Comprehensive Development and Comparison of two Feed Forward Back Propagation Neural Networks for Forward and Reverse Modeling of Aluminum Alloy AA5083; H111 TIG Welding Process

Suneel RamachandraJoshi Int. Journal of Engineering Research and Application www.ijera.com
ISSN: 2248-9622, Vol. 6, Issue 5, ( Part -7) May2016, pp.98-111
www.ijera.com 98 | P a g e
Comprehensive Development and Comparison of two Feed
Forward Back Propagation Neural Networks for Forward and
Reverse Modeling of Aluminum Alloy AA5083; H111 TIG
Welding Process
Suneel RamachandraJoshi *, Dr.J.P.Ganjigatti **
*
Research school,Department of Industrial Engineering and Management,SIT,TumkurKarnataka,India.
**
Professor, Department of Industrial Engineering and Management, SIT, Tumkur, Karnataka, India.
ABSTACT
The development of an intelligent system for the establishment of relationship between input parameters and the
responses utilizing both reverse and forward modeling of artificial neural networks is the main objective of the
present research work. Prediction of quality characteristics such as front width, back width, front height and
back height of the weld bead geometry in Tungsten Inert Gas welding process of AA5083; H111 Aluminum
alloy is the aim in forward modeling from known set of process parameters such as current, %balance, welding
speed, arc gap, gas flow rate, and frequency. Reverse modeling meets the industrial requirements of automatic
welding to predict the recommended weld bead geometry characteristics. Comprehensive approach for the
development of two back propagation networks viz. feed forward back propagation (FFBP) and Elman back
propagation (EBP) neural networks is adopted. 212 Face centered central composite design based experimental
data is utilized for the development of both supervised learning networks with batch mode training approach. A
comparison of performance of FFBPP and EBP neural networks are made with that of stepwise multiple
regression statistical modeling. Analysis of results showed that both neural network modeling outperformed the
statistical approach in making better predictions and the models are efficient in selection of parameters
effectively for the desired responses. FFBP performance found to marginally better than that of EBP neural
network. Also the forward modeling performance was better than that of reverse modeling in both neural
networks.
Key Words: Artificial Neural Networks, TIG welding, weld bead Characteristics, forward feed back
propagation (FFBP) & Elman Back Propagation (EBP) Networks.
I. INTRODUCTION
Some of the quality requirements such as
corrosion resistance, high strength to weight ratio,
toughness and formability have made aluminum as
the best material in most of the fabrication
industries. One of the aluminum alloy, AA5083 ;
H111 famously known as pressure vessel alloy and
strongest non heat treated wrought alloy finds lot of
application in cryogenic corrosion resistant
applications, low temperature transport vessels &
radioactive water waste tanks fabrication, ship
building and general transport vehicles industries.
TIG welding because of its superior weld quality
plays an important role in modern manufacturing
especially in aerospace, automobile and ship
building industries. TIG welding is utilized for the
welding of aluminum, magnesium, stainless steel
and titanium materials. The heat generated by the
electric arc established between the tungsten
electrode and the base metal with inert-gas
shielding produces the coalescence. As it plays an
important role in determining the mechanical
properties of weldment, the weld bead geometry
strongly characterizes the final quality of the TIG
welding. Welding process parameters such as
current, welding speed, stand-off distance and gas
flow rate affect the quality of weld bead geometry.
The weld bead geometry parameters may include
front width, back width, front height, back height,
penetration, and HAZ etc. This clearly indicates the
complex, multivariate, multi response nature of the
TIG welding process. The literature confirms TIG
welding as highly non linear and strongly coupled
process with never ending interest in researching
for input-output relations to obtain high level of
quality under different circumstances. A
manufacturing process like TIG welding needs to
be automated to ensure both high productivity and
good quality which in turn requires a proper tested
model. Many researchers have applied different
statistical techniques successfully for this purpose.
But the automation of any process requires input-
output relationships to be known in both forward
and reverse directions. As the transformation
RESEARCH ARTICLE
OPEN ACCESS

