XML Template (2013) [6.12.

2013–12:22pm] [1–13]
//blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/JVCJ/Vol00000/130570/APPFile/SG-JVCJ130570.3d (JVC) [PREPRINTER stage]

Article
Journal of Vibration and Control
0(0) 1–13

Dynamic non-probabilistic reliability- ! The Author(s) 2013

Reprints and permissions:
sagepub.co.uk/journalsPermissions.nav
based topology optimization of truss with DOI: 10.1177/1077546313514761
jvc.sagepub.com
uncertain-but-bounded parameters

Bin Xu

Abstract
In this paper, non-probabilistic reliability indices for frequency and static displacement constraints are analyzed based on
the ellipse convex model of elastic modulus and mass density. The dynamic non-probabilistic reliability-based topology
optimization model of a truss is built, where the cross-sectional areas and nodal topology variables are taken as design
variables. The objective is to minimize the structural total mass. Constraints are imposed on static stresses and
non-probabilistic reliability indices of static displacement and natural frequency. A genetic algorithm is used as the
optimization method to find optimal solutions in the outer loop and an analysis method is adopted to seek
non-probabilistic reliability index according to implicit forms of the limit state function in the inner loop. Results of
numerical examples show that the optimal mass of a non-probabilistic reliability-based topology optimization is larger
than that of the deterministic topology optimization and the optimal mass increases with the increase of the non-
probabilistic reliability requirement in order to ensure structural safety.

Keywords
Ellipse convex model, natural frequency, non-probabilistic reliability, topology optimization, truss

They used only the first order derivatives of certain

1. Introduction
eigenvalues with respect to design variables to deter-
A truss is a type of structure commonly used in engin- mine whether a specific natural frequency constraint
eering because it is simple and inexpensive to construct. was satisfied. Based on differentiations of Lagrangian
In order to improve structural dynamic characteristics functions, Wang et al. (2004) put forward an optimality
or to reduce the structural total mass under dynamic criteria algorithm for the simultaneous shape and size
constraints, truss dynamic optimization has gained optimizations for three-dimensional truss structures
widespread popularity in academia and industry. with multiple constraints on natural frequencies.
Depending on the types of design variables, truss Obviously, structural topology optimization is an
dynamic optimization is categorized into size optimiza- advanced design method in the field of structural
tion, shape optimization and topology optimization. design. Several researchers have introduced and imple-
A number of papers have dealt with the dynamic opti- mented various algorithms to address this problem.
mization of the truss. Tong and Liu (2001) presented a
two-step optimization procedure for the minimum School of Mechanics, Civil Engineering and Architecture, Northwestern
weight optimization with discrete design variables for Polytechnical University, Xi’an, People’s Republic of China
truss structures subject to constraints on stresses, nat- Received: 12 May 2013; accepted: 21 October 2013
ural frequencies and frequency responses. Furthermore,
Tong et al. (2000) pointed out that the natural fre- Corresponding author:
Bin Xu, School of Mechanics, Civil Engineering and Architecture,
quency constraint was usually the key constraint in Northwestern Polytechnical University, Xi’an, 710072, People’s
determining the solution existence of a dynamic opti- Republic of China.
mization problem of a truss with size design variables. Email: [email protected]
XML Template (2013) [6.12.2013–12:22pm] [1–13]
//blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/JVCJ/Vol00000/130570/APPFile/SG-JVCJ130570.3d (JVC) [PREPRINTER stage]

