Professional Documents
Culture Documents
Artificial Neural Network Simulation and Particle Swarm Optimisation of Friction Welding Parameters of 904L Superaustenitic Stainless Steel
Artificial Neural Network Simulation and Particle Swarm Optimisation of Friction Welding Parameters of 904L Superaustenitic Stainless Steel
Artificial Neural Network Simulation and Particle Swarm Optimisation of Friction Welding Parameters of 904L Superaustenitic Stainless Steel
www.emeraldinsight.com/1573-6105.htm
MMMS
10,2
Artificial neural network
simulation and particle swarm
optimisation of friction welding
250 parameters of 904L
Received 11 July 2013
Revised 9 September 2013 superaustenitic stainless steel
Accepted 9 October 2013
K. Balamurugan
Department of Mechanical Engineering, Periyar Maniammai University,
Thanjavur, India
A.P. Abhilash and P. Sathiya
Department of Production Engineering, National Institute of Technology,
Tiruchirappalli, India, and
A. Naveen Sait
Department of Mechanical Engineering,
Chendhuran College of Engineering and Technology, Pudukkottai, India
Abstract
Purpose – Friction welding (FW) is a solid state joining process. Super austenitic stainless steel is the
preferable material for high corrosion resistance requirements. These steels are relatively cheaper than
austenitic stainless steel and it is expensive than nickel base super alloys for such applications.
The purpose of this paper is to deal with the optimization of the FW parameters of super austenitic
stainless steel using artificial neural network (ANN) simulation and particle swarm optimization (PSO).
Design/methodology/approach – The FW experiments were conducted based on Taguchi L-18
orthogonal array. In FW, rotational speed, friction pressure, upsetting pressure and burn-off length are
the important parameters which determine the strength of the weld joints. The FW trials were carried
out on a FW machine and the welding time was recorded for each welding trial from the computerized
control unit of the welding machine. The left partially deformed zone (L.PDZ) and right partially
deformed zone (R.PDZ) were identified from the macrostructure and their values are considered for the
output variables. The tensile test was carried out, and the yield strength and tensile strength of the
joints were determined and their fracture surfaces were analyzed through scanning electron
microscope (SEM).
Findings – The tensile test was carried out, and the yield strength and tensile strength of the joints
were determined and their fracture surfaces were analyzed through SEM. An ANN was designed to
predict the weld time, L.PDZ, R.PDZ and tensile strength of the joints accurately with respect to the
corresponding input parameters. Finally, the FW parameters were optimized using PSO technique.
Research limitations/implications – There is no limitations, difficult weld by fusion welding
process material can easily weld by FW process.
Originality/value – The research work described in the paper is original.
Keywords Microstructure, Friction welding, 904L, ANN, PSO, Super austenitic stainless steel
Paper type Research paper
2. Experimental details
2.1 FW
FW was carried out using a continuous drive FW machine with a maximum 60 tones
capacity. The FW machine set up is shown in Figure 1. The machine has an advantage
of adjusting the burn-off length (BOL), unlike other FW machines where the BOL was
an output parameter. The base material’s (AISI 904L super austenitic stainless steel)
chemical composition is presented in Table I.
Figure 1.
Friction welding machine
Table I. Elements Si Mn P S Cr Ni Mo C Cu
Chemical composition of
the base material (wt. %) Composition (%) 0.374 1.522 0.018 0.004 19.893 25.557 4.124 0.018 1.650
In this study, the experiments were conducted based on Taguchi L-18 orthogonal Optimisation of
array. The rotational speed (RS), friction pressure (FP), upset pressure and BOL were FW parameters
the input parameters with three levels each. The individual welding parameters and
their levels are presented in Table II. For each experiments two sets trials were
performed and their average values were taken in to the output results. The welded
joints are shown in Figure 2.
The weld profiles were prepared by machining process, and cut into a cross section 253
of 10 10 mm and polished with suitable abrasive and diamond paste. Weld samples
were etched with 10 percent oxalic acid, an electrolyte, to state and increase the
contrast of the fusion zone with the base metal. The typical macrostructure of the
friction welds are shown in Figure 3(a-d). In this macrostructure, the Dexil Imagine ver
1.3.2 software were used and measured the width of weld zone, left partially deformed
zone (L.PDZ) and right partially deformed zone (R.PDZ). The weld zone and partially
deformed zones were distinguished with grain size. The finer grain size is obtained in
all the welds and coarse grains were obtained in both left and R.PDZs .
