Submerged Arc Welding Parameters

International Journal of Mechanical Engineering and Technology (IJMET), ISSN 0976 6340(Print),
ISSN 0976 6359(Online) Volume 1, Number 1, July - Aug (2010), IAEME

95

SUBMERGED ARC WELDING PARAMETERS
ESTIMATION THROUGH GRAPHICAL TECHNIQUE
Aniruddha Ghosh
Dept. of Mechanical Engineering
Govt. College of Engg & Textile Technology, Berhampore

Somnath Chattopadhyaya
Dept of ME&MME
ISM, Dhanbad, India

ABSTRACT:
In submerged arc welding (SAW), selecting appropriate values for process
variables is essential in order to control bead size and quality. Also, condition must be
selected that will ensure a predictable weld bead, which is critical for obtaining high
quality. In this investigation, mathematical models (based on multi regression method)
have been developed and side by side Prediction through artificial neural networks has
been made. A comparative study also has been done. Based on multi regressions and a
neural network, the mathematical models have been derived from extensive experiments
with different welding parameters and complex geometrical features. Graphic displays
also represent the resulting solution on bead geometry that can be employed to further
probe the model. The developed system enables to input the desired weld dimensions and
select the optimal welding parameters. The experimental results have been proved the
capability of the developed system to select the welding parameters in SAW process
according to complex external and internal geometry features of the substrate.
Keywords: submerged arc welding, multi regression method, artificial neural networks,
Graphic displays.
Article outline:
Introduction
Experimental procedure
International Journal of Mechanical Engineering
and Technology (IJMET), ISSN 0976 6340(Print),
ISSN 0976 6359(Online) Volume 1,
Number 1, July - Aug (2010), pp. 95-108
IAEME, http://www.iaeme.com/ijmet.html
IJMET
I A E M E

96
Model development, Results and Discussion
1. multi regression method
2. artificial neural networks
3. Graphical representation
Conclusion
INTRODUCTION:
In the early days, arc welding was carried out manually so that the weld quality
can be totally controlled by the welder ability. The welder when welding can directly
monitor flow pattern in puddle and make immediate adjustments in welding parameters
to obtain a good weldability. To consistently produce high quality of weldability, arc
welding requires welding personnel with significant experience. One reason for this is
need of properly select welding parameters for a given task in order to get good weld
quality which identified by its microstructure and the amount of spatter, and relied on the
correct bead geometry size. Therefore, the use of the control system in arc welding can
eliminate much of the guess work often employed by welders to specify welding
parameters for a given task (Ref.1).in addition of specific importance is the development
mathematical models that can be employed to predict welding parameters about arc
welding parameters about arc welding process with respect to the work piece and bead
geometry to develop a robotic welding system. The submerged arc welding is one of the
major fabrication processes in industry because of its inherent advantages, including deep
penetration and a smooth bead (Refs.2, 3).Critical set of input variables are involved in
this process which are needed to control. For this reason , in the application of SAW,
engineers often face the problem of selecting appropriate and optimum combinations of
input process-control variables for achieving the required weld bead quality or predicting
the weld bead quality for proposed process-control-variable values (Ref.4) .For automatic
SAW, the control parameters must be fed to the system according to the some
mathematical formula it may be come from multi regression method or artificial neural
networks or any other suitable and efficient method to achieve the desired results
(ref.5).These important problem can be solved with the development of mathematical
models through effective and strategic planning, design and execution of experiments.
Already many efforts have been carried out development of various algorithms in the

97
modeling of arc welding process using various technologies (ref.6, 7, 8 and 9).Multiple
regression techniques were used to establish the empirical models for various arc welding
processes(Ref.6,7).However ,the regression techniques cannot describe adequately the
most of the arc welding process as a whole. One of the artificial intelligent (AI)
techniques, a neural network as a tool for incorporating knowledge in a manufacturing
system is massively interconnected networks of simple elements and their hierarchical
organizations. These processes are characterized by welding parameters due to lack of
adequate mathematical models to relate these parameters with bead geometry (Ref. 8,
9).While numerical techniques such as finite element method (FEM) also have several
limitations. The potentially viable processing routes are numerous and, therefore, various
intelligent systems are necessary to identify optimal processing parameters (Ref.10).Now,
it is possible to make this selection with the help of a computer, and complex simulations
become an effective memory for choosing the welding parameters. Also, arc welding
requires a steady hand to keep the electrode at a constant distance from the part being
welded. At the same time that the hand has to move at constant speed, it has to adjust for
distance, as the electrode shortens. This operation requires hundreds of hours of practice,
burning expensive electrodes. There are many systems that simulate a welding machine
and permit significant saving in consumables. The selection of welding parameters for a
given welding process is based on experimental methods and human qualifications
according to fabrication standards and empirical rules (ref.11).This papers represents the
development of intelligent system to obtain detailed information about the bead geometry
in relation to the welding conditions, and to provide the welding engineer with sufficient
information to design the most economic and reliable welded components for a given set
of fabrication conditions.
Experimental Procedure: Process in Action

