TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.5, May 2014, pp
. 3570 ~ 35
7
7
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i5.4915
3570
Re
cei
v
ed O
c
t
ober 2
4
, 201
3; Revi
se
d Decem
b
e
r
15, 2013; Accept
ed Ja
nua
ry 5,
2014
Fault Diagnosis Based on Wavelet Genetic Neural
Network for Motor
Ke
y
ong Shao*, Lijuan Ha
n, Yang Liu,
Xinmin Wang, Feng
w
u
Z
h
ang
Schoo
l of Elect
r
ical a
nd Infor
m
ation En
gi
ne
erin
g, Northea
st Petroleum U
n
iversit
y
,
NO.199, F
a
zha
n
Roa
d
, Daq
i
n
g
, Chin
a, 86-0
459-
650
40
62
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: 1783
81
123
5
@
qq.com
A
b
st
r
a
ct
In the motor fault di
ag
nosis
techno
lo
gy, vibratio
n sig
nals
can fully refl
e
c
t the motor
o
perati
o
n
cond
itions. I
n
this p
aper,
a
l
i
ne
ar
motor
fa
ult d
i
ag
nos
is
meth
od
bas
ed
on
w
a
velet
p
a
cket a
nd
ne
u
r
al
netw
o
rk w
a
s presente
d
. The improv
e
d
ne
ur
al netw
o
rk system w
a
s desi
g
ned w
i
th vari
a
b
le h
i
dd
en l
a
y
e
r
neur
ons. T
he
netw
o
rk chose
different nu
merical v
a
lu
es
d
epe
ndi
ng o
n
d
i
fferent situati
o
ns to reach th
e
requ
ire
m
e
n
ts that i
m
pr
ovin
g
dia
gnostic
acc
u
racy a
nd
sh
o
r
tenin
g
the
dia
gnos
is ti
me. T
he li
ne
ar
moto
r
’
s
nor
mal
a
nd t
w
o common
faults vi
brati
o
n
sign
als
w
e
re
an
aly
z
e
d
a
n
d
the
vibr
atio
n sig
n
a
l
s e
n
e
r
gy
character
i
stics w
e
re extracted
through w
a
vel
e
t packe
t, then ide
n
tified fau
l
t through n
eur
al netw
o
rk. T
h
e
exper
imenta
l
results show
that this meth
od
can effe
ctively
improve th
e motor fault dia
g
n
o
sis accur
a
cy.
Ke
y
w
ords
: w
a
velet pack
e
t, fault dia
g
n
o
sis, g
enetic n
eur
al n
e
tw
ork, vibratio
n sign
als
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
The moto
r
wo
rks a
s
the m
a
in po
wer
equi
pment of
mo
d
e
rn in
du
strie
s
, and its
role i
s
self-
evident [1-3]. As the motor’s st
ru
ctural
cha
r
a
c
te
ri
stics, installation
environme
n
t, load con
d
itio
ns
and
other fa
ctors, the
mo
tor’s si
gnal
s
often contain
a lot
of n
o
ise. Sometime
s the
noi
se
will
make
the
useful si
gnal
s u
n
re
cog
n
ized [
4
]. Someti
me
s
tra
d
itional
Fouri
e
r analy
s
is
metho
d
can
not meet the
requi
rem
ents.
Espe
cially,
whe
n
mo
re
u
s
eful info
rmat
ion in the
si
g
nal is disto
r
te
d,
the Fou
r
ier
a
nalysi
s
is p
o
w
erl
e
ss. Wavelet
analysi
s
has
e
n
joyed increa
singly and
wi
dely
u
s
e
with its uniq
u
e
advantag
es on pro
c
e
ssi
n
g
non-station
a
ry sign
als [5
].
Referen
c
e [6
] propo
se
d feature
extra
c
tion method
for motor’
s f
ault sig
nal b
a
se
d on
optimal wave
let basi
s
. This method id
e
n
tifies the
opt
imal wavelet
basi
s
for spe
c
ific moto
r fault
sign
al but
di
dn’t diag
no
se spe
c
ifically
the fault
s
.
Referen
c
e [7
] prop
osed f
eature
extra
c
tion
method for m
o
tor’s inte
r-tu
r
n sh
ort ci
rcui
t based o
n
wavelet packet
.
It diagno
sed
inter-turn sh
ort
circuit fault according to the
contra
distin
ction of band’
s energy featu
r
e.
