You are on page 1of 10

6th World Congresses of Structural and Multidisciplinary Optimization

Rio de Janeiro, 30 May - 03 June 2005, Brazil

Reliability Based Aircraft Structural Design Optimization with Uncertainty about Probability
Distributions
Erdem Acar and Raphael T. Haftka

Research Assistant, University of Florida, Gainesville, FL, 32611-6250, USA. E-mail: eacar@ufl.edu
Distinguished Professor, University of Florida, Gainesville, FL, 32611-6250, USA. E-mail: haftka@ufl.edu

Abstract
Probabilistic structural design is beset with a credibility problem. Uncertainties associated with errors in structural and aerody-
namic modeling and quality of construction are not well characterized as statistical distributions. Therefore, many engineers are skep-
tical of the accuracy of failure probability predictions. We propose a probabilistic design optimization method, where the probability
of failure calculation is confined to failure stresses, to take advantage of the fact that statistical characterization of failure stresses is
required by FAA regulations. The stress distribution is condensed into a representative single value by utilizing inverse cumulative
distribution function of the failure stress. The method is demonstrated by performing the weight and safety optimization of a repre-
sentative system composed of wing, horizontal tail and vertical tail. It is found that by moving material from the heavy components
to the light components corresponding to a small redistribution of the safety factor, it is possible to reduce the overall weight while
maintaining the same level of safety. The proposed method is also applied to a representative wing structural design problem and
similar results are obtained.
Keywords: reliability-based design optimization, uncertain probability distributions, weight savings

1. Introduction
Aerospace structures have traditionally been designed using a deterministic approach based on FAA regulations. Structural safety
has been achieved by combining safety factors with tests of material and structural components. In design of transport aircraft, the
Federal Aviation Administration (FAA) requires to use a safety factor of 1.5 for loads and conservative material properties (A-basis
value or B-basis value depending on the failure path) to maintain a high level of the safety for aircraft. Kale et al., [6] and Acar et al.
[1] analyzed the safety measures and found that the use of conservative material properties is equivalent to a safety factor of 1.27
when failure stresses have a coefficient of variation of 10 percent.
There is growing interest in replacing safety factors by reliability-based design. (e.g., Lincoln [7], Wirsching [15], SAE Aerospace
Information Report 5080 [2] and Long and Narciso [9]). However, many engineers are skeptical of the accuracy of failure probability
predictions for the following reasons. Data on statistical variability in material properties, geometry and loading distributions are not
always available in full (e.g., joint distributions), and it has been shown that insufficient information may lead to large errors in prob-
ability calculations (e.g., Ben-Haim and Elishakoff [4], Neal, et al. [10]). Uncertainties like variability in failure stress can be charac-
terized well from results of coupon tests. However, some uncertainties are not characterized well such as the distribution of actual
loads, errors in aerodynamic calculation of these loads, statistical characterization of the state of the structure due to errors in con-
struction, damage accumulation and aging, errors in structural modeling and stress calculation for a given state of the structure. In
addition, there is no consensus on how to model these errors in a probabilistic setting.
As a first step to overcome the problem of insufficient statistical data, we propose a probabilistic design optimization procedure
confined to failure stresses, for which the FAA already requires probabilistic characterization. We assume that estimates based on the
historical record are available for the probability of failure of structural components designed deterministically according to the FAA
code. With such estimates the proposed approach can obtain probabilistic designs without the need for the detailed statistical distribu-
tions for uncertainties detailed above. We illustrate the method by two example probabilistic design problems. We first perform
weight and safety optimization of a representative system composed of wing, horizontal tail and vertical tail. Then, we perform the
design optimization of a representative transport aircraft wing modeled by beam theory.
Section 2 discusses the proposed probabilistic design method confined to allowables. The application of the method to a wing and
tail system is given in Section 3. As a second example, the method is also applied to a wing design problem in Section 4, where de-
terministic and probabilistic design optimization of the wing performed. In Section 5, the accuracy of the proposed method is evalu-
ated by utilizing the results of the wing design problem. Finally, concluding remarks are given in Section 6.

2. Probabilistic Design Confined to Stress Allowables


For stress limits, the FAA requires aircraft builders to perform characterization tests, use them to construct a statistical model, and
then select allowables based on this model. For example, B-basis properties are defined to be such that 90% of material batches must
be better with 95% confidence. This initial foray of the FAA to define part of the design probabilistically can be used to introduce
more general probabilistic structural design assuming that the statistical characterization of failure limits is solid.
Structural failure, using most failure criteria, occurs when a stress σ at a point exceeds a failure stress σf. The probability density
distribution of the stress, s(σ) is poorly known, because it depends on the accuracy of structural and aerodynamic calculations, the
knowledge of the state of the structure, damage progression and pilot actions. The probability distribution of the failure stress f(σf),
on the other hand, must be well characterized for estimating allowables. So our goal is to base the design on this second distribution.
Given the stress σ, the probability of failure is
( )
Prob σ > σ f = F (σ ) (2.1)
where F is the cumulative probability distribution of the failure stress σf. The probability of failure Pf is obtained by integrating Eq.
(2.1) over all possible values of the stresses

