Optimization of Surface Roughness of D2 Steels in WEDM using ANN Technique

Attempt has been made to investigate the experimental process and surface roughness (SR) optimization of cold working (high carbon and high chromium) hard die steel (D2) during wire electrical discharge machining processes (WEDM). It is very difficult to determine optimal cutting parameters for improving cutting performance has been reported. Wire electrical discharge machining process relies heavily on the operators ’ technologies and experience because of their numerous and diverse range as using complicated cuts can made through difficult to machine electrically conductive components, WEDM process was developed to generate precise cutting on complicate, hard and difficult to machine materials. Tan-sigmoid and purlin transfer functional with bias based four layered back propagation artificial neural network (BPANN) approach have been used to investigate the effect of six independent parameters namely gap voltage (V g ), flush rate (F r ), Pulse on time (T on ), pulse off time (T off ), wire feed (W f ) and wire tension (W t ) over CLA value of surface roughness (R a ) along with corresponding material removal rate (MRR). A fractional factorial design of experiment of three level were employed to conduct 80 rows of experiment on (D2) steel with chrome coated copper alloy wire electrode. The predicted response, CLA values of SR and corresponding MRR were observed by the approach of BPANN from experimental (55 rows for training, 15 rows for validation and 10 for testing) data. Software instructed programme has been used individually for training, validation and testing in MATLAB 2010a to find the corresponding prediction output. Two fold cross over technique (TFCT) were used to developed distinguish (S1 and S2) models and also developed more models depending on numbers of neurons used in primary and secondary hidden layers. The model adequacy is very satisfactory as correlation coefficient (R 2 ) is found to be 99.1% and adjusted (R adj . 2 ) statistics is 98.5. It is found those spark time ON/OFF, wire feed rate, wire tension, gap voltage and flush rate and few of their interactions have significant effect on SR.


Introduction
Wire electrical discharge machining is the metal removal process by means of repeated spark created between the wire electrode and work piece. It is considered as unique adaptation of the conventional EDM, which used an electrode to create the sparking within kerfs [11]. However, WEDM utilizes a continuously traveling chromium coated copper wire electrode ranging diameter 0.05-0.35 mm, which is capable to achieve very good sharpness of edge [4]. Very high temperature ranging 8000-10,000°C creates within the kerfs gap during machining, so that material removal may takes place by not only melting but directly vaporizations also. WEDM is used for the high precision machining to all type of electrically conductive metallic alloys, tool and die, graphite, and few ceramic and composite materials of any hardness which cannot be machined easily by conventional machining methods [1,5].
Manufacturing processes (WEDM) has been chosen depending on the material characteristics and the type of responses required to be evaluating. The present study aimed to optimization of responses i.e. surface roughness with corresponding MRR of D2 steel by conducting 80 rows of experimental data using frictional factorial (2 6-2 ) design of experiment of five different set at three levels [3]. Four layered BPANN architecture has been used for modeling, where independent process variables are V g , F r T on , T off , W f and W t to get the précised and optimized values of responses R a [6,8,10]. Best model S2 has been found on the basis of correlation coefficient (R 2 ) between observed and predicted responses (SR) [12]. The response (SR) is expressed as the irregularities of material resulted from various machining operations. It is represented as 'R a ' symbol and used to be called center line arithmetic average roughness for the sampling length [2].
The optimum process parameters are much essential to achieve better surface finish with adequate material removal rate (MRR) or shrink of total machining time; lot of research attempts has been reported for modeling and investigation of WEDM process parameters [7], but sum of root mean square error (SRMSE) approach have been used to optimize the process parameters by taking 55 rows of training data [9].