In this
pap
er,
wavel
e
t db
6
whi
c
h
ha
s hig
h
ad
aptability
for th
e fault
s
sig
nal
s i
s
co
mbined
with BP neural network. A method of wavelet and neu
ral n
e
twork sy
ste
m
with Gen
e
tic
Algorithm
s to
optimize
the
weig
hts a
nd
threshold
s
, variabl
e hid
d
e
n
layer n
e
u
r
o
n
s i
s
introdu
ced
at the sa
me
time. So that accu
racy
of the diag
no
sis and p
e
rfo
r
m
ances
of time are
imp
r
oved.
Analyze
norm
a
l sig
nal a
nd
two fault
sign
als of lin
ea
r
motor
with wavelet pa
cket’s go
od a
nalytica
l
perfo
rman
ce
of signal’
s
sli
ght cha
nge a
nd mutati
on. Extract feature vect
ors of sign
al ene
rgy
.
The feature vectors a
r
e u
s
ed as n
eural netwo
rk’
s
inp
u
t vectors.
Train the netwo
rk to re
ach the
requi
rem
ents of fault dia
gno
sis. A
c
co
rding
to the
test, this m
e
thod
can
dia
gno
se the
fa
ult
effec
t
ively.
2. Rese
arch
Metho
d
Multi-re
sol
u
tion a
nalysi
s
can b
e
a
n
eff
e
ctiv
e
time
-freque
ncy de
compo
s
ition of
sig
nal
s.
But becau
se
of its scali
ng function
cha
nge
s ba
sed on bina
ry, it has a highe
r frequ
ency
resolution in t
he low fre
q
u
ency ban
d, while in the high frequ
en
cy band freq
ue
ncy re
solutio
n
is
poor
and th
e sig
nal’s f
r
eque
ncy b
a
n
d
is divide
d
at index eq
ual interval
s.
Wavelet p
a
c
ket
analysi
s
provides a m
o
re m
e
ticulo
us met
hod of anal
y
s
is for the si
gn
al. By dividing the frequ
en
cy
band i
n
to m
u
lti-levels, th
e high
-fre
que
ncy pa
rt whi
c
h i
s
n
o
t bro
k
en
do
wn
by multi-resolut
i
on
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fault Diag
no
sis Ba
sed o
n
Wa
velet Ge
n
e
tic Ne
ural
Network for M
o
tor (Ke
y
ong S
hao)
3571
analysi
s
ca
n be furthe
r de
comp
osed. And adaptivel
y
sele
ct the frequen
cy band
according to the
cha
r
a
c
teri
stics of the sign
al. Make it match wi
th t
he sig
nal sp
ectru
m
and i
n
crea
se the
time-
freque
ncy resolution. The
r
e
f
ore wavel
e
t packet an
alysis ha
s wid
e
r p
r
acti
cal value.
In the multi-resol
u
tion ana
lysi
s, sh
ows that multi-re
so
luti
on analysi
s
divide
s the squ
a
re
-
integra
b
le sp
ace
into orth
ogon
al sum of
all sub
s
p
a
c
e
s
,
of whi
c
h
is
th
e wavel
e
t
su
bspa
ce of
wavelet functi
on. Subdivide
the wavelet sub
s
p
a
ce
s a
c
cording to bin
a
ry form and
so imp
r
ove the
freque
ncy
re
solutio
n
. Sca
l
e su
bspa
ce
and
wavel
e
t sub
s
pa
ce
are
ch
ara
c
te
rize
d by a
n
e
w
s
u
bs
pa
ce
. T
h
e
r
e
f
or
e
,
w
e
ha
ve
:
0
jj
1
j
j
UV
UW
(1)
Then
j-
1
j
j
VV
W
, the orthogon
al d
e
compo
s
ition
of
the
sp
ace
R
L
2
, can b
e
unified
a
s
00
1
1
jj
j
UU
U
by
U
n
j
. Define
the sub
s
pa
ce
U
n
j
as clo
s
u
r
e spa
c
e of
the function
x
u
n
,
U
n
j
2
is
the
clo
s
ure sp
ace of the function
x
u
n
2
, let
x
u
n
satisf
y the followin
g
two-scale e
quation
s
:
Z
k
n
k
n
Z
k
n
k
n
k
x
x
k
x
x
u
g
u
u
h
u
2
2
1
2
2
(2)
Whe
r
e
h
g
k
k
k
1
1
, the
two co
efficie
n
ts also have an ortho
g
o
nal relatio
n
sh
ip. When
n=0, by the a
bove formul
a, we have:
Z
k
k
Z
k
k
k
x
x
k
x
x
u
u
u
h
u
2
2
0
1
0
0
g
(3)
Equation
(3
)
is the t
w
o
-
scale eq
uation
of scaling
fu
nction
x
u
0
an
d
wavelet fun
c
t
i
on
x
u
1
. By
Equation
(2) an
d Equa
tion (3), spa
c
e decompo
sit
i
on ca
n be ob
tained a
s
follows:
22
1
1
nn
n
jj
j
UU
U
(4)
A sequ
en
ce
x
u
n
constructe
d by
Eq. 2 and E
q
. 3 is
calle
d
wavelet p
a
cket determi
ne
d
by basi
s
funct
i
on
x
x
u
0
.