Figure 2.
Welded specimens
MMMS (a) (b)
10,2
0.596 mm
0.637 mm
0.575 mm
254 0.390 mm
0.575 mm
0.534 mm
1 mm 1 mm
0.493 mm
0.534 mm 0.513 mm
0.617mm
0.575 mm
1 mm
Figure 3. 1 mm
Macrograph of the
weld joints (LPDZ=0.617 mm, WELD ZONE=0.534 mm (LPDZ=0.513 mm, WELD ZONE=0.493 mm
and RPDZ=0.575 mm) and RPDZ=0.678 mm)
Figure 4.
Fractured tensile
specimens
FP
OUTPUTS
UP 1) Welding time
2) L.PDZ
3) R.PDZ
BOL
4) Tensile strength
Figure 5.
Schematic diagram of
SPEED
ANN used for simulation
(accuracy), number of epochs and the number of output parameters. In this study, each
of the different outputs was simulated one at a time. So the number of output was taken
as 1. The number of inputs was 4. The maximum number of epochs was taken as 1,000
and the goal was taken as 0.001. So, either the program would be trained for 1,000
iterations or up to an accuracy of 3 decimal places. The graph of the accuracy verses
the number of iterations is shown in Figure 6. The number of hidden layers is the most
important parameter to be determined. This was determined by trial and error method.
The number of hidden layers for weld time, L.PDZ and R.PDZ were selected as four,
six and three respectively. The number of hidden layers for percentage of elongation,
final diameter, yield strength and ultimate strength were three, four and five,
respectively.
MMMS Sum-Squared Network Error for 10 Epochs
100
10,2
10–1
256
10–2
10–3
Figure 6.
10–4
Accuracy vs number
of iterations 0 1 2 3 4 5 6 7 8 9 10
Epochs
The MATLAB codes for the ANN simulation of all the output variables were written
separately with the selected input parameters being given inside the program. The
training inputs and the training outputs were also entered in the program in their
normalized form, to reduce the error. After the MATLAB codes were written, each of
the programs was run in the software. The training was performed for the selected
number of iterations or the selected accuracy value, whichever came first. After the
ANN was trained, the input variables were given to the program to simulate the
respective output parameter. The values were obtained from the program and then
compared with the experimental values.
2.5 PSO
The FW parameters such as RS, FP, upsetting pressure (UP) and BOL highly
influences the tensile properties of the weld which in turn dictate the qualities of the
joints. Increased tensile strength and reduced welding time, L.PDZ and R.PDZ are the
main objectives of this study. The PSO flowchart is as shown in Figure 7. The program
for optimization was written using MATLAB 6.5. The objective function was a sum of
four different ANN functions of the above-mentioned outputs. The weightage for
welding time, L.PDZ and R.PDZ were taken as 0.1 each and the weightage for tensile
strength was taken as 0.7. This is done because the tensile strength is the most
significant parameter to be optimized.
Every time the PSO worked, it started with a random initial value and arrived at the
optimum final value. Since the initialization was random, each time the PSO took
different routes to arrive at the optimum value.
While
termination
END conditions is 257
not met
(a) (b)
Figure 8.
1mm 1mm Microstructures of the
weld zone
Notes: (a) Specimen no. 4; (b) specimen no. 12
also easily visible in the partial deformation zones. A mixture mode of equiaxed
dendrite grain structures and slight long columnar structures were observed in the
weld zone. Relatively high amount of secondary interdendrite phase (darker area) was
also observed in the weld zone of the joints.
MMMS 3.2 Tensile tested fracture surfaces
10,2 The photographs of the fracture surfaces of the tensile tested specimens are presented
in Figure 9. It is clearly seen that, the reduced cross section had occurred in the fracture
region. And further the fracture surfaces were analyzed through SEM. The tensile
tested fracture surface of the SEM structures is presented in Figure 10(a-d).
From Figure 10, it is clearly seen that, the shear flow of material and also miner
258 dimples were observed in the micrographs. Due to shear flow of material the necking
took place in the joints and failure occurred, with ductile mode of fractures.
Figure 9.
