Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010

National Institute of Technology, Agartala, Tripura

Elimination of multi-response correlation while applying Taguchi

philosophy in optimization of submerged arc weld
1 2 3
Dr. Saurav Datta Prof. Siba Sankar Mahapatra Prof. Asish Bandyopadhyay
1
Assistant Professor, Department of Mechanical Engineering,
National Institute of Technology, Rourkela, Orissa-769008, India
Email: [email protected]
2
Professor, Department of Mechanical Engineering,
National Institute of Technology, Rourkela, Orissa-769008, India
Email: [email protected]
3
Professor, Department of Mechanical Engineering,
Jadavpur University, Kolkata-700075, India
Email: [email protected]

Abstract

There are several bead geometry parameters which indicate quality of submerged arc

weldment. These includes, bead height, penetration depth, bead width, percentage

dilution etc. Achieving an optimal weld, with desired quality features, is really a

challenging job. Because, these quality features are highly correlated and are expected to

be influenced directly or indirectly by the direct effect of process parameters or their

interactive effects (i.e. on process environment). However, the extents of significant

influence of the process parameters are different for different responses. Therefore, SAW

is a multi-factor, multi-objective metal fabrication process. It is desired that, an optimal

weld should confirm lesser bead height and width, to reduce excess weld metal

consumption; deeper penetration and higher parentage of dilution, in order to increase

joint strength. Therefore, to solve such a multi-objective optimization problem, it is felt

necessary to identify the optimal parametric combination, following which all objectives

could be optimized simultaneously. In this context, it is essential to convert all the

1
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

objective functions into an equivalent single objective function or overall representative

function to meet desired multi-quality features of the weldment. The required multi-

quality features may or may not be conflicting in nature. The representative single

objective function, thus calculated, would be optimized finally. In the present work,

Design of Experiment (DOE) with Taguchi L16 Orthogonal Array (OA) has been

explored to produce 16 weld specimens on mild steel plates by SAW. Collected data

related to weld bead geometry have been utilized for optimization. Principal Component

Analysis (PCA) has been applied to eliminate correlation among the responses and to

evaluate independent or uncorrelated quality indices called principal components. Based

on quality loss of individual principal components with respect to the ideal condition, an

overall grey relational grade of the weldment has been calculated to serve as the single

objective function for optimization. Finally, Taguchi method has been adopted for

searching optimal process condition to yield desired quality of weld bead geometry.

Result of the aforesaid optimization procedure has been verified through confirmatory

test. The study illustrates the detailed methodology of PCA based grey-Taguchi method

and its effectiveness for multi-response optimization in SAW.

Key words: SAW, PCA, Taguchi method, overall grey relational grade

1. Introduction

Submerged arc welding (SAW) is a useful metal joining process in fabrication industry.

Several process parameters influence directly or indirectly on various aspects of

submerged arc weldment. Work, to a far extent, has already been done to study the

effects of the parameters like voltage, current, electrode stick-out, wire feed rate and

2
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

traverse speed on geometry and quality of the weld bead produced by submerged arc

welding on mild steel. But the search is still being continued. Control of the above

parameters, in a more precise manner, can essentially improve the quality of the

weldment, enhance the possibility of increased deposition rate and economize the process

of submerged arc welding.

In many of the cases, quality of the weld is left dependent on operators past experience

and working skill. But, with the advent of automation, it is now possible to design a

machine capable of selecting optimal process parameters to provide desired quality weld.

However, this requires reliable data of knowledge.

Literature depicts that work has been explored on various aspects of modeling, simulation

and process optimization in submerged arc welding [1, 2, 3, 4, 5] on mild steel and many

other materials. The common approaches to tackle optimization problem in welding

include multiple regression analysis, Response Surface Methodology (RSM), Artificial

Neural Network (ANN) modeling and Taguchi method, [6, 7, 8, 9]. In most of the cases

the optimization was carried out using single objective function. For a multi-response

process, while applying the optimal setting of control factors, it can be observed that, an

increase and (or) improvement of one response may cause change in another response

value, beyond the acceptable tolerance limit. Thus for solving a multi-criteria

optimization problem, it is convenient to convert all the objectives into an equivalent

single objective function. This equivalent objective function, which is the representative

of all the quality characteristics of the product, is to be optimized finally.

