基于优化算法的串联体系可靠度分析
第2l卷第6期
计算力学学报
V01.21,No.62004年12月
ChineseJournalofComputationalMechanics
December
2004
文章编号:1007—4708(2004)06—0665—06
Reliabilityevaluationofseriessystems
‘・-
‘
with0DtlmlZatlon』●
●
・
-J●
technique
Li
Gang“,MeyerJ2
(1.StateKeyLaboratoryofStructuralAnalysisofIndustrial
Equipment
DalianUniversityof
Technology,Dalian116023,China;
2.Departmentof
ChemicalEngineering
ColoradoState
University,FortCollins,CO80523,Colorado,USA)
Abstract:Thefailureprobabilityofa
system
can
beexpressed
asan
integralofthejointprobabilitydensityfunc—
tionwithinthefailure
domaindefinedbythelimit
state
functionsofthe
system.Generally,itis
very
difficultto
solvethisintegraldirectly.Theevaluationofsystemreliabilityhasbeentheactiveresearch
area
duringthe
re—
cent
decades.Some
methodsweredeveloped
tosolve
systemreliabilityanalysis,such
as
MonteCarlomethod,
importancesampling
method,boundingtechniquesandProbabilisticNetworkEvaluationTechnique(PNET).
This
paper
presents
the
implementationofseveraloptimizationalgorithms,modifiedMethodofFeasibleDirec—
tion(MFD),SequentialLinearProgramming(SLP)andSequentialQuadraticProgramming(SQP),inorderto
demonstratethe
convergence
abilitiesandrobustnature
oftheoptimizationtechniquewhenapplied
to
series
sys—
temreliabilityanalysis.Examplestaken
fromthepublishedreferenceswerecalculatedandtheresultswerecom—
paredwiththeanswersofvariousothermethodsandtheexact
solution.Resuhsindicatetheoptimizationtech—
niquehas
a
wide
range
ofapplicationwithgood
convergence
abilityandrobustness,andcan
handleproblems
un—
dergeneralizedconditionsoF
cases.
Key
words:systemreliability;FirstOrderSecondMoment;optimizationalgorithms;MonteCarloMethod;
importance
sampling
od,importancesamplingmethod)andbounding
1Introuction
techniques,Probabilistic
NetworkEvaluation
Theevaluationofsystemreliabilityanalysis
Technique(PNET)werealwaysemployed.
hasbeentheactiveresearcharea
duringthe
recent
Thepresentpaperfocuses
on
the
optimiza—
a
tion
evaluate
decades.Generally,thefailure
probability
of
technique
to
system
reliabilit;r.
system
can
beas[1]
First,themulti—limit
state
functionsoftheseries
expressed
system
are
transformed
into
an
equivalent
limit
孝一
(1)
J
f...k(x)dx
J
state
function,and
to
calculatesystemreliability
上J
whereXis
a
vector
oftherandom
variables,f;
is
to
calculatetheminimumdistancefromtheori—(X)isthejointprobabilitydensityfunctionandDgin
to
thesingleequivalentlimit
statefunctionin
is
the
failuredomain
defined
standardnormalto
thebythe
limit
state
spaceaccording
conceptof
functions
ofthe
system.It
is
very
difficultto
theFirstOrderSecondMoment(FOSM).Three
solvetheemployed
to
solvethis
integralEq.(1)directly,andnumerical
optimizationalgorithms
are
simulationare
modifiedMethod
methods(such
as
MonteCarlometh—
optimizationproblem,which
ofFeasibleDirection(MFD),SequentialLinear
Reeeivedby:2002-11-20;Revised
by:2003-07-18.
Programming(SLP)andSequentialQuadraticProjectsupportedby
NationalNatureScienceFoundationof
China(50008003);863Program(2001Programming(SQP).Todemonstratethe
con—
AA602015):NSFREUProgram.
vergenee
abilitiesandrobust
nature
oftheopti—
Corresponding
authors:Li
Gang’(1966一),Male,Doe.,
mizationforASSOC.Prof.
techniquesystemreliabilityanalysis,
万
方数据
——————————————————————————————————————————————————————一一
sss
计
兰垄
竺兰塑
竺!!兰
the
MonteCarlomethod,theimportancesam—
plingmethodandHL—RFalgorithmare
alsoused,andtheirresults
are
compared
as
well.
