Prediction of Cutting Conditions in Turning AZ61 and Parameters Optimization Using Regression Analysis and Artificial Neural Network

Department of Mechanical Engineering, College of Engineering, King Saud University, P.O. Box 800, Riyadh 11421, Saudi Arabia Department of Mechanical Engineering, Dicle University, Diyarbakir 21280, Turkey Department of Industrial Engineering, College of Engineering, King Saud University, P.O. Box 800, Riyadh 11421, Saudi Arabia Department of Mechanical Engineering, School of Sciences and Engineering, American University in Cairo, AUC Avenue, P.O. Box 11835, New Cairo, Egypt


Introduction
Magnesium alloys are often used in many industrial applications such as the manufacturing of several components used in the aerospace and modern automobiles industry.Also, magnesium block engines have been widely used in some high-performance vehicles.In those applications, the final surface roughness of machined components is playing a major factor in the acceptance of those parts.
Many researchers have investigated the optimization of cutting parameters for the prediction of surface roughness as a key performance measure.Asiltürk used ANN (artificial neural network) and MRM (multiregression models) to predict the surface roughness of steel AISI 1040.ey developed their own models and used ANN to optimize the cutting parameters formulating the surface roughness as objective function.ey used cutting speed, feed rate, depth of cut, and nose radius as optimized parameters.Surface roughness is characterized by the mean (R a ) and total (R t ) values of the recorded roughness at different locations on the produced surface.ey conducted many experiments, each with a different set of the cutting parameters, and the corresponding R a and R t values were reported.Obtained results were then used to train an ANN model.Mean squared error of approximately 0.003% was achieved which outperforms error rates reported in the early literature and are claimed to be suitable for robust prediction of the surface roughness in industrial settings [1].
Another approach can be found in the work of Mokhtariet Homami et al. [2], and they employed a design of experiment (DOE) technique based on a full factorial design to determine the number of experiment and the corresponding parameters.ey represented their results in a statistical analysis, and they used ANN to model the system.Optimization was done using genetic algorithm (GA).e conclusion of their work was that the main factors affecting the flank wear and the produced surface roughness are the feed rate, nose radius, and approach angle, while the cutting speed had the major effect on flank wear.Optimized values of the cutting conditions were attained and showed a significant reduction in the surface roughness values.
Jafarian et al. [3] used GA and particle swarm optimization (PSO) techniques to determine optimal cutting parameters in turning operations with a multiobjective optimization aiming to minimize surface roughness and cutting forces and maximize tool life.ey discussed their results claiming that training ANN using GA gave superior results than those reported in the literature with high accuracy and gives the flexibility of analyzing the effect of each parameter separately on the output.
e PSO technique was also utilized in the work of Karpat and Özel [4], and they used the Pareto optimal frontier to select optimized parameters to maximize material removal rate (MRR) without affecting the induced stresses or the final surface finish of the produced components.ey obtained good results making use of dynamic-neighborhood PSO approach in solving complex turning optimization problems.
A different approach was utilized in the work of Natarajan et al. [5] to test the reliability of ANN in the prediction of surface roughness values when machining Brass C26000 material in dry cutting condition on a CNC turning machine.Surface roughness has been measured and compared to the experimental data and concluded that ANN can be implemented reliably and accurately to predict surface roughness in turning operations of Brass C26000 material.
e applicability of radial base function (RBF) neural networks was investigated in the work of Pontes et al. [6] to predict surface roughness in turning processes of SAE 52100 hardened steel.Networks were trained using different sets.
ey considered several design variables and found that ANN models were capable of providing accurate estimates of surface roughness values in an affordable way.
e turning of Ti-6Al-4V titanium alloy was investigated in the work of Sangwan et al. [7] to minimize surface roughness using ANN-GA approach.A feed forward neural network was proposed for training and testing of the neural network model.e predicted results were found to be in good agreement with the obtained experimental results.
A comparison between linear regression models and ANN approach has been studied in the work of Acayaba and Escalona [8].A target of saving cost, effort, and machining time leads to the necessity of predicting surface roughness prior to performing machining operations.ey used experimental data to validate their claim and found that using ANN outperforms linear modeling.Instead of using GA like other researchers previously listed, this research employed a simulated annealing (SA) optimization algorithm to optimize cutting parameters for minimizing surface roughness.Results show similar findings as reported previously with no major significant improvement.
A more concise investigation focusing only on the three major cutting parameters influencing the surface roughness was presented in the work of Bajić et al. [9].Cutting speed, feed rate, and depth of cut are optimized using regression analysis and ANN.Results obtained show no superiority of one approach over the other, and both gave a good prediction of the surface roughness.
A new approach that integrates artificial intelligence (AI) with ANN and GA has been introduced by Gupta et al. [10], and the paper illustrates the impact of using AI on the quality and type of results obtained for the surface roughness prediction.
ey analyzed the experimental data using support vector regression (SVR) defining the tool wear and power required as output parameters.
Grade-H high-strength steel had its share in the investigation for better surface quality studied by Abbas et al. [11] in their work.ey emphasized that the key factors for the manufacturing of parts produced using Grade-H highstrength steel are parts accuracy and surface roughness MRR.Identifying the final surface roughness of produced parts prior to machining is crucial to ensure that those parts will not be rejected.e rejection of these parts at any processing stage will represent huge problems to any factory because the processing and raw material of these parts are very expensive.ANN was used in this work to determine the optimized cutting parameters to ensure minimum surface roughness during the turning operations.
As a continuation of their work, Abbas et al. [12] investigated the turning of high-strength steel focusing on three main cutting parameters: cutting speed, feed rate, and depth of cut.eir results included a Pareto frontier between surface roughness and machining time of finished components made from high-strength steel using the ANN model that was later used to determine the optimum cutting conditions. is study showed the feasibility of integrating optimization algorithms with computer-aided manufacturing CAM systems using Matlab.
A quantitative approach to evaluate the cutting process and its stability was demonstrated in the work of Yamane et al. [13], and they used the turning operation as a base for their study aiming at identifying the machining system deviation from a perfect process.Such a deviation can be identified by monitoring the machined surface and comparing it with the cutter profile.Adhesion and builtup edge produced during machining operation can then be easily noticed and monitored.Excessive vibration and the accuracy of spindle rotation can also be recorded and is a good indication of system instability and related directly to the quality of parts produced.
eir conclusion was that the proposed method can be successfully implemented to evaluate turning operations.
e influence of the type of inserts used in the machining process on the quality of the surface produced was investigated in the work of D'Addona and Raykar [14].ey compared wiper inserts to conventional ones in the turning operation of oil hardened nonshrinking steel used in the manufacture of strain gauges and measuring instruments.Surface roughness was a major factor in this study as it is a very important aspect in the performance of those devices.

