2122 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 37, NO. 3, JUNE 2022
Fas Algo i hm o Con ac less Pa ial Discha ge
De ec ion on Remo e Ga eway De ice
Tomáš Ma ino iˇcand Jan Fulneˇcek
Abs ac —De ec ion o he high impedance aul caused by eg-
e a ion is one o he bigges challenges b ough by he usage o
co e ed conduc o s in medium ol age o e head powe lines. One
o he accompanying e en s o long- e m con ac o ege a ion
wi h he XLPE insula ion is pa ial discha ge which damages he
insula ion. Cu en sys ems o he de ec ion o pa ial discha ge
ha e wo majo p oblems: he p ice and di icul ins alla ion. In his
pape , an app oach o he de ec ion o he pa ial discha ge om
he da a collec ed by he an enna is desc ibed. This app oach is
ocused on he small compu a ional demand and low alse posi i e
a e. Thanks o he small compu a ional equi emen s, i can be
un on he emo e ga eway de ices which a e collec ing he da a
om he an enna. I is composed o ou s eps: ou lie de ec ion,
ou lie clus e ing, ea u e ex ac ion, and classi ica ion. I is shown
ha his app oach g ea ly imp o es he de ec ion a e and lowe s
alse posi i es compa ed o he p e ious algo i hm used o pa ial
discha ge de ec ion based on he da a om an an enna, making i
i o use in he p oduc ion en i onmen .
Index Te ms—Time se ies, pa ial discha ge, pa e n
classi ica ion, ou lie de ec ion, wi eless, high impedance aul ,
co e ed conduc o .
I. INTRODUCTION
MEDIUM ol age (MV) o e head powe lines a e usually
equipped wi h egula AlFe conduc o s, wi hou any
ex e nal insula ion. In a o es ed a ea, phase- o-g ound and
phase- o-phase aul s o en occu on such powe lines because o
he su ounding ege a ion [1]. To elimina e such aul s, AlFe
conduc o s a e being eplaced wi h co e ed conduc o s (CC).
CC consis s o an aluminum co e, which is co e ed by a hin
laye o XLPE insula ion ma e ial. The bigges disad an age o
CC lies in he p oblema ic de ec ion o high impedance aul
(HIF), caused by ege a ion. Because o he addi ional XLPE
insula ion, he aul cu en is ex emely low, so i canno be
Manusc ip ecei ed Janua y 12, 2021; e ised Ap il 13, 2021 and June
24, 2021; accep ed Augus 8, 2021. Da e o publica ion Augus 13, 2021;
da e o cu en e sion May 24, 2022. This wo k was suppo ed in pa
by The Minis y o Educa ion, You h and Spo s h ough he Na ional P o-
g amme o Sus ainabili y (NPS II) p ojec IT4Inno a ions excellence in sci-
ence - LQ1602, Technology Agency o he Czech Republic unde P ojec
TN01000007 and in pa by The Minis y o Indus y and T ade unde
P ojec CZ.01.1.02/0.0/0.0/20_321/0024308. Pape no. TPWRD-00072-2021.
(Co esponding au ho : Tomáš Ma ino iˇc.)
Tomáš Ma ino iˇc is wi h IT4Inno a ions, VSB – Technical Uni-
e si y o Os a a, 70800 Os a a-Po uba, Czech Republic (e-mail:
omas.ma ino ic@ sb.cz).
JanFulneˇcekiswi h heENETCen e,VSB–TechnicalUni e si yo Os a a,
70800 Os a a-Po uba, Czech Republic (e-mail: jan. ulnecek@ sb.cz).
Colo e sions o one o mo e igu es in his a icle a e a ailable a
h ps://doi.o g/10.1109/TPWRD.2021.3104746.
Digi al Objec Iden i ie 10.1109/TPWRD.2021.3104746
de ec ed by s anda d p o ec ion elays. I a ee o b anch s ays
in di ec long- e m con ac wi h CC, pa ial discha ges (PD) may
appea on he su ace o he XLPE insula ion. PDs a e causing
slow deg ada ion o CC, esul ing in an insula ion ailu e. In
ex eme si ua ions, CC can be damaged a e se e al hou s o
PD ac i i y. Howe e , usually i akes se e al days be o e se e e
insula ion damage occu s. Ea ly de ec ion o PD ac i i y can
p e en he damage o CC and ene gy ansmi in e up ion.
PD ac i i y is always accompanying phenomena o HIF.
The e o e, hep esenceo PDcanbeusedasanindica iono HIF
occu ence. The e a e se e al ways how o acqui e PD pa e ns
om powe lines wi h CC. Mos o he me hods uses gal anic
me hods o he acquisi ion o high equency componen s o
ol age o cu en signals. To ob ain i , he measu ing equipmen
had o be connec ed o he examined powe lines ia senso s: a
capaci i e di ide ( o ol agesignals) o Rogowskicoil (cu en
signal). These senso s mus be sui able o MV usage, which
makes hemexpensi e.This is why he e a ejus a ew online PD
de ec o s o o e headpowe linescomme ciallya ailableon he
ma ke -high cos p e en de ec o s o become widely sp ead.
Lu e al. [2] p o ided a comp ehensi e summa y o di e en
algo i hms used o he de ec ion o pa ial discha ge in a ious
applica ions.
II. PROBLEM DEFINITION
To dec ease p ice o he de ec o as much as possible, i was
decided o eplace he common gal anic senso s by a con ac less
me hod. When PD ac i i y appea s on he su ace o CC, i c e-
a es a ypical pa e n o elec omagne ic ield in he su ounding
space o a CC. This pa e n can be acqui ed by an an enna, as i
was done in ou p e ious wo ks [3], [4]). The main ad an age o
his con ac less app oach is i s p ice - an enna is much cheape
han a high ol age capaci i e di ide o Rogowski coil [5]. On
he o he hand, he e a e also disad an ages. The de ec ion ange
is sho e and he low equency componen o a signal canno
be acqui ed by his me hod.
