TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.5, May 2014, pp
. 3616 ~ 36
2
4
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i5.4054
3616
Re
cei
v
ed
Jul
y
30, 201
3; Revi
sed
De
ce
m
ber 2, 2013
; Accepte
d
Ja
nuary 8, 201
4
A Novel Transformer Fault Diagnosis Approa
ch
Based on Information Fusion Method
Huan
g Xin-b
o
*
，
Song To
ng
，
Lin Xiao-huan, Wang
Ya-na
Coll
eg
e of Elec
tronics an
d Informatio
n
, Xi
’an
Pol
y
t
e
chn
i
c U
n
iversit
y
,
Xi
’a
n 710
04
8, Shaa
n
x
i Provi
n
ce, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: huan
g
x
b
197
5@1
63.com
A
b
st
r
a
ct
T
o
i
m
pr
ove t
h
e
pow
er
transfor
m
er
fau
l
t d
i
ag
n
o
sis
accuracy,
this p
a
p
e
r pr
op
oses
a fa
ult
dia
gnos
is
meth
od
of
infor
m
ati
o
n
fusi
on
w
h
ich is
b
a
sed
on
fu
zz
y
c
o
d
i
n
g
b
o
u
ndary
a
n
d
Bi
as r
egu
lari
z
a
ti
on
L
e
ven
b
e
r
g
-
Marqu
a
rdt (L-
M) netw
o
rk. T
he al
gorith
m
us
es a Bi
as ap
pr
oach to
deter
mine th
e hyp
e
r p
a
ra
meters,
ma
king
the ne
ural
net
w
o
rk adaptiv
el
y adj
ust
the p
a
r
ameter i
n
the
traini
ng pr
oces
s and th
en
get
s the opti
m
i
z
a
t
i
o
n
para
m
eters of
the o
b
j
e
ctive f
unctio
n
. On th
e oth
e
r
han
d, t
he
usin
g
of fu
zz
y
c
o
d
i
ng
b
o
u
ndary
can
re
d
u
ce
the variati
ons
and i
m
prove th
e accuracy of fault di
agn
osis
. T
he contrast of the tw
o
fusion
diag
nosis res
u
lts
draw
s a co
ncl
u
sion. T
h
at is, the
perfor
m
a
n
c
e
of Bi
as
R
e
g
u
l
ari
z
a
t
i
o
n
F
u
zzy L-M N
e
tw
ork is su
peri
o
r to t
he
no fe
ature r
e
d
u
ction
fusio
n
mode
l w
h
ich
is B
i
as R
e
g
u
lar
i
z
a
t
i
on
L-M N
eur
al
Netw
ork, an
d t
he
accuracy
ra
te
of the former is
89.83%.
Ke
y
w
ords
:
p
o
w
er transforme
r, informatio
n
fusio
n
, bias r
e
g
u
lari
z
a
tio
n
; L-
M neur
al n
e
tw
ork; fu
zz
y
c
odi
ng
bou
nd
ary
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Multi-Sou
r
ce
Information
F
u
sio
n
(MSIF) tech
nology i
s
a
theo
ry a
nd a
metho
d
whi
c
h
resea
r
che
s
t
he
comp
re
he
nsive treatm
ent an
d a
ppl
i
c
ation
of u
n
certainty info
rmation, na
m
e
ly,
getting mo
re
accu
rate
an
d credi
ble
concl
u
si
on by
pro
c
e
s
sing
the informati
on fro
m
mult
iple
informatio
n source
s in mul
t
i-level re
cog
n
ition [1
]. In the fault diag
n
o
si
s a
s
pe
ct, MSIF technol
ogy
use
s
the
extracted fe
ature
informat
io
n
of system fail
ure a
nd d
edu
ce the fa
ult type of the o
b
ject
according to
the fault diag
nosi
s
meth
od
s, then
the fu
sion
cente
r
p
r
ocesse
s the
compl
e
me
nta
r
y
and redu
nda
nt information
base
d
on
ce
rtain crite
r
ia
in
the spa
c
e
an
d time, and ul
timately get the
fault deci
s
ion
of object types. The
r
efo
r
e
,
the informat
ion fusio
n
me
thod whi
c
h i
s
applied to fa
ult
diagn
osi
s
ca
n greatly imp
r
ove the
com
p
letene
ss of
fault c
h
arac
ter information
[2]. At present,
the diversity, uncertainty a
nd complexit
y
bring
mo
re
difficulties in
fault diagno
si
s technol
ogy of
the transfo
rm
er. The literat
ure [3] is abo
ut the app
lica
t
ion of neural
network in fa
ult diagno
sis
of
the transfo
rmer, but the rate of con
v
ergen
ce
i
s
slo
w
. The literatu
r
e on [4] uses g
e
n
e
tic
algorith
m
s
(G
A) to improv
e the neu
ral
netwo
rk’
s
wei
ghts an
d thre
shol
d, but the GA method
is
compl
e
x
,
s
o
t
he
net
w
o
r
k
is
ea
sy
t
o
f
a
ll int
o
l
o
cal
opt
imal.
Th
e lit
er
at
ure
[
5
]
ado
pt
s
f
u
zzy
membe
r
ship f
unctio
n
which
co
uld ove
r
co
me the
ab
sol
u
te
situ
ation of
ratio of
co
ding boun
da
ry,
but the parall
e
l processing abilit
y is poor. The literature [6] em
ploy
particl
e swarm optimization
algorith
m
(P
SO) for tran
sformer fa
ult d
i
agno
si
s,
imp
r
oving PSO a
l
gorithm
by linear
de
creasing
strategy. Ho
wever, the n
e
twork i
s
ea
sy to fall in
to local o
p
timum
.
