Está en la página 1de 10

International Journal of Crashworthiness

ISSN: 1358-8265 (Print) 1754-2111 (Online) Journal homepage: http://www.tandfonline.com/loi/tcrs20

Sensitivity analysis-based variable screening and


reliability optimisation for composite fuselage
frame crashworthiness design

Binhua Gao, Yiru Ren, Hongyong Jiang & Jinwu Xiang

To cite this article: Binhua Gao, Yiru Ren, Hongyong Jiang & Jinwu Xiang (2018):
Sensitivity analysis-based variable screening and reliability optimisation for composite
fuselage frame crashworthiness design, International Journal of Crashworthiness, DOI:
10.1080/13588265.2018.1454289

To link to this article: https://doi.org/10.1080/13588265.2018.1454289

Published online: 29 Mar 2018.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


http://www.tandfonline.com/action/journalInformation?journalCode=tcrs20
INTERNATIONAL JOURNAL OF CRASHWORTHINESS, 2018
https://doi.org/10.1080/13588265.2018.1454289

Sensitivity analysis-based variable screening and reliability optimisation


for composite fuselage frame crashworthiness design
Binhua Gaoa,b, Yiru Rena,b, Hongyong Jianga,b and Jinwu Xiangc
a
State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body, Hunan University, Changsha, Hunan, China; bCollege of
Mechanical and Vehicle Engineering, Hunan University, Changsha, Hunan, China; cSchool of Aeronautic Science and Engineering, Beijing
University of Aeronautics and Astronautics, Beijing, China

ABSTRACT ARTICLE HISTORY


During the impact process, fuselage frame structures often experience severe crushing-induced Received 27 June 2017
kinematic deterioration. To improve cabin safety, this paper developed an analysis and design Accepted 12 March 2018
algorithm to optimise 2D triaxially-braided composite (2DTBC) frame under dynamic crush-type KEYWORDS
load. The design methodology integrates three concepts coming from several different Crashworthiness; composite
communities including numerical simulation, sensitivity analysis-based variable screening and fuselage frame; sensitivity
reliability optimisation. Based one continuous medium mechanics theory, a basic finite element analysis; reliability-based
(FE) model coupled with multiple failure modes is built at the macroscopic level previously. design optimisation (RBDO)
Afterwards, the Sobol’ global sensitivity analysis is performed to derive a design variable
importance hierarchy. Finally, differential evolution (DE) algorithm is implemented to identify the
optimum frame geometry that has maximum energy-absorption capacity. The investigation
demonstrates that appropriate redistribution of shape parameters of the frame could enhance its
design reliability and crashworthiness, and the higher number of design variables often performs
better from the energy-absorption viewpoint.

1. Introduction specimens with different trigger geometries like tulip,


bevel, crown, grooves, etc. Lanzi et al. [9] optimised the
Composite material is an increasingly important class of
designs of tapered composite cylinders under axial
structural material in the field of crashworthiness in
impact to increase their crushing energy. To a similar
aerospace industry due to its unprecedented properties
goal, Yin et al. [10] employed multi-objective particle
such as cost-effective, light weight, design flexibility as
swarm optimisation (MOPSO) procedure for crashwor-
well as high energy absorption capability. The material
thiness design of honeycomb-filled thin single and bitub-
has benefited from a multitude of architectural parame-
ular polygonal tubes. It is confirmed by these studies that
ters to tune the mechanical properties to fulfil specific
when the geometric dimensions or material system are
demands, which may indicate pathways to obtain high-
reconfigured, the crashworthiness characteristics of
performance energy absorption concepts to meet the
composite structures can be enhanced.
aircraft crashworthiness requirements with minimum
Thin-walled fuselage frames are typically used as part
weight increases.
of the internal skeletal structures in wide body commer-
In recent years, considerable efforts have been spent
cial transport aircrafts [11–14]. Previous numerical anal-
on the crashworthiness designs of composite structures.
ysis and experiments showed that the structures often
The studies range from material coupons to complete
govern much of the responses of the fuselage substruc-
full-scale structures, from low speed quasi-static to high
tures below the main passenger deck, and absorb (or dis-
speed dynamic tests [1–5]. For instance, Farley and
sipate) a large fraction of the collision energy during the
Jones [6,7] took the lead in attempting to develop a sci-
fuselage section vertical drop testing [15]. The 2D triax-
entific understanding of the physical failure mechanisms
ially-braided composites (2DTBCs) in fuselage frame
of progressive crushing using thin-wall tubular speci-
shows great promise to further improve aircraft crash-
mens, and put forward two distinct crushing modes of
worthiness for their excellent through-thickness proper-
continuous fibre-reinforced composite tube including
ties and impact resistance. Ideally the load response of
transverse shearing and lamina bending. Jimenez et al.
the 2DTBC frame will provide sufficient levels of crash
[8] provided a series of experimental studies of thin-wall

