Sie sind auf Seite 1von 9

Journal of Mechanical Science and Technology 33 (4) (2019) 1751~1759

www.springerlink.com/content/1738-494x(Print)/1976-3824(Online)
DOI 10.1007/s12206-019-0327-9

Reliability based design optimization using


response surface augmented moment method†
Sanghoon Lee*
Department of Mechanical and Automotive Engineering, Keimyung University, Daegu 42601, Korea

(Manuscript Received August 24, 2018; Revised December 1, 2018; Accepted December 1, 2018)

----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Abstract

A new approach toward reliability based design optimization (RBDO) is proposed based on the response surface augmented moment
method (RSMM). In RSMM, the reliability analysis procedure based on design of experiments (DOE) is combined with the response
surface method (RSM) for numerical efficiency. It utilizes the Pearson system with four statistical moments to calculate the failure prob-
ability, and the progressive update of the response surface facilitates the calculation of these four statistical moments. In this study, a
semi-analytic design sensitivity analysis is performed in connection with RSMM for an efficient implementation of RSMM in RBDO.
The sensitivity of failure probability with respect to the design variables is calculated by direct differentiation and finite difference
method with the Pearson system. It is integrated into a mathematical programming for RBDO and applied to several test examples. It was
demonstrated that the proposed method of RBDO based on RSMM is efficient and robust.
Keywords: Design of experiments; Design sensitivity analysis; Moment method; Reliability based design optimization; Response surface method
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

In this study, we investigate the feasibility of using another


1. Introduction
type of reliability analysis, the moment method, in RBDO.
In the conventional formulation of reliability based design The moment method calculates the probability of failure by
optimization (RBDO), the failure probability, which is defined computing the statistical moments of the performance func-
as a multifold probability integral of the joint probability den- tion and fitting the moments in empirical distribution systems
sity function of random variables over the domain of structural such as the Pearson system, Johnson system, and Gram-
failure, serves as a design constraint. As the analytic calcula- Charlier series [13]. For this purpose, it is necessary to com-
tion of the integral is not practicable, a number of approximate pute the performance function for a set of adequately designed
methods and simulation methods have been developed [1, 2]. calculation points, generally referred to as quadrature points or
Among these methods, the first order reliability method designed experimental points. Numerous researchers from
(FORM) is considered to be a highly efficient computational various fields have contributed to the development of the
method [3, 4], and over the past three decades, the method has moment method [14-17]. Those methods and their mathemati-
been successfully integrated with the mathematical program- cal backgrounds are summarized by Lee et al. [18]. The mo-
ing for RBDO [5-12]. As the calculation of probability of ment methods developed until now have advantages over
failure is computationally demanding and may involve errors FORM in that it does not involve the difficulties of determin-
due to the nonlinearity of functions and non-normality of input ing the most probable failure point (MPP) and are robust and
random variables, numerous research contributions have been convenient while working with the non-normal random vari-
made to improve the numerical efficiency and accuracy of the ables. To overcome the common drawback of the moment
reliability analysis and RBDO based on FORM. These include methods—the rapid increase of the computational cost with
the development of numerically efficient RBDO formulations the dimensionality of the problem—Lee and Kwak [17] de-
[6, 7] and solution strategies [8-10], and these efforts are con- veloped a method referred to as response surface augmented
tinuing. Utilization of other reliability analysis methods in moment method (RSMM). RSMM utilizes a progressively
RBDO, such as the dimension reduction method [11, 12], is updating response surface integrated with the full factorial
another notable development in the field. design of experiments. It was demonstrated that the RSMM
*
Corresponding author. Tel.: +82 53 580 5264, Fax.: +82 53 580 5165
improves the numerical efficiency of the moment method
E-mail address: shlee1222@kmu.ac.kr significantly while retaining the advantages of the full facto-

Recommended by Associate Editor Byeng Dong Youn rial moment method.
© KSME & Springer 2019
1752 S. Lee / Journal of Mechanical Science and Technology 33 (4) (2019) 1751~1759

Until now, RBDO using the moment methods has not been the highest integration order of five, which implies that poly-
reported much in literature. Because of its numerical effi- nomials of order up to five can be integrated exactly. Thus, the
ciency, RSMM can be considered as a suitable candidate moments of a linear function g ( x) can be exactly calculated
among the abovementioned moment methods for application up to the fourth order regardless of f X ( x) . For a nonlinear
with RBDO. In this study, the author’s previous formulation, function g ( x) also, the above approximation can provide
the RSMM, is revisited to ascertain its feasibility for applica- accurate results with negligible error if the behavior of g ( x)
tion with RBDO. The advantages of the moment method are within the domain where f X ( x) has non-trivial value, is
demonstrated within the context of RBDO by a comparison approximately linear. From the perspective of experimental
with the conventional FORM-based methods. To apply design technique, replacing m + a is with level li , the pa-
RSMM to RBDO, a semi-analytic design sensitivity analysis rameters can be expressed in terms of levels and weights,
for probability of failure is proposed. As a single evaluation of {l1 , l2 , l3 , w1 , w2 , w3} .
the failure probability requires a significant number of func- When there are multiple random variables, the three level
tion evaluations, an approximate design sensitivity analysis DOE becomes a 3n full factorial DOE from the product
using the finite difference method (FDM) may result in a sub- quadrature rule [19], and the first four moments of the system
stantial increase of the computational cost of RBDO, particu- response function can be calculated as follows:
larly when the number of design variables is large. The
3 3
RSMM is implemented with nonlinear programing algorithms
m g = å w1×i L å wn×i g (l1×i ,L, ln×i ) (2)
in a conventional double loop formulation and applied to the i1 =1
1
in =1
n 1 n

