See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/225846261

An optimization method for radial forging process using ANN and Taguchi
method

Article  in  International Journal of Advanced Manufacturing Technology · February 2009

DOI: 10.1007/s00170-008-1371-2

CITATIONS READS

36 283

3 authors:

Mehdi Sanjari Ali KARIMI Taheri

NanoPhyll Inc. Sharif University of Technology
70 PUBLICATIONS   721 CITATIONS    123 PUBLICATIONS   2,754 CITATIONS   

SEE PROFILE SEE PROFILE

Mohammad Reza Movahhedy

Sharif University of Technology
149 PUBLICATIONS   2,027 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Design and Fabrication of High-Speed Milling spindle View project

Excellent energy absorption capacity of nanostructured Cu-Zn thin-walled tube View project

All content following this page was uploaded by Mehdi Sanjari on 14 October 2014.

The user has requested enhancement of the downloaded file.

Int J Adv Manuf Technol
DOI 10.1007/s00170-008-1371-2

ORIGINAL ARTICLE

An optimization method for radial forging process

using ANN and Taguchi method
M. Sanjari & A. Karimi Taheri & M. R. Movahedi

Received: 24 October 2006 / Accepted: 3 January 2008

# Springer-Verlag London Limited 2008

Abstract In this study, the artificial neural network (ANN) Keywords ANN . FEM Analysis . Inhomogeneity .
and the Taguchi method are employed to optimize the radial Process optimization . Radial forging . Taguchi method
force and strain inhomogeneity in radial forging process.
The finite element analysis of the process verified by the Nomenclature
microhardness test (to confirm the predicted strain distri- RF Radial force
bution) and the experimental forging load published by the k Yield shear stress of the part
previous researcher are used to predict the strain distribu- L1 Length of the sinking zone
tion in the final product and the radial force. At first, a L2 Length of the forging zone
combination of process parameters are selected by orthog- L3 Length of the sizing zone
onal array for numerical experimenting by Taguchi method m Friction shear factor
and then simulated by FEM. Then the optimum conditions Rm Radius of the mandrel
are predicted via the Taguchi method. After that, by using Rn Outer radius of tube at the neutral plane
the FEM results, an ANN model was trained and the R0 Outer radius of the preform
optimum conditions are predicted by means of ANN (using R1 Outer radius of the element at sinking-forging zones
genetic algorithm as global optimization procedure) and interface
compared with those achieved by the Taguchi method. The R2 Outer radius of the product
optimum conditions are verified by FEM, and good t0 Initial wall thickness of the preform
agreement is found between the two sets of results. t Final wall thickness of the element
Vr Absolute radial velocity of dies
α Die angle
s Flow stress of the material in compression
M. Sanjari (*)
Materials and Energy Research Center,
Materials Science and Engineering,
Sharif University of Technology, 1 Introduction
Azadi Ave, Tehran 96941, Islamic Republic of Iran
e-mail: [email protected]
Radial forging is a hot or cold forging process utilizing two
A. K. Taheri or more radially moving anvils, or dies, with the concept of
Department of Materials Science and Engineering, incremental forging generally employed to obtain the
Sharif University of Technology, desired shape from the initially simple round or square
Azadi Ave, Tehran 96941, Islamic Republic of Iran
e-mail: [email protected] billet [1]. The advantages of radial forging are smooth
surface finish, considerable material or weight savings,
M. R. Movahedi preferred fiber structure, minimum notch effect, and increased
Department of Mechanical Engineering, material strength [2, 3]. In order to obtain optimal properties
Sharif University of Technology,
Azadi Ave, Tehran 96941, Islamic Republic of Iran of the final radial forged product, it is necessary to
e-mail: [email protected] investigate many areas such as determination of the required
 Int J Adv Manuf Technol