Taguchi’s philosophy, developed by Dr. Genichi Taguchi, a Japanese quality

management consultant, is an efficient tool available for the design of high quality

3
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

manufacturing system. However, traditional Taguchi method cannot solve a multi-

objective optimization problem [10]. Therefore, Taguchi method coupled with grey

relational analysis has been firmly recommended in literature [11]. In this method, a

multiple response process optimization problem is converted to a single response

optimization problem where overall grey relational grade serves as the single objective

function or response function to be optimized (maximized). Tarng, Y. S. et al. [12]

applied grey-based Taguchi methods for optimization of Submerged Arc Welding

process parameters in hardfacing. They considered multiple weld qualities and

determined optimal process parameters based on grey relational grade from grey

relational analysis proposed by Taguchi method.

The disadvantage of grey based Taguchi approach is the unrealistic assumption of non-

existence of correlation among the responses and they are treated as uncorrelated or

independent. To overcome these shortcomings, the present study explores the use of

Principal Component Analysis (PCA) to convert correlated responses into uncorrelated

quality indices called principal components. PCA is an efficient statistical technique

while studying multi-quality characteristics; those are highly correlated [13, 14]. The

PCA allows data which contain information of multi-quality characteristics to be

converted into several independent quality indicators. Part of these indicators is then

selected to construct a composite quality indicator, which is the representative of multi-

quality features of the process output [13].

In the present work, PCA has been used to eliminate correlation among the responses and

to evaluate independent quality indicators called principal components. These have been

accumulated to calculate the overall grey relational grade which is for replacement of

4
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

four correlated responses viz. bead height, depth of penetration, bead width and

percentage dilution of submerged arc weld. Finally, Taguchi method has been applied to

search an optimal process condition by optimizing the overall grey relational grade.

2. Principal Component Analysis (PCA)

Principal Component Analysis (PCA) is a way of identifying patterns in data, and

expressing the data in such a way so as to highlight their similarities and differences. The

main advantage of PCA is that once the patterns in data have been identified, the data can

be compressed, i.e. by reducing the number of dimensions, without much loss of

information. The methods involved in PCA are discussed below, [15, 16]:

1. Getting some data

2. Normalization of data

3. Calculation of covariance matrix.

4. Interpretation of covariance matrix.

The normalized data have then been utilized to construct a variance-covariance

matrix M , which is illustrated as below:

 N 1,1 N 1, 2 . . . N 1, u 
N N 2, 2 . . . N 2, p 
 2,1
M = . . . . . .  (1)
 
 . . . . . . 
 N q ,1 N q, 2 . . . N q , p 


Cov(Yi *,k , Yi *,l )

Here N k , l = (2)
Var (Yi*,k )Var (Yi *,l )

5
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

In which u stands for the number of quality characteristics and p stands for the number

of experimental runs. Then, eigenvectors and Eigenvalues of matrix M can be computed,

which are denoted by V j and λ j respectively.

In PCA the eigenvector V j represents the weighting factor of j number of quality

characteristics of the jth principal component. For example, if Q j represents the jth

quality characteristic, the jth principal component ψ j can be treated as a quality

indicator with the required quality characteristic.

ψ j = V1 j Q1 + V2 j Q2 + ........................ + V jj Q j = V j' Q (3)

It is to be noted that every principal component ψ j represents a certain degree of

explanation of the variation of quality characteristics, namely the accountability

proportion (AP). When several principal components are accumulated, it increases the

accountability proportion of quality characteristics. This is denoted as cumulative

accountability proportion (CAP). In the present work, the composite principal component

ψ has been defined as the sum/ linear combination of principal components with their

individual Eigenvalues. Thus, the composite principal component represents the overall

quality indicator as shown below:

k
ψ = ∑ψ j (4)
j =1

If a quality characteristic Q j strongly dominates in the jth principal component, this

principal component becomes the major indicator of such a quality characteristic. It

should be noted that one quality indicator may often represent all the multi-quality

characteristics. Selection of individual principal components (ψ j ), those to be included

6
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

in the composite quality indicatorψ , depends on their individual accountability

proportion.

