Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

Materials Today: Proceedings xxx (xxxx) xxx

Contents lists available at ScienceDirect

Materials Today: Proceedings


journal homepage: www.elsevier.com/locate/matpr

Neural network approach for optimizing machining parameters of


aluminium nano composite
K. Vijay Kumar a, T.S.A. Suryakumari b,⇑, S. Ranganathan a
a
Department of Mechanical Engineering, Saveetha School of Engineering, Saveetha Institute of Medical and Technical Sciences, Chennai 602105 Tamilnadu, India
b
Department of mechanical engineering, GMR Institute of Technology, Rajam, Andhra Pradesh, India

a r t i c l e i n f o a b s t r a c t

Article history: In the present study an integrated approach based on Taguchi modelling and Artificial Neural Networks
Received 12 September 2020 (ANN) are used for modelling the experiments and output parameters of Abrasive Water Jet Machining
Received in revised form 10 October 2020 (AWJM). For this process Al7075 nano composite is fabricated by using SiC as reinforcement.
Accepted 14 October 2020
Composite is fabricated using ultrasonic cavitation method. The fabricated composite is machined by
Available online xxxx
AWJM based on Taguchi L18 Orthogonal Array (OA) approach. For ANN modelling the data was selected
according to the Taguchi L18 orthogonal array i.e., the selected input parameters for ANN are Abrasive Jet
Keywords:
Speed, Abrasive Flow Rate, Stand of Distance, Pressure and the outputs from the ANN model are Average
Al7075 nano composite
Ultrasonic cavitation method
KERF width, Material Removal Rate (MRR). Batch Back Propagation is used for Training, Validation and
Taguchi modelling Testing of Artificial neural network. From ANN the correlation coefficient R obtained is above 96% that
Abrasive water jet machining implies the predicted values are 96% in agreement with the experimental values.
Artificial Neural Network Ó 2020 Elsevier Ltd. All rights reserved.
Selection and peer-review under responsibility of the scientific committee of the International Confer-
ence on Mechanical, Electronics and Computer Engineering 2020: Materials Science.

1. Introduction Al7075 reinforced with SiC which was fabricated by stir casting
process. Taguchi technique was carried out for dry sliding wear
The composites where the parent metal is an aluminium alloy test and based on the results it was concluded that sliding distance
are one of the most preferred engineering materials at present con- has highest influence on wear rate followed by sliding speed and
ditions. There are several aluminium alloys, among those Al6061 applied load for 10 wt% SiC reinforcement [2]. Al7075 alloy was
and Al7075 are highly explored due to their mechanical properties. added with varying compositions of SiC and TiC where the higher
Al7075 with SiC are replacing several components like sports com- microhardness is obtained at 90 wt% Al7075 and 8 wt% of TiC and
ponents, under water vehicle components and turbine blades 4 wt% of SiC and also concluded that the incorporation of SiC, TiC
because of their wear resistance and strength. Al7075 reinforced along with Al metal matrix increases the wear resistance property
with Al2O3 are also attracting industrial materials due to their high of the alloy [3]. Research was carried out on Al356/nano-SiC com-
creep resistance as well as high wear resistance. Research was car- posites which were fabricated by stir casting method by varying
ried out on mechanical and tribological properties of hybrid SiC - the volume fractions of SiC particulate reinforcements. From the
Al7075 metal matrix composite where 90 wt% Al7075 alloy, 5 wt study it is concluded that 4.5 vol% of SiC reinforced nano composite
% SiC which is kept as constant and solid lubricants are used for improved the uniform tensile strength among the nano composites
fabrication of composite by stir casting method and based on the [4]. By the addition of 10%SiC and 5% Gr to the parent metal
results it was concluded that graphite reinforcement has signifi- Al2024, it is concluded that hardness improved compared to the
cant impact on static mechanical properties, graphite particles base metal alloy [5]. Al2024 alloy added with 5% B4C was consid-
increases the bonding between the matrix and reinforcement and ered for the study and results indicates the Tensile strength and
the SiC reinforcement increases the ductility of the material [1]. Hardness obtained are maximum [6]. A study conducted on
A study conducted on tribological behaviour of aluminium alloy Al2024 alloy with Tungsten carbide composites and resulted in
improvement of micro hardness and wear characteristics of fabri-
cated composite [7]. Al2024 alloy was taken to study with B4C as
⇑ Corresponding author. reinforcement where UTS and yield strength are improved and
E-mail address: suria.ramana@gmail.com (T.S.A. Suryakumari).

https://doi.org/10.1016/j.matpr.2020.10.407
2214-7853/Ó 2020 Elsevier Ltd. All rights reserved.
Selection and peer-review under responsibility of the scientific committee of the International Conference on Mechanical, Electronics and Computer Engineering 2020:
Materials Science.

