VOL. 7, NO. 4, APRIL 2012 ISSN 1819-6608 ARPN Journal of Engineering and Applied Sciences ©2006-2012 Asian Research Publishing Network (ARPN). All rights reserved. www.arpnjournals.com ARTIFICIAL NEURAL NETWORK PREDICTIVE MODELING OF UNCOATED CARBIDE TOOL WEAR WHEN TURNING NST 37.2 STEEL T. B. Asafa1 and D. A. Fadare2 1Department of Mechanical Engineering, King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia 2Department of Mechanical Engineering, University of Ibadan, Ibadan, Nigeria E-mail: teslimtunde@yahoo.com ABSTRACT We report the development of a predictive model based on artificial neural network (ANN) for the estimation of flank and nose wear of uncoated carbide inserts during orthogonal turning of NST (Nigerian steel) 37.2. Turning experiments were conducted at different cutting conditions on a M300 Harrison lathe using Sandvic Coromant uncoated carbide inserts with ISO designations SNMA 120406 using full factorial design. Cutting speed (v), feed rate (f), depth of cut (d), spindle power (W), and length of cut (l) were the input parameters to both the machining experiments as well as the ANN prediction model while the flank wear (VB) and nose wear (NC) were the output variables. Nine different structures of multi-layer perceptron neural networks with feed-forward and back-propagation learning algorithms were designed using the MATLAB Neural Network Toolbox. An optimal ANN architecture of 5-12-4-2 with the Levenberg-Marquardt training algorithm and a learning rate of 0.1 was obtained using Taguchi method of experimental design. The results of ANN prediction show that the model generalized well with root mean square errors (RMSE) of 3.6% and 4.7% for flank and nose wear, respectively. With the optimized ANN architecture, parametric study was conducted to relate the effect of each turning parameters on the tool wear. The ANN predictive model captures the dynamic behaviour of the tool wear and can be deployed effectively for online monitoring process. Keywords: model, ANN, carbide inserts, Taguchi method, tool wear, NST 37.2 steel, turning, cutting speed, machining. 1. INTRODUCTION In the present study, ANN is adopted because of NST 37.2 is a grade of Nigerian commercial its several advantages. Among these is its capability to steels produced by the Delta Steel Company (Asafa, learn arbitrary nonlinear mappings between noisy sets of 2007). The steel is commonly deployed for the production input and output data and predicting, with substantial of machine components and sometimes as structural accuracy, complex data interactions (Umbrello, et al., members in building construction and other architectural 2008. ANN differs from the traditional modeling edifices. With its wide applications in machining approaches in that it is trained to learn solutions rather industries and the requirements for various machining than being programmed to model a specific problem operations to produce the desired end results, it becomes (Bhatikar and Mahajan, 2002). Also, it is usually used to important to establish an optimized model for the address problems that are intractable or cumbersome to prediction of tool wear during turning of this steel. solve with traditional methods. A number of applications Accurate prediction of the tool wear conditions is an of ANN in tool conditioning monitoring and prediction of essential prerequisite for reliable on-line tool condition tool wear and tool life during non-orthogonal machining monitoring system (Mursec and Cus, 2003; Cus et al., has been reported (Elanayar and Shin, 1990; Elanayar and 1997) and such a system can be deployed for effective tool Shin, 1992; Ghasempoor et al., 1999; Sick, 2002; Dimlar wear monitoring in our local machine tools industries. et al., 1997) or orthogonal turning (Li et al., 1999; Tansel, Without doubt, modern machining system requires tool et al., 2000; and Dimlar, et al., 1998). ANN has equally wear monitoring and prediction systems for higher quality been used for monitoring surface roughness (Asilturk and production. In precision machining, the surface quality of Cunkas, 2011) and induced residual stress (Umbrello, et the manufactured part can be related to tool wear which al., 2008). In addition, ANN has found huge applications contributes to the increase in the industrial interest for in- in other areas of industrial technology including process tool wear monitoring systems. semiconductor industries (Chen et al., 2007), One of the available methods is by the application transportation (Asafa et al., 2010) among others. ANN is of ANN. Prediction of tool wear/tool life and tool often implemented via back propagation, a gradient condition monitoring has been extensively studied using descent algorithm in which the network weights are ANN by many researchers (Sick, 2002). Sunil and Sandra moved along the gradient of the performance function. (2000) considered neural network