1-s2.0-S1526612513000959-main

Page 1

T

Emo

Ra

b

c

a

ARRAA

KDMO

1

cdwAiphtat

r

1h

Journal of Manufacturing Processes 15 (2013) 483–494

Contents lists available at ScienceDirect

Journal of Manufacturing Processes

j ourna l h o me page: www.elsev ier .com/ locate /manpro

echnical Paper

xperimental investigation, intelligent modeling andulti-characteristics optimization of dry WEDM process

f Al–SiC metal matrix composite

eza Kashiry Farda, Reza Azar Afzab, Reza Teimouri c,∗

Department of Mechanical Engineering, Islamic Azad University, Kordestan Branch of Science and Research, Sanandaj, IranDepartment of Mechanical Engineering, Malek Ashtar University of Technology, Tehran, IranDepartment of Mechanical Engineering, Babol University of Technology, Babol, Iran

r t i c l e i n f o

rticle history:eceived 5 January 2013eceived in revised form 7 September 2013ccepted 10 September 2013vailable online 6 October 2013

eywords:ry WEDModelingptimization

a b s t r a c t

Dry wire electrical discharge machining (WEDM) is an environmentally friendly modification of theoil WEDM process in which liquid dielectric is replaced by a gaseous medium. In the present work,parametric analysis has been fulfilled while dry WEDM of Al–SiC metal matrix composite. Experimentswere designed and conducted based on L27 Taguchi’s orthogonal array to study the effect of pulse ontime, pulse off time, gap voltage, discharge current, wire tension and wire feed on cutting velocity (CV)and surface roughness (SR). Firstly, a series of exploratory experiments has been conducted to identifyappropriate gas and wire material based on the values of cutting velocity. After selection of best gasand best wire, they were used for later stage of experiments. Analysis of variances (ANOVA) has beenperformed to identify significant factors. In order to correlate relationship between process inputs andresponses, an adaptive neuro-fuzzy inference system (ANFIS) has been employed to predict the processcharacteristics based on experimental observation. At the end, an artificial bee colony (ABC) algorithmhas been associated with ANFIS models to maximize CV and minimize SR, simultaneously. Then theoptimal solutions that obtained through ANFIS-ABC technique have been compared with numbers of

confirmatory experiments. Results indicated that oxygen gas and brass wire guarantee superior cuttingvelocity. Also, according to ANOVA, pulse on time and discharge current were found to have significanteffect on CV and SR. In modeling of CV and SR by ANFIS, it was resulted that the proposed method hassuperiority in prediction of them in the ranges of factors beyond the training condition. Also, associationof ANFIS with ABC can find the optimal combination of process parameters accurately according to the

s.iety o

confirmatory experiment© 2013 The Soc

. Introduction

The Al/SiC metal matrix composite (MMC) is a hard and toughomposite with noticeable wear resistance that satisfies growingemands of material with higher mechanical properties and lowereight. Despite the superior mechanical and thermal properties ofl/SiC MMC, its poor machinability has been the main deterrent to

ts substitution for metal parts. The hard abrasive-reinforcementhase causes rapid tool wear during machining and, consequently,igh machining costs. For machining of this material with conven-

ional machining process the polycrystalline diamond (PCD) toolsre the only tool material that is capable of providing a usefulool life during the machining of Al/SiC MMC [1]. In the case of

∗ Corresponding author. Tel.: +98 9369098670; fax: +98 1125233899.E-mail addresses: Reza [email protected],

eza [email protected] (R. Teimouri).

526-6125/$ – see front matter © 2013 The Society of Manufacturing Engineers. Publishettp://dx.doi.org/10.1016/j.jmapro.2013.09.002

f Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.

conventional machining of Al/SiC several studies have been investi-gated. El-Gallab and Sklad [2–4] conducted vast research on turningof Al/SiC to find optimum tool and process conditions in whichhigher tool life and surface integrity are obtained. Ozben et al.[5] studied on mechanical properties and machinability of Al/SiC.They indicated that variation of the percentage of reinforcementcan affect on its mechanical properties, tool life and surface rough-ness. Due to high wear resistance and hardness of Al/SiC, it cannotbe machined with conventional machining process simply. Hencenon-conventional machining methods can be applied as an eco-nomical and functional solution for machining of Al/SiC. In thiscase electrical discharge machining (EDM) is extensively used formachining of Al/SiC [6–8]. Also, the electro-chemical machiningprocess is an applicable method that can machine this material

independently of its hardness [9].

Wire electrical discharge machining process (WEDM) is apotential electro-thermal process which is useful for machiningsuch difficult-to-cut electrically conductive materials. The material

d by Elsevier Ltd. All rights reserved.

Page 2

484 R.K. Fard et al. / Journal of Manufacturi

rEoEF

wfloratgpgcahtpihvs

aScf[mmaawncmunc

characteristics and implementation of analysis of variances to

Fig. 1. A schematic diagram of wire EDM process.

emoval mechanism of WEDM process is really like to die sinkingDM process with a little difference in tool electrode shape. Insteadf using pre-shaped tool which used most commonly in die sinkingDM, a thin 0.05–0.3 mm diameter wire acts as a tool in WEDM.ig. 1 shows a schematic diagram of WEDM process.

Like a die sinking EDM, in WEDM all sparks occur in a dielectrichich plays important role in the process. In WEDM, the dielectricushed in gap space and its commonly material is liquid oil. Usef mineral oil-based dielectric liquids is the major cause of envi-onmental concerns that associated with the EDM process and itslternatives. Dry WEDM is an environment-friendly modification ofhe oil WEDM process in which the liquid dielectric is replaced by aaseous medium. Dielectric wastes generated during the oil WEDMrocess are very toxic and cannot be recycled. Also, toxic fumes areenerated during machining due to high temperature and chemi-al breakdown of mineral oils. The use of oil as the dielectric fluidlso makes it necessary to take extra precaution to prevent fireazards. Replacing liquid dielectric by gases is an emerging field inhe environment-friendly EDM technology [10–12]. In dry WEDMrocess high velocity gas that is flowing through nuzzles into the

nter-electrode gap can substitute the liquid dielectric. The flow ofigh velocity gas into the gap facilitates removal of debris and pre-ents excessive heating of the wire and workpiece at the dischargepots.