Particles
are easy to re
ach
prem
ature co
nverge
nce
too.
Relying
on a
singl
e diag
no
sis m
e
thod
of transf
o
rme
r
f
ault ch
ara
c
te
ristics can reflect the
transfo
rme
r
condition
from
only on
e a
s
p
e
ct, an
d it
ca
n not
ma
ke
a
comp
re
hen
si
ve evaluatio
n
on
the overall
health statu
s
of tran
sfo
r
mer.
T
herefore, ba
se
d on the ide
a
of intellige
n
t
compl
e
me
nta
r
y fusion, this pape
r e
s
ta
blish
ed the informatio
n fusion fault dia
gno
sis mo
del
of
fuzzy
co
ding
bou
nda
ry a
nd Bia
s
Reg
u
lari
zation
L
-
M Neural
Ne
twork. Simpli
fy the input
unit
numbe
r of the L-M n
e
u
r
al netwo
rk b
y
usi
ng the
fuzzy codin
g
bound
ary, a
nd improve the
gene
rali
zatio
n
ability of L-M neural net
work by u
s
in
g Bias regul
a
r
izatio
n alg
o
ri
thm. In this way,
the two adva
n
tage
s are
complem
entary, not only
id
entifying the fault types, but also improvi
n
g
the rate of
corre
c
t di
ag
nosi
s
. Fin
a
ll
y, comp
are
d
with
seve
ral fore
ca
stin
g metho
d
s,
the
experim
ental
results
sho
w
that the model
has st
rong a
b
ility of simulation and fore
ca
st.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Novel T
r
an
sform
e
r Fa
ult Diagn
osi
s
Appro
a
ch Based on Inform
ation Fusi
on
… (Hu
ang Xi
n-bo
)
3617
2. The Principle of the Pr
oposed
Algo
rithm
2.1. Regulari
z
atio
n Theor
y
of Bias
The ba
sic id
e
a
of Bias reg
u
lari
zation al
gorithm i
s
as
follows:
Given a
set
of trainin
g
sam
p
le
s
)}
,
(
),...,
,
(
),
,
{(
2
2
1
1
n
n
t
p
t
p
t
p
K
, the ne
ural n
e
twork
learni
ng
obje
c
tive is that lo
ok fo
r a
fun
c
ti
on to
app
roa
c
h th
e
sampl
e
s,
what’
s
m
o
re, mi
nimize
the
error fu
nctio
n
.
Mean
sq
uare erro
r fun
c
ti
on i
s
u
s
ua
lly
used in
ne
u
r
al n
e
two
r
k training
given
by
equatio
n [1]:
n
i
i
i
y
t
n
E
1
2
)
(
1
(1)
In the form
ul
a,
n
as
the total s
a
mple;
i
t
as the expe
cted
output value
s
of network;
i
y
as
the actual output value. Howeve
r, in order to improv
e the gener
al
ization ability of the network,
the Bias regu
larization m
e
t
hod i
n
crea
se
s the
arith
m
e
t
ic average
value
of the
weights’
squa
re in
the obje
c
tive function. The
obje
c
tive function is given b
y
equation [2]:
W
E
E
F
m
i
i
W
m
E
1
2
1
(2)
In the formula,
i
as
co
nne
ction weight
of the n
eura
l
network,
m
as the n
u
mbe
r
of
con
n
e
c
tion
weights i
n
the
neu
ral n
e
twork,
,
as th
e p
a
ram
e
ters of
the obje
c
tive functio
n
. If
, then the
trai
ning
algo
rith
m aim
s
to
minimize
the
ne
twork trainin
g
error;
If
, then the
training al
go
ri
thm aims to
enabl
e the n
e
twork to p
r
o
duce a sm
oot
her respon
se
; It means as far
as po
ssible t
o
red
u
ce the
netwo
rk
para
m
eters e
ffecti
v
ely, then make u
p
for th
e netwo
rk error.
The convent
ional metho
d
of regul
arization i
s
usually difficult to determi
ne the si
ze
of
regul
ari
z
ation
paramete
r
s,
whil
e the
theory of
Bi
as
ca
n a
d
a
p
tively adju
s
t the si
ze
of
regul
ari
z
ation
param
eters and ma
ke the
m
optim
al in the network training p
r
o
c
e
s
s [7].
2.2. Fuzzy
Three Ratio Coding Crite
r
ia
Fuzzy thre
e ratio metho
d
is ba
sed
on
fuzzy theo
ry
to fault diagno
sis of the
powe
r
transfo
rme
r
,
whi
c
h m
a
kes the inte
rval
on th
ree
rat
i
o bou
nda
ry. According t
o
the traditio
nal
codi
ng rule
s
of three
ratio
method , th
e bou
nda
ry of cha
r
a
c
teri
stic g
a
s
ratio
is "0.1", "1", "3"
about
6
2
4
2
2
4
4
2
2
2
/
,
/
,
/
H
C
H
C
H
CH
H
C
H
C
.On the basis of the empiri
cal kno
w
led
ge, the bound
ary
of the "0.1" is fuzzy that "0
.08
~ 0.1
2
", "1" border i
s
f
u
zzy that "0.85 ~ 1.1
5
" and
"0.9 ~ 1.1", "
3
"
is fuzzy that "2.9 ~ 3.1" and "2.85 ~ 3.
15" [8].