2 Journal of Vibration and Control 0(0)

Ohsaki et al. (1999) formulated the topology optimiza- parameters and formulated the optimal design as a
tion problem of trusses for a specified eigenvalue of nested optimization problem. Xu and Jin (2012) inves-
vibration as a semi-definite programming problem, tigated the multi-objective topology optimization
which is especially useful in the case where the optimal of trusses with interval parameters based on Taylor
design has multiple fundamental eigenvalues. Xu et al. expansion and natural interval extension, where the
(2003) presented a practical methodology based on a non-deterministic optimization problem is transformed
topology group concept to find optimal topologies of into a deterministic programming problem. Li et al.
trusses with a constant nodal mass. Norapat and Sujin (2011) developed an efficient nonlinear interval
(2011) applied three methods to the simultaneous optimization-sequential linear programming (NIO-
topology, shape and size optimization of a three- SLP) method for uncertain structures. Obviously,
dimensional truss structure. Obviously, natural fre- when uncertainties are considered in structural opti-
quency constraint is the key constraint for dynamic mization, the deterministic optimization would usually
optimization. Structural dynamic characteristics can result in a severe violation of the reliability con-
be significantly improved by using optimal design. So straints, i.e., the optimal solution for the deterministic
far, most of the papers dealing with dynamic optimiza- optimization problem may be unfeasible for the non-
tion have been concerned with the optimization of deterministic optimization problem. Therefore, uncer-
structures with deterministic parameters. tainties should be taken into account for a rational
However in engineering practice, there are inevit- structural design. In essence, the non-probabilistic reli-
ably many uncertain parameters which possibly affect ability optimization is a nested optimization problem
structural safety and durability. Based on the modes of and there are difficulties in solving this type of opti-
existence, uncertain parameters can be categorized into: mization problem. In addition, the dynamic character-
(1) uncertainty of physical or inherent parameters, istics become much more important for structures
including material constants, structural shape, bound- under dynamic loading. There has been very little
ary conditions and structural damping; (2) uncertainty work in the literature on the dynamic topology opti-
caused by insufficient knowledge; (3) statistical uncer- mization of trusses under non-probabilistic reliability
tainty. When designing and analyzing structures, it is constraints, particularly on the reliability analysis of
necessary to use some mathematical models to describe natural frequency.
or quantify uncertain parameters. Up to now, mathem- The remainder of this paper is organized as follows:
atical models adopted are the probability model, the in Section 2, some dynamic and static non-probabilistic
fuzzy model and the convex model. For engineering reliability indices based on the ellipse convex model are
structures, it is difficult and costly to specify a precise analyzed; in Section 3, the dynamic non-probabilistic
probability distribution or membership functions for reliability-based topology optimization model of
some uncertain parameters so as to restrict their wide trusses is built and the corresponding optimization
application to practice. It is, however, easy to ascertain algorithm is presented; two numerical examples are stu-
the amplitudes or boundaries of uncertain parameters. died in Section 4; and finally concluding remarks are
It is feasible to analyze accurately the performance of provided in Section 5.
structures with uncertain-but-bounded parameters. Qiu
and Wang (2003) considered the dynamical response of 2. Dynamic and static non-probabilistic
structures with uncertain-but-bounded parameters by reliability index based on ellipse
using the interval analysis method and the probabilistic
approach. Further, the non-probabilistic reliability-
convex model
based structure optimization using the convex model There are many kinds of definitions and the corres-
has become a promising research direction in the opti- ponding algorithms about the analysis of the non-
mization of uncertain structures. Lombardi (1998) used probabilistic reliability index. Qiu and Wang (2010)
the probabilistic optimization to deal with the uncertain measured structural non-probabilistic reliability using
optimization problem where only bounds for uncertain the ratio of the volume of the safe region to the total
variables are defined and a closed domain of variation volume of the region associated with the variation
is built. Luo et al. (2009) investigated the topology opti- of the standardized interval variables. Successively
mization of continuum structures with uncertain-but- based on the convex model of uncertain parameters,
bounded parameters using a quantiEed measure the non-probabilistic reliability index was defined
for non-probabilistic reliability. Kang et al. (2011) more perfectly as the minimum distance of a normal-
presented the mathematical deEnition of a non- ized failure surface from the origin of a normalized
probabilistic reliability index for a quantiEed measure infinite space, and the distance is measured in
of the safety margin based on the multi-ellipsoid convex Euclidean norms (L2 norms) or in L1 norms. In this
model description for grouped uncertain-but-bounded case, to seek the non-probabilistic reliability index in
XML Template (2013) [6.12.2013–12:22pm] [1–13]
//blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/JVCJ/Vol00000/130570/APPFile/SG-JVCJ130570.3d (JVC) [PREPRINTER stage]

Xu 3

the normalized infinite space can be solved by posing it where  

Q11 Q12
as an unconstrained multivariate nonlinear optimiza- Q¼
Q21 Q22
tion problem. Jiang et al. (2007) used a semi-analytical
method to search for the non-probabilistic reliability is a orthogonal matrix composed of eigenvectors;
index. Jiang et al. (2013) developed two efficient algo-  
1 0
rithms named first order approximation method ,¼
0 2
(FOAM) and second order approximation method
(SOAM) for the analysis of the reliability index. is a diagonal matrix made of eigenvalues.
In this paper, the unconstrained multivariate nonlinear Assume that
optimization problem is solved by an analytical method
based on the nature of the problem, i.e., the analysis q ¼ ð1="e Þ,1=2 Qd ð5Þ
method is directly adopted to seek the non-
probabilistic reliability index according to the impli- and substituting equation (5) into equation (2) yields
cit form of the limit state function. In some degree, n pffiffiffiffiffiffiffiffi o
it can simplify the corresponding calculation

Ec  q qT q  1 ðor kq
2  1Þ ð6Þ
and reduce the computational time. In this section,
the dynamic and static non-probabilistic reliability
indices corresponding to natural frequency constraint where
and displacement constraint are respectively analyzed. " 1=2 #
   
q1 1 1 0 Q11 Q12 E
q¼ ¼
2.1. Normalization of uncertain parameters q2 "e 01=2
2
Q21 Q22 
8 9
< 1=2 1=2 =
The uncertain structural parameters considered here 1 1 Q11 E þ 1 Q12 
are the elastic modulus and the mass density, which ¼ ð7Þ
"e : 1=2 Q  þ 1=2 Q  ;
can be given by 2 21 E 2 22 

 
E ¼ E^ ð1 þ E Þ,  ¼ ^ 1 þ  ð1Þ Commonly, the non-probabilistic measure in the
case of a circle model with two uncertain parameters
where E^ and ^ are the nominal values of uncertain elas- in the two-dimensional q-space is shown in Figure 1,
tic modulus E and mass density , respectively; E and where the circle
 represents
 the normalized convex set
 are the relative variations of the elastic modulus and q ¼ q1 , q2  qT q  1 .
the mass density, respectively. gðqÞ ¼ 0 is a limit state equation which divides
Then the ellipse convex model is used to describe the the standard q  space into  two parts, i.e., failure
uncertainty of structural parameters, i.e. region  ¼ q  gðqÞ 5 0 and safe region s ¼
  f 
n  o q gðqÞ 4 0 . Based on L2 -norm (or Euclidean norm)
x2E X^ ðI þ dÞdT Wd  "2 ð2Þ for the length measure, the minimal distance from the
e
origin to the limit state surface can be expressed

where

       
E ^ 1 E
x¼ ^¼ E ,
x I¼ , d¼ ð3Þ
 ^ 1 

and W and "e are a dimensionless characteristic

matrix and a positive parameter, respectively; W
determines the principal axis direction of the ellipse;
the size parameter of the ellipse is defined by "e , i.e.,
the uncertainty degree of the uncertain parameter
vector.
The eigenvalue decomposition of W can be deter-
mined by