There are numbers of researches that fulfilled parametric studynd developed empirical models on WEDM process. In this case,arkar et al. [13] developed mathematical model through centralomposite design of experiments to model cutting speed and sur-ace finish of TiAl alloy in trim cutting process. El-Taweel et al.14] applied response surface methodology (RSM) for modeling of

achining parameters WEDM of Inconel 601. They modeled volu-etric material removal rate, wire wear rate and surface roughness

ccording to variation of peak current, pulse on time, wire tensionnd water pressure. Spedding and Wang [15] applied the RSM alongith neural network for modeling of cutting speed, surface rough-ess and surface waviness of WEDM process. Also, Saha et al. [16]onducted extensive experiments and used their observation forodeling of WEDM process using soft computing methods. They

sed mathematical model based on regression analysis and neuraletwork with back-propagation learning algorithm for modeling ofutting speed and surface roughness of WC-Co. Results indicated

ng Processes 15 (2013) 483–494

that the neural network crates more precise prediction rather thanregression technique.

Although the neural network has superiority in modeling ofmanufacturing process rather than statistical models and math-ematical equations, but the main weakness of neural network isits dependency on large amount of data for a problem in whichmany inputs are contributed. Also, in the case of manufacturingprocesses with complex behavior the neural network cannot pre-dict the process characteristics as well. It means that for a processwith complex behavior some linguistic terms are needed to providea precise prediction. Thus, application of fuzzy logic can be ben-eficial for modeling of complex behavior. But construction of anappropriate fuzzy membership function and fuzzy rules is reallydifficult and time consuming job. Thus, for modeling of a complexprocess with small amount of data in a short time, a method withboth concepts of neural network and fuzzy logic is needed. There-fore, an adaptive neuro-fuzzy inference system (ANFIS) is proposedas a hybrid predictive approach that uses both meanings of neuralnetwork and fuzzy logic for modeling of complex processes.

According to above explanation, the advantageous of ANFISmodel is to predict the performance of a given process in factorsranges outside the predefined ranges for the training conditions.This is due to existing linguistic terms of fuzzy theory in the layersof ANFIS. It means that this network can act as an artificial humanbrain and it can estimate the process performance in wider rangesof training condition. This claim will be proved later in the Sec-tion 4.3. The literature [17] is a relative work which was writtenby corresponding author about superiority of ANIFS in modelingmechanical properties of friction stir welding process.

According to the surveyed literatures, there is not a certain pub-lication that uses the ANFIS for modeling of a complex process suchas dry wire EDM process. Hence, application of this method in thepresent work is quite novel.

Recently, using non-traditional optimization algorithm findsits application in single-response and multi-responses of manu-facturing processes. Well-known algorithms such as GA, SA, PSO,and ICA have been widely used for optimization of processeswith multiple quality characteristics. Artificial bee colony (ABC)is a swarm based intelligent optimization algorithm which hasbeen inspired by foraging behavior of honey bees in the caseof finding food. There are not many publications which usedABC for optimization of manufacturing processes. Samanta andChakraborty [18] used the ABC algorithm for parametric opti-mization of some non-traditional machining process includingelectro chemical machining, electrochemical-discharge machiningand electrochemical micro-machining. Teimouri and Baseri [19]applied ABC for simultaneous optimization of material removal rateand surface roughness in dry EDM process. Also, they used ABC forminimizing prediction error of mechanical properties of frictionstir welding process [20]. According to the reviewed works, it canbe inferred that application of ABC for optimization of dry WEDMprocess is quite novel.

The present work consists of experimental investigation, sim-ulation and optimization of WEDM process while using gaseousdielectric. Objectives this work could be summarized as follows:

(I) Implementation of WEDM process by using gaseous medium(II) Selection of the optimum gas based on values of cutting veloc-

ity and surface roughness(III) Selection of appropriate wire material based cutting velocity

and surface roughness(IV) Investigation on effects of process parameters on dry WEDM

find significant factors(V) Simulation of cutting velocity and surface roughness by ANFIS

technique and verification of the results by confirmatory

Page 3

R.K. Fard et al. / Journal of Manufacturing Processes 15 (2013) 483–494 485

Table 1main properties of workpiece.

Nominal Composition Grain size Hardness (HV) Density (g/cm3) Thermal conductivity(W/m K)

Compressivestrength (MPa)

Modules ofelasticity (GPa)

75%Al 6061–25%SiC Fine 1250–1650 3 3100 4200 206

Table 2Process parameter levels for various energy conditions to find best gas and wire.

Condition Pulse on time (�s) Pulse off time (�s) Discharge current (A) Gap voltage (V)

70 60150 40230 20

(

2

2

(mIaAb2opAtTct

ptTrTpt(tc

masefs

2

c

Table 3process parameters and their levels.

Process parameter Unit Level 1 Level 2 Level 3

Pulse on time (Ton) �s 106 116 126Pulse off time (Toff) �s 40 50 60Gap voltage (VG) V 20 40 60Discharge current (I ) A 70 150 230

1 (Low energy) 106 60

2 (Middle energy) 116 50

3 (High energy) 126 40

experiments in the ranges of factors beyond the predefinedranges of factors in training.

(VI) Multi-characteristics optimization of process by ABC toachieve higher cutting velocity and lower surface roughnesssimultaneously.

VII) Validation of optimal results by conducting verification exper-iments.

. Experimentations

.1. Experimental setup and materials

A series of experiments were carried out on WED machineTEHRAN EKRAM CNC AW-500) with Iso-pulse spark generator and

aximum discharge current of 300 ampere manufactured in Iran.nstead of using oil liquid dielectric, the machine was modified by

gas injector to enter the high flow gas with inlet pressure of 2 bar.n air compressor provides gas flow with high pressure, and theny using of a regulator valve, the high pressure flow is broken to

bar flow. For providing oxygen and nitrogen medium, storages ofxygen and nitrogen with high pressure were used and their highressure controlled by regulator valve. The workpiece material isl/SiC metal matrix composite with high hardness and toughness in

he shape of plate with dimensions of 125 mm × 100 mm × 24 mm.he properties of workpiece were presented in Table 1. Also theopper and brass wires with 0.1 mm diameter have been used asool which was made vertical with the help of vertical block.

