j o u r n a l o f m a t e r i a l s p r o c e s s i n g t e c h n o l o g y 2 0 2 ( 2 0 0 8 ) 574–582

journa l homepage: www.e lsev ier .com/ locate / jmatprotec

A study on surface roughness in abrasive waterjetmachining process using artificial neural networksand regression analysis method

Ulas Caydas ∗, Ahmet HascalıkUniversity of Firat, Technical Education Faculty, Department of Manufacturing, Elazig, Turkey

a r t i c l e i n f o

Article history:

Received 11 June 2007

Received in revised form

5 September 2007

Accepted 1 October 2007

Keywords:

Abrasive waterjet machining

Surface roughness

a b s t r a c t

In the present study, artificial neural network (ANN) and regression model were developed

to predict surface roughness in abrasive waterjet machining (AWJ) process. In the devel-

opment of predictive models, machining parameters of traverse speed, waterjet pressure,

standoff distance, abrasive grit size and abrasive flow rate were considered as model vari-

ables. For this purpose, Taguchi’s design of experiments was carried out in order to collect

surface roughness values. A feed forward neural network based on back propagation was

made up of 13 input neurons, 22 hidden neurons and one output neuron. The 13 sets of

data were randomly selected from orthogonal array for training and residuals were used

to check the performance. Analysis of variance (ANOVA) and F-test were used to check the

validity of regression model and to determine the significant parameter affecting the surface

Artificial neural network

Regression analysis

roughness. The statistical analysis showed that the waterjet pressure was an utmost param-

eter on surface roughness. The microstructures of machined surfaces were also studied by

scanning electron microscopy (SEM). The SEM investigations revealed that AWJ machining

produced three distinct zones along the cut surface of AA 7075 aluminium alloy and surface

striations and waviness were increased significantly with jet pressure.

ters (Momber and Kovacevic, 1998; Hashish, 1991). They are

1. Introduction

Manufacturing industry is becoming ever more time-conscious with regard to the global economy, and the needfor rapid prototyping and small production batches is increas-ing. These trends have placed a premium on the use ofnew and advanced technologies for quickly turning rawmaterials into usable goods; with no time being requiredfor tooling (http://www.ttp.net/0-87849-918-0.html). Abrasivewaterjet (AWJ) machining technology has been found to be

one of the most recent developed advanced non-traditionalmethods used in industry for material processing with thedistinct advantages of no thermal distortion, high machining

∗ Corresponding author. Tel.: +90 424 2370000/4229; fax: +90 424 218467E-mail address: ucaydas@firat.edu.tr (U. Caydas).

0924-0136/$ – see front matter © 2007 Elsevier B.V. All rights reserved.doi:10.1016/j.jmatprotec.2007.10.024

© 2007 Elsevier B.V. All rights reserved.

versatility, high flexibility and small cutting forces (Hascalik etal., 2007). Because of these capabilities, it makes an importantcontribution to machining materials with higher performanceand more cost-effective than traditional and some non-traditional machining processes. AWJ is widely used in themachining of materials such as titanium, steel, brass, alu-minium, stone, inconel, any kind of glass and composites(Akkurt et al., 2004). The intensity and the efficiency of themachining process depend on several AWJ process parame-

4.

classified as hydraulic, abrasive, work material and cuttingparameters. Surface roughness, which is used to determineand to evaluate the quality of a product, is one of the major

t e c

qabrepiOfsmAcbw(pdtdmInftd(gana0esiosptahCrafirr

tcitoroHdwcs

j o u r n a l o f m a t e r i a l s p r o c e s s i n g

uality attributes of an AWJ machining product.The use ofrtificial neural networks (ANNs) in machining research haseen extensive and multifaceted. The literature is rich with theelevant investigations on choosing best machining param-ters for low surface roughness during different machiningrocesses. Lee et al. (1998) used abructive network model-

ng for drilling process for predicting the surface roughness.zcelik et al. (2005) have used a statistical three-level full

actorial experimental design with 81 runs to optimize theurface roughness in end milling Inconel 718. A predictiveodel of surface roughness was created using a feed forwardNN exploiting experimental data. Spedding and Wang (1997)ompared the response surface model with neural networky using machining parameters on surface roughness duringire electrical discharge machining. Erzurumlu and Oktem

