RESPONSE PREDICTION IN MACHINING OF AISI 1040 STAINLESS STEEL USING ANN MODEL Shakti kumar, Rabeshkumar singh, AmitRai Dixit, Amitava Mandal and Alokkumar Das Department of Mechanical Engineering, Indian School of Mines Dhanbad, India E-Mail:shaktimech10@gmail.com ABSTRACT AISI 1040 stainless steel is a popular engineering material due to its wide application in the field of manufacturing, automobile and structural engineering. The motive of the research is to find the optimum process parameters for turning AISI 1040 under varying machining environment. Tungsten carbide tip tool is used for the experiment due to its high hardness and wear resistance. This model is used for the prediction of surface roughness and forces act during machining operation in different direction. The root mean square (RMS) it is found to be under expectable range. The surface roughness and forces were examine and it is found that predicted values and experimental values are close to each other which shows that the ANN model is effective for the prediction. Keywords: artificial neural network, forces, surface roughness, turning, machining. INTRODUCTION There are many fabrication processes like casting, welding, forming, machining etc. Out of all the fabrication processes need machining to get the desire surface finish. Metal cutting is one of the most important processes and it is highly used manufacturing process in industries. There are different types of machining process out of which turning is one of them which reduces the diameter of cylindrical workpiece by removing the material in the form of chip. The objective of the machining is to focus on higher quality product and also to maintain the dimensional tolerances. For this lots of research have been done in past and will continue to achieve the most appropriate method of machining. In addition to advances in terms of cutting tool material, machine tool accuracy etc. The present work is concerned with the turning of alloy steel AISI 1040. AISI 1040 steel is a high carbon steel alloy and its strength increases by forming process to achieve high tensile strength up to 150 to 250 ksi. Due to its high toughness and hardness it creates challenge to select appropriate machining parameter to improve surface finish and reduce forces exerted on the tool. Selection of proper cutting parameters is important which influences the quality and economics of machining. During cutting processes, a large amount of heat is generated at the shear zone, work piece, chip and tool interface due to severe plastic deformation. Numbers of published research papers on parametric analysis of metal cutting operations are available in literature. It is reported that during the turning operation higher feed rate produce more soft layers especially with depth of cut of 20micronsdue to more heat generated [1]. The surface roughness was primarily affected by feed rate, depth of cut and cutting speed. With the increase in feed the surface roughness also increases as the cutting speed increases and residual stress tends to become less tensile [2]. Better surface quality is generated when coated carbide insert was used during hard machining of EN31 [3]. Minimum surface roughness occurred during machining of Ni with hardness of 62HRC and 50 HRC are as 0.262 micron and 0.280 micron with CBN cutting tool [4]. With increase in rpm surface roughness decreases and power consumption increases [5].After the preliminary experimentation, it is found that spindle speed, feed and depth of cut mostly affect the surface roughness. The most significant parameters which affect the material removal rate (MRR) are spindle speed, feed rate and depth of cut during turning operation of EN- 31 steel in lathe [6]. Experimental design by Taguchi and ANOVA influence the cutting parameter and it was noticed that feed rate has more influence on surface roughness [7]. Artificial neurological network (ANN) was used for the better prediction of forces and surface roughness. During the turning process the response output was analysed by applying ANN methodology for the EN31 [8] high carbon steel. EXPERIMENTAL SETUP All experiments were performed on HMT NH 22 universal lathe machine in dry condition and 27 experiment were performed, in which 21 experiment were consider for the training purpose and 6 experiment was consider for testing purpose. The workpiece material of AISI 1040 of length 500 mm and diameter of 100 mm and sandvik s carbide CNMG 12 0408 insert was used as cutting tool. All the experiments were performed with single cutting tool used at once and after each experiment surface, roughness (Ra) was measured with the help of contact type Mitutoyo s SJ- 210 surface roughness tester. The roughness testing was performed on three different places on the cylindrical surface at an angle of 120 rotation and value was taken Probe is traveling on the work surface with diamond tip of diameter of 2 micron. Cutting force is measured by Kistler3-component dynamometer of type 9047 CNK. 10117