The faces of each specimen were ground to generate accuratearallel sitting on the machine table. Then a reference point onhe work piece was set for setting work co-ordinate system (WCS).he programming was done with the reference to the WCS. Theeference point was defined by the ground edges of the work piece.he program was made for cutting operation of the work piece and arofile of 5 mm × 5 mm square was cut. Based on the cutting profile,he cutting velocity can be calculated by ratio of machining lengthe.g. 23 mm) on machining time. The WED machine that used inhe present work can monitor the machining time and display theutting velocity on machine indicator.

In order to calculate the surface roughness of WEDMed speci-ens Mahr Marsurf surface profile meter with cut off of 0.5 mm

nd sampling length of 5 mm was used. The value of Ra has beenelected as a main index for measuring the surface roughness. Forach specimen measuring of surface roughness has been repeatedor five times to reduce the measuring error. Then average of mea-urements was reported as surface roughness.

.2. Experimental plan

In the present work, firstly a series of experiment have beenarried out to determine optimum gas and wire based on value

d

Wire feed (WF) m/min 4 8 12Wire tension (WT) g 4 8 12

of cutting velocity. For this purpose, numbers of 18 experimentaltests are carried out under different conditions of discharge ener-gies (various pulse on time, pulse off time, discharge current andgap voltage), various gaseous dielectrics (air, oxygen and nitrogen)and various wires (copper and brass). Table 2 clarifies various con-ditions of discharge energy. Due to lower cutting velocity of dryWEDM process rather than conventional WEDM [10–12], gas andwire that cause higher cutting velocity are selected as a major gasand major tool for later stage of experiment.

As it is known, the wire EDM process is an unconventionalmachining process in which many factors affect on its perfor-mance. Hence, in the next stage, pulse on time, pulse off time,gap voltage, discharge current, wire feed and wire tension areconsidered as the predominant factors of WEDM process. Due towide range of factors, it was decided to use 6 factors-three lev-els L27 Taguchi orthogonal arrays design to optimize number ofexperiments. Selection of higher levels for each factors make theexperimental times longer and it does not satisfy economical aspectof the work. Also, according to the literatures and our laboratoryexperience this design is useful for investigating effect of WEDMprocess factors on cutting velocity and surface roughness.

A MINIITAB 17 package has been utilized to design and analysisof experiments. Table 3 presents process parameters and their lev-els. Also, design matrix and experimental results are presented inTable 4.

3. Methodologies

3.1. Adaptive neuro-fuzzy inference system (ANFIS)

An adaptive neuro-fuzzy inference system is a hybrid predictivemodel which uses both of neural network and fuzzy logic to gen-erate mapping relationship between inputs and outputs [17]. Thestructure of this model consists of five layers which each layer isconstructed by several nodes. Such as a neural network, in an ANFISstructure the inputs of each layer are gained by the nodes from per-

vious layer. Fig. 2 describes an ANFIS structure. It can be inferredfrom Fig. 5 that the network includes m inputs (X1, . . ., Xm), whicheach one consists of n membership functions (MFs). Moreover, alayer with R fuzzy rules and also an output layer are contributed to

Page 4

486 R.K. Fard et al. / Journal of Manufacturing Processes 15 (2013) 483–494

Table 4L27 design matrix and performance measured values.

No Process factors Responses

Ton (�s) Toff (�s) VG (V) Id (A) WF (m/min) WT (g) CV (mm/min) SR (�m)

1 106 40 20 70 4 4 0.68 1.3852 106 40 40 150 8 8 0.71 1.2553 106 40 60 230 12 12 0.505 1.1754 106 50 20 150 8 12 0.48 1.355 106 50 40 230 12 4 0.405 1.2656 106 50 60 70 4 8 0.2 1.1057 106 60 20 230 12 8 0.315 1.358 106 60 40 70 4 12 0.165 1.149 106 60 60 150 8 4 0.165 1.08

10 116 40 20 150 12 8 2.405 2.5711 116 40 40 230 4 12 2.205 2.3912 116 40 60 70 8 4 0.75 1.43513 116 50 20 230 4 4 1.54 2.52514 116 50 40 70 8 8 0.58 1.3815 116 50 60 150 12 12 0.66 1.7216 116 60 20 70 8 12 0.47 1.6317 116 60 40 150 12 4 0.625 1.8518 116 60 60 230 4 8 0.5 1.62519 126 40 20 230 8 12 3.4 2.920 126 40 40 70 12 4 1.26 1.8521 126 40 60 150 4 8 2.2 2.6622 126 50 20 70 12 8 0.91 2.0823 126 50 40 150 4 12 1.66 2.63524 126 50 60 230 8 4 1.54 2.445

ccMnt[

mpfTFi

25 126 60 20 150

26 126 60 40 230

27 126 60 60 70

onstruction of this model. Number of nodes in first layer can bealculated by product of m as number of inputs and n as numberFs (N = m·n). Number of nodes in other layers (layer 2–4) relates to

umber of fuzzy rules (R). For further information about implemen-ation of the ANFIS, the interested readers can read the reference17].

In the present work the ANFIS technique is used to correlateapping relationship between process inputs (e.g. pulse on time,

ulse off time, gap voltage, discharge current, wire tension and wire

eed) and main outputs (cutting velocity and surface roughness).hus, for each output a separate ANFIS structure can be defined.or example for cutting velocity the first layer of ANFIS structures input layer that contains six nodes (for six inputs). And the last

Fig. 2. Basic structure of a

4 4 1.34 2.8558 8 1.345 2.56

12 12 0.29 1.295

layer (output layer) has one node that represents values of cuttingvelocity.

3.2. Artificial bee colony algorithm

An artificial bee colony (ABC) algorithm was introduced byKaraboga [21] in 2005, for optimizing numerical problems. Itwas inspired by the intelligent foraging behavior of honey bees.The model consists of three essential components: employed and

unemployed foraging bees, and food sources. The first two com-ponents, employed and unemployed foraging bees, search forrich food sources, which is the third component, close to theirhive.

n ANFIS model [17].