A. Da a Acquisi ion Ha dwa e Desc ip ion
To ob ain da a om a eal en i onmen , al eady exis ing da a
acquisi ion (DAQ) pla o ms we e used. These pla o ms a e
si ua ed on a ious MV o e head powe lines wi h CC and hei
o iginal pu pose is o p o ide PD de ec ion ia gal anic me hod
( ol age signal e alua ion) [6]. Each pla o m con ains DAQ
ca d wi h ou inpu s. Th ee inpu s a e being used o gal anic
PD de ec ion, he ou h inpu is backup. Fo his expe imen , 4
This wo k is licensed unde a C ea i e Commons A ibu ion 4.0 License. Fo mo e in o ma ion, see h ps://c ea i ecommons.o g/licenses/by/4.0/
MARTINOVI ˇ
C AND FULNE ˇ
CEK: FAST ALGORITHM FOR CONTACTLESS PARTIAL DISCHARGE DETECTION ON REMOTE GATEWAY DEVICE 2123
Fig. 1. High impedance aul on medium ol age powe line.
Fig. 2. An enna moun ed unde o e head powe line.
o hese pla o ms we e addi ionally equipped wi h BONI whip
an ennas. I is a common wide band whip an enna, used by HAM
ope a o s.The o iginal equency ange o his an enna is20 kHz
– 300 MHz. Howe e , because o he limi ed DAQ sample a e,
he equency ange o he acqui ed signals is up o 20 MHz. The
an ennaou pu s a econnec ed o he backupinpu o PD de ec o
DAQ, so he o iginal unc ionali y o he de ec o is p ese ed.
This is an essen ial aspec , because his gal anic me hod is used
o acqui ed an enna da a e i ica ion.
The an enna is moun ed on he pole, pa allel o he CC as i
can be seen in Fig. 2. Once an hou , he ou pu signal om he
Fig. 3. PD de ec ion block diag am.
an ennaisacqui ed a a sample a eo 40MS/s ina o al leng ho
20ms.Inacaseo a ee all,PDac i i yiss ableandcon inuous,
because he ee lies di ec ly on he co e ed conduc o s. This is
he mos dange ous si ua ion om he poin o iew o conduc o
insula ion. Uns able PD ac i i y is ypical o andom con ac s
wi h su ounding ege a ion (usually because o s ong winds),
bu hese andom con ac s ha e a low impac on insula ion
damage. F om ou expe ience, acqui ing da a once pe hou is
su icien o hepu poseo insula iondamagede ec ion.Sho e
acquisi ion pe iod was also es ed, bu i caused ouble wi h
slow da a connec ion. Each acqui ed signal con ains 800 000
samples and ep esen s one pe iod o he powe g id equency.
F equency ange o he DAQ wi h connec ed an enna is 20 kHz
– 20 MHz, he esolu ion is 8 bi s wi h a iable ange up o 1
Vpp. Acqui ed signal is ansmi ed h ough GSM ne wo k o
he da abase. The DAQ pla o m and he cu en p ocess o PD
de ec ion a e shown in Fig. 3. Because o he usage o GSM
and slow da a a es caused by e y low signal le el o dis an
places,i isconside ed omake hecompu a ionsdi ec lyinDAQ
ins ead o a da a se e . The p oposed algo i hm was de eloped
based on his equi emen .
BONI whip is an ac i e an enna, which equi es connec ion o
he powe supply o i s p ope unc ion. The an enna is powe ed
ia a coaxial cable o make he ins alla ion as easy as possible
(no addi ional wi ing is needed). Nominal an enna powe supply
ol age is 12 V, he cu en is up o 150 mA. The whole acquisi-
ionpla o m ispowe ed iains umen ans o me (22/0.1 kV),
which is connec ed di ec ly o he examined powe line. How-
e e , he ampli ie s in DAQ pla o m in oduces DC o se in
he acqui ed signals. Luckily, he ampli ude o his o se is e y
low, so i usually does no cause any signi ican p oblem.
The ene gy (and se e i y) o PD pulse is desc ibed by i s
appa en cha ge. In labo a o y condi ion wi h o e head powe
line model connec ed o he PD calib a o , he sensi i i y o
BONI whip an enna senso is 1 nC/10 mV. Howe e , his alue
2124 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 37, NO. 3, JUNE 2022
Fig. 4. PD pa e n decomposi ion.
depends on ac ual powe line pa ame e s (leng h, heigh abo e
g ound, conduc o ype, ac ual g id con igu a ion e c.). This is
why sensi i i y a ies in a case o eal en i onmen . The only
way o sensi i i y es ima ion in a eal en i onmen is on-si e
calib a ion. Fo his pu pose, he whole dis ibu ion powe g id
would ha e o be shu down. This is no accep able o dis-
ibu ion g ids ope a o s, so he on-si e calib a ion was ne e
pe o med.
B. PD Pa e n Desc ip ion
PD pa e n signal om he gal anic PD de ec ion can be
decomposed in o wo componen s. Low equency (LF) com-
ponen consis s o a 50 Hz ca ie wa e (powe g id equency),
he high equency componen (HF) con ains he PD ac i i y,
see Fig. 4.
Posi ion o he HF PD peak in ela ion o he LF ca ie wa e
is he key ea u e o PD classi ica ion [7], i is desc ibed by so-
calledphase esol edwa e o mdiag ams.Howe e , hisme hod
canno be used o signals acqui ed ia an an enna, because
LF componen is missing [5]. I is a below he lowe cu o
equency o he an enna.
The acqui ed signal con ains only HF componen wi hou a
50 Hz ca ie . This is wha makes he classi ica ion o hese
signals so challenging.
