TELKOM
NIKA
, Vol.14, No
.4, Dece
mbe
r
2016, pp. 12
84~129
1
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v14i4.4033
1284
Re
cei
v
ed Ma
y 23, 201
6; Revi
sed
No
ve
m
ber 1, 2016
; Accepte
d
Novem
b
e
r
16, 2016
Fault Diagnosis in Medium Voltage Drive Based
on Combination of Wavelet Tr
ansform and Support
Vector Machin
e
Xudong Ca
o
*
, Shaozhe Z
hou, Jingze
Li, Shaohua Zhang
F
ault
y
of Geop
h
y
s
i
cs an
d Info
rmation En
gin
e
e
rin
g
,
Chin
a U
n
iversit
y
of Pet
r
ole
u
m, Beiji
ng
, China
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: cao
x
ud
on
g07
07@
163.com
Ab
stra
ct
Now
adays, M
e
diu
m
Volta
g
e
Drive (MV
D) h
a
s b
een
w
i
de
l
y
ap
pli
e
d
in
th
e fie
l
d
of h
i
gh-
pow
ered
motor
sp
eed-r
egu
latio
n
. T
h
e
s
e types
of co
nverter
use
a l
o
t of i
n
sul
a
ted
gate
bip
o
l
a
r transl
a
tors (IGBT
s
).
So it is very i
m
portant to fin
d
an effective w
a
y to
dia
gnos
e I
G
BT
open-circ
uit faults. T
h
is
study descr
ibe
s
a
meth
od
of dia
g
nosis for IGBT
ope
n-circuit fa
ults in MVD
w
hose to
pol
ogy
is cell ser
i
es of
mu
lti-lev
e
l. T
h
is
meth
od
co
mb
i
nes w
a
ve
let tr
ansfor
m
(W
T
)
and
sup
port v
e
ctor
mach
in
e
(SVM). T
he w
a
vel
e
t transfor
m
is
used to
extract
fault featur
es
and SVM
is us
ed to cl
assify
the fau
l
t states
of a si
n
g
le
pow
er un
it. T
hen, the
traine
d SVM cl
assifier is
use
d
to scan
all
p
o
w
e
r uni
ts of
MVD seq
uenti
a
lly. Res
u
lts of
simulati
on
on
the
platfor
m
of MA
TLAB/Simuli
nk
show
that this me
th
od h
a
s a
good
dia
g
n
o
si
s capab
ility. It can di
ag
nose t
he
IGBT open-cir
c
uit faults of the w
hole i
n
ve
rter syst
em, and di
agn
osis
accuracy is u
p
to 96%. So, this
meth
od h
a
s a
goo
d ap
plic
atio
n prosp
e
ct.
Ke
y
w
ords
:
Insulated gate bipolar
transistors, m
u
ltilev
el system
s,
Support vector
m
a
c
h
ines, Wavelet
transform
s
Copy
right
©
2016 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1 Introduc
tion
In re
cent yea
r
s, MV
D ha
s
been
obtain
e
d
lar
ge-scale
appli
c
ation. It has the a
d
vantag
e
s
of high
output
voltage, lo
w
conte
n
t of ha
rmoni
cs, low
swit
chin
g fre
q
uen
cy and
so
on. Th
e thre
e-
pha
se voltag
e outp
u
ts of
MVD a
r
e
gen
erated
by
series power
mo
dule
s
.
Every power modul
e i
s
comp
osed of
rectifier-b
r
id
ge and
H-b
r
i
dge. One
H-
bridg
e
is com
posed of four IGBTs. It me
ans
that there a
r
e
a lot of IGBTs which are u
s
ed i
n
t
he inv
e
rter. But IG
BT is the flim
sie
s
t com
pon
ent
in a MVD sy
stem. So, it
is very important to
kno
w
if they h
a
ve
problem
o
r
not for MVD’
s
reliability and
powe
r
mod
u
l
e
repla
c
e
m
en
t online te
ch
n
o
logy. There alway
s
are
o
pen-ci
rcuit fault
and
sho
r
t-circuit fault stat
es
whe
n
IGB
T
s a
r
e b
r
o
k
e
n
. As for
sho
r
t-ci
rcuit fault
state, there
is
alrea
d
y prop
er
way to di
agno
se. T
h
e
voltage bet
wee
n
colle
ctor a
nd emitt
e
r of IGBT i
s
often
measured to
figure out whether thi
s
ki
nd of
fault happe
ned o
r
not. But for IGBT open
-ci
r
cuit
fault, its imp
a
c
t is not
a
s
bi
g a
s
the
form
er. Th
e m
o
tor may
still ru
n
whe
n
IGBT
o
pen-ci
rcuit fa
ult
occur.
However, the output wa
ve of MVD will get
distortion
，
and it will also introduce DC
comp
one
nt a
nd lead to i
n
crea
sing
of ha
rmoni
cs, insulation failure, heating p
r
o
b
l
e
m and
so
on
. It
even lead
s to
some
othe
r bigge
r proble
m
if this pr
obl
em is n
o
t trea
ted in a timel
y
way [1]. There
is still no
proper
way to deal with thi
s
kind of pr
obl
em for MVD.
So it is
necessary to research
some effe
ctive detectio
n
method
s.
