Truss Shape Optimization With Multiple Displacement Constraints

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

Comput. Methods Appl. Mech. Engrg.

191 (2002) 3597–3612


www.elsevier.com/locate/cma

Truss shape optimization with multiple


displacement constraints
D. Wang a, W.H. Zhang a, J.S. Jiang b,*

a
Sino-French Laboratory of Concurrent Engineering, Department of Aircraft Manufacturing Engineering,
Northwestern Polytechnical University, Xi’an, Shaanxi 710072, China
b
Institute of Vibration Engineering, Northwestern Polytechnical University, Xi’an, Shaanxi 710072, China
Received 9 January 2002; received in revised form 8 March 2002

Abstract

This paper presents an evolutionary node shift method for truss shape optimization of weight minimization prob-
lems. The structure is subject to multiple displacement constraints under multiple load cases. Nodal coordinates are the
design variables. Based on the sensitivity analysis, the optimum shape is achieved gradually from the initially under-
designed configuration by shifting the most efficient nodes with minimum weight increase. Moving directions and in-
tervals of node shifts are determined automatically. The optimum solution is checked by the Kuhn–Tucker optimality
conditions. Limitations of truss shape optimization are studied. The feasibility and effectiveness of the method are
investigated with four typical truss structures.  2002 Elsevier Science B.V. All rights reserved.

Keywords: Truss shape optimization; Node shift; Sensitivity analysis; Multiple displacement constraints

1. Introduction

The layout optimization of a truss structure means the simultaneous optimization of its topology, ge-
ometry (shape) and sizing [1]. The topology and sizing optimization of a truss structure with multiple
displacement constraints have been extensively studied by a variety of methods. Significant progress
has been made. It is widely recognized that an optimal geometry of a structure can greatly improve the
structural performance, e.g. the potential material savings versus the displacement constraints. However,
shape optimization of truss structures subject to multiple displacement constraints under multiple load
cases has received limited attention. Many valuable efforts have been made for truss sizing optimizations
with a fixed configuration, which is usually designed according to the designer’s experiences. If the truss
sizing optimization is performed on a non-optimal geometry, the optimization result may be quite limited.
In contrast, as the sizing optimization is performed on an optimal configuration, great benefits can
be achieved for the design task. Now, the common strategies for truss shape optimization are the

*
Corresponding author. Tel.: +86-29-849-2895.
E-mail address: jiangjs@nwpu.edu.cn (J.S. Jiang).

0045-7825/02/$ - see front matter  2002 Elsevier Science B.V. All rights reserved.
PII: S 0 0 4 5 - 7 8 2 5 ( 0 2 ) 0 0 2 9 7 - 9
3598 D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612

mathematical programming techniques such as the SUMT [2], sequential linear programming [3] or convex
programming methods [4].
This paper is devoted to presentation of a simple and effective method for truss shape optimization. The
Kuhn–Tucker (K–T) optimality conditions are employed to check the obtained result. In addition, the
limitation of the truss shape optimization will be discussed briefly.
The evolutionary node shift method is proposed to perform the truss shape optimization of minimum
weight design. The strategy is following the ideas of evolutionary structure optimization (ESO) method
originated by Xie and Steven [5]. Instead of removing elements from the structure, node (joint) positions are
shifted evolutionarily by means of sensitivity analysis. The process starts from an initially under-designed
configuration, i.e. an unfeasible shape design. Nodes that decrease the displacements maximally while in-
crease the weight minimally are shifted firstly. In each iterative loop, directions and intervals of node shifts
are determined automatically. The optimum shape is achieved gradually until all of the displacement
constraints are satisfied.

2. Problem formulations

In truss shape optimization problems, the topology of the structure and the bar cross-sectional areas are
assumed to be unchangeable. Nodal coordinates are the design variables. For instance, given a truss with a
force P illustrated in solid lines in Fig. 1, the vertical displacement component v3 of node 3 is required to
satisfy the constraint: v3 6 v3 . By intuition, we may try to shift the node 1 to 10 vertically, and then adjust the
length of the element connected to it for compatibility. It is clear that by doing so, the truss shape is
modified and the related element weight is increased at the same time. In other words, such a node shift
would not only reduce the specified displacement, but also change the structural weight simultaneously. The
proposed truss shape optimization approach is devised to decrease the specified displacements maximally
whereas increase the structural weight minimally. To perform the node shifts, the method should determine:
which node, in which direction, with how large interval, would be shifted firstly and subsequently?
The optimum shape of a truss structure implies that all the specified nodal displacements under multiple
load cases satisfy the prescribed constraints while the structural weight is minimum. Therefore, the truss
shape optimization of weight minimization problem can be formulated as
X n
Minimize W ¼ Le qe Ae ; ð1Þ
e¼1

uil 6 ui ðfor i ¼ 1; m; l ¼ 1; pÞ


Subject to Xj 6 Xj 6 Xj ðfor j ¼ 1; kÞ; ð2Þ

Fig. 1. A simple truss.


