Professional Documents
Culture Documents
Abstract
The finite element analysis of the cold flat rolling process is well established. However, the requirement of large computational time
makes it unsuitable for online applications. Recently, there have been some applications of modeling the rolling process by means of
neural networks. In most of the previous works, trained networks predict only roll force and roll torque. The input data for training the
neural network have been obtained either through experiments or from finite element method (FEM) code. In this work, the neural
networks have been used for predicting the velocity field and location of neutral point. The training data are obtained from a rigid-plastic
finite element code. The trained network provides a suitable guess for the velocity field and location of the neutral point, that is further
refined by the finite element code. The post-processor of the FEM code computes roll force, roll torque, strain distribution, etc. This
procedure provides highly accurate solution with reduced computational time and is suitable for on line control or optimization.
r 2006 Elsevier Ltd. All rights reserved.
Keywords: Cold rolling; Neural network; Radial basis function; Finite element method; Deformation field; Rigid plastic
0952-1976/$ - see front matter r 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engappai.2006.10.001
ARTICLE IN PRESS
44 P.P. Gudur, U.S. Dixit / Engineering Applications of Artificial Intelligence 21 (2008) 43–52
improve the prediction ability for rolling force is another where the strain-rate tensor is given by
domain that has attracted the attention of researchers
1 qvi qvj
recently (Lee and Choi, 2004; Son et al., 2005). In most of _ij ¼ þ , (4)
these works, neural networks are trained using the data 2 qxj qxi
obtained by running an FEM code. These trained networks and for material yielding according to von Mises criterion,
have the capability to predict roll force and roll torque. the proportionality factor m is given by (Malvern, 1969)
They do not predict stress, strain and strain-rate fields, that sy
are very much required for the optimization of the process. m¼ . (5)
3~_
In view of this, the present work is an attempt to use neural
networks in conjunction with FEM for the faster and In the above expression, _~ is the equivalent strain defined as
detailed analysis of cold flat rolling process. Here, the rffiffiffiffiffiffiffiffiffiffiffiffi
2
radial basis function neural networks trained by the data _~ ¼ _ij _ij , (6)
obtained from an FEM code have been used for predicting 3
a highly accurate guess of a complete velocity field and the and sy is the flow stress in uniaxial tension or compression.
location of neutral point, which is then used for finding out Neglecting the effect of temperature and strain rate but
the refined velocity and pressure field using the FEM code. considering the strain-hardening effect by a power law, the
The refined velocity and pressure field is post-processed for flow stress is given by
computing stress, strain and strain-rate fields as well as roll
~ n
force and roll torque. The post-processing procedure is sy ¼ ðsy Þ0 1 þ . (7)
same as in FEM. b
Requirement of the huge amount of dataset is considered Here (sy)0, b and n are uniaxial yield stress and material
as one of the drawbacks of neural network modeling. hardening parameters respectively. The equivalent strain ~
Hence, in the present work, attention has been paid to can be obtained by integrating ~_ along the particle path,
develop a scheme, that requires less number of training and Z t
testing data. In order to have faster training, neural ~ ¼ ~_ dt. (8)
networks based on radial basis function have been used. 0
The performance of the scheme has been studied and it In the present work, rectangular elements are used to
indicates the promising application for online prediction discretize the domain, with bi-quadratic approximation for
and optimization. velocity components and geometry and bi-linear approx-
imation for pressure. Fig. 1 shows the mesh consisting of 56