Figure-1. Experimental setup. Table-1. Chemical composition of AISI 1040. Element Iron (Fe) Manganese(Mn) Carbon (C) Sulphur (S) Phosphorus (P) Chemical Composition (%) 98.6-99 0.60-0.90 0.370-0.440 GBP 0.050 GBP 0.040 S. No 1 2 3 Machining parameter Cutting speed,n Feed, f Depth of cut,d Table-2. Critical parameter and their level. Unit Stage 1 Stage 2 Stage 3 RPM mm/rev mm 433 0.08 0.2 737 0.12 0.4 1255 0.16 0.6 SOLUTION METHOLODOGY Artificial Neurological Network (ANN) is a computational system which is inspired by the construction form, Processing Method and Learning capability of a biological brain in which learning method provide robust approach to the real value data. Neural Networks package suggest a number of different options which can be used to change the algorithms. The most useful algorithms were used in the proposed work; back prop algorithm is used as it provides acceptable result. It also allows input output and target. Mean square error (MSE) is evaluated during training/ testing as shown in equation(3). Artificial neural network diagram is used for the analytical process as shown in figure-2.consist of input parameter hidden layer consist of neurons and the output. W ij and W jk show weight between input and output with hidden layer. Each connection between neurons consist some weight, which provide strong link between the inputs and output with hidden layer to the each neuron. If the nodes in the input layer are constitute byx 1, X 2, X 3, the neurons in the hidden layer arey 1,Y 2,Y,Y 4, and W ij is the weight on the link betweeni i (input)and h j (hidden layer)the value of a nodal points in the hidden layer can be shown as Likewise, an output node O k of the neural network can be shown as O k = TF 1 ( ) (2) There is required a non-linear relationship between inputs and outputs. Mean squire error is calculated by MSE = 1 1 = input value Di= desire value 2 Mean absolute percentage error (MAPE) is basically used to predict feature data during forecasting. MAPE is explained in terms of accuracy percentage obtained by the forecasting technique, and it is calculated by the formula MAPE= 1 1 ² (4) ² (3) h = 1 ( ) (1) 10118
Parameters of artificial neural network (ANN) Table-3. Parameters and their types. Name Network type Type Feed-forward back-prop Number of hidden layer 1 Transfer function used LOGSIS Training function TRAINGDX Learning function used LEARNGDM Performed function MSE Number of neurons 12 Number of epoch 10,000 Learning factor 0 Figure-2. Diagram of ANN. RESULT AND DISCUSSIONS In this experiment surface roughness and forces are considered as the response parameters which affect the surface quality of the product. There are number of iteration was checked using different number of neurons during the analytical processes but the most effective value of regression co-efficient and root mean squire value (RMS) 0.062 as shown in table-6 have been found at 10000 epoch with 1 hidden layer and 12 neurons. The regression plot for the training and testing shows that the predicted data are come closer to the base line with regression co-efficient R for the training and testing are0.99481 and 0.98461 as shown in figure-3.the lower valuesof weight between inputs and the hidden layer shows that the predicted values are correct with respect to the input parameteras shown in table-3.graph-1 and graph- 2 shows that the train data and experimental data are more close to each in comparison with testing data which shows the validity of the ANN model for the prediction in future data. Table-4. Weight between input and hidden layer. Input W 1i W 2i W 3i W 4i W 5i W 6i W 7i W 8i W 9i W 10i W 11i W 12i Rpm -5.12-4.83-1.61-1.71-1.26 6.25 2.50 5.51 5.14 6.46-5.53-4.56 Feed 3.32-4.22 0.473-0.64 3.52-0.24-3.93 2.11 3.94-3.12-2.93-3.67 Depth of cut 0.89-2.32-6.18-4.83 6.32-0.231 4.18-2.35 2.64 0.002 3.65 1.69 10119