Wavelet
pa
cket de
comp
osi
t
ion algo
rithm
:
we
get
,2
j
n
l
d
and
,2
1
jn
l
d
by
1,
jn
l
d
, where
d
is
wavelet pac
k
et c
oeffic
i
ent.
,2
1
,
2
,2
1
1
,
2
jn
j
n
lk
l
l
k
jn
j
n
ll
kl
k
dh
d
g
dd
(5)
By the inverse of the above formula
e
, the wa
velet pa
cket re
con
s
tru
c
tion alg
o
rith
m is:
1
,
,2
,2
2
2
2
jn
j
n
j
n
ll
k
l
k
lk
k
g
dh
d
d
(6)
Wavelet
anal
ysis i
s
esse
n
t
ially decomp
o
se
d th
e
si
g
nal into
ap
proximate pa
rt
s of l
o
w-
freque
ncy a
n
d
detail p
a
rts of high-f
r
eq
uen
cy, and t
hen o
n
ly to the lo
w freq
u
ency p
a
rt for the
se
con
d
de
co
mpositio
n, wh
ile the high freque
ncy pa
rt without de
co
mpositio
n. And so o
n
, we
ca
n
get the
co
efficient
s of
wav
e
let de
com
p
o
s
ition. T
he
wavelet pa
cket
analy
s
is is n
o
t only to th
e
low
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: 3570 – 35
77
3572
freque
ncy pa
rt of the decompositio
n, bu
t also to
the seco
nda
ry decompo
sition of
high frequ
en
cy
part. It is
sh
o
w
n i
n
Fig
u
re
1, with th
ree
layers of
wav
e
let pa
cket d
e
com
p
o
s
ition
as an
exam
ple
sho
w
s the
proce
s
s of
wav
e
let pa
cket d
e
com
p
o
s
ition
.
After S is
d
e
com
p
o
s
ed
i
n
to lo
w fre
q
u
ency
A1 and hi
gh
freque
ncy
D1
, A1 and
D1
will be
de
co
mposed into
more
detaile
d, low freque
ncy,
high freq
uen
cy compon
ent
s, and can be
continu
o
u
s
ly decompo
se
d down.
In Figure 1, A is an exploded outlin
e sign
al
of the low-fre
que
n
c
y, D is the explode
d
detail si
gnal
of the hig
h
freque
ncy. Th
e
numb
e
rs
b
e
h
ind in
dicate
the layer nu
mbers
of wav
e
let
packet de
co
mpositio
n.
The de
com
p
o
s
ition ha
s the
following rela
tionshi
p:
S=
AAA3+DAA3+
A
DA3+DDA3+
A
A
D
3+DAD3+
A
DD3+
DDD3
Figure 1. Wa
velet Packet
De
comp
ositio
n
The
sig
nal
e
nergy
feature
vecto
r
whi
c
h got
from
wavelet pa
cket
de
com
p
o
s
ition a
nd
recon
s
tru
c
tio
n
can provid
e a sample f
o
r BP neural
netwo
rk. So that wavelet can com
b
ine
with
neural netwo
rk. BP Netwo
r
k i
s
an erro
r back-p
r
o
pag
ation network. In the forwa
r
d propa
gatio
n
,
Input inform
a
t
ion is h
andl
e
d
from the
in
put layer to t
he hid
den l
a
yer, an
d then i
s
pa
ssed to t
he
output layer.
If
you cann
ot get
the desired outp
u
t in the output
layer, then turn to ba
ck-
prop
agatio
n, return the differen
c
e b
e
twe
en the
actual
output value and the expe
cted value al
ong
the origin
al con
n
e
c
ting chann
els. The
erro
r
re
ach
e
s the allo
wable ra
nge b
y
modifying the
con
n
e
c
tion weights of ne
urons b
e
twe
en
layers. Netwo
r
k trai
ning i
s
compl
e
ted.