CONTACT Yiru Ren renyiru@hnu.edu.cn


© 2018 Informa UK Limited, trading as Taylor & Francis Group
2 B. GAO ET AL.

performances. But the truth is that the structures are Section A-A
A
predisposed to fail by brittle fracture, and that the crash R x1 x2 x5
energy absorption capacities are therefore far from satis-
factory [16]. The daunting complexity of 2DTBC fuse- A
lage frame crashworthiness design is undeniable, which Outer flange Web
involves complex nonlinear mechanical behaviours, and x3
α x6
is largely determined by high number of design parame-
ters at multiple length scales. Despite the practical signif- Inner flange
icance as well as commercial profit of the structure, little x7
attention has been drawn to its design optimisation with x4
crashworthiness criteria.
This paper presents an optimum design of a 2DTBC Figure 2. The 2D triaxially-braided composite fuselage frame.
fuselage frame structure to maximise its energy absorp-
tion efficiency. For an enhanced computational effi- ‘½0o 18k 6 § 60o 6k  39.7% axial’, where ‘39.7% axial’ indi-
ciency, such surrogate model technique as support cates the percentage by volume of axial yarns in the pre-
vector machine (SVM) combined with the nonlinear FE form. The number of braided layers in the web and
method is employed. When the SVM model is validated, inner flange is equal to 8, while outer flange is manufac-
a sensitivity analysis based upon the Sobol’ decomposi- tured by dividing the layers of the web into two groups
tion method is utilised to reveal the relative importance equably and folding them over (see Figure 2). In the
of the individual design variables in determining func- present study, the material properties pertaining to the
tional performance attributes of the frame. Optimisation web of the frame are obtained from [16] and are given in
problems are formulated as both deterministic and reli- Table 1.
ability-based with different number of retained variables. Naik [17] studied the effect of the fibre volume frac-
Differential evolution (DE) algorithm is implemented to tion on the mechanical properties of 2DTBC material.
identify the optimum frame geometries that have maxi- Of particular interest was his observation that the rela-
mum energy absorption characteristics in all cases. tionship between the fibre volume fraction and the mod-
ulus (E1 and E2) are approximately linear, and that the
variation in remaining mechanical properties with the
2. The fuselage frame under impact load fibre volume fraction is negligible. It may provide an
approximate method for calculating the material proper-
2.1. Experimental testing ties of the outer flange and inner flange. Hence the
The structure considered in this study is a circular com- E1 and E2 for the two sections in the frame can be deter-
posite fuselage frame with J-shaped cross-section, which mined by using slops measured from the plots in litera-
is typically used as a part of the internal skeletal structure ture [17], while other properties remain the same with
in wide body commercial transport aircrafts (see web.
Figure 1). The frame was manufactured at Lockheed's Dynamic tests of the textile composite J-section frame
facility in Marietta, Georgia by resin transfer moulding was conducted in the study by Pilkington [18] within the
using 3M PR500 epoxy resin into a textile preform made NASA grant NGT1-03024 funded by NASA Graduate
of AS-4 carbon fibre. The preform is nominally Student Researchers Program. A schematic of the test is
shown in Figure 3. The specimen was subjected to a
radially-inward point loading until it fully fractured.
Fuselage frame

Table 1. Material properties for the web section.


Property Value
Longitudinal modulus, E1 (GPa) 48.6
Transverse modulus, E2 (GPa) 45.4
In-plane shear modulus, G (GPa) 13.1
Poisson's ratio, n12 0.231
Poisson's ratio, n21 0.216
Longitudinal compressive strength, Xc (MPa) 489.0
Skin Longitudinal tensile strength, Xt (MPa) 630.0
Stringer Transverse compressive strength, Yc (MPa) 392.2
Transverse tensile strength Yt (MPa) 504.3
In-plane shear strength Sc (MPa) 210.0
Figure 1. Example of composite fuselage frame on the fuselage Density (g/cm3) 1.52
Fibre volume fraction, Vf (%) 55.26
section.
INTERNATIONAL JOURNAL OF CRASHWORTHINESS 3

Platen
Fuselage frame

Fully clamped boundary


condition at both ends

Figure 5. The finite element model of the fuselage frame.