matrix turns singular and might not be invertible
always, the backward prediction i.e. determination
of process parameters to predict the desired
outputs, might become difficult through the
conventional statistical techniques. Soft computing
techniques like neural networks (NN), genetic
algorithms(GA) and Fuzzy logic(FL) etc. have
made the generation of an integrated system that
estimates two or more responses simultaneously
and reverse prediction modeling , a possibility. In
recent past, application of neural networks for
modeling input - output relationships for
complicated manufacturing processes like casting,
machining and various welding processes has
started to replace the earlier statistical techniques.
This is due to the shear ability of neural networks
to learn and generalize (interpolate) the
complicated input-output relations. Several
researchers have attempted to use neural networks
for various welding process modeling
II. LITERATURE SERVEY
T.G Lim, and H.S cho, [1] proposed a
neural network model for the estimation of Weld
pool sizes in GMA welding of 200×60×4 hot rolled
AISI 1025 Plates. Utilizing the variable polarity
plasma arc data of aluminum welding,
George.E.Cook, Robert Joel Barnett, Kristinn
Andersen, and Alvin M Strauss, [2] used artificial
neural networks (ANN) in its modeling, analysis &
control application. S.C. Juang, Y.S.Tarng, and
H.R. Lii, [3] described both back-propagation
(BPNN) and counter-propagation (CPNN)
networks for modeling TIG Welding process of
pure Aluminum 1100 sheet of 1.6mm with a single
pass with reasonable accuracy & found BPNN
with better generalization & CPNN with better
learning ability. . Kim et al [4] proposed two
different neural networks using two different
training algorithms for predicting the weld bead
width as a function of key process parameters and
found Lavenberg-Marquardt algorithm to perform
better over error back propagation. . D.S Nagesh &
Datta [5] explored the use of Back Propagation
Neural Network to model SMAW process of grey
C.I plates with M.S electrodes to relate welding
process variables with bead geometry. The network
achieved good agreement with the training data &
yielded satisfactory generalization. . Parikshit
Dutta, and Dilip Kumar Pratihar [6] proposed two
neural network-based approaches (i.e., back-
propagation neural network and genetic-neural
system) Both the NN-based approaches were found
to be more adaptive compared to the conventional
regression analysis, for the test cases. Genetic-
neural (GA-NN) system outperformed the BPNN
in most of the test cases (but not all). Taking the
results of submerged arc welding process from
V.Gunarajan & N.Murugan, K.Manikya Kanti, and
P.Shrinivasa Rao [7] developed back propagation
neural network model for the prediction of weld
bead geometry in pulsed gas metal arc welding
process with correlation coefficient of 0.99.
Amarnath & Pratihar [8] solved forward and
reverse mapping problems of the tungsten inert gas
(TIG) welding process using radial basis function
neural networks (RBFNNs). Nagesha & Datta[9]
developed a back-propagation neural network &
Genetic algorithm to optimize the process
parameters for front height to front width ratio and
back height to back width ratio yielded satisfactory
results and it is felt that these are powerful tools
for analysis and modeling of TIG welding process.
Vidyut Dey,Dilip Kumar Pratihar, and G.L.Datta
[10] could find back – propagation neural network
(BPNN) to show better performance than genetic-
neural (GA-NN) in predicting the bead profiles in
Electron beam bead on plates welding. Y.S.Tarng,
J.L.Wu, S.S. Yeh, and S.C. Juang [11] described
application of Neural Network & Simulated
annealing (SA) algorithm to model & optimize the
GTAW process of pure 1.6mm aluminum 1100
sheet. As CPN is equipped with good learning
ability, CPN is selected to model the process & SA
applied to search for welding process parameter
with optimal weld pool features Ghosh and Sarkar
[12] have proposed a neural network model to
predict the yield characteristics of submerged arc
weldments. R.J.Praga-Alejo,L.M.Torres-Trevino,
and M.R.Pina-Monarrez [13] found the
performance of neural network plus GA algorithm
a little better than response surface methodology
with canonical analysis. R.P.Singh, R.C.Gupta, and
S.C.Sarkar [14] used artificial neural network
technique to predict the tensile strength of weld for
the given welding parameters. I.U. Abhulimen &
J.I. Achebo[15] used artificial neural network in the
prediction and optimization of the Tungsten inert
gas weld of mild steel pipes. Neural network model
was generated using the Levenberg-Marquardt
algorithm with feed ward back propagation
learning rule. Results show that the generated
neural network model was able to predict tensile
and yield strength to a mean square error of 34.2.
K. Anand [16] utilized neural networks for
predicting the friction welding process parameters
to weld Incoloy 800H.Lin & Chou [17] adopted
neural network with a Levenberg - Marquardt
back-propagation (LMBP) algorithm was then
adopted to develop the relationships between the
welding process parameters and the tensile-shear
strength of each weldment
The literature survey revealed absence of a
comprehensive and efficient neural network
modeling process with limited application in
reverse modeling and comparison of various