2007) have developed an ANN and response surface model toredict surface roughness in milling mould parts. A statisticalesign consisting of 243 experiments was adopted to collecthe Ra measurement data. An effort has been made to pre-ict surface roughness in end milling process by using ANNodel based on design experiments Oktem et al. (in press).

lhan et al. (1992) have used a [28.3] factorial design with a totalumber of 718 experiments to establish the different objective

unctions corresponding to surface roughness of the elec-rochemical surface grinding process. Nabil and Ridha (2006)eveloped an approach combining the design of experiments

DOE) and the ANN methods to establish accurate models forround surface roughness parameters prediction. Risbood etl. (2003) utilized a neural network to predict surface rough-ess and dimensional deviation based on the cutting forcesnd vibrations in turning of rolled steel bars containing about.35% carbon. Kumar and Choudhury (in press) comparedxperimental investigations and modeling of wheel wear andurface roughness during electro-discharge diamond grind-ng process using DOE and ANN. 31 experiment were carriedut based on the central composite rotatable design to obtainurface roughness for different combination of machiningarameters. Analysis of the data indicated that to achieve bet-er surface finish of workpiece (HSS) pulse current, duty rationd grain size should be chosen from the lower ranges whileigher value of wheel speed should be selected. Chien andhou (2001) presented an ANN approach to predict surfaceoughness of the AISI 304 stainless steel, the cutting forcesnd the tool life. Then the genetic algorithm was introduced tond the optimum cutting conditions for the maximum mate-ial removal rate under the constraints of the expected surfaceoughness.

The literature reveals that ANNs find many applicationso predict surface finish through different machining pro-esses, but very little effort is reported on the use of ANNsn AWJ machining process. Additionally, applied experimen-al methods are requiring large number of trials when numberf machining parameters increase. On the other hand, theegression analysis method has successfully been used forbtaining the machining performance by many researchers.owever for obtaining the suitable mathematical form, a great

eal of experimental data is necessary. The Taguchi method,hich is one of the fractional factorial designs, uses a spe-

ial design of orthogonal arrays to study the entire parameterpace with small number of experiments only (Ko et al., 1999).

h n o l o g y 2 0 2 ( 2 0 0 8 ) 574–582 575

Jegaraj and Babu (2007) attempted to make use of Taguchi’sapproach and analysis of variance (ANOVA) using minimumnumber of experiments for studying the influence of param-eters on cutting performance in AWJ machining consideringthe orifice and focusing tube bore variations to develop empir-ical models. Thus, Taguchi approach was applied to thisstudy. Data for teaching the ANNs can be selected fromthe orthogonal array table with a small number of trial andthe characteristics as output data of the ANN can be trans-formed by the Taguchi method for obtaining more accurateperformance. Therefore, ANNs having properties as learningcapability and adaptation, working property with a few dataand high speed working have been used in this paper. At theend of study, both the ANN and regression analysis resultswere compared with experimental data.

2. Surface roughness

Surface roughness is a measure of the technological qualityof a product and a factor that greatly influences manufac-turing cost. It describes the geometry and surface texturesof the machined parts (Nalbant et al., 2007). There areseveral ways to describe surface roughness, such as rough-ness average (Ra), root-mean-square (rms) roughness (Rq)and maximum peak-to-valley roughness (Ry or Rmax), etc.(http://www.prediv.com/smg/parameters.html). Ra is definedas the arithmetic value of the profile from centreline along thesampling length. It can be expressed by the following mathe-matical relationship (Ozcelik et al., 2005):

Ra = 1L

∫ L

0

|y(x)||dx| (1)

where L is the sampling length, y is profile curve and x is theprofile direction. The average surface roughness Ra is mea-sured within L = 0.8 mm.

3. Artificial neural networks (ANNs)

Neural networks, as used in artificial intelligence, have tradi-tionally been viewed as simplified models of neural processingin the human brain. It is accepted by the most scientists thatthe human brain is a type of computer. The origins of neuralnetworks are based on efforts to model information process-ing in biological systems, which may rely largely on parallelprocessing as well as implicit instructions based on recog-nition of patterns of “sensory” input from external sources(http://en.wikipedia.org/wiki/Neural network).

