On-Line Quality Monitoring in Short-Circuit Gas Metal Arc ...

BY S. ADOLFSSON, A. BAHRAMI, G. BOLMSJÖ AND I. CLAESSON

Results show it is possible to detect changes in weld quality automatically and on-line

On-Line Quality Monitoring inShort-Circuit Gas Metal Arc Welding

ABSTRACT. This paper addresses theproblems involved in the automaticmonitoring of the weId quality producedby robotized short-arc welding. A simplestatistical change detection algorithm forthe weld quality, the repeated SequentialProbability Ratio Test (SPRT), was used.The algorithm may similarly be viewedas a cumulative sum (CUSUM) type test,and is well-suited to detecting suddenminor changes in the monitored teststatistic. The test statistic is based on thevariance of the weld voltage, wherein itwill be shown that the variance de-creases when the welding process is notoperating under optimal conditions. Theperformance of the algorithm is assessedthrough the use of experimental data. Theresults obtained from the algorithm showthat it is possible to detect changes inweld quality automatically and on-line.

Introduction

Gas metal arc welding (GMAW) iswidely used in various industrial weldingapplications because it has certain ad-vantages. A high metal deposition ratemakes this method attractive for high-quantity applications and well-suited toautomatic welding (Ref. 1). There are twostable metal transfer modes in direct cur-rent GMAW: 1) short-circuit metal trans-fer at low arc voltage and 2) spray metaltransfer at high voltage. One cycle of thewelding voltage waveform for optimumweld parameters corresponds to the

transfer of one molten droplet in theshort-circuit transfer mode — Fig. 1.Therefore, it is possible to evaluate thestability or regularity of metal transferusing the welding voltage as measuredduring the welding process (Refs. 2–6).

To assess process stability, standarddeviation and different ratios or indiceshave been calculated for suitable weldparameters, such as arc and short-circuittime, short-circuit rate, short-circuit peakcurrent and mean weld voltage and cur-rent (Refs. 2–17).

Monitoring systems for weld parame-ters such as ADM III, Arc Guard andWeldcheck are commercially available(Refs. 18, 19). They all work in a similarway: voltage, current and other processsignals are measured, presented andcompared with preset nominal values.An alarm is triggered when any differ-ence from the preset values exceeds agiven threshold. It is presently believedthat the performance of these systems hasnot, however, been well-documented.

Experiments have shown that in theshort-circuit mode, optimal stability oc-curs when the short-circuit frequencyequals the oscillation frequency of theweld pool (Refs. 5, 6, 20, 21). This cor-

responds to a maximum in short-circuitfrequency. Deviation from the optimalcondition leads to a larger probability ofspatter, uneven weld bead and other fu-sion defects. In this case, the weldingprocess is said to operate under non-op-timal conditions. Thus, a suitable param-eter for the detection of changes in theweld quality is the variance of the ampli-tude of the weld voltage. This parameteris used to form a test statistic that is fedinto a repeated Sequential ProbabilityRatio Test (SPRT) algorithm (Refs. 17, 22,23). The algorithm may similarly beviewed as a cumulative sum (CUSUM)type test. The SPRT is optimal in that itminimizes the worst mean delay for de-tection, given a specified probability forfalse alarms (Ref. 24). Thus, the algorithmis well-suited to detecting abrupt minorchanges in the monitored test parameters(Ref. 22). In addition, storage and com-putational requirements for the repeatedSPRT are moderate, as compared tofixed, sample-sized tests.

Welding Technology

Short-Circuit Metal Transfer

The GMAW process employs a con-sumable wire electrode passing througha copper contact tube — Fig. 2. Electriccurrent supports an arc flowing from theend of the electrode to the workpiece.The electrode is melted by resistive heat-ing and heat from the arc. The region sur-rounding the weld pool is purged withshield gas to prevent oxidation and con-tamination of the weld joint (Refs. 2, 25,26). The advantage of short-circuit weld-ing is that the mean current (thus the av-erage heat input to the workpiece) islower than in spray arc GMAW (Refs. 11,19, 27, 28). Due to the smaller heat trans-fer, short-circuit gas metal arc welding

KEY WORDS

AlgorithmCumulative Sum TestKullback informationShort-Circuit-GMAWStep DisturbanceSequential Probability

Radio TestT-Joint

S. ADOLFSSON is with the Department ofSignal Processing, University of Karlskrona/Ronneby and the Department of Productionand Materials Engineering, Lund University,Sweden; A. BAHRAMI is with the TechnologyCenter of Kronoberg, Växjö, Sweden, and alsoLund University; G. BOLMSJO is with LundUniversity and I. CLAESSON is with the Uni-versity of Karlskrono/Ronneby.

RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

WELDING RESEARCH SUPPLEMENT | 59-s

(GMAW-S) makes it possible to weldthinner plates than when spray arcGMAW is used.

To limit the heat input to the work-piece, the open circuit voltage is set at alow value compared to that used in sprayarc GMAW (Refs. 11, 19, 27). The cyclebegins with an arc struck between theelectrode wire tip and the workpiece. Thewire electrode melts and a small dropletis formed at the electrode tip. This part ofthe cycle is the arc time, Ta — Fig. 1.

During the short-circuit time, Ts, the

voltage will de-crease to almost 0V and the currentwill increase to itsmaximum value.At this stage, thearc will extinguish.A droplet is thendetached andtransferred fromthe electrode tothe weld pool bythe force of thesurface tension ofthe weld pool, thegravitational forceand electromag-netic pinch force(induced by thecurrent) (Refs. 6,29). After the drop-let is detachedfrom the electrodeand transferred tothe workpiece, thearc is reestablishedand the cycle startsover again.

The weld volt-age Uw, arc voltageUa and the voltage over the wire electrodeextension Ue (Ref. 14) are related by

In GMAW-S welding, Uws and Uwa

denote the weld voltage during the short-

circuit time and weld voltage during thearc time, respectively — Figs. 1, 2.

Optimal Process Stability andWelding Conditions

To produce weld joints of uniformweld quality, the welding process shouldbe stable, which will allow metal trans-

U U Uw a e= + . (1)RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

60-s | FEBRUARY 1999

Fig. 1 — The weld voltage and current during short-circuit welding. Themetal transfer is reflected in the weld voltage as almost zero voltageevents during short-circuit time Ts. The interval between the short circuitsis denoted by short-arc time Ta, while Uws and Uwa are the weld voltageduring short-circuit and arc time, respectively. Ip is the short-circuit peakcurrent.

Fig. 2 — The equipment used in short-arc GMAW. Electric current of theweld process is denoted I. Internal resistance and inductance of the weld-ing source are Ri and Li, respectively. Resistance of the wire electrodestickout, i.e., the part of the electrode between the contact tube and thearc, is Re. Length of the electrode stickout and the arc length are le andla, respectively. Voltage over the wire electrode stickout is Ue. Voltagebetween the electrode tip and workpiece (the arc voltage) is denoted Ua.