as a parallel processing The algorithm computes the weights in the network so as architecture in which knowledge is represented in the form to minimize the output error in a least-squared sense of weights between highly interconnected processing (Howard and Mark, 2005). elements. More details of ANN can be found elsewhere In the past, Boothroyd and Knight (1999) had (Ozel and Nadgir, 2002). observed that wear in metal cutting could be in the form crater (nose wear) or flank (as shown in Figure-1). Crater 396 UNIVERSITY OF IBADAN LIBRARY VOL. 7, NO. 4, APRIL 2012 ISSN 1819-6608 ARPN Journal of Engineering and Applied Sciences ©2006-2012 Asian Research Publishing Network (ARPN). All rights reserved. www.arpnjournals.com wear is the limiting factor for the tool life under very high widely applied, it is however difficult to implement for on- speed cutting conditions because the wear is usually line sensing because of the inaccessibility of the tool severe that the tool edge is weakened and consequently surface during machining (Xu, 2009). fractured. Crater wear occurs on the rake face of the tool, Indirect method of tool wear estimation has been changing the tool-chip interface geometry, thus negatively extensively studied as a solution to the shortcomings of the affecting the cutting process. The most significant factors direct method most especially for in-process monitoring. influencing the crater wear are the temperature at the tool- Matins et al. (1984) used vibration generated during chip interface and the chemical affinity between the tool cutting process to reveal the state of the tool wear. The and the work piece materials (Ezugwu et al., 2005). Flank change in the flank wear has also been monitored by wear occurs on the relief face of the cutting tool and is means of reduction in the work piece dimensions (El- generally attributed to the rubbing of the tool along the Gomayel and Breggar, 1986). The use of acoustic machined surface. At high temperatures, abrasive and/or emission due to change in sound intensity as well as adhesive processes are accelerated, thus affecting tool optical method that relies on the changes in the reflectance material properties as well as work piece surface. Flank characteristics of the worn tool surfaces has also been wear is a mechanically activated wear usually by the reported (Kannately-Asibu and Dornfeld, 1992). Also, abrasive action of the cutting tools on the work piece spindle motor current and power have been used to material (Boothroyd and Knight, 1999). The severity of estimate tool wear for on-line monitoring (Martins et al., abrasion increases in cases where the work piece materials 1984). Literature shows that tool wear prediction during contain hard inclusions, or when there is hard wear debris machining of a wide range of steel grades have been from the work piece or the tool, at the interface (Ozel and reported (Ezugwu et al., 2005). However, it appears that Karpat (2005). Flank wear increases with increase in no study has been conducted in the area of machining of cutting time as well as increases in the axial cutting NST 37.2. Therefore, developing an optimized ANN- distance (Ozel and Nadgir, 2002). The nature of this based predictive model for estimation of flank and nose relationship depends on material and process condition. wear of uncoated carbide cutting tools during turning of NST 37.2 is the objective of this study. It is our opinion that these findings will assist machinists in acquiring a priori knowledge of the magnitude of tool wear without the usual experimental trials and errors approach. 2. MATERIALS AND METHODS NST 37.2 samples were used as the work piece material for the orthogonal turning experiment in a full factorial design. The outputs of the experiment were used to construct the predictive model while the architecture of the model was optimized via Taguchi approach using signal-to-noise ratio and analysis of variance. Details of these steps are discussed in this section. Figure-1. Major regions of tool wear during metal cutting 2.1 Work piece material, cutting tool and tool holder (Boothroyd and Knight, 1999). Samples of fully annealed NST 37.2 steel bars with 25 mm diameter were obtained from Delta Steel Tool wear are measured directly or indirectly Company (DSC). The chemical composition and the (Cuneyt, 2009; Xu et al., 2011). Direct measurement is mechanical properties of the steel sample are given in usually carried out by means of optical, radiometric, Tables 1 and 2, respectively. Uncoated cemented carbide pneumatic or contact sensors in which tool wear is inserts produced by Sandvic Coromant® with ISO measured in term of material loss (Sunil Elanayar and designation SNMA120406 were used as the cutting tool. Sandra, 2000). This technique can be effectively deployed The insert had a square shape with zero clearance angle for on-line measurement. The application