Page 5

acturi

rtp

ewad

R.K. Fard et al. / Journal of Manuf

In ABC, a colony of artificial forager bees (agents) search forich artificial food sources. To apply ABC, the considered optimiza-ion problem is first converted to the problem of finding the bestarameter vector which minimizes an objective function.

In ABC, the colony of artificial bees contains three groups of bees:

mployed bees associated with specific food sources, onlooker beesatching the dance of employed bees within the hive to choose

food source, and scout bees searching for food sources ran-omly. Both onlookers and scouts are also called unemployed bees.

Fig. 3. Flowchart of artificial be

ng Processes 15 (2013) 483–494 487

Initially, all food source positions are discovered by scout bees.Thereafter, the nectar of food sources are exploited by employedbees and onlooker bees, and this continual exploitation will ulti-mately cause them to become exhausted. Then, the employed beewhich was exploiting the exhausted food source becomes a scout

bee in search of further food sources once again. In ABC, the posi-tion of a food source represents a possible solution to the problemand the nectar amount of a food source corresponds to the qual-ity (fitness) of the associated solution. The number of employed

e colony algorithm [19].

Page 6

4 acturing Processes 15 (2013) 483–494

be

sop

XSpaiMe

sfad

V

wpoc

m

f

w

osoisfinwi

P

caaob

sttotfaehf

p

0

0.5

1

1.5

2

2.5

3

0 4

CV (m

m/m

in)

Energy condi�on

Copper wir e and Air

Copper wire and Nitrogen

Copper wir e and Oxygen

Brass wire and Air

Brass wire and Nitrogen

Brass wire and Oxygen

Low Medi um High

88 R.K. Fard et al. / Journal of Manuf

ees is equal to the number of food sources (solutions) since eachmployed bee is associated with one and only one food source.

As mentioned above, the artificial bee colony algorithm con-ists of four main phases, initialize phase, employed bees phase,nlooker bees phase and scout bees phase. The clarification of eachhase is defined as follow.

Initialize phase: All the vectors of the population of food sources,ms are initialized by scout bees and control parameters are set.ince each food source Xm is a solution vector to the optimizationroblem, each Xm vector holds n variables (Xmi, i = 1, . . ., n) whichre to be optimized so as to minimize the objective function. Afternitialization, the solutions is subjected to repeated cycles C = 1, . . .,

CN (maximum cycle number). This is for the search process of themployed bees, onlooker bees and scout bees.

Employed bees phase: Employed bees search for new foodources (Vm) having more nectar within the neighborhood of theood source (Xm) in their memory. They find a neighbor food sourcend then evaluate its profitability (fitness). For example, they canetermine a neighbor food source (Vm) using the formula given by:

mi = Xmi + ˚mi(Xmi − Xki) (1)

here Xk is the randomly selected food source, i is randomly chosenarameter index and ˚mi is a random number within the rangef [−1,1]. After producing the new food source (Vm) its fitness isalculated and a greedy selection is applied between Vm and Xm.

The fitness value of the solution fitm(Xm) might be calculated forinimization problems using the following formula:

itm(Xm) ={

fm(Xm) if fm≥0

abs(fm(Xm)) if fm < 0(2)

here fm(Xm) is the objective function value of solution Xm.Onlooker bees phase: Unemployed bees consist of two groups

f bees: onlooker bees and scouts. Employed bees share their foodource information with onlooker bees waiting in the hive and thennlooker bees probabilistically choose their food sources depend-ng on this information. In ABC, an onlooker bee chooses a foodource depending on the probability values calculated using thetness values provided by employed bees. For this purpose, a fit-ess based selection technique can be used, such as the rouletteheel selection method. The probability value Pm with which Xm

s chosen by an onlooker bee can be calculated by:

m = fitm(Xm)∑SNm=1fitm(Xm)

(3)

After a food source Xm for an onlooker bee is probabilisticallyhosen, a neighborhood source Vm is determined by using Eq. (1),nd its fitness value is computed. As in the employed bees phase,

greedy selection is applied between Vm and Xm. Hence, morenlookers are recruited to richer sources and positive feedbackehavior appears.

Scout bees phase: The unemployed bees that choose their foodources randomly are called scouts. Employed bees whose solu-ions cannot be improved through a predetermined number ofrials, specified by the user of the ABC algorithm and called “limit”r “abandonment criteria” herein, become scouts and their solu-ions are abandoned. Then, the converted scouts start to searchor new solutions, randomly. For instance, if solution Xm has beenbandoned, the new solution discovered by the scout that was themployed bee of Xm. Hence those sources which are initially poor or

ave been made poor by exploitation are abandoned and negative

eedback behavior arises to balance the positive feedback.The flowchart of artificial bee colony algorithm including main

hases is visible in Fig. 3.

Fig. 4. Effects of various types of gases and wires on cutting velocity under differentconditions of discharge energy.

4. Results and discussions

This section describes results that were obtained from experi-ments along with logical discussions based on process behaviors.The section involves 4 subsections. At first, results of exploratoryexperiment are described to find best gas and wire which cause tohigher cutting velocity. Second subsection is allocated to effect ofprocess parameters on performance measures. Modeling of cuttingvelocity, surface roughness and dimensional shift is described onthird subsection. And finally, the fourth subsection defines step bystep optimization of dry WEDM process by ABC algorithm.

4.1. Selection of appropriate gas and wire (exploratoryexperiments)

Fig. 4 shows effect of various gases and various wires on cut-ting velocity under various conditions of discharge energy. It couldbe inferred from the figure that in all three energy condition thebrass wire leads to higher cutting velocity than copper. This is dueto higher specific resistance of brass rather than copper whichmay lead to increasing in the spark intensity and electrode gap[22]. Also Fig. 5 depicts that injecting oxygen gas in machining gapcauses higher cutting velocity than air and nitrogen. While processuses oxygen gas, in addition of discharges, oxidation helps metalremoval process and improves cutting velocity. In other word, bothmelting and chemical reactions contributed in metal removal pro-cess. Also, oxygen gas causes expulsion in interior electrode gapand has a positive effect on cutting velocity. It is shown in Fig. 4 thatthe air outperforms nitrogen in cutting velocity. The main reasonfor this phenomenon is contribution of oxygen gas in air. Accord-ing to above discussions, the oxygen gas and brass wire have beenselected as major dielectric and tool for later stage of experiments.