As multi-hi
dd
en layers net
work i
s
high
accu
r
a
cy
b
u
t
ov
erly
com
p
lex
,
t
hat
makes t
h
e
training
time i
n
crea
sed.
In
this p
ape
r, d
e
sig
n
a
nd
use a
sin
g
le
hid
den l
a
yer ne
ural
net
work
wit
h
variable hi
dd
en layer ne
u
r
on
s and o
p
timized by Ge
netic Algorith
m
s. The abili
ty that network
obtain
s
information from t
he traini
ng i
s
con
c
e
r
ne
d a
bout the n
u
m
ber of n
ode
s on the hi
dd
en
layer. The m
o
re no
de
s the hidden laye
r has, the
stronge
r the abi
lity to
acce
ss information, and
vice versa.
If the nu
mbe
r
o
f
node
s i
n
the
hidd
en l
a
yer
is too
ma
ny, the
compl
e
xity of trai
ning
wi
ll
increa
se
and
so
me
non
co
here
n
t fa
ctors
will
appe
ar
and
affe
ct the e
n
tire
net
work, th
en
cau
s
e
exce
ssive a
n
a
stomo
s
i
s
. So the desi
gn
of hidden
lay
e
r mu
st con
s
i
der multipl
e
factors.
In this pa
per,
firstly determi
ne the rang
e
of
the numb
e
r
of neu
ro
ns
based o
n
exp
e
rien
ce
formula, and
then determine the num
b
e
r of neuron
s in the hidde
n layer by co
mpari
ng the error
ratio.The em
pirical formul
ae are:
2
2
lo
g
21
0.
43
0
.
12
2.
5
4
0
.
7
7
0.
86
lm
lm
lm
n
a
lm
n
n
m
n
(7)
Whe
r
e is the
number of n
euro
n
s in the
input
layer, is the numbe
r of neuro
n
s
in the
output layer, is the numb
e
r
of neuro
n
s in
the hidde
n la
yer, the rang
e of is 1~1
0
.
The e
n
e
r
gy o
f
vibration
sig
nal i
s
chan
ge
d.
The
main
compon
ent of t
he vibration
signal i
s
the non
statio
nary vibratio
n sign
al, noise and lo
w
freque
ncy inte
rfere
n
ce wh
e
n
away from
the
force
of the p
u
lse. Sign
al energy is sm
all. The ene
rgy of the vibration sig
nal i
s
rel
a
tively large
whe
n
ne
ar th
e force
of the
pul
se. So yo
u can u
s
e
th
e chan
ged
e
nergy
of ea
ch freq
uen
cy
band
A1
S
D1
DA2
AA
A3
D
AA3
AD
2
AA
2
AD
D
3
DDA3
DDD3
DD2
DA
D
3
AA
D
3
AD
A
3
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fault Diag
no
sis Ba
sed o
n
Wa
velet Ge
n
e
tic Ne
ural
Network for M
o
tor (Ke
y
ong S
hao)
3573
to extract th
e fault featu
r
es. T
he e
n
e
rgy of
ba
n
d
s
can b
e
obtaine
d by wavelet pa
cket
decompo
sitio
n
coeffici
ents.
Specific
step
s are a
s
follo
ws.
Table 1. Freq
uen
cy Ran
g
e
Signal Freque
nc
y
r
ange
S0 0~0.125f
S1 0.125f~0.250f
S2 0.250f~0.375f
S3 0.375f~0.500f
S4 0.500f~0.625f
S5 0.625f~0.750f
S6 0.750f~0.875f
S7 0.875f~1f
Use the vibration sig
nal
whi
c
h is
de
compo
s
ed i
n
to three l
a
yers by wavel
e
t db6 to
recon
s
tru
c
t the wavelet p
a
cket de
com
positio
n c
oeff
i
cient
s. Extract sign
als of each frequ
en
cy
band. If the f
r
equ
en
cy ra
n
ge of o
r
igin
al
sign
al
i
s
0
~
f, the de
comp
ose
d
si
gnal
s
of each laye
r is
Sj(j=0,1, …,7
)
. The freque
n
c
y rang
e is shown in Tabl
e 1.
Re
con
s
tru
c
t the wavelet p
a
cket co
effici
ents.
Strike t
he total energy of each freque
ncy
band
si
gnal.
