In
te
r
n
ation
a
l Jou
rn
al
o
f Po
we
r
Elec
tron
ic
s an
d
D
r
ive S
y
stem
(IJ
PED
S
)
V
o
l.
11, N
o.
1, Mar
ch 20
20,
p
p.
326~
3
3
2
IS
S
N
: 2088-
86
94,
D
O
I
:
10.11
59
1
/ij
ped
s
.
v11
.
i
1.pp
3
26-
33
2
326
Jou
rn
a
l
h
o
me
pa
ge
:
ht
tp:
//i
j
p
eds.i
a
esco
re
.com
Effects of shorter phase-resolve
d partial disch
a
rge duration on
PD classification accuracy
C
h
on
g Wan
X
i
n
1
, Won
g Jee
Kee
n
R
aymon
d
2
,
Ha
z
l
ee
A
z
il Il
lias
3
,
La
i Wen
g
Kin
4
,
Y
i
a
u
w
Ka
h
H
a
ur
5
1
,
2,
4,
5
Dep
a
rtm
e
nt
o
f
El
e
c
t
r
ical
a
n
d
E
l
ectro
nics
E
ngi
neeri
ng,
Tu
nku
A
b
d
ul
R
ahm
a
n U
n
i
v
ersity
Co
lleg
e
, M
alay
s
i
a
3
Depart
ment
of
E
l
ectr
i
cal
E
n
g
in
e
e
ri
n
g
, Un
i
versit
y o
f
Ma
l
ay
a, Ma
l
ays
i
a
Art
i
cl
e In
fo
ABSTRACT
A
r
tic
le hist
o
r
y
:
Re
ce
i
v
e
d
Ju
l
1
7,
201
9
Re
vise
d O
c
t
2,
201
9
Ac
ce
p
t
ed
No
v
2
3
,
2
019
Parti
a
l
d
i
scharge
(PD)
p
attern
r
ecognit
ion
i
s
u
sef
u
l
to
d
iagnos
e
insulat
i
on
con
d
i
t
i
on.
P
D
meas
urem
ent
data
i
s
comm
only
rep
r
esent
e
d
in
phas
e
-resolved
p
a
r
t
i
a
l
d
i
s
c
h
a
r
g
e
(
P
R
P
D)
f
o
r
m
a
t.
P
R
P
D
is
u
s
e
fu
l
a
s
i
t
pr
o
v
i
d
e
s
a
v
i
si
b
l
e
pat
t
ern
f
o
r
dif
f
e
ren
t
P
D
so
urce
a
n
d
vari
ous
f
ea
t
u
res
can
b
e
ext
ra
cted
f
o
r
P
D
pat
t
ern
reco
gni
tion
.
Sh
ort
e
r
P
R
PD
d
urati
on
will
e
nab
l
e
m
o
re
t
ra
i
n
in
g
dat
a
bu
t
t
h
e
i
n
f
o
rm
atio
n
in
each
d
ata
i
s
l
ess
and
vi
ce
v
ers
a
.
This
w
or
k
s
a
i
m
s
to
in
ves
t
i
g
ate
t
h
e
ef
f
ects
of
u
s
i
n
g
v
ery
s
hort
du
rati
on
P
R
P
D
d
at
a
on
t
h
e
accur
acy
o
f
P
D
p
att
e
rn
r
ecog
n
i
t
i
on.
T
h
e
r
esu
l
t
s
c
on
clu
d
e
t
h
at
m
a
c
hine
learn
i
n
g
m
od
e
l
s
s
u
ch
a
s
Artifici
al
N
eu
ral
Netwo
r
k
(AN
N
)
a
n
d
S
u
p
port
Vect
or
M
ach
in
e
(S
VM
)
are
robu
st
e
n
oug
h
such
t
hat
redu
cti
o
n
of
P
RP
D
du
rati
on
f
r
o
m
1
5-s
econ
d
s
to
1
-secon
d
causes
l
e
ss
th
an
5
%
d
ro
p
in
t
he
clas
sificati
o
n
accuracy
.
H
o
wev
e
r,
t
h
i
s
i
s
only
tru
e
f
or
n
o
i
s
e
f
ree
con
d
i
t
i
on.
Wh
en
t
h
e
s
am
e
P
D
d
ata
i
s
o
v
e
rlap
ped
w
i
t
h
r
andom
n
o
i
se,
t
h
e
clas
sificati
on
accur
acy
s
uf
f
e
rs
a
s
ig
nifi
cant
red
u
ct
io
n
up
t
o
19
%
.
T
h
e
ref
o
re,
lo
n
g
e
r
P
R
P
D
du
rati
on
is
r
ecom
m
e
nd
ed t
o
withs
t
and
t
h
e
eff
ects
of
noise
c
on
t
a
mina
tion
.
K
eyw
ord
s
:
Parti
a
l disch
a
r
g
e
Pat
t
ern
re
c
o
gnit
i
on
PRPD
Th
is
is a
n
o
p
en acces
s a
r
ti
cle u
n
d
e
r t
h
e
CC
B
Y
-S
A
li
cens
e
.
Corres
pon
d
i
n
g
Au
th
or:
Wo
n
g
Je
e
K
ee
n
Ra
ymon
d,
D
e
pa
rtme
nt
o
f
El
e
c
t
rica
l
a
n
d
El
ect
ro
ni
c
s
Eng
in
e
e
ring
,
Tu
nk
u A
b
d
u
l
Ra
h
m
a
n U
n
ive
r
si
t
y
C
ol
le
ge,
S
t
.
G
e
n
t
i
n
g K
e
lan
g
,
S
e
ta
pak,
5
3
3
00, K
ua
la Lum
p
u
r,
Malays
i
a.
Em
ail:
w
o
n
g
jk
@tar
c.
edu.m
y
1.
I
N
TR
OD
U
C
TI
O
N
I
n
su
lat
i
o
n
fa
i
l
u
r
e
in
e
lec
t
rica
l
pow
e
r
s
ystem
com
pone
nts
w
i
ll
c
a
u
s
e
cata
s
tr
op
h
i
c
dam
a
ge
.
There
f
ore
,
it
i
s
i
m
p
o
rta
n
t
t
o
fre
que
n
t
l
y
m
oni
tor
the
i
n
s
u
la
t
i
o
n
q
ua
li
t
y
.
S
ince
P
D
me
a
s
ure
m
e
n
t
is
a
n
on
des
t
ruc
t
i
v
e
te
st
,
i
t
is
w
i
d
e
l
y
use
d
f
or
i
ns
ula
t
i
o
n
c
ond
it
i
o
n
asse
ssm
ent
[1-3]
.
P
D
is
d
e
f
in
ed
a
s
elec
t
r
i
c
a
l
d
i
s
char
ge
t
hat
p
a
rti
a
lly
bri
dges
the
ins
u
la
t
i
on
a
c
c
o
r
d
i
ng
to
I
E
C
6
0
2
70
[4].
D
e
s
p
i
t
e
o
n
l
y
p
a
r
t
i
a
lly
b
ri
d
g
i
n
g
t
h
e
ins
u
l
a
ti
on,
P
D
w
i
l
l
ca
use
e
v
en
t
u
al
i
ns
u
l
a
t
i
on
bre
a
kd
ow
n
if
l
e
f
t
un
de
tec
t
e
d
.
If
P
D
c
an
b
e
de
tec
t
e
d
a
t
a
n
i
n
c
ip
ie
nt
s
ta
ge
,
uti
l
i
t
y
com
p
an
ies ca
n
avo
i
d
e
xpe
ns
ive
e
l
e
c
t
rica
l
eq
ui
pm
en
t
fa
ilur
e
s
[5,
6]
.
Eac
h
in
s
u
la
ti
o
n
d
e
f
e
c
t
h
as
i
ts
o
w
n
un
i
que
di
sc
harge
a
t
t
r
ibu
t
es,
w
h
ic
h
c
a
n
be
u
sed
to
t
ra
i
n
m
achi
n
e
l
e
a
r
ni
ng
m
o
del
s
t
o
ide
n
t
i
fy
t
h
e
d
e
f
ec
t
t
y
pe
b
ase
d
o
n
the
me
asured
P
D
pa
t
t
e
r
n
[
7
].
S
uc
h
P
D
c
lassifiers
w
i
l
l
g
re
a
tly
f
a
c
i
lita
te
t
h
e
i
nsul
a
t
i
on
c
o
nd
iti
o
n
m
on
ito
r
i
n
g
o
f
elec
tr
ical p
ow
e
r
c
ompone
nt
s
a
t
low
c
os
t
an
d
effic
i
e
n
t
ma
nn
er.
P
R
P
D
i
s
t
h
e
m
o
s
t
w
i
d
e
l
y
u
s
e
d
r
e
p
r
e
s
e
n
t
a
t
i
o
n
f
o
r
P
D
[
8
,
9
]
.
