International Journal of Production Research

Vol. 49, No. 17, 1 September 2011, 5079–5097

An intelligent approach to robust multi-response process design

Tatjana V. Sibalija*, Vidosav D. Majstorovic and Zoran D. Miljkovic

Faculty of Mechanical Engineering, University of Belgrade,

Kraljice Marije 16, 11120 Belgrade, Serbia
(Received 29 January 2010; final version received 12 July 2010)

In order to meet strict customer demands in a global highly-complex industrial

sector, it is necessary to design manufacturing processes based on a clear
understanding of the customer’s requirements and usage of a product, by
translating this knowledge into the process parameter design. This paper presents
an integrative, general and intelligent approach to the multi-response process
design, based on Taguchi’s method, multivariate statistical methods and artificial
intelligence techniques. The proposed model considers process design in a general
case where analytical relations and interdependency in a process are unknown,
thus making it applicable to various types of processes, and incorporates
customer demands for several (possible correlated) characteristics of a product.
The implementation of the suggested approach is presented on a study that
discusses the design of a thermosonic copper wire bonding process in the
semiconductor industry, for assembly of microelectronic devices used in
automotive applications. The results confirm the effectiveness of the approach
in the presence of different types of correlated product quality characteristics.
Keywords: Taguchi method; multi-response design; artificial neural networks;
genetic algorithm; multivariate statistical techniques

1. Introduction
Customer driven quality, in terms of meeting and exceeding customers’ expectation and
adding value to the customers, is one of the most emphasised issues in a modern industry.
In order to ensure high-quality of products in a highly competitive market, the transition
from approximation- or experience-based approaches to the knowledge-based approach
in translating customers’ expectations into design of industrial processes is required.
From another perspective, the majority of industrial processes generate products
characterised by several quality characteristics (responses) due to increased demand for
products with high complexity. Hence, the multi-response design has become a significant
topic in today’s manufacturing sector.
Implementation of the advanced cost-effective methodologies for product and/or
process quality improvement, such as Taguchi’s method, has been proven effective to fulfil
or exceed customers’ expectations. Taguchi’s robust design has been successfully used in
many single-response problems, but to date, the original Taguchi method has not proved
appropriate for designing or optimising the multi-response problems; the sole path was
relying on engineers’ judgement.

*Corresponding author. Email: [email protected]

ISSN 0020–7543 print/ISSN 1366–588X online

ß 2011 Taylor & Francis
DOI: 10.1080/00207543.2010.511476
http://www.informaworld.com
5080 T.V. Sibalija et al.

This paper presents an intelligent, general and integrative approach to the multi-
response process design based on Taguchi’s off-line static robust parameter design. The
intelligent system model for the multi-response robust process design (hereafter referred to
as IS_MR_RPD) was developed, composed of three stages: design of experiment,
processing and analysis of experimental data, and process modelling and optimisation.
In the first stage of the IS_MR_RPD application, by using expert system for the design
of experiment (ES_DoE) developed within the model, selection of the experimental plan
was performed with respect to the number of control and noise process parameters and
their levels, enclosing inner and outer orthogonal arrays.
After performing the experimental trials according to the selected experimental plan
and collection of experimental data for all responses, in the second stage experimental data
was transformed into corresponding Taguchi quality losses. Further data transformation
was performed by applying principal component analysis (PCA) to uncorrelated quality
losses of responses and grey relational analysis (GRA), resulting in formation of a
synthetic process performance measure based on the customer’s specifications for all
responses.
In the third stage, artificial neural networks (ANNs) were employed to model the
relationship between the synthetic process performance measure and process parameters,
presenting an objective function for a genetic algorithm (GA). The GA ensures
convergence to the global optimal parameter settings, by searching for the optimal
values among all possible solutions in a continual multidimensional space.
In the proposed model, analytic relationships between responses and process
parameters are unknown; the model does not imply any assumptions regarding the type
of process, type and number of process parameters, type and number of responses,
existence of correlations between responses or process parameters, or their interrelations.
In order to verify the proposed approach, the IS_MR_RPD was applied on the design
of a thermosonic wire bonding process, which is the most important process in
semiconductor assembly used to interconnect the integrated circuitry of the die to the
external world. The observed microelectronic device is used in automotive applications;
hence, the related production and assembly processes must meet the specific and rigorous
demands of the automotive market thus ensuring robust and reliable service of the
automotive device.
The rest of the paper is organised as follows. Section 2 presents a literature review. In
Section 3 the proposed IS_MR_RPD is described, drawing on the Taguchi quality loss
function, multivariate statistical methods and artificial intelligence techniques.
Implementation of the IS_MR_RPD on the design of a thermosonic copper wire bonding
process is explained in Section 4. Final notes regarding this study are included in the last
section.

