Efficient Moment and Probability Distribution
Estimation Using the Point Estimate Method for High-
dimensional Engineering Problems
Liping Wang∗
GE Global Research Center, K1-2A62B, One Research Circle, NY12309
Don Beeson† and Gene Wiggs
‡
GE Aviation, 1 Neumann Way, Cincinnati, OH45215
The point estimate method (PEM) is an alternative to Monte Carlo Simulation
(MCS) and First Order Second Moments (FOSM) for evaluating the moments and
probability distribution of the system or component performance. Although PEM is
a powerful and simple method, it is often limited by the need to make 2n or even 3
n
evaluations when there are n random variables, which is unaffordable for many
engineering applications. In the previous paper by the authors [15], a variable-point
PEM method was proposed to improve the efficiency and accuracy of the existing
approaches. However, when it applied to the problems with large number of design
variables (number of design variables > 10), the method still requires hundreds or
thousands of simulation runs. This paper further improves the efficiency of the
variable-point PEM based upon two fundamental concepts: 1) The Pareto principle;
and 2) The Central Limit Theorem of Statistics, i.e., under common engineering
conditions, a linear combination of random variables can be approximated to first
order by a normal distribution. The efficiency and accuracy of the proposed method
are validated with three benchmark problems
I. Introduction
Performance function probability distributions and their associated first moments are important
statistical unknowns in engineering probabilistic design activities for robustness and reliability. Methods of
predicting this quantities using computer simulations enable engineers to more accurately model the
physical world and make design changes which will produce engineering systems, components or
processes that are less sensitive to manufacturing and environmental variation while at the same time
achieving requirements for performance, reduced costs and other six sigma goals.
Commonly used methods for moment calculations are: Taylor’s series expansion, transfer function or
meta-model construction, Monte Carlo Simulation (MCS), and First-Order Second-Moment (FOSM).
FOSM and Taylor’s series expansion both require accurate gradient calculations that impose excessive
restrictions on the performance functions (smoothness and existence and continuity of the first or first few
derivatives). On the other hand, MCS and traditional transfer functions generated from DOEs (Design Of
Experiments) do not require gradients calculations, but may also have difficulty because the computational
cost for them may be too expensive if high accuracy is required or if the number of input variables is large
(greater than 10). In the recent years, applying MCS to advanced meta-models generated from DACE
(Design Analysis and Computer Experiments) [17] becomes a popular approach. It is good alternative for
∗ Mechanical Engineer, GE Global Research Center, K1-2A62B, One Research Circle, NY12309. AIAA Member
† Program Manager, GE Aviation, 1 Neumann Way, Cincinnati, OH45215.
‡ Principal Engineer, GE Aviation, 1 Neumann Way, Cincinnati, OH45215
48th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference<br> 15th23 - 26 April 2007, Honolulu, Hawaii
AIAA 2007-1874
Copyright © 2007 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved.
2
fast and accurate probabilistic analysis since it can accurately model complex engineering problems with
highly non-linear, non-monotonic and high dimensional design space.
Commonly used methods for probability distribution analysis to generate the Cumulative Distribution
Function (CDF) and Probability Density Function (PDF) are: FOSM, MCS, and MCS with transfer
function or meta-models. Similar to moment estimation, MCS usually requires a minimum of thousands of
runs for CDF and PDF generation. FOSM using AMV (Advanced Mean Value method) can produce CDF
or PDF results with 2-4 orders of magnitude fewer runs than MCS, and the FOSM method usually works
well with linear or almost linear problems. However, FOSM may fail to provide accurate results for any
problem with a noisy performance functions or multiple MPPs (Most Probable failure Points). In addition,
the method becomes less efficient when applied to the problems with multiple performance functions.
The Point Estimate Method (PEM), originally proposed by Rosenblueth [1], is a simple but powerful
technique for evaluating the moments and probability distribution information of performance functions. It
has been widely used in many geo-technical and civil engineering analyses [2-9]. Unlike most FOSM
techniques [10-13], PEM does not require the gradients of the performance functions with respect to the
input variables. Also, it does not use any search techniques. The distribution of performance function is
typically described in terms of a 4-parameter Beta or Lambda distribution. The non-calculus nature of the
algorithm makes it extremely robust to typical engineering simulations that may be noisy, discontinuous or
non-monotonic. Also, several performance functions can be dealt with simultaneously without significant
additional computational effort.
