simulator
has been developed which takes into account all the
transfer functions of the system. The optimization
criterion is the minimum standard deviation of LSF and
SM in the kiln feed as well as the minimum deviation
from the targets of the kiln feed average values. On the
basis of existing data of analysis in the mill outlet and
considering that the limestone and iron ore compo-
sitions are constant, the clay compositions fed to the
mill were back-calculated. The actual percentages of
the raw materials in the weight feeders are also taken
into account. Such a calculated variance of the clay
composition actually summarizes the natural variances
of the three raw materials and can be considered as
natural. The variation of the composition at the mill
outlet is not natural but results from the regulation
applied in the system. The previously applied regulation
of the LSF and SM used target values for the above
moduli as well as upper and lower warning and action
limits. If the values of the moduli of the raw meal
analyzed at the mill outlet were out of these limits,
changes were applied to the three dosing systems to
reach the target values. The derivation of the limits was
based on the previous experience. Figure 8 depicts the
clay LSF values for 700 h of mill running. Each point
of the above data represents the hourly average LSF of
the clay fed to the mill. In the following sections

Figure 6. Application of the Nyquist criterion for the stability of
the LSF control system.
Figure 7. [kp, kd] area for the stable LSF control system using
the Nyquist criterion.
Figure 8. Clay LSF in the mill feed.
concerning the construction and application of the
simulator, the composition of the clay is assumed to be
constant hourly and equal to the hourly average com-
position.
Using these data, the controller simulator calculates
the oxides composition and the corresponding modules
of the raw meal in the mill outlet and in the silo as well.
Some typical simulator results are shown in Figure 9
for the LSF controller using as coefficients the values
of kp ) 0.41 and kd ) 58. Typical results of the SM
Ind. Eng. Chem. Res., Vol. 44, No. 18, 2005 7169
Figure 9. Simulator results of the LSF controller.
Figure 10. Simulator results of the SM controller.
proportional controller are shown in Figure 10. A value
of kA ) 2.7 was selected.
4.2. Estimation of the Optimum Area of the
Controller Parameters. A first estimation of the
optimum coefficients was performed using the clay
analysis data given in Section 4.1. The targets for the
moduli under control were the following: target LSF )
98 and target SM ) 2.5. Using the simulator and
considering different values of kp, kd, and kA as inputs,
the following outputs were derived: (a) average LSF and
SM of the raw meal in the kiln feed, aver LSF and aver
SM, respectively; (b) standard deviation of the moduli
of the of the raw meal in the mill outlet and kiln feed,
std LSF mill, std SM mill, std LSF KF, and std SM KF;
and (c) standard error of the moduli from the target of
the raw meal at the mill outlet and at the kiln feed err
LSF mill, err SM mill, err LSF KF, and err SM KF. The
standard errors, standard deviations, and average
values are connected with the formula presented below
for the LSF of the kiln feed:
err LSF KF2 ) std LSF KF2 +
num S
(aver LSF - LSFtarget) (29)
num S - 1
The same formula applies for all parameters. The
variance of each chemical module is the sum of the
variances arisen from of the simulator and the residual
variance according to eq 30.
7170 Ind. Eng. Chem. Res., Vol. 44, No. 18, 2005

sact2 ) smodel2 + sres2 (30)

One part of the residual variance is caused by the

analysis of reproducibility. To have an exact estimation
of the above variance, Halyps Laboratory includes in
its quality system the following procedure in accordance
with the norm (ISO 5725-2:1994): Once per week, a
fused bead from the same homogenized sample of raw
meal is prepared and analyzed in the XRF. All the
laboratory operators are considered as a group. After
the exclusion of the probable outliers, the reproducibility
sR of the method in the existing laboratory is deter-
mined. In Table 2, the results of sact, smodel, sres, and sR
for the LSF and SM are shown. The table clearly depicts
that the larger part of the sres is attributed to the
laboratory reproducibility. The above is a further indi-
cation that the model fits well the experimental data. Figure 11. Err LSF mill as a function of kp and kd.
Table 2. Standard Deviation Values for LSF and SM
LSF SM
sact 2.357 0.092
smodel 2.230 0.089
sres 0.763 0.022
sR 0.447 0.015

The standard errors of the LSF at the mill outlet and

in the kiln feed as a function of kp and kd are presented
in Figures 11 and 12, respectively. Each curve is drawn
with a constant kd value.