Fig. 3 — Short-circuit rate for 1.2-mm-diam-eter wire vs. open-circuit voltage (Ref. 27).Optimal process stability corresponds to themaximum of the curve. Zones A, B, C and Din the curve and oscillograms correspond tostubbing in, short-circuit, globular and spraymetal transfer modes, respectively.

(A)

(B)

(C)

Fig. 4 — The two steel joints used in the experiments. A — Reference T-joint, front view; B — reference T-joint, side view; C — T-joint with stepdisturbance, front view.

fer from the elec-trode wire to theworkpiece to be asregular as possible.Experiments haveshown that inshort-circuit mode,optimal processstability occurswhen the short-cir-cuit frequencyequals the oscilla-tion frequency ofthe weld pool(Refs. 5, 6, 20,21). Optimal pro-cess stability corre-sponds to

• a maximum short-circuit rate(Number/s)

• a minimum standard deviation ofthe short-circuit rate

• a minimum mass transferred pershort-circuit

• a minimum spatter loss.

The welding conditions in which op-timal process stability occurs are referredto as optimal welding conditions. Devia-tion from the optimal welding conditionis assumed to lead to a higher probabilityof spatter, uneven weld bead and otherfusion defects. In this case, the weldingprocess is said to be operating under non-optimal welding conditions.

The algorithm discussed in this paperis, however, based on the observation that


RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

(A)

Fig. 5 — T-joint with step disturbance No. 1. Photo of: A — Front; B —rear side of a welded joint. Note that the weld joint at the front of the T-joint in the interval 6–10.5 cm, along the scale (where the weld joint ta-pers) deviates from normal weld quality, i.e., the size of the leg lengthand throat dimension is reduced.

(A)

(B)

Fig. 6 — T-joint with step disturbance No. 1. A — Measured voltage; B —measured current.

Fig. 7 — Weld voltage and current. Normal weld: A — Measured volt-age; B — Measured current. During step disturbance: C — Measuredvoltage; D — Measured current.

(C)

(D)

(A)

(B)

(B)

the variance of theweld voltage ampli-tude decreaseswhen the weldingprocess is not oper-ating under optimalconditions, asshown in Fig. 3(Refs. 11, 13). It willalso be shownbelow that the arctime, short-circuitrate and standarddeviation of theshort-circuit rate areless robust para-meters than thevariance of the weldvoltage when de-tecting defectivewelds.

Experiments

Instrumentation

The experimental setup comprised awelding power source, a Motoman robotcarrying a welding torch, a positioner, awelding table and instrumentation forrecording weld voltage and current. Thework angle of the welding torch was fixedat 45 deg and the travel angle was 0 deg.The distance between the contact tube tipand the plate was 11 mm.

The weld voltage was measured be-tween an electrode applied to the contacttube and a reference electrode screwedinto an aluminum plate that served as aninsulated welding table (Ref. 14). The cur-rent was measured by a current sensor,LEM Module LT 500-S, equipped with atransformer. The sensor was mounted

RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT


(A)

(B)

(E)

(C)

(D)

Fig. 8 — T-joint with step disturbance No. 1. A — Measured short-circuittransfer rate; B — estimated mean arc voltage (

–Uwa); C — median filter

of length 100 applied to the estimated mean arc voltage (–Uwa

M ); D — mea-sured arc time (Ta); E — median filter of length 100 applied to the arctime sequence (Ta

M).

Fig. 9 — T-joint with step disturbance No. 1. A — Measured short-circuittime (Ts); B — median filter of length 100 applied to the estimated short-circuit time in part A (Ts

M); C — measured short-circuit current peak (Ip);D — median filter of length 100 applied to the short-circuit current peaksequence in part C (Ip

M).

(A)

(B)

(C)

(D)


RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

around the return conductor. The sam-pling frequency was 8.192 kHz and thetotal lowpass filter had a cutoff frequencyof 3.0 kHz. The data were transferred forpermanent storage to a personal computer.

Two different types of commercialwelding equipment, the Migatronic BDHS550 and the Kemppi P500, were usedin the experiments. The wire feed ratewas measured to be approximately 113–120 mm/s and the nominal weldingspeed was set at 10 mm/s. The weldingwire material used in the experiment wasESAB OK 12.51, with a diameter of1.0 mm. The shielding gas used was Atal:80%Ar-20%CO2, with a gas flow rate setat 15 I/min.

Creating Various Welding Conditions

The object of the experiments was tocreate various welding conditions in acontrolled manner, while monitoring the

weld voltage and current from the pro-cess — Figs. 4–6. Non-optimal weldingconditions were created using a T-jointin which gaps had been cut out from thestanding plate — Fig. 4. This specimenwas denoted “T-joint with step distur-bance.” With the aid of the step dis-turbance plate, the welding processpassed through non-optimal conditions.A second specimen, a T-joint with thestanding plate in perfect contact with thelaying plate, was used as a reference.This specimen was used to producenormal welds and was denoted “refer-ence T-joint.” During normal welding,the welding process was assumed tobe operating under optimal weldingconditions.

The specimens comprised two rectan-gular 200 x 100 x 3-mm plates of SS 1312mild steel. The dimension of the gap inthe T-joint with step disturbance was 2 x

50 mm — Fig. 4C.A photo of a T-joint with step distur-

bance is shown in Fig. 5. Note that theweld joint at the front of the T-joint (in theinterval 6–10.5 cm along the scale wherethe weld joint tapers) deviates from nor-mal weld quality, i.e., the size of the leglength and throat dimension is reduced.

Experimental Data Analysis

Time Domain Analysisof Measurement Data

The object of this experiment was toconfirm, by examination of the wave-form of the weld voltage and current, theassumption that the variance of the weldvoltage and weld current decreasedwhen the welding process deviated fromthe optimal welding conditions. Parame-ters employed in monitoring short arc

(A)

(B)

(C)

(D)

(A)

(B)

(C)

(D)

Fig. 10 — T-joint with step disturbance No. 1. A — Mean of the weldvoltage (m[i]); B — estimated variance of weld voltage (y[i]); C —mean of the weld current (mi[i]); D — estimated variance of weld cur-rent (yi[i]).

Fig. 11 — T-joint with step disturbance No. 1. Standard deviation of:A — Short-circuit time (σs); B — arc time (σa) as a function of position.Reference T-joint. Standard deviation of: C — short circuit time (σs); D —arc time (σa) as a function of position.


RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

GMAW were investigated, e.g., arc time,Ta; short-circuit time, Ts; short-circuitpeak current, Ip; voltage during arc time,Ua; and voltage during short-circuit time,Us. These variables are key parameters inGMAW and can be estimated wheneverthe short-circuit time exceeds 1.0 ms —Figs. 8–11. Note that short-term short-circuits, i.e., those that do not exceed1.0 ms, can result in non-transference ofthe molten metal (Refs. 27, 30–32).

The figures that illustrate the esti-mated short-circuit rate are, however, es-timated with the number of short-timeshort-circuits included. This was done tofacilitate visual interpretation of thepower spectral diagrams — Fig. 12A–D.

To obtain the overall trend for theestimated features, a sliding median filterof length N = 100 was applied to eachparameter sequence. The median filterreplaced the center value in the window,with the median value of all sampleswithin the window. For N even, xM(k) isthe median of

The output of the median filter is de-noted by a superscript M, e.g., when amedian-filter is applied to the arc time,Ta, the resulting sequence is called Ta

M —Figs. 8–10.

Observation

Two examples of actual recordingsof the weld voltages and currents areshown in Fig. 7. Figure 7A and B showsthe results of a normal weld; 7C and Dshow the result of a weld during step dis-turbance. In Fig. 7A and C, the metaltransfers are reflected in the weld voltageas almost zero voltage events of 2 ms.Note that during the step disturbance, thetotal cycle time, T, has increased, ascompared with a normal weld. The arctime Ta, short-circuit time Ts and short-circuit peak current Ip have also in-creased during step disturbance welding,as compared with normal welds, whilethe short-circuit transfer rate and mean

x k x nM

n k N

k N

( ) [ ]./

/

== −

+ −

∑med (2)2

2 1

(A)

(B)

(A)

(B)

Fig. 12 — Power spectral densities of: A — The weld voltage and weldcurrent from a reference T-joint. Power spectral densities of: B — the weldvoltage and weld current during step disturbance. The dotted curve rep-resents 95% confidence limit. Note that the frequency of the spectralmaximum peak decreases from about 80 Hz during normal welding toabout 50 Hz during step disturbance.

Fig. 13 — Application of high-pass-filter (70 Hz) to: A — Weld voltage;B — current from welding a T-joint with step disturbance. DC and fre-quency component of the weld voltage and current below 70 Hz is re-moved and frequency component above 70 Hz is passing through thefilter. Note the difference in the waveforms between the non-filtered andthe filtered weld voltage and current. (Compare to Fig. 6.)

voltage during arc time, Uwa, have de-creased — Figs. 8–10.

A possible physical explanation ofthese phenomena is as follows: When theweld voltage is too low during arc time,the energy in the arc is not sufficient tomelt the electrode and form the dropletthat is required to complete the weldcycle — Fig. 8B and C. When this hap-pens, an excessive amount of time mustbe spent in the short-circuit phase to gen-erate the heat necessary to melt theelectrode and release the droplet. Con-sequently, when the short-circuit time in-creases, the short-circuit peak currentalso increases — Fig. 9B and D. Thesefigures also suggest that a larger dropletis detached from the electrode for eachshort-circuit cycle, to compensate for thedecreasing short-circuit metal transfer

rate during step dis-turbance tobalance thewire melting ratewith the wire feedrate.

The weld voltageduring arc time isapproximately 2 Vlower than theopen circuit volt-age (Ref. 33). Theexact value issystem dependent.Since the weldvoltage during arctime decreasesduring the stepdisturbance, thewelding process isassumed to beoperating in thevicinity of A,i.e., the “stubbing-in” metal transfermode in Fig. 3.

Mean and Variance ofthe Weld Voltageand Current

As discussedabove, the vari-ance of the weldvoltage may be asuitable parameterfor detection ofchanges in theweld quality.

The weld voltageis divided into I sec-tions, with N =1024 samples ineach section. Thevariance is calcu-lated for each sec-tion and given anindex, i, defined by

the position in the sequence. The vari-ance is estimated as follows:

where u[n] is the weld voltage, N isthe number of data points and m[i] is themean of the weld voltage in section i,calculated as

Figures 10A and B show the result ofthe estimated mean, m[i] and variance,y[i] of the weld voltage amplitude taken

from a T-joint with step disturbance.Calculations of mi[i] and yi[i] of the weldcurrent amplitudes are shown in Fig. 10Cand D.

From these diagrams, the followingconclusions can be drawn:

1) There is a decrease in the esti-mated variance of the weld voltage andno change in the mean weld voltage dur-ing step disturbance — Fig. 10A and B.This supports the assumption that theshort-circuit transfer rate has decreased.Since the short-circuit transfer rate hasalso decreased, non-optimal weldingconditions can therefore be assumed —Figs. 3 and 8A.

2) The variance y[i] at the beginningof the welding pass is considerable. Thisis due to the fact that the process is notstabilized, which leads in turn to thenumerous high-voltage transients.

3) Unlike the estimated variance ofthe weld voltage during step disturbance,no decrease in the estimated variance ofweld current yi[i] is observed — Fig. 10C.

Standard Deviation of Short-Circuit Timeand Arc Time

The standard deviations of the arc,short-circuit time, short-circuit peak cur-rent and short-circuit frequency haveoften been used as indicators of thestability and regularity of the weldingprocess (Refs. 4–6, 34, 35).

In this study, only the standard devia-tion of the arc and short-circuit time werecalculated since preliminary resultsshowed that standard deviations inpeak current and short-circuit frequencyyielded no new information. In otherwords, the behavior of the standard de-viation during step disturbance wasapproximately identical for the fourparameters.

The standard deviation was calculatedas follows: Short-circuit time and arc timewere divided into sections, with ten ob-servations in each section (ten observa-tions of the short-circuit time and arc timecorresponded to a weld joint length of ap-proximately 1.0 mm). The standard devi-ation was then calculated for eachsection. The results can be seen in Fig. 11.

Spectral Domain Analysisof Measurement Data

Since variance is an AC power esti-mate (the area below the curve of thepower spectral density), spectra from therecordings of normal welding conditionscan be compared with spectra of therecordings taken during step disturbancewhen searching for relevant characteris-tics (Ref. 36).

The results of the power estimation of

m iN

u n n Ni I

n i N

i N

[ ] [ ] , ,, ,

( )

= ==

= − ⋅

⋅

∑1 1 21 2

1

(4)KK

y iN

u n m i

n Ni I

n i N

i N

[ ] ( [ ] [ ])

, ,, ,

( )

=−

−

==

= − ⋅

⋅

∑11

1 21 2

2

1

(3)KK


RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

(A)

(B)

Fig. 14 — Estimated variance of the filtered: A — Weld voltage; B —weld current. Note the similarity in the appearance of the two wave-forms. The filtered current shows a decrease in the estimated variance,yif[i], while the non-filtered variance estimate of the weld current, yi[i],does not decrease during step disturbance. (Compare to Fig. 10D.)

the weld voltageand current duringboth normal weld-ing and step distur-bance is shown inFig. 12A–D. A vi-sual inspection ofthe four powerspectral densitiesshowed that themaximum spectralpeak during nor-mal welding wasabout 80 Hz, com-pared with themaximum spectralpeak during stepdisturbance, ap-proximately 50 Hz.