learning, training algorithms and transfer functions
was found not adopted for better model
development. The application of Elman back
propagation neural network was rare in weld
process modeling as per literature survey. The
present research work provides a comprehensive
method for developing two different backward
propagation neural networks, feed forward back
propagation and Elman back propagation, both in
forward & reverse modeling and the comparison of
their performance in reverse modeling along with
the comparison of their performance in forward
mapping with stepwise regression modeling
performance by taking into account of upper said
findings in the development of neural network
models.
III. EXPERIMENTAL PROCEDURE
Response surface methodology (RSM)
based design of experiment, Face centered central
composite design with 53 total runs comprising of
32 half factorial points plus 9 center points and 12
star points was employed for conducting the
experiments. The experiment was conducted as per
the above design matrix available in MINITAB 16
at random to take care of systematic errors
infiltration. The experimentation is carried out in
the fabrication shop of Siddhivinayak fabricators,
Bengaluru. Single pass autogenous bead on plate
welding procedure is followed to simulate tight butt
joint welding on 5 mm thick AA 5083; H111
aluminum alloy with current, % balance, welding
speed arc gap, shielding gas flow rate and
frequency as process parameters and front
width(FW), back width(BW), front height(FH),
back height(BH) as weld bead characteristics
(responses). The welding of plate is carried out
normal to the rolled direction. A gas mixture of
75% Helium and 25% Argon is used as shielding
gas along with 0.2% Zirconiated tungsten rod of
3.2 mm diameter as electrode. To simulate the
actual robot welding operation the welding torch
was made to move along the precise aluminum
railings pertaining to ESAB automatic gas cutter
.The experiment was conducted by varying the
current in the range of 145-185 ampere , the %
balance in the range of 32-68 % , the gap in the
range of 1.5-2 mm , welding speed in the range of
230-330 mm/min, the gas flow rate in the range of
15-20 L/min and frequency in the range of 30-110
Hz. The experimental set up is shown in Figure 1.
Then the weld bead geometry quality
characteristics such as, were measured in
millimeters using Project profilometer after
preparing specimen following the standard
metallographic procedures. Four replicates are
taken for each run totaling 212 input data.
Fig. 1: Experimental set up of TIG welding process
Table 1: Design matrix with input parameters at actual values [23]
RUNS Current % balance gap speed Flow. Rate Frequency
DP41 165 50 1.75 280 15 70
DP13 145 32 2 330 15 30
DP22 185 32 2 230 20 110
DP43 165 50 1.75 280 17.5 30
DP50 165 50 1.75 280 17.5 70
DP37 165 50 1.5 280 17.5 70
DP8 185 68 2 230 15 110
DP44 165 50 1.75 280 17.5 110
DP51 165 50 1.75 280 17.5 70
DP35 165 32 1.75 280 17.5 70

DP6 185 32 2 230 15 30
DP23 145 68 2 230 20 110
DP25 145 32 1.5 330 20 30
DP14 185 32 2 330 15 110
DP10 185 32 1.5 330 15 30
DP9 145 32 1.5 330 15 110
DP46 165 50 1.75 280 17.5 70
DP42 165 50 1.75 280 20 70
DP40 165 50 1.75 330 17.5 70
DP19 145 68 1.5 230 20 30
DP48 165 50 1.75 280 17.5 70
DP28 185 68 1.5 330 20 30
DP33 145 50 1.75 280 17.5 70
DP24 185 68 2 230 20 30
DP31 145 68 2 330 20 30
DP36 165 68 1.75 280 17.5 70
DP18 185 32 1.5 230 20 30
DP39 165 50 1.75 230 17.5 70
DP27 145 68 1.5 330 20 110
DP11 145 68 1.5 330 15 30
DP1 145 32 1.5 230 15 30
DP15 145 68 2 330 15 110
DP21 145 32 2 230 20 30
DP7 145 68 2 230 15 30
DP12 185 68 1.5 330 15 110
DP30 185 32 2 330 20 30
DP52 165 50 1.75 280 17.5 70
DP3 145 68 1.5 230 15 110
DP26 185 32 1.5 330 20 110
DP2 185 32 1.5 230 15 110
DP45 165 50 1.75 280 17.5 70
DP49 165 50 1.75 280 17.5 70
DP20 185 68 1.5 230 20 110
DP47 165 50 1.75 280 17.5 70
DP17 145 32 1.5 230 20 110
DP53 165 50 1.75 280 17.5 70
DP34 185 50 1.75 280 17.5 70
DP4 185 68 1.5 230 15 30
DP38 165 50 2 280 17.5 70
DP29 145 32 2 330 20 110
DP16 185 68 2 330 15 30
DP32 185 68 2 330 20 110
DP5 145 32 2 230 15 110
IV. PROCESS MODELING USING
STATISTICAL APPROACH [23]
Utilizing MINITAB Statistical software
version 16 the mathematical models are developed
for all the quality characteristics using stepwise
regression analysis which eliminates the
insignificant model terms automatically with
stepwise selection of terms α to enter = 0.15, α to
remove = 0.15. The method is dealt by
Douglas.C.Montogomory [18]. Considering linear,
square and 2 way interactions the following
response equations are developed for each quality
characteristics.
FW = 2.52 + 0.5874D - 0.1510A - 45.68T -
0.15951P + 3.254M - 0.05035R - 0.002193D2
+ 0.001440A2
+ 13.146T2
+ 0.000028 P2
–
0.07074M2
- 0.000099R2
+ 0.000150DA
+ 0.02667DT + 0.000447DP - 0.001877DM -
0.01595AT - 0.000032AP + 0.001905AM -
0.000223AR + 0.018494TP - 0.4564TM -
0.012805TR + 0.000684PM + 0.000280PR
+ 0.000854MR
FH = - 2.474 - 0.01388D + 0.03874A - 5.761T
+ 0.04493P + 0.1652M - 0.03681R – 0.000284A2
+ 2.128T2
- 0.000102P2
+ 0.000071R2
-
0.000047DA - 0.003188DT + 0.000101DP -
0.000584DM + 0.000024DR - 0.001215AT -
0.000424AM + 0.000087AR - 0.002625TP -
0.01325TM - 0.001062 TR - 0.000161PM
+ 0.000064PR + 0.000095MR
BW = - 9.90 + 0.1653D - 0.0272A - 10.78T -
0.04756P + 2.975M - 0.10527R - 0.000710D2
+ 0.000578A2
+ 1.517T2
- 0.000054P2
-
0.09523M2
+ 0.000278R2
- 0.000246DA
+ 0.01608DT + 0.000178DP + 0.000107 DR -
0.000084AP + 0.001977AM + 0.000067AR
+ 0.012856TP - 0.00432TR + 0.000946PM
+ 0.000123PR + 0.000855MR
BH = 11.38 - 0.1005D - 0.06271A + 5.32T -
0.05544P + 0.310M + 0.02810R + 0.000307D2
+ 0.000445A2
- 2.615 T2
+ 0.000046P2
-

