10.1177@1063293X20908317

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

Standard Article

Concurrent Engineering: Research


and Applications
Use of the stochastic frontier and the 1–20
Ó The Author(s) 2020
grey relational analysis in robust design Article reuse guidelines:
sagepub.com/journals-permissions

of multi-objective problems DOI: 10.1177/1063293X20908317


journals.sagepub.com/home/cer

Mohamed Ali Rezgui and Ali Trabelsi

Abstract
A generic procedure for robust design in developing products and processes, which is referred to as RDPP-SF has been
proposed. The method uses the stochastic frontier model to encompass both stochastic noise (e.g. manufacturing unit-
to-unit variation, and measurement errors) and special-cause variation (e.g. environment, customer use, wearing, and
deterioration noises). Even then, the RDPP-SF method has fallen short of tackling robust design of multi-objective prob-
lems, and its applicability is restrained to the performance characteristics of magnitude type (i.e., ‘‘the larger is the bet-
ter’’ or ‘‘the smaller is the better’’). Aiming at these limitations, the article seeks to address the robust design of the
multi-objective problems using the RDPP-SF method. This is performed by reassessing the procedural scheme of the
RDPP-SF method and the statistical significance of the hypothesis test (H0: g = 0 vs H1: g . 0) at 5% level. Depending
on the statistical significance of the test (H0: g = 0 vs H1: g . 0), the arrays of the extrinsic and/or the intrinsic noise
insensitivity scores are assigned to the grey relational analysis matrix as performance measures. The most robust design
solution for the multi-objective problem is then obtained by sorting the overall grey relational grades. The amended
RDPP-SF method is finally demonstrated using three industrial multi-objective case studies.

Keywords
robust design, Taguchi method, multi-objective optimization, grey relational analysis, stochastic frontier, intrinsic/extrinsic
noise insensitivity, RDPP-SF

Introduction and monitoring of the design parameters to commensu-


rate the system inherent variation. The method pro-
In the literature, many definitions have been given to ceeds according to three major steps: concept design,
the robust design perception. The pivotal point was parameter design, and tolerance design (Kackar, 1985;
how desensitizing the product/process development to Unal and Dean, 1991). Even though the burden of
the environmental sources of variation, even though developing robustness in technology and concepts falls
these sources are not eliminated (Fowlkes and on the concept and the parameter design steps
Creveling, 1995). Therefore, a special emphasis is (Taguchi and Phadke, 1989), the tolerance design is still
placed on reducing variation which emanates from the needed to hinder the effect of the residual variation and
natural and the non-natural sources. The applicability assure the system functionality (Taguchi, 1995).
of the Taguchi off-line procedure for quality engineer- The Taguchi method combines experimental design
ing (QE), which is acknowledged as the most significant techniques with the signal-to-noise ratio (S/N) to assure
contribution of Dr Taguchi is broad in scope, and it is
intended to optimize product and process design in
terms of performance, quality, cost, manufacturability, Ecole Nationale Supérieure d’Ingénieurs de Tunis, Laboratoire de
and reliability longer before the production’s imple- Mécanique, Productique et Energétique (LR18ES01), Tunis, Tunisia
mentation stage (Bagchi, 1993; Taguchi et al., 2004;
Yang and El-Haik, 2003). The Taguchi method for Corresponding author:
Ali Trabelsi, Ecole Nationale Supérieure d’Ingénieurs de Tunis, Universite
robust design is carried out by choosing outperforming de Tunis, 5 Taha Hussein, 1008 Tunis, Tunisia.
concepts and technologies in terms of noise sensitivity Email: Ali.Trabelsi@rnu.ensit.tn
2 Concurrent Engineering: Research and Applications 00(0)

robustness. However, it is admitted that there is no sin- solution. However, the Taguchi tactic and the S/N
gle technique that can solve different kinds of prob- metric have been criticized by countless of statisticians
lems. The RDPP-SF method (Trabelsi and Rezgui, and research fellows. Some discordance issues are
2020) has adddressed robustnes using the stochastic following:
frontier (SF) model and the designed experiments. It
determines the design parameter levels, which yield the 1. The Taguchi method is restrained to single-
least sensitive product/process response. The SF model objective problems (Chen, 1997). Also, the statisti-
is also utilized in the econometric studies to assess the cal bases of the method remain open to discussion
technical inefficiency of firms, which use the same tech- in terms of efficiency and testimony (Box, 1988;
nology. Both the robustness perception and the ineffi- Nair et al., 1992; Tsui, 1992) and the procedural
ciency measurement are made possible because of the scheme of the method in terms of experimentation
two-term error structure of the SF model, that is, one is missing (Antony et al., 2004).
accountes for the neoclassic natural variation (e.g. unit- 2. The capability of the Taguchi S/N metric to be
to-unit, measurement, exogenous parameters) and the receptive of induced noise in developing systems is
other encapsulating the non-natural variation sources. inappropriate since the sample mean and variance
For the econometric studies, the non-natural error term should be analyzed separately for a Gaussian pro-
explains why the actual output of the decision-making cess (Maghsoodloo et al., 2004; Steinberg, 1996).
units (DMUs) cannot meet what it is postulated by the 3. The Taguchi linear graphs, which are used to
production function beyond random disturbance. For assign the design parameters and their interactions
the robustness study, the non-natural error term to the columns of the Taguchi orthogonal arrays
explains why a design solution having a small natural (OAs), are confounding and overlook the effect of
variation is non-robust. Then, a parallel can be drawn the interactions. The choice of the parameters’
between the measurement of the technical inefficiency combination holding the most promise for robust-
of the DMUs and the RDPP-SF method since both ness is then unreliable (Maghsoodloo et al., 2004).
make use of the error structure of the SF model in an Above all, the Taguchi method ignores the interac-
equivalent fashion. tion effect between the design parameters (control-
The remainder of the article is organized into four lable factors) and the noise variables in the
sections. Section ‘‘Literature review’’ reviews the institution of the crossed arrays and the S/N
Taguchi method for robust design. Section metric (Myers et al., 2016).
‘‘Measurement of the TE using the SF technique’’ eluci- 4. The crossed array, which is used to model the
dates the math model of the SF method and draws a design and the noise variables, is not flexible and
parallel between the RDPP-SF method (Trabelsi and computationally ineffective (Myers et al., 1992).
Rezgui, 2020) and the concept of the technical ineffi- 5. The expected loss function and the related process
ciency (TE) of firms in econometrics. Section model may fall short of the robustness expectation
‘‘Parameter design for the multi-objective problem’’ dis- level. Many research works (Dellino et al., 2015;
cusses the amendments which have been performed for Park et al., 2006; Pignatiello, 1988) argue that inac-
the RDPP-SF method to make it amenable to robust curate results may happen.
design of multi-objective problems. Finally, in section 6. Both the OAs and the S/N tools do not account
‘‘Empirical examples,’’ three industrial multi-objective for constrained design problems (Park et al., 2006).
case studies are chosen to elucidate the breadth and the Moreover, because the Taguchi method does not
practical utilization of the modified RDPP-SF method. invest the whole ranges of the design parameters,
the tactic cannot anchor to the reality (Dellino
et al., 2015; Park et al., 2006).
Literature review
7. In the dynamic Taguchi design (i.e. a certain corre-
As we argued earlier, the Taguchi method addresses lation exists between the signal and the process
robustness in developing products/processes using three response), the maximization of the S/N metric may
major steps: concept design, parameter design, and tol- result in minimum average squared error loss;
erance design. The Taguchi’s parameter design thus, in case the target is the ‘‘nominal is the bet-
(Bisgaard, 1992; Box, 1988; Box and Fung, 1986; ter’’ (NTB), the S/N technique is appropriate only
Czitrom, 1989; Leon et al., 1987) is an optimization if the standard deviation and the mean of the
step, which uses a crossed array design to take account response are proportional (Leon et al., 1987). In
of the control and the noise parameters. The signal-to- the static Taguchi design, however, the S/N should
noise ratio (S/N) metric is used to determine the combi- be avoided at just about any cost when the perfor-
nation levels which yield the most robust design mance characteristic (PCH) is of magnitude type
Rezgui and Trabelsi 3

