Seediscussions,stats,andauthorprofilesforthispublicationat:http://www.researchgate.net/publication/236859417

PredictionofbeadgeometryinpulsedGMAweldingusingbackpropagationneuralnetworks

ARTICLEinJOURNALOFMATERIALSPROCESSINGTECHNOLOGY·JANUARY2008

ImpactFactor:2.24

CITATIONS

6

READS

163

2AUTHORS,INCLUDING:

SrinivasaRaoPedapati

UniversitiTeknologiPETRONAS

12PUBLICATIONS29CITATIONS

SEEPROFILE

Availablefrom:SrinivasaRaoPedapati

Retrievedon:17October2015

j o u r n a l o f m a t e r i a l s p r o c e s s i n g t e c h n o l o g y 2 0 0 ( 2 0 0 8 ) 300–305

journa l homepage: www.e lsev ier .com/ locate / jmatprotec

Prediction of bead geometry in pulsed GMA welding usingback propagation neural network

K. Manikya Kanti, P. Srinivasa Rao ∗

Department of Mechanical Engineering, Gayatri Vidya Parishad College of Engineering, Visakhapatnam, Andhra Pradesh, India

a r t i c l e i n f o

Article history:

Received 12 April 2007

Received in revised form

11 September 2007

Accepted 12 September 2007

Keywords:

Artificial neural networks

a b s t r a c t

This paper presents the development of a back propagation neural network model for the

prediction of weld bead geometry in pulsed gas metal arc welding process. The model is

based on experimental data. The thickness of the plate, pulse frequency, wire feed rate, wire

feed rate/travel speed ratio, and peak current have been considered as the input parameters

and the bead penetration depth and the convexity index of the bead as output parameters to

develop the model. The developed model is then compared with experimental results and

it is found that the results obtained from neural network model are accurate in predicting

the weld bead geometry.

© 2007 Elsevier B.V. All rights reserved.

Pulsed GMA welding

Welding parameters

Bead geometry

Convexity index

cess parameters due to the complex and nonlinear nature of

Regression model

1. Introduction

The GMAW-P process has recently gained wide attention inthe welding industry, owing to its comparatively low heatinput and precise control over the thermal cycle. This isbecause, with the pulsed current gas metal arc weldingprocess, spray transfer or more precisely controlled droplettransfer is obtained at a low average current. This provides asmaller and well-controlled weld pool, which allows claddingof thin materials in all positions.

The superiority of the GMAW-P process is mainly related tothe nature of the metal deposition, which is governed by thepulse parameters, namely peak current, pulse frequency, wirefeed rate, wire feed rate/travel speed ratio, etc. Bead geometry

in the arc welding process is an important factor in deter-mining the mechanical characteristics of the weld (SrinivasaRao, 2004; Srinivasa Rao et al., 2005; Rajasekaran, 1999; Senthil

∗ Corresponding author.E-mail address: psrvizag@yahoo.com (P. Srinivasa Rao).

0924-0136/$ – see front matter © 2007 Elsevier B.V. All rights reserved.doi:10.1016/j.jmatprotec.2007.09.034

Kumar and Parmer, 1986). Bead geometry variables, such asbead width, bead height, and penetration depth, are greatlyinfluenced by welding process parameters such as wire feedrate, welding current, welding speed, plate thickness etc. Weldshape and size are represented by bead width (W), bead height(H) and depth of penetration (P) as shown in Fig. 1. Convexityindex (CI) is defined as the ratio of bead reinforcement heightto the bead width. Selection of the most suitable combina-tion of pulse parameters is very difficult owing to the complexinterdependence of the above parameters on the pulsed cur-rent (Srinivasa Rao, 2004; Srinivasa Rao et al., 2005).