2. Finite element formulation elements with 261 nodes for velocity components and 75
nodes for hydrostatic pressure. The inlet and exit velocities
The rigid-plastic finite element formulation (Dixit and U1 and U2, respectively, along AF and DE are related as,
Dixit, 1996) is briefly described in this section. For the U 1 ¼ U 2 ð1 rÞ, (9)
steady-state process of rolling, a mixed pressure–velocity
formulation is used. In the present work, friction is modeled where r is the fractional reduction. The roll velocity
as Coulomb’s model, although Dixit and Dixit (1996) have corresponding to U2 is obtained as the velocity at the
used Wanheim and Bays’a model. It has been observed that neutral point. The neutral point, in this formulation is
in the case of thick strip rolling both the models provide found by minimizing the total power with respect to the
approximately same results, albeit in the case of foil rolling position of neutral point. On the surface, AF, v1 ¼ U1 and
there is a large difference in prediction of various parameters v2 ¼ 0, where v1 and v2 are the components of velocity in x1
by both models (Kumar and Dixit, 2006). and x2 directions, respectively. On the plane of symmetry
For the steady-state rolling process, the Eulerian form of FE and on the surface AB and CD, v2 ¼ 0, t1 ¼ 0, where t1
the continuity and momentum equations is is the traction component in longitudinal direction. On the
roll–strip interface (boundary BC), the normal component
qvi of velocity, vn ¼ 0 and frictional traction ts is related to
¼ 0, (1)
qxi normal traction tn by Coulomb’s friction model subjected
to a limit on its maximum value as
qp qS ij 8
þ ¼ 0, (2) s
qxi qxi < mjtn j if mjtn jp pyffiffi3
jts j ¼ s s , (10)
where vi is the velocity vector at the point with Cartesian : pyffiffi3 if mjtn j4 pyffiffi3
coordinate xi, p is the hydrostatic stress and Sij the
deviatoric stress component. In rigid-plastic material, the where m is the coefficient of friction and ts and tn are the
deviatoric part Sij of the stress tensor sij is related to _ij by tangential and normal components of traction vector t
the relation (Malvern, 1969) defined as
S ij ¼ 2m_ij , (3) ti ¼ sij nj . (11)
ARTICLE IN PRESS
P.P. Gudur, U.S. Dixit / Engineering Applications of Artificial Intelligence 21 (2008) 43–52 45
α
X2
v2= 0
A t1= 0 B v0= 0
v2= 0
ts = tn
1 6 C t1= 0 D
4
s
h1/2
2
v1= U1 h2/2
5
v2= 0
3 v1= U2
v2 = 0
X1
F E
3h1/2 3h2 /2
input vector and emits the output, which is a function of 3.1. Study of data
the Euclidian distance of the input vector and the center
associated with that particular neuron. This function is a The range of the input variables for the modeling of
radial basis function having radial symmetry. The radial neural network is taken as follows:
basis function used in this work is Gaussian function (bell-
R=h1 : 102100; m : 0:0420:14; r : 0:0420:24; b
shaped) of the following form
: 0:0220:08; n : 0:00120:6
2 2
hðxÞ ¼ exp ðx =s Þ, (15) For studying the influence of input parameters, a full
factorial analysis with 25 ¼ 32 dataset was carried out, in
where x is the Euclidian distance between the center and
which each parameter is taken at two levels— low and
input vector, s is known as the spread parameter, that high. As shown in Fig. 1, the metal strip is constructed with
controls the domain of influence of the radial basis a mesh of 56 elements, which consists of 261 nodes with
function. two degrees of freedom each. The study was carried out for
The weighted addition of the outputs of neurons of the six nodes, shown by black dots, in Fig. 1. These are at
hidden layer is sent to the output neuron, which processes different locations in the plastic zone.
it by adding a constant term called bias, b. Thus the output After getting the nodal velocity data for 32 datasets by
of the RBF neural network is approximated as
FEM, the data were studied using SPSS version 12.0.1
X
m software. The correlation analyses and analysis of variance
f ðxÞ ¼ wj hj jjx cj jj þ b, (16) (ANOVA) is carried out to find out critical input variables,
j¼1 which may be significant for predicting velocity field.