Figure 3. Schematic of the fuselage frame test.
simulative assessments and the sophisticated physical-
Impact speed of the 43.6 kg aluminium block is 7.22 m/s. based approach that makes the model on the one hand
The vertical contact load versus the vertical displacement comprehensible and on the other hand convenient for
of the frame during the test is shown in Figure 4. The data acquisition [19]. The failure model has been vali-
area under the load-displacement response curve repre- dated in my previous study [20]. The dynamical
sents the energy absorbed due to crushing of the struc- response characteristics of the frame were analysed
ture. The experimental values of energy absorption Ea from failure sequences and load response curve as well
and peak load Fp, more remarkably, are equal to 31.09 as energy absorption. Figure 5 depicts the representa-
kN and 528.16 J, respectively, which are defined as refer- tive FE models of the frame structure, which was sim-
ence values for following numerical analysis. Typical ulated with the four-node Belytschko–Tsay shell
damage models induced in the polymer-based composite element. The element mesh was refined (average mesh
structure are matrix cracking and fibre breakage/split- size 4, 5, 6, 7 and 8 mm) to accurately represent the
ting/pullout. As a result of this failure, the web and the impact response of the fuselage frame segment. A
inner flange were cracked and torn apart. trade off was made between computational efficiency
and accuracy, and therefore an average mesh size of
6 mm was chosen in the paper. The FE model is kept
2.2. Finite element modelling at rest by constraining all degrees of freedom using
To capture the intricate details of the impact event, nodal single-point constraint boundary condition on
explicit dynamic finite element analysis using Matzen- both ends of frame nodes. In addition, a hexagonal
miller–Lubliner–Taylor (MLT) composite material rigid body with an initial velocity of 7.22 m/s and
model was performed. The material model is based on weight of 43.6 kg was created to crush the frame
the continuum damage mechanics and irreversible ther- model. Contact between the frame and loading platen
modynamics. The main reasons for the choice are the was modelled using the standard penalty formulation
good correlation that is achieved in the previous method. To suppress the hourglass deformation
modes that accompany one-point quadrature, hour-
glass viscosity stresses were added to the physical
stresses at the local element level [21].

rlðQ1 l_e2  Q2 a_eÞ if e_ < 0
q¼ (1)
0 if e_ 0

where Q1 and Q2 are dimensionless input constants


which is equal to 1.5 and 0.06, respectively, l is a char-
acteristic length given as the square root of the area in
two dimensions and a is the local sound speed.
The load-displacement response of frame measured
during the numerical analysis is shown in Figure 4 as
the dot line. Low-pass digital filter (SAE 600 Hz), as rec-
ommended by NASA in 2002, is used for the need of
actual data processing and reduction of high-frequency
noise. The predicted peak crush load value is 30.20 kN,
Figure 4. Comparison of the load-displacement curve of the sim- compared to the experimental 31.09 kN, the difference
ulation with the test. being ¡2.8%. The predicted energy absorption value is
4 B. GAO ET AL.

Deterministic design optimisation


8
> min  Ea ðxÞ
>
>
>
< s:t: Pcr ðxÞ27 kN
(2)
>
> AðxÞ730 mm2
Crack in the fuselage frame >
>
:
xL xxH
Figure 6. Typical morphology of the fuselage frame after
crushing. where xL ¼ ðx1L ; x2L ; . . . ; xnL Þ and xH ¼ ðx1H ; x2H ; . . . ; xnH Þ
are the lower and upper bounds of design variables,
520.87 J, compared to the experimental 528.16 J, the dif- respectively.
ference being ¡1.4%. The predictive capability of the FE RBDO
model was also identified through crushing morphology. 8
As can be seen from Figure 6, the cracks developed in > min mðEa ðxÞÞ
>
>
the outer flange and progressed radially inward into the >
< s:t: mðFp ðxÞÞ þ hsðFp ðxÞÞ27 kN
web, then propagated through the inner flange and the >
(3)
>
> mðAðxÞÞ þ hsðAðxÞÞ730 mm2
frame eventually fractured into two pieces at the load >
:
contact point. All these above show that the model is an xL xxH
effective and accurate failure prediction method for the
composite structure. Thus it is considered appropriate where mðÞ and sðÞ are the mean and standard devia-
for the design optimisation herein. By the nonlinear tion operators, respectively. h is the design level of sigma,
FE technique, the crashworthiness characteristics of e.g. h ¼ 3 denotes 3sRBDO design. RBDO problems
different design samples during the crash process are have same objective as deterministic design optimisa-
captured. tion, but has an additional constraint in the form of
desired Sigma level. First-order reliability method
(FORM) is used to evaluate the failure probability during
3. Optimisation design for crashworthiness reliability assessment of particular design.
In the present work, maximum energy absorption Ea in
the structure is used as a design criteria in crashworthi-
4. Methodology
ness designs as long as the levels of both peak load Fp
and mass M are kept under allowable limits. The bounds Here our design procedure is presented in detail in
on the design variables are given in Table 2 together Figure 7. Sampling of the design space is done using the
with the starting design. All the seven design variables Latin hypercube sampling (LHS) technique, while SVM
(xi, i = 1,2,…,7) are assumed to distribute normally and [22] is used to build surrogate models (metamodels) that
are considered continuous, whose standard deviations relate energy absorption and peak load to the geometric
are defined according to typical manufacturing toler- design variables associated with the fuselage frame
ance. Due to the high ratio of outer flange radius R to model. This operation will largely reduce the number of
other dimensions (xi, i = 1,2,…,7), the mass of the frame finite element (FE) runs, thereby improving the optimi-
is approximately proportional to its cross-section area A sation efficiency. The number of FE simulations in the
(x), (where AðxÞ ¼ ðx1 þ x2 Þ  x5 þ ðx3  x5  x7 Þ LHS is determined by total number of design variables
x6 þ x4  x7 ). Both the deterministic and reliability-based involved. In this study, 5n + 5 sample points are used for
design optimisation (RBDO) problems of the frame are building the metamodels, where n is the number of
therefore more specifically defined as follows, design variables.

