Number 2 Volume 16 June 2010 Journal of Engineering

TAGUCHI EXPERIMENTAL DESIGN AND ARTIFICIAL

NEURAL NETWORK SOLUTION OF STUD ARC
WELDING PROCESS
Prof.Dr. Nabeel K. Abid Al-sahib M.s.C Riyadh M.A Hamza Ass.prof.Dr. Ismail I.Al-kazaz

Baghdad University Baghdad university Baghdad university

Alkhwarzmi college of engineering College of engineering College of Engineering

Mechatronic Eng. Dept Mechanical Eng. Dept Mechanical Eng. Dept

ABSTRACT
Stud arc welding has become one of the most important unit operations in the mechanical
industries. The need to reduce the time from product discovery to market introduction is
inevitable. Reducing of standard deviation of tensile strength with desirable tensile strength
joint as a performance character was use to illustrate the design procedure. The effects of
(welding time, welding current, stud material, stud design, sheet material, sheet thickness,
sheet cleaning and preheating) were studied. Design of Experiment (DOE) is a structured and
organized method to determine relationships between factors affecting a process and output of
the process itself. In order to design the best formulation it is of course possible to use a trial
and error approach but this is not an effective way. Systematic optimization techniques are
always preferable. Tensile strength quality is one of the key factors in achieving good stud
welding process performance. 225 samples of stud welding was tested. Computer aided
design of experiment for the stud welding process based on the neural network artificial
intelligence by Matlab V6.5 software was also explain. The ANN was designed to create
precise relation between process parameters and response. The proposed ANN was a
supervised multi-layer feed forward one hidden layer with 8 input (control process
parameters), 16 hidden and 2 output (response variables) neurons. The learning rule was
based on the Levenberg-Marquardt learning algorithm.

The work of stud welding was performed at the engineering college laboratory, Baghdad
University by using the DABOTEKSTUD welding machine, for 6 mm diameter stud. The

1444
N. K. Abid Al-sahib Taguchi Experimental Design and Artificial
R.M.A Hamza Neural Network Solution of Stud Arc Welding
I.I.Al-kazaz Process

sheet materials are (K14358 and K52355) according to (USN standards, and stud materials
are (54NiCrMoS6 and 4OCrMnMoS8-6) according to (DIN standards).

The eight control parameters (welding time, sheet thickness, sheet coating, welding current,
stud design, stud material, preheat sheet and surface condition) were studied in the mixed L16
experiments Taguchi experimental orthogonal array, to determine the optimum solution
conditions.

The optimum condition was reached for the stud welding process tensile strength, where the
researcher develops a special fixture for this purpose. The analysis of results contains testing
sample under optimum condition, chemical composition of usage materials and micro
structure of optimal condition sample.

According to that:

 Practicality: the influence parameters that affect the stud welding process are welding
time, which have a major effect on stud welding process, followed by sheet material and
stud material.

 The reduction in standard deviation was approximately (30.06 per cent) and for the
range was as approximately (29.39per cent). In the other side the increase in the tensile
strength mean was as approximately (30.84 per cent). The influence parameters that affect
the tensile strength stud welding process are: the factor welding time has a major effect on
stud welding process, followed by factor C (sheet coating) and factor F (stud material).

‫الخالصة‬

‫ٍ ٔي قخارن ْا ا‬ٛ‫ يخاقاخ ياا‬ٙ‫ٍ ٕايام تاإجق ءا‬ٛ‫ى نابار انات ارن عا‬ٛ‫ا ٔتُيا‬ٛ‫كه‬ْٛ ‫ ا‬ٚ‫ى انتدارب ْإ يق‬ًٛ‫تص‬
. ‫ ا ان ارنا‬ٚ‫باط عرن ق‬ٛ‫ ا انتدقعا ٔان اأ ٔنكُاّ ن‬ٚ‫رغ يٍ انًًكٍ اسات ااو يق‬ٛ‫ٍ عأءضم ص‬ٛ‫ى يا‬ًٛ‫ إلخقاخ تص‬،‫األخقاخ‬
.‫ٓر‬ٛ‫ ًَٕذج يٍ ْ ا انهارو ء‬442 ‫رب‬ٛ‫ ٔ ا تى اخت‬،ً‫ أيثم دائًر‬ْٙ ً‫ انًُي‬ٛ‫يقق األءضه‬
‫ا نها رخ‬ٚ‫ ا تًارداً هاٗ انلا كرن انا ا‬ٙ‫ نتدارب أخاقاخ ناارو ان قغا‬ٙ‫ى عًارَٔ انارسٕ نهدرَب اناًها‬ًٛ‫يع انتص‬
‫ٍ انإايام ٔي قخارن‬ٛ‫ ن هق د ا أ ثاق عا‬ٚ‫ صًًط انل كرن انا ا‬،(Matlab V6.5) ‫ عرستاًرل ان قَريح اندرْز‬ٙ ‫انصُر‬
‫ اا‬َٙ‫ جًار‬ٙ‫ا يٕخٓا ٔم اط ءا‬ٛ‫ يتااادب ان ارن أيري‬ٙ‫ا انً تق ا ْا‬ٚ‫ انلا كرن انا ا‬.ٙ‫ى تدارب تار ٕخ‬ًٛ‫اإلخقاخ نتصا‬
ٗ‫ى أ تًارد ها‬ٛ‫ ر اب انتاه‬.) ‫قان االستدرع‬ٛ‫ٍ (يتغ‬ٛ‫ت‬ٛٛ‫ٍ خربخ‬ٛ‫ ٔ ات‬ٛ ‫ اب ي‬43 ،)‫ قب نإلخقاخ‬ٛ‫ياختن ( ٕايم انب‬
.‫( نهتاهى‬Levenberg – Marquardt) ٛ‫آن‬
ٌ‫ ياارد‬.‫ يهاى‬3 ‫ارب اق‬ٛ ٙ‫ ْا ِ اناباسا نهاارو ان إب ان قغا‬ٙ‫( ء‬DABOTEK) ٙ‫عرستاًرل ير ُ نارو ان قاغ‬
‫ َبا‬54NiCrMoS6 4OCrMnMoS8-6 ٙ‫ ْا‬ٙ‫( ٔيااردٌ ان قغا‬USN) ٗ‫ َبا ينا‬K52355ٔ K14358 ‫ا‬ٛ ‫انصا‬
ٌ‫ يااا‬،ٙ‫ى ان قغا‬ًٛ‫ تصا‬،‫ارب انهاارو‬ٛ‫ ت‬، ‫ا‬ٛ ‫ يتخ انص‬، ‫ا‬ٛ ‫ سًك انص‬،‫ (زيٍ انهارو‬ْٙ َٛ‫ قب جًر‬ٛ‫ ٕايم س‬.(DIN) ٗ‫ين‬
‫ٕيهط نابر رن‬ ‫ انً ته‬L16 ٙ‫ى تدرب تر ٕخ‬ًٛ‫ يص ٕء تص‬ٙ‫ٍ انًب ق ٔ رن انب ح) دبسط ء‬ٛ ‫ انتب‬،ٙ‫ان قغ‬
‫ األَااقا‬ٙ‫ ارٌ االَ ارف ءا‬. ‫رَارن انتدقعا‬ٛ‫ام ع‬ٛ‫ خ إب تاه‬ْٙ ٙ‫ تر ٕخ‬ٚ‫ٍ يق‬ٛ‫ق نتاب‬ٕٚ ‫ خ ٕان انت‬،‫انام األءضم‬
‫ يتٕسا‬ٙ‫ردب ء‬ٚ‫ أخقٖ رَط انز‬ٛ ‫ يٍ َر‬.ً ‫ ر‬ٚ‫) ت ق‬%46.66( ٌ‫ انًاٖ ر‬ٙ‫ ر ً ٔ رٌ األَ رف ء‬ٚ‫) ت ق‬%60.03( ٘‫رب‬ٛ‫انًا‬
‫ا‬ٛ َٕ ‫ا جاى‬ٛ ‫ ياااٌ انصا‬ٛ َٕ ّٛ‫ه‬ٚٔ ‫ زيٍ انهارو‬ْٙ ‫قاً هٗ اليخقاخ‬ٛ‫ انإايم األ ثق تأج‬.ً ‫ ر‬ٚ‫) ت ق‬%60.51( ‫ي رٔي انلا‬
.ٙ‫يااٌ ان قغ‬

