Vous êtes sur la page 1sur 13

Comput. Methods Appl. Mech. Engrg.

198 (2009) 631–643

Contents lists available at ScienceDirect

Comput. Methods Appl. Mech. Engrg.


journal homepage: www.elsevier.com/locate/cma

Fuzzy multiobjective optimization of mechanical structures


F. Massa *, B. Lallemand, T. Tison
Laboratoire d’Automatique de Mécanique et d’Informatique industrielles et Humaines, Université de Valenciennes et du Hainaut Cambrésis,
Le Mont Houy B.P. 311, 59313 Valenciennes, France

a r t i c l e i n f o a b s t r a c t

Article history: This paper proposes a method that takes imperfections in multiobjective optimization problems into
Received 12 December 2007 account. The main objective is to determine a feasible design space for fuzzy design variables that
Received in revised form 18 September respects fuzzy restriction rules and optimizes the defuzzified objective functions. The proposed method
2008
uses the Fuzzy Finite Element Method (FFEM) to calculate fuzzy solutions (namely displacements and fre-
Accepted 19 September 2008
Available online 1 October 2008
quencies) and genetic algorithms for exploring the design space of fuzzy design variables. These elements
are explained in detail and then applied to design a drop-tower impactor sled defined with imperfections.
Ó 2008 Elsevier B.V. All rights reserved.
Keywords:
Imperfection
Fuzzy sets
Multiobjective optimization
Genetic algorithm
Computer-aided design

1. Introduction els. Many multiobjective optimization methods are currently used


during the design phase to improve the performance and robust-
Nowadays, numerical simulations have become an essential ness of mechanical structures. However, given the imperfections
part of the design phase for mechanical structures. With determin- described above, these deterministic methods cannot be used clas-
istic procedures, design variables are implicitly considered to be sically to solve multiobjective optimization problems.
precisely known. However, in reality, this is not the case. In fact, Fuzzy set theory has already been used for the multiobjective
the characteristics of manufactured structures are not always the optimization of mechanical systems and structural designs. For
same, and several kinds of deficiency can be identified at different example, Wang and Wang [7] used a level-cuts approach to solve
steps of their production. These deficiencies can be defined by the non-linear structural problems with only fuzzy constraints. Rao
term ‘‘imperfection”. Two classes can be distinguished: Variability [8] used continuous membership functions for fuzzy constraints
refers to the variation inherent to the physical system or the envi- and fuzzy objectives to optimize academic mechanical systems
ronment under consideration (geometric or material characteris- and structural designs. Kelesoglu [9] has proposed a method for
tics and manufacturing tolerances). Uncertainty is a potential fuzzy multiobjective optimization of space trusses, and Ramík
deficiency in any phase or activity of the modelling process that [10] has developed a class of fuzzy optimization problems with
is due to lack of knowledge (for example joints between two sub- an objective function depending on fuzzy parameters. Though
systems). In order to take such imperfections into account, several the above studies have led to many numerical advances, they do
theories, both probabilistic and non-probabilistic, can be applied. not take all of the imperfections associated with the data used in
Among the available non-probabilistic theories, the fuzzy set multiobjective optimization problems into account. In general,
theory allows parametric imperfections to be managed in order these studies apply non-deterministic methods to academic prob-
to quantify the influence of imperfections on mechanical solutions. lems, such as analytical or truss models.
In this context, different fuzzy methods (e.g., TEEM [1,2], PAEM [3], This paper proposes a method that allows the parametric
Transformation [4,5]) have been developed to propagate imperfec- imperfections in multiobjective optimization problems to be taken
tions in static, modal and dynamic analyses and to aggregate the into account via the fuzzy formalism. The optimization problem is
behavioral changes of the finite element structures. These meth- solved by using global method namely genetic algorithms. The
ods, which belong to the category of Fuzzy Finite Element Methods originality of the proposed method is that each generated individ-
(FFEM) [6], combine the fuzzy formalism with finite element mod- ual are not deterministic but defined by fuzzy numbers. These fuz-
zy data represent the design variables of the problem (for example
thicknesses of plates) and the parameters, which present some
* Corresponding author. Tel.: +33 327511459; fax: +33 327511317.
E-mail address: franck.massa@univ-valenciennes.fr (F. Massa).
deficiency during the industrial or experimental process (for