QT WQ ¼ ,, QT Q ¼ I ð4Þ Figure 1. Non-probabilistic reliability measure in case of a

circle model.
XML Template (2013) [6.12.2013–12:22pm] [1–13]
//blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/JVCJ/Vol00000/130570/APPFile/SG-JVCJ130570.3d (JVC) [PREPRINTER stage]

4 Journal of Vibration and Control 0(0)

pffiffiffiffiffiffiffiffi
After deriving #E and # from equation (7) and sub-
as dmin ¼ min qT q . dmin ¼ 1 denotes that the
q:gðqÞ¼0 stituting them into equation (13), equation (13) can be
structure lies in a critical failure state. dmin 4 1 means rewritten as
that all possible values of uncertain parameters lie in
the safe region and the structure has a safety margin 1 þ a1 q#1 þ b1 q#2
accordingly. Therefore, dmin can be used as a non- ¼ c0 ð14Þ
1 þ c1 q#1 þ d1 q#2
probabilistic reliability measure in the case of a circle
model, i.e., a non-probabilistic reliability index is given by where
npffiffiffiffiffiffiffiffio
 ¼ Signðgð0ÞÞ: min qT q "e 1=2
q ð8Þ 2 Q21
a1 ¼ 1=2 1=2
,
s:t: gðqÞ ¼ 0 1 2 ðQ12 Q21  Q11 Q22 Þ
ð15Þ
"e 1=2
1 Q11
where SignðxÞ is the signum function of the real number b1 ¼  1=2 1=2
1 2 ðQ12 Q21  Q11 Q22 Þ
x defined by
8
<1 if x 4 0 "e 1=2
2 Q22
SignðxÞ ¼ 0 if x ¼ 0 ð9Þ c1 ¼ ,
: 1=2 1=2
1 2 ðQ11 Q22  Q21 Q12 Þ
1 if x 5 0 ð16Þ
"e 1=2
1 Q12
d1 ¼ 
1=2 1=2
1 2 ðQ11 Q22  Q21 Q12 Þ
2.2. Dynamic non-probabilistic reliability
For non-probabilistic reliability index analysis cor-
index analysis responding to natural frequency constraint, an aug-
One of the frequency constraints can be written as mented Lagrangian function is constructed as

fj  fj  0 ð10Þ h 2  2 i
L¼ q#1 þ q#2
    ð17Þ
where fj and fj are the jth natural frequency and the þ l 1 þ a1 q#1 þ b1 q#2  c0 1 þ c1 q#1 þ d1 q#2
corresponding allowable upper limit, respectively.
Then for the case without the nodal mass, in the
critical failure domain there follows where l is a Lagrangian multiplier.
h Minimizing the augmented Lagrangian function in
  2  i
equation (17) results
uTj 1 þ #E K#  2fj 1 þ # M# uj ¼ 0 ð11Þ

c0  1
where uj is the jth eigenvector corresponding to fj ; K# q#1 min ¼ ð18Þ
a1 þ b1 e1  c0 c1  c0 d1 e1
and M# are respectively structural deterministic stiff-
ness matrix and mass matrix, which can be written as
c0  1
q#2 min ¼ e1 ð19Þ
X
Ne X
Ne a1 þ b1 e1  c0 c1  c0 d1 e1
# #
K ¼ Ki , M ¼ Mi ð12Þ
i¼1 i¼1 c0 d1  b1
e1 ¼ ð20Þ
c 0 c 1  a1
where Ki and Mi are the deterministic stiffness matrix
and mass stiffness of the ith element, respectively; Ne is Similarly, for the case with the nodal mass there is, in
the number of structural elements. critical failure domain,
Based on structural dynamics, there can be obtained
h   2  i
2   3" # uTj 1 þ #E K#  2fj 1 þ # M# þ M0 uj ¼ 0
E^ 1 þ #E E^ fj2
4 5 ¼ 2 ¼ c0 ð13Þ
^ fj0 ð21Þ
^ 1 þ #

where fj0 is the jth natural frequency of the determinis- where M0 is the additional mass matrix resulted from
tic structure. the nodal mass.
XML Template (2013) [6.12.2013–12:22pm] [1–13]
//blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/JVCJ/Vol00000/130570/APPFile/SG-JVCJ130570.3d (JVC) [PREPRINTER stage]

Xu 5

Then the corresponding augmented Lagrangian The displacement constraint can be written as
function is given by
h 2  2 i juk max j  u ð27Þ
Lm ¼ q#1 þ q#2
n h   2  i o where uk max and u are the maximal displacement and
þ l uTj 1 þ #E K#  2fj 1 þ # M# þ M0 uj the allowable upper limit corresponding to the kth
ð22Þ degree-of-freedom, respectively.
Then in the critical failure domain the follow-
Minimizing the augmented Lagrangian function of ing equation can be obtained for displacement
equation (22) yields constraint as
 
2q#1 u0k  u 1 þ a1 q#1 þ b1 q#2 ¼ 0 ð28Þ
l¼  2  ð23Þ
T #  #
uj a1 K  2fj c1 M uj
where u0k is the displacement corresponding to the kth
degree-of-freedom of the deterministic structure.
q#2 ¼ g1 q#1 ð24Þ Similar to the non-probabilistic reliability index ana-
lysis for natural frequency constraint, there can be
where obtained
   