1444
 Number 2 Volume 16 June 2010 Journal of Engineering

KEY WORDS: Taguchi experimental Design, Stud Welding Optimization, Artificial

Neural Network, Stud Welding.

INTRODUCTION
Stud arc welding is a widely used operation in mechanical structure, where high tensile
strength with minimum variation required. The variation of tensile strength affects the cost of
stud welding unit operations such as rework and time consume. These are often limiting steps
in mechanical manufacturing processes; therefore, significant cost reduction can be realized
by producing the stud welding joint having reliable tensile strength.( Jibson J 1979).

Usually, to find the influence of controlling parameter on welding process a large number of
experiments are needed. In order to avoid this, two statistical methods can be used to design
the optimum number of experiments. Classical design of experiments (DOE) emphasizes
prediction of future behavior of experiments from empirical model while running a fraction of
full factorial design .However; the classical DOE suffers the following limitations: two
designs for the same experiment may yield different results and the designs normally do not
permit determination of the contribution of each parameter. Taguchi DOE method, based on
the classical one, is standardized design methodology that can easily be applied by
investigators. Furthermore, designs for the same experiment by two different investigators
will yield similar data and lead to similar conclusions. (Montgomery D.C 1985)

Allen T.T. and et at 2002, present optimizing process settings method which was developed
and demonstrated for the application in robotic GMAW of sheet metal. The study it include
an objective formulation that addressed variation of noise factors. The method and the
formulation allow direct maximization of the travel speed of the welding robot. As the
formulation was implemented with standard spreadsheet software packages since it was based
on ordinary least-square regression so the method required no special software and minimal
training. Kackar R.N. 1985, introduces the concepts of off-line quality control and parameter
design and discusses the Tguchi method for conducting parameter design experiments. At the
product design stage, the objective of parameter design is to identify settings of product
characteristics, which make the product‟s performance less sensitive to the effects of
environmental variables, deterioration, and manufacturing variations. Because parameter
design reduces performance variation by reducing the influence of the sources of variation
rather than by controlling them, off-line quality control reduces cost-effective for impro ing
product quality. Ottoy K. N. and Antonsson E. K. 1991, Taguchi‟s method was extended to
involve a more design variables together with more ranges for these variables. The method is
also extended to solve design problems with constraints, invoking the methods of constrained
optimization. Finally, the Taguchi method uses a factorial method to search the design space,

1446
N. K. Abid Al-sahib Taguchi Experimental Design and Artificial
R.M.A Hamza Neural Network Solution of Stud Arc Welding
I.I.Al-kazaz Process

with a confined definition of an optimal solution. The method is compared with others for
finding optimal solution. Accordingly, Taguchi method can be used instead of other different
searching techniques. Galdmez E.V.C. and Carpinetti L.C.R. 2004, describe the application of
the experimental of designs and analysis of variance in the process of manufacture of products
for plastic injection modeling. The led experiments brought significant results, the adjustment
considered, only two factors, injection pressure and temperature of the machine, the
researchers presented a significant effect on the quality characteristic considered. Coit D. W.
Jackson B. T. and Smith A. E. 1998, consider practical aspects of building and validating
neural network models of manufacturing processes, and illustrate the recommended
approaches with two diverse case studies. When using a neural network to control and
optimize a manufacturing process, the integrity and balance of the training and validation data
sets dictate the quality of the resultant model. The experimental data was combined with the
production data. and neural networks were trained and validated on the combined data set. Su
c. and Miao C. 1998, apply neural networks to analyze an experiment with singly censored
data (incomplete data). Iwo procedures are developed; the first procedure is quite
straightforward and can be easily used to rapidly determine the optimal condition. Hsu
C.M.2001 ,proposes a four-phased procedure based on neural networks and principal
component analysis to resolve the parameter design problems with multiple responses. The
quality characteristics of a product are first evaluated through Taguchis quality loss function a
neural model is then trained to map out the functional relationship between control factors and
responses‟ quality loss. The functional relationship is then fed into the principal component
analysis procedure to transfer a set of responses into a set of uncorrelated principal
components. A feasible combination of control factors can be obtained through the recalling
function of a neural model.