The membe
r
ship
function of e
a
ch g
a
s ratio
is
fuzzy
distri
b
u
ted by the
method
of
assign
ed. Th
e memb
ershi
p
fun
c
tion of
0, 1 a
nd 2
is
r
e
spec
tive par
t
ial
s
m
all
, middle type
rid
g
e
type a
nd
p
a
rtial la
rg
e
.
The
fo
rmula
s
a
r
e given as
follows. So, whe
n
the
cod
e
is 0
,
1, 2, the corre
s
p
ondi
ng mem
bership fun
c
tion
is
)
(
),
(
),
(
2
1
0
i
i
i
xb
u
xb
u
xb
u
. Then, relyi
ng on th
e principl
e of ma
ximum deg
re
e of memb
ership t
o
determi
ne th
e final
cod
e
.
Therefore, th
e characte
ri
st
ics of ga
s
sa
mple d
a
ta u
s
ed in thi
s
pa
per
cha
nge into t
he co
ding
se
quen
ce of "0, 1, 2" as the in
put of the network.
08
.
0
,
08
.
0
,
1
)
(
)
08
.
0
(
50
0
i
xb
i
i
xb
e
xb
xb
u
i
(4)
1
.
3
,
0
1
.
3
9
.
2
)],
3
(
5
sin[
5
.
0
5
.
0
9
.
2
12
.
0
,
1
12
.
0
08
.
0
)],
1
.
0
(
25
sin[
5
.
0
5
.
0
08
.
0
,
0
)
(
1
i
i
i
i
i
i
i
i
xb
xb
xb
xb
xb
xb
xb
xb
u
(5)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3616 – 36
24
3618
85
.
2
,
1
85
.
2
,
0
)
(
)
85
.
2
(
12
2
i
xb
i
i
xb
e
xb
xb
u
i
(6)
2.3. The L-M
Net
w
o
r
k T
r
ai
ning Principle
The e
s
sen
c
e
of neu
ral n
e
twork m
odeli
n
g is to
fin
d
ou
t the esse
ntia
l con
n
e
c
tion
betwe
en
the input and
output data in the finite sample,
nam
el
y the mappin
g
relation
shi
p
, thus the inp
u
t
without trai
ning ca
n also
give the app
ropriate
outpu
t and has th
e
gene
ralizatio
n function.
Refer
to the do
cu
ments [9, 1
0
], the standa
rd BP alg
o
rit
h
ms
use the
steep
est d
e
s
cent meth
o
d
to
modify wei
g
h
t
s, and th
e training
pro
c
e
s
s fro
m
a p
o
in
t along
with t
he surfa
c
e
of error fu
nctio
n
,
then g
r
ad
uall
y
rea
c
h th
e
minimum
poi
nt to ma
ke th
e erro
r
ze
ro.
Whe
n
the
net
work i
s
co
mp
lex,
the trainin
g
p
r
ocess may b
e
trapp
ed in
a local
mi
nim
u
m,
and
the conve
r
ge
nce spe
ed
is slo
w
.
In
orde
r to ove
r
come
the
s
e
shortcomin
gs i
n
the alg
o
rith
m, usin
g L-M
algo
rithm, al
so
kno
w
n
as
the
dampe
d le
ast squ
a
re
met
hod i
s
used.
It is bette
r th
an the
tra
d
itional BP
and
othe
r imp
r
ov
ed
algorith
m
in the numb
e
r of
iterations, ha
ving highe
r converg
e
n
c
e speed a
nd a
c
cura
cy.
The wei
ghts
adju
s
tments
are given by
equatio
n [7].
e
J
I
J
J
T
T
1
)
(
(7)
In the formula,
e
as error ve
ctor;
J
as the e
rro
r of the we
ight
differenti
a
l Jacobi mat
r
ix;
as
a scalar, whe
n
incre
a
ses, it is clo
s
e
to the steep
es
t de
scent
method of smaller le
arni
n
g
rate; When
droppe
d to 0, t
he al
gorithm
become
s
the
Gau
s
s -
Ne
wton metho
d
.
Therefore, th
e
L-M al
go
rith
m is a
sm
oot
h tran
sition
b
e
twee
n the
st
eepe
st de
sce
n
t method
an
d Gau
s
s
Ne
wto
n
method [11].
The sp
ecifi
c
al
ly iterative steps of L-M al
g
o
rithm a
s
follows:
Step one:
give all i
nput
s t
o
the
network a
nd
com
p
u
t
e the o
u
tput
of the
netwo
rk, th
en
adopt erro
r function to calculate the trai
n
i
ng target’
s
sum of squa
re
error;
Step two: cal
c
ulate the e
r
ror of the
weig
hts’ differenti
a
l Ja
cobi mat
r
ix
J
;
(1) T
he d
e
fini
tion of Marq
u
a
rdt’s
se
nsitiv
ity:
m
i
m
i
n
E
S
,
n
as weight
ed sum of ea
ch laye
r
of the network.
(2)
The
se
nsi
t
ivity of the recursio
n relati
ons i
s
1
1
)
)(
(
m
q
m
m
q
m
q
S
n
E
S
, the sensitivity can
throug
h the la
st layer of the
network ba
ck to the
first layer, and the
n
cal
c
ulate th
e Ja
cobi mat
r
ix.