Experimental setup 2.1 Selection of wire electrode and work piece
A chrome coated cylindrical pure copper wire electrode having 0.25 mm in diameter and high tensile strength were selected for conducting machining operation on 18 mm diameter of D2 steel rod to cut 5 mm thickness of disk using Electronica Maxicut, WEDM process. It is very clear that D2 is hard die steel and conducting material with high carbon and chromium content ( Table 1).
The experiment has carried out on Wire Electrical Discharge Machine, model ELECTRONICA-MAXICUT, SLNO -250, (F:09:0002:01) having the facilities to hold the work piece within the place provided by the help of conductive fixture, so that they can complete the circuit between electrode and work piece. The spark is created depending upon gap voltage applied between the conductive work piece, electrode, and machining performance influence the major independent process parameter which selected for experiment as characteristics of screening test. Commercials grade of deionized water (density = 832 kg/m 3 ) was used as dielectric fluid. 18 mm cylindrical rod of D2 steel was used as the work piece with negative polarity and the power supply has the provision to connect the 0.25 mm chromium coated pure copper tool electrode with positive polarity so that the material removal may takes place by influence of heat generated within kerfs due to applied voltage within it ( Figure 1). The surface roughness R a of the processed material have been measured precisely by using Surftest SJ-210 tester having center line average value (CLA), where least count of the equipment is 0.001 μm for the travel length of 0.85 mm ( Figure 2).

Design of experiment and objective
Five different set of fractional factorial (2 6-2 = 16) experimental design have been selected at two levels, so that 80 rows of experimental data can be observed at three level of replication on D2 using WEDM. In this study the main aim to minimize the surface roughness of D2 on best possible maximum MRR during WEDM ( Table 2).

ANN architecture and training
The hit and trail method based on literature have been adapted to find 7 and 10 neurons in primary and secondary hidden layers respectively, which effects on the R-square statistics for best prediction modeling. Tan sigmoid activation (squashing) function used as the (infinite input to finite output range) learning capability by the  controllable instructed programme in MATLAB 2010a. Steepest descent problem used for the training algorithm to train the multilayer network, where the values of gradient was smallest because of the small changes in weight and biases. p 1 , p 2 , p 3 , p 4 , p 5 and p 6 are the six input layer neurons and O i is the single neurons in output layer, whereas I 11 -I 17 and I 21 -I 29 (7 neurons present in primary and 10 in secondary hidden layers) are the hidden layers ( Figure 3).

Experimentation
Two models (S1 and S2) were developed from 80 rows of experimental data performed of D2. Only training result of best performing model (S2) of 55 rows is Wire tension (Wt): (g) 300 600 900 Table 2.
Factors for screening test. presented here for achieving the aimed to optimization of influencing process parameters (Figures 4 and 5). OPTIMIZATION OF PROCESS PARAMETER R a : D2, 7 Neurons in hidden layer. The best model needs to be predicted among Model-S1 and S2, in D2 steel. Effect of individual input parameters will be observed on the R a (Tables 3-6).

Optimization of process parameters
It is evident from Table 3, that each independent influencing input parameter has corresponding values of their square of residuals at each three levels. Two values at each level (2 Â 3 = 6 rows) has been taken for each inputs, where lowest possible square of residuals are available, to draw the Figure 6(a-f). Figure 6(a-f) shows the relations between individual influencing parameters (V g , F r , T on , T off , W f and W t ) to their optimized response, surface roughness (R a ) with corresponding values of MRR. Table 5 also indicates that unique values of each influencing parameters (corresponding to its serial numbers of Table 5) gives optimum responses, which has been highlighted.    Table 3. D2, S1, 7N, training data (combined parameters).   Again experiment has been conducted on D2 steel using WEDM by setting the individual optimum parametric combinations (V g , F r , T on , T off , W f and W t ) as 90 (V), 8 (Lit./min), 1.05 (μS), 190 (μS), 2 (m/min) and 900 (g) respectively and found the values of Ra = 0.9638 (μm) at MRR = 105 (mg/min) ( Table 7).

Conclusion
It has been concluded that the best fitted model (S2) for material removal rate and surface roughness of D2 steel has been achieved by artificial neural network Correlation coefficient (R 2 ): Training data of D2 steel (best performing model S2) using 7 and 10 neurons in primary and secondary hidden layers.  Table 6. Summary of R 2 values of training validation and testing data: 7 N in 1st and 10 N in 2nd L, R a . Table 7.
Best parametric combination with their possible responses.