Once the variation of tensile strength was chosen as the main performance characteristic (the
measure of quality) then the design factors, which will have an influence on it, have to he
selected. Since most welding experiments usually involve a significant number of factors,
according to the Taguchi method, the number of experiments can be reduced. Using a special
orthogonal array only a small set from all the possible ones is selected. The sense of the
orthogonal arrays method lies in choosing the level combinations of the design factors for
each experiment. A practical definition of experimental design that can be applied to stud of 6
mm diameter arc welding process is presented in this study.

The survey shows this method has some weakness in the required number of experiments
where it equals the number of inner array multiplied by the number of outer array that may
cause higher number of experiments than that which is needed, The survey also shows that it

1441
 Number 2 Volume 16 June 2010 Journal of Engineering

can make a good relationship between the input parameters and the output with minimum
error by using neural network.

Taguchi Approach to Parameter Design

Taguchi method provides a systematic and efficient approach for conducting experimentation
to determine near optimum settings of design parameters for performance and cost. The
method pushes quality back to the design stage, seeking to design a product/process. which is
insensitive to quality problems. The Taguchi method utilizes orthogonal arrays to study a
large number of variables with a small number of experiments. It can reduce research and
development cost by simultaneously studying a large number of parameters. Using orthogonal
arrays the method can significantly reduce the number of experimental configurations. In
order to analyze the results. the Taguchi method uses a statistical measure of performance
called „signal-to-noise‟ ratio, (S/N) After performing the statistical analysis of S/N ratio, an
analysis of variance (ANOVA) needs to be employed for estimating error variance and
determining the relative importance of various factors. From their relative importance and
from the S/N ratio, the optimum condition of factors is chosen. The result at this point is
estimated using equation:

R=T+Σ(Ai-T)

Where:

R= predicate mean response at the optimal condition

T= overall mean of all observation in the data

Ai= average value of significant factors at level i

Cause and Effect Diagram

The total variation in the stud arc process may be due to any or a combination of the six
sources (machine, measurement, method, material, manpower and environment). For this
study of stud arc welding the effect of manpower on variation is limited because the machine
is operating in a semiautomatic process, also the experiments have been executed in the
laboratory environment. so it includes consider the first four ,the other are ignored. Problem
identification is very important for any industrial experiment. One of the most used methods
for identifying the problem is brainstorming. Brainstorming is an activity that promotes team
participation, encourages creative thinking and generates many ideas in a short period of time.
For an investigation into the possible causes of the undesirable variability in stud welding
process, a cause-and-effect diagram that lists several suspected causes of this variability, is
shown in figure (1). Brainstorming in conjunction with cause and effect analysis (CEA) is
used to identify the control factors which are to be considered for the experiment.

1442
 N. K. Abid Al-sahib Taguchi Experimental Design and Artificial
R.M.A Hamza Neural Network Solution of Stud Arc Welding
I.I.Al-kazaz Process

Material
Stud Method
Sheet thickness design Stud (process)
Preheat
)
Sheet
Arc
Stud diameter voltage Power
Sheet Material
supply
Sheet Stud material
coating
Welding time Welding current

Stud
welding
quality
Machine
Polarity of power type Ventilation Measurement
the machine gage
Collect Operator
Gun wire Measurement
wear Performanc Quantity
e tool
Plunge
Machine depth Man Measurement
(Operator) Environment (Testing machine)
tt

Figure (1) the suggested stud welding cause-and-effect diagram

1443
 Number 2 Volume 16 June 2010 Journal of Engineering

MATERIALS

The diameter of stud to be studied was (6 mm) that is widely used in the mechanical
structure and also require low energy for welding from other stud diameter. For sheet, the first
was galvanized (K52355 steel) and the second was non- galvanized (Kl4358steel), with two
dimensions thickness gage 16 (1.6002 mm) and gage 12 (3.175 mm). For stud, the first was
(54N1CrMoS6 steel) and the second was (4OCrMnMoS8-6 steel).

Method (Identification of Process Parameters)

There are (20) factors identified in this study. Eight independent control factors are
considered to improve the stud welding process. These factors are (welding time, sheet
thickness, sheet material. welding current, stud design, stud material, preheat sheet and
surface cleaning). There other factors were classified as noise factors.

Selection of Factor Levels and Range of Factor Setting

Once independent factors are decided, then the number level for each factor is selected.
Selection of levels depends on how the outcome (tensile strength) is affected by different
level settings.

Determining the number levels of selected factors from brainstorming is another major
concern to many researchers in industries. Brainstorming session it was suggested that
suitable to use eight factors on of them in multi-level. Seven of the eight control factors have
two levels, and one has eight levels that is welding time.

After determining the number of levels required for each factor, it is needed to specify
the range of operation for each control factor. It is usually best to experiment with the largest
range feasible, so that the variation inherent in the process does not mask the factor effects on
the response. The levels for welding time is shown in table (1), and the list of seven control
factors and their level are shown in table (2).

Table (1) Levels of Welding Time Control Factor and for the Experiments

Factor Level Level Level Level Level Level Level Level

Factor Unit
label 1 2 3 4 5 6 7 8

Welding
time A second 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
(second)

1444
N. K. Abid Al-sahib Taguchi Experimental Design and Artificial
R.M.A Hamza Neural Network Solution of Stud Arc Welding
I.I.Al-kazaz Process

Table (2) Control Factors and Levels for the Experiments

Factors Factor labels Unit Level1 Level 2

Sheet thickness B mm 1.6 3.175

Sheet material C None K52355 K14358

Welding current D Ampere 350 540

Stud design E None Small stud Flange stud

Stud material F None 54NiCrMoS6 40CrMnMoS8-6

Preheating G None Preheating No preheating

Surface cleaning H None Oil sheet Clean sheet

Measurement (Tensile Testing Technique)

The response measurement should be well defined. This includes choosing the measurement
and processing equipment to be used, how to measure, where to measure and where to
document the data. The stud welding specimen is not the standard specimen tensile test
dimension because one side is screw and the other is sheet, so the tensile testing was made by
developing a special fixture for testing operation, special fixture is shown in figure (2):

Figure (2) the special fixture for the stud welding

Design of Experiments
Normally, in the case of eight factors one of them in eight levels and other in two levels are 8
= 125 experiments should be conducted. In accordance with the Taguchi‟s method the
standard orthogonal array L16, with only 16 experiments (Table 3) could be used.