and the coefficient of variation (CV) is lesser than optimization procedure of the dual response surface
17% (Maghsoodloo et al., 2004). since spurious methodology (DRSM). A data-driven approach is con-
results may occur. ducted to determine the weights for the design para-
meters when the prior information is vague. This is
In view of the limitations of the Taguchi method for performed using an ideal ‘‘efficient curve,’’ which is
robust design, much research has been undertaken. obtained from the sample data set. Shaibu and Cho
Leon et al. (1987) state that the S/N technique takes (2009) retrieve the DR approach, but on the basis of a
advantage of the existence of special design parameters quadratic model for the RSM. They showe that higher-
called ‘‘adjustment parameters.’’ However, in many order polynomial models are more suitable to assure
design problems, the S/N metric is not independent of robust design solutions than ordinary linear models.
the adjustment parameters giving rise to non-optimal The method proves more effective and realistic since it
setting for the design parameters. Leon et al. (1987) takes into account the curvature effect, which emanates
develop a technique, which states the performance mea- from the factors’ interactions. The analysis uses statisti-
sure independently from the adjustment parameters; cal metrics such as the prediction error sum of squares
therefore, the adjustment parameters are managed in a (PRESS), the root mean square error (RMSE), the
more general fashion than the Taguchi S/N technique coefficient of determination (R2 ), the adjusted R-
does. Based on the Leon’s assumption, Box (1988) squared (R2a ), and the predicted R-squared (R2pred ) to
asserted that the same result could be obtained when choose a group of factors in the higher-order form.
logging the system response in order to stabilize the Freeny and Nair (1992) propose a generic approach,
variability that is induced by the engineering and the which can address robust design when the noise vari-
noise parameters. A more suitable transformation of ables are uncontrolled but observable. The noise vari-
the system response is recommended using the l-plot ables are used as covariates, and the model’s location
(Kunert and Lehmkuhl, 1997). parameters and regression coefficients are expressed in
Welch et al. (1990) propose a single-objective design function of the design factors. The coefficients are used
methodology, which is based on a computer simulation as substitute of the Taguchi S/N metric in order to
of the quality characteristic, Y = f(xcontrol, xnoise). The reduce the system variability. Chen et al. (1996) develop
minimum loss statistic is developed so that maximum a modified Taguchi method, which can reduce the var-
S/N ratio leading to the least sensitive design solution iation in the system response because of the noise and
for the engineering parameters (xcontrol) could be rea- the control parameters. The method uses the RSM and
lized. The computational load of the design method, the decision support problem (DSP) techniques and it
however, is low in comparison to the Taguchi crossed is advocated when the design problem has not a closed-
array procedure for robust design. Shoemaker et al. form solution. Bras and Mistree (1994) introduce the
(1991) enhanced the Welch’s method (Welch et al., compromise DSP to model engineering decisions
1990) by developing Welch’s response combined array involving multiple trade-offs. They focus on how using
and investigating the control-by-noise interaction plots, the Suh’s independence, the information axioms and
which yield minimum effects of the noise factors. As a the robust design in a concurrent fashion. The
result, the optimal combination of the engineering para- Taguchi’s S/N ratio is used as a metric for the reduc-
meter levels is readily obtained. Though, the method is tion in the information content and increasing the
disadvantageous since the product–array formulation is robustness of the design.
still ineffective and it estimates many inutile effects. Scibilia et al. (2001) develope a novel procedure for
Vining and Myers (1990) devise an alternative proce- robust design, which is based on cross-product arrays.
dure to the Taguchi approach, which was named as The method estimates the effects of the control factors
‘‘dual response (DR).’’ The model integrates the on the noise factors rather than the S/N metric as a
response surface methodology (RSM) into the robust response. Scibilia’s method is advantageous since it
parameter design (RPD). The mean and the standard provides much information about the system variation
deviation of the system response are stated in function using fewer experiments than the Taguchi method.
of the control variables using the RSM model. Then, Lesperance and Park (2003) use a joint generalized lin-
the regression of the mean and the standard deviation ear regression model (GLM) to analyze robustness in
functions are estimated using the ordinary least square dynamic characteristic problems. Basically, the model
(OLS) estimator. Lin and Tu (1995) develope a more employs the standard regression modeling techniques
effective and simpler optimization procedure for the and uses the whole data rather than summary statistics.
DR approach, which allows more general response Pal and Gauri (2010) devise an original method for
models to be entertained. the robust design of multi-objective problems. The
Ding et al. (2004) devise a weighted mean squared method integrates the multiple regression and the
error (WMSE) approach, which can improve the Taguchi S/N techniques. The total S/N values and the
4 Concurrent Engineering: Research and Applications 00(0)

proximity of the observed responses to the target val- systems which are laced with sensitivity to noise by
ues, which are reflected in the expected mean squared accommodating two variation components: one is the
error (MSE) for the individual responses, are utilized to intrinsic noise insensitivity (INI)—which represents the
determine the most robust design solution. Bashiri et al. neoclassic unit-to-unit manufacturing variation—and
(2016) investigate the multivariate S/N for the multi- the other is the extrinsic noise insensitivity (ENI)—
objective design problems using the dynamic Taguchi which accounts for the non-natural sources because of
procedure. The study can determine the optimal levels environment, customer use, and deterioration. Based
of the controllable design parameters in a continuous on the regression form of the SF function and a series
space and allows for increasing robustness in the of statistical hypothesis tests, the most robust design
dynamic Taguchi method. Pal and Gauri (2017) devel- solution is chosen depending on the ENI and/or the
ope a procedure, which uses the multi-regression (MR) INI scores. Inopportunely, the RDPP-SF method
technique and the Taguchi dynamic S/N ratio (DSN) to (Trabelsi and Rezgui, 2020) cannot address the multi-
optimize the multi-response dynamic systems. First, the objective problems and does not accommodate the per-
multiple regression equations for the dynamic system formance characteristics of the NTB type. Purposely,
are fitted based on the sample data, then, the dynamic the article seeks to alleviate the former limitation of
S/N ratio (known as MRDSN) for the different the RDPP-SF method.
response variables is calculated based on the multiple
regression values. The weighted MRDSN is considered
as the objective function for the optimization problem. Measurement of the TE using the SF
Gauri and Pal (2017) argue that a multiple regression– technique
based weighted S/N ratio (MRWSN) performance
metric can overcome the limitations of the composite The SF technique can be traced back to the works of
desirability function (CDF) for the multi-objective Aigner et al. (1977), Meeusen and Van Den Broeck
problems. The results have shown that the MRWSN (1977), and Schmidt and Lovell (1980). In econo-
method is superior to the CDF method in terms of opti- metrics, the SF methodology is often used to estimate
mization performance. Pal and Gauri (2018) revise the the TE of the DMUs, which work under the same pro-
MRDSN method (Pal and Gauri, 2017) by introducing duction technology. The TE of a DMUi can be defined
a new performance metric called weighted predicted in either input conserving (i.e. given an output, use of
response-based S/N ratio (WPRSNR). The experimen- the inputs in optimal proportions with respect to their
tal findings have shown that the modified MRDSN respective prices) or input expending fashion (i.e. maxi-
method outperforms the MRDSN procedure in search- mization of an output forthcoming from the input com-
ing the optimal design solution with respect to the total binations levels) (Coelli et al., 2005; Farrell, 1957). The
S/N and the MSE of the individual responses. SF model is rooted in two error structures (Forsund
Ozturk et al. (2006) propose a new approach for the et al., 1980): (1) a symmetrical random error, which rep-
robust design of multi-objective optimization based on resents the natural unit-to-unit manufacturing distur-
the Taguchi’s parameter design and the integration of bance, the measurement error and the uncontrolled
neural networks and genetic algorithms to optimize the variables, and (2) another one-sided component—
process of the parameters setting in concurrent engi- which reflects the DMU inefficiency to utilize resources,
neering. The neural network is used to construct the that is, latent sources that may prevent the DMUs to
system response function. Then, a genetic algorithm is achieve their maximum feasible production level
applied to the network for searching the parameter set- beyond natural causes. For a DMUi, the output-
ting with an optimal response. Chen et al. (2011) devise oriented TE is the ratio of the actual to the best practice
a robust product design process, which can help firm output accounting for the DMUi observed inputs,
improve quality, reduce variation, and increase cus- the production technology, and the socioeconomic fea-
tomer satisfaction and enterprise profit. The proposed tures (Murillo-Zamorano, 2004). In the input-oriented
product design methodology proceeds in four interre- TE, the TE may be defined as the ratio of the actual to
lated steps that integrate the grey relation approach to the best practice input levels given the firm’s observed
quality function deployment (QFD) and QE. output, the production technology, and the socioeco-
Recently, Trabelsi and Rezgui (2020) develop an nomic features (Murillo-Zamorano, 2004). Both mea-
original procedure for robust design of products and sures of the TE can be obtained from the econometric
processes using the stochastic frontier methodology estimation of the SF model as suggested by Battese and
(RDPP-SF). To our knowledge, no other parameter Coelli (1993, 1995). The Farrell empirical estimate of
design method has employed the SF mathematics to the output-oriented TE is expressed as the ratio of the
address robustness in developing products and pro- observed production, yi, to the optimal production, yi
cesses. The method uses the SF model to analyze
Rezgui and Trabelsi 5

(Farrell, 1957). Figure 1 illustrates the Farrell output-


oriented TE.
Assume a cross-sectional data of N DMUs being
indexed i (i = 1, ..., n), each using M inputs j (j = 1,
..., m), which are represented by the input vector, xi, to
produce an output, yi. Farrell’s TE of the DMUi is
stated as in equation (1). An equivalent point estimator
of the TE has been proposed by Battese and Coelli
(1988)

