TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 11, Novembe
r
2014, pp. 78
5
4
~ 786
2
DOI: 10.115
9
1
/telkomni
ka.
v
12i11.60
24
7854
Re
cei
v
ed Ma
rch 2
7
, 2013;
Re
vised Ma
y 30, 2014; Accepted June 1
5
, 2014
Fault Diagnosis of Tuning Area Based on Wavelet
Neural Network
Dong Yu*
1
, Li Ya-Lan
1
, Na
n Jie-Nong
2
1
School of Aut
o
matio
n
an
d El
ectrical En
gin
e
e
rin
g
, Lanz
hou
Jiaoton
g Un
iversit
y
, L
anzh
o
u
7300
70, Ch
ina
;
2
Rail De
partme
n
t, Nanji
ng NRI
ET
Industrial C
o
., Ltd. Nanji
n
g
2111
00, Ch
ina
)
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: 6327
45
140
@
qq.com
A
b
st
r
a
ct
W
i
th the r
api
d
dev
elo
p
m
e
n
t
of chi
na r
a
ilw
a
y
, Z
P
W
-
2000A
track circu
i
t h
a
s b
een
w
i
de
l
y
use
d
.
T
unin
g
ar
ea, w
h
ich
is an i
m
p
o
rtant
part
of the Z
P
W
-
2
000A
track cir
c
uit, is n
o
t on
ly the r
e
lati
on
shi
p
betw
een t
he s
i
gn
al tra
n
s
m
is
sion
qu
aliti
e
s,
an
d d
e
ter
m
i
n
es the
effect
of el
ectrical
in
sulati
on
betw
e
en
adj
acent secti
o
ns. T
herefore,
the study of fa
ult dia
g
n
o
sis a
s
pects fo
r tuni
ng ar
ea is ur
ge
nt and si
gn
ifica
n
t.
In this pa
per, usin
g the the
o
r
y
of transmissi
on li
ne a
mo
d
e
l of track circ
uit is bu
ilt, the
comparis
on
of the
actual d
a
ta an
d exper
i
m
ent
al
data of the track surf
ace vo
ltage e
n
ve
lop
e
curve show
s the correctn
e
ss
of
this
mo
del. O
w
ning to
the
goo
d ti
me-fre
q
uency
char
ac
teristics of w
a
v
e
let
and
the
n
onli
n
e
a
r
ma
pp
i
n
g
features of ne
ural n
e
tw
ork,
a
fault dia
g
n
o
s
is of Z
P
W
-
2000A tuni
ng ar
ea bas
ed o
n
the w
a
velet n
e
u
ra
l
netw
o
rk (W
NN
) is
prop
ose
d
.
Co
mbi
n
e
d
w
i
th
the
practic
a
l fa
ilure
situ
atio
n o
f
railw
ay s
i
te, the fa
ult
dia
g
n
o
s
is
meth
od i
n
this pap
er can acc
u
rately
i
dentify failur
e
modes of
tuning area.
Ke
y
w
ords
:
Z
P
W
-
2000A track
circuit, tuning
area, fault di
ag
nosis, w
a
velet
neur
al netw
o
rk
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
With the rapi
d spread of Z
P
W-2
000A track circuit, which ha
s be
e
n
used a
s
th
e se
ction
block
sig
nalin
g eq
uipme
n
t i
n
the
normal
operation
of the
whole
roa
d
was unive
rsal
appli
c
atio
n.
Therefore, th
e no
rmal o
p
e
r
ation of th
e
equipm
ent
is cruci
a
l for transportatio
n
and p
r
od
ucti
on.
From th
e current point of t
he a
c
tual
su
rvey, the
track circuit failu
re
s a
r
e m
a
inly
con
c
e
n
trated
in
the tuning
uni
t disconn
ecte
d, bro
k
e
n
o
r
cap
a
cita
nce value d
e
cre
a
ses of
co
mpen
sation
ca
pa
cito
r
and p
o
o
r
in
sulation of
rail
-to-g
r
ou
nd
ect. Among th
e
m
, the fault
of tuning
unit
will
redu
ce
the
input sig
nal v
o
ltage of tra
c
k ci
rcuit re
cei
v
er ca
us
i
ng a
feint of track occupie
d
wh
ich i
s
se
riou
sly
affect drivin
g
efficien
cy. Moreove
r
, If the train
re
ceiv
ed a
sign
al of
adja
c
ent track ci
rcuit which is
beyond it
s tra
n
smi
ssi
on
ra
nge
sin
c
e th
e
tuning
unit fa
ilu
re, a
nd thu
s
the
sig
nal
d
e
co
ded, thi
s
will
cau
s
e
a eme
r
gen
cy b
r
a
k
in
g be
cau
s
e th
e sp
eed
of
the train ex
cee
d
s the
sp
eed
limit value, a
s
we all kn
ow, this is very da
ngerou
s. In summary,
the impact cau
s
e
d
by the failures of tuning u
n
it
is mo
re
se
rio
u
s o
n
trai
n o
peratio
n, therefore, it
is
ne
ce
ssary to
study the fault
diagn
osi
s
me
thod
of tuning unit.