4.2. Effect of process parameters on performance measures

4.2.1. Analysis of cutting velocityFig. 5 indicates effect of process parameters on cutting velocity.

Also, Table 5 presents ANOVA results of cutting velocity. It could beinferred from the figure and table that pulse on time has greatesteffect on cutting velocity by contribution of 11.7%. Also, the P-valuerepresents that this factor is really significant. Fig. 5(a) indicatesthat higher pulse on time leads to increase in cutting velocity dueto higher thermal energy which transfers from wire to workpiece.

Fig. 5(b) depicts that increasing in pulse off time leads lowercutting velocity due to increasing in non-cutting time. Also, Table 5presents that pulse off time has significant effect on cutting velocitydue to its 0.00 P-value.

Fig. 5(c) indicates that the gap set voltage resulting in lower

cutting velocity. Contribution of gap voltage is about 2.44% and theP-value of 0.033 shows that it is not as significant as pulse on timeand pulse off time.

Page 7

R.K. Fard et al. / Journal of Manufacturing Processes 15 (2013) 483–494 489

Fig. 5. Effect of process parameters on cutting velocity.

Table 5An ANOVA results to explore effects of process parameters on cutting velocity.

Factor Degree of freedom Sum of square Mean of square Contribution (%) P-value Significance

Pulse on time 2 1.9943 0.9974 11.7 <0.01 **

Pulse off time 2 1.5374 0.7687 9.02 <0.01 **

Gap voltage 2 0.4154 0.2074 2.44 0.033 *

Discharge current 2 0.8431 0.4215 5 <0.01 **

Wire feed 2 0.1681 0.09305 1.092 0.08 –Wire tension 2 0.0436 0.0218 0.26 0.1 –Total 26 5.0019 – –Error 14 – – –

Fcpv

Ti

4

Fe

gAnle

ltsa

o

* Significant.** Very significant.

Effect of discharge current on cutting velocity is visible inig. 5(d) which indicates increasing of this factor resulting higherutting velocity due to transferring more thermal energy. Table 5roves that the discharge current has a significant effect on cuttingelocity.

According to graphical results of Fig. 5 and ANOVA results ofable 5, it can be inferred that wire feed and wire tension arensignificant in term of cutting velocity.

.2.2. Analysis of surface roughnessEffect of process parameters on surface roughness is visible in

ig. 6. Also, Table 6 presents ANOVA results to estimate of effect ofach factor on surface roughness.

According to Table 6 it can be inferred that pulse on time hasreatest effect on surface roughness with contribution of 19.6%.lso the P-value of <0.01 indicates that this factor is really sig-ificant. Moreover, Fig. 6 demonstrates that higher pulse on time

eads to higher surface roughness due to transferring more thermalnergy that induces deeper discharge craters on workpiece surface.

As shown in Fig. 6, it can be inferred that higher pulse off timeeads to lower surface roughness due to increasing non-cuttingime. Moreover, Table 6 presents that the pulse off time is not as

ignificant as pulse on time due to its higher P-value (e.g. 0.045)nd lower contribution (e.g. 1.95%).

Gap set voltage is another factor that has influence of 3.17%n surface roughness. Also, the P-value of this factor is <0.01that

proves it is significant. Increasing in gap voltage resulting in higherelectrostatic force and leads to winding of wire during dischargeprocess. So, lower surface roughness is obtainable while gap voltageincreases.

Table 6 presents that discharge current is another significantfactor after pulse on time with contribution of 5.77% and P-valueof <0.01. According to Fig. 6 it could be inferred that higher dis-charge current leads to coarser surface. By increasing dischargecurrent more thermal energy transfers from wire to workpiece anddeeper discharge carter are formed. Hence, poor surface qualityexists while process works at high discharge current.

Wire feed is another factor that has contribution of 2% and P-value of 0.05 on surface roughness. Fig. 6 depicts that increasing inwire feed has negative influence on surface roughness. When wirefeed increases the wire is recovered and renewed during process. Sonon-worn wire can increase spark efficiency and produces deeperdischarge crater. It is for this reason that higher wire feed resultingrougher surface.

Both Fig. 6 and Table 6 prove that the wire tension does not havesignificant effect on surface roughness due to its highest P-valueand lowest percentage of contribution.

4.3. Modeling of cutting velocity and surface roughness using

ANFIS

In order to predict cutting velocity and surface roughness ofthe process and adaptive neuro-fuzzy inference system has been

Page 8

490 R.K. Fard et al. / Journal of Manufacturing Processes 15 (2013) 483–494

Fig. 6. Effect of process parameters on surface roughness.

Table 6An ANOVA results to explore effects of process parameters on surface roughness.

Factor Degree of freedom Sum of square Mean of square Contribution (%) P-value Significance

Pulse on time 2 1.9385 0.9692 19.6 <0.01 **

Pulse off time 2 0.092 0.0463 1.95 0.045 *

Gap voltage 2 0.3138 0.1569 3.17 <0.01 **

Discharge current 2 0.5712 0.2856 5.77 <0.01 **

Wire feed 2 0.1977 0.0988 2 0.05 *

Wire tension 2 0.0042 0.0021 0.004 2.05 –Total 26 3.1174 – –Error 14 – – –

epst

pimw9oc

drswbaeeRtTmnm

Table 7Values of RMSE in testing of cutting velocity and surface roughness.

Type of membershipfunctions

RMSEs of cuttingvelocity

RSMEs of surfaceroughness

Triangular 0.7985 1.0234Trapezoid 0.9122 1.412Generalized bell 0.4563 0.3421

* Significant.** Very significant.

mployed for CV and SR separately. This technique crates a map-ing relationship between process inputs and cutting velocity orurface roughness r. For this purpose a MATLAB R17 package (ANFISoolbox) has been utilized.