I
n
o
r
d
e
r
t
o
o
b
t
a
i
n
a
P
R
P
D
d
a
t
a
repr
esenta
t
i
o
n
,
a
P
D
d
etec
tor
is
r
eq
uire
d
t
o
m
e
a
sure
a
c
on
tin
uo
us
s
trea
m
of
P
D
pu
lse
s
.
Ea
ch
i
n
d
i
v
i
du
a
l
P
D
pu
lse
w
i
ll
b
e
q
u
a
n
tifie
d
in
t
o
t
he
phas
e
a
ng
le
(
ϕ
),
c
harge
m
a
gnit
ude
(
q
)
a
n
d
t
h
e
nu
mb
e
r
o
f
PD
o
cc
u
rren
ce
(
n
).
Be
ca
use
o
f
t
h
i
s,
P
RP
D
i
s
a
lso
kn
ow
n
as
ϕ
-q-
n
pa
t
t
ern
[1
0].
P
R
P
D
can
b
e
re
p
r
esen
ted
a
s
a
3
-dime
n
si
o
n
al
d
at
a
arr
a
y, 3D
figur
e,
or
2D
im
a
ge
w
it
h
co
lor c
o
n
t
o
u
r
.
Evaluation Warning : The document was created with Spire.PDF for Python.
Int J
P
o
w
E
l
e
c
&
D
ri S
yst
IS
S
N
:
2088-
86
94
Ef
fec
t
s
o
f
sh
ort
e
r
ph
ase
-
reso
l
v
e
d
pa
rti
a
l d
i
s
c
ha
rge
d
u
rat
i
o
n on PD
clas
sif
i
ca
t
i
o
n
acc
u
ra
c
y
(Cho
ng
W
a
n Xi
n)
32
7
The
P
D
m
ea
s
u
r
e
m
e
nt
d
urat
i
on
to
g
e
n
era
t
e
a
sing
le
P
RP
D
re
prese
n
t
at
io
n
i
s
n
ot
s
t
a
nd
a
r
di
ze
d
and
di
ffe
re
nt
dura
tio
n
h
a
s
bee
n
u
se
d
b
y
r
e
s
e
a
rc
hers
f
or
P
D
c
l
a
ssifi
c
a
tio
n
relate
d
rese
arc
h
.
F
o
r
exa
m
ple
,
300
seco
nds [1
1
], 120 se
c
o
n
d
s [1
2
]
,
60 seco
n
d
s [9, 13], 50 sec
o
nds
[1
4],
a
nd 3
se
con
d
s o
n
l
y [
1
5
]
. Thi
s w
o
r
k
a
ims
to
i
n
v
e
s
ti
ga
te
t
he
e
ffe
c
t
s
o
f
u
si
n
g
v
e
r
y
s
hor
t
P
R
P
D
dur
a
tio
n
on
P
D
class
i
fica
tio
n
a
c
c
ura
c
y.
A
d
urat
i
o
n
o
f
1
-
seco
nd
t
o
1
5
-sec
onds
w
a
s
c
h
o
se
n
t
o
t
e
s
t
t
h
e
rob
u
s
t
ne
ss
of
m
a
c
h
i
n
e
l
earn
i
ng
m
od
el
s
t
o
r
e
c
ogn
i
z
e
t
h
e
PD
source
w
he
n
pr
ov
i
d
ed w
ith j
u
s
t 1
to
15 sec
onds
o
f P
R
P
D
p
a
t
t
e
r
n
.
Tw
o
gr
ou
p
s
o
f
P
D
d
ata
sour
c
e
w
e
r
e
use
d
f
o
r
t
h
i
s
w
o
r
k
.
Th
e
firs
t
g
r
o
u
p
c
onsis
ts
o
f
3
la
b
fa
br
i
c
a
t
e
d
in
su
la
ti
o
n
m
a
t
e
r
ials,
w
h
ich
p
r
ov
ide
a
m
o
re
c
on
sis
t
e
n
t
P
R
P
D
p
a
t
t
e
r
n
w
h
il
e
t
h
e
se
c
o
nd
g
r
o
up
c
o
n
s
i
st
s
of
5
P
R
P
D
p
a
t
t
e
rn
m
easure
d
f
rom
Cross-l
i
n
k
ed
P
olye
t
hyle
n
e
(X
LP
E)
c
a
b
le
j
oin
t
d
e
f
ec
ts,
w
h
i
c
h
pro
v
ide
s
a
m
ore
inc
o
ns
i
s
te
n
t
P
RP
D
pat
t
ern.
C
ompa
rin
g
the
resul
ts o
f b
o
t
h
grou
ps
w
i
l
l
g
i
v
e a more
c
ompr
ehe
n
si
ve vi
e
w
of
t
h
e
effec
t
o
f r
e
du
ci
ng t
h
e
P
R
P
D
dura
tio
n.
The
P
R
P
D
d
u
r
a
t
i
on
d
i
re
ct
ly
c
orr
e
lates
to
t
he
n
um
ber
of
P
D
occ
u
r
r
i
n
g
p
e
r
m
e
a
s
u
r
e
m
e
n
t
.
W
h
e
n
fe
at
ures
e
x
t
r
acted
f
rom
PRPD
p
at
t
e
rn
w
e
r
e
use
d
f
or
c
la
ssi
f
i
c
a
t
i
o
n
,
th
e
a
ccur
acy
d
e
p
end
s
o
n
a
va
ri
ety
o
f
f
a
ctors.
S
i
n
ce
t
his
wor
k
i
s
f
o
c
u
si
ng
o
n
ex
a
m
i
n
i
n
g
th
e
ef
fect
s
o
f
sh
o
r
ter
P
R
P
D
durat
i
on,
t
he
o
ther
f
ac
t
o
rs
a
r
e
kep
t
c
ons
ta
nt.
In
o
t
h
er
w
ords
,
the
t
ype
o
f
fe
ature
ex
trac
ti
o
n
p
erf
o
rm
ed,
the
n
u
m
b
er
o
f
tr
ain
i
ng
a
n
d
t
e
st
d
a
t
a
,
as
w
ell
as
t
he
c
lass
ifier
h
ype
r
p
ara
m
e
t
e
r
s
re
m
a
ins
t
h
e
sam
e
w
hi
l
e
on
ly
v
aryi
ng
th
e
P
R
P
D
dura
t
i
o
n
to
obs
e
r
ve
its
e
ffec
ts on
PD
p
attern
r
ec
o
gni
ti
o
n
ac
c
ura
c
y.
The
re
ma
i
n
de
r
of
t
he
p
a
p
er
i
s
orga
ni
z
e
d
a
s
f
o
ll
ow
s;
S
e
c
t
i
o
n
II
de
sc
ri
b
e
s
t
h
e
o
v
e
r
all
exper
i
m
e
nt
set
up,
w
h
i
c
h
c
ove
rs
t
he
P
D
me
asur
e
m
e
n
t
s
e
tu
p,
P
D
so
ur
ce
pre
p
ara
tio
n,
a
nd
r
an
dom
n
o
i
se
d
a
t
a
use
d
.
S
e
c
tio
n
III
describes
the
PD
c
las
s
ifica
t
io
n
proce
d
ure,
w
h
i
ch
i
ncludes
f
e
a
t
u
re
e
xtra
cti
o
n
a
nd
P
D
c
lass
i
f
i
e
r.
T
he
r
esu
lts
& disc
ussi
on
a
r
e
incl
ude
d i
n
S
ec
ti
o
n
I
V
w
h
ile
S
e
c
tio
n
V
pr
ov
ide
s
th
e
co
nc
l
u
si
on
f
o
r
t
hi
s wo
rk
.
2.
RESEARCH
M
ETH
O
D
2.1.
PD
m
easur
e
me
nt setup
A
c
o
m
m
e
r
c
i
a
l
P
D
d
e
t
e
c
t
o
r
,
w
h
i
c
h
c
o
m
p
l
i
e
s
w
i
t
h
t
h
e
I
E
C
6
0
2
7
0
s
t
a
n
d
a
r
d
,
w
a
s
u
s
e
d
i
n
t
h
i
s
w
o
r
k
.
The
PD
d
e
t
ect
or
i
s
ab
le
t
o
d
i
sp
lay
the
PRPD
p
atter
n
i
n
rea
l
t
i
m
e
and
th
e
data
c
a
n
b
e
exp
o
rte
d
t
o
a
P
C
f
o
r
furt
her
proc
ess
i
n
g
.
A
blo
c
k d
i
agra
m
of
t
he
m
ea
sur
e
m
e
nt set
up
is
s
how
n
in
F
igure
1.