2
Equivalentlimitstatefunctionof
seriessystem
There
are
many
typesof
systems,allof
which
can
be
categorized
into
twofundamental
svstems,i.e.seriessystemandparallelsystem.
Seriessysterns
can
bethought
ofbya“weakest
link”analogy.If
one
componentfails,theentire
svstemreachesfailure.Parallelsystems
are
also
known
as“redundantsystems”.
If
one
element
fails,thesystemdoesnot
necessarily
reachfail—
ure.Inthe
present
paper,series
systems
are
stlldied.
≮么D}
.q
卜
failure
safeny
/弋≤
毫碍仨
}
..。\己
/..・孟/N甏
Fig.1
Series
system
Fora
seriessystem,thefailureregionisthe
union
ofthe
failureregions
ofeachlimit
state
function,andthe
failure
probability
P,is
ex—
pressed
as
Fig.1
一PmU引
G
痧7
≤o)一卜X)dX(2)
%
Theequivalent
limit
state
function
ofthe
seriessystem
can
bedefined
as
G。。i。。一min(Gl,G2,…,G。)
(3)
Thus.Eq.(2)canbewritten
as
少一P(G。。血。≤O)一
¨
.
,,.f
(XdY
(4)
,
G
耻m≤
0
3
Structuralreliabilityanalysis
Using
the
equivalent
limit
state
function
definedinEq.3,thereliabilityanalysismethodsofthecomponent
casecan
alsobeused
to
solve
the
systemreliability
problem.Manymethods
havebeendevelopedandimplemented
to
solvethe
万
方数据1imit
state
functionsforcomponentfailureproba—
bility[¨.3.1
MonteCarlomethod
孝≈,,=专∑J[G(置)≤o]
(5)
whereⅡ]is
an
indicatorfunction,equals
to
1if厂]is
true
andequals
to
0
if[]isfalse,蜀is
the
拓矗sampleofthevector
ofrandomvariable,Nis
thetotal
numberofsamples.Averylargenum—
berofsamples
are
neededfor
the
MonteCarlo
method
to
theproblemwithlowfailureprobabili—
tv,whichmeanstheMonteCarlomethodistime
consuming.Astrengthofthismethodisitsbroad
problemapplicability.Aslongas
thelimit
state
function
can
be
calculated,continuityand
the
ability
to
derivatethefunctionisnota
problem
in
application.With
suchhigheriterationnumbers’
thesolutionsofthismethod
can
berelatively
ac一
3.2
Importancesamplingmethod
/
每姚一南著N{邸(毫)<o]
jz(、x,
h。(Xj
}㈤
where
h。()iscalled
the
importance
sampling
probabilitydensityfunction.Theproper
selection
of玩()will
generate
moresamplesinthefailure
region,andthuscan
reduce
thenumberofsam—
piesneeded.3.3
FirstOrderSecondMomentmethodThe
First
OrderSecond
Moment
method
(FOSM)is
a
techniqueused
to
transformthein—
tegralinto
a
multi—normaljointprobabilitydensi—
tv
function.Thisisdonebyutilizing
fl
linearrep—
resentationof
a
limit
state
functionbywayofits
firsttwomoments,themeanandstandarddevia—tion.HasoferandLinddefinedreliabilityindex
p
as
theshortestdistancefromtheorigin
to
thelim—
itstate
surface[引,which
can
beexpressed
as
卢一min(∑y;)1/2一rain(yT・y)“2(7)
i一1
subject
to
G(y)-----0
whereG(v)isthelimitstatefunctioninstandard—izednormalspace,Yiistheindependent
random
variable
ofstandard
normaldistribution
on
the
limit
state
surface.Forthelimit
state
function
withgeneralrandomvariablesofnon—normaldis—
第6期
李
刚,等:基于优化算法的串联体系可靠度分析
667
tribution,therandom
variablescan
be
trans—
formed
intothe
independent
equivalent
normal
random
variables
bynormaltailtransformation.