Many PEM methods described in the articles [1-9] used two points per variable. This requires 2n
(n is
the number of input variables) function runs. N. C. Lind [8] developed a face center point method that
requires 4n3
runs and a two-point approximation method that requires 4n2 runs. Furthermore, Milton Harr
and H. P. Hong attempt to reduce the number of runs to 2n [9]. Lind’s, Harr’s and Hong’s methods all
involve replacing the points at the 2n corners of the hypercube with 2n points at or near the intersections of
the circumscribing hypersurface with its principal axes. However, these approaches must move the
evaluation points farther away from the means as n increases since the radius of a hypersphere
circumscribing a unit hypercube of n dimensions is n ; for a bounded variable, the points may easily fall
outside the domain of definition of the variable. Seo and Kwak [14] developed more accurate method for
moment calculations using two points per variable, which requires 3n function runs. This approach becomes
unaffordable if the number of input variables is large (greater than 10).
In the previous paper by the authors [15], a variable-point PEM method proposed two unique and novel
improvements : 1) efficient and accurate moments estimations were obtained by applying a different
number of PEM points for each input variable and 2) a new robust procedure was described for finding the
four parameters of the Beta and Lambda distributions. This approach has been tested and applied to many
engineering applications. It has been demonstrated that the method is able to improve both the efficiency
and accuracy of the 3-point PEM approach given in Ref. [14]. However, when it applied to the problems
with large number of design variables (number of design variables > 100), the method still requires
hundreds of simulation runs which could become very time-consuming and impractical if the engineering
problems require hours or days for a single simulation run.
The objective of this paper is to further improve the efficiency of the 3-point PEM for solving
probabilistic problems with very high number of design variables (100X+). The new method is based upon
two fundamental concepts: 1) The Pareto principle and 2) The Central Limit Theorem of Statistics that
says, under common engineering conditions, a linear combination of random variables can be approximated
to first order by a normal distribution.
3
II. Overview of PEM Techniques
A. 2-point Rosenblueth and 3-point Seo-Kwak PEM Approaches
Many PEM methods, such as Rosenblueth and Seo-Kwak, used the same number of points for each input
variable. For example, the Rosenblueth method used 2 discrete points to represent a continuous distribution
for each variable. By solving the system equations which match the area and first three moments of each
variable,
(1)
the x- locations and probabilities at the two discrete points are computed as
12
2
1
2
2
2
1
1
)2
(1
1
215.0
,)2
(12
,)2
(12
pp
p
x
x
x
x
x
xx
x
x
xx
x
−=
+
−=
+−+=
+++=
υ
υ
συυ
µ
συυ
µ
(2)
where xµ , xσ and xυ are the mean, standard deviation and skewness of the variable x, respectively.
The Seo-Kwak method used three discrete points to represent a continuous distribution for each
variable. By solving the system equations which match the area and first four moments of each variable,
(3)
the x-locations and probabilities at the three points are calculated as
x
x
x
pxpx
pxpx
pxpx
pp
υ
σ
µ
=+
=+
=+
=+
1
3
21
3
1
1
2
21
2
1
1211
21 1
x
x
x
x
pxpxpx
pxpxpx
pxpxpx
pxpxpx
ppp
γ
υ
σ
µ
=++
=++
=++
=++
=++
3
4
31
4
21
4
1
3
3
31
3
21
3
1
3
2
31
2
21
2
1
331211
321 1
4
−++
−−+
=
2
2
2
2
1
3422
3422
xx
x
x
x
x
x
xx
x
x
x
x
x
x
x
υγσ
υσ
µ
µ
υγσ
υσ
µ
(4)
−−
−−−
−
−−
−−
−+−
=
))(34(2
3434
1
))(34(2
3434
22
22
2
2
22
22
3
2
1
xxxx
xxxxx
xx
xx
xxxx
xxxxx
p
p
p
υγυγ
υγυυγ
υγ
υγ
υγυγ
υγυυγ
where xγ is the kurtosis of the variable x. The detailed information on Seo-Kwak approach can be
found in Ref. [14].
B. Variable-point PEM
In order to reduce the number of runs, while at the same time achieving better accuracy, the variable-
point PEM approach [15] applied a different numbers of points for each input variable. The percent
contribution of each variable x to the total variation and the non-linearity of y with respect to x were taken
into account to determine how many points to use for each variable. In general, more points, for example, 3
or 4 points, are used if the variables cause high non-linearity and make significant contributions to y.