In china, the
study in fault diagn
osi
s
of
tuning a
r
ea h
a
s not be
en
expand
ed in-depth, so
it is very valuable to con
duct re
se
arch in th
is are
a
. This pap
e
r
pre
s
e
n
ts a
method for
fault
diagn
osi
s
in
tuning area
based on
wa
velet neural
netwo
rk
(WNN). Currently, there are t
w
o
bindin
g
mod
e
of wavelet
analysi
s
an
d
neural net
wo
rks: one
way
is a combin
ation of wav
e
let
transfo
rm
wi
th co
nventio
nal n
e
u
r
al
netwo
rk,
typically
u
s
in
g wavelet anal
ysis of sign
al
prep
ro
ce
ssin
g, and th
en u
s
e
conve
n
tio
nal ne
ural
net
work l
earn an
d discriminat
e; anothe
r
wa
y is
a
synth
e
si
s o
f
wavelet ana
lysis and
fee
d
forward
n
e
u
r
al n
e
two
r
ks,
namely,
wav
e
let an
alysi
s
is
integrate
d
in
to the neu
ral network
comp
uti
ng. I
n
this p
ape
r both
meth
ods
are u
s
ed.
Experimental
re
sults sho
w
: the two
met
hod
s b
o
th ca
n a
c
hieve fa
u
l
t detectio
n
a
nd lo
cation
with
high a
c
cura
cy.
2. Modeling of ZPW-20
00
A Trac
k Circ
uit
ZPW-2000A t
r
ack
circuit i
s
con
s
tituted
by tw
o pa
rts:
the main tra
ck
ci
rcuit and
tuning
area
whi
c
h i
s
re
gard a
s
a extensio
n
for the se
ct
ion train
run
n
ing. The
r
e
are fou
r
carrier
freque
ncy of ZPW-2000A
track ci
rcuit:
1
700,
2
000,
2
300, 2
600, a
nd that i
s
a
r
range
d alte
rn
ately
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fault Diag
no
sis of Tu
ning
Area Based o
n
Wa
velet Ne
ural Netwo
r
k (Do
ng Yu)
7855
in 170
0 an
d
2300
for th
e do
wnlin
k,
while
ca
rri
er frequ
en
cy 2000
and
2
600 i
s
di
spo
s
ed
alternately in
the uplink. As the rail prese
n
ts
indu
ct
ive prope
rties, the compen
sation
capa
ci
tor
are
arran
ged
at equ
al di
stan
ce
s in
a
parallel
way
to re
du
ce t
he atten
uatio
n of the
sig
nal
transmissio
n
pro
c
e
ss,
wh
a
t
’s mo
re, the
cap
a
cito
r val
ue of
com
pen
sation
ca
pa
ci
tor is varie
s
d
u
e
to the differe
nt ca
rrie
r
fre
q
uen
cy. The le
ngth of
tunin
g
regi
on i
s
2
9
m, and it i
s
comp
osed by
the
F1-type tuni
n
g
unit (BA1
), air
coil, F2
-type t
uning
u
n
it (BA2) a
n
d
the rails
b
e
twee
n them
, as
sho
w
n i
n
Fig
u
re
1. For lo
w carrie
r fre
quen
cy
secti
on de
ploy F
1
-type, while
high frequ
e
n
cy
se
ction d
eplo
y
F2 type [1]. Tunin
g
unit
sho
w
s the
pa
rallel
re
son
a
n
c
e fo
r this se
ction
sign
al t
o
redu
ce the at
tenuation of the sig
nal po
wer, at the same time, a seri
es
re
sona
nce for a
d
ja
cent
segm
ent sig
n
a
l to prevent cro
s
s-
borde
r transmissio
n of signal.