0.02395M2
- 0.000108R2
- 0.000183DA
+ 0.000018DP + 0.000792DM - 0.000073DR
+ 0.000091AP + 0.001536AM - 0.000093AR +
0.007494TP + 0.08512TM + 0.000852TR
+ 0.000504PM - 0.000005PR + 0.000220MR
V. MODELING OF TIG WELDING
PROCESS USING ARTIFICIAL
NEURAL NETWORKS
Figure2: Forward and reverse TIG welding process modeling.
Artificial neural network (ANN) is one of
the computational based artificial intelligence
techniques (AI) that copies the behavior of human
brain. It has the self learning capability, adaptation
characteristic and is basically non linear in nature.
Hence ANN is applied as an excellent tool for
handling complex non linear engineering processes
by R.J.Praga-Alejo, L.M.Torres-Trevino, and
M.R.Pina-Monarrez [13] Development of ANN
consists of 5 basic steps 1) collecting data 2)
preprocessing data 3) creating the network 4)
training the network 5) simulate the network
response to new inputs. As stated by Rakesh
Malviya,and Dilip Kumar Pratihar [19], in order to
automate any process, knowing input-output
relation ships both in reverse and forward direction
is required, on line.
5.1 Forward modeling- Feed forward back
propagation neural network (FFBPNN)
The most common and successful ANN
architecture with feed forward network topology is
multilayer perceptron (MLP).Multilayer back
propagation algorithm (FFBP) is the most common
supervised learning technique used for training
ANNs. FFBP network minimizes the errors
between obtained outputs and desired target values
by feeding back the derivatives of network error
with respect to networks and adjusting the weights
so that the error decreases with each iteration and
the ANN model gets closer and closer to the
desired target values. 212 input and output data
obtained as a result of the real experimentation is
utilized for training and testing of the neural
networks. For training both the back propagation
Neural networks in forward and reverse
modeling, the training parameters utilized are goal-
0,min_grad--1 x 10-7
,max_fail-6,mu-
0.001,mu_dec-0.1,mu_inc-10, and mu-max—1 x
1010
and 1000 epochs. Neural network tool box of
MATLAB R2013a was utilized for the whole
modeling process.
In this research work, preprocessing was
done to scale the inputs and targets to fall within a
specified range (-1 to +1 ) by using minmax
technique so that accuracy of subsequent numeric
computation enhances by avoiding effect of high
valued variables on lower magnitude variable
during training. Creating the network means
finalizing the number of layers and the number of
neurons in hidden layer. S.C Juang et al.[3] have
found that many researchers have confirmed
experimentally that single hidden layer is sufficient
to provide better convergence in the modeling of
TIG welding process, in the present research work
a three layer feed forward neural network
architecture consisting of six input, four output
neurons and a hidden layer is utilized. During the
network construction, the data set was divided into
training, validation and test data in proportion of
0.7, 0.15 and 0.15respectively. For gradient
computation and weights & bias updating, training
set was used. For improving generalization,
validation set and for validating the network
performance, test set was utilized. The selection of
data in each set is done randomly and then the
network was created .As the finalization of number
of neurons in the hidden layer is crucial for
efficient modeling as found by Ill-Soo Kim et al
[20] six data selected randomly, are utilized for