From Figures 11 and 12, the following conclusions can
be drawn:
(1) For a constant kd, an optimum appears in the mill
and silo deviation by increasing the kp. This means that,
for each pair (kp, kd), there is an optimum regulation
frequency. A low kp gives a low frequency. On the con-
trary, a high kp gives a correspondingly high frequency.
Figure 12. Err LSF KF as a function of kp and kd.
The same trend is observed as regards the kd value.
(2) The optimum values of the regulator parameters
are not the same in the two checkpoints. The search is
made in order to optimize the parameters in relation to
the silo outlet deviation. An estimation of the optimum
region for the two parameters can be made from Fig-
ure 12. The proportional part belongs in the region
0.20-0.60. The differential part can be varied from 0
to 2. On the other hand, it can be seen from Figure 12
that a very low kp results in a high deviation at the
mill outlet. It must also be noticed that the optimum
control parameters depend on the deviation of the clay
composition.
The standard error of SM in the kiln feed as a function
of kA is shown in Figure 13. Iron ore with two different
compositions was used. In the case of iron ore rich in
Fe2O3, the optimum area of the kA is found in the space
1.0-3.0. In the second case, the optimum area is larger,
but the space (1.0-3.0) remains valid. The significance
of this is that the optimum coefficient depends on the Figure 13. Err SM KF as a function of kA.
nature and quality of the corrective materials; in any
case, the optimum area can be determined by applying A qualitative explanation of the above result is the
the developed simulator. following: A very slow controller produces small changes
The silo mixing ratio H, given by eq 31, was estimated in the composition, and as a result, the mill outlet raw
as a function of the LSF controller parameters, using meal could be out of the target for more than 4 h, that
the same clay composition. The results are presented is, the residence time in the homogenizing silo. Conse-
in Figure 14. An important conclusion is derived from quently, in the stock silo, thick layers of out-of-target
this figure: The silo mixing ratio is strongly dependant raw meal will appear; because the homogenizing action
on the way of regulation. of the stock silo is low, this raw meal will exit from the
silo. This becomes more obvious when the filling of the
std LSF mill stock silo is low. Using a faster controller, the above
Hsilo ) (31)
std LSF KF layers become thinner and the stock silo can homogenize

Figure 14. Silo mixing ratio as a function of kp and kd.
Figure 15. Optimum area of kp and kd.
them as well. The use of a very fast controller has a
reverse effect: The achievement of the target becomes
difficult and raw meals with very high or very low LSF
and SM are difficult to homogenize. From these figures,
it was concluded that a more precise estimation of the
optimum set (kp, kd) was necessary. To realize this, the
actual results of the mill exit composition were consid-
ered. Each set represents one full month of data. Using
these data with known compositions in the mill feed and
with the aid of eqs 22-24, the clay compositions and
the storage silo residence time τ2 were estimated and
considered as inputs in the simulator. The optimum
values of (kp, kd) for each set of data, given the minimum
standard error of the raw meal in the kiln feed, were
found by applying a nonlinear regression technique.
After this calculation, the area of (kp, kd) which gives
an error 20% higher than the minimum was found for
each data set. The optimum region of (kp, kd) should be
the common area of all the curves. The results are
drawn in Figure 15. A reasonable optimum set of (kp,
kd) can be put in the center of this area.
5. Application of the Controller
The raw meal composition controllers have been
applied in the Halyps plant since October 1997 as
regards the part regulating the LSF and since August
1998 as regards the SM-regulating part. 155 spot raw
meal samples in the kiln feed were collected each month
and were measured as an average number. The spot
clinker samples taken hourly constitute an average
Ind. Eng. Chem. Res., Vol. 44, No. 18, 2005 7171
Figure 16. Monthly standard deviation of the raw meal LSF in
the kiln feed.
daily sample which was also measured. Each month,
27 samples of clinker are measured and averaged. To
examine the improvements attained because of the
controller implementation, the control chart tech-
nique was selected, according to the ISO 8258:1991
standard. First, the following monthly standard devia-
tions were studied: (a) standard deviation of the raw
meal LSF in the kiln feed and in the mill outlet; (b)
standard deviation of raw meal SM in the kiln feed; and
(c) standard deviation of the clinker LSF in the kiln
output. To examine whether this kind of deviation
remains under control, the Shewhart s-control chart was
used.
