Upload
f-gagliardi
View
213
Download
0
Embed Size (px)
Citation preview
ORIGINAL ARTICLE
Design of an optimized procedure to predict oppositeperformances in porthole die extrusion
G. Ambrogio • F. Gagliardi
Received: 9 September 2011 / Accepted: 16 March 2012 / Published online: 1 April 2012
� Springer-Verlag London Limited 2012
Abstract The main objective of advanced manufacturing
control techniques is to provide efficient and accurate tools
in order to control the set-up of machines and manufac-
turing systems. Recent developments and implementations
of expert systems and neural networks support this aim.
This research explores the combined use of neural net-
works and Taguchi’s method to enhance the performance
of porthole die extrusion process; the energy saving and the
quality of the welding line are two conflicting objectives of
the process taken into account. The complexity of the
analysis, due to the number of the involved variables, does
not allow the representation of the specified outputs by
means of a simple analytical approach. The implementa-
tion of a more accurate and sophisticated tool, such as the
neural network, results more efficient and easier to be
integrated into a simple ‘‘ready to use’’ procedure for
predicting the investigated outputs. The main limit to wider
implementation of neural networks is the huge computation
resources (times and capacities) required to build the data
set; a finite element approach was adopted to overcome the
time and money wasting typical of experimental investi-
gations. Satisfactory results in terms of prediction capa-
bility of the highlighted outputs were found. Finally, a
simple and integrated interface was designed to make
easier the application of the proposed procedure and to
allow the generalization to other manufacturing processes.
Keywords Extrusion � Porthole � ANN � DoE � Taguchi
1 Introduction
In a competitive marketplace, good quality and low cost
manufacturing products are, of course, one of the most
sought-after goals. Firstly, the customer agrees to pay if he
recognizes higher characteristics in the product, such as
better mechanical properties or increasing performance; at
the same time, the producers need to find better operating
conditions, which ensure reduction of industrial costs [1].
These aims can rarely be found in a single solution; more
often the production choices represent a compromise
between the customer requests and lower industrial costs.
Due to this compromise, manufacturing processes need to
be closely monitored and designed to ensure the quality
standards and to choose the process conditions, which
ensure the possible lower costs. Beyond the particular
manufacturing process taken into account, there is a great
need for an automatic and effective methodology, which
can indicate its making state. Artificial neural networks
(ANNs) can be used for this purpose due to their ability to
imitate the human brain [2]. This technique is suited for
problems that involve the manipulation of multiple
parameters and nonlinear interpolation [3]. Starting by the
same base of knowledge, the ANNs allow to simulta-
neously evaluate more than a single output of the process.
These tools find applicability in several fields of science
including engineering, medicine, agriculture. The auto-
motive factories can be, for example, used like a case of
study; these factories belong to the manufacturing sector,
which typically presents several complexities to solve.
From this point of view, Lolas and Olatumbosun [4]
demonstrated how a neural network architecture could
G. Ambrogio (&) � F. Gagliardi
Department of Mechanical Engineering,
University of Calabria, 87036 Rende, CS, Italy
e-mail: [email protected]
URL: http://tsl.unical.it
F. Gagliardi
e-mail: [email protected]
123
Neural Comput & Applic (2013) 23:195–206
DOI 10.1007/s00521-012-0916-3
predict the reliability performance of a vehicle at later
stages of its life; same assumptions can be extended to each
item used to realize the car. For example, the polymer
composites are often utilized in the automotive industry,
and the neural network approach can be used to predict the
mechanical properties of these materials [5].
The neural network (NN) approach is obviously appli-
cable to the manufacturing processes too [6]. The manu-
facturing scenario and the material design represent some
of the most promising and natural application fields for this
kind of methodology; due to the intrinsic complexity that
characterizes these contests, the neural networks result
widely utilized in the last years as predictive tool [7–9].
Moreover, Dae-Cheol et al. [10] suggested a method of
preform design in a multistage metal forming processes
based on ANN; they also proposed the workability forecast
limited by ductile fracture.
Lucignano et al. [11] implemented two NNs to optimize
the aluminium extrusion process determining the temper-
ature profile of an AA6060 alloy in two different equip-
ment zones. A conclusion of this study was that the
temperature profiles, predicted by the neural network,
closely agree with experimental values. Previously, Suk-
thomya and Tannock [12] compared methods for training
NNs to analyse complex manufacturing processes; they
provided some interesting guidelines and highlighted the
real implementation in industry by using a case study.
