TELKOM
NIKA
, Vol. 11, No. 7, July 201
3, pp. 3552 ~ 3560
e-ISSN: 2087
-278X
3552
Re
cei
v
ed Fe
brua
ry 12, 20
13; Re
vised
Ma
rch 15, 20
13; Accepted
March 30, 20
13
Application of Support Vector Machine to Reliability
Analysis of Engine Systems
Zhang Xinfe
ng*
1
, ZhaoYa
n
2
1
Ke
y
L
abor
ator
y of Automotiv
e
T
r
ansportatio
n
Safe
t
y
En
ha
nceme
n
t T
e
ch
nol
og
y of the
Ministr
y
of
Commun
i
cati
o
n
, Chan
g’
an U
n
iversit
y
, Xi’
an,
7100
64, Ch
ina
2
School of Eco
nomics & Man
agem
ent, Xi
di
a
n
Univ
ersit
y
,
Xi
’an, 71
00
71, C
h
in
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: zhxfp
e
k@
ya
hoo.com.cn
1
, zhao-
ya
n1
25
@
s
ohu.com
2
A
b
st
r
a
ct
Reli
ab
ility an
al
ysis plays a v
e
ry importa
nt role
i
n
assessi
ng pro
duct per
forma
n
ces a
n
d
maki
n
g
ma
inte
nanc
e
p
l
ans f
o
r
mai
n
tain
abl
e pr
od
uc
tion. T
o
fin
d
e
ffective w
a
ys of forecasti
ng
the rel
i
a
b
il
ity i
n
engine systems, the paper
presents a comparativ
e st
udy
on the pr
ediction perf
ormanc
es using support
vector mac
h
i
n
e (SVM), least squar
e su
ppo
rt vector machi
ne (LSSVM) a
nd ne
ural n
e
tw
ork. The relia
bi
lity
indexes
of engine system
s
ar
e com
p
uted
using the We
i
b
ull probability
paper pr
ogramm
e
d w
i
th Matl
ab.
Illustrative ex
a
m
p
l
es sh
ow
that pr
oba
bil
i
ty d
i
stributi
ons of forecasti
ng o
u
tcomes usi
ng d
i
fferent metho
d
s
are co
nsistent
to the actu
al pr
oba
bil
i
ty distrib
u
ti
on. An
d the
tw
o metho
d
s o
f
SVM and LS
SVM can pr
ovi
d
e
the acc
u
rate
pred
ictions
of
the ch
aracteri
stic lif
e, s
o
S
V
M an
d L
S
S
V
M are
bot
h
effective pr
ed
i
c
tion
m
e
thods
for r
e
liability
anal
ysis
in engine system
s. M
o
reover
, the
predictive pr
ec
ision bas
ed on
LSSVM
is
hig
her than th
a
t
based on SV
M, especia
lly i
n
sma
ll sa
mp
le
s. Because of
i
t
s low
e
r comp
utation costs a
n
d
higher prec
ision, the re
liability prediction using
LSSVM is more popular.
Keyw
ords: rel
i
abil
i
ty a
nalys
is
, supp
ort vect
or
mac
h
in
es, l
east sq
uar
e s
upp
ort vector
mac
h
i
ne, n
eur
al
netw
o
rk, learni
ng metho
d
s
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
The ability
of
accurately predi
cting reli
abilit
y for engi
ne
system
si
s an i
n
valuabl
e
asse
t
for many ma
nufactu
ring
compani
es. E
s
pe
cially fo
r
automobil
e
p
r
odu
ction, th
e main co
nce
r
n is
to
satisfy
the increa
sing de
mand
s
fro
m
cu
stome
r
s
an
d co
nform to
stri
cter a
c
ts a
nd regulatio
n
s
by gove
r
nme
n
ts [1]. As
system reliabil
i
ty indexes
vary
with time
, it is
not e
a
s
y to
pre
d
ict
it
acc
u
rately.
There are m
any forecasti
ng tech
nique
s abo
ut
time serie
s
. Tra
d
i
t
ionally, the method
s
based on sto
c
ha
stic p
r
o
c
e
ss the
o
ry ha
ve been dev
elope
d and u
s
ed
widely for asse
ssi
ng the
system reliability during the whol
e lifetime [2, 3]. Bu
t they impos
e som
e
restri
ctions on fail
ure
cla
s
ses,
so i
t
’s difficult to satisfy and
validate all the assu
mpti
ons. In p
r
a
c
tice, the way of
simplifying th
em is often u
s
ed.