Tables 1 and 2 show some results of correlation analysis.
where x is the input vector, hj(U) is the processing function For node 6, it is observed that significant correlation
of the jth neuron in the hidden layer, ||U||2 denotes occurs for R/h1 and r for v1 as well as v2. The material
Euclidean norm, wj are the weights associated with the hardening coefficient b is found to be insignificant for both
jth neuron in the hidden layer, m is the number of neurons v1 and v2. It is observed that at node 3, there is a change in
in the hidden layer and cj are the RBF centers obtained the trend for v1 and v2. Parameter r makes significant
from the input vector space. For fixed centers and spread correlation for v1 but does not make significance for v2 and
parameter, applying Eq. (16) for each training data R/h1 parameter makes significant correlation for v2 but not
provides a linear system of equations with m unknowns for v1. However, like node 6, strain hardening coefficient b
wj and one unknown b, which can be solved using least is insignificant at node 3. In other four nodes also, b was
square method. Hence, the training procedure takes much found to be insignificant.
less time compared to multi-layer perceptron network. ANOVA was carried out for the six nodes. Table 3
Due to faster training procedure, the present work shows the p-values of v1 and v2, up to three decimal places,
employs RBF neural networks for the estimation of nodal based on F-statistics for 6 nodes. The p-value provides an
velocities as a function of R/h1, m, r, b, n. The number of objective measure of the strength of evidence, which the
nodal variables in a typical problem is quite large; hence parameter supplies against the null hypothesis that the
the reduction of training time is of prime importance. Also, parameter is significant for prediction of velocity. From the
the number of neurons in the hidden layer should be kept table it is observed that parameters R/h1, r and m may be
as small as possible in order to have less prediction time. considered significant at 1% significance level. For some
Thus, before deciding the architecture of the network, a nodes n is found significant at 10% significance level. On
thorough study of the data was carried out in order to the other hand, strain hardening parameter b is not found
understand the dependence of nodal velocities on the five significant at even 10% significance level. Thus, the strain -
input variables. This is described in the next subsection. hardening does not seem to have influence on the velocity
Table 1
Correlation analysis for velocity v1 and v2 (node: 6)
Parameter R/h1 m r b n v1 v2
Table 2
Correlation analysis for velocity v1 and v2 (node: 3)
Parameter R/h1 m r b n v1 v2
Table 3
The p-values based on F-statistics for v1 and v2
Node R/h1 m r b n
v1 v2 v1 v2 v1 v2 v1 v2 v1 v2
1 0.769 0.021 0.713 0.232 0.002 0.143 0.173 0.120 0.270 0.099
2 0.068 0.105 0.253 0.293 0.001 0.000 0.737 0.678 0.140 0.100
3 0.015 0.000 0.008 0.682 0.003 0.019 0.620 0.729 0.971 0.120
4 0.005 0.000 0.003 0.767 0.001 0.000 0.855 0.928 0.437 0.400
5 0.004 0.001 0.001 0.418 0.000 0.000 0.397 0.596 0.419 0.904
6 0.000 0.000 0.017 0.861 0.001 0.000 0.837 0.693 0.795 0.554
distribution. In an earlier paper on FEM analysis of wire and Dixit (2005). In this work, eight random testing
drawing process, Dixit and Dixit (1995) also observed the datasets were chosen for testing.
same. The dimensional values of variables lie in different
Further, the effects of various parameters were evaluated ranges. Hence, normalization has been carried out for these
as per the following formula (Dieter, 1991): input variables such that their values lie from 0.1 to 0.9. In
the present work, centers are chosen randomly. The
Effect of parameter ¼ MATLAB function NEWRB was used for the training of
P P
response at high levels response at low levels the network. The function works on the algorithm of
. ð18Þ
half the number of runs in the experiment starting with zero neurons in the hidden layer and keep on
adding the input vector with greatest error as center in
It was found that for all nodes, the effect of b is least
successive runs till the error goal is achieved. Based on the
followed by n. Based on these observations, it was decided
testing error, the error goal of a sum-squared error of
to develop one type of network models with R/h1, m, r, n as
0.0001 was found to provide a very good performance. The
input parameters and other type of networks with R/h1, m, r
first network that was trained by 24 datasets required 22
as the input parameters.