The
H
V
source
is a
ste
p
-u
p t
r
a
n
sform
e
r
ca
pable
of s
u
p
p
l
yin
g
u
p
to 200 kV
. The
m
ea
suring
c
apacitor
me
asure
s
t
he
v
ol
ta
ge
s
u
p
p
l
ie
d.
T
he
c
ou
p
l
i
n
g
c
a
p
ac
itor
w
i
ll
t
r
an
s
f
er
a
n
a
ppar
e
nt
c
ha
r
g
e
to
t
he
t
es
t
o
b
j
ect
t
o
st
a
b
il
ize
the
v
o
lta
ge
w
he
ne
v
e
r
it
detec
t
s
a
vol
tage
d
ro
p
d
u
e
t
o
P
D
oc
cur
r
e
n
ce.
T
his
da
t
a
i
s
pa
sse
d
t
o
t
he
P
D
detec
t
or
and
t
h
e
US
B
con
t
rol
l
e
r
hand
les the
data
t
ra
nsfe
r b
e
t
w
e
en
t
he
P
D de
t
e
c
t
or a
nd t
h
e
P
C
.
F
i
gur
e 1.
Bl
o
c
k
dia
gr
am
of
PD
m
easure
m
e
n
t
set
u
p
2.2.
PD
s
ou
rc
e p
repar
a
t
ion
Tw
o
grou
ps
o
f
P
D
s
o
u
rc
es
w
ere
pr
epar
ed
a
n
d
c
om
pare
d
in
t
his
w
o
r
k
.
P
D
G
r
o
u
p
1
c
o
n
s
i
s
t
s
o
f
3
classes
o
f
P
D
w
h
ic
h
a
r
e
vo
id,
coro
na
a
n
d
s
urfac
e
d
i
s
c
h
a
r
ge
m
ea
s
ur
ed
f
rom
l
a
b
fa
brica
t
ed
l
ow
-
d
ens
i
t
y
po
lye
t
hyle
n
e
(
L
D
P
E).
T
h
e
de
t
a
i
l
s
o
f
the
sam
ple
pr
e
p
ar
at
ion
an
d
me
asure
m
e
n
t co
n
d
i
tio
n c
a
n be fo
u
nd i
n
[1
6
]
.
P
D
G
r
oup
2
c
o
ns
is
ts
o
f
5
cla
s
se
s
of
P
D
so
u
r
ce
m
ea
sured
fr
om
X
LP
E
c
a
b
l
e
j
o
in
t
s
w
i
t
h
a
rtific
ia
l
de
fe
cts
.
T
he
artific
ia
l
de
fec
t
s
i
n
c
l
ude
i
nc
i
s
i
on
o
n
i
nsu
l
at
io
n
l
a
y
e
r,
m
eta
l
l
ic
p
art
i
c
l
es
on
ins
u
la
t
i
o
n
l
aye
r
,
roug
h
e
d
ges
a
t
sem
i
c
o
nd
uc
tor
la
ye
r,
a
ir
g
ap
a
t
sem
i
c
o
nd
uc
t
o
r
la
ye
r
an
d
o
f
f-ax
i
s
j
o
i
n
t
insta
lla
t
i
o
n
.
M
o
r
e
i
nform
a
ti
on
a
bou
t
the sa
mp
le
p
repa
rat
i
o
n
c
an be
fou
n
d
i
n
[17].
P
D
G
roup
1
cons
is
t
o
f
6
6
P
D
d
ata
w
h
er
e
eve
r
y
3
cla
sses
have
22
da
t
a
eac
h.
P
D
Grou
p
2
co
ns
i
s
t
of
10
0
P
D
d
a
t
a
w
h
er
e
eve
r
y
5
c
l
asses
ha
ve
2
0
data
e
ac
h.
F
ig
ure
2
s
h
o
w
s
o
n
e
e
x
a
m
p
l
e
o
f
v
o
i
d
P
R
P
D
f
r
o
m
P
D
G
r
oup
w
h
i
l
e
F
i
g
u
re
3
s
how
s
one
e
xam
p
le
o
f
i
n
c
i
si
o
n
d
e
f
e
c
t
f
rom
P
D
G
roup
2
at
1
-
s
e
c
on
d
a
nd
15-se
c
ond
s
dura
t
i
o
n.
T
he
x
-
a
xis
r
e
pre
s
e
n
ts
t
he
p
hase
a
ng
le
o
f
t
h
e
P
D
o
cc
urre
n
c
e,
t
he
y
-ax
i
s
r
e
pre
s
en
ts
t
he
c
ha
rg
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N: 2
0
8
8
-
86
94
I
nt
J
P
o
w
E
l
e
c
&
D
r
i
S
yst
V
o
l.
11,
N
o.
1
,
Mar
202
0
:
326
–
33
2
32
8
m
a
gni
t
ude
o
f
t
h
e
P
D
w
hile
t
he
n
um
ber
of
P
D,
a
lso
k
now
n
a
s
t
he
P
D
in
t
e
n
s
it
y
i
s
r
epr
e
sen
t
ed
by
t
h
e
col
o
r
g
a
m
u
t
at
t
h
e
s
id
eb
a
r
.
(a)
(b
)
F
i
g
u
r
e
2
.
V
oid
P
R
P
D
f
r
o
m
P
D
G
r
oup
1,
(
a
)
1-
s
eco
nd
(
b
)
15-
seco
nd
s
The
1
5
-
s
ec
o
n
d
s
d
ur
a
tio
n
P
R
P
D
i
s
ma
de
u
p
of
a
c
o
n
tin
u
ous
c
om
bina
t
i
o
n
o
f
1
5
u
ni
ts
o
f
1
-
se
cond
P
R
P
D
.
The
r
a
ti
o
n
a
l
e
f
o
r
us
i
n
g
t
h
ese
tw
o
gr
ou
ps
o
f
P
D
s
o
u
r
c
e
is
to
b
e
t
t
e
r
obser
ve
t
he
i
mpa
c
t
o
f
u
si
n
g
sho
r
t
e
r
PR
PD
d
u
r
a
t
i
o
n
.
P
D
G
r
oup
1
ha
s
a
m
o
r
e
c
onsis
ten
t
P
D
pa
t
t
e
r
n
and
ju
st
3
d
if
fe
r
e
n
t
cla
s
se
s.
C
on
ve
r
s
el
y,
P
D
G
r
oup
2
ha
s
a
mor
e
i
nc
onsis
te
nt
P
D
pa
t
t
er
n
an
d
5
d
i
ff
er
e
n
t
c
l
a
s
se
s.
W
i
t
h
t
h
e
sa
me
P
RP
D
dur
a
t
ion,
i
t
is
e
x
p
ec
t
e
d
t
h
at
P
D
G
roup
1
will
b
e
easier
t
o
clas
s
i
fy
a
nd
h
ence
S
VM
a
nd
ANN
can
a
c
h
i
e
ve
h
i
gher
c
l
a
s
sif
i
c
a
t
i
on
acc
u
r
ac
y.
Whe
n
r
educ
in
g
t
h
e
P
R
P
D
d
ur
ati
o
n
f
o
r
P
D
G
r
oup
1
,
it
ca
n
be
s
ee
n
t
h
at
a
lt
hou
gh
t
h
e
P
D
in
t
e
n
s
it
y
i
s
d
i
f
f
e
re
nt
f
o
r
1
sec
o
nd
a
nd
1
5
s
e
c
o
n
d
s,
t
he
g
e
n
e
r
al
s
ha
pe
i
s
sim
i
l
a
r
.
S
inc
e
the
oppos
ite
i
s
t
rue
for
G
r
oup
2,
t
hi
s
will
m
a
ke
it
m
o
r
e
c
ha
lle
ng
i
ng
to
b
e
c
l
a
ssif
i
e
d
w
h
e
n
the
P
R
P
D
d
ur
at
io
n is re
d
u
ced.
(a)
(
b
)
F
i
gur
e
3.
I
ncisi
on
de
f
e
c
t
f
r
o
m
P
D
G
r
oup
2,
(
a
)
1
-
s
e
c
ond
(
b
)
15-
s
eco
nd
s
I
n
o
rd
e
r
t
o
i
n
v
e
s
t
i
g
at
e
th
e
r
o
b
u
st
n
e
s
s
o
f
t
h
e
PD
c
l
a
s
s
if
i
e
r
u
n
d
e
r
n
oise
c
o
n
t
am
i
n
a
t
i
o
n,
r
a
n
dom
noise
m
e
a
s
ur
ed
f
r
o
m
gr
ou
nd
in
ter
f
e
r
ence
w
a
s
u
sed
t
o
over
l
ap
t
he
c
lean
P
R
P
D
p
a
t
t
e
rn
.