yi f ðxi ; bÞevi eui


TEi =  = = eui 2 ½0, 1 ð1Þ
yi f ðxi ; bÞevi
where f (xi ; b) and ei = vi  ui are the deterministic
frontier and the composed error term, respectively;
viid 2
i; N(0; sv ) represents the random variation in the out- Figure 1. Farrell’s output-oriented technical inefficiency
put due to exogenous sources; and uiid 2
i; jN (m; su )j represents
(Trabelsi and Rezgui, 2020).
the stochastic shortfall of the ith DMU from the ideal pro-
duction frontier assuming output-oriented TE. production process such that the optimal combination
Coelli et al. (2005) argued that the best predictor of of the parameter levels will lead to the most robust
ui is E(ui jei ) as stated in equation (2) design solution, that is, locate closer to the SF envelope.
For a designed experiment, the regression form of the
Actual output Eðyi jui , Xi Þ SF envelope is obtained from the n outputs of the n
TEi = Eðui jei Þ = =
Potential output Eðyi jui = 0, Xi Þ trials using FRONTIER 4.1. Each trial corresponds to
= expðui Þ a combination of the design parameter levels. The esti-
mates of the natural (Vi) and the non-natural (Ui) varia-
ð2Þ tions are associated with the ENI (ENIi = eui ) and the
In the literature, many functional forms of the SF func- INI (INIi = evi ) scores. The INIi and the ENIi scores
tion have been considered. The Cobb–Douglas (CD) belong to (0, + N) and (0,1), respectively, and the tar-
(linear logs of outputs and inputs) and the TransLog get value (full robustness) for both scores equals 1 (ui =
(TL) (a generalization of the CD function) are likely the vi = 0). Practically, this corresponds to a design solu-
most popular, and this is because of their flexibility and lin- tion wherein the predicted value by the SF model and
earity in parameters. To facilitate the maximum likelihood the observed output locate both on the SF envelope as
(ML) estimation of the SF parameters and the TE scores shown in Figure 1. Therefore, the parameters combina-
of the DMUs, Battese and Corra (1977) proposed a new tion yielding INI = ENI = 1 would hold the most
parameterization, g = s2 =s2 + s2v , which is adopted in promise for robustness (Trabelsi and Rezgui, 2020).
the FRONTIER 4.1 computer program (Coelli, 1996; The major contribution of the article resides in the amend-
Drinkwater and Harris, 1999). In this article, the ment of step 4 in the RDPP-SF flowchart of Trabelsi and
FRONTIER 4.1 program is used to estimate the SF model Rezgui (2020), which is shown in Appendix 1. The multi-
parameters, the ENI scores, and check hypothesis tests con- objective problem has been addressed according to the fol-
cerning the appropriateness of the CD versus TL and the lowing new functional scheme (see Figure 2):
statistical significance of the ENI component.
1. Each of the PCH should be handled separately.
Parameter design for the multi-objective Assume a threshold of 95% and a test significance
level of 5%; check the hypothesis (H0: g = 0 vs
problem
H1: g . 0) to examine the incidence of the non-
The procedural scheme of the RDPP-SF methode is natural sources of variation (ENI component) as
systematic (Trabelsi and Rezgui, 2020). The method expressed by the ENI scores for each trial (design
has borrowed the concepts of the TE of the DMUs, solution). If the test is significant (i.e. the hypoth-
which are implemented in countless of the econometric esis H0: g = 0 is to be rejected), then two cases
studies. Table 1 draws a parallel between the terminolo- should be distinguished: (a) the test is strongly sig-
gies used in the econometric and the RDPP-SF method. nificant (i.e. g ø 95%), then estimate of the ENI
The RDPP-SF benefits from the structure of the score array is set in the grey relational analysis
error term in the SF model. It accommodates the input (GRA) matrix (Deng, 1989) as performance mea-
resources (engineering parameters) and the underlying sure for the PCH; or (b) the test is marginally
6 Concurrent Engineering: Research and Applications 00(0)

Table 1. Terminologies used in the econometric and the RDPP-SF method.

Econometric RDPP-SF method

Production function Technology/process function


Industry technology Process method
Production frontier Process frontier
Decision making unit (DMU) Design solution (any trial in the designed experiment)
Cross-sectional data Planned or non-planned designed experiment
Panel data Replicates of designed experiment
exp(2Ui): DMUi’s TE Extrinsic noise insensitivity (ENI) of the design solution, i (experiment trial)
exp(Vi): DMUi’s random variation Intrinsic noise insensitivity (INI) of the design solution, i (experiment trial)
Input level Parameter level

Figure 2. Functional scheme of the modified RDPP-SF method for multi-objective problems.
Rezgui and Trabelsi 7

significant (i.e. g \ 95%), in which case estimate Table 2. Taguchi L25 for the Eicher 11.10 chassis frame
of both the ENI and the INI score arrays should optimization.
be considered in the GRA matrix as two distinct
Run Design parameters PCHs
performance measures for the PCH. If the hypoth-
esis test (H0: g = 0 vs H1: g . 0) fails to reject x1 (mm) x2 (mm) x3 (mm) Shear Deflection
the H0 assumption, then estimate of the INI score stress (mm)
array is only maintained in the GRA matrix as (MPa)
performance measure for the PCH. 1 3 3 3 150.450 5.015
2. Once the GRA matrix is compounded, the GRA 2 3 4 4 120.550 4.591
procedure is executed to convert the multiple 3 3 5 5 130.240 3.710
responses into a single response optimization prob- 4 3 6 6 114.800 3.436
5 3 7 7 98.638 3.108
lem and obtain the overall optimal solution (Deng,
6 4 3 4 129.730 4.681
1989). The ENI and/or the INI performance arrays 7 4 4 5 127.190 4.238
are normalized and then optimized depending on 8 4 5 6 118.780 3.637
the optimization target, that is, the ‘‘larger is the 9 4 6 7 109.686 3.438
better’’ (LTB), the ‘‘smaller is the better’’ (STB), or 10 4 7 3 123.990 3.675
11 5 3 5 123.300 3.924
the NTB. As it has been outlined earlier, the target
12 5 4 6 115.770 3.482
value for the INI and ENI scores is 1, corre- 13 5 5 7 110.390 3.164
sponding to the ideal robust design solution 14 5 6 3 122.350 3.437
(ui=vi = 0). The most robust design solution is 15 5 7 4 119.430 3.262
obtained by ranking the grey relational grades 16 6 3 6 112.220 3.431
17 6 4 7 99.647 3.027
(GRGs) of the formed matrix. The design solu-
18 6 5 3 111.440 3.330
tion having the highest grade would correspond 19 6 6 4 104.210 3.089
to the most robust design solution for the multi- 20 6 7 5 102.690 2.910
objective problem. 21 7 3 7 107.200 3.138
22 7 4 3 109.910 3.308
23 7 5 4 103.590 3.050
24 7 6 5 98.796 2.871
Empirical examples 25 7 7 6 70.489 2.742
Case study 1: shear stress and bending stiffness PCH: performance characteristic.
optimization for the Eicher 11.10 chassis frame
This case study has been retrieved from the work of Taguchi’s L25 (56) OA strategy where the engineering
Patel and Bhatt (2013). The study aims to optimize the parameters x1, x2, and x3 are evaluated at five levels
design of the Eicher 11.10 chassis frame in the ST52 each. Table 2 shows the experimental layout.
material. The optimization problem consists in maxi- Using FRONTIER 4.1 at 5% level, the tests on the
mizing the shear stress (in MPa) and minimizing the SF model for the shear stress and the deflection PCHs
deflection (in mm) in the Eicher 11.10 chassis frame. suggest TL (log ratio–statistic (LR-stat) = 18.02) and
For the deflection PCH, 1/y transformation is applied CD (LR-stat = 10.58) functional forms, respectively.
to satisfy the prerequisite of the SF condition of maxi- This is outlined in Table 3 (tests 1 and 4). The tests 2
mum production. The non-critical parameters are the and 5 (H0: g = 0 vs H1: g . 0) on the non-natural
web thickness (x1 in mm), the upper flange thickness sources of variation (ENI) in the design process are
(x2 in mm), and the lower flange thickness (x3 in mm). both statistically significant at 5% levels (LR-stat
The designed experiment is conducted according to a equals 16.94 and 4.9, respectively). Therefore, the ENI

Table 3. Hypothesis tests on the SF models for the shear stress and the deflection PCHs.

PCH Test Hypothesis a = 5% LR-stat x 20:95 value Decision

Shear stress Test 1 Cobb–Douglas vs TransLog 18.02 12.59 Reject


Test 2 g=0 16.94 2.71 Reject
Test 3 m = 0 (g 6¼ 0) 1.80 4.91 Fail to reject
Deflection Test 4 Cobb–Douglas vs TransLog 10.58 12.59 Fail to reject
Test 5 g=0 4.90 2.71 Reject
Test 6 m = 0 (g 6¼ 0) 0.00 4.91 Fail to reject

SF: stochastic frontier; PCH: performance characteristic; LR-stat: log ratio–statistic.


8 Concurrent Engineering: Research and Applications 00(0)

Table 4. Estimates of the SF models’ parameters of the shear stress and the deflection PCHs.

Variables Parameter TL-SF model for the shear CD-SF model for the deflection
stress PCH (g 6¼ 0; m = h = 0) PCH (g 6¼ 0; m = h = 0)
Estimate t-test Estimate t-test

Cte. b0 3.204* 6.716 22.525* 2143.430


ln(x1) b1 2.170* 6.302 0.335* 124.562
ln(x2) b2 2.322* 3.389 0.345* 97.272
ln(x3) b3 21.538* 22.713 0.181* 29.069
ln(x1)2 b4 20.722* 28.532
ln(x1) 3 ln(x2) b5 21.268* 24.362
ln(x1) 3 ln(x3) b6 1.174* 3.920
ln(x2)2 b7 20.529* 23.454
ln(x2) 3 ln(x3) b8 0.787* 4.011
ln(x3)2 b9 20.621* 26.403
s2 ¼ s2v þ s2u 5.84 3 1023 4.59 0.75 3 1022 12.77
g 0.999 0.45 3 103 0.999 0.86 3 106
g* 0.997 0.997
m, h – – – –
Log (likelihood) 46.90 44.33
The t-test critical value (5% level) 2.13 2.08
g
g ¼ ½gþ(1g)p=(p2)  (Coelli, 1995)

PCH: performance characteristic; TL: TransLog; SF: stochastic frontier; CD: Cobb–Douglas.

