Approximate methods for calculating

probability of failure
• Working with normal distributions is
appealing
• First-order second-moment method (FOSM)
• Most probable point
• First order reliability method (FORM)
• Section 4.1 in Choi, Grandhi, & Canfield
 The normal distribution is attractive
• It has the nice property that linear functions of
normal variables are normally distributed.
• Also, the sum of many random variables tends to be
normally distributed.
• Probability of failure varies over many orders of
magnitude.
• Reliability index, which is the number of standard
deviations away from the mean solves this problem.
β = −Φ −1 ( Pf ) Pf = Φ (− β ) Φ is the normal CDF
 Approximation about mean
• Predecessor of FORM called first-order
second-moment method (FOSM)
 ( X ) g ( µ X ) + ∇g ( µ X )T ( X − µ X )
g=
Then, easy to show
1/2
 n  ∂g  2

( µ X ) σ g  ∑ 
µ g g=  σ xi 
2

 i =1  ∂xi  
(How can you use that to get rid of unimportant variables?)
The reliability index and probability of failure are
µ g
β = Pf =Φ (− β ) Φ is the normal CDF
σ g
 Beam under central load example
Example 4.2
• Probability of exceeding plastic moment
capacity g ( P, L,W ,=
T ) WT − PL / 4

P L W (plastic section T (yield stress)

modulus)
mean 10kN 8m 0.0001m^3 600,000 kN/m^2
Standard deviation 2kN 0.1m 0.00002m^3 100,000kN/m^2

• All normal
 Reliability index for example
• Using the linear approximation get g ( P , L, W , =
T ) WT − PL / 4
µ g = g ( µ X ) = 0.0001× 600, 000 − 10 × 8 / 4 = 60 − 20 = 40kNm (safety factor of 3!)
∂g ∂g
= W= 0.0001 = T= 600, 000
∂T ∂W
∂g ∂g
=−L / 4 =
−2 = −P / 4 =
−2.5
∂P ∂L
1/2
 n  ∂g  2 
=σ g =∑   σ xi 
 i =1  ∂xi 
2



(0.0001×100, 000) 2 + (600, 000 × 0.00002) 2 + ( −2 × 2 ) + (−2.5 × 0.1) 2 =

2
16.13kNm
µ g
β= = 2.48, Pf =Φ (−2.48) = 0.0066
σ g

• Example 4.2 of CGC shows that if we change to g=T-

0.25PL/W we get 3.48 (0.00025)
• MCS with 10 million samples gave 2.85 (0.00218)
 Top Hat question
• In the beam example, the error in estimating
the reliability index was due to
– Linear approximation
– g is not normal
– both
 Most probable point (MPP)
• The error due to the linear approximation is
exacerbated due to the fact that the expansion
may be about a point that is far from the failure
region (due to the safety margin).
• Hasofer and Lind suggested remedying this
problem by finding the most probable point and
linearizing about it.
• The joint distribution of all the random variables
assigns a probability density to every point in the
random space. The point with the highest density
on the line g=0 is the MPP.
 Response minus capacity illustration
r=randn(1000,1)*1.25+10;
c=randn(1000,1)*1.5+13;
f=@(x) x;
20

fplot(f,[5,20])
hold on 15

plot(r,c,'ro')
c

xlabel('r') 10

ylabel('c')

5
5 10 15 20
r
 Recipe for finding MPP with
independent normal variables
• Transform into standard normal variables
(zero mean and unity standard deviation)
xi − µ xi
ui =
σx i

• Find the point on g=0 of minimum distance to

origin. The point will be the MPP and the
distance to the origin will be the reliability
index based on linear approximation there.
(U U )
1/2
β = min T
U ∈g (U ) =
0
 First order reliability method (FORM)
• Limit state g(X). Failure when g<0.
• Linear approximation of limit state together
with assumption that random variables are
normal.
• Approximate around most probable point.
• Then limit state is also normal variable.
• Reliability index is the distance of the mean of
g from zero measured in standard deviations.
Visual
 Linear Example
• For linear example g = R − C, R= N (10,1.25 ) 2
C=
N(13,1.52 )

• Then u = r1.25
− 10
u =
c − 13
1.5
1 g = 1.25u − 1.5u − 3
2 1 2

• The failure boundary g =0 → u =(5 / 6)u − 2 2 1

3

• Distance from origin 2

u12 + ( 5u1 / 6 − 2 )
β2 =
2 1

dβ 2 25 10
u2
-1
= 2u1 + u1 − = 0
du1 18 3 -2
61
u1 = u2 = −1.1528(r =11.27) β =
c= 1.537 -3
60
Pf =Φ (− β ) = 0.0621 -4

-5
-3 -2 -1 0 1 2 3 4
u1
 Check by MCS
• r=randn(1000,1)*1.25+10;
• c=randn(1000,1)*1.5+13;
• >> s=0.5*(sign(r-c)+1);pf=sum(s)/1000=0.0550
• Six repetitions gave: 0.0660, 0.0640,
0.0670,0.0450, 0.0550, 0.0640
• With million samples got 0.06258
• So even with a million samples, accurate only
to two digits.
 Problems FORM
• Repeat the linear example of Slide 12 with a
slightly different limit state g=R2-C2 with both
FORM and MCS.
• Do the beam problem of slides 4 and 5 with
FORM. Note that you are now in 4
dimensional random space.
 General case
• If random variables are normal but correlated, a
linear transformation will transform them to
independent variables.
• If random variables are not normal, can be
transformed to normal with similar probability of
failure. See Section 4.1.5 of CGC (It is called the
Rosenblatt transformation)
• Murray Rosenblatt, Remarks on a Multivariate
Transformation, Ann. Math. Statist. Volume 23,
Number 3 (1952), 470-472.
 Approximate transformation
• If we want the transformed variable u to have the
same CDF as the original x at a certain value of x we
would require Φ(u ) = F ( x)
• Unfortunately we cannot enforce that everywhere.
• If we focus on MPP we get the following
transformation (Eqs. 4.38, 4.39), which must be used
iteratively, starting from some guess.
X* replaced
x − µx' by xMPP in
u= equations.
σ x'

σ x' =
(
φ Φ −1  F ( x MPP )  )
f ( x MPP )
µ x ' x MPP − Φ −1  F ( x MPP )  σ x '
=