Professional Documents
Culture Documents
To cite this article: Uma Maheshwera Reddy Paturi, Suryapavan Cheruku, Sriteja Salike, Venkat
Phani Kumar Pasunuri & N.S. Reddy (2022) Estimation of machinability performance in wire-EDM
on titanium alloy using neural networks, Materials and Manufacturing Processes, 37:9, 1073-1084,
DOI: 10.1080/10426914.2022.2030875
CONTACT Uma Maheshwera Reddy Paturi maheshpaturi@gmail.com Department of Mechanical Engineering, CVR College of Engineering, Hyderabad,
Telangana 501510, India
© 2022 Taylor & Francis
1074 U. M. R. PATURI ET AL.
to model and predict the performance during WEDM of ● Model the four performance criteria using ANN and
gamma titanium aluminide (γ-TiAl). Further, Hsien et al.[32] compare with the experimental outcome.
planned a BPNN method to analyze and optimize the machin ● Provide a visual representation of the interaction between
ing variables for roughness and cutting speed during WEDM the process variables and their impact on the machinabil
of pure tungsten. The ANN predictions were very accurate and ity performance.
the prediction error is within 3% when matched with the
experimental outcome. Rahim et al.[33] applied neural net
works to establish a correlation among inputs and surface Materials and methods
roughness during WEDM of the copper workpiece. The neural Experimental work
network model predictions demonstrated a good agreement
with experimental data by giving R-value as high as 0.995. WEDM, a non-contact machining process, is widely employed
Recently, Paturi et al.[34] adopted a BPNN to estimate the to cut materials with superior strength and hardness. In the
machined surface quality during WEDM of Inconel 718. The present study, WEDM tests were conducted on EX 400
BPNN model predictions exhibited a remarkable agreement machine with a steady supply of dielectric medium. The experi
with an R-value of 99.997%, signifying a solid fit among model mental system contains a workpiece holding table, wireframe
response and target data. In another work, Titus et al.[35] and feed system, dielectric flow system, and a pulse signal
applied a feed-forward BPNN approach to model the wire- generator power source. A rotating wire drum is used to supply
EDM of aluminum matrix composites. Based on the outcome the wire electrode under constant tension. Deionized water
of the study, ANNs are substantial and accurate in estimating (dielectric fluid) and brass electrode of 0.25 mm diameter
the surface finish and material removal rate (MRR) in the with strength in tension of 900 N/mm2 and 1% elongation
WEDM process. are used in all the experiments. The dielectric fluid temperature
From the overview of the literature surveyed, the reviews is kept steady at 27°C. The WEDM experimental set-up along
reveal that ANNs were capable to model and comprehend the with fabricated samples are shown in Fig. 1.
WEDM of several materials such as TiAl, titanium alloys,
tungsten, copper, nimonic alloys, Inconel 718, Stavax steel,
metal matrix composites, etc. Most of the existing studies on Selection of work material and microstructure analysis
ANN modeling of WEDM process are restricted to a few input The application of titanium and its alloys have been prominent
parameters with a single output. Also, as per the author’s best in various commercial industries due to their distinctive qua
knowledge, the combination of ANN, WEDM, and Ti-6Al-4V lities compared to other pure metals and alloys. Numerous
alloy involving many control factors with multiple responses of types of titanium alloys like Ti-6Al-4V, Ti-6Al-4VELi, Ti-
the WEDM process is not reported. Therefore, the central 5Al-2.5Sn, Ti-5Al-2Sn3Li, etc. are available in the market.
objectives of this research were to: Among them, commercially available Ti-6Al-4 V is a widely
used material. The titanium alloy used in WEDM tests is
● Analyze the influence of eight inputs on WEDM perfor procured in the size of 200 mm x 110 mm x 11 mm from
mance parameters. Bharat Aerospace metals, an ISO certified authorized material
Figure 1. WEDM experimental set-up along with roughness measurement on fabricated sample.
