Sei sulla pagina 1di 10

Journal of Mechanical Science and Technology 32 (4) (2018) 1643~1652

www.springerlink.com/content/1738-494x(Print)/1976-3824(Online)
DOI 10.1007/s12206-018-0319-1

Bi-level optimization of laminated composite structures using particle


swarm optimization algorithm†
Parviz Mohammad Zadeh, Mahdi Fakoor* and Mostafa Mohagheghi
Faculty of New Sciences and Technologies, University of Tehran, North Kargar St., Tehran, Iran

(Manuscript Received April 5, 2017; Revised December 11, 2017; Accepted January 7, 2018)

----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Abstract

This paper presents an efficient bi-level optimization technique to obtain the optimal stacking sequence for symmetric composite struc-
tures. The proposed approach involves two levels of modeling and optimization. The first level of the optimization procedure is used to
minimize the weight of the composite structure. At this level, lamination parameters and the number of plies of specified angles (0, ±45
and 90 degree) are design variables, buckling load factor is treated as a constraint, and the weight of the structure is to be minimized us-
ing continuous-discrete particle swarm optimization algorithm. Next, at the second level the location of each ply orientation through the
thickness (i.e. the layup of the panel) is found. At the second level, optimum stacking sequence is sought to maximize the load bearing
capacity of the structure with respect to the buckling. The proposed methodology is applied to two test cases. Results show that the ap-
proach improves the buckling load factor of the structure without any weight penalty.
Keywords: Bi-level optimization; Buckling; Lamination parameters; Particle swarm optimization
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

biaxial loading. In this work, the minimization of total thick-


1. Introduction
ness was formulated and buckling load was considered as a
In the design of composite structures (e.g. wind turbine constraint. The problem was then solved for uniaxial and biax-
blade) weight plays an important role, while a detailed and ial load cases for various plate aspect ratios. Le Riche and
rigorous analysis is required to ensure the strength of the Haftka [3] studied the problem of buckling load maximization
structure. Using optimization techniques is a prominent ap- of a laminated plate using genetic algorithm (GA). The au-
proach that helps designers to reach the best possible structure, thors also studied the effect of contiguity constraints, in which
satisfying required constraints. In the optimization process, the number of plies with the same ply orientation is not al-
unlike weight evaluation which is of a routine and easy proce- lowed to be more than four plies. Ref. [4] studied optimization
dure, the calculation of mechanical constraints such as buck- of a laminated composite beam with a rectangular cross sec-
ling load factor and strain are usually much more sophisticated, tion. The authors concluded that for the cases in which the
and in general, finite element models are used to reach this dimensions of the skin are small in comparison to the length
evaluation [1]. of the beam, the loads on the skin can be only considered as
However, analysis of these models for the mentioned goals in-plane, and the out-of-plane loads are negligible. Therefore,
can be complicated, entailing time-consuming computations. the classical lamination theory (CLT) can be used for such
Thus, the framework of optimization, which determines the problems.
number of iterations, is of great importance in such problems. In recent years, there has been a growing research interest in
Using multi-level optimization methods in a right manner can multi-level optimization of complex composite structures. In
decrease the iterations required to reach the global optimal this approach the main idea is to separate the optimization
point. problem into two levels for decreasing the complexity of the
Significant research work has been reported in the field of problem and therefore increasing the efficiency of the algo-
optimization of composite structures. Haftka and Walsh [2] rithm. The objective of the first level is to minimize the num-
employed the integer programing approach for the optimum ber of plies in a composite structure. The answer of this level
design of symmetric and balanced composite plates under is not the final solution of the problem but restricts the main
*
Corresponding author. Tel.: +98 21 61115754, Fax.: +98 21 88497324
problem and reduces the unknown design variables. Follow-
E-mail address: mfakoor@ut.ac.ir ing this, the second level is used to calculate the objective of a

Recommended by Associate Editor Kyeongsik Woo more detailed sub-problem (e.g. stacking sequence optimiza-
© KSME & Springer 2018
1644 P. M. Zadeh et al. / Journal of Mechanical Science and Technology 32 (4) (2018) 1643~1652