component is accountable with regard to the total var- Table 5. ENI arrays for the shear stress and the deflection
iation for the shear stress and the deflection PCHs. The PCHs.
tests 3 and 6 in Table 3 on the distributional form of
Run Shear stress Deflection
the ENI components for the shear stress and the deflec-
tion PCHs suggest rather a half-normal distribution ENI ENI
(LR-stat equals 1.80 and 0.00, respectively).
1 0.98788 0.96735
The parameter estimates of the TL and the CD func- 2 0.84403 0.90830
tions for the shear stress and the deflection PCHs are 3 0.99898 0.99956
shown in Table 4. The starred parameters correspond 4 0.97453 0.98075
to the significant parameters using the t-test at 5% 5 0.92965 0.99997
6 0.93278 0.89348
level.
7 0.96547 0.85816
The high g* values of 0.997 for both PCHs indi- 8 0.98006 0.89588
cate that the bulk of variation is due to non-natural 9 0.99535 0.86554
noises (ENI). Moreover, because the g* values are 10 0.87918 0.89492
greater than 95%, the ENI score arrays for both 11 0.96968 0.94979
12 0.96241 0.93795
PCHs are maintained in the GRA matrix as perfor- 13 0.99915 0.92936
mance measures, which is shown in Table 5. The 14 0.99039 0.93650
ENI scores are structurally bounded above by 1; 15 0.99710 0.88827
therefore, the generation of the normalized data uses 16 0.96150 0.98884
the LTB attribute. 17 0.90754 0.98698
18 0.96674 0.96834
The composed GRA table shown in Table 6 sug- 19 0.93846 0.93063
gests that run no. 3 is the least sensitive design solu- 20 0.99835 0.89973
tion since it has the higher GRG of 0.9963. Then, the 21 0.99616 0.99852
robust design solution for the multi-objective prob- 22 0.95674 0.99987
23 0.95607 0.95320
lem consists in setting the web thickness at 3 mm and
24 0.99772 0.91320
the upper flange thickness and the lower flange thick- 25 0.78842 0.87733
ness at 5 mm each. Mean 0.95497 0.93690
Using the 95th percentile of the GRG, the run no. Standard deviation 0.05140 0.04496
21 is also a potential robust design solution, which is
ENI: extrinsic noise insensitivity; PCH: performance characteristic.
worthy of consideration for further functional and ENIi ¼ eui .
Rezgui and Trabelsi 9

Table 6. Grey relational analysis for the shear stress and the deflection PCHs.

Run Grey relational generation of each PCH Grey relational coefficient of each PCH Grey relational grade
Stress Deflection Stress Deflection Grey grade Rank Percentile (%)
ENI ENI ENI ENI

1 0.9465 0.7700 0.9033 0.6849 0.7941 7 76


2 0.2639 0.3536 0.4045 0.4361 0.4203 24 8
3 0.9992 0.9971 0.9984 0.9943 0.9963 1 100
4 0.8831 0.8645 0.8106 0.7868 0.7987 6 80
5 0.6702 1.0000 0.6025 1.0000 0.8013 4 88
6 0.6850 0.2490 0.6135 0.3997 0.5066 22 16
7 0.8402 0.0000 0.7578 0.3333 0.5456 21 20
8 0.9094 0.2660 0.8466 0.4052 0.6259 19 28
9 0.9820 0.0520 0.9652 0.3453 0.6552 17 36
10 0.4307 0.2592 0.4676 0.4030 0.4353 23 12
11 0.8601 0.6462 0.7814 0.5856 0.6835 15 44
12 0.8256 0.5627 0.7415 0.5334 0.6374 18 32
13 1.0000 0.5021 1.0000 0.5010 0.7505 8 72
14 0.9584 0.5525 0.9232 0.5277 0.7254 10 64
15 0.9903 0.2123 0.9809 0.3883 0.6846 14 48
16 0.8213 0.9215 0.7367 0.8643 0.8005 5 84
17 0.5653 0.9084 0.5349 0.8451 0.6900 13 52
18 0.8462 0.7770 0.7647 0.6916 0.7281 9 68
19 0.7120 0.5111 0.6345 0.5056 0.5701 20 24
20 0.9962 0.2932 0.9925 0.4143 0.7034 12 56
21 0.9858 0.9898 0.9724 0.9800 0.9762 2 96
22 0.7988 0.9993 0.7130 0.9987 0.8559 3 92
23 0.7956 0.6702 0.7098 0.6025 0.6562 16 40
24 0.9932 0.3881 0.9866 0.4497 0.7181 11 60
25 0.0000 0.1352 0.3333 0.3663 0.3498 25 4

PCH: performance characteristic; ENI: extrinsic noise insensitivity.


INIi ¼ eui 3yief ðxi bÞ and ENIi ¼ eui .

Case study 2: optimizing weld bead geometry


parameters
This case study is a research work, which is published
by Prasad and Rao (2013). The objective is to optimize
weld bead geometry parameters in a micro-plasma-arc-
welded AISI 304L stainless steel sheet process. The
welding process uses four non-critical parameters: peak
current (x1 in A), back current (x2 in A), pulse rate (x3
in pulses/s), and pulse width (x4 in %). The objective is
to maximize the front-width (FW in mm) and the back-
width (BW in mm) and to minimize the front-height
(FH in mm) and the back-height (BH in mm). A trans-
formation, 1/y, is performed for the FH and BH
responses to agree with the SF condition of maximum
Figure 3. Ranking of the grey relational grade for the shear production. The designed experiment is a central com-
stress and the deflection PCHs. posite design (CCD). Table 7 shows the design matrix
and the observed values of the PCHs at each design
solution.
quality–cost balance as it is shown in Table 6 and Using FRONTIER 4.1, Table 8 suggests that at 5%
Figure 3. level, the suitable functional forms of the SF models for
10 Concurrent Engineering: Research and Applications 00(0)

Table 7. CCD design plan for the welding process optimization of AISI 304L.

Run Design parameters PCHs


x1 (A) x2 (A) x3 (pulses/s) x4 (%) FW (mm) BW (mm) FH (mm) BH (mm)

1 6.5 3.5 30 40 1.448 1.374 0.0609 0.0498


2 7.5 3.5 30 40 1.592 1.522 0.0588 0.0458
3 6.5 4.5 30 40 1.383 1.324 0.0630 0.0490
4 7.5 4.5 30 40 1.504 1.442 0.0569 0.0439
5 6.5 3.5 50 40 1.454 1.401 0.0581 0.0453
6 7.5 3.5 50 40 1.487 1.418 0.0595 0.0466
7 6.5 4.5 50 40 1.469 1.378 0.0599 0.0468
8 7.5 4.5 50 40 1.462 1.402 0.0578 0.0448
9 6.5 3.5 30 60 1.529 1.451 0.0599 0.0470
10 7.5 3.5 30 60 1.591 1.508 0.0571 0.0441
11 6.5 4.5 30 60 1.520 1.447 0.0572 0.0441
12 7.5 4.5 30 60 1.562 1.506 0.0552 0.0423
13 6.5 3.5 50 60 1.442 1.372 0.0605 0.0474
14 7.5 3.5 50 60 1.384 1.306 0.0590 0.0456
15 6.5 4.5 50 60 1.506 1.430 0.0600 0.0470
16 7.5 4.5 50 60 1.420 1.356 0.0584 0.0464
17 6.0 4.0 40 50 1.521 1.451 0.0598 0.0468
18 8.0 4.0 40 50 1.580 1.514 0.0569 0.0439
19 7.0 3.0 40 50 1.452 1.380 0.0575 0.0445
20 7.0 5.0 40 50 1.427 1.358 0.0564 0.0434
21 7.0 4.0 20 50 1.596 1.527 0.0582 0.0453
22 7.0 4.0 60 50 1.466 1.397 0.0564 0.0434
23 7.0 4.0 40 30 1.400 1.337 0.0636 0.0516
24 7.0 4.0 40 70 1.461 1.384 0.0602 0.0472
25 7.0 4.0 40 50 1.531 1.462 0.0606 0.0476
26 7.0 4.0 40 50 1.581 1.512 0.0597 0.0467
27 7.0 4.0 40 50 1.523 1.452 0.0607 0.0477
28 7.0 4.0 40 50 1.519 1.450 0.0606 0.0476
29 7.0 4.0 40 50 1.504 1.432 0.0607 0.0477
30 7.0 4.0 40 50 1.501 1.433 0.0576 0.0446
31 7.0 4.0 40 50 1.401 1.332 0.0597 0.0456

CCD: central composite design; PCH: performance characteristic; FW: front-width; BW: back-width; FH: front-height; BH: back-height.

Table 8. Hypothesis tests on the SF models for the FW, BW, FH, and BH PCHs.

PCH Hypothesis a = 5% LR-stat x 20:95 Decision

FW Test 1 Cobb–Douglas vs TransLog 45.14 18.31 Reject


Test 2 g=0 3.37 2.71 Reject
Test 3 m = 0 (g 6¼ 0) 0.00 4.91 Fail to reject
BW Test 4 Cobb–Douglas vs TransLog 44.00 18.31 Reject
Test 5 g=0 3.12 2.71 Reject
Test 6 m = 0 (g 6¼ 0) 0.00 4.91 Fail to reject
FH Test 7 Cobb–Douglas vs TransLog 36.08 18.31 Reject
Test 8 g=0 0.00 2.71 Fail to reject
BH Test 9 Cobb–Douglas vs TransLog 37.32 18.31 Reject
Test 10 g=0 0.00 2.71 Fail to reject

SF: stochastic frontier; FW: front-width; BW: back-width; FH: front-height; BH: back-height; PCH: performance characteristic; LR-stat: log ratio–statistic.

the four PCHs are a TL (LR-stat equals 45.14, 44.00, respectively); however, it is insignificant for the FH and
36.08, and 37.32, respectively). The test (H0: g = 0 vs the BH PCHs (LR-stat equals 0.00 for both; see tests 2,
H1: g . 0) on the non-natural sources of variation 5, 8, and 10 in Table 8). The tests 3 and 6 in Table 8
(ENI) is statistically significant with regard to the FW indicate that the ENI scores follow a half-normal for
and BW PCHs (LR-stat equals 3.37 and 3.12, the FW and the BW PCHs (LR-stat equals 0.00).
Rezgui and Trabelsi 11

Table 9. Estimates of the TL-SF models’ parameters of the FW, BW, FH, and BH PCHs.