0045-7825/$ - see front matter Ó 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.cma.2008.09.010
632 F. Massa et al. / Comput. Methods Appl. Mech. Engrg. 198 (2009) 631–643

example Young modulus, density or loads). Moreover, the equality sidered as an extension of the classic set theory. A membership de-
and inequality constraints functions applied on mechanical solu- gree, varying between 0 and 1, is associated to the different values
tions (displacements and frequencies) are modelled as fuzzy ramps of the non-deterministic parameters. In this context, each parame-
to traduce the acceptable variability domain. Finally, by including ter is defined using a membership function that can take different
this fuzzy formalism, the objective is now to find the fuzzy design forms depending on the kind of imperfection considered. Fig. 1
variables that will optimize the fuzzy objective functions defined presents the different forms of fuzzy numbers or fuzzy intervals
by the fuzzy constraint functions. currently in use.
In the next section, the formulation of general multiobjective Using the fuzzy set theory, the multiobjective optimization
optimization problems and the definition of fuzzy numbers are re- problem can be rewritten as follows:
viewed, and the fuzzy formalism is introduced into the optimiza-
tion problem. Section 3 describes the different fuzzy data used to Minimize defuzð e F ðe p np ; e
p1 ; . . . ; e q1; . . . ; e
q nq ÞÞ
define the fuzzy design variables, the fuzzy objective functions Subject to e e e e e
g iðp1; . . . ; pn ; q1; . . . ; qn Þ # R e i ; i ¼ 1; . . . ; I ð2Þ
p q
and the fuzzy constraints functions. Section 4 explains how the
membership functions of fuzzy design variables are updated. This p1; . . . ; e
with e p np
update step exploits the concept of feedback from fuzzy solutions e i are, respectively, the fuzzy de-
where e p np , e
p1; . . . ; e q1; . . . ; e
q nq and R
to define the fuzzy design variables and then identifies those up-
sign variables, which defines the initial design space, the fuzzy
dated fuzzy design variables that respect the fuzzy constraint func-
imprecise parameters, which represents the poorly defined data of
tions. Section 5 focuses on the exploration of the design space of
the problem, and the fuzzy restriction rules. These different fuzzy
fuzzy design variables using a genetic algorithm and on the selec-
data are completely defined in Section 3. The goal of this kind of
tion of the best updated fuzzy design variables from the pool of
problem is now to determine the feasible design space that is de-
these design variables. The best updated fuzzy design variables
scribed by the membership functions of fuzzy design variables
are used to optimize the fuzzy objective functions. Section 6 dem-
p1 ; . . . ; e
e p np . These fuzzy design variables optimize the defuzzified
onstrates the efficiency of the proposed method for designing large
objective functions defuzð e p np ; e
p1 ; . . . ; e
F ðe q1; . . . ; e
q nq ÞÞ and respect
structures with imperfections.
the inclusion of the fuzzy constraint functions
g iðe
e p np ; e
p1 ; . . . ; e q1; . . . ; e
q nq Þ in the fuzzy restriction rules R e i . For fuzzy
2. The fuzzy multiobjective optimization problem optimization problem, the global objective function may be defined
as previously after having defuzzified each fuzzy objective function.
A general multiobjective optimization problem can be mathe- The operator ‘‘defuz” refers to the defuzzification step described in
matically formulated as: Section 5. Further developments will use fuzzy logic to define the
Minimize Fðp1 ; . . . ; pnp Þ ¼ ff1 ðp1 ; . . . ; pnp Þ; . . . ; fno ðp1 ; . . . ; pnp Þg fuzzy weights of the different objective functions.
k 6 pk 6 pk with k ¼ 1; . . . ; np
Subject to g i ðp1 ; . . . ; pnp Þ 6 0; i ¼ 1; . . . ; Ip
3. Description of the fuzzy optimization problem’s data
ð1Þ
where f1 ðp1 ; . . . ; pnp Þ; . . . ; fno ðp1 ; . . . ; pnp Þ are the individual objective This section focuses on the different kinds of fuzzy data used in
functions; g i ðp1 ; . . . ; pnp Þ are the inequality (or equality) constraint the fuzzy optimization problem, namely fuzzy design variables,
functions; pk and p k are, respectively, the lower and upper bounds fuzzy imprecise parameters, fuzzy solutions and fuzzy restriction
of parameter pk; and no, np and I are, respectively, the number of rules. The objective of the section is to define these data and to ex-
objective functions, design parameters and constraint functions. plain how imperfections can be identified and/or quantified and
The goal of this problem is to find the design variables p1 ; . . . ; pnp represented by fuzzy numbers.
that optimize a vector of the objective functions Fðp1 ; . . . ; pnp Þ for
the feasible design space defined by the constraint functions. The 3.1. Fuzzy design variables
different objectives functions fk may be antagonistic and that it
may not always be possible to find a common set of design param- Fuzzy design variables p e1 ; . . . ; e
p np (Fig. 2) characterize the initial
eters that minimizes them all simultaneously. In this case, a global design space. A fuzzy design variable is modelled as a triangular
objective function is defined as a linear combination of each objec- fuzzy number, whose crisp value corresponds to a previous or ini-
tive function. tial design and whose support represents the authorized range of
In order to take imperfections into account, the fuzzy formalism variation, defined by the design variable bounds p k and pk with
is used. Introduced by Zadeh [11–14], fuzzy set theory can be con- k = 1, . . ., np. At the design stage, only three data (nominal value