There are so
me method
s
su
ch a
s
expe
rt sy
stem, ne
ural n
e
two
r
ks, current me
a
s
uri
n
g
method an
d
polarity detection meth
o
d
to diagno
se IGBT ope
n-ci
rcuit fault. Expert system
method [2
-3]
is ba
se
d on
accumul
a
tion
of experie
nces. It need
s t
o
list all po
ssible fault stat
es,
summ
ari
z
e
th
e rule
s a
nd
establi
s
h
a
knowl
edge
d
a
taba
se.
Whe
n
failures o
c
cur, it ju
dge
s t
h
e
fault state by
queryin
g
the
databa
se.
Ho
wever, thi
s
m
e
thod h
a
s
so
me we
akne
ss. It is difficult
to
establi
s
h a complete data
base. Artificial neural net
te
chni
que is u
s
ed to identify
fault state in the
neural n
e
two
r
k met
hod. It
doe
sn’t n
eed
the mat
hem
atical m
odel
of the di
agn
o
s
is obj
ect, a
n
d
it
also ha
s som
e
other advan
tages
such a
s
powerful
pa
rallel processi
ng ability, self-lea
rnin
g abili
ty
and g
ood fa
u
l
t-tolera
nt abil
i
ty [4]. But
this meth
od ha
s blin
dne
ss i
n
its st
ru
cture de
signi
ng
and
fall into the
l
o
cal
minim
u
m point
too
easy [5].
Cu
rre
nt me
asuri
ng m
e
thod
i
s
b
a
sed
on
the
analysi
s
of system’s outp
u
t current. It
derive
s
aver
a
ge cu
rrent pa
rk vect
or met
hod [6-7], sin
g
le
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Fault Diagnosis in Medium
Voltage Dri
v
e Bas
ed on Com
b
ination of Wavelet... (Xudong Cao)
1285
curre
n
t tran
sducer
metho
d
[8], cu
rrent
slop
metho
d
[
9
] and
so
on.
Ho
weve
r, th
ese
metho
d
s
are
sen
s
itive to t
he loa
d
. Voltage
cu
rre
nt p
o
larity jud
g
m
ent metho
d
i
s
to u
s
e th
e
polarity of P
W
M
inverter’
s
o
u
tput to diag
no
se IGBT o
p
e
n
-ci
r
cu
it fault. This m
e
thod
has th
e adv
antage
s of ra
pid
diagn
osi
s
, hi
gh reli
ability. But, it is only appli
c
able
to the two
-
lev
e
l or th
ree
-
le
vel inverter.
For
multi-level inverter, it is hard to identify
faul
t states [10]. M.
A. Rodrígue
z-Blan
co and
A.
Vázqu
e
z-Pé
rez [11]
pro
p
o
s
ed
a fa
st fau
l
t detection
scheme fo
r IGB
T
ope
n ci
rcuit usin
g ad
apti
v
e
threshold
s
du
ring the turn
-on tran
sient. This ap
pr
o
a
ch use
s
a lot of compon
ent for dete
c
tion. So,
it is not suite
d
for medium
voltage ca
scaded mult
ilev
e
l inverters. In [12], the author combi
n
ed
the wavelet a
nalysi
s
and t
he su
ppo
rt vector to
dia
g
nose IGBT open-ci
rcuit fault. But they
only
analysed for two-l
e
vel inverter, not for ca
scade multile
vel converte
r.
In this
pape
r, for casca
d
e
multilevel i
n
vert
er sy
ste
m
, wavelet t
r
an
sform
is
use
d
to
extract fault feature, multi-SVM is used
to classify
the states of IGBT fault. In
time and frequ
e
n
cy
domain, th
e
wavelet tran
sform ha
s
go
od lo
cali
za
tio
n
prope
rty. The fre
que
ncy
of co
nverte
r’s
output is va
ri
able, that is t
o
say,
it’s a n
on-stationa
ry sign
al. So
co
nventional F
o
urie
r tran
sfo
r
m is
no lon
ger
sui
t
able for
sp
e
c
trum
analy
s
i
s
he
re. SVM
is a
kind
of machi
ne le
arning al
gorith
m
s
based on
statistical the
o
ry. It can ma
ster the
ch
a
r
acteri
stics of
the sampl
e
by studying t
h
e
training d
a
ta, and then cl
assifies un
kn
own sampl
e
s.
M
u
lti-SVM is used in this pap
er to diagno
se
IGBT open
-ci
r
cuit fault.
2. Problem Formulation
2.1. Fault An
aly
s
is of Cas
cad
e Multilev
e
l In
v
e
rter
Circuit
MVD i
s
m
a
i
n
ly co
mpo
s
e
d
of
pha
se
-shifti
ng tra
n
sf
orme
r
and
p
o
we
r m
odul
e
s
. As is
sho
w
n
in Fi
g
u
re
1, it is th
e po
we
r mo
d
u
les
conne
cti
on mo
del
of
MVD. The
ad
vantage
of th
is
stru
cture is th
at it can avoi
d
high voltag
e addin
g
to IGBT. Figur
e
2 sh
ows that
the po
wer m
o
dule
is co
mpo
s
e
d
of rectifier, f
ilter and
H Bri
dge. Th
e conne
ction p
o
r
t R, S and
T whi
c
h will
be
con
n
e
c
ted to
the second
a
r
y windin
g
of phase-sh
iftin
g
transfo
rme
r
are thre
e p
hase input
s of
power mo
dul
e, and the po
rt U and
V are outputs. U
port of one p
o
we
r mod
u
le
is co
nne
cted
to V
port of anoth
e
r po
we
r mod
u
le. Finally, three p
h
a
s
e hi
gh voltage ou
tput are form
ed in this way.