Let the total
energy of e
a
c
h
ban
d
sign
al
a
s
Ej, the
amplitude
of
the re
co
nst
r
u
c
ted
sign
al Sj at each di
screte
point is re
pre
s
ente
d
by
x
jk
,
we have:
n
1
2
'
j
k
x
E
jk
(8)
Con
s
tru
c
t fea
t
ure vecto
r
wi
th element
s whi
c
h a
r
e en
ergy of ea
ch
band. Fe
ature vecto
r
is co
nst
r
ucte
d as follo
ws:
0
,
1
2345
6
7
=
T
E
EE
E
E
E
E
E
,,
,,
,
,
(9)
Since is a la
rger value
wh
en the ene
rgy
is la
rge
r
, the above feature
vector is no
rmalize
d
in orde
r to re
duce the amo
unt of calcula
t
ions, and the
n
rep
r
e
s
ente
d
by:
0
,
1
2
34
56
7
=
T
E
EE
E
E
E
E
E
,,
,,
,
,
(10)
7
0
2
'
'
j
j
j
E
E
E
j
(11)
In this
pap
er,
the di
agn
ost
i
c o
b
je
ct is t
he f
ault
s
of
linear moto
r.
Colle
ct lin
ear motor’
s
fault signal of
mover misali
gnment a
nd
beari
ng oute
r
race fault. Use the featu
r
e vector o
b
tai
ned
above a
s
a sample inp
u
t for neu
ral n
e
twork an
d set
the expecte
d output. Traini
ng will compl
e
te
whe
n
the
error
rea
c
h
e
s th
e allo
wabl
e
range. T
hen
the te
st samp
le will
be
ent
ered
into
neu
ral
netwo
rk to id
entify faults.
Neu
r
al net
wo
rk de
sig
n
met
hod is a
s
follo
ws:
Energy ei
gen
vectors
whi
c
h
com
e
from
th
e mo
tor statu
s
signal
treat
ed by
wavelet
packet
decompo
sitio
n
an
d recon
s
truction
a
r
e
u
s
ed
a
s
in
put
training
samp
les. Set th
e
desi
r
ed
outp
u
t.
The num
be
r of neuron
s in
the output la
yer is 3. Wh
e
n
the output i
s
(1,0,0
), it indicate
s no
rm
al
con
d
ition. Th
e output (0,1,0) indi
cate
s
mover mi
sali
gnment. Th
e
output (0,0,1
) indicates
be
aring
outer race fault.
The
neu
ral
n
e
twork is a t
h
ree
-
tier net
work con
s
isti
ng of
input
l
a
yer, hi
dden
layer
and
output layer. Since ea
ch in
put feature vector
c
ontai
n
s
eight eleme
n
ts, the numb
e
r of neuron
s in
the input laye
r is 8. Com
p
a
r
e the erro
rs
whe
n
nod
es
are different. Select nod
e 1
7
.
Use Matlab to write the BP neural net
work p
r
og
ram
and train the
neural n
e
two
r
k. Adju
st
the
net
wo
rk para
m
eters according
to
the
situatio
n. If the
re
sults of
the t
r
aining
meet
th
e
requi
rem
ents,
the test sa
mples
co
uld
be ente
r
ed
f
o
r dia
gno
si
s. If the results do
n’t meet
the
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: 3570 – 35
77
3574
requi
rem
ents,
then incre
a
s
e traini
ng sample
s and
repe
at trainin
g
until the output meet the
requi
rem
ents.
3. Results a
nd Analy
s
is
Firstly the collected vib
r
ation sig
nal
s of linear m
o
tor
are d
e
compo
s
ed a
nd reco
nstru
c
te
d
by wavel
e
t p
a
cket. Use
Matlab to
wri
t
e wavel
e
t p
a
cket
sign
al
handl
er. En
e
r
gy featu
r
e
s
are
extracted
after the de
co
mpositio
n an
d re
con
s
tr
u
c
t
i
on of the sample si
gnal
s. Re
con
s
tru
c
ted
sign
als of
each ba
nd
are
sh
own
in
Figu
re
2, Fig
u
re 3
a
nd Fi
gure
4,
6 g
r
ou
ps of
eigenve
c
tors
are
sho
w
n in Ta
b
l
e 2.