3. Experimentation and data collection

Bead-on-plate SAW welding on mild steel plates (thickness 10 mm) has been carried out

as per Taguchi’s L16 Orthogonal Array design with 16 combinations of voltage (OCV),

wire feed rate, traverse speed and stick-out to be varied in four discrete levels within the

scope for factorial adjustments in the setup used. Table 1 represents domain of

experiments. Design of experiment is furnished in Table 2. Domain of experiment has

been selected from the knowledge of the work carried out by Gunaraj and Murugan [2,

3]. Copper coated electrode wire of diameter 3.16 mm (AWS A/S 5.17:EH14) has been

used during the experiments. Welding has been performed with flux (AWS A5.17/SFA

5.17) with grain size 0.2 to 1.6 mm with basicity index 1.6 (Al2O3+MnO2 35%,

CaO+MgO 25% and SiO2+TiO2 20% and CaF2 15%). The experiments have been

performed on Submerged Arc Welding Machine- INDARC AUTOWELD MAJOR

Maker: IOL Ltd., India. Bead geometry (macrostructure) has been observed in Optical

Trinocular Metallurgical Microscope (Make: Leica, GERMANY, Model No. DMLM,

S6D & DFC320 and Q win Software). It consists of an image analyzer. The macrograph

is obtained on the computer interface which has the provision of selecting two points or

some region in order to calculate relative distance between said points as well as area of

the region under consideration respectively. Like this way bead height, penetration depth

and bead width have been calculated. %dilution is the ratio of area of penetration to the

area of reinforcement measured on a bead cross section. These areas have been obtained

7
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

from the said microscope and their ratio has been computed accordingly. The

experimental data of different quality indicators relating bead geometry have been listed

in Table 2. Aforesaid data bank has been utilized in PCA based hybrid Taguchi method

for identification of optimal factors setting.

Table 1: Domain of experiment (limits of factors)

Parameters Unit Notation Level 1 Level 2 Level 3 Level 4

OCV Volt V 32 34 36 38
Wire feed rate m/min Wf 0.93 1.16 1.39 1.62
Traverse speed m/min Tr 0.51 0.59 0.67 0.75
Stick-out mm N 30 32 34 36

Table 2: Design of experiment and collected response data

L16 OA Response values related to bead geometry

Bead height Penetration depth Bead width Dilution
Sl. No. V Wf Tr N
(mm) (mm) (mm) (%)
1 1 1 1 1 1.1730 2.1990 10.6520 46.0090
2 1 2 2 2 1.9370 2.4970 9.4280 46.8320
3 1 3 3 3 2.7510 2.6570 9.3560 45.6350
4 1 4 4 4 3.6150 2.6790 10.4360 42.4180
5 2 1 2 3 1.6220 2.3110 10.2090 43.5030
6 2 2 1 4 2.5120 1.9800 11.4400 45.1840
7 2 3 4 1 1.2080 2.5830 11.2030 49.5450
8 2 4 3 2 2.0980 3.4800 10.1300 47.1860
9 3 1 3 4 1.8170 1.4490 10.0140 44.7810
10 3 2 4 3 1.9420 2.5560 8.9400 48.8120
11 3 3 1 2 1.8930 3.6330 12.7580 52.5390
12 3 4 2 1 1.3420 3.9880 12.2360 52.5300
13 4 1 4 2 2.3210 1.9190 7.9230 53.0510
14 4 2 3 1 1.6580 2.7350 10.4280 57.9400
15 4 3 2 4 1.8850 1.9090 11.2330 54.3930
16 4 4 1 3 1.7980 3.4770 13.1860 55.2420

8
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

2. Methodology for optimization

Assuming, the number of experimental runs in Taguchi’s OA design is m , and the

number of quality characteristics is n . The experimental results can be expressed by the

following series: X 1 , X 2 , X 3 ,.........., X i ,...., X m

Here,

X 1 = { X 1 (1), X 1 (2)......... X 1 (k )..... X 1 (n)}

.
.
.
.
X i = { X i (1), X i (2)......... X i (k )..... X i (n)}
.
.
.
.
X m = { X m (1), X m (2)......... X m (k )..... X m (n)}

Here, X i represents the i th experimental results and is called the comparative sequence in

grey relational analysis.