Figure 1. Power m
odul
es
con
n
e
c
tion m
odel of MVD
Single
power
module
which
is compo
s
e
d
of 4
IG
BT
s i
s
a t
w
o l
e
vel in
verter. A
s
i
s
shown
in Figu
re
2, I
G
BT A+ an
d
B-, B+
and
A-
swit
ch o
n
a
t
the same
time. The
cont
rol
sign
als of
H
Bridge a
r
e g
enerated by the comp
ari
s
on of triangu
lar wave a
n
d
sinu
soid. Sinusoid is call
ed
modulatio
n wave and t
r
ian
gular wave
is calle
d carrie
r. Dead
-time
compen
satio
n
is ad
ded to
th
e
up
and do
wn tube’s co
ntrol
sign
al.
We compa
r
e
on
e
modulatio
n wave to other
different carri
e
rs
whi
c
h have
different ph
a
s
e
s
to ge
nerate co
nt
rol
si
gnal. The
differen
c
e of m
odulatio
n wa
ve
among th
ree
pha
se
s is 12
0°.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1284 – 129
1
1286
Figure 2. Prin
ciple of po
we
r modul
e
Whe
n
IGBT
open
-ci
r
cuit f
ault ha
ppe
n,
the outp
u
t vol
t
age
wave
wil
l
disto
r
t. At th
e same
time, the output current
wa
ve will
distort
too. But the output cu
rrent
wave i
s
sensitive to the l
o
ad.
So, output vo
ltage i
s
sele
cted a
s
cha
r
a
c
teri
stic pa
ra
meter. In
this pap
er,
simul
a
tion m
odule
o
f
MVD i
s
e
s
ta
blish
ed
by Si
mulink.
Fo
r o
ne p
h
a
s
e
out
put of MV
D,
its po
we
r m
o
dule
use
sa
me
modulatio
n
wave. Every powe
r
mo
d
u
le in this
p
hase uses
d
i
fferent ca
rri
er which ph
ase
differenc
e is
Ts
/2N
（
N is the level of MVD and Ts i
s
the
period of
MVD’s outp
u
t wave). As is
sho
w
n i
n
Fi
g.3, the si
mulat
i
on mo
dule
of three
-
le
ve
l
MVD
is
cr
ea
te
d
.
T
h
e mo
du
le
wh
ich
named
“To Fil
e
4” reco
rd o
u
tput
data. RL
C
module
i
s
a
s
MVD’s l
oad.
“Sine
”
, “Sin
e1”
and
“Sin
e2”
module
s
work as
sine
wav
e
gene
rato
r.
Figure 3. Simulation mod
e
l
of cascad
e three
-
level
co
nverter
Simulation m
odel of po
we
r mod
u
le is
comp
osed of
H Brid
ge, DC po
we
r an
d
SPWM
(Sine Pulse Width Mod
u
l
a
tion) wavefo
rm gene
rato
r. We open d
r
i
v
ing signal of
IGBT to simulate
open
-ci
r
cuit fault. MVD’s o
u
tput frequ
en
cy varie
s
fro
m
10HZ to 7
0
HZ. The
n
, the fault simul
a
tion
data whi
c
h
could work a
s
multi-SVM’s training
sam
p
le is a
c
q
u
ired. The
sim
u
lation mo
del
of
SPWM wavef
o
rm ge
nerato
r
is shown as
Figure 4.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Fault Diagnosis in Medium
Voltage Dri
v
e Bas
ed on Com
b
ination of Wavelet... (Xudong Cao)
1287
Figure 4. Model of SPWM waveform ge
nerato
r
2.2. Wav
e
let Trans
f
orm o
f
Po
w
e
r Module’s Output
Voltage
Wavelet transform i
s
very sens
itive to singular poi
nt of
signal. The signal
will
show
some fe
ature
s
unli
k
e noi
se’s to some
extent
when
the output wave disto
r
tio
n
happ
en
s. So,
waveform di
stortion
co
ul
d be dete
c
te
d in the
stro
ng noi
se ba
ckgro
und if
we sele
ct proper
wavelet ba
si
s and scale
p
a
ram
e
ter. Fo
r noi
se, its m
a
xima moduli
of wavelet transfo
rm redu
ce
rapidly. So, this could rai
s
e
anti-jammi
ng
capability of fault diagnosi
s
system.
A prop
er
wav
e
let ba
sis i
s
need
ed to
wa
velet tran
sform. There are
som
e
comm
on u
s
ed
wavelet ba
se
function
s
su
ch a
s
Mo
rlet, Mexico,
Haa
r
, DBN, Symlet and so on.
They mu
st meet
some
ba
sic performan
ce indicators,
such a
s
vanishing mo
ment, regul
a
r
ity, compa
c
tly
sup
porte
d, symmetry. But it is conflict fo
r vani
shi
ng m
o
ment and
co
mpactly su
pp
orted. The hi
gh
ran
k
vanishin
g moment m
ean
s that the value of
wa
velet function
decay
s to zero qui
ckly. This
bring
s
in
a hi
gh
re
solution
ratio,
but le
ad to
suppo
rt
ing le
ngth
growin
g lo
nge
r. So it n
eed
s
to
synthe
size every kind of
situation.
