TELKOM
NIKA
, Vol. 11, No. 5, May 2013, pp. 2330 ~ 2341
ISSN: 2302-4
046
2330
Re
cei
v
ed
Jan
uary 20, 201
3
;
Revi
sed Ma
rch 1, 2
013;
Acce
pted Ma
rch 1
2
, 2013
Resear
ch on Early Fault Diagnostic Method of Wind
Turbines
Zhai Y
ongjie
*
1
, W
a
ng Do
ngfen
g
2
, Zhang Jun
y
ing
3
,
Han
Y
u
ejiao
4
Dep
a
rtme
nt of
Automatio
n
, North Chi
na Ele
c
tric Po
w
e
r Un
i
v
ersit
y
Baod
ing C
h
i
n
a
,
15100
20
82
76
*Corres
p
o
ndi
n
g
author, e-ma
i
l
:
7676
63
514
@
qq.com
A
b
st
r
a
ct
Chal
le
ngi
ng
e
n
viro
nmenta
l
f
a
ctors co
mbin
ed w
i
th
hi
gh
an
d tur
bul
en
t w
i
nds
mak
e
seri
ou
s
de
ma
nds o
n
w
i
nd turbin
es
and res
u
lt in s
i
gnific
ant
co
mpon
ent fault ra
tes. In this paper, an e
a
rly fau
l
t
dia
gnostic r
e
s
earch is co
nd
u
c
ted up
on w
i
n
d
turbin
es
. F
i
rstly, the SCADA (Supervis
o
ry
Control a
nd D
a
ta
Acquis
i
tion) sy
stem is us
ed
to analy
z
e
the
units
’
lo
ng-
ho
ur oper
atin
g d
a
ta, prep
arin
g
for the further
mo
de
lin
g w
o
rk. T
hen the MSET
(Multivari
a
te State Es
timati
on T
e
ch
ni
que) is a
d
o
p
te
d to estimate
the
temp
eratur
e of
the ge
ar b
o
x
and to
obta
i
n a
result of h
i
gh
accuracy;
w
i
th
the Movi
ng W
i
ndow
Ca
lcul
ati
o
n
(MW
C), the residu
al va
lue
b
e
tw
een the est
i
mated v
a
lu
e
a
nd the r
eal v
a
l
ue is st
u
d
ie
d to get the
dyn
a
mi
c
trend of its ave
r
age va
lu
e; ac
cordi
ng to this
trend in tr
a
i
ni
n
g
, w
e
define th
e thresh
old r
e
g
i
on of the r
e
sid
u
a
l
me
an v
a
lu
e. C
onsi
deri
ng a
ma
n-
ma
de d
e
v
iatio
n
in th
e
observ
a
tion v
e
ctors, faults
of the ge
ar box
are
simulat
ed an
d
studie
d
. W
hen
the resid
ual
mean va
lu
e curv
e excee
d
s the
setting thresh
ol
ds, an alert w
ill
be
give
n to r
e
mi
n
d
the
o
perator
s of h
i
dd
en
pr
obl
e
m
s i
n
th
e
unit. R
e
searc
h
show
s that
thi
s
ear
ly d
i
ag
no
stic
meth
od is q
u
ite
effective in det
ecting the
abn
ormal p
e
rfor
ma
nce of w
i
nd turbin
es in a re
al-
t
ime
ma
nn
er.
Ke
y
w
ords
:
SCADA, F
ault di
agn
ostic, MW
C, MSET
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Wind p
o
wer,
one of the gree
n, safe
and lo
w-
ca
rb
on ene
rgy, is so fast-deve
l
oping in
gene
rating
electri
c
ity that it has be
com
e
the f
ourth
major p
o
wer
sou
r
ce after
coal, water a
n
d
nucl
ear. It i
s
also the
onl
y rene
wa
ble
po
wer re
so
urce
that owns over one
hun
dre
d
mil
lion
kilo
watt glo
b
a
l in
stalled
capa
city apa
rt
from
wa
ter.
The
devel
o
p
ment
of wi
n
d
po
we
r
bro
ught
about a
serie
s
of p
r
o
b
lem
s
at the
sam
e
time, with t
he mai
n
tena
nce
of wi
nd t
u
rbin
es bei
ng
the
foremo
st. As
the main
co
mpone
nts of
a larg
e-sc
ale
wind tu
rbin
e
are fixed at
a heig
h
t of o
v
er
one h
und
red
meters, sp
e
c
ial e
quipm
e
n
t like
cra
n
e
s
a
r
e ne
ede
d in the
repa
iring of im
pel
lers,
gear b
o
xes a
nd gen
erato
r
s. Whe
n
it co
mes to the
u
n
its located a
t
sea, other i
m
porta
nt factors
like the b
oats’ cha
r
terin
g
a
nd we
athe
r should al
so
b
e
co
nsid
ered.
As to a win
d
plant of whi
c
h
the de
sig
ned
life-span
is
twent
y years,
the mai
n
ten
ance
co
st
ta
ke
s u
p
1
0
-1
5% of the to
tal
incom
e
; while
the ratio is 2
0
-25% to the
one on the
se
a.
Owin
g to th
e ign
o
ra
nce
of wi
nd tu
rbine
s
’ featu
r
es
and
the
lack of m
a
n
ageme
n
t
experience, the testing
and repairi
ng
system of thermal plant
s
are still wi
dely
used in the
wind
plants in o
u
r country. The
maintenan
ce of therma
l
power eq
uip
m
ent mainly covers its
sta
t
us
sup
e
rvisi
on a
nd diagn
osi
n
g method
s (life-span of th
e
metal, cavitations, scalin
g
etc.); while the
faults in
a wi
nd turbine
are ca
used
by mech
ani
ca
l
stress a
nd the
aging
of ele
c
troni
c
part
s
for
they
are
th
e major co
mpo
nents of
a wi
nd
turbine
un
it. In fact, a wind tu
rbi
ne
approximate
s
to
electroni
c e
q
u
ipment
run
n
i
ng un
de
r po
or
con
d
ition,
so
it i
s
rather ration
al to di
scuss th
e fa
ult
diagn
osin
g
m
e
thod
s
an
d maintena
nce system co
mb
ing its
own e
l
ectro
n
ic
characteri
stics. T
h
e
con
d
ition m
o
nitoring
and
fault diag
no
si
ng of
wind
tu
rbine
s
i
s
a
proce
s
s to
sup
e
rvise,
estim
a
te
and a
nalyze
the ope
rating
data of the
major
co
m
p
o
nents
(imp
ell
e
rs, g
e
a
r
box
es, ge
nerators,
transdu
ce
rs e
t
c.), so a
s
to
detect fa
ults
prom
pt
ly. Usi
ng related
co
ndition m
onit
o
ring
techniq
ues,
we can ma
ster
the
tu
rbin
es’ runni
ng state
in
a
real
-time way. Thus,
se
riou
s
damag
es t
o
the
equipm
ent ca
n be avoide
d in advan
ce a
nd the
mainte
nan
ce cost wi
ll be greatly redu
ced.
