7/29/2019 09-2003-KANG-s.pdf

1/10

WELDING RESEARCH

SEPTEMBER 2003-S238

ABSTRACT. This study was conducted todevelop the best model to estimate spatterrate for the short circuiting transfer modeof the gas metal arc (GMA) weldingprocess. Utilizing an artificial neural net-

work. The spatter rate generated duringwelding is a barometer of process stability

for metal transfer, and it depends on theperiodic waveforms of the welding currentand arc voltage in the short circuitingmode. Twelve factors representing thecharacteristics of the waveforms as inputsand the spatter rate as an output were em-ployed as variables for the neural network.Two neural network models were evalu-ated for estimating the spatter rate: onemodel did not consider arc extinction; theother model did. The input vector and thenodes of hidden layers for each model

were optimized to provide an adequate fit,and estimated performance of each opti-

mized model to the spatter rate was as-sessed and compared with the previouslyproposed model. It was, in addition,demonstrated in this study that the com-bined neural network model was more ef-fective in predicting the spatter rate thanother models through evaluation of the es-timated performance of each optimizedmodel.

Introduction

In GMA welding, process quality canbe represented by bead shape, implying

the bead width and penetration depth, orby spatter rate. Prediction models to esti-

mate process quality in terms of quantity,nevertheless, are required since the quan-tities cannot be measured during a weld-ing process. The short circuiting transfermode has periodic characteristics of arcand short circuit between the welding wireand the weld pool. The metal droplet

grows at the electrode tip when the arc ismaintained, and it is transferred to the

weld pool while in contact with the tip. Ir-regular waveforms of the welding currentand arc voltage indicate variations in themetal droplet size and imbalance betweenseveral forces on the droplet. The likeli-hood of spatter generated, therefore, issignificantly greater when the short-circuittime or arc time is irregular (Ref. 1) withthe spatter rate closely related to the reg-ularity of the arc and short-circuit time. Astable arc leads to a low spatter rate andregular waveform while an unstable arc

causes a high spatter rate. An evaluationof the arc state through correlation analy-sis between the standard deviation of arctime and stability was attempted for thethe short circuiting transfer mode (Ref. 2).It was also reported that the short circuitpeak current and standard deviation of thearc time affect the arc stability to a greatextent (Ref. 3), and Suban (Ref. 4) has at-tempted to analyze weldability adoptingthe probability distribution function andFFT analysis of current waveform, thusshowing that the arc becomes stable withhigher intensity and narrow distribution.

Shinoda and Nishikawa (Ref. 5) have re-ported that arc stability is related to theperiod and size of the area wherein thecurrent-voltage diagram (I-V diagram) is

drawn in the short circuiting mode. Theyhave proposed an index expressing arc sta-bility by adopting the welding current andthe arc voltage waveforms in short circuit-ing transfer. Gupta (Ref. 6), in addition,has examined how wire size, gas flow rate,and electrode extensions affect the spatter

loss showing that a small-sized wire has ahigher spatter loss at a lower arc voltage

while a large-sized wire has a higher spat-ter loss at a higher arc voltage with thephenomena depending on the metaltransfer mode. Hermans et al. (Refs. 7, 8)showed that arc voltage and wire feed rateaffect the short-circuiting frequency inGMA welding and that the arc becomesmore stable as the standard deviation ofthe short-circuit frequency gets smaller.They have, in addition, proved that the arcstability is affected by weld pool oscilla-tion, and the arc reaches the maximum

stability state when the short-circuit fre-quency equals weld pool oscillation.Ogunbiyi and Norrish (Ref. 9) have classi-fied metal transfer mode and arc stabilityusing several indices induced from mini-mum current, average current, maximumcurrent, and average voltages during thesampling period. Mita et al. (Ref. 10) haveextracted the waveform elements, includ-ing arc time, short-circuit time, averagearc current, and average short-circuit cur-rent, and their standard deviations, fromthe measured signals of the welding cur-rent and arc voltage. They have proposed

several regression models composed ofthe waveform factors comprising thewaveform elements and their standard de-viations, and developed the optimal re-gression model based on the evaluation of

welders subjective judgment, leading to aconsideration of the model as the arc sta-bility estimation index. They have demon-strated that the arc becomes more stableas the index moves down and vice versa.Most of the results reported hitherto havebeen induced from the arc stability or themetal transfer mode using the regularityof the waveform factors or particular rela-tionships between the welders experience