D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612 3599

where W is the structural weight, Le , qe and Ae are the length, material density and cross-sectional area of
the eth element, respectively. n denotes the total number of elements, of which the structure consists. uil is
the ith nodal displacement under lth load case and ui the corresponding limit for all load cases, respectively.
m and p are the total numbers of the constrained displacements and the load cases, respectively. X j and Xj
represent the lower and upper bounds on the jth node coordinates, respectively. k is the total number of the
side constraints.

3. Sensitivity analyses

3.1. Effects of node shifts on a displacement

First, the sensitivity number of a specific nodal displacement with respect to a node shift is derived. For
simplicity of presentation but without loss of generality, we assume that the jth node is shifted along x-axis.
The force equilibrium equation of a loaded truss structure is

½Kfug ¼ fP g; ð3Þ

where [K] is the global stiffness matrix, fP g the load vector and fug the unknown nodal displacement
vector, respectively. The change of the global stiffness matrix [DK] with respect to the jth node shift is
expressed with linear approximation [6]

X
nj Xnj
o½ke 
½DK ¼ ½Dke  ¼ Dxj ; ð4Þ
e¼1 e¼1
oxj

where nj is the number of elements connected to the jth node and [Dke ] is the change of the related element
stiffness matrix. Dxj is the interval of the jth node shift. The force vector fP g is commonly assumed un-
changeable during the optimization process, and the change of nodal displacement vector fDug is evaluated
approximately [5]

fDug ¼ ½K 1 ½DKfug: ð5Þ


T
Premultiplying Eq. (5) by a virtual unit load vector fF i g , in which, only the term corresponding to the ith
displacement is equal to a unit and the others are zeroes, then the change of the ith displacement Dui is
expressed as
T 1
Dui ¼ fF i g ½K ½DKfug
T
ð6Þ
¼ fui g ½DKfug;

where fui g is the virtual nodal displacement vector caused by fF i g. The change of the ith displacement with
respect to the jth node shift Duij is evaluated by substituting Eq. (4) into Eq. (6). This gives
!
X
nj
i T
Duij ¼ fu g ½Dke  fug
e¼1
nj   ð7Þ
X  i T o½ke 
¼ ue fue g Dxj ðj ¼ 1; kÞ:
e¼1
oxj
3600 D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612

In the above equation, fuie g and fue g are the nodal displacement vectors of the related element caused by
virtual and real forces, respectively. From formula (7), it is seen evidently that Duij can be calculated at the
element level around the jth node. Additionally, the most effective node for changing the ith displacement ui
can be identified easily. It is worth noting that Duij may be either positive or negative, which implies that ui
may increase or decrease with respect to different node shift schemes. To reduce the specified displacement
ui towards the limit ui , one requires
Duij < 0 ðj ¼ 1; kÞ: ð8Þ
Therefore, the direction along which the jth node is shifted can be determined from formulae (7) and (8)
X nj  !
 i T o½ke 
signðDxj Þ ¼ sign ue fue g ðj ¼ 1; kÞ: ð9Þ
e¼1
oxj

However, it should be noted that the most effective node for displacement reduction does not mean the
most efficient one for weight minimization problems. When the jth node is shifted, the elements connected
to it may change in length, which therefore leads to the change of structural weight
X
nj X
nj
oLe
DWj ¼ Dwe ¼ qe Ae Dxj ; ð10Þ
e¼1 e¼1
oxj
where DWj and Dwe represent the changes of the structural and element weights due to the jth node shift,
respectively.
In the optimization problem, the structural weight is referred to as the objective while the displacement is
subject to constraint. Therefore, the sensitivity number is defined for each shifted node as [6]
 