Otherwise, fewer points (1 or 2 points) are used if the variables are linear or close to linear and have low
contributions to y. Recently the minimum number of points has been set to 2 to obtain more accurate
standard deviations and higher moments. The percent contribution of the ith
variable to the total variation,
Pcti , is computed as
22
2
2
2
1
1
2
)(...)()(
)(
n
n
i
ii
x
y
x
y
x
y
x
y
Pct
σσσ
σ
∂
∂++
∂
∂+
∂
∂
∂
∂
= (5)
A Two-Point Adaptive Nonlinear Approximation (TANA) [16] is used to determine the non-linearity
index ir . At the pre-selected two points, 1X and 2X , the approximation is given as
∑∑=
−
=
−+−∂
∂+=
n
i
r
i
r
i
r
i
r
i
i
n
in
i
iiii xxxxr
x
x
XgXgXg
1
2
1,1,
1
1,
1 1
1
2 )(2
1)(
)()()(~ ε (6)
The non-linearity index ir is solved by matching the function value and gradients at 2X , that is,
5
nigg
rxxxxX
x
x
xX
i
r
i
r
i
r
i
i
ri
i
i
i
iii ,...,2,1,)()(
)()(
1
2,1,2,
11
1,
2,1 =−+= −− ε∂
∂
∂
∂ (7a)
∑∑==
−
−+−+=n
i
ri
i
ri
i
ri
i
ri
i
n
ii
ri
i
i xxxxx
xX
XX r
g
gg1
2
1,2,1,2,1
1
1,1
12)(
2
1)(
)(
)()( ε∂
∂
(7b)
(See Ref. [16] for more details on TANA)
The number of points for each variable is determined based on the non-linearity and percentage
contribution to Ys of each variable . If the number of the points is one, the point is located at the mean and
probability is 1.0, that is,
0.1== px xµ (8)
where xµ is the mean of variable x. If the number of the points is two, the Rosenblueth equation (2) is
used for the x-location and probability calculations. If the number of the points is three, the Seo-Kwak
equations (4) are applied to find the locations and probabilities of the three points. If the number of points
equals to four, the 4-point locations and probabilities are derived from the area and the first 7 moments
given below, that is,
74
7
43
7
31
7
21
7
1
64
6
43
6
31
6
21
6
1
54
5
43
5
31
5
21
5
1
44
4
43
4
31
4
21
4
1
34
3
43
3
31
3
21
3
1
24
2
43
2
31
2
21
2
1
144331211
4321 1
mpzpzpzpz
mpzpzpzpz
mpzpzpzpz
mpzpzpzpz
mpzpzpzpz
mpzpzpzpz
mpzpzpzpz
pppp
=+++
=+++
=+++
=+++
=+++
=+++
=+++
=+++
(9)
where 43,21 ,, zzzz and 43,21 ,, pppp are the point locations in the standardized space and
corresponding probabilities. 71 ,...,mm are the standardized moments
By taking into account of the fact that all the probabilities )4,3,2,1( =ipi are linear factors in the
above system of equations (9), the probabilities can be simply expressed as shown below
MZZZP TT 1)( −= (10)
where
6
=
4
3
2
1
p
p
p
p
P
=
=
7
6
5
4
3
2
1
7
4
7
3
7
2
7
1
6
4
6
3
6
2
6
1
5
4
5
3
5
2
5
1
4
4
4
3
4
2
4
1
3
4
3
3
3
2
3
1
2
4
2
3
2
2
2
1
4321
11111
m
m
m
m
m
m
m
M
zzzz
zzzz
zzzz
zzzz
zzzz
zzzz
zzzz
Z (11)
The errors for the area and moments can then be calculated as
PZMzzzzE ⋅−=),,( 4,321 (12)
Solving for the 4-point locations and probabilities from Eq. (9) can be expressed as the following
optimization problem,
4321
4,3214,321
..
),,(),,(
zzzztS
zzzzEzzzzEMinT
<<< (13)
The optimum search for the above problem is much more robust and accurate since the problem has
only 4 variables instead of 8 parameters. This can be applied for solving 4-point locations of any
distributions.
For the distributions like normal, lognormal, uniform and exponential, more robust and simpler
methods can be used to find point locations and probabilities. Instead of performing optimization searches
for point locations every time PEM is run, a pre-calculated table that stores the point locations for
standardized distributions can be generated. In order to obtain the pre-calculated table, the transformation
functions of the distributions must be used to convert the first 7 moments to standardized moments, and the
four point locations in the standardized space Z are computed. Based on ,,, 221 ZZZ and ,4Z the 4- point
locations in the original space X can be easily computed from the transformation functions. For example,
for a normal distribution, the transformation function is given as
x
xxxz
σ
µ−=)( (14)
The standardized 7 moments are given as
7
=
0
15
0
3
0
1
0
1
7
6
5
4
3
2
1
1
M
M
M
M
M
M
M
(15)
The four point locations in the standardized space Z and probabilities are computed as
=
99212.33441397
30950.74196315
48530.74196411-
17192.33441421-
4
3
2
1
Z
Z
Z
Z
(16a)
=
62270.04587588
4210.45412444
22670.45412384
32550.04587585
4
3
2
1
P
P
P
P
(16b)
For skewed distributions like weibull and extreme value, a one dimensional table of the point locations and
probabilities as a function of the shape parameter β can be generated by solving the optimization problem
given by Eq. (13).