neurons in the hidden layer to achieve the error goal of
0.0001, whilst the network trained by 16 datasets needed 15
3.2. Neural network (NN) modeling neurons in the hidden layer to reach the error goal.
Spread parameter is another important parameter for
Two types of RBF network architectures were tried out constructing an RBF neural network. The following simple
in the present work. The first type of networks is trained by relationship is available in the literature (Haykin, 1996) to
24 datasets, out of which 16 datasets correspond to the calculate the spread parameter
corners of four-dimensional input space and eight datasets
d max
are randomly chosen. Neural networks behave poorly in s ¼ pffiffiffi , (20)
extrapolation. Therefore, it is essential to include in the J
training dataset the data corresponding to the corners of where dmax is maximum Euclidean distance between the
the input space. This ensures that neural networks centers and J is the number of center. However, as this
interpolate during the prediction. The second network relation is a heuristic relation, to fine-tune the value of
was trained in the similar way with total 16 datasets. The spread parameter, the value calculated by Eq. (20) is
size of testing (or cross-validation dataset) is very multiplied by a constant factor b. A number of numerical
important from the viewpoint of reliability of the network. experiments were carried out to find out the optimum b. In
Some guidelines regarding this have been provided in Kohli most of the cases, the optimum b was obtained as 1.5.
ARTICLE IN PRESS
48 P.P. Gudur, U.S. Dixit / Engineering Applications of Artificial Intelligence 21 (2008) 43–52
Table 4 shows the sensitivity of testing errors with b for a neutral point, the non-dimensionalization being done with
typical nodal velocity (at node 4 of Fig. 1) in the the total contact arc length. The network predicted
deformation zone. Although the variation of percentage information is supplied as the initial guess to the FEM
root mean squared fractional error and maximum percen- code. At the beginning, no roll flattening is considered and
tage error is not unimodal with b, both error measures the pressure and modified velocity fields are calculated. The
provide minimum value at b ¼ 1.5. Observation of the post-processing module of the FEM code calculates the roll
sensitivity of testing errors with b for about 25 nodal force, knowing which the deformed roll radius is calculated
velocities, led to the conclusion that either the errors are using Hitchcock’s formula. With the deformed roll radius,
the least or very near to the least errors for b ¼ 1.5. FEM code is run again to obtain new pressure and velocity
Once the networks are trained and best fitted for least fields and deformed roll radius. This procedure is repeated
effective error, velocity field of any given input data can be till the roll radius converges within 1% accuracy. Only 2–3
found out. This complete information of velocity field iterations of FEM code are needed. After that, the roll
predicted by RBF neural networks (one network for each torque is calculated as described earlier in Section 2. Roll
velocity component) can be utilized for further calculation force, roll pressure distribution, stress–strain field, strain-
of strain rate, stress, pressure field along with roll force and rate field, etc. can be found by the post processing module
roll torque that is the main task of the present work. of the FEM. A general overview of this procedure is shown
Similarly, a RBF neural network for predicting the location in Fig. 3.
of neutral point is also modeled. Here, the output neuron
corresponds to arc length from exit to neutral point divided 4. Results and discussion
by the total arc length. It was observed that this ratio does
not depend significantly on hardening parameters. In the present work, radial basis function neural
networks have been implemented for the modeling of cold
flat rolling process, which predicts velocity field and
3.3. Prediction of the parameters using the trained neural
location of neutral point. These predicted values have
network and FEM
been used as guess values for FEM code based on rigid-
plastic finite element formulation (Dixit and Dixit, 1996).
A number of neural networks are trained for predicting
The velocity field that is in the form of nodal velocity
the components of velocity at different nodes. One network
values and the location of neutral point is further refined in
predicts the non-dimensionalized arc length from exit to
the FEM code to generate the roll force, roll torque and
strain rate. From the result of statistical analyses of present
Table 4
Variation of testing error with b for a typical nodal velocity prediction work, it has been observed that strain hardening para-
meters have little effect on determination of the velocity
b % rms fractional error max. % error field. Therefore, in the first model, strain-hardening
0.5 5.22 8.32
parameter b is excluded whereas in the second model both
0.75 2.06 7.70 b and n are excluded.
