Reliability Analysis Methods

Emilio Bastidas-Arteaga, Abdel-Hamid Soubra

To cite this version:

Emilio Bastidas-Arteaga, Abdel-Hamid Soubra. Reliability Analysis Methods. Michael
A. Hicks; Cristina Jommi. ALERT Doctoral School 2014 - Stochastic Analysis and In-
verse Modelling, pp.53-77, 2014, 978-2-9542517-5-2. <http://alertgeomaterials.eu>. <in2p3-
01084369v2>

HAL Id: in2p3-01084369

http://hal.in2p3.fr/in2p3-01084369v2
Submitted on 19 Feb 2015

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
_____________________________________________________________________________________

Reliability Analysis Methods

E. Bastidas-Arteaga and A.-H. Soubra

University of Nantes – GeM Laboratory – UMR CNRS 6183 – Nantes

– France
____________________________________________________________________

There are numerous sources of uncertainties that should be considered in engineer-

ing design. Reliability analysis methods provide a framework to account for these
uncertainties in a rational manner. This chapter presents the First Order Reliability
Method (FORM) and the Second Order Reliability Method (SORM). These methods
are illustrated by academic examples.

1. Introduction
Engineering design aims at providing minimum levels of serviceability and safety
during the structural lifetime. This is a difficult task because there are important
sources of uncertainty that could lead to over- or under-design solutions. For exam-
ple, there are uncertainties related to environmental exposure, loading, material
properties, engineering models, etc. Reliability analysis methods offer the theoretical
framework for considering uncertainties in a comprehensive decision scheme. The
main goal of reliability analysis methods is to evaluate the ability of systems or
components to remain safe and operational during their lifecycle.

The main objective of this chapter is to present and illustrate various reliability anal-
ysis methods that can be used for engineering or research purposes. The chapter
starts with a description of fundamental concepts for reliability analysis. After, we
present the First Order Reliability Method (FORM) and the Second Order Reliabil-
ity Method (SORM). These methods are illustrated by academic examples.

2. Basic concepts for reliability analysis

Reliability methods have been established to take into account, in a rigorous manner,
the uncertainties involved in the analysis of an engineering problem. The failure
probability and the reliability index are used to quantify risks and therefore evaluate
the consequences of failure. In this approach, the governing parameters of the prob-
lem are modeled as random variables. Random variables can be grouped in a ran-
dom vector X where fX(x) is the joint probability density function (PDF).

For reliability analysis, the space D of random variables may be divided into the
failure and the safety regions. The failure region Df is defined by Df = {X | g(X) ≤ 0}
and the safety region, Ds, by Ds = {X | g(X) > 0} where g(X) represents the perfor-
mance function. Notice that g(X)=0 is the boundary between failure and safety re-
gions and it is called the limit state surface.

In the simplest case, the performance function g(X) is expressed as the difference
between the resistance R(X) and the demand or solicitation on the system S(X) – i.e.,
g(X) =R(X) – S(X). In reliability engineering analysis, g(X) is usually expressed in
terms of displacement, strain, stress, etc. The performance functions can be related
to the following structural conditions:

1. Serviceability limit state: under this condition, ‘failure’ is related to a ser-

viceability loss that does not imply a significant decay of structural safety.
For example, if the reliability analysis of a given structural component fo-
cuses on a maximum displacement vmax, the performance function can
write:

� � = �!"# − � � (1)

where vmax could be fixed by standards or particular serviceability con-

straints and v(X) is the displacement of the point of interest that depends on
X random variables (material strength, geometry, load, etc.). In the case of
failure, v(X) > vmax, but the structural component is still considered safe.

2. Ultimate limit state: this condition describes the state at which structural
safety is highly affected and may lead to total failure or collapse. For in-
stance, if the reliability analysis focuses on the bending moment of a beam,
the performance function is:

� � = �! � − �! (2)

where Mr(X) is the resistant bending moment of the beam that depends on
X random variables (material strength, sectional geometry, etc.), and Ms is
the soliciting bending moment. Notice that although Ms is assumed to be
deterministic in eq. (2), this variable may also be considered as a random
variable. In the case of failure, Ms>Mr(X), leading to the collapse of the
beam.

