a r t i c l e
i n f o
Article history:
Received 22 May 2008
Accepted 11 March 2009
Available online 19 March 2009
Keywords:
Rotary cement kiln
Neural networks
Stress distribution
Shell deformation
Finite element method
a b s t r a c t
We present a new approach to the fast determination of structural deformations and stresses in the
refractory-reinforced body of a cement rotary kiln. The proposed approach builds on a comprehensive
neuro-nite element simulation of the kiln shell. Three-dimensional stresses and deformations in the
rotating tubular shell are rst determined for a nite number of input vectors using a validated nite element model of the kiln. The resulting data are then used to train a Multi-Layer Perceptron (MLP) Neural
Network which would predict accurately enough values of stresses and deformations throughout the
kiln body for any given input vector. The resulting neural simulator would serve as a replacement for the
computationally expensive cost-function evaluators that are traditionally used in numerical optimization
algorithms. To demonstrate the applicability of the proposed approach, we analyze a typical rotary kiln
using the Neuro-FE method and compare the results with those obtained from traditional method.
2009 Elsevier Ltd. All rights reserved.
1. Introduction
A rotary cement kiln is basically a giant tubular shell structure
with a non-stop, very low rotational speed which is supported by a
number of rolling supports unevenly spaced along the shell. Raw
material (clinker) enters the kiln from one end and travels the
whole length of the nearly horizontal kiln towards the hot end
where the torch is located. The baked material then leaves the
kiln for the cooler where it loses its heat and is made ready for
further processing.
In a uniform-section rotary kiln (as opposed to a variable-section one in which the diameter varies along the shell) the tubular
structure would comprise multiple shell segments of the same
diameter but of various thicknesses welded together to form a
multi-thickness cylindrical shell (Fig. 1).
As the gure shows, the cylindrical body of the kiln is enfolded
at multiple (usually 3 or 4) points by thick steel rings whose inner
diameters exceed the external diameter of the kiln body by some
200400 mm. The rings are attached to the body via a number of
circumferentially spaced radial connectors that operate as torsional shock absorbers. The rings, in turn, are supported by pairs
of rollers symmetrically located under the kiln body [1,2] (Fig. 2).
Maximum deformation in the body occurs in two regions: at
support points where the circular cross-section changes to an oval
one, and between supports where the cross-section remains fairly
circular but the shell deects downwards.
* Corresponding author. Tel.: +98 912 5273015; fax: +98 21 8801 3029.
E-mail address: kpazand@gmail.com (K. Pazand).
0261-3069/$ - see front matter 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.matdes.2009.03.016
To insulate the high-temperature (up to 1600 C) baking chamber and to guard the steel shell against overheating, the inner surface of the shell is reinforced with a refractory lining of variable
thickness. Also in the baking zone of the kiln, a coating layer gradually builds up throughout the life of the kiln. These two layers not
only add to the dead-load (weight) of the shell, but also contribute
to the rigidity of the structure and hence affect the magnitude and
distribution of stresses along the shell body [3,4].
3469
TIRE 1
TIRE 2
TIRE 3
Material inlet
Body
Material outlet
Fig. 1. Schematic of a uniform-section rotary kiln.
3. FE Modeling
This section elaborates on the construction of the FE model and
how its boundary conditions and loadings are set in order to determine the distribution of stress and deformation along the kiln
shell.
3.1. Modeling the shell, the lining, the coating and the passing
material
Fig. 2. Typical kiln rings and supporting rollers.
Table 1
General specications of the sample kiln.
Magnitude
Valve
Unit
71.5
4.8
3
4
30
Meter
Meter
%
Degree
Table 2
Various zones along the kiln.
Region
From (mm)
To (mm)
1
2
3
4
5
6
7
8
9
1000
2000
5000
6000
7000
8000
10,000
10,600
11,440
1000
2000
5000
6000
7000
8000
10,000
10,600
Coating PRESENCE
28
28
28
28
28
28
50
65
65
250
250
250
250
250
250
250
250
250
3470
1
cosh=2
r
cosa g
P r q0 g 1
cos 2h
cos 2h
l sin g
cosa g
cos a
This, when added to the weight of the refractory itself, makes the
total load on the shell:
Fig. 4. Individual load and temperature distributions along the sample kiln.