1 1.19 2.51
1.25 4.07 4.86 4.1. Validation of the NN model
1.5 1.10 2.09
1.75 1.61 3.40
2 2.07 4.02
Two different types of neural network models are
studied. The neural network model with four input neurons
Velocity field
Trained
FEM model
RBF NN model
Location of
neutral point
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
was trained with 24 training datasets and cross-validated uN
uP 2
(tested) with eight datasets. The performance of the neural u jei j
ti¼1
network was assessed using 32 random validation datasets. Error 2 ¼ , (22)
These validation datasets were generated within the ranges N
of the process parameters that are used for generating
training and testing datasets. The trained network was used Error 3 ¼ max jei j, (23)
for predicting the output for these datasets. The prediction
time was less than 5 s. Finite element code would have where ei is the difference between the nodal velocity value
required about 30 min for the same task, in a Pentium IV of FEM code and NN model and N is the total number of
processor with a compiler that optimizes the code. These nodes (261 in the present case) in the finite element mesh.
computation times are for the mesh shown in Fig. 1. If a Figs. 4(a) and (b) show the errors in network prediction for
refined mesh having the total velocity nodes n four times different datasets. Here, the non-dimensional horizontal
the existing nodes is employed, FEM computation time exit velocity is taken as 1. The Error 1 and Error 2 are less
varying approximately as o(n3), will increase by 64 times, than 2% and the maximum error is less than 7% of the exit
whereas the velocity prediction by neural network varying velocity in both the horizontal and vertical components of
at most as o(n2) will increase only 16 times. Thus, as the velocity (v1 and v2).
mesh refinement increases, computational gain obtained The second network has been modeled without con-
with the assistance of neural networks increases. sidering b and n. Thus, the strain hardening behavior is
The performance of the neural network modeling has excluded in the modeling of velocity field. It is observed
been checked for three different error criteria, which are that the non-inclusion of strain-hardening parameters does
termed as Error 1, Error 2 and Error 3: not deteriorate the prediction capability of the network for
the prediction of velocity field, as apparent from Fig. 5.
X
N
jei j The statistical analysis led to the conclusion that strain-
Error 1 ¼ , (21) hardening parameter n is of minor influence on the
N
i¼1 prediction of the velocity field. However, after fitting the
a 0.07 a 0.07
0.06 0.06
0.05 0.05
Errors
Errors
0.04 0.04
0.03 0.03
0.02 0.02
0.01 0.01
0 00 5 10 15 20 25 30 35
0 5 10 15 20 25 30 35
Validation Data Validation Data
b 0.07 b 0.07
0.06 0.06
0.05 0.05
Errors
Errors
0.04 0.04
0.03 0.03
0.02 0.02
0.01 0.01
0 0
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
Validation Data Validation Data
Fig. 4. Errors in four input neural network prediction for (a) horizontal Fig. 5. Errors in three input neural network prediction for (a) horizontal
velocity component, (b) vertical velocity component. velocity component, (b) vertical velocity component.
ARTICLE IN PRESS
50 P.P. Gudur, U.S. Dixit / Engineering Applications of Artificial Intelligence 21 (2008) 43–52
networks, it is observed that strain-hardening parameter n calculates very accurate location of the neutral point in
can be excluded from the input parameters of NN model only 2–3 iterations.
along with b, leaving only three parameters viz. R/h1, m, r
required for obtaining the velocity field. This also reduces
4.2. Prediction of the roll force and roll torque
the requirement of training data as the requirement of
training datasets increases with increasing number of input
The NN predicted velocity field has been used to
neurons.