tion problem with a given number of plies as constraints). This duces a set of constraints on the lamination parameters to de-
procedure will reduce the computational cost as well as the termine the feasible region of search in the optimization algo-
design uncertainties. Levels of the problem can be defined in rithm. Although some research work used lamination parame-
different manners. For example, using high fidelity and low ters as variables in PSO algorithm to optimize stacking se-
fidelity models in two levels is an approach implemented in quence of composite plates [21, 22], few studies integrate the
Ref. [5]. Another bi-level method for composite structural PSO into the bi-level procedure. Some researchers utilized
problems is to optimize one set of variables (often lamination gradient based methods [23] and some others implemented the
parameters) in a global level (first level) and other set (that is a solution via one of the evolutionary optimization algorithms
representative of lay-up) in the local one (second level). As a such as a GA. In the PSO algorithm, used in the present study,
pioneer in the use of a two-level optimization strategy for a condition is added to eliminate the similar stacking se-
composite structures in the latter manner, B. Liu used a two- quences that are generated in one generation. This will in-
level optimization algorithm with response surface approxi- crease the efficiency of the algorithm.
mations [6]. Furthermore, a bi-level lamination parameter- This paper focuses upon the development of an efficient
based algorithm considering the continuity constraint between multi-level particle swarm optimization for complex compos-
adjacent panels was proposed in Refs. [7, 8]. In this regard, ite structures and the main novelties can be stated as:
Liu [6] developed a permutation GA for lower level optimiza- (1) Defining two actual and design related objectives
tion and continuity measure for improving continuity con- (weight and buckling load factor) in two levels and using
straints among adjacent panels. While GAs have been popular lamination parameters as a guide to impose constraints (in
for optimizing composite laminated plates, they usually suffer some previous works the difference between lamination pa-
from time-consuming computations [9]. It should be noted rameters of two levels, that has no physical result, have been
that some authors use “bi-level” for approaches in which the considered as objective function).
procedure is implemented in two levels, while both terms “bi- (2) Calculating new velocity and position of particles for
level” and “multi-level” are used interchangeably through this continuous variables (lamination parameters) and discrete
paper. variables (the number of plies) simultaneously (in a single
Particle swarm optimization (PSO) is a meta-heuristic op- run).
timization technique that may require less computational ef-
fort to obtain the same high-accuracy solution than a GA [10, In order to avoid the undesired effects of coupling, a sym-
11]. Using PSO for a multi-level structural optimization prob- metric stacking sequence is supposed. These assumptions
lem can be traced back to Ref. [12] for sever numerical noises. together with some other design and manufacturing considera-
In Ref. [13] the applicability and reliability of the proposed tions are added to define the feasible region of lamination
algorithm is also indicated. Unlike other evolutionary algo- parameters. Additionally, in the population generation step,
rithms, individuals of PSO population retain good solutions stacking sequences with more than four same orientations are
during algorithm implementation. Fast convergence rate, the eliminated to obtain a solution observing the continuity condi-
ease of implementation and the number of parameters set are tion.
other advantages of PSO which have led to its wide applica-
tion [14, 15]. Suresh et al. [11] has adopted PSO as a multi- 2. Theoretical background
agent search method to maximize elastic coupling of a com-
2.1 Laminated composite structures analysis
posite helicopter rotor blade. Other proposed application in-
cludes the use of both gradient based and PSO algorithms in In the present work, CLT is used to analyze the mechanical
the optimization of a helicopter rotor blade [16]. It has been behavior of composite structures. According to this approach,
found that being a stochastic method, the PSO does not need the relationship between forces and displacements is described
to start from different initial points. Chang et al. [17] make as:
some changes to adapt standard PSO for stacking sequence
design, which is a discrete problem. They developed a discrete
ì N ü é A B ù ìe ü
PSO with new rules for abstraction operator in the formula. Li í ý=ê úí ý (1)
and Chandrashekhara [18] applied this method to the design î M þ ë B D û îk þ
of composite hydrokinetic turbine blades. They also devel-
oped a permutation discrete PSO model which maximizes the where N and M are forces and moments per unit width, A, B
out-of-plane load carrying capacity of the composite blade. and D are stiffness matrices, e and k are strain and curvature
Although composite structures can be tailored using appro- vectors, respectively.
priate ply orientations and ply stacking sequences [13] to meet To evaluate stiffness matrix D in the optimization algorithm,
specific design requirements [11], some research work (e.g. [8, lamination parameters are used as design variables. Tsai and
19, 20]) focused on the use of lamination parameters in the Hahn presented the stiffness properties of a single-material
composite optimization procedure. The emphasis of Ref. [8] is laminated composite in terms of 12 lamination parameters
on the solution of top level optimization while Ref. [19] intro- [24]:
P. M. Zadeh et al. / Journal of Mechanical Science and Technology 32 (4) (2018) 1643~1652 1645

æ A11 ö é1 x1A x 2A 0 0ù
ç ÷ ê A ú æ U1 ö
ç A22 ÷ ê1 -x1 x 2A 0 0ú ç ÷
U2
ç A12 ÷ ê0 0 -x 2A 1 0 ú çç ÷÷
ç ÷ = hê ú U3 (2)
ç A66 ÷ ê0 0 -x 2A 0 1ú ç ÷
çA ÷ ê0 x A / 2 x A çU 4 ÷
ç 16 ÷ 0 0ú ç ÷
ê 3 4
ú U5
çA ÷
è 26 ø êë0 x3A / 2 -x 4A 0 0 úû è ø
æ D11 ö é1 x1D x 2D 0 0ù
ç ÷ ê D ú æ U1 ö
ç D22 ÷ ê1 -x1 x 2D 0 0ú ç ÷
U2
ç D12 ÷ h3 ê0 0 -x 2D 1 0 ú çç ÷÷
ç ÷= ê ú U3 (3)
ç D66 ÷ 12 ê0 0 -x 2D 0 1ú ç ÷
çD ÷ ê0 x / 2 x D çU 4 ÷
ç 16 ÷
D
0 0ú ç ÷
ê 3 4
ú U5
çD ÷
è 26 ø ê
ë 0 x 3
D
/ 2 -x D
4 0 0 úû è ø