Lipschit
z ind
e
x
α
coul
d b
e
used to
de
scribe
si
gnal
f
eature
of pa
rtly sing
ularit
y [13]. It
mus
t
meet the following formula.
|f(x0+h)-Pn
(
h
)
|
≤
A*|h|
α
(
1
)
n
＜
α≤
n
+
1
(
2
)
n
≥
0, n
∈
Z
(
3
)
A
≥
0, A
∈
Z
(
4
)
h0
≥
0, h0
∈
Z
(
5
)
There is a th
eore
m
whi
c
h
states that wavele
t transfo
rm ca
n’t dete
c
t sing
ularity of signal
if it just has n rank vani
shi
ng mome
nt. So, t
he rank
of wavelet ba
sis mu
st be h
i
gh eno
ugh to
get
ability of detecting si
ngul
ari
t
y of signal. Generally, to
identify disco
ntinuity of nth order d
e
rivative
,
we sele
ct at least n o
r
de
r
disa
ppe
are
d
regul
ar
wavel
e
t. Some stu
d
ies di
scove
r
that Daub
echies
wavelet ba
si
s is suita
b
le to
pro
c
e
ss
ele
c
tronic
ci
rcuit sign
al. Wavel
e
t basi
s
DB
2
5
is sele
cted
to
decompo
se f
ault sign
al into 8 levels. M
a
llat
algorith
m
wa
s sel
e
ct
ed to decom
pose the out
put
wave of p
o
wer mo
dule i
n
to 8 levels.
One IGBT d
r
ive sig
nal i
s
discon
ne
cte
d
at the out
put
freque
ncy of 50HZ.
2.3. Energ
y
Eigenv
alue
of Wav
e
let Transform
First, eval
uat
e en
ergy
va
lue of
wavel
e
t de
com
p
o
s
ition coeffici
ent. The
n
, p
u
t the
evaluated
val
ue in
a
col
u
mn ve
ctor
which
coul
d
b
e
a
s
ei
genve
c
tor of a
faul
t. The e
nerg
y
of
wavelet is d
e
fined a
s
follows:
EWj=
2
2
1
|(
)
|
(
)
|
|
n
j
jk
k
St
d
t
d
(
6
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1284 – 129
1
1288
S(t) is the f
unctio
n
of a signal, an
d
EW
j is the
powe
r
of wavelet deco
m
positio
n
coeffici
ent. Specifi
c
step
s
are a
s
follows:
1.
De
comp
ose output voltag
e sign
al into 8 levels with
wavelet tran
sform.
2.
Evaluate total energy of ea
ch ba
nd si
gn
al.
If the power o
f
S (8
，ｊ
) is
E (8
，ｊ
) (j=1
,2,3,4,5,6,7,8), then
2
1
(8
,
)
|
|
n
jk
k
Ej
x
. Xjk
is
the amplitude
of recom
p
o
s
ed sig
nal.
3.
Con
s
tru
c
t eig
envecto
r ba
sed on en
ergy
of every band
signal.
T1=[E(8,1)/E E(8,2)/E E(8,
3)/E E(8,4)/E E(8,5)/E E(
8,
6)/E E(8,7)/E]
(7)
Whe
n
som
e
IGBTs op
en
-ci
r
cuit fault
s
occur, the
r
e will
be some
cha
nge
s of eigen
vector.
So, this eigen
vector
could
be the input o
f
SVM classifi
er.
3. Fault Diag
nosis Ba
sed
on Multi-SVM
SVM
is a kin
d
of
m
a
chine
learni
ng algo
rithm
which
i
s
based on stat
istic studyin
g theory
and structu
r
e
risk mi
nimizi
ng prin
cipl
e. It has m
any advantag
es i
n
solving
sm
all size sam
p
le,
nonlin
ear an
d high dimen
s
ion
a
lity in pattern identifi
c
ation p
r
obl
e
m
s. It has a solid theo
reti
cal
basi
s
an
d sim
p
le mathem
atical mod
e
l.
3.1. Principle of SVM’s Multi-clas
s Classific
a
tio
n
The b
a
si
c al
g
o
rithm of SV
M is to
put li
near in
sepa
rable p
r
o
b
lem
be ma
ppe
d i
n
to high
-
dimen
s
ion
a
l space by the kernel fun
c
tion
.
If there are m
sample
s a
s
follows:
1
,
1
,
,
,
,
,
,
2
2
1
1
y
R
x
y
x
y
x
y
x
d
m
m
(
8
)
In low dimen
s
ional spa
c
e, linear di
scrimi
nant f
unctio
n
usu
a
lly is exp
r
esse
d as foll
ows:
b
x
w
x
g
(9)
Equation of cl
assificatio
n
is as follows:
0
w
b
x
(10)
The co
ndition
of correct cl
a
ssifi
cation i
s
t
hat it must meet the followi
ng equ
ation:
0
1
)
(
y
b
x
w
i
i
(11)
The
algo
rith
m ab
ove ju
st
appli
e
s to
bi
nary
cla
s
sification p
r
o
b
le
ms. F
o
r t
h
is
pape
r,
we
need to u
s
e
multicla
ss
cla
ssifi
cation m
e
thod whic
h i
s
based on
ba
sic SVM abov
e. There are two
ways to
co
n
s
tru
c
t multi-SVM classifi
er. On
e
way is called
di
rect m
e
thod.