of the tool and inbuilt chip breakers. It was rigidly mounted on a tool maker microscopes and the radiometric method help us to holder with ISO designation PSBNR 3225P15 while the measure the tool wear directly (Ozel and Nadgir, 2002). holder was clamped to the tool post in an orthogonal For example, Ezugwu et al. (2005) measured tool wear on arrangement. ceramic inserts during machining of Inconel 718 by means of a travelling microscope. Though direct method is Table-1. Chemical composition of NST 37.2 (Asafa, 2007). Element C S Si Mn P Fe Composition (%W) 0.33 0.01 0.150 0.69 0.02 98.80 397 UNIVERSITY OF IBADAN LIBRARY VOL. 7, NO. 4, APRIL 2012 ISSN 1819-6608 ARPN Journal of Engineering and Applied Sciences ©2006-2012 Asian Research Publishing Network (ARPN). All rights reserved. www.arpnjournals.com Table-2. Mechanical properties of NST 37.2 2.3.1 Data collection and processing (Asafa, 2007). Input data into the neural networks were obtained from the machining parameters used in the full factorial Properties Average value experiments. 112 input/output data were used for the Yield strength (MPa) 245 network training, model validation and testing in a relative Tensile strength (MPa) 342 proportion of 2:1:1. The input vectors were first normalized with Matlab function, prestd, in order to obtain Elongation (%) 18 inputs with zero mean and unity variance. In addition, Reduction in area (%) 15 principal component analysis was done with Matlab function, prepca, to eliminate those components that Young modulus (GPa) 199 contribute less than 99% to the variation in the datasets. Brinell Hardness 49 The outputs of the network were later converted to the 3 original data format with Matlab function postd. Density (g/m ) 8.15 2.3.2 Neural network design and optimization 2.2 Machining experiment The design of the network architecture requires Straight turning was done on M300 Harrison-type the selection of a number of hidden layers and those of the lathe driven by 3.0 Hp Kapak inductions motor with speed neurons in each of the hidden layer, the training algorithm range of 40-2500 rpm. Cutting conditions typical of those and the learning rate that would minimize the prediction available in the machining industries were used for the error. Figure-2 shows a general architecture of the ANN machining trials. These cutting conditions are listed in model having five inputs and two outputs as used in this Table-3. The cutting parameters - cutting speed (v); feed work. Taguchi method is used to optimize the number of rate (f) and depth of cut (d) - were investigated at three neurons in each of the hidden layers, the training different levels in a 33 full factorial experiment. Full algorithm and the learning rate based on the method factorial design is chosen to study the interactions between proposed by Khaw et al. (1995). Three training algorithms the turning parameters. which are Levenberg-Marquardt algorithm, Scaled Conjugate Gradient and Bayesian Regularization were Table-3. Summary of turning conditions. included in the optimization because of their similar performance in some previous studies (Bealle at al., 2010). Cutting parameters Two hidden layers were proposed since ANN model of Level Cutting speed Feed rate Depth of one layer is usually too weak to accurately predict non- (m/min) (mm/rev) cut (mm) linear function (Kaw et al., 1995). 1 20.4 1.0 0.2 2 29.1 1.8 0.4 Cutting speed (v) 3 42.4 2.2 0.8 Feed rate (f) Flank wear For each of the machining conditions, eight Depth of cut (d) passes of 50 mm length of cut were made. At the end of Nose wearLength of cut (l) each pass, the spindle motor current, nose wear and flank wear were measured. Spindle current was measured with Spindle power (P) digital multimeter connected to the electric motor of the lathe while the nose and flank wear were measured by Input Hidden Output means of a machine vision system earlier developed at the layer layer layer University of Ibadan, Nigeria (Oni, 2007). The tool Figure-2. Schematic illustration of the ANN structure. rejection criteria for roughing operation according to ISO 3685 Standard were used. The criteria directed that an The four important parameters of ANN model are insert must be rejected and further machining discontinued arranged using the Taguchi’s orthogonal array (Taguchi, when any or combination of the following criteria is 1993). These parameters are: (A) number of neurons in reached: flank wear ≥ 0.7 mm, nose wear ≥ 0.5 mm, or hidden layer 1, (B) number of neurons in hidden layer 2, catastrophic failure. These values serve as constrains into (C) type of training algorithm and (D) learning rate. The the predictive model. number of neurons in each of the hidden layers is selected based on the mathematical relationship presented by Chen 2.3 Neural network model description et al. (2007) (Table-4). Here, we discuss the processes involved in the