MATERIALS AND MANUFACTURING PROCESSES 1075
Table 1. The major elements of titanium alloy substance (wt.%). Selection of WEDM parameters and machinability
Element C Fe Al V Ti Others measures
Composition 0.012 0.257 6.54 4.46 88.717 0.057
Surface roughness is one of the significant machining attri
butes in any machining process, which substantially
supplier. The composition of titanium alloy substance is tested impacts the performance of mechanical parts and affects
at Jyothi Spectro Analysis, Hyderabad, India. The composition the component’s operational life. In WEDM, the surface
of the titanium alloy substance presented in Table 1 is con finish quality of the machined parts is highly influenced
firmed to be a Grade 5 of titanium. by the range of process parameters and machining condi
To enhance the material’s mechanical properties, titanium tions. The occurrence of overcut on the workpiece during
alloys are heat-treated, then quenched and aged during the machining is measured using the kerf width. The varia
manufacturing to maintain their β-phase and decrease the tion in kerf width directly indicates the amount of material
residual stress. It is essential to understand the α-phase and utilization. In general, the larger the value of kerf width, the
β-phase transformations; therefore, the microstructure of the greater the possibility of material wastage. Cutting speed
Ti-6Al-4V has been examined. It is known that the titanium and MRR are the other two primary parameters that influ
properties are primarily influenced by the α-platelet thick ence determining the work piece’s functional performance.
ness, β-particle size, and colony size.[36,37] As the α-platelet MRR indicates how quickly the workpiece is being
thickness increases, the yield strength diminishes continually. machined. A high machining rate is usually desirable, clo
In contrast, the thickness of the α-platelets affects the mate sely connected to productivity and machinability perfor
rial elongation more complexly. The specimen’s yield mance. Further, machinability performance largely relies
strength and extensibility are also enhanced by reducing on the choice of WEDM parameters such as pulse on and
the β-crystal’s particle size and colony size. As a result, β- off duration, wire feed rate, arc on and off duration, servo
grain size and the colony mass have a greater impact on the voltage, flushing pressure, and wire tension. Therefore, in
tensile characteristics of Ti-6Al-4V alloy than the α-platelets the present study, machinability performance in terms of
thickness.[38] Similar observations of α and β-phase transfor surface roughness, cutting speed, kerf width, and MRR were
mations have been observed in the current work. Figure 2 examined under varying process parameters. A surface pro
represents the microstructure of the workpiece. The micro file meter is used to determine the surface finish on fabri
structure shows the transformed β-phase with α-phase in the cated samples. The average of three measurements on each
matrix of solid titanium solution. machined surface sample is recorded.
1X N �2
MSE ¼ Tip Oip (1)
N i¼1
1X N �
�Tip
��
AE ¼ Oip � (2)
N i¼1
Figure 4. MSE and AE variation corresponding to single and double hidden layer with different neurons.
A measure of linear association between observed and pre datasets to observe the model’s overall accuracy. Figure 8
dicted values can be observed through R-value. An R-value of demonstrates the correlation graph obtained for train, test,
1 indicates an ideal fit for the model data. The model needs to validation, and all data. R-values of 0.99985 for train data,
perform well on all the datasets, i.e., train, test, and validation. 0.99898 for test data, and 0.99861 for validation data in Fig. 8
Hence, R-value was calculated for train, test, and validation indicate the proposed neural network model’s good prediction
MATERIALS AND MANUFACTURING PROCESSES 1079
Figure 6. Validation performance of the developed ANN model. Figure 7. The histogram for calculated errors between targets and outputs.
Figure 8. ANN model prediction efficacy with training, validation, test, and all datasets.
1080 U. M. R. PATURI ET AL.
Table 3. Comparison between WEDM test results and ANN model estimated between many input and output parameters. Evaluating the
values.
model’s performance only on a single parameter does not
Test WEDM test results ANN model estimates Absolute provide enough evidence to clarify its high accuracy. Hence,
No. SR CS KW MRR SR CS KW MRR error (%) in the present study, absolute error in percentage (Δ) is also
01 3.205 4.322 0.68 123.28 3.167 4.064 0.674 115.92 3.489385 considered to observe the prediction error at each value in the
02 3.159 5.238 0.66 149.41 3.197 5.496 0.665 156.77 2.978015
03 3.224 4.238 0.65 120.91 3.224 4.240 0.65 120.91 3.71E-14 data set and the model’s overall efficiency. The absolute error in
04 2.793 4.182 0.67 119.30 2.832 4.440 0.675 126.66 3.636366 percentage between experimental values and ANN model pre
05 3.304 4.695 0.65 133.94 3.304 4.695 0.65 133.94 1.97E-14 dictions is estimated based on Eq. (5).