Fig. 1 Schematic representation I Sinking

of radial forging for tubes and II Forging
variables III Sizing
Vr

t0
I
α II
III
R0
R1
R2 Rn
Rm

L1 L2 L3

process, die design, and the selection of the process inhomogeneity in radial forging process and the effect of
condition. In industry, however, the conventional approaches various process parameters on its magnitude have not yet
to reach an optimal process condition are based on empirical been considered broadly. It should be noted that although
or approximate analysis, requiring a time consuming and the simulations based on the FEM method are capable of
expensive trial-and-error [4]. In the literature, there are some predicting many variables such as radial force, strain, strain
numerical or analytical models to assess the radial forging rate and stress distributions more accurately, but those are
process. For example Lahoti et al. [5, 6] analyzed the time consuming and computationally expensive for a com-
mechanics of radial forging process for single and compound pletely optimization procedure. Therefore in this work,
angle dies using slab method. Subramanian et al. [7] statistical design of experiments is used to get information
modeled the metal flow in die cavity in radial forging for such as main effects of design parameters from the minimum
rifling of gun barrels under the plane strain condition. Tszeng number of experiments or, in our case, the minimum number
and Kobayashi [8] were the first to model the tube forging of simulations. The objective of parameter design in the
using the finite element method (FEM). Domblesky et al. [9, statistical design of experiments is to find the best combina-
10] have considered a model of the tube-forging process tion of controllable factors. The Taguchi method, which is one
based on the rigid-thermo viscoplastic finite-element method. of the fractional factorial designs, has a good reappearance of
Jang and Liou [11] used a 3-D non-linear finite-element experiment or simulation concerned only with the main effects
method to study the residual stress distributions in radial of design parameters. Therefore, in this study, in order to get
forging. Lee et al. [12] proposed a new numerical scheme of the optimal process parameters in a short time the following
the finite element method, the automatic expansion of steps was applied:
domain (AED), to reduce computation time and computer
1. The orthogonal array for numerical experimenting is
memory. Ghaei et al. [13] used the finite element method to
determined by Taguchi method.
model the radial forging process and to study the effects of
2. The combinations of process parameters are selected by
die shape in the cross-section area on deformation. Some
orthogonal array, simulated by FEM.
other researchers [14–18] also proposed numerical model for
3. The optimum process parameters to provide minimum
radial forging process.
inhomogeneity and radial force were obtained via the
Inhomogeneous deformation is one of the most impor-
Taguchi method.
tant parameters in the forming process since it affects the
4. An ANN model using the FEA results obtained in 2 is
final properties of product such as the distribution of grain
trained.
size and grain boundary after recrystallization in final
5. Using an ANN model, we predicted the optimum
annealing [19, 20]. Although the inhomogeneity of defor-
process parameters by means of ANN (using genetic
mation in different metal forming processes has been
investigated [21, 22], but in the literature the deformation
Table 2 Experimental conditions for swaging

Table 1 Chemical compositions of investigated alloys (in wt. %) R0(mm) R1(mm) L2(mm) L3(mm) α Axial Back Front
feed push pull
Alloy Al Si Mg Mn Cu Fe Other element (mm) (KN) (KN)

in wt.% 95.8 1.4 0.6 0.1 0.6 0.6 0.3 13 7 7 5 12 15 10000 0

Int J Adv Manuf Technol

0.45
Thus, the small clearances existing between the hammer
0.4 Experimental data dies at the end of a forging stroke are neglected. Based on
4th degree this assumption, no material spreading would be allowed
0.35
between the dies at the end of the forging strike and the
0.3 deformation could be modeled axisymmetrically. To locate
Effec. Strain

0.25
in the center and maintain the workpiece position, two
chuck-heads grip the workpiece at both ends during
0.2 forging. Also these chuck-heads can apply back-push or
0.15 front-pull forces to the workpiece [8].

