Oguz C olak a, * , Cahit Kurbanoglu b , M. Cengiz Kayacan a a University of Suleyman Demirel, CAD/CAM Research and Application Center, 32300 Isparta, Turkey b University of Suleyman Demirel, Mechanical Engineering Department, 32300 Isparta, Turkey Received 31 January 2005; accepted 7 July 2005 Available online 26 August 2005 Abstract CNC milling has become one of the most competent, productive and exible manufacturing methods, for complicated or sculp- tured surfaces. In order to design, optimize, built up to sophisticated, multi-axis milling centers, their expected manufacturing out- put is at least benecial. Therefore data, such as the surface roughness, cutting parameters and dynamic cutting behavior are very helpful, especially when they are computationally produced, by articial intelligent techniques. Predicting of surface roughness is very dicult using mathematical equations. In this study gene expression programming method is used for predicting surface rough- ness of milling surface with related to cutting parameters. Cutting speed, feed and depth of cut of end milling operations are col- lected for predicting surface roughness. End of the study a linear equation is predicted for surface roughness related to experimental study. 2005 Elsevier Ltd. All rights reserved. Keywords: Surface roughness; CNC end milling; Genetic expression programming 1. Introduction To maintain appropriateness to global rivalry condi- tions of manufacturing conditions, customer related, fast and agile manufacturing strategies are becoming an unelectable manufacturing philosophy. To maintain agile and responsive manufacturing, relationships in manufacturing steps should be determined in a fast, ex- act and reliable way. Therefore it is a known fact that there are some limitations in agile and responsive man- ufacturing. To overcome limitations for agile and responsive manufacturing, all manufacturing parame- ters must be supervised on electronically environment. This is called electronically manufacturing. To achieve this process, all eective manufacturing parameters must be transferred to electronically environment by help of intelligent systems. If it is to be evaluated in the view of business management, determination and control of performance values of elements used in process period and determination of conjectural economical life period on electronically environment are important to maintain agile and responsive manufacturing. Especially, inspec- tion and determination of surface roughness in metal working which has an important place in manufacturing industry has very high importance in the view of eco- nomical manufacturing. In surface roughness, cutting parameters (cutting velocity, feed, depth of cut, cutting force, etc.), types of cutting material and tool are eec- tive parameters. Surface nish is important factor in evaluating the quality of products. Surface roughness (Ra) most used 0261-3069/$ - see front matter 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.matdes.2005.07.004 * Corresponding author. Tel.: +90 246 211 1674. E-mail addresses: ocolak@mmf.sdu.edu.tr (O. C olak), kurban@ mmf.sdu.edu.tr (C. Kurbanoglu), ckayacan@mmf.sdu.edu.tr (M.C. Kayacan). www.elsevier.com/locate/matdes Materials and Design 28 (2007) 657666 Materials & Design index to determine the surface nish. Many dierent attributes of the product, including surface friction, wearing, heat transmission, the ability to distribute and hold a lubricant, the ability to accept a coating, and the ability to resist fatigue, are at least partially dis- tinguished by how well the surface nish is produced, due to the fact that surface roughness aects several functional attributes of products. Consequently, the de- sired surface roughness value is usually specied for an individual part, and specic processes are selected in or- der to achieve the specied nish. Surface specication can also be a good reference point in determining the stability of a production process, because the stability of the machine is contingent on the quality of the oper- ating part [1]. Standard roughness measurement procedures depend heavily on stylus instruments which have only limited exibility in handling