Neu
r
al n
e
tworks a
r
e u
n
i
v
ersal fu
ncti
on app
roxim
a
tors th
at ca
n map a
n
y non-li
nea
r
function
with
out a prio
ri a
s
sumption
s a
bout the
pro
pertie
s
of the data [4]. The theory is m
o
re
powerful in describi
ng the
dynamics of reliability in
compari
s
on to tradition
al stat
istical model
s.
With the eme
r
gen
ce a
nd d
e
velopme
n
t of statistica
l le
arnin
g
theory
,
the resea
r
ch of applicati
on
of neural net
works in reli
a
b
ility enginee
ring ha
s b
e
e
n
made g
r
ad
ually. Zheng
[5] illustrated
a
non-pa
ramet
r
ic software re
liability
predi
ction system b
a
se
d on ne
ural
netwo
rk en
sembl
e
s. And
it
improved
the system predi
c
tab
ility by combing m
u
ltipl
e
neural
networks. Lolas
and Olatunbosun
[6] demon
strated h
o
w
a t
ool li
ke n
e
u
r
al net
works
can
be
de
sig
ned
and
opti
m
ized
for u
s
e in
reliability perf
ormance prediction
s. Xu
et al. [1] applied
Radi
al
Basis Functi
on (RBF) neural
netwo
rks
to
forecast engi
ne system reliability.
The
com
p
a
r
ative
study
amon
g
feed
-forwa
rd
multilayer pe
rce
p
tro
n
(ML
P
) and RBF
wa
s pre
s
e
n
te
d. The re
sult
s sh
owed th
at the model
is
more a
c
curate than those.
Another l
e
a
r
n
i
ng meth
od i
s
sup
p
o
r
t vect
or
ma
chi
ne
(SVM) [7-1
0] prop
osed by
Vapnik,
whi
c
h is ba
sed on the st
ructu
r
e
d
risk minimiza
tion
(SRM) p
r
in
ciple and stat
istical lea
r
ni
n
g
theory. SVM has bette
r gene
rali
zatio
n
perfo
rma
n
ce than othe
r neural network m
odel
s. The
solutio
n
of SVM is uniqu
e, optimal and
absent fr
om l
o
cal mini
ma, unlike other n
e
tworks’ traini
ng
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Applicatio
n of Support Vect
or Ma
chin
e to Re
liabilit
y Analysis of Engi
ne … (Zh
ang
Xinfeng)
3553
whi
c
h re
quires no
n-lin
ear optimization
thus ru
n
n
in
g the dang
e
r
of getting stuck in a lo
cal
minima. Ho
wever, the co
mputation co
st of SVM is
very high. Suyken
s et al. [11] introdu
ce
d the
modified l
e
a
s
t squ
a
re
s l
o
ss fun
c
tion i
n
to SVM, whi
c
h is
kn
own
as th
e lea
s
t
squ
a
re
supp
ort
mac
h
ine (LSSVM). Unlike SVM, LSSV
M turns
inequa
lity c
o
ns
traints
into equality c
ons
traints
,
whi
c
h makes computation efficiency hi
gher. At
the same time, LSSVM consi
ders the training
errors
of all th
e trai
ning
sa
mples.
So fa
r SVMs
we
re
su
ccessfully
applie
d in
ma
ny fields,
such a
s
pattern
re
cog
n
ition proble
m
s, fun
c
tion
estimation,
ti
me se
rie
s
foreca
sting, di
sease dete
c
tin
g
in
medici
ne [12, 13]. But the research
on the
appli
c
ation
of SVM and
LSSVM in reli
ability
predi
ction i
s
very limited [14, 15].
This paper ai
ms at validating the effect
iveness of SVM and LSSVM for the reliability
predi
ction of
engine
syst
ems. T
he
co
mparative study of the pr
edi
cted resu
lts of the en
gine
s
y
s
t
ems
by S
V
M, LSSVM, MLP, and
RBF is made. B
a
s
e
d on the pr
edic
t
ion,
the affec
t
ion on
the
life characte
ri
stics of t
he engine
system
is analyzed.
2.
SVM and LSSVM
2.1 SVM
Let data
set
,whe
re
,
n
ii
x
Ry
R
,
1,
,
il
. For reg
r
e
ssi
on a
nalysi
s
in
SVM, there a
r
e two ba
si
c
obje
c
tives. T
he first is to
find real a
p
p
r
oximating fun
c
tion
, which
make
s the
structu
r
e ri
sk o
f
the function
esti
mation le
st in the inse
nsitive loss functio
n
. The
s
e
c
o
nd is
to mak
e
func
tion
flat.