Step three: u
s
ing the form
ula [7] find out
;
Step four:
cal
c
ulate th
e
su
m of squa
re
error
rep
eate
d
ly. If the ne
w
sum
small
e
r than
the
cal
c
ulatio
n in
step o
ne, the
use
of
)
1
(
divided
by
and the
r
e
are
.Then g
o
to ste
p
one; Otherwi
se,
multiplied by
, then g
o
to step three. Wh
en the sum of
squ
a
re e
r
ror
decrea
s
e
s
to a target erro
r, the algor
ithm
is con
s
id
ere
d
conve
r
ge
nce [12].
3. Multi-feature
Fusion Faul
t Diagno
sis
3.1. Fusion Principle
Due to the ch
ara
c
teri
stics
of large qu
an
tity
, high dim
ensi
on, co
rrel
ation and rep
eat with
each other,
when u
s
ing the
Bias reg
u
la
ri
zation L
-
M
ne
ural net
wo
rk t
o
re
cogni
ze t
he fault type of
the extracted feature, there w
ill be large net
work
computation,
slow trai
ning speed, and the
cla
ssifi
cation
effect is n
o
t good be
cau
s
e of th
e existen
c
e
of redun
da
nt informatio
n’s
interferen
ce. While, usi
ng
t
he
fu
zzy bou
ndary co
ding
method not
o
n
ly
to
re
du
ce feature, remo
ve
redu
nda
nt in
formation, b
u
t also to
maintain
t
h
e
sam
e
cl
as
sif
i
cat
i
o
n
abi
lit
y
by
esse
nt
ial
cha
r
a
c
teri
stics. Input
the
Bias
reg
u
lari
zation
L-M n
e
twork train
again,
red
u
ce the
amo
u
n
t
of
cal
c
ulatio
n, retain the key attributes, im
prove t
he
rat
e
of corre
c
t d
i
agno
si
s ultimately. The ste
p
s
of information
fusion of ga
s f
eature are shown as Fig
u
r
e 1.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Novel T
r
an
sform
e
r Fa
ult Diagn
osi
s
Appro
a
ch Based on Inform
ation Fusi
on
… (Hu
ang Xi
n-bo
)
3619
Figure 1.
The
Proce
s
s Dia
g
ram of Multi
p
le Informatio
n Fusio
n
3.2. The Esta
blishment of Bias Re
gula
r
ization Par
a
meters
A key of the
transfo
rme
r
fault diagno
si
s model i
s
to establish the
paramete
r
s
and
.
That
is ho
w to
establi
s
h
t
he si
ze
of the paramete
r
,
to ma
ke
E
E
W
,
,
stabl
e. What’
s
m
o
re,
ensurin
g th
e
netwo
rk to
a
c
hieve
the
o
p
timal. By th
e Baye
sian
formul
a
cal
c
ul
ating
,
given a
s
follows
.
The po
steri
o
r distributio
n o
f
and
base
d
on
Bias theore
m
are given b
y
equation [8]:
)
|
(
)
|
,
(
)
,
,
|
(
)
,
|
,
(
S
D
P
S
P
S
D
P
M
D
P
(8)
If the prio
r
distrib
u
tion
)
|
,
(
S
P
is a wid
e
di
st
ribution,
and
,
two va
riable
s
of the
poste
rio
r
p
r
o
bability is i
n
d
epen
dent
with the n
o
rm
al
ization fa
cto
r
)
|
(
S
D
P
.Therefore, it
is only
necessa
ry to
make the li
kelihoo
d fun
c
tion
)
,
,
|
(
S
D
P
maximum, whi
c
h
will make the posterior
distrib
u
tion of
,
maximum.
Bias m
e
thod
s focu
s
on th
e
prob
ability di
stributi
on
of weights in th
e
spa
c
e. S
on
b
ehalf of
the netwo
rk structu
r
e. Firstly, the network stru
ctu
r
e ha
s bee
n dete
r
mined an
d no
sampl
e
data.
If
a p
r
iori
di
stribution
of
weig
hts
)
,
|
(
S
P
is giv
en, the
po
sterior di
strib
u
tion of
wei
ght is
)
,
,
,
|
(
S
D
P
whe
n
a
sam
p
le data
D
ha
s b
een
set. Acco
rding t
o
the B
i
as the
o
re
m, the form
ula
is given by eq
uation [9]:
)
,
,
|
(
)
,
|
(
)
,
,
|
(
)
,
,
,
|
(
S
D
P
S
P
S
D
P
S
D
P
(9)
From th
e formula [9]
sho
w
s, i
n
o
r
de
r t
o
obtain
the
poste
rio
r
di
stribution
)
,
,
,
|
(
S
D
P
,
the prior di
st
ribution
)
,
|
(
S
P
and t
he likelih
ood
function
)
,
,
|
(
S
D
P
sho
u
ld be kn
own firstly.
The followi
ng
is the con
c
re
te solv
ing pro
c
e
ss of the two function
s.
)
exp(
)
(
1
)
,
|
(
W
W
E
Z
P
2
)
2
(
)
(
Z
(10)
W
E
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3616 – 36
24
3620
)
exp(
)
(
1
)
,
,
|
(
E
Z
S
D
P
D
2
)
2
(
)
(
N
D
Z
(11)
Note that
)
,
,
|
(
S
D
P
ha
s nothin
g
to
d
o
with
the
weight ve
ctor
, thus
s
u
bs
titut
i
ng the
prio
r di
stributi
o
n
)
,
|
(
S
P
and the li
kelihoo
d fun
c
tion
)
,
,
|
(
S
D
P
can d
e
rive
the wei
ghts
of the
poste
rio
r
distribution which
is given by Equation (12
)
.
d
E
E
Z
F
Z
S
D
P
W
F
F
)
exp(
)
,
(
)]
(
exp[
)
,
(
1
)
,
,
,
|
(
(12)
Due to th
e in
depe
ndent
b
e
twee
n
)
,
(
F
Z
and
, the maximum
of poste
rio
r
distrib
u
tion
can
be obtai
ned by mini
mizing
)
(
F
.And the corre
s
pon
ding weight i
s
re
qui
red at
this time. By
the formula (9) and
(12
)
g
e
t the formula
[13]:
)
(
)
(
)
,
(
)
,
,
|
(
D
W
F
Z
Z
Z
S
D
P
(13)
In orde
r to determin
e
the
)
,
(
F
Z
, making the
)
(
F
e
x
pand at the
minimum po
int
*
.