06 3.188 3.675 0.6 104.85 3.150 3.417 0.594 97.49 4.043123
07 3.265 2.890 0.65 82.46 3.265 2.890 0.65 82.46 2.31E-14 � �
�Exp Pred�
08 3.133 4.444 0.62 126.77 3.094 4.1863 0.614 119.41 3.433983 Δ¼� � � � 100 (5)
09 3.363 3.259 0.64 92.96 3.401 3.517 0.645 100.32 4.461653 Exp �
10 3.372 2.943 0.61 83.95 3.410 3.201 0.615 91.31 4.896787
11 3.318 3.340 0.64 95.29 3.318 3.340 0.64 95.29 2.14E-14 Calculated absolute error in percentage between experimental
12 3.156 4.583 0.66 130.73 3.118 4.325 0.654 123.37 3.330149 values and ANN model predictions is presented in Table 3. The
13 3.263 5.534 0.65 157.87 3.263 5.534 0.65 157.87 1.08E-14
14 3.57 4.592 0.62 131.01 3.540 4.334 0.614 123.64 3.301964 last column of Table 3 indicates the average absolute error in
15 3.511 5.641 0.57 160.90 3.549 5.899 0.575 168.26 2.804787 percentage for all the machinability measures. From Table 3, it
16 3.300 5.789 0.66 165.14 3.262 5.531 0.654 157.78 2.730374 is perceived that WEDM test results and ANN estimated values
17 3.349 4.746 0.65 135.39 3.387 5.004 0.655 142.75 3.219132
18 3.332 5.605 0.63 159.88 3.332 5.605 0.63 159.88 1.57E-14 are observed to be in a significant harmony with each other.
19 3.526 5.116 0.65 145.93 3.526 5.116 0.65 145.93 1.37E-14 The peak absolute percentage error has been recorded as
20 3.901 5.736 0.65 163.63 3.863 5.478 0.644 156.27 2.709271 5.05%, indicating the high prediction capability of the ANN
21 3.906 2.811 0.63 80.19 3.944 3.069 0.635 87.55 5.055854
22 3.756 4.746 0.65 135.39 3.717 4.488 0.644 128.03 3.188023 model. Moreover, the mean absolute error in percentage of the
23 3.554 4.476 0.62 127.67 3.593 4.734 0.625 135.03 3.37707 developed ANN model is 2.332%, which validates the precision
24 3.478 4.782 0.64 136.42 3.478 4.782 0.64 136.42 2.64E-14 of the model. High R-value and very less absolute percentage
25 3.610 4.746 0.66 135.391 3.648 5.004 0.665 142.75 3.195119
26 3.619 5.729 0.65 163.42 3.619 5.729 0.65 163.42 1.31E-14 error contribute to the ANN’s applicability in modeling non
27 4.091 4.798 0.67 136.86 4.052 4.540 0.664 129.50 3.131359 linear complex processes, especially in problems involving
more than one output parameter.
performance is calculated using ANOVA (Table 4). The results MRR is arc on duration with 27.21%, while the influence of
determined from ANOVA for measured surface roughness, other parameters is not very dominant as arc on. It is because
cutting speed, kerf width, and MRR and corresponding weigh of the high discharge energy at high arc on duration value that
tage of each input parameter is shown in Fig. 9. From Fig. 9, it allows more time to draw the current. As a result, the spark
can be noticed that pulse on duration is the utmost steering energy increases, due to which more work material is melted
variable on the surface finish with percentage contributions of and eliminated by flushing, thereby yielding high MRR.