2.1.1
Liner Suppo
rt Vector Ma
chine
Whe
n
the
rel
a
iton of th
e ip
ut and
the
ou
tput
data i
s
li
near, th
e
reg
r
ession
fun
c
tio
n
can
be sh
own in Eq. (1).
(,
)
f
xw
w
x
b
(1)
To mak
e
the func
tion flat, the paramet
e
r
w
in Eq.(1
)
sho
u
ld
be at
most l
e
ss.
So the
solutio
n
is tra
n
sformed into
the following
optimazation
probl
em.
2
1
mi
n
|
|
|
|
2
..
|
(
)
|
,
1
,
2
,
i
w
s
ty
w
x
b
i
l
(2)
Con
s
id
erin
g the po
ssible
e
rro
rs an
d introdu
cing t
w
o
sla
ck va
riabl
es
,and pe
na
lty
para
m
eter
,
the above o
p
timazatio
n
obj
ective functio
n
can b
e
writt
en as
sho
w
n i
n
Eq.(3)[8]:
2*
1
*
*
1
mi
n
|
|
|
|
(
)
2
..
,0
,
1
,
2
,
l
ii
i
ii
i
ii
i
ii
wC
yw
x
b
st
w
x
b
y
il
(3)
The
solution
method i
s
co
mmonly La
grange
Multipli
er te
chniq
ue.
So the du
al
form of
the initial optimazatio
n que
stion is exp
r
e
s
sed a
s
follo
w.
**
*
*
,1
1
1
*
1
*
1
mi
n
(
)(
)
(
)
(
)
2
()
0
..
()
[
0
,
]
,
1
,
2
,
ll
l
ii
j
j
i
j
i
i
i
i
i
ij
i
i
l
ii
i
ii
xx
y
st
Ci
l
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN:
2087
-27
8
X
TELKOM
NIKA
Vol. 11, No
. 7, July 2013
: 3552 – 356
0
3554
Furthe
r, from
Eq. (4), we
can obtain
:
*
1
()
l
ii
i
i
wx
(5)
Submitting Eq. (5) into Eq
.(1), the linea
r reg
r
e
ssi
on e
quation i
s
followe
d.
*
1
()
(
)
l
ii
i
i
f
xx
x
b
(6)
2.1.2
Nonlinear S
upport Ve
ctor Machine
If any ag
rith
m can
be
expre
s
sed
by t
he d
o
t p
r
od
u
c
t, its
gen
era
lization
form
can
be
achive
d by
kernel
fun
c
tio
n
s.
Whe
n
th
e relation
of
the inp
u
t an
d
output
data
is n
onlin
ear,
the
nonlin
ear fu
n
c
tion a
pproximations c
an
be got by
repl
acin
g the d
o
t prod
uct
of inp
u
t vectors
with a
kernel fun
c
tio
n
. It can be re
pre
s
ente
d
by
,where
and
are ea
ch inp
u
t vectors.
Repl
aci
ng the
dot produ
ct o
f
input vectors in Eq.(4
)
wit
h
the ke
rnel fucntio
n
, we o
b
tain
**
*
*
,1
1
1
**
1
1
mi
n
(
)
(
)
(
)
(
)
(
)
2
.
.
(
)
0,
,
[
0,
]
,
1
,
2,
ll
l
ii
j
j
i
j
i
i
i
i
i
ij
i
i
l
ii
i
i
i
kx
x
y
st
C
i
l
(7)
At the same
method,
repl
acin
g the
dot
prod
uct
of inp
u
t vecto
r
s in
Eq.(6) with th
e kern
e
l
fucntion, the
nonlin
ear
reg
r
essio
n
equ
ation is follo
win
g
.
*
1
()
(
)
(
)
l
ii
i
i
f
xK
x
x
b
(8)
2.2 LSSVM
Accordi
ng to the LSSVM theory, the data se
t can be written as
shown in Eq.(9).
()
()
f
xx
b
(9)
Whe
r
e
den
otes the
weight
vector;
rep
r
e
s
ent
s the
non
linear fun
c
tio
n
that map
s
t
h
e
input
spa
c
e
to a
high
-dim
e
n
sio
n
featu
r
e
spa
c
e
an
d p
e
r
form
s lin
ea
r
reg
r
e
ssi
on; a
nd
b
is the
bi
as
term.
Unli
ke SVM,
LSSVM turns i
nequality const
r
aints into the equality. For
function
estimation, t
he origi
nal o
p
timization p
r
oble
m
, con
s
eque
ntly, change
s a
c
cording to the
SRM
prin
ciple. Th
e
algorithm i
s
the sol
u
tion to
conv
ex qua
d
r
atic p
r
og
ram
m
ing, as follo
wing formula.