Let, X 0 be the reference sequence:

Let, X 0 = { X 0 (1), X 0 (2)......... X 0 (k )..... X 0 (n)}

The value of the elements in the reference sequence means the optimal value of the

corresponding quality characteristic. X 0 and X i both includes n elements, and X 0 (k ) and

X i (k ) represent the numeric value of k th element in the reference sequence and the

comparative sequence, respectively, k = 1, 2,........, n . The following illustrates the

proposed parameter optimization procedures in detail [18].

9
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

Step 1: Normalization of the responses (quality characteristics)

When the range of the series is too large or the optimal value of a quality characteristic is

too enormous, it will cause the influence of some factors to be ignored. The original

experimental data must be normalized to eliminate such effect. There are three different

types of data normalization according to whether we require the LB (lower-the-better),

the HB (higher-the-better) and NB (nominal-the-best). The normalization is taken by the

following equations.

(a) LB (lower-the-better)

min X i (k )
X i* (k ) = (5)
X i (k )

(b) HB (higher-the-better)

X i (k )
X i* (k ) = (6)
max X i (k )

(c) NB (nominal-the-best)

min{ X i (k ), X 0b (k )}
X i* (k ) = (7)
max{ X i (k ), X 0b (k )}

i = 1, 2,........, m;
Here,
k = 1, 2,........., n

X i* (k ) is the normalized data of the k th element in the i th sequence.

X 0b (k ) is the desired value of the k th quality characteristic. After data normalization, the

value of X i* (k ) will be between 0 and 1. The series X i* , i = 1, 2,3,........, m. can be viewed

as the comparative sequence used in the grey relational analysis.

10
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

Step 2: Checking for correlation between two quality characteristics

Qi = { X 0* (i ), X 1* (i ), X 2* (i ),............, X m* (i )}
Let, (8)
where, i = 1, 2,......., n.

It is the normalized series of the i th quality characteristic. The correlation coefficient

between two quality characteristics is calculated by the following equation:

Cov(Q j , Qk )
ρ jk = , (9)
σ Q ×σ Q
j k

j = 1, 2, 3......, n.
here, k = 1, 2,3,........, n.,
j≠k

Here, ρ jk is the correlation coefficient between quality characteristic j and quality

characteristic k ; Cov(Q j , Qk ) is the covariance of quality characteristic j and quality

characteristic k ; σ Q j and σ Qk are the standard deviation of quality characteristic j and

quality characteristic k , respectively.

The correlation is checked by testing the following hypothesis:

 H 0 : ρ jk = 0 (There is no correlation)
 (10)
 H1 : ρ jk ≠ 0 (Thereis correlation)

Step 3: Calculation of the principal component score

(a) Calculate the Eigenvalue λk and the corresponding eigenvector

β k (k = 1, 2,......, n) from the correlation matrix formed by all quality

characteristics.

11
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

(b) Calculate the principal component scores of the normalized reference sequence

and comparative sequences using the equation shown below:

n
Yi (k ) = ∑ X i* ( j ) β kj , i = 0,1, 2,......., m; k = 1, 2,........, n. (11)
j =1

Here, Yi (k ) is the principal component score of the k th element in the i th series.

X i* ( j ) is the normalized value of the j th element in the i th sequence, and β kj is

the j th element of eigenvector β k .

Step 4: Calculation of the individual grey relational grades

Calculation of the individual grey relational coefficients

Use the following equation to calculate the grey relational coefficient between
X 0 (k ) and X i (k ) .