Taguchi approach is essentially a statistical data collection, pre-processing and partitioning of data as technique used in experimental study for analyzing the the preliminary stages in ANN development. Then, the relationship between large numbers of design parameters steps in the architecture design, training as well as testing with the smallest number of experimental runs (Chen et of the neural network performance are explained. al., 2007). Taguchi used an engineering approach to plan 398 UNIVERSITY OF IBADAN LIBRARY VOL. 7, NO. 4, APRIL 2012 ISSN 1819-6608 ARPN Journal of Engineering and Applied Sciences ©2006-2012 Asian Research Publishing Network (ARPN). All rights reserved. www.arpnjournals.com and design optimum experimental runs using orthogonal OA and analyze the results using signal-to-noise ratio arrays and signal-to-noise ratios. His method is widely (S/N) approach to determine the optimal combination of applicable because of its strengths; it has however been parameters and using analysis of variance (ANOVA) to criticized for the poor interaction among the processing rank the parameters in order of influence/significance and variables. A reviewed article on the strengths and (4) conducting a confirmatory experiment using the limitations of Taguchi’s experimental design approach had optimal ANN architecture (Hinkelmann and Kempthorne, been published (Maghsoodloo et al., 2004). Generally, 2005). The viability of this approach has been implementation of Taguchi method requires four basic demonstrated for selection of ANN parameters for steps. These include (1) brainstorming on the design designing high quality and robust networks (Asilturk and parameters that are important to the process and identify Cunkas, 2011; Chen et al., 2007). In this work, the number the factors as well as the levels of each factor, (2) selecting of input variables (N) is 5 and that of the output variable the appropriate orthogonal array (OA) from the published (OP) is 2, each with three levels. The results are presented table, (3) conducting experiments based on the selected in Table-5. Table-4. Estimation of number of neurons in the hidden layers. ANN parameter Level Number of neurons in the Number of neurons in the hidden hidden layer 1 layer 2 Level 1 N OP N + OP+ 2 2 N + 1 Level 2 2 N + 1 2 N + 1 + 3 OP x (N +1) OP x (N 1) OP x (N + 1)Level 3 + + 3 Table-5. Factor and level for the ANN parameters. Factor Level Number of neurons in Number of neurons in Training Learning the hidden layer 1 (A) the hidden layer 2 (B) algorithm (C) rate (D) 1 3 4 LM 0 2 11 15 RP 0.05 3 12 16 SCG 0.1 Table-6. L9 (34) orthogonal arrays and S/N ratio. Test run Orthogonal array S/N ratio 1 A1 B1 C1 D1 -25.766 2 A1 B2 C2 D3 -31.496 3 A1 B3 C3 D2 -57.066 4 A2 B1 C2 D2 -23.837 5 A2 B2 C3 D1 -29.758 6 A2 B3 C1 D3 -21.393 7 A3 B1 C3 D3 -22.726 8 A3 B2 C1 D2 -24.123 9 A3 B3 C2 D1 -23.897 399 UNIVERSITY OF IBADAN LIBRARY VOL. 7, NO. 4, APRIL 2012 ISSN 1819-6608 ARPN Journal of Engineering and Applied Sciences ©2006-2012 Asian Research Publishing Network (ARPN). All rights reserved. www.arpnjournals.com Table-7. S/N ratio and ANOVA. Level (dB) Range Rank DOF1 SS2 Variance % contr3 1 2 3 A -38.109 -24.99 -23.58 14.52 1 2 384.98 192.49 39.34 B -24.109 -28.46 -34.12 10.01 4 2 151.14 75.57 15.44 C -23.761 -26.41 -36.52 12.76 2 2 271.87 135.96 27.78 D -26.473 -35.01 -25.22 9.80 3 2 170.58 85.29 17.43 Total 8 7601.54 100 Figure-3. Response chart for the S/N ratio. Table-8. The optimized ANN model. Network Training Learning MSE MSE architecture algorithm rate (flank (nose wear) wear) 5-12-4-2 LM 0.1 0.0365 0.0467 Table-6 shows the arrangement of 4-factor, 3- reaches 10-10, (b) when the number of iterations reaches level design proposed to determine the effect of ANN 2000 and (c) when validation data begin to over fit. A variables on the network performance. For this type of an typical example of the convergence of the testing and arrangement, 34 (or 81) sets of experimental runs are training networks for flank wear is shown in Figure-4. The needed for full factorial design while only 9 suffice based best validation performance of 0.041 is obtained after 27 on Taguchi method. For every experimental run, the iterations. The coefficients of regression are 0.98 and 0.99 training section is terminated if one of the following for training and testing data, respectively. These stopping criteria is reached: (a) when the mean square coefficients are statistically satisfactory. error (MSE) between the actual and predicted output 400 UNIVERSITY OF IBADAN LIBRARY VOL. 7, NO. 4, APRIL 2012 ISSN 1819-6608 ARPN Journal of Engineering and Applied Sciences ©2006-2012 Asian Research Publishing Network (ARPN). All rights reserved. www.arpnjournals.com (a) (b) Figure-4. Network performance (a) Network validation with number of epochs (b) Regression analysis for training and validation datasets for flank wear. The