dierent parts. Furthermore, the procedure is a post-process approach, which is not ame- nable for automation, and the measurement is also rela- tively slow. In recent years, the modeling and prediction problems of surface roughness of a workpiece by com- puter vision in turning operations have received a great deal of attention [2,3]. Although it has been shown that the surface roughness is strongly characterized by the surface image, practical surface roughness measure- ments based on computer vision technology are still dif- cult [2]. The main problem is how to accurately obtain the actual surface roughness of a workpiece using sur- face images and various parameters of cutting operations. Many researchers have studied the end milling pro- cess in the past few years. Kline et al. [5] investigated the eect of vibration, deection and chatter of the workpiecetool system in the end milling process on surface roughness. Alauddin et al. [6] established a mathematical model that predicts the surface roughness of 190 BHN steel after end milling. The prediction model was expressed via cutting speed, feed rate and depth of cut. The researchers also used response sur- face methodology (RSM) to explore the eect of such cutting parameters as cutting speed, feed rate and depth of cut on surface roughness. Alauddin et al. [7] also established a mathematical model for predicting the tool in the end milling process of 190 BHN steel under dry cutting conditions. The model included the following variables: cutting speed, feed rate and axial depth of cut. It also veried the suitability of the pre- diction model via ANOVA. Fuh and Hwang [8] used RSM to construct a model that can predict the milling force in end milling operations. They considered the speed of spindle rotation, feed per tooth and axial and radial depth of cut as the three major factors that aect the milling force. Comparison between the exper- imental data and the values predicted by this prediction model showed the models accuracy to be as high as 95%. Lin [9] studied the issue of inverse heat conduc- tion in end milling process. He used the inverse nite element method to estimate the transient tool face tem- perature and heat dissipation to workpiece during the end milling processes. The ndings indicate that the re- sults of the end milling of both steel and aluminum al- loy derived by the numerical prediction method are a close match to the experimental result. Chiang et al. [10] used two dierent kinds of neural networks to cal- culate the optimal cutting conditions online. They used a Back-Propagation Network with three inputs and four outputs to simulate the cutting process, and an- other kind of neural network to calculate the optimal cutting parameters to achieve the goal of maximizing the material removal rate. Luo et al. [11] used a neural network to simulate the cutting force and contour error in an end milling experiment. They used two sets of neural networks in the study. The rst set of networks derived the feed rate by a parametric interpolation algorithm along the direction of feed axes to obtain the desired cutting force and workpiece geometric shape. The second set is used to make corrections to the feed rate components specied by the parametric interpolation algorithm to minimize the contour error caused by the dynamic lag of the closed-loop servo sys- tems controlling the feed drives. Alauddin et al. [12] developed a mathematical model that predicts the aver- age tangential cutting force in the uncoated carbide in- set end milling process of Inconel 718 workpiece under dry conditions. The predictive cutting force model has been developed in terms of cutting speed and axial depth of cut by RSM. Lou and Chen [13] studied the eect of spindle speed, feed rate and depth of cut on the surface roughness of end milling processes. They used in-process surface roughness recognition (ISRR) and a neural fuzzy system to predict the workpiece sur- face roughness [14]. This research investigates to predicting surface roughness with a new approach method called gene expression programming (GEP). Three milling parame- ters have been selected. The parameters and their mean- ings are listed in