Pnj  i T o½ke 
e¼1 ue fue g
Duij oxj
aij ¼ ¼ ðj ¼ 1; kÞ; ð11Þ
DWj Pnj oLe
q A
e¼1 e e
oxj
aij defined in the above indicates the ratio of the effect on the ith displacement to the effect on the structural
weight due to the jth node shift. If there are two different node shift schemes producing the same reduction
for a certain displacement, it is more efficient to shift the node with the minimum weight increment. On
the other hand, for the same weight increment, such a node shift scheme with the largest reduction of the
displacement is preferred to those with smaller ones. From this viewpoint, to obtain the minimum weight
design, among all the shifted nodes we should identify and shift the most efficient one, which has the
smallest (most negative) value of sensitivity number
minfaij ; j ¼ 1; kg: ð12Þ
Sometimes, a node shift would introduce reductions in both the ith displacement and the structural
weight simultaneously. The values Duij and DWj are all negative so that
Duij
aij ¼ > 0: ð13Þ
DWj
In this case, the most efficient node may be the one with the smallest positive value of aij since it can render
the largest weight reduction for a certain displacement reduction. However, we would prefer to shift all the
nodes with positive sensitivity numbers for reducing both the displacement and structural weight. There-
fore, the strategy to implement the truss shape optimization procedure is

• shifting preferentially the nodes with positive sensitivity numbers and then node with the most negative
sensitivity number.
D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612 3601

3.2. Multiple displacement constraints in one load case

In practical design, we usually have to consider multiple displacement constraints in one load case and
the Lagrange multiplier technique is often used for such situation [7]. However, the main difficulty of this
technique is the calculation of the Lagrange multipliers [7]. In this paper, we adopt a simpler approach by
considering all the displacement constraints as integrity and utilizing the weighted sum of the corresponding
sensitivity numbers [5]
Xm
gj ¼ ki aij ðj ¼ 1; kÞ; ð14Þ
i¼1

where gj is defined as the global sensitivity number in the case of multiple displacement constraints with
respect to the jth node shift and ki is the weighting coefficient. aij is determined by Eq. (11) for each dis-
placement constraint. Eq. (14) should represent such a fact that the more seriously an individual constraint
is violated, the more it contributes to the global sensitivity numbers. Xie and Steven [5] have proposed the
following relation:
 b
ui
ki ¼ : ð15Þ
ui
If a displacement constraint is violated greatly, it is possible that the displacement may be reduced
primarily in the current loop of shape optimization. Contrarily, if a displacement is far below the prescribed
limit, the corresponding constraint will be of little significance to gj . The exponent b is a penalizing factor. A
large value of b will make greater the weighting coefficient ki corresponding to the violated constraint and
smaller the one corresponding to the satisfied constraint. A typical value of the exponent b ¼ 3 is used in
our examples.
Substituting Eqs. (11) and (15) into Eq. (14), gj is calculated by
!
Xm  b
1 X
nj
ui  i T
gj ¼ u ½Dke fue g
i¼1
ui DWj e¼1 e
!T ð16Þ
m  b
Dxj X X
nj
ui  i  o½ke 
¼ ue fue g ðj ¼ 1; kÞ;
DWj e¼1 i¼1 ui oxj
where DWj is calculated with Eq. (10).
It should be pointed out that the advantage of introducing global sensitivity number gj is that it delivers
a unified approach to handle problems with single or multiple displacement constraints. Besides, the most
efficient node shift is always the one with the smallest value of gj .

3.3. Multiple displacement constraints under multiple load cases

When a truss structure is subject to multiple displacement constraints under multiple load cases, the
sensitivity number of the ith displacement with respect to the jth node shift in the lth load case is defined as
!
Dulij 1 Xnj
 i T  l
l
aij ¼ ¼ u ½Dke  ue ðj ¼ 1; k; l ¼ 1; pÞ; ð17Þ
DWj DWj e¼1 e
where fule g is the element displacement vector under the lth load case fP l g. According to Ref. [8], the global
sensitivity numbers for m displacement constraints under p load cases are defined as
X p X m
gj ¼ kli alij ðj ¼ 1; kÞ: ð18Þ
l¼1 i¼1
3602 D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612

The weighting coefficient kli takes the exact form of Eq. (15) and is defined as
 b
l uil
ki ¼ : ð19Þ
ui
Substituting Eqs. (17) and (19) into Eq. (18) yields
!
X p X m  b
1 X
nj
uil i T
 l
gj ¼ fu g ½Dke  ue ;
l¼1 i¼1
ui DWj e¼1 e
0 !T 1 ð20Þ
m  b
Dxj X X X
nj p
u o½k   
¼ @ il
fuie g
e
ule A ðj ¼ 1; kÞ:

DWj e¼1 l¼1 i¼1
u i ox j

To reduce computational costs of finite element reanalysis and sensitivity number calculations, more
than one node with the smallest sensitivity numbers are shifted in each optimization loop. The evolutionary
process is performed until all of the displacement constraints are satisfied.