Variables Parameter FW (g 6¼ 0; m = h = 0) BW (g 6¼ 0; m = h = 0) FH (g = 0, OLS) BH (g = 0, OLS)


Est. t-test Est. t-test Est. t-test Est. t-test

Cte. b0 218.055* 218.055 217.729* 217.729 2.039 0.369 26.326 20.880


ln(x1) b1 3.807* 3.807 3.196* 3.196 23.576 21.022 21.972 20.433
ln(x2) b2 0.119 0.119 20.548 20.548 24.143* 22.535 22.279 21.072
ln(x3) b3 2.873* 2.873 3.101* 3.101 1.428 1.774 2.487* 2.374
ln(x4) b4 4.837* 4.837 4.972* 4.972 2.125 1.979 4.028* 2.883
ln(x1)2 b5 1.253 1.253 1.375 1.375 1.257 1.674 1.431 1.464
ln(x1) 3 ln(x2) b6 20.474 20.474 20.100 20.100 0.802 1.395 0.345 0.461
ln(x1) 3 ln(x3) b7 21.109 21.109 21.149 21.149 20.529 21.875 20.785* 22.137
ln(x1) 3 ln(x4) b8 20.978 20.978 21.031 21.031 20.063 20.177 20.229 20.494
ln(x2)2 b9 20.726 20.726 20.765 20.765 0.738* 3.110 0.900* 2.915
ln(x2) 3 ln(x3) b10 0.361 0.361 0.276 0.276 20.104 20.647 20.303 21.451
ln(x2) 3 ln(x4) b11 0.368 0.368 0.462 0.462 0.253 1.248 0.083 0.315
ln(x3)2 b12 20.003 20.003 20.004 20.004 0.098 1.938 0.110 1.675
ln(x3) 3 ln(x4) b13 20.326 20.326 20.335 20.335 20.248* 22.489 20.343* 22.652
ln(x4)2 b14 20.285 20.285 20.303 20.303 20.180 22.109 20.306* 22.748
s2 ¼ s2v þ s2u 5.1 3 1024 5.1 3 1024 5.6 3 1024 5.6 3 1024 4.3 3 1024 – 7.2 3 1024 –
g 0.710 0.710 0.700 0.700 – – – –
g* 0.471 0.459
m, h – – – – – – – –
Log (likelihood) 84.57 82.73 86.49 78.33
The t-test critical
value (5% level) = 2.12
g
g ¼ ½gþ(1g)p=(p2)  (Coelli, 1995)

TL: TransLog; SF: stochastic frontier; FW: front-width; BW: back-width; FH: front-height; BH: back-height; PCH: performance characteristic; OLS:
ordinary least square.

The parameter estimates of the TL functions for the 40 pulses/s, and the pulse width at 30%, in natural
four PCHs are shown in Table 9. The starred para- coding.
meters are found significant using the t-test at 5% level. Allowing for 95th percentile of the design solutions,
At 5% level, the hypothesis test (H0: g = 0 vs H1: g Table 11 and Figure 4 indicate that the run no. 18 is also
. 0) is statistically significant for the FW and the BW a potential robust design solution, which should be eval-
PCHs, and their g* values equal 0.471 and 0.459, uated in terms of functional and quality cost.
respectively, suggest that both the arrays of the ENI
and the INI scores should be maintained in the GRA
matrix as performance measures for the FW and the Case study 3: enhancement of the mechanical
BW PCHs. The test (H0: g = 0 vs H1: g . 0) for the properties of bone powder and coir fiber-reinforced
FH and the BH PCHs is insignificant at 5% level (i.e. polyester composites
the OLS model holds); therefore, only the INI score
arrays are retained in the GRA matrix as performance This case study is relevant to the work of Vignesh et al.
measures for the FH and BH PCHs. Table 10 shows (2016). The objective is to optimize the process para-
the matrix of the GRA analysis, and the generation of meters so that the mechanical properties of bone pow-
the normalized data uses the LTB attribute for the ENI der and coir fiber–reinforced polyester composites
arrays and the NTB for the INI arrays. could be improved. Four engineering parameters are
The GRA table for the FW, BW, FH, and BH PCHs examined: coir fiber diameter (x1 in mm), coir fiber
is shown in Table 11. It is suggested that run no. 23 is length (x2 in mm), bone powder content (x3 in %), and
the least sensitive (most robust) design solution since it bone powder size (x4 in mm). In addition, four PCHs
has the highest GRG (=0.6458). Accordingly, the most are examined, that is, tensile strength (TS in MPa),
robust solution is obtained while setting the peak cur- flexural strength (FS in MPa), compressive strength
rent at 7 A, the back current at 4 A, the pulse rate at (CS in MPa), and impact energy (IE in kJ/m2), which
12 Concurrent Engineering: Research and Applications 00(0)

Table 10. ENI and INI arrays for the FW, BW, FH, and BH equals 20.10, 20.26, 20.68, and 21.60, respectively).
PCHs. The replication effect is insignificant as the tests 3, 7,
11, and 15 indicate.
Run FW BW FH BH
The parameter estimates of the CD functions for the
ENI INI ENI INI INI INI four PCHs—TS, FS, CS, and IE—are shown in Table
14, and the starred parameters are significant using the
1 0.9868 0.9905 0.9834 0.9924 1.0111 1.0039
t-test at 5% level.
2 0.9838 0.9877 0.9843 0.9932 1.0058 1.0149
3 0.9869 0.9908 0.9891 1.0002 0.9848 0.9893 At 5% level, the g* values pertaining to the hypoth-
4 0.9868 0.9911 0.9843 0.9933 1.0167 1.0138 esis test (H0: g = 0 vs H1: g . 0) for the four PCHs—
5 0.9834 0.9860 0.9848 0.9947 1.0024 1.0080 TS, FS, CS, and IE—equal 81.1%, 87.3%, 65.8%, and
6 0.9846 0.9875 0.9825 0.9924 0.9776 0.9659 73.7%, respectively. Since the g* values are all lesser
7 0.9881 0.9914 0.9840 0.9940 0.9924 0.9838
8 0.9845 0.9877 0.9860 0.9962 0.9974 0.9994
than 95%, the arrays of the ENI and INI scores should
9 0.9866 0.9882 0.9861 0.9943 0.9956 0.9989 be included in the GRA matrix as performance mea-
10 0.9844 0.9863 0.9851 0.9930 1.0065 1.0032 sures for the four PCHs. Table 15 displays the matrix of
11 0.9878 0.9899 0.9863 0.9947 1.0227 1.0230 the GRA analysis where the ENI and INI scores are
12 0.9878 0.9903 0.9875 0.9963 0.9930 0.9932 obtained by averaging the predicted values over the two
13 0.9856 0.9860 0.9846 0.9934 0.9816 0.9719
14 0.9834 0.9844 0.9814 0.9905 1.0079 1.0076 replications. The generation of the normalized data uses
15 0.9881 0.9893 0.9885 0.9989 0.9843 0.9794 the LTB attribute for the ENI arrays and the NTB for
16 0.9859 0.9870 0.9859 0.9950 0.9844 0.9780 the INI arrays.
17 0.9851 0.9867 0.9851 0.9941 1.0023 1.0005 The GRA matrix for the TS, FS, CS, and IE PCHs
18 0.9904 0.9952 0.9895 1.0009 0.9935 0.9915 is shown in Table 16. It is indicated that the run no. 4
19 0.9904 0.9948 0.9903 1.0025 0.9951 0.9940
20 0.9822 0.9849 0.9809 0.9903 1.0029 1.0002 has the highest GRG (=0.7008). The robust design
21 0.9865 0.9901 0.9859 0.9939 0.9824 0.9775 solution is then obtained by setting the coir fiber dia-
22 0.9896 0.9923 0.9894 1.0014 1.0363 1.0446 meter at 0.5 mm, the coir fiber length at 30 mm, the
23 0.9883 0.9936 0.9882 0.9996 0.9975 0.9925 bone powder content at 15%, and the bone powder size
24 0.9861 0.9866 0.9849 0.9930 1.0004 1.0040 at 120 mm.
25 0.9920 0.9990 0.9917 1.0065 0.9839 0.9746
26 0.9965 1.0270 0.9962 1.0362 0.9985 0.9930 Accounting for 89% percentile, run no. 6 is also a
27 0.9906 0.9953 0.9898 1.0015 0.9824 0.9726 potential robust design solution, which is worthy
28 0.9897 0.9935 0.9894 1.0006 0.9839 0.9746 consideration as it is underscored in Table 16 and
29 0.9856 0.9878 0.9843 0.9933 0.9824 0.9726 Figure 5.
30 0.9846 0.9868 0.9846 0.9936 1.0343 1.0390
31 0.9393 0.9654 0.9377 0.9698 0.9985 1.0165
Mean 0.9855 0.9901 0.9849 0.9964 0.9980 0.9962
Standard 0.0091 0.0088 0.0093 0.0096 0.0149 0.0193
Conclusion
deviation The RDPP-SF method (Trabelsi and Rezgui, 2020) has
ENI: extrinsic noise insensitivity; INI: intrinsic noise insensitivity; FW: been proposed to address parameter design in develop-
front-width; BW: back-width; FH: front-height; BH: back-height; PCH: ing products and processes for single-objective prob-
performance characteristic. lems. This article has amended the RDPP-SF method
INIi ¼ eui 3yief (xi :b) and ENIi ¼ eui . to commensurate multi-criteria decision-making prob-
lems. This is carried out by revising the step 4 in the
are to be maximized. The experimentation is conducted RDPP-SF method and using the GRA technique. To
using an L9 (34) OA, which is replicated once. The lay- assess the variation, which resides in the developed
out of the designed experiment is shown in Table 12. product/process, the test (H0: g = 0 vs H1: g . 0) is
Using FRONTIER 4.1, the tests 1, 5, 9, and 13 in conducted based on three levels of significance: strong
Table 13 suggest that at 5% level, the CD model fits (g ø 95%), marginal (g \ 95%), and insignificant
for the four PCHs (LR-stat equals 253.78, 267.64, (g = irrelevant). The GRA matrix is composed of the
281.40, and 282.18, respectively). The test (H0: g = 0 ENI and/or the INI score arrays for the different PCHs
vs H1: g . 0) on the non-natural sources of variation as performance measures depending on the outcome of
(ENI) is statistically significant for the four PCHs: TS, the hypothesis (H0: g = 0 vs H1: g . 0) at 5% level. If
FS, CS, and IE. The LR-stat equals 8.30, 11.12, 2.87, the test is insignificant, then the INI score arrays are
and 9.72, respectively, as it is indicated in tests 2, 6, 10, only maintained in the GRA matrix. When the test is
and 14 (Table 13). The tests 4, 8, 12, and 16 in Table 13 strongly significant, then the ENI score arrays are only
suggest that the ENI scores follow rather a half-normal maintained, and in case the test is marginally signifi-
distribution for the TS, FS, CS, and IE PCHs (LR-stat cant, then both the INI and the ENI score arrays are
Table 11. Grey relational analysis for the FW, BW, FH, and BH PCHs.