in which the lamination parameters are,


Fig. 1. A bi-axial loaded composite plate.
A
x [1,2,3,4] =
(4) In general, a composite laminate can have following cou-
1 n
å[cos 2qi ,cos 4qi ,sin 2qi ,sin 4qi ]( zi - zi -1 )
h i =1
plings:
D
(i) The condition of Bij ≠ 0 that is named in-plane coupling
x[1,2,3,4] = (extension or membrane) and out-of-plane (bending or flex-
4 n
3
(5) ural) actions,
å[cos 2q ,cos 4q ,sin 2q ,sin 4q ]( z
i i i i i - z 3i -1 )
h3 i =1 (ii) The situation of A16 = A26 ≠ 0 that is in-plane shear and
extension, and
where qi is the ply angle of the i-th layer, and zi is the dis- (iii) When D16 = D26 ≠ 0 that is called out-of-plane bending
tance from mid-plane to the bottom of the ith layer. Based on and twisting [25].
this definition and with the assumption of using only 0, ±45
and 90 degree orientations, it can be inferred that: In this work, the laminate is forced to be balanced, that is,
every ply with a positive fiber orientation angle is of a corre-
(n0 - n90 )t n0 - n90 sponding ply with the negative fiber orientation angle. This
x1A = = (6)
Nt n0 + n90 + 2n±45 leads to no normal-shear extensional couplings [26]. As seen
n0 + n90 - n±45 in the Eq. (1) non-zero B matrix implies extensional-flexural
x 2A = (7) couplings. In other words, if B is not zero, a small axial force
n0 + n90 + 2n±45
may create a curvature (out-of-plane deformation). Therefore,
x3A = x 4A = x 4D = 0 . (8)
to prevent this problem symmetric stacking sequence is con-
sidered.
In Eqs. (6)-(8), n0, n±45, n90 are the number of plies of each
orientation. Since only laminates with ply orientations of 0o,
2.2 Critical buckling load calculation
±45o and 90o are considered, the minimum and maximum
values of x1D can be determined. These bounds with some In this paper, two case studies are investigated. The first one
other relations between lamination parameters can be seen in is a composite panel under bi-axial compression loading
Ref. [1]. (shown in Fig. 1). One of the failure modes of such laminated
The material intervals are shown by U1, U2, U3, U4 and U5 composite panel is buckling that may occur in m and n half
in Eqs. (2) and (3) and are defined as below, waves in the x and y directions. For example, the simply sup-
ported composite plate, shown in Fig. 1, has buckled with
U1 = [3Q11 + 3Q22 + 2Q12 + 4Q66 ] / 8 (9) three half waves (m = 3) in the x-direction. The dimensions of
U 2 = [Q11 - Q22 ] / 2 (10) the panel in x and y directions are named a and b, respectively.
U 3 = [Q11 + Q22 - 2Q12 - 4Q66 ] / 8 The buckling load for a symmetric composite plate is ob-
(11) tained through Eq. (14) [17]:
U 4 = [Q11 + Q22 + 6Q12 - 4Q66 ] / 8
(12)
U 5 = [Q11 + Q22 - 2Q12 + 4Q66 ] / 8 (13) lb( m ,n )
=
p2
(14)
in which Qij denotes reduced stiffness elements for unidirec- D11 (m / a ) 4 + 2( D12 + 2 D66 )(m / a ) 2 (n / b) 2 + D22 (n / b) 4
.
tional lamina. ( m / a ) 2 N x + ( n / b) 2 N y
1646 P. M. Zadeh et al. / Journal of Mechanical Science and Technology 32 (4) (2018) 1643~1652

Suppose that m and n are integer numbers and should be for i = 1 : n do


evaluated to obtain the smallest value of lb( m ,n ) , which is if X(i) ≤ n0 then
called critical buckling load factor. X(i) is decoded to 0;
The second case study is minimization of a composite wing else if X(i) ≤ n0 + n±45 then
box, in which the buckling load factor is calculated using fi- X(i) is decoded to ±45;
nite element software. else
X(i) is decoded to 90;
end if
3. Bi-level optimization approach
end for
The proposed multi-level particle swarm optimization algo-
Fig. 2. The procedure of conversion.
rithm is described in detail in Secs. 3.2 and 3.3. The single
population approach of PSO algorithm as a practical and effi-
cient strategy in structural problems [27-29] is used in both and n90). These variables initially are of broad range and there-
levels. It should be added that, multi-swarm variant of PSO is fore the corresponding search space is wide. While for PSO as
efficient specifically in dynamic problems, in which optimum a heuristic optimization method, the less extent of the search
may be changed over time [30]. In composite structural opti- space, the less number of iterations is required to perform a
mization problems, however, single swarm PSO is commonly global search, and to obtain the optimum point. Hence, the
used as a suitable approach to deal with the problem, and to bounds of the search space, which are determined by design
obtain the optimal solution [14, 18, 31]. variable limits, are of considerable importance. In the pro-
The proposed bi-level approach based on PSO algorithm posed method, before starting the optimization process, the
presents an efficient way of solving composite optimization upper limits of the design variables are redefined so that the
problems. First, in the proposed bi-level method, modeling unwanted space is removed. This technique can be considered
and optimization are separated into two levels. The first level as a pre-processing phase and contains the following tasks:
is utilized to minimize weight in which there are six design (1) Calculating the buckling load factor for n0 = n0i, n±45 =
variables. The second level is used to maximize buckling load n±45i, n90 = n90i named λi. The parameters n0, n±45 and n90 are
factor where the first level design variables are frozen as con- initial upper bounds that are usually too conservative;
stant values. Second, two optimizers using PSO algorithm are (2) Obtaining the buckling load factor for three group of
used both at the first and second levels. These characteristics layers: λ1 for (n0 = n0i-6, n±45 = n±45i, n90 = n90i), λ2 for (n0 = n0i,
of the algorithm is reflected in computational time reduction n±45 = n±45i-3, n90 = n90i), λ3 for (n0 = n0i, n±45 = n±45i, n90 = n90i-
and computational efficiency increase. 6).
The algorithm contains a swarm of particles where each (3) Determine the effect of reduction in the number of lay-
particle shows a potential solution to the optimization problem. ers of each orientation (Δ1= λi - λ1 , Δ2= λi - λ2, Δ3= λi – λ3), and
The position of the particle i is modified by adding a velocity obtain the least effective case:
vector ( Vki +1 ) to the current position, i. e.: Least effective case = index (min (Δ1, Δ2, Δ3));
(4) Update the number of layers and repeat the procedure.
Xik +1 = Xik + Vki +1 t (15) For example if Δ3 is less than Δ1 and Δ2, then n0i,new = n0i,
n±45i,new = n±45i, n90i,new = n90i – 6.
where Vki +1 is the velocity vector at iteration k+1 and depends
on both the intended particle experience and the experience of It is worth noting that for obtaining the buckling load factor,
other particles in the swarm: six lamination parameters are required. Three first one,
x1A , x 2A , x3A are obtained using Eqs. (6)-(8) and, x1D , x 2D ,
x3D are randomly evaluated satisfying feasibility constraints
Vki +1 = wVki + c1r1 (Pbest i - x ik ) + c2 r2 (Gbest kg - x ik ) . (16)
(stated in Ref. [7]). Another constraint, buckling load factor, is
applied using a penalty function. For this purpose, if this fac-
In Eq. (15) t is time interval and is considered equal to 1. tor is less than 1 (the minimum allowable value), a term is
In Eq. (16), r1 and r2 are random numbers [32]; c1 and c2 are added to the objective function. Next, the final values of the
user defined parameters showing how much the next position number of layers are fed to the bi-level optimization procedure.
of the particle depends on itself or on the swarm. The parame- The main steps involved in the proposed approach are de-
ter w is the inertia weight and controls the space to explore scribed below.
[31].
3.2 First level optimization
3.1 Pre-optimization phase
In the present work a bi-level framework is proposed for the
As used in the procedure of Fig. 2, there are three variables, optimization of composite structures, in which at the first level
associated to the number of layers of each orientation (n0, n±45 the number of plies of each fiber orientation (n0, n±45, n90) and
P. M. Zadeh et al. / Journal of Mechanical Science and Technology 32 (4) (2018) 1643~1652 1647