 2  a1 u0k  u b1 u0k  u
T #  #
uj b1 K  2fj d1 M uj q#1 min ¼  2 , q#2 min ¼  2  ð29Þ
u a1 þ b21 u a1 þ b21
g1 ¼  2  ð25Þ
uT a1 K#  2f c1 M# u
j j j
However for the static stress constraint, the
stress value of structural element does not depend
on the nominal value of the elastic modulus under
Then for the case with the nodal mass, the calcula- given cross-sectional areas of all truss elements
tion procedure of the dynamic non-probabilistic reli- (Appendix A).
ability index for frequency constraint is presented
as follows:
3. Dynamic non-probabilistic reliability-
Step 1: To calculate the jth eigenvector uð0Þ j when E ¼ E
^
and  ¼ ; ^
based topology optimization of truss
Step 2: To calculate qð0Þ 1 based on equation (18) and The deterministic structural optimization is essentially
qð0Þ ð0Þ ð0Þ
2 ¼ g1 q1 ; a single-loop optimization problem.
Step 3: To compute uðNÞ j and fj ðNÞ when In contrast, the non-deterministic structural
E ðN1Þ ¼ E^ 1 þ a1 qðN1Þ ðN1Þ
and ðN1Þ ¼ optimization is a double-loop optimization problem
1 þ  b1 q2 because some constraints are commonly applied to
ðN1Þ ðN1Þ
^ 1 þ c1 q1 þ d1 q2 ; the reliability index. In this paper, in order to solve
Step 4: To judge whether
 the convergence condition the latter optimization problem, a genetic algorithm
 
is satisfied, i.e., fj ðN1Þ  fj  fj  ". If not, go to (GA) is adopted in the outer loop and the analysis
Step 2. method is used to seek the non-probabilistic reliability
according to implicit forms of the limit state function
in the inner loop. The dynamic non-probabilistic
2.3. Static non-probabilistic reliability reliability-based topology optimization model is built
and the corresponding optimization algorithm in the
index analysis
outer loop is introduced in this section.
The relationship between the static displacement and
the external force can be expressed as
3.1. Optimization problem statement
The dynamic non-probabilistic reliability-based top-
KðEÞu ¼ F ð26Þ
ology optimization of the truss can be stated as follows:
for a given ground structure, find the nodal layout and
where K, u and F are the system stiffness matrix, the the cross-sectional areas so that the structural mass is
displacement vector and the external force vector, minimized under the condition that the non-
respectively. probabilistic reliability indices of displacement and
XML Template (2013) [6.12.2013–12:22pm] [1–13]
//blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/JVCJ/Vol00000/130570/APPFile/SG-JVCJ130570.3d (JVC) [PREPRINTER stage]

6 Journal of Vibration and Control 0(0)

natural frequency satisfy the constraints. The optimiza- frequency will approach zero (Appendix B).
tion problem can be formulated as Therefore, further effort should be made towards
nodal removal. A feasible method based on a GA is
introduced in this section.
X
Ne
min fðAÞ ¼ ^ i Ai Li
A 3.2.1. Code. In order to accelerate the optimization pro-
i¼1
  cess and improve the optimization efficiency, the nodal
s:t:  gu,j ðA, qÞ  u,j ðj ¼ 1, 2, . . . , Nu Þ
    ð30Þ topology variables are taken as another kind of design
 gf,k ðA, qÞ  f,k k ¼ 1, 2, . . . , Nf variable, i.e., when a node is removed, tp ¼ 0, or else,
  tp ¼ 1. The decimal value of the element cross-sectional
l ðA, qÞ   u ðl ¼ 1, 2, . . . , Ne Þ
i
area is also encoded into a gene represented by the
Al  0 ðl ¼ 1, 2, . . . , Ne Þ
binary code. Then all corresponding chromosomes are
joined together to represent a design variable vector as
where the objective
 function fðAÞTis the nominal struc- shown in Figure 2.
tural mass; A ¼ A1 , A2 , . . . , ANe is a vector composed
of element cross-sectional areas; q is a normalized 3.2.2. Fitness evaluation. An optimization problem with
uncertain vector denoted by a circle model; Li and Ai constraint functions is commonly transformed into an
are the length and the cross-sectional area of the ith unconstrained optimization problem by the penalty
element, respectively; gu,j and gf,k are the limit state function method. The individual fitness is directly
functions of the jth displacement constraint and the formed by the objective function and the penalty func-
kth natural frequency constraint, respectively; u,j tions for GA.
and f,k are the given lower limits of displacement Define constraint violation degrees for all kinds of
and natural frequency reliabilities, respectively; Nu constraints as
and Nf are the numbers of displacement and frequency
reliability constraints, respectively. 8  
> 0  gu,j ðA, qÞ  u,j
<  
CVDu,j ¼ u,j   gu,j ðA, qÞ  
3.2. Optimization algorithm >
:  gu,j ðA, qÞ 5 u,j
u,j
The "-relaxed method can be successfully applied to the
static topology optimization of truss with stress con- ð31Þ
straint, i.e., when the cross-sectional area of an element
in the optimization procedure approaches zero, a tiny 8  
> 0  gf,k ðA, qÞ  f,k
value " is assigned to it to keep the dimension of <  
the finite element (FE) model unchanged. However CVDf,k ¼ f,k   gf,k ðA, qÞ  
>
:  gf,k ðA, qÞ 5 f,k
for the topology optimization problem with frequency f,k
constraint, the feasible domains of the optimization
problems with different dimensions are disconnected. ð32Þ
The "-relaxed method may affect the evaluation of fre-  
8 l ðA, qÞ   u
quency constraint. In the "-relaxed method, the cross- <0   l
sectional areas of all elements connected to a removed CVDl ¼ lu  l ðA, qÞ  
l ðA, qÞ 4  u ð33Þ
node should be equal to ". Correspondingly, the truss : l
lu
will become a mechanism, i.e., the fundamental

Figure 2. Chromosome constructions of design variables (a) cross-sectional area chromosome; and (b) nodal topology
chromosome.
XML Template (2013) [6.12.2013–12:22pm] [1–13]
//blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/JVCJ/Vol00000/130570/APPFile/SG-JVCJ130570.3d (JVC) [PREPRINTER stage]

Xu 7

Then the individual fitness for minimizing the object- where fne ðAÞand fne þb ðAÞ denote the objective values of
ive function can be written as evolutionary generations ne and ne þ b, respectively;
1
evalðA, qÞ ¼ h P PN f PN e i ð34Þ
Nu ne
fðAÞ 1 þ ðpenal=2Þ  j¼1 CVD  u,j
þ k¼1 CVD f,k
þ l¼1 CVD l

where penal and ne are a penalty factor and the evo- e and n0 are a prescribed small value and a prescribed
lutionary generation, respectively. number, respectively.