Table 1. Based on these three milling parameters and another important parameter on surface roughness are investigating how to use GEP for surface roughness prediction. Table 1 Parameters used in GEP Parameter Meaning Sp Spindle speed (rpm) Fe Feed rate (mmpm) Dep Depth of cut (mm) Ra Roughness (lm) 658 O. C olak et al. / Materials and Design 28 (2007) 657666 2. Measurement of surface roughness There are various simple surface roughness amplitude parameters used in industry, such as roughness average (Ra), root-mean-square (rms) roughness (Rq), and max- imum peak-to-valley roughness (Ry or Rmax), etc. (PDI Webmaster, 2000). The parameter Ra is used in this study. The average roughness (Ra) is the area between the roughness prole and its mean line, or the integral of the absolute value of the roughness prole height over the evaluation length (Fig. 1) [4]. Therefore, the Ra is specied by the following equation: Ra 1 L Z L 0 jY xj dx; 1 where Ra is the arithmetic average deviation from the mean line, L is the sampling length, and Y is the ordi- nate of the prole curve. There are many methods of measuring surface rough- ness, such as image processing, microscopes, stylus type instruments, prole tracing instruments, etc. A Pocket Surf stylus type instrument (produced by Hommel Verke T500) was used in this study. In Fig. 2 several types of surface roughness forma- tions given related to DIN 4760 standard. 3. Prediction of surface roughness for end milling operations Surface quality of machined part depends on stability of CNC, cutting parameter, cutting tool, workpiece Y Roughness center line X Y: Profile curve X: Profile direction Z: Average roughness height L: Sampling length H: Profile height L Z H Cutoff Length Fig. 1. Surface roughness prole [4]. Fig. 2. Classication of surface roughness formation related to DIN 4760 Standart. O. C olak et al. / Materials and Design 28 (2007) 657666 659 material and tool path. In this study Hadrford VMC 1020 Vertical machining center is used for experimental study. When the cutting speed increases, surface quality also increases in CNC end-milling operations. Depth of cut is indirectly eect surface quality. Cutting forces, vibration and cutting temperature are increase with higher depth of cut. Therefore surface roughness of machined part is increased with dept of cut. Experimental study shows that surface roughness is increase with increasing feed. Cutting edge of cutting tool and cutting tool radius is eective factor for sur- face roughness variation as depth of cut. Tool wear is also increase surface roughness. In Fig. 3 eective parameters shbone diagram on surface roughness is seen. [15] Dynamic parameters of machining, like cutting forces are eect indirectly surface roughness variation [15]. 4. Genetic expression programming (GEP) GEP algorithm is a solution method which makes a global function search for the problem, developed as a resultant of genetic algorithm(GA) and genetic program- ming (GP) algorithms. Characteristic of GA algorithms is a linear array of constant length chromosomes. Despite of they could be manipulated by genetic operators easily these are not functional in non-linear problems. GP algo- rithms try to nd a suitable solution using parse three which they create to dene relations between dierent size and shape non-linear variables. Advantages of GA and GP algorithms are jointed in GEP algorithm. Relation- ships of non-linear variables which are characteristically Process Kinemtics Cutting Fluid Depth of Cut Feed Rate Stepover Tool Angle Cutting Speed Vibrations Friction in the cutting zone Cutting Force Variation Accelerations Chip Formation Workpiece Hardness Workpiece Lenght Tool Material Tool Shape Nose Radius Ranout Errors Surface Roughness Machining Parameters Cutting Tool Properties Workpiece Properties Cutting Phenomena Workpiece width Fig. 3. Fishbone diagram with the parameters that aect surface roughness [15]. Fig. 4. (a) Example crossovers for GEP algorithm: (a-1) program ET, (b-2) program ET, (c) crossover ET, (d) result program ET. (b) Experimental setup. 