However, costly and time-consuming experiments arerequired in order to determine the optimum welding pro-

the welding process. Therefore, a more efficient method isneeded to determine the optimum welding process param-eters. Technique of neural networks offers potential as

j o u r n a l o f m a t e r i a l s p r o c e s s i n g t e c h n o l o g y 2 0 0 ( 2 0 0 8 ) 300–305 301

Fig. 1 – Weld bead geometry.

atdtufawd

fmTobipltDce1

2

BiadmsridiwftpieniY

n alternative to standard computer techniques in controlechnology and has attracted a widening interest in theirevelopment and application (Rajasekaran, 1999). The advan-ages of the neural networks is that the network can bepdated continuously with new data to optimize its per-

ormance at any instance, the networks ability to handlelarge number of input variables rapidly, and the net-

orks ability to filter noisy data and interpolate incompleteata.

The back propagation network (BPN) system is one of theamily of artificial neural network techniques used to deter-

ine welding parameters for various arc welding processeshe network is a multi layer network that contains at leastne hidden layer in addition to input and output layers. Num-er of input layers and number of neurons in each hidden layer

s to be fixed, based on the application, the complexity of theroblem, and the number of inputs and outputs. Use of non-

inear log-sigmoid or tan-sigmoid transfer function enableshe network to simulate non-linearity in practical systems.ue to its numerous advantages, back propagation network ishosen for present work (Kim et al., 2001, 2002, 2005; Chant al., 1999; Lee and Um, 2000; Krose and Van der Smagt,996).

. Feed forward back propagation network

ack propagation learning is a supervised learning wheret needs to know the inputs and the desired outputs indvance. This network is well established as a method forata mapping. In this work, the welding parameters areapped to the weld dimensions through the internal repre-

entation of BPNs. Data obtained from the experiments andegression models are provided to a network at the learn-ng stage, e.g., welding parameters and weld bead geometryimensions. During network learning, the network output

s compared with the desired output and the connectoreights inside the network adjust to minimize the dif-

erence. The error is then propagated backwards throughhe network and weights are changes based on the backropagation-learning algorithm. This learning process is an

terative process and learning will stop once an acceptable

rror is achieved. When the trained network is presented withew input (beyond training), the network responds accord-

ng to the knowledge it has acquired (Hagan et al., 1996;agnanarayana, 1999; Hammerstrom, 1993; Rukmini Srikant,

Fig. 2 – A back propagation network.

2003). Fig. 2 shows the architecture of the proposed sys-tem.

The process parameters of the pulsed GMA welding processare fed as input to the input layer of the network. Each of theinputs, Ii is multiplied by weights on inter-layer connectionsof a hidden layer neuron, Wij

1 and added to bias, � to produceactivation, a

a1 = Wij1IT + � (1)

where Ii is the output of the input layer, Wij1 is the weight

structure between input and hidden layers.A log-sigmoid activation function that transforms activa-

tion of hidden layer neurons to a scaled output O1 can bewritten as

O1 = 1ea1

+ 1 (2)

Outputs from hidden layer neurons are treated as inputsto output layer neurons. Summation of product of all hiddenlayer outputs and weights between hidden and output layersadded to bias constitutes activation a2 of output layer neu-rons. Log-sigmoid transfer function is applied and output O2 iscomputed. Error is estimated as difference between actual andcomputed outputs. This procedure constitutes forward flowof back propagation phase and error computed is back prop-agated through same network to update weights. Weights areupdated using generalized delta rule as given below:

Wnew = Wold − �ETI (3)

where Wnew is weight after modification, Wold is the weightstructure before modification, � is the learning rate, usuallytaken between 0 and 1, ET is the error obtained.