In this paper, the porthole die extrusion was considered
and the prediction of the process performances at varying
of all the geometrical and technological variables was
taken into account by means of an ANN. The extrusion by
porthole die is one of the most used manufacturing pro-
cesses to produce hollow components; its feasibility is
directly dependent by the availability of a complex die
structure, which allows to obtain the desired shape. A lot of
factors involved in the process with effects that conflict on
the measured outputs. Therefore, the process design is
quite difficult due to the geometrical parameters that need
to be set-up, such us the bridge shape, the width of the
bridge and the height of the welding chamber. Actually, the
problem is that these parameters are derived only after few
trial and error experiments; vice versa, the availability of a
tool that helps the designer during the preliminary set-up
definition could make easier the process applicability and
reduce the start-up time. A pure analytical approach results
unsuitable due to the large number of variables to take into
account and the nonlinearity of the phenomenon to analyse;
none of the finite element analysis (FEA) is competitive
due to some drawbacks that penalize its use, such as the
high computational time and the remeshing problem [13].
The approach proposed in this paper is based on the use
of an optimized artificial neural network that, starting from
a given process condition, supplies an effective and fast
prediction of two output variables: (1) the maximum load
required to carry out the process and (2) the quality of the
welding line. In a reliable range, by an inverse approach,
the process designer can, easily, find the better configura-
tion to execute the process reducing the load and increasing
the quality of the extruded part. The described procedure
has given satisfactory results for a set of investigated cases,
as accurately discussed in the following sections.
2 Optimized procedure for neural network design
From an industrial point of view and for any type of pro-
cess, the prediction capability of suitable output perfor-
mance is considered strategic for everyone [14]. The
proposed methodology tries to satisfy this need in a more
efficient and effective way reducing time and resources.
The approach is based on the combined use of the NN, as
highly performance predictive tool [15, 16], and the Design
of Experiments (DoE), to reduce the test number and to
ensure the robustness of the analysis [17].
As it is well known, multilayer feed-forward networks
(commonly referred as multilayer perceptrons) are an
important class of NNs. Multilayer perceptrons networks
suggested by Rumelhart et al. [18] have been successfully
applied to solve some difficult and diverse problems in
many disciplines of science and technology, when non-
linear solution is required. The procedure normally
involves training of the network in a supervised or unsu-
pervised manner. The network typically consists of a set of
sensory units (source nodes) that constitute the input layer,
one or more hidden layer of computation nodes and an
output layer of computation nodes. The input signal
propagates through the networks in a forward direction in a
layer-by-layer basis. The data for teaching the artificial
neural network need to be carefully selected, and a wide
range of data, more than those to be found, should be
trained. As a consequence, there is the necessity to have an
adequate quantity of experimental data available since the
NN predictive capability is related to the data set dimen-
sion: higher is the number of the training data set, higher is
the NN predictive capability to generalize. In the manu-
facturing field, the number of experimental data for the NN
design cannot be easily increased being largely time- and
resource consuming; at the same time, a careful NN design
results time-consuming due to the number of NN mor-
phologies to test. Due to these aspects, a simple search for
the optimal NN does not supply additional information in
matter of factor significance, nor it guarantees the result
quality. To overcome these drawbacks, an optimized pro-
cedure, based on the hybrid approach of both finite element
method (FEM) and DoE, was proposed in the following
study. First of all, the FEM allowed to improve the
196 Neural Comput & Applic (2013) 23:195–206
123
knowledge of the process itself and to acquire more
information on the process mechanics. In general, to
decrease development times, it is necessary to apply finite
element analysis (FEA) even in the very first step of a
design process, since it serves the opportunity for a sig-
nificant speed-up of the design process. However, when the
simulation process results to be too much complex, model
simplification has to be taken into account, like in this
study.
Contemporarily, the statistical DoE is used to get
information such as main effects and their interaction,
starting from the minimum number of experiments or, in
this case, the minimum number of NN run. The design
objective in the statistical DoE is to find a good combi-
nation of controllable factors. The Taguchi’s method,
which is one of the fractional designs, has a good repre-
sentation of experiment concerning only the main effects of
the design parameters including noise factors [19]. The
Taguchi’s DoE allows to highlight the main factors that
have to be considered into the investigation, and to verify
the quality of the analysed networks starting from the
evaluation of some performance indexes. Contemporarily,
it drives the designer in the selection of the NN parameter
set-up, which minimizes the network sensitivity from an
external noise [20].
The proposed procedure tries to address this issue fol-
lowing the below-discussed steps (Fig. 1):
1. Key process variables and constrain identification:
Each manufacturing process is characterized by a set
of variables, both numerical and categorical. To build
a predictive tool, it is necessary to well understand
which possible input variables really affect the desired
outputs; a preliminary step is focalized on a better
definition of the problem, involved factors and possible
bonds.
2. Data collection: In general, both the experimental
investigation in the laboratory and the numerical
model can be applied to build the data set.
3. Efficient search for the optimal NN configuration by
means of the statistical analysis: The DoE was used to
determine the NN, which ensures better prediction of.
4. Toolbox design: Implementation of a simple and
‘‘ready to use’’ procedure in order to predict the
process output.