the out-of-plane lamination parameters related to stiffness for i=1 : n do


matrix D ( x1D , x 2D , x3D ) are design variables. Hence, the num- for j=1 to n do
ber of design variables, which is expressed as particle size in if particle(i) == velocity(j) then
PSO [33], is equal to 6. In order to eliminate the extensional- Temp = particle(i) ;
flexural couplings, symmetric stacking sequence is assumed particle(i) = particle(j) ;
and therefore x1B , x 2B , x3B are fixed and equal to zero. As well, particle(j) = Temp ;
more constraints than unsymmetrically laminated plates can end
be used [34, 35]. end for
There exist some improvements in PSO method to be appli- end for
cable in the ply stacking sequence optimization of composite
Fig. 3. The procedure of calculating new particle positions.
laminates. In the present paper, PSO is used at both levels of
the bi-level optimization approach. Some consistent con- 3.3 Second level optimization
straints are derived to be used in the first level.
The objective of this level is to minimize the weight of the There exist some improvements in PSO method to make it
structure (Eq. (17)), which means to minimize total thickness applicable in the ply stacking sequence optimization of com-
of panels because the density and also other dimensions of the posite laminates. In Ref. [37] it was shown that for continuous
structure are assumed invariable. problems of layup optimization and also for problems in
which the number of possible ply orientations is considerable,
PSO outperforms other meta-heuristic techniques such as ant
W º rV = r (n0 + 2n±45 + n90 )abt . (17)
colony optimization (ACO) and GA.
In the second level of the present work, the numbers of plies
In the above equation r is the density of the considered from the first level are considered fixed, and ply orientations,
composite material, and V is the total volume, t is the thick- θ, are the design variables. Therefore, particle size in this level
ness of one ply. As a result, the first level optimization can be will be n0 + n±45 + n90.
formulated as: The optimization formulation is given:

Minimize(n0 + 2n±45 + n90 ) . (18) Minimize :buckling load factor (lb ) . (19)

Subject to: Subject to: Laminates are symmetric and balanced, ply con-
Feasibility of lamination parameters (laminate is symmetric tinuity requirements and manufacturing limitations.
and balanced) (buckling) lb ³ 1 where lb indicates the Here, the stacking sequence of the structure is determined
buckling load factor as calculated by Eq. (14) (for an individ- by implementing PSO algorithm, which is specific to compos-
ual rectangular panel), or by finite element analysis (for com- ite structure problems. This procedure is based on the follow-
plex structures such as a wing model including spars, ribs, and ing steps:
upper and lower panels). (1) The algorithm parameters including w, c1, c2 are set to
In the beginning of the algorithm, three first variables that proper values (1, 0.6, 0.4, respectively);
represent the number of plies are randomly generated integers (2) A random permutation of the integers 1 to n, denoted by
satisfying corresponding constraints. Next, updating the posi- X, is generated where n = n0 + n±45 + n90. For instance, for n0 =
tion of each particle (number of plies and also lamination pa- 5, n±45 = 4, n90 = 2, (n = 11), X as a randomly generated vector
rameters) using Eqs. (15) and (16), the new position may not can be: X = [7 / 4 / 6 / 10 / 11 / 3 / 2 / 5 / 1 / 9 / 8]; this string is
be integer. In this situation the three first ones, which specify identical to a unique lay-up: [±45/0/±45/90/90/0/0/0/0/±45/
the number of plies, are rounded to the nearest integers. Also ±45]s. The procedure of conversion can be stated as:
if the new position is out of bounds, it is replaced by the near- (3) Calculate objective function (buckling load factor that is
est limit to maintain in the desired search space. a representative of the load bearing capability of the structure)
The first level algorithm is repeated until it reaches the for all of the particles;
maximum number of generations. The output of this level is (4) Evaluate particles and obtain the best positions of each
three integers specifying optimum number of plies (for mini- individual Pbesti and the best position of the whole swarm
mum weight), that is fed to the next level. Although lamina- Gbestk (i denotes the particle i and k denotes the iteration
tion parameters are also design variables, they are not used in number).
the second level. This is because different lay-ups generate (5) Particle velocities are determined using Eq. (16),
some discrete lamination parameters, while they are treated as rounded to integer values, and checked to be in the right do-
continuous variables before, and it has been shown that the main;
optimal point of continuous lamination parameters is not nec- (6) Then new particle positions (lay-ups) are calculated as
essarily the closest one to the optima of discrete space [36]. shown in Fig. 3.
1648 P. M. Zadeh et al. / Journal of Mechanical Science and Technology 32 (4) (2018) 1643~1652

Table 1. Parameters of PSO algorithm used in this study. Table 2. Result comparison for a rectangular bi-axial loaded panel.