1445
 Number 2 Volume 16 June 2010 Journal of Engineering

Table (3) code design matrix orthogonal array L16 2781 .

run welding sheet sheet welding stud stud preheat surface
time thickness material current design material cleaning
1 1 1 1 1 1 1 1 1
2 1 2 2 2 2 2 2 2

3 2 1 1 1 1 2 2 2
4 2 2 2 2 2 1 1 1
5 3 1 1 2 2 1 1 2

6 3 2 2 1 1 2 2 1
7 4 1 1 2 2 2 2 1
8 4 2 2 1 1 1 1 2
9 5 1 2 1 2 1 2 1
10 5 2 1 2 1 2 1 2
11 6 1 2 1 2 2 1 2
12 6 2 1 2 1 1 2 1
13 7 1 2 2 1 1 2 2
14 7 2 1 1 2 2 1 1
15 8 1 2 2 1 2 1 1
16 8 2 1 1 2 1 2 2

Experimental Preparation and Process Run

The experimental preparation involves those activates that occur prior to actual running of the
experiment. Poor preparation is the most frequent cause of inconclusive results. Errors in the
experimental procedure this step can affect the experimental validity. The experiment of the
study was conducted in a laboratory at which air conditioning is similar to outdoor
environment. Also all the important material, machine was prepared in similar of an
experimental area. In this step. the main task was to construct the uncoded design matrix for
the experiment. The uncoded design matrix is shown table (4).

1446
 N. K. Abid Al-sahib Taguchi Experimental Design and Artificial
R.M.A Hamza Neural Network Solution of Stud Arc Welding
I.I.Al-kazaz Process

Table (4) uncoded design matrix array L162781

sheet welding
welding time sheet stud surface
Run thickness current stud material preheat
(second) material design cleaning
(mm) (Ampere)

1 0.15 1.6 K14358 350 Small 54NiCrMoS6 Preheat Clean sheet

2 0.15 3.175 K52355 540 Large 40CrMnMoS8-6 No preh. Oil sheet

3 0.2 1.6 K14358 350 Small 40CrMnMoS8-6 No preh. Oil sheet

4 0.2 3.175 K52355 540 Large 54NiCrMoS6 Preheat Clean sheet

5 0.25 1.6 K14358 540 Large 54NiCrMoS6 Preheat Oil sheet

6 0.25 3.175 K52355 350 Small 40CrMnMoS8-6 No preh. Clean sheet

7 0.3 1.6 K14358 540 Large 40CrMnMoS8-6 No preh. Clean sheet

8 0.3 3.175 K52355 350 Small 54NiCrMoS6 Preheat Oil sheet

9 0.35 1.6 K52355 350 Large 54NiCrMoS6 No preh. Clean sheet

10 0.35 3.175 K14358 540 Small 40CrMnMoS8-6 Preheat Oil sheet

11 0.4 1.6 K52355 350 Large 40CrMnMoS8-6 Preheat Oil sheet
12 0.4 3.175 K14358 540 Small 54NiCrMoS6 No preh. Clean sheet
13 0.45 1.6 K52355 540 Small 54NiCrMoS6 No preh. Oil sheet
14 0.45 3.175 K14358 350 Large 40CrMnMoS8-6 Preheat Clean sheet
15 0.5 1.6 K52355 540 Small 40CrMnMoS8-6 Preheat Clean sheet
16 0.5 3.175 K14358 350 Large 54NiCrMoS6 No preh. Oil sheet

The outer array of (12) noise factors with three combinations will be L16, so the total
number of runs to be conducted in this case would be
l6  12  2 = 384 experiments as minimum. Performing of many unimportant experiments is
costly and time consuming; the operating characteristic (OC) curve was used to develop the
sample size. The experiment tensile outputs are shown in table (5)

1450
 Number 2 Volume 16 June 2010 Journal of Engineering

Table (5) tensile strength runs

Standard
actual run Mean
Run Tensile strength (N/mm² ) deviation
order N/mm²
N/mm²

1 5 175.73 213.23 143.66 195.09 210.50 155.60 182.302 28.860

2 9 288.70 251.20 330.40 284.99 225.90 300.70 280.315 36.946
3 13 284.39 198.56 225.89 245.87 276.24 263.54 249.082 32.539
4 3 359.99 420.50 428.42 300.03 387.38 367.54 377.310 46.790
5 12 190.70 245.87 235.90 298.46 164.33 289.46 237.453 52.977
6 11 370.45 392.68 191.74 360.38 288.70 383.26 331.202 77.637
7 8 321.60 139.00 349.05 310.00 362.93 457.50 323.375 104.318
8 1 331.96 326.32 331.15 401.60 387.26 314.78 348.828 36.095
9 4 388.10 233.60 372.20 287.95 225.43 278.00 297.547 68.611
10 2 530.00 460.72 549.85 375.12 410.53 375.89 450.352 76.343
11 15 305.40 383.20 456.00 378.00 478.00 375.00 395.933 62.388
12 7 152.09 160.74 170.76 166.80 250.88 132.45 172.287 40.835
13 16 219.19 152.97 250.85 257.16 266.78 198.75 224.283 43.258
14 10 155.65 180.45 289.40 220.68 225.35 248.78 220.052 47.705
15 14 289.36 215.62 318.43 256.84 288.23 145.63 252.352 62.900
16 6 185.32 178.45 223.21 155.82 298.33 188.43 204.927 50.651

Analysis of variance
Equations for conducting the variance are presented in this section. Sum of squares (Si)
of factor i at level k was calculated according to the equation:
2 2
 N   N 
 Y j    YJ 
   
L
Si   J
 J

K N K N

Where, N is the total number of experiments. Nk the number experiments of each level and Yj
the mean response.
The total sum of squares (ST) was calculated using equation:
2
 N 
  YJ 
 
N
ST  YJ
J
2
 J

Experimental error (Se) was calculated:

Se  ST   Si

1454
N. K. Abid Al-sahib Taguchi Experimental Design and Artificial
R.M.A Hamza Neural Network Solution of Stud Arc Welding
I.I.Al-kazaz Process

Mean square of factor i (Vi) vas computed using the following equation:
Si
Vi 
fi