3.2.3. Operator. Parents are selected according to their

4. Numerical examples
fitness. The better the chromosomes are, the more
chances to be selected they have. Roulette wheel selec- To demonstrate the validity of the proposed approach,
tion is applied to select parents into offspring. two numerical examples of the non-probabilistic relia-
Crossover and mutation are two basic operators bility-based topology optimization of truss structures
of GA. To a large extent, the performance of GA are given in this section. The cross rate, the mutation
depends on the crossover and the mutation. One rate and the maximal evolutionary generation number
point crossover selects genes from parent chromosomes are 0.8, 0.02 and 100, respectively. The upper and lower
and creates a new offspring, i.e., one crossover point is limits on cross-sectional areas are 129cm2 and 1cm2 ,
selected, the binary string from the beginning of the respectively. The population sizes are respectively 80
chromosome to the crossover point is copied from and 20 for 25-bar plane truss and 12-bar plane truss.
one parent, the rest is copied from the second parent. "e ¼ 0:1," ¼ 1  104 , n0 ¼ 10, penal ¼ 100.
After the crossover is performed, mutation takes place.
This is to prevent all solutions in the population from
falling into a local optimum. Mutation changes ran-
4.1. 25-bar plane truss
domly the new offspring. For binary encoding a few The geometric dimensions and loads of the ground
randomly chosen bits from 1 to 0 or from 0 to 1 can structure for a 25-bar plane truss are shown in
be switched, i.e., selected bits are inverted. Figure 3. The nominal values of the elastic modulus
and the mass density are 199815:8MPa and
3.2.4. Convergence criterion. The optimization steps 7850kg=m3 , respectively. P1 ¼ P3 ¼ 1779:2KN,
are repeated until the convergence criterion is P2 ¼ 2224KN, P4 ¼ 1334:4KN. The stress constraint
satisfied, where the convergence criterion can be requires that the stress of each element be less
expressed as than 172:37MPa. f1 ¼ 10Hz, u ¼ 3  102 m.The low
8  n þb  limit on the non-probabilistic reliability index for fun-
> f e ðAÞ  fne ðAÞ  " b ¼ 1, 2, . . . , n0 damental natural frequency constraint is f ¼ 1:0.
>
>  
<  gu,j ðA, qÞ   j ¼ 1, 2, . . . , Nu Three cases with different indices for the displacement
  u,j ð35Þ constraint are discussed, i.e., u ¼ 0:1, 0:2, 0:3.
>
>  gf,k ðA, qÞ  f,k k ¼ 1, 2, . . . , Nf 
>
:   4 1
l ðA, qÞ  lu l ¼ 1, 2, . . . , Ne W¼ .
1 1

Figure 3. Ground structure for topology optimization of 25-bar plane truss.

XML Template (2013) [6.12.2013–12:22pm] [1–13]
//blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/JVCJ/Vol00000/130570/APPFile/SG-JVCJ130570.3d (JVC) [PREPRINTER stage]

8 Journal of Vibration and Control 0(0)

Table 1. Optimal solutions of the 25-bar plane truss.

Cross-sectional area (cm2 )

No. bar Deterministic f ¼ 1:0, u ¼ 0:1 f ¼ 1:0, u ¼ 0:2 f ¼ 1:0, u ¼ 0:3

1 \ 52.891 36.226 19.578

2 \ 42.352 \ \
3 12.602 \ 14.539 116.109
4 \ \ 56.992 \
5 103.070 124.711 99.844 126.219
6 93.297 69.969 27.953 38.867
7 \ 32.773 \ 25.727
8 29.539 63.195 \ \
9 \ \ \ 19.992
10 36.109 40.406 94.938 116.258
11 81.219 80.031 102.734 126.695
12 78.961 122.164 95.563 117.867
13 96.094 86.266 98.281 106.734
14 105.844 114.086 123.266 128.477
15 92.469 127.656 116.398 113.109
16 17.648 \ 16.297 29.641
17 118.016 78.539 85.469 122.211
18 92.56 96.750 123.461 107.445
19 54.156 92.344 92.109 125.586
20 68.070 74.813 100.000 67.563
21 69.796 90.383 123.984 124.820
22 83.438 95.039 95.313 63.781
23 76.055 95.602 125.836 127.844
24 94.633 82.570 113.828 107.664
25 82.961 95.969 83.211 108.289
Total mass (kg) 21539 25012 26603 29291

