ieee_application of design of experiment techniques to measurement procedures

8/3/2019 IEEE_Application of Design of Experiment Techniques to Measurement Procedures

http://slidepdf.com/reader/full/ieeeapplication-of-design-of-experiment-techniques-to-measurement-procedures 1/6

IEEE lnstumentation an d MeasurementTechnology ConferenceOttawa, Canada, May 19-21, 1997

Application of Design Of Experiment techniques to measurementprocedures. An exam le of optimisation applied to the digital

measurement of partial discharges.

R.Bozzo, G.Coletti, C.Gemme, F.GuastavinoElectrical Engineering Department

University of Genova- 16145 Genova - ITALYPhone +39 10 3532 725 Fax +39 10 3532 700 E-Mail: [email protected]

Abstract - In the last decade, multistage digitalmeasuring systems of partial discharges (PDs)have been introduced, allowing to support thediagnostic of defects (sites of PDs) in power

electric components. They transfer informationabout defects from full data sets of PD patterns,obtained from a Phase Resolved Partial DischargeAnalyser (PRPDA) to reduced data sets, byimplementing pattern recognition techniques. Thelatter data sets are then classified versus similarreference data sets, The validity of the abovediagnostic requires that the measuring process,which is influenced by several factors, is optimised.The three main settings of the PRPDA are amongsuch factors of influence, but so far, as a simplemathematical model of this measuring process isnot available, it has not been possible toquantitatively assess heir “weight” on the validity

of above diagnostic.This work presents a successful “Design OfExperiments” (DOE) approach to solve the latterproblem. The DOE analysis of the results o f 81 PD

tests performed on a simple physical model of aninsulation system quantified the weight and theinteraction of the three factors and allowed toderive criteria for selecting the “optimal” values ofsuch factors and the “optimal9’ composition of thereduced data sets.

I.INTRODUCTION.

An electrical discharge which affects only a part of aninsulating gap, without giving rise to an immediatebreakdown is named a partial discharge (PD) and canbe detected through non destructive techniques.The presence of PDs in the insulating system of apower component is very often associated to thepresence of defec ts, i.e. of sites where the massdensity is, or becomes, similar to the one of a gas andwhere the local electric field is higher than the local

dielectric strength. Usually in high voltage power (HVP)components the PD measurement is performed bymeans of electrical techniques, using standardizedanalog procedures [I] r non-standardized digital

approaches.As the named defects become potential causes of HV Pcomponents early failures in service (when the PDsshow an amplitude higher than a threshold level whichis specific to each type of component), the PDsmeasurement is usually part of the manufacturingquality evaluation plan of such components.The implementation of the latter measurementsthrough digital techniques (usually a Phase ResolvedPartial Discharges Analyzer - PRPDA) allows to storedata sets (PD patterns) which contain implicitinformation about the type and location of the relevantdefects (sites of PDs). Different advancedpreprocessing techniques (based on statistical

procedures or on expert system s or on N eural Networkapproaches, etc.) can be applied to such da ta in orde rto efficiently render explicit the above inform ation. Thedata recorded for implementing the above qualityevaluation procedures can be complemented with theoutput of the latter techniques application, to supportthe implementation of diagnostics aimed at identifyingthe causes of the HVP components defects. Fig.1

unknown se/ected factorsdefect of nf/uence

defect ypeand loccltion

current voltage PD reducedPulse I I I pa- dataset

Itage1 stage2 st-3 st-4

referencereduceddata set

Fig.1. Block diagram of the P D s measuring process supporting thediagnostic of high voltage power components

0-7803-3312-8/97/$5.0001997 EEE 470



illustrates the block diagram o f the m easuring process,from the PDs manifested by defects to the finaldiagnosis.While the solutions to the first two stages problems ofaccuracy are well known, the validity of the finalinferences generally depends also on other factors (as,for instance, the settings of the PRPDA), whoseinfluence has been proved, but could not, so far, bequantified.This work presents an application of "Design OfExperiment" (DOE) techniques which allowed toquan tify the PRPDA settings influence on the validity ofHVP components diagnostics, based on the PDpatterns ana lysis.