By accounting for these definitions, the failure probability, Pf, is determined by:

Pf = P[g(X) ≤ 0] = ∫ fX (x)dx1 ... dxn (3)

g(X )≤0
Notice that the limit state function can be a linear or a nonlinear function of the basic
variables. FORM can be used to evaluate eq. (3) when the limit state function is a
linear function of uncorrelated normal variables or when the nonlinear limit state
function is represented by a first-order (linear) approximation with equivalent nor-
mal variables. SORM estimates the probability of failure by approximating the non-
linear limit state function (including a linear limit state function with correlated non-
normal variables) by a second-order representation.

3. First Order Reliability Methods (FORM)

The First Order Reliability Method (FORM) makes use of the first and second mo-
ments of the random variables. This method includes two approaches [Hal00]. These
are First-Order Second-Moment (FOSM) and Advanced First-Order Second-
Moment (AFOSM) approaches. In FOSM, the information on the distribution of
random variables is ignored; however, in AFOSM, the distributional information is
appropriately used.

3.1. First Order Second Moment (FOSM)

The First Order Second Moment (FOSM) method makes use of only second-
moment statistics (i.e. mean and standard deviation) of the random variables and it
requires a linearized form of the performance function at the mean values of the
random variables. A first-order Taylor series approximation is used to linearize the
performance function at the mean values of the random variables.

Cornell [Cor69] proposed the original FOSM formulation. Let us consider an ele-
mentary reliability case where a weight is hung by a cable. The load-carrying ca-
pacity or the resistance of the cable is R and the load effect is S. The resistance and
the load will be modeled as independent Gaussian random variables with N(µR, σR)
and N(µS, σS), respectively. In this case, the failure probability Pf is related to the
failure event R < S, and is computed as:

 
Pf = P(R < S) = P(R – S < 0) (4)

A new random variable Z (called performance function) can be introduced:

Z =R – S (5)

The performance function Z is characterized by a mean µZ = µR – µS and a standard

deviation σZ2 = σR2 + σS2. Since R and S are Gaussian, it can be demonstrated that Z
also follows a Gaussian distribution. Figure 1 presents the PDF of Z. It is observed
that the failure probability is related to the event P(Z < 0). Consequently, Pf could be
estimated directly from:

Pf = Φ[(0 – µZ)/σZ] = Φ[– µZ /σZ] =Φ[– β] (6)

where Φ[.] represents the standard normal cumulative distribution function (CDF)
and β = µZ /σZ is the ‘reliability index’ that is also used to quantify risks of failure.
Probability"density"func;on""

fZ(z)" μZ,"σZ"

μZ=β"σZ"
Z"
Pf"
Figure 1: Probability density function of Z

As may be easily seen from Figure 1, the reliability index computed by FOSM rep-
resents the number of standard deviations that separate the mean value of the per-
formance function from the limit state surface Z = 0.

For lognormal random variables, an alternative formulation to eq. (6) could be de-
rived as follows: Assume that R and S are statically independent lognormal varia-
bles, that is, LN(λR, ξR) and LN(λS, ξS). In this case, another random variable Y can be
introduced as

Y=R/S (7)
or

lnY = Z =lnR – lnS (8)

The failure event can be defined as Y < 1.0 or Z < 0. Since R and S are lognormal,
lnR and lnS are normal; therefore, lnY or Z is a normal random variable with mean
λR – λS and standard deviation ξ R 2 + ξ S 2 . The probability of failure can be defined
(similar to eq. (6)) by

(9)
!! !! !!! !!
�! = Φ =Φ − = Φ −�
! ! !!
!! !!!

In the general case where the performance function Z is a function of a vector of n

random variables, i.e.,

Z = g(X) = g(X1, X2,…, Xn) (10)

the Taylor series expansion about the mean value gives:

 n
∂g 1 n n ∂2 g
Z = g(µX )+ ∑ (Xi −µXi ) + ∑ ∑ (Xi −µXi )(X j −µX j )+ ... (11)
i=1 ∂ Xi 2 i=1 j=1 ∂ Xi ∂ X j

where the derivatives are evaluated at the mean values of the random variables (X1,
X2,…, Xn), and µXi is the mean value of Xi. By truncating eq. (11) at the linear
terms, it is possible to obtain first-order approximations of the mean and variance of
Z as follows:

µZ ≈ g(µX1 ,µX2 ,...,µXn ) (12)

and

n n
∂g ∂g
σZ2 ≈ ∑ ∑ Cov(Xi , X j ) (13)
i=1 j=1 ∂ Xi ∂ X j

where Cov(Xi, Xj) is the covariance of Xi and Xj. For uncorrelated random variables,
the variance becomes:

2
n ⎛
∂g ⎞⎟
σZ2 ≈ ∑⎜⎜⎜ ⎟⎟ Var(Xi ) (14)
i=1 ⎝ ∂ Xi ⎠
⎟

Consequently, by estimating µZ and σZ from eqs. (12) and (13), respectively, the
reliability index can be computed as β = µZ /σZ.