design optimization of several problems involving probabilis- 12


é 3 3 2 ù
tic design constraints. In Sec. 2, the moment methods and s g = ê å w1×i L å wn×i ( g (l1×i ,L, ln×i ) - m g ) ú
1 n 1 n
(3)
RSMM are shortly reviewed, and in Sec. 3, the semi-analytic ë i1=1 in =1 û
design sensitivity analysis with RSMM is described. Exam- é 3 3 3 ù
b1g = ê å w1×i L å wn×i ( g (l1×i ,L, ln×i ) - m g ) ú s g 3 (4)
ples are described in Sec. 4, and this is followed by the con- ë i1=1
1
in =1
n 1 n
û
clusions.
é 3 3 4 ù
b 2 g = ê å w1×i L å wn×i ( g (l1×i ,L, ln×i ) - m g ) ú s g 4
1 n 1 n
(5)
2. Review of moment method and RSMM ë i1 =1 in =1 û

2.1 Moment estimation using full factorial design of experi-


where b1g and b 2 g are the skewness and kurtosis of
ments
g ( x ) , respectively, and li× j and wi× j imply the j-th level and
For a random variable x , the k-th order statistical moment weight of i-th variable, respectively.
about the origin of a system response function, g ( x) , can be
approximated by using a quadrature formula with m nodes 2.2 Probability calculation using the Pearson system
[19]:
After the first four moments of the system response function
+¥ k
are obtained, the proper member of the Pearson system [13]
E { g k } = ò ëé g ( x)ûù f X ( x)dx can be identified, and the probability of failure, Pf is calcu-

k k k
; w1 éë g ( m + a1s )ùû + w2 éë g ( m + a 2s )ùû + L + wm éë g ( m + a ms )ùû lated as expressed in Eq. (6):
(1) 0
Pf = Pr éëG ( x ) < 0ùû = ò f x ( x )dx = ò f G ( x ) ( G ( x ) ) dG ( x )

G ( x )<0
where m , s and f X ( x) are the mean, standard deviation,
and probability density function of x . m + a is and wi (6)
imply the location of the i-th node and the i-th probability
weight, respectively. The accuracy of the above approxima- where f x ( x ) is the joint probability density function (PDF)
tion highly depends on the number of nodes, m , and it is of the vector of random variables x and f G ( x ) (G ( x )) is the
necessary to select m with caution considering the nonlin- PDF of G ( x ) identified by the Pearson system. The Pearson
earity of g ( x) within the domain where f X ( x) is defined. system is an empirical system of distribution, in which the
In general, to obtain an accurate approximation of the first PDF of a random variable x is the solution of a differential
four moments of g ( x) , which are frequently required by the equation (Eq. (7)):
empirical distribution systems such as the Pearson system, at
least a three node quadrature rule is necessary. For this rule, 1 df ( x) a+x
the optimal parameters {a1 ,a 2 ,a 3 , l1 , l2 , l3} are those from the =- . (7)
f ( x) dx c0 + c1 x + c2 x 2
Gauss-type quadrature formula defined in the weighted space
by f X ( x) and can be determined through various methods as
discussed by Lee et al. [18]. It is mathematically established The coefficients a , c0 , c1 and c2 are determined from
that the parameters thus determined are unique and provide the first four statistical moments of the random variable x
S. Lee / Journal of Mechanical Science and Technology 33 (4) (2019) 1751~1759 1753

and f ( x) is the PDF of x to be determined by solving the x2


equation. The shape of f ( x) varies considerably with the 1 1 1
36 9 36
coefficients, and Pearson classified the different shapes into
seven groups according to the form of the solution of Eq. (8):
1 4 1
9 9 9
2
c0 + c1 x + c2 x = 0 . (8)
x1
1 1 1
Other probability distribution systems such as the Johnson 36 9 36
system and Gram-Charlier series can be used for the PDF
estimation in RSMM. Among all empirical distribution sys-
tems, the Pearson system is known to provide the tail prob-
ability estimation of the highest accuracy [13], and conse- Fig. 1. Experimental layout of RSMM at initial approximation.
quently, this system is adopted in RSMM. However, probabil-
ity estimation by the Pearson system of order less than 10-5 ing g% ( x ) .
might not be reliable, as discussed in Lee and Kwak [17], and (e) Calculate the first four statistical moments of g ( x )
it is noteworthy that the Pearson system is not capable of pre- with Eqs. (2)-(5) and obtain the probability of failure,
dicting a PDF with a multimodal shape [20]. The adequacy of Pr[ g ( x ) < 0] , using the Pearson system in the same manner as
using 4 moments for reliability estimation is discussed in Ref. in FFMM.
[21]. (f) Calculate the influence index at the un-experimented
points. The influence index k i at i-th experimental point is
2.3 Response surface augmented moment method (RSMM) defined as follows:

The 3n full factorial DOE becomes highly inefficient as


the number of random variables increases. This is critical dPf
when working with structures that require extensive analysis ki = . (11)
dg% ( xi )
to evaluate the system response function. The RSMM is pro-
posed to overcome this shortcoming of the full factorial mo-
ment method (FFMM). In RSMM, a response surface is con- The influence index is a measure used to determine the rela-
structed based on the full factorial design of experiment which tive importance of experimental points and devised to deter-
replaces significant amount of computation with approxima- mine the point that is likely to cause the highest variation in
tion. The detailed procedure is as follows: the probability when the experiment is performed at that point.
(a) Establish the 3n full factorial DOE with the levels and (g) Perform an additional experiment at the point with the
weights derived from the distribution of each variable. highest k i .
(b) Calculate system response function g ( x ) at the (h) With the data obtained in step (g), update response sur-
2n + 1 experimental points located on the axis of the mid- face approximation and recalculate data at the un-
level of each variable. Fig. 1 is an example of the two normal experimented points with g% ( x ) . During this stage, the coeffi-
variable case. The numbers in the figure are the probability cients of response surface approximation are updated, and a
weights imposed on the experimental points calculated as new cross product term can be added into the formulation as
follows: in Eq. (12):
n
wi = w1i × w2i × L × wni = Õ wij (9) n n nmix

j =1 g% ( x ) = a + å bi xi + å ci xi2 + å d k xi ( k ) x j ( k ) (12)
i =1 i =1 k =1

where wi is the overall probability weight imposed on the i-


th experimental point, and wij is the weight of the j-th vari- where nmix is the number of cross product terms included in
able at the i-th experimental point. The circled points are those the formulation and i (k ) and j (k ) are the indexes of the
at which the experiments for the initial approximation are first and second variable in the k-th cross product term, re-
performed. spectively, where i (k ) < j (k ) . Candidates for the new cross
(c) With the 2n + 1 data obtained in step (b), build a quad- product terms to be added are determined from the coordinate
ratic response surface without cross product terms, as ex- of the newly added experimental point in step (g), and among
pressed by Eq. (10), using the least square estimation: them, the one that minimizes the regression error is selected.
(i) With newly updated data, calculate the probability of
n n
failure as in step (e).
g% ( x ) = a + å bi xi + å ci xi2 . (10)
i =1 i =1 (j) Repeat the steps from (f) to (i) until the value for the
probability of failure converges.
(d) Complement the data at the un-experimented points us- More details on RSMM can be found in Ref. [17].
1754 S. Lee / Journal of Mechanical Science and Technology 33 (4) (2019) 1751~1759

3. Design sensitivity analysis of probability constraint ¶s g 3 3 3


= å w1,i1 L å wk -1,ik -1 å wk +1,ik +1 L
calculated by RSMM ¶lk ,ik i1 =1 ik -1 =1 ik +1 =1

3.1 Design sensitivity formulations 3


ë 1
(
é g l1,i ,K, ln ,i - m g ù ¶g l ,K, l
n û 1,i1 n ,in ) ( )
One of the advantages of RSMM is that it can provide the
åw
in =1
n ,in
sg ¶lk ,ik
× wk ,ik

design sensitivity of the probabilistic constraint, dPf dd , (20)


without additional evaluations of the g ( x ) by the semi- 2

analytic design sensitivity analysis introduced in this section. ¶s g 3 3 3 3


ë ( )
é g l1,i ,K, ln ,i - m g ù
û
= å w1,i1 L å wk -1,ik -1 å wk +1,ik +1 Lå wn ,in
1 n

Here d is a vector of design variables. RSMM calculates the ¶wk ,ik i1 =1 ik -1 =1 ik +1 =1 in =1 2s g


design sensitivity using the data and response surface previ- (21)
ously obtained during the reliability analysis. The procedure ì é g l ,K , l - m ù 2
¶ b1g 3 3 3 3
ï ( )
= å w1,i1 L å wk -1,ik -1 å wk +1,ik +1 Lå wn ,in í ë û
1,i1 n ,in g
utilizes the finite difference method with the Pearson system ¶lk ,ik s 3
i1=1 ik -1=1 ik +1=1 in =1 ï g
and chain rule differentiation. The procedure is as follows: î
As the probability of failure, Pf , is a function of the four (
¶g l1,i1 ,K , ln ,in ) ×w ü 3
ï æ ¶m g ¶s g ö
´ ý- ç + b1g ÷
statistical moments, m g , s g , b1g and b 2 g , the design ¶lk ,ik
k ,ik
ïþ s g
ç ¶lk ,i ¶lk ,ik ÷
è k ø
sensitivity can be expressed as follows:
(22)
3

dPf ¶Pf d mg ¶Pf ds g ¶Pf d b1g ¶Pf d b2 g ¶ b1g 3 3 3 3


ë ( )
é g l1,i ,K , ln ,i - m g ù
û
= å w1,i1 L å wk -1,ik -1 å wk +1,ik +1 Lå wn ,in
1 n