4.2. 12-bar plane truss

The optimal solutions of the 25-bar plane truss are
shown in Table 1, where the second column is the deter- The topology optimization of a 12-bar plane truss with
ministic optimization solution and other columns are a2:5kg nodal mass, where the corresponding ground
the optimal solutions under non-probabilistic reliability structure is shown in Figure 8, is considered. The nom-
constraints. The optimal topology obtained by the inal values of the elastic modulus and the mass density
deterministic optimization is shown in Figure 4 and are 2:1  1011 Pa and 7800kg=m3 , respectively. A stress
some optimal topology layouts obtained by the present constraint ji j  172:37MPa ði ¼ 1, 2, . . . , 12Þ is
reliability-based approach are shown in Figures 5–7, imposed, where the subscript denotes the element
respectively. The numbers of remaining bars in optimal serial number. P1 ¼ P2 ¼ 1000KN,
 f1 ¼ 10Hz,
topologies under non-probabilistic reliability con- 2 10 0:5
u ¼ 5  10 m, W ¼ . Seven cases with
straints are larger than that in the deterministic opti- 0:5 20
mization topology. As can be seen from Table 1, the different specified lower limits u ¼ 0:3, 0:4, 0:5,
total mass obtained by the non-probabilistic reliability 0:6, 0:7 and f ¼ 1, u ¼ 0:5 and f ¼ 2, 5 are con-
optimization is larger than that corresponding to the sidered, respectively. The achieved optimal solutions
deterministic optimization and increases with the are summarized in Table 2, including the optimal
increase of the non-probabilistic reliability constraint total structural mass and non-probabilistic reliability
index. In addition, the optimal topology also changes indices. The deterministic optimization result is also
according to the non-probabilistic reliability constraint presented for comparison, where the elastic modulus
index.
XML Template (2013) [6.12.2013–12:22pm] [1–13]
//blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/JVCJ/Vol00000/130570/APPFile/SG-JVCJ130570.3d (JVC) [PREPRINTER stage]

Xu 9

Figure 4. Optimal layout of 25-bar plane truss obtained by deterministic optimization.

Figure 5. Optimal layout obtained by present reliability-based approach ðf ¼ 1:0, u ¼ 0:1Þ.

Figure 6. Optimal layout obtained by present reliability-based approach ðf ¼ 1:0, u ¼ 0:2Þ.

Figure 7. Optimal layout obtained by present reliability-based approach ðf ¼ 1:0, u ¼ 0:3Þ.

and the mass density take nominal values. As can be optimal mass increases with the increase of the limit
seen from Table 2, the optimal mass for the non- on the displacement non-probabilistic reliability index,
probabilistic reliability optimization is larger than where there is the same limit on the frequency reliability
that for the deterministic optimization. Also, the index. For the cases with the same limit on the
XML Template (2013) [6.12.2013–12:22pm] [1–13]
//blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/JVCJ/Vol00000/130570/APPFile/SG-JVCJ130570.3d (JVC) [PREPRINTER stage]

10 Journal of Vibration and Control 0(0)

displacement reliability index, the optimal structural layouts are shown in Figures 9 and 10 for the deterministic
mass increases with the increase of the limit on the fre- optimization and the non-deterministic optimization.
quency reliability index. From the viewpoint of reliabil- It is observed that the introduction of reliability require-
ity, the worst case must be taken into account in the ments into a topology optimization problem may result in
non-probabilistic reliability optimization. For the deter- different optimal layouts compared to conventional deter-
ministic optimization design, each non-probabilistic reli- ministic solutions. The number of remaining bars in opti-
ability index is zero. The higher the reliability mal topologies obtained by the present reliability-based
requirements are, i.e., the higher the performance require- optimization is seven or eight. This signifies that the opti-
ments are, the larger the optimal structural total mass will mal topology layout for the non-deterministic structural
be. Therefore, the uncertainties should be considered in optimization is statically indeterminate. Obviously, this
structural optimization so as to ensure structural safety feature aligns with the reliability principle.
under extreme loading conditions. The optimal topology

Figure 9. Optimal layout of 12-bar plane truss obtained by

deterministic optimization and present reliability-based opti-
Figure 8. Ground structure for topology optimization of mization ðf ¼ 1:0, u ¼ 0:3; f ¼ 1:0, u ¼ 0:4; f ¼ 1:0,
12-bar plane truss. u ¼ 0:5; f ¼ 1:0, u ¼ 0:6; f ¼ 1:0, u ¼ 0:7Þ.

Table 2. Optimal solutions of the 12-bar plane truss.

Cross-sectional area (cm2 )

f ¼ 1:0 f ¼ 1:0 f ¼ 1:0 f ¼ 1:0 f ¼ 1:0 f ¼ 2:0 f ¼ 5:0

No.bar Deterministic
u ¼ 0:3 u ¼ 0:4 u ¼ 0:5 u ¼ 0:6 u ¼ 0:7 u ¼ 0:5 u ¼ 0:5

1 23.204 27.018 72.144 59.528 57.501 99.163 37.658 59.034

2 21.302 20.009 22.594 26.662 51.385 48.945 48.822 40.665
3 40.992 63.849 62.832 111.670 115.875 60.820 122.420 116.442
4 \ \ \ \ \ \ \ \
5 \ 46.011 34.230 109.527 113.871 90.897 \ \
6 \ \ \ \ \ \ \ \
7 65.433 100.223 36.097 89.416 67.452 106.361 104.683 127.112
8 28.645 34.128 56.303 26.640 86.605 74.076 58.177 52.304
9 \ \ \ \ \ \ \ \
10 54.167 66.195 110.021 33.111 67.459 83.358 57.508 15.033
11 20.597 23.502 11.010 15.295 56.971 19.646 17.496 17.329
12 \ \ \ \ \ \ \ \
Total mass (kg) 1255.5 1571.0 1739.4 2053.7 2750.0 3232.3 2063.6 2213.8
f ¼ 0 f ¼ 2:0527 f ¼ 4:6849 f ¼ 1:0204 f ¼ 2:7610 f ¼ 2:0759 f ¼ 3:0371 f ¼ 6:1830
Non-probabilistic
reliability index u ¼ 0 u ¼ 0:3207 u ¼ 0:4009 u ¼ 0:5277 u ¼ 0:6093 u ¼ 0:7068 u ¼ 0:5196 u ¼ 0:5079
XML Template (2013) [6.12.2013–12:22pm] [1–13]
//blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/JVCJ/Vol00000/130570/APPFile/SG-JVCJ130570.3d (JVC) [PREPRINTER stage]

Xu 11

Fund for Doctoral Program of Higher Education (grant

number 20096102120024) and the Fund for Basic Research
from Northwestern Polytechnical University (grant number
JC20120232).