Table 2. Eige
nvector
Signal
eigenvectors
E0
E1 E2 E3
E4 E5 E6
E7
Work
status
T1
0.7403
0.2202
0.2431
0.2349
0.2186
0.2336
0.2186
0.2207
Normal
T2
0.7687
0.2321
0.2236
0.2126
0.2351
0.2270
0.2232
0.2283
Normal
T3
0.3424
0.6425
0.2225
0.2591
0.2149
0.2245
0.2690
0.2132
Mover
misalignment
T4
0.3466
0.6480
0.2229
0.2505
0.2737
0.2240
0.2067
0.1965
Mover
misalignment
T5
0.2268
0.4270
0.2849
0.5909
0.2257
0.2070
0.1988
0.2821
Outer
ring
fault
T6
0.1907
0.4524
0.2631
0.6627
0.1900
0.2186
0.2212
0.2383
Outer
ring
fault
Figure 2. Normal Signal
Figure 3. Mover Misalignm
ent
0
200
400
600
800
1000
1200
1400
1600
1800
-5
0
5
S1
0
500
1000
1500
2000
-2
0
2
S
130
0
500
1000
1500
2000
-1
0
1
S
131
0
500
1000
1500
2000
-1
0
1
S
132
0
500
1000
1500
2000
-1
0
1
S
133
0
500
1000
1500
2000
-1
0
1
S
134
0
500
1000
1500
2000
-1
0
1
S
135
0
500
1000
1500
2000
-1
0
1
S
136
0
500
1000
150
0
2000
-1
0
1
S
137
0
100
20
0
300
400
-2
0
2
S
130
0
100
20
0
300
40
0
-2
0
2
S
131
0
100
20
0
300
400
-1
0
1
S1
3
2
0
100
20
0
300
40
0
-2
0
2
S1
3
3
0
50
100
15
0
200
250
30
0
350
40
0
-5
0
5
S1
0
100
20
0
300
400
-1
0
1
S1
3
4
0
100
20
0
300
40
0
-1
0
1
S1
3
5
0
100
20
0
300
400
-1
0
1
S
136
0
100
20
0
300
40
0
-1
0
1
S
137
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fault Diag
no
sis Ba
sed o
n
Wa
velet Ge
n
e
tic Ne
ural
Network for M
o
tor (Ke
y
ong S
hao)
3575
Figure 4. Outer Rin
g
Fault
Use Matla
b
t
o
write the
BP neu
ral n
e
twork p
r
og
ram,
the de
sired
o
u
tput an
d the
resulting
feature ve
cto
r
s
whi
c
h a
r
e
use
d
as t
r
ain
i
ng sa
mple
s
are e
n
tere
d i
n
to the neu
ra
l netwo
rk. T
r
ain
the net
work
with vari
able
hidde
n laye
r
neuron
s.
Th
e
com
p
a
r
ison
of the a
c
tual
output a
nd th
e
desi
r
ed o
u
tpu
t
is sho
w
n in
Table 3.
Table 3. The
Actual Output
and The Desired Outp
ut
The desired out
p
u
t
The actual outpu
t
Motor status
1 0
0
0.9846
0.0062
0.0057
Normal
1 0
0
0.9813
0.0079
0.0067
Normal
0
1
0
0.0341
0.8393
0.0298
Mover
misalignment
0
1
0
0.0287
0.8528
0.0277
Mover
misalignment
0 0
1
0.0067
0.0051
0.9936
Outer
ring
fault
0 0
1
0.0048
0.0038
0.9869
Outer
ring
fault
Analysis the
other thre
e
group
s of
test
signal
s
with
Wavelet Pa
cket De
co
mpo
s
iti
on, the
norm
a
l sign
al
, mover misa
lignment si
gn
al and oute
r
ring fault sig
nal are
sho
w
n in Figure 5
,
Figure 6 and
Figure 7.
Figure 5. Normal Signal of the Test Grou
p
0
50
100
150
200
-1
0
1
S
130
0
50
100
150
200
-2
0
2
S
131
0
50
100
150
200
-1
0
1
S
132
0
50
100
150
200
-2
0
2
S
133
0
20
40
60
80
100
120
140
160
180
200
-5
0
5
S1
0
50
100
150
200
-1
0
1
S
134
0
50
100
150
200
-1
0
1
S
135
0
50
100
150
200
-1
0
1
S
136
0
50
100
150
200
-1
0
1
S
137
0
20
0
40
0
60
0
80
0
1
000
120
0
140
0
160
0
18
00
-5
0
5
S1
0
50
0
100
0
15
00
20
00
-2
0
2
S
130
0
50
0
100
0
15
00
20
00
-1
0
1
S
131
0
50
0
100
0
15
00
20
00
-1
0
1
S1
3
2
0
50
0
100
0
15
00
20
00
-1
0
1
S1
3
3
0
50
0
100
0
15
00
20
00
-1
0
1
S1
3
4
0
50
0
100
0
15
00
20
00
-1
0
1
S1
3
5
0
50
0
100
0
15
00
20
00
-1
0
1
S1
3
6
0
50
0
100
0
15
00
20
00
-1
0
1
S1
3
7
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: 3570 – 35
77
3576
Figure 6. Mover Misalignm
ent of the Test Group
Figure 7. Outer Rin
g
Fault
of the Test Group
The te
st eige
nvectors
are
sho
w
n i
n
Ta
b
l
e 4.