Human body consists of trillions of cells. A portion ofthem is the nerve cells called “neurons”. These neuronshave different shapes and sizes (Tosun and Ozler, 2002). Aneuron collects signals from others through fine structurescalled dendrites. The neuron sends out spikes of electricalactivity through a long, thin stand known as axon, which

splits into thousands of branches. At the end of each branch,a structure called a synapse converts the activity from theaxon into electrical effects that inhibit or excite activity in theconnected neurons. When a neuron receives excitatory input

n g t

576 j o u r n a l o f m a t e r i a l s p r o c e s s i

that is sufficiently large compared with its inhibitory input,it sends a spike of electrical activity down its axon. Learn-ing occurs by changing the effectiveness of the synapsesso that the influence of one neuron on another changes(http://www.doc.ic.ac.uk/nd/surprise96/journal/vol4/csll/report.html).

3.1. Backpropagation algorithm

Even though several learning methods have been developed,the back propagation (BP) method has been proven to be suc-cessful in applications related to surface finish prediction (Tsaiet al., 1999; Azouzi and Guillot, 1999; Zouaghi et al., 1996). Inthis study, BP learning algorithm, which has a unique learningprinciple, generally called delta rule, is used. Fig. 1 depicts aschematic illustration of BP networks. The three layer of thenetwork architecture include the input layer, hidden layer andoutput layer. Layers include several processing units knownas neurons. They are connected with each other by variableweights to be determined. In the network, the input layerreceives information from external source and passes thisinformation to the network for processing. The hidden layerreceives from the input layer, and does all information pro-cessing. The output layer receives processed information fromthe network, and sends the results to an external receptor(Singh et al., 2006). In the network, each neuron receives totalinput from all of the neurons in the proceeding layer as:

netj =∑

i

W(n)ji

Xi(n−1) (2)

where netj is the total or net input, Xj(n) is the output of the

node j in the nth layer, and W(n)ji

represents the weights fromnode i in the (n − 1)th layer to node j in the nth layer.

A neuron in the network produces its input by processingthe net input through an activation (transfer) function which

Fig. 1 – Schematic illustration of artificial neural

e c h n o l o g y 2 0 2 ( 2 0 0 8 ) 574–582

is usually nonlinear. There are several types of activation func-tions used for BP. However, the sigmoidal activation functionis most utilized. Three types of sigmoid functions are usuallyused, as follows (Liu et al., 2006):

f (x) = 11 + e−x

range (0, 1) (3)

f (x) = 21 + e−x

− 1 range (−1, 1) (4)

f (x) = ex − e−x

ex + e−xrange (−1, 1) (5)

the weights are dynamically updated using the BP algorithm.The difference between the target output and actual output(learning error) for a sample p is (Tosun and Ozler, 2002)

Ep = 12

K∑k=1

(dpk − opk)2 (6)

where dpk and opk are the desired and calculated output for kthoutput, respectively. K denotes the number of neuron in outputof network. The average error for whole system is obtained by:

Ep = 12

P∑p=1

K∑k=1

(dpk − opk)2 (7)

where P is the total number of instances. For the purposeof minimizing Ep, the weights of the inter-connections areadjusted during the training procedure until the expectederror is achieved. To adjust the weights of the networks, theprocess starts at the output neuron and works backward to the

hidden layer. The weights in BP based on the delta learningrule can be expressed as follows:

wnewij = wold

ij + �wij (8)

network model for the surface roughness.

j o u r n a l o f m a t e r i a l s p r o c e s s i n g t e c h n o l o g y 2 0 2 ( 2 0 0 8 ) 574–582 577

Table 1 – Chemical composition of Al 7075 alloy

Al 91.02Cu 1.65Mg 2.0

�

wptwm

4

4

Ia�

acivvtmm

4

Fmtpsestvftauf

Table 3 – Experimental design using L27 orthogonalarray and experimental results

Experiment no. V P h d m Surface roughness,Ra (�m)

1 1 1 1 1 1 2.1242 1 1 1 1 2 2.7533 1 1 1 1 3 3.3524 1 2 2 2 1 4.3115 1 2 2 2 2 4.5416 1 2 2 2 3 5.1237 1 3 3 3 1 6.7898 1 3 3 3 2 7.5249 1 3 3 3 3 9.123