5.1. Standard Deviation of the LSF in the Kiln
Feed and in the Mill Outlet. The average monthly
standard deviation, s av was calculated using all the
data from October 1997 up to December 2003, 75 points
total, and was 1.63. The upper control limit, UCL, is
calculated according to the formula:
sav
c4 x
UCL ) sav + 3 1 - c42 (32)
The constant c4 is estimated using eq 39 for n ) 155
n
-1! ( )
x
2 2
c4 ) (33)
n-1n-1
2 (-1! )
Using eqs 32 and 33 results in UCL ) 1.91. The results
are demonstrated in Figure 16.
The corresponding results as regards the mill outlet
are presented in Figure 17.
The difference between the two periods as illustrated
in Figure 18 is obvious. 46 out of 54 points were out of
control before applying the controller in the mill produc-
tion. Because of the regulating action of the LSF
controller, only 6 out of 75 points are out of control. The
under-control points increased from 15% to 92%. On the
other hand, the standard deviation of the LSF in the
mill outlet was decreased to a lower degree, e.g., from
9.3 to 8.0. This means that the controller acts as it was
designed, by influencing the raw meal layers in the silo.
The plant quality manager conducted an analysis of
the operation of the process to determine the causes of
7172 Ind. Eng. Chem. Res., Vol. 44, No. 18, 2005
Figure 17. Monthly standard deviation of the raw meal LSF in
the mill outlet.
Figure 18. Monthly standard deviation of the raw meal SM in
the kiln feed.
out-of-control signals, identified these causes, and has
taken corrective actions to prevent these causes from
recurring. The implementation of the controller made
easierthe identification of the causes contributing to an
out-of-control signal, increasing the effectiveness of the
quality system.
5.2. Standard Deviation of the SM in the Kiln
Feed. The average monthly standard deviation, s av,
was calculated using all the data from August 1998 up
to December 2003, 65 points total, and was 0.056. The
UCL estimated with the aid of eqs 32-33 is 0.066. The
results before and after implementation of the SM con-
troller are shown in Figure 18. Before the implementa-
tion of the controller, only 4 out of 70 points were under
control, which corresponds to 6% of the total signals.
After the controller implementation, 12 out of 65 points
are out of control, which means that the percentage of
the under-control signals is increased to 82%.
One of the major causes of the out-of-control signals
was the fact that, some days, the iron ore was not able
to correct the SM: This module was less than the target,
and the percentage of the additive in the weight scale
was zero. In this case, a correction with silica was
needed. The quality manager has studied the case to
correct with this kind of material, and some industrial
tests were performed.
5.3. Standard Deviation of the Clinker LSF in
the Kiln Outlet. The calculations and the time periods
Figure 19. Monthly standard deviation of the clinker LSF in the
kiln outlet.
Figure 20. Monthly average LSF of the clinker in the kiln outlet.
are exactly the same as in the Section 5.1. The values
used are sav ) 1.26 and UCL ) 1.78. The results are
depicted in Figure 19, from which the following conclu-
sions can be drawn. Before the controller implementa-
tion, 26 out of 54 points were out of control and the
corresponding percentage of the “in”-control points was
49%. After the implementation, the out-of-control sig-
nals were diminished to 6 out of 75, with a correspond-
ing increase of the in-control points to 92%. This result
is identical to the previous one found from the control
chart of the raw meal LSF. Especially during the last 5
years, only 2 out of 60 points produced an out-of-control
signal and the causes were assignable. Because of the
controller action, there is a clear improvement verified
at both sides of the kiln as concerns the conformity of
the clinker. Furthermore, the clinker can be managed
as regards its chemical analysis according to the results
of the kiln feed analysis. The Halyps plant quality
system includes technical instructions related to the
above topic.