III. CLASSIFICATION ALGORITHM
The acqui ed da a con ains a high le el o backg ound noise.
As i was obse ed in he da a, mos o he backg ound noise
is wi hin no mal de ia ion. I is impo an o no e ha he
p esence o ou lie s does no imply he p esence o he HIF.
The e a e o he ypes o discha ges, which a e no ha m ul o
he CC insula ion and need o be disce ned as such. Among
he o he ypes o discha ges usually p esen a e he co ona
discha ge, which is no iceable by he symme ic discha ges in
bo h ampli ude di ec ions and andom soli a y ou lie s.
The p oposed algo i hm consis s o ou pa s: ou lie de-
ec ion, ou lie clus e ing, ea u e ex ac ion, and classi ica ion.
The whole p ocess is ocused on making he compu a ion as
as as possible wi h minimal ha dwa e equi emen s. This is o
enable compu a ion on si e wi h al eady ins alled ga eway de-
ices, which a e no mally used o measu ing and con ol o he
dis ibu ion powe line. Wi h as compu a ion ime, he ga eway
de icecanconduc i so iginal askwhilebeingable ode e mine
he aul by hemsel es wi h minimum da a ansmissions. A
he momen , he ime se ies a e being sen o he se e o he
analysis and his leads o delays in deli e ing he esul s o he
a eas wi h slow da a connec ion. Only he ou lie de ec ion uns
on he whole ime se ies o he signal and has a compu a ional
complexi y o O(n). This means ha he e o needed o make
he compu a ion scale linea ly wi h he p ocessed da a size. All
subsequen ope a ions a e done on a much smalle subse o
he ou lie s, sa ing a lo o compu a ional esou ces. De ec ed
ou lie s a e hen clus e ed based on hei empo al dis ance.
A e wa ds, he ea u esa ecompu ed o indi idualclus e sand
hen combined in o inal ea u es common independen ly om
he numbe o clus e s. Finally, he classi ica ion is lea n o line
on he selec ed ea u es and he inal model can be deployed
e en on embedded de ices wi h smalle compu ing powe . The
ea u esa e based mos ly ons a is ics and he classi ica ioni sel
isdoneby heXGBoos [8]model. Thewholesolu ionis ocused
on as compu a ion e en o long ime se ies and maximiz-
ing he p ecision o he algo i hm, i.e., minimizing he alse
posi i es.
A. Ou lie De ec ion
The signal measu ed by he wi eless an enna con ains a high
amoun o noise a e ac s gi en mos ly by AM adio and he
esis ance in he ci cui y i sel . Gi en ha he de ice is always
measu ing he highes equency ecei ed, he noise usually
hides a la ge pa o he o iginal signal and c ea es an almos
compac body when isualized. Fu he mo e, his compac body
changes i s magni ude o e ime. Due o his, i is no possible
o de ine a single alue which would disce n ou lie s and noise.
Ano he cha ac e is ic o he signal is ha i is p ocessed by he
8 b analog- o-digi al con e e (ADC) which e u ns alues in
he ange <−128,128 >.
Based on his knowledge, he ou lie s may be de ined as all
poin swhich dono ha ea leas cmin neighbou s ina gi ena ea.
Limi s o his a e gi en by wo pa ame e s εand k.Thisa eais
limi ed by a maximal empo al dis ance ε/2in bo h di ec ions
om he poin . The alue ange o a poin wi h alue xis [x−
k,128] o X≥0and [−128,x+k] o x<0.
F om he compu a ional iew, his can be easily implemen ed
by coun ing a olling his og am o e he window o leng h εand
summing all alues g ea e han x−k o posi i e xo summing
all alues less han x+k o nega i e x.
Many ypes o discha ges which a e no PD ha e a la ge num-
be o obse a ions in a sho ime span, wi hin he whole ange
o alues, see Fig. 5. One addi ional ea u e o his app oach is
ha discha ges such as co ona a e o en igno ed when a wide
ime window is used o ou lie de ec ion.
MARTINOVI ˇ
C AND FULNE ˇ
CEK: FAST ALGORITHM FOR CONTACTLESS PARTIAL DISCHARGE DETECTION ON REMOTE GATEWAY DEVICE 2125
Fig. 5. Co ona discha ge example.
In ou case, cmin =20,k=30, and ε= 1001 ga e he bes
esul s and will wo k wi h he same se up as desc ibed in
subsec ion II-A. These pa ame e s could be modi ied i di e en
ha dwa e is used, o di e en kind o ou lie s a e o be ound.
To make a gene al ule on se ing hese pa ame e s is no easy
since i always depends on he use case, howe e i is possible o
p o ide guidance on how o app oach his. The easies ule could
be said o k, which is se as a li le o e 11% o he maximum
ange o he signal, and his app oach should be good in mos
cases. In he case when 8 b ADC is used, his ange is easy o
be de e mined based on i s p ecision, howe e , e en in he case
o gene al signal a 10% o i s ange is a good place o s a .
The εis se o 1001 based on he idea ha some a ea a ound he
examined poin should be in es iga ed, howe e no oo wide, so
i will ca ch oo many o he discha ges. This pa ame e should
be based mos ly on he sampling equency o he signal. In ou
case, he sampling equency was 40MS/s. The las pa ame e
cmin wo ks as a sensi i i y pa ame e . The lowe alue o cmin
e u ns a smalle numbe o ou lie s since i is mo e es ic i e
on o he spikes ha may occu in he neighbou hood o he
examined poin . One o he hing ha a ec s his pa ame e is
he amoun o noise in he da a. I he da a con ain less noise, i
is sa e o use he lowe alue, since he e a e a lowe numbe o
spikes p oduced by noise.