Parameter Value Optimal design λb


Initial velocity Random [0 1] Present work [±45/902/±45/903/±452/904/±452/90/0/90]s 1.019
Inertia weight (w) 1 Lopez [38] [902/±45/(902/±45)2/±455]s 0.990
c1, c2 0.6, 0.4
First level: 6
Particle size
Second level: n0 + n±45 + n90 Table 3. Material properties for carbon/epoxy.
Number of swarms 1
Material IM7/8552 (case 1) Graphite epoxy (case 2)
Population size 60
E11 (GPa) 141 128
Number of iterations 100
E22 (GPa) 9.03 13
E12 (GPa) 4.27 6.4
(7) If the iteration reaches the ultimate value, save the last ν 0.32 0.3
Pbesti and Gbestk and finish the loop, else go to step 3. t (mm) 0.191 0.127

In the proposed technique the particle positions (ply orienta-


tions) are discrete values. Based on the definition (Eq. (16)), 100 generations, for laminated composite structure optimiza-
the longer the velocity vector is, the farther the particle posi- tion problems in the scale of the above case, gives the optimal
tion is from the optimum point. Next, as shown in the proce- answer in appropriate number of function calls.
dure of step 6, the orientations of previous lay-up is replaced In addition, because stochastic optimization algorithms may
by corresponding ones of the velocity vector. In other words, converge to different results in each run, the optimization of
if the particle position is close to the optimum point, nearly all the case was performed 100 times.
elements of the velocity vector will be zero, and therefore will With above mentioned settings, the success rate was over
have a little effect on particle position updating. 90 percent. Therefore, these values are taken for the following
As mentioned above, new population is a string of numbers, computations to ensure the convergence to the right global
which in this step is converted to ply angles. These data to- optima.
gether with material properties are sufficient to calculate buck- As the result shows, the buckling load factor that is obtained
ling load factor. So lamination parameters only are used in the from present approach is about 3 percent higher than the result
first level to ensure the feasibility of the solutions. of Ref. [38]. This parameter has been considered as objective
function in the present work, while is a constraint in some
4. Evaluation and discussion of the results previous researches [1].
4.1 Validation
4.2 Test case 1: Individual composite panel
For validation of the proposed procedure, the bi-level
framework has been applied to a rectangular, simply sup- To show the performance of the proposed procedure, two
ported panel that is subjected to compressive in-plane loads test problems are selected. The first one is a set of graphite-
[38]: Nx = 1716 N/mm and Ny = 858.1 N/mm. In both ap- epoxy panels that are considered individually and is investi-
proaches the optimum value for the number of plies (the first gated in Ref. [38]. The second one is a composite wing box
level objective function) is 48. The obtained stacking se- model presented in Refs. [6, 7]. The material properties for
quence is presented in Table 1. these two cases are given in Table 2.
In updating the particle position, the velocity is obtained us- In both examples the weight (proportional to the number of
ing Eq. (16) in each iteration. Based on the study of Ref. [39], plies) is minimized subject to lamination parameters con-
setting w in range [0.8 1.2] yields a faster convergence. In the straints and buckling constraint. Furthermore, the stacking
present work w = 1 has been set. Furthermore, for initial ve- sequence is assumed to be symmetric (no membrane-bending
locity three convectional ways [40] were examined and ran- coupling).
domly generated values from [0 1], that gives a proper con- Before implementing the wing structure design optimization
vergence behavior [16], has been set. using the PSO based bi-level approach, the process is applied
In selecting algorithm parameters, also, it is required to to several individual panels. The resulting number of plies and
achieve an appropriate balance between population size and buckling load factor are listed in Table 3 and the optimal
number of generations to minimize the number of function stacking sequences are shown in Table 4. In this case study the
calls. Considering this fact, we run the program for some val- laminates are assumed to be balanced (no stretching-shearing
ues of population size and investigated the convergence dur- coupling). The length and the width of panels 1, 2, 9-12 are
ing generations. Based on the observations, it was obtained 610 mm and 457 mm, respectively, and other panels are 508
that selecting population size of 60 and recording results after mm by 305 mm.
P. M. Zadeh et al. / Journal of Mechanical Science and Technology 32 (4) (2018) 1643~1652 1649

Table 4. Result comparison for individual panels. Table 5. Stacking sequence of rectangular panels.