2. Literature review
Multi-response design has become an increasingly emphasised issue in complex industrial
processes, particularly in situations where more than one correlated response must be
assessed simultaneously. However, most available procedures based on Taguchi’s method
have not proved fully functional for optimising the multi-response problem, especially in
a case of correlated responses.
Many recent studies were centred on solving multi-response optimisation problems.
 International Journal of Production Research 5081

Pignatiello’s (1993) regression approach considers a multivariate quadratic quality loss

function, but it was subjected to certain concerns: the proposed procedure does not
necessarily lead to the global optimum (Tsui 1999); the approach based on the regression
technique increases the computational process complexity, and the possible correlations
among the responses may still not be considered; in addition, a factor that is significant in
a single-response case may not be significant when considered in a multi-response case
(Su and Tong 1997).
Tsui (1999) proposed two-step procedures for the multiple characteristics problem
under a general class of models, but it was assumed that responses are uncorrelated with
each other, for both constrained and unconstrained minimisation with non-additive
models.
There are a variety of proposed methods based on the transformation of Taguchi’s
quality losses or signal-to-noise (SN) data for several responses by using the PCA method
(Su and Tong 1997, Antony 2000). However, they considered only principal components
with an eigenvalue greater than or equal to one, enclosing the larger portion of variance
but not the total variance of responses.
Wang and Tong (2005) used PCA and GRA to transform quality losses for several
responses into a single measure. Wu’s (2004) approach also employs PCA and GRA, but
it is based on proportion of quality loss with respect to the known starting conditions.
Various authors employed ANNs to optimise the multi-response process (e.g., Hsieh
and Tong 2001). Hsu (2001) employed ANN to model the relation between process
parameters and quality losses for several correlated characteristics and applied PCA
to uncorrelate the model by adopting only components with an eigenvalue greater or equal
to one.
Besides the stated limitations, the general shortcoming of the above methods is that
they consider only discrete parameter values used in the experiment, hence only parameter
levels used in the experimental trials could be selected for the optimal parameters setting.
In addition, the above methods could not solve multi-response problems where
optimisation requires the implementation of knowledge of experts into the formulae.
Response surface methodology (RSM) is a well-defined technique and the most
common approach for process design and optimisation, proven to be effective in many
applications. However, there are certain concerns regarding RSM application in multi-
response problems: the RSM does not enable simultaneous optimisation of both mean and
variance of the responses; when a process is influenced by a large number of variables and
is highly non-linear with multiple outputs, an RSM model may not find the overall best
solution and might be trapped easily in a local minimum (Antony 2000, Tong et al. 2004).
Detailed discussion regarding the above and other related approaches for multi-
response optimisation can be found in reference Sibalija and Majstorovic (2009).
The approaches based on a GA application could select the optimal parameter setting
in a continual multidimensional space for the multi-response problem, but they have
certain limitations for general application. Roy and Mehnen (2008) used desirability
function analysis (DFA) in Pareto front genetic optimisation, assuming that the analytical
model of the process is known which is not always the case in practice. Drain et al. (2004)
proposed a method that combines RSM and GA (shortcoming of RSM was commented
on above). Lau et al. (2005) used GA for the optimisation of moulding operations. Mok
et al. (2000) presented an intelligent system based on case-based reasoning, TOPSIS, ANN
and GA, to optimise an injection moulding process. Jeong et al. (2005) employed GA for
shadow mask manufacturing. Hou et al.’s (2006) method based on RSM, ANN and GA
5082 T.V. Sibalija et al.

presents an integrated system for wire bonding process optimisation. Tong et al.’s (2004)
approach is based on case-based reasoning, ANN and GA, dedicated to optimise transfer
moulding of electronic packages. The mentioned GA-based approaches are designed to
optimise a particular manufacturing process; hence, they are not suitable for general
application.
The proposed IS_MR_RPD attempts to overcome the deficiencies of the above
methods, owing to the novel procedure that employs:
. Taguchi’s quality loss function that adequately represents relative financial
significance of each response and simultaneously assesses the response mean and
variation;
. Multivariate statistical methods PCA and GRA to uncorrelate and synthesise
responses in such a way to ensure that the weights of responses in a synthetic
performance measure are based on the total variance of the original responses,
thus improving objectivity of the analysis;
. Artificial intelligence techniques to provide correct process modelling (ANN) and
ensure the global optimal solution (GA) in terms of the optimal process
parameters setting found in a continual multidimensional space that meets
customer demands for all responses.

3. Intelligent approach to multi-response robust process design

The IS_MR_RPD presents an intelligent, generic and integrative approach to the multi-
response design (or optimisation) of manufacturing processes based on Taguchi’s static
off-line robust design, composed of the following three stages (Sibalija 2009).

3.1 Stage I: design of experiment

By using orthogonal arrays and fractional factorial instead of full factorial, Taguchi’s
approach allows for an easy set-up of experiments with a very large number of factors
varied on few levels. In the presented model, the design of the experimental plan is
performed with respect to the number of control and noise process parameters and their
levels, by using the expert system (ES_DoE) designed for this purpose within the model.
According to the user input for the number of control factors and their levels and the
number of noise factors and their levels in the experiment, ES_DoE selects an inner
orthogonal array (for control factors) and outer orthogonal array (for noise factors),
forming the plan of experiment.

3.2 Stage II: processing and analysis of experimental data

The approach is based on Taguchi’s quality loss function, because it provides the right
metric for multi-criteria decision making due to its physical meaning. The quality loss
function directly represents a financial measure of a customer’s dissatisfaction with a
product’s performance as it deviates from a target value. Unlike the conventional
weighting methods, the quality loss function is a direct way to indicate the decision
maker’s preference and is simple to apply. The quality loss function is based on the SN
ratio that assesses simultaneously the mean value of the quality characteristic and its
 International Journal of Production Research 5083

variation, which is especially important in a case of multi-response optimisation.