Fig. 1. Current membership function forms: (a) triangular; (b) trapezoid.


F. Massa et al. / Comput. Methods Appl. Mech. Engrg. 198 (2009) 631–643 633

modulus experimentally, measurement errors are inevitable,


reaching nearly 10% in vibrometric experimental analyses, for
example. Fuzzy imprecise parameters can also be defined using or-
dinary intervals (Fig. 4b), which allow an identical range of param-
eter variation to be imposed for each degree of membership a.
Different membership functions can so be chosen (triangular, p-
shape, trapezoidal...) to model the perception of the user. It is
important to note that the choice of the membership function is
a way to influence the final feasible design space. The form of
membership functions of fuzzy solutions is dependant of the form
of membership functions of input parameters.

3.3. Fuzzy solutions

Fuzzy output solutions es 1 ; . . . ; es ns define the fuzzy constraint


functions e g iðe p np ; e
p1 ; . . . ; e q1; . . . ; e
q nq Þ. These data are calculated di-
rectly from the fuzzy design variables e p1; . . . ; e
p np and fuzzy impre-
cise parameters e q1; . . . ; eq nq for the different mechanical analyses
(e.g., static or dynamic) using a fuzzy finite element method
Fig. 2. Description of fuzzy design variables.
(FFEM) whose objective is to propagate the imperfections identi-
fied in the input data towards the output data.
and bounds of each design variables) are available, triangular Computing with fuzzy numbers requires a convention on how
membership functions are so the simplest and common choice. to handle the combination of fuzzy numbers that’s why an appro-
Therefore, these fuzzy numbers allow to introduce more flexi- priate t-norm has to be chosen. This work applies the t-min norm,
bility in the optimization process. The form of the membership which is by far the most widely adopted t-norm in fuzzy numer-
function can authorise more or less important variations on the de- ical analysis. Its main benefit is that by applying the t-min norm
sign parameter. By using the different forms presented in Fig 3, the within the Zadeh’s extension principle [11], the numerical result
designer can control the optimization process. For example, if the of a combination of fuzzy variables can be calculated using the
previous membership functions are associated to dimensions of a a-sublevel technique. The membership functions of the fuzzy
structure, the membership function (Fig. 3a) corresponds to a parameters were transformed into a set of intervals. In a previous
dimension, which is difficult to realize around the nominal value. article [3], we proposed an FFEM, called the Padé approximants
The second case (Fig. 3b) corresponds to a dimension, which can with extrema management (PAEM) method, to propagate the
be realized more precisely and easily. The last case (Fig. 3c) corre- imperfections for each a-cut level. This method is summarized
sponds to a intermediary case; that is to say that the designer do briefly in Annexe A. The PAEM method uses sensitivity techniques
not prejudge the realization. and high-order approximations to calculate fuzzy solutions
During the optimization process, these data are first updated, es 1 ; . . . ; es ns for static and modal analyses. This method requires that
thus generating the updated fuzzy design variables, which verify the membership function be discretized according to the degree
the constraint functions. Then, the different updated fuzzy design of membership. The problem is then transformed into an interval
variables are evaluated by genetic algorithms in order to select problem for each a-cut level to define the initial discretized de-
the best design. These best updated fuzzy design variables are used sign space.
to optimize the fuzzy objective functions. The updating and selec- The PAEM strategy follows a two-step procedure:
tion steps are described in Sections 4 and 5, respectively.
 A search for parameter combinations, which implies minimum
3.2. Fuzzy imprecise parameters and maximum variations, is performed for each a-cut level.
The sensitivity of the studied solutions is evaluated between
Fuzzy imprecise parameters q e1 ; . . . ; e
q nq are used to represent each level to determine the evolution of the response function.
poorly defined data, notably material properties or geometric char-  The modified solutions are approximated for each selected com-
acteristics (Fig. 4a). For example, the Young modulus for standard bination. A high-order approximation using Padé rational func-
steel is generally fixed at 2  1011 N/m2, but can vary between tions is required in order to decrease the calculation time and
1.9  1011 N/m2 and 2.1  1011 N/m2. For this reason, it is more maintain a good level of accuracy. (More information about
acceptable to describe this kind of data with fuzzy numbers rather Padé rational functions can be found in the previous study by
than use a deterministic description. When determining the Young Massa et al. [3].)