MSE of the nine ANN architectures are taken This is done by taking the average value of the S/N ratios as the outputs of the experimental runs and as the inputs to for each level and the corresponding factor level. The the S/N ratios. Depending on the optimization results are presented as a response table (Table-7) and a requirements, different S/N ratios may be applicable, response graph (Figure-3). The optimum parameter including “lower is better” (LB) - minimum performance combination is A3B1C1D3 as highlighted in Table-7. Under characteristics, “nominal is best” (NB) - medium this condition, the BPNN architecture is 5-12-4-2 with a performance characteristics, and “higher is better” (HB) - Levenberg-Marquardt training algorithm and a learning maximum performance characteristics. Since the value of rate of 0.1 as presented in Table-8. The relative MSE is desired to be as small as possible for both flank contribution of each ANN parameter on the performance and nose wear, LB is selected. The S/N ratio is therefore characteristic of the predictive model expressed as a calculated from the Eq. (1) (Chen et al., 2007). The results percentage is obtained via ANOVA (Hsu et al., 2008). The are presented in Table-6. results, shown in Table-7, indicate that the first hidden layer is the most significant parameter contributing ~39% ⎛ N i y 2 ⎞ to the change in the network performance while the second (S / N ) i = −10 log ⎜⎜∑ j ⎟ ⎟ (1) N hidden layer is the least with ~15%. The learning rate and ⎝ u =1 i ⎠ the training algorithm contribute ~17% and ~27%, Where i = experimental number, j = number of trials, N = respectively. i number of trial for experiment i and y is the trial output. After estimating the S/N ratio for each of the 3. RESULTS AND DISCUSSIONS experimental runs (Table-6), the average S/N value is Once the optimal level of the design parameters calculated for each factor and the corresponding level. has been selected with Taguchi method, we thereafter run 401 UNIVERSITY OF IBADAN LIBRARY VOL. 7, NO. 4, APRIL 2012 ISSN 1819-6608 ARPN Journal of Engineering and Applied Sciences ©2006-2012 Asian Research Publishing Network (ARPN). All rights reserved. www.arpnjournals.com a confirmatory test to verify the improvement of the rest. We thereafter used the optimal ANN architecture to performance characteristic using the optimal level of the predict the outcome of the testing data. The result of the design parameters for the flank wear. The optimal model is now compared with the experimental value. combination of the factors level is the test with highest value of S/N ratio. The estimated optimum output 3.1 Model validation y predicted using the optimal level of the design Three typical cutting conditions were selected to test the accuracy of the model. The results of these tests parameters is calculated from (Lin and Chang, indicate that ANN is a viable tool for prediction of tool 2003). behaviour. Figure-5(a) shows that the flank and nose wear k ( ) increased as the cutting length increased. This behaviour is y predicted = y m + ∑ y i − y m (2 ) well captured by the ANN model. It is however obvious i =1 that increase in flank wear is greater than that of the nose y y wear. This observation can be attributed to the tool nose Where m is the global mean S/N ratio, i is the mean radius (0.6 mm) being less than the depth of cut (0.8 mm). S/N ratios at the optimal level, and k is the number of the The analogy can be correlated to the direct proportionality design parameters. Equation (2) gives -~20dB which is between flank wear and the depth of cut as presented by greater than -~21dB, the maximum value obtained from Yongjin and Fischer (2002). However, the nose wear the orthogonal array. Experiment conducted at the exceeded that of flank wear at the condition cutting optimum design parameters gives S/N ratio of -~20dB condition when v = 20.42 mm/s, f = 2.2 mm/rev and d = which is the same as the predicted value and higher than - 0.4 mm as showing in Figure-5(b). This behaviour is 21dB. The average prediction errors were 3.7% and 4.7% attributed to the depth of cut (0.4 mm) being smaller than for flank and nose wear respectively. This gives a good the nose radius which subsequently lead to the partial confidence that the optimal parameters are truly optimal. engagement of tool nose during the turning operation. At With this level of accuracy, the model performance is higher cutting speed (42.42 mm/s), the flank wear satisfactory. The higher MSE of the nose wear is attributed becomes higher than the nose wear Figure-5(c) due to high to the fact that the two outputs were being predicted temperature generated at high cutting speed. Such a high simultaneously and in such a case the output of the ANN temperature could easily weaken the tool materials and model is often more accurate for the first output than the thereby enhances tool wear. 0.8 0.08 