This
pap
er p
r
opo
se
s
a
state e
s
timatio
n
of the
ge
a
r
box’
s
tem
p
eratu
r
e
ba
se
d on
the
SCADA d
a
ta
and the
MS
ET. Then the
MWC resid
u
a
l statisti
cal
method i
s
a
d
opted to a
nal
yze
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2330 – 234
1
2331
the re
sult
s
and o
b
tain i
t
s mea
n
val
ue
curv
e.
Wheneve
r
thi
s
cu
rve ex
ce
eds the
setting
thresholds, a system alarm
will occur.
2. Condition
Monitoring a
nd Faul
t Dia
gnosing of
Wind Turbin
e
s
A wind tu
rbin
e, a co
mplex
system,
con
s
ist
s
of several su
bsy
s
tem
s
: towe
r, wi
n
d
wh
eel,
whe
e
l hub, p
i
tch sy
stem, geari
ng
syste
m
, yaw,
bra
k
e, gene
rato
r, variable
-
fre
q
uen
cy syste
m
,
ma
s
t
er
c
o
n
t
ro
l s
y
s
t
e
m
, var
i
a
b
l
e-
vo
lta
g
e
a
n
d
gr
id
-
t
ied s
y
s
t
em. Each
su
bs
ys
te
m is ma
de
up
o
f
several pa
rts.
These su
bsy
s
tem
s
co
ope
rate with ea
ch
other to co
n
duct compli
cated ope
ratio
n
s.
Beside
s, the dynamic
cha
r
acteri
stics of
a wind
turbi
ne incl
ude b
o
th contin
uo
us an
d dispe
r
se
parts. Bei
ng
greatly influe
ntial by som
e
external
u
n
co
ntrolla
ble
factors, su
ch as th
e win
d
’s
spe
ed, di
re
ction a
nd
alterin
g
fre
quen
cy, t
he tu
rbine
is l
o
cate
d in
so
compl
e
x a
worki
ng
co
nditi
on
that it performs differe
ntly accordi
ngly.
The pre
s
e
n
t fault diagno
si
ng method
s applie
d in wi
nd turbin
es a
r
e listed a
s
follows [1]:
the dia
gno
sti
c
m
e
thod
ba
sed o
n
statistical data,
th
e o
ne b
a
sed
on t
i
me
sequ
en
ce p
r
edi
ction, t
he
one m
odel
controllin
g, the
one
ba
sed
o
n
vibratio
n
a
nalysi
s
, and
the on
e a
dopt
ing othe
r te
sting
techni
que
s
(sound t
r
an
smi
tting [3], ultra
s
oni
c- ele
c
tri
c
cap
a
city liq
uid level te
st
[6]). Do
cu
m
ent
[3] adopts th
e BP neu
ral
netwo
rk to co
nstru
c
t the m
odel of g
e
a
r
box and
gen
erato
r
, and
u
s
e
s
the Multi-age
nt metho
d
to
analy
z
e th
e
diag
no
stic
result
s of
different
compo
nents in
orde
r to
demon
strate the overall op
erating
state
of t
he unit. Howeve
r, the modelin
g pro
c
e
ss u
pon n
e
u
ral
netwo
rk the
o
ry takes rath
er long a time for trainin
g
, while trainin
g
sample
s are al
ways difficult
to
sele
ct. And t
he si
gnal
s’ a
c
qui
sition
sp
eed
can
ha
rd
ly meet the n
eed
s of the
analysi
s
of hi
gh-
freque
ncy vib
r
ation. Docu
ments [4
-6] summari
ze th
e variou
s wa
ys of con
d
itio
n monitori
ng
for
wind tu
rbine
s
in the re
cent years. Docum
ent
[7-1
0] con
s
tru
c
ts the hard
w
a
r
e expe
rimen
t
al
platform fo
r g
ear b
o
x and
gene
rato
r. Th
ough it a
naly
z
e
s
the vib
r
at
ion si
gnal
s u
s
ing the
wavel
e
ts
analysi
s
m
e
thod, the
mo
d
e
l is quite
different
from
th
e p
r
a
c
tical
st
ate. Do
cu
me
nt [11] dia
g
n
o
se
s
the faults b
a
sed on
an a
u
tomatic a
naly
s
is
of
the SCADA data. It doe
s not
co
rrelate some
key
factors, su
ch
as vibration
sco
pe, temp
eratu
r
e,
po
wer an
d sta
r
t-stop record
s,
giving rise to a
relatively sep
a
ration of stu
d
y content
s a
nd diag
no
stic result
s.
To monito
r the condition
of wind tu
rbi
nes,
a
dyna
mic mo
del of
its normal o
peratio
n
sho
u
ld be
co
nstru
c
ted, ba
sing o
n
whi
c
h
the early
sig
n
s of abn
orm
a
l acts a
r
e te
sted. Co
nsi
d
e
r
ing
the ran
dom
cha
nge
of wind spee
d, the great
turbulen
ce of
external su
rroundi
ng
fa
ctors
(tempe
ratu
re), the great differen
c
e
s
bet
wee
n
di
fferen
t
units, the close coupli
ng
relation am
o
n
g
all the mech
a
n
ical a
nd ele
c
tri
c
com
pon
ents, it is
difficult to apply tradition
al mo
nitoring the
o
ri
es
and metho
d
s
in wind turbin
es.