Table 2. Definition of design variables for the crashworthiness optimisation.


Design variables Baseline designs Lower bound Upper bound Probabilistic distribution Standard deviation
Outer flange width x1 /mm 30.18 24.14 36.22 Normal 0.10
Outer flange width x2 /mm 30.18 24.14 36.22 Normal 0.10
Cross-section height x3 /mm 121.92 97.54 146.30 Normal 0.16
Inner flange width x4 /mm 31.76 25.41 38.11 Normal 0.10
Outer flange thickness x5 /mm 2.25 1.72 2.79 Normal 0.03
Web thickness x6 /mm 4.04 3.43 5.58 Normal 0.03
Inner flange thickness x7 /mm 5.18 3.43 5.58 Normal 0.03
INTERNATIONAL JOURNAL OF CRASHWORTHINESS 5

The sum of all the first-order Sobol' indices is then


obtained by

X
n
S1T ¼ Si (8)
i¼1

It should be noted that whenj1  S1T j is a small num-


ber, the first-order Sobol' indices could provide a nearly
complete representation of the response, and all the
higher Sobol’ indices can be negligible [26]. The signifi-
cant variables are further ranked based on their respec-
tive first-order sensitivity index (main effect), and placed
into groups in the construction of dimension-reduced
Figure 7. Diagram of the flowchart used in the research: (a) full optimisation problems.
flow diagram; (b) constructing surrogate model process in more
detail.
4.2. Differential evolution (DE) algorithm
To date, various single-objective optimisation methods
have been well developed, which include both local opti-
4.1. The Sobol' decomposition method misation (sequential quadratic programming, nonlinear
programming quadratic line search, generalised reduced
As the number of variables is directly related to the cost
gradient, adaptive region optimisation method etc.) and
of performing an optimisation study, a crucial matter in
global optimisation (DE, self-adaptive evolution, simu-
simulation-based design optimisation problems is the
lated annealing, efficient global optimisation). Local
screening of insignificant variables [23]. The Sobol’
optimisation methods are typically gradient-based, and
decomposition method is a quantitative sensitive analy-
they therefore find a local optimum near the starting
sis method based on the variance decomposition theory,
position with the fast convergence speed normally. Con-
which can provide reliable quantitative sensitivity infor-
versely, global optimisation methods are usually proba-
mation of the input variables [24,25]. Based on the
bility-based, and their task is to find the absolute best
Sobol' method, the variance of response function DðyÞ
optimum solution based on global information of the
can be decomposed as
objective function [27].
X X Due to its conceptual simplicity, good convergence
DðyÞ ¼ Di þ Dij þ  þ D12n (4)
1in 1i < jn
properties, easy implementation and suitability for par-
allelisation, DE algorithm is deemed well-suited to set
up and solve all optimisation problems in this paper.
The Sobol' indices can be calculated by
The algorithm is typical for stochastic global search algo-
rithm and has ability to handle nondifferentiable, non-
Si1 is ¼ Di1 is 6 DðyÞ 1i1 < i2 <    < is n (5)
linear and multimodal objective functions. The more
details of it can be consulted from literature [28]. Exam-
where the integer s is called the order or dimension of
ples of the application of DE algorithm to structural
the index, and Di1 is is the portion of DðyÞ contributed
optimisation can be found in the work by Loja et al.
by interactions between design variable fxi1 ; xi2 ; . . . ; xis g.
[29], Ho-Huu et al. [30] and Fiore et al. [31].
The value of each index Si1 is can vary from 0 to 1, and
their sum is
X X 5. Sensitivity analysis-based variable screening
Si þ Sij þ  þ S12n ¼ 1 (6)
1in 1i < jn
5.1. Surrogate model
There are seven potential design variables and each
Specifically, the first-order sensitivity index (main varies within its domain. FE models corresponding to 40
effect) of ith variable can be calculated as sample points are created and values of the response
functions (Ea and Pcr) are acquired through FE analysis
Si ¼ Di 6 DðyÞ (7) (Table 3). The former 35 sample points are used to train
the SVM models, while the predictive capabilities of
6 B. GAO ET AL.