The fundamentalfrequency (80 Hz)

of the power spectral density of the weldvoltage during normal welding did notseem to be in agreement with the short-cir-cuit transfer rate curve during a normalweld, as shown by a comparison of Figs.8A and 12A and B. The fundamental fre-quency (80 Hz) differed from the numberof short-circuits (approximately 110/s). Thedifference may have been due to the factthat short-term short-circuits wereincluded in the total number of shortcircuits. The maximum spectral peak inFig. 12A and B represents the true short-circuit rate, where metal is transferred fromthe electrode tip to the workpiece. The dif-ference between the two results mayrepresent the number of short-term short-circuits/s, estimated as 110 – 80 = 30.

As stated above, a decrease in thevariance was reflected in a decrease inthe area in the power spectral density.


RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

(A)

(B)

(C)

(D)

(E)

Fig. 15 — T-joint with step disturbance No. 2. A — Measured short-cir-cuit transfer rate; B — estimated mean arc voltage (

–Uwa); C — median fil-

ter of length 100 applied to the estimated mean arc voltage (–Uwa

M); D —arc time (Ta); E — median filter of length 100 applied to the arc time se-quence (Ta

M).

(A)

(B)

(C)

(D)

Fig. 16 — T-joint with step disturbance No. 2. A — Measured short-cir-cuit time (Ts); B — median filter of length 100 applied to the estimatedshort-circuit time in part A (Ts

M); C — measured short-circuit currentpeak (Ip); D — median filter of length 100 applied to the short-circuitcurrent peak sequence in part C (Ip

M).

However, it is not clear in which fre-quency band the fall in power occurred.Compare Fig. 12A and C. Further inves-tigations must be undertaken.

Filtering the Data

The main difference in the powerspectra of the weld voltage and currentwas the decrease in the frequency ofthe spectral maximum peak from 80 Hzduring normal welding to about 50 Hzduring step disturbance. To enhance thedifference between normal welds andwelds during step disturbance, the weldvoltage and current were high-pass-filtered with an elliptic discrete-time fil-ter. The cutoff frequency filter was set atf2 = 70 Hz.

To avoid any phase distortion of theoutput, the phase-shift of the filter shouldbe zero. One technique for achieving thisis to process the data forward and thenbackward through the same filter (Ref.37). is shown in fig. 13A and B. The resultof applying the high-pass filter to the weldvoltage and weld current in this way.

Variance of Filtered Data

The variances in the amplitude of thehigh-pass-filtered weld voltage and cur-rent may be suitable parameters for de-tection of changes in the weld quality.

The high-pass-filtered weld voltageand current are divided into N sections,with 1024 samples in each section. De-noting the section of the filtered voltageuf[i] and section of the filtered weld cur-rent cf[i] where i = 1, …, I, for each sec-tion the variance of the weld voltage yf[i]and current yif[i] can be calculated.

Figure 14A depicts the result of yf[i]taken from a T-joint with step distur-bance. Note the decrease in the esti-mated variance of the weld voltage, yf[i]during step disturbance, indicating non-optimal welding conditions.

The same algorithm was also appliedto the filtered weld current to obtain anestimate of the variance in filtered weldcurrent amplitude. The estimated vari-ance for weld current yif[i] is shown inFig. 14B. Note that the filtered currentshows a decrease in the estimated vari-ance yif[i] while the non-filtered varianceestimate of the weld current yi[i] does notdecrease during step disturbance. Com-pare Figs. 10D and 14B.

Note also the similarity in the appear-ance of the two waveforms between theestimated variance in the filtered weldvoltage and the current.

Selected Test Parameter for Monitoring ShortArc GMAW

The observations described above

are typical of the welds, thoughdeviations from normal behavior canoccur. The normal pattern for a T-jointwith step disturbance is shown in Figs.5–12 as follows:

• a decrease in short-circuit rate andarc voltage

• an increase in arc time, short-circuit time and short-circuit cur-rent peak

• no increase or decrease in themean of the weld voltage andcurrent

• a variance in weld voltage de-crease, though there is no variancein weld current

• increase in standard deviation ofarc and short-circuit time.

A probable explanation of the aboveobservations is discussed below.

Examples of deviations from the nor-mal pattern for a T-joint with step dis-turbance are shown in Figs. 15–19 asfollows:

• an increase in short-circuit rateand no change in the amplitude ofarc voltage

• no increase or decrease in arc timeand short-circuit time, but a de-crease in short-circuit current peak

• no increase or decrease in themean of the weld voltage and cur-rent

• variance of the weld voltage de-creases, while the variance of theweld current remains unchanged

• no increase or decrease in stan-dard deviation of arc and short-cir-cuit time.

The above observations indicate thatoptimal process stability can also occurduring step disturbance (the short-circuitrates increase and the standard deviationof arc and short-circuit time remain con-stant). Note also that the variance of theweld voltage decreases, despite the factthe short-circuit rates increase duringstep disturbance. (Compare Figs. 16Band 17A.) Hence, it is still possible todetect step disturbance in the weld joint,even where no optimal process stabilityhas occurred.

The observations in this sectionsuggest that a detection algorithm thatuses the variance of the weld voltage orfiltered-weld voltage and current showspromise due to its robustness. The vari-ance of the non-filtered weld voltage ischosen as a suitable parameter for the de-tection of deviations from optimal weld-ing conditions. The high-pass-filtered weldvoltage and current are a little more dif-ficult to obtain, and yield no significantimprovement in the detection of the rela-

tive decrease in the amplitude level of thevariance during step disturbance. Figures10B and 14A confirm this impression.

Fault Detection Algorithm

Test Parameter

The variance of the weld voltage waschosen as the test parameter for detectionof changes in weld quality. When thevariance of the weld voltage is larger thanthe variance during normal conditions,spatter or other severe weld defects giv-ing rise to large voltage transients haveoccurred. When the variance of the weldvoltage is less than the variance duringnormal conditions, the welding processhas been disturbed. To avoid any confu-sion in terms, the variance of the weldvoltage, y[i] is denoted “AC power.”

The AC power is estimated as de-scribed in Equation 3 and is shown inFigs. 10 and 17B. Note the decrease inmean of the AC power estimate y[i] dur-ing step disturbance, indicating non-op-timal welding conditions.

Algorithm

The algorithm presented below is de-signed to detect sudden small changes inthe monitored parameter.