Spatter Rate Estimation in the Short-CircuitTransfer Region of GMAW

A study shows the optimal model for estimating spatter rate using an artificial

neural network in the short circuit transfer region of GMAW

BY M. J. KANG, Y. KIM, S. AHN, AND S. RHEE

M. J. KANG is a Senior Researcher (moojin@kitech.re.kr), Advanced Welding and JoiningTechnology Team, KITECH, Chonan, Korea. Y.KIM is a Researcher (yongjaekim@doosanheavy.com), Welding Technology ResearchTeam, Doosan Heavy Industries and Construc-tion Co., Ltd., Changwon, Korea. S. AHN is anAssociate Professor (shahn@chuldo.krc.ac.kr),Dept. of Train Operation and Mechatronics En-gineering, Korea National Railroad College, Eui-wang, Korea. S. RHEE is an Associate Professor(srhee@hanyang.ac.kr), Dept. of Mechanical

Engineering, Hanyang University, Seoul, Korea.

KEY WORDS

Gas Metal Arc WeldingShort Circuit TransferNeural NetworkSpatter RateComputer Data Analysis

7/29/2019 09-2003-KANG-s.pdf

2/10

WELDING RESEARCH

-S239WELDING JOURNAL

and certain models to estimate the arcstate. The models, nevertheless, are notexplained quantitatively and are deficientin statistical verification. The authors(Refs. 11, 12) have proposed a statisticalmodel to predict the process stabilityquantitatively by applying a relationshipbetween the spatter rate and the wave-form factors in GMA welding to multipleregression analysis, but there is still a large

error in predicting the spatter rate withthese models.

If the stability of the welding process isnot assessed precisely during welding, it be-comes very difficult to comprehend

whether welding is performed under properconditions. The majority of spatters occurunder improper welding conditions leadingto an understanding that it is possible tochange the process conditions to get ideal

weld quality if the process state is preciselyassessed, and this is dependent on the esti-mation accuracy of a prediction model as-sessing the state of the welding process.

This study attempted to develop opti-mal models, which can predict process sta-bility (or metal transfer stability) for shortcircuit transfer in GMA welding in realtime quantitatively through the use of anartificial neural network. And theseneural network models can be adopted tocontrol the arc when an improper weldingcondition is established or the arc status ischanged by any disturbances. Ideal weldquality with minimal spatter can specifi-cally be obtained by an appropriate ad-

justment of the arc voltage wherein thecontrol performance is dependent on the

estimation quality of the model. This studyfocuses on the development of estimationperformance.

Experimental Procedure

The experimental devices used in thisstudy to extract the waveform factors af-fecting spatter generation included a

welding power source, a spatter capturingdevice, welding process parameter sen-sors, A/D and D/A converters, and a com-puter, as shown in Fig. 1. The weldingpower source used was a transistorized in-

verter-control led type with a maximumoutput of 350 A. It had only constant volt-age characteristics without any waveformcontrol functions.

The welding current and voltage wereused as set variables, and the welding cur-rent was measured with a hall sensor that

was attached to the ground cable, whilethe arc voltage was measured from theoutput terminals of the welding powersource. The measured signals were putinto the computer through an A/D con-

verter with a maximum sampling rate of200 kHz.

A gathering device for spatter was

made of brass and in the shape of a box.The welding gun and workpiece werecompletely enclosed by the device before

welding to prevent the spatter fromsplashing out during welding. The weldingexperiment was conducted for one minutefor each welding condition, and the spat-

ter in the gathering device was subse-quently collected and screened using asieve with 300 (grid/in.2) mesh. The spat-ter on the sieve was weighed using an elec-tronic balance. The characteristics of spat-ter, such as the size or type, were notconsidered. The welding variables were100% CO2 shielding gas with a flow rate of20 L/min, a 1.2-mm-diameter AWSER70S-6 welding wire, 6-mm-thick

ASTM A36M workpiece, and a weldingspeed of 5 mm/s.