The performance function values at all the combinations of the point locations for all the variables are
computed by running the simulation code. The total number of runs is
nNNNNofRun ...21= (17)
where iN is the number of points for the th
i variable. The first four moments of the performance
function Y, which is non-linear in general, are then calculated from the following equations,
2
2
4
1
11
4
2
3
2
3
1
11
3
2
1
11
2
11
1
/)),...,((...
/)),...,((...
)),...,((...
),...,(...
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
MMxxyppM
MMxxyppM
MxxyppM
xxyppM
n
n
n
n
n
n
n
n
n
n
n
n
n
n
n
n
i
N
i
ii
N
i
i
i
N
i
ii
N
i
i
i
N
i
ii
N
i
i
i
N
i
ii
N
i
i
−=
−=
−=
=
∑∑
∑∑
∑∑
∑∑
==
==
==
==
(18)
where ,,, 321 MMM and 4M are the mean, variance, skewness and kurtosis of the performance function.
8
SEO-KWAK’s 3-point PEM used the Pearson system to solve the above system equations to find the four
parameters of the Beta distributions. The variable-point PEM found the analytical solution of the four
parameters of Beta distribution and performed optimization search to find the four parameters of Lambda
distribution. The flow-chart of the variable-point PEM approach is shown in Figure 1. (See more details on
the variable-point PEM in [15])
Analytical equations and numerical methods
to ensure robustnessand accuracy
3. Compute four moments of
responses Ys using
,
a b
4. Finding 4 parametersof Beta or Lambda distribution based on
four moments of Ys
. λ1 λ2
Key Steps
1. Determine how many pointsto use for each x
1) Percent contribution of x to the total variation y
2) Non-linearity of y with respect to x
xp1_1
x = location
p = probability
.
(P1) (g1k) + (P2) (Z2K) + (P3) (g3K) + (P4) (g4K) = MK
Mean (k=1), Std dev (K=2), Skewness (K=3), Kurtosis (k=4)
2. Run simulation code to calculate function values g and
corresponding probabilities
gi is Y value at (x1_i,x2)
pi is product of p1_i and p2
Engineering
Simulation
X1_1, x1_2X1_3, x1_4
X2
g1, p1g2, p2g3, p3g4, p4
For example, x1 uses 4 points andx2 uses 1 point
xp2
x = location
p = probability
.
x1
x2
p1_2 p1_3 p1_4
Beta or Lambda
distribution
Figure 1. Flow-Chart of the Variable-point PEM
III. Proposed PEM Approach
Although the variable-point PEM approach is able to improve the efficiency and accuracy of the
existing 3-point PEM approach given in Ref. [14]. However, when it applied to the problems with large
number of design variables (number of design variables > 10), the method still requires hundreds of or
thousands of simulation runs that could become very time-consuming and impractical if the engineering
problems require hours or days for a single simulation run. The objective of this paper is to further
improve the efficiency of the modified PEM for solving probabilistic problems with high number of design
variables ( > 10 or 100X+). The new method is based upon two fundamental concepts: 1) The Pareto
principle and 2) The Central Limit Theorem of Statistics that says, under common engineering conditions, a
linear combination of random variables can be approximated to first order by a normal distribution.
The detailed steps of the proposed approach are:
9
1. Perform a sensitivity analysis centered on the mean values of all the X factors.
2. Based on the Pareto Principle concept, it can be expected that approximately 80% of the
variance will be caused by 20% of the design factors. Therefore, 20% of the design factors
will likely determine the overall shape of the probability distribution of the output variable
3. However, the variation effects of the remaining 80% minor design factors cannot be ignored.
These effects can be approximated by running the simulation at the nominal values of the
minor design factors and rolling the total variation contribution up into a single equivalent
normal distribution that has a mean of zero and a standard deviation that is determined from
the sensitivity analysis..
4. Finally, a reduced PEM analysis is run based upon the significant design factors (20%) with a
single extra normal distribution added on (representing the other 80%).
IV. Numerical Examples
The efficiency and accuracy of the proposed PEM approach are demonstrated with the following three
benchmark problems. The results are compared with Monte Carlo, variable-point and 3-point PEM
approaches.