While the first two steps are strongly related to the
considered problem, the optimal NN identification can be
formalized into a more general way. As concern the third
phase, a careful NN design requires first of all a pre-
liminary phase that establishes the analysis perimeter; this
is necessary to determine all the possible network config-
urations and the algorithms that have to be considered and
tested. At the same time, the choice and the number of
parameters that have to be taken into account affect the
design performance. The methodology proposed is based
on a more formal approach derived by the Taguchi’s DoE
[19–21]. The Taguchi’s method drives the designer in the
selection of an optimal network multilayer perceptron
(MLP), which well describes the connection between input
and output factors and minimizes the network sensitivity to
an external noise. This is possible through the analysis of
variance (ANOVA) and the analysis of means (ANOM);
the ANOVA establishes the relative significance of factors
Fig. 1 Implemented procedure
for a robust process control
Neural Comput & Applic (2013) 23:195–206 197
123
in terms of their percentage contribution to the response,
while the ANOM estimates the main effects of each factor.
The ANOVA is also necessary for estimating the error
variance for the effects and variance of the prediction error
[22]. The analysis is performed on signal-to-noise ratio
(S/N ratio) to obtain the contribution of each factor.
A break point into the analysis is the choice of the
learning algorithm; several studies identify in the error
back propagation (EBP) a good solution [23–25]. The main
idea of the EBP is that the errors propagate backwards from
the output nodes to the inner nodes. During each training
iteration, the gradient of the error of the network is mea-
sured and used in a simple stochastic gradient descent
algorithm to find weights that minimize the error. In this
paper, the Levemberg-Marquardt (LM) algorithm was
applied because it gives better results, especially during the
training phase [26]. Specifically, the LM is also based on
the back-propagation approach, but it allows to overcome
the loss of convergence, which typically affects the EBP
[11, 27]. Actually, the LM algorithm is a blend of the
gradient descendent and Gaussian-Newton iteration, since
these methods are complementary in the advantages they
provide; the parameter updating in the gradient descendent
approach is performed by adding the negative of the scaled
gradient at each step, that is,
xiþ1 ¼ xi � krf ð1Þ
where xi is the parameter value at the actual step, rf is the
gradient of a nonlinear least square function f(x) and k is a
factor used to influence the gradient descent.
Levenberg and Marquardt instead proposed an algo-
rithm whose update rule is given as
xiþ1 ¼ xi � ðH þ kdiag½H�Þ�1rf ðxiÞ ð2Þ
where H and rf are the Hessian matrix and the gradient
evaluated at xi, respectively, and k is a factor that increases
or decreases to influence the descent. If the error goes down
following an update, it implies that quadratic assumption
on f(x) is working, and in this way, the factor k can be
reduced (usually by a factor of 10) to decrease the influence
of gradient descent. If the error goes up, k is increased by
the same factor, to follow the gradient more. Since the
Hessian is proportional to the curvature of f (x), the Eq. (2)
implies a large step in the direction with low curvature (i.e.
an almost flat terrain) and a small step in the direction with
high curvature (i.e. a steep incline).
According to this theoretical aspect, the LM algorithm is
in no way optimal but is just a heuristic and it works
extremely well in practice [11]. The only flaw is its need
for matrix inversion as part of the update. Even though the
inverse is usually implemented by using clever pseudo-
inverse methods such as singular value decomposition, the
cost of the update becomes prohibitive after the model size
increases to a few thousand parameters. For models of a
few hundred parameters, this method is much faster than
gradient descent; as a consequence, it was introduced to
solve problem in manufacturing scenario where the avail-
able data are generally poor [28].
Finally, considering the LM as a learning algorithm for
the NN training, the k factor is expressed by two parame-
ters: (1) the damping factor (Mu) and (2) the boost/drop
factor (Bdf); they influence the convergence velocity as
explained. While the learning algorithm was univocally
determined, the experimentation for the optimal NN search
was based on the variation of its intrinsic parameters.
Furthermore, according to the state of the art, other two
factors were utilized to generate the investigation plane for
the network design: the number of the hidden layer (NHL)
and the number of the neurons in the hidden layer (NNHL).
Different studies also highlighted the significance of this
factors and the necessity to include them into the analysis
[29, 30]. The NNHL increases the amount of connections
and weights to be fitted. This number cannot be increased
without limit because it is possible to reach a situation
where the number of connections to be fitted is larger than
the number of data pairs available for training [3]. Some
heuristics allow to find a lower and an upper bound for the
NNHLs [23]. These approaches prescribe the better way to
fix the number of neurons into the next hidden layers
dependent on the number of the input (NIN) and the output
(NON) neurons, properly decreased or increased according
to a constant. Three heuristics reported in Table 1 have
been considered in the analysis.
Finally, the range of the utilized experimental plane is
shown in Table 2, which determines 27 network configu-
rations that have to be tested, according to the Taguchi’s
theory of the orthogonal array [19]. In fact, let k = 4 be the
number of considered parameters, a not-optimized inves-
tigation leads to analyse a fully factorial plane (conven-
tional trials and errors approach), which corresponds to 3k
NN configurations at least. Instead, a fractional factorial
plane, corresponding to only 27 network configurations,
was enough according to the Taguchi’s method to reduce
the computational time for training and testing. According
to that, the time required for the optimal NN search is
reduced up to the 70 %.