∫ F (σ )s(σ )dσ = F (σ ) ∫ s(σ )dσ = F (σ )


* *
Pf = (2.2)
where the second equality is obtained by using the mean value theorem and the fact that the integral of s(σ) is one. Equation (2.2)
basically states that the effect of the poorly characterized probability of the stress can be boiled down to a single representative stress
allowable value σ*. This value can be obtained by estimating s(σ) and integrating as specified in Eq. (2.2). However, it is equally
possible to use historical data to do the reverse. That is, given an estimate of the probability of failure, we can obtain the allowable
σ* that corresponds to this data when airplanes are designed using the deterministic FAA process.
Now consider the possibility that we will deviate from the deterministic process by reducing structural margins on some compo-
nents and increasing them on others. We assume that if we increase stresses by a factor of (1+∆) the characteristic stress σ* increases
by the same factor and changes to σ*(1+∆). The accuracy of this approximation is demonstrated in Appendix 1.
Next, consider two structural components designed deterministically with failure probabilities corresponding to σ1* and σ2*. If
failure of the two components is uncorrelated, the probability that one of them will fail is
Pf =1- [1-F(σ1*)][1-F(σ2*)] (2.3)
If the two are correlated, the calculation is still simple for a given correlation coefficient. We can now entertain a probabilistic design
where material is added to one component and removed from the other so that stresses are scaled by (1+∆1) and (1+∆2), respectively.
The problem can be formulated, for example as
Minimize W(∆1, ∆2) (2.4)
Such that Pf* =1- [1-F(σ1*(1+∆1))][1-F(σ2*(1+∆2))] = Pf0=1- [1-F(σ1*)][1-F(σ2*)]
where Pf* is the approximate probability of failure, Pf0 is the probability of failure of the components for the deterministic design, and
W is their total weight.
Alternatively, instead of saving weight from the components, the safety of the system can be improved while keeping the weight
of the structure the same. The optimization problem then can be formulated as
Minimize Pf*(∆1, ∆2) (2.5)
Such that W(∆1, ∆2) = W0 = W(0, 0)

3. Historical Record for Probability of Failure


Before starting with the demonstration of the method, we look to the historical record on the probability of failure of traditionally
designed aircraft structures. Tong [13] performed a wide literature review on aircraft structural risk and reliability analysis. He refers
to the work by Lincoln [8] that reports the overall failure rate for all systems due to structural faults is one aircraft lost in more than
ten million flight hours, i.e. Pf =10-7 per flight hours. The Boeing Company released a report on statistical summary of commercial
jet airplane accidents from 1959 and 2001. The number of accidents that occurred between 1992 and 2001 due to structural failure,
the total number of accidents and the accident rate corresponding to different aircraft generations are given below Table 1.

Table 1. Aircraft accidents and probability of failure of aircraft structures. Examples of first generation airplanes are Comet 4, 707,
720, DC-8. Boeing 727, Trident, VC-10, 737-100/-200 are examples of second generation airplanes. Early widebody airplanes are 747-
100/-200/-300/SP, DC-10, L-1011 and A300. Examples of current generation airplanes are MD-80/-90, 767, 757, A310, A300-600,
737-300/-400/-500, F-70, F-100, A320/319/321.
Aircraft Generation* Accident Rate per Total Number of Accidents due to Structural failure rate per
million departures* accidents* structural failure* departure
(A) (B) (C) (A×C / B)
First 27.2 49 0 0
Second 2.8 130 2 4.31 x10-8
Early widebody 5.3 53 2 2.00 x10-7
Current 1.5 161 2 1.86 x10-8
Total --- 393 6 ---
* These columns are taken from the Boeing report

We see from Table 1 that failure probability of second generation airplanes is 4.31x10-8, whereas the failure probability of early
widebody airplanes and current generation airplanes are 2.0x10-7 and 1.86x10-8, respectively.

4. Demonstration of the method by wing and tail example


For a typical transport aircraft, the structural weight of the horizontal tail is about 20% of that of wing. The weight of the system
can be optimized by moving some material from wing to tail while still keeping the safety of the system unchanged. The weights of
the wing and the tail before probabilistic optimization are taken as 100 and 20, respectively. For illustrative purposes, we assume that
the historical record showed that both components had lifetime probability of failure of 10-7. The optimization problem given in
Eq.(2.4) is solved and the results are presented in Table 2.
To check the accuracy in probability of failure calculation, we assumed a normally distributed stress with 20% coefficient of varia-
tion for each component. The failure stress is taken lognormally distributed with mean value of 100 and 10% coefficient of variation.
The mean value of the stress in each component is computed as 39.77 such that the probability of failure before probabilistic optimi-
zation is 10-7. The value of Pf =10-7 corresponds to stress allowable σ*=59.24. The optimization is performed by using the proposed
method and the probabilities of failure calculated by σ* are compared to actual probabilities of failure. The correlation coefficient for
probabilities of failure of wing and tail is assumed zero.

