Professional Documents
Culture Documents
Determination of Robustness For A Stiffe
Determination of Robustness For A Stiffe
Determination of Robustness For A Stiffe
Composite Structures
journal homepage: www.elsevier.com/locate/compstruct
a r t i c l e i n f o a b s t r a c t
Available online 19 March 2008 It is a common practice to only consider the nominal means as input variables for both classical solid
mechanics and finite element (FE) analysis problems. A single solution based on the mean values is then
Keywords: used in design. In reality all input variables are stochastic, existing within a range of possible values. Dif-
Stochastics ferent combinations of these stochastic input variables will lead to differing output responses and the
Robustness introduction of variability will cause each structure to have a response that deviates from the original
Buckling specification, sometimes with catastrophic consequences. In this paper two variables, influence and sen-
Stiffened structures
sitivity, have been identified as parameters affecting structural robustness. Variability and uncertainty in
Composite yacht design
loads, geometry and lamina stiffness are introduced via a stochastic finite element analysis (SFEA) proce-
dure. The procedure is applied to the design of composite yacht hulls comparing the robustness of
designs aimed at satisfying a range of performance and cost requirements. It is shown that influence
and sensitivity are useful in identifying designs that lead to imperfection tolerant structures.
Ó 2008 Elsevier Ltd. All rights reserved.
1. Introduction sign and manufacture of products. It can be noted that the aim of
all current methodologies lead toward a similar goal – insensitivity.
Robust design has been defined by Phadke [1] as a concept ‘to Unfortunately one feature that is lacking in most approaches is the
make a product’s performance insensitive to variations in material, ability to quantify the robustness of a design, as well as the
geometry, manufacture and operating environments’. The manu- improvement in robustness that is achieved once changes are
facture of products often involves a balance between the controlla- made.
ble factors such as member sizing and uncontrollable noise such as Fig. 1 is a simple representation of the difference between a ro-
material and processing variations or the inability to accurately bust and unstable response in a design. In the unstable maxima, it
predict the boundary conditions experienced by a structural com- is observed that the performance drops significantly once the input
ponent. The concepts robustness and reliability may appear to be moves away from the nominal mean. In the robust case, the perfor-
similar as structures or products designed to be reliable often have mance remains fairly constant even when there is deviation from
an element of robustness built in. The difference is that in robust the nominal mean. Robustness is an especially important consider-
design, changes in the design are made such that variability is ation in situations where there is difficulty in predicting loadings
accommodated. Much of the foundations for the robust design con- on structures. Examples of these include the slamming loads on
cept can be attributed to Taguchi [2]. The ideology of robust design yacht hulls and the gust loads on aircraft wings.
came about in post-war Japan where there was a need to produce Buckling is a nonlinear phenomenon that has high dependency
high quality products despite the lack of quality input into the on geometry as well as stiffness and boundary conditions. Buckling
manufacturing facilities. of shell structures can be sensitive to imperfections and this can
Park et al. [3] have conducted a thorough literature review on have catastrophic consequences with structures failing at loads
robust design and various theories such as the Taguchi method, lower than the designed collapse load. Studies into the effect of
robust design optimization and robust design using the axiomatic boundary conditions and the stiffness matrix have been performed
approach are discussed in their publication. Another overview has by Yadav and Verma [5], Raj et al. [6] and Singh et al. [7] using first-
been provided by Allen et al. [4], where examples are provided for order perturbation. These studies involve the first buckling mode
the different levels at which robust design can be applied to the de- of various unstiffened curved composite panels with different as-
pect ratios. It was found that the factors driving the buckling fail-
ure were the aspect ratio and boundary conditions. These in turn
* Corresponding author. Fax: +61 2 9663 1222.
E-mail address: merrill.lee@student.unsw.edu.au (M.C.W. Lee). determined which laminate properties were significant.
0263-8223/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compstruct.2008.03.036
M.C.W. Lee et al. / Composite Structures 86 (2008) 78–84 79
of
DY ¼
Output Response Unstable Maxima oX 1 x2 ¼ a ð1Þ
xn ¼ b
where Y is the output response, X1 is an input variable.