Figure 1 MEMCO Semi Automatic Submerged Arc Welding machine at the workshop of the
Indian School of Mines, Dhanbad, India.

98
Specifications:
Input voltage supply- 380/440 volts Welding speed Trolley-30 to 1200 mm/min
3 Phase,50/60 Hz cycle, Air cooled Wire feed speed-100 to 8000mm/min
Output current 600 amps Wire diameter -2 to 5 mm
Duty cycle 100% Head movement vertical/horizontal -135
mm
Open circuit voltage 56volts, 35Kva Deposition rate- 4 to 6 kg/hr
Flux hopper capacity 12.5 kg Wire flux ratio-1:1
Flux used: ADOR Auto melt Gr II AWS/SFA 5.17(Granular flux)
Electrode: ADOR 3.15 diameter copper coated wire
Test Piece: 300 x 300x20 mm square butt joint
Weld position flat
Electrode positive and perpendicular to the plate
Process flow chart:

SELECTING THE EXPERIMENTAL DESIGN
The experiments were conducted as per the design matrix at random to avoid
errors due to noise factors. The job 300x300x15 mm (3 pieces) was firmly fixed to a base
plate by means of tack welding and then the welding was carried. The welding
parameters were noted during actual welding to determine the fluctuations if any. The
slag was removed and the job as allowed to cool down. The job was cut at three sections
V-Groove
preparation on
Shaper Machine
Face Grinding on Automatic
Grinding Machine
Cutting of 20 mm Mild
Steel Plate by Gas Cutting

Welding by
Submerged Arc
Welding
Cutting of Samples by
Power Hack Saw
Grinding of the Face where
further study is to be carried out
Carbon coating on the
surface
Removal of wax and
cleaning of surface
Wax Coating of the Ground
Surface
Study of bead geometry,
estimation of dimensions of bead
geometry

99
by hacksaw cutter and the average values of the penetration, reinforcement height and
width were taken using venire caliper of least count 0.02mm

Figure 2 Bead geometry, P-Penetration, H-Reinforcement height,W-Bead Width
Table-1: Observed Values for Bead Parameters
Heat input(kj/cm)
Wirefeed
rate(cm/min)
Penetration(mm
)
Reinforcement
height(mm) Width(mm)
6.15 60 3.4 1.5 13.5
12.31 60 3.5 1.7 10.2
6.15 120 4 1.9 15.6
12.31 120 4.7 2.3 11
3.20 60 3.2 1.2 10.2
6.40 60 3.24 1.4 8.3
3.20 120 3.3 1.8 9.9
6.40 120 3.52 1.5 9.2
MODEL DEVELOPMENT, RESULTS & DISCUSSION:
Multi regression model:
The response function representing any of the weld bead dimensions can be
expressed as y=f (Q, F).The relationship selected, being second degree response surface,
is expressed as follows (Ref.12): Y=b
0
+ b
1
Q+ b
2
F+ b
3
Q
2
+ b
4
F
2
+ b
5
QF.Where Q=Heat
input, F-Wire feed rate.
Table-2: Regression coefficients of model
Sl.No. Coefficient For the case of
Penetration
For the case of
Reinforcement Height
For the case
of Width
1 b
0
6.9714 1.2218 88.1254
2 b
1
0.0601 0.0852 2.6686
3 b
2
-0.1081 -0.0111 -2.1975
4 b
3
0.0079 -0.0016 0.1816
5 b
4
0.0008 0.0001 0.0161
6 b
5
-0.0015 0.0000 -0.0708