II . THE MEASURAND AND THE MEASUREMENTS

Basically [2], a P D can be considered a sudden transferof actual charge, which produces a fast current pulse,modifies the voltage across a partial capacitance (asmall region of an IS), propagates along the IS and canbe detected as a voltage pulse (the PD signal) acrossthe terminals of the power component. An unknowntimevariant transfer function F I , which depends also onthe p osition of the P D site, can be attributed to this firstsignal processing stage.The PD signal is then fed into a detection circuit (see asimplified version in Fig.2) and the relevant currenti(t)produces a voltage pulse Vz(t) across an impedanceZm. While a second transfer function F2 can beattributed to this stage and is "measurable" (through aso called "calibration* procedure), the peak of the latterpulse is proportional to the time integral of i(t), named"apparent charge" Q.Q can be shown to be independent of the pulse shape,is related to the original current pulses through F1 andF2, can be analogically measured with a degree ofuncertainty (and a traceability) acceptable for qualityassurance purposes and it represents the pulseamplitude.At the next stage the signal is processed through aPRDPA, where the voltage Vz(t) is amplified, sent to anA/D converter, quantized and stored, together with thequantized phase of the supply voltage at the instant ofthe pulse occurrence. At each cycle, under acconditions, each defect can originate several differentpulses, whose irregular distribution with time isconsidered to be related to the type of defect.Therefore, in order to acquire statistically significantdata, it is necessary to process in this way all the

pulses detected during a sufficiently long(>

about 100cycles) acquisition time .

HIGH VOLTAGE

One such PD acquisition leads, taking in due accountthe relevant problem s of accuracy, to the storage of thefollowing three basic quan tities, per each am plitude andphase cell (iJ), where i and j identify one of theavailable h amp litude and k phase windows:

NI = number of discharges stored in the -ij cell

Qi = i-th amplitude of the discharge occurred in the j-thphase window

qj = phase of occurrence of Q i

The complete set of such data acquired during a presetacquisition time is the PR pattern.

At this third stage takes place an information transfer,which is influenced by several factors (e.g. the PRDPAsettings, the specific power component, the testingconditions and/or unknown factors) which can alter[3,4], in an unknown way, also the contents,regardingtype and location of defects.

The PD pattern data set becomes an intermediatemeasurand, heuristic pattern recogn ition techniques areintroduced and the measuring process ends with thecomparison of an "unknown" PD pattern with"reference" PD patterns (obtained for differentcomponents having known defects, or for a similarcomponent under different conditions).

Therefore the validity of the measuring process isdetermined by its ability to correctly classify the actualmeasurands (defects) in different classes: a completePD measuring system which produces correctclassifications of defects can be considered"optimized".

The latter stage of the measuring process includes apreprocessing (data compression) stage, where somesignificant features are extracted from the measurand,implementing one (or more) of the techniques reportedin the introduction, and are gathered in a reduced dataset (usually about 5-20 data).

If such techniques are "non-lossy" (i.e. if their outputcontains all the sought inform ation which is present inthe input), the comparisons between such reducednumerical sets can be equivalent to the comparisonsbetween the relevant PD patterns full data sets.

Experience has shown that the "optimal" choice of thequantities composition of such reduced sets dependson the type of power component [Z ] and on thepurposes of the diagnostics: sometimes differentcomp ositions are equally valid.

A number of suitable "computational strategies", arethen available to implement the last stage of themeasuring process: the classification. Any such

Fig.2 Schematic diagram of a PD detection circuit

47 1



strategy needs a fine tuning to select the mostappropriate classification criteria.

A. Optimisation of a PD Measuring System: a simple

Experience has shown [3] that, in order to achieve avalid diagnostic when using digital PD measuringsystems, the “dead time” (td) of the PRDPA is to behigher than the “settling time” of the PD signal (mainlydeterm ined by Zm). If this condition is not satisfied thePRDPA output (the PD patterns) will not contain acorrect information about the inve stigated defect@).

example

Fig. 3 illustrates a case where a same defect (anelectrical tree), processed by the same measuringcircuit (same Zm and same PRDPA settings, excepttd), originated two different PD patterns for twodifferent values of td (note that in the original picturesthe PDs amplitude data were colour coded): one typicalof an electrical tree (td = I O ps ) and one typical of avoid (td = 5 ps).

The approach to correct this situation, after itsdetection, was very simple because only the effects (onthe diagno stic assessment) of the varia tion of the “deadtime” factor was considered and because a direct(visual) comparison of the PD patterns was sufficientlysignificant. While, in general, the Zm isto be modified(within limits) in orde r to render the “settling time” lowerthan td, in the specific PRPDA instrument used in thiswork, td could be easily chosen in a range of presetvalues. Therefore a series of tests could be run in avery short time, the influence of the td fac toron the PDmeasurements output was promptly quantified and,choosing the lowest possible value o f td, a step towards

the measuring system “optimisation” (which

I

disregarded the possible influence of factors differentfrom the PRPDA ”dead time”) was implemented.