22
1
1
mi
n
|
|
|
|
22
..
(
(
)
)
,
1
,
2
,
l
i
i
ii
i
st
y
w
x
b
i
l
(10
)
Whe
r
e
denot
es the regula
r
ization
con
s
t
a
nt,
rep
r
e
s
ent
s the dat
a e
r
rors.
The sol
u
tion
method is
co
mmonly Lag
range Mu
ltipli
er tech
niqu
e. So the Lagra
nge
polynomial i
s
sho
w
n a
s
Eq. (1
1
)
.
22
11
1
(,
,
,
)
|
|
|
|
(
)
22
ln
ii
i
i
i
ii
Lb
x
b
y
(1
1
)
Furthe
r
,
we
can obtain the
set of linear e
quation
s
(1
2).
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Applicatio
n of Support Vect
or Ma
chin
e to Re
liabilit
y Analysis of Engi
ne … (Zh
ang
Xinfeng)
3555
T
T
0
00
0
00
0
1
0
00
10
IZ
b
II
y
ZI
(12
)
whe
r
e
,
,
,
,
The sol
u
tion i
s
also expre
s
sed a
s
Eq.(1
3
)
T
01
0
1
b
y
K
I
(13)
The sol
u
tion to Eq. (13) i
s
given by Eq.(14).
1
T1
T1
(1
)
1
11
Ay
b
Ay
b
A
(14)
In Eq. (14),
AK
I
.
Submitting Eq. (14) into
1
l
ii
i
x
, in accordan
ce
with Mercer
con
d
ition, the regre
s
sion
equatio
n is fo
llowe
d.
1
()
(
)
l
ii
i
f
xK
x
x
b
(15
)
Whe
r
e
is call
ed the kernel functions. Eq
. (15) is the desired LSSVM model.
3
Reliabilit
y
Anal
y
s
is of E
ngine S
y
stems Using SVM and LSSVM
System relia
b
ility is the function varyin
g with
time. For engine
syste
m
s, the time
can b
e
time between failures
,
time to failure,
or
the to
tal failure num
bers,
which
ca
n b
e
consi
dered
as
a
colle
ction
of rand
om vari
a
b
les. So
relia
bility pr
edi
ctive can
be fin
i
she
d
by the
traditional tim
e
seri
es a
nalysi
s
method.
This paper proposes the
appli
c
ation of
SVM
and
LSSVM
described in secti
on 2to
predi
ct system reliability. And the com
parativ
e study of algorithm performance amongSV
M
,
LSSVM and neural networks i
s
made. In order to
assess the predictive
errors among the above
method
s, the norm
a
lized ro
ot mean sq
ua
re
erro
r mea
s
ure (NRMSE) is introdu
ce
d
.
2
2
ˆ
((
)
(
)
)
()
x
tx
t
NR
M
S
xt
(16
)
In Eq.(16),
denotes the p
r
e
d
iction of
.
The predi
ctio
n is cl
assified
into the long
-term a
nd sh
ort-te
rm. Due
to the accu
mulation
of erro
rs,the perfo
rman
ce
of the former
is poo
re
r tha
n
that one of the latter. For engine
syste
m
,
the short te
rm predictive
re
sults are
more effe
ct
ive, which p
r
ovide timely
inform
ation
for
preventive m
a
intena
nce and co
rre
ct
ive maintena
nce plan
s. So her
e only the single-step
-ah
e
a
d
predi
ction
s
wi
ll be con
s
id
ered.
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN:
2087
-27
8
X
TELKOM
NIKA
Vol. 11, No
. 7, July 2013
: 3552 – 356
0
3556
For the
reliability predi
cti
on of engi
ne system
s by
SVM
and LSSVM,it’s important
t
o
sele
ct
t
he
suit
ke
rn
el f
unct
i
on.As
ke
rnel functio
n
s satisfy
the Mercer
co
ndi
tion,they ena
ble
thedot pro
d
u
c
t to be com
puted in high
-dime
n
si
on space usin
g lo
w-di
men
s
ion
spa
c
e dat
a input
without th
e transfe
r fu
nctio
n
[16]. We
will
ma
ke th
e
sel
e
ction
of th
e
radial
ba
sis fu
nction
(RBF)
as th
e
kernel
function,
whi
c
h is commo
nl
y useful
in fu
nction
e
s
tima
tion. The
RB
F kern
el fun
c
t
i
on
is
r
e
pr
es
en
ted
a
s
Eq
.(
17
)
.