0.9 1 0.7 0.07 0.8 0.9 0.6 0.06 0.7 0.8 0.7 0.5 0.05 0.6 0.6 0.4 (a) 0.04 0.5 (b) 0.5 0.4 0.3 0.03 0.4 Experimental f lank w ear 0.3 0.2 0.02 Experimental f lank w ear 0.3Predicted Flank w ear 0.2 Experimental Nose w ear Predicted f lank w ear0.1 0.01 0.2 Predicted Nose Wear 0.1 Experimental nose w ear 0 0 Predicted nose w ear 0.1 0 0 0 200 400 600 0 100 200 300 400 500 Cutting length (mm) Cutting length (mm) 2 0.7 Experimental f lank w ear 1.8 Predicted Flank w ear 0.6 1.6 Experimental Nose w ear 1.4 Predicted Nose w ear 0.5 1.2 0.4 1 0.8 0.3 0.6 0.2 0.4 (c) 0.1 0.2 0 0 0 100 200 300 400 500 Cutting length ( mm) Figure-5. Comparison of ANN prediction and measured tool wear for various cutting conditions (a) v = 20.42 mm/s, f = 1.0 mm/rev, d = 0.8 mm (b) v = 20.42 mm/s, f = 2.2 mm/rev, d = 0.4 mm and (c) v = 42.42 mm/s, f = 2.2 mm/rev, d = 0.8 mm The feature that is common to both the wear violates the standard conditions for tool disposal. experiment and the ANN predictive model is the ability to Few machining experiments were conducted such that the capture the cutting conditions where the resulting tool resulting tool wear were higher than the accepted values. 402 U FNlank Wear (mm) Flank Wear (mm)IVERSITY OF Nose wear (mm) Nose wear (mm) I Flank Wear ( mm) BADAN LIBRAR Nose Wear (mm) Y VOL. 7, NO. 4, APRIL 2012 ISSN 1819-6608 ARPN Journal of Engineering and Applied Sciences ©2006-2012 Asian Research Publishing Network (ARPN). All rights reserved. www.arpnjournals.com In the same manner, the predictive model equally predicts temperature. The tool wear are still within the acceptable the cutting conditions at which tool wear constrains are level for the cutting conditions shown in Figure-6. violated. This will surely guide machinist on the selection Generally, when one of the cutting parameters is of the appropriate turning parameters without violating the increased, wear mechanism due to diffusion and adhesion wear constraints. is activated. A similar correlation between feed rate and dept-of-cut, and process parameters such as cutting force, 3.2 Effects of cutting parameters on tool wear flank and nose wear has been reported (Ezugwu et al., Cutting speed is one of the most important 2005). The results of the ANN prediction for the influence parameters that influence the development of tool wear of feed rate on tool wear are shown in Figure-7 (a) and (b). during machining. Figure-6 (a) and (b) show the Increase in feed rate raises the thermal state of the tool progressive increase in both the flank and nose wear with with subsequent softening and eventual rise in the wear increase in cutting speed within the experimental rate. The condition for continuous tool application is only consideration. Both wear types are linearly related to the satisfied at feed rate of 1 mm/rev for both wear. Higher cutting speed with the correlation coefficients of 0.99 for wear is recorded for a feed rate greater than 1 mm/rev. both wear. The increase in tool wear at higher cutting This behaviour is also confirmed by Sivasakthivel et al. speed can be explained by the enhancement of tool (2010). material diffusion and thermal stress inducement at higher (a) (b) Figure-6. Effects of cutting speed on tool wear (a) flank wear and (b) nose wear. 1.1 0.9 Flank 1 w ear,V=29.06mm/s,DOC=0.2mm, 0.8 L=300mm 0.7 0.9 (a) (b) Nose w ear, 0.6 V=29.06mm/s,DOC=0.2mm, L=300mm 0.8 0.5 0.7 0.4 0.3 0.6 0.5 1 1.5 2 2.5 0.5 1 1.5 2 2.5 Feed (mm/rev) Feed (mm/rev) Figure-7. Neural network prediction on the influence of feed on tool wear: (a) flank wear (b) nose wear. The effect of depth of cut (DOC) on tool wear is More study is required to identify the reasons for this shown in Figure-8 (a and b). The behaviour of tool wear behaviour. A typical tool wear image generated at cutting under DOC modulation is in two folds. Between 0.2 mm speed of 42.4 m/min is shown in Figure-2. Presence of and 0.4 mm, both the flank and nose wear decreased and chipping, attrition and abrasion were observed on the thereafter increased between 0.4 mm and 0.8 mm. Thus carbide tools during machining under the conditions the wear has minimum values at a depth of cut of 0.4 mm. investigated in this study. 403 UN Flank Wear (mm)IVERSITY OF I Nose w ear (m m ) BADAN LIBRARY VOL. 7, NO. 4, APRIL 2012 ISSN 1819-6608 ARPN Journal of Engineering and Applied Sciences ©2006-2012 Asian Research Publishing Network (ARPN). All rights reserved. www.arpnjournals.com 1 0.5 0.95 0.45 Nose0.9 Flank Wear, Wear,V=42.41mm/s,f=1.8mm/rev, 0.85 V=42.42mm/s,f=1.8mm/rev, 0.4 L=200mm L=200mm 0.8 0.35 0.75 0.3 0.7 0.25 0.65 (a) (b) 0.6 0.2 0.55 0.15 0.5 0.1 0 0.2 0.4 0.6 0.8 1 0.1 0.3 0.5 0.7 0.9 Depth Of Cut (mm) Depth Of Cut (mm) Figure-8. Effect of depth of cut on tool wear: (a) flank wear (b) nose wear. condition of v = 42.42 mm/s, f = 1.8 mm/rev and l = 200 mm. However, minimum wear were obtained for flank and nose wear at d = 0.8 mm and 0.2 mm respectively for condition v = 29.06 mm/s, f = 2.2 mm/rev and l = 400 mm. The ANN predictive model captures the dynamic behaviour of the tool wear and can be effectively deployed for online monitoring process. REFERENCES Asilturk I. and Cunkas M. 2011. Modeling and prediction of surface roughness in turning operations using artificial neural network and multiple regression method. Expert Systems with Applications. 