Robust Maxima
The behaviour of the system in question can be quantified
by determining the gradient. This may not reflect the true
behaviour of the system under actual operating conditions,
as it is not appropriate to consider input variables as being
deterministic. Hence a new stochastic analysis procedure has
Input Function
been devised in order to account for this variability. With
Fig. 1. Comparison between robust and unstable maxima. the introduction of variation into the input variables, the input
variables can now be considered stochastic in nature as
follows:
The philosophy of robust design was introduced by Marcyzk [8].
The employment of stochastic techniques proved useful in reveal- X 1 ¼ f1 ðl; rÞ
ing the range of possible responses in a system. These techniques X n ¼ fn ðl; rÞ ð2Þ
can show how certain combinations of input variables may cause
a variation in the expected result that was previously not designed where X is the random input variable, l is the mean, r is the stan-
for. Another positive factor is that variables driving the design of dard deviation, and n is the number of input variables.
the system can be found. Hence the analysis can be described as multi-variant and each
Lee et al. [9] have previously presented a stochastic finite ele- output result should consist of a combination of input variables
ment analysis (SFEA) procedure in order to introduce variability as follows:
into structural analyses. SFEA involves the use of stochastic load- Y 1 ¼ f1 ðX 1 ; X 2 ; X z Þ
ing, geometry, boundary conditions and material properties. This Y m ¼ fm ðX a ; X b ; X c Þ ð3Þ
procedure was applied in a European Commission Framework
6th Framework Project COCOMAT [10]. The SFEA procedure was where Y is the output response, X is the input variable, and m is the
able to determine the factors causing a difference in postbuckling sample size.
mode shapes exhibited by a curved composite stiffened panel. Therefore a stochastic analysis with m number sample size and
One of the issues that arise from the SFEA is that although the n number input variables can be described as follows:
variables influencing the behaviour can be found, the influence var-
Sm
n ¼ fn ðY 1 ; Y 2 ; . . . ; Y m Þ ð4Þ
iable in itself is not able to provide a complete quantitative descrip-
tion of the structural robustness. Even if the influence of the variable The variation of each sample point from the nominal design
on the structural response is high, it is also the gradient or sensitivity mean can be determined using:
of that function that determines the robustness of the structure.
of
of
of
Payne and Kelly [11] have previously conducted research into DY m ¼ DX 1 þ DX 2 þ; . . . ; þ DX n : ð5Þ
oX 1 oX 2 oX n
the design of composite yachts. The key feature that was identified
in the design process was that the yachts could be designed for From the stochastic analyses, m number of metamodels can be ob-
speed, safety or cost. With these design drivers in place, two yacht tained, where the output response, Ym is plotted against the input
hull designs were achieved. The first yacht hull was designed for variables, X1,2,n. The figures show the difference between a multi-
speed with a monocoque hull, while the second designed with variable and multi-variant analysis.
durability as the design driver had multiple frames in the forward Once the metamodels are generated, the first variable can be
portion of the hull, from the back bow to amidships. It was identi- derived in order to determine the robustness of the system. The
fied however that safety was difficult to quantify and an attempt first measure of robustness is the influence of the input with re-
was made to identify the robustness of the designs in the face of spect to the output. A Spearman correlation [13] is performed in
uncertainty in seaway loads. order to find the influence of the input variable. The Spearman
It was noted by the authors that a common failure mode for rank correlation is a non-linear correlation which can be used
yachts is the buckling of the deck when the hull is subjected to slam- at the ordinal level. The formulation for the Spearman correla-
ming loads. Here the frames act as redundant members, transferring tion, q, is:
loads across the hull by providing alternative load paths [12]. Pn nþ12
In the following sections the formulation for a stochastic meth- i¼1 Rðxi ÞRðyi Þ n 2
q ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
nþ12 Pn qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
nþ12 ð6Þ
Pn 2 2
odology will be presented. This will include derivation of the two i¼1 Rðxi Þ n 2 i¼1 Rðyi Þ n 2
variables of influence and sensitivity for identification of robust-
ness. A simple hand calculation is given to demonstrate this con- where R is the ordinal rank, y is the output response, x is the input
cept. Following this two composite yacht hull designs will be variable, and n is the number of input variables.
evaluated. From the analyses conducted, it was found that in struc- An influence factor of unity can be interpreted as an input var-
tures incorporating redundancy, damage had lower influence on iable being proportional to the output. Conversely if the influence
the response and the sensitivity to damage was reduced. The re- factor is 1, it can be concluded that the input variable is inversely
sults are being used to try to define a measure of robustness that proportional to the output response. A high influence also means
will allow qualitative tradeoffs between cost, performance and that the input variable lies on the sensitivity curve and has domi-
robustness to be part of the design process. nance over the response of the output.