Prediction of cutting velocity and surface roughness of therocess by ANIFS consists of two main stages, training and test-

ng. Hence, number of 27 data sets which were cited in designatrix (Table 4), have been selected for training of ANFIS net-ork. Then the trained network has been tested by other remaining

data sets which not contributed in training. Also, the rangesf inputs for these 9 data sets are beyond the ranges of trainingondition.

There are some important factors which are contributed to pro-uce an accurate prediction by ANFIS; they are type of fuzzy basedule, number of membership functions (MFs) and type of member-hip functions. In this paper a first order TSK type fuzzy based ruleas used for creation of predictive model. Then the various num-

ers of membership functions have been tried. In order to comparell existing networks and select most accurate one, the value ofrror goal (RMSE) was set 0.01 and the iteration number was 200pochs. It means that the training epochs are continued until theMSE fell below 0.01 or the epochs go up 200. As the RMSE cri-erion for all networks is the same, their actions are comparable.

hen their testing performances were compared and the optimizedodel is selected based on its predictive accuracy in response to

ew input data in the testing phase when compared with experi-ental values.

Gaussian 0.5671 0.6112

The bold values refers to the lowest RMSE that means the most accurate prediction.

By testing of various structures of ANFIS model for each response(cutting velocity and surface), it was obtained that structures withnumbers of 12 membership functions (2 MFs for each input or 2-2-2-2-2-2 topography) has the lowest values of RMSE. Selection ofnetwork with larger number of MFs leaded to over-fitting and didnot generate desired value of RMSE. Another factor which is influ-ential in accuracy of ANFIS model is type of membership functions.In this work various types of MFs namely triangular, trapezoid,generalized bell and Gaussian have been practiced.

Table 7 presents RMSEs of ANFIS models in testing for cuttingvelocity and surface roughness. It could be seen from the table thatfour types of membership functions under 2-2-2-2-2-2 structurehas been trained and their RMSEs were calculated. Results indicated

that for both cutting velocity and surface roughness the generalizedbell membership function leads to lowest values of RMSE.

Table 8 presents the comparison of measured values of cut-ting velocity and surface roughness which were obtained through

Page 9

R.K. Fard et al. / Journal of Manufacturing Processes 15 (2013) 483–494 491

Table 8Comparison between measured and predicted values of CV and SR for testing data.

No Inputs with the ranges beyond training condition CV (mm/min) SR (�m)

Ton (�s) Toff (�s) VG (V) Id (A) WT (g) WF (mm/min) Measured Predicted Error (%) Measured Predicted Error (%)

1 90 70 80 50 2 2 0.18 0.1682 6.56 0.93 1.0010 72 140 30 10 250 16 16 4.58 4.42 3.49 4.22 4.2513 0.7423 70 90 100 50 2 18 0.38 0.4000 5.26 1.2 1.1288 5.934 140 70 100 250 16 14 3.61 3.7234 3.05 2.34 2.2987 1.715 70 70 80 250 14 16 2.39 2.5247 5.44 2.11 2.1916 4.126 90 30 100 50 2 2 3.13 2.9811 4.79 2.43 2.5122 3.29

erfadetvdec

4r

rmsi

iaccmd

7 140 70 10 50 16 2

8 70 30 80 250 2 14

9 140 90 100 50 14 16

xperiments, with predicted values of cutting velocity and surfaceoughness which were obtained through ANFIS model. It is seenrom this table that although the ranges of inputs of testing datare beyond the ranges of training condition, but the ANFIS can pre-ict the cutting velocity and surface roughness with an acceptablerror. Also, Fig. 7(a) and (b) shows these results graphically for bet-er understanding of agreement between measured and predictedalues of CV and SR. It can be inferred from Table 7 and Fig. 7 thateveloped ANFIS models has superiority in prediction of CV and SRven in the ranges of inputs beyond predefined ranges of trainingondition.

.4. Response surface analysis of cutting velocity and surfaceoughness based on developed ANFIS models

Here, response surface analysis of cutting velocity and surfaceoughness are presented using surfaces obtained through ANFISodels. These surfaces prove results obtained by experiments and

how that the developed ANFIS models could predict cutting veloc-ty and surface roughness as well.

Fig. 8(a)–(c) depicts ANFIS surfaces of cutting velocity for thenteraction terms. Response surface of CV versus pulse on timend pulse off time is presented in Fig. 8(a). From this figure it

an be seen that higher pulse on time and lower pulse off timeombination resulting higher cutting velocity due to higher ther-al energy which transfers from wire to workpiece. Also, Fig. 8(b)

epicts surface of cutting velocity versus gap voltage and discharge

Fig. 7. Comparison measured values of testing data trough experime

2.98 3.1011 4.03 2.29 2.2314 2.622.37 2.3161 2.27 2.5 2.4679 1.63.22 3.2918 2.17 2.81 2.7359 2.85

current. From the figure it can be seen that higher discharge currentand lower gap voltage combination resulting higher cutting veloc-ity due to higher thermal energy and non-winded wire situation.Moreover, Fig. 8(c) depicts cutting velocity versus wire feed andwire tension. The figure presents that both wire feed and wire ten-sion are insignificant in term of cutting velocity. Variation of themhas not great influence on CV (just about 0.15 mm/min).

Fig. 9(a)–(c) depicts ANFIS surfaces of surface roughness for theinteraction terms. Response surface of SR versus pulse on time andpulse off time is shown in Fig. 9(a). It could be found that lowerpulse off time and higher pulse on time combination resulting lowersurface roughness due to increase in non-cutting time which cre-ate shallower discharge crater. Also, surface roughness versus gapvoltage and discharge current is visible in Fig. 9(b). It can be seenthat higher gap voltage and lower discharge current combinationproduces smoother surface because of transferring lower thermalenergy from wire to workpiece (because of lower discharge cur-rent) and winding condition of wire (because of winding conditionof wire). Fig. 9(c) indicates surface roughness versus wire feed andwire tension. It can be seen from the figure that lower wire feedand lower wire tension combination resulting better surface finish.On one hand, increasing in wire feed can recover the wire duringprocess and increases the spark efficiency; therefore, it has nega-

tive influence on surface quality. On the other hand increase in wiretension can prevent the wire to be wound and produce rigid sparks;thus, surface roughness increases while values of higher wire ten-sion selected. According to above descriptions it can be inferred

nts with predicted values through ANFIS for (a) CV and (b) SR.