Table 2. Weight optimization of the representative wing and tail system. Pf* is the approximate probability of failure and Pf is the ac-
tual probability of failure. Based on (Wwing)0=100, (Wtail)0=20, (Pf wing)0=(Pf tail)0=10-7, σwing*=σtail*=59.24
W % ∆W Pf *(x10-7) Pf (x10-7) % ∆P f σ* % ∆ σ*
Wing 99.04 -0.96 1.673 1.408 40.8 59.81 0.97
Hor. Tail 20.41 2.06 0.327 0.477 -52.3 58.04 -2.02
System 119.45 -0.46 2.0 1.885 -5.8

We see from Table 2 that by moving 0.96 % of the wing weight from the wing to tail the approximate probability of failure Pf * of
the tail is reduced by one third of its original value, while the exact probability of failure is reduced by only about one half. The
moved material from the wing increased failure probability of the wing, however the overall level of safety of the system, which was
to remain unchanged, is reduced by about 6%. The total weight of the system is reduced by 0.46%, the stress level at the tail is re-
duced by about 2%, and the stress level of the wing is increased by only about 1%.
Alternatively, instead of saving weight from the wing and tail system, the safety of the system can be improved for fixed weight by
moving weight from the wing to the tail. The initial weights, probabilities of failures and stress levels are taken as the same before.
The optimization problem given in Eq.(2.5) yields the results in Table 3. It is seen that 0.5% of the wing weight is moved to tail.
Although the probability of failure of the wing is increased by 20% of its original value, the probability of failure of the tail is re-
duced by 59%. On overall, the probabilistic optimization corresponds to 20% safety increase for the wing and tail system. The value
of system probability of failure is only 2% higher than the approximate probability of failure, so the proposed method is quite accu-
rate.

Table 3. Safety improvement of the representative wing and tail system. Pf* is the approximate probability of failure and Pf is the ac-
tual probability of failure. Note that (Wwing)0=100, (Wtail)0=20, (Pf wing)0=(Pf tail)0=10-7, σwing*=σtail*=59.24
W % ∆W Pf* (x10-7) Pf (x10-7) % ∆Pf σ* % ∆ σ*
Wing 99.5 -0.5 1.309 1.196 19.6 59.54 0.50
Hor. Tail 20.5 2.5 0.256 0.407 -59.3 57.79 -2.44
System 120 0.0 1.565 1.603 -19.9

Next, we added a vertical tail to the wing and horizontal tail system. For typical transport aircraft, the structural weight of the hori-
zontal tail is about 10% of that of wing. The weights of wing, horizontal tail and vertical tail before probabilistic optimization are
taken as 100, 20 and 10, respectively. We assume that the historical record showed that all components were designed for lifetime
probability of failure of 10-7. The results of weight and safety optimizations are presented in Tables 4 and 5.

Table 4. Weight optimization of wing, horizontal tail and vertical tail system
W % ∆W Pf* (x10-7) Pf (x10-7) % ∆Pf σ* % ∆ σ*
Wing 98.43 -1.572 2.322 1.751 75.1 60.18 1.60
Hor. Tail 20.29 1.459 0.453 0.593 -40.7 58.39 -1.44
Ver. Tail 10.28 2.741 0.225 0.373 -62.8 57.66 -2.67
System 129.0 -0.77 3.0 2.717 -9.4

Table 5. Safety optimization of wing, horizontal tail and vertical tail system
W % ∆W Pf*(x10-7) Pf (x10-7) % ∆Pf σ* % ∆ σ*
Wing 99.21 -0.79 1.532 1.327 32.7 59.71 0.80
Hor. Tail 20.44 2.2 0.299 0.451 -54.9 57.95 -2.17
Ver. Tail 10.35 3.5 0.148 0.284 -71.6 57.24 -3.37
System 130.0 0 1.979 2.062 -31.3

Table 4 and 5 demonstrate that by including the vertical tail in the system, weight saving is increased from 0.46% to 0.77%. In the
same way, the probability of failure reduction is increased from 20% to 31%. The increase in number of components may thus in-
crease the weight saving and safety improvement.

5. Application to wing design example


We next consider a representative wing structural design problem by using beam theory. The wing is chosen as typical long-range
transport aircraft wing in the Boeing 767 class, the geometric details of which are obtained from Venter and Sobieszczanski-Sobieski
[14]. Some parameters of the reference wing are listed in Table 6 and a simple sketch of the reference wing geometry is given Fig. 1.
The wing is subject to elliptical load distribution (see Fig. 2) of magnitude equal to the 40% (for each side, totally 80%) of 2.5 times
(load factor) take-off gross weight (TOGW) of the aircraft. The effective (load carrying) chord length is taken as 45% of the actual
chord corresponding to front spar located at 15% chord and the rear spar is located at 60% of the chord (Niu, 1988, p.254). The wing
depth/chord ratio is taken as 85% of its maximum value, since the depth/chord ratio changes chordwise.
The chord length, wing depth, moment of inertia, wing loading, the wing area and bending moment at any x-location can be ex-
pressed as
 c   c  x   h   c   c  x  1
c( x) = c r  t  + 1 − t 
  c  b/ 2 h( x) = c r    t  + 1 − t
  c

 b/ 2 I Z ( x, t ) = c( x) t (h( x) )2 (5.1)
 c r   r    c   c r   r   2

 x2 
A( x) =
1
[c( x) + ct ] x P( x) = Pr 1 −
 (b / 2 )2


(5.2)
2  
 c  x4  c  x5 c  x6 
M ( x) = 4 Pr c r b  t 
 2
+ 1 − 2 t 
 +  t − 1
3 c 4
 (5.3)
 c r  12b  cr  20b  r  30b 

where Pr is the value of loading at the root, which is calculated as 1.67 x10-2 lb/in3.