10 2 1 2 3 1 3.57511 2 1 2 3 2 4.45712 2 1 2 3 3 5.62813 2 2 3 1 1 7.01014 2 2 3 1 2 7.53515 2 2 3 1 3 7.89316 2 3 1 2 1 8.12117 2 3 1 2 2 8.31218 2 3 1 2 3 9.16319 3 1 3 2 1 4.32820 3 1 3 2 2 5.12021 3 1 3 2 3 5.85222 3 2 1 3 1 6.14323 3 2 1 3 2 6.72124 3 2 1 3 3 7.78025 3 3 2 1 1 8.89026 3 3 2 1 2 9.120

Cr 0.23Zn 5Mn 0.1

wij = −�∂Ep

∂wijoutj (9)

here outj the jth neuron output. � is the learning ratearameter controlling stability and rate of convergence ofhe network, which is a constant between 0 and 1. Once theeights of all the links of the network are decided, the decisionechanism is then developed.

. Experimental studies

.1. Materials choice

n this study, AA 7075-T6 wrought alloy (AlZnMgCu1.5) withn ultimate strength of �ult = 675 MPa and yield strength of

ult = 610 MPa was selected. Because of its low specific weightnd high strength to weight ratio and also its high electri-al and thermal conductance, this alloy is widely used inndustry and in particular in aircraft structure and pressureessels (Majzoobi and Jaleh, 2007). The material stock was pro-ided from ETIALUMINYUM (Turkey). The stock was milledo 100 mm × 100 mm × 100 mm in dimensions under the same

achining conditions. The chemical composition of the usedaterial is given in Table 1.

.2. AWJ cutting procedure and design of experiments

or conducting the experiments, an abrasive waterjet cuttingachine (WJ-S42 E, Germany) was employed. The erodent par-

icle was garnet abrasive. In the cutting tests, the jet waserpendicular to the sample surface. The orifice assembly con-isted of a carbide nozzle insert 1 mm in diameter. In thexperimental plan, the most dominant process parametersuch as traverse speed (V), waterjet pressure (P), standoff dis-ance (h), abrasive grit size (d) and abrasive flow rate werearied at three levels. The five process parameters and theiractor levels are summarized in Table 2. In order to measure

he average surface roughness (Ra) of AWJ machined samples,SJ-201 portable surface roughness measurement device wassed. The measurements were taken at a distance of 5 mm

rom the top of the cut surface. A LEO 32 scanning electron

Table 2 – Machining settings used in the experiments

Symbol AWJ cutting parameter

V Traverse speed (mm/min)P Waterjet pressure (MPa)h Standoff distance (mm)d Abrasive grit size (�m)m Abrasive flow rate (g/s)

27 3 3 2 1 3 10.035

microscope (SEM) was used to investigate the machined sur-face.

Normally, one need to conduct 35 (243) experiments whenfive factors, each varied at three levels considered, using fullfactorial experimental design. In order to save on experimen-tal cost and time, Taguchi’s orthogonal array was applied toobtain the surface roughness of the AWJ process. A L27 orthog-onal array was found to be appropriate and it was chosen. Thelayout of the L27 orthogonal array and the measured surfaceroughness values are shown in Table 3. Among these datasets in orthogonal array, 13 data were selected randomly astraining data of neural network, and the residuals were usedto verify the predicted results. For training, a commercialMicrosoft Windows-based ANN software, Matlab Version6.1 (The MathWorks, Natick, MA) was used throughout the

study with a P-4 personal computer. Several iterations wereconducted with different numbers of nodes of hidden layerin order to determine the optimal ANN structure. Duringpre-trials, the minimum least mean squared error was inter-

Level 1 Level 2 Level 3

50 100 150125 175 250

1 2.5 460 90 120

0.5 2 3.5

n g t e c h n o l o g y 2 0 2 ( 2 0 0 8 ) 574–582

Fig. 2 – Iteration number vs. mean square error.

578 j o u r n a l o f m a t e r i a l s p r o c e s s i

estingly achieved with 22 hidden nodes. The learning rate andmomentum values were selected 0.9 and 0.2, respectively.