Where, f i is degree of freedom, which is one less than the number of levels. The total degree
of freedom of the result (fT) is one less than the total number of experiments. The degree of
freedom for error variance (fe) is the total degree of freedom minus sum of degree of freedom
of factors. The next step was the calculation of the variance ratio (Fi), which is the quotient of
mean square of factor and error. The fraction of importance of each factor (in percents) was
calculated according to the equation:

va
Fa 
ve

The variance ratio, commonly called F statistic (named after Sir Ronal A. Fisher). is the
ratio of variance due to the effect of a factor and variance due to the error term. This ratio is
used to measure the significance of the factors included in the error term. The F value
obtained in the analysis of variance is compared with a value from standard F tables for a
given statistical level of significance. Confidence interval, C.I., of the factor effect and
estimated value of the result at the optimum condition was computed using the following
equation:

F  ,1, f 2   ve
C.I. of m  mˆ 
Ne

F (, 1, f 2) = table value of F,  significant level with 1 degree of freedom for the numerator
and f 2 degrees of freedom for the error term.
ve = error variance = MSE , Ne = is effective number of replications ,Where each factor can
be calculate from:

 sa   se 
a  
 s 
  100 , e     100
 t   st 
Results and Discussion
After creating a Taguchi orthogonal array, the selected experiments were performed. A
statistic analysis summary of the tensile strength, called S/N ratio, is employed to find the
optimum level of the selected factors. The average s/n ratio of each run is shown in (Table 6).

 1 
 2 
 10 log 
S yi
N LTB  n 
 
 

1454
 Number 2 Volume 16 June 2010 Journal of Engineering

Table (6) The SNR values for experimental trials

Trial
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
no.
S/N
44.9 48.7 47.7 51.3 46.9 49.4 48.1 50.7 48.9 52.7 51.6 44.2 46.5 46.3 47.0 45.7
(dB)

After obtaining the SNR values, the next step was to obtain the average response values of
SNR at low and high levels of each factor and hence the effect of each factor on the SNR. The
results are shown in table (7) and table (8).

Table ( 7) Average SNR Table for factor A

Factor Average Average Average Average Average Average Average Average Effect rank
A SNR at SNR at SNR at SNR at SNR at SNR at SNR at SNR at of the
level 1 level 2 level 3 level 4 level 5 level 6 level 7 level 8 factor
Factor 46.83 49.53 48.19 49.43 50.84 47.96 46.41 46.38 4.52 1
Effect dB

Table (8) Average SNR Table for other factor

Average SNR at Average SNR at

Effect of the factor dB
Factors level 1 dB level 2 dB rank

B 47.73 48.69 0.96 6

C 47.10 49.31 2.21 2
D 48.18 48.23 0.05 8
E 48.23 48.46 0.23 7
F 47.41 49.00 1.69 3
G 48.98 47.43 -1.65 4
H 47.55 48.86 1.31 5

Tables (7) and (8) show that factors A and C have a dominant effect on the SNR, followed by
factors F, G, H, B, F, and D. The main effects plot for the SNR is shown in figure (3).

1456
 N. K. Abid Al-sahib Taguchi Experimental Design and Artificial
R.M.A Hamza Neural Network Solution of Stud Arc Welding
I.I.Al-kazaz Process

Welding Sheet Sheet Welding Stud Stud Surface

Preheat
time thickness material current design material condition
51

50

49
SNRA1

48

47

1 23 45 67 8 1 2 1 2 1 2 1 2 1 2 1 2 1 2

Figure (3) the main effects plot for S/N ratio

The relative magnitude of the effect of different factors can be obtained by decomposition of
variance, called ANOVA (Table 9). The Sum of Squares column in Table (9) was calculated
using equations (2), (3) and (4), the Mean Square column with equation (5) and the F-ratio
column as calculated with equation (6). The ANOVA table has shown that the most dominant
factor effects arc D (welding current), E (stud design) and A (welding time). The optimal
conditions setting of factors. which will maximize the SNR is (i.e. the best control factor
settings) depend on the SNR arc AS, 132. C2, D2, E2, F2, G1 and H2.The calculations of
Analysis of Variance for the factors by using Matlab software are:

Table (9) ANOVA for the SNR

Source of variation Sum of Squares df Mean Square F-ratio

A 37.384 7 5.341 0.88

B 3.529 1 3.529 0.58
C 19.769 1 19.769 3.26
D 0.004 1 0.004 0.00
E 1.129 1 1.129 0.19
F 9.899 1 9.899 1.63
G 9.402 1 9.402 1.55
H 6.679 1 6.679 1.10
error 6.070 1 6.070 1
Total 93.865 15 6.257

1451
 Number 2 Volume 16 June 2010 Journal of Engineering

Stage (2): Performing the SNR analysis and (S.D.) analysis, then the next step was to
identify the factor effects that have significant impact on the mean response. The average
response values at each level of the factor A and the effects are present in table (10), and the
average response values at low and high level for the other factors and their effects are present
in table (11).
Table (10) the average response of welding time control factor
Factor Average Average Average Average Average Average Average Average Effect rank
A mean at mean at mean at mean at mean at mean at mean at mean at of the
level 1 level 2 level 3 level 4 level 5 level 6 level 7 level 8 factor
Factor Effect 231.3 313.1 284.3 336.1 382.3 284.1 222.1 228.6 160.6 1
2
N/mm

Table (11) the average response values at each level of the factors and their effects

Mean response at Mean response at Effect

Factors rank
level 1 N/mm 2
level 2 N/mm 2
N/mm2

B 270.29 298.16 29.96 6

C 257.07 313.47 56.4 3
D 278.73 291.81 13.08 8
E 278.43 292.11 13.68 7
F 255.61 314.93 59.32 2
G 310.17 260.37 -49.8 4
H 269.55 300.99 31.44 5

The main effects plot factor effects are illustrate in Figure (4).
Welding Sheet Sheet Welding Stud Stud Surface
2 Preheat condition
N/mmtime thickness material current design material
380

340
strength
tensile

300

260

220
12 34 56 78 1 2 1 2 1 2 1 2 1 2 1 2 1 2
Figure (4) main effects plot for the mean response

1452
N. K. Abid Al-sahib Taguchi Experimental Design and Artificial
R.M.A Hamza Neural Network Solution of Stud Arc Welding
I.I.Al-kazaz Process

Figure (4) shows factors A, C, E and F have a significant impact on the mean
response (i.e. mean tensile strength). This will be followed by factors B, H, D and
E.
The variance ratio (F-value statistic) represents the ratio of variance due to the effect
of a factor and variance due to the error term. This ratio is used to measure the
significance of factors included in the error term [17, 18]. The F value obtained in the analysis
of variance is compared with a value from standard F tables, to decide significance of
statistical level. It can be seen from table (12) that factor A (welding time) has a large affect
on the mean of stud welding tensile strength (43.37° of fraction of importance). Value of
factor C (sheet material) and F (stud material) are (13.84%). (13.53%) respectively.