The exact failure probability could be derived from the reliability index only in few
cases:
1. if all the Xi’s are statically independent normal variables and if Z is a linear
function of the Xi, then, Z is normal and the probability of failure is given
by eq. (6);
2. if all the Xi’s are statically independent lognormal variables and if g(X) is a
multiplicative function of the Xi’s, then Z = ln g(X) is normal and the prob-
ability of failure is also given by Pf = Φ(–β).

To conclude, in most cases it is not likely that all the variables are statically inde-
pendent normal or lognormal. Nor is it likely that the performance function is a
simple additive or multiplicative function of these variables. In such cases, the relia-
bility index cannot be directly related to the probability of failure; nevertheless, the
equation Pf = Φ(–β) does provide a rough idea of the level of risk. Notice finally that
FOSM approach has the following shortcomings:
• the information about the distribution of the independent variables is not con-
sidered,
• the truncation errors may be significant if g(.) is non-linear,
 • the assessment of reliability index varies when the limit state function is ex-
pressed in different but mechanically equivalent formulations – e.g., (R – S
= 0) or (R / S = 1).

Example 1 Reliability index of a steel beam using FOSM

Let us consider a steel beam subjected to a deterministic bending moment of Ma
=130 kNm. The yield stress Fy and the plastic modulus Zp of the beam are consid-
ered as random variables following normal distributions with the following parame-
ters:

µFy = 250 MPa, σFy = 25 MPa

µZ p = 9×10−4 m 3 , σZ p = 4.5×10−5 m 3

Question 1: Write two different performance functions by considering strength and

stress formulations.

For the strength formulation, the resistance of the beam is a random variable defined
as R = FyZp and the solicitation S is deterministic where S = Ma. Then, the perfor-
mance function becomes

g(Fy , Z p ) = R − S = Fy Z p − M a (15)

For the stress formulation, the yield stress becomes the resistance of the perfor-
mance function, i.e. R = Fy and the solicitation is computed as S = Ma/Z. In this
case, both R and S are random variables and the performance function becomes

Ma
g(Fy , Z p ) = R − S = Fy − (16)
Zp

Question 2: Estimate the reliability index using the strength formulation.

We assume that the random variables are independent. We will first estimate µR and
σR by using eqs. (12) and (14) respectively:

µR ≈ µFy µZ p = (250)(9×10−4 ) = 225kN ⋅m

1
⎡ ⎛ ∂R ⎞⎟
2
⎛ ⎞ ⎤2
2 1
⎢
σR ≈ ⎢Var(Fy )⎜⎜⎜ ⎟⎟ + Var(Z p )⎜⎜ ∂R ⎟⎟⎟ ⎥⎥ = ⎡⎢ 25 2 (9×10−4 ) + ( 4.5×10−5 ) 250 2 ⎤⎥ 2
2 2

⎢ ⎜⎝ ∂Fy ⎟⎟⎠ ⎜⎜ ∂Z ⎟⎟ ⎥ ⎣ ⎦
⎢⎣ ⎝ p⎠ ⎥
⎦
= 25.16 kN ⋅m
The sollicitation is deterministic, then, µS = Ma and σS = 0. Consequently, the relia-
bility index is given by :

µR −µS 225 −130

β= = = 3.77
2 2
σ +σ R S 25.15 2 + 0 2

Question 3: Estimate the reliability index using the stress formulation.

According to the performance function given by eq. (16), we obtain directly the
mean and the standard deviation of the resistance: µR = µFy and σR = σFy . For the
solicitation, µS and σS are computed from eqs. (12) and (14) respectively:

µS ≈ M a / µZ p = 130 / (9×10−4 ) = 144.4 MPa

1
⎡ ⎛ M ⎞⎟ ⎤⎥ 2
2
⎢ ⎜⎜ a⎟ M 130
σS ≈ ⎢Var(Z p )⎜− 2 ⎟⎟ ⎥ = σZ p 2 a = 4.5×10−5 2 = 7.22 MPa
⎢ ⎜
⎜⎝ µZ p ⎟⎠ ⎥ µZ p (9×10−4 )
⎢⎣ ⎥⎦

And the reliability index becomes

µR −µS 250 −144.4

β= = = 4.06
2 2
σ +σ
R S 25 2 + 7.22 2

By comparing the reliability indexes, it is noted that the result depends on the for-
mulation of the performance function. This lack of invariance motivated the devel-
opment of other reliability methods such as that presented in the following sections.