The next step is to find the sensitivity of output to the input var-
iable. Using the least-squares method, the sensitivity can be found.
2. Example of stochastic methodology
The formulation is as follows (see Figs. 2 and 3):
In a traditional multi-variable sensitivity analysis, the analysis Pn
is conducted by varying only one input variable and observing i¼1 ðX i XÞðY i YÞ
b¼ Pn ð7Þ
the change in response. This can be seen in the equation below: i¼1 ðX i XÞ
80 M.C.W. Lee et al. / Composite Structures 86 (2008) 78–84
1.8 Table 1
Stochastic boundary for Beam A
1
the inverse cube effect of the height in the deflection equation.
0.8 Figs. 5–7 show some of the metamodels obtained from the stochas-
tic analyses conducted on Beams A and B.
0.6 The next set of analyses is for Beam B. The load, length and
85 90 95 100 105 110 115
Input Young’s modulus have been kept the same, while the breadth
and height have been modified to show the difference in robust-
Fig. 2. Example of multi-variable system. ness. The stochastic boundary for Beam B can be found in Table 3.
2
1.6
Deflection (mm)
1.8
1.4 1.6
Output
1.4
1.2
1.2
1 1
0.8
0.8
0.6
60000 62000 64000 66000 6 8000 70000 72000 74000 76000 78000 80000
0.6 Young's Modulus (MPa)
85 90 95 100 105 110 115
Input Fig. 5. Metamodel of deflection against Young’s modulus for Beam A.
2
An example of this methodology is demonstrated below. Consider a
Deflection (mm)
1.1
1
0.9
P b 0.8
0.7
0.6
h 0.5
0.4
L 110 115 120 125 130 135 140
Height (mm)
Fig. 4. Cantilevered beam with edge loading. Fig. 7. Metamodel of deflection against height for Beam B.
M.C.W. Lee et al. / Composite Structures 86 (2008) 78–84 81
Table 5
Stochastic boundary for the composite yachts
Fig. 8. Region of structure that damage is simulated.
Input variable Mean Defined range
Carbon fibre
Young’s modulus, E11 (MPa) 142 000 120 700 – 163 300
Young’s modulus, E22 (MPa) 10 300 8 755 – 11 845
Poisson’s ratio, t 0.27 0.2295 – 0.3015
Shear modulus, G12 (MPa) 7 200 6 120 – 8 280
Shear modulus, G23 (MPa) 7 200 6 120 – 8 280
Shear modulus, G31 (MPa) 7 200 6 120 – 8 280
0° Lamina orientation (deg) 0 3.375 – 3.375
45° Lamina orientation (deg) 45 41.625 – 48.375
45° Lamina orientation (deg) 45 41.625 – 48.375
Lamina thickness (mm) 0.20 0.185 – 0.215
Aluminium core
Young’s modulus, E (MPa) 117 99.45 – 134.55
Poisson’s ratio, t 0.4 0.34 – 0.46
Core thickness (mm) 10 9.25 – 10.75
Fig. 9. Isometric view of monocoque hull.
82 M.C.W. Lee et al. / Composite Structures 86 (2008) 78–84
monocoque yacht, the multiple frame yacht has the laminate stiff-
ness D11 as having the highest influence on the deck buckling both
before and after damage is included. The values can be seen in
Table 6 below. However, unlike the monocoque yacht, the sensitiv-
ity has not changed significantly, and this is reflected in the sensi-
tivity lines seen in Fig. 14. They appear to be almost parallel. In this
design, it can be seen that the buckling factors of all samples are
decreased once damage is introduced into the hull skins.
The yacht seen below in Fig. 12 has the same hullform as the
monocoque, with the exception of transverse members, or frames,
being included in the forward region of the hull. The purpose of the
frames is to increase resistance to the deck buckling once a slam-
ming load is applied. The laminate specification for the multiple
frame yacht is as follows:
½457 ; Cs
Fig. 13. Buckling occurring on the top deck of multiple frame yacht.