2
Advances in Materials Science and Engineering ey used analysis of variance (ANOVA) and analysis of means (AOM) plots to evaluate their results.
Pu et al. [15] reported that magnesium alloys are gaining a lot of intension from researchers in the literature due to their advanced properties over conventional materials used in the automotive industry as well as medical applications such as biodegradable implants.One of the major factors looked at in machining of this alloy is the surface integrity.Pu et al. [15] investigated the effect of machining AZ31B Mg alloy under dry conditions as well as using liquid nitrogen as a lubricant.ey concluded that using cryogenic machining with large nose radii improved several material performance criteria such as surface finish and grain size refinement.
Increasing productivity and maximizing material removal rate (MMR) have been also investigated for the machining of magnesium alloys.Using very high cutting speeds has its drawbacks and has been analyzed by Tomac et al. [16] in their work.ey concluded that using speeds in excess of 600 m/min will result in buildup edge on the flank face of the cutting tool.ey supported their argument with microstructure pictures of tool inserts as well as the machined surface of three different Mg-Al-Zn alloys.
Different coating materials have been used in industry to reduce the builtup edge effect appearing on the tool flank face during turning operations of magnesium alloys at high speeds.Tönshoff and Winkler [17] reported the different interactions happening between the cutting tool inserts, the coating, and the workpiece materials in turning of AZ91 HP at very high speeds ranging between 900 and 2400 m/min.ey concluded that cutting tools with polycrystalline diamond (PCD) inserts can significantly reduce the cutting forces and hence the frictions at the tool-workpiece interface.
e optimization of cutting parameters is another venue pursued by researchers to improve surface quality of machined magnesium alloys.Wojtowicz et al. [18] studied the effect of changing cutting parameters on the turning of AZ91 HP.Parameters explored include cutting speed, feed rate, depth of cut, and tool nose radius.Surface integrity and increasing fatigue life were the major optimized parameters of the machined components, and other reported parameters include microstructure, grain size and residual stresses improve fatigue life.ey also supported the argument provided by Tönshoff and Winkler [17] regarding the superior performance of PCD coating tool inserts at high cutting speeds and feed rates.
In this paper, an ANN model has been employed to estimate and optimize the produced surface roughness of AZ61 material during turning operations. is method proves to be more efficient and provided the manufacturing engineers with a good tool to be utilized to effectively predict the quality of the surface produced in an economical and time saving manner.us this eliminates the possibility of part rejection due to manufacturing process errors that costs the factory time and money and wasted raw material.