Figure 1. System Configu
r
ati
on of ZPW-2000A Track
Circuit
As sh
own in
Figure 1, co
mpen
sation
capa
ci
tor i
s
se
t up for equi
d
i
stant metho
d
,
and the
first com
pen
sation cap
a
cit
o
r from the b
eginni
ng of the track circuit
as
δ
/ 2m, a
s
well as
the las
t
one from the
track ci
rcuit termin
al is
δ
/
2m , while the distan
ce be
tween the oth
e
r co
mpen
sat
i
on
cap
a
cit
o
r is
δ
m. Assumin
g
that the
δ
/ 2
m
track circui
t is equivalent
to a four-terminal network, to
comp
en
satio
n
cap
a
cita
nce for the sa
me equivale
nt. The dash
ed box in Figure 1 i
s
a fou
r
terminal network
unit(T
ZF
b
)
for the m
a
in t
r
ack
circuit, if
the num
be
r
of com
pen
sat
i
on capa
cito
r
is
N, then the e
quivalent net
work of mai
n
track
b is
ca
scade
d by the sam
e
num
ber fou
r-te
r
mi
nal
netwo
rk [2
-3]. The equivale
nt param
eters of
the four termin
al network u
n
it is:
()
(
)
22
b
ZF
gd
c
g
d
TT
T
T
(1)
W
h
er
e T
gd
is the tran
smi
ssi
on
pa
ram
e
ters
o
f
tr
ack
fo
ur
-
t
er
mina
l n
e
t
w
o
r
k
,
T
C
is the
transmissio
n para
m
eters o
f
compen
sati
on ca
pa
citor four-termi
nal
netwo
rk.
c
c
c
o
sh(
)
Z
s
in
h(
)
sinh
(
)
/Z
c
o
s
h
(
)
gd
ll
T
ll
0
1
0
2
1
c
T
jf
C
(2)
In Equation
(2),
l
repre
s
e
n
ts the lengt
h of the track qua
drip
ole,
r
e
p
r
es
en
ts th
e
prop
agatio
n con
s
tant of track ci
rcuit, Z
C
is the cha
r
acteri
stic im
p
edan
ce of th
e rail, C de
n
o
tes
the capa
cito
r valu
e of
comp
en
satio
n
capa
cito
r,
0
f
d
enote
s
the carrie
r f
r
equ
en
cy of
transmissio
n sign
al. So the equivalent n
e
twork of the
main rail b i
s
:
T
Z
b
=(T
ZF
b
)
N
(3)
For th
e eq
uivalent net
wo
rk of tuning
are
a
b,
ea
ch
of t
he tunin
g
u
n
it and th
e ai
r
coil will
be
con
s
ide
r
e
d
a
s
a
qua
dri
pole, which
combin
with th
e tra
c
k four-t
ermin
a
l net
work bet
wee
n
t
hem
make u
p
the tuning a
r
ea. T
he tran
smi
ssi
on paramete
r
s is:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 78
54 – 786
2
7856
T
TX
b
=T
F1
×T
gd
(
T
l
/2)×
T
SVA
×T
gd
(
T
l
/2)×
T
F2
(4)
Whe
r
e T
F1
is
the tran
smission pa
ramete
rs of F1
-type
tuning unit (B
A1), T
gd
deno
tes the
transmissio
n
para
m
eters o
f
trac
k four-te
r
minal
network bet
wee
n
tu
ning a
r
ea
an
d air
coil,
T
l
is
the len
g
th of
tuning
a
r
ea
, T
SVA
is the
tran
smi
ssi
o
n
pa
ram
e
ters of
air coil
qua
drip
ole,
T
F2
rep
r
e
s
ent
s the transmissio
n para
m
eter
s of F2-type tuning unit (BA
2
).
1
1
1
0
1
1
F
BA
T
Z
1
0
1
1
SV
A
SVA
T
Z
2
2
1
0
1
1
F
BA
T
Z
(5)
Whe
r
e Z
BA1
is the p
o
le im
p
edan
ce
of F1
-type tunin
g
u
n
it (BA1),
Z
SVA
is the
imp
e
d
ance of
air coil, Z
BA2
is the ze
ro im
peda
nce of F2-type tuning
unit (BA2).
Take th
e section in 170
0
H
Z carrie
r freque
ncy for
example, its
rail impe
dan
ce(Z
) and
balla
st
resi
st
o
r
(
R
d
) are as f
o
llows:
Z=
15.1
8
2.3
Ω
/k
m R
d
=1
Ω
k
m
(6)
The p
r
op
agat
ion
con
s
tant
and
cha
r
a
c
te
ristic impe
da
nce
of the
rai
l
are
sho
w
n b
e
low [4
-
5]:
2.
92
6
j
2.
55
7
2
Z
d
Z
R
2.926
j
2
.557
2
Z
cd
ZZ
R
(7)
Since the len
g
th of tuning area i
s
29m, i
n
equatio
n (4
),
T
l
is
29m, s
u
bs
tituting into
Equation (2), the coeffici
en
t of T
gd
(
T
l
/2) is
follows
:
0.01
45
(
2
.9
26
j
2
.55
7
)
0
.0
145
0.042
4
j
0
.037
1
cosh(
0
.014
5
)
=1.000
2 + j0.0016
sinh(
0
.0145
)
=0.0424 + j0.0371
gd
1.000
2
0
.001
6
0.02
92
0.21
70
T=
0.014
5
1
.