It chang
es
the
obje
c
tive function directly
, incorp
orate
s
multip
le cl
assificatio
n
surface pa
ra
meters into an
optimizatio
n
probl
em. It re
alize
s
m
u
lti-cl
assificati
o
n
b
y
solving th
e
optimizatio
n
probl
em. But
this
kind
of m
e
th
od i
s
com
p
le
x and
difficult
to u
s
e.
T
he
other way
is
calle
d in
dire
ct method. It g
e
ts
the ability of
multi-cl
assification by
com
b
ining multip
l
e
cl
assifiers.
This method i
s
divided i
n
to
two
kind
s. O
ne i
s
call
ed
one
-versus-rest
(O
VR) SVM
s. T
he oth
e
r i
s
called
one
-versu
s-o
ne
(OV
O
)
SVMs. The
b
a
si
c meth
od
is to
de
sign
a SVM bin
a
ry cla
ssifie
r
b
e
twee
n a
n
y two
sam
p
le
s.
So
sampl
e
s of
m catego
ry
should
be
de
signed
m*
(m
-
1
)
/ 2
SVM
bina
ry cla
ssi
fiers. Wh
en we
classify unknown samples, the
num
ber of
one classification
will be added one if it
is
corre
s
p
ondin
g
categ
o
ry. Finally, the sample
whi
c
h h
a
s the maxim
u
m numb
e
r will be recogni
ze
d
as
the final c
l
ass
i
fic
a
tion res
u
lt.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Fault Diagnosis in Medium
Voltage Dri
v
e Bas
ed on Com
b
ination of Wavelet... (Xudong Cao)
1289
3.2. The Clas
sification M
e
thod of IG
BT
Open-circui
t
Fault of MV
D
As
we
can
see from Fi
gu
re 2
that fo
ur IGBTs are u
s
ed
in
a p
o
wer m
odul
e. So, for
a
power mod
u
l
e
, there a
r
e
16
kind
s of I
G
BT fault
s
(
0
4
C
+
1
4
C
+
2
4
C
+
3
4
C
+
4
4
C
=16).
Combinin
g the
above an
alysi
s
, one to one
cla
ssifi
cation
algorith
m
(O
VO SVMs) is
more a
pprop
riate.
There a
r
e t
w
o sta
g
e
s
of t
h
is m
e
thod.
First,
we
sho
u
ld
sele
ct th
e optimal
pa
rameters,
train th
e m
u
lti-SVM cl
a
ssifie
r
with
the eig
enve
c
tors
whi
c
h
are extra
c
t
ed from
wa
velet
decompo
sitio
n
. Second, u
s
e trained
cla
s
sifier to cl
a
s
si
fy test sampl
e
s. Detail
s are as follo
ws:
1.
Select the ke
rnel functio
n
Acco
rdi
ng to the prin
cipl
e o
f
the Hilbert -
Sc
hmidt, a function
can b
e
a kern
el function as
long a
s
it satisfies the M
e
rcer
con
d
ition.
Kernel
fun
c
tions
whi
c
h are usu
a
lly use
d
are a
s
follo
ws:
Linea
r inne
r p
r
odu
ct functio
n
y
x
y
x
K
T
,
(12)
Polynomial kernel fun
c
tion
d
y
x
y
x
K
)
1
(
,
(13)
RBF ke
rnel fu
nction
2
exp
,
y
x
r
y
x
K
(14)
In this
pape
r,
we
select th
e RBF
kern
e
l
functio
n
whi
c
h
ca
n be
carri
ed
out no
nlinea
r
mappin
g
ea
si
ly.
2.
Pre-treatment
of sample da
ta and trainin
g
cla
ssifie
r
First,
we
sele
ct the ei
genv
ector which can refl
ect the
feature
of fa
ults. The
n
, the fault
s
whi
c
h have t
he sam
e
ch
a
r
acte
ri
stics are remove
d.
The trainin
g
sa
mples a
r
e g
o
t
by normalizi
ng
the feature ve
ctor.
3.
Cycli
c
al sca
n
the output wa
ve data of eac
h po
we
r mo
dule with the
trained
cla
ssif
i
er
In a com
p
let
e
cycle, all
p
o
we
r mod
u
le
s c
ould b
e
d
i
agno
se
d by the sam
e
cla
ssifie
r
.
Becau
s
e
con
t
rol sign
als
of IGBTs are gene
rate
d
by the compare of the
same frequ
e
n
cy
modulatio
n wave and the
carrie
r. We u
s
e the
cla
ssif
i
er to cy
clical
scan all the
power mo
dul
e
s
whi
c
h
we
kno
w
thei
r
se
que
nce
s
. T
hen
we kno
w
whi
c
h
IGBT a
nd
which
po
we
r m
odule
o
c
cur th
e
open
-ci
r
cuit fault. It is fea
s
ible to the
re
al MV
D, b
e
ca
use
every p
o
w
er mod
u
le
h
a
s a
unit
cont
rol
board to
co
ntrol IGBT’
s
o
p
e
rating
state
s
. This
boa
rd i
s
al
so
ca
n b
e
used to
gath
e
r the
value
of
power
modul
e’s
output d
a
t
a
an
d
send
it to the m
a
st
e
r
controller which
contain
s
the
cla
ssifie
r
in
it. Then IGBT open-ci
rcuit faults of the whole MV
D
system can be
diagn
osed by
this cla
ssifie
r
.