5.4. Monthly Average LSF of the Clinker in the
Kiln Outlet. Except the comparison of the standard
deviations of the different moduli before and after the
application of the controller, the monthly attainment
of the clinker LSF target is illustrated in Figure 20. The
upper and lower control limits (UCL and LCL) are also
demonstrated. These values are calculated as

sav
UCL ) LSFtarg + 3 (34)
c4xn
sav
LCL ) LSFtarg - 3 (35)
c4xn
The LSFtarg value is a management decision. To calcu-
late UCL and LCL, the previously given sav and n are
taken into account. The months of kiln general overhaul,
with little available data, were excluded. Without a
doubt, the control limits are very strict as they are
calculated after the implementation of the controller.
As can be observed in the figure, the out-of-control
signals decreased drastically after the controller imple-
mentation. Before October 1997, 47% of the monthly
results were out of the control limits. This percentage
was decreased to 12.5% after this date. When a point is
out of control, the structure of the controller permits a
fast adjustment by only modifying the LSF target in the
mill. For this reason, when an out-of-control signal
appears, the next point is always found inside the
control area.
6. Conclusions
The development and application of a PD controller
to regulate the raw meal composition is presented. From
this study, the following conclusions can be drawn. The
mill transfer function estimation is based on the tech-
nique of the step change of a characteristic raw meal
property. The selection of the above property is based
on definite criteria. This technique and the correspond-
ing experimental procedure are not feasible in the case
of the silo transfer function calculation, without causing
an enormous and costly disturbance of the kiln’s normal
operation. Instead of this technique, the actual analysis
of the material in the inlet and outlet of the silo is
utilized, by applying the convolution theorem. This
procedure proves that, in many cases, current industrial
data, which usually include a lot of noise, can give a
perfect description of the process if the correct math-
ematical treatment is applied and can be utilized to
adjust and calibrate complicated mathematical models.
There is a large influence of the previous feeder’s
composition on the composition of the sampled raw meal
in the mill output, because of the delay time between
sampling and analysis result. For this reason, the
application of stability criteria is necessary to find an
initial optimum area for the controller parameters. To
optimize and to obtain a continuous follow-up of the
controller coefficients, a simulator of the system is
constructed and applied to actual data. The criterion of
optimization is the minimum standard deviation of LSF
and SM in the kiln feed as well as the minimum
deviation from the targets of the kiln feed average
results. Using this analysis, the optimum area of [kp,
kd] pairs is derived.
The implementation of the controller can be followed
up by applying the control charts technique. The
combination of both techniques makes easier the iden-
tification of the causes contributing to an out-of-control
signal, increasing the effectiveness of the quality sys-
tem.
In summary, it is proved that the standard deviation
of the kiln feed composition not only depends on the raw
materials variations and the mixing capacity of the silos
Ind. Eng. Chem. Res., Vol. 44, No. 18, 2005 7173
but also strongly depends on the effectiveness of the
regulating action.
Nomenclature
%addcurr ) current percentage of the corrective material
in the weight feeder, %
%addnext ) next percentage of the corrective material in
the weight feeder derived from the controller, %
AM ) alumina modulus
aver LSF, aver SM ) average LSF and SM of the raw meal
in the kiln feed
C2S ) belite, dicalcium silicate
C3S ) alite, tricalcium silicate
C3A ) aluminate
C4AF ) ferrite
c4 ) factor given in the equations 32 and 33 to calculate
an unbiased estimation of the population standard
deviation
D(s) ) characteristic equation of the feedback control
system
err LSF KF, err SM KF ) standard error of the raw meal
LSF and SM in the kiln feed
err LSF mill, err SM mill ) standard error of the raw meal
LSF and SM in the mill outlet
Gc[LSF](s) ) LSF controller transfer function
Gc[SM](s) ) SM controller transfer function