Fo
r
ex
a
m
p
l
e
,
T
dur
ati
on
of
P
R
P
D
w
i
ll
be
over
l
a
p
pe
d
w
i
t
h
T
dur
a
tion
o
f
r
andom
n
o
i
se
t
o
ge
ner
a
te
a
n
oi
se
c
on
tam
i
na
t
e
d
P
R
P
D
d
at
a. T
h
e
P
D
cl
a
ssifi
er w
i
l
l b
e
train
e
d
usin
g
clea
n
P
R
P
D
d
at
a b
u
t tes
t
ed
agai
ns
t c
ont
a
m
inat
ed
P
RPD
data.
An
e
xample
of
r
a
n
d
o
m
nois
e
P
RPD
is
s
how
n
in
F
ig
ur
e
4.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t
J
P
o
w
Elec
&
D
r
i
S
y
st
I
S
S
N
:
2088-
86
94
Ef
f
e
cts o
f
sh
or
t
e
r ph
ase-
re
so
lv
ed pa
rt
i
a
l
d
i
sc
ha
rge
d
u
r
a
t
i
o
n
on PD
clas
si
f
i
ca
t
i
o
n
acc
u
r
a
c
y
(
C
ho
ng
W
a
n Xi
n)
32
9
F
i
g
u
r
e
4
.
Ran
d
o
m
noi
se
f
or
15-
seco
nd
s
3.
PD
C
LAS
S
IF
ICAT
ION
3.
1.
F
e
at
u
re
ext
r
act
i
on
The
ra
w
P
D
d
a
t
a
a
n
d
PRPD
i
s
to
o
large
t
o
b
e
use
d
a
s
i
n
p
u
t
f
ea
tu
re
t
o
trai
n
the
P
D
c
las
s
i
f
ier.
H
ence
,
f
eat
ur
e
e
x
t
r
act
i
on is
r
eq
uir
e
d
t
o
o
bta
i
n
a
use
f
ul
r
e
p
r
e
se
nta
t
i
o
n of
the P
RP
D
patt
er
n.
T
he
e
xtr
a
c
t
e
d
fe
a
t
u
r
e
s ar
e
a
l
s
o
k
now
n
a
s
“
P
D
f
i
nge
r
p
r
i
n
t
”.
T
he
P
RP
D
can
b
e
sor
t
e
d
i
nt
o
tw
o
p
r
i
m
a
r
y
d
is
t
r
ibu
t
io
ns
H
n(
φ)
a
nd
H
q
n(
φ
)
.
H
n
(
φ
)
is
a
2
-
D
p
lo
t
of
P
D
i
n
te
ns
i
t
y
vs
pha
se
o
cc
ur
r
e
nc
e
w
h
i
l
e
H
qn(
φ)
i
s
a
2-
D
plo
t
o
f
P
D
c
har
g
e
m
a
g
n
i
t
ude
vs
phase
o
cc
u
r
r
e
nc
e.
T
hese
t
w
o
d
istr
ib
u
t
i
o
ns
c
a
n
t
he
n
b
e
d
iv
i
d
ed
i
nto
an
ot
h
e
r
tw
o
sepa
r
a
te
d
i
s
tr
i
b
u
t
i
o
ns
ba
se
d
o
n
t
he
p
os
i
t
i
v
e
a
n
d
ne
gat
i
v
e
ha
lf
o
f
t
h
e
p
h
ase
c
y
cle
.
F
ou
r
st
at
i
s
ti
ca
l
fe
atu
r
es
s
u
c
h
a
s
M
e
a
n
,
V
ari
a
n
c
e,
K
u
r
t
os
is,
a
nd
S
k
e
w
ness
c
a
n
be
c
a
l
c
u
la
te
d
f
r
om
a
ll
f
o
u
r
d
i
str
i
bu
t
i
o
n
s
t
o
f
o
r
m
a
t
o
t
a
l
o
f
1
6
f
e
a
t
u
r
e
s
f
o
r
e
a
c
h
PR
PD
d
at
a
.
T
h
e
K
u
r
t
o
s
i
s
a
n
d
Sk
e
w
n
e
s
s
c
a
n
b
e
c
a
l
cu
l
a
t
e
d
b
y
u
si
n
g
t
he
f
o
l
l
o
w
i
ng
f
o
r
m
ula
s
:
∑
∑
3
(
1
)
∑
∑
(
2
)
wh
ere
N
i
s
t
h
e
t
o
ta
l
d
a
t
a
s
iz
e
,
f(x
i
)
i
s
t
h
e
fu
nc
ti
on
o
f
in
ter
e
st,
a
n
d
x
i
i
s
t
h
e
in
d
i
v
i
dua
l
di
scr
e
t
e
v
a
l
ue
o
f
t
h
e
di
str
i
b
u
t
i
on.
A
c
omple
t
e
m
a
th
em
atica
l
d
escr
i
p
t
i
on
o
f
K
ur
to
sis
a
n
d
S
kew
n
e
s
s
can
b
e
fo
un
d
in
[
1
8-
20]
.
3.
2.
P
D
class
ifier
Tw
o
com
m
o
n
l
y
use
d
m
a
c
hi
n
e
l
ea
r
n
i
ng
clas
sif
i
er
w
a
s
u
sed
for
t
h
is
w
or
k
a
s
t
he
P
D
c
l
assi
f
i
er
,
A
N
N
[
2
1
-
2
3
]
an
d
S
V
M
[
24-
27]
.
U
s
ual
l
y,
t
he
t
ot
a
l
i
np
u
t
d
ata
w
i
l
l
b
e
di
v
i
d
e
d
i
n
t
o
t
r
a
in
i
n
g
&
te
st
in
g
da
ta.
Th
e
c
l
a
ssif
i
e
r
w
il
l
be
t
r
a
i
n
e
d
u
si
ng
the
tr
a
i
n
i
ng
da
ta
a
n
d
t
e
s
t
e
d
u
s
i
ng
t
h
e
te
st
i
n
g
da
ta.
F
o
r
t
h
is
w
or
k,
t
he
pe
r
f
or
m
a
nc
e
o
f
t
he
P
D
cla
ssi
fier
w
a
s
e
va
lua
t
ed
u
s
i
n
g
K
-
f
o
l
d
c
r
o
ss-
v
al
ida
tio
n.
T
he
i
n
p
u
t
data
w
e
r
e
r
a
nd
oml
y
di
v
i
de
d
in
to
K
numbe
r
o
f
s
ets,
t
he
f
irs
t
s
e
t
w
ill
be
u
se
d
for
t
e
s
tin
g
wh
i
l
e
t
h
e
o
t
h
e
r
s
e
t
s
w
i
l
l
b
e
u
s
e
d
f
o
r
t
r
ai
n
i
ng
.
T
h
is
p
ro
ce
ss
was
rep
e
at
e
d
K
num
ber
of
tim
es
w
her
e
e
ac
h
se
t
w
i
l
l
t
a
k
e
a
tur
n
t
o
be
u
se
d
onc
e
as
tes
t
i
n
g
da
t
a
.
The
ave
r
a
g
e
cla
ssi
f
i
ca
t
i
o
n
a
cc
ur
acy
i
s
t
h
e
n
c
alcu
la
te
d.
F
or
P
D
Gr
oup
1,
1
1-
f
o
ld
c
r
o
ss-
v
a
lida
t
i
o
n
w
a
s
use
d
w
h
i
l
e
1
0-
f
o
ld
c
r
o
ss
-
v
al
ida
t
io
n
w
a
s
use
d
f
or
P
D
G
r
oup
2
.
This
K
numbe
r
w
a
s
c
h
osen
s
o
t
h
a
t
e
a
c
h
fo
ld
c
on
ta
ins
t
h
e
sa
me
num
ber
of
d
ata from
ea
ch
c
lass.
The
be
ne
f
it
of
u
s
i
ng
K
-
f
o
l
d
c
r
oss-
val
i
d
at
io
n
is
t
o
a
v
o
i
d
over
f
i
t
t
i
n
g
a
n
d
se
l
e
c
t
ion
b
i
as
.
In
o
rd
er
t
o
obs
er
ve
t
he
p
e
r
for
m
anc
e
o
f
the
P
D
c
lassi
f
i
e
r
w
he
n
usi
n
g
n
o
i
s
e
c
o
n
tam
i
na
ted
da
t
a
,
the
cl
assifier
w
a
s
t
r
a
in
e
d
us
in
g
cle
a
n
in
pu
t
data,
an
d
the
tes
t
d
ata
w
a
s
ove
r
l
ap
pe
d
w
ith
n
o
i
se
d
a
t
a
pr
i
o
r
to
t
es
ting.
T
h
i
s
w
i
l
l
p
r
oper
l
y
ga
u
g
e the
ca
pa
bi
l
i
t
y
of the P
D
cla
ssi
f
ier
to r
ecog
n
iz
e
c
o
nta
m
in
at
ed
i
npu
t
d
a
t
a
t
h
a
t
wa
s
n
o
t
seen
b
efo
r
e
d
u
r
ing
the
tr
a
i
n
i
ng
pr
oc
ess.