= × + × + × + ×
dd ¶m g dd ¶s g dd ¶ b1g dd ¶b 2 g dd ¶wk ,ik i1=1 ik -1=1 ik +1=1 in =1 s g3

d b1g 3 æ ¶m g ¶s g ö
DPf d mg DPf ds g DPf DPf d b2 g - ç + b1g ÷
; × + × + × + × . s g çè ¶wk ,i ¶wk ,i ÷
Dm g dd Ds g dd D b1g dd Db 2 g dd k k ø
(23)
(13) 3
ì é g l ,K , l
¶b 2 g 3 3 3 3
ï ë 1,i1 n ,in - m g
ù
û ( )
= å w1,i1 L å wk -1,ik -1 å wk +1,ik +1 Lå wn ,in í
The terms DPf Dm g , DPf Ds g , DPf D b1g and ¶lk ,ik i1 =1 ik -1 =1 ik +1 =1 in =1 ï s g4
î
DPf Db 2 g can be calculated using the finite difference
method with the Pearson system program, and the rest of the ´
(
¶g l1,i1 ,K, ln ,in ï
× wk ,ik ý -
)
ü 4 æ
ç b1g
¶m g
+ b2 g
¶s g ö
÷
ç ¶lk ,ik ÷ø
terms can be obtained from Eqs. (2)-(5) as follows: ¶lk ,ik ïþ s g è ¶lk ,ik

(24)
d mg 3 ¶m g dlk ×ik 3 ¶m g dwk ×ik 4

dd k

¶lk ,ik dd k

¶wk ,ik dd k
(14) ¶b 2 g 3 3 3 3
ë ( )
é g l1,i ,K, ln ,i - m g ù
û
= å w1,i1 L å wk -1,ik -1 å wk +1,ik +1 Lå wn ,in
1 n
ik =1 ik =1
¶wk ,ik i1 =1 ik -1 =1 ik +1 =1 in =1 s g4
ds g 3 ¶s g dlk ×ik 3 ¶s g dwk ×ik
=å +å (15) 4 æ ¶m g ¶s g ö
dd k ik =1 ¶lk ,ik dd k ik =1 ¶wk ,ik dd k - ç b1g + b2 g ÷.
s g çè ¶wk ,ik ¶wk ,ik ÷
ø
d b1g 3 ¶ b1g dlk ×ik 3 ¶ b1g dwk ×ik (25)
=å +å (16)
dd k ik =1 ¶lk ,ik dd k ik =1 ¶wk ,ik dd k
d b2 g 3 ¶b 2 g dlk ×ik 3 ¶b 2 g dwk ×ik
=å +å (17) It is noted that for calculating the partial derivatives of mo-
dd k ik =1 ¶lk ,ik dd k ik =1 ¶wk ,ik dd k ments with respect to lk ×i , the partial derivative of limit state
k

function g ( x ) with respect to lk ×i is required. It is calcu- k

where d k is the design variable related with the k-th random lated using the previously obtained response surface model
variable. For a parameter design problem, d k becomes the g% ( x ) without additional g ( x ) evaluations as follows:
mean values of a random design variable while it becomes
multiple of standard deviation, e.g. 3s k in a tolerance syn- ì0 if ik ¹ m
thesis problem. The partial derivatives in Eqs. (14)-(17) can be (
¶g l1,i1 ,L, ln ,in ) = ïï ¶g ( x ) ¶g% ( x )
í ;
calculated by directly differentiating Eqs. (2)-(5) as follows: ¶lk ,m ï ¶x ¶xk
ïî k
(
x = l1,i1 ,L,lk , m ,L,ln ,in ) (
x = l1,i1 ,L,lk , m ,L,ln ,in ).
(26)
¶m g 3 3
= å w1×i1 L å wk -1×ik -1 å wk +1×ik +1 Lå wn×in
3 3 (
¶g l1×i1 ,K , ln×in ) ×w
¶lk×ik i1 =1 ik -1 =1 ik +1 =1 in =1 ¶lk×ik
k ×ik The derivatives dlk ,i dd k and dwk ,i dd k in Eqs. (14)-
k k

(17) can be obtained from the relationship between lk ×i , wk ×i


(18) k k

and d k . As explained in Sec. 2.1, lk ×i and wk ×i are derived


¶m g 3 3 3 3 k k

¶wk×ik
= å w1×i1 L å wk -1×ik -1 å wk +1×ik +1 Lå wn×in g l1×i1 ,K , ln×in ( ) from the Gauss-type quadrature determined from the distribu-
i1 =1 ik -1 =1 ik +1 =1 in =1
tion of xk . The algorithms to determine the Gauss-type quad-
(19) rature for arbitrary random distributions can be coded into a
S. Lee / Journal of Mechanical Science and Technology 33 (4) (2019) 1751~1759 1755

Table 1. Input variables of Fortini’s clutch example.