3.4. Kuhn–Tucker optimality conditions

In general, the method can offer an optimal result. To ensure the weight minimum design, the K–T
optimality conditions are employed to check the solution. For a single displacement constraint, the Lag-
rangian function is defined as
LðX Þ ¼ W ðX Þ þ kðui ui Þ: ð21Þ
In which X is a vector representing the design variables and k is the Lagrange multiplier.
At the optimum, the constraint is active and the design is dominated by K–T conditions
8
< rX LðX Þ ¼ rX W ðX Þ þ krX ðui ui Þ ¼ 0;
u ui ¼ 0; ð22Þ
: i
k P 0:

For constrained optimization problems, it is known that oW =oxj 6¼ 0. From the first expression of Eq.
(22), one gets
oui
1 oxj oui
¼ ¼ ¼ aij ðj ¼ 1; kÞ: ð23Þ
k oW oW
oxj
Eq. (23) denotes that at the optimum, all the sensitivity numbers ought to be a negative constant. It is seen
that the proposed strategy to implement the optimization process can guarantee that k is positive.
However, there may be difficulty to determine the optimum design via aij in Eq. (23) when there are
many design variables in the problem. An alternative checking condition is suggested by means of the
gradient of the Lagrange function.
T
Premultiplying the first expression of Eq. (22) by ðrui Þ yields
T T
k¼ ðrui Þ rW =ðrui Þ ðrui Þ: ð24aÞ
If the gradient satisfies
jrX LðX Þj 6 10 2 : ð24bÞ
D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612 3603

The design point is believed to be convergent and the process may be terminated. Example 6.1 gives the
comparisons of the two convergent conditions in detail.
For multiple displacement constraints, ru is calculated approximately according to Eq. (20)
0 !T 1
Xnj Xp X m  b
u  i  l
ru ¼ @ il
ue r½ke  ue A: ð25Þ
e¼1 l¼1 i¼1
ui

If the convergent condition is not satisfied, the design point is moved in the opposite direction of the
gradient of the Lagrange function (which is also the gradient projection direction). If the design point is
moved away from the feasible domain, the node shift is undertaken again.

4. Intervals of node shift

In truss shape optimization process, the node shift interval i.e. the step length is an important parameter.
The smaller the node shift interval is, the more accurate the optimum design is at a cost of large compu-
tation time while the large interval may cause oscillations or convergence difficulties for the final design. As
the shape optimization progresses, the element lengths connected to a shifted node are changed, so the
interval should be adapted to this variation.
When the change of nodal displacement vector fDug is expressed in formula (5), a higher order of tiny
term ½DKfDug has been ignored. To ensure the precision of fDug and the correction of Eq. (9), ½DK is
required to be controlled under a tiny range of [K], for instance, 2–5% of [K]. In this paper, 2% is adopted. If
the trace of the matrix is considered simply, one can get the following expression:
jT ð½DKÞj 6 0:02jT ð½KÞj: ð26aÞ
By substituting Eq. (4) into above expression, we obtain
 !
 Xnj 
 
T ½Dke   6 0:02jT ð½KÞj: ð26bÞ
 e¼1


This can be satisfied in element level


jT ð½Dke Þj 6 0:02jT ð½ke Þj: ð27Þ
From Eq. (4), we can get the following inequality:
  
  
Dxj T o½ke   6 0:02jT ð½ke Þj: ð28Þ
 ox  j

Zhu et al. [9] have derived the expression of o½ke =oxj , and the traces are
  
   
T o½ke   ¼ 2Ee Ae  3a3 2a þ 3ab2 þ 3ac2  ¼ 2Ee Ae jaj; ð29Þ
 ox  L2 L2
j e e

2Ee Ae  2  2Ee Ae
jT ð½ke Þj ¼ a þ b 2 þ c2 ¼ ; ð30Þ
Le Le
where fa; b; cg is the direction cosines of the eth element with respect to global xyz-coordinate system.
Then, we get
Le
jDxj j 6 0:02 ) jDxj j 6 0:02Le : ð31Þ
jaj
3604 D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612

For each shifted node, we choose


 
jDxj j ¼ 0:02 min Le ; e ¼ 1; nj ; ð32Þ

then, inequality (21) is satisfied.