In the proposed approach, Taguchi’s robust design was not applied directly, as not
every response may have the same measurement unit and may not be of the same category
in the signal-to-noise (SN) ratio analysis. The procedure consists of the following steps.
Step 1: Calculate the signal-to-noise (SN) ratio and the average quality loss for all
responses
After determination of the type of responses according to SN ratio, it is necessary to
calculate the SN ratios and average quality losses for all responses (Taguchi 1986):
8 !
> 1 Xn
>
> 10 log 2
y , for STB
>
> n i¼1 i
>
>
>
>  2   2
<
y 1 y
 ¼ 10 log 2   10 log 2 , for NTB
>
> s n s
>
> !
>
>
>
> 1X 1
n
>
: 10 log , for LTB:
n i¼1 y2i

As defined by Taguchi, the average quality loss is L ¼ K  MSD, where L is the existing
average loss per unit, K is the coefficient, and MSD is the sample mean square deviation
when n units of a product are measured (Taguchi 1986, Peace 1993) as follows:
8 n
>
> 1X 2
>
> y, for STB
>
> n i¼1 i
>
>
<1X n
n1 2
MSD ¼ ðyi  mÞ2 ¼ s þ ðy  mÞ2 , for NTB
>
> n n
> i¼1
>
>
> 1X n
>
> y2 , for LTB,
:
n i¼1 i

where y is measurable statistic of response; STB, NTB, and LTB are smaller-the-better,
nominal-the-best, and larger-the-better responses, respectively; y is the sample mean, and
s2 is the sample variance of n units.
Step 2: Transform the average quality loss into normalised quality loss (NQL)
Quality loss QLik of the ith (i ¼ 1, 2, . . . , p) response in the kth (k ¼ 1, 2, . . . , m) trial is:
0
ik
QLik ¼ K  MSD0ik ¼ 10 10
and it can be transformed into the normalised value NQLi(k) (NQLi(k) 2 [0, 1]):
min QLik
QLik  |{z}
i
NQLi ðkÞ ¼ :
max QLik  |{z}
|ffl{zffl} min QLik
i i

Step 3: Perform PCA on the NQL data and obtain the principal component scores Yi(k)
Principal component analysis (PCA) is considered as an effective means of determining
a small number of uncorrelated linear combinations which account for most of the
5084 T.V. Sibalija et al.

variance in the original number of responses. All principal components are uncorrelated
with each other. The sum of variances of the principal components (eigenvalues) is equal
to the sum of variances of the original responses.
Since this approach considers the general case where correlations among responses
exist, PCA is performed on the NQL data resulting in a set of uncorrelated components.
In contrast to the usual practice where only components with eigenvalues greater than or
equal to one are considered (Su and Tong 1997, Antony 2000), here, in order to involve the
total variance of the original data, principal component scores include all components.
The number of principal components and the number of components of an eigenvector
correspond to the number of responses. If a component of an eigenvector of the first
principal component PC1 is denoted as I1i, (i ¼ 1, . . . , p), the multi-response performance
statistics corresponding to PC1 for NQL could be expressed as:
X
p
Y1 ðkÞ ¼ I1i  NQLik :
i¼1

The larger the Yi(k) value, the better is the performance of the product/process.
Step 4: Perform GRA on principal component score Yi(k)
Grey relational analysis (GRA) provides an effective means of dealing with one event
that involves multiple decisions and deals with poor, incomplete and uncertain data.
Here, GRA is performed on the transformed principal scores, resulting in a single
multi-response performance measure that adequately takes into account all, possibly
correlated, response values with respect to the customer’s specifications. The weights used
in the presented method for determining the multi-response performance statistic are
based on the variance of the original responses (from PCA), which results in improved
objectivity of the experimental analysis.
GRA is performed on the absolute value of the principal component scores Yi(k). The
linear data pre-processing method is employed to transform the principal component
scores jYi(k)j into a set of standardised multi-response performance statistics Zi(k),
expressed as follows:    
max Yi ðkÞ  Yi ðkÞ
|ffl{zffl}
i
   
Zi ðkÞ ¼
max   min Yi ðkÞ ,
|ffl{zffl} Yi ðkÞ  |{z}
i i
where:
   
max Yi ðkÞand |{z}
|ffl{zffl} min Yi ðkÞ
i i

are the maximal and the minimal value of jYi(k)j for the ith response, respectively.
The grey relational coefficient i ðkÞ is:
   
min Zi ðkÞ  Z0 ðiÞ þ & |ffl{zffl}
min |{z}
|{z} max Zi ðkÞ  Z0 ðiÞ
max |ffl{zffl}
i i
i ðkÞ ¼  k  k  ,
Zi ðkÞ  Z0 ðiÞ þ & max max Zi ðkÞ  Z0 ðiÞ
|ffl{zffl} |ffl{zffl}
i k

where Z0(i) are ideal sequences with a value of 1, and & is called the distinguishing
coefficient (& 2 ½0; 1).
 International Journal of Production Research 5085

The grey relational grade k is calculated by a weighted mean, where the weight is
determined by the percentage of variance of the response NQL in PCA:
X
p
k ¼ !i i ðkÞ,
i¼1

where !i is the percentage of variance of the ith component in PCA:

X
p
!i ¼ 1:
i¼1

In this approach, the grey relational grade k is adopted as the synthetic performance
measure of a multi-response process.
Step 5: Calculate the factor effects and obtain the optimal parameter conditions
Knowing the synthetic performance measure values and process parameter (factor)
values for all experimental trials, it is possible to calculate the effects of factors on the
synthetic performance measure for all factor levels used in the experiment. By selecting the
maximum of factor effects on the multi-response performance statistic k , the optimal
parameter conditions can be obtained (hereafter, the above procedure described in Stage II
is referred to as the factor effects approach) (Sibalija and Majstorovic 2009).
The shortcoming of the factor effects approach is that it considers only those discrete
factor values that were used in experimental trials.