Example 1
The performance function is
7577.8),( 2
2
121 −+= xxxxg (19)
where 1x and 2x are the random variables with normal distributions (mean=10, standard deviation =
3). The sensitivity analysis information for the output variable g with respect to the two variables (see
Table 1) shows that 99.75% of the variation is contributed by 1x . The remaining variable 2x contributes
only 0.25% to the total variation of the output variable. The significant variation effects of the minor
factor is converted into a normal distribution that has a mean of zero but a non-zero standard deviation of
0.04994 (square root of 0.002494). The modified PEM method can now be run with design factor 1x that
has its original mean and standard deviation, and one extra variable x with a normal distribution N(0,
0.04994) which is added on. That is, the output random, g( 1x , 2x ), which is function of the 1major input
random variable at the mean of 2x , ie., 2xµ , is modified by the x normal random variable as shown by
the equation
xxgxg x += ),()(ˆ211 µ
Table 2 shows that the four moments obtained from the new PEM approach are almost exactly matched
the Monte Carlo results. Figure 2 shows the CDF comparisons from the Monte Carlo (pink), the new PEM
(blue), variable-point PEM (green) and 3-point PEM (black). The new PEM and variable-point PEM
methods match the Monte Carlo’s (pink) exactly, however, 3-point Seo-Kwak PEM is off from the correct
solution at the tail. Since this example only has two design variables, the numbers of function evaluations
for both variable-point and new PEM methods are the same and actually higher than 3-point PEM due to
extra runs required by sensitivity analysis.
10
Table1. Sensitivity Information of Example 1
Table2. Four Moments Comparison of Example 1
Figure 2. CDF Comparison of Example 1
Effect of Factor Variation on Response Variation
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Variable Var Contrib Std Dev %Contribution
X1 3600.0360 60.0003 99.7506%
X2 9.0000 3.0000 0.2494%
-------------
Total: 100%
Method 3-point Seo-
Kwak PEM
Variable-point
PEM
New PEM Monte
Carlo
Mean 110.2423 110.2423 110.2423 110.1703
Std Dev 61.40847 61.40846 61.33516 61.6242
Skewness 0.845778 0.864667 0.86777 0.86804
Kurtosis 3.2433 3.99533 4.00012 4.01879
No. of Runs 9 13 13 100,000
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
0.00 100.00 200.00 300.00 400.00 500.00 600.00
g
Pro
bab
ilit
y
New PEM Variable-point PEM Monte Carlo 3-point PEM
11
Example 2- 10 Variables I-Beam Problem
In the example of an I-Beam CDF/PDF analysis shown in Figure 3, 10 input variables are taken into
account (see Figure 3 for detailed inputs, in fact only 8 variables are active since only 1 output Y is
considered), and the performance function is computed as
SY −= maxσ
where S is the strength, which is a random variable, and maxσ is the maximum stress, which is given as
IL
daLap
⋅⋅
⋅−⋅⋅=
2
)(maxσ
where
12
)2)(( 33
fwff tdtbdbI
−−−=
Figure 3. 10-Variable I-Beam
The sensitivity analysis information for the output variable Y with respect to the ten variables (see
Table 1) shows that 51.59% of the variation is contributed by tf, 26.11% is contributed by L and 7.85% is
contributed by a. The remaining variables d, tw, bf, E, rho, P and S contribute only 14.455% to the total
variation of the output variable Y. The significant variation effects of the minor factors is rolled into a
normal distribution that has a mean of zero but a non-zero standard deviation of 0.380197 (square root of
0.144549797). The modified PEM method can now be run with design factors tf, L and a that have their
original means and standard deviations, and one extra variable x with a normal distribution N(0, 0.380197)
which is added on. That is, the output random, Y, which is function of the 3 major input random variables
at the means of d, tw, bf, E, rho, P and S is modified by the x normal random variable as shown by the
equation
aP
L
tw
bf
d
tf
E, rho
Variable Type Mean Std
P Normal 6070 200
L Normal 120 6
a Normal 72 6
S Normal 170E3 4760
E Normal 30E6 3E6
rho Normal 0.28 0.028
d Normal 2.3 1/24
bf Normal 2.3 1/24
tw Normal 0.16 1/24
tf Normal 0.26 1/24
aP
L
tw
bf
d
tf
E, rho
Variable Type Mean Std
P Normal 6070 200
L Normal 120 6
a Normal 72 6
S Normal 170E3 4760
E Normal 30E6 3E6
rho Normal 0.28 0.028
d Normal 2.3 1/24
bf Normal 2.3 1/24
tw Normal 0.16 1/24
tf Normal 0.26 1/24
12
xSY +−= maxˆ σ
Table 4 shows the four moments obtained from the new PEM approach are almost exactly matched the
Monte Carlo results. The new PEM takes 33 runs, however, the variable-point PEM takes 384 runs and 3-
point PEM will need 59049 runs. In [15], the number of runs required by the variable-point PEM was 213
if the minimum number of points for the variables with fewer contributions to the outputs was set to 1. In
order to obtain more accurate standard deviations and higher order moments of the outputs, the minimum
number of points used in the variable-point PEM is set to 2 in this paper. Figure 4 shows the CDF
comparisons from the Monte Carlo (pink), the new PEM (blue) and variable-point PEM (green). The new
PEM and variable-point PEM methods match the Monte Carlo’s (pink) exactly, however the new PEM
only takes 33 runs which reduces the number of runs significantly.