Table 6. Parameters of the reference wing


Parameter Value
Span, bref 120 ft
Chord length at root, cr 25 ft
Depth/chord ratio, h/c 0.12
Tip/chord ratio, ct/cr 0.4
Aspect ratio 6.8571
Area 2100 ft2
Take-off Gross Weight 300,000 lbs

(a) (b)

(c)

Figure 1. Reference wing; (a) cross-sectional view (b) top view and (c) loading on the wing

We first perform deterministic weight optimization of the reference wing model based on FAA safety requirements. Then, the prob-
ability of failure of the structure is optimized by probabilistic design methodology while either having the same weight or having the
same probability of failure.

5.1. Deterministic Design Optimization


A safety factor of 1.5 for loads and conservative material properties are used in the design. There are multiple failure paths due to
structural redundancy; hence we use B-basis material properties (see Appendix 2 for details of B-basis value calculation). Each sec-
tion is designed based on bending stress failure. Then, the design thickness of the section is calculated from
 h( x ) 
1.5 M ( x) 
g ( x, t design ) = σ a −  2  ≥0 (5.4)
I Z ( x, t design )
where σ a is the B-basis allowable stress value, M(x) is the bending moment, h(x) is the wing depth and Iz(x,t) is the moment of iner-
tia. The wing material is aluminum having 59.1 ksi B-basis stress value.
Application of Eq. (5.4) to all five sections yields five constraints. Then the deterministic optimization problem is written as
min Wdet
s.t. g ( xi , ti ) ≥ 0 for i = 1, 2, K ,5 (5.5)
ti ≥ t gauge = 0.005 in
where Wdet is the deterministic optimum structural weight of the wing and tgauge is the minimum gauge thickness for the material
(taken as 0.005 in for aluminum).
The optimization problem stated in Eq. (5.5) is solved and the optimum values of thickness and the weight for each section are
presented in Table 7 total skin weight of the wing is calculated as 9,647 lbs, 3.2% of TOGW of the aircraft. Based on the results of
deterministic optimization, the probability of failure of the panels and the system probability of failure of the wing is calculated. The
calculation of probability of failure of wing sections is given in the next sub-section.

Table 7. Results of Deterministic Optimization Problem


Wing Section Optimum Thicknesses (in) Weights (lb)
1 (tip) 0.005 9.1
2 0.050 114.1
3 0.178 494.4
4 0.415 1349.6
5 (root) 0.766 2856.3
Total Skin Weight (one wing) 4823.6 (1.6% of TOGW)

Probability of Failure Calculation of the Wing Sections


The first stage of calculation of probability of failure is to define limit-state, which is the equation defining the failure state. The true
value of the stress is different than its calculated value due to errors. We can define the true value in terms of the calculated value by
introducing a single error factor e
σ calc = (1 + e )σ true (5.6)
Then, the limit-state function for wing sections is obtained by combining Eqs. (5.4 and 5.6) as
 h( x) 
M ( x ) 
 1   2 
g ( x, t design ) = σ f −   (5.7)
 1 + e  I Z ( x, t design )
where g ≤ 0 is the failure state. The random variables defining the limit-state function, their distribution type and distribution pa-
rameters are given in Table 8.

Table 8. Random variables defining the limit-state function


Random Variable Distribution Type Mean Scatter
Pr Lognormal 1.67 x10-2 lb/in3 10% c.o.v.
cr Uniform 25 ft ± 1% bounds
bref Uniform 120 ft ± 1% bounds
ct/cr Uniform 0.4 ± 1% bounds
h/c Uniform 0.12 ± 1% bounds
σf Lognormal 73 ksi 10% c.o.v.
e Uniform 0 ± 10% bounds
The variables appearing in the limit-state function g are Pr, cr, ct/cr, h/c, b, σf and e. We assume that the central limit theorem is
applicable and hence the limit state function g is normally distributed. Mean value, g and standard deviation σg of limit-state func-
tion can be calculated using AFOSM (advanced first order second moment) approximation as presented in Appendix 3.
The probability of failure of wing sections can be calculated as
 g i 
P f = Φ −
i
( )
 σg 
= Φ (− β i )
( ) (5.8)
 i 
where β is the reliability index and Φ is the cumulative distribution function for standard normal distribution.
We further assume that the failure probabilities are correlated with the same correlation coefficient ρ. The probability of failure of
the system can be calculated from (Gupta [5])
β1 β 2 βn ∞  n  β + ρ y 
PFS =
∫∫ ∫
−∞ −∞
L
−∞
fg
1, g 2 ,L, g n
dg1dg 2 L dg n = 1 −

−∞
 ∏ Φ i  φ ( y ) dy
i =1  1 − ρ 
 
(5.9)

where β1 , β 2 , L, β n are reliability indices corresponding to individual wing sections and φ is probability density function for stan-
dard normal distribution. Appendix 4 gives some details about this integration.
In addition, the redundancy in the wing is taken into account. Each section is considered as a parallel system made of two compo-
nents. The failure is assumed to occur upon simultaneous failure of the two components. The correlation coefficient between the
probabilities of failure of the two components is assumed to be equal to 0.5. For parallel systems of equally correlated elements with
equal probabilities of failure and normal distributions for limit-state functions, the system probability of failure can easily be calcu-
lated from Eq.(5.10) (Thoft-Cristensen and Baker [12], p. 124)

n 2
βS = β =β = 1.155 β (5.10)
1 + ρ (n − 1) 1 + 0.5
where β and βS are the reliability indices for components and for the system, respectively. After calculating the reliability index βS,
the probability of failure can be calculated from Eq. (5.8).