38: 5826-5832. Figure-9. A typical flank wear image generated at cutting Asafa T.B. 2007. Investigation on tool wear prediction and speed of 42.4 m/min, feed rate f 1.8 mm/rev and depth-of- optimization of cutting conditions in Machining Aladja cut of 0.8 mm as observed on a tool insert NST 37.2. Unpublished M.Sc. Dissertation of Mechanical (Fadare and Asafa, 2009). Engineering Department, University of Ibadan, Nigeria. CONCLUSIONS Asafa T. B., Ajayeoba A. O. and Adekoya L. O. 2010. We have used ANN to create a predictive model Prediction of some anthropometric data of Bus Passengers for the estimation of flank and nose wear during using Artificial Neural Network: A case study of Molue machining NST 37.2. The optimized architecture, obtained bus in Lagos State, Nigeria. International Journal of with Taguchi experimental design, was 5-12-4-2 with the Biological and Physical Sciences. 15(3): 345-356. Levenberg-Marquardt training algorithm and a learning rate of 0.1. The results of ANN prediction show that the Beale M. H., Hagan M. T. and Dewuth H. B. 2010. Neural model generalized well with root mean square errors Network Tool: Users' Guide in MATLAB Documentation, (RMSE) of 3.55% and 4.67% for flank wear and nose by MATLAB. Mathworks Corporation. pp. 160-170. wear, respectively. With the optimized ANN architecture, parametric study was conducted to show the effects of the Bhatikar S. R. and Mahajan R. L. 2002. Artificial neural- cutting parameters on the tool wear. Generally, the network-based diagnosis of CVD barrel reactor. IEEE magnitude of flank wear was significantly higher than that Transactions on Semiconductor Manufacturing. 15(1): 71- of the nose wear in all the cutting conditions considered 78. except at cutting condition (v = 20.42 mm/s, d = 2.2 mm and f = 0.4 mm/rev) where the magnitude of the nose wear Boothroyd G. and Knight W.A. 1989. Fundamentals of nd was higher than that of the flank wear. From the machine and machine tools. 2 Ed. Marcel Dekker Inc. parametric study, increase in cutting speed accelerates tool New York, USA. p. 230. wear. Increase in the feed rate increased both flank and nose wear. Depth of cut significantly affects the tool wear. Caldas L. G. and Norford L. K. 2002. A design of A DOC of 0.4 mm produces least wear for cutting optimization tool based on a genetic algorithm. Automatic Constrain. 11(2): 173-184. 404 UN Flank wear ( mm)IVERSITY OF IBA Nose Wear (mm)DAN LIBRARY VOL. 7, NO. 4, APRIL 2012 ISSN 1819-6608 ARPN Journal of Engineering and Applied Sciences ©2006-2012 Asian Research Publishing Network (ARPN). All rights reserved. www.arpnjournals.com Cook D. F., Ragsdale C. T. and Major R. L. 2000. Goldgerg D.E. 1999. Genetic algorithm in search of Combining a neural network with a genetic algorithm for optimization and machine learning. Addison Wesley, process parameter optimization. Eng. Appl. Artif. Intell. USA. pp. 2-5. 13: 391-396. Howard D. and Mark B. 2005. Neural network tool box: Cuneyt A. 2009. Tool wear condition monitoring using a MATLAB, Language of Technical Computing. p. 120. sensor fusion model based on fuzzy inference system. Mechanical Systems and Signal Processing. 23(2): 539- Hsu C.Y., Ko T.F. and Huang Y.M. 2008. Influence of 546. ZnO buffer layer on AZO film properties by radio frequency magnetron sputtering. Journal of the European Cus F., Sokovic J., Kopac J. and Balic J. 1997. Model of Ceramic Society. 28: 3065-3070. complex optimization of cutting conditions. International Journal of material processing Technology. 64: 41-52. International Standard Organization (ISO). 1977. Tool life testing with single-point turning tool. International Daren Z. 2001. QSPR studies of PCBs by the combination Organization for Standardization. Ref No: ISO 3685-1977. of genetic algorithms and PLS analysis. Comput. Chem. 25: 197-204. Islier A. A. 1998. A genetic algorithm approach for multiple criteria facility layout design. Int. J. Prod. Res. Dimla D. E., Lister P. M. and Leighton N. J. 1997. Neural 36(6): 1549-1569. network solution to the tool condition monitoring problems in metal cutting - a critical review of methods. Kannately-Asibu J.E. and Dornfeld D.A. 1992. A study of International Journal of Machine Tool and Manufacture. tool using statistical analysis of metal cutting acoustic 37: 1219-1241. emission. Wear. 76: 247-261. Elanayar S.V.T. and Shin Y.C. 1990. Machining condition Khaw J. F. C., Lim B. S. and Lim L. E. N. 1995. Optimal monitoring for automation using neural networks. ASME design of neural network using the Taguchi method. Winter Annual Meeting, Dallas, TX, USA. pp. 85-100. Neurocomputing. 