Present work Panel Optimal stacking sequence


Panel Nx, Ny (N/mm)
Number of plies l buckling 1 [±454/902/02/(90/0)2]s
1 122570, 70040 32 1.110 2 [±452/45/90/-45/906/0]s
2 65662.5, 63036 28 1.046 3 [03/±453/90]s
3 47277, 56907.5 20 1.037 4 [04/±452/902]s
4 43775, 35020 20 1.268 5 [03/±452/902]s
5 36771, 17510 18 1.191 6 [03/±45/0/±45/903]s
6 53405.5, 63036 22 1.204 7 [04/±452/902]s
7 50779, 34144.5 20 1.334 8 [03/±454/902]s
8 105060, 84048 26 1.036 9 [±453/902/±452/0/90/0/90/03]s
9 192610, 105060 38 1.210 10 [±452/903/±452/902/04]s
10 157590, 70040 34 1.161 11 [±455/90/02/902]s
11 65662.5, 91927.5 30 1.009 12 [±452/90/±45/902/02/90/02]s
12 70040, 56032 28 1.072 13 [04/±45/90/±45/902]s
13 57783, 57783 22 1.169 14 [02/±453/90]s
14 33269, 35895.5 18 1.084 15 [02/±452/02/±45/902]s
15 52530, 106811 24 1.286 16 [03/±45/04/±45/902/±45]s
16 142706.5, 175100 30 1.058 17 [03/±453/90]s
17 56032, 31518 20 1.043 18 [03/±453/902]s
18 52530, 71791 22 1.180
Number of function evaluations (mean)

7000 Without pre-optimization


With pre-optimization
6000

5000

4000

3000

2000

1000

0
0 5 10 15 20
Panel number

Fig. 5. Comparison of number of function evaluations with and with-


out pre-optimization phase.
Fig. 4. Weight reduction vs. iteration.
increase the bending stiffness of the panel.
In the proposed approach, since the buckling load factor is The objective function reduction by generation is illustrated
maximized in the second level, in the final stacking sequence in Fig. 4 for the first panel. In most cases the algorithm con-
the buckling load factor not only satisfies the constraint, but verges before fifty generations, but a challenge of heuristic
maximizes the load bearing capacity of the structure. algorithms such PSO is premature convergence. The parame-
From a structural point of view, the first ply orientation of ters of w, r1, r2 in the main relation, Eq. (15), are set to prevent
obtained optimal stacking sequence for panels 1, 2, 9-12, that this issue.
are of same dimensions, is ±45 degree (Table 4), while in Defining a pre-optimization phase is an idea that is used in
other panels the first ply orientation is 0 degree. This can be this work to reduce the required structural analysis, which is
due to the fact that a (the length of the panel perpendicular to the time consuming part of optimization process, especially in
Nx) is greater than b, and also Nx is greater than Ny in panels 1, complicated structures. For illustrating the effect of the sug-
2, 9-12. gested phase, the number of constraint evaluations, that is
In other panels that a is less than b and Nx is more than Ny in equal to the number of structural analysis, are shown in Fig. 5.
most of them, the critical buckling mode is in x direction and For panels with more buckling load factor (Table 3) the influ-
the optimal stacking sequence (outer plies) are 0 degree that ence of using pre-optimization phase is more considerable.
1650 P. M. Zadeh et al. / Journal of Mechanical Science and Technology 32 (4) (2018) 1643~1652

Table 6. First level results (total number of plies). Table 7. Second level results (stacking sequence).

Objective function Buckling


n0 n±45 n90
(Total number of plies) Obtained Stacking sequence load factor
Present work (With and (λb)
without the assumption 34 8 26 152 [±45/02/45/90/-45/0/902/02/45/902/
of balanced laminate) Present -45/(0/90)2/90/02/90/0/±45/45/0/
Ref. [7] 34 8 26 152 work -45/45/902/-45/0/45/902/02/-45/ 1.0868
No assump-
(PSO) 45/90/0/90/04/(90/0)2/0/
Ref. [6] 32 12 28 168 tion of
903/04/(90/0)3/0/90/03/-45/0/902]s
balanced
layup [(±45)2/03/45/902/-45/903/45/902/-45/
0/904/45/02/-45/0/903/02/904 /0/45/0/
Ref. [7] 1.0837
-45/04/90/04/90/03/902/04/45/02/−45/
02/45/02/-45/903/02/90]s
Present [±45/04/902/±452/02/904/±45/02/±452/
With the
work 02/±452/04/904/(02/902)2/(04/902)2/ 1.0867
assumption
(PSO) (902/02)2/904/02]s
of balanced
layup [(±45)12/(904/02)2/(902/04)2/
Ref. [6] 1.0985
902/02/904/02/(02/902)2/(04/902)3]s

cessitates a large number of layers. In other words, for about


150 layers, this assumption has no significant effect in buck-
Fig. 6. Wing box model.
ling strength of the structure.