Rosenblatttransformationor
Nataftransforma—tion.
The
relation
between
reliability
index
and
failureprobabilityis
P,一垂(一f1)
(8)
where垂representsthestandardnormaldistribu—tion.Eq.8isaccurate
onlyforthespecificcase
of
random
variables
withnormal
distributionand
linearlimitstate
functions.
ItshouldbenotedthattheFOSMmethodis
infact
a
type
ofoptimization,asshowninEq.7.
algorithms
can
beemployed
to
solvesuchoptimizationproblem.Thefamous
HL—RFal—
is
an
efficientchoiceformanyproblems,
whichthereliabilityindexisfoundbyan
itera—process
after
a
linear
approximation
ofthe
state
function,definedat
the
designpoint
byEq.7)[2’引.
∥一一∑y。。(aG/ayi)/[∑(aGlayi)2]l/2(9)
Optimization
algorithms
A
general
optimization
problem
can
be
ex—
as
find.X
rain.
W(X)s.t.
Gj(X)≤0J=1,…,朋
XL≤X≤Xu
(10)
Three
optimizationalgorithmsavailable
in
program(DesignOptimizationTools)are
inthepresentpaper,which
are
the
Modified
ofFeasibleDirection(MFD),Sequential
Programming(SLP)andSequential
Programming(SQP)t4|.
ModifiedmethodofFeasibleDirection
Thefeasibledirectionmethodis
intheclassdirectsearchalgorithms,which
can
bestated
x‘+1一x‘+叠dk(11、and掣+1
are
thekth
and(是+1)£^
designspace,d‘is
the
directionand∥isthedistanceoftravelbe—
万
方数据tweenthesetwodesignpoints.Thecriticalpartsof
the
optimization
task
are
finding
a
usable
searchdirectionandthetraveldistance.Herewe
use
the
Fletcher—Reeves
conjugate
direction
method,thesearchdirection
as:
2
d‘一一VⅣ(X‘一1)+
VW(X卜1)∥_1(12)
VW(X6—2)
—2
Having
determined
a
usable—feasible
search
direction,theproblemnowbecomesone—dimen—
sionalsearch
thatminimizes
W(X卜1+o?d‘),
whichcan
besolved
by
many
available
algo—
rithms.4.2
SequentialLinearProgramming(SLP)Thebasic
concept
ofSI。Pis
quitesimple.
First,create
a
TaylorSeriesapproximation
to
the
andconstraintfunctions
形(x)=Ⅳ(∥1)+Vw(x‘一1)T(X‘一X6—1)
(13a)
;,(x)=g,(x卜1)+Vg,(x‘一1)T(x6一x‘一1)
(13b)
Then,usethisapproximationforoptimiza—
oftheoriginalnonlinearfunctions.theoptimizationprocess,definemovelira—on
thedesignvariables.Typically,during
onedesign
variables
will
be
allowed
to
by
20%---40%,butthisisadjustedduring
cycles.
SequentialQuadraticProgramming(SQP)Thebasic
concept
isverysimilar
to
thatof
a
TaylorSeriesapproximation
a
quadraticapproximateobjectivefunctionand
constraints,with
whicha
direction
problemisformed
as
follows:
1
min.Q(d)一Wo+VwTd+去dTBd
s.t.g;+VgYd<O,歹一1,…,m
(15)
Thissub—problemissolvedusingMFD.The
BinEq.15is
a
positive
definitematrix,
is
initiallythe
identitymatrix.Onsubse—iterations,B
is
updated
to
approach
the
oftheLagrangianfunction.