0.1

0.05 3 Mathematical models

0
100 105 110 115 120 125 130 135 140 145 150 3.1 Finite element method
Hardness (HV)
Fig. 2 The hardness values (experimental) versus effective strain It is known that several simulations are needed to
(FEM) and approximated fourth order polynomial
investigate the effects of process parameters on the
inhomogeneity deformation of radial forging process. In
algorithm as global optimization procedure) and com-
order to reduce the computational time, the simplifying
pared with those achieved by the Taguchi method.
assumptions may be made to analyze the radial forging
6. Under two sets of optimum condition, the FEA was
process in an axisymmetric deformation model [10].
done to verify these results.
Neglecting the rotational feed, the variation of die shape
7. The two set of optimum condition and their FEA
in the cross-section and the small clearance between the
verifications were compared together. A good agree-
dies are the assumptions used in this work.
ment between these two predicted results is found.
Domblesky et al. [9] used different methods to impose
For evaluating the FEM simulation, strain distribution in the front-pull and back-push forces on their model.
the radial forged sample is computed and compared with However, in this study, a pressure boundary condition is
the simulation results. To do this, the local values of successfully imposed at tube ends. The following steps are
hardness measured on a radial forged sample and the applied to the model: (1) the tube end is first fixed in the
relationship between the hardness and strain are utilized. axial direction before the die touched the tube; (2) the die
Finally, the accuracy of the radial forging load predicted by touched the deforming tube and a firm contact is estab-
the FEM model is verified by comparing the predicted lished; (3) the moving die is stopped; (4) the fixed end of
forging loads with experimental results published by the tube is released; and (5) finally the pressure boundary
previous researcher. condition is imposed at both ends of the tube.
To model the radial forging process, the commercially
available FEM software ABAQUS is used. The finite
2 Radial forging process element code is based on a Lagrangian formulation with
elastic-plastic material behavior. The Von Mises yield
In Fig. 1 the general model of the radial forging process for criterion and isotropic hardening are assumed for the
tubes, considered in this study, is illustrated. It is assumed model. First order quadrilateral elements were used to
that in the innermost position of the hammer dies, the inlet mesh the tube. To avoid shear locking problem associated
section of the dies form a perfect conical surface and the die with this element, a reduced integration strategy with
land is perfectly cylindrical. hourglass control was used. Due to the axisymmetric of

Fig. 3 Effective strain distribu-

tion obtained from measured
hardness (above) and simulation
(down)
 Int J Adv Manuf Technol

Table 3 Comparison loads

predicted by FEM with the Sample Billet Product Length of Length of Predicted Load from
experimental results of number diameter diameter forging zone sizing zone maximum load experiment
reference [6] (mm) (mm) (mm) (mm) (KN) (KN)

1 15.97 13.18 18.55 18 177 172

2 15.97 13.25 18.09 18 174 167
3 15.03 13.11 12.77 18 137 124
4 13.99 13.03 6.83 18 80 74.5

the process, the finite-element meshes were generated on 3.3 Taguchi method
the half cross-section of the workpiece. The type of the
element used in the model is a quadratic reduced matrix The Taguchi method, which is one of the fractional factorial
element (CAX4R). In cold forging, the process is usually designs, utilizes orthogonal arrays (OAs) to study a large
assumed to be under isothermal conditions although the number of variables with a small number of experiments or,
actual process is non-isothermal. This assumption can in this study, the minimum number of simulations.
provide good preliminary results when applied to processes, Furthermore, the Taguchi method can be used to determine
which can occur so quickly that die chilling effects are the values of the processing parameters optimizing the
small. Because in radial forging process the contact time quality characteristics [27]. Also analysis of variance
between tools and workpiece is less than 10–15% of each (ANOVA) can be applied to investigate which parameters
forging cycle an isothermal condition is assumed [9]. influence the radial forging process most significantly.
Penalty formulation is employed to model the friction in Taguchi method is employed using the following procedure
the contact surfaces. It is assumed that the limiting shear [28]:
stress is obtained by m psffiffi3, where s is the flow stress and
(i) Select the factors affecting on the performance of the
m=0.15 is the friction coefficient commonly used for the
inhomogeneity factor or radial force (RF).
cold forging conditions with phosphate lubricant [6, 10].
(ii) Select the appropriate levels that each factor may hold
within the simulation.
3.2 Neural network modeling
(iii) Select the orthogonal array to determine the number of
simulation to run.
Artificial neural networks (ANNs) are basically a data-
(iv) Conduct the simulations.
driven black-box model capable of solving highly non-
(v) Analyze the results; determine the optimum factor level
linear complex problems. Among the various kinds of
combination.
ANNs, the back propagation (BP) learning algorithm has
(vi) Predict the optimum performance and influence of
become the most popular in engineering applications
factors.
[23–25]. By providing a network of this type with a set of
(vii) Confirm the experimental design.
training data, the network is able to learn and then adjusts
the interconnection weights between layers. This process is
repeated until the network performs well on the training set
and can subsequently be used to classify or predict from 4 FEM model validation procedures
previously unseen data output. However, the data for
training the artificial neural network is needed to be 4.1 Experiments
carefully selected, to provide the wide range and well
distribution. Therefore, in this work the data are selected The 6082 aluminum alloy rods received in hot extrusion
using the orthogonal array table in the statistical design of condition were used. In this study the chemical composition
experiments proposed by Taguchi [26]. of alloy is shown in Table 1.