Moreover, the conventional trials and error approach do
not supply additional information to the designer in matter
of factors significance, nor it guarantees a good quality of
Table 1 Used heuristics to determine the NNHLs
Lippmann (LIP) NON 9 (NIN ? 1)
Kolmogorov (KOL) 2NIN ? 1
Her Majesty’s Department of Trade and
Industry (MTI)
(NIN ? NON)/2
198 Neural Comput & Applic (2013) 23:195–206
123
the results determining very high performance cost; as a
consequence, the analysis could result incomplete. The
Taguchi’s Design of Experiments (TDoE) allows both to
highlight the main factors to be considered into the
investigation phase and to verify the quality of the inves-
tigated NNs starting from the evaluation of some perfor-
mance indexes (i.e. ANOVA, Normal Probability Plot).
The solution results more robust, and the performance costs
low [19].
3 Porthole die extrusion
The extrusion by porthole die is a process that is always
more used in order to produce hollow components; parts
with different sections can be obtained.
The process takes place through the complex die
structure composed of container, porthole, mandrel, weld-
ing chamber and bearing part (Fig. 2); the working
sequence consists of three stages that can be identified like
dividing, welding and forming stage.
The mandrel legs divide the material before that it flows
inside the welding chamber, where due to simultaneous
action of temperature and pressure, it joins again forming
welding lines [31]. Solid joining mechanics occur during
their formation similar, for instance, to the friction stir
welding [32]; two typologies of welding lines, especially,
can be found on the extruded parts, namely longitudinal
and transverse lines [33]. In order to manufacture hollow
components with better quality, understanding of mandrel
deflection and metal flow through porthole, welding
chamber and die bearing is of great importance [34]. Dif-
ferent researches have been carried out to highlight the
influences that process parameters have on the metal flow
inside the die, extrusion load and so on [35, 36].
3.1 Numerical model
The complexity of the equipment construction generates
great difficulties to carry out pure experimental investiga-
tions; the developed studies were founded, even, on FEM
analyses aimed to highlight strain, strain rate and stress
distributions in the whole porthole die. Its geometry is just
one of the variables that strongly affects the effectiveness
of the extruded part; the extrusion ratio, the porthole
number, the extrusion speed, the billet and die temperature
are further variables to be considered for the improvement
of the worked part quality [31]. However, numerical sim-
ulation presents some drawbacks, too; the analysis requires
a 3D simulation, and due to the process complexity, the
necessary computational time is high. The metal flow
inside the welding chamber is characterized by large sur-
face generation with very heavy numerical simulations.
Moreover, the strong reduction in area deeply impacts on
the remeshing–rezoning code capability, and this becomes
a possible reason of result inaccuracy. Significant mesh
distortion occurs inside the porthole die; consequently, the
simulation convergence can be reached, step by step,
through a tidier mesh obtained by numerical interpolation
from the old to the new discretization. Simplified experi-
ments and numerical model were proposed [13, 37], which
represent a good compromise between the robustness of the
analysis and the lower simulation time. They have to be
designed in order to physically reproduce the actual pro-
cess mechanics allowing, in the meantime, simpler and
quicker numerical simulations. The 2D simulations of
Table 2 Investigated factors and levels
Factors Levels
Number of hidden layers (NHL) 1
2
3
Number of hidden neurons (NHN) LIP
KOL
MTI
Damping factor (Mu) 0.001
0.05
0.25
Boost/drop factor 10
20
40
Bearing
Chamber
Mandrel
Welding
Porthole
Container
Fig. 2 Sketch of a traditional equipment for porthole die extrusion
Neural Comput & Applic (2013) 23:195–206 199
123
porthole die extrusion were especially set considering the
section reported in Fig. 3. A section plane is utilized: it cuts
the bridge of the mandrel along its axis and the welding
chamber in the zone where the two material flows weld.
The effectiveness of the welding algorithm can be tested
through 2D numerical analyses [31] with plane strain
geometry.
Despite this simplification, the porthole die extrusion
keeps high complexity. Some process and geometric vari-
ables have to be taken into account during this study: the
punch velocity (V), the die width (B), the bridge width
(BW), the bearing length (BL) and the height of the
welding chamber (HWC) were investigated. Moreover, the
bridge shape was changed too; its upper (U) and lower
(L) part were considered with both right (90�) and sharp
(45�) angles. All the investigated ranges are reported in
Table 3.
The strength of welding lines is a fundamental variable
that has to be taken into account for judging the process
goodness [32]; in literature, several criteria were intro-
duced to quantitatively evaluate this characteristic [13, 37,
38]. In this paper, the welding criterion (Q) proposed by
Plata and Piwnik [39] was taken into account; the criterion
formulation is based on the integral on time of contact
pressure (P), normalized on the actual effective stress (r),
along a generic path for a welding element according to:
Q ¼Z
t
P
rdt�C; ð3Þ
where C is the critical limit for the material welding,
function of material and process conditions. However,
higher is the value that it is possible to reach during the
extrusion, in the welding chamber, and better should be the
mechanical characteristics of the joint. This criterion was
implemented in the commercial finite element code,
DEFORM 2DTM [40] by using a customized user-
subroutine.