Weight change is calculated for all connections. Errors forall patterns are summed and the algorithm is run till errorfalls below a specified value. Back propagation algorithmsattempts to minimize the error of mathematical system rep-resented by neural network’s weights and thus walk downhillto the optimum values for weights. Unfortunately due to

the mathematical complexity of even simplest neural net-work, there are many minima, some deeper than others.In order to reach the global minimum, the aid of heuris-tic optimization techniques is generally sought (Hagan et

n g t e c h n o l o g y 2 0 0 ( 2 0 0 8 ) 300–305

Fig. 3 – A schematic diagram of BPN used for predicting

performance goal set to 0.0002, the network was trained for

302 j o u r n a l o f m a t e r i a l s p r o c e s s i

al., 1996; Yagnanarayana, 1999; Hammerstrom, 1993; RukminiSrikant, 2003). The present work uses the concept of adop-tion learning rate to overcome local minima and speed up theprocess of attaining stable weight structure, also known asconvergence.

2.1. Variable learning rate (traingda) algorithm

With standard steepest descent, the learning rate is heldconstant throughout training. The performance of the algo-rithm is very sensitive to the proper setting of the learningrate and can be improved if we allow the learning rate tochange during the training process. An adaptive learningrate will attempt to keep the learning step size as large aspossible while keeping learning stable. First, the initial net-work output and error are calculated. At each epoch newweights and biases are calculated using the current learn-ing rate. New outputs and errors are then calculated. If thenew error exceeds the old error by more than a predefinedratio max perf inc (typically 1.04), the new weights and biasesare discarded. In addition, the learning rate is decreased (bymultiplying by lr dec). Otherwise, the new weights, etc., arekept. If the new error is less than the old error, the learn-ing rate is increased (by multiplying by lr inc). This procedureincreases the learning rate, but only to the extent that thenetwork can learn without large error increases. The gradi-ent is computed by summing the gradients calculated at eachtraining example, and the weights and biases are only updatedafter all training examples have been presented (MATLAB 6.1,2001).

3. Model development

3.1. Process parameters used in the study

Five input process parameters, plate thickness (A), pulsefrequency (B), wire feed rate (C), wire feed rate to travelspeed ratio, i.e. WFR/TS ratio (D) and peak current (E) areused in the present study to predict two output parame-ters, the depth of penetration (PD) and the convexity index(CI). Experiments have been carried out for different com-binations of inputs and the bead penetration depth andconvexity index have been found (Srinivasa Rao et al.,2005).

The process parameters included in performing the exper-iment are three levels of plate thickness ‘A’ (6, 8, 10 mm),three levels of frequency ‘B’ (50, 101, 152 Hz), three levels ofwire feed rate/travel speed ratio ‘D’ (15, 20, 25), three lev-els of peak current ‘E’ (440, 480, 520) and varying wire feedrate (C) at 3.0, 4.0, 4.5, 5.0, 5.5, 6.0, 6.5, 7.0 and 8 m/min.All other parameters except these under consideration arefixed.

3.2. Development of neural network model

Various experiments have been carried out for different com-binations of inputs and the bead penetration depth andconvexity index have been found. Fifty-four samples are takento develop the neural network model, of which, 48 samples are

both depth of penetration and convexity index of the bead.

used for training the neural network model and 6 samples areused for testing.

4. Network experimentation

In this study, an attempt is made to construct a single multioutput BPN model to predict both weld dimensions with onenetwork, as it is believed that the welding parameters andweld dimensions are interrelated in such a way that the solu-tion should always be considered as a set (Manikya Kanti,2006). Therefore solving all with one network is a more logicalapproach in this case. The neural network development andthe training are carried out using MATLAB 6.1 application tool.

A schematic representation of the BPN model is shown inFig. 3. The hidden network structure contains two layers anda bias node. The neural network employed in this case is afeed forward back propagation rule with ‘Traingda’ as learningalgorithm.

The input layer has five neurons (since five input param-eters) and the output layer has two neurons (depth ofpenetration and convexity index). Various network structureswith one and two hidden layers with varying number of neu-rons in each layer were examined. Out of all the networkstested, a network with two hidden layers with five and fourhidden neurons in each layer reached the best performancegoal when compared with other networks. This particular net-work also gave accurate results for test data compared to othernetworks. Hence the network with two hidden layers with fiveand four hidden neurons is selected in this work.