The no
n-lin
e
a
r mod
e
ling i
s
a metho
d
based on th
e ope
rating
data of the o
b
ject. By
analyzi
ng a
n
d
processin
g
these
data, t
he dyna
mic-c
hara
c
te
risti
c
model
i
s
con
s
tru
c
ted, whi
c
h
i
s
the so-call
e
d
data-d
r
iven modelin
g me
thod. Wi
th pl
enty of operating data from the SCADA
system, the turbin
es’
con
d
i
tion can b
e
monitored to discover e
a
rl
y faults.
3. SCADA Data Anal
y
s
is
in Wind Farms
A
large win
d
farm
is always
eq
uipp
ed with
the SCADA s
y
s
t
em. Its bas
ic
func
tion is
to
record the
massive o
r
igi
nal data
at a fixed time
interval (ge
nerally 1
0
s
or 1
0
min
)
i
n
the
sup
e
rvisory compute
r
s
of central cont
rol
room.
T
h
e
s
e
data
main
ly cove
rs out
put en
ergy,
state
and ala
r
m informatio
n, fault information, transdu
ce
r pa
ramete
rs a
n
d
so forth. The
quantity of th
e
SCADA
data
is
so
big th
at
the mo
nthly reco
rd
s of
a
si
ngle
unit i
s
a
s
m
any a
s
hu
ndre
d
s MB.
At
pre
s
ent, th
e
SCADA
data
in wi
nd fa
rm
s are me
rely
u
s
ed
in
monito
ring
the
data,
gen
eratin
g t
h
e
repo
rt form
s
and recalling
acci
dent
s after a fault o
c
curs.
Re
corde
d
in com
pute
r
s, the ma
ssi
v
e
data are r
e
g
u
larly co
pied
to discs
with
out being or
g
anized and a
nalyze
d
, the rea
s
on
s ar
e as
follows
:
(1) The la
rg
e numb
e
r of
SCADA data.
For
in
stan
ce
, the daily records
of a si
ngle unit
are ove
r
1
0
MB whe
n
re
cording
every 10s. Fo
r a
large
win
d
farm that
con
s
ist
s
of hu
nd
red
s
units, the hug
e quantity of data will impo
se hig
her
sta
ndard on the
SCADA sy
ste
m
’s efficien
cy
.
(2) The
feat
ure
s
of
wi
nd turbine
s
’ ope
rating.
In
wi
n
d
-po
w
e
r
gen
erating,
the
source
of
energy is the
natural win
d
,
random a
nd
unpredi
ctabl
e
.
With the change
s of win
d
, nearly all the
data recorded by SCADA
will chang
e
accordingly,
such as the
rotating
speed of wheel, the
vibration accelerate
d spe
ed,
the
g
ene
rating po
wer,
the temp
erature
of ge
a
r
boxe
d
. Th
ese
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
sea
r
ch on
Early F
ault Di
agno
stic Met
hod of Win
d
Turbi
n
e
s
(Zh
a
i Yongjie
)
2332
rand
om data
bring g
r
e
a
t obsta
cle
s
to
the acq
u
isitio
n and furth
e
r pro
c
e
ssi
ng
of informatio
n.
Beside
s, me
a
s
uri
ng
errors
of tran
sdu
c
e
r
s a
nd oth
e
r
e
quipme
n
t ma
ke it eve
n
h
a
rder to
corre
c
tly
analyze the d
a
ta.
(3) Effective
theorie
s a
n
d
method
s a
r
e
ne
ede
d to se
parate a
nd extra
c
t the clo
s
e
relevan
c
e
am
ong
paramet
ers.
Since th
e chan
ge
of
a si
ngle
pa
ra
meter i
s
ra
nd
om a
nd i
r
reg
u
lar,
it is im
po
ssibl
e to p
r
ovide
e
noug
h o
perating info
rmatio
n only
by ob
serving
ea
ch
p
a
ram
e
ter i
n
a
n
isolate
d
way.
(4) Wi
nd turb
ines’
cha
r
a
c
teristi
cs
are di
fferent from o
ne to anothe
r. Even locate
d in the
same
win
d
farm, two turbi
nes of the
sa
me type
may have totally different feat
ure
s
, for they
are
fixed and in
stalled in diffe
rent po
sition
s. Take
the vibration sign
als
in
the sa
me
tran
smitting
chai
n-a
unit’s wide
-rang
e vibration
may
be a
c
cept
abl
e acco
rdin
g t
o
ope
rating
e
x
perien
c
e,
while
a little vibration i
s
li
kely to
ca
use a
bno
rmalities to
a
n
o
ther unit
of the
sam
e
type
. So it i
s
difficult
to con
c
lu
de
a commo
n
rule
or
eq
uation to
an
alyze the
SCADA d
a
ta, whi
c
h m
a
kes it
c
o
ns
ide
r
ab
le
w
o
rk
.
In fact, the
wi
nd turbine
s
’
o
peratin
g
state
and
their dyn
a
mic
ch
aract
e
risti
c
s a
r
e
shown in
the massive
SCADA info
rmation. Thi
s
pape
r extra
c
t
s
the fault
co
de an
d the rel
a
ted re
co
rd
s f
r
o
m
the processe
d SCA
D
A dat
a. By
studyin
g
the releva
n
c
e amon
g
S
C
ADA
data, i
t
then
con
s
tructs
the inherent n
on-lin
ea
r mod
e
l with multipl
e
variable
s
u
nder n
o
rm
al operating stat
e.
Whe
n
a
n
ab
norm
a
lity occurs in
the
uni
t, the inn
e
r relevan
c
e
am
ong
multiple
variable
s
will be broken. The non-linear m
u
lti-variable
stat
e estimated value
will
deviate from
the
measured val
ue, whi
c
h will
incre
a
se the
resid
ual
valu
e. In order to
monitor the
unit’s state,
we
must dete
c
t e
v
en the slight
est abn
orm
a
li
ties or
cha
n
g
e
s p
r
omptly. Figure 1 sho
w
s the
co
ncrete
flow.