Materials and Methods
Table 1 presents the chemical composition of magnesium alloy AZ61 which contains zinc and aluminum with 1 and 6 percent content, respectively.e microstructure of the composition is analyzed at the previously mentioned aluminum concentration, and the phase diagram shows a magnesium-rich phase interacting with an Al 12 Mg 17 composite.Zinc along with other traces is found to have no effect on the alloy microstructure.
e machining of test specimens is done using Emco mill concept 45 CNC turning machine equipped with Sinumeric 840-D.e diameter of the workpiece is equal to 40 mm with a length of 100 mm.Tool holder specification is SVJCL2020K16, while the insert is VCGT160404 FN-ALU.
e cutting edge angle, nose radius, and clearance angle are set at 35 °, 0.4 mm, and 5 °, respectively.All experiments were conducted in wet conditions while the cutting parameters are controlled via CNC part program.e surface roughness tester TESA Rugosurf 90-G is used to evaluate the produced surface roughness.A sketch of the test specimen is shown in Figure 1. e test plan was implemented through 64 turning runs.ese runs were divided into 16 groups.Each of four groups was subjected to one common cutting speed (125, 150, 175, and 200 m/min).Each group was machined using four levels of cutting depth (0.30, 0.60, 0.90, and 1.12 mm).Each depth was processed using feed rate having four levels (0.05, 0.010, 0.15, and 0.20 mm/rev).Full listing of all samples and the resulting measured surface roughness are provided in Appendix A.
Multivariable regression analysis was used to build a mathematical model relating the process outcome (surface roughness R a ) with the three studied input parameters (cutting speed (V), depth of cut (d), and feed rate (fr)).56 experiments were conducted that cover the input parameter range described previously.Eight extra experimental runs were carried out to be used in testing both the regression and ANN models.
Regression was conducted using Minitab 17 software with stepwise technique to eliminate the insignificant terms from the model.e model was fitted in the form given by the following equation [19]: where β o is the constant term, β i represents the linear effects, β ii represents the pure quadratic effects, β ij represents the second level interaction effects, β ijk the third level interaction effects, β iij represents the effect of interaction between linear and quadratic terms, and ε i represents the error in predicting experimental surface roughness.e material removal rate (MRR) was calculated using (2) for each run.Desirability function approach was used to maximize MRR maintaining R a below 0.4 as a maximum limit for the surface roughness value: where MRR is the volume removed per unit time (mm 3 /min.),V is the cutting speed (m/min.),fr is the feed Advances in Materials Science and Engineering rate (mm/rev), and d is the depth of cut (mm).Multivariable regression analysis was used to build a mathematical model relating the process outcome (surface roughness R a ) with the three studied input parameters (cutting speed (V), depth of cut (d), and feed rate (fr)).56 experiments were conducted that cover the input parameters range described previously.Eight extra experimental runs were carried out to be used in testing both the regression and ANN models.

Results and Discussions
e regression-tted mathematical model is given by (3).Box-Cox transformation was used to normalize the residuals with λ 0 (natural log) for R a .Anderson-Darling test was conducted to check the normality of residuals with a result of p value 0.885 > 0.05.e null hypothesis of such a test is that the data are normal, and a p value < 0.05 proves nonnormality.
Determination coe cient (R 2 ), mean square error (MSE), and mean absolute error (MAE) were calculated to be 0.975, 0.02, and 0.12, respectively.Figure 2 shows a scatter plot for the predicted R a versus measured R a .It is clear from the gure that the relation between them is very close to linear with calculated R 2 equals 0.97.ln R a −2.6420 + 0.1874 d + 16.516 fr. ( Desirability function approach was used to estimate the values of studied process parameters that maximize the MRR keeping R a at levels not exceeding a practical value of 0.4 μm.
e optimization plot, illustrated in Figure 3, shows that an optimum MRR of 13,928 mm 3 /min with R a 0.4 μm is obtained at cutting speed 200 m/min., depth of cut 1.12 mm, and feed rate 0.09 mm/rev.
ANN modeling was used to predict the surface roughness of AZ61 magnesium alloy.e three input parameters cutting speed (V), depth of cut (d), and feed rate (fr) were taken into account to predict the surface roughness as an output parameter.e data were taken from the experimental result.e transfer function was selected as a TanhAxon.e data were divided in two parts as training and test data.80% of data were used for training stage while the left 20% of data were used for test stage to understand the performance of the developed ANN model.e best ANN model predicting the R a value between developed trial models was obtained according to the values of R 2 and MSE. Figure 4 gives the experimental and ANN-predicted results.It can be seen from the gure that a good agreement was obtained between experimental and ANN-predicted results.It is also seen from Figure 5 that the value of R 2 is 0.9629 for experimental surface R a and ANN-predicted surface R a , while R 2 is 0.9931 for experimental MRR and ANN-predicted MRR in Figure 6.

