2011 Third International Conference on Computational Intelligence, Modelling & Simulation

Comparison of Intelligent Optimization algorithms for Wire Electrical Discharge

Machining Parameters
Abolfazl Golshan

Soheil Gohari

Amran Ayob

Faculty of Mechanical Engineering

Faculty of Mechanical Engineering

Faculty of Mechanical Engineering

Isfahan University of Technology

Universiti Teknologi Malaysia

81310

Universiti Teknologi Malaysia

81310

Skudai, Johor, Malaysia

Skudai, Johor, Malaysia

Email: soheil.gohari7@gmail.com

Email: amran@fkm.utm.my

Isfahan, Iran
Email: gabolfazl@gmail.com

Choosing the appropriate parameters in order to

achieve the corresponding surface roughness and
maximum material removal rate will be possible with
having the knowledge of the way these parameters
influence on mentioned factors which is also prioritized
by this study.
In the recent years, diverse theorical and experimental
methods have been used in order to model and optimize
wire electrical discharge machining (WEDM) process.
Scott and his associations formulized and solved a
multifunctional optimization problem aimed at choosing
the best adjustment of wire electrical discharge machining
(WEDM) machining parameters [1]. Their corresponding
performance was material removal rate and surfacefinished quality. Spedding and Wang optimized this
process with use of nerves network. They considered
surface roughness, value of being wavy in a surface and
material removal speed as outputs [2]. Rozenek and his
associations used a piecework made of composite material
with metal matrix composite and investigated the
variation in feed rate and surface roughness led by
changing the corresponding parameters [3]. Tosun and his
associations used a statistical model for determining
optimal parameters in order to minimize the holes led on
the wire during the process [4]. Tosun and Cogun
conducted a research regarding the effect of machining
parameters on the rate of wire corrosion considering
lessened weight from wire while being machined [5].
In the researchers conducted, the optimal conditions
are led by piecework property and machining conditions
and they cannot be used for other materials or different
manufacturing conditions.
In this study, for the first time, optimal machining
conditions of wire electrical discharge machining
(WEDM) are introduced in one sort of applicable coldwork steal 2601 using comparison of non-dominated
sorting genetic algorithm (NSGA-II) and Tabu search
algorithm both aimed at achieving the appropriate

Abstract-In this research the influence of wire electrical

discharge machining on surface roughness and volumetric
material removal rate is conducted. With use of
experimental result analysis, design of experiments method
and mathematical modeling, the correlation between
corresponding
parameters
and
process
output
characterization are studied. The investigated input
parameters include electrical current, pulse-off time, opencircuit voltage and gap voltage. With use of experimental
results and, subsequently, with exploitation of variance
analysis, importance and effective percentages of each
parameter are studied. In order to find optimal conditions,
outputs extracted from Non-dominated Sorting Genetic and
Tabu search algorithms compared with each other led in
achieving appropriate models. Tabu search algorithm and
Non-dominated Sorting Genetic Algorithm were compared
with each other proving the superiority of Non-dominated
Sorting Genetic Algorithm over Tabu search algorithm in
optimizing machining parameters.
Keywords: Wire Electrical Discharge Machining; Surface
roughness; volumetric material removal rate; optimization;
Tabu search algorithm; Non-dominated Sorting Genetic
Algorithm.

I. INTRODUCTION
These days, with advent of new material exploitation
and era of technological advancement, particularly in
manufacturing industries, the use of machining and
methods of shaping the components has necessarily been
increasing. Meanwhile, methods of machining concerning
wire electrical discharge machining (WEDM) have high
importance and increasingly high applications in
industries. WEDM process is one of EDM machining in
which by creation of alternative spark between tool (wire)
and piece work, machining process of piece work is
implemented. In WEDM process, it is vital to choose best
machining parameters to economize choosing process
whereas WEDM is a nonconventional, applicable and
required machining process with high initial investment.