Mean STD Distribution


Input Distribution
(mm) (mm) parametersa
x1 Beta 55.29 0.0793 g 1 = h1 = 5.0

x2 Normal 22.86 0.0043 -

x3 Normal 22.86 0.0043 -

x4 Rayleigh 101.60 0.0793 sˆ 4 = 0.1211


Fig. 2. Overrunning clutch assembly (Fortini’s clutch). a: Notations for distribution parameters follow [23].

computer program [18], and the simplest method to calculate Table 2. Result of design sensitivity analysis (case 1).
the derivatives dlk ,i dd k and dwk ,i dd k is to use the finite
k k
FDM with MCS DSA with RSMM FDM with RSMM
difference method with the computer codes to calculate lk ×i k ( Dm = 10-5 m ) (Proposed) ( Dm = 10-5 m )
and wk ×i . As these levels and weights can be determined for
k

arbitrary distributions, the above sensitivity calculation can dPf d m1 1.3269 1.3156 1.3248
also be applied to arbitrary distributions. Another advantage of dPf d m2 1.3206 1.3249 1.3185
this method is that it can be applied to tolerance synthesis
dPf d m3 1.3206 1.3249 1.3185
problems without modifying the derived equations, wherein
the design variable d k is expressed as a multiple of s k . For dPf d m4 -1.3068 -1.3086 -1.3050
a parameter design problem, dlk ,i dd k and dwk ,i dd k be-
k k

come dlk ,i d mk and dlk ,i d mk , respectively, while for a


k k Table 3. Result of design sensitivity analysis (case 2).
tolerance synthesis problem with tolerances specified as three
times of standard deviation, they become dlk ,i d (3s k ) and FDM with MCS DSA with RSMM FDM with RSMM
k

dwk ,i d (3s k ) , respectively. In either case, those derivatives ( Dt = 10-5 t ) (Proposed) ( Dt = 10-5 t )
k

can be conveniently calculated using the finite difference dPf dt1 0.5230 0.5240 0.5364
method.
dPf dt2 0.0277 0.0249 0.0233
By substituting all the calculated terms into Eq. (13), the de-
sign sensitivity of the failure probability can be calculated. dPf dt3 0.0277 0.0249 0.0234
The accuracy of the proposed method is demonstrated in the dPf dt4 0.4046 0.4056 0.4153
following examples.

3.2 Example of design sensitivity analysis variables. Compared to the results obtained by the MCS and
FDM, the error is within a maximum of 1 % in all the cases
The first example is the overrunning clutch assembly examined. The results are also benchmarked with the results
known as the Fortini’s clutch (Fig. 2) [22]. The contact angle obtained by RSMM and FDM. The proposed method provides
y is specified in terms of the independent component vari- stably accurate results. The calculation of the failure probabil-
ables, x1 , x2 , x3 and x4 as follows: ity required 13 calculations of y , while the design sensitivi-
ties are calculated without additional y calculations.
é x + 0.5 ( x2 + x3 ) ù
y = arccos ê 1 ú. (27)
ëê x4 - 0.5 ( x2 + x3 ) ûú 4. RBDO using RSMM with semi-analytic design sen-
sitivities
The distribution of input variables are summarized in Table 4.1 RBDO formulation and implementation
1. The sensitivity of probability Pf = Pr ëé y < 6°ûù is calculated
with respect to μ x and the results are summarized in Table 2. In conventional parameter design, a RBDO problem can
In Table 3, the sensitivity with respect to 3σ x is summarized. generally be formulated as follows:
Both results are compared with those obtained by FDM and
Monte Carlo simulation (MCS) with 1000000 samples for Minimize W (d )
verification. subject to Pr éë g i ( X , P ) < 0ùû £ pi i = 1,L nc (28)
From Tables 2 and 3, it is observed that the proposed
d = m ( X ) , d L £ d £ dU
method of design sensitivity analysis provides highly accurate
results irrespective of the distribution types of input random
variables. Moreover, the sensitivities are calculated accurately where W and gi are the objective function and limit state
with respect to both the mean values and tolerances of input function, respectively, and X and P are the vector of ran-
1756 S. Lee / Journal of Mechanical Science and Technology 33 (4) (2019) 1751~1759

Table 4. Results of simple mathematical RBDO example. Table 5. Distributions of input variables in truss example.

RSMM RIA Variable Distribution Mean STD


Results x1 x2 Obj. x1 x2 Obj. 1 E1 Log-normal 2100000 kg/cm2 210000 kg/cm2
2.000 29.00 41.56 2.000 28.89 41.46 2 E2 Log-normal 2100000 kg/cm2 210000 kg/cm2
2
3 A1 Log-normal 20 cm 2 cm2
fn Obj g1 g2 Obj. g1 g2 2
4 A2 Log-normal 10 cm 1 cm2
calls 27(11)a 342 342 49(11)a 1011 972
5 P1 Gumbel 5000 kg 750 kg
Pf Final Final
Initial Final Initial Final 6 P2 Gumbel 5000 kg 750 kg
MCS MCS
7 P3 Gumbel 5000 kg 750 kg
Pf 1 1.792e-1 2.511e-6 8.500e-6 1.709e-1 7.658e-6 1.000e-5
8 x1 Normal 200 cm 2 cm
Pf 2 2.609e-1 9.998e-4 1.012e-3 2.520e-1 9.996e-4 1.106e-3 9 x2 Normal 200 cm 2 cm
a: Enclosed number is the number of sensitivity calls 10 x3 Normal 200 cm 2 cm