References
Jiang C, Bi RG, Lu GY, et al. (2013) Structural reliabil-
ity analysis using non-probabilistic convex model.
Computer Methods in Applied Mechanics and Engineering
254: 83–98.
Jiang T, Chen JJ and Xu YL (2007) A semi-analytic method
for calculating non-probabilistic reliability index based on
Figure 10. Optimal layout of 12-bar plane truss obtained by interval models. Applied Mathematical Modelling 31(7):
present reliability-based optimization ðf ¼ 2:0, u ¼ 0:5; 1352–1370.
f ¼ 5:0, u ¼ 0:5Þ. Kang Z, Luo YJ and Li A (2011) On non-probabilistic relia-
bility-based design optimization of structures with uncer-
5. Conclusions tain-but-bounded parameters. Structural Safety 33(3):
196–205.
1. A measure of dynamic and static non-probabilistic Li D, Jiang C, Han X, et al. (2011) An efficient optimization
reliabilities is analyzed based on the ellipse convex method for uncertain problems based on non-probabilistic
model with uncertain-but-bounded parameters. interval model. International Journal of Computational
Methods 8(4): 837–850.
Specifically, for the topology optimization of a
Lombardi M (1998) Optimization of uncertain structures
truss with nodal mass, the calculation procedures
using non-probabilistic models. Computers & Structures
of the dynamic non-probabilistic reliability index 67(1–3): 99–103.
for frequency constraint is also presented. Luo YJ, Kang Z, Luo Z, et al. (2009) Continuum topology
2. Based on the nested nature of the non-probabilistic optimization with non-probabilistic reliability constraints
reliability-based topology optimization and the dis- based on multi-ellipsoid convex model. Structural and
connected feasible domain for frequency constraint, Multidisciplinary Optimization 39(3): 297–310.
a GA is adopted in the outer loop and the analysis Norapat N and Sujin B (2011) Simultaneous topology, shape
method is used to seek the non-probabilistic reliabil- and sizing optimization of a three-dimensional slender
ity index, based on the implicit form of the limit state truss tower using multiobjective evolutionary algorithms.
function in the inner loop. Computers & Structures 89(23–24): 2531–2538.
3. It is well recognized that uncertainties should Ohsaki M, Fujisawa K, Katoh N, et al. (1999) Semi-definite
be taken into account for a more rational struc- programming for topology optimization of trusses under
multiple eigenvalue constraints. Computer Methods in
tural design. The optimal mass for non-deterministic
Applied Mechanics and Engineering 180(1–2): 203–217.
structural optimization problems is obviously larger
Qiu Z and Wang J (2010) Reliability study of fracture mech-
than that for the deterministic optimization problem. anics based non-probabilistic interval analysis model.
The larger the non-probabilistic reliability require- Fatigue & Fracture of Engineering Materials & Structures
ment is, the larger the optimal structural mass will be. 33(9): 539–548.
4. Though valuable work has been done in the prob- Qiu ZP and Wang XJ (2003) Comparison of dynamic
abilistic framework, system reliability analysis is a response of structures with uncertain-but-bounded param-
more challenging task and the corresponding non- eters using non-probabilistic interval analysis method and
deterministic structural optimization problem has probabilistic approach. International Journal of Solids and
been studied in the present paper. The paper has Structures 20(40): 5423–5439.
focused on the pure convex set model of uncertainty. Tong WH and Liu GR (2001) An optimization procedure for
However in a real-life engineering design problem, truss structures with discrete design variables and dynamic
structures may exhibit both probabilistic and constraints. Computers & Structures 79(2): 155–162.
Tong WH, Jiang JS and Liu GR (2000) Solution existence of
bounded uncertainties and the corresponding prob-
the optimization problem of truss structures with fre-
ability and convex set mixed model will require fur-
quency constraints. International Journal of Solids and
ther research in the future. Structures 37(30): 4043–4060.
Wang D, Zhang WH and Jiang JS (2004) Truss optimization
on shape and sizing with frequency constraints. AIAA
Funding Journal 42(3): 622–630.
This work was supported by the Natural Science Foundation Xu B and Jin YJ (2012) Multiobjective dynamic topology
of China (grant number 11072197), the Specialized Research optimization of truss with interval parameters based on
XML Template (2013) [6.12.2013–12:22pm] [1–13]
//blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/JVCJ/Vol00000/130570/APPFile/SG-JVCJ130570.3d (JVC) [PREPRINTER stage]

12 Journal of Vibration and Control 0(0)

interval possibility degree. Journal of Vibration and where

Control DOI: 10.1177/1077546312456725.
x2  x1 y2  y1 z2  z1
Xu B, Jiang JS, Tong WH, et al. (2003) Topology group ¼ , ¼ , ¼ ðA4Þ
concept for truss topology optimization with frequency lAB lAB lAB
constraint. Journal of Sound and Vibration 261(5):
After the action of static loading, the coordinate
911–925.
values of node A vary from ðx1 , y1 , z1 Þ to
ðx1 þ u1 , y1 þ v1 , z1 þ w1 Þ, i.e.,
Appendix A dx1 ¼ u1 , dy1 ¼ v1 , dz1 ¼ w1 ðA5Þ
The nodal coordinates of an element before and after
the action of static loading are shown in Figure A1. Similarly, the following can be obtained
Correspondingly, the length of element AB can be
obtained dx2 ¼ u2 , dy2 ¼ v2 , dz2 ¼ w2 ðA6Þ