B. Ou lie Clus e ing
A e he ou lie s a e ound, a simple p ocedu e is used o
clus e he ou lie s in o g oups. In essence, he idea is o ind
a clus e o ou lie s which a e close oge he . The clus e ing is
done sepa a ely o he posi i e and nega i e ou lie s, since hey
a e pa o di e en discha ge beha iou .
The pa ame e s dand ma e chosen, whe e ds ands o
he maximal dis ance o wo poin s o be conside ed as pa
o a clus e and mis he minimal numbe o poin s g ouped
oge he o be conside ed a clus e . The ou lie s a e conside ed
a clus e i hei dis ance is less han do he e is a pa h be ween
wo ou lie s c ea ed by such ou lie s. This simple me hod is
simila o he DBSCAN clus e ing [9], bu i is compu ed only
in one dimension, so i can be compu ed in O(n) compu a ional
complexi y. Example o clus e de ec ion in a ime se ies is
shown in Fig. 6
Fig. 6. Ou lie de ec ion and clus e ing example.
Se ing o dand mshould depend on he applica ion and
sampling equency o he signal. In his case, mwas se o 5,
because he PDs should be de ined by a leas se e al ou lie s
and smalle numbe s would mos likely mean some o he kind
o discha ge was de ec ed. The dpa ame e was se o 30 000
ha equals o 3/4 o millisecond.
C. Fea u es Compu a ion
When he clus e s a e de e mined, he ea u es o he classi i-
ca ion algo i hm a e compu ed. The i s s age is conce ned wi h
c ea ing ea u es o each clus e . A e wa ds, he ea u es a e
agg ega edac oss he clus e s.On he clus e le el, he ollowing
10 ea u es a e compu ed:
Time ange
S anda d de ia ion
Minimum alue
Median alue
Mean alue
Maximum alue
Signal ange
Skewness
Ku osis
Numbe o obse a ions
These ea u es we e selec ed since based on he p e ious
esea ch [3]. I was ound ha he ac o s o he iden i ica ion
o pa ial discha ges a e he du a ion o PD pulse, i s magni ude,
and shape. Howe e , i is no possible o apply hese ea u es in
he con ac less me hod, because o he missing phase- esol ed
in o ma ion (LF componen ) in he signal. This is why he
ea u eswe e applied o hewholepulse- ain in hisexpe imen .
Due o he di e en numbe o clus e s in each se ies, i is
necessa y o make an agg ega ion o hese ea u es ac oss he
clus e s o c ea e he equi alen ea u es o e e y ime se ies.
The agg ega ion was done by compu ing he minimum, 1s
quan ile, mean, median, 3 d quan ile, and maximum o each
ea u e ac oss he clus e s. Fo example, ins ead o ‘Time ange’
he e would be ‘Minimum ime ange,’ ‘1s quan ile ime ange,’
‘Mean ime ange,’ ‘Median ime ange,’ ‘3 d quan ile ime
ange,’ and ‘Maximum ime ange’ a iables in he agg ega ed
ea u e se . These a e compu ed om he ‘Time ange’ a iables
o each clus e . This p ocess is hen execu ed o each o he 10
ea u es compu ed o each clus e . This leads o 60 ea u es o
2126 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 37, NO. 3, JUNE 2022
Fig. 7. Se ies con aining oo much noise example.
he classi ica ion model. Addi ionally, a numbe o clus e s and
loca ions a e added o hese ea u es. The loca ion is a quali a i e
a iable and he e o e i needs o be ans o med o quan i a i e
a iables. A dummy a iable ans o ma ion was used [10],
whe e o each unique alue a new a iable is c ea ed wi h 1 in
ows whe e he alue is ue and 0 o he wise. One a iable is
always excluded o p e en he singula i y in he ea u e ma ix.
Since he measu emen s we e aken on 4 loca ions, he inal
model has 65 ea u e columns (60 + 4 loca ions and numbe o
clus e a iables).
D. Classi ica ion
The las s age is he classi ica ion and i is di ided in o h ee
pa s making p eselec ion o he ime se ies based on some
cha ac e is ics. The i s selec ion is whe he any ou lie clus e s
we e ound in he ime se ies. I he e a e none, i is conside ed as
he signal wi hou any discha ges, since he e a e no no iceable
alues poin ing owa ds he discha ge. I is no gi en ha he e
will be no discha ges, because hey could be hidden in he
noise c ea ed by ex e nal sou ces, howe e in such a case he e
is no way o iden i y he pa ial discha ge. The second es
il e s he se ies which ha e oo noisy measu emen s and i
is impossible o disce n be ween he una ec ed se ies and he
se ies con aining pa ial discha ge. Based on ex ensi e analysis,
i was de e mined ha any se ies which con ains a clus e wi h
mo e han 500 obse a ions is conside ed as an ex emely noisy
se ies. This cha ac e is ic was obse ed by many poin s in he
posi i e spec a being conside ed an ou lie in he case i he e is
an ex eme amoun o noise. I happens in he posi i e spec um
due o he signal o se caused by he powe supply. An example
o such ime se ies wi h clus e s is p o ided in Fig. 7.
I he se ies does no belong o nei he o hese ca ego ies,
i s ea u es a e e alua ed by he model c ea ed by he XGBoos .
XGBoos is a lib a y based on he g adien boos ing amewo k,
which is known o be used in sol ing classi ica ion p oblems
in indus y and compe i ions such as Kaggle [11]. The g adien
boos ing is based on aining mul iple models and weigh ing he
esul s c ea ing an ensemble model.
E. Usage in O he Applica ions
The whole p ocess is composed o se e al s eps and mos o
heseha esomepa ame e s.Thepa ame e s o ou lie de ec ion
and clus e ing a e p o ided o inc ease he lexibili y o he
solu ion. Fo example, i one would like o de ec discha ges on
he cables inside he ac o y, hese would ha e a di e en pa e n
han he ones c ea ed by he ege a ion on he o e head powe
lines. The e o e, i would make sense o adjus he pa ame e s.