4.3. Determination of regression analysis model forsurface roughness

Regression analysis method includes the experimental inves-tigations, mathematical methods and statistical analysis. Inthe present investigation, a whole analysis was done using theexperimental data in Table 3. A multilinear stepwise regres-sion analysis was performed to predict the surface roughnessusing MINITAB 14 software. Specially, with a sample of n obser-vations of the dependent variable Y (Ra), the regression model(Lin et al., 2007) can be expressed as:

Ra = ˇ0 +k∑

i=1

ˇiXi +k∑

i=1

ˇiiXi2 +

∑∑i<j

ˇijXiXj + εi (10)

where k is number of factors (5) ˇ0 is the free term, ˇi is thelinear effect, ˇii is the squared effect and ˇij is the interac-tion effect. The second-order polynomial regression equationrepresenting the surface roughness (Ra) can be expressed as afunction of AWJ machining parameters such as traverse speed(V), waterjet pressure (P), standoff distance (h), abrasive gritsize (d) and abrasive flow rate (m). Eq. (10) can be rewritten tobuild the relationship between the AWJ process parametersand surface roughness as follows:

Ra = b0 + b1V + b2P + b3h + b4d + b5m + b11V2 + b22P2

+ b33h2 + b44d2 + b55m2 + b12VP + b13Vh + b14Vd

+ b15Vm + b23Ph + b24Pd + b25Pm + b34hd

+ b35hm + b45dm (11)

5. Results and discussions

Results of ANN and regression analysis, used to establishinput–output relationships in AWJ machining process, areshown and discussed below.

5.1. Estimation of surface roughness by ANN

The optimal neural network architecture used in this studyis indicated in Fig. 1. The network consists of one input, onehidden and one output layer. Hidden layer has 22 neurons,whereas output layer has one neuron. 13 neurons with fivefeatures have been used as an input of ANN. Iteration numberversus mean square error (MSE) is shown in Fig. 2. It can beseen that training of neural networks can be achieved quickly.After 331 cycles of training (Epochs), the training error of net-work reaches stabilization value. The mean error is 3.0072%for surface roughness. The error is lower than 10%, which

show that the well-rained network model takes on optimalperformance and has a great accuracy in predicting surfaceroughness (Ozel and Karpat, 2005). The results predicted fromthe ANN model are compared with experimental measure-

Fig. 3 – Comparison of ANN results with experimentalmeasurements.

ments results in Table 3 for 13 check sets. Fig. 3 depicts thecomparison of results between them. Good agreement can beseen.

5.2. Estimation of surface roughness by regressionanalysis

From the results (Table 3), the final regression model for sur-face roughness obtained is as follows:

Ra = −5.07976 + 0.08169V + 0.07912P − 0.34221h − 0.08661d

− 0.34866m − 0.00031V2 − 0.00012P2 + 0.10575h2

+ 0.00041d2 + 0.07590m2 − 0.00008Vm − 0.00009Pm

+ 0.03089hm + 0.00513dm (12)

note that some interaction terms are removed from full modelbecause of their nonsignificant effect. Regression statistics

j o u r n a l o f m a t e r i a l s p r o c e s s i n g t e c h n o l o g y 2 0 2 ( 2 0 0 8 ) 574–582 579

Table 4 – ANOVA test results for regression model

Source DF SS MS F P

Regression 14 124.100 8.86432 182.19 0.000*

Linear 5 117.429 2.15934 44.38 0.000*

Square 5 5.973 1.19467 24.55 0.000*

Interaction 4 0.698 0.17444 3.59 0.038Residual error 12 0.584 0.04866Total 26 124.684

R

T9tsTsasaFa

Fe

Fig. 5 – The relative effect of AWJ parameters on the surface

R2 = 0.95. DF—degree of freedom. SS—sum of squares. MS—meansquare.∗ Significant.

2 and R2Adj are obtained equal to 99.5 and 99%, respectively.

he R2 value indicates that the machining parameters explain9.5% of variance in surface roughness. This value indicateshat the presented model fits the data very well. The analy-is of variance (ANOVA) for regression analysis is shown inable 4. The p-value shows that the model, linear terms andquared terms are significant at ˛-level of 0.005, whereas inter-ction terms seems to have not of significant influence on

urface roughness. The results predicted by regression modelre compared with experimental measurements results inig. 4. It can be seen from Fig. 8 that model prediction presentsgood agreement with the experimental data.

ig. 4 – Comparison of regression model results withxperimental measurements.

