Beruflich Dokumente
Kultur Dokumente
Chapter 12
Reliability-Based Design
12.1 Introduction
In this chapter, we will present the general RBD model, discuss approaches of solving
RBD problems, and demonstrate the implementation of RBD through examples.
1
Probabilistic Engineering Design
variables could be random design variables such as the dimension variables that define
the geometry of a component. We use X to denote a random design variable and X to
denote a vector of random design variables ( X 1 , X 2 ,L, X n ) , where n is the number of
random design variables. Examples of random design variables include the dimensions of
a mechanical component. The uncertainties of the design variables, which represent the
dimensions of the component, may come from the variations introduced from
manufacturing processes, installation, wear, and many other reasons.
It should be noted that a random design variable is characterized by its distribution and
more precisely its distribution type and parameters such as means and standard
deviations. We cannot control a random variable itself; all we can control is the
distribution types and parameters. Therefore, the actual design variables are the
distribution parameters. In this chapter we only consider the mean values
µ X = ( µ X 1 , µ X 2 , L, µ Xn ) of random design variables. Therefore the actual design
variables are d = (d1 , d 2 , L , d m ) and µ = ( µ1 , µ2 , L, µ n ) . However, the method
discussed in this chapter is applicable for the general situation where other distribution
parameters (for example, standard deviation) are also treated as design variables.
Random parameters are those random variables that can not be controlled by designers,
for example, the environmental parameters, noises, and operation conditions. The vector
of random parameters is denoted by P = ( P1 , P2 ,L, Pr ) , where r is the number of random
parameters. Random parameters are also called noise factors in robust design.
Design objectives
In reliability-based design, the focus is on achieving the reliability requirement, and the
objective is usually calculated at the mean values of all the random variables. Therefore,
the objective function is express by
f (d, µ X , µ P ) , (12.1)
2
Chapter 12 Reliability-Based Design
Design constraints
There are two types of constraints. The first type is the reliability constraint. A reliability
constraint is given by
where the constraint function g i (d, µ X , µ P ) is evaluated at the means of random variables
X and P, and ng is the number of deterministic constraints.
The purpose of reliability-based design is to meet the expectation of high reliability and
quality. In general, a higher reliability implies a greater cost, and a lower reliability also
implies a greater cost due to failure consequences. The following typical reliability-based
design model formulates the trade-off between a higher reliability and a lower cost.
Min f (d, µ X , µP )
( d ,µX )
s.t.
P{Gi (d, X, P) ≤ 0} ≥ Ri , i = 1,2,L , nG
(12.4)
g j (d, µ X , µP ) ≤ 0, j = 1, 2,L , ng
d k ≤ d k ≤ d k , k = 1, 2,L , m
l u
µs ≤ µs ≤ µs , s = 1, 2,L , n
l u
where d kl and d ku are lower and upper bounds of deterministic design variable d i ,
respectively, and µ sl and µ su are lower and upper bounds of the mean of X s , respectively.
3
Probabilistic Engineering Design
shown in the model in Eq. 12.4 and Fig. 12.1, the optimizer calls reliability analysis
repeatedly in the optimization loop. A reliability analysis is also an iterative process if
FORM or SORM is used since the MPP search is a special minimization problem (see
Chapter 7). Therefore, two nested loops are involved in a reliability-based design process.
The outer loop is the optimization loop, and the inner loop is the reliability analysis loop.
Since the reliability analysis also needs to repeatedly call the limit-state functions,
Gi (d, X, P), i = 1,2,L, nG , the total number of function evaluations will be very large,
especially when the number of random variables and number of reliability constraints are
large. Even though moment matching method is more efficient for the reliability analysis,
it is not generally used because of the low accuracy.
Optimization Loop
Design Variables Reliability Constraints
Reliability Analysis
Monte Carlo
FORM
SORM
In the above design model, the actual reliability of each reliability constraint is evaluated.
We call this model direct reliability-based design model. Because of the efficiency and
accuracy, FORM is commonly used for reliability calculation in solving the direct
reliability-based design model. The flowchart of the direct reliability-based design model
is illustrated in Fig. 12.2.