Table 3. Design matrix and objective and constraint values SVM models are calculated based on the remaining 5
obtained from FE simulation for ‘Top 7’designa. sample points. Following the sampled FE responses, the
Design variables
surrogate models of Pcr and Ea are constructed based on
No. x1 x2 x3 x4 x5 x6 x7 Ea (J) Pcr (kN)
SVM technique over the entire design space.
1 30.03 24.62 124.88 30.81 2.33 4.44 4.07 403.78 29.31
2 32.70 31.49 103.24 37.93 2.35 4.10 3.58 436.09 28.00 Two error metrics R2 and R2pred are used to assess
3 34.32 33.22 132.33 30.01 1.82 4.67 3.53 380.26 27.53 the quality of the metamodels obtained (Equation 9).
4 35.88 34.47 122.27 26.58 2.66 3.55 4.16 397.76 27.12
5 30.94 29.43 101.34 34.48 2.10 4.24 4.44 399.70 26.07 The former assesses the accuracy of reproducing the
6 31.58 28.44 119.82 34.14 2.22 3.65 4.87 379.42 25.58 observed data and the latter assesses the accuracy of pre-
7 25.11 29.75 117.02 26.34 2.18 5.15 4.59 392.43 23.35
8 25.72 24.15 108.91 28.30 1.91 4.05 4.55 329.32 18.54 dicting the original model at unobserved locations. The
9 33.60 29.88 110.41 26.98 2.05 5.24 4.26 428.16 26.56 values of the two error metrics are presented in Figure 8,
10 34.07 31.13 136.01 37.65 2.77 5.04 4.47 543.94 37.26
11 33.50 32.18 129.09 34.67 1.88 3.83 3.63 390.59 25.02
which indicate that the accuracies of the two surrogate
12 27.29 26.42 144.09 36.40 2.75 4.31 4.62 375.09 35.77 models (Ea and Pcr) are high precision.
13 31.90 24.97 115.43 36.01 2.08 5.11 5.52 471.03 29.85
14 28.69 29.04 142.34 33.09 2.59 4.72 4.32 468.19 34.53
15 31.09 32.93 105.66 30.21 2.49 3.93 5.15 457.09 25.66 X
N X
N
16 36.13 30.51 111.71 29.24 2.44 4.75 4.82 445.62 30.11 R2 or R2pred ¼ 1  ðyi  ~yi Þ26 ðyi  yi Þ2 (9)
17 26.01 26.09 99.31 33.60 1.97 5.57 4.99 364.48 22.51 i¼1 i¼1
18 26.89 34.83 118.80 30.73 2.74 5.31 5.46 482.57 31.24
19 32.11 26.81 114.12 34.97 1.94 4.58 3.91 416.87 25.64
20 29.31 28.70 117.78 31.89 2.41 4.16 5.30 403.67 28.66 where N is the number of sample points, yi is the simu-
21 30.20 31.86 121.29 31.60 2.24 4.94 4.76 488.68 30.29 lated result at sample point xi , ~yi is the corresponding
22 32.43 25.89 129.65 28.71 2.51 5.28 4.23 478.06 30.46
23 24.83 27.49 132.95 25.78 2.30 3.89 3.72 385.10 21.34
result predicted by the two metamodels, yi is the mean
24 33.03 35.08 140.54 25.56 2.16 4.35 3.81 446.24 28.69 of all simulated results.
25 28.66 25.38 131.18 37.41 1.85 5.39 3.92 422.15 32.25
26 27.63 28.28 125.78 27.81 1.76 3.45 3.69 290.43 17.92
27 24.55 36.02 101.03 35.75 2.69 4.52 5.55 477.89 28.30
28 24.28 33.60 107.22 27.96 2.63 3.63 5.08 341.18 23.81 5.2. Variable screening
29 30.67 30.44 124.26 31.25 2.15 3.79 4.37 373.64 25.17
30 35.33 35.49 135.13 33.76 2.04 3.99 4.01 454.40 30.60
31 25.40 31.00 98.36 32.80 2.55 4.83 5.25 368.53 26.29
As different design variables are expected to have differ-
32 35.06 25.13 143.49 29.11 2.46 4.21 4.11 444.75 30.48 ent contributions or sensitivities in the allowable design
33 29.16 35.83 128.00 32.23 2.37 5.09 5.36 456.88 32.25 space, it would be advantageous to have an automated
34 26.68 27.80 137.11 36.66 2.00 4.48 5.04 393.23 29.99
35 26.29 26.91 140.09 27.27 1.89 4.92 5.15 411.14 27.46 screening method available that would reduce the
36 29.60 27.34 113.21 32.53 1.72 5.50 4.70 408.22 26.06 parameter set such that only significant parameters
37 28.17 33.96 145.53 27.41 2.57 3.71 5.37 456.35 26.43
38 28.00 32.69 107.46 37.02 2.26 4.81 3.77 427.45 27.83 become design variables. Based on obtained surrogate
39 34.81 32.34 137.80 35.30 1.80 3.52 3.48 381.22 24.12 models, the Sobol’ method was used to identify the rela-
40 34.69 34.22 104.16 29.62 2.64 5.45 4.90 453.80 34.40
tive importance of individual input parameters over the
a
Here the terminology of ‘Top 7’ refers to seven-dimensional design
variables. entire domain. The bar charts representing the relative
importance of the seven variables are shown in Figure 9.