978-0-7695-4562-2/11 $26.00 2011 IEEE

DOI 10.1109/CIMSim.2011.32

134

conditions of surface roughness (Ra) and volumetric

material removal rate (VMRR).

Where F is the average feed rate [mm/min], l is the

value of length which is cut in t second and VMRR is
volumetric material removal rate [mm3 / min] , Dw is wire
diameter [mm] and H is piecework thickness [mm].
After obtaining the experimental samples, surface
roughness was measured using a mobile roughness
measurement (Mahr Perthometer M2) on each piece work
3 times and subsequently, the average of it was selected.

II. EXPERIMENTAL SETUP AND EQUIPMENT

In this study, the experiments are done using ONA
R250 Series 5-axis CNC Wire EDM on a
piecework ,made of cold-work steel 2601 with thickness
of 30 mm. Chemical synthesis of this steel is
X165CrMoV12. For machining, brass wire (Cu Zn37)
without cover with diameter of 0.25 [mm] and yield
strength of 900 [MPA] is used.
In this experimental analysis, sections with the length
of 20 [mm] (to depth of piecework thickness) are made.
Since the exact amount of machining period was recorded
using chronometer, therefore, this value of time is used
for assessing average feed rate and subsequently, for
volumetric material removal rate according to Eq.1 and
Eq. 2, respectively.
F=

60 l
t

III. EXPERIMENTAL DESIGN

The experiments were planed based on Taguchis
orthogonal array for the design of experiments (DOE). It
can help to reduce the number of experiments. In addition,
four cutting parameters were chosen including electrical
current, pulse-off time, open-circuit voltage and gap
voltage. The three level tests for each factor was used
whereas the considered factors are multi-level variables
whose outcome effects are nonlinear related rather than
linearly. The machining parameters which were used and
their levels are presented in Table 1 and the experimental
results are presented in Table 2.

(1)

VMRR = F Dw H

(2)
TABLE 1.
MACHINING PARAMETERS AND THEIR LEVELS
Levels

Control
Parameters

Unit

Symbol

Current

[A]

Pulse-Off Time

[S]

Toff

14

10

22

Volt
Servo

[volt]
[volt]

V
S

140
28

130
32

110
30

120
26

1
11

2
10

3
9

4
12

TABLE 2.
EXPERIMENTAL RESULTS
S.No.

I
[A]

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

11
11
11
11
10
10
10
10
9
9
9
9
12
12
12
12

Toff
[S]
14
10
8
22
14
10
8
22
14
10
8
22
14
10
8
22

VMRR

V
[volt]

S
[volt]

Ra
[S]

[mm3 / min]

140
130
110
120
130
140
120
110
110
120
140
130
120
110
130
140

28
32
30
26
30
26
28
32
26
30
32
28
32
28
26
30

3.169
3.410
3.229
2.707
3.018
3.035
3.046
2.754
2.421
2.845
3.270
2.575
3.462
3.293
3.528
3.408

12.555
13.785
12.645
7.995
11.108
12.255
10.913
6.698
7.725
8.205
9.593
6.000
11.033
12.435
14.288
8.250

135

for surface smoothness are taken into account. For both of

them, the correlation coefficients became 99.99%
representing appropriate fitting between these models and
experimental data. The results are shown in Table.3.
The normal probability plots for surface roughness and
volumetric material removal rate are illustrated in fig.1. It
is noticeable that residuals fall on a straight line. It
basically shows that the errors are dispersed and the
regression model completely matches the observed values.
Table.4 and Table.5 show that test results are valid.
Predicted machining factors performance was compared
with the actual machining performance and, subsequently,
a good agreement was made. Since the amount of errors
was proved to be acceptable, so these models can be
selected as the best ones and use them in optimization
level.