4.
RESU
L
T
S
A
ND DIS
C
U
S
S
I
ON
The
e
ffe
c
t
s
of
r
educ
i
n
g
P
R
P
D
dur
a
t
i
o
n
o
n
b
o
t
h
P
D
G
r
o
u
p
1
a
nd
P
D
G
r
o
u
p
2
a
s
w
e
l
l
a
s
t
h
e
e
f
f
e
c
t
s
of
n
o
i
se
c
o
n
ta
minat
i
on
ar
e
s
how
n
in
F
ig
ur
e
5
a
nd
F
i
g
u
r
e
6
w
he
r
e
t
h
e
x
-
a
xis
r
e
pr
e
s
e
n
ts
t
he
P
R
P
D
d
u
r
a
tio
n
wh
ile
t
h
e
y
-
a
x
i
s
r
e
pr
ese
n
t
s
t
h
e
a
ver
a
ge
c
l
a
ss
i
f
i
c
a
t
i
on
ac
cur
a
cy.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN: 2088-
8694
I
nt
J
P
ow
Elec
& Dr
i
S
y
st V
ol.
11,
N
o.
1
, Ma
r
202
0
:
326
–
33
2
33
0
(a)
(b)
F
i
g
u
r
e
5
.
A
N
N
a
vera
g
e
c
l
a
ssifi
c
a
t
ion
ac
cura
cy for (
a)
P
D
g
r
ou
p
1
and (
b
) P
D
gr
o
up
2
U
nde
r
no
ise
fre
e
con
d
i
tio
n,
u
s
i
n
g
s
hor
ter
P
R
P
D
dura
t
i
on
bare
ly
a
ffect
s
t
h
e
ANN
cl
assi
fi
cati
o
n
ac
cura
cy
o
f
PD
G
r
oup
1
.
T
h
e
ave
r
age
a
c
c
u
r
acy
i
s
94
%
w
i
t
h
o
n
l
y
1
.
36
%
s
t
a
n
d
a
rd
d
ev
i
a
ti
on
.
Fo
r
PD
G
ro
up
2,
t
h
e
c
lass
i
f
i
c
ati
on
ac
c
u
ra
cy
s
u
f
fer
s
a
m
i
n
or
r
educ
tio
n
o
f
5
%
w
h
e
n
t
h
e
P
R
P
D
dura
t
ion
redu
c
e
s
fr
o
m
15-
seco
nds
t
o
1-sec
o
n
d
.
Th
is
c
a
n
b
e
ex
p
l
ai
ne
d
by
t
h
e
rela
ti
v
e
ly
c
o
n
s
iste
nt
P
R
P
D
patter
n
o
f
PD
g
rou
p
1
,
he
n
c
e
the
sh
or
t
e
r
P
R
P
D
durati
on
doe
s
n
o
t
s
i
g
ni
fi
cant
l
y
a
ffe
c
t
th
e
P
D
c
l
assi
f
i
ca
ti
o
n
a
ccur
a
c
y
c
om
par
e
d
to
P
D
grou
p
2.
T
he
overa
l
l
s
m
a
ll
reduc
tio
n
i
n
c
lass
ifica
tio
n
ac
cura
c
y
s
how
s
t
h
e
r
o
bustn
e
ss
of
ANN
a
n
d
SVM
i
n
d
e
a
li
ng
w
ith
s
h
o
rter
P
RP
D
durati
o
n a
s
i
npu
t
da
ta.
Wh
en
n
oi
se
c
on
t
a
min
a
ti
on
i
s
t
a
k
e
n
i
n
t
o
c
on
si
d
e
ra
t
i
o
n
,
t
h
e
ANN
c
lassific
a
t
i
o
n
acc
urac
y
of
P
D
gro
u
p
1
de
ter
i
ora
t
e
s
m
ore
seve
rel
y
c
ompa
red
t
o
P
D
gro
u
p
2.
D
u
e
t
o
t
h
e
l
o
w
v
a
r
i
a
t
i
o
n
o
f
P
R
P
D
p
a
t
t
e
r
n
i
n
P
D
g
r
o
u
p
1,
t
he
P
D
clas
sifier
f
or
P
D
G
r
oup
1
i
s
n
o
t
g
o
o
d
i
n
ge
nera
l
i
z
i
ng
.
H
e
nc
e,
a
ny
v
ar
i
a
t
i
on
in
t
he
i
n
p
u
t
da
ta
w
i
l
l
ca
use
a
large
r
r
e
duc
tio
n
in
c
l
a
ssi
fica
t
i
o
n
a
c
c
ura
c
y.
F
or
P
D
gro
u
p
2
,
t
h
e
r
e
is
a
n
o
b
v
i
o
u
s
tre
nd
w
h
ere
hi
ghe
r
PRPD
dura
tio
n
r
e
sul
t
s
in
b
e
tte
r
classifica
ti
o
n
ac
c
ura
c
y.
A
sim
i
l
a
r
be
ha
vior
i
s
obse
r
ved
for
t
h
e
S
V
M
class
i
fier
w
here
s
ho
r
t
er
P
RP
D
dura
t
i
o
n
affec
t
s
P
D
G
r
oup
2 m
o
re sever
el
y com
p
ar
ed t
o G
r
ou
p 1.
H
ow
e
v
er,
th
e overa
l
l
acc
u
r
acy
of
SVM i
s
l
o
w
e
r t
h
a
n
ANN.
F
o
r
P
D
G
roup
1
a
n
d
G
r
ou
p
2,
S
V
M
h
as
a
n
a
v
era
g
e
o
f
1
3
%
and
1
9
%
l
o
w
er
a
cc
u
r
ac
y
comp
a
r
ed
t
o
ANN
u
n
d
e
r
n
o
i
s
e co
nt
ami
n
a
t
ion
.
(a)
(
b
)
Fi
g
u
re
6
. SVM
av
e
r
ag
e cl
assi
fi
cati
o
n
a
c
c
u
ra
c
y
fo
r
(a
)
P
D g
r
o
u
p
1
and (
b
) P
D
group
2
5.
CONCL
U
S
ION
Th
e
eff
ect
s
of
u
si
ng
s
h
o
r
t
e
r
d
u
r
a
t
i
o
n
PR
PD
f
o
r
P
D
c
l
as
sif
i
c
a
t
i
o
n
has
bee
n
s
ucc
e
ssfu
lly
i
n
v
es
t
i
ga
t
e
d.
Ba
se
d
on
the
resu
lts
o
b
t
a
i
ne
d
,
it
ca
n
be
c
onc
lude
d
t
h
at
P
D
clas
s
i
fica
t
i
o
n
acc
urac
y
o
f
P
D
source
m
ea
sure
d
from
l
a
b
f
a
b
ri
ca
ted
i
n
su
la
t
i
o
n
m
ateria
ls
w
i
l
l
no
t
be
s
ig
ni
fica
n
t
l
y
affec
t
e
d
b
y
usi
ng
sh
or
t
e
r
P
R
P
D
durati
on.
H
o
w
e
ve
r,
t
h
i
s
is
o
n
l
y
true
f
o
r
l
ab
f
a
b
rica
te
d
m
a
te
ria
l
s.
F
or
m
o
re
r
e
a
l
i
sti
c
a
n
d
p
ract
i
c
al
P
D
mea
s
u
r
ed
f
ro
m
pow
er
sys
t
e
m com
p
o
n
e
n
t
s
su
c
h as X
LP
E c
a
ble
joi
n
ts, usin
g
lo
nge
r
P
RP
D
dur
a
tio
n can
i
mprove
c
l
a
ssifi
c
a
t
i
o
n
accu
racy
o
f
ANN
an
d
SVM.
U
si
ng
l
ong
er
P
R
P
D
d
u
r
a
t
i
o
n
al
so
e
n
a
bl
e
s t
h
e PD classi
fier to
be
less
s
u
sc
e
p
ti
ble
to cla
ssi
fica
ti
o
n
a
cc
urac
y re
d
u
c
t
ion
in
d
ea
l
i
ng w
i
th n
o
i
se
c
on
t
am
ina
t
i
o
n.
Evaluation Warning : The document was created with Spire.PDF for Python.