4. Simulation and Result Anal
y
s
is
As is shown in Figure 3, it
is a three
-
le
v
e
l MVD’s
sim
u
lation mod
e
l
.
Every phase of the
MVD is com
p
ose
d
of three power mod
u
l
e
s. We op
en IGBT’s drivin
g signal to si
mulate its failure.
Table 1
sh
ows eig
enve
c
tors of en
ergy v
a
lue
wh
ich a
r
e de
comp
ose
d
by wavelet
transfo
rm
wh
en
output freq
ue
ncy of MVD i
s
50
HZ. Th
e first
colum
n
lists the IGBT
states: “NO”
m
ean
s that the
r
e
is no IGBT fai
l
ure; “A
+” m
e
ans th
at IGBT A+ ha
s
op
e
n
-ci
r
cuit fault; “A+ A- B
+
B-” mea
n
s that
all
of them o
c
cur IGBT ope
n-circuit fa
ult. Th
e se
co
nd
col
u
mn li
sts la
b
e
ls of IGBT
o
pen-ci
rcuit fault.
“E0” in the third column
mean
s wavel
e
t transfo
rm’
s
ene
rgy value of the eighth level in low
freque
ncy pa
rt. “E1” in the fourth colum
n
mean
s wav
e
let transfo
rm’s ene
rgy value of the eight
h
level in high frequ
en
cy part. “E3” in the
fifth co
lumn mean
s wavel
e
t transfo
rm’s energy value
of
the seventh
level and
so
on. F
r
om
T
able
1,
we
get informati
on that th
e
energy of p
o
w
er
module’
s out
put is mainly focu
sed on lo
w frequ
en
cy
part. There al
so are som
e
other informa
t
ion
can b
e
got th
at the first cla
ss a
nd the fo
rth cla
s
s, the se
con
d
cla
ss and the third
class, the fifth
cla
s
s an
d the
ninth
cla
s
s, t
he
sixth cl
ass a
nd th
e te
n
t
h cla
s
s, the
eleventh
cla
s
s a
nd thi
r
tee
n
th
cla
s
s have
n
early the
sa
me en
ergy value of
f
r
eq
u
ency band. So
we
me
rg
e
the sam
e
two
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1284 – 129
1
1290
cla
s
ses i
n
to o
ne cl
ass. A
s
we
can
see from the
follo
wing table, th
e
most e
nergy focu
se
s o
n
first
four ra
nks. So, we sel
e
ct “E0”, “E1”
, “E2
”
and “E3
”
a
s
the eigenve
c
t
o
r.
Table 1. wav
e
let transfo
rm
’s ene
rgy value of one po
wer m
odul
e’s voltage outp
u
t
Fault
Label
E0 E1
E2 E3 E4
E5 E6 E7
E8
NO
0
0.7184
0.0281
0.0186
0.1093
0.1188
0.0759
0.0544
0.0381
0.0318
A+
1
0.6882
0.0550
0.0767
0.1172
0.1058
0.0687
0.0504
0.0330
0.0277
A-
2
0.2902
0.0393
0.0799
0.1169
0.1088
0.0655
0.0492
0.0338
0.0284
B+
3
0.2367
0.0481
0.0657
0.0970
0.1041
0.0595
0.0456
0.0328
0.0275
B-
4
0.6883
0.0456
0.0883
0.1221
0.1037
0.0688
0.0494
0.0341
0.0284
A+A-
5
0.1978
0.0658
0.1138
0.1287
0.0864
0.0587
0.0440
0.0279
0.0226
A+B+
6
0.1087
0.0591
0.0811
0.1284
0.0905
0.6562
0.0433
0.0276
0.0241
A+B-
7
0.7391
0.0956
0.0562
0.0930
0.1019
0.0637
0.0463
0.0308
0.0254
A-B+
8
0.8619
0.0909
0.0588
0.0817
0.1009
0.0536
0.0410
0.0303
0.0251
A-B-
9
0.2061
0.0742
0.1332
0.1107
0.0814
0.0561
0.0400
0.0277
0.0234
B+B-
10
0.1097
0.0693
0.1250
0.1196
0.0828
0.0520
0.0387
0.0279
0.0223
A+A-
B+
11
0.8382
0.1082
0.0915
0.1085
0.0626
0.0419
0.0319
0.0213
0.0178
A+A-
B-
12
0.3393
0.1156
0.1149
0.1069
0.0711
0.0531
0.0357
0.0241
0.0215
A-B+
B-
13
0.8389
0.1130
0.1307
0.0915
0.0614
0.0419
0.0283
0.0238
0.0212
A+B+
B-
14
0.3025
0.1075
0.0902
0.0886
0.0603
0.0428
0.0324
0.0217
0.0178
A+A-
B+B-
15
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
The o
u
tput freque
ncy
of
MVD
wa
s
ch
ange
d fro
m
10HZ
to 7
0
HZ. So, there
are
71
sampl
e
s for e
v
ery type of IGBT faults. We sele
cted the sam
p
le da
ta of one power modul
e as
the
SVM’s trainin
g
data. The d
a
ta of other p
o
we
r mod
u
le
s we
re dia
g
n
o
se
d by this traine
d SVM.