Please cite this article as: K. Vijay Kumar, T.S.A. Suryakumari and S. Ranganathan, Neural network approach for optimizing machining parameters of alu-
minium nano composite, Materials Today: Proceedings, https://doi.org/10.1016/j.matpr.2020.10.407
K. Vijay Kumar, T.S.A. Suryakumari and S. Ranganathan Materials Today: Proceedings xxx (xxxx) xxx

decrease in % elongation of the composite when there is an


increase in boron carbide percentage [8].
Abrasive Water Jet Machining is an unconventional machining
process used to remove unwanted material from a given work
piece. This process makes use of an abrasive jet with high velocity,
to remove material and provide smooth surface finish to hard
metallic work pieces. Surface of the work piece is cleaned automat-
ically in this process with water as medium. The effect of traverse
speed on abrasive water jet machining of Ti6Al4V alloy was stud-
ied. Different traverse speeds were used for machining the alloy by
AWJM. Surface profilometry and SEM were used to examine the
microstructures and kerf geometry. It was concluded that the
experimental findings indicated a great significance of traverse
speed on the machined surface profile [9]. To study the influence
of process parameters on MRR and kerf width an experiment was
conducted on glass which was taken as workpiece because of its
homogenous properties and observed that there is direct propor-
tionality between nozzle diameter and MRR and the divergence
of the produced hole was reduced with a decrease in the nozzle
tip distance from the workpiece [10]. A performance test was con-
ducted on different abrasive materials in abrasive water jet
machining of glass and the results reveals that the width of the Fig. 1. Ultrasonic cavitation method experimental setup.
cut increases while increasing the SOD and also demonstrates that
the taper of cut decreases with increase in jet pressure and it is the
Abrasive Flow Rate, Stand of Distance, and Pressure. Machining
case for all the types of abrasives used [11]. The effect of shape and
(Holes of 6 mm) of the work piece is done based on the design of
particle hardness during the abrasive water jet milling process on
experiments as demonstrated in Table 2 (Fig. 2. Fig. 3.).
titanium alloy was studied and concluded that MRR and surface
The process parameters are taken at three levels based on L18
roughness were increased with increase in abrasive particle hard-
OA and are demonstrated in Table 1 and the design of experiments
ness [12]. AWJM experiments were conducted on marble and in
are demonstrated in Table 2. Average KERF width is calculated
the process it is noted that nozzle traverse speed is the significant
using the average of top KERF width and bottom KERF width
factor which affects the kerf taper angle and top kerf width [13].
respectively. Material Removal Rate is calculated using the formula
By the literature study, it is concluded that the mechanical
and it is calculated in (cm3/sec). Measured depth of cut is 12 mm.
properties of the Aluminium metal matrix composites improved
MRR = Depth of cut  Traverse rate  Average kerf width
by reinforcing with ceramic additions. This study presented the
abrasive jet machining on aluminium 7075 metal matrix nano
composite fabricated by ultra sonic cavitation method. The correla-
tion among input and output machining parameters were obtained 3. Ann modelling
using unconventional modelling technique, Artificial Neural Net-
works (ANN). Artificial Neural Networks is having predominant features of
network structure and network training algorithm. ANN is used
to develop the mathematical relations between the process param-
2. Experimental setup
eters. Proper training develops the network model and correlation
of output value with input values. The development techniques
2.1. Ultrasonic cavitation method
will be used for mathematical details and implementation of
ANN. Batch back propagation used for the developing the network,
In the initial stage of this method Aluminium metal is melted in
which is an advanced variant of back propagation where network
the furnace of 900 °C and the used reinforcement SiC nano particles
weight update takes place once per iteration is selected as network
are preheated in the furnace at 450 °C to improve the wettability of
type for ANN modelling and it is also widely used in engineering
the particles. This preheated reinforcement is poured into the Alu-
applications. First layer of the network represents inputs and mid-
minium slurry and stirred using motorized stirrer at constant fur-
dle layer is known as hidden layer and output signals represented
nace temperature. The dispersion time is 5 min. Stirrer cannot
by last layer (Fig. 4.).
disperse the nano particles through out the slurry this may result
The above figure represents the architecture of the neural net-
in properties of the composite which will not same through out
work model which is used for ANN modelling and is same for both
the composite, therefore a vibrator is used for proper dispersion.
the response parameters where ‘c’ indicates the input parameters
The combined metal slurry is poured into the die where the cast-
and ‘N’ indicates the output or response parameters. ANN mod-
ings (Aluminium nano composites) can be removed after proper
elling is done separately for two different response parameters,
cooling. Specimen can be prepared from the fabricated castings.
therefore there will 2 outputs i.e., 1 for each ANN modelling and
Architecture search report is demonstrated in below table
2.2. Abrasive water jet machining (Table 3).
LOGISTIC: An activation function that has a sigmoid curve and is
The experiments are conducted using the AWJM set up has calculated using the following formula: F(x) = 1/(1 + e-x). Its output
shown in Fig. 1 for the fabricated Aluminium nano composite. Gar- range is [0–1]. This function is used most often in multi-layer
net 80 mesh abrasives are used as abrasive flow through the car- perceptrons.
bide nozzle. Unthreaded nozzle is used to supply the abrasive SUM OF SQUARES: The network error function is calculated as
with the water as medium and at 900 operating angle. The process the sum of the squared differences between the actual value (tar-
parameters that selected for the process are Abrasive Jet Speed, get column value) and neural network output.
2
K. Vijay Kumar, T.S.A. Suryakumari and S. Ranganathan Materials Today: Proceedings xxx (xxxx) xxx