B. The Addressed Problem

In practice, the approaches to shorten the d istance of aPD measuring system from its optimal operation,unfortunately, can be more complex: the factors ofinfluence (and their interaction) are to be iden tified andthe relevant influence is to be quantified, in order toenable the implementation (and the validation) ofcorrecting measures.For instance, some of the authors have demonstrated[3] that the following factors, among other ones, caninfluence the PD patterns and so can affect the validityof the final PD based diagnostics of an insulationsystem:a) the acquisition time tacq , whose value can affect thePRPDA outputb) the PRPDA amplifier gain: the choice of its valuecan be critical, owing to the opposite effects of theam plifier saturation and of the quantisation error on thePD p atternc) when a Low Leve l Discriminator (LLD) d igital filterintroduced to avoid “saturation” of the number ochannels due to noise, the choice of its cut-off level,obviously does affect the PD pattern

However no method was available to determine, “apriori” and quantitatively, the extent of the above threefactors influence on the final re sults and to evaluate thepossible relevance of their interaction.

In the present work, the above mentioned PD patternsdata compression was operated by means of statisticalprocedures applied to the following “de rived” quantitiescalculated from the basic quantities of the P D patterns:

Figure 3 Examples of PD patterns. On the left it is reported a correctly acquired pattem (td = 10 ms)for an electrical treeing phenomenon. On the right it is reported a “wrong” pattern (td = 5 ms),

obtained on the same site, which suggests the existence of a void

472



i=h+l =12

total charge, positive and negative

(3 )

average positive and negative charge (referred to theacquisition time) or PD current.

While from each PD pattern, several frequencyhistograms can be com puted together with the relevantstatistic parameters such as Mean, Median, Mode,

Variance, Skewness, Kurtosis, Cross Correlation Factoretc., for the purpose of the present work just the Hn(Q)-, frequency histogram of the number of negativedischarges Ni(Q) as function of the amplitude Qi, andthe Hn(@)-, frequency histogram of the number ofneg ative discharges Ni(CD) as function o f the phase @,have been considered, as they allow to identify and tocharacterise several de fects in insu lation systems [5,6].

C. The testing arrangement

A very simplified “model” of an insulation system wasobtained realizing a PD test cell, detailed in [7], whichhosted thin polymer films. For the purposes of thiswork, surface PD tests have then been performed onpolyethilentereph talate (PET) films using such a ce ll. A50 pm thick film was placed, in a sphere-planegeom etry, between the spherical electrode (CD= 6 mm,set, in air, at 0.1 +/- 0.01 mm distance from the film)and the plane electrode, a 2.4 kV ac voltage wasapplied and the PD patterns were acquired undercontrolled conditions.

Ill. THE ADOPTED DO E APPROACH

A solution to the addressed problem has been foundimplem enting Design Of Exp erimen t (DOE) techniques.

Since decades, industrial problems of criticalparameters optimisation, connected to total quality, toproducts development and to cost-and-time reductionsare often addressed applying the DOE techniques [8].In a general situation, the output variables of anidealized process depend on the input variables, on aset of controllable factors (factor of influence) and on aset of non -controllable a ctors (noise factors).The here adopted DOE approach can be summarisedas follows:- the Input variables, the Output variables and the set

of Controllable Factors (together with the relevant

473

ranges of variation) have been selected, while a ll theother influences on the actual process have beenassumed to be “noise”.

- a plan (“design”) of the “experimen ts” to be performe dhas been outlined, following the rules of DOE, whichevidence the minimu m num ber of combinations of theControllable Factors values

- the analysis of re sults has been ”shaped” to evidencethe validity of the Controllable Factors selection, thedegree of influence (the weight) of each factor on theoutput and the weight of the selected factorssynerge tic action on the ou tput.

The latter analysis allowed to quantitatively estimatethe conditions necessary, but not sufficient, to achievethe optimisation of the PD measuring system fordiagnostic purposes.