Gd(s) ) delay transfer function
Gmill(s) ) mill transfer function
GO(jω) ) frequency response of the open loop
Gsilo(s) ) silo transfer function
Gsilo(t) ) silo pulse response function
Gs[LSF](s) ) feedback control system including the control-
ler, grinding process, and delay time for the LSF
Gs[SM](s) ) feedback control system including the control-
ler, grinding process, and delay time for the SM
Gs,tot[LSF](s) ) feedback control system including the
controller, grinding process, and delay time for the LSF
Gs,tot[SM](s) ) total system transfer function for the LSF
Gc[SM](s) ) total system transfer function for the SM
Hsilo ) silo mixing ratio
kA ) proportional coefficient of the SM controller
kd ) derivative coefficient of the LSF controller
kp ) proportional coefficient of the LSF controller
%limnext ) next percentage of the limestone in the weight
feeder derived from the controller, %
%limcurr ) current percentage of the limestone in the
weight feeder, %
LCL ) lower control limit
LSF ) lime saturation factor
LSF clinker ) monthly average clinker LSF
LSFcurr ) value of the LSF in the current sample
LSFin(t) ) input LSF function to the system under consid-
eration
LSFIN(s) ) Laplace transform of the input LSF function to
the system under consideration
LSFmin ) minimum value of the LSF
LSFmax ) maximum value of the LSF
LSFout(t) ) output LSF function to the system under
consideration
LSFOUT(s) ) Laplace transform of the output LSF function
to the system under consideration
LSFprev ) value of the LSF in the previous sample
LSFs ) sample LSF
LSFtarget ) target value of the LSF
num S, n ) number of samples
s ) Laplace independent variable or standard deviation
sact ) actual standard deviation
sav ) average monthly standard deviation
s2act ) actual variance
7174 Ind. Eng. Chem. Res., Vol. 44, No. 18, 2005
s LSF, s SM ) monthly standard deviation of the LSF and
SM
smodel ) standard deviation derived from the model
s2model ) variance estimated from the model
sR ) laboratory reproducibility
s2res ) residual variance
SM ) silica modulus
SMcurr ) value of the SM in the current sample
SMtarget ) target value of the SM
std LSF KF, std SM KF ) standard deviation of the raw
meal LSF and SM in the kiln feed
std LSF mill, std SM mill ) standard deviation of the raw
meal LSF and SM in the mill outlet
stdev LSF clay ) standard deviation of the clay LSF
t ) time, given in minutes in the case of the mill system
and in hours in the case of the silo system
td ) delay time, min
ti ) time in the I point of partitioning of the total time
interval
ts ) sampling period, min
UCL ) upper control limit
X(t) ) dimensionless value of the LSF as a function of the
time
Xmill(t) ) mill step response function
Xsilo(t) ) silo step response function
Greek Symbols
∆t ) time interval, h
τ ) time, h
τ1, τ2 ) time constants of the second-order transfer function,
given in minutes in the case of the mill transfer function,
they are given in minutes; and given in hours in the case
of the silo transfer function they are given in hours
τmill ) average residence time of the material in the mill
circuit, min
φ ) phase, deg
ω ) frequency, rad-1
Literature Cited
(1) Lee, F. M. The Chemistry of Cement and Concrete, 3rd ed.;
Chemical Publishing Company, Inc.: New York, 1971; pp 171-
174 and 384-387.
(2) Keviczky, L.; Hetthessy, J.; Hilger, M.; Kolostori, I. Self-
Tuning Adaptive Control of Cement Raw Material Blending.
Automatica 1978, 14, 525
(3) Swain, A. K. Material Mix Control in Cement Plant
Automation. IEEE Control Syst. Mag. 1995, 15, 23.
(4) Banyasz, C.; Keviczky, L.; Vajk, I. A Novel Adaptive Control
System for Raw Material Blending. IEEE Control Syst. Mag. 2003,
23, 87.
(5) Lee, F. M. The Chemistry of Cement and Concrete, 3rd ed.;
Chemical Publishing Company, Inc.: New York, 1971; pp 164-
165.
(6) Bogue, R. H. The Chemistry of Portland Cement, 2nd ed.;
Reinhold Publishing Corporation: New York, 1955; pp 245-268.
(7) Kidd, I., R. Stability of the Laboratory Control Loop in
Cement Raw Meal Production. Proceedings of the 9th APCChE
Congress and CHEMECA 2002, Christchurch, U.K., Sept. 29-Oct.
3, 2002.
(8) Perry, R. H.; Green, D. Perry’s Chemical Engineers’ Hand-
book, 6th ed.; McGraw-Hill International Editions: New York,
1984; pp 22-6 - 22-7.
Received for review August 9, 2004
Revised manuscript received July 2, 2005
Accepted July 6, 2005
IE049284C