Nowassumewehavesolvedtheapproximate
ofminimizingQsubject
to
thelinearized
theoptimumforthisproblem,wethe
Lagrange
multipliers,tj.We
now
Variousan
objectivegorithmintive
tion,insteadDuringits
limit
(expressed4
cycle,the
changepressed
later4.3
SLP.First,createof
DOT
linearizedfindingusedmethodLinear
matrixwhichquent
Quadratic4.1
ofaS
Hessianwhere,Fproblempoint,respectively,inconstraints.Atcalculate
search
668
计算力学学
报第21卷
searchindirectiond
using
theapproximateLa—
grangianfunction.Thatis.wefind口tominimize
西一Ⅳ(x)+∑甜Jmax
E0,gJ(x)]
(16)
』=1
where
X—X卜1+ad
(17)
Afterthe
one—dimensional
search
is
completed,
theBmatrixisupdatedusing
theBFGSformu—
la[43.
5
Numericalexamples
TheDesignOptimizationTools(DOT)pro—
andthereliability
part
ofProgramsforReli—
AnalysisandDesignofStructuralSystems(PRADSR)are
employed
inthe
present
pa—
isa
programthatsolvesforthe
offailure,coefficientofvariance,andreliabilityindexbytheMonteCarlomethod,the
method,andimportancesampling.
A1ltheexamplesevaluatedinthepresentpa—
are
takenfrom
the
publishedpapers.The
tWOexamplesare
componentreliabilityprob—
tO
showthe
general
applicability,conver—
abilityandrobust
nature
oftheoptimization3——5
are
reliabilityanalysis
seriessystems.Theresultsforeachoptimiza—methodwillberepresentedbySQPalgorithmall
three
optimization
algorithmsproduce
identicalresults.Thenumberofsamples
MonteCarlomethodandimportancesampling
are
100000
and500,respectively.Error
representedbyrelative
error
withrespecttotheresults
or
theresults
inthe
reference
exact
values
are.not
known.
Example1[6]
and“2
are
standardnormalrandomvariables.
Tab.1
Resultsofexample1
万
方数据5.2
Example2[7]
g—z;+z:一18,
zl,z2~Ⅳ(10,5)
Tab.2
Resultsofexample2
5.3
Example3L“
g,一0.15x1+0.3xs+0.3x4—0.17x5一O.5x6+3g。一0.15Xz+0.3xs+0.15x4—0.5x6+3
g。一0.15x1+0.15x2+0.15x4—0.17x5+2
91,92
and93
are
limit
state
functionsof
ase—
system,zl~z6
are
standardnormalrandomequivalent
limit
state
functionof
frameisG。。,i。。=min(91,92,93)
Tab.3
Resultsofexample
3
n
Fig.2
One
bayand
one
storyframe
Example4[8]
ThespanandheightofthisframeFig.2
are
ft
and
15ft.Its
significant
potentialfailure
are
g,一M1+3M2+2M3—15S1—10S2
g。一2M,+2%一15Sl
g,一M1+Mz+4Ms一15Sl一10S2g。一2M1+M2+M3—15S1
g。=尬+%+2Ms一15S1g。一M十2%+%一15S1
AUtherandomvariableshavelognormaldis—
withthestatisticalparametersshowninequivalent
limit
state
functionofthis
isG。i。=min(gl,92,93,94,95,96)
Tab.4
Randomvariablesofexample4
gramabilityriesperE4’5J.PRADSRvariables.TheprobabilitythisFOSMper
firstlems
gence
technique.Example
fortionsince5.4
20nearlyformodes
method
isaccurate
where5.1
g,=0.1(“;+“;一2“l“2)一(“l+甜2)/√2+2.592一一O.5(“;+“;一2“1“2)一(“l+“2)/J2+3.0
“1
tributiontable.Theframe
第6期
李
刚,等:基于优化算法的串联体系可靠度分析
669
Tab.5
Resultsofexample4
Fig.3
Two—bayandtwo—story
frame
Example5C8]Thespan
andheightofthisframe(Fig.3)20ft
and
15
ft,and
itssignificantpotential
modes
are
g,一2%+2%+2坛一15S1—15S2g。=M6+M7+2慨一10S3
g。=Ms+3Ms一10S4g。=M7+3M8—10Ss
g。一2M1-'I-2M2+%+坛一15S1—15S2
96=M6+3M7+2Ms一15S,
Alltherandomvariableshavelognormaldis—withthestatisticalparametersshowninequivalentlimitstate
functionofthis
is
G。。,洒=min(91,92,93,94,95,96)Tab.6
Randomvariablesofexample5
Tab.7
Resultsofexample5
Discussionandconclusions
The
results
of
the
examples
demonstrated
convergenceabilitiesandrobust
nature
ofthe
万
方数据optimizationtechniquewhenapplied
tO
structural
reliabilityanalysis.Forbothcomponentandsys—tern
reliability
analysis,the
optimization
algo—
rithms
employedinthe
presentpaper(MFD,
SLP,SQP)canproducesatisfyingsolutionswith
negligibleerror.TheFOSMmethod
can
alsogivesatisfyingresultsforsomecases(suchas
example4and
5)andmayproduce
high
error
bylinear
representationofhighlynonlinearfunctionsat
the
designpointforothercases,suchas
theresultsofexample
3
withtherelative
error
of49.12%.