Wi hou p ecise knowledge how such discha ges p esen hem-
sel es, i is di icul o p o ide a gene al ule o such se ings.
The bes app oach is o de e mine he bes pa ame e s is o use
hype pa ame e sea ch. Tha means o sea ch he expe imen
space wi h mul iple di e en con igu a ions, un il a se ing wi h
good esul s is ound. This should no be a p oblem in mos cases
since he algo i hms a e o linea compu a ional complexi y and
he e o e a e compu ed e y as .
Ano he possible d awback o he p oposed app oach is us-
age o he machine lea ning model XGBoos , which elies on
lea ning. This means ha i is necessa y o ha e a se o labelled
da a o p o ide o lea ning. In his case, we ound ha a leas
h ee hund ed di e en ime se ies had o be used o lea ning o
p o ide good esul s. On he o he hand, he ad an age is ha he
mo e da a a e p o ided o lea ning, he be e he model should
be and i can imp o e when new da a a e a ailable.
One o he easons o making he whole p ocess con ain
mul iple s eps is he modula i y o he app oach. Ac ually, i is
possible o change, o example, he me hod o ou lie de ec ion
o a di e en one and hen p oceed wi h he es o he p oposed
app oach. O qui e he opposi e, he ou lie de ec ion may be
used as a s andalone pa in comple ely di e en ields such as
economics, musics, medicine signal analysis, e c. In he same
way he clus e ing algo i hm, ea u e c ea ion, o he XGBoos
model, each componen may be eplaced by a di e en one.
Tha means he in e es ed eade s can ake inspi a ion om ou
app oach and eplace some pa s i hey a e unsui able o hei
use case.
All unc ions used can be ound in an R package called
WIPADD.1This package con ains indi idual unc ions o each
s ep o he p esen ed me hod as well as helpe unc ions o make
he whole p ocess easie .
IV. RESULTS
To e alua e he p oposed app oach,we es ed he pe o mance
on manually labelled da a measu ed by Bony whip an enna
loca ed on o e head powe lines o he Czech Republic powe
g id.The esul s o he p oposed app oacha e compa ed wi h he
cu en app oach o he PD de ec ion om he wi eless signal
p esen ed in [4].
The Fulnecek‘s app oach begins wi h da a ans o ma ion.
The p ep ocessing consis s o compu ing he signal s eepness
ins ead o using he signal ampli ude. The s eepness is a sim-
ple compu a ion o he slope o each indi idual inc easing o
dec easing pa o he ampli ude. Le assume Pis a se o n
indices o u ning poin s o he signal ampli ude se ies X. Then
he s eepness Y(Pi) o i∈[1,2,3,...,n]is de ined as:
Y(Pi)=X(Pi)−X(Pi−1)
Pi−Pi−1
.(1)
1[Online]. A ailable: h ps://code.i 4i.cz/ADAS/wi eless-pd-de ec ion

MARTINOVI ˇ
C AND FULNE ˇ
CEK: FAST ALGORITHM FOR CONTACTLESS PARTIAL DISCHARGE DETECTION ON REMOTE GATEWAY DEVICE 2127
Fig. 8. Example o ans o ma ion in o he s eepness se ies.
Example o he signal ampli ude and signal s eepness se ies is
shown in Fig. 8.
The Fulnecek‘s app oach o PD de ec ion hen de ec ed
ou lie s by means o z-sco e es ing. Finally, i checks whe he
he e is a clus e o ou lie s which a e close o each o he and he
size o he clus e is la ge han a se pa ame e . In he case ha
such a clus e is ound, he se ies is assumed o con ain he PD.
I i is no ound, hen he se ies is assumed o be wi hou PD.
Many measu es may be used o e alua e he e ec i eness
o he algo i hm classi ica ion capabili ies. These measu es a e
usually based upon di e en a ios o ue posi i e (TP), alse
posi i e (FP), ue nega i e (TN), and alse nega i e (FN) p e-
dic ions. In he case o pa ial discha ge de ec ion, he mos
impo an ea u e o he algo i hm is o educe he alse posi i e
p edic ion, e en a he cos o inc easing he alse nega i e
p edic ions. This is due o he ac ha sending a eam o check
on he cables in emo e loca ions is cos ly and i he algo i hm
makes a lo o alse posi i e calls, he ope a o will dis ega d he
wa nings gi en by such a sys em. The obse ed measu es a e:
P ecision - TP / (TP + FP)
Sensi i i y - TP / (TP + FN)
Speci ici y - TN / (FP + TN)
Accu acy - (TP + TN) / (TP + TN + FP + FN)
The aim is o maximize he sensi i i y and p ecision. O
cou se, hese wo measu es a e con adic ing each o he , since in
mos cases i is possible o inc ease one o he cos o dec eas-
ing he o he . The speci ici y and accu acy a e less impo an
and p e y high due o he ac ha he gi en da ase is e y
imbalanced as pa ial discha ges do no occu o en. Valida ion
o he concep was done using he R s a is ical so wa e [12] and
using he packages idy e se [13] and [14] o compu a ion and
isualiza ion.
A. Da a
The da ase consis s o 1700 ime se ies o signal ampli ude,
wi h 1520 se ies wi hou pa ial discha ge and 180 se ies wi h
pa ial discha ge. All signals we e checked by he human ope -
a o o he p esence o pa ial discha ges o make su e ha he
ime se ies a e labelled co ec ly. This alida ion was done based
on da a acqui ed by he gal anic me hod o PD de ec ion. The
chosen signals con ained a a ie y o di e en cha ac e is ics o
make he es s as gene al as possible. Fo example, he signals
TABLE I
PARAMETER SETTINGS USED FOR THE TEST
TABLE II
RESULTS FOR THE DATA EXCLUDED FROM THE LEARNING PART DUE TO
INSUFFICIENT KNOWLEDGE.(ABSOLUTE NUMBERS)
wi hou PD con ained o he ypes o discha ges, such as co ona,
so he model had o lea n o di e en ia e be ween PDs and
o he discha ges. Single se ies is 20 ms o da a measu ed wi h a
equency o 40 MS/s, esul ing in 800 000 obse a ions. These
se ies a e om 4 di e en places in Czech Republic esul ing in
sligh ly di e en cha ac e is ics o each loca ion.