0
002
0.00
16
jj
j
As the t
r
ain
e
n
tering
the
section, th
e track
circuit di
splays
shu
n
t status. When
the trai
n
enters the i
-
th four-termi
n
a
l network, n
a
mely t
he lo
cation
of the
i-th co
mpe
n
sation
ca
pa
ci
to
r
(T
ZFi
b
), then the main track is e
quivale
nt to a fou
r-t
ermin
a
l network
ca
scad
ed
by the i-th four-
terminal
net
work
and
th
e re
st of fo
u
r-te
r
minal
net
work
num
ber for
N-i. P
r
e
s
umin
g that
the
locatio
n
of branching p
o
int
from the be
g
i
nning of t
he i
-
th four-termi
nal network i
s
x, then the i
-
th
four-te
r
min
a
l
network i
s
divided i
n
to
the follo
wi
ng
thre
e
con
d
itions a
s
sh
o
w
n i
n
Fi
gu
re
2:
shu
n
ting poi
nt on the left of compen
sation capa
ci
tor, the shu
n
ting point a
t
compen
sati
on
cap
a
cito
r, the shuntin
g poi
nt on the
right
of compen
sa
tion cap
a
cito
r.
2
2
2
2
2
2
x
x
x
Figure 2. Dist
ribution of Sh
unting Point in the i-th Qua
d
ripol
e
From the Fig
u
re 2, the tra
n
smi
ssi
on pa
rame
ters
of i-th four-terminal network
is
:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fault Diag
no
sis of Tu
ning
Area Based o
n
Wa
velet Ne
ural Netwo
r
k (Do
ng Yu)
7857
(
)
(
)
22
2
(
)
22
(
)
2
b
ZF
i
gd
c
g
d
cg
d
gd
Tx
T
T
x
TT
T
x
Tx
x
(8)
At this point, the entire m
a
i
n
rail equival
ent
to a four-t
ermin
a
l network
ca
scade
d
by the i-
th four-te
r
min
a
l netwo
rk
(T
ZFi
b
) with the rest four-termi
nal
networks i
n
N-i nu
mbe
r
s. The
transmissio
n para
m
eters is:
Ni
ZZ
F
i
Z
F
bb
b
TT
T
(9)
The len
g
th of
10km tran
smissi
on
cabl
e
can b
e
seen
as a u
n
iform
transmission
line [6],
its tran
smissi
on matrix is the sam
e
as f
o
rmul
a (2
).
In a four-terminal net
wo
rk, the relatio
n
shi
p
bet
wee
n
the two
po
rt voltage an
d cu
rrent
can b
e
expre
s
sed by the followin
g
equ
a
t
ion:
12
2
1
2
2
12
2
1
UA
U
B
I
UA
B
U
I
IC
U
D
I
I
C
D
(10)
Whe
n
cal
c
ul
a
t
ing the voltage at any poi
nt in the track surfa
c
e
,
the
entire sectio
n
is
divided into two pa
rts, left and rig
h
t, such as Figu
re 3.
a
Figure 3. Equivalent Netwo
r
k at any Poin
t (a point)
The input imp
edan
ce to the
right of the four-te
r
min
a
l n
e
twork T
R
is:
11
1
2
21
2
2
RR
RR
z
a
z
TZ
T
Z
TZ
T
(11)
Whe
r
e Z
z
is t
he input impe
dan
ce of the receive
r
. For t
he left four-te
r
minal n
e
two
r
k T
L
,
there a
r
e voltage-cu
rrent relation
ship:
11
12
11
2
1
SL
a
L
a
SL
a
L
a
UT
U
T
I
I
TU
T
I
(12)
Acco
rdi
ng to voltage-cu
rre
nt relation
shi
p
of a point: Ua
=Ia×Za, a point cu
rrent is
cal
c
ulatin
g as:
11
12
S
a
La
L
U
I
TZ
T
(13)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 78
54 – 786
2
7858
Thus, voltage
of rail surfa
c
e in any point
is:
aa
a
UI
Z
(14)
This pa
pe
r calcul
ate the voltage of ra
il
surfa
c
e ba
sed o
n
the above form
ul
a, the
simulatio
n
re
sults of rail
surface voltag
e and re
sid
u
a
l voltage of shu
n
ting are sho
w
n in Fig
u
re 4
and Figu
re 5.