IV.PROCESS MODELLING AND ANALYSIS OF

VARIANCE
In this study, regression method was used to determine
the relationship between input and output variables of
cold-work steel 2601 using WEDM. For modeling the
process, different types of exponential and second-order
mathematical functions over 14 sets of outputs acquired
from experimental results were fitted. Subsequently, these
models were modified using statistical method of stepwise
elimination in the Minitab environment and with
confidence level of 95%.
Eventually, equations
coefficients for volumetric material removal rate and
surface roughness were measured.
Correlation coefficients calculated for each of the
equations are used to choose the model. Meanwhile,
model of second-order polynomial equations for
volumetric material removal rate and exponential model

TABLE 3.
FITTING EQUATIONS WITH THEIR CORRELATION PERCENTAGE
Respond Value

Model Type

Volumetric

Second-order

Material Removal

polynomial

Rate

equation

Surface
Roughness

Correlation

Fitting Equation

VMRR=15.8 + 22.9 I - 9.19 S - 0.742

(%)

0.0319

Toff2

+ 0.19

V + 0.00188

Toff

S - 0.00220VS

I2-

S2-

0.1I

Toff

99.99

0.00599 I V - 0.143 I S + 0.0132

Ra=exp (2.22-0.120

Toff

Toff

+ 0.523 I- 0.259S- 0.00519 I + 0.00602 S - 0.000677 I

99.99

Exponential
V-0.00949 I S + 0.000606 Toff V + 0.00125 Toff S + 0.000078 V S )

TABLE 4.
RESULTS OF CONFORMATION TEST FOR Ra
I
[A]

Toff

10

12

Run

V
[volt]

S
[volt]

Results of Model

Results of Experiments

Error (%)

120

28

2.98

3.05

-2.3

14

120

32

3.33

3.46

-3.76

[S]

TABLE 5.
RESULTS OF CONFORMATION TEST FOR VMRR
I
[A]

Toff

10

12

Run

V
[volt]

S
[volt]

Results of Model

Results of Experiments

Error (%)

120

28

10.67

10.91

-2.2

14

120

32

12.21

11.03

9.66

[S]

136

(a)

(b)

Fig.1. Normal plot of residuals: (a). Normal plot of residuals for average surface roughness (b). Normal plot of residuals for volumetric material removal rate

After choosing the appropriate models for

investigation of each input parameter effect on outputs
using variance analysis method, distribution percentage
for each of parameters was calculated considering Table.6
and Table.7. According to these calculations, the most
effective parameters concerning surface roughness (Ra)
and volumetric material removal rate (VMRR) are
demonstrated to be electrical current and pulse-off time,
respectively. Moreover, gap voltage doesnt have any
influence on volumetric material removal rate (VMRR)
statistically.

V. MULTI-OBJECTIVE OPTIMIZATION
In this study, two objectives are put into consideration,
volumetric material removal rate (VMRR) and surface
roughness (Ra). It is noticed that if VMRR rises, Ra also
increases. But our objective is aimed at maximizing
VMRR and minimizing surface roughness. A single
optimal solution does not help to achieve our goals as our
purpose, since these objectives are opposing in nature.
Choice of VMRR and surface roughness is also dependent
on user and environment of the problem. Optimizing both
of the output parameters requires multi-objective
optimization [6].

TABLE 6.
RESULTS OF VARIANCE ANALYSIS ON VOLUMETRIC MATERIAL REMOVAL RATE
Factors

d.f.

Sum of Squares

Variance

Distribution Percentage

37.52164687

12.50721562

149.92

37.13

Toff

55.13897813

18.37965938

220.31

54.56

7.51227188

2.50409062

31.01

7.43

0.62510625

0.20836875

2.5

0.08

Error

0.2502844

0.0834281

0.8

Total

15

101.0482875

100

TABLE 7.
RESULTS OF VARIANCE ANALYSIS ON SURFACE ROUGHNESS
Factors

d.f.