4
Advances in Materials Science and Engineering Predictability of both regression and ANN models were compared using eight extra experimental cases that were not included in the modeling phase.Table 2 illustrates the results of this comparison.Figures 4-6 and Table 2 clearly show the good agreement between the results.From the table, both ANN and regression-predicted R a and MRR values can be acceptable when they are compared with experimental results.
Feed rate has a direct impact on product surface nish.Although increasing feed rate can result in a signi cant increase in the material removal rate (MMR) and increase productivity, it will also increase cutting forces resulting in higher tool-workpiece friction associated with poor surface nish.Horizontal markings as well as vertical ones in surface roughness pro le can be detected when examining the surface pro le produced using high feed rates.MRR and surface roughness are two contradicting objectives in determining an optimized value for the feed rate.Optimization algorithms are often used to come up with optimized cutting parameters for di erent machining processes.Figure 7 represents an optical microscopy view of machined surface, while Figure 8 shows the surface roughness pro le produced by   Advances in Materials Science and Engineering the surface roughness tester under the following cutting conditions: a speed of 125 m/min, a depth of cut of 0.3 mm, and a feed rate of 0.20 mm/rev.Another set of views and graph is also provided in Figures 9 and 10 for di erent cutting conditions summarized as follows: a cutting speed of 125 m/min, a depth of cut of 0.3 mm, and a feed rate of 0.05 mm/rev.From those results, we can conclude that reducing feed rate will produce thinner surface roughness markings, and increasing the feed rate is associated with the presence of distant thick surface roughness markings.

Conclusions
Optimization and estimation of R a and MRR in cutting conditions of AZ61 magnesium alloy were realized by ANN modeling and regression analysis.
e results of the developed ANN-predicted model and regression model were   6 Advances in Materials Science and Engineering compared with experimental results.e results showed that a good agreement was obtained for both developed ANNpredicted and regression analysis results.In addition, the compatibility between the developed ANN model and experimental results showed that ANN approach is an accurate method to estimate surface R a and MRR.Optical microscopy views and the corresponding surface roughness pro le for di erent sets of cutting parameters were utilized on two machined surfaces to showcase the direct e ect of increasing the feed rate on surface nish.A hypothetical analysis relating the higher surface roughness values associated with the increase in feed rate is reported.

Appendix
e listing of all samples and the resulting measured surface roughness are provided in Table 3.

Figure 1 :Figure 2 :
Figure 1: Working drawing of the test workpiece.

Figure 3 :
Figure 3: Optimization plot for R a and MRR.

Figure 4 :R 2 =
Figure 4: Desired output and actual network output.

Figure 5 :
Figure 5: Comparison between ANN predicted and desired output for R a .

R 2 =Figure 6 :
Figure 6: Comparison between ANN predicted and desired output for MRR.

Figure 7 :
Figure 7: Optical microscopy for machined surface with V 125 m/min, d 0.30 mm, and fr 0.20 mm/rev.

Figure 8 :Figure 9 :
Figure 8: Pro le of surface roughness graph by the surface roughness tester for machined surface with V 125 m/min, d 0.30 mm, and fr 0.20 mm/rev.

Figure 10 :
Figure 10: Pro le of surface roughness graph by the surface roughness tester for machined surface with V 125 m/min, d 0.30 mm, and fr 0.05 mm/rev.

Table 2 :
Comparison between the regression model and ANN predictions.

Table 3 :
Listing of four-level full factorial samples.