Table 1
Range of process parameters.

Machining or Process Parameters Units Levels (1 2 3) Response or Output parameters


Pressure MPa 340 360 - Average KERF width (mm) Material Removal Rate (cm3/sec)
Jet Traverse Speed mm/min 100 200 300
Abrasive Flow Rate g/min 300 400 500
Stand of Distance mm 2 2.5 3

Table 2
Design of experiments (Taguchi L18 OA) and experimental results.

Run Pressure Jet traverse speed (mm/ Abrasive flow rate (g/ Stand of distance Average kerf width Material removal rate (cm3/
(MPa) min) min) (mm) (mm) sec)
1 340 100 300 2 5.807432432 0.116148649
2 340 100 400 2.5 5.800675676 0.116013514
3 340 100 500 3 5.472972973 0.109459459
4 340 200 300 2 5.861486486 0.234459459
5 340 200 400 2.5 5.797297297 0.231891892
6 340 200 500 3 5.864864865 0.234594595
7 340 300 300 2.5 5.675675676 0.340540541
8 340 300 400 3 5.888513514 0.353310811
9 340 300 500 2 5.969594595 0.358175676
10 360 100 300 3 5.898648649 0.117972973
11 360 100 400 2 5.847972973 0.116959459
12 360 100 500 2.5 5.783783784 0.115675676
13 360 200 300 2.5 5.77027027 0.230810811
14 360 200 400 3 5.804054054 0.232162162
15 360 200 500 2 5.952702703 0.238108108
16 360 300 300 3 5.810810811 0.348648649
17 360 300 400 2 5.817567568 0.349054054
18 360 300 500 2.5 5.709459459 0.342567568

Fig. 2. AWJM Experimental set up. Fig. 3. Initial stage of AWJM process.

The data required to develop the ANN model is taken from the by water jet machining process and output represents whereas
Taguchi L-18 Orthogonal Array design of experiments which are the output represents the simulated results obtained by the cre-
demonstrated in Table 2. The modelling is done separately for 2 ated ANN model for average kerf width (mm). AE stands for Abso-
different outputs. Because of 2 different ANN modelling for 2 dif- lute error i.e., error value that indicates the ‘‘quality” of a neural
ferent outputs Data analysis results are demonstrated separately network training. Calculated by subtracting the current output val-
in Table 4. ues with the target output values of the neural network. and ARE
Stands for Absolute Relative Error. ARE is an error value that indi-
4. Results and discussion cates the ‘‘quality” of the neural network training. This index is cal-
culated by dividing the difference between actual and desired
Table 5 represents the performance measures based on the output values by the module of the desired output value. The val-
design of experiments presented in Table 2. Here Target represents ues of AE and ARE obtained are very small that indicates that the
the experimental values of the average kerf width (mm) obtained network trained efficiently.

3
K. Vijay Kumar, T.S.A. Suryakumari and S. Ranganathan Materials Today: Proceedings xxx (xxxx) xxx

Table 5
Experimental and results obtained from created ANN model for response parameter
(Average kerf width (mm)).