Table (12) ANOVA for the response

Source of Percent contribution

Sum of Squares df Mean Square F-ratio
variation (ρ)
A 42304.33 7 6043.48 42.35 43.37
B 3089.23 1 3089.23 21.6 3.17
C 13493.92 1 13493.92 95.14 13.84
D 519.19 1 519.19 3.35 0.53
E 1005.63 1 1005.63 63.25 1.04
F 13203.33 1 13203.33 91.04 13.53
G 8817.68 1 8817.68 63.25 9.05
H 3271.65 1 3271.65 23.94 3.35
error 11829.65 81 143.84 1 12.12
total 97534.58 95 1119.02 - 100

Added the factors B,D,E,G and H can be pooled. A new table without the above factors was
constructed (table13)

Table (13) the pooled ANOVA for response

Source of Sum of df Mean Variance ratio Percent
variation Squares Square (F-ratio) contribution (ρ)
A 42644 7 6092 14.2 40.36
C 13686 1 13686 31.91 13.5
F 13095 1 13095 30.53 12.89
error 28779 86 428.86 1 33.25
total 98204 95 1033.72 100

1453
 Number 2 Volume 16 June 2010 Journal of Engineering

The sum of squares of pooled factors was added to the error term, and then new mean square
of the error term was calculated using equation:

s i
p
 se
, Where superscript p indicates the pooled factors.
Ve  i

f
i
i
p
 fe

Since the degree of freedom of the factor A is 7 and for error term is 86, so, F7 = 2.11
at level of significance (950 o confidence), see ( F-table )(Fisher -table).

As the computed value of variance ratio (14.2), (table 13), is bigger than the value from
(F table), so there is in 95% of confidence this factor A (welding time) has an effect on stud
welding process. For (C and F) factors, the degree of freedom is 1 so the F1, 3.97, since
computed F-ratio are 31.91 and 30.53 respectively for each (table 13) which is higher than
from F-table, then the above two factors also have an effect in the stud welding process, as
well as for A.

After identifying the significant factor effects, the next step was to determine the
optimal setting for these factors which will bring the mean response as close as possible to the
target. He optimum condition (i.e. the best control factor settings) based on the mean response
figure (4) was:
A5, B2, C2, D2, E2, F2, G1 and H2.
Here the factors B, C, F and H are the same with the last setting (see stage one). While
for factor A this is difference, when we choose A5 or A6, if choose A5 (the welding time
is0.35 second), then the tensile strength and standard division will be (382.341N/mm2 and
72.47 N/mm2) respectively, while when choose A6 (the welding time is 0.4 second), the
tensile strength and standard division will he (284.110 N/mm2 and 51.61N/mm2). So, in this
study an estimated of factor A (A6=0.38 second) is considered. For factor D, the mean and
standard deviation of this factor, in level D1, is (278.73 N/mm2, 50.56 N/mm2) respectively,
while them, in level D2, are (292.11 N/mm2, 58.04 N/mm2) respectively, so D1 would be
considered. The same thing for factor E1. For factor G, the mean and standard division of this
factor, in level G . are (310.17 N/ni2, 51.75 N/mm2) respectively, while them, in level C , are
(260.37 N/mm2, 56.84 N/mm2) respectively, so C I would be considered. The factors levels
are:

Ẩ6, B2, C2, D1, E1, F2, G1 and H2.

In order to reach the optimal factor settings, the factor setting that yields minimum
quality loss can he study. The quality loss function for larger the better is:

1454
N. K. Abid Al-sahib Taguchi Experimental Design and Artificial
R.M.A Hamza Neural Network Solution of Stud Arc Welding
I.I.Al-kazaz Process

L y   o 2o
1
y2

The summarized calculation is shown in table (14).

Table (14) loss function calculation

(yˆ)² L(y)/K (money unit/piece)

Run
1 299094.4 3.3×10-6
2 202216.8 4.9×10-6
3 230924.9 4.33×10-6
4 125509.6 7.97×10-6
5 243914.1 4.1×10-6
6 163858.9 6.1×10-6
7 174992.3 5.71×10-6
8 145438.5 6.89×10-6
9 190410.7 5.25×10-6
10 83183.4 1.2×10-5
11 160655.2 6.22×10-6
12 290938.1 3.43×10-6
13 256371.3 3.9×10-6
14 260776.3 3.83×10-6
15 230653.3 4.33×10-6
16 276680.6 3.61×10-6

From table (14), run (1) which represented in bold yield the rnininmm loss. Settings
based on the loss-function analysis was therefore obtained as:

A1, F1, C1, G1 and H1

For factor A, level 1 will yield a veiy low tensile strength (182.302N/mm2), so this
level is not taken. ibr the three factors F, C and G the level is the same, for factor H in level 1
the tensile strength is (269.55N/rnrn2), while in level 2 it is (300.99N/mm2) the reduction is
also high, so the final optimum stetting is:

Ẩ 6, B2, C2, D1, E1, F2, G1 and H2.

1455
 Number 2 Volume 16 June 2010 Journal of Engineering

Predicted Mean Response at the Optimal Condition

The predicted mean response at the optimal condition is estimated only from the
significant main effects. The main factor effects, which has a significant impact on the mean
response were A. F. C, C and H. The predicted mean response based on the optimal factor
levels of A, F, C, C and II is given by:

R= T+ (Ẩ 6-T) + (C2-T) + (F2-T) + (G1-T) + (H1-T)

Where
R= predicted mean response at the optimal condition
T = overall mean of all observation in the data
R=284.225+(310.5-284.225)+(313.47-284.225)+(314.93-284.225) +(310.17-
284.225)+(300.99-284.225)
2
R=413.185 N/mm
Interpretation, Experimental Conclusions and Confidence Interval for the Predicted
Mean Response
After interpreted the results of the analysis, it is advisable to ensure that the
experimental conclusions are supported by the data. The confidence interval is the variation of
the estimated result at the optimum condition.