The AC power sequence y = (y[0],y[1], … y[k]) is assumed to be identical,independent and Gaussian distributed asfollows (Ref. 38):

The welding process is known to op-erate under either optimal (θ = m0) ornon-optimal (θ = m1) welding conditionswhere m0 > m1. Furthermore, it is as-sumed that prior to t = 0, θ = m0 and mayonly change to θ = m1 at one of the i sam-pling instants. Consider the problem oftesting k + 1 hypotheses H0, H1, … Hk,where Hj is defined as

and Hk is the null hypothesis. If the in-stant of change j is fixed, then the se-quential probability ration test (SPRT)between Hj and Hk is based on a com-parison of the likelihood ratio (Refs. 22,23, 39)

Λjk

i j

ki=

=∑λ[ ] (7)

H m i j

m j i kj : θ

θ= ≤ ≤ −= ≤ ≤

0

1

0 1for

for (6)

p y i eT

y i

Tθ

θσ

σ π( [ ])

( [ ] )

=− −

1

2

2

22 (5)


RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

with

to a threshold h. At the sampling instantk, Λj

k is computed. If Λjk ≥ h, then a defect

in the weld joint is detected. In the scalarindependent case, Λj

k is recursively up-dated as

In the case of a change in the meanvalue m of an independent Gaussian ran-dom sequence y[i] with known varianceσT

2, the sufficient statistics λ[i] are calcu-lated as (Ref. 22)

which is written as

where

is the change in magnitude. The SPRT isoptimal with respect to the worst meandelay, when the error probability for falsealarms goes to zero. The instant ofchange j is, in fact, unknown. This may,however, be estimated using the maxi-mum likelihood principle (Ref. 40), lead-ing to the decision function and alarminstant

The algorithm has been formulated asa set of parallel SPRTs, but may equallybe viewed as a repeated SPRT or a cu-mulative sum (CUSUM) type test. Theconnection between these alternativepoints of view has been investigated (Ref.22). The decision function g[k] intro-duced in Equation 13 becomes in re-peated SPRT formulation

and in the Gaussian case

where (x)+ = sup(0, x). The alarm thresh-old h is chosen by a tradeoff betweenworst mean delay time for detection τand false alarm probability α. TheCUSUM algorithm (Ref. 24) is optimalwhen α goes to zero:

where

is the Kullback information. In the Gaus-sian case, the Kullback information is

Due to Wald’s equality, the probabilityof false alarms α and the alarm thresholdh satisfies following equation:

when the probability of non-detectiongoes to zero. The alarm threshold h istherefore easy to obtain for fixed α (Ref.41). The complete fault detection algo-rithm may be summarized as follows:

For each section k of 1024 datasamples:

1) calculate AC power y[k]2) calculate g[k] = [g[k – 1] + λ[k]]3) if g[k] ≤ 0 then g[k] = 04) if g[k] ≥ h then set Alarm.

Estimation of the Mean and Varianceof the AC Power Parameter

The AC power of weld voltage y[i] inEquation 10 is assumed to be identical,independent and Gaussian-distributed,with a mean value m0 and m1 undernormal and fault welding conditions, re-spectively. The variance σT

2 of the ACpower is assumed to be constant, duringboth normal weld and step disturbance.However, the mean and variance of theAC power are not known and must,therefore, be estimated. Since there areboth within-record variations (see Fig.21C) and between-record variations ofthe AC power (probably due to slightlydifferent welding conditions occurringduring the experiments when weldingthe different T-joints), the between-record variations are incorporated intothe total variance σT

2 — Fig. 20.

To estimate the mean value of the ACpower factor, the between- and within-record variance from five experiments(each with 32 observations originatingfrom weld voltages recorded during nor-mal conditions and welds during stepdisturbance) was used. The processmodel and estimation procedure of themean, within-record variations, be-tween-record variations and total vari-ance are described in Ref. 17. The resultsare given in Table 1. For this data, themean for normal welds is m0 = 56.60 andduring step disturbance,m1 = 47.56. Theestimated variances between record are^σR

2 = 0.66 and ^σR2 = 6.19, respectively.

The estimated variance within the recordfor a normal weld is ^σ I

2 = 6.26 and ^σ I2 =

8.84 during step disturbance. The totalvariance for the data is ^σT

2 = 6.92 and ^σT2

= 14.75, respectively.In the experiment, the total variance

σT2 was set at 6.92. Due to an underesti-

mation of the total variance during thestep disturbance, the worst mean delaytime for detection τ will increase (Refs.42, 43).

Tuning

In the proposed algorithm, the onlytuning parameter is the threshold h (Table2). Using Equation 20, we computed theworst mean delay for detection τ andchose a false alarm probability α. Thesemay then be used to determine a relevantalarm threshold, h.

If the false alarm probability α is set at10–6 and assuming that the AC power se-quence y[i] is Gaussian and statisticallyindependent, the alarm threshold h iscalculated to be h = 13.8. The reverse ar-rangement and the χ2 test was applied tothe AC power to test whether or not theAC power was Gaussian and statisticallyindependent (Ref. 44). The outcome ofthese tests — not included in this paper— shows the AC power is likely to be sta-tistically independent, but non-Gaussianduring both normal and step disturbancewelding (Ref. 17). Since the AC power se-quence y[i] cannot be assumed to beGaussian, the alarm threshold h is set at16 to maintain the false alarm probabil-ity, α ≥ 10–6. In real industrial applica-tions, it is recommended that the welderin charge has the option of changing thealarm threshold in accordance with thetype of welding mode and application.

Test of the Repeated SPRT Algorithm

The recursive SPRT algorithm wastested on 31 specimens. A total of 15 ex-periments were conducted for reference

α = −e h (20)

K m mm m

T

( , )( )

.1 01 0

2

22= −

⋅σ(19)

K m m Ep y i

p y imm

m

( , ) ln( [ ])

( [ ])1 0 1

1

0

=

(18)

τ α α~ln( , )

−

→1

1 00

K m mwhen (17)

g k g k m y kT

[ ] [ ] [ ]= − + − −

+

122 0

υσ

υ(16)

g k g k k[ ] [ [ ] [ ]]= − + +1 λ (15)

t k g k ha = ≥min : [ ] (14)

g kj k j

k[ ] max=≤ ≤0

Λ (13)

υ = −( )m m0 1 (12)

λ υσ

υ[ ] [ ]i m y i

T= − −

2 0 2

(11)

λσ

[ ] [ ]im m m m

y iT

= − + −

0 12

0 1

2(10)

Λ Λjk

jk k+ = + +1 1λ [ ]. (9)

λ[ ] ln( [ ])( [ ])

ip y ip y i

m

m= 1

0(8)


RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT


RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

(A)

(B)

(C)

(D)

Fig. 17 — T-joint with step disturbance No. 2. A — Mean of the weldvoltage (m[i]); B — estimated variance of the weld voltage (y[i]); C —mean of the weld current (mi[i]); D — estimated variance of the weld cur-rent (yi[i]).