Table-5. Predicted value with percentage error obtains after training ANN. Exp. No Ra Fx (KN) Fy(KN) Fz(KN) Ra% Fx% Fy% Fz% 1 0.466 1.42 0.07 25.16 0.12 2.73 0.24 1.28 2 0.377 7.95 0.09 5.86 0.12 0.53 0.22 5.27 4 1 1.55 0.14 1.13 0.19 3.33 0.32 3.21 5 0.481 0.12 4.42 0.15 2.71 0.28 6.98 6 1 2.88 0.14 1.13 0.19 7.87 0.32 7.25 8 0.430 12.40 0.09 7.67 0.13 15.67 0.28 0.67 9 0.526 0.89 0.13 3.67 0.22 7.57 0.36 10.84 10 0.633 1.61 0.24 3.90 0.32 1.18 0.48 2.36 12 0.686 1.23 0.22 0.25 2.35 0.38 8.34 13 0.465 1.87 0.07 10.44 0.12 28.73 0.24 22.73 14 0.631 0.62 0.21 2.08 0.37 1.38 0.51 0.63 15 0.571 2.46 0.14 10.69 0.21 1.07 0.38 4.51 17 0.526 3.62 0.13 2.19 0.22 4.04 0.36 2.03 18 0.487 1.05 0.09 10.09 0.18 0.51 0.25 2.27 19 0.581 5.49 0.12 0.34 0.18 1.75 0.27 0.48 20 0.671 10.22 0.19 51.91 0.27 24.82 0.39 17.46 22 0.571 1.92 0.14 7.02 0.21 0.31 0.38 3.8 23 0.663 1.51 0.22 1.51 0.29 0.99 3.65 24 0 7.95 0.14 1.03 0.19 5.64 0.31 10.71 26 0.518 18.12 0.08 1.95 0.17 8.84 0.26 4.22 27 0.495 2.53 0.08 10.67 0.18 0.38 0.31 3.77 Table-6. Predicted data obtained after testing through ANN. Exp. Exp. Ra Exp. Fx Exp. Fy(% Exp. Ra Fx Fy Fz Fz% No Ra (%) Fx (%) Fy ) Fz 3 9 0.431 6.00 0.14 0.15 10.61 0.15 0.12 18.14 0.23 0.23 4.16 7 0.624 0.663 6.36 0.20 0.22 11.44 0.26 0.29 12.56 0.41 10.07 11 0.651 0.689 14.77 0.14 0.16 18.52 0.12 0.13 9.90 0.24 0.26 17.07 16 0.625 0.689 10.36 0.15 0.16 7.05 0.12 0.13 8.12 0.25 0.26 4.16 21 0.461 0.430 6.70 0.08 0.09 10.21 0.12 0.13 3.02 0.26 0.27 6.14 25 0.637 0.648 1.84 0.18 0.18 3.29 0.27 0.28 1.72 0.44 0.47 6.93 Table-7. Root mean square, training time and plot interval. RMS (Root Mean Square) Time taken for iteration in second Plot interval Gradient 0.062 0.18 1 epoch 0.000150 Table-8. Mean absolute percentage error (MAPE) for training and testing. Ra(micron) Fx(KN) Fy(KN) Fz(KN) Training 4.188 7.70 5.83 5.84 Testing 7.778 10.19 8.89 8.09 10120
Regression plot obtain after applying ANN Figure-3. Regression plots. 0.75 TrainRa 0.70 0.65 0.60 Ra(micron) (a) (b) ExperimentalFx ExperimentalFy ExperimentalFz TrainFx TrainFy TrainFz Forces Fx,Fy,Fz(KN) 0.25 0.20 0.15 0.10 Forces Fx,Fy,Fz(KN) 0.25 0.20 0.15 0.10 0.05 (c) Exp. no 0.05 Figure-4. Plot (a) and (c) shows experimental plot for (surface roughness and forces), (b) and (d) shows training plot through ANN for (surface roughness and forces) (d) 10121
0.70 0.65 ExperimentalRa ANNRa Ra(micron) 0.60 Forces Fx,Fy,Fz(KN) 0.25 0.20 0.15 0.10 ExperimentFx ANNFx ExperimentFy ANNFy ExperimentFz ANNFz 0.05 (a) Figure-5. Plot between measured data and testing data, (a) Test and experimental Ra (b) Test and experimental force. (b) CONCLUSIONS 1. Surface roughness (Ra), is primarily affected by feed. It is clear from Figure-4 (a) and (b) that both experimental and predicted data are almost on the same path which shows the validity of ANN model with MAPE 4.188and 7.771 during training and testing as shown in table-8. 2. Cutting forces, It is clear from Figure-4 (c) and (d) that the experimental and predicted value of forces Fx, Fy, Fz are almost on the same path which shows that the predicted and experimental values are close to the actual values. Hence it shows better consistency between predicted and experimental values which confirm the existence of this model. From table-8 it is seen that MAPE for the train forces are less as compared to test which shows that the train values are more accurate as test values. 3. From Table-5 it is clear that ANN model help in selecting the cutting parameter to get the required quality of surface within the tolerance limits REFERENCES [1] Choi, Y., 2015. Influence of feed rate on surface integrity and fatigue performance of machined surfaces. International Journal of Fatigue, 78, pp.46-52. [2] Kumar, R. and Chauhan, S. 2015. Study on surface roughness measurement for turning of Al 7075/10/SiCp and Al 7075 hybrid composites by using response surface methodology (RSM) and artificial neural network (ANN). Measurement, 65, pp.166-180. [3] Beatrice, B.A., Kirubakaran, E., Thangaiah, P.R.J. and Wins, K.L.D., 2014. Surface Roughness Prediction using Artificial Neural Network in Hard Turning of AISI H13 Steel with Minimal Cutting Fluid application. Procedia Engineering, 97, pp. 205-211 [4] Sharma, V.S., Dhiman, S., Sehgal, R. and Sharma, S.K., 2008. Estimation of cutting forces and surface roughness for hard turning using neural networks. Journal of Intelligent Manufacturing. 19(4), pp.473-483. [5] Valera, H.Y. and Bhavsar, S.N., 2014. experimental investigation of surface roughness and power consumption in turning operation of en 31 alloy steel.procedia Technology. 14, pp.528-534. [6] Bhushan, R.K. 2013. Optimization of cutting parameters for minimizing power consumption and maximizing tool life during machining of Al alloy SiCparticle composites. Journal of cleaner production. 39, pp. 242-254. [7] Gowd, G.H., Goud, M.V., Theja, K.D. and Reddy, M.G. 2014. Optimal Selection of Machining Parameters in CNC Turning Process of EN-31 Using Intelligent Hybrid Decision Making Tools. Procedia Engineering, 97, pp. 125-133. [8] Prasad, M.V.R.D. and Janardhana, G.R. 2011. Effect Of Input Parameters On Residual Stress In Dry Machining Of Hardened Steel (EN31) With CBN Cutting Tool-Coactive Neuro-Fuzzy Interface System Approach. I-Manager s Journal on Mechanical Engineering. 1(2), p. 44. 10122