The SVM’s p
a
ram
e
ters
we
re o
p
timize
d
with g
r
id-re
s
e
a
rch meth
od.
Finally, the v
a
lue of
penalty facto
r
wa
s set to 3
2768. Th
e pa
ramete
r g
wa
s set to 0.5. The final resu
lt of IGBT open-
c
i
rc
uit fault c
l
ass
i
fic
a
tion in
powe
r
mod
u
l
e
who
s
e a
c
cu
racy is
96.92
62% is sh
own as Figu
re 5.
Figure 5. Dia
gno
sis result of IGBT open
-ci
r
cuit fault
The final dia
gno
sis of the
whole MVD
is finish
e
d
by this SVM classifier dia
g
n
o
sin
g
all
the po
wer m
odule
s
in tu
rn. The a
c
curacy of the remainin
g po
wer
modul
es are 9
6
.950
8
%
,
96.930
5%, 9
6
.7754%, 9
6
.8433%
and
96.720
5%. It can
be se
e
n
that
the accuracy of
all the
power mo
dul
es are high e
noug
h.
0
50
100
150
200
250
300
350
400
450
500
0
1
2
3
4
5
6
7
S
a
m
p
l
e
o
f
te
s
t
s
e
t
C
l
assi
f
i
ca
t
i
on
l
a
b
e
l
R
e
a
l
i
t
y
an
d
T
e
s
t
i
n
g
S
e
t
of
P
o
w
e
r
M
odu
l
e
R
e
a
l
i
t
y se
t
T
e
sti
n
g
se
t
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Fault Diagnosis in Medium
Voltage Dri
v
e Bas
ed on Com
b
ination of Wavelet... (Xudong Cao)
1291
5. Conclusio
n
This pa
per p
r
opo
se
s a method for IGBT
open-circuit fault detection
in Medium Voltage
Drive
wh
ose t
opolo
g
y is cel
l
se
rie
s
m
u
ltileve
l. The
pro
posed m
e
tho
d
combin
ed
WT
analy
s
is
and
SVM cla
ssifie
r
. The
ba
sic p
a
ram
e
ters of
WT
a
r
e obtai
ned
by analy
s
ing output of
power mod
u
l
e
s
in MVD
and
multi-SVM cl
assifiers a
r
e t
r
aine
d by
th
e
eigenve
c
to
rs of this
outpu
t. The remain
ing
power m
odul
es a
r
e di
agn
ose
d
by the t
r
aine
d mu
lti-SVM classifie
r
one
by one
. The sim
u
lat
i
on
results
demo
n
strate
the p
o
ssibility and
effectivene
ss with thi
s
m
e
thod to d
e
tect IGBT o
p
en-
circuit fa
ults i
n
MVD. It i
s
a
l
so
ada
ptable
for
so
m
e
MV
Ds which h
a
ve different
co
ntrol
algo
rith
ms
su
ch a
s
vecto
r
cont
rol and
voltage to frequen
cy cont
ro
l, if we put the relate
d sa
mples to train
the
multi-SVM cl
assifier.
Referen
ces
[1]
Yu Yon
g
, Jia
n
g
She
ngc
hen
g.
IGBT
Open Ci
rcuit F
ault Di
ag
nosis M
e
thod f
o
r Inverter
. Pro
c
eed
ings
of
the CSEE. 2011; 31(9): 30-35
．
[2]
Shan
g W
u
j
un,
He Z
h
e
n
g
y
o
u
.
An IGBT
Output
Po
w
e
r-B
a
s
ed D
i
ag
nos
is
of Open-
Circ
u
it F
ault i
n
Inverter.
Power System
Technology.
201
3; 4(201
3): 110
4-11
45.
[3]
Che
ngg
ua
ng
H
uan
g, Jin Z
h
ao
, Chaor
on
g W
u
.
Data-b
ase
d
inverter IGBT
ope
n-circu
i
t fault di
ag
nosi
s
in vector c
o
n
t
rol in
ductio
n
motor
driv
es
.
In IEEE Confere
n
ce
on
Industria
l El
ectronics
and
Appl
icatio
ns. 2
013: 10
39-
104
4.
[4]
MA Rodríg
uez-
B
lanc
o, A Váz
quez-P
érez, L
Hern
ánd
ez-Go
n
zál
e
z, A Pech
-Carbo
ne
ll, M
Ma
y
A
larc
ón.
IGBT
F
ault Dia
gnos
is usi
ng A
daptiv
e T
h
resh
olds d
u
ri
ng the
T
u
rn-on T
r
ans
ient.
In IEEE International
S
y
mp
osi
u
m on
Diag
nostics fo
r Electric Mach
ines. Po
w
e
r El
ectronics a
nd
Drives. 20
13: 2
41-2
48.
[5]
Estima JO, Marqu
e
s CAJ. A ne
w
ap
pro
a
c
h for real
-tim
e multip
le o
p
e
n
-circuit fau
l
t dia
gnos
is in
voltag
e-sourc
e
inverters.