Table3. Sensitivity Information of I-Beam
Table4. Four Moments of I-Beam
Effect of Factor Variation on Response Variation
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Variable Var Contrib Std Dev %Contribution
d 13259100 3641.304 2.66%
tw 4340278 2083.333 0.87%
bf 6617570 2572.464 1.33%
tf 256821200 16025.64 51.59%
E 0 0 0.00%
rho 0 0 0.00%
P 25082440 5008.237 5.04%
L 129960000 11400 26.11%
a 39062500 6250 7.85%
S 22657600 4760 4.55%
-------------
Total: 100%
Method 3-point Seo-
Kwak PEM
Variable-point
PEM
New PEM Monte
Carlo
Mean - -17966.22 -17966.22 -17967.12
Std Dev - 22110.19 22110.97 22111.07
Skewness - 0.55 0.55 0.55
Kurtosis - 3.15 3.15 3.14
No. of Runs 59049 384 33 100000
13
Figure 2. CDF Comparison of I-Beam
Example 3- 18-Variable Torsion Vibration Problem
The new PEM approach is also demonstrated with a conceptual 18-variables torsion vibration problem
shown in Figure 5. The mean values of the eighteen design variables are given in Table 5. The three output
variables include the weight, low and high natural frequencies (see Table 6).
Figure 5. 18-Variables Torsion Vibration Problem
1 2 3
1
2
1 2 3
1
2
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
-80000.
00
-60000.
00
-40000.
00
-20000.
00
0.00 20000.
00
40000.
00
60000.
00
80000.
00
New PEM Variable-point PEM Monte Carlo
14
Table5. 18 Design Variables
Table 6. Three Output Variables
... weight density (lb/cubic inch)λ3 0.10=
... modulus or rigidity (lb/sq inch)G3
3.9 106
⋅=
... length (inches)L3
8=
... diameter (inches)d3
2.25=Shaft 3
(Aluminum)
... weight density (lb/cubic inch)λ2 0.16=
... modulus or rigidity (lb/sq inch)G2
6.2 106
⋅=
... length (inches)L2
12=
... diameter (inches)d2
1.825=Shaft 2
(Titanium)
... weight density (lb/cubic inch)λ1 0.28=
... modulus or rigidity (lb/sq inch)G1
11.7 106
⋅=
... length (inches)L1
10=
... diameter (inches)d1
2=Shaft 1
(Steel)
Design Parameters (18 total)
Disk 1
(Steel)D
112= ...diameter (inches)
t1
3= ...thickness (inches)
ρ 1 0.28= ... weight density (lb/cubic inch)
Disk2
(Aluminum)D
214= ...diameter (inches)
t2
4= ...thickness (inches)
ρ 2 0.10= ... weight density (lb/cubic inch)
... weight density (lb/cubic inch)λ3 0.10=
... modulus or rigidity (lb/sq inch)G3
3.9 106
⋅=
... length (inches)L3
8=
... diameter (inches)d3
2.25=Shaft 3
(Aluminum)
... weight density (lb/cubic inch)λ2 0.16=
... modulus or rigidity (lb/sq inch)G2
6.2 106
⋅=
... length (inches)L2
12=
... diameter (inches)d2
1.825=Shaft 2
(Titanium)
... weight density (lb/cubic inch)λ1 0.28=
... modulus or rigidity (lb/sq inch)G1
11.7 106
⋅=
... length (inches)L1
10=
... diameter (inches)d1
2=Shaft 1
(Steel)
Design Parameters (18 total)
Disk 1
(Steel)D
112= ...diameter (inches)
t1
3= ...thickness (inches)
ρ 1 0.28= ... weight density (lb/cubic inch)
Disk2
(Aluminum)D
214= ...diameter (inches)
t2
4= ...thickness (inches)
ρ 2 0.10= ... weight density (lb/cubic inch)
Kiπ Gi⋅ di⋅
32 Li⋅= Mj
ρ j
gπ⋅ tj⋅
Dj
4⋅=
bK1 K2+
J1
K2 K3+
J2+
−=
cK1 K2⋅ K2 K3⋅+ K3 K1⋅+
J1 J2⋅=
22
1
23
1
3
22
4
2
22
4
1
)2
()2
(
2
2
j
j
j
ji
i
i
i
a
acbb
a
acbb
Dt
dLY
Y
Y
πρπλ
π
π
∑∑==
−+−
−−−
+=
=
=
where
a = 1
Jj1
2Mj⋅
Dj
2
2
⋅=
- Low natural frequency
- High natural frequency
- Weight
(i=1,2,3)
(j=1,2)
15
Table 7. Sensitivity Information Of Low Frequency w.r.t 18 Design Factors