3.2. Advanced First Order Second Moment (AFOSM)

AFOSM is also called ‘Hasofer-Lind’ method. In this method, the assessment of the
reliability index is mainly based on the transformation/reduction of the problem to a
standardized coordinate system. Thus, a random variable Xi is reduced as:

Xi' = (Xi −µX1 ) / σ X1 (i = 1,2,...,n) (17)

where Xi’ is a random variable with zero mean and unit standard deviation. Thus,
Eq. (17) is used to transform the original limit state surface g(X) = 0 into a reduced
limit state surface g(X’) = 0. Consequently, X denotes ‘original coordinate system’
and X’ describes the ‘transformed or reduced coordinate system’. In the standard-
ized coordinate system, the Hasofer-Lind reliability index βHL corresponds to the
minimum distance from the origin of the axes (in the reduced coordinates system) to
the limit state surface:
 βHL = (x'*)T (x'*) (18)

The minimum distance point on the limit state surface is called the ‘design point’. It
is denoted by vector x* in the original coordinate system and by vector x’* in the
reduced coordinate system. These vectors represent the values of all the random
variables, i.e. X1, X2, …, Xn at the design point corresponding to the coordinate sys-
tem being used.

Figure 2 illustrates the reduction of the random variables R and S for a linear per-
formance function such as that described by eq. (5). According to eq. (17), R and S
can be reduced as:

R′ = (R −µR ) / σR and S ′ = (S −µS ) / σS (19)

The substitution of R’ and S’ into the limit state surface (Z=0) gives the new limit
state surface in the reduced coordinate system (Figure 2b):

Z = σR R′ −σS S ′ + µR −µS =0 (20)

(a)" (b)"
S" S’"
Design"point"
R"–"S"="0" Failure'region'
s*" (0,(μR–μS)/σS)"
Failure'region'
Design"point"
μS" (r’*,s’*)" Safety''
region'
βHL"
Safety'region' θR"
R" " αR" R’"
r*" μR" (–(μR–μS)/σR,0)" θS" " α"
α S"
Figure 2: Reduction of coordinates: (a) original coordinates, (b) reduced coordinates

The reliability of the problem is estimated by using eq. (18). It corresponds to the
minimum distance between the limit state surface and the origin (in the reduced
coordinates system). By using simple trigonometry, this distance (reliability index)
can be estimated as:

µR −µS
βHL = (21)
σR2 + σS2

It should be emphasized here that in the present case of a linear limit state surface,
the Hasofer-Lind reliability index corresponds exactly to the reliability index com-
puted from FOSM if both R and S are normal variables. However, this is not the case
for other limit state surfaces or random variable distributions.

From Figure (2b), it is apparent that if the limit state line is closer to the origin in the
reduced coordinate system, the failure region is larger, and if it is farther away from
the origin, the failure region is smaller. Thus, the position of the limit state surface
relative to the origin in the reduced coordinate system is a measure of the reliability
of the system. Notice that the Hasofer-Lind reliability index can be used to calculate
the failure probability as Pf = Φ(–βHL). This is the integral of the standard normal
density function along the ray joining the origin and x’*. It is obvious that the nearer
x’* is to the origin, the larger is also the most probable failure point. The point of
minimum distance from the origin to the limit state surface, x’*, represents the worst
combination of the stochastic variables and is appropriately named the ‘design
point’ or the ‘most probable point MPP’ of failure.

Finally, it should be noted that the Hasofer-Lind reliability index is invariant, be-
cause regardless of the form in which the limit state equation is written [e.g., (R–
S=0) or (R/S=1)], its geometric shape and the distance from the origin remain con-
stant.