The welding parameters are shown inTable 1. The welding conditions lackingthe pure short circuit metal transfer mode

(250 A and 28 V or higher) were excludedand all of the data were handled solely forthe short circuit transfer mode. The weld-ing experiments on plate without a groove

were repeated five to six times under thesame welding conditions to reduce exper-imental errors.

The sampling rate to gather the wave-forms of the arc voltage and welding cur-rent was 5000 samples/s, and the wave-form signals were collected 10 secondsafter the welding start at 20-second inter-

vals. A digital low-pass filter with a 200-Hzcutoff frequency removed the noise on thesignals, and six waveform factors and theirstandard deviations influencing the spat-ter generation were extracted from the

welding current and arc voltage wave-forms. The 12 extracted factors are as fol-low (Fig. 2): short-circuit period (T), arctime (Ta), short-circuit time (Ts), short-circuit peak current (I

p

), short-circuit in-

Fig. 1 Configuration of experimental setup.

Table 1 Welding Conditions and the Number of Experiments per Setting Condition

Wire Feed Rate Contact Tube-to-Workpiece Welding Voltage No. of Welding(m/min) Distance, CTWD (mm) (V) Experiments Per

Setting Condition

3.4 15 19 ~ 25 66.0 15 20 ~ 26 6

20 21 ~ 27 325 21 ~ 27 3

7.3 15 21 ~ 27 3

20 22 ~ 28 425 21 ~ 27 38.6 20 23 ~ 27 6

7/29/2019 09-2003-KANG-s.pdf

3/10

WELDING RESEARCH

SEPTEMBER 2003-S240

stantaneous current (Is), average currentduring short-circuit period (I), standarddeviation of the short-circuit period(s[T]), standard deviation of arc time(s[Ta]), standard deviation of short-circuittime (s[Ts]), standard deviation of short-circuit peak current (s[Ip]), standard devi-ation of short-circuit instantaneous cur-rent (s[Is]), standard deviation of averagecurrent during short-circuit period (s[I]).

Neural Network Structure

The purpose of this study is to developmodels for evaluating the spatter ratebased on the conventional feed-forwardmultilayer perceptrons with the errorback-propagation as the learning algo-rithm. The neural network with typicalfeed-forward multilayer perceptrons iscomposed of an input layer, hidden layer,

and output layer. The neural networkstructure used in this study is shown in Fig.3. The variables of the input layer are theaforementioned 12 factors, whereas theoutput variable is the spatter rate. Ofcourse, there are many other factors, suchas power source characteristics, induc-tance, contact tube-to-workpiece distance(CTWD), shielding gas, and the size andchemical composition of the electrode, af-fecting the spatter generation in short cir-cuiting metal transfer with GMAW. Butsince the variation of these factors inducesthe variation of the welding current and

voltage waveform, it is very efficient to usethe waveform factors instead of all affect-ing factors as the input variables of theneural network model for spatter rate es-timation. The neural network model wasprogrammed using the C++ program code

without using other commercial softwarepackages.

An output of the hidden layer is pro-duced using the values of the input vari-ables and the connection weight. The out-put of the output layer is obtained byprocessing the output of the hidden layerand the relevant connection weight, and

this process can be expressed as shown inEquations 13.

(1)

(2)

(3)

o f w xt tt

n

=

=0

x f u y t nt tj jj

m

= ( )=0

=1,2,...,

y f z j mj ji ii

l

= ( )=

n

0

=1,2,...,

Fig. 2 Waveforms of arc voltage and welding current in short-circuit trans-

fer mode.

Fig. 3 Multilayer feed-forward neural network.

Fig. 4 The spatter rate with respect to the welding voltage at different wire feed rates. A Wire feed

rate 3.4 m/min; B wire feed rate 6.0 m/min; C wire feed rate 7.3 m/min; D wire feed rate 8.6

m/min.

7/29/2019 09-2003-KANG-s.pdf

4/10

WELDING RESEARCH

-S241WELDING JOURNAL

Where zi is the input vector of the inputlayer,yj, xt, ando are the output vectors ofthe hidden layers and output layer, re-spectively. nji, utj, andwt are the arrays ofconnection weight, andl, m, andn are thenumber of the neurons in the input andhidden layers, respectively. Bipolar sig-moid function was adopted as an activa-tion function in Equation 4 due to its pop-ularity.