Figure 8. Statistic parameters of the surrogate models of both peak load and energy absorption.
INTERNATIONAL JOURNAL OF CRASHWORTHINESS 7

6. Results of the optimisation


Before running the optimisation, a sample of 30, 20 and
15 FE simulations were executed for ‘Top 5’, ‘Top 3’ and
‘Top 2’ design, respectively. Based upon the sampled FE
results, the new surrogate models (SVM models) with
different number of retained variables were constructed.
From the statistical parameters presented in Figure 8, we
can find that the new SVM models are accurate and thus
are used in the subsequent optimisation work. Both
deterministic and reliability-based design optimisation
were then conducted through DE algorithm. It is of note
that all of the omitted variables were fixed to the initial
design value and never got changed over the optimisa-
tion process.
Figure 9. The first-order Sobol’ indices for the three crashworthi-
ness indicators of the fuselage frame.
6.1. ‘Top 7’ design
Table 5 details the case of ‘Top 7’ design where both
It can be revealed that the web thickness x6, outer flange
deterministic and reliability-based design optimisation
width x1 and cross-section height x3 have a significant
results are presented. One can immediately note that
influence on the crushing energy absorption, while the
compared with its deterministic counterparts, the energy
inner flange width x4 and thickness x7 are almost negligi-
absorption capability of reliability-based design optimi-
ble. The corresponding values of j1  S1T jfor the three
sation (1:5s and 3s level) has worsened. A closer look at
responses are equal to 0.13, 0.10 and 0.07, respectively,
Table 5 would indicate that the web thickness (x6)
which indicate that for this problem the set of first-order
reaches its lower bounds while the outer flange width
Sobol’ indices could provide a nearly complete represen-
(x2), the inner flange width (x4) and the inner flange
tation of the response. From here onwards, the research
thickness (x7) reach their upper limit, respectively, in all
will focus entirely on the first-order Sobol’ indices in
cases. For deterministic case, the optimum Ea of the
light of the insignificant interactions.
fuselage frame is increased by 5.43% relative to the base-
It is common practice to reduce the dimension of the
line design, the minimum failure probability for any
input space such that only variables appearing in Sobol'
constraint observed is equal to 0.188. This probability is
index above a threshold value will be retained in the
deemed unacceptable, and the deterministic design
reduced model [25]. The acceptable cutoff threshold
therefore is viewed as insufficient quality. In the cases of
value here is set to different levels of 0, 0.05, 0.10 and
1:5sRBDO, the Ea of the thin-walled structure increased
0.20. The corresponding retained design variables
by 4.67%, and the maximum failure probability is equal
and the sum of first-order Sobol’ indices are summarised
to 0.072. In the cases of 3sRBDO, the Ea increased by
in Table 4. Here the terminology of ‘Top n’ refers to
3.81% and the maximum failure probability is approxi-
n-dimensional design variables.
mately equal to 0. It is natural to anticipate that a higher
Sigma level design is more ‘safer.’
Table 4. Dimension reduction guided by Sobol’ indices for
energy absorption.
Terms Threshold Retained variables S1T Normalised S1T 6.2. Dimension-reduced design
Top 2 0.20 x1, x6 0.557 0.640
Top 3 0.10 x1, x3, x6 0.688 0.791 Table 6 lists the obtained optimal solutions of ‘Top 5’
Top 5 0.05 x1, x2, x3, x5, x6 0.835 0.960 design. In the case of ‘Top 5’ design, the maximum Ea of
Top 7 0.00 x1, x2, x3, x4, x5, x6, x7 0.870 1.000
the fuselage frame under radially-inward loading for

Table 5. Optimisation results of maximising energy absorption for ‘Top 7’ design.


Optimal design variables Failure probability
Description x1 x2 x3 x4 x5 x6 x7 Max. Ea (J) Pcr (kN) A (mm2) Pf (Pcr) Pf (A)
Deterministic 34.37 36.22 108.50 38.11 2.39 3.43 5.58 556.86 27.00 726.65 0.495 0.188
1:5slevel 34.15 36.22 108.23 38.11 2.34 3.43 5.58 552.86 26.68 721.41 0.072 0.009
3slevel 33.89 36.22 108.14 38.11 2.28 3.43 5.58 548.28 26.36 716.99 0.000 0.000
8 B. GAO ET AL.

Table 6. Optimisation results of maximising energy absorption for dimension-reduced design.