Table 4 Process parameters of

radial forging at four levels Parameter destination Process parameters Range Level 1 Level 2 Level 3 Level 4
used in Taguchi method
A Die angle(degree) 3–12 3 6 9 12
B Back pull force (KN) 0–75000 0 25000 50000 75000
C Forward push force(KN) 0–75000 0 25000 50000 75000
D Friction factor 0.1–0.25 0.1 0.15 0.2 0.25
E Axial feed (mm) 5–20 5 10 15 20
Int J Adv Manuf Technol

Table 5 Process parameters were used for simulation comparison of local values of strain evaluated using the
R0(mm) R1(mm) Rm(mm) L3(mm) numerical simulation of the compression test, and the local
values of hardness measured on a real specimen can be
150 115 110 30 utilized to construct the local strain-local hardness curve

2.5 Level 1
Cylindrical specimens with initial dimensions of h0 =17 Level 2
and d0 =11 mm were compressed to different final heights Level 3
2 Level 4
h, using a lubricated contact between the specimen and the
flat platens. The strain rate of about 0.01 1/s was used

IF (inner)
during the test. 1.5
After compression, samples were cut along the longitudi-
nal direction, and about 30 microhardness tests were 1
performed on the area of the cross-section. The tests were
carried out on an INSTRON machine. To investigate the 0.5
strain field in the radial forging samples, the swaging process
was used. Experimental conditions for swaging were
0
illustrated in Table 2. After swaging, the samples were cut Die angle Back push Front pull Friction Axial feed
along the longitudinal direction and several microhardness force force factor
tests were carried out on the regions. All experiments were a
done in the room temperature. Using these values and the 2.5
strain–hardness reference curve, we constructed and used Level 1
Level 2
the strain field in the radial forging samples for the ver- Level 3
2
ification of the strain field predicted by the finite element Level 4
method.
1.5
IF (outer)

4.2 Determination of relationship between strain

and hardness 1

The hardness test can be used to assess the local properties,

0.5
as well as the determination of the local strain [29]. Amongst
a number of hardness measurement methods, the micro-
hardness method is considered to be the most appropriate. 0
Die angle Back push Front pull Friction Axial feed
Determination of the strain, ɛ, from the hardness measure- force force factor
ments requires the function HV ¼ f ð"Þ, known as the
strain–hardness reference curves. These curves are obtained b
usually by the assessment of compression tests on a given 80000
material under relevant thermo mechanical conditions. Level 1
70000 Level 2
Conventionally, cylindrical specimen is compressed into Level 3
Radial fore (KN)

the specific strain. Then, the specimen is cut along the 60000 Level 4
longitudinal axis of symmetry and a number of hardness 50000
tests are carried out on a mesh of n testing points in the
40000
cross-sectional plane [30]. The mean value of hardness is
computed from all the measured points, and this value is 30000

related to the mean value of strain, ɛm. Therefore, using 20000

different compression samples upset to various final
10000
heights, the strain–hardness curve can be constructed. The
main problem of this procedure is the fact that the local 0
Die angle Back push Front pull Friction Axial feed
values of hardness and strain are transformed into the force force factor
global mean values of hardness and strain prior to the curve
c
construction. So that, the actual distribution of both strain
Fig. 4 The average effect of each factor at each level for three
and hardness may show a great deviation from the mean optimization criteria a I.F.inner b I.F.outer and c RF obtained by the
value. To overcome this problem in this study, direct Taguchi method
 Int J Adv Manuf Technol