(4)

The least square method was used in theerror calculation as shown in Equation 5

(5)

whereE is the cost function measuring themean squared error andp is the number of

learning patterns. dk is the kth desiredvalue andok is thekth output value of theneural network. 2p was used in order tonormalize the cost function in the range of[0, 1].

The values of the connection weights inthe learning process are changed to mini-mize the mean squared error,E, based onEquations 6 through 8, and thus the valuesof the connection weights are adjusted tothe opposite direction of the gradient ofthe mean squared error,E, and learning isperformed in the direction in which theerror decreases. The rate of the decrease

depends on the learning rate h.

(6)

(7)

(8)

The Learning Procedure for theSpatter Rate Estimation

A fixed input of 1 was added to theinput and hidden layers while the initialnumber of neurons of each layer was es-tablished asl = 12,m = 20, andn = 20 in

the learning process for the spatter rate es-timation while that of the output layer was1. The connection weights nji, utj, and wt

were expressed in an array with the sizesof [m (l + 1)], [n (m + 1)], and [1 (n+ 1)], respectively. These were initializedas the random values between [1, 1] be-fore learning, and the input vectors, wave-form factors in other words,z,were madeto have values normalized as 1 to theminimum value and as 1 to the maximum

value. In addition, the output vectoro, the

spatter rate estimated through learning,was normalized to have a value betweenthe range of [1, 1], and the learning rate

was 0.1 and the error boundary was estab-lished as 0.002 to the cost function,E, be-tween the estimated results of the normal-ized output variable and the normalized

values to spatter rate.

Optimization of the NeuralNetwork

In constituting the multiple layers neuralnetwork model, the number of hidden lay-

ers and the number of nodes of each layermay influence the output. The number ofhidden layers of the neural network in thisstudy was fixed at two. One hidden layerneural network model retained the diver-

gence problem and a three or more hiddenlayers model retained the convergent timeproblem. The three-layer model with twohidden layers, therefore, was determined asoptimal in terms of the convergence andconvergent time. The number of neurons isalso a very important factor for estimationperformance. Networks with too few train-able parameters for the given amount oftraining data fail to learn the training data.With too many trainable parameters, thenetwork learns well but does not generalize

well and performs very poorly on the new

Dn hn

ji

ji

E= -

DuE

utj

tj

= -

h

DwE

wt

t

= -

h

Ep

d ok kk

p

= -( )=

1

2

2

1

f xx

( ) =+ -( )

-2

1 exp1

Table 2 Correlation Coefficients between Each Factor and Spatter Rate in the Various Regions

Region T Ta Ts Ip Is I s[T] s[Ta] s[Ts] s[Ip] s[Is] s[I]

Overall 0.9198 0.723 0.109 0.1746 0.2615 0.163 0.7457 0.7361 0.1749 0.5019 0.1986 0.0423Arc extinction 0.0593 0.0437 0.1235 0.230 0.351 0.006 0.4312 0.3666 0.6504 0.7650 0.8886 0.8816No arc 0.8165 0.8267 0.164 0.1757 0.2592 0.273 0.7914 0.7876 0.2565 0.6562 0.1398 0.3033extinction

Fig. 5 Waveforms of the arc voltage and the welding current under the lower and upper welding volt-

age condition. A 150 A, 19 V; B 250 A, 28 V.

Fig. 6 Number of arc extinctions with respect

to the welding voltages at a CTWD of 15 mm.

7/29/2019 09-2003-KANG-s.pdf

5/10

WELDING RESEARCH

SEPTEMBER 2003-S242

data not used in training. This phenomenonis usually called overfitting (Ref. 13). Thereare several generalization methods: regular-ization technique using modified perfor-mance function, automated regularizationmethod, early stopping technique, etc. (Ref.14). One of the generalization methods is touse a network that is just large enough toprovide an adequate fit. In this study, by op-timizing the number of input and hidden

nodes, the optimal-sized neural networkscould be obtained. In optimizing the neu-rons of the input layer, the factors with thelowest correlation coefficients (Table 2) be-tween each variable of the input layer andthe spatter rate were removed in turn withinthe convergence condition of an errorboundary of 0.002. In optimizing the neu-rons of the hidden layers, the number ofneurons (20 20) was reduced as long asthey converge within the limited number ofrepetitions (10,000 times).