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Then the strain of the truss element is written as

lAB ¼ ðx2  x1 Þ2 þðy2  y1 Þ2 þðz2  z1 Þ2 ðA1Þ
dlAB 
"AB ¼ ¼ ðu1 þ u2 Þ
lAB lAB
ðA7Þ
where lAB is the length of a truss element; x1 , y1 and 
z1 are the coordinate values of node A in x, y and þ ðv1 þ v2 Þ þ ðw1 þ w2 Þ
lAB lAB
z directions, respectively; x2 , y2 and z2 are the coord-
inate values of node B in x, y and z directions, The stress of the truss element can be expressed as
respectively.
Implementing the differential operation on both AB ¼ E"AB
sides of equation (A1), there follows " 
#

lAB ðu1 þ u2 Þ þ lAB ðv1 þ v2 Þ
¼E ðA8Þ
þ lAB ðw1 þ w2 Þ
lAB dlAB ¼ ðx2  x1 Þðdx2  dx1 Þ
E
þ ðy2  y1 Þðdy2  dy1 Þ þ ðz2  z1 Þðdz2  dz1 Þ ¼ ½      de
lAB
ðA2Þ
However, each component of the vector de can be
calculated by the following equation
Further, equation (A2) can be rewritten as
1X n
i1 ¼ Si1j Fj ðA9Þ
dlAB ¼ ðdx2  dx1 Þ þ ðdy2  dy1 Þ þ ðdz2  dz1 Þ E j¼1
ðA3Þ
where Si1j is the element of system flexibility matrix
which lies in the i1th row and the jth column when
the value of the elastic modulus is equal to one.
Substituting equation (A9) into equation (A8) yields
2 n 3
P
6 j¼1 Si1j Fj 7
6 7
6P n 7
6 S F 7
6 i2j j 7
6 j¼1 7
6P 7
6 n 7
6 Si3j Fj 7
1 6 j¼1 7
AB ¼ ½      6 6 Pn
7 ðA10Þ
7
lAB 6 Si4j Fj 7
6 j¼1 7
6 7
6P n 7
6 S F 7
6 i5j j 7
6 j¼1 7
6P 7
4 n 5
Figure A1. Nodal coordinates of an element before and after Si6j Fj
j¼1
the action of static loading.
XML Template (2013) [6.12.2013–12:22pm] [1–13]
//blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/JVCJ/Vol00000/130570/APPFile/SG-JVCJ130570.3d (JVC) [PREPRINTER stage]

Xu 13

As can be seen from equation (A10), the stress value Further, there can be obtained
of the truss element does not depend on the elastic
modulus. SX ¼ l~ X ðB4Þ

where X is the Fourier transformation of X, and

Appendix B
2 3
When a node with mass is removed in topology s11 s12    s1n
optimization, the cross-sectional areas of all elements 6s s22    s2n 7
6 21 7
connected to this node should be equal to " in the S ¼ M1 K ¼ 6
6 .. ... . .. 7 7
"-relaxed method. However, the nodal mass has a 4 . . . . 5
relatively large value and the truss will become a mech- sn1 sn2    snn
anism, i.e., the fundamental frequency will approach 2 3
k11 =m11 k12 =m11    k1n =m11
zero. 6 k =m
6 21 22 k22 =m22    k2n =m22 7
7
Proof: when the dynamic displacement vector is
¼6
6 .. ... . .. 7
7
denoted by X ¼ ðu1 , v1 , w1 , . . . , uj , vj , wj , . . . , un=3 , 4 . . . . 5
vn=3 , wn=3 ÞT , where uj , vj and wj are respectively the dis-
kn1 =mnn kn2 =mnn    knn =mnn
placements of node j in x, y and z directions, the system
stiffness matrix K of a space truss can be written as ðB5Þ

2 3 where m11 ¼ ms11 þ ma11 .

k11 k12 ... k1n
Assume that the removed node in the optimization
6 k21 k22  k2n 7
K¼6
4 ... .. .. .. 7 ðB1Þ process is node j, the cross-sectional areas of all elem-
. . . 5 ents connected to this node are assigned a tiny value
kn1 kn2  knn " ! 0. The components of the system matrix S related
to node j shp ! 0 or sph ! 0ðh ¼ 3j  2, 3j  1, 3j Þ.
where n is the number of degrees-of-freedom. Based on the Gerschgorin theorem, the following for-
The global mass matrix M can be expressed as mula can be derived
 
2 3 ~  0
ms11 0 ... 0 lðSÞ  shh   jSh j1 ðB6Þ
6 0 ms22  0 7
6 7
M ¼ Ms þ Ma ¼ 6
6 .. .. .. .. 7
7
0

4 . . . . 5 where Sh ¼ ðsh1 , . . . , sh,h1 , sh,hþ1 , . . . , shn ÞT ; k: k1

denotes 1-norm.; l~ ðSÞ is the eigenvalue of system
0 0    msnn matrix S.
2 3 ðB2Þ
ma11 0 ... 0 Then equation (B6) can be rewritten as
6 0 ma22  0 7
6 7  
þ6
6 .. .. .. .. 77
0
0 5 l~ ðSÞ  shh þ kSh k1 ðB7Þ
4 . . . . 5
0 0    mann However, there is the formula
0
kSh k1 ! 0 ðB8Þ
where M, Ms and Ma are the system mass matrix
and the mass matrix without nodal mass, and the Therefore, the following equation is found to be true
additional mass matrix induced by nodal mass,
respectively. l~ ðSÞ ! 0 ðB9Þ
Then the system free vibration equation can be writ-
ten as

MX€ þ KX_ ¼ 0 ðB3Þ