Input th
e thre
e group
s of eig
enve
c
tors to
the trained n
e
u
ral net
wo
rk.
The output of
test is sho
w
n
in Table 5.
Table 4. The
Test Eigenve
c
tors
Signal
eigenvectors
E0
E1 E2 E3
E4 E5 E6
E7
Motor
status
T’1
0.7514
0.2220
0.2256
0.2439
0.2466
0.2562
0.2458
0.2290
Normal
T’2
0.4128
0.6532
0.2726
0.2520
0.2566
0.2501
0.1941
0.2385
Mover
misalignment
T’3
0.2408
0.4574
0.2517
0.6167
0.1865
0.2460
0.2133
0.2254
Outer
ring
fault
Table 5. The
Test Outp
ut
Test output
Motor status
0.9557
0.0046
0.0051
Normal
0.0306
0.9311
0.0298
Mover
misalignment
0.0113
0.0062
0.9874
Outer
ring
fault
The 6
g
r
ou
p
s
of t
r
ainin
g
sampl
e
s a
r
e
enter
ed i
n
to
the n
eural
netwo
rk with
variabl
e
hidde
n laye
r
neuron
s a
n
d
optimized
b
y
Geneti
c
Al
gorithm.
Trai
n the
network, Th
e o
u
tpu
t
is
sho
w
n in Ta
b
l
e 6.
0
50
10
0
150
200
25
0
300
350
40
0
-5
0
5
S1
0
100
20
0
300
40
0
-2
0
2
S1
3
0
0
100
20
0
300
40
0
-2
0
2
S1
3
1
0
100
20
0
300
40
0
-2
0
2
S
132
0
100
20
0
300
40
0
-1
0
1
S
133
0
100
20
0
300
40
0
-1
0
1
S
134
0
100
20
0
300
40
0
-1
0
1
S
135
0
100
20
0
300
40
0
-1
0
1
S1
3
6
0
100
20
0
300
40
0
-1
0
1
S1
3
7
0
20
40
60
80
100
120
140
16
0
180
200
-5
0
5
S1
0
50
10
0
150
200
-1
0
1
S130
0
50
10
0
150
200
-2
0
2
S131
0
50
10
0
150
200
-1
0
1
S1
32
0
50
10
0
150
200
-2
0
2
S1
33
0
50
10
0
150
200
-1
0
1
S1
34
0
50
10
0
150
200
-1
0
1
S1
35
0
50
10
0
150
200
-1
0
1
S136
0
50
10
0
150
200
-1
0
1
S137
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fault Diag
no
sis Ba
sed o
n
Wa
velet Ge
n
e
tic Ne
ural
Network for M
o
tor (Ke
y
ong S
hao)
3577
Table 6. The
Output of The
Optimized
Network
The desired out
p
u
t
The actual outpu
t
Motor status
1 0
0
0.9994
0.0003
0.0023
Normal
1 0
0
0.9994
0.0003
0.0024
Normal
0
1
0
0.0024
0.9993
0.0014
Mover
misalignment
0
1
0
0.0020
0.9994
0.0014
Mover
misalignment
0 0
1
0.0034
0.0004
0.9979
Outer
ring
fault
0 0
1
0.0031
0.0004
0.9979
Outer
ring
fault
As can
be
se
en fro
m
the
T
able th
at the
system
i
s
mo
re stable,
th
e diagn
osi
s
a
ccura
cy
is
further imp
r
o
v
ed. Enter th
e test
sample
s to
the
optim
ize
network.
The
actu
al o
u
t
puts a
r
e
sho
w
n
in Table 7.
Table 7. The
Test Outp
ut of The Optimized Net
w
ork
Test output
Motor status
1.0000
0.0015
0.0001
Normal
0.0003
0.9963
0.0009
Mover
misalignment
0.0002
0.0035
0.9982
Outer
ring
fault
The re
sult
s show that u
s
in
g the neu
ral
netwo
rk
whi
c
h has va
riabl
e hidde
n laye
r neu
ron
s
and optimi
z
e
d
by Geneti
c
Algorithm
s to extr
act fault signal
feature an
d
diagno
se f
ault
effectively. Its
stability and accuracy
can
meet the eng
ineeri
ng ne
ed
s.