Table 5 – Result of the ANOVA for the surface roughness

Machining parameter Degree of freedom (f) Sum of squ

Traverse speed 2 22.21Waterjet pressure 2 88.39Standoff distance 2 2.84Abrasive grit size 2 0.89Abrasive flow rate 2 9.08Error 24 4.39Total 34 127.8

roughness.

It was also statistically studied the relative effect of eachAWJ parameters on the surface roughness by using ANOVAand F-test (Appendix A) (Huang et al., 1999). Table 5 shows theresults of ANOVA for the surface roughness. Larger F valueindicates that the variation of the process parameter makes abig change on the surface roughness (Tosun and Cogun, 2003)and P denotes its percent contribution on surface roughness.Fig. 5 shows the relative importance of AWJ parameters usedin this study on the surface roughness. As clearly shown fromthe figure, the waterjet pressure has an utmost importance onthe surface roughness and the effects of standoff distance andabrasive grit size on surface roughness are insignificant. Here,the effect of grit size on surface roughness is surprisingly. Noplausible reason is available in the open literature and furtherinvestigation is needed for this. Traverse speed is the secondranking factor, whereas the abrasive flow rate has a statisticinfluence (7.28%).

5.3. Comparison of ANN and regression results forsurface roughness

Results of ANN and regression analysis are compared withexperiments in Table 3 for 13 check sets. The comparison

results are depicted in Table 6. The maximum test errorsfor ANN and regression model are about 3.0072 and 1.03%,respectively. Both the methods are suitable for estimatingsurface roughness in an acceptable error ranges. But, the

ares (SSA) Variance (VA) FA0 Contribution (%)

11.11 2.60 17.8144.20 29.23 70.89

1.42 0.28 2.280.44 0.09 0.714.54 0.94 7.282.195 1.03

100

580 j o u r n a l o f m a t e r i a l s p r o c e s s i n g t e c h n o l o g y 2 0 2 ( 2 0 0 8 ) 574–582

Table 6 – Comparison of ANN and regression model results with experimental measurements

Experiment no. Surface roughness (�m)

Experimental measurements Regression model ANN model

2 2.753 2.62915 2.74454 4.311 4.00520 3.95266 5.123 5.42532 5.13928 7.524 7.69815 7.8421

10 3.575 3.66819 3.877112 5.628 5.55233 5.057414 7.535 7.36548 7.490916 8.121 7.96455 7.974218 9.163 9.21330 8.645420 5.120 4.98615 5.045022 6.143 6.07837 5.8056

7.79815 7.84779.23448 9.3647

24 7.78026 9.120

model generation and training procedure of ANN model tookmore time than regression model, which took just a couple ofseconds.

6. The characteristics of machined surface

The AWJ machined AA 7075 surfaces were observed underSEM after air cleaning to analyse the machined surface char-acteristics. Typical macroscopic features of an AWJ cut surfaceand kerf geometry has been studied previously by Arola andRamulu (1997). A typical AWJ machined surface has three dis-tinct zones along the kerf wall, i.e. an initial damage region(IDR), a smooth cutting region (SCR) and a rough cuttingregion (RCR) from the jet entry to the exit of workpiece. Typ-ical microstructures of these regions are shown in Figs. 6–8.The IDR is the shallow dark wear tracks created by shal-low angles of attack. The SCR exists between IDR and RCRwith a small area, which is created at large angle of attack.

The surface roughness gets worse through the RCR becauseof the jet upward deflection. In RCR, the surfaces generallyexhibit striate or wavy characteristics. Fig. 9 shows this surfacewaviness for different jet pressures. As shown, an increase

Fig. 6 – SEM micrograph of IDR.

Fig. 7 – SEM micrograph of SCR.

in jet pressure led to striations or waviness with large feeton the wall of the cut surface to deteriorate the surfaceroughness.