(A)

(B)

Fig. 18 — T-joint with step disturbance No. 2. Standard deviation of: A— Short-circuit time (σs); B — arc time (σa) as a function of position.

(A)

(B)

Fig. 20 — The between-record variation of the AC power σR2. This vari-

ation is probably due to slightly different welding conditions between theexperiments, while the within-record variation σI

2 is associated with fluc-tuations along the weld joint during an experiment. The between-recordvariation is exaggerated in the diagram for illustrative purposes.

Fig. 19 — T-joint with step disturbance No. 2. Power spectral densitiesof: A — The weld voltage; B — weld current during step disturbance. Thedotted curve represents the 95% confidence limit.

T-joints and 16 experiments were con-ducted for T-joints with step disturbance.The recording time of the measured sig-nals was 15 s.

The test was designed as follows: Whenthe alarm turns on and there is a step dis-turbance, the test results in a detection. Ifthe alarm turns on and there is no step dis-turbance, the result is a false alarm.

Two different changes in magnitude,u = m0 – m1 and u = 1/2sT, were used.The choice of u was dictated by theexperimental results in Table 1. The firstchoice of u is the estimated change in-magnitude; the second choice is theminimum change in magnitude. Theminimum change in magnitude is boundto be positive. It is known, however, thatthe SPRT algorithm is quicker at detect-ing a magnitude of change betweenabout 1/2sT and 3/2sT from the targetvalue m0, as reflected in the Shewhartchart (Refs. 22, 39, 45).

Since the welding process was notworking under optimal welding condi-tions at the start and end of the weldingpass, the alarm was inhibited during thefirst and last centimeter of the weldingpass.

Test Results

The results of the test shown in Tables

3 and 4 indicate that it is possible todetect changes in the weld quality auto-matically and on-line. To illustrate the re-sults of the data, the test result for a T-jointwith a step disturbance is shown in Figs.21 and 22.

Discussion

All the tests showed consistent results:during a step disturbance, the AC powerof the weld voltage decreased. In the ex-periments, the decrease in the AC powerduring step disturbance welding was thesame for the two different brands of weld-ing equipment used in the experiments.It is likely, therefore, that other equip-ment of different brands will also exhibitthe same behavior. A suitable change inmagnitude u and target value m0 must beadjusted for each piece of equipment andfor each type of welding.

The proposed SPRT algorithm was de-signed to detect sudden changes in theaverage level of the AC power. Four-stepdisturbances were not detected, how-ever, when the change in magnitude υwas set at 9. These non-detected step dis-turbances occurred on the same day. Aspecial physical cause could not befound to account for this variation; there-fore, these records cannot be excludedfrom the tests. A probable explanation isthat the experimental conditions on theday in question were not identical to theother three. Nevertheless, the recordingsfrom that day all show a decrease in ACpower during step disturbance, but theAC power is unusually large — both be-fore as well as during step disturbance.(Compare Figs. 21C and 23.) Since theminimum value that the SPRT algorithmcan detect is m0 – υ/2 and the AC power

value during step disturbance is abovethis critical value, these two factors com-bined explain the lack of detection of thestep disturbance.

In industrial welding applications, it issometimes necessary to weld withmixed-mode transfer to increase thewelding speed, thereby increasing pro-ductivity. The mix-mode transfer, whichcontains a mixture of short-circuiting,globular and spray transfer, is related to aworking point in the vicinity of C in Fig.3. In this case, it is relevant to use twoSPRT algorithms together: the first for de-tecting an increase in the mean of the ACpower, and the second for detecting a de-crease in the mean of the AC power. Anew target value m0, which correspondsto the new working point, must be esti-mated together with a robust tuning ofthe minimum magnitude of change interms of the Kullback information and of


RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

Table 1 — Parameter Estimates During Normal Welds and During Step Disturbance

Estimated Parameters

During Normal Weld During Step Disturbance

Estimated within-record variance 6.26 8.84

Estimated between-record variance 0.66 5.91

Estimated total variance 6.92 14.75

Estimated overall mean 56.60 47.56

Table 2 — Design Parameters Selected for the Fault Detection Algorithm for Two DifferentValues of the Change Magnitude, υ = 9 and υ = 2.63, Respectively (Welding Speed is Set at 10mm/s)

Design Parameters

Test 1 Test 2υ Magnitude of change 9 2.63

h Alarm threshold 16 16α False alarm probability 10–6 10–6

h Alarm threshold 16 16Total variance 6.92 6.92

m0 Mean during normal condition 56.60 56.60m1 Mean during non-optimal condition 47.56 53.97

σT2

m

σT2

σR2

σI2

Table 3 — The Experimental Results ofTest 1 of the SPRT Algorithm (Magnitude ofChange υ, is set at 9 and Welding Speed isSet at 10 mm/s)

Results

Type of T-Joint Reference Step

Number of specimens 15 16Detection 0 12Non-detection 0 4False alarm 0 0

Table 4 — The Experimental Results ofTest 2 of the SPRT Algorithm (Magnitude ofChange υ is set at 2.63 and Welding Speedis Set at 10 mm/s)

Results

Type of T-Joint Reference Step

Number of specimens 15 16Detection 0 15Non-detection 0 1False alarm 0 0

the threshold. The task of the SPRT algo-rithm is, therefore, to detect when thewelding process deviates from the work-ing point.

The decrease in the maximum of theshort-circuit frequency during step dis-turbance may be physically explained bythe oscillation behavior of the weld poolas follows: As the oscillation frequency ofthe weld pool decreases with increasingweld pool width, the maximum short-circuit frequency must also decrease(Refs. 6, 20, 21). However, this explana-tion is not confirmed by experimentaldata. The photo of the T-joint in Fig. 5shows that during step disturbance, theweld joint tapers. Consequently, theshort-circuit frequency should increaseinstead of decrease.

In some cases, the AC power presents

large transients inwaveform behav-ior — Fig. 24. Acloser examinationof the weld jointshowed that a severe defect or spatterwas generated at the corresponding po-sition on the weld joint. At the pointwhen the molten electrode tip was incontact with the workpiece, a small neckbetween the electrode and the weld pooldeveloped, due to the surface tension. Asa result of this small neck, the currentdensity increased. When the current den-sity in the neck becomes too large, an ex-plosion occurs and a droplet forms. Thisexplosion of the metal bridge gives rise tothe high-voltage transients. In this case,the two-sided detection algorithm alsoseems to be a suitable means of detect-

ing spatter and other severe defects.To improve the performance of the

proposed detection algorithm, other pa-rameters such as short-circuit time andarc time can be incorporated into a com-posite SPRT detection algorithm (Ref.22). The difficulty with the compositeSPRT detection algorithm is that it givesno indication as to which parameter orparameters are causing the problem. Thiscan be dealt with by using an SPRT algo-rithm for each parameter, which mentorsand simply takes action as soon as thefirst alarm signal occurs.