67.70%, while other parameters do not have a great influence Therefore, to achieve high MRR, arc on duration must be
on surface roughness. As pulse on time increases, the average high.[40] With a percentage contribution of 16.85%, arc off
discharge energy per pulse on time also increases, resulting in duration is the dominant variable on kerf width. During high
globules of debris formation, wide and overlapping craters. arc off duration, the time interval between two consecutive
Their presence affects the surface topography thereby, increas discharges increases, as a result, flushing of molten material
ing the surface roughness. Therefore, less pulse on duration increases significantly, leaving less debris in the working
can result in low surface roughness. A similar trend has been gap.[41] This results in an increase in kerf width. Therefore,
observed in the study.[39] For cutting speed, arc on duration is arc off duration must be lowered to have low kerf width.
the key influencing variable with percentage contributions of Thus, the parameters with a high impact on output response
27%, while the other inputs influence is not as effective as arc variation have been identified. Also, 3D surface plots were
on. More spark duration results in high discharge energy that developed to relate the variation of output response concerning
causes the material to melt and vaporize, thus making the different input parameters. 3D surface plots are shown in
material easy to cut. This results in an increase in cutting Fig. 10 present each output response’s interpretation against
speed. Therefore, a high arc on duration is preferable for high the most and least influencing input parameters. From Fig. 10,
cutting speed. Similarly, the significant variable affecting the it can be noticed that the variation of all the responses has
1082 U. M. R. PATURI ET AL.
Figure 10. 3D surface plots for machinability performances against input parameters.
followed the same trend that is observed from the percentage percentage contribution and 3D surface plots, pulse on dura
contribution. Surface roughness, cutting speed, kerf width, and tion in surface roughness variation, arc on duration in cutting
MRR have shown higher variation concerning the highest speed variation and MRR, and arc off duration in kerf width
influencing parameters, whereas they have shown lesser varia variation are the most significant parameters affecting the out
tion against least affecting parameters. Therefore, from the put response.
MATERIALS AND MANUFACTURING PROCESSES 1083
[22] Gohli, V.; Puri, Y. M. Turning by Electrical Discharge Machining: [33] Jafari, R.; Kahya, M.; Oliaei, S. N. B.; Ünver, H. O.; Özyurt, T. O.
A Review. Proc IMechE part B: J Eng. Manuf. 2017, 231 (2), Modeling and Analysis of Surface Roughness of Microchannels
195–208. DOI: 10.1177/0954405415590560. Produced by μ-WEDM Using an ANN and Taguchi Method.
[23] Ramakrishnan, R.; Karunamoorthy, L. Modeling and J. Mech. Sci. Technol. 2017, 31, 5447–5457. DOI: 10.1007/s12206-
Multi-Response Optimization of Inconel 718 on Machining of 017-1039-7.
CNC WEDM Process. J. Mater. Process. Technol. 2008, 207, [34] Paturi, U. M. R.; Devarasetti, H.; Reddy, N. S.; Kotkunde, N.;
343–349. DOI: 10.1016/j.jmatprotec.2008.06.040. Patle, B. K. Modeling of Surface Roughness in Wire Electrical
[24] Yeganefar, A.; Niknam, S. A.; Asadi, R. The Use of Support Vector Discharge Machining of Inconel 718 Using Artificial Neural
Machine, Neural Network, and Regression Analysis to Predict and Network. Mater. Today Proc. 2020, 38, 3142–3148. DOI: 10.1016/
Optimize Surface Roughness and Cutting Forces in Milling. Int. j.matpr.2020.09.503.
J. Adv. Manuf. Technol. 2019, 105, 951–965. DOI: 10.1007/s00170- [35] Thankachan, T.; Soorya Prakash, K.; Malini, R.; Ramu, S.;
019-04227-7. Sundararaj, P.; Rajandran, S.; Rammasamy, D.; Jothi, S.
[25] Paturi, U. M. R.; Cheruku, S.; Pasunuri, V. P. K.; Salike, S.; Prediction of Surface Roughness and Material Removal Rate in
Reddy, N. S.; Cheruku, S. Machine Learning and Statistical Wire Electrical Discharge Machining on Aluminum Based Alloys/
Approach in Modeling and Optimization of Surface Roughness Composites Using Taguchi Coupled Grey Relational Analysis and
in Wire Electrical Discharge Machining. Mach. Learn. Appl. 2021, Artificial Neural Networks. Appl. Surf. Sci. 2019, 472, 22–35. DOI:
6, 100099. DOI: 10.1016/j.mlwa.2021.100099. 10.1016/j.apsusc.2018.06.117.