Int J
P
o
w
E
l
e
c
&
D
ri S
yst
IS
S
N
:
2088-
86
94
Ef
fec
t
s
o
f
sh
ort
e
r
ph
ase
-
reso
l
v
e
d
pa
rti
a
l d
i
s
c
ha
rge
d
u
rat
i
o
n on PD
clas
sif
i
ca
t
i
o
n
acc
u
ra
c
y
(Cho
ng
W
a
n Xi
n)
33
1
ACKNOW
LEDG
E
MEN
T
S
The
a
u
th
or
w
ou
l
d
l
ike
t
o
e
xpress
gra
t
i
t
ude
t
o
Tu
n
k
u
A
b
d
u
l
Ra
hm
a
n
U
n
i
v
e
r
s
ity
C
o
l
le
ge
f
or
sup
por
tin
g
th
i
s
w
ork
thr
o
u
g
h
T
A
R
U
C
I
n
t
ern
a
l
Re
sea
r
ch
G
ra
nt
(
U
C
/
I/G
2
0
18-
00
0
26)
a
n
d
N
v
id
i
a
C
orpor
a
tio
n
for
spo
n
sor
i
n
g
t
he
G
P
U
used
for
this w
ork.
REFE
RENCES
[1]
J.
K
.
W
o
ng,
H
.
A.
I
ll
ias,
H
.
M
okh
lis
an
d
A.
H
.
A.
B
akar,
"Inv
es
ti
ga
t
i
o
n
o
f
p
a
rti
a
l
di
scharg
e
s
e
ver
i
t
y
a
t
XL
PE
cab
le
wi
t
h
out
termi
n
a
tion,"
Po
wer a
n
d
En
erg
y
(
P
ECo
n
),
20
14
IEEE Int
e
rn
atio
nal Conference
,
pp.
1
3
-
16,
20
1
4
.
[2]
X.
P
eng,
et al.,
"
A
con
v
o
l
ut
ional
neu
r
al
n
etw
o
rk
b
as
ed
d
eep
l
earni
ng
m
eth
o
d
o
l
o
g
y
f
o
r
recogni
t
ion
o
f
p
artial
dis
c
harg
e
pat
t
erns
f
rom
h
i
g
h
v
o
ltag
e
cabl
e
s,
"
IEEE Transacti
o
ns
on
P
o
wer De
li
v
e
ry
,
pp
.
1-1
,
2
019.
[3]
E.
M
.
Lal
i
t
h
a
and
L
.
S
a
t
i
s
h,
"
W
a
vel
e
t
analys
is
f
or
c
lass
if
icat
i
o
n
o
f
mu
lt
i-
s
o
ur
c
e
P
D
p
a
t
t
e
r
n
s,
"
IE
EE Tr
ansact
i
o
n
s
on
Diel
ectrics an
d
E
l
ectrica
l In
sulati
on
,
Vol
.
7
, pp
.
4
0-4
7
, 2
00
0.
[4]
"
I
EC
i
nte
r
na
tion
a
l
sta
n
da
rd
6
027
0:
H
igh
v
o
l
t
a
g
e
te
st
t
e
c
hn
iq
ue
s
-
p
art
i
al
d
i
s
charge
m
e
a
s
u
rem
e
nt
s
,
"
Internat
ion
a
l
Elect
rotechn
i
ca
l Co
mmissio
n
.
[5]
W.
J
.
K
.
R
ay
mo
nd
,
L
.
T
.
S
i
ng
,
L
.
W
.
K
i
n
,
G
.
K
.
M
en
g,
H
.
A.
I
ll
i
as
a
nd
A
.
H.
A
.
Bakar,
"
F
e
a
t
ure
pru
n
in
g
f
o
r
p
a
rtial
dis
c
harg
e
clas
sificati
o
n
u
s
in
g
i
ndf
eat
a
n
d
relieff
a
lgo
r
it
h
m
,"
2
0
1
8
IE
EE
2n
d
In
ter
n
a
t
i
onal
Conferen
ce on
D
i
elect
ri
cs
(ICD)
, pp
. 1-4
,
2
0
1
8
.
[6]
M.
A
llah
b
ak
hs
h
i
a
nd
A
.
A
kbari
,
"A
m
eth
od
f
o
r
di
scrimi
na
t
i
n
g
o
ri
gina
l
p
u
l
se
s
in
o
n
l
ine
pa
rtia
l
di
sc
ha
rg
e
meas
urem
ent
,
"
Me
asure
m
e
n
t
,
v
o
l.
44
, p
p.
14
8
-
1
5
8
, 20
1
1
.
[7]
H.
J
anan
i,
B
.
K
o
rdi
and
M
.
J
.
J
o
zani
,
"
Class
i
fi
catio
n
o
f
s
im
u
l
t
an
eou
s
m
ulti
pl
e
part
ial
d
i
sch
a
rg
e
so
urces
b
as
ed
o
n
p
r
o
b
a
b
il
is
ti
c
in
t
e
r
p
r
e
t
a
ti
on
u
s
i
n
g
a
t
w
o
-
s
t
e
p
lo
g
i
s
t
i
c
r
e
g
r
e
s
s
i
on
a
lg
ori
t
h
m
,"
IEEE Tran
sa
cti
ons
o
n
Die
l
ectr
i
cs
a
n
d
Elect
r
i
cal Insu
l
a
ti
on
,
v
o
l
.
2
4
,
p
p.
54-6
5
,
20
17.
[8]
H.
S
o
ng,
J
.
Dai,
G
.
S
h
en
g
an
d
X
.
J
iang
,
"GIS
p
arti
al
d
is
charge
patt
ern
reco
gn
iti
o
n
vi
a
deep
c
on
vo
lu
tio
n
al
n
eu
ral
netw
ork
un
der
com
p
lex
d
a
ta
s
ou
rce,
"
IE
EE
Tran
sa
cti
o
n
s
on
Di
electrics
an
d
E
l
ectrica
l Insulatio
n
,
Vo
l
.
2
5,
pp.
6
7
8
-6
85,
2018
.
[9]
M.
K
arimi,
M
.
M
a
ji
d
i
,
M
.
Etezad
i-Am
oli
and
M
.
Osk
uo
ee,
"
P
a
rt
ial
d
i
s
ch
arge
c
l
a
ss
ificat
io
n
usi
n
g
deep
b
el
ie
f
netw
ork
s
,
"
2018 IEEE/PE
S
Tr
ansmission and D
i
st
r
i
bution
C
o
n
f
erence and E
x
position
(
T
&
D
)
,
pp
. 1
06
1-1
070
,
2
018.
[10]
W.
J
.
K.
R
aym
ond,
H
.
A
.
I
lli
a
s,
A
.
H.
A
.
Bakar
and
H.
M
ok
hl
is,
"P
arti
al
d
is
charg
e
c
las
s
i
f
icati
ons:
Revi
ew
o
f
recen
t
prog
ress
,
"
M
e
asur
emen
t
,
Vo
l. 68
, pp
. 1
64
-1
81
, 2
01
5.
[11]
B.
K
arthik
eyan,
S.
G
op
al
a
nd
M
.
V
im
ala,
"
Concep
ti
on
o
f
com
p
l
e
x
pr
obabilist
i
c
neural
n
etwork
s
ystem
for
class
i
fi
catio
n
of
p
art
i
al
d
is
char
ge
p
atterns
using
mult
ifari
o
us
i
npu
ts,
"
Expert
Syst
ems w
i
t
h
App
licati
ons
,
vol.
29
,
pp.
9
5
3
-9
63,
2005
.
[12]
E.
G
u
l
sk
i
,
"
Di
g
i
t
a
l
analys
is
o
f
parti
a
l
discharges,
"
IE
EE
Tr
ansa
c
tio
n
s
on D
i
electrics
and
El
ectr
i
cal Ins
u
l
a
t
i
o
n
,
Vol
.
2
,
p
p
.
8
2
2
-
83
7,
199
5.
[13]
L.
S
ati
s
h
an
d
W.
S
.
Zaeng
l
,
"Artif
ici
a
l
neu
r
al
n
et
work
s
f
o
r
rec
og
ni
tio
n
of
3
-d
p
arti
al
d
i
s
charg
e
p
att
e
rns,
"
IE
E
E
T
r
an
sa
cti
ons o
n
D
i
elect
ri
cs
a
n
d
El
ectrical Ins
u
l
a
t
i
o
n
, Vo
l
. 1
, pp
. 2
6
5
-2
75
, 19
9
4
.
[14]
J.
A
.
Hun
t
er,
et
a
l
.,
"
P
a
rti
a
l
dis
c
harge
i
n
m
edi
u
m
v
o
ltage
t
hree-p
h
ase
c
a
b
l
es,
"
E
l
ectr
i
ca
l In
sul
a
t
i
o
n
(
I
S
E
I)
,
Conf
eren
ce Recor
d
o
f
th
e 20
10
IEEE Intern
a
t
io
n
a
l S
y
mpo
s
i
u
m
, p
p. 1-5
,
2
0
1
0
.
[15]
B
.
K
o
r
d
i
a
n
d
H
.