To accelerate the displacements approaching to the corresponding limits, the node shift intervals can be
taken largely at the beginning of the optimization. Hence, the following node shift interval is adopted in the
proposed method:

  0:02 X p X m  
  
Dxj  ¼  uil  min Le ; e ¼ 1; nj : ð33Þ
mp l¼1 i¼1 ui  

5. Limitation of truss shape optimization

As is well known, the shape optimization of a truss structure can be performed independently. It can also
offer an optimal geometry for further sizing optimization. However, it should be pointed out that although
the shape optimization can reduce the nodal displacements significantly, it cannot replace the sizing op-
timization, especially when stress constraints are imposed on the structure. The nodal displacement is, as we
know, inversely proportional to the cross-sectional area of the element. Sizing optimization can reduce the
nodal displacement arbitrarily if there are no upper bounds for sectional areas, but shape optimization
cannot. If the prescribed limits of the nodal displacements are set too small, the solution of the shape
optimization may not exist. A simple two-bar truss is used below to illustrate the fact.
As shown in Fig. 2, the two-bar planar truss is loaded at point 3. The material is the same for both bars.
Node 1 and 2 can be shifted vertically and h1 , h2 are the two design variables. The vertical displacement of
the loaded point d3 is subject to a constraint. The axial forces in each bar are expressed as
8
> PL1
>
< N1 ¼ h þ h ;
1 2
ð34Þ
>
> PL2
: N2 ¼ :
h1 þ h2

The displacement is
 3 
N12 L1 N22 L2 P L1 L32
d3 ¼ þ ¼ þ ; ð35Þ
PA1 E PA2 E Eðh1 þ h2 Þ2 A1 A2
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
in which L1 ¼ d 2 þ h21 , L2 ¼ d 2 þ h22 .
To find the minimum value of d3 , we set od3 =oh1 ¼ 0, od3 =oh2 ¼ 0, namely,
8  3 
>
> 3 2 L1 L32
< A1 L1 h1 ðh1 þ h2 Þ
> 2ðh1 þ h2 Þ
A1 A2
þ ¼ 0;
 3  ð36Þ
>
> 3 L L3
>
: L2 h2 ðh1 þ h2 Þ2 2ðh1 þ h2 Þ 1 þ 2 ¼ 0:
A2 A1 A2
If A1 ¼ A2 , we obtain
pffiffiffi
h1 ¼ h2 ¼ 2d: ð37Þ
D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612 3605

Fig. 2. A two-bar planar truss.

The Hessian matrix is required to be positive definite at the stationary point


 pffiffiffi
o2 d3  7 3P
¼ > 0;
oh21 h1 ¼h2 ¼pffiffi2d 16EA1 d
 2 2  ð38Þ
o2 d3 o2 d3 o d3  15P 2
 ¼ > 0:
2
oh1 oh2 2
oh1 oh2  pffiffi 32E2 A21 d 2
h1 ¼h2 ¼ 2d

Therefore,

2P pffiffiffi 3 3pffiffi3ffiPd
dmin
3 ¼  pffiffiffi 2 3d ¼ : ð39Þ
A1 E 2 2 d 4A1 E
pffiffiffi
If d3 < ð3 3PdÞ=ð4A1 EÞ is set, the shape optimization has no solution. In this case, sizing optimization has
to be applied. Fig. 3 shows the vertical displacement of node 3 ðlnðlnðd3 ÞÞÞ together with the contour lines
with respect to h1 and h2 (for A1 ¼ A2 ¼ E ¼ d ¼ P ¼ 1).

6. Illustrative examples

The feasibility and the effectiveness of the proposed method are illustrated by four typical examples.
These examples are often used to demonstrate the fitness of the optimization algorithms by researchers. In
this paper, only the shape optimization is considered.

6.1. Michell arch

The popular Michell’s semi-circular arch design has often been attained by topology optimization [5,6].
It is also used for sizing optimization [5]. In this paper, this semi-circular configuration would be gained by
truss shape optimization with a displacement constraint. The initially designed structure with the load is
3606 D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612

Fig. 3. Vertical displacement of node 3.