Table 6 Parameters values and

performance of the optimum Optimization Optimum condition for factors Optimum Confirming
conditions for three optimiza- criteria performance results
tion criteria were obtained by Die Forw. pull Back Friction Axial (KN) (KN)
Taguchi method angle force push factor feed
(deg.) (KN) force(KN) (mm)

RF 12 25000 25000 0.1 5 6127 6334

I.F.inner 12 0 25000 0.15 5 0.156 0.167
I.F.outer 12 0 25000 0.1 5 0 0

[31]. The reliability of this reference curve can be checked the loads measured for cold swaging of AISI 1015 steel
by predicting the strain in another real specimen. For this rods and used by Lahoti and Altan [6] are utilized to verify
aim, another sample is upset to a different final height. the load predicted by the present analysis. The process
Then it is cut along its cross-section area to measure the parameter for samples, the predicted and experimental loads
local value of its hardness. These hardness values are are shown in Table 3. As it is observed, there is a good
transformed to the local strain value using the strain– agreement between the experimental and predicted results
hardness reference curve. Comparing this value (using so that the predicted results are 5 to 10 percent greater than
strain–hardness reference curve) with the value of strain the experimental data.
obtained from numerical simulation of the compression
test, the reliability of the reference curve can be examined.
In this work about 50 points in different regions of 5 Apply the models and results
compression samples are considered. By this way, individ-
ual points of the strain–hardness reference curve are 5.1 Inhomogeneity factor
obtained and the curve is approximated by a fourth order
polynomial as shown in Fig. 2. The local values of strain on It is necessary to define a criterion as the Inhomogeneity
the radial forged samples are determined using strain– Factor, to investigate the strain distribution in final product
hardness reference curve and the local values of hardness and the effect of various parameters on it. Because of
measured on radial forged samples. For evaluating the axisymmetric deformation in tube an inhomogeneity factor
strains on the entire points of sample, triangle-based cubic (I.F.) is defined as follows:
interpolation with the aid of these determined local values "outer
"min
of strain, are used. Then these values of strains based on the I:F:outer ¼ ð1Þ
"min
experiments are used for validation of FEM simulation.
Figure 3 shows strain contour for computed and measured
"inner
"min
strain obtained from FEM and microhardness test of I:F:inner ¼ ð2Þ
swaging sample, respectively. We found good correlation "min
between these the two set of results. where I.F.inner and I.F.outer are the inhomogeneity factors at
the inner diameter and outer diameter of the tube,
4.3 Confirmation of predicted radial forging load respectively, ɛinner and ɛouter are effective strains at the
surfaces contacting the mandrel and the die, respectively.
In present research for the radial forging of tubes, the ɛmin is the minimum strain in the radial direction of the
experimental results of forging loads are not available. So tube.

Table 7 The analysis of variance

Parameter DOF Sum of square Variance F-ratio Percent influence

16 16 16
RF×10 I.F.inner I.F.outer RF×10 I.F.inner I.F.outer RF×10 I.F.inner I.F.outer RF I.F.inner I.F.outer

A 3 4.16 48.00 27.38 13.87 16.00 11.99 2.44 2.63 4.19 49.51 28.74 30.58
B 3 2.33 33.35 12.59 7.76 11.12 7.05 1.36 1.83 2.47 12.55 14.57 14.06
C 3 2.30 35.12 12.50 7.66 11.71 7.02 1.35 1.93 2.46 11.96 16.27 14.00
D 3 2.37 40.93 21.14 7.90 13.64 9.90 1.39 2.25 3.47 13.39 21.85 23.58
E 3 2.33 37.52 15.93 7.77 12.51 8.17 1.36 2.06 2.86 12.59 18.57 17.78
Tot. 15 13.49 194.92 89.54 44.96 64.98 44.13 7.90 10.70 15.40 100 100 100
Int J Adv Manuf Technol