Detectable differences were found inthe power spectral densities of the weld


RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

(A)

(B)

(C)

(D)

Fig. 21 — Illustration of the detection of step disturbance No. 1. A —Measured weld voltage; B — weld current; C — the corresponding ACpower y[i] and the actual position of the step disturbance; D — the be-havior of the decision function g[i] and Alarm time. The Alarm time is in-dicated by a vertical line in the figure. The magnitude of change is set at9.2 and the welding speed at 10 mm/s.

(A)

(B)

Fig. 22 — T-joint with step disturbance No. 1. A — The AC power y[i]and the actual position of the step disturbance; B — the behavior of thedecision function g[i] and Alarm time. The Alarm time is indicated by avertical line in the figure. The magnitude of change is set at 2.63 and thewelding speed is set at 10 mm/s. (Compare with Figs. 21C and D.)

Fig. 23 — The unusually large AC power value before and during stepdisturbance as compared with a typical T-joint with step disturbance.(Compare with Fig. 21C.) As a consequence of the large AC value dur-ing step disturbance, no alarm is indicated in the diagram.

voltage and current between a normalweld and welds during step disturbance(Ref. 36). The maximum spectral peakduring normal welding is about 80 Hz asopposed to the maximum spectral peakduring step disturbance, which is ap-proximately 55 Hz. A detection algo-rithm based on the detection of changesin the fundamental frequencies, togetherwith detection of changes in the ampli-tude in the fundamental frequenciesshows promise. However, preliminaryresults described above (Ref. 36) haveshown that the variance of the weld volt-age is the most robust parameter for de-tecting changes in the weld quality.

Acknowledgments

The authors wish to express theirgratitude to the Industrial DevelopmentCentre of Olofström and its personnel fortheir support in arranging and assistingwith the experimental measurements.We also wish to acknowledge the finan-cial support given by The Foundation forKnowledge and Competence Develop-ment (KKS).

References

1. Cary, H. B. 1994. Modern WeldingTechnology. Englewood Cliffs, N.J.Regents/Prentice-Hall.

2. Amin, M. 1983. Pulse current parame-ters for arc stability and controlled metal trans-fer in arc welding. Metal Construction 15(5):272–278.

3. Cornu, J., and Weston, J. 1995.Advanced Welding System, Vol. 2, London,England, IFS.

4. Mite, T., Sakabe, A., and Yokoo, T. 1985.The estimation of arc stability of CO2 gasshielded arc welding. Proceedings of the

Welding InstituteConference: Ad-vanced Welding Sys-tems, pp. 261– 271.

5. Gupta, S. R.,Gupta, P. C., andRehfeldt, D. 1988.Process stability andspatter generationduring dip transferin MAG welding.Welding Review, pp.232–241.

6. Hermans, M. J.M., and den Ouden,G. 1997. A stabilitycriterion for shortcircuiting gas metalare welding. Pro-ceedings of the Inter-national Conferenceon the Joining of Ma-terials, Helsingør, pp.112–119.

7. Jennings, C.1951. Dynamic char-acteristics of DC

welding machines. Welding Journal 30(2):117–138.

8. Popkov, A., et al. 1977. Reducing thespatter of liquid metal in CO2 welding bymeans of optimisation of the welding param-eters. Welding Production 3: 26–27.

9. Leino, K., Nikkola, A., and Varti-ainen, K. 1984. Prediction of weld defectsusing welding condition data. Research report264, Technical Research Center of Finland.

10. Adam, G., and Siewert, T. A. 1990.Sensing of GMAW droplet transfer modesusing an ER 100S-l electrode. Welding Journal69(3): 103–108.

11. Johnson, J. A., Carlson, N. M., Smartt,H. B., and Clark, D. E. 1991. Process controlof GMAW: Sensing of metal transfer mode.Welding Journal 70(4): 91–99.

12. Wang, W., Liu, S., and Jones, J. E.1995. Flux cored arc welding: Arc signals,processing and metal transfer characteriza-tion. Welding Journal 74(11): 369– 377.

13. Ogunbiyi, B., and Norrish, A. 1996.GMAW metal transfer and arc stability assess-ment using monitoring indices. Proceedings ofthe 6th International Conference on ComputerTechnology in Welding, TWI, Cambridge,England, Abington Publishing.

14. Adolfsson, S., Ericson, K., and Agren,B. 1995. On automatic detection of burn-through in GMA welding — weld voltageanalysis. Research report TULEA 1995:15,Sweden, Division of Signal Processing, LuleåUniversity of Technology.

15. Adolfsson, S., Ericson, K., andGrennberg, A. 1996. Automatic detection ofburn-through in GMA welding using a para-metric model. Mechanical Systems and SignalProcessing 10(5): 633–651.

16. Adolfsson, S., Bahrami, A., and Claes-son, I. 1996. Quality monitoring in robotisedwelding using sequential probability ratio test.Proceedings of TENCO ’96, Digital Signal Pro-cessing Applications, Vol. 2, New York, N.Y.,IEEE. pp. 635–640.

17. Adolfsson, S., Bahrami, A., Bolmsjö,G., and Claesson, I. 1997. Automatic quality

monitoring in robotised GMA welding using arepeated sequential probability ratio testmethod. International Journal for the Joining ofMaterials 9(1): 2–8.

18. Blakeley, P. J. 1992. Developments inmonitoring systems for resistance and arcwelding. Proceedings of the InternationalConference on Automated Welding Systemsin Manufacturing, Paper 40. Gateshead, Eng-land, Woodhead Publishing, Ltd.

19. Agren, B. 1995, Sensor integration forrobotic arc welding. Ph.D. thesis, Sweden,Lund University..

20. Hermans, M. J. M., Spikes, M. P., andden Ouden, G. 1993. Characteristic featuresof the short circuiting arc welding process.Welding Review International 12(2): 80–86.

21. Rehfeldt, D., and Schmitz, T. 1995. In-vestigation and measurement of weld pool os-cillation in GMAW. Technical report, IIW-doc.212-882-95.

22. Basseville, M., and Nikiforov, I.V.1993. Detection of Abrupt Charges: Theoryand Application. Englewood Cliffs, N.J., Pren-tice-Hall.

23. Grainger, R. W., Hoist, J., Isaksson, A.J., and Ninness, B. M. 1994. A parameticstatistical approach to fdi for the industrialactuator benchmark. Research reportLUTFD2/TFMS — 3106 — SE LUTFD2/TFMS-3106-SE, Sweden, Dept. of MathematicalStatistics, Lund Institute of Technology.