Fig. 3. Different forms of fuzzy design variables.


634 F. Massa et al. / Comput. Methods Appl. Mech. Engrg. 198 (2009) 631–643

Fig. 4. Description of fuzzy imprecise parameters.

Fig. 5. Graphical representation of the PAEM method in the design space.

The calculation of the parameter value combinations in the


PAEM method is represented graphically in Fig. 5 for the case of
two parameters.

3.4. Fuzzy solutions

e i (Fig. 6) are constraints applied to fuzzy


Fuzzy restriction rules R
solutions and are defined in terms of the structural specifications.
These restrictions cannot be handled deterministically, since the
input data (i.e., design variables and imprecise parameters) are
themselves fuzzy. For example, the yield stress of standard steel
is generally set at 350 MPa, but can vary between 340 MPa and
360 MPa. For this reason, the restriction rules are defined using a
fuzzy formalism that allows a certain degree of acceptability to
be incorporated into the design. Fig. 6 illustrates a situation in
which there are three domains of acceptability: poor performance Fig. 6. Definition of fuzzy restriction rules.
corresponds to a degree of membership of 0, acceptable perfor-
mance corresponds to a degree between 0 and 1, and good perfor-
mance corresponds to a degree of 1. This situation represents a fuzzy restriction rules. The information included in the fuzzy solu-
gradation around a reference value. tions’ membership functions is exploited directly.
The membership functions of the fuzzy solutions represent
4. Updating the fuzzy design variables structural behavior variations. Thus, an a-cut level of design vari-
ables can be associated to an a-cut level of fuzzy solutions and vice
This section explains how the membership functions of fuzzy versa. In this way, the process exploits the concept of feedback
design variables can be updated according to fuzzy solutions and from fuzzy solutions to define the fuzzy design variables.
F. Massa et al. / Comput. Methods Appl. Mech. Engrg. 198 (2009) 631–643 635

Practically, this updating process is organized in three steps: are as close as possible to 1 and 0, respectively, (Fig. 9) an iterative
process is employed.
 Superposition of the fuzzy restriction rules on fuzzy solutions. The updated fuzzy design variable’s membership functions
The fuzzy restriction rules are superimposed on the membership (Fig. 10) represent the feasible design space that respects the inclu-
functions of the fuzzy solutions. The intersection of the fuzzy sion of fuzzy constraint functions in the fuzzy restriction rules.
restriction rules and the fuzzy solutions defines two a-cut lim- Fig. 11 illustrates the result of this updating step for the case of
its: LGA (good-acceptable limit) and LAP (acceptable-poor limit) two design variables. The larger square presents the initial design
for each fuzzy solution (Fig. 7). These a-cut limits represent space defined by the bounds of the design variables p k and pk with
the boundaries of the different degrees of acceptability. k = 1, . . ., np. In this design space, the range of solution acceptability
 Determination of the final degrees of acceptability. (good solutions, acceptable solutions, poor solutions) is presented
Since the limits LGA and LAP can be different for each ‘‘rule/solu- in shades of gray, and the solution acceptability is defined using
tion” couple, some final acceptability limits, noted LFGA and LFAP , the fuzzy restriction rules. The central squares represent the feasi-
must be defined. The maximum limits, chosen so that all the ble design spaces of the updated fuzzy design variables. These
restrictions are respected, are defined using Eq. (3)
n o
LFGA ¼ max L1GA ; . . . ; LIGA
n o ð3Þ
LFAP ¼ max L1AP ; . . . ; LIAP

 Determination of the updated fuzzy design variable sets.