100
Checking the adequacy of the developed model:
(a) The adequacy of the models was then tested by the analysis of variance
techniques. The calculated value of the F ratio of model developed does not exceed the
tabulated value of F ratio for a desired level of confidence (selected as 95%).
(b) Adding a variable to the model will always increase the value of coefficient of
multiple determination R
2
, regardless of whether the additional variable is statistically
significant or not. Fulfillment of above condition means model is adequate. Here models
are adequate.
Table-3: Calculation of Variants for Testing the Adequate of the Models
Bead
Parameters
SS
R
SS
E
DF
R
DF
E
D
F
T
MS
R
MS
E
F
0
R
2
Whether
model is
adequate
Penetration 84.22 6.48 6 1 7 14.03 6.48 2.1
6
0.9
2
adequate
Reinforcem
ent height
14.16 1.061
5
6 1 7 2.36 1.061
5
2.2
2
0.9
3
adequate
Width 826.83 17.9 6 1 7 137.8 17.9 7.6
9
0.9
7
adequate
Where SS
T
is total sum square,SS
R
is sum square due to model (or to regression) and sum
square due to error or residual.DF-degree of freedom.
Development of final mathematical model:

The final mathematical models developed are given below. The process control variables
are in their coded form.
Penetration, mm= 6.9714+0.0601Q-0.1081F+0.0079Q
2
+0.0008F
2
-0.0015QF -----------(1)
Reinforcement height, mm=1.2218+0.0852Q-0.0111F-0.0016Q
2
+0.0001F
2
------------ (2)
Width of weld bead,
mm =88.1254+2.6686Q-2.1975F+0.1816Q
2
+0.0161F
2
-0.0708QF ---------------- (3)
These mathematical models furnished above can be used to predict bead geometry by
substituting the values of the respective process parameters.
Conducting conformity tests:
To determine the accuracy of the mathematical models developed, conformity test
runs were conducted with same experimental setup. After collecting experimental results
a comparison was made between the actual and predicted values of bead parameters, and
the results are 97% accurate.

101
Artificial neural networks: Two primary elements make up a neural network-processing
elements (called nodes or units) and interconnections. The network mimics the human
brain, which contains more than 10 billion (biological) neurons. Hence, the processing
elements in ANNs are also called artificial neurons. An ANN node model of a biological
neuron is shown in Figure 2.
W
1
Transfer function

Inputs W
2 Output

W
3
T
J
Weights

Figure 2 Artificial neural network node
In this model, the j-th processing element computes a weighted sum of its inputs and
outputs yj according to whether this weighted input sum is above or below a certain
threshold Tj:
y
j
= f( )- )-------------------------------------------------------------------------------(4)
Where the function f is called transfer function. The most commonly used transfer
function is the sigmoid (S-shape) function. A typical sigmoid function is:
f(x)= ------------------------------------------------------------------------------------------(5)
Other types of functions such as hard limit, symmetrical hard limit, linear, and hyperbolic
tangent are commonly used.
Neural Network Structures: The structure of the neural network is defined by the
interconnection architecture between the processing elements. The basic types of
structures are feed forward and recurrent nets (shown inFig.3). Others are combinations
of the two types (Multilayer feed forward network and Multilayer recurrent network).

Figure 3 the structure of neural network employed for the prediction of bead geometry.

Sum of
(x
i
w
i
)

102

Rules adopted for learning: The primary training method commonly used is Error-
Correction Learning. It is a form of supervised learning where the weights are adjusted in
proportion to the output-error vector, . The output error from the k-th node on the output
layer is defined as:
=d
k
c
k
-------------------------------------------------------------------------------(6)
Where d
k
is the desired output and ck is the calculated output, for the k-th node on the
output layer only. The total squared error on the output layer, E, is:
E= = - )
2
---------------------------------------------------------------(7)
Knowing E, we can calculate the change in the weight factor for the i-th connection to the
j-th node, Wij:
w
ijnew
- w
ijnew = ijnew
=
j
a
i
E ----------------------------------------------------------(8)
Where j is a linear proportionality constant for node j, called the learning rate (typically,
0 < j < 1), and a1 is the i-th input to node j.
Neural Network Design
With more than 40 functioning models to choose from, it is important to know which
models have had the most success and to understand their similarities and differences.
After choosing the model, you then have to decide the number of hidden layers and the
nodes for each layer. The sizes (number of nodes) of input and output layers are fixed by
the number of inputs and outputs used. The sizes of middle (hidden) layers are
determined by trial and error. It is better to choose the smallest number of neurons
possible for a given problem to allow for generalization. If there are too many neurons,
the net will tend to memorize patterns. The number of neurons may be dictated by the
number of input training examples, or facts. In other words, the number of training
examples should be greater than that of trainable weights. In an ideal world, having 10 or
more facts for each weight are required. For instance, in a 10101 architecture there are
110 (= 10 x 10 + 10 x 1) weights, so you should have about 1,100 facts (example data).
BACK PROPAGATION NETWORK:
The back propagation neural network (BPN) is the most widely used feed forward
neural network system. The term back propagation refers to the training method by which