dom design parameters and random system parameters, re- P1 P2 P3 P3 P2 P1


E1, A1
spectively. The vectors of design variables, d , are the mean
values of the design parameter X , with the lower bound and E2, A2
upper bound denoted as d L and dU . For a tolerance synthe- x1 x2 x3
sis problem, the design variables can be expressed in terms of E2, A2
the standard deviations. nc denotes the number of probabil- E1, A1
istic constraints, and pi is the design limit for failure prob- disp1
400
ability of the i-th constraint. In RBDO based on FORM, the
Fig. 3. Truss example with 23 members.
probabilistic constraints are frequently expressed in terms of
the reliability index or performance measure [5-7].
In this work, RSMM is utilized to solve RBDO problems ginal difference in the values of d 2 and the objective func-
with conventional double-loop implementation. The proposed tion is noticed. The probabilities at the solutions were calcu-
semi-analytic design sensitivity analysis is implemented lated by MCS with 1000000 samples for verification, and the
within the optimization loop and the results are compared with actual value of Pf 2 is calculated as 0.001012 for RSMM and
those obtained by the conventional reliability index approach 0.001106 for RIA. The result of FORM contains higher error
(RIA) using FORM. Hasofer-Lind method [3] and Rackwitz- than RSMM and RIA violates the constraint marginally more
Fissler transformation [4] are used for the calculation of reli- than RSMM does. Consequently, the two methods converge
ability index. It is noted that comparison with up-to-date to different solutions. The numbers of function calls are also
RBDO methods based on FORM is not the scope of this paper. listed in Table 4. It is noteworthy that during the optimization
with RIA, the design sensitivity is calculated with FDM;
4.2 Simple mathematical example therefore, forming an opinion on the efficiency of the method
based on the number of function calls appears biased in this
The first example is stated as follows [24]: case. However, we can estimate the number of function calls
Minimize p d12 + d 2 of RIA by assuming that the analytic design sensitivity of
probability constraint is available. As a single evaluation of
subject to Pf 1 º Pr éë g1 º x13 x2 - 95.5 £ 0 ùû £ 0.0010
the probability requires approximately 20 calls of g1 and g 2 ,
Pf 2 º Pr éë g 2 º x12 x2 - 70.7 £ 0 ùû £ 0.0010 (29) an estimated 220 function calls could be saved for g1 and
1.0 £ d1 £ 2.0 g 2 if an analytic design sensitivity is available. Considering
20.0 £ d 2 £ 50.0 .
this, RSMM with the proposed sensitivity analysis exhibits
superior performance compared to the RIA in terms of both
The design variables d1 and d 2 are the mean values of accuracy and efficiency.
random variables x1 and x2 , respectively, and x1 and x2
follow normal distribution with standard deviations of 0.1 and 4.3 RBDO of 23 bar truss structure
3.0, respectively. The initial design is (1.5, 35.0), and optimi-
zation is performed with sequential quadratic programming The last example is that of a practical engineering problem
(SQP). The results together with those obtained by RIA are which was addressed in Ref. [17]. A truss structure that con-
summarized in Table 6. For both methods, the SQP terminated sists of 23 members, as shown in Fig. 3, is considered. It is
according to the criteria that the magnitude of the search direc- symmetric, and there are 10 random variables in the system,
tion vector is less than the predefined threshold 10-4. namely, three random design parameters x = ( x1 , x2 , x3 ) and
It is observed that the two methods determined similar solu- seven random system parameters p = ( E1 , E2 , A1 , A2 , P1 , P2 , P3 ) .
tions with the second constraint activated. However, a mar- The distributions of these random variables are summarized
S. Lee / Journal of Mechanical Science and Technology 33 (4) (2019) 1751~1759 1757

Table 6. Sensitivity analysis results of truss example. Table 7. Distribution of input variables in truss example.