Photographic view of the
fracture surfaces
of tensile specimens
(a) (b)
(c) (d)
Figure 10. SE 06-Jun-11 NITT WD20.7 mm X50 1 mm SE 06-Jun-11 NITT WD19.1 mm X500 100 um
SEM microstructures
of the joints
Notes: (a) Exp. no. 3; (b) exp. no. 4; (c) exp. no. 13; (d) exp. no. 18
the joints. The simulation was carried out using MATLAB 6.5. For training the ANN, Optimisation of
two-third of the experiment sets were fed into the ANN model. The technique of FW parameters
random selection was employed to select the experiments for training purpose. In this
work, the experiments were selected for training purpose, and the experiments numbers
were 1, 2, 4, 5, 7, 8, 10, 11, 14, 15, 16 and 17. The remaining experiments were directly
taken for testing the ANN model. The results obtained from experiments and simulation
results for welding times were tabulated and also compared with both the results and the 259
calculated percent of error is presented in Table IV. The comparison graphs for welding
time with respect to the experimental and simulated result are shown in Figure 11.
Exp. no. Measured welding time (s) ANN simulated welding time (s) Percentage error (%)
Training experiments
1 62.78 63.18 0.64
2 33.5 33.69 0.58
4 196.43 196.55 0.061
5 36.64 36.97 0.91
7 128.63 129.19 0.43
8 24.47 24.87 1.66
10 81.05 81.00 0.04
11 24.41 24.89 1.99
14 35.64 36.17 1.49
15 17 17.04 0.25
16 209.07 208.18 0.42
17 31.89 32.44 1.74
Testing experiments
3 17.25 15.63 9.39
6 15.63 16.27 4.09
9 20.5 21.59 5.35 Table IV.
12 15.02 15.9 5.85 Comparison of
13 119.17 115.87 2.77 experimental and ANN
18 20.52 20.38 0.67 simulated welding time
140
Experimental
120
ANN Simulated
100
Welding Time (sec)
80
60
40
Figure 11.
20
Plot of experimental
welding time and
0 simulated welding time vs
0 3 6 9 12 13 18 experiment number
Experiment Number
MMMS The results obtained from experiments and simulation results for both partial
10,2 deformed zones (left and right) are tabulated and also compared with both the results
and the calculated percent of error is presented in Table V. The comparison graphs for
both partial deformed zones (left and right) with respect to the experimental and
simulated results are shown in Figures 12 and 13.
The results obtained from experiments and simulation results for tensile tests are
260 tabulated and also compared with both the results and the calculated percent of error is
Training experiments
1 0.575 0.577 0.36 0.637 0.577 9.42
2 0.534 0.533 0.074 0.596 0.633 6.21
4 0.658 0.640 2.64 1.006 0.94 6.56
5 0.493 0.494 0.20 0.411 0.394 4.14
7 0.617 0.621 0.76 0.575 0.621 8.00
8 0.658 0.650 1.12 0.678 0.65 4.13
10 0.596 0.595 0.03 0.678 0.695 2.51
11 0.513 0.514 0.29 0.678 0.614 9.44
14 0.37 0.370 0.05 0.493 0.47 4.67
15 0.534 0.533 0.05 0.555 0.533 3.96
16 0.493 0.493 0.06 0.431 0.453 5.10
17 0.555 0.555 0 0.596 0.555 6.88
Testing experiments
3 0.308 0.327 6.17 0.452 0.494 9.40
Table V. 6 0.431 0.445 3.38 0.575 0.549 4.36
Comparison of 9 0.78 0.742 4.80 0.76 0.771 1.53
experimental and 12 0.514 0.500 2.70 0.596 0.588 1.34
ANN-simulated 13 0.534 0.530 0.73 0.699 0.631 9.62
bead profile 18 0.555 0.593 6.97 0.699 0.671 3.87
1.0
0.8
L.PDZ (mm)
0.6
0.4
0.6
261
0.4
0.2
Figure 13.
Experimental
Plot of experimental
ANN Simulated
R.PDZ and simulated
0.0 R.PDZ vs experiment
0 3 6 9 12 13 18 number
Experiment Number
presented in Table VI. The comparison graphs for tensile test with respect to the
experimental and simulated results are as shown in Figure 14.