Optimal design variables Failure probability
Description x1 x2 x3 x5 x6 Max. Ea (J) Pcr (kN) A (mm2) Pf (Pcr) Pf (A)
Top 5 Deterministic 31.45 27.82 97.54 2.35 4.74 513.53 27.00 730.00 0.512 0.498
1:5slevel 31.69 27.57 97.54 2.30 4.71 499.60 26.19 725.11 0.070 0.066
3slevel 31.87 27.37 97.54 2.25 4.68 486.93 25.42 720.21 0.002 0.003
Top 3 Deterministic 33.43 / 97.54 / 4.68 444.08 26.29 730.00 0.000 0.481
1:5slevel 33.32 / 103.24 / 4.36 430.14 26.31 725.55 0.046 0.000
3slevel 33.36 / 103.23 / 4.32 427.34 25.57 721.10 0.001 0.002
Top 2 Deterministic 30.61 / / / 3.74 406.08 25.54 730.00 0.000 0.500
1:5slevel 30.91 / / / 3.69 400.94 25.20 724.80 0.000 0.066
3slevel 30.82 / / / 3.65 395.82 24.96 719.67 0.000 0.001

crashworthiness occurs at the value of cross-section variables could miss the true optimum and lead to
height (x3) at its lower boundary. The energy absorption unfavourable designs. Additionally, this diagram also
capability of the thin-walled structure is slightly reveals that the reliable optimisation compared with its
enhanced in comparison with that of the baseline design. deterministic counterpart pushes the design further
Specifically, the optimum Ea is decreased by 7.8% rela- away from the constraints.
tive to the baseline design and the maximum computed
reliability for any constraint observed is 0.997 in case of
the 3sRBDO. 7. Conclusions
For comparison purposes, the optimisation results
We present a design methodology that maximises the
obtained from the ‘Top 3’ and ‘Top 2’ design are also
energy absorption capability of the 2DTBC fuselage
given in Table 6. Similar to the findings in the case of
frame in a sequential manner by considering the contri-
‘Top 5’ design, optimisation with less than seven design
bution of individual variables. The methodology integra-
variables could miss the true optimum and result in
tes three concepts that come from several different
unfavourable designs. In case of the 3sRBDO, the
communities: numerical simulation, sensitivity analysis-
energy absorption can be reduced by 19.09% as com-
based variable screening and reliability optimisation.
pared to the baseline design for ‘Top 3’ design and by
By employing the Sobol’ decomposition method, we
25.05%, respectively, for ‘Top 2’ design.
can find that the web thickness, outer flange width and
To gain more insight into the optimisation results, the
cross-section height are the primary design parameters
relationship of the normalised total first-order Sobol’
for determining the energy absorption capacity of the
index S1T to the energy absorption are shown in Figure 10.
2DTBC fuselage frame, while the inner flange width and
As can be seen, Ea increases almost linearly as the nor-
thickness are almost negligible. Following variable
malised S1T increases. Thus, it is concluded that optimisa-
screening process, both the deterministic and reliability-
tion without simultaneous consideration of full set of
based design optimisation problems with different num-
ber of retained variables were conducted. Results
obtained suggest that the appropriate redistribution of
the shape parameters of the fuselage frame can enhance
its crashworthiness as well as improve design reliability,
and that higher number of design variables often per-
forms better from energy absorption viewpoint.
Although significantly different final designs may be
obtained, an appropriately selected subset of design vari-
ables is effective while significantly reducing computa-
tional cost of the optimisation. The research methods
presented might be applied to the design of analogous
structures, especially in aeronautical crashworthiness
applications with costly simulations.

Disclosure statement
Figure 10. Deterministic and reliability-based design optimisa-
tion results of maximising energy absorption. No potential conflict of interest was reported by the authors.
INTERNATIONAL JOURNAL OF CRASHWORTHINESS 9