60
OA column divided by the number of tests (data points) for
Die angle (degree)
Back pull force (KN) that level provide the appropriate averages. The results of
50 Forward push force (KN)
Friction factor
five factors of this study are evaluated for three optimiza-
Axial feed (mm) tion criteria as shown in Fig. 4. In this work a lower
Percent(%)

40
average response is better (LB) for I.F.inner , I.F.outer and
30
radial force (RF), therefore the optimum combination of the
levels of the factors is predicted as exhibited in Table 6.
20 To estimate the performance at the optimum condition
the following terms must be evaluated [30]:
10
i) The average of all trial results
 is called the grand
average of performance T
0
RF IF(inner) IF(outer) ii) The improvement (contribution) for preferred level of
Fig. 5 Percent influence of factors (die angle, back pull force, each factor is calculated by subtracting the  grand
forward push force, friction factor and axial feed) on radial forging, average from the average effect of that level Ap . For
I.F.inner and I.F.outer the state of LB, this can be obtained as follows:
 
5.2 Application of the Taguchi method to radial forging CA ¼
AP
T 
of 6082 aluminum alloy

Generally, the import factors of radial forging include the Then, the optimum performance can be calculated as
die angle and die shape, axial feed and tool velocity, follows:
forward pull and back push forces and friction factor. In this
Yopt ¼ T þ CA þ CB þ ::::
study, die angle, forward pull and back push forces, friction
factor and axial feed as the most influence parameters on The optimum performance for the I.F.inner , I.F.outer and RF
three optimization criteria (RF, I.F.inner and I.F. outer) are as three optimization criteria, are shown in Table 6. Because
considered. the optimum condition is not one of the experiments, other
Table 4 shows four levels of five process factors. These simulations run in the conditions of the three optimum
levels commonly are used in industry and other researches conditions are performed to confirm the predicted results.
[6, 9]. When the levels of the factors are determined the There is a good agreement between the predicted results by
suitable orthogonal array is selected. Orthogonal array of Taguchi method and those achieved by the FEM simulation.
L32 is selected to provide the sufficient training data for
ANN model instead of orthogonal array of L16 [28]. Then, 5.2.2 Analysis of variation (ANOVA)
by using FEM model each simulation are carried out.
Process parameters were used for simulation was illustrated The ANOVA for the different factors, including total
in Table 5. variation, sum of square and percent influence of factors,
is determined (Table 7). The results reveal that die angle
5.2.1 Average effects of factor levels made the major contribution to overall performance for the
three optimization criteria (Fig. 5). And the effect of it for
To evaluate the average effect of each factor at each level IFs criteria is less than RF criteria. The effect of friction
(for example A1 is the average effect of factor A at level 1), especially for IFs criteria is considerable. The effect of axial
the average results for each level must be calculated. To do feed is in the later turn. The back pull and front push force
this, the sum of the data associated with each level in the have the less influence three criteria.

Table 8 Parameters values and performance of the optimum conditions for two optimization criteria were obtained by ANN model

ANN Optimum condition for factors Hidden Optimum Confirming

output layers performance (KN) results (KN)

Die angle Forw. pull force Back push force Friction Axial feed First Second
(deg.) (KN) (KN) factor (mm)