2
2
(,
)
e
x
p
(
)
2
ij
ij
xx
Kx
x
(17
)
In Eq.(17),
is the kernel fun
c
tion pa
ram
e
ter.
All data are
divided into
training d
a
ta and
p
r
edi
ctive data. To obtain the
optimal
para
m
eter co
mbination
in es
tablishing
the SVM and L
SSVM models
, this
res
e
arch
us
ed
the gri
d
sea
r
ch al
go
rithm
with a
k
-fol
d cro
s
s-valid
ation meth
od
[17]. Furthe
r, the SVM
and
LSSVM predi
c
tive models will be built
with t
he
opti
m
al combination. The
predictive data
are
sub
s
tituted in
to the model; the reliability predi
ction
will
be comp
uted
.
3.1
Reliabilit
y
Pr
ediction of Engine S
y
stems b
y
SVM
and LSSVM
Table
1 [1] g
i
ves the
ori
g
i
nal te
st data
of time to fail
ure fo
r 4
0
su
its of turbo
c
h
a
rge
r
s.
The first col
u
mn in Table
1 denote
s
the failure orde
r. The se
con
d
colum
n
in Table 1 de
no
tes
time to failure of the turbochargers
.
Comm
only, the estimations of reliability
like
this are achi
eved by median
ranki
ng. The
formula [18] is
as
follows
.
0.
3
()
1
0.
4
i
i
RT
n
(18
)
Table 1. Failure data and
reliability of turbocharges
F
a
ilur
e
Or
der
(
i
)
Timetofailure
(
T
/1000
h
)
(%)
F
a
ilur
e
Or
der
(
i
)
Timetofailure
(
T
/1000
h
)
(%)
1 1.6
99.303
21
6.5
79.382
2 2
98.307
22
6.7
78.386
3 2.6
97.311
23
7
77.390
4 3
96.315
24
7.1
76.394
5 3.5
95.319
25
7.3
75.398
6 3.9
94.323
26
7.3
74.402
7 4.5
93.327
27
7.3
73.406
8 4.6
92.331
28
7.7
72.410
9 4.8
91.335
29
7.7
71.414
10 5
90.339
30
7.8
70.418
11 5.1
89.343
31
7.9
69.422
12 5.3
88.347
32
8
68.426
13 5.4
87.351
33
8.1
67.430
14 5.6
86.355
34
8.3
66.434
15 5.8
85.359
35
8.4
65.438
16 6
84.363
36
8.4
64.442
17 6
83.367
37
8.5
63.446
18 6.1
82.371
38
8.7
62.450
19 6.3
81.375
39
8.8
61.454
20 6.5
80.378
40
9
60.458
Whe
r
e
i
is failure order.
Substituting the failure orders in table 1
in
to Eq.(18), the corresp
onding reli
ability will be
obtaine
d. The
result
s are listed in the thi
r
d col
u
mn in
Table 1.
All the data
are
divided
i
n
to traini
ng
d
a
ta an
d p
r
ed
ictive data. T
he
single
-
ste
p
-ah
ead
predi
ction
s
a
r
e ado
pted. T
he num
ber
of the lagge
d
variabl
es i
s
3
5
. Initially, the
former
35 d
a
ta
s
e
t of time to
failure and
are con
s
ide
r
ed
as the t
r
ainin
g
data; the la
tter one a
r
e v
i
ewe
d
a
s
predi
ctive data. The time to failure is subs
tituted i
n
to the SVM and LSSVM
model, and t
h
e
reliability predictionsare computed.
Other predi
ctive result
s are
a
ttainted with the sam
e
method.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Applicatio
n of Support Vect
or Ma
chin
e to Re
liabilit
y Analysis of Engi
ne … (Zh
ang
Xinfeng)
3557
The re
sults a
r
e tabulate
d
in the
sixth and seventh col
u
mn in table 2. The figure of the reliabili
ty
of the turbo
c
h
a
rge
r
traini
ng
and pre
d
ictiv
e
results is
sh
own a
s
Figu
re 1(a
)
and Fi
gure
1
(b
).
To evalu
a
te
the p
e
rfo
r
mance of S
V
M
and
LS
SVM, the p
r
edi
ctive results
with
otherm
e
thod
s, su
ch
a
s
M
L
P (lo
g
isti
c
activation),
MLP (Gau
ssi
an a
c
tivation
), an
d
RBF, are
tabulated into
table 2. For the detail
s
ab
out t
heir co
m
putation process se
e Xu K et al.[1].
Substituting t
he predi
ctive results into
Eq
.(16),
thei
r
perspe
c
tive NRMS are a
ttained,
whic
h
are tabulated in table 2.