The inputs and the outputs are normalized in the range[0,1]. A log-sigmoid transfer function is used for both the layersas the outputs are ranging between zero and one. The net-work was created in MATLAB 6.1 using the function net1 =newff(minmax(p),[5,4,2],{‘logsig’,‘logsig’,‘logsig’},‘traingda’).

Then it was imported to GUI network/data manager fortraining and testing the results. The MATLAB command win-dow consisting of sample code and workspace containing theresults of the present work is shown in Appendix A. The num-bers of the samples for training and testing are 48 and 6,respectively as discussed above. With a learning rate of 0.06,lr dec of 0.5, lr-inc of 1.05 and max perf inc of 1.04, and a

150,000 iterations and an error of 0.0006 was reached as con-nection weights increased and decreased as a neural networksettled down to a stable cluster of mutually excitatory nodesas shown in Fig. 4

j o u r n a l o f m a t e r i a l s p r o c e s s i n g t e c h n o l o g y 2 0 0 ( 2 0 0 8 ) 300–305 303

Fp

5

Abrdtprtrpipo

Br

oirrt

Fpm

Fig. 6 – Comparison of predicted and actual results inpredicting convexity index while training using BPN model.

Fig. 7 – Comparison of % error of predicted results usingBPN model for depth of penetration.

ig. 4 – Training the network to predict depth ofenetration and convexity index.

. Results

back propagation neural network (BPN) model to predictead geometry is developed in this work. To ensure the accu-acy of the BPN model developed to predict bead penetrationepth and the convexity index, the experimental results andhe predicted results using the developed BPN model are com-ared. A correlation coefficient is calculated to measure theelationship between experimentally measured output andhe output predicted by the BPN model. However, a high cor-elation coefficient is not necessarily equivalent to accurateredictions since the slope of the measured vs. predicted plot

s not reflected by this parameter (Chan et al., 1999). Therefore,ercentage difference is also included to measure the spreadf prediction error.

The following sections show the results obtained by thePN model and their comparison with the experimentalesults.

A comparison plot of the analysis data and verification dataf the penetration depth and the convexity index of the bead

n the multiple regression analysis is shown in Figs. 5 and 6,espectively. It can be observed from Figs. 5 and 6 that theesults obtained by BPN model are in close agreement withhe experimental results for training samples.

ig. 5 – Comparison of predicted and actual results inredicting bead penetration depth while training using BPNodel.

Fig. 8 – Comparison of % error of predicted results using theBPN model for convexity index.

Fig. 9 – Line of best fit for predicted PD and the actual PD byBPN model for training samples.

304 j o u r n a l o f m a t e r i a l s p r o c e s s i n g t e c h n o l o g y 2 0 0 ( 2 0 0 8 ) 300–305

Fig. 10 – Line of best fit for predicted CI and the actual CI byBPN model for training samples.

Fig. 11 – Line of best fit for predicted CI and the actual CI by

Fig. 12 – Line of best fit for predicted depth of penetration

BPN model for testing samples.

The percentage difference to measure the spread of pre-diction error obtained by the BPN model for all the samples inpredicting the depth of penetration and the convexity indexof the bead is calculated. It can be observed from Figs. 7 and 8that the ability of the BPN model to predict bead geometry iswell within the allowable error.

The regression analysis is performed to find out the correla-tion coefficient. The correlation coefficient is used to measure

the relationship between the measured and predicted values.And it is observed that a regression coefficient of 0.989 and0.983 is obtained while predicting the bead penetration depthand the convexity index by BPN model, respectively for train-

and the actual depth of penetration by BPN model fortesting samples.

ing samples proving that accurate bead prediction is possiblethrough the welding input parameters using this model. Theline of best fit using the plotted points is calculated using theregression and is shown in Figs. 9 and 10.