4.3 Test case 2: Wing box structure (finite element analysis) 5. Conclusions
The benchmark wing box model, which is shown in Fig. 6, In this study, a two-level composite optimization procedure
is used to illustrate the bi-level optimization procedure pre- was presented. The procedure was demonstrated using two
sented in Sec. 3. In Ref. [7] the structure was assumed to be composite structure examples. This approach includes a con-
balanced, while in Ref. [6] this assumption was not included. ventional PSO algorithm at the first level, that uses both inte-
Hence, to compare the results with these references, the opti- ger and continuous variables, and a permutation discrete PSO
mization is done two times with two different assumptions. at the second level. In the first level, the number of 0o, ±45o,
Top skin panels are treated in the composite design optimi- 90o stacks are determined. Then, at the second level problem
zation whereas other parts are considered with the fixed stack- becomes a combinatorial optimization problem, with fixed
ing sequence, i, e, spars and ribs with [±4511]s and the bottom number of variables. Shuffling of layers is performed to
skin panels with [±45/(90/0)3/06]s. The actual loading on the maximize the buckling load factor. The process decreases
structure is replaced by four concentrated forces: F1 = 90,009 computations and increases the efficiency of the optimization
N, F2 = F3 = 187888 N, F4 = 380176 N. problem of selecting the optimal stacking sequence. This ap-
The results of optimization are compared with those pre- proach results in a realistic objective function and it was
sented in Refs. [6, 7] in the table above. shown that rounding off at the mentioned step does not influ-
Although the number of each layer (n0, n±45, n90) as design ence the final stacking sequence as the optimal point.
variables which contribute to the objective function are of Regarding the effect of the size of search space in the com-
unique convergence point in different runs, the lamination putation time of the algorithm, a pre-optimization step was
parameters are not converged to a single point. So in the pre- added to the first level, in which the upper bound of the num-
sent study, unlike some previous work, only the number of ber of layers for each orientation is determined.
layers is transferred to the second level and lamination pa- The number of layers in each orientation, as design vari-
rameters are yet variable in the second level. As stated in Ta- ables at the first level, is transferred to the second one and is
ble 6, this allows to reach a higher value of buckling load fac- used as fixed parameters at this level, but lamination parame-
tor in the latter level. Table 5 shows a considerable weight ters are not transferred. Lamination parameters in the first
reduction in comparison to Ref. [7]. level are of continuous values, while variables in the second
The optimal weight, which is obtained in first level, is simi- level (the arrangement of orientations) are actually discrete,
lar to the results of Ref. [6], and the buckling load factor, and sometimes the optimum discrete designs are not equal to
which is maximized in the second level, is a little higher than the continuous optima in the lamination parameter space. In
buckling load factor calculated in Ref. [7]. large structures with a large number of layers, the distance
The difference between optimal lay-up with balanced stack- between discrete design points decreases. Case studies in the
ing sequence and without this assumption is very small in this paper show this conclusion and in wing box example there is a
case. This could be due to the magnitude of loading that ne- minor difference between present result and Ref. [7].
P. M. Zadeh et al. / Journal of Mechanical Science and Technology 32 (4) (2018) 1643~1652 1651

In conclusion, using a two level approach including two ac- composite box structure, Advanced Materials Research, 430
tual objective functions and pre-optimization phase offers a (2012) 470-475.
practical procedure to optimize laminated composite struc- [15] I. C. Trelea, The particle swarm optimization algorithm:
tures with a reasonable accuracy and confidence. Convergence analysis and parameter selection, Inf. Process.
Lett., 85 (6) (2003) 317-325.
[16] R. Kathiravan and R. Ganguli, Strength design of
References
composite beam using gradient and particle swarm
[1] D. Liu, V. V. Toroporov, O. M. Querin and D. C. Barton, optimization, Composite Structures, 81 (4) (2007) 471-479.
Bilevel optimization of blended composite wing panels, [17] N. Chang, W. Wang, W. Yang and J. Wang, Ply stacking
Journal of Aircraft, 48 (1) (2011) 107-118. sequence optimization of composite laminate by permutation
[2] R. T. Haftka and J. L. Walsh, Stacking-sequence discrete particle swarm optimization, Structural Multidisci-
optimization for buckling of laminated plates by integer plinary Optimization, 41 (2) (2010) 179-187.
programming, AIAA Journal, 30 (3) (1992) 814-819. [18] H. Li and K. Chandrashekhara, Particle swarm-based
[3] R. Le Riche and R. T. Haftka, Optimization of laminate structural optimization of laminated composite hydrokinetic
stacking sequence for buckling load maximization by turbine blades, Engineering Optimization, 47 (9) (2015)
genetic algorithm, AIAA Journal, 31 (5) (1993) 951-956. 1191-1207.
[4] S. A. Ragon, Z. Gürdal, R. T. Haftka and T. J. Tzong, [19] M. W. Bloomfield, C. G. Diaconu and P. M. Weaver, On
Bilevel design of a wing structure using response surfaces, feasible regions of lamination parameters for lay-up
Journal of Aircraft, 40 (5) (2003) 985-992. optimization of laminated composites, Proceedings of the
[5] P. M. Zadeh, A. Mehmani and A. Messac, High fidelity Royal Society of London A: Mathematical, Physical and
multidisciplinary design optimization of a wing using the Engineering Sciences (2104) 1123-1143.
interaction of low and high fidelity models, Optimization [20] A. E. Assie, A. M. Kabeel and F. F. Mahmoud, Effect of
and Engineering, 17 (3) (2015) 503-532. loading and lamination parameters on the optimum design of
[6] B. Liu, R. T. Haftka and M. A. Akgün, Two-level composite laminated plates, Journal of Mechanical Science and
wing structural optimization using response surfaces, Technology, 25 (5) (2011) 1149-1158.
Structural and Multidisciplinary Optimization, 20 (2) (2000) [21] M. W. Bloomfield, Efficient optimization of laminated
87-96. composites, Ph.D. Thesis, University of Bristol (2010).
[7] D. Liu, V. V Toropov, D. C. Barton and O. M. Querin, [22] M. W. Bloomfield, J. E. Herencia and P. M. Weaver,
Weight and mechanical performance optimization of Enhanced two-level optimization of anisotropic laminated
blended composite wing panels using lamination parameters, composite plates with strength and buckling constraints,
Structural and Multidisciplinary Optimization, 52 (3) (2015) Thin-Walled Structures, 47 (11) (2009) 1161-1167.
549-562. [23] M. Bloomfield, J. E. Herencia and P. M. Weaver,
[8] D. Liu and V. V. Toropov, A lamination parameter-based Optimization of anisotropic laminated composite plates
strategy for solving an integer-continuous problem arising in incorporating nonconventional ply orientations, Proceeding
composite optimization, Computers & Structures, 128 of the 49th AIAA/ASME/ASCE/AHS/ASC Structures, Struc-
(2013) 170-174. tural Dynamics, and Materials Conference, Schaumburg, IL
[9] S. Chen, Z. Lin, H. An, H. Huang and C. Kong, Stacking (2008) 1-15.
sequence optimization with genetic algorithm using a two- [24] J. E. Herencia, P. M. Weaver and M. I. Friswell,
level approximation, Structural and Multidisciplinary Optimization of long anisotropic laminated fiber composite
Optimization, 48 (4) (2013) 795-805. panels with T-shaped stiffeners, AIAA Journal, 45 (10)
[10] A. P. Engelbrecht, Fundamentals of computational swarm (2007) 2497-2509.
intelligence, John Wiley & Sons (2006). [25] F.-X. Irisarri, D. H. Bassir, N. Carrere and J.-F. Maire,
[11] S. Suresh, P. B. Sujit and A. K. Rao, Particle swarm Multiobjective stacking sequence optimization for laminated
optimization approach for multi-objective composite box- composite structures, Composites Science and Technology,
beam design, Composite Structures, 81 (4) (2007) 598-605. 69 (7) (2009) 983-990.
[12] G. Venter and J. Sobieszczanski-Sobieski, Multidisci- [26] B. Liu, Two-level optimization of composite wing
plinary optimization of a transport aircraft wing using structures based on panel genetic optimization, Ph.D. Thesis,
particle swarm optimization, Structural and Multidisci- University of Florida (2001).
plinary Optimization, 26 (1) (2004) 121-131. [27] R. E. Perez and K. Behdinan, Particle swarm approach for
[13] Z. Jing, X. Fan and Q. Sun, Global shared-layer blending structural design optimization, Computers & Structures, 85
method for stacking sequence optimization design and (19) (2007) 1579-1588.
blending of composite structures, Composites Part B: [28] M. H. Sadr and H. G. Bargh, Fundamental frequency
Engineering, 69 (2015) 181-190. optimization of angle-ply laminated plates using Elitist-
[14] P. Y. Jiang, Z. M. Lin, J. Xu and J. Q. Sun, A particle Genetic algorithm and finite strip method, ASME 2010 10th
swarm optimization algorithm for minimizing weight of the Biennial Conference on Engineering Systems Design and
1652 P. M. Zadeh et al. / Journal of Mechanical Science and Technology 32 (4) (2018) 1643~1652