Fig. 4. Initial Michell’s arch shape design.

shown in Fig. 4, and the vertical displacement component at the loaded node is constrained v1 6 1:168 mm.
The Young’s modulus is assumed to be E ¼ 2:1  1011 Pa and material density q ¼ 7800 kg/m3 .
Suppose all the bars are of the same cross-sectional areas A ¼ 10 cm2 , nodes 3 and 7 are shifted
in horizontal direction and nodes 4, 5 and 6 in vertical direction, respectively. During the optimization
process, the symmetry of the structure is maintained. Therefore, only three sensitivity numbers for nodes 5,
6 and 7 are evaluated for sensitivity analysis independently. Two nodes are shifted in each iterative loop.
Fig. 5 shows the optimum design of the structure. It is close to a semi-circle. Figs. 6 and 7 demonstrate
the evolutionary histories of the specified displacement and the structural weight, respectively. From
Figs. 6 and 7, it is found that the displacement decreases monotonously but the weight changes non-
monotonously. The structural weight decreases at the beginning and the displacement reduces sharply. This
phenomenon corresponds to the positive sensitivity numbers of node 7 (and 3). Near the optimum solution,
the weight increases sharply while the displacement decreases tardily, which means that the efficiencies of
the shape optimization become much lower, and the sensitivity numbers approach to zeroes. The sudden
decline of the weight is caused by the node moving along the opposite direction of the gradient of the
Lagrange function. But the solution converges rapidly. Table 1 compares the sensitivity number values of
each node with the set convergent requirement Eq. (24b) at the optimum design.
D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612 3607

Fig. 5. Optimum Michell’s arch configuration.

Fig. 6. Displacement history of node 1.

Fig. 7. Structural weight history.

This example reveals again the high efficiency of truss shape optimization. With 10% increment of
structural weight, from 71.4 to 78.6 kg, the specified displacement reduces 76%, from 4.805 to 1.168 mm.
3608 D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612

Table 1
Comparison of sensitivity numbers with the convergent requirement at the optimum
jrLðX Þj a15 a16 a17 Maximal error (%)
0.991051E 2 1.167572E 5 1.180982E 5 1.202276E 5 3.0
0.479529E 2 1.174874E 5 1.180636E 5 1.191348E 5 1.4
0.167435E 2 1.178730E 5 1.180844E 5 1.184539E 5 0.5

6.2. 25-bar space truss

The 25-bar space truss structure is shown in Fig. 8. The structure is imposed by two load cases listed in
Table 2, and all the nodal displacements are constrained to be 6 8.89 mm [10]. The structure is required to

Fig. 8. 25-bar space truss.

Table 2
Two load cases for space truss (N)
Load case Node X Y Z
1 1 4448 44,4482 5000
2 0 44,4482 5000
3 500 0 0
6 500 0 0
2 1 0 88,964 5000
2 0 88,964 5000
D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612 3609

remain symmetry with respect to x–z plane and y–z plane. So, only the displacements of nodes 2 and 4 in x,
y and z-axes are considered. The bar cross-sections are divided into two groups with A1 ¼ 5 cm2 and A2 ¼ 4
cm2 , respectively. The Young’s modulus is E ¼ 6:89  1010 Pa and material density q ¼ 2768 kg/m3 . The
coordinates of nodes 4 and 8 are design variables and two independent coordinates are shifted in each loop.
The node coordinates and structural weight at the initial and the optimum are shown in Table 3.

Table 3
Values of design variables (m)
Design variables Initial coordinates Optimum coordinates
X4 0.952 0.608
Y4 0.952 1.957
Z4 2.54 2.756
X8 2.54 1.575
Y8 2.54 3.020
Weight (kg) 109.0 114.4

Fig. 9. Dome structure.


3610 D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612

6.3. Dome structure

Lipson and Gwin [10] used the dome structure for shape and sizing optimization simultaneously (seeing
Fig. 9). In this example, all the elements are assumed to have the same cross-sections A ¼ 10 cm2 and the
vertical displacements of node 1 are required not to exceed 1 cm in the four load cases, seeing Table 4. The
Young’s modulus is E ¼ 2:1  1011 Pa and material density q ¼ 7850 kg/m3 . Two independent coordinates
are shifted in each loop. The coordinates of nodes 1, 2 and 6 at the initial and optimum are listed in Table 5
respectively and the vertical displacement variations of node 1 under the four load cases are shown in
Fig. 10.