A second output was monitored during the numerical
campaign; this is the punch load whose value goodness
changes in antithesis with the Q value. Higher Q is usually
obtained in conjunction with higher punch load, which
instead has to be lower possible in order to reduce the die
wear, the required capacity of the press and, as conse-
quence, the energy.
The simulations were carried out modelling, as billet
material, an AA1050 whose plastic behaviour was derived
from the FE code library; the die and the punch were
modelled as rigid bodies in order to reduce the simulation
time. Mechanical analyses were executed, and the material
temperature was set equal to 500 �C. The simulations were
2D SectionSection Plane
Fig. 3 2D section extracted by a real porthole die equipment
Table 3 The process and geometric variables taken into account for
the analyses
V (mm/s) B (mm) U–L (�) BW (mm) LB (mm) HWC (mm)
1 100 45–45 12 0.5 20
90–45 3
5 130 90–90 18 6 40
200 Neural Comput & Applic (2013) 23:195–206
123
set by using an irregular initial mesh with 4500 Deform 2D
standard elements; several mesh boxes were introduced in
order to thicken the billet discretization close to the die
fillets and, in the zone of the welding chamber, where the
flows of the two material join (Fig. 4); the die and the
punch, instead, were considered rigid bodies in order to
reduce the simulation time. Adhesion between die and
deforming material was assumed; in detail, the constant
shear model was used with the frictional stress defined as:
s = mk, where s is the frictional stress, k is the shear yield
stress and m is the friction factor [40]. When there is
adhesion between material and die, the factor m is set equal
to 1. The shear stresses are probably overestimated at low
values of contact pressure; however, this condition was
carefully tested and validated [41, 42].
Naturally, the goodness of the utilized model was
already highlighted in previous research carried out by the
authors [13]: in fact, it was shown that despite the intro-
duced simplification, the 2D model allows to qualitatively
analyse the investigated variables (load and quality of the
welding line), with satisfactory results. In particular, the
comparison between experimental and numerical loads
highlighted similar trends, from a qualitative point of view
[13].
According to that, 144 numerical simulations by using
two-dimensional analyses were executed according to a
full orthogonal plane. The computational time can be found
in few hours for each simulation using a PC-Dual- Xeon
2,8 GHz with 4-GB RAM. This information has to be
highlighted considering that 3D investigations need days to
complete the numerical calculations.
Fig. 4 A numerical simulation
a at its beginning and b when
the material totally fill the
welding chamber
Table 4 The complete experiment settings
NHL NHN Mu Bdf
1 1 KOL 0.001 10
2 1 LIP 0.050 20
3 1 MTI 0.250 40
4 2 KOL 0.050 20
5 2 LIP 0.250 40
6 2 MTI 0.001 10
7 3 KOL 0.250 40
8 3 LIP 0.001 10
9 3 MTI 0.050 20
10 1 KOL 0.050 40
11 1 LIP 0.250 10
12 1 MTI 0.001 20
13 2 KOL 0.250 10
14 2 LIP 0.001 20
15 3 MTI 0.050 40
16 3 KOL 0.001 20
17 3 LIP 0.050 40
18 3 MTI 0.250 10
19 1 KOL 0.250 20
20 1 LIP 0.001 40
21 1 MTI 0.050 10
22 2 KOL 0.001 40
23 2 LIP 0.050 10
24 2 MTI 0.250 20
25 3 KOL 0.050 10
26 3 LIP 0.250 20
27 3 MTI 0.001 40
Neural Comput & Applic (2013) 23:195–206 201
123
4 Discussion of the results
The NN-Tool provided by Matlab [43] was properly
implemented to train and test the NN; the DoE analysis was
executed by using Minitab. It is well known that during a
training phase, to avoid an overfitting, it is recommended to
use a validation set for measuring the value of the error
function during the learning. The NN is then fixed when the
error value is minimum (early stopping method). This
approach requires a large data set [44]; however, in such
manufacturing process, like the one here investigated, it is
not easy to build a large data set due to the time and
resources required to make the experiments or to run a
numerical simulation. Specifically, taking into account the
porthole die extrusion, the experimental investigation
requires further energy consumption and material waste
[45], while the numerical simulation needs several hours to
converge [13]. To solve this problem, without sacrificing
the robustness of the analysis and the NN generalization
capability, Maren et al. [23] proposed to split the original
data set in training and test set.