∆ min + ξ∆ max
r0,i (k ) = , i = 1, 2,.........., m; k = 1, 2,....., n. (12)
∆ 0,i (k ) + ξ∆ max

Here, r0,i (k ) is the relative difference of k th element between sequence X i and the

comparative sequence X 0 (also called grey relational grade), and ∆ 0,i (k ) is the

absolute value of difference between X 0 (k ) and X i (k ) .

 X 0* (k ) − X i* (k ) , no significant correlation between quality characteristics

∆ 0,i (k ) = 
 Y0 (k ) − Yi (k ) , Significant correlation between quality characteristics

(13)
 max max X 0* (k ) − X i* (k ) , no significant correlationbetween quality characteristics

∆ max = i k

max
 i max Y0 (k ) − Yi (k ) , Significant correlation between quality characteristics
k

(14)

12
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

 min min X 0* (k ) − X i* (k ) , no significant correlation between quality characteristics


∆ min = i k

 i k Y0 (k ) − Yi (k ) ,
min min Significant correlation between quality characteristics

(15)
Note that ξ is called the distinguishing coefficient, and its value is in between 0 to 1. In
general it is set to 0.5, [17].

(5) Calculation of the overall grey relational grade

After the calculation of the grey relational coefficient and the weight of each quality

characteristic, the grey relational grade is determined by:

n
Γ0,i = ∑ wk r0,i (k ), i = 1, 2,..................., m. (16)
k =1

In this paper, the multiple quality characteristics are combined to one grey relational

grade, thus the traditional Taguchi method can be used to evaluate the optimal parameter

combination. Finally the anticipated optimal process parameters are verified by carrying

out the confirmatory experiments.

(6) Optimization using Taguchi method

The overall grey relational grade is then optimized (maximized) using Taguchi method

[11]. Taguchi’s HB (Higher-the-Better) criterion has been explored to maximize the

overall grey relational grade (Equation 17).

1 t 1 
SN ( Higher − the − better ) = −10 log  ∑ i =1 2  (17)
t yi 

Here t is the number of measurements, and y i the measured i th characteristic value i.e.

i th quality indicator.

13
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

4. Evaluation of optimal setting

Experimental data (Table 2) have been normalized first. Normalized data have been

furnished in Table 3. For reinforcement, bead width Lower-the-Better (LB) and for

penetration, percentage dilution Higher-the-Better (HB) criterion has been selected

(Equation 5 and 6 respectively). After normalization the data have been checked for

correlation. Table 4 shows existence of correlation among the responses (coefficient of

correlation evaluated using Equation (8) and (9) became non-zero value). Principal

component analysis has been applied to eliminate correlation among the responses. Table

5 represents results of analysis of correlation matrix. Correlated responses have been

converted to four independent quality indices denoted as principal components:ψ 1 , ψ 2 ,

ψ 3 and ψ 4 respectively by using equation 11. Principal components in all L16 OA

experimental observations have been shown in Table 6. Quality loss estimates ∆ 0i (k ) for

all principal components have been computed using equations 13, 14 and 15. The values

have been furnished in Table 7. Individual grey relational coefficients have been

calculated using equation 12.

Table 3: Normalized data

Normalized Response data related to bead geometry

Sl. No. Bead height Penetration depth Bead width Dilution
Ideal condition 1.0000 1.0000 1.0000 1.0000
1 1.0000 0.5514 0.7438 0.7941
2 0.6056 0.6261 0.8404 0.8083
3 0.4264 0.6662 0.8468 0.7876
4 0.3245 0.6718 0.7592 0.7321
5 0.7232 0.5795 0.7761 0.7508
6 0.4670 0.4965 0.6926 0.7798
7 0.9710 0.6477 0.7072 0.8551
8 0.5591 0.8726 0.7821 0.8144
9 0.6456 0.3633 0.7912 0.7729
10 0.6040 0.6409 0.8862 0.8425

14
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

11 0.6197 0.9110 0.6210 0.9068

12 0.8741 1.0000 0.6475 0.9066
13 0.5054 0.4812 1.0000 0.9156
14 0.7075 0.6858 0.7598 1.0000
15 0.6223 0.4787 0.7053 0.9388
16 0.6524 0.8719 0.6009 0.9534