Probability of Failure of the Wing Corresponding to Deterministic Optimization


Probability of failure of the wing sections of the deterministic design and the overall failure probability of failure of the wing are
calculated first for zero correlation coefficients between sections (Table 9). We see that the gauge thickness constraint is active for
the tip section. The overall failure probability of the wing system is the sum of probabilities of failure of the sections since the corre-
lation coefficient of probabilities of failure is zero. In the next section we analyze the effect of correlation coefficient.

Table 9. Probability of failure of deterministic optimization problem. (ρ=0)


Section Optimum Thicknesses (in) (Pf /10-6)
1 (tip) 0.005 0.421
2 0.050 4.117
3 0.178 4.047
4 0.415 4.021
5 (root) 0.766 4.013
Wing (one side) 16.62

The gauge thickness constraint is active at the tip, because the stress there is not critical leading to a small probability of failure.
The other sections of the wing are subject to the same stress constraint. However, slight differences in the probabilities of failure
given in Table 9 are due to the AFOSM approximation.

5.2. Reliability-Based Design and Optimization

Saving weight for fixed level of safety


We optimize the structure following reliability based design approach. We minimize weight keeping the probability of failure of
the wing at the same value of that of deterministic optimum. The optimization problem is stated as
min f = W pro
s.t. (PFS ) pro = (PFS )det (5.11)
t i ≥ t gauge = 0.005 in

The optimal thicknesses and probabilities of failure of the wing sections are compared with the deterministic optimum values in
Table 10 for the case of zero correlation coefficient between the sections.

Table 10. Comparison of deterministic and probabilistic optimization (fixed PFS) results. Zero correlation coefficient between
the Pf of the sections
Wing Sec- (topt)det (topt)pro (Wopt)det (Wopt)pro (Pf)det (Pf)pro
tion (in) (in) (lb) (lb) (x 10-6) (x 10-6)
1 0.0050 0.0054 9.1 9.9 0.421 0.031
2 0.0496 0.0542 114.1 122.9 4.117 0.404
3 0.1781 0.1831 494.4 508.3 4.047 1.719
4 0.4150 0.4130 1349.6 1343.1 4.021 4.653
5 0.7662 0.7439 2856.3 2773.0 4.013 9.812
System 4823.6 4757.2 16.62 16.62

Compared to the probabilistic optimum material is transferred from heavy sections to light sections (i.e. from root to tip) so that
light sections have smaller failure probabilities than heavy sections. The total bending material weights is reduced by 1.3%.
Next, we introduce a common correlation coefficient of 0.75 between sections. Results in Table 11 indicate that the weight saving
reduced to 1.2%. As expected overall wing failure probability is smaller than zero correlation case.

Table 11. Comparison of deterministic and probabilistic optimization (fixed PFS) results. The correlation coefficient between the Pf
of sections is 0.75.
det pro
Wing Sec- (topt) (topt) (Wopt)det (Wopt)pro (Pf)det (Pf)pro
tion (in) (in) (lb) (lb) (x 10-6) (x 10-6)
1 0.0050 0.0053 9.1 9.7 0.421 0.054
2 0.0496 0.0531 114.1 122.1 4.117 4.939
3 0.1781 0.1828 494.4 507.4 4.047 1.818
4 0.4150 0.4133 1349.6 1344.2 4.021 4.544
5 0.7662 0.7457 2856.3 2779.9 4.013 9.111
System 4823.6 4763.3 14.85 14.85

Increasing safety while keeping the same weight


Next, we optimize the overall probability of failure of the wing under the constraint that the total weight of the wing is the same as
that of deterministic optimum. The optimization problem is then written as
min f = PFS
s.t. W pro = Wdet (5.12)
t i ≥ t gauge = 0.005 in

where PFS is the probability of failure of the wing, Wpro is the optimum weight of the wing calculated via probabilistic approach.

Table 12 compares the thicknesses and probabilities of failure of the wing sections with the deterministic optimum values. As be-
fore, material is transferred from heavy sections to light sections. There is 34.3% reduction in the probability of failure.