24. Lorden, G. 1971. Procedures for re-acting to a change in distribution. Ann. Math.Statistics 42: 1897–1908.

25. Allum, C. J. 1985. Metal transfer in arcwelding as a varicose instability: I. Varicose in-stabilities in a current-carrying liquid cylinderwith surface charge. British Journal of AppliedPhysics 18(7): 1447–1468.

26. Amin, M. 1981. Synergic pulse MIGwelding. Metal Construction 13(6): 349– 353.

27. Smith, A. A. 1966. Characteristics ofthe short-circuiting CO2-shielded arc. Physicsof the Welding Arc, Institute of Welding, pp.75–91.

28. Lancaster, J. F., ed. 1986. The Physicsof Welding. Oxford, England, Pergamon Press .

29. Kim, J. A., and Eager, N. M. 1991.Analysis of metal transfer in gas metal arcwelding. Welding Journal 70(6): 91–99.

30. Mita, T., Sakabe, A., and Yokoo, T.1988. Quantitative estimates of arc stabilityfor CO2 gas shielded arc welding. Welding In-ternational 12(2): 152–159.

31. Piuchuk, I. S., et al. 1980. Stabiliza-tion of transfer and methods of reducing thespattering of metal in CO2 welding with shortarc. Automatic Welding 6.

32. Shinoda, T, Nishikawa, H., andShimizu, T. 1996. The development of dataprocessing algorithms and assessment of arcstability as affected by the titanium content ofGMAW wires during metal transfer. Proceed-ings of the 6th International Conference onComputer Technology in Welding, TWI, Cam-bridge, England, Abington Publishing, .

33. Lundin, R., and Widfeldt, M. 1988.Mätning, registrering och övervakning avsvetsdata. Technical report 264, IVF.

34. Liu, S., and Siewert, T. 1989. Metaltransfer in gas metal arc welding: droplet rate.Welding Journal 68(2): 52–58.


RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

Fig. 24 — An example of the AC power waveform when spatter occurs.High voltage transients give rise to high AC power transients at 13 cm inthe diagram.

35. Lucas, W. 1985. Computers in arcwelding — the next industrial revolution, part3: instrumentation and process analysis. MetalConstruction 17(7): 431– 436.

36. Adolfsson, S., Bahrami, A., Bolmsjö,G., and Claesson, I. 1997. Quality monitoringin robotized short circuiting GMA welding.Research report HK-R 1997:3, Sweden, De-partment of Signal Processing, University ofKarlskrona/ Ronneby.

37. Oppenheim, A. V., and Schafer, R. W.1989. Discrete-Time Signal Processing. Engle-wood Cliffs, N.J., Prentice-Hall.

38. Papouis, A. 1984. Probability, Ran-dom Variables, and Stochastic Processes. NewYork, N.Y., McGraw-Hill.

39. Montgomery, D. C. 1985. Introduc-tion to Statistical Quality Control. New York,N.Y., J. Wiley & Sons.

40. Rao, C. R. 1973. Linear Statistical In-ference and Its Applications. New York, N.Y.J.,Wiley & Sons.

41. Wald, A. 1947. Sequential Analysis.New York, N.Y., J. Wiley & Sons.

42. Bagshaw, M., and Johnson, R. A.1975. The effect of serial correlation on theperformance of CUSUM-tests part II. Techno-metrics 17(1): 73–80.

43. Bagshaw, M., and Johnson, R. A.1975. The influence of reference values andestimated variance on the ARL of the CUSUMtests. Jal Royal Statistics Society, 37(B) (3):413–420.

44. Bendat, J. S., and Piersol, A. G. 1986.Random Data: Analysis and MeasurementProcedures. New York, N.Y., J. Wiley & Sons.

45. Wetherill, G. B., and Brown, D. W.1991. Statistical Process Control. England,Chapman and Hall, London.

List of Symbols

α False alarm probabilityfs Sampling rate (Hz)g[i] Decision functionh Threshold for alarmi, j, k, l, m, n IntegersI Total number of sections within a

recordI Source current (A)

Ip Short-circuit peak current (A)j Record index, instant of changeK(m1, m0) Kullback informationλ[k] Increment of Λk

j

Λjk Log-likelihood ratio for observation

from y[j] until y[k]Li Internal inductance of the power

source (H)la Length of arc (mm)le Length of wire electrode stickout(mm)m Overall mean for the AC powerm[i] Mean value, weld voltage, section

i (V)mRMean value, AC power, weld voltage,

record rm0Mean value, AC power, weld voltage,

normal weldm1Mean value, AC power, weld voltage,

step disturbanceΝ (0, 1) Normal distribution function

with zero mean and unit vari-ance

PAC[i] AC power of the weld voltage ofsection i

PAC[r, i] AC power of the weld voltageof the section i of the record r

∆PAC,R[r] Variation between the records∆PAC,I[r, i] Variation within the recordpθ (y[i]) Probability density functionR Total number of recordsRi Internal resistance of the power

source (Ω)σa Standard variation of the arc timeσ A

2Variance of total number of reverse ar-rangements, A

σ I2 Within-record variance of the AC

power^σ I

2 Estimated within-record varianceof the AC Power

σ R2 Between-record variance of AC

power^σ R

2 Estimated between-record Vari-

ance of AC powerσs Standard variation of the short circuit

time

σ T2 Total variance, σ W

2 + σ B2, of the AC

power PAC[r, i]^σ T

2 Estimated total variance e, ^σ W2 + ^σ B

2,

of the AC power PAC[r, i]ta Alarm instantT Total cycle time, T = Ta + Tsh (s)Ta Arc time (s)Ts Short-circuit time (s)τ Mean delay time for detectionθ0 Parameter before change

θ1 Parameter after change

u[n] Weld voltage at the sampling in-stant (V)

Ua Arc voltage (V)Ue Wire electrode stickout voltage (V)Uoc Open circuit voltage (V)Up Peak voltage (V)Uw Weld voltage, Uw = Ue + Ua (V)Uwa Weld voltage during arc time (V)Uws Weld voltage during short-circuit

time (V)υ Change magnitudeWb Wire melting rate (mm/s)Wf Wire feed rate (mm/s)WsWelding speed (mm/s)y[i] Variance, weld voltage, section iy[i] AC power, weld voltage, section i


RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

/RE

SE

AR

CH

/DE

VE

LO

PM

EN

T/R

ES

EA

RC

H/D

EV

EL

OP

ME

NT

74-s| D

ECEM

BER

1997

RESEARCH/DEVELOPMENT/RESEARCH/DEVELOPMENT/RESEARCH/DEVELOPMENT/RESEARCH/DEVELOPMENT

On-Line Quality Monitoring in Short-Circuit Gas Metal Arc ...

Documents