Beruflich Dokumente
Kultur Dokumente
i
1
W x
5
Fig. 4. NN errors on the training and test sets.
242 C. Bisagni, L. Lanzi / Composite Structures 58 (2002) 237247
where x represents the DNA of the structure, W x the
panel weight, mx; c
i
i
, i 1; . . . ; 3 three dierent expo-
nential penalty functions used to evaluate the con-
straints upon the buckling load, c
1
, the collapse load, c
2
,
and the pre-buckling stiness, c
3
, respectively, and
Q
mx; c
i
i
represents the product of the three penalty
functions.
The GA used in the present investigation are imple-
mented in Matlab [14]. Only three basic genetic opera-
tors are used: selection, crossover and mutation. The
selection operator produces or eliminates the members
characterised by high or low tness values, respectively.
A common approach is to assign a probability of se-
lection on the basis of the tness function to each
member of an initial population. The crossover process
allows the exchange of the characteristics among a se-
lected set of the population. In general two members,
called parents, are chosen among the initial population
and their DNA is randomly combined bearing two new
members, called children. The mutation operator is in-
troduced to safeguard the process from a premature loss
of genetic information during crossover. In particular a
few members of the population are chosen and their
DNA is randomly changed according to a probability of
mutation.
A single binary string 16 bits long codies the genetic
information of any member of the population in a given
generation. The rst two bits describe the lay-up of the
skin, then three bits represent the number of layers, nine
bits the side dimension of the stieners and the last two
bits take into account their number.
7. Optimisation results
The optimisation problem is formulated in order to
nd the number of layers at 45 in the skin of the
panel, the number of layers in the stieners, the side
dimension of the stieners as well as their number so to
minimise the weight of the panel subject to the following
constraints:
P
cr
X > 75 kN
P
co
X > 135 kN
S
pb
X > 120 kN=mm
8
<
:
6
The values of the minimum buckling load and collapse
load are evaluated considering panels made of alumin-
ium alloy and actually used in helicopter eld. With
these constraints, a panel conguration characterised by
a post-buckling eld with local skin buckling between
the stieners till the collapse load and by an high ratio
between the collapse load and the buckling load is ex-
pected.
7.1. Optimised conguration
The genetic search is performed with an initial popu-
lation of 40 members, generated randomly inside the
optimisation domain. The probability of crossover and
mutation are xed to 0.75 and 0.05, respectively.
The GA converged after 19 generations and required
760 tness function evaluations.
The optimised panel is characterised by a weight of
18.98 N and presents three layers at 45 in the skin, six
stieners, seven layers in the stieners and side dimen-
sions of the stieners equal to 24.5 mm.
The solution obtained using the NN system is veried
by means of nite element analyses using ABAQUS.
The buckling load is calculated by an eigenvalue ana-
lysis while the load versus displacement curve is ob-
tained by a dynamic analysis.
The optimised conguration is characterised by a
ratio between the collapse load and the buckling load
equal to 2.05 and satises the imposed constraints as
reported in Table 4, where the dierences between the
NN system and the nite element analysis are also re-
ported. The NN response shows a good approximation
level if compared to the values obtained performing
the validation nite element analyses. A direct com-
parison of the load versus displacement curves obtained
Table 4
Optimised conguration
Panel weight (N) 18.98
Panel behaviour Con-
straints
FE ana-
lysis
NN sys-
tem
% Dier-
ence
Buckling load (kN) >75 79.94 79.3 0.8
Collapse load (kN) >135 164.5 162.5 1.2
Pre-buckling
stiness (kN/mm)
>120 124.7 129.8 4.1
Displacement at the
buckling load (mm)
0.641 0.611 4.7
Fig. 5. Comparison between the load versus displacement curves ob-
tained by NN system and nite element analysis.
C. Bisagni, L. Lanzi / Composite Structures 58 (2002) 237247 243
respectively by the NN system and by the non-linear
dynamic analysis is presented in Fig. 5.
Fig. 6 shows the deformed shape evolution, obtained
by the non-linear nite element analysis, with a magni-
cation factor equal to 4. It is possible to see that the
post-buckling eld is characterised by a local skin buck-
ling between the stieners before the collapse.
The adopted optimisation procedure based on the
GA coupled with the NN system has allowed a con-
siderable reduction of the total computational costs.
Indeed, to nd the optimal conguration, 55 nite ele-
ment non-linear analyses were performed in a total CPU
time of about 660 h using a parallel machine HP N4000
with eight processors and 8 GB of RAM. To this it must
be added the CPU time required by the neural networks
training processes and all the optimisation runs, that is
comparable to a single nite element analysis. In com-
parison, a direct optimisation, that uses directly the -
nite element analyses to evaluate the objective function
coupled with GA would have required about 760 dif-
ferent simulations, that means about 9480 h of CPU
times.