The min–max normalization was used to transform the
original real data set. The hidden layer neurons have sig-
moidal transfer functions, and the output neuron uses a
linear activation function. Two stopping criteria were
adopted, that is sufficient accuracy and the maximum
number of iterations (the first activated). For checking the
signal of noise, a threefold cross-validation was used: the
data set was randomly split into different training set and
test set for three times. More specifically, 2/3 of the whole
data set was used as training set and the remaining 1/3 as
test set [23]. All the NNs were trained by using the con-
sidered learning algorithms and tested on each set. In this
way, starting from 27 network configurations, the whole
experimentation required the execution of 162 tests. The
complete experimental plane was synthesized in Table 4.
Finally, according to Taguchi’s method, the ANOVA
was executed to identify which factors strongly influence
the output and, as a consequence, which factors maxi-
mize the S/N ratio (i.e. minimizing the prediction error
network).
From the results of ANOVA, it was possible to identify
two control factors: the number of hidden layer (NHL) and
the number of neurons in the hidden layer (NNHL) are
highly significant for the NN performance. The factors of
the Levemberg-Marquardt’s algorithm are less significant
for the NN foresight capability. Similarly, the interaction
between NHL and NNHL is also significant; this result can
be justified taking into account the specific heuristics
implemented in the analysis to determine the NNHL [46].
Table 5 presents the results of ANOVA.
The ANOM outcomes for the investigated case can be
derived by the response diagrams as shown in Fig. 5.
Table 5 Analysis of variance for SN ratios
Source DF F P-value
NHL 2 68.72 \0.001
NHN 2 63.51 \0.001
Mu 2 2.51 0.161
BdF 2 1.30 0.340
NHL 9 NHN 4 47.44 \0.001
NHL 9 MU 4 2.76 0.128
NHL 9 BdF 4 0.49 0.747
Fig. 5 Response diagram of
S/N ratio
202 Neural Comput & Applic (2013) 23:195–206
123
The level of a factor with the highest S/N ratio is the
optimal one. According to that, the better configuration was
detected for the pattern {LIP – 1 – 0.05 – 40}, which
corresponds to the NN morphology 6-14-2 trained by set-
ting Mu = 0.05 and BdF = 40. Starting by this result, the
NN capability for predicting new data was tested and a
prediction error Y = 4.64 % was found.
Aimed at testing the generalization capability of the
network, it was decided to generate a new set of examples;
this set does not belong to any combination of input pro-
vided to the network during training and validation phase.
It was possible to test the network performance with
respect to unknown data and, at the same time, to compare
the output from the proposed network with the real one
associated with the combination of the provided input. Ten
new examples, numerically derived and characterized by
new combinations of geometric and process parameters,
were proposed to the network (see Table 6); these patterns
were not used in previous steps allowing to understand the
optimal NN generalization capability.
The NN prediction capability is graphically highlighted
in Fig. 6 with respect to the actual process load, Fmax (a),
and the welding line quality, Q (b).
According to these outcomes, a mean error prediction
equal to Y = 7.16 % was measured confirming again both
the suitability of the analysis and the possibility to apply
the NN for controlling the main process outputs.
5 Interface design between neural network
and manufacturing process
The lack of knowledge on artificial intelligent tools usually
represents a significant limitation to their use. A ‘‘user
friendly’’ toolbox (named Advanced Manufacturing Con-
trol Toolbox, AMCT) was introduced to overcome this
problem and also to complete the general procedure dis-
played in Fig. 1.
This automatic toolbox was designed by implementing a
Java code, compiled in Eclipse, and by integrating it to the
standard NN toolbox in Matlab, by means of proper user-
subroutine. The Taguchi method, instead, is not integrated
in the toolbox and needs to be managed by Minitab.
Some customized windows of the designed tool are
displayed in Fig. 7.
For a general study, the proposed procedure can be
implemented by using the same toolbox sequentially, as
shown in Fig. 8. Firstly, a problem needs to be defined, in
terms of number of input, output and data by choosing the
Generate option in the main menu. The optimal NN search
can be carried out directly through this simple interface,
avoiding the knowledge of the Matlab toolbox and by using
the Simulation option (Step 1). The results of the optimal
NN search are stored in an Excel file, which can be used in
the statistical step (Step 2). After that, the optimal NN
Table 6 Ten novel cases
Test # 1 2 3 4 5 6 7 8 9 10
V (mm/s) 5 1 1 3 5 2 2 1 5 1
B (mm) 100 130 100 90 100 100 130 90 90 100
U–L 45�–45� 90�–90� 45�–45� 45�–90� 45�–45� 90�–90� 45�–90� 45�–45� 90�–90� 45�–90�BW (mm) 18 18 18 15 15 15 12 12 12 15
LB (mm) 6 3 3 0.5 0.5 6 3 2 6 2
HWC (mm) 40 20 20 30 40 10 30 30 20 30
(a)
(b)
10000
15000
20000
25000
1 2 3 4 5 6 7 8 9 10
Loa
d [N
]
Test Number
Fmax
Fmax_pred
0.00
5.00
10.00
15.00
20.00
25.00
30.00
35.00
1 2 3 4 5 6 7 8 9 10
Q [
-]
Test Number
Q
Q_pred
Fig. 6 Prediction capability of the optimal NN with respect to new
data
Neural Comput & Applic (2013) 23:195–206 203
123
needs to be trained, through the Optimised NN Design
option, and properly saved (Step 3). At that point, the
optimal NN can be used automatically as a predictive tool
for the problem being investigated (Step 4).