Table 12. Comparison of deterministic and probabilistic optimization (fixed W) results. Zero correlation coefficient between the Pf
of the sections
Wing Sec- (topt)det (topt)pro (Wopt)det (Wopt)pro (Pf)\det (Pf)pro
tion (in) (in) (lb) (lb) (x 10-6) (x 10-6)
1 0.0050 0.0055 9.1 10.0 0.421 0.020
2 0.0496 0.0540 114.1 124.2 4.117 0.268
3 0.1781 0.1854 494.4 515.2 4.047 1.129
4 0.4150 0.4186 1349.6 1361.7 4.021 3.054
5 0.7662 0.7547 2856.3 2812.2 4.013 6.442
System 4823.6 4823.6 16.62 10.91

Using the thickness values of wing panels presented in Table 12, we calculate an effective safety factor for each section. Effective
safety factor for deterministic case is equal to the safety factor of 1.5 for loads times the additional safety factor (of 1.235 for this
problem) due to the use of B-basis material properties. Effective safety factor for deterministic optimum is obtained by multiplying
the effective safety factor for deterministic case with the thickness ratio of the two optima. The effective safety factors for wing pan-
els are given in Table 13.

Table 13. Comparison of effective safety factors of deterministic and probabilistic design. Zero correlation coefficient between the Pf
of the sections
Wing Section (SF, eff)det (SF, eff)pro
1 1.5 x 1.236 = 1.853 1.5 x 1.357 = 2.035
2 1.5 x 1.236 = 1.853 1.5 x 1.348 = 2.021
3 1.5 x 1.236 = 1.853 1.5 x 1.287 = 1.931
4 1.5 x 1.236 = 1.853 1.5 x 1.247 = 1.870
5 1.5 x 1.236 = 1.853 1.5 x 1.216 = 1.825

The lowest material safety factor of probabilistic design corresponding to wing root is 1.216. For other sections the material safety
factor is greater than 1.236. Therefore, if FAA gives aircraft companies freedom to slightly change conservative material properties
from their B-basis values, they can improve safety of airplanes without adding extra weight.

5.3. Approximate Calculation


We next repeat the calculations using the proposed approximate method. First we obtain the dependency of stress to the weight.
We consider one section of the wing. The maximum bending stress developed in the section can be written as

h
σ = M  IZ (5.13)
2

where M is the bending moment, h is the wing height and Iz is the moment of inertia. The moment of inertia can be approximated as
I Z = 2 × c t (h / 2 )2 and the structural weight of the section is approximated as
W = 2ρ ct L (5.14)

where ρ is the density, c is the mean chord length for the section, t is the thickness and L is the longitudinal length of the section.
Combining Eqs. (5.13-14) we can express the stress in the section in terms of the weight of the section as

c L 1 
σ = 2M ρ   (5.15)
c h W 

Thus the stress is inversely proportional to the weight of the structure, and to change the stress value σ to (1+∆) σ, the weight of the
structure is required to change from W to (1/(1+∆)) W.
Using the weights and the probabilities of failure, we solve the optimization problem stated in Eq. (5.12) with the approximate
method yielding the results presented in Tables 14 and 15.

Table 14. The weights of the wing sections (lbs)


Section W1 W2 W3 W4 W5 Wtotal
Det. opt. 9.1 114.1 494.4 1349.6 2856.3 4823.6
Prob. opt. 9.9 122.9 508.3 1343.1 2773.0 4757.2
Prob. opt. (σ*) 9.6 119.8 503.8 1345.6 2800.3 4779.9

Table 15. The probabilities of failure of the wing sections (Pf /10-6)
Section (Pf)1 (Pf)2 (Pf)3 (Pf)4 (Pf)5 (Pf)system
Det. opt. 0.421 4.117 4.047 4.021 4.013 16.62
Prob. opt. 0.031 0.404 1.719 4.653 9.812 16.62
Prob. opt. (σ*) 0.080 0.898 2.271 4.400 7.317 14.97
Prob. opt. (σ*) 0.027 0.371 1.654 4.615 9.951 16.62
Approx. values

We see from Tables 14 and 15 that the proposed approximation is conservative in that the weight is reduced by 0.9% although
with exact calculation we can save 1.4%. However, the probability of failure is now 10% lower.

6. Conclusion
A probabilistic optimization method is proposed in which the probability of failure calculation is confined to stress allowables and
the stress distribution is condensed into a representative single value by an inverse transformation. The main goal of this method is to
base the design to stress allowables whose probability distributions are more accurate. This method therefore serves a way to take a
first step of probability of failure calculations when probabilistic data is not scarce. The method is illustrated with a wing, horizontal
tail and vertical tail system weight and safety optimization. It is found out that probabilistic design renders about 0.77% weight sav-
ing for the same level of safety and 31% safety improvement while keeping the weight unchanged. The method is found to predict
slightly lower (around 6%) failure probabilities when optimizing for weight and slightly higher (around 2%) failure probabilities
when optimizing for safety.
The method is also applied to a representative wing design problem. First, the deterministic design optimization of the wing is per-
formed based on FAA requirements. Next, probabilistic design optimization of the wing is carried out by assuming probability distri-
butions for the geometric and loading parameters. The probabilistic design optimization is performed once more using the proposed
method to check for the efficiency of the method. The similar results as in the case of wing and tail system are observed, the method
predicts slightly higher probabilities of failure when optimizing for weight.

Acknowledgements
This work has been supported by the NASA Constellation University Institute Program (CUIP), Ms. Claudia Meyer program
monitor.