7.2. Failure considerations on the optimised conguration
One of the most critical aspects of structural opti-
misation problems involving composite materials con-
cerns the introduction of verication and design criteria
considering failure and progressive damage mecha-
nisms. So, even if during the optimisation procedure no
failure and damage constraints were directly considered,
the optimised panel conguration is then veried to
check that there was no failure in the post-buckling eld
before the collapse load.
A non-linear dynamic nite element analysis is then
performed on the optimised structure adopting the Tsai
Hill failure criterion [2022] as available in ABAQUS
[13].
According to the TsaiHill criterion, a failure index
(here indicated as F) is dened, once known the strength
values of the material rr
ij
:
F
r
2
11
rr
2
11
r
11
r
22
rr
2
11
r
2
22
rr
2
22
r
2
12
rr
2
12
7
Local failures happen when the failure index F becomes
>1. The verication analysis shows that the failure index
of the nal conguration is lower than 1 till the collapse
load as reported in Fig. 7, where the contour of the
failure index F is plotted. Consequently the obtained
conguration results able to work in the entire post-
buckling eld without any local failure. From Fig. 7, it is
also possible to see that the most critical areas in the
post-buckling eld moves from the skin to the stieners.
The structural collapse happens for buckling of the
Fig. 6. Deformed shape evolution of the optimised panel.
244 C. Bisagni, L. Lanzi / Composite Structures 58 (2002) 237247
stieners and the failure occurs in the stieners blades
and anges due to high bending and shear deformations.
7.3. Weight reduction considerations
The main advantage of the optimisation procedure
here presented is the complete separation between the
system modelling and the optimisation problem. In this
way it is possible to run dierent optimisations changing
either constraints or objective functions without any
other nite element analysis. Indeed, the trained system
of parallel NN is able to reproduce the behaviour of the
structural component.
This advantage is used here to compare the obtained
optimised conguration, able to work in the post-
buckling eld with a local skin buckling between the
stieners, to a second conguration obtained under the
hypothesis that the panel remains in the pre-buckling
eld until the collapse load.
This second optimisation is performed in order to
minimise the panel weight, subjected to two dierent
constraints. The rst constraint is on the buckling load
that must be higher than the collapse load, equal to 135
kN, and the second constraint is on the initial stiness
that must be higher than 120 kN/mm. The genetic search
is performed with an initial population of 40 members
and probability of crossover and mutation equal to 0.75
and 0.05 respectively.
This second panel optimised conguration is charac-
terised by a weight of 23.26 N and presents four layers at
45 in the skin, six stieners, 10 layers in the stieners
and side dimensions of the stieners equal to 22 mm.
As reported in Table 5, the conguration meets the
initial design requirements and also in this case the NN
Fig. 7. Failure identication with TsaiHill criterion.
C. Bisagni, L. Lanzi / Composite Structures 58 (2002) 237247 245
response shows a good approximation level if compared
to the values obtained by the validation nite element
analyses. Comparing the rst optimised conguration,
where a local skin buckling between the stieners is
allowed, to this second conguration, the reduction
in weight is equal to 18%. It is so possible to take
advantage from the post-buckling eld to reduce sig-
nicantly the structural weight of stiened composite
structures.
8. Conclusions
A procedure for the post-buckling optimisation of
composite stiened panels is here presented. Due to the
presence of integer variables, as the number of stieners
and the panel lay-up, the use of genetic algorithms ap-
pears particularly protable, but their direct application
together with non-linear nite element analyses result
unfortunately too expensive from a computational point
of view. The proposed procedure, based on a global
approximation, strongly reduces the total computa-
tional time required for the strategy of the optimisation,
where the non-linear nite element analyses are replaced
by a system of MLP neural networks. Besides, the use of
a system of several parallel NN allows to obtain not
only specic structural indices, such as the buckling
load, the collapse load and the corresponding displace-
ments, but also the whole approximate load versus dis-
placement curve.
Another advantage oered by this approach is the
complete separation between the system modelling and
the optimisation problem. In this way it is possible to
run dierent optimisations changing the minimisation
algorithms, the constraint values or the objective func-
tion without any other nite element analysis. This
advantage is here used to compare the optimised con-
guration, able to work in the post-buckling eld with a
local skin buckling between the stieners, to a second
conguration obtained under the hypothesis that the
panel remains in the pre-buckling eld until the collapse
load. The results show that, where a local skin buckling
between the stieners is allowed, the reduction in weight
is equal to 18%. Allowing a local skin buckling seems so
a protable way to reduce signicantly the weight of
stiened composite structures.
The optimised conguration has been then veried
using the TsaiHill failure criterion to see that there
was no failure in the post-buckling eld before the col-
lapse load, even if during the optimisation procedure
no failure and damage constraints were directly con-
sidered.
The optimisation procedure appears exible and ap-
plicable to various engineering problems in which the
results are given by complex and expensive models and
a high number of analyses are necessary to reach a
suitable optimum.