7: 225-245. Elanayar S.V.T. and Shin Y.C. 1992. Robust tool wear Lin Z. C. and Chang D. Y. 2003. Tool wear investigation estimation via radial basis function neural networks. on the precision progressive die for the IC dam-bar cutting ASME Winter Annual Meeting, Anaheim, C. A, USA. pp. process. Int. Journal of Advanced Manufacturing 37-51. Technology. 22: 344-356. El-Gomayel J. I. and Breggar K. D. 1986. On-line tool Li X.P., Iynkaran K. and Nee A.Y.C. 1998. A hybrid wear sensing for turning operations. Tans. ASME, J. Eng. machining simulator based on predictive machining theory Ind. 108: 44-47. and neural network modeling. Proceedings of CIRP International Workshop on Modeling and Machining Ezugwu E. O., Bonney J., Fadare D. A. and Wales W. F. Operations. Atlanta, Georgia, U.S.A. pp. 417-428. 2005. Machining of Nickel-based Inconel 718, alloy with ceramic tool under finishing conditions with various Martins K.F., Brandon J.A., Grrosvenor R.I. and Owen A. coolants Supply pressures. Journal of International 1984. A comparison of in-process tool wear measurement Processing Technology. 162-163: 609-614. methods in turning. Proc. 26th inter. MTDR Conf. p. 121. Ezugwu E.O., Fadare D.A., Bonney J., Silva R.B. and Mursec B. and Cus F. 2003. Integral model of selection of Wales W.F. 2005. Modelling the correlation between optimal cutting conditions from different databases of tool cutting and process parameters in high-speed machining of makers. International Journal of material processing Inconel 718 alloy using an artificial neural network. Technology. 133: 158-165. International Journal of Machine Tools and Manufacture. 45: 1375-1385. Oni A. O. 2007. Development of a vision system with respect to tool wear and surface roughness measurement. Fadare D.A and Asafa T. B. 2009. Performance evaluation Unpublished M.Sc. Dissertation. Mechanical Engineering of uncoated carbide. Proceedings of 22nd AGM and Department, University of Ibadan, Nigeria. International conference of the Nigerian Institute of Mechanical Engineers, Osun. pp. 20-24. Ozel T. and Nadgir A. 2002. Prediction of flank wear by using back propagation neural network modeling when Ghasempoor A., Jeswiet J. and Moore T.N. 1999. Real cutting hardened H-13 steel with chamfered and honed time implementation of on-line tool condition monitoring CBN tool. International Journal of Machine Tool and in turning. International Journal of Machine Tool and Manufacture. 42: 287-297. Manufacture. 39: 1883-1902. 405 UNIVERSITY OF IBADAN LIB ARY VOL. 7, NO. 4, APRIL 2012 ISSN 1819-6608 ARPN Journal of Engineering and Applied Sciences ©2006-2012 Asian Research Publishing Network (ARPN). All rights reserved. www.arpnjournals.com Richards E., Bessant C. and Saini S. 2002. Optimization of a neural network model for calibration of valtammetric data. Chemometrics and Intelligent Laboratory Systems. 61: 35-49. Sick B. 2002. On-line and indirect tool wear monitoring in turning with artificial neural network: a review of more than a decade of research. Mechanical System Signal Processing. 16: 487-546. Sivasakthivel P. S., Vel-Murugan V. and Sudhakaran R. 2010. Prediction of tool wear from machining parameters by response surface methodology in end milling. International journal of engineering science and technology. 2(6): 1780-1789. Summil Elanayar V.T. and Saundra K. 2000. Machining conditions monitoring for automation using neural networks. International Journal of Material Processing Technology. 105: 218-228. Tang K. and Li T. 2002. Combining PLS with GA-GP for QSAR. Chemom. Intell. Lab. Syst. 64: 55-54. Tansel I.N., Arkan T.T., Bao W.Y, Mahendrakar N., Shisler B., Smith D. and McCool M. 2000. Tool wear estimation in micro-machining, Part I: too-usage-cutting force relationship. International Journal of Machine Tool and Manufacture. 40: 599-608. Umbrello D., Ambrogio G., Filice L. and Shivpuri R. 2008. A hybrid finite element method-artificial neural network approach for predicting residual stresses and the optimal cutting conditions during hard turning of AISI 52100 bearing steel. Materials and Design. 29(4): 873- 883. Xu C.W. 2009. Condition monitoring of milling tool wear based on fractal dimension of vibration signals. Journal of Mechanical Engineering. 55(1): 15-25. Xu C. W., Xu T., Zhu Q. and Zhang H. 2011. Study of Adaptive Model Parameter Estimation for Milling Tool Wear. Journal of Mechanical Engineering. 57(7-8): 568- 578. Yongjin K. and Fischer G.W. 2003. A novel approach to quantifying tool wear and tool life measurements for optimal tool management. International Journal of Machine Tool and Manufacture. 43: 359-368. Zbigniew M. and David B.F. 2000. How to solve it: Modern Heuristic. Springer-Verlad, Germany. pp. 30-60. 406 UNIVERSITY OF IBADAN LIBRARY