References
1. Acar, E., Kale, A. and Haftka, R.T. Effects of Error, Variability, Testing and Safety Factors on Aircraft Safety. NSF workshop on
Reliable Engineering Computing, Savannah, Georgia, 15-17 September, 2004.
2. Aerospace Information Report 5080. Integration of Probabilistic Methods into the Design Process. SAE, 1997.
3. Ang, A., H-S. and Tang, W.H. Probability Concepts in Engineering Planning and Design, Volume I: Basic Principles. John Wiley
& Sons, New York, 1975.
4. Ben-Haim, Y., and Elishakoff, I. Convex Models of Uncertainty in Applied Mechanics. Elsevier, Amsterdam, 1990.
5. Gupta, S.S. Probability Integrals of Multivariate Normal and Multivariate t1, The Annals of Mathematical Statistics, Vol. 34, No.3,
pp.792-828, 1963.
6. Kale, A., Acar, E. Haftka, R.T., and Stroud, W.J. Why Airplanes are so Safe Structurally? Effect of Various Safety Measures on
Structural Safety of Aircraft. 45th Structures, Structural Dynamics and Materials Conference, 19-22 April 2004, Palm Springs
California.
7. Lincoln, J.W. Method for Computation of Structural Failure Probability for an Aircraft. ASD-TR-80-5035, July 1980.
8. Lincoln, J.W. Aging Aircraft Issues in the United States Air Force. 41st International SAMPE Symposium and Exhibition, Ana-
heim, California, 1996.
9. Long, M.W. and Narciso. J.D. Probabilistic Design Methodology for Composite Aircraft Structures. DOD/FAA/AR-99/2, Final
Report, June 1999.
10. Neal, D. M., Matthews, W. T., and Vangel, M. G. Uncertainties in Obtaining High Reliability from Stress-Strength Models. Pro-
ceedings of the 9th DOD-/NASA/FAA Conference on Fibrous Composites in Structural Design, Lake Tahoe, NV, 1991,
DOT/FAA/CT 92-95, I, pp. 503-521, 1992.
11. Niu, M.C.Y. Airframe Structural Design. Conmilit Press Ltd., 1988.
12. Thoft-Cristensen, P. and Baker, M.J. Structural Reliability Theory and Its Applications”. Springer-Verlag, 1982.
13. Tong, Y.C. Literature Review on Aircraft Structural Risk and Reliability Analysis, DSTO Aeronautical and Maritime Research
and Laboratory Technical Report, DSTO-TR-1110, 2001.
14. Venter, G. and Sobieszczanski-Sobieski, J. Multidisciplinary Optimization of a Transport Aircraft Wing using Particle S warm
Optimization. AIAA Paper, AIAA 2002–5644.
15. Wirsching, P.H. Literature Review on Mechanical Reliability and Probabilistic Design. Probabilistic Structural Analysis Methods
for Select Space Propulsion System Components (PSAM). NASA Contractor Report 189159, Vol. III, 1992.

Appendix 1. Accuracy of the approximation on σ*


Recall that we assume if stresses increase by a factor of (1+∆) the characteristic stress increases approximately by the same factor and changes to
σ*(1+∆). We check the accuracy of this approximation for a number of cases and the results are presented below.

a) Failure stress is lognormal, stress distribution is uniform


In order to avoid having a material dependent illustration, we take the mean value of failure strength 100 and assume 10% coefficient of variation.
Stress is assumed uniformly distributed with ± 5 bounds. The mean value of the stress is taken 58.20 that yields to a Pf =10-7 for ∆=0. The first col-
umn of Table A1.1 shows the assumed ∆ increase of the characteristic stress. The second column shows probability of failure associated with the
corresponding ∆ increase. The third column shows the probability of failure corresponding to (σ*)0 (1+∆), where (σ*)0 is the value of σ* when ∆ is
equal to zero. The last column shows the % error rising from our assumption. We see that even 10% value of ∆ has an error of 0.3%. This is a good
accuracy considering that our illustrative problem results we have ∆ values considerably less than 10%.

Table A1.1. Evaluating the accuracy of the approximation on σ* with lognormal failure stress and uniform stress distributions
%∆ Pf (x10-7) Pf*(x10-7) σ* (1+∆) (σ*)0 % error
-10 0.0023 0.0020 53.44 53.31 -0.24
-5 0.0594 0.0553 56.34 56.28 -0.12
-3 0.5864 0.5787 57.50 57.46 -0.07
0 1.000 1.000 59.24 59.24 0.00
3 4.547 4.719 60.97 61.02 0.07
5 11.65 12.37 62.12 62.20 0.12
10 98.39 109.9 65.00 65.16 0.25

b) Failure stress and stress distributions are both normal


Next, we consider normal distributions for both failure stress and stress distributions. The mean value and c.o.v. of failure stress are again 100 and
10%, correspondingly. For this case, c.o.v. of stress is assumed 5%, which we change to 20% for part (c). Similar to part (a), we assume a Pf value of
10-7 for ∆=0, yielding a mean stress value of 46.89. The errors due to our assumption on σ* is checked and we found out that for ∆=10% results in
0.01 error, which is very small. The results are given in Table A1.2.