MSE = error variance =143.84 N/mm2

F1, 96 = 3.96
96
Ne  8
7 11111
Therefore, the 99 per cent confidence interval for the mean tensile strength is given by:

3.96  143.84
99 percentCI  413.185 
8
=413.185 ±8.43 N/mm2
The result at the optimal condition is 413.185±8.43 N/mm2 at the 99 percent
confidence level. After determination the confidence level for the predicate mean response,
makes a confirmation experiment or run. The confirmation experiment/run is used to verify
whether the predicated mean response based on the optimal combination of factor levels give
process response within the confidence limits or not. If conclusive results are obtained from
the confirmation run, a specific action on the process may be taken for improvement.

Confirmation Run
A confirmatory run/experiment (or follow-up experiment) is necessary in order to
verify the results from the statistical analysis. This is to demonstrate that the factors and levels
chosen for the influential factors do provide the desired results. The insignificant factors

1456
N. K. Abid Al-sahib Taguchi Experimental Design and Artificial
R.M.A Hamza Neural Network Solution of Stud Arc Welding
I.I.Al-kazaz Process

should be set at their economic level during the confirmation run/experiment. If

conclusive results have been obtained, improvement action on the product or process under
investigation is recommended. On the other hand, if the result does not turn out as expected,
further investigation may be required.

In industrial experiments, once the solution has been implemented, it is recommended

to monitor the process by constructing control charts on the experiment‟s response variable
(s) and critical factors that influence the response. Control charting will ensure that the
problem does not reoccur [133]. For the study, the sample taken contains ten pieces were
produced under the optimal condition that is in table (15):

Table (15): the optimum stud welding condition based on Taguchi

methodology optimization
factor level
Ẩ 6: welding time 0.38 second
B2 :sheet thickness 3.175 mm
C2 :sheet material non- galvanized (K14358steel)
D1: welding current 350 Ampere
E1: stud design Small stud
F2: stud material 40CrMnMoS8-6 steel
G1 Preheating
H2: Surface cleaning Clean sheet

The results are shown in table (16):

Table (16) the sample tensile strength based on Taguchi methodology optimization

Tensile strength
Sample
N/mm2
1 443.52
2 421.32
3 410.63
4 390.48
5 472.40

6 422.67

7 398.93
8 431.88
9 408.33

10 524.55

1460
 Number 2 Volume 16 June 2010 Journal of Engineering

The mean tensile strength from the confirmation run was 432.47 N/mm2 the standard division
is 39.950 N/mm2 and the range is 134.07 N/mm2.

The effect of every factor of the study can be summarized as:

Welding Time
This factor strongly effects on tensile strength measure. the mean value of tensile strength in
levels (0.15 ,0.2 ,0.25 ,0.3 ,0.35 ,0.4 ,0.45 ,0.5) second is (231.3 ,313.196 ,284.32 ,336.l
,373.95, 284.11 , 222.16 . 228.64) N/mm2 respectively. The effect of factor on the mean is
(42.3 9perccnt) which shows 110W much the variation of stud welding tensile strength from
one level of welding time to another, The welding time has a relationship with the input
energy rate; there is when the welding time increases the average input energy increases that
lead to increase in tensile strength until value it decrease due to over energy.

Sheet Material
Macrograph pictures show that the sheet material had two effects. First, galvanizing appeared
to result in greater porosity in the joints. The mean of tensile strength in level l(K52355) is
157.07 Nmm2 and in level 2 (K14358) is 313.47 N/mm2. Also, there appeared to be
considerably less heat and retained liquid metal in the joints on coated sheet. Second, the non-
galvanized sheet (K14358stec1) indicates higher tensile strength, this may be due to the
percentage of carbon contain (O.l44°o) is higher than for galvanized (K52355) sheet
(0.0689%). The effect of factor on the mean tensile strength is (13.78percent).

Stud Material
This factor also effects in the stud welding process, the different value of tensile strength call
he shown from one level to another. The mean tensile strength in level 1(54NiCrMoS6) is
255.61 N/mm2 and in level 2 (4OCrMnMoS8-6) is 314.93 N/mm2.The effect of factor on the
mean is (l3.l8percent). The higher value of strength for (4OCrMnM0S8-6) from the strength
for (54NiCrM0S6) may be due to the containing of carbon where for the first (0.229%) while
for the second (0.139%) , as described previously for sheet material, and also due to other
alloy elements for example tile percentage of Mg is (1.07°) ill 4OCrMnMoS8-6 and (0.405%)
in 54NiCrMoS6.

Preheating
This factor gives a positive effect on both the increase of the tensile strength and a decrease in
the variation of process. The mean tensile strength in level l(preheating) is 310.17 N/mm2 and
in level 2 (no- preheating) is 260.37 N/mm2 .The factor effect on the mean is (9.1 percent).

1464
N. K. Abid Al-sahib Taguchi Experimental Design and Artificial
R.M.A Hamza Neural Network Solution of Stud Arc Welding
I.I.Al-kazaz Process

The base metal must he preheated to prevent the formation of cracks. This is similar to the
effect on arc welding process for reducing heat effect (heat tear) that reducing the cooling rate
for tile welding area and HAL which reducing the hardness of these areas especially when the
carbon percentage more than 0,25 that yield hardness phases (without preheating). All
oxyfuel gas heating torch is used for heating because only a localized preheated zone is
needed; the preheating temperature is between (31 5-3 70) °C.

Stud Design
The design of tile stud influenced the working area of the stud surface. This factor was found
to completely dominate the tensile results. Despite of the fact that flange studs are going to
have a greater area for welding and subsequently greater strengths, flange stud joints
susceptible to porosity compared to smaller studs as micrographic pictures show, this appears
to be due to a geometry effect. Tile mean of tensile strength in level l(small stud) is 278.43
N/mm2 and in level 2 (flange stud) is 292.1 IN/mm2 Flange stud actually appeared to
increase tensile strength performance and this is showing in many specimens but the variety is
more may be due to preparing of flange stud is not at accuracy enough that causes porosity.
The effect of this factor on the mean is (9.11 percent).