103
the weights of the network connection are adjusted. The calculations procedure is feed
forward, from input layer through hidden layers to output layer. During training, the
calculated outputs are compared with the desired values, and then the errors are back
propagated to correct all weight factors. The whole calculation procedure (for a three-
layer BPN) is summarized as follows eqn.no.15:
1. Randomly assign values between 0 and 1 to weights Wi,j (l) for each layer, l. All
input-layer thresholds are assigned to zero, i.e. T
i,1
= 0; all hidden- and output-layer
thresholds are assigned to one, i.e., T
i,3
= 1.
2. Introduce the input Ii into the neural network, and calculate the output from the first
layer according to the equations:
x
i
= Ii + Ti,1 -------------------------------------------------------------------------------------(9)
a
i ,1
= f (x
i
) ------------------------------------------------------------------------------------(10)
Where f ( ) is the transfer function mentioned in the previous section.
3. knowing the output from the first layer, calculate outputs from the second layer, using
the equation:
a
i2
=f[ ----------------------------------------------------------------------(11)
4. Given the output from the second layer, calculate the output from the output-layer,
using the equation:
a
i3
= ---------------------------------------------------------------------- (12)
y
i
= a
i,3
---------------------------------------------------------------------------------------------(13)
Steps 1 to 4 represent the forward activation flow; that is, the given input values Ii
move forward in the network, activate the nodes, and produce the actual output values y
i

based on the initially assumed values of interconnecting weights, W
i,j
(l) and internal
threshold, T
i
,l. Obviously, the initial calculation will not produce the desired output
values (d
i
). The next few steps of the back propagation algorithm represent the backward
error flow in which the errors between the desired output d
i
and the actual output y
i
flow
backward through the network and try to find a new set of network parameters (W
i,j
(l)
and T
i
).
5. Now back propagate the error through the network, starting from the output layer and
moving backward toward the input layer. Calculate the gradient-descent term (
1,3
) using
the equations:

104
x
i3
= ---------------------------------------------------------------------- (14)
i3
= ------------------------------------------------------------------------------ (15)
6. Knowing the output-layer,
1,3
,
1,2
has been calculated the gradient-descent term for
the hidden layer (layer 2) using these equations:
x
i2
= ---------------------------------------------------------------- (16)
i2
= --------------------------------------------------------------- (17)
7.Knowing the deltas for the hidden and output layers, calculate the weight changes,
Wi,j , using the equation: Wi,j(l),
new
= + Wi,j(l)
,old
Where h is the learning rate, and a is the momentum coefficient. The momentum term is
added to speed up the training rate. The momentum coefficient, a, is restricted to 0< < 1.
8.Knowing the weight changes, update the weights as: Wi,j(l),
new
= Wi,j(l),
old
+
Wi,j(l),
new
---------------------------------------------------------------------------------(18)
One iteration has now been completed. This feed forward calculation and error
back propagation procedure is repeated until the sum of errors is less than the specified
value. This is the whole learning process for the neural network. The new weight factors
are calculated from the old weight factors from the previous training iteration by the
following general expression:
[W
i
]
new
= [W
i
]
old
+[learning rate][input term][gradient descent corrected
term][momentum coefficient][previous weight change]--------------------------------(19)
Table-4: Predicted parametric values through neural network
Heat input(kj/cm)
Wire feed
rate(cm/min)
Penetration
(mm)
Reinforcement
height(mm) Width(mm)
6.15 60 3.3 1.4 13.0
12.31 60 3.5 1.6 10.3
6.15 120 3.8 1.8 15.1
12.31 120 4.5 2.2 11.2
3.20 60 3.2 1.2 10.1
6.40 60 3.3 1.4 8.8
3.20 120 3.2 1.8 10.3
6.40 120 3.6 1.5 10.0

105

Table-5: % of error for prediction through neural network
For the case of Penetration For the case of
Reinforcement height
For the case of Width
3.66 5.94 3.70
-1.32 6.49 -0.98
4.09 7.84 3.21
4.26 4.35 -1.82
-0.10 0.00 0.98
-1.34 -1.31 -6.02
2.10 1.97 -4.03
-2.27 0.00 -8.70

Figure 5 Comparison between predicted and experimental values of output parameter
GRAPHICALLY PREDICTION:
Graphically prediction technique is a new and very simple technique; previously
it was not seen in any literature. It is more appropriate technique w.r.t regression and
neural networks model. In this technique by taking input variables with in there range at
first values of out put variables values have been found out then it has been graphically
plotted with the help of MATLAB-7. By clicking on these graphs, values of variables can
be gotten very quickly.