The two final limits, LFGA and LFAP , are superimposed on the mem-
bership functions of the fuzzy design variables. If no fuzzy impre-
cise parameter has been defined, the membership functions of the
updated fuzzy design variables can be obtained directly. These
functions are constructed so that the interval of membership for
a = 0 corresponds to the interval of limit LFAP and the interval of
membership for a=1 to the interval of limit LFGA (Fig. 8). However,
if the problem includes imprecise parameters, the solution cannot
be obtained directly because the obtained limits do not take all the
bounds of fuzzy imprecise parameters into account. To update the
design variable membership functions so that the LGA and LAP limits Fig. 9. Updating the fuzzy solution’s membership functions.

Fig. 7. Application of fuzzy restriction rules to fuzzy solutions.

Fig. 8. Updating the fuzzy design variable’s membership functions.


636 F. Massa et al. / Comput. Methods Appl. Mech. Engrg. 198 (2009) 631–643

feasible design set) that verify the constraint functions, while also
taking the possible imperfections of fuzzy imprecise parameters
into account.
Different updated fuzzy design variables are built by changing
the nominal values of initial fuzzy design variables and then select-
ing the design variables that best optimize the fuzzy objective
functions (e.g., minimum cost, minimum mass or maximum
robustness). The best design space is then defined using the best
updated fuzzy design variables.
Using the genetic algorithm proposed by Holland [15], the
above strategy can be automated to allow a great number of pos-
sible feasible design spaces to be explored. Genetic algorithms sim-
ulate natural evolutionary processes by using the main classic
operators: crossover, mutation and selection. The use of fuzzy data
does not completely modify the classic genetic algorithm and oper-
Fig. 10. Description of updated design variable’s membership function. ators (see Appendix B for a brief review of GA). For the case pre-
sented here, the different generations are composed of
design spaces can be qualified as acceptable and/or good in terms individuals defined with fuzzy design variables, which constitute
of the membership degree of the solution. These feasible design the individual’s chromosomes. The genes of each chromosome
spaces contain all the values of design variables for which the fuzzy are the different values that can be used as the nominal values of
constraints functions have been verified. fuzzy design variables. Since the selection operator is defined with
a deterministic value, the fuzzy objective functions must necessar-
5. Selecting the updated fuzzy design variables ily be defuzzified, which is accomplished classically using the
method ‘‘center of gravity” or ‘‘area” of the membership functions
As Fig. 12 shows, the location of feasible design space is directly [16].
dependant on the nominal value of the fuzzy design variables. So, The Fig. 13 superimposes the feasible design space obtained
by using a variety of nominal values to define the fuzzy design with the initial nominal value and one of feasible design space ob-
variables and by applying the updating step presented above, the tained after genetic algorithm. The new feasible design space is lo-
initial design space can be completely explored. The objective of cated in the lower left of the initial design space. This figure
this section is to determine the best feasible design space in order highlights the evolution of the feasible design space in function
to provide the designer with a set of design values (i.e., the best of updated fuzzy design variables and nominal value of the fuzzy

Fig. 11. Illustration of the design space of fuzzy design variables and updated fuzzy design variables.
F. Massa et al. / Comput. Methods Appl. Mech. Engrg. 198 (2009) 631–643 637

Fig. 12. Design space exploration of fuzzy design variables.

Fig. 13. Description of the feasible design space obtained with the genetic algorithm.
638 F. Massa et al. / Comput. Methods Appl. Mech. Engrg. 198 (2009) 631–643

design variables. By using different nominal values to define the manufacturing costs. In fact, the designer wanted the new ver-
fuzzy design variables, we explore the initial design space with sion of the impactor sled to respect certain specifications, nota-
success to find the best updated design space. bly a maximal displacement less than 1 mm and a first
The next section presents an industrial test case in which the eigenfrequency greater than 250 Hz. In addition, the designer
proposed fuzzy multiobjective optimization method was applied. wanted the impactor sled to have a minimal mass and be robust
to imperfections. Since the material characteristics and loads
6. Numerical application were not sufficiently well-known, the Young modulus, density
and static loads had to be considered as fuzzy. The two design
A study of the strategy presented above was completed on a variables were the plate thicknesses that had to be optimized
drop-tower impactor sled currently used for dynamic compres- during the process.
sion and bending tests. The general goal was to improve the The fuzzy multiobjective optimization problem was defined as
overall mechanical performance of this structure while reducing follows:

Fig. 14. Definition of fuzzy design variables.