A. The DOE plan and the Experimental Results

In the present case the input variables were the PDthemselves. 11 quantities (statistical parameters)

derived from the PD patterns have been selected asoutput variables and the controllable factors were:A) the acquisition time fo r each PD patternB) he amp lifier gainC) the setting of the Low Level Discriminator (LLD)

A complete factorial33 plan has been chosen, where 3possible levels for each controllable factor have beenconsidered (see Table I). As three tests (replication = 3)were to be performed far each planned combination ofsuch factors, the “experiment“ regarded a total of 81tests.

filter

TABLE I

SELECTED LEVELS FOR THE CONTROLLING FACTORS

FACTORS I LOW ICENTRALIHIGHTime (s) I 10 I 95 I 180

Each test was represented by a PD pa ttern and by therelevant values of the 11 output variables. All the testsresults, i.e. the value s of each selected output variablefor each planned combination of factors and levels,have been processed by m eans of analysis of variance(ANOVA) techniques (implemented through “off the

she lf” software packages).

Hereby 3 examples relevant to 3 different outputvariables are reported.

Example 1) Output variable: I+ . Th; ANOVAprocedure evidence d that the highest R indicator(which can be considered representative of the ratiobetween the variability of results due to the assumedstatistical model and the actual variability of results)was ob tained for the second order statistical model. Its



value (0.95) indicated that the selected factors wereactually the main factors of influence.Then the weight of the controllable factors (A =acquisition time; B = gain; C = LLD level) , of therelevant quadratic terms (AA, BB, CC) and of the firstorder interactions (AB, BC, AC) was computed andvisualised. A Pareto chart (see Fig. 4) evidenced boththat the e ffects of tacq were far less impo rtant than theeffects of the other two m ain factors and that the effectof interactions was negligible. Besides a plot of themain effects (of these three factors on the I+ outputvariable) vs. the relevant ranges of variation outlined(see Fig. 5) both the unexpected presence (probablydue to the effects of the quantization error) of amaxim um for a gain value of about8 and the expected(almost linear) decrease of the LLD leve l effect.Therefore, in this case, the presence of the variable I+in a set of output variables, was to be "optimised" bysetting the gain at 8 and the LLD level as low aspossible. Besides, while the influence of the interactionbetween the investigated factors could be disregarded,the influence on the final diagnostics of tacq deserved

further investigation, because, during the previousexperiment the variation of factors B and C could have"masked" the effects of factorA.Actually further PD tests ru n under identica l conditions,varying only tacq, keep ing the o ther factors constant atthe ab ove optimised values, did revea l that the factor Ahad a "modest" influence on the output I+, up to a

2C:LLD

I 1.59 I 9.5 I 4

B:Gain

BB

A:Time

0 4 8 12 16 20 24

Standardized effect

saturation value o f 100 s. As higher tacq values d id notcause any change in such an output, the "o ptima l" tacqvalue was to be set at 100s.

In general, the latter reasoning is to be checked againstthe possibility that the phenomena underlying the PDsare time-varying. In this specific case, concerningsurface PDs acting on a specific polymer film, auxiliarPD aging tests showed that such phenomena could beconsidered not varying, provided that the tacq wasassumed in the range of 40 s.

Example 2) Output variable: Skewness of the Hn(Q)-distribution (SkHnQ-). The ANOVA procedureevidenced that the highest R2 ndicator was obtainedfor the second order statistical model. Its value (0.84)indicated that the selected factors were acceptablydescriptive of the phenomena involved.An interaction between the factors gain and LLD levelwas at first evidenced by the crossing of the BC curvesin Fig. 6. Its relevance was q uantified by com puting theweight of the factors B, BB and BC influenceONB,WBB

Wac) on this output variable, the lim it significance leveof such weight (WLS) and the highest weight of theother factors (WO~), hich were reported in Table 11.

TABLE II

WEIGHTOF 2ND ORDER STATITICAL MODELFACTORS ON OUT PUT VARIABLE SkHnQ-

I WLS I WE I WB0 I

Such results indicated that:

- the factor gain was not linearly related to the outpuvariable, as confirmed by the m ain effects p lot inFig. 7, which revealed that the influence of the factogain had a minim um at a value of about 8- the choice of the LLD level and of the tacq value(within the here considered ranges) would haveproduced little or no significant effect on the shapeof

the Hn(Q)- distribution, therefore would not havaffected the validity of a relevant diagnostics, based onthe output variable SkHnQ-

Fig. 4 Pareto chart for the output variable I*

Standa rdized main e ffects (arbitrary units) Degree of Intera ction (arbitary units)

Fig. 5 Main effects plot for the output variable I+

47 4

Fig. 6 Interaction plot for output variable SkHnQ-



Main Effec ts (arbitrary units)

2 3

2.1

1.9

1. 710.0 180.0 4.0 10.0 1.0 10.0

Tacq Gain LLD

Fig. 7 Main effects plot for the output variable SkHnQ-

- the adoption of different pairs of B and of C valuesduring a comparison between PD measurements onequal HVP components would have “artificially“(although slightly) affected the measurement of theHn(Q)- distribution, thus, possibly, the relevant PDbased diagnostic would ha ve classified the two objectsas differen t ones.