Fig. 8 – SEM micrograph of RCR.

j o u r n a l o f m a t e r i a l s p r o c e s s i n g t e c

Fw

7

TaaTad

r

machinability of 304 stainless steel. J. Mater. Process. Technol.

ig. 9 – Increase of surface waviness and surface roughnessith jet pressure (a) P: 125 MPa, (b) P: 175 MPa, (c) P: 250 MPa.

. Conclusion

o determine the relationship between machining parametersnd surface roughness in AWJ machining process, an ANN

nd multiply regression analysis were carried out based onaguchi’s orthogonal array. Comparisons were made of thebove approaches, after testing their performances on 13 ran-omly selected test cases. The machined surfaces were also

h n o l o g y 2 0 2 ( 2 0 0 8 ) 574–582 581

examined using scanning electron microscopy (SEM). Sum-marizing the mean features of the results, both the neuralnetwork and regression approaches were seen to be sufficientfor estimating surface roughness in AWJ machining with avery small test errors. The predicted process parameters onvalidation were found to be close correlation with the actualperformance results. However, the regression model showeda slightly better performance compared to the ANN model.The training of developed neural networks can be achievedquickly after 331 epoches, which could reduce the computa-tional cost of ANN because of its iterative calculations. Fromthis, the predictive models can be used for predicting surfaceroughness in AWJ process with a higher reliability. The perfor-mance can further be enhanced with large experimental datafrom full factorial experimentation and considering the addi-tional performance characteristics. Based on the ANOVA andF-test, the most dominant parameter on the surface rough-ness was found as waterjet pressure, while the second rankingfactor was traverse speed. Abrasive flow rate and standoff dis-tance were less effective on surface roughness, while effect ofabrasive grit size can be negligible. Microstructure evaluationof cutting surfaces of samples revealed that an AWJ processproduce three distinct region along the cut wall surface as: aninitial damage region (IDR), a smooth cutting region (SCR) anda rough cutting region (RCR).

Appendix A. Analysis of variance and F-test

Sm =(∑

�i

)2

27, ST =

∑�2

i − Sm

SA =∑

�A2i

N− Sm, SE = ST −

∑SA

VA = SA

fA, FA0 = VA

VE

where ST is the sum of squares due to total variation; Sm is thesum of squares due to mean; SA is the sum of squared due tofactor A (A = (V, P, h, d, m); SE is the sum of square due to error;�i is the � value of each experiment (i = 1–27); �Ai is the sum ofi level of factor A (i = 1, 2, 3); N is the repeating number of eachlevel of factor A; fA is the degree of freedom of factor A; VA isthe variance of factor A and FA0 is the F-test value of factor A.

e f e r e n c e s

Akkurt, A., Kulekci, M.K., Seker, U., Ercan, F., 2004. Effect of feedrate on surface roughness in abrasive waterjet cuttingapplications. J. Mater. Process. Technol. 147, 389–396.

Arola, D., Ramulu, M., 1997. Material removal in abrasive waterjetmachining of metals surface integrity and texture. Wear 210,50–58.

Azouzi, R., Guillot, M., 1999. On-line prediction of surface finishand dimensional deviations in turning using neural networkbased sensor fusion. Int. J. Mach. Tool Manuf. 37 (9),1201–1217.

Chien, W.T., Chou, C.Y., 2001. The predictive model for

118, 442–447.Erzurumlu, T., Oktem, H., 2007. Comparison of response surface

model with neural network in determining the surface qualityof moulded parts. Mater. Des. 28, 459–465.

n g t

582 j o u r n a l o f m a t e r i a l s p r o c e s s i

Hascalık, A., Caydas, U., Gurun, H., 2007. Effect of traverse speedon abrasive waterjet machining of Ti-6Al-4V alloy. Mater. Des.28, 1953–1957.

Hashish, M., 1991. Optimization factors in abrasive waterjetmachining. Transact. ASME J. Eng. Ind. 113, 29–37.

Huang, J.T., Liao, Y.S., Hsue, W.J., 1999. Determination of finish –cutting operation number and machining – parameterssettings in wire electrical discharge machining. J. Mater.Process. Technol. 87, 69–81.

Ilhan, R.E., Sathyanarayanan, G., Storer, R.H., Liao, T.W., 1992.Offline multiresponse optimization of electrochemical surfacegrinding by a multi objective programming method. Int. J.Mach. Tool Manuf. 32 (3), 435–451.

