First-Order Reliability Methods (FORM)

-First Order Second Moment Method

The perIormance Iunction oI a system can be written as

'

<

>
state Iailure 0
state limit 0
state saIe 0
) , , , (
2 1 n
X X X g !
where g(X) 0 is known as a limit state surIace and each X indicates the basic load or
resistance variable. Usually, a number oI limit states can be identiIied Ior a system,
with each representing a state oI either ultimate system Iailure, system
unserviceability, or operational malIunction. X such that g(X) 0 is the level set oI
g(-) at level 0.
Eor ease oI analytical development, all the variables are transIormed into their
standardized Iorms.
i
i i
i
X X
X

'
Obviously, the expected value and standard deviation oI X
I
' are, respectively, zero and
unity. The limit state equation must then correspoindingly be written in terms oI the
standardized variables as Iollows;

'

<

>
state Iailure 0
state limit 0
state saIe 0
) ' , , ' , ' (
2 1 n
X X X g !
In general, g(X') is a nonlinear Iunction. However, it can be linearized at some point
X
0
.
( )
0 0 0 0
' ' ) ' ' ( ) ' ( ) ' ( ) ' ( x X x X x x X + + o g g g
T
where
T
n
x
g
x
g
g
1
]
1

) ' ( ) ' ( ) ' (

0 0
1
0
x x x !
Neglecting higher order terms and assuming that X's are statistically uncorrelated, the
expected value and standard deviation oI g(X') is given by
0 0 0
0
' ) ' ( ) ' ( x x x
T
g g g
) ' ( ) ' (
0 0 2

x x g g
T
g

0
0
g
g

II X's are correlated, we can use the principal component transIormation beIore
standardization.
Mean Value First Order Second Moment (MVFOSM)
The MVEOSM method derives its name Irom the Iact that it is based on a Iirst-order
Taylor series approximation oI the perIormance Iunction linearized at the mean values
oI the random variables, and because it uses only second-moment statistics oI the
random variables. In such a case, x'
0
0, and
) ' ( 0 x g g
) ' ( ) ' (
2
0 x 0 x g g
T
g

g
g

Deficiencv of MJFOSM
The Iunction g(-) is linearized at the mean values oI X
i
variables. When g(-) is
nonlinear, signiIicant error may be introduced by neglecting higher order terms.
More importantly, the saIety index Iails to be constant under diIIerent but
mechanically equivalent Iormulations oI the same perIormance Iunction. Such an
invariance problem is circumvented iI the Iirst-order approximations are evaluated at
a point on the Iailure surIace.
Example (Example 7.3 in Haldar & Mahadevan, 2000)
A W16X31 steel section made oI A36 steel is suggested to carry an applied
deterministic bending moment oI 1140 kip-in.
38

F
ksi, 8 . 3

F
ksi
54
Z
in
3
, 7 . 2
Z
in
3
It is quite logical to assume that E
y
and Z are statistically independent
Strength Eormulation
( )( ) 1140 ' ' 1140 ) ( + +
Z Z F v F v
Z F Z F g


Using the MVEOSM method, 975 . 3
Stress Eormulation
( )
( )
Z Z
F v F v
Z
F
Z
F g



'
1140
'
1140
) (
+
+
Using MVEOSM method, 282 . 4
The limit state and the linearized limit state oI the perIormance Iunction are shown in
Eigure 1. Note that diIIerent equivalent Iormulations oI perIormance Iunction will
not change the Iailure (limit state) surIace because the "equivalency" is based on g(-)
0. However, linearized limit state depends on what Iormulation oI perIormance
Iunction is used in mean value Taylor series expansion. This is because the mean
values are not on the Iailure surIace and two perIormance Iunctions are diIIerent away
Irom Iailure surIace. Such an invariance problem is circumvented iI the Iirst-order
approximations are evaluated at a point on the Iailure surIace.
-20 -15 -10 -5 0 5 10 15 20
-15
-10
-5
0
5
10
15
20
Z'
F
y
'
Limit State
Linearized-Strength Formulation
Linearized-Stress Formulation
Eigure 1. Limit State and Linearized Limit State oI the PerIormance Eunction.
Advanced First Order Second Moment Method (AFOSM)
The invariance problem is circumvented iI the Iirst-order approximations are
evaluated at a point on the Iailure surIace. However, there are inIinitely many points
on the Iailure surIace. The value oI
0
obviously depends on the point x'