106

Figure 6A:Change of Penetration w.r.t
change of input variables
Figure 6B: Change of Reinforcement height
w.r.t change of input variables

Fig.6C: Change of width w.r.t change of
input variables

Figure 6 Change of output variables w.r.t change of input variables.
Value of input & output variables have some limit. Limit of output variables can
be detected through this graphically prediction technique. For any welding machine,
input parameters have some range. Beyond this range, particular machine cannot work, it
is practically true but theoretically variables values can be calculated beyond these limits.
The output parameters values beyond the input variable values range can be calculated
through multi regression method, artificial neural networks, but their have no practical
visibility. Suppose output variable value is selected whose corresponding input variables

107
values are beyond the range of input variables. So it is not applicable. In this method, the
range of possible output variables can be predicted and corresponding input variables
values can be easily predicted so there is no chance of above mentioned mistake for this
method. This graphically prediction technique gives maximum, minimum range of
possible output and input variables, side by side only just clicking on the graphs idea of
variables can be gotten quickly and accurately. It is the main advantage & difference
from other methods of this method.
CONCLUSION:
The performance of the developed system has been tasted experimental for certain
welding conditions for a particular bead dimensions. The experimental data were proved
a clear correlation between welding parameters and the weld bead dimensions, and
showed the geometrical features. The neural network model is capable of making bead
geometry prediction of the real-time quality control based on observation of bead
geometry and for on-line welding process control. In this paper a graphically prediction
technique has been described which can able to give maximum, minimum range of
possible output and input variables, side by side only just clicking on the graphs idea of
variables can be gotten quickly and accurately. It is the main advantage & difference
from other method of this method.
REFERENCES:
1) G.E. Cook, Feedback and adaptive control in automated arc welding system. Met.
Construct.139 (1990), pp. 551 556.
2) P.D. Houldcroft (1989), Submerged Arc Welding Abington Publishers, U.K.
3) Annon. (1978). Principal of Industrial Welding. The James F.Lincoin Arc
Welding Foundation, Cleveland, Ohio.
4) S.Balckman. (1981).Welded fabrication of subsea pipelines in the North Sea.
Welding and Metal Fabrication.
5) N.Murugan, R.S.Paramar, S.K.Sud.1993.Effect of submerged arc process
variables on dilution and bead geometry in single wire surfacing. Journal of
Materials Processing Technology 37:767-780.
6) J.Ravindra, R.S.Pramar, Mathematical models to predict weld bead geometry for
flux cored arc welding. Met. Construct.192 (1987), pp. 31R-35R.

108
7) L.JYang, R.S.Chandel and M.J.Bibby, The effects of process variables on the
weld deposit area of submerged arc welds.Weld.J.721 (1993), pp. 11s-18s.
8) P.Li,M.T.C. Fang and J.Lucas,Modelling of submerged arc welding bead using
self-adaptive offset neural network.J.Mater.Process Technol.71(1997),pp.228-
298.
9) L.T.Srikanthan,R.S.Chandal,Neural Network based modeling of GMA welding
process using small data sets,in:Proceedings of the fifth international conference
on control, Automation, Robotics and Vision,Singapore,1988,pp.474-478.
10) I.S.Kim, C.E.Park, Y.G.Cha, J.Jeong and J.S.Son, A study on development of a
new algorithm for predicting process variables in GMA welding processes.JSME
Int.442(2001),pp.561-566.
11) R.J.Salter,R.T.Deam,A Practical Front face Penetration Control System for TIG
Welding, Developments in Automated and Robotic Welding ,Cambridge,UK,The
Welding Institute,1987,pp.38-1-38-12.
12) A.I.Khuri and J.A.Cornel, 1996, Response Surfaces, Design and Analysis.Marcol
Dikker Inc., New York, N.Y.

Submerged Arc Welding Parameters

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Submerged Arc Welding Parameters

Uploaded by

Copyright:

Available Formats

International Journal of Mechanical Engineering and Technology (IJMET), ISSN 0976 6340(Print),

ISSN 0976 6359(Online) Volume 1, Number 1, July - Aug (2010), IAEME

You might also like