Acknowledgements
The authors would like to thank Prof. Vittorio
Giavotto for his generous advises and Prof. Sergio Ricci
for the useful discussions on the formulation of the
optimisation problem. The support provided by the
European Commission, Competitive and Sustainable
Growth Programme, Contract No. G4RD-CT-1999-
00103, project Improved Postbuckling Simulation
for Design of Fibre Composite Stiened Structures
(POSICOSS) is gratefully acknowledged.
References
[1] Romeo G, Frulla G. Post-buckling behaviour of graphite/epoxy
stiened panels with initial imperfections subjected to eccentric
biaxial compression loading. Int J Non-Linear Mech 1997;
32(6):101733.
[2] Bisagni C. Numerical analysis and experimental correlation of
composite shell buckling and post-buckling. Composites: Part B
2000;31:65567.
[3] Meyer-Piening H-R, Farshad M, Geier B, Zimmermann R.
Buckling loads of CFRP composite cylinders under combined
axial and torsion loadingexperiments and computations. Comp
Struct 2001;53:42735.
[4] Park O, Haftka RT, Sankar BV, Starnes Jr. JH, Nagendra S.
Analytical-experimental correlation for a stiened composite
panel loaded in axial compression. J Aircr 2001;38(2):37987.
[5] Weller T, Singer J. Durability of stiened composite panels under
repeated buckling. Int J Solids Struct 1990;26(9):103769.
[6] Butler R, Williams FW. Optimum design using VICONOPT, a
buckling and strength constraint program for prismatic assemblies
of anisotropic plates. Comput Struct 1992;43(4):699708.
[7] Wiggenraad JFM, Arendsen P, da Silva Pereira JM. Design
optimization of stiened composite panels with buckling and
damage tolerance constraints. American Institute of Aeronautics
and Astronautics, AIAA-98-1750, 1998. p. 42030.
[8] Muc A, Gurba W. Genetic algorithms and nite element analysis
in optimization of composite structures. Comp Struct 2001;
54:27581.
[9] Tabakov PY. Multi-dimensional design optimisation of laminated
structures using an improved genetic algorithm. Comp Struct
2001;54:34954.
Table 5
Comparison conguration without post-buckling eld
Panel weight (N) 23.26
Panel behaviour Con-
straints
FE
analysis
NN
system
% Dier-
ence
Buckling load (kN) >135 138 142 2.9
Pre-buckling stiness
(kN/mm)
>120 133 136 2.2
Displacement at the
buckling load (mm)
1.039 1.040 0.1
246 C. Bisagni, L. Lanzi / Composite Structures 58 (2002) 237247
[10] Nagendra S, Jestin D, G uurdal Z, Haftka RT, Watson LT.
Improved genetic algorithm for the design of stiened composite
panels. Comput Struct 1996;58(3):54355.
[11] Kaletta P, Wolf K. Optimisation of composite aircraft panels
using evolutionary computation methods. In: Proceedings of
ICAS 2000 Congress, Harrogate, UK, 27 August1 September
2000. p. 411.110.
[12] Lillico M, Butler R, Hunt GW, Watson A, Kennedy D, Williams
FW. Optimum design and testing of a post-buckled stiened
panel. American Institute of Aeronautics and Astronautics,
AIAA-2000-1659, 2000. p. 110.
[13] ABAQUS Theory and users manuals. Hibbitt, Karlsson &
Sorensen. Pawtucket USA, 1998.
[14] MATLAB. Using MATLAB manual. Version 5. The Math
Works Inc., Natick, MA. 1999.
[15] McCulloch WS, Pitts WH. A logical calculus of ideas immanent in
nervous activity. In: Bulletin of Mathematical Biophysics, Selected
Papers on Optical Neural Networks, vol. 96. 1994. p. 73.
[16] Cammarata S. Reti neuronali: dal Perceptron alle reti caotiche e
neuro-fuzzy. Milano: Etas libri; 1997 [in Italian].
[17] Raq MY, Bugmann G, Easterbrook DJ. Neural network design
for engineering applications. Comput Struct 2001;79(17):1541
52.
[18] Demuth H, Beale M. Neural Network TOOLBOX. Users Guide.
Version 3.0. The Math Works Inc., Natick, MA. 1998.
[19] Goldberg DE. Genetic algorithms in search, optimization and
machine learning. London: Addison-Wesley; 1989.
[20] Singh SB, Kumar A. Postbuckling response and strength of
laminates under combined in-plane loads. Comp Sci Technol
1999;59:72736.
[21] Naik NK, Chandra Sekher Y, Sailendra M. Damage in woven-
fabric composites subjected to low-velocity impact. Comp Sci
Technol 2000;60:73144.
[22] Park JH, Hwang JH, Lee CS, Hwang W. Stacking sequence
design of composite laminates for maximum strength using
genetic algorithms. Comp Struct 2001;52:21731.
C. Bisagni, L. Lanzi / Composite Structures 58 (2002) 237247 247