Table 4: Test for correlation among the responses

Sl.
Correlation between responses Coefficient of correlation Comment
No.
1 Penetration and reinforcement +0.293 Both are correlated
2 Penetration and bead width -0.171 Both are correlated
3 Penetration and dilution +0.462 Both are correlated
4 Reinforcement and bead width -0.038 Both are correlated
5 Reinforcement and dilution +0.384 Both are correlated
6 Bead width and dilution +0.035 Both are correlated

Table 5: (Analysis of correlation matrix)

Eigenvalues, eigenvectors, accountability proportion (AP) and cumulative
accountability proportion (CAP) computed for the four major quality indicators

ψ1 ψ2 ψ3 ψ4
Eigenvalue 0.051316 0.026943 0.012819 0.004896
Eigenvector −0.720 0.680 −0.065 0.120
−0.653 −0.716 0.132 0.207
0.079 0.153 0.975 0.146
−0.219 −0.047 0.169 −0.960
AP 0.536 0.281 0.134 0.051
CAP 0.536 0.815 0.949 1.000

Table 6: Principal components in all L16 OA experimental observations

Principal componentsψ 1 (1st PC) to ψ 4 (4th PC)

Sl. No.
(1st PC) ψ 1 (2nd PC) ψ 2 (3rd PC) ψ 3 (4th PC) ψ 4
Ideal condition -1.5130 0.0680 1.2110 -0.4870
1 -1.1952 0.3602 0.8672 -0.4196
2 -0.9555 0.0524 0.9993 -0.4510
3 -0.8476 -0.0962 1.0190 -0.4434
4 -0.7727 -0.1801 0.9315 -0.4140
5 -1.0022 0.1588 0.9131 -0.4007

15
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

6 -0.7765 0.0300 0.8423 -0.4887

7 -1.2535 0.2631 0.8564 -0.4671
8 -1.0889 -0.1648 0.9790 -0.4199
9 -0.8088 0.2620 0.9080 -0.4738
10 -0.9679 0.0461 1.0518 -0.4743
11 -1.1906 -0.1797 0.8387 -0.5163
12 -1.4297 -0.0664 0.8597 -0.4639
13 -0.7996 0.1071 1.1604 -0.5727
14 -1.1162 0.0578 0.9543 -0.6222
15 -0.9105 0.1428 0.8691 -0.6245
16 -1.2004 -0.1347 0.8197 -0.5688

Overall grey relational grade has been calculated using equation 16. While calculating

overall grey relational grade it has been assumed that all responses are equally important.

Therefore 25% weightage has been assigned to prioritize four responses. Table 9 shows

overall grey relational grade for all experimental runs and corresponding S/N ratio

calculated using Taguchi’s Higher-the-Better (HB) criterion (Equation 17). Optimal

parameter setting has been determined from Figure 1. The predicted optimal setting

becomes V3 Wf2 Tr4 N2.

Table 10 represents mean values of overall grey relational grade. It indicates the order of

factors (ranking) representing the extent of significance on the overall grey relational

grade. After evaluating the optimal parameter settings, the next step is to predict and

verify the enhancement of quality characteristics using the optimal parametric

combination. Table 11 reflects the satisfactory result of confirmatory experiment.

Table 7: Calculation of ∆ 0i (k ) for all principal components

Sl. No. ∆ 0i (1st PC ) ∆ 0i (2nd PC ) ∆ 0i (3rd PC ) ∆ 0i (4th PC )

1 0.3178 0.2922 0.3438 0.0674
2 0.5575 0.0156 0.2117 0.0360
3 0.6654 0.1642 0.1920 0.0436
4 0.7403 0.2481 0.2795 0.0730
5 0.5108 0.0908 0.2979 0.0863

16
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

6 0.7365 0.0380 0.3687 0.0017

7 0.2595 0.1951 0.3546 0.0199
8 0.4241 0.2328 0.2320 0.0671
9 0.7042 0.1940 0.3030 0.0132
10 0.5451 0.0219 0.1592 0.0127
11 0.3224 0.2477 0.3723 0.0299
12 0.0833 0.1344 0.3513 0.0231
13 0.7134 0.0391 0.0506 0.0857
14 0.3968 0.0102 0.2567 0.1352
15 0.6025 0.0748 0.3419 0.1375
16 0.3126 0.2027 0.3913 0.0818