4
Chapter 12 Reliability-Based Design
Optimization
Deterministic limit-state
Objective and constraint functions Gi , i = 1, 2,L, nG
function evaluation
The reliability for each of the constraint functions Gi (d, X, P), i = 1,2,L, nG is evaluated
by FORM as follows. Let the transformed standard normal variables from X be UiX and
the transformed standard normal variables from P be UiP , where the subscript i indicates
i-th reliability constraint. The MPP is then searched with the following model,
Min β i = Ui
U = ( U ,U )
i iX iP (12.5)
s.t. G i (d, UiX , UiP ) = 0.
The above model can be solved by the algorithm presented in Chapter 7, or by any
optimization algorithm. The solution is the MPP (U*iX , U*iP ) and the reliability index β i .
Then the reliability is calculated by
5
Probabilistic Engineering Design
It should be noted that the MPP search has to be conducted for every reliability constraint
and that generally the MPP of one reliability constraint is different from that of the other
reliability constraints.
Using the direct reliability-based design model, we need to evaluate the reliability
Pr{Gi (d, X, P) < 0}, i = 1,2,L, nG for each of functions Gi (d, X, P) . The outer loop
optimization calls reliability analysis not only at each iteration, but also for computing
derivatives, if finite-difference derivatives are used. If there are multiple reliability
constraints, some of them may have very high reliabilities and will never be active during
the optimization process. Although those never-active reliability constraints are the least
critical, the evaluations of their reliabilities will unfortunately dominate the
computational effort in the RBD process. The solution to improving the computational
efficiency is to perform the reliability assessment only up to the necessary level. Hence,
a formulation of percentile performance (inverse reliability) has been proposed to replace
the reliability constraint formulation. Recall that in Chapter 7, the percentile value, G R , of
a limit-state function G (d, X, P) , corresponding to the required reliability R, is defined as
P{G(d, X, P) ≤ G R } = R . (12.7)
Eq. 12.7 indicates that the probability of G (d, X, P) less than or equal to the R- percentile
performance G R is exactly equal to the desired reliability R. The concept is demonstrated
in Fig. 12.3. If the shaded area, the probability at the left-hand side of Eq. 12.7, is equal
to the desired reliability R, then the point G R on G axis is the R-percentile value of
function G R . From Fig. 12.3, it is seen that that, G R ≤ 0 indicates
Pr{G(d, X, P) ≤ G R } ≥ R , meaning that the reliability constraint is feasible. With the
transformation to inverse reliability, the original constraints that require reliability
assessments can now be converted into equivalent constraints that evaluate the R–
percentile performance. Instead of evaluating the actual reliability Pr{G(d, X, P) ≤ 0} ,
the location of G R will now determine the feasibility of the reliability constraint. The
equivalent constraint is then given by
G R (d, X, P) ≤ 0 . (12.8)
It has also been shown that with the above percentile formulation we can avoid
singularity problems which may occur in solving a direct reliability-based design model
during the iterative reliability assessment procedure.
6
Chapter 12 Reliability-Based Design
GR 0 G
Min f (d, µ X , µP )
( d ,µX )
s.t.
GiR (d, X, P) ≤ 0, i = 1, 2,L , nG
(12.9)
g j (d, µ X , µP ) ≤ 0, j = 1,2,L , ng
d k ≤ d k ≤ d k , k = 1,2,L , m
l u
µs ≤ µs ≤ µs , s = 1, 2,L , n
l u
The percentile value GiR (d, X, P), i = 1,2,L , nG can be calculated by FORM as follows.
Let the transformed standard normal variables from X be UiX and the transformed
standard normal variables from P be UiP . The MPP is then searched with the following
model,
The above model can be solved by the algorithm presented in Chapter 7, or by any
optimization algorithm. The solution will be the MPP (U*iX , U*iP ) . After the MPP
(U*iX , U*iP ) in U-space is transformed to the original space as ( X*i , Pi* ) , the percentile
value is calculated by
7
Probabilistic Engineering Design
It should be noted that the MPP search has to be conducted for every reliability constraint
and that the MPP of one reliability constraint is different from that of the other reliability
constraints.
Min f (d, µ X , µP )
( d ,µX )
s.t.
Gi (d, X *i , Pi* ) ≤ 0, i = 1, 2,L , nG
, (12.12)
g j ( d , µ X , µ P ) ≤ 0, j = 1, 2, L , n g
d k ≤ d k ≤ d k , k = 1, 2,L , m
l u
µs ≤ µs ≤ µs , s = 1, 2,L , n
l u
The flowchart of solving the inverse reliability-based design model is given in Fig. 12.4.