, around
which the perIormance Iunction is expanded. The minimum value oI
0
is reIerred
to as the reliability index. A corresponding point x'

is denoted by x'

and designated
either the design point or the checking point.
It can be shown that the reliability index deIined by the linear perIormance Iunction
with uncrrelated random variables equals to the distance oI the Iailure surIace to the
origin oI the reduced (or standardized) variable. Eor non-linear perIormance
Iunction with uncorrelated random variables, the reliability index equals to the
distance oI the linearized Iailure surIace to the origin oI the reduced (or standardized)
variable, as shown in Eigure 2.
Eigure 2. Determination oI Design Point when Components oI X are Uncorrelated
The design point, x'

can be Iound by the Iollowing constrained optimization problem:

Minimize: ( )
2 / 1
0
' ' x x
T

Subjected to the constraint: 0 ) ' ( x g
The reliability index becomes
* *
*
' ) ' ( x x
T
g g
) ' ( ) ' (
* * 2

x x g g
T
g

) ' ( ) ' (
' ) ' (
* *
* *
*

x x
x x
g g
g g
T
T
g
HL

The reliability index such deIined is called HasoIer-Lind reliability index Ior the
credit oI their Iirst invention. Using the above
HL
, the design point can be written as
HL
T
HL
g g
g

!
x x
x
x

) ' ( ) ' (
) ' (
'
* *
*
*
Where
i
is the direction cosine along the axes x
i
'*
) ' ( ) ' (
) ' (
* *
*
x x
x
g g
x
g
T
i
i

An algorithm was Iormulated by Rackwitz (1976) to compute
HL
and x
i
* as Iollows:
1. DeIine the appropriate limit state equation
2. Assume initial values oI x
i
*; I 1, 2, ., n and obtain the reduced variates
x
i
* (x
i
* -
Xi
) /
Xi
. Typically, the initial design point may be assumed to be
at the mean values oI the random variates.
3. Evaluate g(x'*)and at x'*.
4. Express the new design point, x'*, in terms oI
HL
,
HL
! x
*
' .
5. Substitute the new x'* in the limit state equation g(x'*), and solve Ior
HL
.
6. Using the
HL
value obtained in Step 5, reevaluate
HL
! x
*
'
7. Repeat Step 3 through 6 until
HL
coverges.
This algorithm is shown geometrically in Eigure 3. The algorithm constructs a linear
approximation to the limit stat at every search point and Iinds the distance Irom the
origin to the limit state. In Eigure 3, Point B represents the initial design pint,
usually assumed to be at the mean values oI the random variables, as noted in Step 2.
Note that B is not on the limit state equation g(X') 0. The tangent to the limit state
at B is represented by the line BC. Then AD will give an estimate oI
HL
in the Iirst
iterations, as noted in Step 5. As the iteration continues,
HL
converges.
Eigure 3 Algorithm Ior Einding
HL
.
II the components oI X' are correlated and have a covariance (correlation) matrix,
(symmetric), a standardized random vector, Z, whose components are uncorrelatd can
be constructed by principal component transIormation
' X Z
where I
T
or
T
In this case, the limit state equation must also be transIormed Ior the estimation oI the
limit state probabilities into
0 ) ( ) (
1
Z Z g g
T
Then, the random vector, Z, with uncorrelated components and the limit state
equation in terms oI Z, Iorm exactly the same analytical problem as solved Ior X' and
g(X') 0.
It can be shown that
'* '* * *
1
x x : :
T T
HL


Hence, the saIety index Ior the cases involving x with correlated components can also
be derived without transIorming X' into Z. The design point, x'

can be Iound by the

Iollowing constrained optimization problem (as shown in Eigure 4):
Minimize: ( )
2 / 1
0
' ' x x
T

Subjected to the constraint: 0 ) ' ( x g

II a random vector X' is Gaussian with uncorrelated (independent) components, the
joint pdI is given by
( )

,
_

' '
2
1
exp
2
1
) ' ( x x x
T
n
f

while iI X' is Gaussian with correlated components

( )

,
_

' '
2
1
exp
2
1
) ' (
1
2 / 1
x x x
T
n
f

It Iollows Irom the deIinition oI the reliability index that the design point x'* is the
point oI maximum likelihood (Eigure 2 and Eigure 4); and the reliability index
represents the shortest distance between the origin and the limit state surIace in Z
space.
Eigure 4 Determination oI Design Point when Components oI X are Correlated
Estimation of Limit State Probabilities
Eigure 5 Tangent plane to g(X)0 at x'*
Limit State under Multiple Limit Conditions
Eigure 6 Limit State under Multiple Limit Conditions