Figure 4. Voltage of Rail S
u
rface
Figure
5. Re
sidual Voltage
of Train Shun
ting
In Figure 1, whe
n
the tuning unit BA1 brea
kd
ow
n, for tra
ck
circui
t a, the corre
spo
ndin
g
parall
e
l re
so
nan
ce relatio
n
shi
p
is d
e
st
royed, an
d the se
rie
s
re
son
a
n
c
e effe
ct of seg
m
e
n
t a
formed
by th
e BA2
still exi
s
ts,
so
the
si
gnal i
n
tr
ack
circuit
a i
s
n
o
t
tran
smitted
across its
ran
ge.
Whe
n
the failure o
c
curred
in tuner uni
t BA2,
for track circuit b,
the corre
s
p
ondin
g
parall
e
l
resona
nce re
lationship i
s
destroyed, b
u
t the se
rie
s
resona
nt rel
a
tionshi
p con
s
tituted by BA1 is
being, a
s
a re
sult, the p
hen
omeno
n in
se
ction
b
of b
e
yond the
tra
n
smissi
on
doe
sn’t occu
r; whil
e
se
ction a, its corre
s
p
ondin
g
seri
es a
nd
parall
e
l re
son
ance relatio
n
s
have bee
n destroyed, which
lead to
the
si
gnal
of sectio
n a t
r
an
smitted into
sectio
n b
until it
re
ach
e
s the tu
n
i
ng a
r
ea
BA2
of
se
ction b
pre
s
entin
g its
se
ries
re
so
nan
ce ch
ar
a
c
te
rist
ics fo
r the
si
gnal. When t
hese conditio
n
s
occured, the tran
smi
ssi
on
para
m
eters fo
r the entire netwo
rk h
a
ve
been chan
ge
d.
3. Fault Diag
nosis Ba
sed
on WNN
3.1. Basic Pr
inciples of
WNN
In rece
nt years, sig
nal pro
c
e
ssi
ng and
f
ault diagno
si
s tech
nologi
e
s
ba
sed on
wavelet
analysi
s
h
a
ve achieved
g
ood results
with the
co
nsta
nt improvem
ent and
rapi
d
developm
ent
of
wavelet th
eo
ry. The
wavel
e
t tran
sform i
s
a
n
e
w conv
ersi
on
metho
d
devel
ope
d
on the
ba
si
s
of
the short
-
time fouri
e
r tran
sform
which
have the
ch
a
r
acte
ri
stic of multi-re
sol
u
tion
a
nalysi
s
and
stron
g
ability of chara
c
teri
zing the lo
cal
featur
es of sign
al both in time domai
n and frequ
e
n
cy
domain.
Wave
wavel
e
t is p
r
e
s
ent
in a smalle
r are
a
. The
mathemati
c
al
definition of
wavelet
func
tion is
: let
t
is a squa
re
-integ
rabl
e fu
nction
s, that i
s
2
tL
R
, if its fourier transform
meet the following
con
d
itions:
2
R
Cd
(15)
In this
c
a
s
e
,
t
is a ba
si
c wavelet or mot
her
wavelet,
saying the fo
rmula
(15
)
i
s
permi
ssible
con
d
ition
of
wavelet fu
nction. Ta
king
stret
c
hing
a
n
d
tra
n
sl
ation
of the
mot
her
wavelet func
tion
t
, then gen
erate
s
a wav
e
let seq
uen
ce. Namely:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fault Diag
no
sis of Tu
ning
Area Based o
n
Wa
velet Ne
ural Netwo
r
k (Do
ng Yu)
7859
1
2
,
,
;
0
ab
tb
ta
a
b
R
a
a
(16)
Whe
r
e
a
is
a
stretchin
g
fa
ctor
(al
s
o
kno
w
n a
s
scal
e fa
ctor),
b
is the t
r
an
slation
factor,
,
ab
t
is
a wavelet func
tion.
In the case of the discrete,
wavelet sequ
ence is:
/2
,
(
)
2
(
2
)
,
jj
jk
tt
k
j
k
Z
(17)
The co
ntinuo
us wavelet transfo
rm for a
n
y function
2
()
(
)
f
tL
R
is
:
1/
2
,
,,
fa
b
R
tb
Wa
b
f
a
f
t
d
t
a
(18)
Figure 6 i
s
th
e mo
st sig
n
ificant a
nd ve
rsatile
rep
r
e
s
e
n
tative for 3-l
a
yer fee
d
forward BP
neural network model, which is com
p
o
s
e
d
by i
nput layer, output layer and
seve
ra
l hidden laye
rs.