Row Target(Exp) Output(ANN) AE ARE


TRN 0 5.807432 5.80305 0.004383 0.075467
TRN 1 5.800676 5.800766 0.00009 0.001557
TRN 2 5.472973 5.510235 0.037262 0.680836
TRN 3 5.861486 5.872719 0.011232 0.191629
TRN 4 5.797297 5.796614 0.000683 0.011789
TRN 5 5.864865 5.858528 0.006337 0.108044
VLD 6 5.675676 5.783814 0.108138 1.905287
TST 7 5.888514 5.809405 0.079108 1.343435
TRN 8 5.969595 5.936978 0.032617 0.546383
TST 9 5.898649 5.544303 0.354346 6.007236
TRN 10 5.847973 5.844867 0.003106 0.053107
TRN 11 5.783784 5.781728 0.002056 0.035541
TRN 12 5.77027 5.771483 0.001212 0.02101
TRN 13 5.804054 5.801494 0.00256 0.044108
VLD 14 5.952703 5.918488 0.034214 0.574769
TRN 15 5.810811 5.809821 0.000989 0.017026
TRN 16 5.817568 5.829494 0.011926 0.205007
TRN 17 5.709459 5.709455 0.000004 0.000071

below in Table 6. The values of AE and ARE obtained are very small
which indicates that the training of network model and its
efficiency.
Fig. 4. Architecture model for ANN modelling for both Response parameters.
Fig. 6 represents the scatter plot between target and output val-
ues of MRR (cm3/sec) and the correlation (R2) is obtained as
Table 3 0.997256 which is almost equivalent to 1 that indicates good cor-
Architecture report for ANN modelling. relation between the experimental and simulated values. Fig. 9 and
Architecture selected manually[10– Architecture selected manually[10–
Fig. 10 shows the data set error variation of experimental and mod-
5–1] architecture selected for 5–1] architecture selected for elling and network error variations of MRR simultaneously. Net-
trainingHidden layers activation trainingHidden layers activation work error for MRR gradually decreased and then Increased.
function: LogisticOutput function: LogisticOutput parameters:
parameters:AVERAGE KERF MATERIAL REMOVAL RATE(cm3/sec)
WIDTH (mm)Error function: Sum- Error function: Sum-of-
4.1. Comparison plots
of-squaresActivation function: squaresActivation function: Logistic
Logistic Fig. 11 and Fig. 12 represents the relation between experimen-
tal and output results of ANN from the tables which compares the
experimental values and simulated values from ANN model for the
Response parameters. From figure it is observed that the close rela-
Table 4 tion between experimental and modelled values of kerf width as
ANN Data analysis results.
well as material removal rate. From the figures it has observed that
5 columns and 18 rows analyzed 5 columns and 18 rows analyzed the ANN modelling technique predicted very well the responses
5 columns and 18 rows accepted for 5 columns and 18 rows accepted for required.
neural network training neural network training
4 categorical columns:PRESSURE 4 categorical columns:PRESSURE
5. Conclusion
(MPa)JET TRAVERSE SPEED (mm/ (MPa)JET TRAVERSE SPEED (mm/
min)ABRASIVE FLOW RATE (g/ min)ABRASIVE FLOW RATE (g/min)
min)STAND OF DISTANCE (mm) STAND OF DISTANCE (mm) The proposed integrated approach i.e., Taguchi design for con-
1 numeric columns:AVERAGE KERF 1 numeric columns:MRR (cm3/sec) ducting experiments and ANN modelling are used for the mod-
WIDTH (mm) elling purpose of Response parameters obtained from Abrasive
Data partition method: random Data partition method: random
Data partition results:14 records to Data partition results:14 records to
Water Jet Machining process. Modelling of parameters of AWJM
Training set (77.78%)2 records to Training set (77.78%)2 records to process which was designed using Taguchi L18 Orthogonal Array.
Validation set (11.11%)2 records Validation set (11.11%)2 records to ANN modelling used for development of mathematical relations
to Test set (11.11%) Test set (11.11%) among parameters and prediction accuracy of responses.
Experimental values and the output values obtained from ANN
modelling are compared and the values of AE and ARE obtained are
Fig. 5 represents the scatter plot between target and output for very small which indicates the best performance of ANN model for
Average kerf width (mm) and shows most of the points lie on line. the given process parameters and it also indicates that this mod-
This shows the precision of the experimental values. The correla- elling approach can be used to predict the response parameters
tion (R2) is obtained as 0.977284 which is almost equivalent to 1 in Abrasive Water Jet Machining which improves the decision
that indicates good correlation. Fig. 7 and Fig. 8 shows the data making in process planning.
set error variation of experimental and modelling and network
error variations of kerf width simultaneously. Network error for CRediT authorship contribution statement
kerf width has increased initially and further has decreased.
The experimental values of MRR (cm3/sec) (Target) and the sim- K. Vijay Kumar: Methodology, Software, Formal analysis, Inves-
ulated values obtained from ANN model (Output) are tabulated tigation, Resources, Data curation, Visualization. T.S.A. Suryaku-
4
K. Vijay Kumar, T.S.A. Suryakumari and S. Ranganathan Materials Today: Proceedings xxx (xxxx) xxx

Fig. 5. Scatter plots of Target and output for response parameter (Average kerf width (mm)) and (MRR (cm3/sec)).

