Vous êtes sur la page 1sur 6

Reliability Analysis 303

2.5 b = 1, a = 0.5 b = 1, a = 1 b = 1, a = 1.5 b = 1, a = 5

2.0

1.5

1.0

0.5

0.0 0.0 0.5 1.0 1.5 2.0 2.5

FIGURE 5.10: Weibull distribution function fX with several mean values and standard deviations.

5.4 Reliability Analysis Methods


We have reached the key topic of this chapter, reliability analysis methods. Our focus is on those that are the most commonly employed: Monte Carlo simulation, the rst-order reliability method (FORM), the second-order reliability method (SORM), importance sampling, and the response surface method. In addition, we discuss the transformation of random variables into standard normal distribution that is necessary to employ some of these methods. We assume a single failure mode in a product component in this discussion. Multiple failure modes or system failure are discussed in Section 5.5.

5.4.1 The Limit State Function


The rst step in reliability analysis is to identify the failure mode in which performance exceeds design limitdthat is, how the product or component fails. The mode must be written in a mathematical form that involves random variables. Such a mathematical representation is called the limit state function. A vector of random variables X X1 ; X2 ; .; Xn T with joint probability density function fX(x) can be used to model the uncertainties of a physical problem. For a structural problem, X can be used to model uncertainties in loads, yield strength, geometric dimensions, material properties, and so forth. Realization of X X1 ; X2 ; .; Xn T is denoted x x1 ; x2 ; .; xn T ,

304

Chapter 5

which is a point in the n-dimensional space. A limit state function of these variables can be written as gx gx1 ; x2 ; .; xn 0 which divides the x-space into two regions, safe region Rs and failure region Rf : gx  > 0; x Rs  0; x Rf (5.41)

(5.42)

If in the limit state function x is replaced by random variables X, the so-called safety margin M is dened as M gX (5.43)

where M is a random variable. The probability of failure Pf of a structure with this failure mode is thus Z Z (5.44) fX x d x fX xdx Pf P M  0 P g X  0
gx0 Rf

and the reliability of the structure can be simply obtained as PR 1 Pf . The probability integration in Eq. 5.44 is visualized with a two-dimensional case in Figure 5.11a, which shows the joint PDF fX(x) and its contour projected on the x1-x2-plane. All points on the projected contours have the same values of fX(x) or the same probability density. The limit state function g(x) 0 is also shown. The failure probability Pf is the volume underneath the surface of the joint PDF fX(x) in the failure region g(x)  0. To show the integration more clearly, the contours of the joint PDF fX(x) and the limit state function g(x) 0 are plotted on the x1-x2-plane, as shown in Figure 5.11b. The direct evaluation of the probability integration of Eq. 5.44 is very difcult if not impossible. First, the integration is multidimensional since often multiple random variables are involved in engineering problems. Second, the joint PDF fX(x) is in general a nonlinear function. Third, the limit state function g(x) is often nonlinear without an analytical form, in which case a numerical method, such as nite element analysis (FEA), is employed for a solution. For this reason, methods other than direct integration have been developed, which are discussed next.

5.4.2 Monte Carlo Simulation


A powerful statistical analysis tool that has been widely used in both engineering and nonengineering applicationsdMonte Carlo simulation, or, simply, simulationdis the simplest and most reliable analysis method among many others. After a set of random

Reliability Analysis 305

FIGURE 5.11: Probability integration using a two-dimensional example: (a) isometric view and (b) projected view on the x1-x2-plane.

variables is identied and a failure mode and associated limit state function are dened, Monte Carlo simulation involves three major steps: Step 1: Sampling on random input variables X. Step 2: Evaluating the limit state function g(x). Step 3: Statistical analysis of the outcome of the limit state function. We assume that the CDFs of these respective random variables are known. To simplify our discussion, we further assume that these random variables are independent. Note that Monte Carlo simulation is not limited to problems of independent random variables.

306

Chapter 5

The purpose of sampling the input random variables is to generate samples that represent distributions of the input random variables from their respective CDFs FXi xi ; i 1; n. For each random variable, a set of random variable values z z1 ; z2 ; .; zm T between 0 and 1 is generated rst. Note that m is the prescribed number of sample points. These samples z [0,1] are then transformed into sample values of random variable Xi following a given CDF FXi xi by 1 zj ; j 1; m xij FX i (5.45)

1 z is the inverse of the CDF of the ith random variable X . For example, if the where FX j i i random variable Xi is normally distributed with N mxi ; sxi , then

  xij mXi ; j 1; m zj FXi xij F sXi Thus, xij mXi sXi F1 zj ; j 1; m

(5.46a)