Page 10

492 R.K. Fard et al. / Journal of Manufacturing Processes 15 (2013) 483–494

ts

4

rmmt

F

wapdiwv

C

where CVmin and CVmax are the minimum and maximum valuesof MRR, respectively. Also SRmin and SRmax are the minimum andmaximum values of SR, respectively.

Table 9Setup parameters for implementation of ABC.

Parameters Value/function Remark

X0 Li + rand(0,1) × (Ui − Li) Equation used for initializationpurpose

Np 20 Number of population (swarm size)NEB 50% of Np Number of employed beesN 50% of N Number of onlooker bees

Fig. 8. (a–c) Obtained surfaces of cutting velocity through ANFIS model.

hat lower wire tension and lower wire feed produces smootherurface.

.5. Multi-characteristics optimization of process by ANFIS-ABC

For simultaneous optimization of cutting velocity and surfaceoughness by artificial bee colony algorithm the developed ANFISodels are used as objective functions. Hence, the following opti-ization function is used to convert the multi-objective problem

o a single objective optimization problem

= −W1CV + W2SR (4)

here W1 and W2 are the weighing factor related to each responseccording to its importance in the process. As discussed in theresent work 0.6 < W1 < 0.9 and 0.1 < W2 < 0.4 are considered due tory WEDM process behavior. Also, the minus sign for cutting veloc-

ty is due to the fact that the cutting velocity must be maximized

hile the ABC has minimize nature. CV and SR are the normalizedalues of CV and SR which obtained by following equations:

V = CV − CVmin

CVmax − CVmin(5)

Fig. 9. (a–c) Obtained surfaces of surface roughness through ANFIS model.

SR = SR − SRmin

SRmax − SRmin(6)

OB p

NSB 1 Number of scout bees per cycleMCN 1000 Maximum cycle numberH(X) y = evalfis(x,net)

H(X) = −y(1) + y(2)Objective function which uses “evalfis”to simulate the ANFIS network

Page 11

R.K. Fard et al. / Journal of Manufacturing Processes 15 (2013) 483–494 493

Table 10Obtained optimal solutions through ANFIS-ABC for various weight factors.

Weight Optimal factors Responses

W1 W2 Ton (�s) Toff (�s) VG (V) Id (A) WF (mm/min) WT (g) CV (mm/min) SR (�m)

0.6 0.4 146 33 15 245 14 4 4.4315 3.19310.7 0.3 148 32 14.2 247 15 3.4 4.6391 3.53770.8 0.2 149.2 31 13.5 249 15.3 3 4.8876 3.72010.9 0.1 149.7 30.4 13 249.8 15.8 2.7 4.9929 3.8629

Table 11Results of confirmatory experiments to validate optimal results which were obtained through ANFIS-ABC.

Optimal factors Responses

Ton (�s) Toff (�s) VG (V) Id (A) WF (mm/min) WT (g) CV (mm/min)Predicted/measured SR (�m)Predicted/measured

146 33 15 245 14 4 4.4315/4.35 3.1931/3.15148 32 14.2 247 15 3.4 4.6391/4.73 3.5377/3.42

3

2.7

d••••••

mpso

fitaaioSftvT

miAmfiat

5

lWHtTLoo

149 31 13.5 249 15

150 30 13 250 16

Moreover the range of each factor for optimization can beefined as follow

Pulse on time = from 50 to 150 �s (Available range on WEDM machine) Pulse off time = from 30 to 90 �s (Available range on WEDM machine) Gap voltage = from 10 to 100 V (Available range on WEDM machine) Discharge current = from 50 to 250 A (Available range on WEDM machine) Wire feed = from 2 to 16 mm/min (Available range on WEDM machine)

Wire tension = from 2 to 16 g (Available range on WEDM machine)The ABC algorithm needs to some setup parameters for imple-

entation and finding global optima. Table 9 defines the main setuparameters for ABC algorithm. These parameters have not beenelected stochastically, they were selected based on trial and errorsn some benchmark functions such as Rastrigin and Rosenbrock.

Table 10 presents obtained optimal solutions for various weightactors. It can be seen from this table that the optimal pulse on times about 147–150 �s, the optimal pulse off time is about 30–32 �s,he optimal gap voltage is about 10–12 V, the optimal current isbout 247–250 A, the optimal wire feed is about 14–16 mm/minnd the optimal wire tension is about 2–3 g. According to this tablet is seen that by increasing W1 and decreasing W2 (weight factorsf cutting velocity and surface roughness, respectively) the CV andR also increase. This procedure seems logical because with majorocus on CV, the ABC algorithm searches solutions for improvinghe CV. Hence, the higher pulse on time, lower pulse off time, loweroltage, and higher current are the results that cause to higher CV.herefore, by selection of these factors the SR also increases.

In order to verify the results of Table 10, confirmatory experi-ents have been performed and their results have been presented

n Table 11. From this table, it can be inferred that the developedNFIS-ABC is a potential method for solving modeling and opti-ization of dry WEDM process. It has also acceptable accuracy in

nding optimal solutions. Therefore, the proposed method can bepplied for modeling and optimization of other types of manufac-uring processes.

. Conclusion

The present work consisted of experimental investigation, intel-igent modeling and multi-characteristics optimization of dry

EDM process while machining of Al–SiC metal matrix composite.ere, firstly a series of exploratory experiments has been conducted

o find appropriate gas and wire based on higher cutting velocity.

hen numbers of 27 experiments were carried out based on Taguchi27 orthogonal array to investigate effects of pulse on time, pulseff time, gap voltage, discharge current, wire tension and wire feedn cutting velocity and surface roughness. Afterward, an adaptive

4.8876/5.02 3.7201/3.654.9929/5.18 3.8629/3.72

neuro-fuzzy inference system (ANFIS) was employed to crate map-ping relationship between process inputs and main responses. Atthe end of the work, a multi-characteristic optimization was ful-filled to select optimal setting of factors by using of artificial beecolony algorithm. A summary of obtained results is concluded asfollow:

• From the exploratory experiments it could be found that oxy-gen gas and brass wire resulting higher cutting velocity, oxygengas can create a chemical reaction and increases corrosion rate ofworkpiece. Also, the brass wire leads to increasing in spark effi-ciency and electrostatic forces due to its higher specific resistancethan the copper wire.