This
shows
th
at predic
t
ive
res
u
lts
by
SVM and LSSVM can be
approved
with co
mpa
r
e of
the othe
r re
sults. So
reli
ability predi
ct
ion of en
gine
system
by SVM
and LSSVM is effective.
1
2
3
4
5
6
7
8
9
0.
5
0.
5
5
0.
6
0.
6
5
0.
7
0.
7
5
0.
8
0.
8
5
0.
9
0.
9
5
1
T
i
m
e
t
o
f
a
i
l
ur
e (
x
10
00h)
R
e
li
a
b
ili
t
y
A
c
t
ual
dat
a
S
V
M
tr
a
i
n
i
n
g
d
a
ta
S
V
M
f
o
r
e
ca
st d
a
t
a
Figure 1(a). Reliability of the
turbocharger trai
ning and pr
edictive
results with S
V
M
1
2
3
4
5
6
7
8
9
0.
5
0.
5
5
0.
6
0.
6
5
0.
7
0.
7
5
0.
8
0.
8
5
0.
9
0.
9
5
1
T
i
m
e
t
o
f
a
i
l
ur
e (
x
10
00h)
R
e
lia
b
i
lit
y
A
c
t
ual
d
a
t
a
LS
S
V
M
t
r
a
i
n
i
ng
da
t
a
L
S
S
V
M
f
o
r
e
ca
st d
a
t
a
Figure 1(b). Reliability of t
he turbocharger
trai
ning and predictive
results with LSSVM
Table 2. Forecasti
ng result
of turbochargers reli
ability using different method
Number
Reliability
(ac
t
ual
)/%
MLP
(logistic
activation)/%
MLP
(Gaussian
activation)/%
RBF
(Gaussian
activation)/%
SVM
/%
LSSVM
/%
36 64.442
66.01
65.39
64.40
64.37
66.190
37 63.446
65.42
64.76
63.31
63.86
65.480
38 62.450
64.71
64.41
62.14
63.14
64.830
39 61.454
64.19
63.89
61.10
62.36
63.420
40 60.458
63.57
63.60
60.04
61.65
62.770
NRMS
0.0383
0.0338
0.0046
0.0088
0.0336
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN:
2087
-27
8
X
TELKOM
NIKA
Vol. 11, No
. 7, July 2013
: 3552 – 356
0
3558
3.2
Reliabilit
y
Anal
y
s
is of E
ngine S
y
stems .
In orde
r to verify the effect of the pre
d
ictive re
sult
s on reliabilit
y indexes, probability
distrib
u
tion n
eed
s to be
d
e
termin
ed. Weibull reliabilit
y paper is
plo
tted in Matlab
7.0. The a
c
t
ual
reliability an
d
the predi
ctive re
sult
s by
SVM and
LS
SVM are
ana
lyzed
with th
e help
of it. The
data pro
c
e
s
s results sho
w
that t
heir distri
bution
s
are consi
s
tent, whi
c
h both follo
w weibull failure
distrib
u
tion. F
o
r detail
s
se
e
the Fi
gure 3,
Figure 4, an
d Figure 5.
1
1.
5
2
2.
5
3
3.
5
4
4.
5
5
5.
5
6
6.
5
7
8
9
10
15
20
0.
1
0.
2
0.
3
0.
4
0.
5
1
2
3
4
5
10
20
30
40
50
60
70
80
90
95
99
99.
9
t
/1
0
3
h
D
i
s
t
r
i
b
u
ti
o
n
F
u
n
c
ti
o
n
F
(
t
)%
W
e
ib
u
ll P
r
o
b
a
b
ilit
y
P
a
p
e
r
S
h
ape P
a
r
a
m
e
t
e
r
b
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
1.
4
1.
6
1.
8
2
2.
2
2.
4
2.
6
2.
8
3
3.
2
3.
4
3.
6
3.
8
4
4.
2
Figure 3. Reli
ability analysi
s
of turbo
c
ha
rgers (a
ctual d
a
ta)
1
1.
5
2
2.
5
3
3.
5
4
4.
5
5
5.
5
6
6.
5
7
8
9
10
15
20
0.
1
0.
2
0.
3
0.
4
0.
5
1
2
3
4
5
10
20
30
40
50
60
70
80
90
95
99
99
.
9
t
/1
0
3
h
D
i
s
t
r
i
but
i
on F
unc
t
i
on
F
(
t
)%
W
e
i
bul
l
P
r
ob
ab
i
l
i
t
y
P
a
p
e
r
S
hap
e P
a
r
a
m
e
t
e
r
b
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
1.