calculating simulating error of each architecture
during the finalization of neurons in hidden layer
and remaining 6 data are used in calculating
absolute % of prediction error (APPE) for the
model finalized through comprehensive method.
Out of 212 data, 200 data has been utilized for the
modeling. During finalization of architecture, the
conducted parametric analysis is shown in Table 1.
During the parametric analysis, trainlm, learngdm
and mse are used as training; adaptation learning
and performance function respectively. Tan
sigmoid was used as transfer function for hidden
and output layers. Five trainings were conducted
and the architecture with minimum mean square
error (MSE) with minimum percentage of
simulation error was chosen to avoid over fitting.
The analysis of the table 2 yielded the final
architecture for forward mapping as 6-20-4. This
architecture was again trained using 12 different
back propagation training algorithms to select an
efficient algorithm for training the ANN. Then
afterwards, various transfer functions and
adaptation learning algorithms are considered for
creating the network. The neural model with best
prediction capacity having Pearson Correlation
Coefficients above 0.98 for the training, validation
and test sets was taken as performance criteria. The
analysis is shown in table 3, 4 and 5 respectively.
Table2: Parametric analysis for architecture finalization
Analysis of above table yielded a neural network as
shown in Figure 3. It is a feed forward back
propagation network of 6-20-4 neuron
configuration with a Lavenberg-Marquardt training
algorithm (trainlm), gradient descent BP with
momentum as adoptive learning algorithm
(traingdm) and tan-sigmoid (tan-sig) as transfer
function for both hidden & output layers. The
Levenberg-Marquard approximation algorithm was
found to be the best fit for application. Similar
results found in literature by P.Sreeraj, T.kannan,
and S.Maji [21]
Architecture MSE
%Simulating
error
Average
absolute
error
Architecture MSE
%Simulating
error
Average
absolute
error
6--5—4
0.0359 2.8489
7.34745 6--20--4
0.00187 0.4841
0.6655075
0.0396 17.786 0.0017 1.16
0.0371 -7.5426 0.00177 -0.1033
0.0393 -1.2123 0.00188 -0.91463
0.0378 0.00168
6--10—4
0.00947 -0.4169
2.07601 6--25--4
0.0019 0.49815
0.80853
0.0109 -4.365 0.00176 1.72026
0.011 -0.2814 0.0019 -0.18139
0.00845 3.2407 0.00203 -0.83432
0.00823 0.00184
6--15--4
0.00203 -1.2439
1.6248
0.00226 4.564
0.00215 0.02723
0.00239 0.66417
0.00222

Figure 3: Configuration of back propagation neural network for forward modeling of TIG welding
5.2 Reverse modeling- Feed forward back
propagation neural network (FFBPNN)
Following the similar procedure as that of forward
modeling, reverse modeling is performed. The
finalized architecture is a feed forward back
propagation network of 4-20-6 neuron
configuration with a Lavenberg-Marquardt training
algorithm (trainlm), gradient descent BP with
momentum as adoptive learning algorithm
(traingdm) and tan-sigmoid (tan-sig) as transfer
function for both hidden & output layers.
6 ELMAN’S BACK PROPAGTION
NEURAL NETWORK (EBPNN)
Jeffrey Elman proposed this network. It is a neural
network with semi recursive character which
recognizes patterns from a sequence of values by
back propagation through time learning algorithm.
It is basically a recurrent neural network that
enables sequential learning and identification of
patterns in series of values or events that unfold
over time and can be predicted. Elman’s neural
network consists of a recurrent first layer opposed
to a conventional two layer network. Values from a
previous step can be stored as context and used in
the current time step. The stored information can be
used in future and this enables temporal and spatial
pattern learning. Following the similar procedure as
that of earlier back propagation neural network
generation, Elman’s back propagation neural
networks are generated both for forward and
reverse modeling. During network finalization with
respect to various training & adaptation learning
algorithms and transfer functions, mean square
error (MSE) is used as the performance criteria.
6.1 Forward modeling (EBP)
Finalized network is Elman back propagation
network of 6-25-4 neuron configuration with a
Lavenberg-Marquardt training algorithm (trainlm),
gradient descent BP with momentum as adoptive
learning algorithm (traingdm) and log-sigmoid
(log-sig) as transfer function for hidden & pure lin
for output layer as shown in Figure 4.
6.2 Reversed modeling (EBP)
The analysis yielded the final architecture for
reverse mapping as 4-25-6. This architecture was
again trained using different learning algorithms,
transfer functions and training algorithms Finalized
network is Elman back propagation network of 4-
25-6 neuron configuration with a Lavenberg-
Marquardt training algorithm (trainlm), gradient
descent BP with momentum as adoptive learning
algorithm (traingdm) and log-sigmoid ( log-sig) as
transfer function for hidden & pure lin for output
layer.

Figure 4: Configuration of Elman back propagation neural network for forward modeling of TIG welding
7 RESULTS AND DISCUSIONS
Forward and reverse modeling of TIG welding of
aluminum alloy AA5083: H111 have been carried
out using the artificial neural networks developed
through FFBP and EBP. Results of the modeling
are stated and discussed below.
7.1 Results of forward modeling
Once the networks are trained with the finalized
parameters, algorithms and transfer functions, the
networks are simulated with six unused data,
network outputs are noted and the predicted vs. the
actual plots for all the four quality characteristics
FW, BW, FH, and BH are drawn respectively. The
entire predicted vs. actual plots give an indication
that the models developed are adequate as points
are scattered randomly and closure to the 45 degree
line as shown in Figure 6
9 1 0 1 1 1 2 1 3 1 4
9
1 0
1 1
1 2
1 3
1 4
Predictedvalues
A c tu a l v a lu e s
-2 .0 -1 .8 -1 .6 -1 .4 -1 .2 -1 .0
-2 .0
-1 .8
-1 .6
-1 .4
-1 .2
-1 .0
Predictedvalues
A ctu a l va lu e s
a) b)