• From analysis of variances it could be found that pulse on timeand discharge current are most significant factors rather than theothers. Also, wire tension appears the most insignificant factorbased on its percentage of contribution.

• The ANFIS model could predict the cutting velocity and surfaceroughness as well due to low values of RMSE in testing. Also,based on obtained ANFIS surfaces it could be found combinationof high pulse on time, low pulse off time, low gap voltage, highcurrent resulting higher cutting velocity regardless wire feed andwire tension. Moreover, combination of low pulse on time, highpulse off time, high gap voltage, low discharge current, low wirefeed and low wire tension leads to lower surface roughness.

• The optimal results which are obtained through ANFIS-ABC havebeen verified by confirmatory experiment to show the efficiencyof proposed method

References

[1] Lin JT, Bhattacharyya D, Fergusson WG. Chip formation in machining of SiC-particle-reinforced aluminum–matrix composites. Composites Science andTechnology 1998;58:285–91.

[2] El-Gallab M, Sklad M. Machining of Al/SiC particulate metal matrix com-posite: Part I. Tool performance. Journal of Materials Processing Technology1998;83:151–8.

[3] El-Gallab M, Sklad M. Machining of Al/SiC particulate metal matrix compos-ite: Part II. Surface integrity. Journal of Materials Processing Technology831998:277–85.

[4] El-Gallab M, Sklad M. Machining of Al/SiC particulate metal matrix composite:Part III. Comprehensive tool wear. Journal of Materials Processing Technology2000;101:10–20.

[5] Ozben T, Kilickap E, Cakir O. Investigation of mechanical and machinability

properties of SiC particle reinforced Al-MMC. Journal of Materials ProcessingTechnology 2008;98:220–5.

[6] Mohan B, Rajadurai A, Satyanarayana KG. Electric discharge machining ofAl–SiC metal matrix composite using rotary tube electrode. Journal of MaterialsProcessing Technology 2004;153–154:978–85.

Page 12

4 acturi

[

[

[

[

[

[

[

[

[

[

[

[

94 R.K. Fard et al. / Journal of Manuf

[7] Mohan B, Rajadurai A, Satyanarayana KG. Effect of SiC and rotation of electrodeon electrical-discharge machining of Al–SiC composite. Journal of MaterialsProcessing Technology 2002;127:297–304.

[8] Hocheng H, Lei WT, Hsu HS. Preliminary study of material removal in electrical-discharge machining of SiC/Al. Journal of Materials Processing Technology1997;63:813–8.

[9] Senthikumar C, Ganesan G, Karthikeyan R. Study of electrochemical machin-ing characteristics of Al/SiCp composites. International Journal of AdvancedManufacturing Technology 2009;43:256–63.

10] Yoshida M, Taniguchi N. Electrical discharge machining in gbtained throuas.CIRP Annals – Manufacturing Technology 1997;46:143–6.

11] Choudhury SK. Experimental investigation and empirical modeling of the dryelectric discharge machining process. International Journal of Machine Toolsand Manufacture 2009;49(3–4):297–308.

12] Tao J, Shih AJ. Near dry electrical discharge machining. International Journal ofMachine Tools and Manufacture 2007;47(15):2273–81.

13] Sarkar S, Sekh M, Mitra S, Bhattacharyya B. Modeling and optimization of wireelectrical discharge machining of �-TiAl in trim cutting operation. Journal ofMaterials Processing Technology 2008;205:376–87.

14] Hewidy MS, El-Taweel TA, El-Safty MF. Modeling the machining parametersof wire electrical discharge machining of Inconel 601 using RSM. Journal of

Materials Processing Technology 2005;169:328–36.

15] Spedding TA, Wang ZQ. Study on modeling of wire EDM process. Journal ofMaterials Processing Technology 1997;69:18–28.

16] Saha P, Singha A, Pal SK, Saha P. Soft computing models based prediction ofcutting speed and surface roughness in wire electro-discharge machining of

[

ng Processes 15 (2013) 483–494

tungsten carbide cobalt composite. International Journal of Advanced Manu-facturing Technology 2008;39:74–84.

17] Babajanzade-Roshan S, Behboodi-Jooibari M, Teimouri R, Asgharzade-Ahmadi G, Falahati-Naghibi M, Sohrabpoor H. Optimization of frictionstir welding process of AA7075 aluminum alloy to achieve desirablemechanical properties using ANFIS models and simulated annealing algo-rithm. International Journal of Advanced Manufacturing Technology 2013,http://dx.doi.org/10.1007/s00170-013-5131-6.

18] Samanta S, Chakraborty S. Parametric optimization of some non-traditionalmachining processes using artificial bee colony algorithm. Engineering Appli-cations of Artificial Intelligence 2011;24:946–57.

19] Teimouri R, Baseri H. Improvement of dry EDM process characteristics usingartificial soft computing methodologies. Production Engineering Research andDevelopment 2012, http://dx.doi.org/10.1007/s11740-012-0398-2.

20] Teimouri R, Baseri H. Forward and backward predictions of the fric-tion stir welding parameters using fuzzy-artificial bee colony-imperialistcompetitive algorithm systems. Journal of Intelligent Manufacturing 2012,http://dx.doi.org/10.1007/s10845-013-0784-4.

21] Karaboga D. An artificial bee colony (ABC) algorithm for numeric functionoptimization. In: Proceedings of the IEEE swarm intelligence symposium.

2005.

22] Teimouri R, Baseri H. Experimental study of rotary magneticfield-assisted dry EDM with ultrasonic vibration of workpiece. Inter-national Journal of Advanced Manufacturing Technology 2012,http://dx.doi.org/10.1007/s00170-012-4573-6.