4
1.
6
1.
8
2
2.
2
2.
4
2.
6
2.
8
3
3.
2
3.
4
3.
6
3.
8
4
4.
2
Figure 3. Reli
ability analysi
s
of tu
rbo
c
ha
rgers (fo
r
ecast
data by SVM)
The shap
e p
a
ram
e
ters an
d ch
ara
c
te
ristic
life of bot
h actu
al dat
a and th
e predictive
results by SVM and LSSVM are shown
as Tabl
e 3. Their go
odnesses of fit are all 1.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Applicatio
n of Support Vect
or Ma
chin
e to Re
liabilit
y Analysis of Engi
ne … (Zh
ang
Xinfeng)
3559
The relative error between
the actual reliab
ility and
the predi
ctive
reliability by
SVM is
only 14.9%. The rel
a
tive error bet
ween the act
ual
reliability and the predi
ctive reliability by
LSSVM is
only 4.1%, whic
h s
h
ows
that t
he predic
t
ive res
u
lt
s
are more perfec
t.
1
1.
5
2
2.
5
3
3.
5
4
4.
5
5
5.
5
6
6.
5
7
8
9
10
15
20
0.
1
0.
2
0.
3
0.
4
0.
5
1
2
3
4
5
10
20
30
40
50
60
70
80
90
95
99
99.
9
t
/1
0
3
h
D
i
s
t
ri
bu
t
i
on
F
u
nc
t
i
on
F
(
t
)%
W
e
i
bul
l
P
r
o
babi
l
i
t
y
P
a
p
e
r
S
hap
e P
a
ram
e
t
e
r
b
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
1.
4
1.
6
1.
8
2
2.
2
2.
4
2.
6
2.
8
3
3.
2
3.
4
3.
6
3.
8
4
4.
2
Figure 5. Reli
ability analysi
s
of turb
ochargers (forecast
data by LSSVM)
Table 3. Reli
ability analysi
s
re
sult by we
ibull pro
babilit
y paper
b
T
(
×
10
3
h
)
Good
ne
ssof fit
Actual r
e
liability
2.423
12.155
1
Predictivereliabilit
y
b
y
SVM
2.397
13.966
1
Predictivereliabil
i
t
y
b
y
LSSVM
2.290
12.653
1
4 Conclu
sions
This research applied the SVM and LSSVM
in the reliability
predi
ction of engine
s
y
s
t
ems
.
The predic
t
ive performa
nce of the SVM and LSSVM were
c
o
mpared
with that of the
neural n
e
two
r
ks of MP
L
(logi
stic act
i
vation)
, ML
P (Ga
u
ssian
activation
), and
RBF.
The
s
i
mulation ex
periment out
c
o
mes
s
h
ow that t
he predic
tive results
by SVM and LSSVM are
perfe
ct. The f
a
ilure
distri
bu
tion analysi
s
of the pr
e
d
icti
ve reliabilitie
s of engine
sy
stem
s with S
V
M
and LSSVM was close to that of the actual. So
SVM
and LSSVM
are the alternative choices of
the reliability predi
ction of
engin
e
syste
m
s.
The n
u
me
rical re
sult
s al
so sho
w
that
the p
r
edi
ct
pre
c
ise of th
e metho
d
b
a
s
ed
on
LSSVM is higher than that
of SVM. Especially in
small
samples, the predi
ction by
LSSVM will
be
more p
opul
ar, becau
se its
comp
ution co
st is
lowe
r a
nd the pre
c
ise can b
e
more satisfie
d.
Ackn
o
w
l
e
dg
ments
This resea
r
ch
has b
een
su
pporte
d in pa
rt by the Fun
damental
Re
sea
r
ch Fun
d
s for the
Central
Unive
r
sitie
s
(CHD2
012
JC048, K
5051
2060
11
), the Fun
dam
ental Resea
r
ch Su
ppo
rt Plan
Fund
s for
Cha
ng’an
University, an
d Openi
ng
Fund of Ke
y Laboratory
of Automotive
Tran
sp
ortatio
n
Safety Enhancement Te
chn
o
l
ogy of the Minist
ry of Commu
nication.
Referen
ces
[1]
Xu K,
Xi
e M, T
ang
LC, Ho S
L
. A
pplic
ation
of neur
al n
e
tw
orks in forecasti
n
g en
gin
e
syste
m
s re
lia
bil
i
ty
.
Appl
ied S
o
ft Computi
ng. 20
0
3
; 2: 255-2
68.
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN:
2087
-27
8
X
TELKOM
NIKA
Vol. 11, No
. 7, July 2013
: 3552 – 356
0
3560
[2] Ross
SM.