FDM (with RSMM) RSMM with DSA RIA with FDM


DSA with RSMM Dm = 10-3 m ( x8 , x9 , x10 )
dPf dxi d (115.14, 156.21, 217.18) (110.14, 166.21, 204.37)
(Proposed)
Dt = 10-3 t (the rest)
W / W0 0.9796 0.9775
1 2.4941e-8 2.5603e-8
Pf 0.0499 0.0499
2 9.8481e-10 1.0635e-9
3 2.6215e-3 2.6883e-3 Pf (MCS) 0.0510 0.0665
4 2.0681e-4 2.2333e-4 Fn calls 458 4004
5 2.5811e-7 2.6667e-7
6 2.5520e-6 2.5733e-6
The mean values of the random design parameters x
7 6.0369e-6 6.0533e-6
serve as the design variables of the optimization, whose initial
8 -9.5220e-6 -9.5100e-6
values are (200, 200, 200), and the upper limit for the failure
9 -3.3426e-4 -3.3447e-4 probability is set as 0.05. This problem was solved using both
10 -8.9230e-4 -8.9196e-4 RSMM and RIA for comparison. While applying RIA, the
Rackwitz-Fissler transformation was used along with the
in Table 5. It is assumed that the random design parameters Hasofer-Lind algorithm to work with the non-normal random
follow normal distribution, while the other variables follow variables. For both the solution methods, optimization was
non-normal distributions. The design requirement imposed on performed using the SQP, and the results are summarized in
this structure is that the displacement of center span, denoted Table 7.
as disp1 in Fig. 3, is not to exceed 11.5 cm. Thus, the limit The RSMM converged to the solution after four iterations
state function g ( x , p) is defined as follows: involving 458 calculations of g ( x , p) , while RIA determined
the solution with 4004 evaluations of g ( x , p) over 10 itera-
g ( x , p) = 11.5 - disp1 . (30) tions. As observed in Table 7, the solution determined by
RSMM and RIA are dissimilar; and this is due to the dissimi-
Firstly, the failure probability Pr éë g ( x , p) £ 0ùû is calculated lar accuracy of RSMM and FORM in failure probability cal-
using RSMM with the specified input distributions listed in culation. In both cases, the optimization procedure terminated
Table 5. The displacement disp1 was calculated using the with the activation of the design constraint; however, the con-
commercial finite element analysis code, ANSYS. Forty eight firmatory examination with MCS using 50000 samples reveals
analyses were performed until a converged value of failure that the constraint is violated by approximately 33 % in RIA
probability was obtained by RSMM. The distribution of and 2 % in RSMM. Thus, the accuracy of RSMM in the pres-
g ( x , p) was determined to be the Pearson type I, the beta ence of non-normal random variables is adequately demon-
distribution, and the failure probability is calculated as strated through this example. A single evaluation of reliability
0.008421. The failure probability obtained by a MCS with index requires approximately 70 function calls; for this exam-
100000 samples is 0.009081, which demonstrates that the ple, there are 3 design variables, and there were 10 gradient
probability calculation by RSMM is very close to the refer- calls during the optimization. Thus, approximately 10 ´ 3 ´ 70
ence solution by MCS. = 2100 function calls could be saved if analytic design sensi-
Then, the sensitivity of the failure probability with respect tivities were provided, which approximately halves the total
to all the random variables x and p were performed. For computational cost of RIA. Nonetheless, the computational
x , the sensitivity was calculated with respect to the mean efficiency and accuracy of RSMM with the semi-analytic
values of x , while for p , it was calculated with respect to design sensitivity analysis is demonstrated to be superior to
the tolerances defined as three times the standard deviation of that of RIA in this example. However, the number of function
each variable. The results are summarized in Table 6. It is evaluations of RBDO using RSMM does not appear to be
demonstrated that the semi-analytic DSA provides approxi- inexpensive and necessitates the development of a more effi-
mately similar sensitivity results with differences less than cient solution strategy such as single loop or sequential loop
3 % when compared to those obtained by FDM. approach using RSMM.
For the truss structure example, a RBDO problem was for-
mulated as follows:
5. Discussions and conclusions
Minimize Weight = W ( d1 , d 2 , d 3 ) For an efficient implementation of RSMM within a frame-
work for RBDO, a semi-analytic design sensitivity analysis
Subject to Pr éë g ( x , p) < 0ùû £ 0.05
(31) procedure is proposed and verified through several examples.
where xi ~ N ( d i , 2 ) , i = 1, 2,3 It is noted that the sensitivity of failure probability can be ac-
100 £ d i £ 400, i = 1, 2,3 . curately calculated without additional limit state function
1758 S. Lee / Journal of Mechanical Science and Technology 33 (4) (2019) 1751~1759