Optimization
Deterministic limit-state
Objective and constraint functions Gi , i = 1, 2,L, nG
function evaluation
d, µ X , f GiR , i = 1,2,L, nG d, µ X
New d, µ X
Deterministic objective
g j , j = 1,2,L, ng
Convergence function f (d, µ X , µ P ) and
check constraint functions
No
g j (d, µ X , µ P ) < 0, j = 1,2,L, ng
Yes
Stop
8
Chapter 12 Reliability-Based Design
Solving the inverse reliability-based design model is more efficient than solving the
direct reliability-based design model. As shown in Fig. 12.4, solving the inverse
reliability-based design model still involves a double loop procedure. Many efficient
reliability-based methods have been proposed. They employ a single-loop strategy where
a serial of cycles of optimization and reliability assessment is employed. In each cycle
optimization and reliability assessment are decoupled from each other; no reliability
assessment is required within optimization and the reliability assessment is only
conducted after the optimization. The key concept is to formulate the optimization model
if the reliability requirement is not satisfied such that the reliability will be improved in
the next cycle of optimization. Hence the design is quickly improved from cycle to cycle
and the computational efficiency is improved significantly.
12.4 Examples
Reliability-Based Cantilever Beam Design
L Py
h Px
The objective of the problem is to minimize the weight or equivalently the cross-area
f (d) = bh ,
where b and h are width and height of the cross section, respectively, and the design
variables are d = ( b, h ) .
Two constraints are considered. The first constraint is that the maximum stress at the
fixed end of the cantilever beam is less than the yield strength S.
6 L Px Py
g1 (d) = ( + )−S ≤0,
bh b h
where Px and Py are external forces; L = 100'' is the length of the beam.
The second constraint is that the tip displacement does not exceed an allowable value D0,
9
Probabilistic Engineering Design
2
4 L3 Px Py
2
g 2 ( d) = + − D0 ≤ 0 ,
E b3h bh3
dMin f (d )
= ( b ,h )
s.t
P
g1 (d ) = 6 L ( Px + y ) − S ≤ 0
bh b h
2
4 L3 Px Py
2
g 2 (d ) = E b3h + bh 3 − D0 ≤ 0
1 ≤ b ≤ 10
1 ≤ h ≤ 20
The distributions of random variables are given in Table 12.1. The required reliability of
each of the reliability constraints is R1 = R2 = 0.999 . The reliability-based design model
is given by
dMin f (d )
= ( b ,h )
s.t.
P 6 L ( Px + Py ) − S ≤ 0 ≥ R
bh b h
1
3 2
4 L Px Py
2
P +
3 3 − D 0 ≤ 0 ≥ R2
E b h bh
1 ≤ b ≤ 10
1 ≤ h ≤ 20
10
Chapter 12 Reliability-Based Design
The direct reliability-based design method is used to solve the problem. The results are
shown in Table 12.2. For comparison, the deterministic optimal results are also displayed
in the table.
1) The objective of the deterministic design is better than that of reliability-based design
since the former has a lower cross-sectional area. However, the reliability is very low
in the deterministic design.
2) The reliability-based design satisfies the reliability requirement for both constraints.
The reliability of the second constraint is equal to the required value, and this
constraint is active at the optimal design. Therefore, the reliability-based design has
much higher probability in satisfying the design feasibility under uncertainty. This
will result in higher quality and safety.
3) The reliability-based design is much computationally expensive. It calls the analysis
functions (objective and constraint functions) 5312 times, while the deterministic
design only needs 49 function evaluations.
12.5 Conclusion
11
Probabilistic Engineering Design
Reference
Du, X. and Chen, W., 2004, “Sequential Optimization and Reliability Assessment for
Probabilistic Design,” ASME Journal of Mechanical Design, Vol. 126, No. 2, pp. 225-
233, 2004.
Wu, Y.-T., Shin Y., Sues, R., and Cesare M., 2001, Safety-Factor based Approach for
Probabilistic-based Design optimization. 42nd AIAA/ASME/ASCE/AHS/ASC
Structures, Structural Dynamics and Materials Conference and Exhibit, Seattle,
Washington.
Wu, Y.-T. and Wang, W., “A New Method for Efficient Reliability-Based Design
Optimization,” Probabilistic Mechanics & Structural Reliability: Proceedings of the 7th
Special Conference, American Society of Civil Engineers, New York, 1996, pp. 274-277.
12