calculate the roll force per unit width of strip Fr, roll
Another network has been modeled with RBF function
torque per unit width T and equivalent strain by providing
neural network for predicting the location of the neutral
this field as a guess field in the FEM code. The refined
point. Here, a maximum of 10% deviation from that of
velocity and pressure fields are obtained by carrying out
FEM results is observed. This much accuracy in the
2–3 iterations of FEM code. Fig. 6 compares the roll force
prediction of neutral point is enough for providing a
and roll torque obtained by NN assisted FEM with those
suitable guess to FEM code. Taking this guess, FEM
obtained by the FEM code directly. It is observed that
predicted values of roll force and roll torque are almost
a 4
same as obtained by FEM code alone. The maximum
deviation in roll force prediction is 3% and in roll torque
prediction 2%. NN assisted FEM takes about one tenth of
3
the time taken by the FEM code. It is estimated that if a
refined mesh with 4 times the velocity nodes is used, the
Fr by NN assisted
FEM (MN/m)
NN assisted FEM will take less than 1/40 times the time
2
taken by the unassisted FEM code. Although the refine-
ment of this order is not required just for roll torque and
force prediction, it may be desirable when a highly accurate
stress field needs to be predicted for optimizing the process.
1
2 hardening material.
Fig. 7 shows the equivalent-strain contours of the high
1 strain-hardening material during the rolling process. Con-
tours in Fig. 7(a) have been obtained by running FEM code
till all the nodal velocities converges to within 0.01%
0
accuracy. On the other hand, contours in Fig. 7(b) have
been obtained by NN assisted FEM code, in which FEM
0 code refined the NN predicted velocity field by carrying out
0 1 2 3 4
three iterations. It is seen that both the contours are similar.
T by FEM (kN-m/m)
There is about 10% difference in the maximum strain. This
Fig. 6. NN assisted FEM predictions versus FEM predictions for (a) roll difference can further be reduced if a few more iterations of
force, (b) roll torque. NN-assisted FEM code are carried out.
a b
3
3 11 11
Level Strain Level Strain
10 10
11 0.20 11 0.18
5
7
5 8 9 0.18 9 0.16
1 7 0.15 7 0.13
9
4 5 0.13 5 0.09
1
8
3 0.09 3 0.04
3
1 0.02 9 1 0.01
3
Fig. 7. Equivalent strain contours for high strain-hardening material (b ¼ 0.02, n ¼ 0.6) plotted by (a) FEM code, (b) NN-assisted FEM code.
ARTICLE IN PRESS
P.P. Gudur, U.S. Dixit / Engineering Applications of Artificial Intelligence 21 (2008) 43–52 51
a b
5
11 Level Strain 11 Level Strain
3
2 10 10
11 0.20 11 0.21
9
6 9 0.17 9 0.17
5
9 7 0.16 7 0.14
5 0.14 5 0.10
7
3 0.02 3 0.06
8
1 0.01 1 0.02
Fig. 8. Equivalent strain contours for low strain-hardening material (b ¼ 0.08, n ¼ 0.001) plotted by (a) FEM code, (b) NN-assisted FEM code.
Li, G.J., Kobayashi, S., 1982. Rigid plastic finite element analysis of plane Reddy, J.N., 1993. An Introduction to the Finite Element Method, second
strain rolling. ASME Journal of Engineering for Industry 104, 55–64. ed. McGraw-Hill, New York.
Liu, C., Hartley, P., Sturgess, C.E.N., Rowe, G.W., 1985. Elastic-plastic Son, J.S., Lee, D.M., Kim, I.S., Choi, S.G., 2005. A study on on-line
finite element modeling of cold rolling of strip. International Journal of learning neural network for prediction for rolling force in hot-
Mechanical Science 27, 531–541. rolling mill. Journal of Materials Processing Technology 164–165,
Malvern, L.E., 1969. Introduction to the Mechanics of a Continuous Medium. 1612–1617.
Chapter 6, Section 6.5 and 6.6. Prentice-Hall, Englewood Cliffs, NJ. Yang, Y.Y., Linkens, D.A., Talamantes-Silva, J., 2004. Roll load
Mori, K., Oskada, K., Oda, T., 1982. Simulation of plane strain rolling by prediction-data collection, analysis and neural network
rigid plastic finite element method. International Journal of Mechan- modeling. Journal of Materials Processing Technology 152,
ical Science 24, 519–527. 304–315.