sum of squares

Variance

Distribution percentage

0.89816618

0.29938873

23538.8

53.79

Toff

0.36541701

0.12180507

9576.7

21.88

0.20304333

0.06768111

5321.28

12.16

0.20300948

0.06766983

5320.39

12.15

Error

0.00003816

0.00001272

0.02

Total

15

1.66967416

100

137

VI. NONE-DOMINATED SORTING GENETIC

ALGORITHM
One of the powerful and comprehensive algorithms is
what was introduced by Srinivas and Deb [7] named as
the non-dominated sorting genetic algorithm. It deals with
a possible solution regarding a population and, therefore,
it can have some applications in problems of multiobjective optimizations. It leads to have a number of
simultaneous solutions. Despite, this algorithm is fast,
but it has been either a controversial method or opposed
due to have some difficulties and complexities when it
comes to computational approach. The elitism is also
disregarded in this method. The selection operator differs
from simple genetic algorithm (SGA). Crowded
comparison is the operator in which selections can be
achieved considering ranking and crowding distance. The
solution of initially parent population is checked with
other solutions and eventually, put into consideration to
make aware of solutions validation. They must satisfy
rules given below [8]:

Obj.1[i] f Obj.1[ j ] and Obj.2[i ] Obj.2[ j ] ,Or

(3)

Obj.1[i] Obj.1[ j ] and Obj.2[i] f Obj.2[ j], i j (4)

Where, chromosome numbers can be shown as i and j,
respectively. Subsequently, it can be noticeable that the
selected solution is validated by rules introduced in Role.3
and Role.4 and makes it be marked as dominated. If the
rule doesnt satisfy the corresponding equations above, it
will be marked as non-dominated. The corresponding
process must continues until all solution selected are
ranked. Fitness which is as equal as its non-dominated
level assigns to each solution.
There is no result to demonstrate none of the solutions
is better compared with other solutions. They are
considered as part of a special rank or the non-dominated
level. The crowding distance is considered to be as an
average distance between two points on both sides of
selected solution point along each objectives function.
Subsequently, each objective functions boundary
solution with the largest and smallest values is assigned as
an infinity value in this step.
The algorithm flowchart is illustrated in Fig.2. For
solving optimization problem using GA, fitness value is
required. It connects the objective with decision variable.
MATLAB codes have been developed in order to obtain
the optimal output results based on none-dominated
sorting genetic algorithm (NSGA-II).

Fig.2. Flow chart for the NSGA-II algorithm [8]

small possible changes in current solutions values. In

each iteration, after creation of each neighborhood, the
values of purpose functions are calculated and,
subsequently, compared with each other. In this step, the
movement to the best solution in Tabu search is

VII.TABU SEARCH ALGORITHM

Tabu search algorithm (Ts) was first proposed by
Glover in the end of 80th decade [9]. The definition of
neighborhood here is the entire solutions obtained via the

138

implemented among current neighborhood solutions. This

process continues until some stopping criterion has been
satisfied. After each movement, the previous solution lies
on Tabu list. Tabu list is apparently characteristics of
Tabu search containing number of previous accepted
solutions. Movement to those solutions is not possible in
current iteration. Tabu search highly avoids creation of
close loop and convergence of algorithm to local
optimizations and, subsequently, provides possibilities of
searching in more space. Size of Tabu search algorithm
depends on properties and dimensions of problem and is
usually determined experimentally [10-12].
In this study, suggested algorithm is coded using
MATLAB scripts and uses this software for
corresponding problem optimization. The propose
function for optimizing two output factors of surface
roughness and volumetric material removal rate can be
stated using Eq.5 and Eq.6, respectively.
In this process, the aim is to maximize Eq.5 and in
contrast, minimize Eq.6. These equations represent the
absolute difference between expected values and
approximated ones (using proposed models in section IV).