Run Grey relational generation of each PCHs Grey relational coefficient of each PCHs Grey relational grade
FW BW FH BH FW BW FH BH Grey grade Rank Percentile (%)
ENI INI ENI INI INI INI ENI INI ENI INI INI INI
Rezgui and Trabelsi

1 0.831 0.724 0.781 0.791 0.694 0.914 0.497 0.441 0.432 0.456 0.375 0.677 0.4798 18 45
2 0.778 0.645 0.795 0.812 0.840 0.666 0.428 0.374 0.449 0.483 0.543 0.343 0.4366 21 35
3 0.833 0.733 0.877 0.995 0.580 0.761 0.499 0.450 0.576 1.000 0.302 0.422 0.5418 10 71
4 0.831 0.743 0.795 0.816 0.540 0.691 0.497 0.46 0.449 0.488 0.283 0.36 0.4229 22 32
5 0.771 0.593 0.804 0.852 0.934 0.822 0.421 0.341 0.460 0.545 0.761 0.497 0.5042 15 55
6 0.793 0.638 0.764 0.789 0.383 0.235 0.446 0.369 0.414 0.454 0.226 0.184 0.3489 29 10
7 0.853 0.75 0.791 0.834 0.792 0.637 0.532 0.468 0.444 0.515 0.473 0.323 0.4593 20 39
8 0.791 0.643 0.824 0.894 0.927 0.986 0.444 0.373 0.487 0.628 0.741 0.950 0.6039 5 87
9 0.827 0.66 0.826 0.842 0.88 0.976 0.491 0.385 0.49 0.528 0.619 0.899 0.5686 8 77
10 0.788 0.605 0.809 0.806 0.82 0.928 0.440 0.347 0.466 0.475 0.512 0.717 0.4930 16 52
11 0.848 0.709 0.831 0.855 0.374 0.485 0.523 0.426 0.496 0.549 0.224 0.252 0.4118 24 26
12 0.848 0.720 0.850 0.897 0.807 0.847 0.523 0.437 0.526 0.636 0.492 0.537 0.5252 14 58
13 0.81 0.595 0.800 0.817 0.494 0.37 0.467 0.341 0.455 0.491 0.264 0.215 0.3722 28 13
14 0.771 0.548 0.746 0.736 0.781 0.83 0.421 0.315 0.396 0.398 0.46 0.509 0.4167 23 29
15 0.854 0.691 0.867 0.968 0.569 0.538 0.533 0.410 0.556 0.865 0.297 0.273 0.4890 17 48
16 0.815 0.623 0.823 0.863 0.571 0.506 0.474 0.359 0.485 0.564 0.297 0.260 0.4066 26 19
17 0.800 0.615 0.809 0.837 0.936 0.99 0.454 0.354 0.466 0.519 0.769 0.968 0.5883 6 84
18 0.893 0.862 0.885 0.976 0.821 0.809 0.609 0.642 0.592 0.900 0.513 0.479 0.6225 2 97
19 0.894 0.849 0.898 0.93 0.866 0.865 0.612 0.615 0.621 0.725 0.589 0.569 0.6219 4 90
20 0.750 0.564 0.737 0.733 0.920 0.995 0.400 0.324 0.388 0.395 0.717 1.000 0.5373 11 68
21 0.826 0.713 0.822 0.831 0.515 0.497 0.489 0.43 0.484 0.51 0.272 0.256 0.4070 25 23
22 0.879 0.777 0.883 0.961 0.000 0.000 0.579 0.501 0.588 0.835 0.152 0.147 0.4669 19 42
23 0.857 0.816 0.862 0.989 0.931 0.832 0.538 0.556 0.547 0.967 0.754 0.513 0.6458 1 100
24 0.819 0.612 0.806 0.805 0.989 0.910 0.480 0.352 0.462 0.474 1.000 0.668 0.5727 7 81
25 0.922 0.971 0.922 0.82 0.558 0.431 0.681 1.000 0.682 0.494 0.291 0.233 0.5635 9 74
26 1.000 0.218 1.000 0.000 0.958 0.844 1.000 0.206 1.000 0.147 0.850 0.530 0.6222 3 94
27 0.896 0.863 0.89 0.958 0.514 0.386 0.617 0.642 0.602 0.82 0.272 0.22 0.5288 13 61
28 0.881 0.812 0.882 0.984 0.558 0.431 0.584 0.55 0.586 0.936 0.291 0.233 0.5302 12 65
29 0.81 0.646 0.796 0.814 0.514 0.386 0.467 0.375 0.45 0.485 0.272 0.22 0.3781 27 16
30 0.793 0.617 0.802 0.823 0.056 0.126 0.446 0.355 0.457 0.499 0.160 0.165 0.3469 30 6
31 0.000 0.000 0.000 0.164 0.958 0.629 0.143 0.167 0.143 0.171 0.850 0.319 0.2989 31 3

FW: front-width; BW: back-width; FH: front-height; BH: back-height; PCH: performance characteristic; ENI: extrinsic noise insensitivity; INI: intrinsic noise insensitivity.
13
14 Concurrent Engineering: Research and Applications 00(0)

maintained in the GRA matrix. The modified RDPP-


SF method is actually broad in scope since: (1) it can
virtually tackle all design strategies (Taguchi OA, CCD,
factorial, etc.) without any restriction; (2) it can handle
multi-criteria decision making problems; and (3) it has
brought an alternative to the ineffective Taguchi S/N
metric for robust design.
However, the modified RDPP-SF method still needs
additional research efforts: (1) to solve robustness prob-
lems when the PCHs are of the NTB type since the SF
does not support optimization with nominal attribute
and (2) to determine the optimal levels of the controlla-
ble design parameters in a continuous space, thus allow-
ing for robustness growth.
Figure 4. Ranking of the grey relational grade for the FW,
BW, FH, and BH PCHs.

Table 12. Replicate of the Taguchi L9 OA for the TS, FS, CS, and IE optimization.

Run Design parameters PCHs


x1 (mm) x2 (mm) x3 (%) x4 (mm) Replication 1 Replication 2
TS FS CS IE TS FS CS IE
(MPa) (MPa) (MPa) (kJ/m2) (MPa) (MPa) (MPa) (kJ/m2)

1 0.2 30 10 80 27.35 54.42 55.69 62.4 28.32 54.62 56.18 63.83


2 0.2 50 15 100 28.40 57.46 59.33 64.32 28.93 58.38 60.62 64.94
3 0.2 70 20 120 32.49 57.66 61.72 65.32 31.36 58.26 62.14 63.71
4 0.5 30 15 120 30.39 59.04 61.56 60.91 31.21 58.16 62.17 62.13
5 0.5 50 20 80 30.44 58.73 53.2 68.28 32.22 58.12 56.73 66.63
6 0.5 70 10 100 31.78 56.94 55.34 65.43 31.05 57.63 56.43 65.29
7 0.8 30 20 100 29.24 54.65 55.72 66.23 29.56 55.87 55.84 65.65
8 0.8 50 10 120 28.78 55.28 54.35 64.93 27.64 55.36 54.79 63.04
9 0.8 70 15 80 27.39 54.28 55.29 62.57 27.56 54.99 55.40 62.46

OA: orthogonal array; TS: tensile strength; FS: flexural strength; CS: compressive strength; IE: impact energy; PCH: performance characteristic.

Table 13. Hypothesis tests on the SF models for the TS, FS, CS, and IE PCHs.

Hypothesis a = 5% LR-stat x 20:95 Decision

Tensile strength Test 1 Cobb–Douglas vs TransLog 253.78 11.07 Fail to reject


Test 2 g=0 8.30 2.71 Reject
Test 3 h = 0 (g 6¼ 0) 0.00 4.91 Fail to reject
Test 4 m = 0 (g 6¼ 0) 20.10 4.91 Fail to reject
Flexural strength Test 5 Cobb–Douglas vs TransLog 267.64 11.07 Fail to reject
Test 6 g=0 11.12 2.71 Reject
Test 7 h = 0 (g 6¼ 0) 26.20 4.91 Fail to reject
Test 8 m = 0 (g 6¼ 0) 20.26 4.91 Fail to reject
Compressive strength Test 9 Cobb–Douglas vs TransLog 281.40 11.07 Fail to reject
Test 10 g=0 2.87 2.71 Reject
Test 11 h = 0 (g 6¼ 0) 24.60 4.91 Fail to reject
Test 12 m = 0 (g 6¼ 0) 20.68 4.91 Fail to reject
Impact energy Test 13 Cobb–Douglas vs TransLog 282.18 11.07 Fail to reject
Test 14 g=0 9.72 2.71 Reject
Test 15 h = 0 (g 6¼ 0) 0.00 4.91 Fail to reject
Test 16 m = 0 (g 6¼ 0) 21.60 4.91 Fail to reject

SF: stochastic frontier; TS: tensile strength; FS: flexural strength; CS: compressive strength; IE: impact energy; PCH: performance characteristic; LR-
stat: log ratio–statistic.
Table 14. Estimates of the CD models’ parameters of the TS, FS, CS, and IE PCHs.