MonteCarlomethod
can
alwaysproducegood
re—suits
as
long
as
thenumberofthesamplesis
e—
nough,whichresultsinexpensivecomputationalefforts.Importancesamplingmethodcan
reduce
the
numberofsamples
a
lot
byintroducingthe
importancesamplingprobabilitydensityfunction,
however,thedesignpointshouldbedetermined
at
firSt.
妇■、太~G。o。N
\l
∥l\
。皖o。
Fig.4
FOSM
Error
Inreliabilityanalysis,tWOstandard
indices
used;theprobabilityoffailureandtherelia—
index.They
can
betransformedintoeach
byEq.8,whichisaccurateonlyforthespe—case
ofrandomvariableswithnormaldistri—
andlinear
limit
statefunctions.The
Carlomethodsolvestheformerwhiletheandoptimizationmethodssolveforthelat—
shownbytheaboveexamples,thistrans—
can
introduce
error
whensolving
forspecific
index.A
graphical
explanation
is
inFig.4.Thefailureregion,andthustheoffailure,isgreaterforG2(X).Thedistance,however,iSthesameforboth
errors
betweenthedif-
methodsofMonteCarlo,importancesam—
and
optimization
algorithms
are
by(1)the
algorithms(such
as
FOSM
5.5
are
failuretributiontable.Theframeare
bilityothercific
butionMonteFOSMter.Asformation
one
shownprobabilityshortest6
functions.Therefore。the
ferentpiing,FOSMcausedthe
670
计算力学
学报
第21卷
method)and(2)thetransformationbetweenreli—
abilityindexandfailurenonlinearlimitablesother
as
state
Prediction[M].JohnWiley&Sons,1999.[23
HasoferA
probabilitybyEq.8
for
M.LindNC.An
exactandinvariantEng
functionwithrandomvari—
firstorderreliabilityformat[刀.J
MechDiv,
thannormalrandomvariables(such
ASCE,1974,100(EMl):111一】21。
MonteCarlomethodandimportancesampling)
Optimizationtechniqueis
a
[3]RackwitzR,FiesslerB.Structuralreliabilityundercombinedrandomload
goodmethodfor
a
sequences[J].Computers
and
Structures,1978,9:489—494.
findingthereliability
indexofstructuralrelia—
bilityproblem.Aswithallmethods,applicationshould
use
[4]
DOT
user
manual.VanderplaatsResearchandDe—
Springs,CO,1999.
part
velopment[M].Inc.,Colorado[5]
Sorensen
fittheresearch
or
designproblem.The
JD.PRADSR:Reliability
ofPRA—
ofothermethodsinconjunctionwithoptimi—
recommendedfor
referencecouldfocus
to
tionmethodiScomparativeoptmizationandhow
DSS[R].AalborgUniversity,Denmark,1994.