machining (W1=1, W2=0), semi polishing (W1=W2=0.5)

and finish polishing (W1=0, W2=1) is illustrated in the
Table.8.
In the next step, a non-dominated sorting genetic
algorithm (NSGA-II) can be exploited for optimization
issues including multi performance of non-linear models.
The individuals are ranked by the use of NSGA-II
concerning dominance. In order to achieve the high
performance, the controlled factors in NSGA-II are
adjusted. These factors are: crossover probability= 0.8,
mutation probability 0.2 and population size 100. It was
shown that for better convergence and optimal solution
distribution, above controlled factor must be produced.
The 100 generations were generated to acquire the true
optimal solution. The none-dominated set obtained over
the entire optimization is shown in Fig.3.
For comparison of two mentioned algorithms used for
surface roughness optimization and volumetric material
removal rate for three levels of rough machining, semipolishing and finish polishing are shown in Table.8.
Considering the experimental results shown in the Table 2,
the parameters of trial number 11 resulted to surface
roughness of 3.27 [m] and VMRR of 9.593 [ mm 3 / min ].
After optimizing machining parameters through
NSGA-II and Tabu search algorithm, considering NSGAII the value of surface roughness fell to 3.22 [m] and
volumetric material removal rate , in contrast, soared to
14.21 [ mm 3 / min ]. However, regarding Tabu search
algorithm these mentioned values fell to 3.22 [m] and
increased to 14.19 [ mm 3 / min ], respectively (Refer to
Table.8, Section semi polishing). Thus, considering the
data given, as open-circuit voltage setting is kept
constant, by changing pulse-off time, electrical current
and gap voltage, it can be observed that higher VMRR
and lower Ra can be achieved which both are more
desirable.
It is noticed that results in two mentioned algorithms
both concerning optimization in the level of finish
polishing were exactly same. However, in the level of
semi polishing and rough machining, better results were
achieved with use of NSGA-II. According to the level of
semi polishing, the reason why use of NSGA-II is better
is that despite both algorithms lead in same values for
surface roughness, but values of 14.21[ mm 3 / min ] and
14.19 [ mm 3 / min ] were attributed to NSGA-II and Tabu
search algorithms , respectively, demonstrating
superiority of NSGA-II over Tabu search algorithm.
Moreover, according to level of rough machining, results
for both algorithms were same concerning VMRR.
However, surface roughness achieved by use of NSGA-II
and Tabu search algorithm resluts in 3.60 [m] and 3.62
[m] which again show superiority of NSGA-II algorithm
over Tabu search algorithm.

SC

SC

F1 F1 ( x )

(5)

F1 ( x )

F2 ( x) F

F2 ( x )

(6)

Where,
  is output variable values (calculated by

models),
 is expected value of output variable (chosen
by operator).
In this process of optimization, in order to indicate the
importance of every output parameter in different levels
of machining process, Eq.7 can be applicable.

Z = w1 F1SC + w2 F2SC

(7)

According to Eq.7, with changing the weighting

coefficients including w1 (coefficient of volumetric
material removal rate) and w2 (coefficient of surface
roughness), the appropriate conditions for outputs can be
achieved.
VIII. DISCUSSION
Machining processes are often followed in two or three
levels of rough machining, finish polishing and semi
polishing. Level of rough machining is usually aimed at
rising machining speed in which surface roughness is not
necessary. On the other hand, in level of semi polishing
and finish polishing, the aim is raising the surface
roughness. According to the forgoing analysis, the Tabu
search algorithm used for three conditions of rough
139

TABLE 8.
OPTIMUM MASHINING PARAMETERS ACQUIRED USING NSGA-II AND TABU SEARCH ALGORITHM
Level of Machining

Rough Machining

Semi-Polishing

Finish Polishing

Algorithm Type

Current (A)

Pulse-Off
Time (S)

Volt (v)

Servo (v)

Ra(m)

VMRR (mm3/min)