5. Conclusion

Grey based Taguchi method is generally adopted for solving multi-attribute decision

making problems (multi-response optimization). The method is based on the assumption

that all response features must be uncorrelated or independent. However, this assumption

may not be valid in practical situation. For example, in the present case, it has been

observed that quality features thus selected viz. bead height, penetration depth and bead

width are highly correlated with %dilution. So, traditional grey-Taguchi technique fails to

overcome this problem. In order to solve this shortcoming the present work highlights

application of Principal Component Analysis (PCA) in combination with Taguchi

technique to solve this correlated multi-response optimization problem. The correlated

quality indices have been transformed into uncorrelated independent quality indices

called principal components. These independent multi-indices have been accumulated to

calculate the composite principal component which has been optimized finally by

Taguchi technique. From the foregoing study the following conclusions can also be

made.

1. PCA can provide a representative quality indicator which can replace correlated

quality characteristics of the process output.

17
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

2. Grey based Taguchi method coupled has been found appropriate to tackle a multi-

objective optimization problem with PCA to eliminate response correlation and to

convert a multi-objective optimization problem to a single response optimization by

accumulating the principal components into the overall grey relational grade.

3. It can be recommended that the PCA based hybrid Taguchi method is good, for

example, in case of processes (chemical and pharmaceutical) industries when there are

hundreds of response variables.

Table 8: Calculation of individual grey relational coefficients

Individual grey relational coefficients

Sl. No.
(1st PC ) (2nd PC ) (3rd PC ) (4th PC )
1 0.6591 0.3566 0.4565 0.5174
2 0.4888 0.9668 0.6045 0.6725
3 0.4379 0.5037 0.6352 0.6270
4 0.4083 0.3965 0.5183 0.4969
5 0.5147 0.6599 0.4989 0.4544
6 0.4098 0.8490 0.4363 1.0000
7 0.7201 0.4581 0.4475 0.7943
8 0.5709 0.4125 0.5759 0.5187
9 0.4221 0.4595 0.4939 0.8596
10 0.4954 0.9301 0.6939 0.8646
11 0.6548 0.3969 0.4336 0.7140
12 1.0000 0.5571 0.4502 0.7671
13 0.4185 0.8440 1.0000 0.4561
14 0.5912 1.0000 0.5444 0.3454
15 0.4662 0.7076 0.4581 0.3416
16 0.6642 0.4481 0.4195 0.4681

Table 9: Calculation of overall grey relational grade and corresponding S/N ratio

Sl. No. Γ0,i S/N Ratio

1 0.4974 -6.06588
2 0.6832 -3.30904
3 0.5509 -5.17854
4 0.4550 -6.83977
5 0.5320 -5.48177
6 0.6739 -3.42809
7 0.6050 -4.36489

18
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

8 0.5195 -5.68829
9 0.5588 -5.05487
10 0.7460 -2.54522
11 0.5498 -5.19591
12 0.6936 -3.17782
13 0.6796 -3.35493
14 0.6203 -4.14796
15 0.4934 -6.13602
16 0.5000 -6.02060

Figure 1: S/N ratio plot for overall grey relational grade

Table 10: Mean value table of overall grey relational grade

Level V Wf Tr N
1 0.546625 0.566950 0.555275 0.604075
2 0.582600 0.680850 0.600550 0.608025
3 0.637050 0.549775 0.562375 0.582225
4 0.573325 0.542025 0.621400 0.545275
Delta 0.090425 0.138825 0.066125 0.062750
Rank 2 1 3 4

19
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

Table 11: Results of confirmatory experiment

Optimal setting
Prediction Experiment
Level of factors V3 Wf2 Tr4 N2 V3 Wf2 Tr4 N2
S/N ratio of Overall grey relational grade -1.76623 -1.75600
Overall grey relational grade 0.91495 0.91594

5. References

1. Gupta, S. R. and Arora, N., 1993, “Influence of Flux Basicity on Weld Bead

Geometry and HAZ in Submerged Arc Welding”, Journal of Material Processing

Technology, Volume 39, Issues 1-2, pp. 33-42.