Buckling Factor againstLaminate Stiffness
1.25
Table 6
Influence and sensitivity for monocoque hull
Buckling Factor
1.2
Undamaged Relationship of buckling factor to Influence Sensitivity
Damaged Without damage
1.15 Sensitivity (undamaged)
Laminate stiffness A11 0.426 7.421 107
Sensitivity (damaged)
Laminate stiffness D11 0.468 1.380 108
1.1 With damage
Laminate stiffness A11 0.638 1.023 106
Laminate stiffness D11 0.678 1.827 108
1.05 Damaged area length 0.276 1.791 103
21000000 22000000 23000000 24000000 25000000 26000000
Stiffness (D 11)
1.2
Buckling Factor
1.15 Undamaged
Damaged
1.1 Sensitivity (undamaged)
Sensitivity (damaged)
1.05
0.95
10000000 10500000 11000000 11500000 12000000 12500000
Stiffness (D 11)
Fig. 12. Isometric view of multiple frame hull. Fig. 14. Difference in buckling factor with and without damage.
M.C.W. Lee et al. / Composite Structures 86 (2008) 78–84 83
Change in Buckling Load againstDamaged Area The yacht example has shown the multiple frame yacht hull to
0.03 be more robust compared to the monocoque yacht hull. In the mul-
tiple frame hull, the frames acted as redundant members, allowing
Change in Buckling Load
References
4. Discussion of results
[1] Phadke MS. Quality engineering using robust design. Englewood Cliffs,
The hand calculation in Section 2 was a demonstration of how NJ: Prentice-Hall; 1989.
the influence and sensitivities of the input variables vary with a [2] Taguchi G. Introduction to quality engineering: designing quality into products
and processes. Quality Res 1986(June).
simple change to a design. It involved the deflection of a cantile- [3] Park G-J, Lee T-H, Lee KH, Hwang K-H. Robust design: an overview. AIAA J
vered beam subjected to an end loading. This showed that even 2006;44:181–91.
with a similar cross-sectional area, there are differences in the [4] Allen JK, Seepersad C, Choi HJ, Mistree F. Robust design for multiscale and
multidisciplinary applications. J Mech Des Trans ASME 2006;128:832–43.
influence and sensitivity for each variable. As the height of the
[5] Yadav D, Verma N. Buckling of composite circular cylindrical shells with
beam was increased, the sensitivity decreased. This decrease in random material properties. J Compos Struct 1997;37:385–91.
sensitivity means that if the same amount of variability were to [6] Raj BN, Iyengar NGR, Yadav D. Response of composite plates with random
material properties using FEM and Monte Carlo simulation. Adv Compos Mater
be applied in both beams, Beam B would experience less change
1998;7:219–37.
in deflection. One of the positive outcomes in the analyses was that [7] Singh BN, Iyengar NGR, Yadav D. Stability of curved composite panels with
the deflection decreased as the robustness in Beam B improved. random material properties. J Aerospace Eng 2002;15:46–54.
84 M.C.W. Lee et al. / Composite Structures 86 (2008) 78–84
[8] Marczyk J. In: 8th AIAA/USAF/NASA/ISSMO symposium on multidisciplinary [11] Payne R, Kelly D. Knowledge based engineering and yacht design. In: 2nd
analysis and optimization, Long Beach, CA, 6–8 September; 2000. High performance yacht design conference, Auckland, 14–16 February;
[9] Lee M, Kelly D, Orifici A, Thomson R. Postbuckling mode shapes of stiffened 2006.
composite fuselage panels incorporating stochastic variables. In: 1st CEAS [12] Kelly DW, Hsu P, Asudullah M. Load paths and load flows in finite element
European aerospace conference – century perspectives, Berlin, 10–13 analysis. Eng Comput 2001;18:304–13.
September; 2007. [13] Spearman C. Demonstration of formulae for true measurement of correlation.
[10] Degenhardt R, Rolfes R, Zimmermann R, Rohwer K. COCOMAT – improved Am J Psychol 1907;18:161–9.
material exploitation of composite airframe structures by accurate simulation [14] <www.youtube.com> One Australia Sinks. Accessed 01/10/2007.
of postbuckling and collapse. J Compos Struct 2006;73:175–8.