Figure 1. The
fault diagno
stic method ba
sed o
n
the techni
que of st
atistical movi
ng win
d
o
w
4. Basic Prin
ciple of Early
A
l
ert Meth
od
4.1.
MSET Model Cons
truc
tio
n
The MSET is
a multi-vari
ab
le state e
s
tim
a
ting tech
niq
ue first p
r
op
o
s
ed
by Singe
r [13]. It
is no
w wi
del
y used in th
e nucl
e
a
r
po
wer
plant
se
nso
r
calibration, elect
r
ic
p
r
odu
ct life-sp
an
predi
ction a
n
d
softwa
r
e a
g
ing re
se
arch [14-16]. Th
e prin
ciple of
MSET is as follows. It first
studie
s
the hi
story data of
norm
a
l
wo
rki
ng state; then
it defines the
relation
s am
ong pa
ram
e
ters;
after that, an
inherent nonli
near mo
del
with multiple
correl
ated variable
s
is con
s
tru
c
ted. Thi
s
is
how the
state
estimation works.
A certain phy
sical pro
c
e
s
s or ope
rating
data
of a device ca
n be re
p
r
esented by a
matrix.
This p
r
o
c
e
s
s
or devi
c
e
con
s
ist
s
of n vari
able
s
and m states (m
mo
ments).
The colum
n
vecto
r
is
the ope
rating
data of all t
he rel
a
ted v
a
riabl
es
at a
fixed mome
nt and the
row ve
ctor
sh
ows
certai
n varia
b
l
e’s value
wh
en the pro
c
e
s
s or devi
c
e
is at State m. Let us su
ppo
se, the n relat
e
d
variable
s
ob
served at Time
i are refe
rre
d
to as the observation vect
or:
T
n
i
x
i
x
i
x
i
X
)]
(
,
),
(
),
(
[
)
(
2
1
(1)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2330 – 234
1
2333
Und
e
r no
rmal
operatin
g state of the proce
ss
o
r
device, m historica
l
obse
r
vation vectors
are colle
cted
to con
s
tru
c
t the memo
ry matrix D, denot
ed as:
11
1
22
2
[(
1
)
(
2
)
(
)
]
(1
)
(
2
)
(
)
(1
)
(
2
)
(
)
(1
)
(
2
)
(
)
nn
n
nm
m
xx
x
m
xx
x
m
xx
x
m
DX
X
X
(2)
Each
ob
serv
ation vecto
r
i
n
the me
mory matr
ix rep
r
ese
n
ts a
no
rmal op
eratin
g state
of
the pro
c
e
s
s o
r
device. By selectin
g m hi
stori
c
al
o
b
servation vectors from
an ext
ende
d pe
riod
of
norm
a
l state prop
erly,
the
sub
s
et sp
ace
sp
ann
ed
by matrix
D
ca
n be
taken t
o
re
present t
h
e
whol
e dynam
ic working
co
ndition of the
pro
c
e
ss
or
device. Th
us,
the co
nstruction of memo
ry
matrix D is
substa
ntially a pro
c
ed
ure of
learni
n
g
an
d
memori
zin
g
the norm
a
l b
ehaviors of the
p
r
oc
es
s
or
d
e
vic
e
.
Duri
ng su
bse
quen
ce ope
ration,
the
in
put
to
the M
SET at each
time step
is a ne
w
observation v
e
ctor Xo
bs a
nd the output
from the M
SET is a pre
d
i
c
tion Xest for this input vector
for the same
moment in time. For ea
ch
input vector Xobs, MSET will pro
d
u
c
e an m-dim
e
n
s
i
onal
weig
ht vector W
T
12
[]
m
ww
w
W
(3)
With
est
1
2
(1
)
(
2
)
(
)
m
ww
w
m
XD
W
X
X
X
(4)
Equation
(4
)
mean
s th
at th
e e
s
timate
of
MSET
is a lin
ear combin
ation of the m histori
c
al
observation v
e
ctors in the
memory matri
x
D. Then
the weight vecto
r
is cal
c
ul
ated
and optimize
d
.
The re
sid
ual
betwe
en MS
ET estimate
and the inp
u
t is
obs
e
s
t
ε
XX
(5)
The wei
ght vector [1
7, 18] is co
nst
r
ucte
d as follo
ws:
)
(
)
(
obs
1
X
D
D
D
W
T
T
)
1
0
(
(6)
is a
nonli
near op
erato
r
u
s
ed to
re
place t
he
re
gular multiply
ing op
erator
in matrix
multiplication.
There a
r
e
m
any optio
nal
nonlin
ear op
erato
r
s to
ch
oose fro
m
[1
9], with the
Euclide
an
Norm (DIST),
the City Block Di
stan
ce
(CITY)
and th
e
Linea
r Correl
ation Coeffici
ent (L
CC) b
e
i
n
g
the foremo
st. In this pape
r, the nonline
a
r operato
r
is
chosen a
s
the Euclide
an di
stance b
e
twe
e
n
the two vecto
r
s
2
1
(,
)
(
)
n
ii
i
x
y
XY
(7)
Whe
n
two o
b
se
rvation v
e
ctors are th
e same
or similar, the
di
stan
ce
between th
e
vectors will b
e
zero or ne
a
r
ze
ro. Whe
n
one vecto
r
is very different
from the other, the distan
ce
betwe
en the
m
will be gre
a
t and the re
sult of the
no
nlinea
r ope
ra
tor will be la
rge. The wei
g
h
t
vector i
n
(6)
reflect
s
the
si
milaritie
s
bet
wee
n
the MS
ET input ve
ctor Xob
s
and
the m hi
stori
c
al
observation v
e
ctors in the
memory mat
r
ix D.