Fig. 15. Definition of fuzzy imprecise parameters.


F. Massa et al. / Comput. Methods Appl. Mech. Engrg. 198 (2009) 631–643 639

e max ; ef 1 Þ e max and e


Minimize defuzð mðe ee1; e e2; e e ÞÞ & Maximize rsp ð U
E; q modulus and the density; U f 1 are, respectively, the fuzzy
e e e e maximal displacements and the fuzzy first eigenfrequency; and
Subject to U max ðe
e1; e
e 2 ; E; F Þ # R U e U and R
R e f are the restriction rules applied, respectively, to the fuzzy
ef 1 ðe e2; e
e1; e ef
eÞ # R
E; q maximal displacements and the fuzzy first eigenfrequency.
ð4Þ The optimization problem required the following fuzzy data:

where m e and rsp are, respectively, the fuzzy mass of the structure e 1 and e
 Design variables: The thicknesses e e 2 had to be updated,
and robustness criterion of the fuzzy solutions; e e2; e
e1 ; e F; e e
E and q with an authorized variability of ±20% in the nominal values
are, respectively, the plate thicknesses, the static load, the Young (Fig. 14).

Fig. 16. Finite element model of the impactor sled.

Fig. 17. Definition of fuzzy restriction rules.

Fig. 18. Different cases of static loading.


640 F. Massa et al. / Comput. Methods Appl. Mech. Engrg. 198 (2009) 631–643

 Fuzzy solutions: Fuzzy solutions U e max and ef 1 were calculated


using PAEM method summarized in Section 3. The finite ele-
ment model associated with the impactor sled (Fig. 16) con-
tained 37,962 degrees of freedom (shell and brick elements).
Restraint conditions were applied to eight holes.
 Restriction rules: Restriction rules were applied (Fig. 17) to the
maximum fuzzy displacements U e max2 calculated for
e max1 and U
the different loading cases (Fig. 18) and to the fuzzy first eigen-
frequency e f 1 . The limits were defined based on previous accept-
able designs that corresponded to the study specifications.
 Objective functions: The fuzzy sets were optimized to produce a
minimum mass and a maximum robustness for the structure.
The mass criterion m e was evaluated by defuzzifying the corre-
sponding fuzzy number using the center of gravity method. A
solution was considered to be robust with respect to a parame-
ter if a large parameter variation implied a small solution varia-
tion. The dimensionless criterion rsp was evaluated using Eq. 5,
which can be stated as follows: the smaller the criterion, the
more robust the solution,
Ds pc
rsp ¼ ð5Þ
Dp sc
where Ds and Dp represent the areas of the membership func-
tions for the solution s and the parameter p, respectively. pc
and sc are the crisp values.
The proposed methodologies are implemented with MATLAB
and the Structural Dynamics Toolbox. The CPU time for this fuzzy
Fig. 19. Fuzzy solution’s membership functions (before and after the updating
optimization was 180mn (Pentium 4 2000 MHz).
step). First, the updating step was performed on the fuzzy data de-
fined above. Fig. 19 shows the fuzzy solutions before and after
updating. The restriction rule for the first eigenfrequency being
more restrictive than the ones for the maximum displacements
(first load), no specific updating was necessary for the second load
because the rule is always verified. The updated fuzzy design vari-
ables are shown in Fig. 20.
Second, the selection step was performed to obtain the different
adjusted design variables’ membership functions. Fig. 21 shows the
different feasible design spaces (gray) generated for l = 0 and the
best set (black), corresponding to the highest criterion values of
defuzzified objective functions. The best updated fuzzy design var-
iable’s membership functions and the corresponding fuzzy solu-
tions are illustrated in Figs. 22 and 23, respectively. Fig. 24
shows the stagnation of the robustness of the structure whereas
Fig. 25 highlights an improvement of the performance of the fuzzy

Fig. 20. Design variable’s membership functions (before and after the updating
step).