2. Gunaraj, V. and Murugan, N., 1999, “Application of Response Surface

Methodology for Predicting Weld Bead Quality in Submerged Arc Welding of

Pipes”; Journal of Materials Processing Technology, Volume 88, pp.266-275.

3. Gunaraj, V. and Murugan, N., 1999, “Prediction and Comparison of the Area of

the Heat Affected Zone for the Bead-no-Plate and Bead-on-Joint in Submerged

Arc Welding of Pipes”, Journal of Materials Processing Technology, Volume 95,

Issues 1-3, pp. 246-261.

4. Gunaraj, V. and Murugan, N., October 2000, “Prediction and Optimization of

Weld Bead Volume for the Submerged Arc Process- Part-1”, Welding Research

Supplement, pp. 286s-294s.

5. Gunaraj, V. and Murugan, N., November 2000, “Prediction and Optimization of

Weld Bead Volume for the Submerged Arc Process- Part-2”, Welding Research

Supplement, pp. 331s-338s.

20
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

6. Unal, R., and Dean, Edwin B., 1991, “Taguchi Approach to Design Optimization

for Quality and Cost: An Overview”, Presented at the Annual Conference of the

International Society of Parametric Analysis”.

7. Rowlands, H., Antony, J. and Knowles, G., 2000, “An Application of

Experimental Design For Process Optimization”, The TQM Magazine, Volume

12, Number 2, pp. 78-83.

8. Antony, J. and Antony, F., 2001, “Teaching the Taguchi Method to Industrial

Engineers”, Work Study, Volume 50, Number 4, pp. 141-149.

9. Maghsoodloo, S., Ozdemir, G., Jordan, V. and Huang, C-H., 2004, “Strengths and

Limitations of Taguchi’s Contributions to Quality, Manufacturing, and Process

Engineering”, Journal of Manufacturing Systems, Volume 23, Number 2, pp. 73-

126.

10. Jeyapaul, R., Shahabudeen, P. and Krishnaiah, K., 2005, “Quality Management

Research by Considering Multi-Response Problems in the Taguchi Method-A

Review”, International Journal of Advanced Manufacturing Technology, Volume

26, pp. 1331-1337.

11. Datta S, Bandyopadhyay A, Pal PK (2008) Grey based Taguchi method for

optimization of bead geometry in submerged arc bead-on-plate welding.

International Journal of Advanced Manufacturing Technology 39:1136-1143.

12. Tarng, Y. S., Juang, S. C., Chang, C. H., 2002, “The Use of Grey-Based Taguchi

Methods to Determine Submerged Arc Welding Process Parameters in

Hardfacing”, Journal of Materials Processing Technology, Volume 128, pp. 1-6.

21
 Joining Processes: Challenges for Quality, Design and Development, March 5-6, 2010
National Institute of Technology, Agartala, Tripura

13. Antony J (2000) Multi-response optimization in industrial experiments using

Taguchi’s quality loss function and Principal Component Analysis. Quality and

Reliability Engineering International 16:3-8.

14. Su CT, Tong LI (1997) Multi-response robust design by principal component

analysis. Total Quality Management 8(6):409-416.

15. Johnson, Richard, A. and Wichern, Dean, W., “Applied Multivariate Statistical

Analysis”, Prentice-Hall, Inc., Englewood Cliffs, New Jersey 07632.

16. MINITAB (Release 13.1) State College, PA, USA. User manual.

17. Deng J (1989) Introduction to grey system. Journal of grey system 1(1):1-24.

18. Yu-Min Chiang and Hsin-Hsien Hsieh, (Article in press), The use of the Taguchi

method with grey relational analysis to optimize the thin-film sputtering process

with multiple quality characteristic in color filter manufacturing, Computers &

Industrial Engineering, DOI:10.1016/j.cie.2007.12.020.

22