To determine the accuracy of the BPN model for predict-ing the outputs when an unknown input is given a regressionanalysis is performed on the network and line of best fit isdetermined (Figs. 11 and 12) for testing samples of NN model.A correlation coefficient of 0.99 and 0.954 are obtained for thetesting samples while predicting the depth of penetration andconvexity index respectively.

6. Conclusions

The effects of process parameters on bead penetration depthand convexity index in pulsed GMA welding using artificialneural networks has been studied, and the following conclu-sions have been reached:

• The neural network model developed in this work fromthe experimental data (for bead penetration depth andconvexity index) can be employed to control the processparameters in order to achieve the desired bead penetrationdepth and convexity index.

• It is observed from the results that a correlation coeffi-cient of 0.99 is obtained between the results obtained by theexperimental and the BPN model developed. Also the per-centage error obtained for desired bead penetration depth

and convexity index is relatively very less. This shows thatthe developed neural network model is capable of makingthe prediction of the weld bead dimensions with reasonableaccuracy.

Ar

r

C

H

H

K

K

K

K

L

j o u r n a l o f m a t e r i a l s p r o c e s s i n g t e c h n o l o g y 2 0 0 ( 2 0 0 8 ) 300–305 305

ppendix A. MATLAB command window consisting of sample code and workspace containing theesults

e f e r e n c e s

han, B., Pacey, J., Bibby, M., 1999. Modeling gas metal arc weldgeometry using artificial neural network technology. Can.Metall. Quart. 38 (1), 43–51.

agan, M.T., Demuth, H.B., Beale, M., 1996. Neural NetworkDesign, Thomson Learning. Vikas Publishing House.

ammerstrom, D., 1993. Working with neural networks. IEEESpectrum, 46–53, July.

im, I.S., Park, C.E., Jeong, Y.J., Son, J.S., 2001. Development of anintelligent system for selection of the process variables in gasmetal arc welding processes. Int. J. Adv. Manuf. Technol. 18,98–102.

im, I.S., Son, J.S., Park, C.E., Lee, C.W., Prasad, Y.K.D.V., 2002. Astudy on prediction of bead height in robotic arc welding usinga neural network. J. Mater. Process. Technol. 130–131, 229–234.

im, I.S., Son, J.S., Park, C.E., Kim, I.J., Kim, H.H., 2005. Aninvestigation into an intelligent system for predicting beadgeometry in GMA welding process. J. Mater. Process. Technol.159, 113–118.

Manikya Kanti, K., Prediction of bead geometry in pulsed GMAwelding using artificial neural networks. MTech Thesis. G.V.P.College of Eng., Visakhapatnam, 2006.

MATLAB 6.1, Help Navigator, 2001.Rajasekaran, S., 1999. Weld bead characteristics in pulsed GMA

welding of Al–Mg alloys. Weld. J. 78 (12), 397s–407s.Rukmini Srikant, R., Tool wear monitoring using artificial

intelligence. MTech Thesis. GITAM College of Engineering,Visakhapatnam, 2003.

Senthil Kumar, R., Parmer, R.S., 1986. Weld bead geometryprediction for pulse MIG welding. In: Proceedings of anInternational Conference on Trends in Welding Technology,USA, May 18–22, pp. 647–652.

Srinivasa Rao, P., Development of arc rotation mechanism andexperimental studies on pulsed GMA welding with andwithout arc rotation. PhD Thesis. I.I.T. Kharagpur,2004.

Srinivasa Rao, P., Gupta, O.P., Murty, S.S.N., 2005. Influence ofprocess parameters on bead geometry in pulsed gas metal arcwelding. In: IIW International Congress 2005 on Frontiers ofWelding Science and Technology, Mumbai.

rose, B., Van der Smagt, P., 1996. An Introduction to NeuralNetworks, eighth ed, November.

ee, J.I., Um, K.W., 2000. A prediction of welding processparameters by prediction of back bead geometry. J. Mater.Process. Technol. 108, 106–113.

Yagnanarayana, B., 1999. Artificial Neural Networks.Prentice-Hall, India.