B. Compa ison Wi h he S a e-o - he-A
Making a ai compa ison o all a ian s is qui e di icul
due o he undamen al di e ences o he wo app oaches. Ful-
necek‘s app oach is dependen only on se e al pa ame e s, bu
does no ely on lea ning, while he p oposed algo i hm equi es
o lea n he model be o e he possibili y o make a classi ica ion
on he new da a. The lea ning was made on 70% o he da a
which had ea u es om ou lie g oups and we e no classi ied
as oo noisy. To make he esul s mo e in o ma i e, es s we e
made wi h 300 di e en andom combina ions o aining and
es ing da a.
In Table I he e a e he se ings o all pa ame e s o bo h
me hods. These we e expe imen ally ound o wo k he bes
h ough he hype pa ame e sea ch.
Pa ame e αis a cu o o he z-sco e alues compu ed in
he Fulnecek‘s model, all he alues o z-sco e g ea e han α
a e conside ed o belong o discha ges, o ou lie in ou sense.
Pa ame e dis he same as in ou case and i de e mines he
dis ance o ou lie s which can be conside ed as pa o he same
clus e . In Fulnecek‘s model, i a se ies con ains a clus e wi h a
empo al leng h o a leas lmin i is conside ed ha i con ains a
pa ial discha ge. Pa ame e s cmin,k,εand ma e desc ibed in
de ail in Sec ion III. The de aul se ings o ee model lea ning
wi h logis ic bina y ou pu we e used o XGBoos models.
Fi s , in Table II he e is compa ison o he se ies which we e
classi ied as oo noisy (Excluded) and he ones which had no
ou lie g oups(No G oups).In hesecases,wewillassume ha
2128 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 37, NO. 3, JUNE 2022
TABLE III
RESULTS FOR THE PREDICTIONS.(COUNTS)
TABLE IV
RESULTS FOR THE PREDICTIONS.(MEASURES)
he e is no PD in he se ies as an ou pu o he p oposed model. I
is clea ha in he case o Excluded se ies nei he model could
de ec any o he 33 PD in he se ies, howe e Fulnecek‘s model
had 18 alse nega i e esul s. On he o he hand, o No G oup
se ies, he Fulnecek‘s app oach co ec ly de ec ed 6 PD ou o
35 and had only 1 alse posi i e. O e all, his p ocess excluded
68 se ies wi h he PD ei he due o he PD being in he noise o
being oo small o no ice in he o iginal da a.
The s abili y o he model is checked by he andomly ini-
ialized c oss- alida ion o 300 di e en se s o aining and
alida ion da a on he 921 se ies ha we e le . The aining
se con ained 645 se ies and he alida ion se con ained 276
se ies. Table III shows he coun s o ue posi i e, alse posi i e,
alse nega i e, and ue nega i e o he uns wi h bes and wo s
p ecision o each model. I is clea ha in he bes case, he
Fulnecek‘s model de ec ed 16 se ies wi h PD ou o 36 and has
15 alse posi i e p edic ions. The p oposed model could de ec
13 se ies wi h PD ou o 33 and had 0 alse posi i e p edic ions.
In he wo s case, he Fulnecek‘s model de ec ed 4 PDs ou o
26 and 27 alse posi i es p edic ions. The p oposed model in
he wo s case de ec ed 9 PDs ou o 31 and 13 alse posi i es
p edic ions.
Nex , he esul s a e p esen ed in e ms o he measu es shown
in Table IV. In he i s ou ows, he e is p ecision, sensi i i y,
speci ici y, and accu acy o he alues om he Table III.
The i h and six h ows a e he mean o he measu es om
he 300 samples. These a e also called c oss- alida ed measu es
and hey gi e an in ui ion abou he mos likely pe o mance in
p ac ice. The speci ici y and accu acy a e always high, due o
he high numbe o se ies wi hou PDs. In eali y, i he model
p edic ed ha he eisnoPDinallse ies, hesenumbe swouldbe
s ill qui e high due o he unbalance in he da ase . The e o e, he
ocus is on he p ecision and sensi i i y. Wi h he same model,
adjus ing i s pa ame e s usually leads o ade-o be ween hese
wo measu es, since he di e ence is ha one depends on alse
posi i es and he o he on alse nega i es. Usually, inc easing
one leads o dec easing he o he unless he model would ge
TABLE V
RESULTS FOR THE DATA EXCLUDED FROM THE LEARNING PART DUE TO
INSUFFICIENT KNOWLEDGE.(DATA TRANSFORM,ABSOLUTE NUMBERS)
TABLE VI
RESULTS FOR THE PREDICTIONS.(DATA TRANSFORM,COUNTS)
be e o e all. As was al eady men ioned, he mos impo an
measu e in his case is he p ecision, because i is impo an o
ha e as li le alse posi i es as possible.
F om he c oss- alida ed measu es, we can see ha on a e -
age, he p ecision o he Fulnecek‘s model was 32.77 and he
p oposed model had a p ecision o 68.9. The p oposed model
also had a be e sensi i i y o 42.6 on a e age compa ed o
30.98. Howe e , in case o he bes p ecision scena ios, he
Fulnecek‘s model had a li le be e sensi i i y.