The Advanced Manufacturing Control Toolbox is easy
to use, even if the user is not aware of all the artificial
intelligence techniques and particularly the neural network
tools.
Fig. 7 Toolbox interfaces for data prediction
Fig. 8 Advanced manufacturing control toolbox steps
204 Neural Comput & Applic (2013) 23:195–206
123
6 Conclusion
Nowadays, it is a common point of view that the avail-
ability of such predictive tools, easy to be implemented and
used, is a point of strength for a widespread application of
new and more complex production alternative. When the
trial and error investigation results particularly time-con-
suming, the coupled use of FEM analysis and optimization
method can help to improve the process knowledge and/or
to perform the design phase. In this study, the specified
techniques were invoked, and a simple and more general
procedure was firstly developed and then applied to the
porthole die extrusion. More deeply, a decision maker can
predict the main process performances of the considered
process by using the designed tool; a very fast response can
be obtained avoiding expensive trial and error experimen-
tation. Moreover, to simplify the application of the pro-
posed procedure and especially the optimal NN search, a
‘‘user friendly’’ technique was developed and tested for the
investigated problem. Finally, the toolbox can be easily
used to measure the performances of manufacturing
processes.
References
1. Ulrich KT, Eppinger SD (2008) Product design and development,
4th edn. Mc Graw Hill, NY, USA
2. Wang H, Chen P (2011) Intelligent diagnosis method for rolling
element bearing faults using possibility theory and neural net-
work. Comput Ind Eng 60(4):1457–1471
3. Sha W, Edwards KL (2007) The use of artificial neural networks
in materials science based research. Mater Design 28:1747–1752
4. Lolas S, Olatunbosun OA (2008) Prediction of vehicle reliability
performance using artificial neural networks. Expert Syst Appl
34(4):2360–2369
5. Zhang Z, Friedrich K (2003) Artificial neural networks applied to
polymer composites: a review. Compos Sci Tech 63(14):2029–
2044
6. Peng Y, Liu H, Du R (2008) A neural network-based shape
control system for cold rolling operations. J Mater Proc Technol
202:54–60
7. Saberi S, Yusuff RM (2010) Neural network application in pre-
dicting advanced manufacturing technology implementation
performance. Neural Comput Appl. doi:10.1007/s00521-010-
0507-0
8. Serhat Y, Armagan AA, Erol F (2011) Surface roughness pre-
diction in machining of cast polyamide using neural network.
Neural Comput Appl. doi:10.1007/s00521-011-0557-y
9. Venkatesan D, Kannan K, Saravanan R (2009) A genetic algo-
rithm-based artificial neural network model for the optimization
of machining processes. Neural Comput Applic 18(2):135–140
10. Dae-Cheol K, Dong-Hwan K, Byung-Min K (1999) Application
of artificial neural network and Taguchi method to preform
design in metal forming considering workability. Int J Mach
Tools Manuf 39:771–785
11. Lucignano C, Montanari R, Tagliaferri V, Ucciardello N (2010)
Artificial neural networks to optimize the extrusion of an alu-
minium alloy. J Intel Manuf 21:569–574
12. Sukthomya W, Tannock J (2005) The training of neural networks
to model manufacturing processes. J Intel Manuf 16:39–51
13. Ceretti E, Fratini L, Gagliardi F, Giardini C (2009) A new
approach to study material bonding in extrusion porthole dies.
CIRP Ann Manuf Technol 58:259–262
14. Kalpakjian S, Schmid SR (2003) Manufacturing processes for
engineering materials, 4th edn. Addison-Wesley, Menlo Park, CA
15. Di Lorenzo R, Ingarao G, Micari F (2006) On the use of artificial
intelligence tools for fracture forecast in cold forming operations.
J Mater Proc Technol 177:315–318
16. Khotanzad A, Chung C (1998) Application of multi-layer per-
ceptron neural networks to vision problems. Neural Comput Appl
7(3):249–259
17. Bahloul R, Mkaddem A, Dal Santo P, Potiron A (2007) Sheet
metal bending optimisation using response surface method,
numerical simulation and design of experiments. Int J Mech Sci
48(6):997–1003
18. Rumelhart DE, Hinton GE, Williams RJ (1986) Learning internal
representations by error propagation, 2nd edn. Parallel Distrib-
uted Processing MIT Press, Cambridge, MA
19. Khaw JFC, Lim BS, Lim LE (1995) Optimal design of neural
networks using the Taguchi methods. Neurocomputing 7(3):
225–245
20. Ambrogio G, Filice L, Guerriero F, Guido R, Umbrello D (2011)
Prediction of incremental sheet forming process performance by
using a neural network approach. Int J Adv Manuf Technol. doi:
10.1007/s00170-010-3011-x
21. Sukthomya W, Tannock J (2005) The optimization of neural
network parameters using Taguchi’s design of experiments
approach: an application in manufacturing process modeling.