3.3 Stage III: process modelling and optimisation

Step 1: Training of ANNs with different architecture
The ANN is a powerful technique to generate complex multi-response models without
referring to a particular mathematical model, proven as effective in various applications
(Cook and Shannon 1992, Hsieh and Tong 2001, Hsu 2001). By applying an ANN to learn
and model the process behaviour, the process is considered a ‘black-box’; it does not
require any prior knowledge about the process interrelationships. This feature essentially
contributes to generality of the presented approach, because the model does not depend on
the type of relations between responses and process parameters or their correlations, thus
making it applicable to different processes.
Multilayer feed forward artificial neural networks (ANNs) were employed to model the
relationship between the synthetic performance measure (k ) and process parameters,
which presents an input (objective function) for the genetic algorithm (GA).
The number of neurons in the input layer corresponds to the number of process
parameters; the output layer has only one neuron. The neurones in the hidden layer are
computational units that perform non-linear mapping between inputs and outputs. For
the training of ANNs, the input data set contains process parameter values for all
experimental trials; the output set accommodates the synthetic multi-response perfor-
mance measure k (k ¼ 1, . . . , n; n is the number of experimental trials). The error back-
propagation (BP) learning method, improved by the Levenberg–Marquardt algorithm,
was adopted in this approach. The transfer functions for all hidden neurones are tangent
sigmoid functions, and for the output neurones are linear functions. It was proven that this
choice of transfer functions makes ANN capable of performing successful approximation
5086 T.V. Sibalija et al.

of various complex functions (i.e., Tong et al. 2004). BP learning employs a gradient
descent algorithm to minimise the mean square error (MSE) between the target data and
the ANN predictions.
During the training process, the learning rate controls the amount by which weights are
changed and the momentum avoids a major disruption of the direction of learning in a
presence of outliers in the training set. A smaller learning rate and larger momentum
reduce the likelihood that ANN will find weights that are a local, but not global minimum
(Hsu 2001). Thus, the adopted values for training parameters are: learning rate  ¼ 0:01
and momentum factor  ¼ 0:9.
Step 2: Selection of the best ANN
Determining the number of hidden neurones is critical in the design of an ANN. Since
the process modelling is the most sensitive part of this model, various ANNs with different
architecture (number of hidden neurons) are developed in MatlabTM until MSE of 103 is
achieved. The best ANN is chosen according to the minimum MSE criterion (mean square
error between the original (target) data and actual network output). The coefficient of
correlation between the original data and the actual network output (R value) is also
considered, with the acceptance level of 0.9 (Sibalija 2009). The training, testing and
verification data were obtained from the experimental results.
Step 3: Development of GAs with different selection and crossover functions
The trained neural model presents an objective (fitness) function for the GA. The GA,
by maximising the objective function finds the optimal parameter values among all
possible solutions in a continual multidimensional space and hence ensures convergence to
the global optimum. A GA was chosen for optimisation due to the following reasons: GA
is proven as a potent multiple-directional heuristic search method for optimising highly
non-linear, non-convex and complex functions; it is less likely to become trapped at a local
optimum than traditional gradient-based search methods (Ortiz et al. 2004, Roy and
Mehnen 2008).
In order to obtain the optimal performance of a GA, a large number of parameters
must be tuned.
According to the results of analysis (Ortiz et al. 2004), the choice of the basic GA’s
operations (selection and crossover functions) depends on the application (optimisation
problem). In order to consider the specifics of each particular problem and to enhance the
generality of the proposed model, nine GAs are developed in MatlabTM, employing the
most commonly used types of selection function (‘stochastic uniform’, ‘roulette wheel’
and ‘tournament’) and crossover function (‘single point’, ‘two point’ and ‘arithmetic’)
(Sibalija 2009).
Since the parameters setting obtained by the factor effects approach presents a
potentially good solution of the observed multi-response problem, it serves as a basis to
form the initial population in GAs. This feature is of essential importance because it allows
a GA to converge to the global optimum faster and enhance its capability to find the
global optimum in the given number of generations (Sibalija 2009).
The rest of the GA’s operating parameters are: natural chromosomes presentation;
population size equal or larger than five times dimensionality (number of process
parameters); reproduction parameters: elite count ¼ 2, crossover fraction ¼ 0.9; mutation
function is ‘adaptive feasible’ (Sibalija 2009).
 International Journal of Production Research 5087

A GA considers all continual parameter values between corresponding bounds, in

contrast to traditional experimentation methods that consider only those (discrete) values
that have been used in the experimental trials or in historical data. Relaying on this and
setting a GA’s parameters properly, the proposed method ensures optimal performance
of a GA to converge to the global rather than local optimum.
Step 4: Selection of the best GA
The nine tuned GAs were run for 2000 generations. The best GA is chosen according
to the best fitness value (on-line performance criteria), presented by the synthetic multi-
response performance measure. An additional criterion is the best off-line performance
(the mean of the best fitness values through the whole run) (Morrison 2003). Finally, the
solution of the best GA is adopted as the final solution of a multi-response problem:
optimal process parameter settings found in a continuous multidimensional space.