With (4
) and
(6), the final estimate of the
MSET model
for the pro
c
e
ss o
r
device is
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
sea
r
ch on
Early F
ault Di
agno
stic Met
hod of Win
d
Turbi
n
e
s
(Zh
a
i Yongjie
)
2334
T1
T
e
s
t
obs
()
(
)
XD
D
D
D
X
(8)
Whe
n
the
p
r
oce
s
s o
r
devi
c
e
wo
rks no
rmally, the in
p
u
t ob
se
rvatio
n vecto
r
of
MSET is
most likely to
be locate
d in
the normal
worki
ng sp
a
c
e
that is rep
r
e
s
ented by the memory mat
r
ix
D, in that it is similar to some
histo
r
i
c
all
y
measu
r
e
d
vectors in the
memory m
a
trix. As a re
su
lt,
the estim
a
te
of MSET will
have a
high
accu
ra
c
y
. W
h
en
pr
o
b
l
ems
ar
ise
w
i
th
th
e
pr
oc
ess
or
device,
its dynamic ch
ar
acteristics
will
chang
e, and t
he ne
w o
b
servation vector
will deviate from
the normal
worki
ng
spa
c
e.
In this case, the linea
r co
mbination
of the histo
r
ical
vectors in th
e
memory m
a
tri
x
will not p
r
ov
ide an
accu
ra
te estimate
of the input a
n
d
the re
sid
ual
will in
cre
a
se i
n
magnitud
e
.
4.2.
Mov
i
ng Windo
w
Re
sidu
al Statistic
a
l Metho
d
The big
g
e
s
t advantag
e of
the Moving
Wind
ow
Re
sidu
al Statistical (MWRS)
method lie
s i
n
that it enabl
es di
strib
u
tio
n
of re
sidu
al to be sho
w
n
continuo
usly, basi
ng on
wh
ich
wheth
e
r a va
riable value i
s
normal
or n
o
t
is judge
d. Under th
e sa
m
e
accu
ra
cy level, the MWRS
method can provide the
earlie
st
sig
n
of developin
g
faults.
Through this m
e
thod, the paper
eliminate
s
un
kno
w
n fa
ctors and
ran
d
o
m
disturban
ces (su
c
h a
s
t
r
an
sdu
c
e
r
s’ measuri
ng e
r
rors)
of an ope
rati
ng win
d
turbi
ne and
prom
otes its
relia
bility as well.
By a prope
r sele
ction of
the
wind
ow’
s
wi
dth, the su
ccessive
re
sid
ual statistical
ch
a
r
acte
ri
st
ics are monitored
prom
ptly, which
improve
s
the
stability of the device a
n
d
dedu
ce
s
the
chan
ce
s that
erro
r ala
r
m
s
happe
n. Wh
en
an ab
normalit
y occurs to th
e unit, these
dynamic
mod
e
ls
can
dete
c
t even the sli
ghtest
cha
n
g
e
s
of param
eters, so a
s
to diagno
se the fa
ults at a early
stage.
If, during a
certain period
of time, the res
i
d
ual sequence of the g
ear box’s
temperature
from the MSET model is:
GT
1
2
[]
N
ε
(9)
A time windo
w with width
N is ad
opted
to calculate the moving av
erag
e or me
a
n
value
and sta
nda
rd
deviation for the N succe
ssive resid
ual
s in the wind
ow
N
i
i
i
N
X
1
1
(10
)
Then
a
s
sume
that the
re
si
dual
avera
g
e
fault threshol
d is
Y
E
, the max
i
mum of
re
sid
ual
averag
e of MSET model u
nder n
o
rm
al condition i
s
V
E
, so the fault thresh
old of gea
r box EY is:
V
Y
E
k
E
1
(11
)
In this equati
on, k1 can be
cho
s
en b
a
se
d on ope
ratin
g
experie
nce.
Whe
n
the
re
sidu
al of MS
ET model
excee
d
s
a set thre
shol
d, an
alert will
be
given to
remin
d
the op
erato
r
of the potential thre
atens to the g
ear box’
s
saf
e
operating.
5. Model Co
nstru
c
tion o
f
Gear Bo
x’s Tempera
t
ure
Based o
n
SCADA Da
ta
5.1. Selection
of
Variables
All of the operating
data
of wind turbi
nes a
r
e reco
rded i
n
the SCADA sy
ste
m
. It is
a
comp
uter-ba
s
ed syste
m
which i
s
aime
d to r
eali
z
e the autom
atic sch
eduli
ng
and pla
nnin
g
of
working process. By supervisin
g and
controlling the devices’
st
at
e, the S
C
ADA achieves t
he
function
s of data acqui
sition,
device
controllin
g, pa
ramete
r me
a
s
uri
ng an
d regulatin
g, an
d
informatio
n al
erting.
In this pa
per,
the wind tu
rbine i
s
man
u
factured by V
e
sta
s
an
d its
con
c
rete pa
rameters
are: rated p
o
w
er
0.9M
W, cut-in wi
nd
sp
eed 3m
ps
,
cu
t-out win
d
sp
eed 2
5
mp
s, rated wi
nd
sp
eed
15 mp
s, ove
r-voltag
e
pro
t
ection
settin
g
value
1.2p
.u, low –voltage p
r
ote
c
tio
n
setting val
u
e
0.85p.u.. The
SCADA re
cord
s 12
6 ope
rating p
a
ra
m
e
ters
and
sta
t
e informatio
n every 10 m
i
n:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2330 – 234
1
2335
the forme
r
inclu
d
e
s
time tag, active powe
r
, rea
c
tive power,
bearin
g te
mperature a
nd oil
temperature i
n
gear boxe
s
, cabin tem
peratu
r
e, ex
ternal temp
erature, fault code
s, hydrau
lic-
pre
s
sure oil tempe
r
ature, three
stator v
o
ltage
an
d current, rotate
d spe
ed of g
enerator a
nd
so
on; whil
e the
latter con
s
ist
s
of sta
r
t and
halt of
unit, o
v
erheatin
g of
gene
rato
r, pitch
system
fa
ult,
gene
rato
r fau
l
t, frequen
cy-conve
r
tor fa
ul
t, hydrauli
c
-p
ressure
syste
m
fault, gea
r
box fault and
so
on.
Followi
ng a
review
re
co
rded by SCA
D
A,
the parameters rela
ted to the beari
n
g
temperature
of a gear box
are cho
s
en to
con
s
tru
c
t its observation v
e
ctor Xo
bs of
MSET.
(1) Active p
o
w
er (P
): P is clo
s
ely
relat
ed to the
be
aring
tempe
r
ature
of a g
ear
box.