[26] Paturi, U. M. R.; Cheruku, S. Application and Performance of [36] Yi, H. J.; Kim, J. W.; Kim, Y. L.; Shin, S. Effects of Cooling Rate on
Machine Learning Techniques in Manufacturing Sector from the the Microstructure and Tensile Properties of Wire-Arc Additive
past Two Decades: A Review. Mater. Today Proc. 2020, 38, Manufactured Ti-6Al-4V Alloy. Met. Mater. Int. 2020, 26,
2392–2401. DOI: 10.1016/j.matpr.2020.07.209. 1235–1246. DOI: 10.1007/s12540-019-00563-1.
[27] Carlone, P.; Aleksendrić, D.; Rubino, F.; Ćirović, V. Artificial Neural [37] Ji, H.; Peng, Z.; Huang, X.; Wang, B.; Xiao, W.; Wang, S.
Networks in Advanced Thermoset Matrix Composite Manufacturing. Characterization of the Microstructures and Dynamic
Springer Int. Publ. 2018. DOI: 10.1007/978-3-319-89563-5_5. Recrystallization Behavior of Ti-6Al-4V Titanium Alloy through
[28] Singh, B.; Misra, J. P. Surface Finish Analysis of Wire Electric Experiments and Simulations. J. Mater. Eng. Perform. 2021. DOI:
Discharge Machined Specimens by RSM and ANN Modelling. 10.1007/s11665-021-06044-0.
Measurement. 2019, 137, 225–237. DOI: 10.1016/j. [38] Shi, X.; Zeng, W.; Sun, Y.; Han, Y.; Zhao, Y.; Guo, P.
measurement.2019.01.044. Properties Correlation for the Ti-6Al-4V Titanium Alloy.
[29] Phate, M. R.; Toney, S. B. Modeling and Prediction of WEDM J. Mater. Eng. Perform. 2015, 24, 1754–1762. DOI: 10.1007/
Performance Parameters for Al/SiCp MMC Using Dimensional s11665-015-1437-x.
Analysis and Artificial Neural Network. Eng. Sci. Technol. Int. J. [39] Chalisgaonkar, R.; Kumar, J.; Pant, P. Prediction of Machining
2019, 22, 468–476. DOI: 10.1016/j.jestch.2018.12.002. Characteristics of Finish Cut WEDM Process for Pure
[30] Yusoff, Y.; Mohd Zain, A.; Sharif, S.; Sallehuddin, R.; Titanium Using Feed Forward Back Propogation Neural
Ngadiman, M. S. Potential ANN Prediction Model for Network. Mater. Today. Proc. 2019, 25, 592–601. DOI:
Multiperformance WEDM on Inconel 718. Neural Comput. Appl. 10.1016/j.matpr.2019.07.260.
2018, 30, 2113–2127. DOI: 10.1007/s00521-016-2796-4. [40] Rehman, M.; Khan, S. A.; Naveed, R. Parametric Optimization in
[31] Sarkar, S.; Ghosh, K.; Mitra, S.; Bhattacharyya, B. An Integrated Wire Electric Discharge Machining of DC53 Steel Using Gamma
Approach to Optimization of WEDM Combining Single-Pass and Phase Coated Wire. J. Mech. Sci. Technol. 2020, 34, 2767–2773.
Multipass Cutting Operation. Mater. Manuf. Process. 2010, 25, DOI: 10.1007/s12206-020-0609-2.
799–807. DOI: 10.1080/10426910903575848. [41] Lalwani, V.; Sharma, P.; Pruncu, C. I.; Unune, D. R. Response
[32] Chen, H. C.; Lin, J. C.; Yang, Y. K.; Tsai, C. H. Optimization of Wire Surface Methodology and Artificial Neural Network-Based
Electrical Discharge Machining for Pure Tungsten Using A Neural Models for Predicting Performance of Wire Electrical Discharge
Network Integrated Simulated Annealing Approach. Expert Syst. Machining of Inconel 718 Alloy. J. Manuf. Mater. Process. 2020, 4.
Appl. 2010, 37, 7147–7153. DOI: 10.1016/j.eswa.2010.04.020. DOI: 10.3390/jmmp4020044.