(5.46b)

Note that some software, such as MATLAB, generates sample points for xij directly, in which case the steps just described may not be necessary. The MATLAB function normrnd(MU, SIGMA, n, m) returns a set of sample points in an n m matrix chosen from the normal distribution with mean value m and standard deviation s. Once the sample values of all random variables are generated, the limit state function yj g(xj) is solved for each sample point xj x1j ; x2j ; .; xnj T ; j 1; m in step 2. Note that y y1 ; y2 ; .; ym T gx1 ; gx2 ; .; gxm T is the vector of m limit state function values for the respective m sample points. After m samples of output y are obtained, a statistical analysis can be carried out to estimate the failure probability (in addition to characteristics of the output such as mean value and standard deviation) using
m mf 1 X I g xj pf m m j1

(5.47)

where I() is an indicator function dened as I gx  1; if gx  0 0; otherwise (5.48)

and mf is the number of sample points that yield a nonpositive limit state function. We are essentially counting the number of failures among the total sample points generated.

Reliability Analysis 307 EXAMPLE 5.4


We employ Monte Carlo simulation to estimate the failure probability of the cantilever beam discussed in Section 5.2. Recall that the stress s and yield strength Sy are two random variables of normal distribution with s w N ms ; ss 40; 8 ksi and Sy w Nmsy ; ssy 50; 6 ksi. The limit state function for the stress failure mode of the beam was dened as g x S y s The failure probability calculated was Pf 0:159 in Section 5.2, which is analytical. Solution If we simply use 10 sample points, m 10, we can use the MATLAB function normrnd (MU,SIGMA,1,10) to generate sample points for s and Sy directly, as shown in Table 5.2 (left two columns). Note that with only 10 sample points, three values of the g function are less than 0; hence, the failure probability Pf is 0.3. It is obvious that the failure probability calculated using only 10 sample points is inaccurate compared with the analytical solution. We need more sample points. Table 5.2a: Sample points for s, Sy and g function
s 27.9961 35.6724 43.7378 36.8566 41.2556 50.2375 6.8483 50.5365 36.1466 32.7855 Sy 60.0879 42.0859 48.5422 45.0047 51.3974 37.4567 45.6715 42.6835 60.8851 48.5281 g Sy s 32.0918 6.4135 4.8044 8.1481 10.1418 12.7808 1.1769 7.8530 24.7385 15.7425

Table 5.2b: Increasing sample points to improve accuracy of probability estimate


m 10 100 1000 10,000 100,000 1,000,000 Pf 0.3000 0.1300 0.1510 0.1546 0.1598 0.1588

If we increase the number of sample points, for example from 10 to 1,000,000 as shown in Table 5.2b, the failure probability becomes 0.1588, which is very close to the analytical solution of 0.159. The table clearly shows that when we increase the number of sample points the failure probability calculated using Monte Carlo simulation approaches the analytical solution. Calculating failure probability using Monte Carlo simulation can be implemented using the following MATLAB script.

308

Chapter 5

%Matlab script m=10 s=normrnd(40,8,1,m) Sy=normrnd(50,6,1,m) n=0 for i=1:1:m g=Sy(i)-s(i) if g < 0 n=n+1 end end Pf=n/m

5.4.3 The First-Order Reliability Method


Monte Carlo simulation is straightforward and reliable. However, as demonstrated in Example 5.4, it requires a large number of sample data to ensure an accurate estimate of failure probability. Engineering applications; for example, complex structural problems that require nite element analysis, often requires nontrivial computational time. For such applications, Monte Carlo simulation is infeasible because of the substantial computational effort it requires. A number of methods have been developed to reduce the computational effort involved in Monte Carlo simulation while still offering sufciently accurate failure probability estimates. These methods aim to provide acceptable estimates for the integral form of failure probability Pf dened in Eq. 5.44, which can be achieved in two important steps: Step 1: Simplify the joint probability density function fX(x) by transforming a given joint probability density function, which may be multidimensional, into a standard normal distribution function of independent random variables of the same dimensions. Step 2: Approximate the limit state function g(x) 0 by the Taylor series expansion and by keeping the rst few terms for approximation. This is mainly to ease calculation of the failure probability; it has nothing to do with analysis of product performance. If only linear terms are included in the calculation, the method is referred to as rst-order reliability method. When the second-order terms are also considered, the method is referred to as second-order reliability method. In this subsection we introduce FORM. SORM is discussed briey in Section 5.4.4. For both methods, we assume that the random variables are independent and are normally distributed. Transforming dependent random variables of nonnormal distribution is discussed in detail in Section 5.4.5.

Vous aimerez peut-être aussi