Whe
n
P in
creases, the l
o
ad of a g
e
a
r
box will a
ggrandi
ze
s which lead
s to a
n
increa
se in
the
gear b
o
x. P is influenced b
y
wind sp
eed,
rotated spee
d of gear box,
yaw angle.
(2)
Wind spe
ed (u): Th
e variabl
e sp
eed
turbine
s
are
studie
d
in this pape
r, whi
c
h pursue
the be
st
usa
ge of
wi
nd
p
o
we
r
by a
c
hi
eving the
o
p
timal tip
sp
ee
d ratio. Th
e
highe
r th
e
wi
nd
spe
ed is, the
faster the g
e
a
r
box rotate
s, and the hig
h
e
r
its tempe
r
at
ure will b
e
.
(3)
Rotate
d
spe
ed of
ge
ar b
o
x (U):
The g
ear bo
x has th
e fu
nction to
a
c
celerate
or
decelerate th
e spe
ed. As
U is
clo
s
ely related to
turb
ines’ P, a hig
her
U is al
wa
ys accom
pan
ied
with a bigg
er
P.
(4) Ya
w angl
e (A): The direction
s
of nat
ural
wi
nd
s are cha
nge
able
and unp
redi
ctable. In
orde
r to
enh
ance wi
nd
p
o
we
r’s efficie
n
cy, a fun
c
ti
on a
s
a
si
g
n
ificant p
a
ra
meter to
adj
ust
turbine
s
’ dire
ction to mee
t
the wind, this ha
s a profound influe
nce o
n
syste
m
’s safety a
nd
effic
i
enc
y
.
(5) Bea
r
ing te
mperature of
gear b
o
x (Tg
ear
): Operating unde
r sev
e
re working
condition
and he
avy loads fo
r lon
g
hours, the
beari
n
g
s
of
gear
boxe
s
are li
kely to suffer fault
s
and
damag
es. Th
e freque
nt da
mage
s ari
s
e
mainly from
noises, temp
eratu
r
e, vibra
t
ions, lubri
c
at
ion
probl
em
s and
other bad
sta
t
es.
(6) Oil temp
e
r
ature of gear boxes (Toil
)
:
With
a temperatu
r
e se
nsor in the gea
r box, th
e
Toil mu
st b
e
highe
r tha
n
0
(it va
rie
s
a
c
cordi
ng to
th
e re
qui
reme
n
t
s of lu
bri
c
ati
on oil
)
, an
d t
hen
heated to ov
er 10
to op
erate. In no
rmal wo
rking
states, the
oil pump
contin
uou
sly eject
s
oil
into gears an
d beari
n
g
s
. Whe
n
Toil is highe
r than 6
0
°C
, the oil cooling sy
ste
m
starts to function
and the
heate
d
oil is t
r
an
smitted to an e
x
ternal e
xcha
nger to be
co
oled by n
a
tural win
d
or wa
ter.
Whe
n
Toil i
s
belo
w
45
, the oil
cooli
n
g
loop i
s
cut, and the
cooli
ng p
r
o
c
e
ss
stops. Th
e ov
er-
heated Toil i
s
always
cau
s
ed by the lo
n
g
hours of full-load
ed op
erating.
(7)
Cabin te
mperature (T
cabi
n): Tcabi
n is al
so a factor that infl
uen
ce
s the gear box
temperature.
Whe
n
Tcabin
is t
oo low, the mech
ani
cal
compo
nent
s can h
a
rdly op
erate p
r
op
erl
y
;
while too hi
gh
a Tcabin
will sho
r
ten the el
ectri
c
compo
nents’ life spa
n
.
(8) Extern
al
temperature
(Tc): Becau
s
e t
he local tempe
r
ature that the win
d
turbine
experie
nces
cha
nge
s grea
tly in t
he short term (from d
a
y to night for example
)
a
nd in the long
er
term (we
e
ks t
o
month
)
d
u
e
to pa
ssi
ng
weather sy
ste
m
s a
nd
se
asons it mu
st
b
e
taken expli
c
itly
into accou
n
t. Thus, differen
t
Tc will pr
o
d
u
ce different gear b
o
x temperatu
r
e.
5.2.
SCA
DA
Da
ta
Recor
d
Ana
l
y
s
is
As wi
nd tu
rbi
nes
are g
r
ea
tly influence
d
by ex
ternal
factors (temp
e
ratu
re,
wi
nd
spe
ed
etc.), this
pa
per
pro
b
e
s
in
to the SCA
D
A data
of
Ja
n.2011 of
a certai
n wind turbine.
Fi
gu
re
2
sho
w
s the 72
1 10-mi
nute d
a
ta from 17:2
0
:
00 22n
d Ja
n. 2011 to 17:
20:00 27th
Ja
n. 2011.
Table
s
1-6
li
st the
SCA
D
A ope
rating
reco
rd
s
of whi
c
h th
e
power is le
ss than
0 du
ring
this pe
riod. In
these
cha
r
ts,
every fault code is
co
rrespondi
ng to a fault rea
s
o
n
. The 0 fault co
de
rep
r
e
s
ent
s n
o
fault. As to the state of
the wind tu
rbine, 0 refers to sh
ut-d
o
w
n an
d 1 m
ean
s
norm
a
l ope
ra
ting.
In Table 1, the high
-wi
n
d
fault code is
alerted
whe
n
the wind
speed i
s
23.7
mps a
nd
19.7 mp
s. Bu
t unde
r neith
er
con
d
ition
doe
s the
win
d
sp
eed
re
ach
the cut-in
wind sp
eed. The
rea
s
on
lie
s i
n
that the
wi
n
d
spee
d i
s
2
4
.1
mp
s at
5:
20:00 23
rd Jan.