Neural Comput Appl 14:337–344
22. Phadke MS (1989) Quality engineering using Robust Design.
Prentice Hall, New Jersey
23. Maren AJ, Jones D, Franklin F (1990) Configuring and opti-
mizing the back-propagation network, 2nd edn. Handbook of
neural computing applications. Academic Press, Santiago
24. Udo GJ (1992) Neural networks applications in manufacturing
processes. Comput Ind Eng 23(1–4):97–100
25. Hou TH, Su CH, Chang HZ (2008) An integrated multi-objective
immune algorithm for optimizing the wire bonding process of
integrated circuits. J Intel Manuf 19:361–374
26. Marquardt D (1963) An algorithm for least-squares estimation of
nonlinear parameters. J Soc Ind Appl Math 11(2):431–441
27. Ding L, Matthews J (2009) A contemporary study into the
application of neural network techniques employed to automate
CAD/CAM integration for die manufacture. Comput Ind Eng
57:1457–1471
28. Costa MA, Braga A, De Menezes BR (2007) Improving gener-
alization of MLPs with sliding mode control and Levenberg-
Marquardt algorithm. Neurocomputing 70:1342–1347
29. Mutasem KSA, Khairuddin BO, Shahrul AN (2009) Back prop-
agation algorithm: the best algorithm among the multi-layer
perceptron algorithm. Int J Comput Sci Netw Secur 9(4):378–383
30. Zhenyu J, Lada G, Zhong Z, Klaus F, Alois KS (2008) Neural
network based prediction on mechanical and wear properties of
short fibers reinforced polyamide composites. Mater Design
29(3):628–637
31. Jo HH, Jeong CS, Lee SK, Kim BM (2003) Determination of
welding pressure in the non-steady-state porthole die extrusion of
improved Al7003 hollow section tubes. J Mater Proc Technol
139:428–433
32. Valberg H (2002) Extrusion welding in aluminium extrusion. Int
J Mater Product Tech 17:497–556
33. Valberg H, Loeken T, Hval M, Nyhus B, Thaulow C (1995) The
extrusion of hollow profiles with a gas pocket behind the bridge.
Int J Mater Product Technol 10(3–6):222–267
Neural Comput & Applic (2013) 23:195–206 205
123
34. Lee JM, Kim BM, Kang CG (2005) Effects of chamber shapes of
porthole die on elastic deformation and extrusion process in
condenser tube extrusion. Mater Design 26:327–336
35. Jo HH, Lee SK, Jung CS, Kim BM (2006) A non-steady state FE
analysis of Al tubes hot extrusion by a porthole die. J Mater Proc
Technol 173:223–231
36. Kim YT, Ikeda K, Murakami T (2002) Metal flow in porthole die
extrusion of aluminium. J Mater Proc Technol 121:107–115
37. Bariani PF, Bruschi S, Ghiotti A (2006) Physical simulation of
longitudinal welding in Porthole-Die extrusion. CIRP Ann Manuf
Technol 55(1):287–290
38. Donati L, Tomesani L, Minak G (2007) Characterization of seam
weld quality in AA6082 extruded profiles. J Mater Proc Technol
191(1–3):127–131
39. Plata M, Piwnik J (2000) Theoretical and experimental analysis
of seam weld formation in hot extrusion of aluminum alloys. In:
Proceedings of 7th international aluminum extrusion technology,
Chicago USA, pp 205–211
40. Deform TM (2010) User’s manual (Version 10.1). Columbus, OH
41. Chanda T, Zhou J, Duszczyk J (2000) 3D FEM simulation of
thermal and mechanical events occurring during extrusion trough
a channel-shaped die. In: Proceedings of 7th international alu-
minum extrusion technology, Chicago, USA, pp 125–134
42. Fitta I, Sheppard T (2000) Material flow and prediction of
extrusion pressure when extruding through bridge dies using
FEM. In Proceedings of 7th international aluminum extrusion
technology, Chicago, USA, pp 141–147
43. Demuth H, Beale M, Hagan M (2005) Matlab neural network
toolbox user’s guide. The MathWorks, Inc, Natick, MA
44. Arbib MA (1995) Handbook of brain theory and neural networks.
The MIT Press, Cambridge, MA. ISBN:0-262-01148-4
45. Ingarao G, Di Lorenzo R, Micari F (2010) Sustainability issues in
sheet metal forming processes: an overview. J Cleaner Product
19(4):337–347
46. Freeman JA, Skapura DM (1991) Neural networks: algorithms,
applications, and programming techniques. Addison-Wesley,
Reading MA
206 Neural Comput & Applic (2013) 23:195–206
123