4. Conclusio
n
Vibration p
h
e
nomen
on is
prevale
n
t in t
he machine
r
y
and equi
pm
ent durin
g op
eration.
Linear motor
will have
different vibr
ation
when in different operati
ng states.
When the motor has
internal fault
s
or p
a
rts d
e
fect, the en
ergy
and a
m
plitude of the
vibration sig
nal will chan
ge.
Different fa
ults cau
s
e diffe
rent ch
aracte
ri
stics of
the vi
bration
sig
nal
. In this pa
per, a linea
r mot
o
r
fault diagn
osi
s
meth
od b
a
sed on
wavele
t packet
a
nd
neural n
e
two
r
k was presen
ted. Extract the
sign
al feature
with wavelet
packet which is se
nsit
ive to
mutation and
slight ch
ang
e in the sign
a
l
.
Then id
entify faults by ne
ural n
e
two
r
k
system
with
variable
neu
rons i
n
the hi
dden l
a
yer a
nd
optimize
d
by
Geneti
c
Alg
o
rithm
s
. The
re
sult
s
sh
o
w
the
co
rre
ctness an
d fe
asibility of the
prop
osed met
hod.
Provide a sta
t
ement that what
is exp
e
c
ted, as
stated in t
he
"Introdu
ction" ch
apter can
ultimately re
sult in "
R
e
s
ult
s
a
n
d
Di
scu
s
sion"
c
hapte
r
, so
there i
s
comp
atibility. More
over, it
can
also
be
add
e
d
the p
r
o
s
pe
ct of the devel
opment
of
re
sea
r
ch results an
d a
ppli
c
a
t
ion prospe
cts of
further
studie
s
into the nex
t (base
d
on result an
d discussion
)
.
Referen
ces
[1]
Jianj
un H
e
, Rui Z
hao. H
y
dr
oel
ectric gen
e
r
ating sets fau
l
t diag
nosis b
a
sed o
n
infor
m
ation fusi
o
n
technology
.
Jo
urna
l of Centra
l South Un
iver
s
i
ty (Science a
n
d
T
e
chno
lo
gy)
. 2007; 3
8
(2): 3
33-3
38.
[2]
Song
lin Wu, Fumin
g
Zhan
g, Xi
ao
don
g Li
n. Fault
y
dia
g
n
o
si
s of rollin
g be
a
r
ing b
a
sed
on
w
a
vel
e
t ne
ural
net
w
o
rk.
Jour
n
a
l of Air F
o
rce Engi
neer
in
g U
n
iversity (Nat
ur
al Scie
nce Edit
ion)
. 20
07; 9(1)
: 50-53.
[3]
Leve
n
t Eren, Micha
e
l J Dev
ane
y. Bear
in
g Dama
ge Detec
t
ion Via W
a
ve
l
e
t Packet Dec
o
mpos
ition o
f
the Stator Curr
ent.
IEEE Transactions on Ins
t
rum
e
ntation and Meas
urem
ent
. 2004; 53(
2): 431-4
36.
[4]
Yunl
ong
Yu
an,
Jun
Chi. F
a
ult
y
Dia
gn
osis
of Roll
in
g Be
arin
g Bas
ed
on W
a
vel
e
t Ana
l
ysis
.
Me
ch
an
i
c
al
& Electrical En
gin
eeri
ng Ma
g
a
z
i
n
e
. 200
8; 25(6): 31-3
4
.
[5]
Lizhi
Ch
eng,
Hon
g
x
ia W
a
ng
, Yong L
uo.
W
a
velet
T
heor
y a
nd A
ppl
icat
ion. Be
iji
ng: S
c
ienc
e Press.
200
4.
[6]
Z
hong
pei T
e
n
g
, Yi Gou. Research of F
eat
ure
Extr
action
from Electromotor F
ault Sig
nal Bas
ed o
n
Optimal Wavelet Basis.
Mode
rn Machi
nery
. 200
9; 33(3): 33
-35.
[7]
Kun Z
hao, Yin
ghu
i Li, Pengs
ong Ya
ng. F
eature Ex
tr
actio
n
of BLDC Inter-turn Short Circuit F
aul
t
s
Based on Wavelet Packet.
La
rge Electric Ma
chin
e an
d Hydr
aulic T
u
r
b
in
e
. 200
9; 20(4): 20
-23.
Evaluation Warning : The document was created with Spire.PDF for Python.