IEEE Transactions
on Industry Ap
plicati
ons
. 20
11
.
[6]
Gu Y, Ni F
,
Yang D, et al. S
w
itchin
g state p
has
e s
h
ift method for thr
ee-p
hase c
u
rrent r
e
constructi
o
n
w
i
t
h
a sin
g
le D
C
-link curr
ent s
ensor.
IEEE Transacti
ons o
n
Industri
a
l Electr
onics
. 20
11
[7]
Cho
u
SF
, Lee
CT
, Ko HC, et al. A Lo
w
-
voltag
e Ri
de
throug
h Metho
d
w
i
th T
r
ansformer F
l
u
x
Comp
ensati
o
n
Capa
bil
i
t
y
of
Rene
w
a
b
l
e
Po
w
e
r Grid-si
de Co
nverters
.
IEEE
Trans, on Power
Electron
ics
. 20
14.
[8]
Ki-Hon
g K
i
m,
Yoon-
Che
u
l
Je
ung,
Do
ng-C
h
oon
L
ee.
LVR
T
Scheme of
PMSG W
i
nd P
o
w
e
r
S
y
stems
Based o
n
F
eed
back Li
ne
arizat
ion.
IEEE Transactions on Power Electronics
. 2012.
[9
]
X
u
Yo
u
,
Zh
an
g j
i
a
ny
ong
.
One-cycle Control S
t
rategy
of Three-phase PV Inve
rters Based on Negativ
e
Sequ
enc
e Vol
t
age F
e
ed for
w
ard Co
mp
en
sation
un
der
Unb
a
la
nce
d
Input V
o
ltag
e
Con
d
itio
ns
.
Proceeding of the C
SEE. 2012; 12(25): 22-30.
[10]
W
ang F
e
i, Du
arte JL, He
ndr
ix MAM. Pli
a
n
t
acti
ve an
d r
eactive
po
w
e
r
control for gr
i
d
-inter
active
converters u
n
d
e
r unb
ala
n
ce
d voltag
e di
ps.
IEEE Transactions on Power
Electronics
. 20
11.
[11]
Z
uoli
ang Z
h
an
g. F
ault dia
g
n
o
sis s
y
stem of
VF
D based
o
n
W
a
velet T
r
ansform an
d Ne
ural n
e
t
w
ork.
Master’s de
gre
e
. Chan
g sha:
Centra
l South
Univers
i
t
y
; 2
0
0
8
.
[12]
Hu Qing, W
a
n
g
Ron
g
ji
e.
F
a
u
l
t Diag
nosis T
e
chn
o
lo
gy Bas
ed on SVM i
n
Pow
e
r Electronics Circu
it.
Procee
din
g
s of
the CSEE. 2008; 28(3
1
): 107
-111.
[13] Caseir
o
JAA
，
Card
oso AJ
M
．
Fault
diagnosis on a
P
W
M rectifier AC driv
e system
with fault
tolera
nce us
in
g the av
era
g
e
current Park
’
s
vector ap
pro
a
c
h
. IEEE International E
l
ectric Machines
and Dr
ives Co
nferenc
e. Mia
m
i, F
L
, USA. 2
009: 69
5-7
0
1
．
[14]
Suroso
Suros
o
, Daru
T
r
i Nug
r
oho, T
o
shih
ik
o
N
oguc
hi. H-
Bridg
e
b
a
se
d
F
i
ve-Lev
el
Cur
r
ent-Sourc
e
Inverter for Gr
id C
o
n
necte
d
Photovo
l
taic P
o
w
e
r
Co
nditi
o
ner.
TELKOMNIKA Telec
o
mmu
n
ic
ation,
Co
mp
uting, El
ectronics a
nd
Contro
l
. 201
3; 11(3): 48
9-4
9
4
.
[15]
T
ang Qingq
u
an, Yan
Shi
c
hao.
Ope
n
-c
ircuit Fau
l
t Diag
nos
is of
Transistor i
n
Three-
leve
l
Inverter
．
Proc
eed
ings
of the CSEE. 2008; 2
8
(21): 26-
27
．
[16]
A Jidi
n, S S
a
n
u
si, T
Sutikno.
Improve
d
Out
put
Vo
ltag
e Q
ualit
y
usi
ng
Sp
ace V
e
ctor Mo
dul
ation
for
Multilev
e
l
Inve
rters.
TELKOMNIKA Telec
o
mmu
n
icati
on,
Co
mp
uting,
El
ectronics
a
nd
Contro
l
. 2
016
;
14(2): 38
7-3
8
9
.
[17]
Stepha
ne Mal
l
a
t. A
w
a
vel
e
t tour of sign
al p
r
oc
essi
ng. Lih
ua Yan
g
. Beiji
ng: Chi
na Mac
h
in
e Press.
200
2: 122-
142.
[18]
Yunj
ie
Xu, Sh
udo
ng
Xi
u. Di
agn
ostic Stud
y Ba
sed o
n
W
a
vel
e
t Packet Entrop
y an
d
W
ear Loss of
Supp
ort Vecto
r
Machin
e.
TELKOMNIKA Teleco
mmu
n
ic
ati
on, Co
mp
utin
g
,
Electronics a
nd Co
ntrol
.
201
4; 12(4): 84
7-85
4.
Evaluation Warning : The document was created with Spire.PDF for Python.