Surface Cleaning
This factor has little effect on the measured tensile strength compared with tile previous
factors; the effect of this factor on the mean is (3.36 percent). The mean of tensile strength in
level 1(oil sheet) is 269.55 N/mm2 and in level 2 (clean sheet) is 300.99 N/mm2 .The clean
sheet already shows the greatest tensile strength and this is logic, but limited and which may
be due to the wielding area is small that lead to little effect.

Sheet Thickness
Increasing sheet thickness has two effects; the first; a thicker sheet is stiffer during
mechanical testing and this minimizes the peel characteristic of the tests and increases
strength. The second thicker sheet present increase in the area of heat diffusion that lead to
high cooling rate which creating inherently stronger welds. The mean of tensile strength in
level 1(1.6 mm) is 270.29 N/mitt and in level 2(3.175 mm) is 298.16 N/mm2 The effect of
this factor on the mean is (3.0 percent).

Welding Current
This factor has the smallest effect factor where the effect is (0.34percent). The mean of tensile
strength in level 1(350 ampere) is 278.73 N/mm2 and in level 2 (540 ampere) is 291.81
N/mm2. I his result is far from the expected result where the welding current play important

1464
 Number 2 Volume 16 June 2010 Journal of Engineering

role in arc welding process. But this happen here may be due to the two levels of welding
current that choosing represent the boundaries of welding current, and there no
intermediate grade between this two levels in the welding current selector of stud arc
machine.

CONCLUSIONS
The study has showed a significant improvement (approximately 30.84 percent) increase in
stud joint tensile strength and (approximately 30.06 percent) decrease in stud joint tensile
strength variation.

Measures of weld quality in this study included tensile strength testing and some macrograph
photos. Statistical techniques used to produce a series of main effect plots for factors and
results are analyzed. These robustness plots allowed direct observation of how weld quality
measure was affected by each factor of interest. Specific conclusions from this study are as
follows:

 Dominant factors in the Performance of Stud Welds - the dominated effective factors
of stud welds performance are (welding time), (plate material) and (stud material) study.

 Effect of preheating plate - preheating has positive effects on the increasing of the
tensile strength with reducing variability.

 Effect of Stud design - increasing stud area appeared to decrease of measures of

tensile strength. This was true where the levels of internal porosity also increased with the
larger studs.

 Effect of Plate Thickness - increasing thickness led to increases in mechanical

measure (tensile strength) of weld quality. The benefits appeared to come from increased
stiffness of the joint as well as increased peel strengths associated with the thicker
material.

 Effect of Plate Material - Welding onto galvanized plates appears substantial porosity
in the joint, so the non-galvanized plates obtain better tensile strength.

 Effect of Other Factors - weld quality measurements (tensile strength) as well as

macrograph sections show the other factors in the study, welding current and the presence
of surface cleaning, all had little effect.

1466
N. K. Abid Al-sahib Taguchi Experimental Design and Artificial
R.M.A Hamza Neural Network Solution of Stud Arc Welding
I.I.Al-kazaz Process

REFERENCES
 Jibson J.”Advance Welding”, John Wiley & Sons, 1997.

 Montgomery D.C” Design and Analysis of Experiments” Second Edition John Wiley &Sons,
Inc., 2001.

 Montgomery D.C. "The Use of Statistical Process Control and Design of Experiments in
Product and Process Improvement" JIB Transactions, Vol.24 No. 5. PP. 4-17, 1992.

 Montgomery D.C “Introduction to Statistical Quality Control” John Wiley & Sons, Inc.1985.

 Allen IT. and et al” Statistical Process Design for Robotic GMA We dipg of Sheet Metal”
Welding Journal, PP.69-s_77-s, May 2002.

 Kackar R.N.” Qft4Jng Quality Control, Parameter Decigppnd the Tagnehi „VIe/hod”. .1. Qual.
Techn. , Vol. 17, No. 4, PP 176-188, 1985.

 Ottoy K. N. and Antonsson E. K.” Extensions to the Taguchi Method of Product ASME
Journal of Mechanical Design. January 6, 1991.

 Galdmez E.V.C. and Carpinetti L.C.R. “Application as Design oLpgrimentu7 the Process of
Manufhcturing of Plastic Products”, ASME Journal of Manufacturing Science and
Engineering, Vol. 122, PP. 360-369, 2004.
 Coit D. W. Jackson B. T. and Smith A. E.” Static Neural IVetwork Process Models:
Considerations and Case Studies”. mt j. Prod. Res., Vol. 36. No. 11, 2953- 2967, 1998.

 Su C. and Miao C. “Neural Network_Procedures fbr ExperpgjtilAigil.sisJ/h Censored Data”,

International Journal of Quality Science, Vol.3, No.3, PP.239-253. 1998.

 Hsu C.M.” Solving Multi-Response Problems through Neural Networks aJgLj‟iI7cipp7

J22gnepd‟ApglJis”. Journal of the Chinese Institute of Industrial Engineers, Vol. 18. No. 5,
pp. 47-54, 2001.

 Roy R., "Design of Experimental Using the Tagpchi Approach‟ Wiley, New York, 2001.

 Montgomery D. C. “Design andAnalsis QjExperiments”. New York. Wiley. 1991.

 Buyske S. and Trout R.” Robust Design and Taguchi Methods” .Journal of Quality
Technology, vol. 22, No. 1, PP. 15 22, Jan. 2003.

 Steiner 5.1-1. and MacKay R. J.” Statistical Engineering: an Algorithm far Reducing
Variation in Majfpcturing Processes” American Society for Quality ASQ, 2005.

 Taguchi G. and Yokoyama Y.” Taguchi Method‟s: Design of Experiments “Quad ity
Engineering, Vol.4, Dearbon, MI, 1993.

 Mukhopadhyay S. K. and Chakraborty D. “Optimal Process Variance under Tqggjj Loss”

International Journal of Quality & Reliability Management. Vol. 12 No. 9, PP. 14-29, 1995.

 Antony J. and et al “Process Optimization using Taguchi Me1fpfExperimeJ7/ul çin” Work

Study, Volume 50, Number 2, PP. 51-57, 2001.

 Lofthouse T. “The Taguchi Loss Function” Work Study, Volume 48, Number 6. PP. 218-222,
1999.

1461