660 O. C olak et al. / Materials and Design 28 (2007) 657666 in dierent size and shape are derived in order to convert constant size and linear arrays using suitable function ge- netic operators [10]. GEP algorithm can be handled as a common applica- tion of GA and GP algorithms. The genetic algorithm (GA) is applied to solve the expression tree. GEP can be applied to the application of conventional genetic algorithm and genetic programming (GP). GEP genes are composed of a list of operators, functions, constants and variable names as chromosomes. Firstly GEP is produced randomly program depends on this operators and data sets. The derived programs mutate until the program with best tness value among them is found. The program with the best tness value is taken as the last result. The results can be compared by applying the mathematical relation obtained from the program with the best tness value in the proposed computer program [10]. Crossover is determined by choosing two ET (expres- sion tree) based on tness and generating for each ET the crossover point (node) at random. For example: consider the following ETs (Fig. 4) with crossover points 2 and 3. The sub-tree of ET 1 starting from crossover point 2 will be swapped with the sub-tree of ET 2 at crossover point 3. In this study 80% of test data are se- lected for training GEP algorithm and the other 20% data are selected for test value. Table 2 Experimental cutting tool Spindle speed (rpm) Feed (mm/min) Depth of cut (mm) 750 152 0.25 1000 228 0.76 1250 305 1.26 1500 381 457 533 610 Table 3 Workpiece composition Workpiece Composition %Component Value Aluminum 6061-T8 Dimensions: 35 50 30 mm Aluminum, Al 98 Chromium, Cr 0.040.35 Copper, Cu 0.150.4 Iron, Fe 0.7 Magnesium, Mg 0.8 Using area: Aircraft ttings, camera lens mounts, couplings, hydraulic pistons, appliance ttings, valves and valve parts, etc. Manganese, Mn 0.15 Silicon, Si 0.40.8 Titanium, Ti 0.15 Zinc, Zn 0.15 Fig. 5. GEP algorithm model of experimental surface roughness study. Table 4 Used GEP parameter Experimental GEP optimization values Settings Number of terminals: 3 Number of Chromosomes: 50 Number of training values: 56 Number of genes: 7 Number of test values: 28 Head size: 8 Linking function: + Mutation rate: 0.044 One-point recombination rate: 0.3 Two-point recombination rate: 0.3 Gene recombination rate: 0.1 Gene transposition rate: 0.1 IS transposition rate: 0 Number of IS elements: 0 IS elements: 0 RIS transposition rate: 0 Number of RIS elements: 0 RIS elements: 0 Error type: relative Precision: 0 Selection range: 100 Fitness cases: 56 Max. tness: 5600 O. C olak et al. / Materials and Design 28 (2007) 657666 661 5. Implementation of surface roughness to the GEP The experimental design of this study consisted of cutting parameters is given in Table 2. Experimental val- ues are consisted of 84 pieces of machined part. Surface roughness value (Ra) of each machined part measured by Hommel Verke T500 Stylus prolometer. Hardware of this study: Hardford VMC 1020 Vertical Machining Centre. Computer boards DAS-6071 32/8 DAQ A/D PCI Converter CARD. Pentium III, 800 MHz PC computer. 0 1 2 3 4 5 6 150 200 250 300 350 400 450 500 550 600 650 Feed (mm/min) S u r f a c e
R o u g h n e s s
( m i c r o m m ) D.O.C=0.25 mm Expr. Ra D.O.C=0.25 mm GEP Ra D.O.C=0.76 mm Expr. Ra D.O.C=0.76 mm GEP Ra D.O.C=1.27 mm Expr.Ra D.O.C=1.27 mm GEP Ra 0 1 2 3 4 5 6 150 200 250 300 350 400 450 500 550 600 650 Feed (mm/min) S u r f a c e
R o u g h n e s s
( m i c r o m m ) D.O.C=0.25 mm Expr. Ra D.O.C=0.25 mm GEP Ra D.O. C= 0. 76 mm Expr. Ra D.O.C=0.76 mm GEP Ra D.O.C=1.27 mm Expr .Ra D.O. C= 1. 27 mm GEP Ra a b Fig. 6. Comparison of experimental and GEP predicted surface roughness value related to spindle speed and DOC. (a) Experimental and GEP predicted surface roughness values (spindle speed = 750 rpm). (b) Experimental and GEP predicted surface roughness values (spindle speed = 1000 rpm). (c) Exprimenatl and GEP predicted surface roughness values (spindle speed = 1250 rpm). (d) Experimental and GEP predicted surface roughness values (spindle speed = 1500 rpm). 