Table A1.2. Evaluating the accuracy of the approximation on σ* with normal failure stress and stress distributions
%∆ Pf (x10-7) Pf*(x10-7) σ* (1+∆) (σ*)0 % error
-10 0.0669 0.0676 43.19 43.21 0.04
-5 0.2664 0.2674 45.60 45.61 0.01
-3 0.4454 0.4562 46.56 46.57 0.01
0 1.000 1.000 48.01 48.01 0.00
3 2.149 2.148 49.45 49.45 0.00
5 3.538 3.537 50.41 50.41 0.00
10 11.80 11.84 52.80 52.81 0.01

c) Failure stress and stress distributions are both normal


Finally, we increase the variability in material properties, the c.o.v. of failure stress is 20%. For this case the stress 35.96 yields to a Pf value of 10-7
for ∆=0. The errors due to our assumption on σ* is checked and we see that for ∆=10% results in a very small error, namely 0.4 error. The results are
presented in Table A1.3.

Table A1.3. Evaluating the accuracy of the approximation on σ* with normal failure stress and stress distributions. (Higher variability in failure
stress)
%∆ Pf (x10-7) Pf*(x10-7) σ* (1+∆) (σ*)0 % error
-10 0.0680 0.0676 43.22 43.21 -0.02
-5 0.2716 0.2674 45.64 45.61 -0.06
-3 0.4620 0.4562 46.59 46.57 -0.05
0 1.000 1.000 48.01 48.01 0.00
3 2.103 2.148 49.41 49.45 0.08
5 3.397 3.537 50.33 50.41 0.16
10 10.67 11.84 52.60 52.81 0.40

Appendix 2. B-basis property calculation


B-basis value is the value exceeded by 90% of the population with 95% confidence. This is given by
B-basis = µ - σ kB (A2.1)
where µ is the mean, σ is the standard deviation and k1 is the tolerance coefficient for normal distribution given by Equation (A2.2)

z1− p + z12− p − ab
k1 =
a (A2.2)
z12−γ z12−γ
a = 1− ; b = z12− p −
2( N − 1) N
where N is the sample size and z1-p is the critical value of normal distribution that is exceeded with a probability of 1- p. The tolerance coefficient kB
for a lognormal distribution is obtained by first transforming the lognormally distributed variable to a normally distributed variable. Equation (A2.1-
2) can be used to obtain an intermediate value. This value is then converted back to the lognormally distributed variable using inverse transformation.
In order to obtain the B-basis values, we assume that 15 panels are randomly selected from a batch and tested. The uncertainty in material property
is confined to allowable stress. The mean and standard deviation of 15 random values of allowable stress is calculated and used in determining the B-
basis value of allowable stress.

Appendix 3. Calculation of mean and standard deviation of the limit state function
Recall that the limit-state function is defined in Eq. (5.7) as

 h( x) 
M ( x ) 
 1   2 
g ( x, t design ) = σ f −   (A3.1)
 1 + e  I Z ( x, t design )

which is a random variable. Mean and standard deviation of g can be approximated using AFOSM (advanced first order second moment) as (Ang and
Tang, 1975, pp. 196-197)
h
M  
2
g =σ f −   (A3.2)
IZ
2 2
2 2
 ∂g  2  ∂g  2  ∂g  2  ∂g  2 2
 ∂g  2
σ g2 =   σM +  σh +  σI +  σ +   σ (A.3.3)
 ∂M   ∂h    ∂σ f  σ f  ∂e  e
 ∂I Z  Z
 
2 2 2 2
and σ M , σ h , σ I Z and σ c can be obtained in a similar manner. In addition, the other standard deviations are calculated as

σσ
f
( f )⋅ σ f
= c.o.v. σ and σ e =
1
be (A.3.4)
3

( )
where c.o.v. σ f =0.10 and be =0.10.

Appendix 4. Probability integral calculation of multivariate normal


Probability integral calculation of multivariate normal is presented in a paper by Gupta [5]. For some special cases of correlation matrix, Gupta has
shown that the probability integral calculation becomes easy and it has been stated in that paper as follows.
Let
h1 h2 hn
( { }) ∫ ∫ L ∫ f (x1, x2, L, x n ; {ρij }) dx1dx2 Ldxn
Fn h1 , h2 ,L, hn ; ρ ij = (A4.1)
−∞ −∞ −∞
{ }
If the correlation matrix ρ ij of the xi’s has the structure ρ ij = α iα j (i ≠ j ) where −1 ≤ α i ≤ +1 , then these variates Xi can be generated from
n+1 independent standard normal variates Z1 , Z 2 , L , Z n ; Y by transformation

X i = (1 − α i )1 / 2 Z i + α i Y (A4.2)
and it follows that
∞   hi + ρij y 
( { }) ∫  ∏ F 
n

Fn h1, h2 ,L, hn ; ρij = f ( y ) dy (A4.3)
i =1 1 − ρij 
− ∞   
Also, if h1 = L = hn = h and ρ ij = ρ for all i and j values, then Eq. (A4.3) reduces to

  h + ρ y 
Fn (h, h,L, h; ρ ) =
∫ F n  i  f ( y ) dy
  1 − ρ 
−∞  
(A4.4)

You might also like