Results and Discussion

Neural Network Models for the Spatter

Rate Estimation

Spatter rate was measured for all thewelds using the conditions shown in Table1. As shown in Fig. 4, there was an optimal

voltage wherein the minimum spatter ratewas generated.

Figure 5 shows the typical waveforms ofthe arc voltage and welding current whereone is for high current-voltage and theother is for low current-voltage. In Fig. 5A,it can be seen that the arc is reignited after

the short circuit and the arc is then com-pletely extinct, showing the open-circuit voltage. This arc extinction is dif-ferent from the arc extinguishment causedby the short circuit where the metal dropletis contacted to the weld pool. This leads toan understanding that the arc extinguishesand the circuit is electrically disconnectedat the moment of arc reignition after theshort-circuit period. Figure 6 shows thenumber of arc extinctions generated, not atthe moment of short circuit but at the mo-ment of arc reignition. This number wasgradually reduced as the welding voltageincreased, and the arc extinction subse-quently did not occur as the voltage be-came higher than a certain value.

The correlation analysis between 12waveform factors and the spatter rate wasperformed for the three cases as follows:the overall range without considering thearc extinction, the welding conditions

wherein the arc extinction occurs, and thewelding conditions without the arc extinc-tion. The second row of Table 2 shows thecorrelation coefficients between the spat-ter rate and waveform factors obtainedunder all welding conditions overallrange in other words regardless of the

Table 3 Input Variables of Each Neural Network Model

Model T Ta Ts Ip Is I s[T] s[Ta] s[Ts] s[Ip] s[Is] s[I]

Overall O O O O O O O O O OArc O O O O O O O O OextinctionNo arc O O O O O O O O Oextinction

Fig. 7 Relationship between the estimated results and the spatter rate with the trained data set using

several models. A Nonlinear regression model in overall range; B combined nonlinear regression

model; C neural network model in overall range; D combined neural network model.

Fig. 8 Comparison of estimation performance between model of overall range and combined model

with a new data set. A Neural network model in overall range; B combined neural network model.

7/29/2019 09-2003-KANG-s.pdf

6/10

WELDING RESEARCH

-S243WELDING JOURNAL

arc extinction. The third row shows underwelding conditions where arc extinctionoccurred, while the fourth row showsunder welding conditions where the arcextinction did not occur. According to thesecond and fourth row of Table 2, the be-havior of the correlation coefficients in allcases was very similar, and the primaryfactors affecting the spatter rate, in par-ticular, were the short-circuit period (T),

the standard deviation of the short-circuitperiod (s[T]), the standard deviation ofthe arc time (s[Ta]), and arc time (Ta), etal. The two factors, the short-circuit time(Ts) and average current during a shortcircuit (I), were eliminated to prevent themulticolinearity. The standard deviationof average current during a short-circuitperiod (s[I]), the standard deviation ofshort-circuit instantaneous current (s[Is]),the standard deviation of short-circuitpeak current (s[Ip]), and the standard de-

viation of short-circuit time (s[Ts]), never-theless, affected the spatter rate to a great

extent according to the third row of Table2. It is shown that the primary waveformfactors for the spatter generation whenthe arc extinction occurs differ from those

when the arc extinction does not occur. Amodel in overall range without consider-ing the arc extinction and another modelconsidering the arc extinction were pro-posed to develop a more precise model forthe spatter rate estimation and to comparethe estimation performance. The modelconsidering the arc extinction combinestwo conditions: one under welding condi-tions where the arc extinction occurs and

the other where it does not occur.

The Neural Network Model for the Spatter

Rate Estimation in the Overall Range

The hidden layer was first set at 20 20,and the number of neurons of the hiddenlayer was reduced in turn as long as theyconverge within 10,000 repetitions, in de-

veloping the neural network model for thespatter rate estimation in the overall range.Finally, the size of neurons in the hiddenlayer of 5 5 with an error of 0.00196 wasobtained in 9976 repetitions. The 12 wave-form factors were then gradually eliminatedin optimizing the input variables, beginning

with those with the lowest correlation coef-ficient on the spatter rate based on the cor-relation analysis of Table 2. Last, the input

variables remaining within the convergenceerror bound during 10,000 repetitions werethe 10 factors in Table 3.