J
a
n
a
n
i
,
"
T
o
w
a
r
d
s
a
u
to
m
a
ted
statis
ti
cal
p
art
i
al
d
i
s
c
harge
s
o
u
r
ce
cl
ass
i
ficat
io
n
usi
n
g
pattern
r
eco
gn
it
io
n
tech
ni
ques
,
"
Hi
gh
Volta
ge
.
2
018.
A
vai
l
able:
h
t
t
p
://
d
igital-
lib
rary.t
hei
e
t.
o
r
g
/
cont
ent/j
o
u
r
nals/1
0.
1
0
4
9
/
hve.
2018.
50
48
[16]
H.
I
lli
as
,
G
.
A
lta
m
im
i,
N
.
M
o
k
h
t
a
r
and
H.
A
rof
,
"
Cl
assificat
ion
o
f
m
u
l
t
iple
p
a
r
ti
a
l
d
isch
arge
s
ou
rces
i
n
d
i
elect
ric
i
n
su
la
t
i
o
n
m
a
t
e
r
ia
l
us
in
g
c
e
p
s
t
r
u
m
analysi
s
-artificial
neural
n
et
wo
rk
,"
IE
EJ
Transactions
on Elect
r
i
cal a
nd
Elect
ro
n
i
c En
g
i
n
eeri
n
g
, v
o
l
. 1
2,
p
p
. 35
7
-3
64
,
2
0
17
.
[17]
W
.
J
.
K
.
R
a
y
m
o
n
d
,
H
.
A
.
I
l
l
i
a
s
a
n
d
A
.
H
.
A
.
B
a
k
a
r
,
"
H
i
g
h
n
o
i
s
e
t
o
l
e
ran
ce
f
eatu
r
e
ex
tract
io
n
f
o
r
p
a
rti
a
l
dis
c
harg
e
cl
assificat
io
n
in
X
LPE
cabl
e
j
o
i
nts,"
IEEE T
r
ansa
c
tio
ns o
n
D
i
electrics
and
Elect
r
i
ca
l
In
sul
a
tio
n
,
vol.
24,
pp.
6
6
-
74
,
20
17
.
[18]
F
.
H
.
K
r
eug
e
r,
E
.
Gu
l
s
ki
a
nd
A
.
K
r
iv
da,
"Cl
a
ss
ifi
catio
n
o
f
p
a
r
t
ia
l
discharges
,
"
IEE
E
Tra
n
sa
c
t
i
ons on
Electrica
l
Insulati
on
,
Vol
.
2
8, p
p.
91
7
-9
3
1
,
19
93
.
[19]
E
.
G
u
l
s
k
i
a
n
d
A
.
K
r
i
v
d
a
,
"
N
e
u
r
a
l
n
e
t
w
o
r
k
s
a
s
a
t
o
o
l
f
o
r
r
e
c
o
g
n
i
tio
n
of
p
arti
al
d
is
charges
,
"
IEEE Trans
a
ct
i
o
ns on
Elect
r
i
cal Insu
l
a
ti
on
,
Vo
l. 2
8
, pp
.
9
8
4
-1
00
1, 1
99
3
.
[20]
E.
G
uls
k
i,
"
Comp
uter-ai
d
ed
m
easu
r
em
ent
o
f
p
artial
dis
c
harges
i
n
HV
eq
uip
m
en
t
,
"
IEE
E
Tra
n
sa
cti
o
n
s
on
E
l
ectric
a
l
Insulati
on
,
Vol
.
2
8, p
p.
96
9
-9
8
3
,
19
93
.
[21]
A.
A
.
M
a
s’u
d
,
B.
G
.
Stewart
an
d
S
.
G
.
McM
eek
in,
"An
i
n
v
e
sti
g
a
t
i
v
e
s
tudy
in
to
t
he
s
ensit
i
vit
y
o
f
d
i
ff
eren
t
parti
a
l
dis
c
harg
e
φ
-
q-n
p
a
ttern
r
eso
l
u
t
i
o
n
si
zes
on
st
atisti
cal
n
e
u
ral
n
e
t
w
o
r
k
patt
ern
c
l
assifi
cation,
"
M
e
as
ur
emen
t
,
Vo
l
.
92
,
p
p
. 4
97
-50
7
, 2
01
6.
[22]
A
.
A
.
Z
a
h
e
d
,
A
.
H
.
E
l
-
H
a
g
,
N
.
Q
a
d
d
o
u
m
i
,
R
.
H
u
s
s
e
i
n
a
n
d
K
.
B
.
S
h
aban,
"Com
parison
o
f
d
i
fferent
f
ourt
h
ord
e
r
hil
b
ert
f
r
act
a
l
a
nt
enn
a
s
f
o
r
part
ia
l
d
i
sch
a
rge
m
easu
r
e
m
en
t,"
IEEE Tra
n
sa
ction
s
on
D
i
e
l
ectrics a
nd Electrica
l
Insulati
on
, v
o
l
.
2
4
, pp
.
17
5
-1
82
,
20
17
.
[23]
L.
G
uomin
a
nd
Z
.
D
a
min
g
,
"Re
c
og
ni
ti
on
of
p
artia
l
d
i
sch
a
rge
u
s
i
n
g
wa
v
e
le
t
e
n
tr
op
y
a
n
d
n
e
ura
l
n
e
t
work
f
or
T
EV
meas
urem
ent
,
"
Po
wer S
y
st
em
Tech
nol
og
y
(
P
OWER
CON)
,
2012
IEEE Int
e
rn
a
t
ion
a
l
Confer
ence
, pp
. 1
-6,
2
0
1
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N: 2
0
8
8
-
86
94
I
nt
J
P
o
w
E
l
e
c
&
D
r
i
S
yst
V
o
l.
11,
N
o.
1
,
Mar
202
0
:
326
–
33
2
33
2
[24]
Y.
X
u,
Y
.
Q
i
an
,
F
.
Y
ang
,
Z
.
Li
,
G
.
S
heng
a
nd
X
.
J
ian
g
,
"DC
cabl
e
fe
a
t
ure
e
x
trac
tio
n
ba
se
d
on
the
PD
i
ma
g
e
i
n
th
e
no
n-s
u
b
s
am
p
l
ed cont
ou
rlet
t
ran
s
fo
rm dom
ain,
"
IEEE T
r
an
s
a
ctio
ns on
Dielect
ri
cs and
Electr
i
cal In
su
la
ti
on
,
V
o
l
.
2
5
,
pp
.
5
33
-540
,
2
0
1
8
.
[25]
R.
Y
ao,
M.
H
u
i
,
J.
L
i,
L
.
Ba
i
and
Q
.
W
u,
"
A
ne
w
d
i
s
c
harg
e
pat
t
e
rn
f
or
t
he
c
h
a
ract
erizat
ion
and
i
d
ent
i
fi
cati
on
of
insulation de
f
ects in GI
S
,"
E
n
ergies
,
Vo
l
.
1
1,
p.
9
7
1
, 2
01
8.
[26]
I.
M
i
tiche,
G.
M
ori
s
on
,
A.
N
esbitt,
M
.
Hug
h
es-Narborough,
B.
G
.
S
t
e
w
a
r
t
a
n
d
P
.
B
o
r
e
h
a
m
,
"
C
l
a
s
s
i
f
i
c
a
t
i
o
n
of
p
artial
d
i
scha
rg
e
si
gn
als
by
c
om
b
i
nin
g
a
d
a
ptiv
e
lo
cal
iterat
iv
e
f
i
lterin
g
a
n
d
e
ntro
py
f
eatu
r
es,"
S
e
ns
ors
(
B
a
s
el)
,
Vol. 1
8,
2
01
8.
[27]
J.
A
.
H
u
n
t
er,
P.
L
.
Lew
i
n,
L
.
Hao,
C
.
W
a
lt
on
a
nd
M
.
Michel
,
"
A
u
to
nom
ou
s
c
l
ass
i
fi
catio
n
o
f
P
D
s
o
u
r
ce
s
wit
h
in
th
ree-ph
ase
1
1
k
V
P
IL
C
cabl
e
s,"
IEEE T
r
a
n
sa
cti
ons o
n
Diel
ect
rics an
d
Elect
ri
cal
In
sula
ti
on
,
Vo
l.
2
0,
pp
.
2
11
7-21
24
, 2
01
3.
BIOGRAPHI
E
S
OF
AUT
HORS
Cho
ng
W
a
n
Xin
o
b
tai
n
ed
t
h
e
B
En
g
(H
on
ou
rs)
E
l
ectri
cal
a
n
d
E
lect
r
on
ics
in
2018,
m
ajo
r
ing
i
n
e
l
e
c
t
ric
a
l
e
ng
in
e
e
r
ing
,
f
ro
m
Tu
nk
u
Ab
du
l
R
a
hma
n
U
nive
rsity
C
ol
l
ege
(TA
R
U
C
),
K
u
a
la
Lu
mp
ur,
M
a
l
a
y
s
ia.