The sensitivity analysis information for the low frequency with respect to the eighteen design variables
(see Table 7) shows that 77.8% of the variation is contributed by D1, D2, T2 and R2. The remaining 14
variables contribute only 22.2% to the total variation of the low frequency. The significant variation
effects of the 14 minor factors are rolled up into a single normal distribution that has a mean of zero but a
non-zero standard deviation. Working through the sensitivity table numbers indicates that a standard
deviation of 6.921373 should be used for that equivalent normal distribution variable. The modified PEM
method can now be run with only the 4 design factors (D1, D2, T2, R2) that have their original means and
standard deviations, and one extra variable x with a normal distribution N(0, 6.921) which is added on.
That is, the output random, y1, which is function of the 4 major input random variables, is modified by the
x normal random variable as shown by the equation
xyy += 11ˆ
The CDF results generated from the new PEM approach (Blue) is compared with the one calculated
from Monte Carlo with 100000 simulations (Red).
EFFECT OF FACTOR VARIATION ON RESPONSE VARIATION
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Factor Partial Deriv St Dev Var Contrib Pct of Total
-------- ------------- ------------ ------------ ------------
DS1 1.205000e+01 2.000000e-01 5.808100e+00 2.6956
L1 -2.530000e+00 1.000000e+00 6.400900e+00 2.9707
G1 2.059829e-06 1.170000e+06 5.808100e+00 2.6956
Lmd1 -1.442288e-11 2.800000e-02 1.630873e-25 0.0000
DS2 1.350000e+00 2.000000e-01 7.290000e-02 0.0338
L2 -2.166667e-01 1.200000e+00 6.760000e-02 0.0314
G2 3.870968e-07 6.200000e+05 5.760000e-02 0.0267
Lmd2 -3.707967e-12 1.600000e-02 3.519749e-27 0.0000
DS3 1.177778e+01 2.250000e-01 7.022500e+00 3.2592
L3 -3.225000e+00 8.000000e-01 6.656400e+00 3.0893
G3 6.794872e-06 3.900000e+05 7.022500e+00 3.2592
Lmd3 -1.791619e-11 1.000000e-02 3.209900e-26 0.0000
D1 -6.000000e+00 1.200000e+00 5.184000e+01 24.0596
T1 -7.066667e+00 3.000000e-01 4.494400e+00 2.0859
R1 -7.571429e+01 2.800000e-02 4.494400e+00 2.0859
D2 -7.725806e+00 1.240000e+00 9.177640e+01 42.5946
T2 -8.650000e+00 4.000000e-01 1.197160e+01 5.5562
R2 -3.460000e+02 1.000000e-02 1.197160e+01 5.5562
------------ -------
Total Var = 2.154650e+02 100.0000
Std Dev = 1.467873e+01
16
Figure 6. CDF Curves from the MCS and New PEM Methods
Table 8 shows the number of simulation runs required by MCS, 3-level PEM, variable-level PEM and
the new PEM approaches. It demonstrates the new PEM significantly reduces the number of simulation
runs for the 18-variable problems. The 78 runs of the new method includes 19 runs for sensitivity analysis
and 59 runs for performing PEM with 4 major factors with 1 extra normal distribution added on.
Table 8. Number of Simulations Required by the
MCS, 3-level PEM, Variable-level PEM and new PEM Approaches
V. Summary
The new PEM provides nearly identical results as Monte Carlo simulation with a large number of runs,
while the number of runs is significantly fewer than the number of Monte Carlo runs, as well as the
previous variable-point PEM. Benchmark examples demonstrate that the new PEM significantly improves
the efficiency of the existing 3-point and variable-point PEM approaches with high accuracy of the four
moments and CDF/PDF results for highly non-linear problems.