662 O. C olak et al. / Materials and Design 28 (2007) 657666 Software of this study: NET Platform for collecting signal and processing. Automatic Problem Solver Academic version V2.0 for gene expression programming running. Aluminum 6061 T8 is used for workpiece. Specica- tion of workpiece is given in Table 3. Ten millimeter diameter HSS end-mill cutting tool used for this study. A new end mill is used after each experiment. A cutting tool is used for only nine pieces of workpiece to eliminate tool wear eect. There fore nine new tools are used for this study. The view of experimental study is shown in Fig. 4. Measured surface roughnesses of parts are used for output of algorithm. 84 experimental data is used to establish GEP algorithm. Eighty percent of experimen- tal data is used for training of system the other values used for testing. GEP modeling of surface roughness prediction is shown in Fig. 5. GEP algorithm uses some setting parameters such as chromosomes, gene, mutation rate, one-point recombi- nation rate, etc. Setting parameter of this study shown in Table 4. Experimental data has been evaluated in APS 2.0 software and surface roughness function derived for end-milling operations. To derive surface roughness function for end milling with 92% straightness is derived using GEP algorithm at the end of 1,121,737 iterations of 50 chromosomes size. The GEP function after optimization of GEP operator parameters are given below. The C++ code which is produced after the training using GEP algorithm: 0 1 2 3 4 5 6 150 200 250 300 350 400 450 500 550 600 650 Feed (mm/min) S u r f a c e
R o u g h n e s s
( m i c r o m m ) D.O.C=0.25 mm Expr. Ra D.O.C=0.25 mm GEP Ra D.O.C=0.76 mm Expr. Ra D.O.C=0.76 mm GEP Ra D.O.C=1.27 mm Expr.Ra D.O.C=1.27 mm GEP Ra 0 1 2 3 4 5 6 150 200 250 300 350 400 450 500 550 600 650 Feed (mm/min) S u r f a c e
R o u g h n e s s
( m i c r o m m ) D.O.C=0.25mm Expr. Ra D.O.C=0.25 mm GEP Ra D.O.C=0.76 mm Expr. Ra D.O.C=0.76mm GEP Ra D.O.C=1.27 mm Expr.Ra D.O.C=1.27 mm GEPRa c d Fig. 6 (continued) O. C olak et al. / Materials and Design 28 (2007) 657666 663 0 1 2 3 4 5 6 150 200 250 300 350 400 450 500 550 600 650 Feed (mm/min) S u r f a c e
R o u g h n e s s
( m i c r o m m ) Spindle=750 rpm Expr. Ra Spindle=750 mm GEP Ra Spindle=1000 rpm Expr. Ra Spindle=1000 rpm GEP Ra Spindle=1250 rpm Expr.Ra Spindle=1250 rpm GEP Ra Spindle=1500 rmp Expr. Ra Spindle=1500 rpm GEP Ra 0 1 2 3 4 5 6 150 200 250 300 350 400 450 500 550 600 650 Feed (mm/min) S u r f a c e
R o u g h n e s s
( m i c r o m m ) Spindle=750 rpm Expr. Ra Spindle=750 mm GEP Ra Spindle=1000 rpm Expr. Ra Spindle=1000 rpm GEP Spindle=1250 rpm Expr.Ra Spindle=1250 rpm GEP Ra Spindle=1500 rmp Expr. Ra Spindle=1500 rpm GEP 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 150 200 250 300 350 400 450 500 550 600 650 Feed (mm/min) S u r f a c e
R o u g h n e s s
( m i c r o m m ) Spindle=750 rpm Expr. Ra Spindle=750 mm GEP Ra Spindle=1000 rpm Expr. Ra Spindle=1000 rpm GEP Ra Spindle=1250 rpm Expr.Ra Spindle=1250 rpm GEP Ra Spindle=1500 rmp Expr. Ra Spindle=1500 rpm GEP Ra a b c Fig. 7. Comparison of experimental and GEP predicted surface roughness value related to dierent spindle speed and DOC. (a) Experimental and GEP predicted surface roughness values (DOC = 0.25 mm). (b) Experimental and GEP predicted surface roughness values (DOC = 0.76 mm). (c) Experimental and GEP predicted surface roughness values (DOC = 1.26 mm). 664 O. C olak et al. / Materials and Design 28 (2007) 657666 where; input variables which is in code: d[0] = cutting speed, d[1] = feed rate, d[2] = DOC, output of this function gives predicted surface roughness. 6. Results and conclusion By using GEP algorithm non-linear relation between cutting forces and cutting parameters could be expressed as a function. This would facilitate prediction of surface roughness and cost in manufacturing experimental data is evaluated in GEP algorithm. Then the function in C++ language is determined. Results of experimental data and values that function evaluated are compared in graphical representations. The evaluated C++ func- tion, gives the relation between cutting parameters and surface roughness in a high accuracy, as a rate of 91%. This study simplies surface roughness process monitoring. In Fig. 6, the eect of feed rate on surface roughness is given for various cutting speed (750, 1000, 1250 and 1500 rpm) and at depth