The Neural Network Model for the Spatter

Rate Estimation Considering

Arc Extinction

In welding conditions where the arc ex-tinction occurs, it was confirmed that the

hidden layer of the neural network modelfor estimating the spatter rate was opti-mized to a size of 7 5 in 9983 repetitions

with a normalized output error of 0.00199.The input variables were optimized based

on the results of the correlation analysis inTable 2. As a result, the factors of the inputlayer were the nine factors in Table 3.

In welding conditions wherein there isno arc extinction, the normalized output

Table 4 Regression Characteristics of Several Estimation Models

Nonlinear Regression Neural NetworkModel Model

Model in Combined ModelOverall Range

Model Combined Trained New Data Trained New Datain Overall Model Data Set Set Data Set Set

Range

Multiple correlation 0.9109 0.8424 0.9441 0.9156 0.9986 0.9467coefficients

Adjusted R2 0.8258 0.7084 0.9058 0.8368 0.9965 0.8953Standard error 0.4562 1.2954 0.3134 1.109 0.0691 0.8884

Fig. 9 Comparison of the estimated results by model in the overall range and combined model with

spatter rate (wire feed rate of 3.4 m/min). A Model in overall range; B combined model.

Fig. 10 Error between the estimated results from model in the overall range and combined model

with spatter rate (wire feed rate of 3.4 m/min). A Model in overall range; B combined model.

7/29/2019 09-2003-KANG-s.pdf

7/10

WELDING RESEARCH

SEPTEMBER 2003-S244

Fig. 11 Comparison of the estimated results from model in overall range and combined model with spatter rate (wire feed rate of 6.0 m/min). A, B, and C

Model in overall range; D, E, and F combined model.

Fig. 12 Error between the estimated results from model in overall range and combined model with spatter rate (wire feed rate of 6.0 m/min). A, B, and C

Model in overall range; D, E, and F combined model.

7/29/2019 09-2003-KANG-s.pdf

8/10

WELDING RESEARCH

-S245WELDING JOURNAL

error was 0.00198 in 9994 repetitions whenthe hidden layer was 5 5 in size. The

waveform factors were optimized basedon the results of the correlation analysis

shown in Table 2. As a result, the factorsof the input layer were determined as thenine factors in Table 3.

The combined neural network modelitself cannot distinguish whether arc ex-tinction occurs or not. The neural networkmodel for the arc extinction region is ap-plied when the arc voltage retains theopen circuit voltage value, the occurrenceof the arc extinction in other words, andthe model for nonarc extinction region isapplied otherwise.

Experimental Results

The proposed models were testedusing two data sets. One was the traineddata set shown in Table 1 and the other wasa new set of 108 data. Figure 7 is the testedresult with the trained data set. Figure 7Aand B shows the estimated results ob-tained using the overall and combinednonlinear regression models proposed byKang and Rhee (Ref. 12). Figure 7C andD shows the results of the estimated spat-ter rate using the neural network modelsproposed in this study.

A large error was shown in predicting

the spatter rate in the regression models.

The results with the neural network mod-els, in contrast, were linearly related to theactual spatter rate in ranges where thespatter rate was over 5 g/min, in particu-

lar. The estimated results with the neuralnetwork model in the overall ranges, nev-ertheless, retained a slightly larger errorrate than that with the combined neuralnetwork model in ranges where the spat-ter rate was under 5 g/min. Figure 8 is thetested result with a new set of 108 data. Itcould be seen that the dispersion of the es-timated results with the combined neuralnetwork model was less than that with theneural network model in the overall range.

Table 4 shows the compared results be-tween the regression models and theneural network models, and the multiple

regression coefficients obtained using re-gression models show significantly lower

values than those obtained using twoneural network models. The multiple cor-relation coefficients obtained using thecombined neural network model show re-sults close to 1, and the standard error

with the combined neural network modelwas determined to be 0.0691, showing agood estimation performance.