In the future work, non-normal distributions of input variables will be taken into account. The
transformation from non-normal to normal distributions need to be considered.
Y1 (Low Freq)
Probability
Method MCS3-Level PEM
(Gen 1)
Variable-Level PEM
(Gen 2)
New PEM
(Gen 3)
No. of Runs 100,000 387,420,489 393,216 78
17
VI. References
1. Emilio Rosenblueth, “Point Estimates for Probability Moments”, Applied Mathematics Modeling,
Vol. 5, October 1981, pp. 329-335, 1981.
2. Miltone Harr, “Reliability Based Design in Civil Engineering”, McGraw-Hill Book Company, New
York, 1987.
3. John T. Christian and Gregory B. Baecher, “The Point-Estimate Method With Large Numbers of
Variables”, International Journal for Numerical and Analytical Methods in Geo-mechanics, 26,
pp1515-1529, 2002.
4. Chengqing Wu, Hong Hao and Yingxin Zhou, “Distinctive and Fuzzy Failure Probability Analysis of
An Anisotropic Rock Mass to Explosion Load”, International Journal for Numerical Methods In
Engineering, 56, pp767-786, 2003.
5. Geethanjali Panchalingam, “Modeling of Many Correlated and Skewed Random Variables”, Appl.
Math. Modeling, Vol. 18, Nov. pp. 635-640, 1994.
6. K.S.Li, “Point Estimate Method for Calculating Statistical Moments, Journal of Engineering
Mechanics, Vol. 118, No. 7, July 1992.
7. Nagaraman Sivakugan and Ali Al-Harthy, “ Probabilistic Solutions to Geotechnical Problems”,
Probabilistic Mechanics & Structural Reliability: Proceeding of the Seventh Specialty Conference,
Worcester Polytechnic Institute, Worcester, Massachusetts, Aug. 7-9, 1996.
8. N. C. Lind, “Modeling of Uncertainty In Discrete Dynamical Systems”, Appl. Math. Modeling, Vol. 7,
June, 1983.
9. H. P. Hong, “An Efficient Point Estimate Method for Probabilistic Analysis”, Reliability Engineering
and System Safety, 1998, 59, pp. 261-267.
10. H. O. Madsen, S. Krenk, N. C. Lind, “Methods of Structural Safety”, Prentice-Hall International
Series in Civil Engineering and Engineering Mechanics, 1986, pp. 44-101.
11. R. E. Melchers , “Structural Reliability Analysis and Prediction”, Ellis Horwood Limited Publishers,
Halsted Press, a Division of John Wiley & Sons, 1987, pp. 104-141.
12. Y. T. Wu and O. H. Burnside , “Efficient Probabilistic Structural Analysis Using An Advanced Mean
Value Method”, Proceeding of 5th
ASCE speciality Conference – Probabilistic Methods in Civil
Engineering, ASCE, New York, pp. 492-495.
13. Liping Wang and Ramana Grandhi, “Efficient Safety Index Calculation for Structural Reliability
Analysis”, Journal of Computers and Structures, Vol. 52, Nov.1, 1994, pp. 103-111.
14. Hyun Seok Seo and Byung Man Kwak, “Efficient Statistical Tolerance Analysis for General
Distributions Using Three-Point Information”, Int. J Prod. Res. 2002, Vol. 40, No. 4, pp931-944.
15. Liping Wang, Don Beeson and Gene Wiggs, “Efficient and Accurate Point Estimate Method for
Moments and Probability Distribution Estimation”, 10th AIAA/ISSMO Symposium on
Multidisciplinary Analysis and Optimization”, August 30 – Sept 1, 2004, Albany, New York.
18
16. Liping Wang, and Ramana Grandhi, “Improved Two-point Function Approximation for Design
Optimization”, AIAA Journal, Vol. 33, No. 9, 1995, pp. 1720-1727
17. Liping Wang, Don Beeson, Gene Wiggs, and Mahidhar Rayasam, “A Comparison Of Meta-modeling
Methods Using Practical Industry Requirements”, AIAA 2006-1811, 47th
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, 1 - 4 May
2006, Newport, Rhode Island.
18. Liping Wang, Don Beeson and Gene Wiggs , “A Robust and Efficient Probabilistic Approach for
Challenging Industrial Applications with High-dimensional and Non-monotonic Design Spaces”,
AIAA 2006-7014, 11th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, 6 - 8
September 2006, Portsmouth, Virginia