Fig. 6. Scatter plots of Target and output for response parameter (Average kerf width (mm)) and (MRR (cm3/sec)).

5
K. Vijay Kumar, T.S.A. Suryakumari and S. Ranganathan Materials Today: Proceedings xxx (xxxx) xxx

Fig. 7. Data set errors for the response parameter (Average kerf width (mm)).

Fig. 8. Network errors for the response parameter (Average kerf width (mm)).

Table 6
Experimental and results obtained from created ANN model for response parameter (MRR (cm3/sec)).

Row Target(Exp) Output(ANN) AE ARE


TRN 0 0.116149 0.115506 0.000643 0.553661
TRN 1 0.116014 0.116106 0.000092 0.079684
TRN 2 0.109459 0.115534 0.006075 5.549584
TRN 3 0.234459 0.234492 0.000032 0.013696
TRN 4 0.231892 0.23203 0.000138 0.059455
TRN 5 0.234595 0.234456 0.000138 0.058935
TRN 6 0.340541 0.34257 0.002029 0.59585
TST 7 0.353311 0.344576 0.008734 2.472141
VLD 8 0.358176 0.342329 0.015847 4.424338
TRN 9 0.117973 0.115525 0.002448 2.075417
TRN 10 0.116959 0.116297 0.000662 0.566159
TRN 11 0.115676 0.115489 0.000187 0.16175
TRN 12 0.230811 0.230451 0.00036 0.156087
TRN 13 0.232162 0.232349 0.000187 0.080577
VLD 14 0.238108 0.237994 0.000114 0.048069
TST 15 0.348649 0.342562 0.006087 1.74584
TRN 16 0.349054 0.345324 0.00373 1.06863
TRN 17 0.342568 0.343347 0.00078 0.227618

Fig. 9. Data set errors for the response parameter (MRR (cm3/sec)).

Fig. 10. Network errors for the response parameter (MRR (cm3/sec)).

6
K. Vijay Kumar, T.S.A. Suryakumari and S. Ranganathan Materials Today: Proceedings xxx (xxxx) xxx

Fig. 11. Comparison of target and output values for Average kerf width (mm).

Fig. 12. Comparison of target and output values for MRR (cm3/sec).

mari: Conceptualization, Validation, Supervision, Project adminis- [2] S. Madhavarao, Ch. Ramabhadri Rju, J. Madhukiran, N. Sudheerkumar Varma, P.
Ravi Varma, Mater. Today Proc. 5 (9) (2018) 20013–20022.
tration. S. Ranganathan: Resources, Data curation.
[3] Pradeep Devaneyan, Ganesh Ramalingam, T. Senthilvelan, Indian J. Mater. Sci.
(2017), 1-6.
[4] Ali Mazahery, Mohsen Ostad Shabani, J. King Saud Univ. - Eng. Sci. 25 (2013)
Declaration of Competing Interest 41.
[5] Marlon Jones Louis, Int. J. Res. Aeronaut. Mech. Eng. 2 (2014) 29.
The authors declare that they have no known competing finan- [6] Bhargavi Rebba, N. Ramanaiah, Adv. Mater. Manuf. Charact. 4 (2014).
[7] R. Abhijith, T.M. Harish, Int. J. Eng. Res. Technol. (IJERT), 5(2016).
cial interests or personal relationships that could have appeared [8] M. Nagaral, R. Pavan, P.S. Shilpa, V. Auradi, FME Trans. 45 (1) (2017) 93–96.
to influence the work reported in this paper. [9] M. Wakuda, Y. Yamauchi, S. Kanzaki, J. Mater. Process. Technol. 132 (1–3)
(2003) 177–183.
[10] D.V. Srikanth, M. SreenivasaRao, Procedia Mater. Sci. (2014), 6, 1303-1311.
References [11] A.A. Khan, M.M. Haque. J. Mater. Process. Technol. (2007), 191(1-3), 404-407.
[12] G. Fowler, I.R. Pashby, P.H. Shipway, Wear 266 (7–8) (2009) 613–620.
[1] S. Devaganesh, P.K. Dinesh kumar, N. Venkatesh, R. Balaji, J. Mater. Res. [13] V. Gupta, P.M. Pandey, M.P. Garg, R. Khanna, N.K. Batra. Procedia Mater. Sci. 6,
Technol. (2020), 9(3), 3759-3766. 140-149.

You might also like