Funding [14] Ren Y, Xiang J. Energy absorption structures design of


civil aircraft to improve crashworthiness. Aeronaut J.
National Natural Science Foundation of China [grant number 2014;118(1202):383–398.
11402011]. [15] Carden HD, Boitnott RL, Fasanella EL. Behavior of
composite/metal aircraft structural elements and compo-
nents under crash type loads: What are they telling us ?
Washington (DC): NASA;1990. (NASA Technical
References Memorandum 102681).
[16] Perez JG. Energy absorption and progressive failure
[1] Christensen RM. The World Wide Failure Exercise II
response of composite fuselage frames [doctoral disserta-
examination of results. J Reinf Plast Compos. 2013;
tion]. Virginia Polytechnic Institute and State University,
32(21 ):166–1672 .
Blacksburg; 1999.
[2] Guo S, Li D, Zhang X, et al. Buckling and post-
[17] Naik RA. Analysis of woven and braided fabric reinforced
buckling of a composite C-section with cutout and
composites, J Compos Mater. 1994, 29, 17, 2334–2363.
flange reinforcement. Compos Part B. 2014;60(60):
[18] Pilkington LO. Impact response and failure of a textile
119–124.
composite fuselage frame [master's thesis]. Virginia
[3] Farooq U, Myler P. Finite element simulation of damage
Polytechnic Institute; 2004, Blacksburg.
and failure predictions of relatively thick carbon fibre-
[19] Matzenmiller A, Lubliner J, Taylor RL. A constitutive
reinforced laminated composite panels subjected to flat
model for anisotropic damage in fiber-composites. Mech
and round noses low velocity drop-weight impact. Thin-
Mater. 1995;20(2):125–152.
Walled Struct. 2016;104:82–105.
[20] Gao BH, Ren YR. Impact dynamic characteristics of
[4] Hesse SH, Lukaszewicz DHJA, Duddeck F. A method to
braided composite fuselage frame. Acta Mat Compos
reduce design complexity of automotive composite struc-
Sin. 2017;34(8):1780–1787.
tures with respect to crashworthiness. Compos Struct.
[21] Hallquist JO. LS-DYNA keyword user's manual [M]. Liv-
2015;129:236–249.
ermore Software Technol Corp. 2013:612–613.
[5] Bikakis GSE. Simulation of the dynamic response of
[22] Suykens JAK, Vandewalle J. Least squares support vector
GLARE plates subjected to low velocity impact using a
machine classifiers. Neural Proc Lett. 1999;9(3):293–300.
linearized spring–mass model. Aerosp Sci Technol.
[23] Craig KJ, Stander N, Dooge DA, et al. Automotive crash-
2017;64:24–30.
worthiness design using response surface-based variable
[6] Farley GL, Jones RM. Energy-absorption capability of
screening and optimization. Eng Comput. 2005;22(1):
composite tubes and beams.[Thesis].Virginia Polytech-
38–61.
nic Institute and State University. 1989, Blacksburg.
[24] Sobol IM. Global sensitivity indices for nonlinear mathe-
[7] Farley GL, Jones RM. Crushing characteristics of contin-
matical models and their Monte Carlo estimates. Math
uous fiber-reinforced composite tubes. J Compos Mater.
Comput Simul. 2001;55(1):271–280.
1992;26(1):37–50.
[25] Li J, Duan QY, Gong W, et al. Assessing parameter
[8] Jimenez MA, Miravete A, Larrode E., et al. Effect of trig-
importance of the Common Land Model based on quali-
ger geometry on energy absorption in composite profiles.
tative and quantitative sensitivity analysis. Hydrol Earth
Compos Struct. 2000;48(1):107–111.
Syst Sci. 2013; 17(8):3279–3293.
[9] Lanzi L, Castelletti LML, Anghileri M. Multi-objective
[26] Arwade SR, Moradi M, Louhghalam A. Variance decom-
optimisation of composite absorber shape under crash-
position and global sensitivity for structural systems. Eng
worthiness requirements. Compos Struct. 2004;65(3):
Struct. 2010;32(1):1–10.
433–441.
[27] Optimusver 10.15. Noesis Solutions. 2014. Reference
[10] Yin M, Wen G, Hou S, et al. Crushing analysis and mul-
manual.
tiobjective crashworthiness optimization of honeycomb-
[28] Storn R, Price K. Differential evolution–a simple and effi-
filled single and bitubular polygonal tubes. Mat Design.
cient heuristic for global optimization over continuous
2011;32(8):4449–4460.
spaces. J Global Optim. 1997;11(4):341–359.
[11] Tay YY, Bhonge PS, Lankarani HM. Crash simulations
[29] Loja MAR, Soares CM, Barbosa JI. Optimization of mag-
of aircraft fuselage section in water impact and compari-
neto-electro-elastic composite structures using differen-
son with solid surface. Int J Crashworthiness. 2015;
tial evolution. Compos Struct. 2014;107:276–287.
20(5):464–482.
[30] Ho-Huu V, Do-Thi TD, Dang-Trung H, et al. Optimization
[12] Huculak RD, Lankarani HM. Methods of evaluating ES-2
of laminated composite plates for maximizing buckling load
leg flail in dynamic evaluation and certification tests of
using improved differential evolution and smoothed finite
side-facing aircraft seats. Int J Crashworthiness. 2015;20
element method. Compos Struct. 2016;146:132–147.
(6):613–628.
[31] Fiore A, Marano GC, Greco R, et al. Structural optimiza-
[13] Ren Y, Xiang J. A comparative study of the crashworthi-
tion of hollow-section steel trusses by differential evolu-
ness of civil aircraft with different strut configurations.
tion algorithm. Int J Steel Struct. 2016;16(2):411–423.
Int J Crashworthiness. 2010;15(3):321–330.

También podría gustarte