For the general case of a non-linear limit state surface, the assessment of the mini-
mum distance can be written as an optimization problem:

minimize D = x′T x′
(22)
subject to g(x′) = 0

By using Lagrange’s multipliers, the minimum distance (for n random variables)

could be estimated as:

*
n ⎛ ∂g ⎞⎟
∑ xi′* ⎜⎜⎜ ⎟⎟
i=1 ⎝ ∂ Xi′ ⎟⎠
βHL = − (23)
2*
⎛ ∂g ⎞⎟
n
⎜⎜
∑ ⎜
⎟⎟
i=1 ⎝ ∂ X ′ ⎠
⎟i

where (∂g/∂X’i)* is the ith partial derivative evaluated at the design point (x1’*,
x2’*,…, xn’*). The design point in the reduced coordinates is:

xi′* = −αi βHL (i = 1,2,...,n) (24)

where αi are the direction cosines along the coordinate axes X’i. They are given by:
 *
⎛ ∂g ⎞⎟
⎜⎜ ⎟
⎜⎝ ∂ X ′ ⎟⎟⎠
i
αi = (25)
2*
n ⎛ ∂g ⎞⎟
⎜⎜
∑ ⎜
⎟⎟
i=1 ⎝ ∂ X ′ ⎠
⎟ i

By using eq. (17), the design point (in the original space) is given by:

xi* = µXi −αi σ Xi βHL (26)

An algorithm was formulated by [Rac76] to compute βHL and xi’* as follows:

Step 1: Define the appropriate performance function g(X).

Step 2: Assume initial values for the design point x*i. The initial design point
is usually assumed to be at the mean value of the n –1 random variables.
For the last random variable, its value is obtained from the performance
function to ensure that the design point is located on the limit state surface
g(X) = 0.

Step 3: Obtain the design point in the reduced space x’*= [x1’*, x2’*,…, xn’*] as

xi′* = (xi* −µXi ) / σ Xi (i = 1,2,...,n) (27)

Step 4: Estimate the partial derivatives of the performance function (∂g/∂X’i)*

with respect to the variables in the reduced space and evaluate them at xi’*.
These derivatives can be estimated from the performance function in the
original space by using the chain rule of differentiation

∂g ∂g ∂ Xi ∂g
= = σX (28)
∂ Xi′ ∂ Xi ∂ Xi′ ∂ Xi i

Define the column vector A such that

*
⎛ ∂g ⎞⎟
Ai = ⎜⎜ ⎟ (29)
⎜⎝ ∂ Xi′ ⎟⎟⎠

Step 5: Compute the reliability index as

A T x ′*
βHL = − (30)
AT A
 Step 6: Determine a vector of directional cosines as

A
α= (31)
AT A

Step 7: Obtain the new design point xi’* for the n –1 random variables.

Step 8: Determine the coordinates of the new design point in the original
space for the n –1 random variables considered in Step 7 as

xi* = µXi + xi′*σ Xi (i = 1,2,...,n) (32)

Step 9: Determine the value of the last random variable in the original space
such that the estimated point belongs to the limit state surface g(X) = 0.

Step 10: Repeat Steps 3 to 9 until convergence of βHL.

Finally, notice that the Hasofer-Lind reliability index βHL can be used to estimate a
first-order approximation of the failure probability as Pf ≈ Φ(–βHL).

Example 2 Assessment of the reliability index by using FORM

(Adapted from [San10])
Let us suppose that the performance function of a problem is defined by

g(X1, X2, X3) = 6.2 X1 – X2 X32

where the random variables X1, X2, X3 follow normal distributions with means µX1 =
20, µX2 = 5, and µX3 = 4; and standard deviations σX1 = 3.5, σX2 = 0.8, and σX3 = 0.4.

Question 1: Estimate the reliability index by using FORM.

Once the performance function is defined (Step 1), we define the coordinates of the
design point in the original space (Step 2):

⎡ 6.2x1* ⎤⎥
x* = ⎢⎢ x1* = µX1 , x2* = µX2 , x3* = = [ 20,5, 4.979 ]
⎢⎣ x2* ⎥⎥
⎦

The Step 3 consists of obtaining the design points in the reduced space:
 x1′* = (x1* −µX1 ) / σ X1 = 0
x2′* = (x2* −µX2 ) / σ X2 = 0
x3′* = (4.979 −µX3 ) / σ X3 = 2.448

Then, (x’*)T = [0, 0, 2.448]. In the Step 4, we estimate the vector containing the
derivatives of the performance function evaluated at the design point (eqs. 28-29):