RF 12 14236 16983 0.12 5 5 7 5826 6136

I.F.inner 12 9838 17658 0.13 5 4 6 0.115 0.136
 Int J Adv Manuf Technol

4
x 10 function is used for neurons in the hidden layers, whereas
10
a linear transfer function used for the output layer. Training
of the neural network is done in MATLAB, using
8 TRAINLM function. TRAINLM is a network training
function that updates weights and bias values in a back
Predicted results

propagation algorithm according to Levenberg-Marquardt

6
optimization. Levenberg–Marquardt algorithm is a highly
efficient method for solving non-linear optimization prob-
4 lems. The network is trained until the mean squared error
(MSE) between the data and network output reduced to
10−3. After training, by using 16 testing data provided by
2 FEM, the ability of ANNs for predicted new situations is
examined. The plot of predicted results by ANN versus
FEM results are shown in Fig. 6. Then, genetic algorithm as
0
0 2 4 6 8 10 a global optimization technique is utilized for finding the
4
Experimental results x 10 optimum conditions for two networks (Table 8). In this
a work the “genetic algorithm and direct search” toolbox of
4 MATLAB are utilized and the trained network was used as
a fitness function. Initial population is the range that used in
3.5 the Taguchi method for factors. As can be seen, the results
show a good agreement with the Taguchi method results
3
but as expected the optimum performances show the lower
Predicted results

2.5 response because the ANN model can search between more
points than the Taguchi method.
2

1.5

1 6 Conclusions

0.5 From the results of the finite element analysis and optimi-
0
zation procedure using the Taguchi method and ANN, we
0 0.5 1 1.5 2 2.5 3 3.5 4 make the following conclusions:
Experimental results
b
Fig. 6 Comparison between measured and ANN predicted values for 1) There is a good correlation among the calculated strain
a radial force and b I.F.inner field simulated by the FEM and those achieved by
using microhardness distribution and strain–hardness
5.3 ANN model reference curves. As can be seen the maximum and
minimum value of strain and distribution of strain field
The data provided by OAs are used to train the two ANNs. was similar together.
For both ANNs, inputs comprised die angle; back pull 2) Because of higher friction force at inner diameter of
force, forward push force, friction factor and axial feed tube the inhomogeneity factor at the inner diameter of
while RF or I.F.inner , as the output for each networks tube is higher than that of the outer diameter of tube.
respectively. 3) An ANN model using OA as training data can predict a
Before training, all data are normalized within the better and more accurate optimum performance than
interval [−1, 1] to enable comparison of the relative the Taguchi method.
significance of each input. A neural network with five 4) The optimum conditions for three optimization criteria
neurons in the input layer (each neuron receiving one of the are slightly different; therefore the importance of the
five inputs), a single neuron in the output layer (RF or criteria can help one to select the optimum condition.
I.F.inner), and two hidden layer (number of neurons depicted 5) The most important parameter is the die angle that
in the Table 8) proved to be the most suitable for the made the major contribution to overall performance for
present study. A hyperbolic tangent sigmoid transfer the three optimization criteria.
Int J Adv Manuf Technol