6.4. Simply supported bridge

A simply supported bridge structure is performed for shape optimization. To represent practical design
of the bridge members, beam elements are used for the element group 2 with the same rectangular cross

Table 4
Four load cases for the dome (N)
Load case Values Nodes
5
1 3:0  10 1
2 3:0  104 1–13
3 1:5  105 1
1:0  105 4, 5
4 1:5  105 1
7:0  104 2–4

Table 5
Values of design variables (m)
Node Initial coordinates Finial coordinates
X Z X Z
1 0 9.25 0 9.62
2 5.00 8.22 2.10 7.41
6 10.00 5.14 7.21 4.08
Weight (kg) 3459.2 3174.8

Fig. 10. Displacement variations of node 1 under four load cases.


D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612 3611

Fig. 11. Initial design of simply supported bridge.

Fig. 12. Optimum configuration of simply supported bridge.

section B ¼ 8 cm and H ¼ 5 cm. Elements in group 1 are bars with A ¼ 5 cm2 . The Young’s modulus is
E ¼ 2:1  1011 Pa and material density q ¼ 7800 kg/m3 . The vertical displacements of node 8 and 10 under
two load cases are constrained within 1 cm (1/1000 of the span), respectively. The initial configuration is
designed as in Fig. 11.

Load case 1: p ¼ 10 KN at the all nodes on the lower chord.


Load case 2: P ¼ 100 KN at the node 8.

The nodes on the lower chord remain fixed while nodes on the upper chord are shifted in the vertical
direction. The symmetry of the structure is remained during the optimization process and two nodes are
shifted in each loop. The optimum shape is shown in Fig. 12 with the corresponding y-coordinates of the
nodes. The structural weight changes from 433.5 to 489.6 kg.

7. Conclusions

When a truss structure is subject to multiple displacement constraints under multiple load cases, shape
optimization is an effective and highly efficient approach for minimum weight design. However, it turns out
that the shape optimization cannot reduce the displacement arbitrarily.
In this paper, an evolutionary optimization method is proposed to perform the shape optimization of a
truss structure subject to multiple displacement constraints for minimum weight design. The global sen-
sitivity numbers are employed to handle both single and multiple displacement constraints in a unified way.
By means of the sensitivity analysis, the most efficient nodes are shifted and the structural shape evolves
towards the optimum configuration. Directions and intervals of node shifts are determined automatically
and the solution is checked by the K–T optimality conditions. The method has proved to be effective and
reliable.
3612 D. Wang et al. / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3597–3612

Acknowledgements

This research work is supported by the National Natural Science Foundation of China under the
guarantees no. 10072050 and 10172072.

References

[1] G.I.N. Rozvany, M. Zhou, Advances in overcoming computational pitfalls in topology optimization, in: Proc. of the Sixth AIAA/
NASA/ISSMO Symp. on Multi-disc. Anal. and Optim., 1996, pp. 1122–1132.
[2] L. Gil, A. Andreu, Shape and cross-section optimization of a truss structure, Comput. Struct. 79 (2001) 681–689.
[3] N.L. Pedersen, A.K. Nielsen, Optimization of practical trusses with constraints on eigenfrequencies, displacements, stresses and
buckling, report no. 664, Technical University of Denmark, 2001.
[4] W.H. Zhang, M. Domaszewski, C. Fleury, A new mixed convex approximation method with applications for truss configuration
optimization, Struct. Optim. 15 (1998) 237–241.
[5] Y.M. Xie, G.P. Steven, Evolutionary Structural Optimization, Springer, Berlin, 1997.
[6] D.N. Chu, Evolutionary structural optimization method for systems with stiffness and displacement constraints. Ph.D. Thesis,
Department of Civil and Building Engineering, Victoria University of Technology, Melbourne, Australia, 1997.
[7] R.T. Haftka, M.P. Kamat, Elements of Structural Optimization, Martinus Nijhoff Publishers, Hague, 1985.
[8] D.N. Chu, Y.M. Xie, A. Hira, G.P. Steven, Evolutionary structural optimization for problems with stiffness constraints, Finite
Element Anal. Design 21 (1996) 239–251.
[9] B.F. Zhu, Z.M. Li, B.C. Zhang, Principle and Applications of Structural Optimization, Waterpower Publisher, Beijing, 1984, in
Chinese.
[10] S.L. Lipson, L.B. Gwin, The complex method applied to optimal truss configuration, Comput. Struct. 7 (1977) 461–468.

You might also like