4. Implementation
4.1 Problem description
This study discusses the design of a thermosonic copper wire bonding process for the
assembly of microelectronic devices used in automotive applications.
Thermosonic wire bonding is the most widely used assembly technique in the
semiconductor industry, used to interconnect the internal circuitry of the die to the
external world. This method uses bond force, bond power, time, temperature and
ultrasonic energy to form the ball bonds at the die pads and welds at the output leads.
The main elements of the microelectronic device are: the die with internal circuitry and
output leads. In the microelectronic assembly, a gold wire is the most commonly used
interconnection material in the wire bonding technique. The high cost and potential
reliability degradation of a gold wire demand alterative interconnection materials. The
advantages of copper have rapidly established itself as a main material for gold
replacement: the price of copper wire is approximately six times lower than the price of
gold wire (50 mm cross-section diameter); mechanical properties of copper wire (elongation
and tensile strength) are superior to gold wire for this application; copper wire has 40%
more electrical conductivity than corresponding gold wire which turns up electronic signal
transmission rate thus improving performance of the device. By applying copper wire
bonding, a microelectronic device’s quality and reliability could be improved, simulta-
neously achieving cost reduction. However, significant oxidation of copper when exposed
to high temperatures (the wire bonding process is performed at 260–300 C) might cause
corrosion micro cracks inside the die pad, which decreases the interfacial shear strength
and weakens the bonds between aluminium pads and copper ball bonds at the die surface
(Sibalija 2004).
In order to meet high customer demands for quality and reliability of microelectronic
devices, an experiment was conducted to design one part of the copper wire bonding
process which refers to forming the bonds between copper wire balls and aluminium die
pads. The main purpose of the experiment was to ensure a reliable connection between the
copper balls and aluminium die pads, thus ensuring a reliable performance of the device.
Since the considered microelectronic device is used as a power amplifier in automotive
products, the quality and reliability issue is of particular importance due to the specific
conditions of automotive applications.
5088 T.V. Sibalija et al.

Table 1. Process parameters with different operating levels.

Level

Process parameter Unit Symbol Level ‘1’ Level ‘2’

Contact Power_bond mW CP1 0 30

Contact Force_bond gf CF1 250 400
Base Force_bond gf BF1 200 350
Base Power_bond mW BP1 40 80

4.1.1 Process parameters

The part of a thermosonic wire bonding process that refers to forming a bond between the
aluminium die pad and a copper wire could be described as follows (Sibalija 2004):
(1) A copper wire is fed down through the tool (capillary) inserted in the ultrasonic
transducer, which converts electrical energy and transmits resonant energy to the
tip of the tool.
(2) The special electronic flame off element (EFO) generates the electrical flame, which
is transmitted to the tip of the tool to form the copper wire ball.
(3) The tool with a copper wire moves down to the die pad surface, applying Contact
Force_bond and Contact Power_bond, to mesh copper ball and make the contact
between the copper ball and the aluminium pad surface.
(4) Base Power_bond and Base Force_bond are applied for the final inter-metallic
aluminium-copper bond formation.
(5) After forming the first bond at the die pad, the tool with a wire moves up forming
a loop and then goes down to the lead surface to form the weld between a copper
wire and the output lead.
This cycle is repeated until the microelectronic device is fully assembled (bonded). The
mentioned four process parameters were identified as significant in affecting the quality of
the copper-aluminium inter-metallic bond, and they were used as control factors in the
experiment. The rest of the wire bonding parameters (preheat and bond-site temperatures)
are not discussed in this study, since change in their values would affect the quality of the
welds between the copper wire and leads. All four process parameters are continual
variables. List of factors (process parameters) and levels used in the experiment are shown
in Table 1 (Sibalija 2004).

4.1.2 Quality characteristics (responses)

The quality of the considered part of a thermosonic copper wire bonding process is
determined by the strength of the ball bond at the die pad. The so called ball shear test was
performed to show the integrity of the bond between the copper ball and the aluminium
die pad. Since one microelectronic device contains several ball bonds, the ball shear test
was performed on all bonds in one device. The mean (average) ball shear test value of a
tested device was considered as the first response. The target specified value for the ball
shear test (for a copper wire 50 mm in diameter) is 160 [gf]. After performing the shear test,
it is necessary to consider the shear test failure mode. Copper has significantly higher
strength than aluminium, so the shear surface of a good copper-aluminium bond is
 International Journal of Production Research 5089

through the aluminium bond pad. The copper ball and inter-metallic layer are both
significantly stronger than the aluminium pad; but, a harder, stiffer copper ball could
result in bond defects, such as cratering bond failure, which can cause reliability problems
in the microelectronic device. Dark areas on a sheared pad surface indicate oxide damage
(bond failure), meaning that the silicon-oxide inside the die pad was damaged during wire
bonding process. The number of oxide damages (NoOD) found as shear test failure mode
in a tested device was considered as the second response (Sibalija 2004).
Characteristic BS (ball shear) is a continuous measurable response of the nominal-the-
best (NTB) type, because the objective is to achieve the nominal (target) value. Since the
existence of oxide-damage indicates a potential reliability problem, characteristic NoOD
belongs to the smaller-the-better (STB) type. By its nature, characteristic NoOD is of
attribute type. Here, NoOD actually presents the frequency of oxide-damage occurrences
in one sample expressed numerically (it can take discrete numerical values from zero to the
total number of bonds in the device). In the literature that discusses the Taguchi method,
such characteristics are considered attribute and their optimisation is performed by mean
of the number of occurrences in a sample (Peace 1993). The responses BS and NoOD are
directly correlated.