Becau
s
e the
gradi
ent is 0, the approxim
ation of
)
(
F
is give
n by Equation
(14).
)
(
)
(
2
1
)
(
)
(
*
*
*
H
F
F
(14)
The Hessia
n matrix is symmetri
c
positive semi definite,
so
2
1
2
1
)
(
H
H
H
T
. Let
)
(
*
2
1
H
u
, makin
g
the f
o
rmul
a (14) subsume i
n
to (10)
, the
n
the
integral
of the
both
side
s o
f
the equatio
n to get formula
(15
)
.
2
1
1
*
)
(
2
)]
)
(
[det(
)
2
(
)
,
(
*
H
e
Z
F
N
F
(15)
Makin
g
the formul
a (15
)
sub
s
um
e into (
13), then
the logarithm
is used for
the new
equatio
n. An
d the
u
s
e
of the first-ord
e
r
co
nditi
on
of optimal
worth
ca
n o
b
tain the
opti
m
al
regul
ari
z
ation
param
eter:
)
(
2
*
*
W
E
)
(
2
*
*
E
N
(16)
In the form
ula,
1
)
(
2
H
tr
N
,
N
is the t
o
tal num
ber of the n
e
twork
wei
ght
s;
)
,
0
(
N
,
as the re
al e
ffective para
m
eters in ge
neral
parame
t
ers
N
,which ref
l
ects the
actu
al
size of the ne
twork.
Thro
ugh
the
establi
s
hm
en
t of pa
ram
e
ters
and
，
ma
ki
ng the
L
-
M
a
l
gorithm
trai
n
the
neural net
work by u
s
ing th
e error
obje
c
t
i
ve function
with weig
hts. In this
way, e
n
su
re the
su
m of
squ
a
re
e
rro
r
minimum
abo
ut the n
e
two
r
k, an
d effe
ctively control t
he n
e
two
r
k complexity, wh
ich
will help to im
prove the generali
zation ability [13].
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Novel T
r
an
sform
e
r Fa
ult Diagn
osi
s
Appro
a
ch Based on Inform
ation Fusi
on
… (Hu
ang Xi
n-bo
)
3621
4. Resul
t
s and
Analy
s
is
4.1. Data So
urces
In the pap
er, the sim
u
lation test
data come
from the main tran
sfo
r
mer oil
chromatography monitori
ng Intelligent Electri
c
Dev
i
ce
(IED) normal op
eration data i
n
±660kV
conve
r
tor stat
ion in th
e e
a
s
t Yin
Chu
an,
Nin
g
Xia
pro
v
ince
and
±2
20kV
su
bstati
on in A
n
Sha
n
,
Liao Ni
ng pro
v
ince, Chi
na. The main tra
n
sformer
oil
chrom
a
tograp
hy monitorin
g
IED is install
e
d
on the cabi
ne
t of main transform
er intelli
gent co
m
pon
ents. IED not only impleme
n
ts the functi
on
of transfo
rm
er di
ssolved
gas mo
nito
ring a
nd
dat
a remote t
r
a
n
smi
ssi
on, b
u
t its con
d
ition
monitori
ng m
a
ster
station and su
bstatio
n
throug
h the optical fiber
comm
uni
cati
on system, a
nd
follows the
IEC618
50
co
mmuni
cation
proto
c
ol
s.
Th
e mo
nitorin
g
data i
s
store
d
an
d di
spl
a
yed
throug
h the station level software
of the monitori
ng
cante
r
in ±
6
60kV
conve
r
tor statio
n in the
east Yin Chu
an, Ning Xia
provin
ce, Chi
n
a. An inst
all
a
tion pictu
r
e
of DGA monit
o
ring i
s
sho
w
n as
Figure 2. Sof
t
ware
interfa
c
e is
sh
own
a
s
Fig
u
re
3. Real-time
data
fault diagn
o
s
is i
s
sh
own
as
Figure 4.
Figure 2. Installation Pictu
r
es of DGA
M
onitorin
g
IED
Figu
re 3. Re
al-time Data Diag
ram
4.2. The Res
u
lts of Simulation Tes
t
in
g Data
Based o
n
th
e operation
data of main
trans
fo
rme
r
IED monitoring an
d the
typical
c
h
arac
teris
t
ics
of the transformer fault
data, t
here a
r
e 12
0
sets of
data [1
4]. T
he
sele
cted
120
grou
ps of da
ta include
s
5 kind
s of chara
c
te
risti
c
s gas and th
e corre
s
po
nd
ing tran
sform
e
r
runni
ng state
,
in which 9
0
grou
ps’ d
a
ta are a
s
trainin
g
sampl
e
s of
diagno
stic m
odel, while th
e
remai
n
ing
3
0
group
s’
d
a
ta are a
s
test
sa
m
p
l
e
s. Ea
ch
group
data
correspon
ding
the
c
h
ar
ac
te
r
i
s
t
ics
ga
s su
ch
as
2
2
H
C
,
4
2
H
C
,
4
CH
,
2
H
and
6
2
H
C
o
f
th
e
s
e
lec
t
e
d
d
a
t
a a
r
e as
the
input
sign
al of diag
nosti
c mod
e
l. The co
de
s “1
”, ”2”, “
3
”, ”4
”,
”5” r
epr
es
ent
for ”no
r
mal
”
, “ov
e
r
h
eatin
g
in
low-te
mpe
r
at
ure
”
, “overhe
a
ting in high-temperature
”
,
“ spark di
scharg
e” an
d “arc di
scha
rge
”
of
five kinds transfo
rme
r
ru
nning
status,
and the
s
e
five kinds
of operation state as outp
u
t.