It shoul
d
be
noted th
at n
e
u
ron
s
in th
e
same
laye
r d
on’t conn
ect
each oth
e
r a
nd the
a
d
jacent
layers co
nne
cted by
weig
hts. In Figu
re
6,
j
x
rep
r
e
s
ent
s a in
put of t
he j-th i
nput
l
a
yer
nod
e,
ij
rep
r
e
s
ent
s the hidden laye
r weig
hts bet
wee
n
j-th inp
u
t layer node
and i-th hid
d
e
n
layer nod
e,
i
rep
r
e
s
ent
s th
e threshold
o
f
i-th hi
dde
n
layer n
ode,
()
x
rep
r
e
s
ent
s a
c
tivation fun
c
tion in th
e
hidde
n laye
r,
ki
represents
the weight
s b
e
twee
n
k-th
output layer n
ode
and
i-th
hidde
n
laye
r
node,
k
a
means the thresh
ol
d of k-th outp
u
t layer node
,
()
x
denote
s
activation function in the
output layer,
k
O
denote
s
the o
u
tput of the k-th output layer node.
1
x
M
x
j
x
ij
1
i
j
ki
1
a
k
a
L
a
1
O
k
O
L
O
Figure 6. The
Model of 3-la
yer BP Neura
l
Netwo
r
k
The trai
ning
pro
c
e
ss
of BP netwo
rk i
s
: firstl
y, calculating the
o
u
tput of ea
ch nod
e
positively, then calculatin
g the
e
rro
r b
a
se
d on
actu
al output, fin
a
lly, adjustin
g
wei
ghts
bet
wee
n
the hid
den
la
yer an
d o
u
tp
ut layer as
well a
s
the
wei
ghts
between
the in
put lay
e
r a
nd t
he
hidden
layer in turn a
c
cordi
ng to B
P
erro
r rul
e
, in ord
e
r to red
u
ce the
error,
makin
g
the n
e
twork o
u
tput
s
meet expecta
tions.
Whe
n
BP n
e
twork traini
ng, the nu
m
ber
of
hidd
e
n
layer
and
node
s o
n
each, the
activation fun
c
tion a
nd sa
mples fo
r in
put/output
m
u
st have b
e
en spe
c
ified.
This i
s
be
cause
these p
a
ram
e
ters
will aff
e
ct the conv
erge
nc
e rate
and effectiv
ene
ss of the
BP network.
The
numbe
r of hid
den no
de
s is
giv
en by the empiri
cal formula:
H
ML
a
(19)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 78
54 – 786
2
7860
Whe
r
e
M is the nu
mbe
r
o
f
input ne
uro
n
s, L i
s
th
e
numbe
r of
o
u
tput neu
ro
n
s
,
a
is a
con
s
tant bet
wee
n
1 and 1
0
.
Wavelet tra
n
s
form
ha
s ad
vantage of time-fre
que
ncy
cha
r
a
c
teri
sti
cs
of local
a
nd zoom
feature
s
, while neu
ral
networks can a
c
hi
ev
e satisfa
c
tory perfo
rmances,
su
ch a
s
stron
g
ch
aracter of robu
stn
e
ss,
fault
tol
e
ra
n
c
e a
nd st
udie
s
ability
indep
ende
ntly,
and
h
o
w
t
o
combi
ne the
advantag
es of both ha
s be
en an
issue that p
eople
ca
re
about. In fa
ct,
They
com
b
in
e with
each other in
two
ways: one
ki
nd is auxiliary combinatio
n way, which
use
wavelet anal
ysis pre
p
rocess sign
al
p
r
epro
c
e
s
s,
th
en the
task
of learning
a
nd di
scrimin
a
t
ion
transmit to th
e ne
ural
net
work;
anot
her
approa
ch i
s
compa
c
t
stru
cture
(o
r
neste
d combin
atio
n),
that is
blen
di
ng the
wavel
e
t algo
rithm i
n
ne
ural
net
work to fo
rm
a
wavelet
neu
ral net
work. In
this
pape
r, two m
e
thod
s we
re
use
d
re
spe
c
ti
vely for fault
diagn
osi
s
of tuning a
r
ea.
3.2. Diagnos
tic Meth
ods
and Re
sults
For th
e first bin
d
ing
mode,
th
e
cha
r
a
c
teri
st
ic
vecto
r
of
rail
su
rface
voltage
is
extrac
ted by wavelet
trans
f
orm with
time-fre
que
ncy
locali
zatio
n
f
eature
s
, the
n
maki
ng
use
o
f
nonlin
ear
ma
pping
of BP netwo
rk to cl
assify featur
e
vector i
n
various
state to
impleme
n
t fault
diagn
osi
s
of tuning a
r
ea. Its structu
r
al p
r
incipl
e is sho
w
n in Figu
re
7.