4.2 Process design by implementing IS_MR_RPD

4.2.1 Stage I
By using ES_DoE, the experimental plan was generated. Since there were four control
factors varied on two levels and no noise factors in the experiment, the output from
ES_DoE was an inner orthogonal array L8. The experimental plan contained eight trials
and five repetitions were added. The part of the experiment plan and observations are
shown in Table 2 (Sibalija 2004).

4.2.2 Stage II
Step 1: SN ratios, MSD and QL (quality loss) values were computed according to the
corresponding formulae.
Step 2: Transformation of QL values into normalised values (NQLi ðkÞ 2 ½0; 1) was
performed with respect to the maximal QL value in k experimental trials and the ideal case
where QL ¼ 0. The NQLs for the two responses are listed in Table 2.
Step 3: PCA was performed on NQL values. The principal component scores Yi(k)
corresponding to each trial are shown in Table 2. Table 3 lists the eigenvalues and
proportions of NQL of each response for the principal components. As stated before, all
principal components were considered in this approach, in contrast to the common
approach where only PC1 would be taken into account enclosing only 68.4% of the total
variance of responses. According to the eigenvectors data shown in Table 3, principal
component scores were computed by using the following formulae:
Y1 ðkÞ ¼ 0:707  NQLBS þ 0:707  NQLNoOD
Y2 ðkÞ ¼ 0:707  NQLBS  0:707  NQLNoOD :

Step 4: The principal component scores Yi(k) were first taken from the absolute value
and then transformed into a set of comparable sequences Zi(k). Next, the grey relational
 5090

Table 2. The part of experiment plan and experimental observations, NQLs, principal component scores and data of grey relational analysis.

i ðkÞ
Control factor Quality Normalised quality Principal component i ¼ 1, 2;
levels characteristics losses (NQLs) scores Yi(k) k ¼ 1, . . . , 13
k
Trial no. CP CF BF BP BS NoOD NQLBS NQLNoOD Y1(k) Y2(k) 1 2 k ¼ 1, . . . , 13

1 1 1 1 1 110.15 1 0.8460 0.0100 0.5911 0.6052 0.3742 0.3984 0.38304

2 1 1 1 2 175.81 10 0.1290 1.0000 0.6158 0.7982 0.3647 0.3343 0.35362
3 1 2 2 1 145.91 0 0.1200 0.0000 0.0848 0.0848 0.8064 0.8253 0.81331
4 1 2 2 2 213.59 0 0.9826 0.0000 0.6947 0.6947 0.3372 0.3659 0.347667
5 2 1 2 1 148.73 0 0.0947 0.0000 0.0669 0.0669 0.8408 0.8569 0.846643
T.V. Sibalija et al.

6 2 1 2 2 213.01 2 0.9636 0.0400 0.6530 0.7095 0.3512 0.3610 0.354767

7 2 2 1 1 146.94 0 0.1142 0.0000 0.0807 0.0807 0.8141 0.8324 0.820758
8 2 2 1 2 209.41 0 0.8452 0.0000 0.5976 0.5976 0.3717 0.4014 0.382519
9 1 1 1 1 109.42 1 0.8699 0.0100 0.6079 0.6220 0.3677 0.3918 0.376466
10 1 1 1 2 175.81 9 0.1290 0.8100 0.4815 0.6639 0.4234 0.3765 0.406291
11 1 2 2 1 143.48 0 0.1457 0.0000 0.1030 0.1030 0.7743 0.7955 0.782042
12 1 2 2 2 213.69 0 0.9846 0.0000 0.6961 0.6961 0.3368 0.3654 0.34720
13 2 1 2 1 146.57 0 0.1121 0.0000 0.0792 0.0792 0.8169 0.8349 0.82347
 International Journal of Production Research 5091

Table 3. Results of PCA performed on NQLBS and

NQLNoOD.

Principal component

PC1 PC2

Eigenvalues 1.3674 0.632

Proportions 0.684 0.316
Eigenvectors
NQLBS 0.707 0.707
NQLNoOD 0.707 0.707

Table 4. Summary of factor effects.

Control factor

Level CP1 CF1 BF1 BP1

1 0.4762 0.5063 0.4538 0.6922

2 0.6456 0.5823 0.6164 0.3653

coefficient i ðkÞ was calculated. Finally, the grey relational grade k was computed, where
the weights !i are [0.684; 0.632] for the first and the second component respectively
(proportions from Table 3). The results of GRA are listed in Table 2.
Step 5: Knowing the k values and the parameter (factor) levels corresponding to each
trial, the factor effects can be tabulated. The results are listed in Table 4. In multi-response
problems, the optimal setting of each factor is the one that yields the highest multi-
response synthetic performance measure. The optimal parameter conditions obtained
from the factor effects approach were: CP1 ¼ 30; CF1 ¼ 400; BF1 ¼ 350; BP1 ¼ 40. The
presented factor effects approach discusses only discrete parameter values used in the
experiment. The above set of parameter values was adopted as a basis to form the initial
population in the GA, in order to find the optimal solution in a continual multi-
dimensional space.

4.2.3 Stage III

Step 1: The set of BP ANNs was developed and trained in order to model the
relationship between the synthetic performance measure k and process parameters. Each
of the developed networks has four neurons in the input layer corresponding to four
process parameters and one neuron in the output layer corresponding to a single
performance measure. The results of the training of ANNs are presented in Table 5.
Step 2: From the data in Table 5, it could be concluded that the network 4-7-1 showed
the least error (MSE ¼ 0.00044) and therefore it was selected to present the process model
(Figure 1). The presentation of the training, validation and testing process of the selected
4-7-1 network is displayed in Figure 2.
5092 T.V. Sibalija et al.