8 .0 8 .5 9 .0 9 .5 1 0 .0 1 0 .5 1 1 .0
8 .5
9 .0
9 .5
1 0 .0
1 0 .5
1 1 .0
Predictedvalues
1 .2 1 .4 1 .6 1 .8 2 .0 2 .2 2 .4 2 .6 2 .8 3 .0
1 .2
1 .4
1 .6
1 .8
2 .0
2 .2
2 .4
2 .6
2 .8
Predictedvalues
c) d)
Fig 5 : predicted vs. the actual values plots for all the four quality characteristics a) FW b)FH c)BW and d) BH
respectively (BPNN-forward modeling)
Then the absolute percentage of prediction error
(APPE) is calculated for the simulated results and
compared with corresponding RSM statistical
model results and the comparison is shown in table
6. The analysis of the table reveals that the both
back propagation neural networks predict the
results accurately. The percentage of prediction
errors found in FFBP is smaller than that of EBP.
And the performance of both the neural networks
found to be better than that of statistical method in
three quality characteristics except front height.
Experimental errors and measurement errors could
be the reason in case of front height.
Table 6: Comparison of APPE for FFBP and EBP forward modeling of welding process
FW FH
Experimental RSM EBPP FFBP Experimental RSM EBPP FFBP
11.12 11.035 11.043 11.0618 -1.31 -1.316 -1.346 -1.35
11.42 11.519 11.52 11.5282 -1.38 -1.351 -1.34
-
1.354
11.28 11.274 11.24 11.2348 -1.64 -1.611 -1.611
-
1.594
11.22 11.173 11.2 11.1985 -1.05 -1.048 -1.06
-
1.052
11.58 11.519 11.52 11.5282 -1.36 -1.3509 -1.34
-
1.354
11.91 11.982 11.89 11.8943 -1.68 -1.6749 -1.656
-
1.651
APPE 0.54% 0.46% 0.44% APPE 1.40625 1.8728 1.7
BW BH
Experimental RSM EBPP FFBP Experimental RSM EBPP FFBP
9.26 9.217 9.246 9.2374 1.96 1.9713 1.924 1.917
9.92 9.977 9.991 9.9522 2.08 2.083 2.102 2.081
9.27 9.292 9.26 9.2456 2.21 2.1793 2.229 2.255
9.01 9.106 9.015 9.0158 1.7 1.75 1.73 1.729
9.96 9.977 9.991 9.9522 2.08 2.15 2.102 2.081
10.13 10.09 10.15 10.181 2.24 2.2368 2.282 2.242
APPE 0.48 0.253 0.2466 APPE 1.4266 1.405 1.013

Fig 6 predicted vs. the actual values plots for all the four quality characteristics a) FW b) FH c) BW and d) BH
respectively (EBPNN-forward modeling)
7.2 Reverse modeling results
On the same lines as that of forward modeling,
following results are obtained for reverse modeling
and are depicted in Figure 7 and Figure 8
respectively for FFBPNN and EBPNN
Table 7: Comparison of APPE for FFBP and EBP reverse modeling of welding process.
Current % Balance Gap
actual BPNN EBP actual BPNN EBP actual BPNN EBP
165 183.29 184.96 50 45.855 62.4542 1.75 1.7542559 1.6284
165 162.97 158.8 50 52.5567 44.049 1.75 1.7791983 1.7357
185 184.91 179.954 68 56.2772 51.2951 1.5 1.7613236 1.6059
185 185 184.998 68 67.8162 67.9879 1.5 1.622845 1.9578
145 180.98 145.212 68 66.5043 52.4638 1.5 1.7131567 1.6093
165 163.56 164.8 50 52.0795 56.6171 1.75 1.7818084 1.7612
APPE 6.3417 3.141769 APPE 6.212 16.2459 APPE 7.2585 8.8787
Speed Gas flow rate Frequency
actual BPNN EBP actual BPNN EBP actual BPNN EBP
280 305.03 249.988 15 15.3171 16.5187 70 74.909148 38.2077
280 271.23 275.753 17.5 17.6578 18.1394 70 67.746647 70.945
330 330 257.977 20 16.3975 16.6112 30 31.483593 77.6979
330 330 329.703 15 15.0151 15.3297 110 107.78938 109.098
230 230 235.849 15 18.3615 18.9418 110 109.27331 108.44
280 272.86 276.815 17.5 17.4879 18.0923 70 65.085587 69.4436
APPE 2.437 6.305 APPE 7.268 10.43 APPE 4.1447 34.798