 Imprecise parameters: The material tested was standard alumin-


ium, but no other information was known. Variability was intro-
duced into the Young modulus e E and the density qe values in
order to represent the entire variation interval [13]. Variability
was also introduced into the load range e F , which was repre-
sented as an ordinary interval in order to take the most pessi-
mistic case into account (Fig. 15). Fig. 21. Different feasible design spaces obtained with the genetic algorithm.
F. Massa et al. / Comput. Methods Appl. Mech. Engrg. 198 (2009) 631–643 641

Fig. 24. Robustness of the solutions.

Fig. 22. Design variable’s membership functions (before and after the selection
step).

Fig. 25. The mass structure’s membership function.

Fig. 23. Fuzzy solution’s membership functions (before and after the selection
step).

Fig. 26. Determination of the membership functions of fuzzy solutions.


mass. The center of gravity of membership functions of the fuzzy
mass is lower after the selection process.
This study allowed the designer some latitude in defining the thickness e1 between 7.45 mm and 9.1 mm and a thickness e2 be-
thicknesses of each plate. For a thickness e1 between 8.4 mm and tween 8.55 mm and 10.7 mm is considered acceptable. Thick-
8.7 mm and a thickness e2 between 9.1 mm and 9.45 mm, the nesses over or under these limits do not respect the fuzzy design
structure meets and exceeds its performance requirements; a specifications.
642 F. Massa et al. / Comput. Methods Appl. Mech. Engrg. 198 (2009) 631–643

Fig. 27. Crossover and mutation operators in a genetic algorithm.

7. Conclusion function and define the combinations of discrete fuzzy parame-


ter values for the following a-cut level, which could supply the
A multiobjective optimization method, which takes parametric minimum and maximum variations (Step 1).
imperfections into account, was developed to facilitate the design For each a-cut level:
of more robust structures. The imperfections are modelled using  Evaluate the first-order derivatives of the quantities for the com-
fuzzy theory. The fuzzy solutions were optimized by updating binations of discrete fuzzy parameter values determined at the
the fuzzy design variables to verify the fuzzy constraint functions previous a-cut level (Step 2).
and by selecting the best updated fuzzy design variables to opti-  Compare the signs of the derivatives with those obtained at the
mize the fuzzy objective functions. A large model numerical appli- previous a-cut level.
cation was tested to show the efficiency of the method.  If the derivatives have the same signs, the response function
The proposed method allows both input data variability and de- is considered to be locally monotonic, and the determined
signer judgment and/or experience to be taken into account in the combinations provide the minimum and maximum varia-
design phase in an effort to produce robust optimal mechanical tions of the studied quantities for the current a-cut level
behavior that satisfies the product specifications. Since we con- (Step 3a).
sider computer-aided design (CAD) methodologies to be relatively  If the derivatives have different signs, the response function
close to updating procedures, our future experiments will consist cannot be considered as monotonic, giving rise to an extreme
of adapting our updating methodology for modal updating. This value between these two a-cut levels. The combination near-
will allow both the notion of expert judgment with respect to est the extreme value is chosen and the search is stopped for
experimental modal parameters and data sets obtained under dif- this variation (Step 3b).
ferent experimental conditions to be introduced into the updating
procedure.
Calculate the modified solutions for the selected combinations
Acknowledgments of discrete fuzzy parameter values. The MAC criterion is applied
for fuzzy modal analysis to verify the shape of the modes. To de-
The present research work has been supported by International crease the calculation time and maintain a good level of accuracy,
Campus on Safety and Intermodality in Transportation the Nord- the ‘‘exact” calculation (corresponding to a deterministic finite ele-
Pas-de-Calais Region, the European Community, the Regional Del- ment simulation) is replaced by a high-order approximation using
egation for Research and Technology, the Ministry of Higher Edu- Padé rational functions.
cation and Research, and the National Center for Scientific
Research. The authors gratefully acknowledge the support of these Appendix B. Genetic algorithms
institutions.
Genetic algorithms consist of optimization procedures based
Appendix A. PAEM algorithm on principles inspired by natural evolution. First, many individual
solutions are randomly generated to form an initial population,
The PAEM algorithm (Fig. 26) consists of the following steps: with the population size depending on the nature of the problem
For the crisp values (a = 1): (minimum 15 members). Traditionally, the population is gener-
ated randomly to cover the entire design space. Each individual
 Determine the studied quantities and their first-order sensitivi- in a population is a chromosome. Then, to create the next gener-
ties for each fuzzy parameter. The signs of the first-order sensi- ation (minimum 10 generations), three classic operators are
tivities indicate the functional dependence of the response used:
F. Massa et al. / Comput. Methods Appl. Mech. Engrg. 198 (2009) 631–643 643