Predi
cting an
d actual fault
type curve a
r
e sho
w
n a
s
follows.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3616 – 36
24
3622
Table 1. The
T
w
o M
odel
s Simulation Result
s of Power
T
r
an
sform
e
r Fault Di
ag
nosi
s
The model simulation results of
L-M net
w
o
rk
The model simulation results of L-
M based on Bias regularization
The model simulation results of
fuzzy
L
-
M based
on Bias
regularization
The
simulatio
n
compa
r
ison
re
sults of th
e thre
e mo
de
ls o
n
the
nu
mber of iterations an
d
the corre
c
t di
agno
si
s num
ber of sample
are sh
own a
s
follows:
Table 2. The
Simulation Compa
r
ison Ta
ble
model
the model name
number of ite
r
ations
the correct diagn
osis number of samples
a
L-M neu
ral net
work
15
21
b
The fusion result
s of no feature
re
duced
23
25
c
The fusion result
s of feature r
edu
ced
15
29
4.3. The Ana
l
y
s
is of Comparing
w
i
th
other M
e
tho
d
s
In ord
e
r to fu
rther
explain
the advanta
g
e
s of
th
e Bia
s
regul
ari
z
ati
on of the fu
zzy L-M
neural net
wo
rk in fa
ult dia
gno
sis, p
u
t it in the following several
p
r
edi
ction met
hod
s and
ma
ke
comp
arative analysi
s
ba
se
d on the 30 te
sting sample
s.
Table 3. The
Simulation Compa
r
ison Ta
ble
Method
The model name
Training steps
Accurate rate
1
Gene
ral gradient
descent algorithm
175
72.88%
2
L-M neu
ral net
work algorithm
15
76.27%
3
Bias regularization fuzzy
L-M algo
r
i
thm
15
89.83%
The tran
sfo
r
mer fault sim
u
lation re
sult
s of the abov
e 3 method
s are sho
w
n a
s
follows:
By comp
arin
g
the trai
ning
steps
of metho
d
s
1 a
nd
2: th
e BP net
work achieve
s
the
target
error in th
e 1
75th ste
p
. In contrast, the
L-M al
g
o
rithm
which is
ba
sed on the
ada
ptive adjustm
ent
to optimize
n
e
twork
weig
h
t
s by the ste
epe
st
gra
d
ie
nt method a
n
d
Gau
s
s Ne
wton meth
od
just
need
s
15
ste
p
s;
What’
s
m
o
re,
com
p
a
r
i
ng the
si
mul
a
tion
re
sults
of metho
d
s 2
and
3, th
e L
-
M
neural net
wo
rk mo
del ha
s
a larg
e ga
p b
e
twee
n the
a
c
tual o
u
tput a
nd the exp
e
ct
ed outp
u
t in the
0
5
10
15
10
-4
10
-3
10
-2
10
-1
10
0
10
1
15 E
p
oc
hs
T
r
ai
ni
ng-
B
l
u
e
G
o
a
l
-
B
l
a
c
k
P
e
r
f
or
m
a
nc
e
is
0
.
0
008
90
83
3,
G
o
al
i
s
0.
0
0
1
10
0
Tr
-
B
l
u
e
T
r
a
i
n
i
n
g
SSE =
0
.
00964
696
10
0
SSW
S
quar
ed W
e
i
ght
s
=
3.
74
222
0
50
100
15
0
200
250
300
350
400
450
500
50
100
150
200
500 E
poc
hs
#
P
a
ra
m
e
t
e
rs
E
f
f
e
c
t
i
v
e N
u
m
ber
of
P
a
r
a
m
e
t
e
r
s
=
20.
1944
0
5
10
15
20
25
30
1
1.
5
2
2.
5
3
3.
5
4
4.
5
5
Te
s
t
S
a
m
p
l
e
T
h
e t
r
ans
f
o
r
m
er
f
a
ul
t
t
y
p
e
s
L-
M
net
w
o
r
k
di
agnos
i
s
c
u
r
v
e bas
ed
on B
a
y
e
s
P
r
edi
c
t
i
n
g
O
u
t
put
A
c
t
ual
O
u
t
put
10
0
10
2
T
r-B
lu
e
T
r
a
i
n
i
n
g
SSE =
1
.
88654
10
0
10
2
10
4
SSW
S
quar
ed W
e
i
ght
s
=
15.
72
37
0
5
10
15
20
25
30
35
40
45
50
20
40
60
80
100
120
140
50 E
poc
hs
#
P
a
ra
m
e
t
e
rs
E
f
f
e
c
t
i
v
e
N
u
m
ber
of
P
a
r
a
m
e
t
e
r
s
=
7.