[6]
BorriA,SperanziniE.Structuralreliabilityanalysisusing
a
and
on
use.Futureresearchtechnique
when
standarddeterministicfiniteelementcode
appliedparallel
on
[J].StructuralSafety,1997,19(4):361—382.[7]
XuL,ChengGD.INscussionfor
structural
on
combinedsystemsconsideringespecially
to
momentmethods
applyoptimizationtechniques
tO
suchsys—
reliabilityanalysis[J].Structural
t户ms.
Safety,2003,25:193-199.
[83ZhaoYG,OnoT.Systemreliabilityevaluationductile
frame
of
References:
[13
Melchers
RE.Structural
ReliabilityAnalysis
and
structures[J].Journalof
Structure
Engineering,1998,124(6):678-685.
基于优化算法的串联体系可靠度分析
李
刚“,
MeyerJ2
(1.大连理工大学工业装备结构分析国家重点实验室,辽宁大连116024;
2.科罗拉多州立大学化学工程系,科林斯堡,C080523,美国)
摘要:结构体系的失效概率数学上可以表示为结构体系失效域上联合概率密度函数的积分,一般情况下很难直接积分求解。近几十年来,结构体系可靠度分析一直是可靠度领域的一个研究热点,人们提出许多方法,如:Monte—Carlo法、重要性抽样法与界限法和概率网络估算技术等,这些算法在求解精度、计算效率、收敛性和易使用性等方面是不同的。本文采用优化算法(改进的可行方向法、序列线性规划和序列二次规划法)进行串联体系可靠度分析,并且与其他算法(HL—RF法、Monte—Carlo法和重要性抽样法)的结果以及一些精确解进行了比较。结果表明,相对于其他算法,基于优化算法的可靠度分析适用性广,在收敛性和健实性等方面具有明显的优势。关键词:体系可靠度;一次二阶矩法;优化算法;Monte—Carlo法;重要性抽样法中图分类号:TU311.2
收稿日期:2002一ll一20;修改稿收到日期:2003—07—18.
基金项目:国家自然科学基金(50008003);863计划(2001AA602015);美国NSFREU资助项目.作者简介:李刚。(1966一).男,博士,副教授.
万方数据
基于优化算法的串联体系可靠度分析
作者:作者单位:刊名:英文刊名:年,卷(期):被引用次数:
李刚, Meyer J
李刚(大连理工大学,工业装备结构分析国家重点实验室,辽宁,大连,116024), Meyer J(科罗拉多州立大学,化学工程系,科林斯堡,CO,80523,美国)计算力学学报
CHINESE JOURNAL OF COMPUTATIONAL MECHANICS2004,21(6)4次
参考文献(8条)
1. ZhaoYG;Ono T System reliability evaluation of ductile frame structures[外文期刊] 1998(06)2. XuL;Cheng G D Discussion on moment methods for structural reliability analysis[外文期刊] 2003(2)3. BorriA;Speranzini E Structural reliability analysis using a standard deterministic finite elementcode 1997(04)
4. SorensenJD PRADSR: Reliability part of PRA-DSS 19945. DOTusermanual Vanderplaats Research and Development 1999
6. RackwitzR;Fiessler B Structural reliability under combined random load sequences 19787. HasoferAM;Lind N C An exact and invariant first order reliability format 19748. MelchersRE Structural Reliability Analysis and Prediction 1999
引证文献(4条)
1. 秦力. 张学礼. 陶颐格 500 kv输电铁塔结构体系可靠性分析[期刊论文]-中国电力 2008(12)
2. 秦力. 张学礼. 张杰. 朱军 浅析500 kV输电铁塔结构体系的可靠性[期刊论文]-东北电力大学学报 2008(2)3. 陈海波. 廖宗高. 肖洪伟 受风荷载控制的杆塔结构体系可靠度分析[期刊论文]-电力建设 2007(7)
4. 蒋友宝. 冯健. 孟少平 求解结构可靠指标的线性可行方向算法[期刊论文]-东南大学学报(自然科学版) 2006(2)
本文链接:http://d.g.wanfangdata.com.cn/Periodical_jslxxb200406005.aspx