Table 5. Results of ANNs training (MSE and R values for ANNs with different topology).

ANN architecture

4-2-1 4-3-1 4-4-1 4-5-1 4-6-1 4-7-1 4-8-1 4-9-1

MSE 0.000264 0.000147 0.000063 0.000144 0.000093 0.000044 0.000106 0.000150

R 0.99754 0.99798 0.99693 0.99812 0.99789 0.99769 0.99787 0.99810

Figure 1. Topology of the selected neural network (4-7-1).

Figure 2. Convergence of the selected 4-7-1 ANN (MSE versus learning iterations).

Step 3: The selected network presents an objective function for a GA. Nine GAs
were developed; the initial population was seeded close to the set suggested by the
factor effects approach; population size was 20. The results of the GAs are presented
in Table 6.
Step 4: GA 1 showed the best on-line performance or the best fitness value (0.96871),
which is confirmed by the off-line performance (0.96794), giving the optimal parameters
setting: CP1 ¼ 28; CF1 ¼ 400; BF1 ¼ 299; BP1 ¼ 40. This set is adopted as a final solution
Table 6. GA settings and results.

GA

GA 1 GA 2 GA 3 GA 4 GA 5 GA 6 GA 7 GA 8 GA 9

Stochastic Roulette Stochastic Roulette Stochastic Roulette

Selection uniform wheel Tournament uniform wheel Tournament uniform wheel Tournament
Crossover Single point Two point Arithmetic

On-line performance 0.96871 0.96860 0.92552 0.96837 0.96848 0.92552 0.95312 0.96807 0.92552
Optimal set
CP1 28.41 28.64 15.00 29.84 28.85 15.00 20.59 30.13 15.00
CF1 400.00 400.00 400.00 400.00 400.00 400.00 400.00 400.00 400.00
BF1 299.24 300.72 300.00 299.95 300.53 300.00 301.87 306.36 300.00
BP1 40.00 40.00 45.00 40.77 40.55 45.00 45.47 42.58 45.00
International Journal of Production Research

Off-line performance 0.96794 0.96727 0.92552 0.96772 0.96755 0.92552 0.94517 0.94772 0.92552
5093
5094 T.V. Sibalija et al.

Figure 3. The best and the mean fitness function values versus generations for GA 1.

Table 7. Comparative analysis of optimal parameter settings results obtained by

different methods.

Method

RSM Factor effects IS_MR_RPD

Optimal parameter settings [20; 350; 300; 45–55] [30; 400; 350; 40] [28; 400; 299; 40]
Synthetic performance 0.8514–0.7166 0.9613 0.9687
measure ()

of the observed multi-response problem. Figure 3 gives presentation of a fitness function

across generations for GA 1. It is visible that GA 1 converged early to the best fitness value
(in the 260th generation), which is the consequence of a good-seeded initial population.

4.3 Discussion
The implementation of IS_MR_RPD on the observed multi-response problem was
compared to RSM and the factor effects method application. In the RSM application,
using a combined response technique the superposition plot was formed by superposing
contour plots of all responses in order to find the specific area on the superposition plot
that meets specifications for all responses (the correlations among responses are not
discussed) (Sibalija 2004).
Table 7 provides a comparison of the simulated synthetic performance measure values
obtained from the three methods for the analysis. It could be concluded that the factor
effects approach, that considers correlations among responses, showed better results than
the RSM method. Application of IS_MR_RPD resulted in a better solution than the
factor effects approach, due to the search over a continual space. It is important to note
that the obtained multi-response performance value ( ¼ 0.9687) is very close to the ideal
value of 1 ( 2 ½0; 1).
Regarding the GA settings, from Table 6 it is visible that GAs with the ‘tournament’
selection function show significantly lower fitness function values and different optimal
 International Journal of Production Research 5095

parameter set than the other six GAs, which could mean that the ‘tournament’ selection is
not adequate for the observed optimisation problem. This proves the necessity to consider
a different GA’s functions for each optimisation problem.

5. Conclusion
The majority of today’s industrial products are defined by several quality characteristics,
hence, the multi-response design of complex industrial processes has become an
increasingly important and demanding task.
The shortcomings of the traditional Taguchi robust parameter design, as well as of
the other approaches discussed in Section 2, could be overcome by implementing the
IS_MR_RPD knowledge-based approach for the multi-response process design based on
a customer’s specifications for several, possible correlated, quality characteristics of a
product.
The analysis of several experimental results confirms the effectiveness of the model and
emphasises universality of application for static multi-response problems (Sibalija 2009).
Analysis of IS_MR_RPD application on the presented experiment and its comparison
to the other two methods for multi-response optimisation showed that the proposed
approach can yield a better solution in terms of the optimal parameters setting and the
multi-response performance measure, but also proved that it could successfully include
both numerical and attribute nature of responses. The implementation of IS_MR_RPD
could expand the application of Taguchi’s robust design for solving the multi-response
design problems in industrial practice.
The significance and the prospect of the presented IS_MR_RPD could be summarised
as follows:
. In the factor effects approach, the relative significances of responses are
adequately presented via Taguchi’s quality loss function and no assumptions or
assigned weights are required. Further, responses are synthesised into a
single performance measure based on the total variance of the original data,
which maximise the objectivity of the analysis and the effectiveness of the
approach.
. In contrast to GA-based approaches mentioned in the literature review,
IS_MR_RPD offers two novel perspectives. First, since the solution obtained
by the factor effects approach presents a potentially good solution, it was adopted
as a starting point for GA to search for the global optimum in a continual space.
This feature significantly improves the speed of the convergence to the global
optimum, therefore maximising the probability that the model will find the actual
global solution in the given number of generations. Second, various types of the
main GA’s operators (selection and crossover function) were considered for each
problem in order to find the GA’s setting that is the most adequate for the
observed problem. In the presented study, it was proven that the choice of
selection and crossover type could affect the quality of the final solution.
Therefore, this feature significantly contributes to a generality of the model,
meaning that it could consider specifics of each particular multi-response
problem.
. As a consequence of the above mentioned advantages, it could be concluded that
the proposed IS_MR_RPD offers a generalised and effective solution for the
5096 T.V. Sibalija et al.