Figure 7. The
Diagn
osti
c Pr
ocess of First Structure
The se
con
d
combi
nation
way
i
s
to re
pl
ace
the nonli
near
a
c
tivation
fun
c
tion of neuron
s
of nonli
nea
r
wavelet
ba
sis functio
n
, in t
h
is
articl
e, thi
s
a
r
ticle
Mo
rl
et mothe
r
wa
velet functio
n
wa
s
sele
cted a
s
t
he activation
function of hi
dden laye
r.
The structu
r
al
pro
c
e
ss
of the se
con
d
wa
y is
s
h
ow
n
in
F
i
gu
r
e
8
.
Figure 8. The
Diagn
osti
c Pr
ocess of Second Stru
cture
BP algorith
m
ha
s the
e
s
sence of
solvi
ng the
minim
u
m e
r
ror fu
n
c
tion, b
u
t the
r
e i
s
a
probl
em
of lo
w le
arning
ef
ficien
cy, slo
w
co
nver
gen
ce
and
e
a
sy to
fall into
lo
ca
l minimal
stat
e
becau
se it u
s
es
stee
pest
d
e
scent m
e
tho
d
in n
onlin
ea
r p
r
og
rammi
n
g
which
modi
fy the weig
hts
according
to
the neg
ative
gradi
ent
direction of
the e
r
ror fu
nctio
n
.
For a
bove
sh
ortco
m
ing
s
, the
followin
g
improvements h
a
v
e two metho
d
s: additi
o
nal
momentum a
nd ada
ptive learni
ng rate.
Reg
u
lating fo
rmula for
wei
ghts an
d thre
shol
ds
with a
dditional mo
mentum facto
r
is:
(1
)
(
1
)
(
)
(1
)
(
1
)
(
)
ij
i
j
ij
ii
i
km
c
p
m
c
k
bk
m
c
m
c
bk
(20)
Reg
u
lating fo
rmula for a
d
a
p
tive learnin
g
rate is:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fault Diag
no
sis of Tu
ning
Area Based o
n
Wa
velet Ne
ural Netwo
r
k (Do
ng Yu)
7861
1.05
(
)
(
1
)
(
)
(
1
)
0
.7
(
)
(
1
)
1
.0
4
(
)
(
)
othe
r
s
kE
k
E
k
k
k
Ek
Ek
k
(21)
Whe
r
e
()
E
k
is the sum of the squares of the
errors in k-th step,
is the learni
ng rate.
Whe
n
usi
ng t
he additio
nal
momentum f
a
ctor m
e
thod
, BP algorithm can find th
e global
optimal
soluti
on, while
red
u
ce
traini
ng ti
me in
ado
ptin
g ad
aptive le
arnin
g
rate.
Whe
n
the
s
e
two
method
s co
m
b
ined, mom
e
ntum-a
daptiv
e learni
ng
rate adju
s
tment algorith
m
is p
r
odu
ce
d.
In this pa
per,
BP ne
ural
n
e
twork, im
pro
v
ed BP n
eural net
wo
rk co
mbined
with
wavelet
sep
a
rately
b
y
the two
structure u
s
e
d
for tro
ubl
e
s
h
ooting. Com
pari
s
on
of di
agno
stic
re
sult is
sho
w
n in Fig
u
re 9 an
d Fig
u
re 10.
Figure 9. The
Improved BP Netwo
r
k
Combi
ned
with Wavelet in
First Way
Figure 10. Th
e Improved B
P
Netwo
r
k
Combi
ned
with Wavelet in
Secon
d
Way
As
c
a
n be seen from the figures
,
with the firs
t
st
e
p
st
r
u
ct
ur
e of
W
N
N o
n
ly 35
st
e
p
s
nee
d
to meet the requireme
nts
of the predet
ermin
ed er
ro
r 0.002; while
it only takes 33 step
s to re
ach
the pre-m
eet
req
u
ire
m
ent
s e
r
ror
of 0
.
002 w
hen
adoptin
g the
se
con
d
structure of
WNN.
By
compari
s
o
n
, rega
rdle
ss of the first stru
ctur
e or t
he se
co
nd
structu
r
e, bot
h of them can
achi
eves th
e
error re
quirement
with f
a
st
c
onve
r
ge
nce
speed.
Simultaneo
usly, compa
r
ed
the
spe
ed of im
proved
co
mb
ination al
gori
t
hm with
the
origin
al
one, as sho
w
n
i
n
Figure
11
a
nd
Figure 12.