evaluations using the procedure proposed in this study. Also it MCS : Monte Carlo simulation
was successfully demonstrated that it can be applied to toler-
ance synthesis problems as well as parameter design problems.
References
The proposed approach with RSMM is implemented in a con-
ventional double loop formulation of RBDO and applied to [1] H. O. Madsen, S. Krenk and N. C. Lind, Methods of Struc-
popular examples in literature, including a case from the engi- tural Safety, Prentice-Hall, Englewood Cliffs (1986).
neering field. Thus, the feasibility of applying RSMM to [2] A. D. Kiureghian, Structural reliability methods for seismic
RBDO is adequately established by this work. However, it is safety assessment: A review, Engineering Structures, 95
noteworthy that the accuracy of the proposed method is lim- (1996) 412-24.
ited by the accuracy of the probability distribution system, the [3] A. M. Hasofer and N. C. Lind, Exact and invariant second
Pearson system. As noted in Sec. 2 and the authors’ previous order code format, ASCE Journal of Engineering Mechanics,
publications [17, 20], the estimation of failure probability of 100 (1) (1974) 111-21.
an order less than 10-5 might not be reliable. [4] R. Rackwitz and B. Fiessler, Structural reliability under
In this study, the computational cost of the developed combined random load sequences, Computers and Struc-
method was compared with that of the conventional double tures, 9 (1978) 489-94.
loop strategy using RIA. Although the performance of the [5] B. M. Kwak and T. W. Lee, Sensitivity analysis for reliabil-
proposed method is demonstrated to be higher than that of the ity based design optimization using an AFOSM method,
conventional RIA, it is not the intention of this study to state Computers and Structures, 27 (1987) 399-406.
that this method outperforms FORM based RBDO approaches. [6] T. W. Lee and B. M. Kwak, A reliability-based optimal
The comparison is limited to the double loop implementation, design using advanced first order second moment method,
and the comparison with single loop or serial loop methods Mechanics of Structures and Machines, 15 (1987-1988) 523-
are not covered. Moreover, the feasibility of implementing the 542.
moment based reliability method in a single loop or serial loop [7] B. D. Youn, K. K. Choi and Y. H. Park, Hybrid analysis
formulation has not been addressed. As the RSMM or other method for reliability-based design optimization, Journal of
moment based methods do not rely on the information of most Mechanical Design, 125 (2003) 221-232.
probable point (MPP) for reliability analysis, implementation [8] X. Du and W. Chen, Sequential optimization and reliability
of these methods in single or serial loop RBDO formulations assessment method for efficient probabilistic design, Journal
is not straightforward. It is widely recognized that the effi- of Mechanical Design, 126 (2004) 225-233.
ciency of the conventional double loop implementation is not [9] X. Chen and T. K. Hasselman, Reliability based structural
comparable to that of a single loop or serial loop formulation, design optimization for practical applications, Proceedings
and this is apparently a disadvantage of the proposed RBDO of 38th AIAA/ASME/ASCE/AHS/ASC Structures, Structural
method using RSMM. However, the advantages and feasibil- Dynamics and Materials Conference and Exhibit and
ity of RSMM in RBDO are adequately demonstrated in this AIAA/ASME/AHS Adaptive Structural Forum, Kissimmee,
study, and it is considered that the applicability of the devel- Florida, USA (1997).
oped method to practical engineering applications can be fur- [10] J. Lim, B. Lee and I. Lee, Second-order reliability method-
ther enhanced through subsequent research. based inverse reliability analysis using Hessian update for
In this paper, all the random variables are considered inde- accurate and efficient reliability-based design optimization,
pendent. For correlated random variables, the conventional International Journal for Numerical Methods in Engineering,
methods using transformation can be utilized; this will be 100 (2014) 773-792.
discussed in the author’s subsequent publications. [11] I. Lee, K. K. Choi and D. Gorsich, Sensitivity analyses of
FORM-based and DRM-based performance measure ap-
proach (PMA) for reliability-based design optimization
Acknowledgments
(RBDO), International Journal for Numerical Methods in
This study was supported by the Keimyung University Re- Engineering, 82 (2010) 26-46.
search Grant of 2015 (20150571). [12] B. D. Youn, Z. Xi and P. Wang, Eigenvector dimension
reduction (EDR) method for sensitivity-free probability
analysis, Structural and Multidisciplinary Optimization, 37
Nomenclature------------------------------------------------------------------------
(2008) 13-28.
FORM : First order reliability method [13] N. L. Johnson, S. Kotz and N. Balakrishnan, Continuous
RSMM : Response surface augmented moment method Univariate Distributions, John Wiley & Sons, New York
FFMM : Full factorial moment method (1995).
MPP : Most probable point [14] Y. G. Zhao and T. Ono, Moment methods for structural
RIA : Reliability index approach reliability, Structural Safety, 23 (2001) 47-75.
RBDO : Reliability based design optimization [15] J. R. D’Errico and N. A. Zaino, Statistical tolerancing using
FDM : Finite difference method a modification of Taguchi’s method, Technometrics, 30 (4)
S. Lee / Journal of Mechanical Science and Technology 33 (4) (2019) 1751~1759 1759

(1988) 397-405. Architecture and Building Engineering, 6 (1) (2007) 151-


[16] H. S. Seo and B. M. Kwak, Efficient statistical tolerance 158.
analysis for general distributions using three-point informa- [22] C. M. Creveling, Tolerance Design: A Handbook for De-
tion, International Journal of Production Research, 40 (4) veloping Optimal Specification, Addison-Wesley, Cam-
(2002) 931-44. bridge, MA (1997).
[17] S. H. Lee and B. M. Kwak, Response surface augmented [23] G. J. Hahn and S. S. Shapiro, Statistical Models in Engi-
moment method for efficient reliability analysis, Structural neering, Wiley, New York, USA (1994).
Safety, 28 (2006) 261-272. [24] Q. Xiao, R. Sues and G. Rhodes, Computational strategies
[18] S. H. Lee, W. Chen and B. M. Kwak, Robust design with for reliability based multidisciplinary optimization, Proceed-
arbitrary distributions using Gauss-type quadrature formula, ings of the 13th ASCE EMD Conference (1999).
Structural and Multidisciplinary Optimization, 39 (2009)
227-243.
[19] H. Engels, Numerical Quadrature and Qubature, Aca- Sanghoon Lee is an Assistant Professor
demic Press, New York (1980). at Keimyung University, Korea. He got
[20] S. H. Lee and W. Chen, A comparative study of uncertainty his Ph.D. in mechanical engineering at
propagation methods for black-box-type problems, Struc- Korea Advance Institute of Science and
tural and Multidisciplinary Optimization, 37 (3) (2009) 239- Technology in 2006 and served as a
53. Senior Researcher at Korea Atomic
[21] Y. G. Zhao and Z.-H. Lu, Applicable range of the fourth- Energy Research Institute before joining
moment method for structural reliability, Journal of Asian Keimyung University.

Das könnte Ihnen auch gefallen