* *
⎛ ∂g ⎞⎟ ⎛ ∂g ⎞⎟
A1 = ⎜⎜ ⎟ = ⎜⎜ ⎟ σ = (6.2)(3.5) = 21.7
⎜⎝ ∂ X1′ ⎟⎟⎠ ⎜⎝ ∂ X1 ⎟⎟⎠ X1
* *
⎛ ∂g ⎞⎟ ⎛ ∂g ⎞⎟
A2 = ⎜⎜⎜ ⎟⎟ = ⎜⎜ ⎟⎟ σ X = −(x3* )2 σ X2 = −19.84
⎝ ∂ X2′ ⎟⎠ ⎜⎝ ∂ X2 ⎟⎠ 2
* *
⎛ ∂g ⎞⎟ ⎛ ∂g ⎞⎟
A3 = ⎜⎜ ⎟ = ⎜⎜ ⎟ σ = −2x2* x3*σ X3 = −19.92
⎜⎝ ∂ X 3′ ⎟⎟⎠ ⎜⎝ ∂ X 3 ⎟⎟⎠ X3

Therefore, AT = [21.7, –19.84, –19.92]. The first estimate of the reliability index
becomes (Step 5):

A T x ′*
βHL = − = 1.374
AT A

We estimate the vector of directional cosines in Step 6:

⎡ 0.611 ⎤
A ⎢ ⎥
α= = ⎢⎢ −0.559 ⎥⎥
AT A ⎢ −0.561 ⎥
⎣ ⎦

The new design point in the reduced space is estimated for the n – 1 random varia-
bles (Step 7):

x1′* = −α1βHL = (−0.611)(1.374) = −0.840

x2′* = −α2 βHL = (0.559)(1.374) = 0.768

Consequently, the corresponding values in the original space are (Step 8):

x1* = µX1 + x1′*σ X1 = 20 + (−0.840)3.5 = 17.06

x2* = µX2 + x2′*σ X2 = 5 + (0.768)0.8 = 5.614
We estimate afterwards the value of x3* such that it belongs to the limit state func-
tion (Step 9):

6.2x1*
x3* = = 4.34
x2*

The algorithm is repeated until convergence of the reliability index (Step 10). Table
1 presents the results of various iterations. Convergence is reached after three itera-
tions for this example. We found a reliability index βHL = 1.413 that corresponds to a
failure probability Pf ≈ 0.079 computed based on a first-order approximation.

Table 1: Summary of the iterative process to estimate βHL

Variable Iteration Number
1 2 3
x 1* 20 17.06 16.732
x 2* 5 5.614 5.519
x 3* 4.98 4.34 4.335

βHL 1.374 1.413 1.413

x 1* 17.06 16.732 16.709

x 2* 5.614 5.519 5.521
x 3* 4.34 4.335 4.332

3.2.1. Extension of AFOSM to the case of non-normal random varia-

bles
To extend the Hasofer-Lind method to the case of non-normal random variables,
Rackwitz and Fiessler [Rac78] proposed to transform each non-normal random
variable into an equivalent normal random variable with a mean µXe i and standard
deviation σ Xe i . This transformation allows estimating a solution in the reduced space
by using the procedure explained in the previous paragraphs. The equivalent param-
eters evaluated at the design point xi* are given by:

µXe i = xi* −Φ−1 ⎡⎢ FXi ( xi* )⎤⎥ σ Xe i (33)

⎣ ⎦

φ ⎡⎢Φ−1 ⎡⎢ FXi ( xi* )⎤⎥ ⎤⎥

σ = ⎣
e ⎣ ⎦⎦ (34)
f Xi ( xi* )
Xi

where Φ[.] and ϕ[.] are the CDF and the PDF of the standard variate, respectively,
and FXi (.) and f Xi (.) are the CDF and PDF of the original non-normal random vari-
able (some useful Matlab commandes may be found in Appendix). Notice that eqs.
(33-34) are derived by equating the cumulative distribution functions and the proba-
bility density functions of the actual variables and the equivalent normal variables at
the design point on the limit state surface.

Since the equivalent parameters are evaluated at the design point, each iteration
should include the assessment of the equivalent parameters. The before presented
algorithm is thus modified as follows:

• Steps 1 and 2 remain similar.

• Step 3 should include the assessment of the equivalent parameters µXe i and
σ Xe i at the design point for each non-normal random variable. These
equivalent parameters will be used to determine the design point in the re-
duced space as follows

xi′* = (xi* −µXe i ) / σ Xe i (35)

• Steps 4 to 7 remain exactly similar.