NSGA-II

11.94

110

26

3.60

14.80

Tabu Search

12

110

26

3.62

14.80

NSGA-II

10.97

13.22

140

26

3.22

14.21

Tabu Search

11

14

140

26

3.22

14.19

NSGA-II

22

110

26

2.05

3.39

Tabu Search

22

110

26

2.05

3.39

In conclusion, by comparison of Tabu search algorithm

and NSGA-II it was noticed that in spite of the fact both
algorithms have good results in optimization issues, but it
was shown that NSGA-II had slightly superiority over
Tabu search algorithm whereas NAGA-II results were
more satisfactory than Tabu search algorithm in terms of
optimizing machining parameters.
REFERENCES
[1] D.Scott , S.Boyna , K.P.Rajurkar ., "Analysis and optimization of
parameter combinations in WEDM", Int. J. Prod. Res., Vol. 29, pp.
21892207, 1991.
[2] T.A.Spedding, Z.O.Wang , "Parametric optimization and surface
characterization of wire electrical discharge machining process", Int. J.
Precision Eng., Vol.20, pp. 515, 1997.
[3]M.Rozenek.M,J.Kozak,L.Dabrovwki,K.Lubkovwki,
Electrical
discharge
machining
characteristics
of
metal
matrix
composites,J.Mater.Process.Technol.109, pp.367-370, 2001.
[4] N.Tosun, C.Cogun , H.Pihtili , "The effect of cutting parameters on
wire crater sizes in WEDM", int. J . Adv. Manuf. Techonl., Vol. 21, pp.
857-865, 2003.
[5] N.Tosun,C.Cogun, An investigation on wire wear in WEDM,
j.Mater.Process.Technol.1349 (3) , pp. 273-278, 2003.
[6] J.Y. Kao, Y.S. Tarng, A neutral-network approach for the on-line
monitoring of the electrical discharge machining process, J. Mater.
Process. Technol. 69, pp.112119, 1997.
[7] K. Deb, P. Amrit, A. Samir, and Meyarivan, IEEE Trans. Evol.
Comput. 6, 182 , 2002.
[8] M.Debabrata, K. Pa.Surjya, S.Partha, Modeling of electrical
discharge machining process using back propagation neural network and
multi-objective optimization using non-dominating sorting genetic
algorithm-II, J. Mater. Process. Technol. 186, pp.154162, 2007.
[9] F.Glover, Tabu search: Part I, ORSA J Comp, 1989.
[10] A.Hertz A., D. De Werra, "The tabu search metaheurestic: how we
used it", Annual Mathematics in Artificial Intelligence, Vol. 1, pp. 111
121, 1991.
[11] F.Kolahan, M.Liang, "Optimization of hole-making operations: a
tabu-search approach", Euro. J. Operational
Res., Vol. 109, pp. 142- I59, 1998.
[12] F.Kolahan, M.Liang, "An adaptive TS approach to JIT sequencing
with variable processing times and
sequence-dependent setups", Int. J. Machine Tools Manuf., Vol. 40, pp.
17351753, 2000.
[13] K. Palanikumar, B. Latha , V.S.Senthilkumar ,R.Karthikeyan,
Multiple Performance Optimization in Machining of GFRP Composites
by a PCD Tool using Non-dominated Sorting Genetic Algorithm
(NSGA-II),Met. Mater. Int.,Vol.15, No. 2, pp. 249-258, 2009.

Fig 3. Pareto optimal set with use of NSGA-II

IX.CONCLUSION
As follows from forgoing analysis, the study based on
the influence of WEDM on surface roughness and
volumetric material removal rate was carried out. The
nonlinear polynomial models were developed for
volumetric material removal rate and average surface
roughness were used for optimization. In this study, two
multi-objective evolutionary algorithms concerning
efficient methodology including NSGA-II and Tabu
search algorithm were used to optimize machining
parameters in cold-work steel 2601.
The emphasis must be put on providing a preferred
solution for the process engineer in the short period of the
time. The choice of one solution over other ones is
dependent on the requirements of process engineer [13].

140