 Crossover: There are many different kinds of crossovers. The References


most common type is the single-point crossover. This operator
combines genes from two chromosomes and creates new ones. [1] F. Massa, B. Lallemand, T. Tison, P. Level, Fuzzy eigensolutions of mechanical
structures, Engrg. Comput. 21 (1) (2004) 66–77.
In single-point crossovers, a locus is chosen at which the [2] F. Massa, T. Tison, B. Lallemand, A fuzzy procedure for the static design of
remaining alleles from one parent are swapped for the other. imprecise structures, Comput. Methods Appl. Mech. Engrg. 195 (2006) 925–941.
As shown in Fig. 27, the children take one section of the chromo- [3] F. Massa, K. Ruffin, T. Tison, B. Lallemand, A complete method for efficient
fuzzy modal analysis, J. Sound Vib. 309 (1–2) (2008) 63–85.
some from each parent. The point at which the chromosome is [4] M. Hanss, The transformation method for the simulation and analysis of
broken depends on the randomly selected crossover point. The systems with uncertain parameters, Fuzzy Sets Syst. 130 (3) (2002) 277–289.
probability of a crossover occurring is usually 60–70%. A random [5] S. Donders, D. Vandepitte, J. Van de Peer, W. Desmet, Assessment of
uncertainty on structural dynamic responses with the short transformation
two-point crossover is also possible.
method, J. Sound Vib. 288 (2005) 523–549.
 Mutation: After the crossover step, some individuals are copied [6] S.S. Rao, P. Sawyer, Fuzzy finite element approach for the analysis of
directly to create the new generation; others undergo mutation imprecisely defined systems, AIAA J. 33 (12) (1995) 2364–2370.
in order to ensure that the individuals are not all exactly the [7] G.Y. Wang, W.Q. Wang, Fuzzy optimum design of structures, Engrg. Optimiz. 8
(1985) 291–300.
same. As shown in Fig. 27, a loop is made through all the alleles [8] S.S. Rao, Multi-objective optimization of fuzzy structural systems, Int. J.
of all the individuals. If an allele is selected for mutation, it may Numer. Methods Engrg. 24 (1978) 1157–1171.
either undergo a small change or be replaced by a new value. [9] O. Kelesoglu, Fuzzy multiobjective optimization of truss-structures using
genetic algorithm, Adv. Engrg. Softw. 38 (2007) 717–721.
The probability of a mutation occurring is usually 1–5%. [10] J. Ramík, Optimal solutions in optimization problem with objective function
 Selection: During each successive generation, a proportion of the depending on fuzzy parameters, Fuzzy Sets Syst. 158 (2007) 1873–1881.
existing population is selected to breed a new generation. Indi- [11] L.A. Zadeh, Fuzzy sets, Inform. Control 8 (1965) 338–353.
[12] L.A. Zadeh, The concept of a linguistic variable and its application to
vidual solutions are selected through a fitness-based process. approximate reasoning – Part I, Inform. Sci. 8 (1975) 199–249.
Generally, the selection methods rank the fitness of each solu- [13] L.A. Zadeh, The concept of a linguistic variable and its application to
tion and select the best solutions. approximate reasoning – Part II, Inform. Sci. 8 (1975) 301–357.
[14] L.A. Zadeh, The concept of a linguistic variable and its application to
approximate reasoning – Part III, Inform. Sci. 8 (1975) 43–80.
These different operators are used to maintain a sufficiently [15] J. Holland, Adaptation in Natural and Artificial Systems, University of Michigan
large population diversity, preventing premature convergence on Press, Ann Arbor MI, 1975.
[16] D. Dubois, H. Prade, Fuzzy sets and systems: theory and applications, in: R.
poor solutions. More details about genetic algorithms for multiob-
Bellman (Ed.), Mathematics in Science and Engineering, Academic Press,
jective optimization problems are available in the literature. London, 1980.

Vous aimerez peut-être aussi