Variables Parameter TS (g 6¼ 0; m = h = 0) FS (g 6¼ 0; m = h = 0) CS (g = 0; m = h = 0) IE (g = 0; m = h = 0)
Est. t-test Est. t-test Est. t-test Est. t-test

Cte. b0 2.631* 4.700 3.638* 9.824 3.096* 10.804 4.283* 16.928


Rezgui and Trabelsi

ln(x1) b1 0.009 0.253 20.004 20.205 20.032 22.169 0.013 1.157


ln(x2) b2 0.064 1.585 0.010 0.404 20.019 20.843 0.025 1.238
ln(x3) b3 0.060 1.271 0.049 1.734 0.065 1.830 0.039 1.810
ln(x4) b4 0.091 1.087 0.056 1.101 0.187* 3.734 20.062 21.419
s2 ¼ s2v þ s2u 6 3 1023 1.944 1.5 3 1023 1.881 2 3 1023 1.854 0.1 3 1023 2.044
g 0.922 17.124 0.950 27.986 0.841 6.813 0.885 12.439
g* 0.811 0.873 0.658 0.737
m, H – – – – – – – –
Log (likelihood) 33.98 48.37 41.40 45.56
The t-test critical value (5% level) = 2.18
g
g ¼ ½gþ(1g)p=(p2)  (Coelli, 1995)

CD: Cobb–Douglas; TS: tensile strength; FS: flexural strength; CS: compressive strength; IE: impact energy; PCH: performance characteristic.

Table 15. ENI and INI arrays for the TS, FS, CS, and IE PCHs.

Run TS FS CS IE
ENI INI ENI INI ENI INI ENI INI

1 0.9590 0.9797 0.9642 0.9966 0.9767 0.9914 0.9805 0.9881


2 0.9160 0.9858 0.9858 0.9913 0.9849 0.9881 0.9875 0.9964
3 0.9630 1.0149 0.9604 0.9932 0.9747 0.9915 0.9796 1.0117
4 0.9830 0.9907 0.9931 1.0096 0.9920 1.0032 0.9551 0.9876
5 0.9850 0.9775 0.9933 1.0076 0.9564 0.9613 0.9908 1.0160
6 0.9860 1.0192 0.9927 0.9954 0.9792 0.9863 0.9912 1.0054
7 0.9400 0.9920 0.9380 0.9875 0.9382 0.9898 0.9897 1.0071
8 0.8970 1.0153 0.9566 0.9978 0.9371 0.9867 0.9866 1.0161
9 0.8680 0.9897 0.9445 0.9918 0.9907 1.0041 0.9229 0.9963
Mean 0.9441 0.9961 0.9698 0.9967 0.9700 0.9892 0.9760 1.0027
Standard deviation 0.0423 0.0161 0.0218 0.0074 0.0211 0.0124 0.0228 0.0111

ENI: extrinsic noise insensitivity; INI: intrinsic noise insensitivity; TS: tensile strength; FS: flexural strength; CS: compressive strength; IE: impact energy.
INIi ¼ eui 3yief (xi :b) and ENIi ¼ eui .
15
16 Concurrent Engineering: Research and Applications 00(0)

Grey grade Rank Percentile (%)

100
33
67
11

78
89
44
22
56
Grey relational grade

7
4
9
1
3
2
6
8
5
0.4401
0.5407
0.4178
0.7008
0.5490
0.6940
0.4661
0.4264
0.5337
0.400
1.000
0.406
0.387
0.309
0.756
0.611
0.308
0.972
INI
Grey relational coefficient of each performance characteristics

0.444
0.700
0.423
0.191
0.959
1.000
0.849
0.649
0.111
Figure 5. Ranking of the grey relational grade for the TS, FS,
ENI

TS: tensile strength; FS: flexural strength; CS: compressive strength; IE: impact energy; ENI: extrinsic noise insensitivity; INI: intrinsic noise insensitivity.
IE

CS, and IE PCHs.


0.600
0.479
0.603
1.000
0.185
0.433
0.534
0.443
0.902
INI

Acknowledgement
The authors are grateful to the editor of the CERA journal
0.309
0.490
0.284
1.000
0.162
0.349
0.113
0.111
0.838

and the four anonymous reviewers for their useful suggestions


ENI
CS

and criticism to the article.


0.754
0.366
0.450
0.337
0.412
0.607
0.268
1.000
0.387
INI

Declaration of conflicting interests


The author(s) declared no potential conflicts of interest with
0.192
0.479
0.174
0.963
1.000
0.914
0.111
0.159
0.124

respect to the research, authorship, and/or publication of this


ENI
FS

article.
0.469
0.637
0.612
0.897
0.428
0.492
1.000
0.598
0.824
INI

Funding
The author(s) disclosed receipt of the following financial sup-
0.353
0.174
0.391
0.831
0.937
1.000
0.243
0.142
0.111
ENI

port for the research, authorship, and/or publication of this


TS

article: The financial support which is allocated to the


"Laboratoire de Mécanique, Productique et Energétique
0.260
0.779
0.271
0.231
0.003
0.667
0.559
0.000
0.769

(LR18ES01)" by the Tunisian Ministry of High Education


INI

and Scientific Reserch is gratefully acknowledged.


Run Grey relational generation of each performance characteristics
Table 16. Grey relational analysis for the TS, FS, CS, and IE PCHs.

0.843
0.946
0.830
0.471
0.995
1.000
0.978
0.932
0.000
ENI
IE

References
Aigner DJ, Lovell CAK and Schmidt P (1977) Formulation
0.779
0.692
0.781
0.917
0.000
0.646
0.736
0.657
0.895

and estimation of stochastic frontier production function


INI

models. Journal of Econometrics 6: 21–37.


Antony J, Somasundarum V, Fergusson C, et al. (2004)
0.721
0.870
0.684
1.000
0.352
0.767
0.019
0.000
0.976
ENI

Applications of Taguchi approach to statistical design of


CS

experiments in Czech republican industries. International


Journal of Productivity and Performance Management
0.725
0.303
0.456
0.230
0.394
0.629
0.000
0.824
0.347
INI

53(5): 447–457.
Bagchi TP (1993) Taguchi Methods Explained: Practical Steps
to Robust Design. New Delhi, India: Prentice Hall.
0.473
0.864
0.405
0.995
1.000
0.988
0.000
0.337
0.118
ENI

Bashiri M, Amiri A and Jalili M (2016) Taguchi design opti-


FS

mization using multivariate process capability index. Jour-


nal of Industrial and Systems Engineering 9(1): 57–78.
0.097
0.369
0.337
0.588
0.000
0.146
0.643
0.320
0.540
INI

Battese GE and Coelli TJ (1988) Prediction of firm-level tech-


nical efficiencies with a generalized frontier production func-
0.771
0.407
0.805
0.975
0.992
1.000
0.610
0.246
0.000

tion and panel data. Journal of Econometrics 38: 387–399.


ENI
TS

Battese GE and Coelli TJ (1993) A stochastic frontier produc-


tion function incorporating a model for technical inefficiency
effects. Working Papers in Econometrics and Applied
1
2
3
4
5
6
7
8
9
Rezgui and Trabelsi 17