Figure 11 shows the comparison of experimental and simulated values of welding
time. It can be seen from the figure that the experimental and simulated welding times
are in equal agreement with each other with a maximum error of 710 percent.
Figure 12 shows the comparison of experimental and simulated values of L.PDZ. It can
be seen from Figure 12, that the experimental and simulated values of L.PDZ are in
equal agreement with each other with a maximum error of 710 percent. Figure 13
shows the comparison of experimental and simulated values of R.PDZ. It can be seen
from the figure that the experimental and simulated welding values of R.PDZ are in
Training experiments
1 660.97 660.40 0.08
2 709.67 701.05 1.21
4 621.87 625.28 0.54
5 692.47 688.75 0.53
7 648.82 647.18 0.25
8 669.88 684.07 2.11
10 656.9 653.74 0.47
11 677.18 684.69 1.10
14 685.28 685.45 0.02
15 709.67 703.28 0.89
16 639.14 644.63 0.85
17 701.34 696.96 0.62
Testing experiments
3 695.4 708.60 1.89
6 709.67 699.91 1.37 Table VI.
9 746.48 691.11 7.41 Comparison of
12 709.93 697.61 1.73 experimental and
13 647.13 649.28 0.33 ANN-simulated
18 750.18 705.09 6.01 tensile properties
MMMS 800
10,2
700
500
equal agreement with each other with a maximum error of 710 percent. Figure 14
shows the comparison of experimental and simulated values of tensile strength. It can
be seen from the figure that the experimental and simulated values of tensile strength
are in equal agreement with each other with a maximum error of 710 percent.
3.4 PSO
The welding parameters are optimized using PSO by coding in MATLAB software.
The properties to be optimized are welding time, L.PDZ, R.PDZ and tensile strength.
Since tensile strength was the significant parameter, it was given more weightage for
optimization. Tensile strength was given a weightage of 0.7, and the remaining
properties such as L.PDZ, R.PDZ and welding time were given an equal weightage of
0.1 each. The results obtained are shown in Table VII.
The near optimum FW parameters are: speed – 1,850 rpm, FP – 80 MPa, UP –
175 MPa and BOL – 4.5 mm. The confirmation test was carried out using the same
ETA FW machine and the results are tabulated in Table VIII.
4. Conclusions
The following important conclusions are drawn from this work:
. The dendrite and interdendrite grain structures were observed in the weld zone.
The fine dendrite arm spacing resulted in higher tensile strength of the weld joints.
Table VII. Speed (rpm) Frictional pressure (MPa) Upsetting pressure (MPa) Burn-off length (mm)
Optimum parameter
values 1,827.27 78.252 173.53 4.4118
Table VIII.
Output values Welding time (s) L.PDZ (mm) R.PDZ (mm) Tensile strength (MPa)
corresponding to optimum
parameter values 43.17 0.516 0.609 685.71
. The tensile tested samples were fractured at the nearby joint zone and partly Optimisation of
penetrated through the parent material. FW parameters
. In all the tensile samples, fracture appeared like shear flow type pattern and
revealed small dimples. It is confirmed that, the ductile mode fractures occurred.
. The simulated results obtained are in good agreement with the experimental results.
. The FW parameters are optimized and the values are: speed – 1,827.2 rpm, 263
FP – 78.252 MPa, UP – 173.53 Mpa, BOL – 4.41 mm.
. The welding was performed with optimized parameters and their results are as
follows: welding time – 43.17 s, L.PDZ – 0.516 mm, R.PDZ – 0.609 mm, tensile
strength – 685.71 MPa.
References
Akbari Mousavi, S.A.A. and Rahbar Kelishami, A. (2008), “Experimental and numerical analysis
of the friction welding process for the 4340 steel and mild steel combinations”, Welding
Research, Vol. 87, pp. 178-186.
Bendzsak, G.B., North, T.H. and Smith, C.B. (2000), “An experimentally validated 3D model for
friction stir welding”, Proceedings of the Second International Symposium on Friction Stir
Welding, Department of Metallurgy and Materials Science, University of Toronto, Ontario.
Benjounis, K.Y., Olabi, A.G. and Hasmi, M.S.J. (2005), “Effect of laser welding parameters on
the heat input and weld-bead profile”, Journal of Materials Processing Technology,
Vol. 164/165, pp. 978-985.