Analysis (2010) 1-10. [38] R. H. Lopez, M. A. Luersen and E. S. Cursi, Optimization


[29] M. H. Sadr, H. G. Bargh, M. K. Nejadi and H. Pourzand, of laminated composites considering different failure criteria,
Free vibration analysis of rotating laminated composite Composites Part B: Engineering, 40 (8) (2009) 731-740.
panels using finite strip method with modified shape [39] Y. Shi and R. Eberhart, A modified particle swarm
functions, ASME 2011 International Mechanical Engineering optimizer, Proceedings of IEEE World Congress on
Congress and Exposition (2011) 787-793. Computational Intelligence Evolutionary Computation
[30] T. Blackwell and J. Branke, Multi-swarm optimization in (1998) 69-73.
dynamic environments, Applications of Evolutionary [40] A. Engelbrecht, Particle swarm optimization: Velocity
Computing (2004) 489-500. initialization, IEEE Congress on Evolutionary Computation
[31] M. G. Alonso and P. Duysinx, Particle swarm optimization (CEC) (2012) 1-8.
(PSO): An alternative method for composite optimization,
10th World Congress on Structural and Multidisciplinary Parviz Mohammad Zadeh is an Assis-
Optimization (2013) 1-10. tant Professor in the Faculty of New
[32] X. Guo, X. Fu, H. Shang and J. Chen, Integrated aero- Sciences and Technologies at the Uni-
structural optimization design of pre-bend wind turbine versity of Tehran. He has an extensive
blades, Journal of Mechanical Science and Technology, 30 research record on various aspects of
(11) (2016) 5103-5113. multidisciplinary design and optimiza-
[33] J.-F. Chang, S. C. Chu, J. F. Roddick and J.-S. Pan, A tion, complex systems design, integra-
parallel particle swarm optimization algorithm with tion, concurrent design, computational
communication strategies, Journal of Information Science modelling, robust and reliability-based design optimization,
and Engineering, 21 (2005) 809-818. meta-modelling, design of experiments, optimization methods
[34] J. E. Herencia, P. M. Weaver and M. I. Friswell, Local that leading to variety of engineering applications.
optimisation of long anisotropic laminated fibre composite
panels with T shape stiffeners, Forty-seventh AIAA/ASME/ Mahdi Fakoor is currently an Associate
AHS/ACS Structures, Structural Dynamics and Materials Professor in the Faculty of New Sci-
Conference, Newport, RI, 1-4 May (2006) 1-25. ences and Technologies, University of
[35] D. Liu, V. V Toropov, M. Zhou, D. C. Barton and O. M. Tehran, Tehran, Iran. His research inter-
Querin, Optimization of blended composite wing panels ests include the fields of fracture me-
using smeared stiffness technique and lamination parameters, chanics and composite materials and
51st AIAA/ASME/ASCE/AHS/ASC Structures, Structural optimization.
Dynamics, and Materials Conference (2010).
[36] J. E. Herencia, R. T. Haftka, P. M. Weaver and M. I.
Friswell, Lay-up optimization of composite stiffened panels Mostafa Mohagheghi is currently Ph.D.
using linear approximations in lamination space, AIAA student in the Faculty of New Sciences
Journal, 46 (9) (2008) 2387-2391. and Technologies, University of Tehran,
[37] M. W. Bloomfield, J. E. Herencia and P. M. Weaver, Iran. His interests lie in the field of com-
Analysis and benchmarking of meta-heuristic techniques for posite structures, bi-level design optimi-
lay-up optimization, Computers & Structures, 88 (5) (2010) zation, multi-objective optimization, and
272-282. engineering design.

Potrebbero piacerti anche