Introductio
n
to prob
abil
i
ty mode
ls
. (10e
d). San Di
ego: Acad
emic
Press. 2010.
[3]
Ascher H, F
e
i
ngo
ld H.
R
epa
irabl
e syste
m
s
relia
bil
i
ty: mo
deli
ng, i
n
fere
n
c
e, misc
onc
ept
ions
and th
ei
r
causes
. Ne
w
Y
o
rk: Marcel De
kker, Inc. 1984
.
[4] Hay
k
in
S.
Ne
ur
al netw
o
rks: a compre
hens
ive
foundati
o
n
. En
gle
w
o
o
d
CliKs,
NJ: Prentice H
a
ll, 19
99.
[5] Z
heng,
Ju
n.
Predicti
ng soft
w
a
re reli
abi
lity
w
i
th neura
l
netw
o
rk ense
m
b
l
es.
E
x
pert S
y
stems
w
i
t
h
Appl
icatio
ns. 2
009; 36: 2
116-
212
2.
[6]
Lol
as S, Olatu
nbos
un OA.
Pr
edicti
on
of ve
hi
cle re
lia
bi
lity p
e
rformanc
e us
i
ng
artificia
l
n
e
u
r
al n
e
tw
orks
.
Exp
e
rt S
y
stem
s
w
i
th A
ppl
icati
ons. 200
8; 34 (
4
): 2360-
23
69.
[7] Vapnik
V.
Statistical le
arni
ng theory
, Ne
w
Y
o
rk: John W
ile
y
and So
ns; 199
8.
[8] Vapnik
V.
T
he nature of statis
tical le
arni
ng th
eory
. Ne
w
Yor
k
: Springer-V
er
lag; 20
00.
[9]
Han, F
eng
qi
n
g
, Li, Hongm
ei, and etc
A ne
w
incr
e
m
ental su
pp
ort vector machin
e alg
o
rithm
.
Te
lkom
n
i
ka
.
20
12; 10(6): 1
171
-117
8.
[10]
Yu, Yang, Z
h
o
u
, Lian
g. Acou
stic emissio
n
sign
al class
i
fic
a
tion b
a
se
d o
n
supp
ort vector machi
ne.
Te
lkom
n
i
ka
. 20
12; 10(5): 1
027
-103
2.
[11]
Su
y
k
ens JAK, Gestel T
V
, Brabanter JD, Va
n
d
e
w
a
l
l
e
J.
Lea
st Square Su
p
port Vector Ma
chin
es
. World
Scientific, Sin
g
apor
e.20
02.
[12]
S Dumais, J Platt, J Heckerman and M Sahami.
Inductive l
earn
i
ng a
l
g
o
rit
h
ms a
nd re
pre
s
entatio
ns for
text categori
z
a
t
ion
. In Proce
e
d
i
ngs of ACM-CI
K
M98. 199
8.
[13]
Cao LJ, T
a
y
F
r
ancis EH.
Su
pport vector machi
ne w
i
th ad
aptive p
a
ra
met
e
rs in fina
ncia
l
time seri
es
forecasting
. IEEE
T
r
ansactio
n
s on Ne
ural N
e
t
w
o
r
ks. 200
3; 14: 150
6-1
518.
[14]
Moura
Md
C, Z
i
o E
an
d
ect.
Failur
e
and
re
lia
bility
pre
d
ictio
n
by s
upp
ort ve
ctor mach
ines
regressi
on
o
f
tim
e
series data
. Reli
abi
lit
y
E
ngi
neer
in
g & System Safet
y
. 201
1; 96(1
1
): 1527-
153
4.
[15]
Sapa
nkev
yc
h NI, Sankar R.
T
i
me ser
i
es pr
edicti
on usi
ng
supp
ort vector mach
in
es: a survey.
IEEE
Comp
utation
a
l Intelli
genc
e
Ma
gazi
ne. 20
09; 4: 24-38.
[16] Vapn
ik
V.
T
he nature of statis
tica
l le
arni
ng th
eory (firsted)
. Sprin
ger Ne
w
York, 1995.
[17]
Dua
n
K, Keerthi SS, Poo AN. Evaluati
on
of simple
performanc
e
measures fo
r tuning SVM
h
y
per
param
ete
r
s.
Neuroco
m
p
u
ting
. 20
03; 50
: 41-59.
[18] W
Nelson.
App
lied
Life Dat
a
Analysis
. Wiley
,
Ne
w
York. 1988.
Evaluation Warning : The document was created with Spire.PDF for Python.