S
h
e
is
c
urrent
ly
p
urs
u
i
ng
h
e
r
M
a
s
t
er
d
eg
ree
a
t
T
A
RU
C.
H
er
r
es
ear
ch
i
n
t
e
r
e
s
t
s
in
c
l
ud
e
hi
g
h
v
o
l
ta
ge
,
p
a
r
t
i
a
l
d
i
sc
h
a
r
g
e
in
d
ie
l
e
c
t
r
i
c
s
i
n
sulation,
a
nd
c
l
a
ssificati
o
n
o
f
part
ial
discharg
e
.
Wo
ng
J
ee
K
een
R
ay
m
o
n
d
w
as
b
orn
i
n
P
erak,
M
a
lay
s
ia
i
n
1987.
H
e
r
ecei
ved
t
h
e
Bach
elo
r’s
Deg
r
ee
i
n
El
ectrical
&
E
lectro
n
i
cs
E
ngin
eeri
n
g
and
M
a
st
er’s
D
eg
ree
i
n
E
lect
rical
E
ngin
eerin
g
f
r
o
m
U
ni
versi
t
y
Ten
a
ga
N
asi
o
n
a
l
i
n
2
01
0
an
d
2
0
1
3
.
He
o
b
t
ained
h
is
P
hD
i
n
El
ectri
ca
l
En
gin
eerin
g
f
r
o
m
U
niv
e
rsity
o
f
Mal
a
ya
i
n
20
16.
H
e
work
ed
a
s
a
P
rod
u
ct
E
ngin
e
er
i
n
M
o
t
o
rola
M
a
lays
ia
from
20
10
t
o
2
011.
H
e
has
been
w
o
r
ki
ng
a
s
a
Senior
L
ec
tu
r
e
r
in
t
h
e
D
e
p
a
r
tme
n
t
o
f
El
ectrical
a
n
d
E
l
e
c
t
ron
i
cs
E
ng
i
n
eeri
n
g
,
T
un
ku
A
b
d
u
l
R
ahm
a
n
Un
iv
ers
i
ty
C
o
l
le
ge
s
ince
2
016.
Hi
s
m
a
in
r
esearch
i
nt
erest
i
n
c
l
udes
cl
ass
i
fi
c
a
tion
of
p
art
i
a
l
d
is
c
h
arge,
f
a
u
l
t
l
o
cati
on
an
d
f
a
u
l
t
clas
sifi
cati
on usin
g
m
ach
ine learni
ng
.
Hazl
ee
A
zil
Illias
was
born
i
n
Ku
ala
Lumpur,
M
a
l
a
ysia.
He
r
ecei
v
e
d
t
h
e
Bachel
or'
s
D
egree
i
n
El
ectrical
E
n
g
i
n
eeri
ng
f
r
o
m
t
he
U
niv
e
rsi
t
y
o
f
M
al
aya,
M
al
ays
i
a
i
n
M
a
y
2
0
0
6
.
T
h
e
n
,
h
e
imm
e
di
a
t
el
y
join
ed
F
rees
cale
S
e
mi
con
d
u
c
tors
(
M
)
S
dn
.
Bh
d
.
a
s
a
prod
uct
en
gin
eer
b
ef
ore
pu
rsu
i
ng
h
is P
h
D
s
t
udies
i
n
January
2
008
.
H
e
o
b
t
ai
ned
th
e
P
h
D
D
e
g
ree
i
n
E
lect
ri
cal Engin
eerin
g
f
r
o
m
t
he
U
ni
vers
it
y
o
f
S
outh
a
mpto
n,
U
nit
e
d
K
i
ngd
om
i
n
M
a
y
20
11
.
H
e
was
a
S
e
n
i
or Lectu
r
er
i
n
th
e
Dep
a
rtm
e
nt
o
f
El
ectri
cal
E
ngin
e
e
r
i
n
g
,
U
ni
versity
o
f
M
a
lay
a
f
r
om
J
ul
y
20
11
to
J
a
n
uary
2
017
.
S
i
n
ce
J
a
nu
ary
2
017
,
he
h
as
b
e
e
n
an
A
sso
cia
t
e
P
r
of
ess
o
r
i
n
t
h
e
D
ep
a
r
tm
en
t
o
f
E
l
ectri
cal
En
gin
eerin
g
,
U
ni
versity
o
f
M
a
lay
a
.
He
i
s
a
regi
st
ered
C
hartered
E
n
g
i
n
e
e
r
(
C
.
E
n
g
)
a
n
d
a
P
r
of
ess
i
o
n
a
l
E
n
g
in
eer
(
P.
Eng
)
.
Hi
s
m
a
i
n
r
esearch
i
nt
erest
s
i
n
c
l
ude
modelling
and
mea
s
urement
of
p
art
i
al
d
isch
arge
p
h
e
no
men
a
i
n
soli
d
d
i
elect
ric
ins
u
l
a
ti
on
,
condition
mon
i
tor
i
ng,
insulat
i
on
sy
st
e
m
d
iag
nos
is
,
li
ghtnin
g
o
v
e
rvo
ltag
e
,
tran
smissi
on
li
ne
m
od
e
lli
ng,
optimis
ation
m
e
th
od
s
an
d
artif
ic
ial
intelli
g
e
nce techniques.
Lai
W
e
ng
K
in
i
s
an
A
ssocia
te
P
r
o
f
e
ss
or
w
ith
th
e
Depart
m
e
nt
o
f
E
l
ectrical
&
E
lect
ronic
En
gin
eerin
g
,
F
acul
t
y
of
E
n
g
i
n
eeri
ng,
T
un
ku
Ab
d
u
l
Rahm
an
U
n
i
vers
it
y
Co
l
l
eg
e.
P
ri
or
t
o
this
,
he
has
serv
ed
a
s
t
h
e
Direct
or
f
o
r
A
dvan
ced
I
n
f
ormat
i
cs
i
n
a
n
a
ti
on
al
r
es
ear
ch
cen
tre
f
o
r
ICT
f
o
r
nearl
y
t
wo
d
ec
ades,
cent
e
r
e
d
prim
arily
i
n
the
ar
eas
o
f
patt
ern
re
cogn
it
io
n
and
m
a
chi
n
e
in
te
l
l
ig
ence.
H
e
receiv
e
d
h
i
s
Doct
orat
e
(by
research)
i
n
E
ng
ine
e
r
in
g
fro
m
t
h
e
U
nive
rsity
o
f
Au
cklan
d
,
New
Zealan
d
an
d
a
M
a
s
t
ers
in
S
cien
ce
(Elect
ron
i
cs
)
f
r
om
Q
u
eens
Un
iv
ersit
y
o
f
Belf
as
t
(N
or
thern Ire
l
a
nd).
D
r.
L
ai
i
s
a F
e
ll
o
w
o
f
the IET,
r
eg
ist
e
red
Ch
artered En
gi
neer
w
ith
t
he
En
gin
eerin
g Co
u
n
cil
o
f
U.
K
.
,
s
e
n
i
or m
em
ber of
IE
E
E,
and
a P
rofe
ssio
na
l
En
gin
e
e
r
,
re
g
i
s
t
e
r
e
d
on
the
Internationa
l
Pr
of
es
siona
l
E
n
g
ine
e
r
regist
er
(
U.K.).
H
e
is
a
l
s
o
a
B
o
a
r
d
m
e
m
b
e
r
o
f
t
h
e
A
s
i
a
-
P
acific
Neural N
et
wo
rk S
o
c
iet
y
(A
P
NN
S
)
.
Yi
auw
Kah
Kau
r
w
as
b
o
r
n
i
n
M
al
aysi
a
in
1
9
7
5
.
H
e
receiv
e
d
t
h
e
B.
E
and
P
h
.D.
d
e
grees
f
rom
Un
iversit
y
o
f
W
a
les
,
S
wan
s
ea,
U
K
i
n
1
99
8
a
n
d
20
05
r
esp
ectiv
ely
.
H
e
j
o
i
n
ed
T
un
ku
A
b
dul
Ra
h
m
a
n
U
nive
rsity
C
o
lle
ge
,
Ku
a
l
a
Lu
mp
ur,
Ma
la
y
s
ia
s
inc
e
20
06
,
w
h
e
r
e
h
e
i
s
c
u
r
r
e
n
t
l
y
a
P
r
in
c
i
pa
l
L
ecturer.
His
m
a
i
n
a
reas
o
f
research
i
n
t
e
r
es
t
are
rea
cti
v
e
h
a
rm
on
ic
f
ilt
e
rs,
sm
art
energ
y
m
a
nagem
e
n
t
s
y
s
tem
s
a
nd
th
e
ir
c
o
n
t
r
ol
s
ystems
.
Evaluation Warning : The document was created with Spire.PDF for Python.