Jegaraj, J.J.R., Babu, N.R., 2007. A soft approach for controlling thequality of cut with abrasive waterjet cutting systemexperiencing orifice and focusing tube wear. J. Mater. Process.Technol. 185, 217–227.

Ko, D.C., Kim, D.H., Kim, B.M., 1999. Application of neural networkand Taguchi method to perform design in metal formingconsidering workability. Int. J. Mach. Tool. Manuf. 39, 771–785.

Kumar, S., Choudhury, S.K. Prediction of wear and surfaceroughness in electro-discharge diamond grinding, J. Mater.Process. Technol., in press.

Lee, B.Y., Liu, H.S., Tarng, Y.S., 1998. Modeling and optimization ofdrilling process. J. Mater. Process. Technol. 74, 149–157.

Lin, B.T., Jean, M.D., Chou, J.H., 2007. Using response surfacemethodology for optimizing deposited partially stabilizedzirkonia in plasma spraying. Appl. Surf. Sci. 253, 3254–3262.

Liu, T.C., Li, R.K., Chen, M.C., 2006. Development of an artificialneural network to predict lead frame dimensions in anetching process. Int. J. Adv. Manuf. Technol. 27,1211–1216.

Majzoobi, G.H., Jaleh, M., 2007. Duplex surface treatments on Al

7075-T6 alloy against fretting fatigue behavior by applicationof titanium coatings plus nitriding. Mater. Sci. Eng. A 452/453,673–681.

Momber, A., Kovacevic, R., 1998. Principles of Abrasive WaterjetMachining. Springer–Verlag, London.

e c h n o l o g y 2 0 2 ( 2 0 0 8 ) 574–582

Nabil, B.F., Ridha, A., 2006. Ground surface roughness predictionbased upon experimental design and neural network models.Int. J. Adv. Manuf. Technol. 31, 24–36.

Nalbant, M., Gokkaya, H., Sur, G., 2007. Application of Taguchimethod in the optimization of cutting parameters for surfaceroughness in turning. Mater. Des. 28, 1379–1385.

Oktem, H., Erzurumlu, T., Erzincanlı, F. Prediction of minimumsurface roughness in end milling mold parts using neuralnetwork and genetic algorithm, Mater. Des., in press.

Ozcelik, B., Oktem, H., Kurtaran, H., 2005. Optimum surfaceroughness in end milling Inconel 718 by coupling neuralnetwork model and genetic algorithm. Int. J. Adv. Manuf.Technol. 27, 234–241.

Ozel, T., Karpat, Y., 2005. Predictive modeling of surfaceroughness and tool wear in turning using regression andneural networks. Int. J. Mach. Tool Manuf. 45, 467–479.

Risbood, K.A., Dixit, U.S., Sahasrabudle, A.D., 2003. Prediction ofsurface roughness and dimensional deviations by measuringcutting forces and vibrations in turning process. J. Mater.Process. Technol. 132, 203–214.

Singh, A.K., Panda, S.S., Pal, S.K., Chakraborty, D., 2006. Predictiondrill wear using an artificial neural network. Int. J. Adv. Manuf.Technol. 28, 456–462.

Spedding, T.A., Wang, Z.Q., 1997. Study on modeling of wire EDMprocess. J. Mater. Process. Technol. 69, 18–28.

Tosun, N., Cogun, C., 2003. An investigation on wire wear inWEDM. J. Mater. Process. Technol. 134, 273–278.

Tosun, N., Ozler, L., 2002. A study of tool life in hot machiningusing artificial neural networks and regression analysismethod. J. Mater. Process. Technol. 124, 99–104.

Tsai, H.Y., Chen, C., Lou, S.J., 1999. In in-process surfacerecognition system based on neural networks in end millingcutting operations. Int. J. Mach. Tool Manuf. 39, 583–605.

Zouaghi, N., Ichida, Y., Ben Fredj, N., Kimura, N., 1996. Grindingmode identification of silicon carbide by using neuralnetwork. In: Proceedings of the third International conferenceon progress of cutting and grinding, JSPE, Osaka 3,pp. 342–347.