of cut at 0.25, 0.76 and 1.27 mm. Due to this graphic in high cutting speed val- ues, there is good relation between functional data and experimental data, but in lower cutting speed values some dierences are seen in this relation. In nish pro- cess of machining operations, where low surface rough- ness is aimed, always high cutting speed values is preferred, therefore the tting problem of the function determined and experimental data is not much impor- tant at all. High precision tting of GEP algorithm and experimental data on surface roughness related to cutting speed will decrease need have high cost experi- mental studies. Some dierences is shown between experimental and GEP predicted in low spindle speed and DOC. But there is a good correlation in high speed and high DOC as seen in Fig. 7. In manufacturing environment prediction of surface roughness is important for product quality and produc- tion time. In this study, as a conclusion, using GEP algorithm, surface roughness prediction has been done using a few experimental data. GEP is coming from its ability to generate mathematical equations that can be easily programmed even into programming for use in monitoring of surface quality. Acknowledgement This study is supported by Prime Ministry State Plan- ning Organization (TR) (Project Nos: DPT 2003 K121020 and 98K123040). References [1] Chen JC, Huang LH, Lan AX, Lee S. Analysis of an eective sensing location for an in-process surface recognition system in turning operations. J Ind Technol 1999;15(3). [2] Lee BY, Tarng YS. Surface roughness inspection by computer vision in turning operations. Int J Mach Tools Manuf 2001;41:125163. [3] Kiran MB, Ramamoorthy B, Radhakrishnan V. Evaluation of surface roughness by vision system. Int J Mach Tools Manuf 1998;38(5):68590. [4] Yang JL, Chen JC. A systematic approach for identifying optimum surface roughness performance in end-milling opera- tions. J Ind Technol 2001;17(2). [5] Kline WA, DeVor RA, Shareef IA. The prediction of surface geometry in end milling. ASME J Eng Ind 1982;104:2728. [6] Alauddin M, El Baradie MA, Hashmi MSJ. Computer-aided analysis of a surface-roughness model for end milling. J Mater Process Technol 1995;55:1237. [7] Alauddin M, El Baradie MA, Hashmi MSJ. Prediction of tool life in end milling by response surface methodology. J Mater Process Technol 1997;71:45665. [8] Fuh KH, Hwang RM. Predicted milling force model for high- speed end milling operation. Int J Mach Tools Manuf 1997;37(7):96979. double APSCfunction(double d[]) { double dblTemp = 0; dblTemp += ((d[1] 1/((sqrt(d[0]) + pow(d[2],d[1])))) d[1]); dblTemp += (tan(1/((d[0]/d[1])))* log 10((cos(d[2]) * (d[1] + d[1])))); dblTemp += tan(tan(((tan(d[2])/(d[1]/d[2]))/tan(sin(d[1]))))); dblTemp += 1/((1/(fmod((d[1] d[2]),log(d[2]))) > oor(pow(d[0],d[2]))?1/(fmod((d[1] d[2]),log(d[2]))): oor(pow(d[0],d[2])))); dblTemp += sin(cos(tan(log10(((log(d[2]) + d[1])*(d[0] + d[0])))))); dblTemp += tan(log10(1/((fmod((d[0]+d[0]),(d[1] + d[1])) + (d[0] d[2]))))); dblTemp += sin(sqrt((d[0]/(((d[1]*d[2])*d[2]) + ceil(d[1]))))); return dblTemp; } O. C olak et al. / Materials and Design 28 (2007) 657666 665 [9] Lin J. Inverse estimation of the toolwork interface temperature in end milling. Int J Mach Tools Manuf 1995;35(5):75160. [10] Chiang ST, Liu DI, Lee AC, Chieng WH. Adaptive control optimization in end milling using neural networks. Int J Mach Tools Manuf 1995;35(4):63760. [11] Luo T, Lu W, Krishnamurthy K, McMillin B. Neural network approach for force and contour error control in multi-dimen- sional end milling operations. Int J Mach Tools Manuf 1998;38:134359. [12] Alauddin M, El Baradie MA, Hashmi MSJ. Modelling of cutting in end milling Inconel. J Mater Process Technol 1996;58:1008. [13] Lou SJ, Chen JC. In-process surface roughness recognition (ISRR) system in end-milling operation. Int J Adv Manuf Technol 1999;15:2009. [14] Ferreira C. Gene expression programming: new adaptive algo- rithm for solving problems. Complex Systems 2001;13(2):87129. [15] Benardos PG, Vosniakos G. Predicting surface roughness in machining: a review. Int J Mach Tools Manuf 2003;43:83344. 666 O. C olak et al. / Materials and Design 28 (2007) 657666