Analysis of the table 7 reveals that absolute
percentage error of prediction of both networks
found accurate enough except frequency prediction,
which was slightly out of limit. BPNN found better
than EBP in prediction of five characteristics
except current. And also comparisons of forward
and reverse modeling found that the forward
modeling accuracy was far better than that of the
reverse modeling.
8 CONFIRMATION EXPERIMENTS
RUNS (FORWARD MODELING)
Using composite desirability approach, optimal
parameter setting obtained was current 145 A,%
balance 37.0909,gap of 1.6667 mm, welding speed
of 330 mm/min,20 L/min gas flow rate and
38.0808 Hz frequency. And the optimal setting
yielded the experimental results which were
compared with that predicted by FFBP and EBP
forward modeling approaches and the results are
shown in table 8 with absolute percentage
prediction error(APPE) as evaluation criteria.
Accuracy of prediction found better in FFBP
network in most cases than RSM and EBP
approaches.
Table 8: Comparison of APPE for FFBP and EBP forward modeling at optimum condition
140 150 160 170 180 190
140
150
160
170
180
190
Predictedvalues
A ctual values
30 35 40 45 50 55 60 65 70
30
35
40
45
50
55
60
65
70
Predictedvalues
A ctual values
1 .5 1 .6 1 .7 1 .8 1 .9 2 .0
1 .5
1 .6
1 .7
1 .8
1 .9
2 .0
Predictedvalues
A ctu a l va lu e s
a) b) c)
2 2 0 2 4 0 2 6 0 2 8 0 3 0 0 3 2 0 3 4 0
2 2 0
2 4 0
2 6 0
2 8 0
3 0 0
3 2 0
3 4 0
Predictedvalues
1 5 1 6 1 7 1 8 1 9 2 0
1 5
1 6
1 7
1 8
1 9
2 0
Predictedvalues
A c tu a l v a lu e s 2 0 4 0 6 0 8 0 1 0 0 1 2 0
2 0
4 0
6 0
8 0
1 0 0
1 2 0
Predictedvalues
A ctu a l va lu e s
d) e) f)
Fig 7: predicted vs. the actual values plots for all the six quality characteristics a) current b) % balance c) gap
d) speed e) gas flow rate f) frequency respectively (FFBPNN-reverse modeling)
FW FH
Expmtl. RSM FFBP EBP Expmtl RSM FFBP EBP
9.93 9.21 10.0261 9.7291 -1.48 -1.6 -1.5616 -1.5743
APPE 7.81% 0.95% 2.06% APPE 7.50% 5.23% 5.99%
BW BH
Expmtl RSM FFBP EBP Expmtl RSM FFBP EBP
8.22 8.83 8.514 8.6243 1.58 1.68 1.4655 1.5695
APPE 6.90% 3.40% 4.68% APPE 5.95% 7.80% 0.67%

Fig 8: predicted vs. the actual values plots for all the six quality characteristics a) current b) % balance c) gap d)
speed e) gas flow rate f) frequency respectively (EBPNN-reverse modeling)
9 CONCLUSIONS
The present work proposes two artificial
intelligence techniques, feed forward back
propagation and Elman’s back propagation
artificial neural network as effective methods of
conducting both forward and reverse modeling of
TIG welding process of aluminum alloy
AA5083:H111 to enable the automation of the
process. The prediction results found in this work
are in good agreement with the actual
measurements with low absolute percentage of

error performance index. The good results indicate
that both the artificial neural networks are capable
of accurately modeling weld bead geometry. The
construction, training and simulating process of
theses ANN models was very complicated so as the
architecture finalization. A comprehensive way
adopted in this work was to use some trail and error
method and thoroughly understand the theory of
back propagation for designing the neural networks
efficiently to generate accurate predicting results.
Both the approaches found to have more adoptive
nature than the statistical approach which may be
due to their ability to carry out interpolation within
the parameter ranges. Both the neural network
models found to possess better predictive ability
than the step wise regression analysis based
statistical approach. These two ANNs were found
to be viable methods of predicting the parameters
in both forward and reverse modeling as their
accuracy has been tested by the comparison of the
simulated results with that of the real experimental
data of TIG welding process. Modeling by BPNN
found to be more accurate in more cases in both
reverse and forward modeling than EBP.
Confirmation test during forward modeling
emphasizes this superiority. Prediction accuracy in
forward modeling found to be more than reverse
modeling in both neural networks. Similar results
are found by many authors in literature like Billy
Chan ,Jack Pacey, and Malcolm Bibby [22]
REFERENCES
[1]. T.G Lim, and H.S cho, “Estimation of weld
pool sizes in GMA welding process using
neural Networking” Journal of system &
control Engineering Proc Instn. Mech Engrs
Korea vol 207- IMechE 1993.
[2]. George.E.Cook, Robert Joel Barnett,
Kristinn Andersen, and Alvin M Strauss,
“Weld modeling and control using artificial
Neural Networking,” IEEE Transactions on
industry applications, vol31, No.6 Novklee-
1995
[3]. S.C. Juang, Y.S.Tarng, and H.R. Lii, “A
comparison between the back-propagation
and counter-propagation networks in the
modeling of the TIG welding process,”
Journal of Materials processing Technology
75(1998)54-62
[4]. I.S. Kim, J.S. Son, C.E. Park, C.W. Lee,
Yarlagadda, and K.D.V.Prasad, “A study on
prediction of bead height in robotic arc
welding using a neural geometry,” Journal
of materials processing Technology 130-
131(2002) 229-234
[5]. D.S. Nagesha, and G.L. Datta, “Genetic
algorithm for optimization of welding
variables for height to width ratio and
application of ANN for prediction of bead
geometry for TIG welding process,” Applied
Soft Computing 10 (2010) 897–907
[6]. Parikshit Dutta, and Dilip Kumar Pratihar,
“Modeling of TIG welding process using
conventional regression analysis and neural
network-based approaches,” Journal of
Materials Processing Technology 184
(2007) 56–68
[7]. K.Manikya Kanti, and P.Shrinivasa Rao,
“Prediction of bead geometry in pulsed
GMA using back propagation neural
networks,”Journal of Materials Processing
Technology 200 (2008) 300-305
[8]. M V Amarnath, and D K Pratihar, “Forward
and reverse mappings of the tungsten inert
gas welding process using radial basis
function neural networks,” Proceedings of
the Institution of Mechanical Engineers,
Part B: Journal of Engineering Manufacture
2009 223: 1575
[9]. D.S. Nagesha, and G.L. Datta, “Genetic
algorithm for optimization of welding
variables for height to width ratio and
application of ANN for prediction of bead
geometry for TIG welding process,” Applied
Soft Computing 10 (2010) 897–907
[10]. Vidyut Dey,Dilip Kumar Pratihar, and
G.L.Datta, “Prediction of weld bead profile
using neural network,”
[11]. Y.S.Tarng, J.L.Wu, S.S. Yeh, and S.C.
Juang, “Intelligent modeling & optimization
of the gas tungsten arc Welding process,”
Journal of intelligent manufacturing (1999)
10, 73-79
[12]. A.Ghosh, S.Chattopadhyay, and P.K.Sarkar,
“Effects of input parameters on weld bead
geometry of SAW process,” ICME2007,
Dhaka, Bangladesh
[13]. R.J.Praga-Alejo,L.M.Torres-Trevino, and
M.R.Pina-Monarrez,“Optimization of
welding process parameters through
response surface, neural network and
genetic algorithms”
[14]. R.P.Singh, R.C.Gupta, and
S.C.Sarkar,“Application of artificial neural
network to analyze and predict the tensile
strength of shielded metal arc welded joints
under the influence of external magnetic
field,” Int J. of engineering and science
ISBN: 23189-6483,ISSN: 2278-4721,Vol. 2
Issue 1 Jan.2013,pp 53-57
[15]. I.U. Abhulimen, and J.I. Achebo,
“Application Of Artificial Neural Network