C. Resul s Using T ans o med Da a
Based on he esul s om Table II, whe e he Fulnecek‘s
model could iden i y 6 PD while he p oposed could no , ano he
es was made. The p oposed model was un on he da a wi h he
same da a ans o ma ion as he Fulnecek‘s model. The esul s
con i med he heo y om Fulnecek‘s wo k, ha he s eepness
da a a e be e o wo k wi h han he aw da a om he senso .
A e some es ing o he pa ame e s, i was de e mined ha he
op imal pa ame e s o he p oposed model on he ans o med
da a a e he same as on he aw da a. Due o he di e en
se ies being excluded and ha ing no g oups, i is no possible o
make he di ec compa ison o he model pe o mance wi h and
wi hou da a ans o ma ion. In Table V he e is he compa ison
o he p oposed and Fulnecek‘s model on he Excluded and
No G oups se ies as was in Table II. In his case, Fulnecek‘s
model de ec ed one PD ou o 38 in he Excluded g oup and
had 8 alse posi i es. No G oupscon ained 28 se ies wi h PD
which nei he model de ec ed. Fulnecek‘s model had 2 alse
posi i es ou o 789 se ies wi hou PD. O e all, he e a e 66
se ies wi h PD which a e no de ec ed by he p oposed model in
con as o 68 when no da a ans o m was made.
In his case, 1036 se ies a e excluded and only 664 a e used
in he nex phase. Tha means 465 se ies we e used o aining
he models and 199 was used o alida ion. Table VI con ains
he coun s o he classi ica ion o he bes p ecision and wo s
p ecision samples as was in he Table III. The p oposed model
now has much be e esul s wi h he bes being 28 de ec ed
PDs ou o 43 wi h no alse posi i es. The wo s esul s o he
MARTINOVI ˇ
C AND FULNE ˇ
CEK: FAST ALGORITHM FOR CONTACTLESS PARTIAL DISCHARGE DETECTION ON REMOTE GATEWAY DEVICE 2129
TABLE VII
RESULTS FOR THE PREDICTIONS.(DATA TRANSFORM,MEASURES)
p oposed model a e 18 de ec ed ou o 24 and 9 alse posi i es.
The Fulnecek‘s model esul s a e simila o hose in Table III.
Table VII p esen s he same esul s as Table IV bu o he da a
ans o ma ion e sion. The a e age p ecision o he p oposed
modeljumped om68.9 o noda a ans o ma ion o87.52.The
sensi i i y was also inc eased signi ican ly om 42.6 o 72.7.
I is impo an o no e ha he inal p ecision and sensi i i y
a e lowe due o he se ies which we e excluded. Es ima ing he
eal p ecision in p oduc ion is di icul since i depends g ea ly
on he a io o signals wi h high noise and in his pape only
30% o he se ies which ge h ough he elimina ion p ocess
we e used o alida ion. O e all, accoun ing o he Excluded
and No G oup ime se ies in he inal esul s would ha e no
e ec on he p oposed model p ecision since he e is no ue
posi i e o alse posi i e obse a ions and i would educe he
sensi i i y sligh ly.
D. De ec ion Range
In his expe imen , all acqui ed HIFs we e si ua ed up o 2 km
om he de ec o . I is no easy o es ima e he ange o de ec ion.
Ampli ude o HF componen depends on wo main ac o s.
Humidi y o he ege a ion and aul -de ec o dis ance. Di ec
con ac be ween CC and esh b anch wi h ee sap gene a es
pulses wi h high epe i i e a es and ampli ude, which can be
easily de ec ed a a dis ance up o 1 km. On he o he hand,
con ac wi h a d y/dead ee canno be de ec ed, e en in a case
o close dis ance o he senso . I jus does no gene a e any
PD ele an pulses. The e is also a s ong backg ound noise
in luence. Ampli ude o he backg ound noise is highly a iable
in he examined equency band. This is caused by p ocesses
in he ionosphe e, so i a ies du ing he day ime. Usually, he
highes le el o backg ound noise is p esen ed du ing sun ise
and sunse . In such si ua ion, PD pa e n is “bu ied” in he noise
and i is di icul o de ec i .
V. CONCLUSION
In his pape , a new me hod o pa icle discha ge de ec ion o
he wi eless senso ocused on low compu a ional complexi y
and high p ecision was p esen ed. One o he ad an ages o his
app oach is ha i can be un di ec ly on he ga eway de ices
collec ing da a om he an enna, wi hou he need o send da a
o he se e . This me hod con ains a no el app oach o ou lie
de ec ion, which is alid o noisy da a wi h a iable s anda d
de ia ion o he noise, and a ou -s ep p ocess o aul de ec ion.
In he i s s ep, he ou lie s o he se ies a e de ec ed. The second
pa is he exclusion o all se ies which a e ound o be unsui able
o he classi ica ion model and a e au oma ically labelled as no
con aining pa ial discha ge. In his s age, a li le o e 6% o all
excluded se ies con ained PD. These PDs a e usually no s ong
enough o be isible in he se ies. In he hi d s age, he ime
se ies is classi ied based upon he ea u es compu ed om he
ou lie clus e s in he se ies. Due o he na u e o he inpu ,
he se ies is ans o med in o he s eepness se ies a he han
signal ampli udes. I was shown ha his p ocess inc eased he
p ecisiono classi ica ionbymo e han50%incompa ison o he
p e ious s a e-o - he-a me hod. The p ecision is 87.52% and
he sensi i i y is 72.7% on he alida ion se . When he excluded
ime se ies a e conside ed, he sensi i i y would be lowe , while
he p ecision would emain he same. This means ha he wi e-
less de ec o can be used as a cheap al e na i e o he mo e ex-
pensi e solu ions o suppo a la ge a ea co e age. In he u u e,
addi ional ea u es will be conside ed in he pu sue o u he
imp o e he quali y o classi ica ion. Addi ionally, i he co e -
age inc eases, a mul inode model could be conside ed o accoun
o se e al ime se ies ou pu s which a e geog aphically close.