References 16. Thompson EG, Hamzeh O, Jackman LA, Srivatsa SK (1992) A

quasi-steady-state analysis for radial forging. J Mater Process
Techno 34:1–8
1. Altan T, Oh SI, Gegel H (1983) Metal forming fundamentals and
17. Jackman LA, Ramesh MS, Forbes-Jones R (1992) Development of a
applications. American Society for Metals Materials Park 17:
finite element model for radial forging of superalloys. Proc Int Sym
31–35
on the Metallurgy and Applications of superalloys, Pittsburgh:
2. Yano H, Akshi T (1994) Application of analytical simulation to
103–112
forged part design. Toyota Tech Rev 43(2):20–25
18. Domblesky JP, Shivpuri R, Mohamdein N (1994) FEM simulation
3. Jang DY, Liou JH (1998) Study of stress development in axi-
of multiple pass radial forging of pyromet. Proceeding of the Int
symmetric products processed by radial forging using a 3-D non-
Symposium, Pittsburg:26–29
linear finite-element method. J Mater Process Technol 74:74–82
19. Humphreys FJ, Hatherly M (1995) Recrystallization and
4. Ko D-C, Kim D-H, Kim B-M (1999) Application of artificial
related annealing phenomena, vol 11. Elsevier Science, Oxford,
neural network and Taguchi method to preform design in metal
p 15
forming considering workability. Int J Mach Tools Manuf 39:
20. Lindh E, Hutchinson B, Ueyama S (1993) Effect of redundant
771–785
deformation on recrystalization behaviour of copper. Scripta
5. Lahoti GD, Altan T (1976) Analysis of the radial forging process
Metall 29:347–352
for manufacturing of rods and tubes. J Eng Ind 98:265–271
21. Chin RK, Steif PS (1995) A computational study of strain
6. Lahoti GD, Dembowski PV, Altan T (1976) Radial forging of
inhomogeneity in wire drawing. Int J Mach Tools Manuf 35:
tubes and rods with compound-angle dies. Proceeding of
1087–1098
NAMRC-IV Columbus Ohio May 17–19:87–95
22. Kazeminezhad M, Karimi Taheri A (2006) Deformation Inhomoge-
7. Subramanian TL, Venkateshwar R, Lahoti GD, Lee FM (1978 and neity in Flattened Copper Wire. DOI 10.1016/j.matdes.2006.06.009
1979) Experimental and computer modeling of die cavity fill in 23. Chun MS, Biglou J, Lenard JG, Kim JG (1999) Using neural
radial forging of rifling. Proceeding of process modeling sessions, networks to predict parameters in the hot working of aluminum
Process modeling - Fundamentals and applications to metals alloys. J Mater Process Techno 86:245–251
USA:66–72 24. Bahrami A, Mousavi Anijdan SH, Madaah Hosseini HR, Shafyei A,
8. Tszeng TC, Kobayashi S (1986) Determination of residual stresses Narimani R (2005) Effective parameters modeling in compression of
in radial forging, Manufacturing Processes Simulation. PED- an austenitic stainless steel using artificial neural network. Comput
20:31–45 Mater Sci 34:335–341
9. Domblesky JP, Shivpuri R, Painter B (1995) Application of the 25. Kim DJ, Kim BM (2000) Application of neural network and FEM
finite-element method to the radial forging of large diameter tubes. for metal forming Processes. Int J of Mach Tools Manuf 40:911–
J Mater Process Technol 49:57–74 925
10. Domblesky JP, Shivpuri R (1995) Development and validation of 26. Taguchi G (1986) Introduction to quality engineering, designing
a finite-element model for multiple-pass radial forging. J Mater quality into products and processes. Asian Productivity Organization,
Process Technol 55:432–441 Japan
11. Jang DY, Liou JH (1998) Study of stress development in axi- 27. Duan X, Sheppard T (2002) Influence of forming parameters on
symmetric products processed by radial forging using a 3-D non- the final subgrain size during hot rolling of aluminium alloys.
linear finite-element method. J Mater Process Technol 74:74–88 J Mater Process Technol 130–131:245–249
12. Lee KH, Lee SR, Yang DY (2004) Rigid-plastic finite element 28. Roy RK (2001) Design of experiments using the Taguchi
analysis of incremental radial forging process using the automatic approach. John Wiley and Sons, USA
expansion of domain scheme. Eng Computations 21(5):470–487 29. Sadok L, Luksza J, Majta J (1994) Inhomogeneity of mechanical
13. Ghaei A, Movahhedy MR, Karimi Taheri A (2005) Application of properties in stainless steel rods after drawing. J Mater Process
3D FEM to die design in the radial forging process. TICME2005, Technol 44:129–141
Tehran, Iran :112–120 30. Petruska J, Janicek L(2003) On the evaluation of strain inhomo-
14. Rodic T, Owen DRJ (1987) Basic concepts of numerical geneity by hardness measurement of formed products. J Mater
simulation of a radial forging process. Zezb 214:167–174 Process Tech 143–144:300–305
15. Shimizu T, Ohuchi K, Sano T (1990) Reducing of circular bars by 31. Ross PJ (1996) Taguchi techniques for quality engineering, 2nd
a multi ram forging machine. J JSTP 31:392–397 edn. Tata McGraw-Hill, New Delhi

View publication stats