Bull, C.E., Stacey, K.A. and Calcraft, R. (1993), “On line weld monitoring using ultrasonic”,
Journal of Non-destructive Test, Vol. 35 No. 2, pp. 57-64.
Can, A., Sahin, M. and KücüK, M. (2010), “Modelling of friction welding”, International Scientific
Conference, Vol. 2, pp. 135-142.
Cook, G.E. (1990), “Feedback and adaptive control in automated arc welding system”, Metal
Construction, Vol. 13 No. 9, pp. 551-556.
Jeng, J.Y., Mau, T.F. and Leu, S.M. (2000), “Prediction of laser butt joint welding parameters using
back-propagation and learning vector quantization networks”, Journal of Materials
Process Technology, Vol. 99, pp. 207-218.
Juang, S.C., Tarng, Y.S. and Lii, H.R. (1998), “A comparison between the back propagation and
counter-propagation networks in the modeling of the TIG welding process”, Journal of
Materials Process Technology, Vol. 75, pp. 54-62.
Kim, I.S. and Park, C.E. (2000), “Use of a neural network to control bead width in GMA welding”,
Australian Welding Journal, Vol. 45, pp. 33-43.
Kubiszyn, I. and Pietras, A. (2003), “Numerical modeling of the friction-welding process”,
Welding International, Vol. 17 No. 6, pp. 425-430.
Li, P., Fang, M.T.C. and Lucas, J. (1997), “Modeling of submerged arc welding bead using self
adaptive offset neural network”, Journal of Materials Process Technology, Vol. 71, pp. 228-298.
Montgomery, D.C. (1984), Design and Analysis of Experiments, 2nd ed., Wiley, New York, NY.
Nagesh, D.S. and Datta, G.L. (2002), “Prediction of weld bead geometry and prediction in shielded
metal-arc welding using artificial neural networks”, Journal of Materials Process
Technology, Vol. 79, pp. 1-10.
Sathiya, P., Aravindan, S. and Noorul Haq, A. (2008), “Tensile properties of similar AISI 304
austenitic and AISI 430 ferritic stainless steels joined by friction welding”, Multidiscipline
Modelling in Mat. and St., Vol. 4, pp. 141-154.
MMMS Sathiya, P., Aravindan, S., Noorul Haq, A. and Paneerselvam, K. (2009), “Optimization of friction
welding parameters using evolutionary computational techniques”, Journal of Materials
10,2 Processing Technology, Vol. 209, pp. 2576-2584.
Srikanthan, L.T. and Chandel, R.S. (1988), “Neural network based modeling of GMA welding
process using small data sets”, Proceedings of the Fifth International Conference on
Control, Automation, Robotics and Vision, Singapore, pp. 474-482.
264 Tang, Y.S., Tsai, H.L. and Yeh, S.S. (1999), “Modeling, optimization and classification of weld
quality in tungsten inert gas welding”, International Journal of Machine Tools &
Manufacture, Vol. 39, pp. 1427-1438.
Tarng, Y.S. and Yang, W.H. (1998), “Optimization of the weld bead geometry in gas Tungsten Arc
welding by the Taguchi Method”, Journal of Advanced Manufacturing Technology, Vol. 14,
pp. 549-554.
Torres-Treviño, L.M., Reyes-Valdes, F.A., Lopez, V. and Praga-Alejo, R. (2011), “Multi-objective
optimization of a welding process by the estimation of the Pareto optimal set”, Expert
Systems with Applications, Vol. 38, pp. 8045-8053.
Veera Babu, K., Ganesh Narayanan, R. and Saravana Kumar, G. (2009), “An expert system based
on artificial neural network for predicting the tensile behaviour of tailor welded blanks”,
Expert Systems with Applications, Vol. 36, pp. 10683-10695.
Wang, Y., Li, B., Weise, T., Wang, J., Yuan, B. and Tian, Q. (2011), “Self-adaptive learning based
particle swarm optimization”, Information Sciences, Vol. 181 No. 15, pp. 4515-4538.
Zhang, Y.M., Kovacevic, R. and Li, L. (1996), “Characterization and real time measurement of
geometrical appearance of the weld pool”, International Journal of Machine Tools &
Manufacture, Vol. 36 No. 7, pp. 799-816.
Corresponding author
Dr A. Naveen Sait can be contacted at: naveensait@yahoo.co.in