In gene al, he p oposed con ac less me hod is less e ec i e
when di ec ly compa ed wi h commonly used adi ional me h-
ods. Fo example, he gal anic PD de ec ion me hod (unpub-
lished, bu deployed in p oduc ion a Czech ene gy dis ibu o s),
used o da a alida ionin hisexpe imen ,p o idesbe e esul s
(p ecision: 70%, sensi i i y: 98.4%, speci ici y: 99.8%, accu-
acy: 96.7%). The e ec i e ange o he con ac less de ec ion
is also sho e (4 km s. 1 km). The main ad an age o he
p oposed me hod lies in i s low cos ( 10% o cos in compa ison
o gal anic de ec o senso s), because he e a e no high ol age
pa s. This enables o co e a la ge a ea o dis ibu ion powe
g id, whe e adi ional me hods o PD online moni o ing a e no
cos -e ec i e.
REFERENCES
[1] P. Pakonen, “De ec ion o incipien ee aul s on high ol age co e ed
conduc o lines,” Ph.D. disse a ion, Tampe e Uni . Technol., Tampe e,
Finland, 2007. [Online]. A ailable: h ps:// epo. uni. i/bi s eam/handle/
10024/114541/pakonen.pd
[2] S. Lu, H. Chai, A. Sahoo, and B. T. Phung, “Condi ion moni o ing
based on pa ial discha ge diagnos ics using machine lea ning me hods: A
comp ehensi e s a e-o - he-a e iew,” IEEE T ans. Dielec . Elec . Insul.,
ol. 27, no. 6, pp. 1861–1888, Dec. 2020.
[3] S. Misak, J. Fulnecek, T. Jezowicz, T. Van uch, and T. Bu ianek, “Usage
o an enna o de ec ion o ee alls on o e head lines wi h co e ed
conduc o s,” Ad . Elec . Elec on. Eng., ol. 15, no. 1, 2017.
[4] J. Fulnecek and S. Misak, “A simple me hod o ee all de ec ion on
medium ol age o e head lines wi h co e ed conduc o s,” IEEE T ans.
Powe Del., ol. 36, no. 3, pp. 1411–1417, Jun. 2021.
[5] J.M. Ma ínez-Ta i a,J. A.A dila-Rey,A. P ie o-Duque,J.M. F esno,and
G. Robles, “S a is ical analysis o signals emi ed by pa ial discha ges,”
in P oc. IEEE In . Con . Dielec ., 2016, pp. 536–539.
[6] S.MišákandV.Poko ný,“Tes ingo aco e edconduc o ’s aul de ec o s,”
IEEE T ans. Powe Del., ol. 30, no. 3, pp. 1096–1103, Jun. 2015.
[7] N. D. Jacob and W. M. McDe mid, “Expe ience wi h pa ial discha ge
measu emen s on ins umen ans o me s in high ol age labo a o y ac-
cep ance es s,” in PES T. D. 2012, pp. 1–4.
[8] T. Chen and C. Gues in, “XGBoos : A scalable ee boos ing sys em,”
in P oc. ACM SIGKDD In . Con . Knowl. Disco . Da a Mining, 2016,
pp. 785–794.
[9] M. Es e , H.-P. K iegel, J. Sande , and X. Xu, “A densi y-based algo i hm
o disco e ing clus e s in la ge spa ial da abases wi h noise,” in P oc. 2nd
In . Con . Knowl. Disco . Da a Mining, 1996, pp. 226–231.
2130 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 37, NO. 3, JUNE 2022
[10] M. Kuhn and K. Johnson, Fea u e Enginee ing and Selec ion.NewYo k,
NY, USA: Chapman and Hall/CRC, 2020.
[11] “Kaggle,” Tampe e, Finland, 2020. [Online]. A ailable: h ps://www.
kaggle.com/
[12] R Co e Team, R: A Language and En i onmen o S a is ical Compu ing,
R. Founda ion o S a is ical Compu ing, Vienna, Aus ia, 2020. [Online].
A ailable: h ps://www.R-p ojec .o g/
[13] H. Wickham e al., “Welcome o he idy e se,” J. Open Sou ce So w.,
ol. 4, no. 43, 2019, A . no. 1686. [Online]. A ilable: h ps://joss. heoj.
o g/pape s/10.21105/joss.01686#
[14] T. Chen e al.,xgboos : Ex eme G adien Boos ing, 2020. [Online].
A ailable: h ps://CRAN.R-p ojec .o g/package=xgboos
Tomáš Ma ino iˇc ecei ed hePh.D. deg eein com-
pu e science om he VSB – Technical Uni e si y o
Os a a, Mo a skoslezsky, Czech Republic, in 2018.
He is cu en ly a Resea che wi h eam o Ad anced
Da a Analysis and Simula ion Lab, IT4Inno a ions,
VŠB-Technical Uni e si y o Os a a. His esea ch
in e es s include ope a ional esea ch, ime se ies
analysis, machine lea ning, deep lea ning, and non-
linea analysis o dynamical sys ems.
Jan Fulneˇcek was bo n in Ka iná, Czech Republic,
in 1989. He ecei ed he Ph.D. deg ee om he De-
pa men o Elec ical Powe Enginee ing, Technical
Uni e si y o Os a a, Mo a skoslezsky, Czech Re-
public, in 2018. He is cu en ly a Junio Resea che
wi h he ENET Resea ch Cen e and as an Assis an
P o esso wi h he Depa men o Elec ical Powe
Enginee ing, VŠB Technical Uni e si y o Os a a.
His esea ch in e es s include on elec ical equipmen
diagnos ics.