When the wire feed rate was set at 3.4m/min, the estimated results with theneural network model in overall range withthe combined model was compared to the

spatter rate according to the changes in

welding voltage as shown in Fig. 9. As a re-sult, the estimated results of the combinedmodel were considered more accuratethan the results of the model in the overall

range. The error between the estimatedvalues for each model and the spatter rateto compare the results more accurately isshown in Fig. 10. According to Fig. 10, theestimation error with the neural networkmodel in the overall range is larger thanthat with the combined model. The esti-mation error with the combined model wasbelow 0.1 g/min, in particular, under lower

voltage welding conditions where the arcextinction occurs, confirming the superior-ity of the combined models estimationperformance.

Figure 11 shows the estimated results

with the neural network model in the over-all range and the estimated results withthe combined model according to thechanges in welding voltage when the wirefeed rate is 6 m/min. Figure 12, in addi-tion, shows the estimation error with eachmodel, and the error with the model in theoverall range shows the maximum value of0.78 g/min at CTWD of 15 mm and weld-ing voltage of 19 V, where arc extinctionfrequently occurs, whereas the estimationerror with the combined model was below0.1 g/min. The estimation error with themodel in the overall range at CTWD of 25

mm and the welding voltage of 27 V was

Fig. 13 Comparison of the estimated results from the model in the overall range and the combined model with spatter rate (wire feed rate of 7.3 m/min). A,

B, and C Model in overall range; D, E, and F combined model.

7/29/2019 09-2003-KANG-s.pdf

9/10

WELDING RESEARCH

SEPTEMBER 2003-S246

above the optimal voltage and over 0.9g/min, while the estimation error with thecombined model was below 0.3 g/min. Inaddition, the estimation errors in theranges where the arc extinction occurs andthe combined model retained an errorbelow 0.2 g/min and an average error

below 0.4 g/min under the welding condi-

tions of 25~27 V.Figure 13 shows the estimated results

with the neural network model in the over-all range and the estimated results withthe combined model when the wire feedrate is 7.3 m/min, according to the changesin the welding voltage, and Fig. 14 shows

the estimation error with each model. In

CTWD of 20 mm, 25 mm, and lower volt-age conditions, where the arc extinctionfrequently occurs, the estimation error

with the model in the overall range re-tained a maximum estimation error ofmore than 0.8 g/min. The estimation error

with the combined model, on the contrary,shows a value below 0.1 g/min under thelow voltage conditions where arc extinc-tion frequently occurs. Furthermore, theestimation error with the combined model

was considered better than that with themodel in the overall range even in thehigher voltage conditions where there is ahigh spatter rate.

Figure 15 shows the estimated resultswith the neural network model in the over-all range and the estimated results with thecombined model when the wire feed rate is8.6 m/min, according to the changes in the

welding voltage. Figure 16 shows the esti-mation error with each model. The estima-tion error with the model in the overallrange showed the maximum value of 0.3 ~0.7 g/min, depending on the voltage condi-tion. The error with the combined model,nevertheless, showed a value below 0.2g/min over all voltage conditions.

This study confirmed that the com-bined neural network model where arc ex-tinction was considered demonstrated theideal estimation performance in the short

circuit transfer mode of GMA welding.

Fig. 14 Error between the estimated results by model in the overall range and the combined model with spatter rate (wire feed rate of 7.3 m/min). A, B, and

C Model in overall range; D, E, and F combined model.

Fig. 15 Comparison of the estimated results from the model in the overall range and the combined

model with spatter rate (wire feed rate of 8.6 m/min). A Model in overall range; B combined

model.

7/29/2019 09-2003-KANG-s.pdf

10/10

WELDING RESEARCH

-S247WELDING JOURNAL

Conclusions

A neural network model for the over-all range and a combined neural networkmodel that considered arc extinction areproposed in this study. The neural net-

work structure for each model was opti-mized as [10 5 5 1] and [9 7 5 1] or[9 5 5 1], respectively. As a result, thegeneralization of the neural network wasaccomplished and the proposed models

were determined to be ideal for the spat-ter rate estimation compared with thepreviously proposed results. The com-bined neural network model, which takesinto consideration the arc extinction,showed a linear correlation coefficient of0.9986 and a standard error of 6.91% re-garding the spatter rate. A significant im-provement in estimation performance ata lower voltage was demonstrated in thisstudy where arc extinction occurs.