Therefore the DOE analysis pointed out that, insertingthe output variable SkHnQ- in the reduced data setcomposition would have required to fix the LLD leveland the gain value for all the future diagnosticdeterminations on similar objects (i.e. to cancel suchfactors from the group of the controllable variables), inmin imize the p ossibility of an incorrect diagnosis.

Example 3) Output variable: Kurtosis of the Hn(@)-distribution (KuHncD-). The ANOVA procedureevidenced that the highest R2 indicator was obtainedfor a third order statistical model. Its value (0.43)

indicated that the phenomena involved in the averagedistribution of PDs occurence along a half-cycle of thefrequency were driven also by other factors, differentfrom the three selected ones. Therefore the DOEanalysis evidenced that the KuHncD- output variablewas not be enclosed in the final reduced data set of aPD based diagnostic, without investigating theinfluences of other factors.

B. Comments

The information obtained through the implementationof the DOE techniques during PD measurements inpower components can be useful in several ways. Fo rinstance:- it can suggest of setting the PD digital acquisitiondevice at a certain “working point“ [e.g. the maximumof gain or the acquisition time saturation value inexample I), s well as the m inimum of gain in example2)], when it is desirable to improve the stability of thePD patterns for sm all variations of a specifc factor ofi fI en ce- given a range of variation of the investigated factors,it allows to check wether the classification criteria of

475

defects [5,6,9], assumed in the last stage of themeasuring process, are appropriate- allows to formulate some “criteria of selection of thefactors of influence” which would minim ize the possibleintroduction of artificial errors when operating acomparison between PD patterns obtained in differenttesting conditions

IV. CONCLU SIONS

The present work indicates that the adoption of DOE(Design Of Experiment) techniques allows to obtainquantitative assessments about the influence of thesettings of a partial discharges (PD) digital measuringsystem on the results of the PD patterns analysis. Suchinformation can help the optimisation procedure of aPD measurement system aimed at the diagnostic ofpower components.

Although these techniques have been successfullytested only in simple cases (surface PDs on polymerfilms), its application to the diagnostic of actual

insulating systems (power electrical components) bymeans of PD digital acquisitions appears to bepos itively promising.

Furthermore the DOE techniques, which can beconsidered particularly efficient applications ofstatitistics, appear to have the potentiat for a widerapplication, especially to the “optimisation” ofmeasurement systems, when no simple models of thephysic processes are available.

ACKNOWLEDGMENTS

The authors acknowledge the support of MURST 40%

funds to the present work.

REFERENCES

[ I ] IEC Publication 270[2] GZingales: discussion about article “Digital processing of PDpulses”, by P.OsvBth, IEEE Trans. on Dielectrics and ElectricalInsulation, vol. 2, august 1995, pp.688-691[3] R. B w o . C. Gemme, F. Guastavino: “The Influence of DigitalMeasurement System Characteristics on PD Patterns”; Proc of CEIDP95, Conf. Proceeding, Virginia Beach, VA, USA, October 22-25, 1995,

[4] R.Bono, CGemme, F.Guastavino: ”The dependence of fingerprintsrelevant to PDs on degradation time and test voltage”; Proc. of 1996CEIDP, San Francisco, USA, October 20-23, 1996, pp. 500-503.[5] E.Gulski: “Diagnosis of HVcomponents by digital PD analyzer“; IEEE

Trans. on DEI, Vol. 2, August 1995, pp.630-640[SI R.E.James, B.T.Phung:”Development od computer-basedmeasurements and their application to PD pattern analysis”; IEEE Trans.on DEI, Vol. 2, October 1995, pp.838-856m R. Bono , L. Centurioni, F. Guastavino: “Measuring the endurance offilms in Partial Discharges”. IEEE Trans. on Electrical Insulation, Vol.28, Num. 5, Dicembre 1993, pp. 1050-1056[8] D.C. Montgomery: “Design and analysis of experiments”,l991, J.Wiley&Sons, ISBN0-471 52994-X[9] E.Gulski: “Digital analysis of partial discharges”; IEEE Trans. on DEI,Vol. 2, October 1995, pp.822-837

pp. 347-352

ieee_application of design of experiment techniques to measurement procedures

Documents