3471
PTotal
8
h
>
< 1 pt qgp g q0 gr l sing cos 2 1
2
cosa
>
:
1
2
ptqgp g
cos 2h
9
>
=
a 2h 6 g 6 a 2h
cos ag
>
g in the other point ;
12
Min.stress
Stress(MP)
10
Max.stress
8
6
4
2
0
0
10
20
30
40
50
60
70
Length (m)
Fig. 6. Boundary conditions for the FE model.
80
3472
Deflection
Deflection(mm)
2.5
Point
Results
FEM
EXP.
Error
1.5
0.5
Stress (MPa)
Displacement (mm)
Stress (MPa)
Displacement (mm)
Stress (MPa)
Displacement (mm)
9.86
7.89
9.37
5.03
11.6
4.3
10.3
8.2
9.81
5.8
12.5
4.7
0.044
0.039
0.046
0.150
0.077
0.090
10
20
30
40
50
Length(m)
60
70
80
tor which would be employed as the evaluation engine of the optimization algorithm (see Table 3).
4. Neural approximation
4.1. The inevitability of approximation
The purpose of this research is to provide fast and reliable
means for the optimum design of cement rotary kilns. These kilns
are basically tailor-made for specic working conditions and
therefore virtually no two kilns are the same from a thermomechanical viewpoint. This means that experimental data from
one existing kiln cannot be reliably used to predict the exact thermo-mechanical behavior of another; and simulation is the only
affordable means the designer has at his/her disposal in the design
phase.
As mentioned earlier, on the other hand, the traditional simulation methods (such as FEM itself) are costly and tediously timeconsuming. Since most optimization algorithms require that the
simulated system be analyzed many times at each iteration of
the algorithm,1 it would be virtually impossible to use the simulated system and one would have to resort to an approximation
technique that could predict, quickly and accurately enough, the
stress and deformation quantities at any given point along the
body shell. To achieve this, we propose the use of a neural simula-
1
For instance, in a typical run of a simple, unconstrained Genetic Algorithm with a
population size of 100 running for 1000 generations, the model has to be modied
and analyzed 105 times.
5. Analysis results
As mentioned earlier, a total of 90 (inputoutput) data sets
were used to train and test the constructed neural network; 70 sets
for training the network and 20 sets for testing it. Once the training
was concluded, an overall Mean Square Error of less than 4% was
achieved over the test data sets; meaning that the neural net
was now capable of fairly accurately replicating the numerical
model of the kiln.
To get a avor of how reliable the neural approximation could
be, results from the FE model and the trained neural net are compared in Table 5 for three randomly selected input vectors. The
comparison shows that the network has a good generalization
Table 4
Characteristics of the constructed network.
Momentum
Rate learning
Average error
0.08
0.0323
3473
FEM
ANN
Error
Stress (MPa)
Displacement (mm)
Stress (MPa)
Displacement (mm)
Stress (MPa)
Displacement (mm)
9.86
7.89
9.37
5.03
11.6
4.3
11.1
8.1
9.6
5.6
12.01
4.2
0.12
0.026
0.024
0.1
0.035
0.02
10.2
10.2
10
10
9
9
43.6
43.6
42.4
42.4
44.2
44.2
64.6
64.6
62.6
62.6
63
Table 6
Time comparison of FEM and neural simulation results.
Description
FEM
ANN
Time of modeling
Time of running
Time of new modelling with variation in ring
120140 h
13 h
12 h
2060 h
0.00030.0005 h
0.060.09 h
power which means it is well capable of replicating the FE simulation and hence predicting the maximum stress and deection at
any point along the shell for unprecedented input vectors.
Solving the validated neural model of the kiln body (complete
with its refractory, coating layers and the passing material), the
distribution of the Von-Mises equivalent stress on the kiln shell
for the fore-mentioned loading and boundary conditions was
determined and is illustrated in Fig. 5.
6. Summary and conclusions
A new methodology for the fast and reliable determination of
the performance characteristics of rotary kilns was presented.
The methodology builds on the following postulations.
(1) No numerical model of the kiln shell will be able to yield
realistic results unless it takes into account the effect of
refractory and coating layers on the rigidity of the structure
and the distribution of stress along it.
(2) An accurate approximation of the simulation results is inevitable if the global optimality of the nal design is not to be
compromised for the computational cost of the newer, more
efcient optimization tools.
Based on these, a complete model of the kiln body was created
and validated using data from a local cement plant. The validated