References

1. Gupta, S. R., Gupta, P. C., and Rehfeldt,

D. 1988. Process stability and spatter genera-

tion during dip-transfer MAG-welding. Weld-

ing Review, pp. 232241.

2. Arai, T., Kobayashi, M., Yamada, T.,

Rokujyo, M., Hirakoso, K., and Kaneko, T.

1983. The investigation of arc phenomena by

means of a computer. Quarterly Journal of the

Japan Welding Society 1(3): 1520.

3. Lucas, W. 1987 : Microcomputer system,

Software and expert system for welding engi-

neering, Welding Journal 66(4): 1930.

4. Suban, M. 2001. Determination of stabil-

ity of MIG/MAG welding. Quality and Reliabil-ity Engineering International 17: 345353.

5. Shinoda, T., and Nishikawa, H. 1995.

Monitoring and signal processing of short cir-

cuiting metal transfer of metal active gas weld-

ing process.Proc. of the Int. Conf. on the Join-

ing of Materials, JOM-7, pp. 558565.Helsingor, Denmark: The European Institute

for the Joining of Materials.

6. Gupta, S. R., and Gupta, P. C. 1984. Ef-

fect of some variables on spatter loss. Welding

and Metal Fabrication 52(9): 361364.

7. Hermans, M. J. M., Sipkes, M. P., and

Ouden, G. D. 1993. Characteristic features of

the short circuiting arc welding process. Weld-

ing Review International 12(2): 8086.

8. Hermans, M. J. M., and Ouden, G. D.

1999. Process behavior and stability in short cir-

cuit gas metal arc welding. Welding Journal

78(4): 137-s to 141-s.

9. Ogunbiyi, B., and Norrish, J. 1996.

GMAW metal transfer and arc stability assess-

ment using monitoring indices. Computer Tech-

nology in Welding, Sixth International Confer-

ence. Cambridge, U.K.: The Welding Institute.

10. Mita, T., Sakabe, A., and Yokoo, T. 1988.

Quantitative estimates of arc stability for CO2

gas shielded arc welding. Welding International2(2): 152159.

11. Kang, M. J., and Rhee, S. 2001. A study

on the development of the arc stability index

using multiple regression analysis in the short

circuit transfer region of gas metal arc welding.

Proc. Inst. Mech. Engrs. Part B 215: 195205.

12. Kang, M. J., and Rhee, S. 2001. The sta-

tistical models for estimating the spatter rate in

the short circuit transfer mode of GMAW.

Welding Journal 80(1): 1-s to 8-s.

13. Lin, C. T., and Lee, C. S., George. 1996.

Neural Fuzzy Systems: A Neuro-Fuzzy Synergism

to Intelligent Systems, Upper Saddle River, N.J.:

Prentice Hall.

14. Demuth, H. B., and Beale, M. H. 1997.

Neural Network Toolbox. Miami, Fla.: The

MathWorks.

Fig. 16 Error between the estimated results from the model in the overall range and the combined

model with spatter rate (wire feed rate of 8.6 m/min). A Model in overall range; B combined

model.

2004 Poster SessionCall for Entries

The American Welding Society announces a Call for Entries for the 2004 Poster Session to be held as part of Welding Show2004 on April 6-8, 2004, in Chicago, Ill. Students, educators, researchers, engineers, technical committees, consultants, andanyone else in a welding- or joining-related field are invited to participate in the world's leading annual welding event by visuallydisplaying their technical accomplishments in a brief graphic presentation, suitable for close, first-hand examination by interestedindividuals.

Posters provide an ideal format to present results that are best communicated visually, more suited for display than verbalpresentation before a large audience; new techniques or procedures that are best discussed in detail individually with interestedviewers; brief reports on work in progress; and results that call for the close study of photomicrographs or other illustrativematerials.

Submissions should fall into one of the following two categories and will be accepted only in a specific format. Individualsinterested in participating should contact Dorcas Troche, Manager, Conferences & Seminars, via e-mail at dorcas@aws.org forspecific details. Deadline for submission of entries is Monday, December 1, 2003.

1. Student DivisionCategory A: 2 Year or Certificate ProgramCategory B: Undergraduate DegreeCategory C: Graduate Degree

2. Professional/Commercial Divison