parameter design of static multi-response processes. Moreover, since no assump-

tions regarding the nature of the process, parameters, responses or their
correlations or interrelations have been imposed in IS_MR_RPD, it can be
applied to various production processes in diverse industrial sectors.

References

Antony, J., 2000. Multi-response optimization in industrial experiments using Taguchi’s quality loss
function and principal component analysis. Quality and Reliability Engineering International,
16 (1), 3–8.
Cook, D.F. and Shannon, R.E., 1992. A predictive neural network modelling system for
manufacturing process parameters. International Journal of Production Research, 30 (7),
1537–1550.
Drain, D., et al., 2004. A genetic algorithm hybrid for constructing optimal response surface designs.
Quality and Reliability Engineering International, 20 (7), 637–650.
Hou, T.H., et al., 2006. An integrated system for setting the optimal parameters in IC chip-
package wire bonding processes. International Journal of Advanced Manufacturing Technology,
30 (3–4), 247–253.
Hsieh, K.L. and Tong, L.I., 2001. Optimization of multiple quality responses involving qualitative
and quantitative characteristics in IC manufacturing using neural networks. Computers in
Industry, 46 (1), 1–12.
Hsu, C.M., 2001. Solving multi-response problems through neural networks and principal
component analysis. Journal of the Chinese Institute of Industrial Engineers, 18 (5), 47–54.
Jeong, B., Lee, J., and Cho, H., 2005. Efficient optimization of process parameters in shadow mask
manufacturing using NNPLS and genetic algorithm. International Journal of Production
Research, 43 (15), 3209–3230.
Lau, H.C.W., et al., 2005. Design and implementation of a process optimizer: a case study on
monitoring molding operations. Expert Systems, 22 (1), 12–21.
Mok, C.L., Kwong, C.K., and Lau, W.S., 2000. An intelligent hybrid system for initial process
parameter setting of injection moulding. International Journal of Production Research, 38 (17),
4565–4576.
Morrison, R.W., 2003. Performance measurement in dynamic environments. In: Proceedings of
GECCO 2003: workshops, genetic and evolutionary computation conference, 12–16 July,
Chicago, Illinois, 99–102.
Ortiz, F., et al., 2004. A genetic algorithm approach to multiple-response optimization. Journal of
Quality Technology, 36 (4), 432–450.
Peace, S.G., 1993. Taguchi methods: a hands-on approach. Reading, MA: Adison-Wesley.
Pignatiello, J.J., 1993. Strategies for robust multiresponse quality engineering. IEE Transactions,
25, 5–15.
Roy, R. and Mehnen, J., 2008. Dynamic multi-objective optimisation for machining gradient
materials. CIRP Annals - Manufacturing Technology, 57 (1), 429–432.
Sibalija, T., 2004. Attaining process robustness through design of experiment and statistical process
control. In: Proceedings of 11th CIRP international conference on life cycle engineering, 20–22
June, Belgrade, Serbia, 161–168.
Sibalija, T., 2009. Development of intelligent designer of experiment model for application of Taguchi
method. Thesis (PhD). Faculty of Mechanical Engineering, University of Belgrade.
Sibalija, T. and Majstorovic, V., 2009. Multi-response optimisation of thermosonic copper wire-
bonding process with correlated responses. International Journal of Advanced Manufacturing
Technology, 42 (3–4), 363–371.
 International Journal of Production Research 5097

Su, C.T. and Tong, L.I., 1997. Multi-response robust design by principal component analysis. Total
Quality Management, 8 (6), 409–416.
Taguchi, G., 1986. Introduction to quality engineering: designing quality into products and processes.
Tokyo: Asian Productivity Organization.
Tong, K.W., Kwong, C.K., and Yu, K.M., 2004. Intelligent process design system for the transfer
moulding of electronic packages. International Journal of Production Research, 42 (10),
1911–1931.
Tsui, K.L., 1999. Robust design optimization for multiple characteristic problems. International
Journal of Production Research, 37 (2), 433–445.
Wang, C.H. and Tong, L.I., 2005. Optimization of dynamic multi-response problems using grey
multiple attribute decision making. Quality Engineering, 17 (1), 1–9.
Wu, F.C., 2004. Optimising robust design for correlated quality characteristics. International Journal
of Advanced Manufacturing Technology, 24 (1–2), 1–8.
Copyright of International Journal of Production Research is the property of Taylor & Francis Ltd and its
content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for individual use.