In Predicting The Weld Quality Of A
Tungsten Inert Gas Welded Mild Steel Pipe
Joint,” International journal of scientific &
Technology Research Vol.3 Issue 1, January
2014 ISSN 2277-8616
[16]. K. Anand, Birendra Kumar Barik , K.
Tamilmannan, and P. Sathiya , “Artificial
neural network modeling studies to predict
the friction welding process parameters of
Incoloy 800H joints,” Engineering Science
and Technology, an International Journal 18
(2015) 394e407
[17]. Hsuan-Liang Lin, and Chang-Pin Chou,
“Optimization of the GTA welding process
using combination of the Taguchi method
and a neural-genetic approach,” Materials
and manufacturing processes, 25:631-636,
2010
[18]. Douglas.C.Montogomory,“Design and
analysis of experiments,” Seventh edition
published by John Wiley& Sons, INC,UK.
[19]. Rakesh Malviya,and Dilip Kumar
Pratihar,“Tuning of neural networks using
particle swarm optimization to model MIG
welding process,” Swarm and Evolutionary
Computation 1 (2011) 223–235
[20]. Ill-Soo Kim, Joon-Sik Son ,Sang-Heon
Lee,Prasad K.D, and V.Yaralagadda ,
“Optimal design of neural networks for
control in robotic arc welding,” Robotics
and computer integrated manufacturing
20(2004) 57-63
[21]. P.Sreeraj, T.kannan, and S.Maji, “Genetic
algorithm for the optimization of welding
variables for percentage of dilution &
application of ANN for prediction of weld
bead geometry in GMAW process,”
Mechanical confab vol.2 No.1 Jan 2013
[22]. Billy Chan ,Jack Pacey, and Malcolm Bibby,
“Modeling gas metal arc weld geometry
using artificial neural network Technology,”
Canadian Metallurgical Quarterly , Vol 38
No 1, pp. 43-51
[23]. Suneel.R.Joshi, and Dr.J.P.Ganjigatti,
“Simultaneous optimization of multiple
quality characteristics in TIG welding of
AA5083; H111 Aluminum Alloy using
Response Surface Methodology coupled with
composite desirability function”
International Journal of Applied Engineering
Research ISSN 0973-4562 Volume 11,
Number 9 (2016) pp 6525-6541

Comprehensive Development and Comparison of two Feed Forward Back Propagation Neural Networks for Forward and Reverse Modeling of Aluminum Alloy AA5083; H111 TIG Welding Process

More Related Content

What's hot (17)

Viewers also liked (20)

Similar to Comprehensive Development and Comparison of two Feed Forward Back Propagation Neural Networks for Forward and Reverse Modeling of Aluminum Alloy AA5083; H111 TIG Welding Process (20)

Recently uploaded (20)

Comprehensive Development and Comparison of two Feed Forward Back Propagation Neural Networks for Forward and Reverse Modeling of Aluminum Alloy AA5083; H111 TIG Welding Process