• In Step 8, the assessment of the coordinates in the original space becomes

xi* = µXe i + xi′*σ Xe i (36)

• Finally, Steps 9 and 10 remain similar.

3.3. Ellipsoid Approach

The Hasofer-Lind reliability index can be formulated in a matrix form:

βHL = min (X − µ)T C−1 (X − µ) (37)

g(X )=0

where X is a vector containing the n random variables, µ is a vector containning

their mean values, and C is the covariance matrix.

An intuitive interpretation of the reliability index was suggested in Low and Tang
(cf. [Low97] and [Low04]), where the concept of an expanding ellipse (Figure 3) led
to a simple method for computing the Hasofer-Lind reliability index in the original
space of the random variables using an optimization tool available in most spread-
sheet software packages.
When there are only two uncorrelated non-normal random variables X1 and X2, these
variables span a two-dimensional random space, with an equivalent one-sigma dis-
persion ellipse (corresponding to βHL=1 in Eq. (37) without the min.), centered at the
equivalent mean values �! ! and �! ! and whose axes are parallel to the coordinate
axes of the original space. For correlated variables, a tilted ellipse is obtained.

Low and Tang (cf. [Low97] and [Low04]) reported that the Hasofer-Lind reliability
index βHL may be regarded as the codirectional axis ratio of the smallest ellipse
(which is either an expansion or a contraction of the 1−σ ellipse) that just touches
the limit state surface to the 1−σ dispersion ellipse. They also stated that finding the
smallest ellipse that is tangent to the limit state surface is equivalent to finding the
most probable failure point.

Figure 3: Ellipsoid approach for the computation of the Hasofer-Lind reliability

index

3.3.1. Ellipsoid Approach via Spreadsheet

Low and Tang (cf. [Low97] and [Low04]) showed that the minimization of the Has-
ofer-Lind reliability index can be efficiently carried out in the Excel spreadsheet
environment. The spreadsheet approach is simple and easy to understand because it
works in the original space of random variables and does not require the additional
step of transforming X to X’ where X’ is a transformed vector of the random varia-
bles in the uncorrelated Gaussian space. Notice however that the optimization in
original space is not preferred from a computational perspective. This is because the
optimization in a standardized space is mathematically more desirable in nonlinear
optimization. For example, when minimizing the quadratic form of Eq. (37) in the
original space; in some cases, the correct solution is obtained only when the solver
option “use automatic scaling” is activated. As an alternative, Cholesky factorization
of the covariance matrix can be used.
When the random variables are non-normal, the Rackwitz-Fiessler equivalent nor-
mal transformation can be used to compute the equivalent normal mean � ! and the
equivalent normal standard deviation � ! . The iterative computations of � ! and � !
for each trial design point are automatic during the constrained optimization search.
For non-normal random variables, Eq. (37) may be rewritten as

T
⎛ X − µ e ⎞⎟ −1 ⎛ X − µ e ⎞⎟
βHL = min ⎜⎜ ⎟⎟ R ⎜⎜ ⎟ (38)
⎝ σ ⎟⎠
g(X )=0 ⎜ e
⎜⎝ σ e ⎟⎟⎠

where µe and σe are vectors containning the equivalent mean and standard deviations
values, respectively, and R−1 is the inverse of the correlation matrix. This equation
will be used instead of Eq. (37) since the correlation matrix R displays the correla-
tion structure more explicitly than the covariance matrix C.

Additional information on Solver’s options and algorithms can be found in the Mi-
crosoft Excel Solver’s help file and at www.solver.com. The implementation proce-
dure of the ellipsoid approach in the spreadsheet is described in [Low97] and
[Low04] among others. Some Excel files are available at
http://alum.mit.edu/www/bklow.

3.4. Response Surface Method (RSM)

In case of analytically-unknown system response (such as the responses computed
using a finite element/finite difference method), several approaches based on the
Response Surface Method (RSM) can be found in the literature with the aim of cal-
culating the reliability index and the corresponding design point. We present herein
the approach by Tandjiria et al. [Tan00]. The basic idea of this method is to approx-
imate the system response Y(x) by an explicit function of random variables, and to
improve the approximation via iterations. The system response can be approximated
(in the original space of random variables) by:

n n
Y (x) = a0 + ∑ ai xi + ∑ bi xi2 (39)
i=1 i=1

where xi are the random variables (µi and σi being their mean and standard deviation
values, respectively); n is the number of random variables; and (ai, bi) are coeffi-
cients to be determined. The RSM algorithm is summarized as follows:

Step 1: Evaluate the system response Y(x) at the mean value point µ and at the
2n points each at µ ± kσ where k can be arbitrarily chosen (k = 1).
 Step 2: The above 2n+1 values of Y(x) are used to solve eq. (39) and find the
coefficients (ai, bi). Then, the performance function g(x) can be constructed
to give a tentative response surface function.