20
11a
nd rea
c
he
s
25
mps
durin
g 5:2
0
:0
to 5:30:00,
whi
c
h m
a
kes the turbine
cut out. Becau
s
e of th
e
self
prote
c
tion
a
n
d
system d
e
lay, a 144 fault code is al
erte
d
at 5:30:00 an
d 5:40:00 23
rd Jan. 20
11.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
sea
r
ch on
Early F
ault Di
agno
stic Met
hod of Win
d
Turbi
n
e
s
(Zh
a
i Yongjie
)
2336
Figure 2. Hist
orical
cu
rve from 17:20:00
22
th
Jan to 1
7
:20:00 27
th
Jan
Table 1. No
d
e
spe
c
ific offli
ne paramete
r
SCADA data
reco
rd
s 1
Num Data
Time
Wind
speed
Active
pow
er
Fault code
Fault reason
1 2011/01/23
5:20:00
24.1
0
2 2011/01/23
5:30:00
23.7
11.8
144
Over w
i
nd
speed
3 2011/01/23
5:40:00
19.7
148.7
144
Over w
i
nd
speed
Table 2. SCA
D
A data re
co
rds 2
Num
Data
Time
Wind speed
Active pow
er
Fault code
1 2011/01/24
9:40:00
5.9
-19.7
0
2 2011/01/24
9:50:00
7.9
-22
0
Table 3.
SCA
D
A data re
co
rds 3
Num
Data
Time
Wind speed
Active pow
er
Fault code
1 2011/01/24
11:40:00
5.4
-7.6
0
2 2011/01/24
11:50:00
5.1
-16
0
3 2011/01/24
12:00:00
4.5
-15.1
0
Table 4.
SCA
D
A data re
co
rds 4
Num
Data
Time
Wind speed
Active pow
er
Fault code
1 2011/01/24
17:20:00
4.9
-20.5
0
2 2011/01/24
17:30:00
3.6
-22
0
3 2011/01/24
17:40:00
3.9
-21.2
0
4 2011/01/24
17:50:00
6.9
-21.4
0
5 2011/01/24
18:00:00
6
-3.6
0
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2330 – 234
1
2337
From T
able
s
2-4, we can see that faults
cod
e
s a
r
e
0. It infers that the thre
e shut
-do
w
n
s
are
cau
s
e
d
b
y
manned
re
aso
n
s in
stea
d of faults
an
d that is the
so-call
ed ma
nned
shut
-do
w
n.
Gene
rally sp
eaki
ng, the re
aso
n
of a manned
shut-d
o
w
n is the po
wer gri
d
impo
ses a limit on the
gene
rating
a
m
ount
of wi
n
d
farm
s,
whi
c
h sto
p
s the
n
o
rmal
op
erati
ng of th
e tu
rb
ines. In
Ta
ble
4,
both the wind
spee
ds at 1
7
:30:00 an
d 17:40:00 2
4
th Jan. 20
11
are lo
we
r than the cut-in
speed
and the turbi
n
es are not ge
neratin
g.
In Table 5, the first two sh
ut-do
w
n
s
are
caused by manne
d ope
rat
i
ons an
d the rest a
r
e
becau
se that the automatic
yawing make
s the active p
o
we
r less tha
n
0. Chart 6’s fault code 14
4
is b
e
ca
use
that the wi
n
d
s
peed
at
7:30:00
27t
h Jan. 20
11
rea
c
h
e
s 24
.1 mps,
whi
c
h
approximate
s
the cut-o
u
t speed
and
trig
gers the
f
ault
co
de at th
e
next time du
e to the
syst
em
delay.
Table 6 sho
w
s
that
fou
r
shut-do
w
n
s
o
c
cur f
r
om
17:
20:00
22nd
Jan. 20
11 to
17:20:0
0
27th 201
1. All of them are
manne
d shut
-do
w
n
s
.
In the mean time,
there a
r
e n
o
gear
box fau
l
ts
and re
pai
ring
records in
Ja
n. and after Jan.
Table 5.
SCA
D
A data re
co
rds 5
Num Data
Time
Wind
speed
Active
pow
er
Fault code
Fault reason
1 2011/01/25
17:50:00
5.4
-19.4
0
2 2011/01/25
18:10:00
4.4
-16.5
0
3 2011/01/25
18:20:00
5.9
-22
275
Automation
y
a
w
4 2011/01/25
18:30:00
9.3
-21.4
275
Automation
y
a
w
5 2011/01/25
18:40:00
8.6
-21.2
275
Automation
y
a
w
6 2011/01/25
18:50:00
8.9
-3.4
275
Automation
y
a
w
Table 6. SCA
D
A data re
co
rds 6
Num Data
Time
Wind
speed
Active
pow
er
Fault code
Fault reason
1 2011/01/27
7:40:00
22.7
-21.1
144
Over
w
i
nd
speed
2 2011/01/27
7:50:00
21.7
-20.9
144
Over
w
i
nd
speed
3 2011/01/27
8:00:00
22.5
-21.1
144
Over
w
i
nd
speed
4 2011/01/27
8:10:00
20.3
-3
144
Over
w
i
nd
speed
To build the
MSET model
, operating d
a
ta of normal
state are sel
e
cted a
nd a
pro
c
e
ss
matrix
is co
n
s
tru
c
ted. We aban
don
th
e data
of whi
c
h
the po
we
r i
s
less than
0, a
s
sume th
e d
a
t
a
of which the wind
spee
d is lower than
3 mps is
3
m
ps an
d the data of which the win
d
spe
e
d
is
highe
r than 2
5
mps is 25
mps. Othe
r data are
all from the norm
a
l operating informatio
n and
referred to
as the pro
c
e
s
s
memory m
a
trix D. The
ulti
mate process memory m
a
trix D con
s
ist
s
of
685 o
b
servat
ion vecto
r
s. After co
nstru
c
ting the
D, we can
predi
ct
the
n
e
w i
n
put ob
se
rvation
vector of MS
ET temperatu
r
e mod
e
l usi
n
g Equation 8.
5.3.
Cons
tru
c
tio
n
And Ch
eck
ing of The M
odel
669
histo
r
i
c
al
ope
ratin
g
d
a
t
a from
17:2
0
:
00 22
nd
Ja
n
.
2011
to
17:
20:00
27th
Jan. 20
11
confirm the
correctn
ess of
the MSET
model.
Du
ri
n
g
this
pe
riod,
the maximu
m and
minim
u
m
values of ge
a
r
box bea
ring
s’ tempe
r
ature are 70
℃
an
d 50
℃
res
pec
tively.