81699
0
5
10
15
20
25
30
1
1.
5
2
2.
5
3
3.
5
4
4.
5
5
Te
s
t
S
a
m
p
l
e
T
h
e
tr
ans
for
m
er
faul
t t
y
pes
L-
M
ne
t
w
or
k
d
i
agn
os
i
s
c
u
r
v
e b
a
s
ed o
n
B
a
y
e
s
P
r
e
d
i
c
ti
n
g
O
u
tp
u
t
A
c
t
ual
O
u
t
p
ut
0
5
10
15
20
25
30
0
0.
5
1
1.
5
2
2.
5
3
3.
5
4
4.
5
5
Te
s
t
S
a
m
p
l
e
T
h
e
tr
ans
f
or
m
e
r
fa
ul
t ty
pe
s
T
he
ne
t
w
o
r
k
o
f
L
-
M
di
ag
nos
i
s
c
u
r
v
e
P
r
edi
c
t
i
n
g
O
u
t
p
ut
Ac
t
u
a
l
O
u
tp
u
t
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Novel T
r
an
sform
e
r Fa
ult Diagn
osi
s
Appro
a
ch Based on Inform
ation Fusi
on
… (Hu
ang Xi
n-bo
)
3623
6~9 te
st sam
p
les, what’s
more, the
r
e
a
r
e fault judgm
ents of tran
sf
orme
r type.
In additio
n
, compa
r
ed
to o
t
her intelli
gen
t algorith
m
, firstly, the Bia
s
regul
ari
z
ati
on fu
zzy
L-M fu
sion
m
odel n
o
t only
overcome th
e
ab
solute
rati
o of codin
g
b
ound
ary,
but
also i
m
prove
the
parall
e
l processing ability
.
As for the
23th poin
t, its fault type belongs to the spark di
scharge
whi
c
h is n
o
t an effective d
i
agno
si
s of the fault ty
pe,
becau
se of the less fault types of traini
ng
sampl
e
s. Th
e
actual o
u
tput
and expe
cte
d
output of
transfo
rme
r
are con
s
i
s
tent
with the resid
ual
test point
s. Secon
d
ly, by increa
sing t
he inp
u
t dim
ensi
on of
sa
mple data
s
,
the L-M
net
work
overcome
s th
e slo
w
co
nvergen
ce rate a
n
d
low
convergen
ce preci
s
i
on of
the stan
dard BP neu
ral
netwo
rk. Fina
lly, the GA method is so complex that
the netwo
rk is easy to fall into local opti
m
al.
More
over, the PSO method is ea
sy to fall into
local optimum an
d particl
es a
r
e easy to re
ach
prem
ature
co
nverge
nce. T
herefo
r
e,
usi
ng the Bi
as regula
r
ization
method i
s
su
perio
r to th
e
GA
algorith
m
and
PSO algorith
m
in optimizin
g the weig
ht and thre
sh
old
.
Thus, Bia
s
re
gulari
z
atio
n fuzzy L-M fu
si
on mod
e
l is
a optimizatio
n algo
rithm in trainin
g
rate and a
c
cu
racy.
5. Conclusio
n
Thro
ugh th
e
contrast of th
e above t
w
o f
u
sio
n
diag
no
sis re
sults,
it can be con
c
l
uded
th
at
th
e
p
e
r
f
o
r
man
c
e o
f
Bias
R
e
gu
la
r
i
z
a
tion
F
u
zz
y L-
M N
e
tw
or
k is
s
u
pe
r
i
or
to th
e
no
fea
t
u
r
e
redu
ction fu
si
on mod
e
l whi
c
h is Bia
s
Re
gulari
z
at
io
n L
-
M Ne
ural Ne
twork. The fo
rmer
diag
no
si
s
model
rem
o
ves th
e
red
u
n
dant featu
r
e
i
n
formatio
n, n
o
t only
retaini
ng
key attri
b
u
t
es b
u
t al
so
fully
reflectin
g
the
characte
risti
cs of input
s after
the co
mbination a
n
d
optim
izatio
n of the feature
informatio
n. What’
s
mo
re,
the forme
r
di
agno
si
s
mod
e
l achieve
s
b
e
tter cl
assification re
sult
s
and
greatly in
cre
a
se
s the a
c
curate rate i
n
the
aspe
ct of diagno
sis re
sult
s. For the form
er
optimizatio
n algorith
m
whi
c
h ma
ke the
network error re
ach the expecte
d value only after 15
iteration
s
of t
r
ainin
g
, the
a
c
cura
cy
rate
of f
ault diag
n
o
si
s i
s
89.8
3
%
. And the p
r
edi
ction
effect
wa
s
far
sup
e
rio
r
to
the
gene
ral gra
d
i
ent
de
scent method and
L-M algo
rith
m.
Thro
ugh
th
e
analysi
s
of e
x
amples, the
information
fusion fault d
i
agno
si
s method ba
se
d o
n
Fuzzy co
di
ng
boun
dary an
d
Bias reg
u
lari
zation L
-
M ne
ural net
wo
rk i
s
effective an
d feasibl
e
.
Ackn
o
w
l
e
dg
ements
This p
ape
r is sup
p
o
r
ted by Nati
onal Basi
c
Re
sea
r
ch Progra
m
of China(973
Program)(2
0
09CB7
245
07
-3)
，
S
c
ien
c
e
and T
e
ch
nol
ogy Re
se
arch and
Devel
opment p
r
o
g
rams
of Shaanxi P
r
ovinci
al
Co
mmittee(20
11
KJXX09) a
n
d
the Mi
nistry
of edu
catio
n
about "Prog
r
am
for Ne
w Ce
ntury Excellent
talents" (NCE
T-11
-10
4
3
)
．
Referen
ces
[1]
Li Bi-c
he
ng, H
uan
g Ji
e, Gao
Shi-h
a
i.