Statistics No. 69. Armidale, NSW, Australia: Department Ding R, Lin DKJ and Wei D (2004) Dual response surface
of Econometrics, University of New England. optimization: a weighted MSE approach. Quality Engi-
Battese GE and Coelli TJ (1995) A model of technical ineffi- neering 16(3): 377–385.
ciency effects in a stochastic frontier production function Drinkwater S and Harris R (1999) Frontier 4.1: a computer
for panel data. Empirical Economics 20(2): 325–332. program for stochastic frontier production and cost func-
Battese GE and Corra GS (1977) Estimation of a production tion estimation. The Economic Journal 109(456):
frontier model: with application to pastoral zone of east- F453–F458.
ern Australia. Australian Journal of Agricultural Economics Farrell MJ (1957) The measurement of productive efficiency.
21(3): 167–179. Journal of the Royal Statistical Society, Series A: Statistics
Bisgaard S (1992) A comparative analysis of the performance in Society 120(3): 253–290.
of Taguchi’s linear graphs. Center for Quality and Produc- Forsund FR, Lovell CAK and Schmidt P (1980) A survey of
tivity Improvement (CQPI) Reports (Report no. 82). frontier production function and of their relationship to
Madison, WI: University Wisconsin–Madison. efficiency measurement. Journal of Econometrics 13(1):
Box GEP (1988) Signal-to-noise ratios, performance criteria, 5–25.
and transformations. Technometrics 30(1): 1–17. Fowlkes WY and Creveling CM (1995) Engineering Methods
Box GEP and Fung CA (1986) Studies in quality improvement: for Robust Product Design: Using Taguchi Methods in
Minimizing transmitted variation by parameter design. Cen- Technology and Product Development. Reading, MA:
ter for Quality and Productivity Improvement (CQPI) Addison-Wesley.
Reports (Report no. 8). Madison, WI: University Wiscon- Freeny AE and Nair VN (1992) Robust parameter design
sin–Madison. with uncontrolled noise variables. Statistica Sinica 2(2):
Bras B and Mistree F (1994) Concurrent axiomatic and robust 313–334.
design using compromise decision support problems. Con- Gauri SK and Pal S (2017) Optimization of multi-response
current Engineering: Research and Applications 8(1): 17–31. dynamic systems using multiple regression-based weighted
Chen LH (1997) Designing robust products with multiple signal-to-noise ratio. International Journal of Industrial
quality characteristics. Computers & Operations Research Engineering Computations 8(1): 161–178.
24(10): 937–944. Kackar RN (1985) Off-line quality control, parameter design,
Chen M-S, Lin C-C, Tai Y-Y, et al. (2011) A grey relation and the Taguchi method. Journal of Quality Technology
approach to the integrated process of QFD and QE. Con- 17(4): 176–188.
current Engineering: Research and Applications 19(1): Kunert J and Lehmkuhl F (1997) The generalized [beta]-
35–53. method in Taguchi experiments. Technical Report No.
Chen W, Allen JK, Tsui KL, et al. (1996) A procedure for 1997.07. Dortmund: Sonderforschungsbereich 475—Kom-
robust design: minimizing variations caused by noise fac- plexitätsreduktion in Multivariaten Datenstrukturen, Uni-
tors and control factors. Journal of Mechanical Design versität Dortmund.
118(4): 478–485. Leon RV, Shoemaker AC and Kacker RN (1987) Perfor-
Coelli TJ (1995) Estimators and hypothesis tests for a stochas- mance measures independent of adjustment: an explana-
tic frontier function: a Monte Carlo analysis. Journal of tion and extension of Taguchi’s signal-to-noise ratios.
Productivity Analysis 6(3): 247–268. Technometrics 29(3): 253–265.
Coelli TJ (1996) A guide to FRONTIER 4.1: a computer pro- Lesperance ML and Park S-M (2003) GLMs for the analysis
gram for stochastic frontier production and cost function of robust designs with dynamic characteristics. Journal of
estimation. CEPA Working Papers No. 7/96. Armidale, Quality Technology 35(3): 253–263.
NSW, Australia: Department of Econometrics, University Lin DKJ and Tu W (1995) Dual response surface optimiza-
of New England. tion. Journal of Quality Technology 27(1): 34–39.
Coelli TJ, Rao DSP, O’Donnell CJ, et al. (2005) An Introduc- Maghsoodloo S, Ozdemir G, Jordan V, et al. (2004) Strengths
tion to Efficiency and Productivity Analysis. 2nd ed. New and limitations of Taguchi’s contributions to quality, man-
York: Springer. ufacturing, and process engineering. Journal of Manufac-
Czitrom V (1989) An application of Taguchi’s methods recon- turing Systems 23(2): 73–126.
sidered. Communications in Statistics—Theory and Meth- Meeusen W and Van Den Broeck J (1977) Efficiency estima-
ods 18(11): 4105–4119. tion from Cobb–Douglas production functions with com-
Dellino G, Kleijnen JPC and Meloni C (2015) Metamodel- posed error. International Economic Review 18(2): 435–444.
based robust simulation-optimization: an overview. In: Murillo-Zamorano LR (2004) Economic efficiency and fron-
Dellino G and Meloni C (eds) Uncertainty Management in tier techniques. Journal of Economic Surveys 18(1): 33–77.
Simulation-Optimization of Complex Systems (Operations Myers RH, Khuri AI and Vining G (1992) Response surface
Research/Computer Science Interfaces Series 59). New alternatives to the Taguchi robust parameter design
York: Springer, pp. 27–54. approach. The American Statistician 46(2): 131–139.
Deng JL (1989) The introduction of grey system. Journal of Myers RH, Montgomery DC and Anderson-Cook CM (2016)
Grey System 1(1): 1–24. Response Surface Methodology: Process and Product
18 Concurrent Engineering: Research and Applications 00(0)

Optimization Using Designed Experiments (Wiley Series in design. The International Journal of Advanced Manufactur-
Probability and Statistics, 4th ed.). Hoboken, NJ: John ing Technology 41: 631–641.
Wiley & Sons. Shoemaker AC, Tsui K-L and Jeff Wu CF (1991) Economical
Nair VN, Abraham B, MacKay J, et al. (1992) Taguchi’s experimentation methods for robust design. Technometrics
parameter design: a panel discussion. Technometrics 34(2): 33(4): 415–427.
127–161. Steinberg DM (1996) Robust design: experiments for improv-
Ozturk N, Yildiz AR, Kaya N, et al. (2006) Neuro-genetic ing quality. In: Ghosh S and Rao CR (eds) Handbook of
design optimization framework to support the integrated Statistics, vol. 13. Amsterdam: Elsevier Science, pp.
robust design optimization process in CE. Concurrent 199–240.
Engineering: Research and Applications 14(1): 5–16. Taguchi G (1995) Quality engineering (Taguchi methods) for
Pal S and Gauri SK (2010) Multi-response optimization using the development of electronic circuit technology. IEEE
multiple regression-based weighted signal-to-noise ratio Transaction on Reliability 44(2): 225–229.
(MRWSN). Quality Engineering 22(4): 336–350. Taguchi G and Phadke MS (1989) Quality engineering
Pal S and Gauri SK (2017) Optimization of multi-response through design optimization. In: Dehnad K (ed.) Quality
dynamic systems integrating multiple regression and Tagu- Control, Robust Design, and the Taguchi Method. Boston,
chi’s dynamic signal-to-noise ratio concept. International MA: Springer, pp. 77–96.
Journal of Engineering, Science and Technology 9(1): Taguchi G, Chowdhury S and Wu Y (2004) Taguchi’s Quality
16–33. Engineering Handbook. Hoboken, NJ: John Wiley & Sons.
Pal S and Gauri SK (2018) Simultaneous optimization of Trabelsi A and Rezgui MA (2020) Robust design of processes
quantitative and ordinal responses using Taguchi method. and products using the mathematics of the stochastic fron-
International Journal of Research in Industrial Engineering tier methodology (SF). The International Journal of
7(2): 184–205. Advanced Manufacturing Technology 106(7-8): 2829–2841.
Park G-J, Lee T-H, Lee KH, et al. (2006) Robust design: an Tsui KL (1992) An overview of Taguchi method and newly
overview. AIAA Journal 44(1): 181–191. developed statistical methods for robust design. IIE Trans-
Patel TM and Bhatt NM (2013) Shear stress prediction using actions 24(5): 44–57.
FEA-ANN hybrid modeling or Eicher 11.10 chassis frame. Unal R and Dean EB (1991) Taguchi approach to design
IOSR Journal of Mechanical and Civil Engineering 9(2): optimization for quality and cost: an overview. In: Annual
22–32. conference of the International Society of Parametric Ana-
Pignatiello JJ (1988) An overview of the strategy and tactics lysts, New Orleans, LA, 21–24 May, Document ID:
of Taguchi. IIE Transactions 20(3): 247–254. 20040121019.
Prasad KS and Rao Ch-S (2013) Application of grey rela- Vignesh K, Ramasivam G, Natarajan U, et al. (2016) Optimi-
tional analysis for optimizing weld bead geometry para- zation of process parameters to enhance the mechanical
meters of pulsed current micro plasma arc welded AISI properties of bone powder and coir fiber reinforced polye-
304L stainless steel sheets. International Journal of ster composites by Taguchi method. ARPN Journal of
Advanced Design and Manufacturing Technology 6(1): Engineering and Applied Sciences 11(2): 1224–1231.
79–86. Vining GG and Myers RH (1990) Combining Taguchi and
Schmidt P and Lovell CAK (1980) Estimating stochastic pro- response surface philosophies: a dual response approach.
duction and cost frontiers when technical and allocative Journal of Quality Technology 22(1): 38–45.
inefficiency are correlated. Journal of Econometrics 13(1): Welch WJ, Yu T-K, Kang SM, et al. (1990) Computer experi-
83–100. ments for quality control by parameter design. Journal of
Scibilia B, Kobi A, Chassagnon R, et al. (2001) Robust Quality Technology 22(1): 15–22.
design: a simple alternative to Taguchi’s parameter design Yang K and El-Haik B (2003) Design for Six Sigma: A Road-
approach. Quality Engineering 13(4): 541–548. map for Product Development. New York: McGraw-Hill.
Shaibu AB and Cho BR (2009) Another view of dual response
surface modeling and optimization in robust parameter
Rezgui and Trabelsi 19

Appendix 1

Figure 6. Functional scheme for the RDPP-SF method (Trabelsi and Rezgui, 2020).

Author biographies
Mohamed Ali Rezgui, Dr., is assistant professor at the High National School of Engineering
of Tunis (ENSIT, Tunisia). He gained a PhD in Mechanical Engineering from the ‘‘Ecole
Centrale de Lyon’’, France in 1994. He is interested in study and analysis of forming processes
and friction stir welding of metal and polymer materials using experimental design strategies.
The use of the statistical tools and the regression models to better understand and analyze
microscopic and macroscopic phenomena occurring in friction stir welding and forming pro-
cesses are its major research activity.
20 Concurrent Engineering: Research and Applications 00(0)

Ali Trabelsi, Dr., is assistant professor at the High National School of Engineering of Tunis
(ENSIT, Tunisia). In 1995 he gained a PhD in Mechanical Engineering from the ‘‘Ecole
Centrale de Paris’’ in association with the university of Dundee, Scotland (U.K.), His research
activity concerns the tolerance representation and propagation in rigid and non rigid assem-
blies. Also, Dr. Trabelsi has actively worked on the use of the statistical tools to analyze sensi-
tivity and robustness in developing products and processes from the simultaneous engineering
perspective.

You might also like