Figure 11. Co
mpari
s
o
n
of First WNN for the
Origin
al Algorithm and Improved Algorith
m
Figure 12. Co
mpari
s
o
n
of Secon
d
WNN for
the Origin
al Algorithm a
nd Improve
d
Algorithm
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 78
54 – 786
2
7862
As ca
n be se
en from the figure
s
, the re
ductio
n
of
initial error in both two structures i
s
quickly, but a
bout 15
0 to
200
step
s, d
o
wn
wa
rd tr
en
d in the t
r
aini
ng e
rro
r b
e
comes very
sl
ow,
however, th
e
improved
al
gorithm
to
re
duce the
e
r
ror
rate
ha
s
b
een ve
ry fa
st. Mean
whil
e, the
introdu
ction
o
f
momentu
m
item avoid
th
e o
c
curren
ce
of sho
c
k eff
e
ctively an
d
has a
sig
n
ificant
smoothi
ng effect.
4. Conclusio
n
This pa
per
combine
s
wavelet transfo
rm
and neu
ra
l network, produ
cing the
wavelet
neural n
e
two
r
k
(WNN), a
n
d
ma
king full
u
s
e
of
the adv
antage
s of
th
em
to diagn
o
s
e ZPW-20
0
0
A
track ci
rcuit tuning a
r
ea. F
i
rst
of
all, this
thesis
b
u
ild the model
of
ZPW-2000A t
r
ack
circuit u
s
ing
transmissio
n
line theo
ry a
nd the
voltage of
rail
su
rface
alo
ng
with resi
dual
voltage of t
r
ai
n
shu
n
ting
are
simulate
d. T
hen th
e BP
neural n
e
two
r
k i
n
co
njun
ct
ion
with im
proved BP n
e
u
r
al
netwo
rk a
r
e
combi
ned
wit
h
wavelet an
alysis in
two
way
s
, formi
ng different
method
s of f
ault
diagn
osi
s
. Finally, use ea
ch metho
d
to diagn
ose
tuning are
a
. Wha
t
’s more, exp
e
rime
ntal re
sults
sho
w
that th
e WNN
whi
c
h formed
by
improve
d
BP netwo
rk a
n
d
wavelet an
a
l
ysis is
a mo
re
effective diag
nosti
c metho
d
s. In additio
n
, combi
ned
with on
-site a
cci
dent cases, the accura
cy of
this diag
no
stic method u
p
to 98%.
Referen
ces
[1]
Fei
Xika
ng. Pr
i
n
cipl
e
and
An
a
l
y
s
is of J
o
intl
e
ss T
r
ack Circui
t. Beijin
g: C
h
in
a Ra
il
w
a
y Pr
es
s. 199
3: 17-
91.
[2]
Z
hao L,
Liu B
,
Xu
X.
Us
in
g
the w
a
vel
e
t ana
lysis b
a
se
d no
ise c
ance
ling
techn
i
q
u
e
to detect t
h
e
integr
ity of th
e
co
mp
ensati
n
g
cap
a
citors of UM71
track
cir
c
uit
. IEEE 2005 International S
y
m
pos
ium
,
on Micro
w
a
v
e,
Antenna, Pro
p
agati
on a
nd E
M
C
T
e
chnol
og
i
e
s for W
i
reless
Communic
a
tio
n
s. 2005; 2:
126
8-12
74.
[3]
Kim MS, Lee JW
, Ko JS.
A study on th
e co
mp
ens
ation c
a
pacitor i
n
bal
la
st track circuit
.
International
Confer
ence
on
Electrical Mac
h
in
es
an
d S
y
st
ems, IEEE. 2008: 418
2-4
187.
[4]
Jin Y, Z
hen
g X, Di
ng T
.
A h
i
gh-
accuracy
p
a
ra
meter esti
mation
al
gorithm for jointl
ess frequency-shif
t
track circuit
. Internati
o
n
a
l C
o
llo
qu
ium on
Comp
uting,
C
o
mmunic
a
tio
n
, Control, an
d Mana
geme
n
t,
IEEE. 2008; 1: 750-
753.
[5]
Minin VA, Jo
h
n
ston DA, Le
p
s
k
y
NA.
A ge
nera
l
mode
l fo
r optimi
z
a
t
i
on
of track circuit para
m
eters
.
Proceedings of
the 1995 IEEE/ASME Joint
Railroad Confer
ence.
1995: 113-117.
[6]
Nikol
ov N, Ne
d
e
lch
e
v N. Stud
y o
n
centre-fe
d
boun
dl
ess track circuits.
IEE Proceedings-Electric
Power
Appl
icatio
ns
. 2
005; 15
2(5): 10
49-1
054.
Evaluation Warning : The document was created with Spire.PDF for Python.