Infor
m
ation
fusio
n
te
chno
logy
a
nd
i
t
s app
licati
on.
First
Editio
n.
Beiji
ng: Nati
on
al Defe
nce Ind
u
str
y
Press. 20
10: 13-1
8
.
[2]
Han Ch
on
g-zh
ao, Z
hu Hon
g
-
y
a
n
, Dua
n
Z
h
a
n
-she
ng.
Multi-
source inf
o
rma
tion fusio
n
.
Second Ed
ition.
Beiji
ng: T
s
ingh
ua Un
iversit
y
P
r
ess. 2010: 1
5
-
20.
[3]
Peng
Nin
g
y
u
n
, W
en
Xish
an,
Che
n
Jia
n
g
bo.
Rese
arch o
n
Po
w
e
r T
r
ansfo
rmer F
ault Di
a
gnos
is
w
i
th
BPNN Method.
High Vo
ltag
e Appar
atus.
200
4; 40(3): 17
3-1
76.
[4]
YAN
Xu
e-son
g
, W
U
Qing-
H
ua, Z
H
ANG C
an. An
Improv
ed Ge
netic A
l
gorithm
an
d Its App
licati
o
n
.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2012; 1
0
(5): 1
081-
108
6.
[5]
Li L
i
n, W
a
n Z
h
i
c
ong. T
he res
e
arch o
n
po
w
e
rt ransformer
ins
u
lati
on fa
ult d
i
a
gnos
is b
a
se
d o
n
the f
u
z
z
y
three ratio met
hod.
Z
he ji
an
g Electric Pow
e
r.
2011; (2): 12-
14.
[6]
HE Jie, GUO Hui. A Modifified
Particle Sw
arm Optimization Algorit
hm.
T
E
LKOMNIKA Indon
esia
n
Journ
a
l of Elec
trical Eng
i
ne
eri
n
g
. 201
3; 11(1
0
): 5842-
58
50.
[7]
W
ang Y
o
n
g
-qi
ang,
Lv F
a
ng
-chen
g, L
i
H
e
-ming.
Synth
e
tic F
a
u
l
t Di
a
gnos
is Met
h
o
d
of
Pow
e
r
T
r
ansformer B
a
sed
on
Ro
ug
h Set T
h
e
o
ry
and
Bayes
i
a
n
Netw
ork.
Proceedings
of the CSEE. 2006;
26(8): 13
8-1
4
1
.
[8]
Yang
Li
，，
．
Sha
ng Yo
ng
Z
h
o
u
Yuefe
n
g
P
r
oba
bil
i
ty reas
oni
ng a
ndfu
z
zy techni
qu
e
app
lie
d fo
r
ide
n
tifying pow
er
transformer
m
a
l
functi
on
．
P
r
oceedings of the CSEE. 2000; 20(7): 19-23.
[9]
CHENG Jia-tang, X
IONG Wei, XU Shao-kun. Po
w
e
r T
r
a
n
sformer F
ault
Diag
nosis B
a
sed o
n
Ne
ura
l
Net
w
orks
w
i
t
h
Improve
d
Part
ic
le S
w
arm Op
timization.
High Voltag
e
App
a
r
a
tus
. 201
2; 48(
2): 42-45.
[10]
JIA Rong, XU Qi-hui, LI Hu. Po
w
e
r T
r
ansformer
F
ault Dia
gnos
is via Ne
u
r
al Net
w
ork B
a
sed on Partic
le
S
w
a
rm Optimiz
a
tion
w
i
t
h
Nei
g
hbor
hoo
d Oper
ator.
High Vo
ltage Ap
par
atus.
2008; 4
4
(1): 8
-
10.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3616 – 36
24
3624
[11]
Z
H
ANG
De-fe
ng.
MAT
L
AB n
eura
l
netw
o
rk
app
licati
on
des
ign.
First Editi
on. Bei
jin
g: Ch
ina M
a
chi
n
e
Press. 2009: 1
07-1
13.
[12]
SONG Gong-
yi
, G
uo Qing-tao
,
T
u
F
u
-rongl.
Novel Met
hod f
o
rT
ransfor
m
er
F
aults Dia
gnos
is Based o
n
T
heory of F
u
zzy Bayesia
n
Net
w
orks
．
Proceedin
g
s of the C
S
U-EPSA. 201
2; 24(2): 10
3-1
0
5
．
[13]
Cao
Dun-k
u
i.
Gas in tra
n
sfo
rmer oil
an
al
ysis f
ault
dia
g
n
o
sis a
nd
insp
e
c
tion. F
i
rst Edi
t
ion. Bei
j
i
ng:
Chin
a Electric
Po
w
e
r Press. 2
005: 11
7-1
24.
[14]
W
A
NG Xia
o
-
x
i
a
, W
A
NG
T
ao. Po
w
e
r T
r
ansformer F
ault Di
agn
osis Bas
e
d
on Ne
ural N
e
t
w
o
r
k Evolv
e
d
by
Particle S
w
arm Optimization.
Hig
h Volta
ge Eng
i
n
eeri
n
g
.
2008; 34(
11): 236
2-23
67.
Evaluation Warning : The document was created with Spire.PDF for Python.