In this
pap
er,
the first 50
0
data
are
cho
s
en
to
con
s
truct the
matri
x
D, an
d the
beari
ng
temperature
colum
n
of th
e re
st 19
9 d
a
ta is
ch
ose
n
to be th
e i
nput Xob
s
. T
he first pictu
r
e of
Figure 3 is t
he ob
se
rved
value and
e
s
timate curve
throug
h sim
u
lation test,
and the
se
co
nd
picture sh
ows the resid
ual
s betwee
n
ob
served value a
nd estimate v
a
lue.
Usi
ng the mo
ving windo
w
statistical method to
furthe
r analyze the
resi
dual
s ab
o
v
e, we
con
c
lu
de its cha
r
a
c
teri
stics
cu
rve as Figure
4. We a
s
sume
the wind
ow width N=2
0
and
cal
c
ulate th
e
thre
shol
d E
v
’s value. T
he maxi
mu
m
absolute
re
sidu
al value
is 0.49
3 (k1
=
3,
479
.
1
Y
E
).
5.4.
Simulation Test o
f
Early
Diagno
stic Alert
After the
co
nstru
c
tion
a
nd
corre
c
tne
s
s che
cki
ng
of MSET
model,
we
add th
e
temperature
offset to imitate the situ
a
t
ion
wh
en a
gear
box fa
ult lead
s to
its tempe
r
atu
r
e
increa
se. Starting from Po
int 51, a step temper
atu
r
e offset of 0.25 is add
ed to the 199 data.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
sea
r
ch on
Early F
ault Di
agno
stic Met
hod of Win
d
Turbi
n
e
s
(Zh
a
i Yongjie
)
2338
Figure 5
is th
e e
s
timate
si
mulation
re
su
lt of M
SET m
odel
with te
m
peratu
r
e
offsets a
dde
d in.
The
first pi
cture i
s
the
co
mpa
r
iso
n
of
ob
served va
l
ue
and e
s
timate
value a
nd t
he
se
cond
is its
resi
dual
cu
rve. From the
seco
nd pi
ct
ure
’
s cu
rve, we
can se
e that t
he errors at th
e first 50 p
o
in
ts
are very sm
al
l, while it increases g
r
ad
ua
lly from
the 51st point, and
the deviation
s are m
a
inly the
temperature offsets.
Figure 3. Curves of observ
ed value an
d estimate valu
e and re
sid
u
a
l
Figure 4. Re
sidual’
s
movin
g
wind
ow
statistical
cha
r
a
c
teristi
c
s
Figure 5. Estimate re
sults
with tempe
r
at
ure ad
ded
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2330 – 234
1
2339
Then the
re
sidual
s after t
e
mpe
r
ature o
ffsets a
r
e ta
ken into a
c
co
unt and
a si
mulation
test is cond
ucted ba
se
d o
n
the
mo
vin
g
w
i
n
d
o
w
s
t
a
t
is
tica
l me
th
od
. We
as
su
me
the w
i
nd
ow
w
i
d
t
h
N=10, and its characte
ri
stics
cu
rve is sh
own in Fig
u
re
6.
Figure 6. Re
sidual moving
wind
ow
statistical ch
ara
c
te
ristics curve
a
fte
r temperature offset
In Figure 6,
after ad
ded
with ma
nne
d o
ffset
s an
d processe
d
by moving
wind
ow
statistical m
e
thod, the
resi
d
ual
cu
rve of
b
earin
g tem
p
e
r
ature in
crea
ses
co
nsta
ntly, and
ove
r
tops
its up thresho
l
d at Point 45
, which trigg
e
r
s the
ea
rly fault alert. Th
e dista
n
ce be
tween thi
s
p
o
i
nt
and the first point at whi
c
h mann
ed offset is a
dded
(Point 51
) is
45+20-51
=14
(
20 refers to
the
wind
ow
width
)
. Thu
s
, at th
e 65th
(51
+
1
4
=6
5)
point,
we
can
dete
c
t the ab
no
rmal a
c
ts of g
ear
box’s b
eari
n
g
temperature. As to Point 6
5
, we
can
also cal
c
ul
ate th
e deviation
b
e
twee
n ori
g
in
al
state and
manne
d-offse
tting state according to
temperatu
r
e offset ste
p
s an
d bea
ring
temperature,
that is 14*0.2
5
=3.5
.
Figure 7. Similarity curve
with temperatu
r
e offset
Figure 7
sho
w
s the
bea
ri
ng temp
eratu
r
e’s si
m
ilarity
cu
rve afte
r
manne
d offsets h
a
ve
been
ad
ded
i
n
. We
can tel
l
that the
simi
larity at
Point
65 i
s
0.96, a
nd the
si
milarity at Point 1
4
7
is 0.76. A small simila
rity stands fo
r an abno
rm
al act of turbine
s
’ ope
rat
i
on. In Figure 8
,
con
s
id
erin
g the mann
ed
offsets, the whol
e tempe
r
ature cu
rve are divide
d into three pa
rts----
norm
a
l ope
ra
ting state, de
vice ea
rly wa
rning
st
ate a
nd ale
r
ting st
ate. Whe
n
the gea
r box i
s
workin
g no
rm
ally and its te
mperature
re
sidu
al
doe
s n
o
t exceed th
e
mean thresh
old, the turbi
n
e
is in the n
o
rmal ope
rating
state. Wh
en
the re
sidu
al
e
x
ceed
s the th
reshold, it will
be in the e
a
rly
warning state
.
When device’s
tem
peratu
r
e ove
r
top
s
th
e maximum v
a
lue
set by its ma
nufa
c
turer,
an alert will occur (The normal operating te
mperature
of this type of wind turbine:
91
gear
T
).
In Figure 7,
at Point 147, the beari
ng tempe
r
at
ure rea
c
he
s 92.25
, wh
ich exceed
s the
maximum op
erating val
u
e
and tri
gge
rs
the syste
m
al
ert. If the turbine move
s
o
n
witho
u
t taki
ng
necessa
ry m
easure
s
, the
gear
box will be
da
mage
d and
the unit will
not
fun
c
ti
on
p
r
ope
rly
a
n
d
regul
arly.
Evaluation Warning : The document was created with Spire.PDF for Python.