Step 3: Solve eq. (37) to obtain a tentative design point and a tentative βHL
subjected to the constraint that the tentative performance function of step 2
be equal to zero.

Step 4: Repeat Steps 1 to 3 until convergence of βHL. Each time, Step 1 is re-
peated, the 2n + 1 sampled points are centered at the new tentative design
point of Step 3.

Concerning the numerical implementation of the RSM algorithm described above,

the determination of βHL requires (i) the resolution of eq. (39) for the 2n+1 sampled
points, and (ii) the minimization of βHL given by Eq. (37). These two operations
constitute a single iteration and can be done using the optimization toolbox available
in Excel. Several iterations should be performed until convergence of βHL. Conver-
gence is considered to be reached when the absolute difference between two succes-
sive values of the reliability index becomes smaller than 10–2.

4. Second Order Reliability Method (SORM)

Reliability assessment is relatively simple if the limit state function is linear. How-
ever, most of the limit state functions are nonlinear. The nonlinearity is due to non-
linear relationship between random variables, to the consideration of non-normal
random variables, and/or to the transformation from correlated to uncorrelated ran-
dom variables. Indeed, a linear limit state in the original space becomes non-linear
when transformed to the standard normal space if any of the variables is non-normal.
Also, the transformation from correlated to uncorrelated variables might induce
nonlinearity.

Figure 4 presents examples of linear and nonlinear limit state functions in the re-
duced space. Both limit state functions have the same minimum distance point β, but
the failure regions are different in both cases. The failure probability of the nonlinear
limit state should be less than that of the linear limit state. The FORM approach
approximates the limit state function with a linear function and will therefore pro-
vide the same assessment of the probability of failure for both cases. This approxi-
mation introduces errors in the assessment of the probability of failure. Consequent-
ly, it is preferable to use a higher order approximation for the failure probability
computation.

The SORM method improves the assessment given by FORM by including infor-
mation about the curvature (which is related to the second-order derivatives of the
limit state function with respect to the basic variables). The Taylor series expansion
of a general nonlinear function g(X1, X2,…, Xn) at the design point (x*1, x*2,…, x*n) is
given by:
 n
∂g
g(X1 , X2 ,..., Xn ) = g(x1* , x2* ,..., xn* )+ ∑ (xi − xi* )
i=1 ∂ Xi
(40)
1 n n ∂2 g
+ ∑∑ (xi − xi* )(x j − x *j )+ ...
2 i=1 j=1 ∂ Xi ∂ X j

where the derivatives are evaluated at the design point.

x’2$
Failure'region'

Design$point,$x’*$

β$ g(X’)$=$0$
Safety'region'
x’1$
Figure 4: Linear and nonlinear limit state functions

For reliability analysis, the space of standard normal variables is more convenient
for a second order approximation of g(). In the following, Xi and Yi will refer to
random variables in the original and equivalent uncorrelated standard normal spaces,
respectively. If all the variables are uncorrelated, Yi = (Xi −µXe i ) / σ Xe i where µXe i and
σ Xe i are the equivalent normal mean and standard deviation of Xi at the design point
x i* .

In the Taylor series approximation given by eq. (40), FORM ignores the terms be-
yond the first order term, and SORM ignores the terms beyond the second-order
term (involving second-order derivatives).

Breitung [Bre84] proposed a simple closed-form solution for the probability compu-
tation using the theory of asymptotic approximation as:

n−1
Pf ≈ Φ(−βFORM )∏ (1+ βFORMκi )−1/2 (41)
i=1

where κi represents the principal curvatures of the limit state function at the mini-
mum distance point, and βFORM is the reliability index computed by the FORM
method.
The assessment of Pf requires the computation of κi. Towards this aim, the random
variables Yi (in the Y reduced space) are rotated to another set of variables Yi’, such
that the last Yi’ variable coincides with the vector α where α is the unit gradient
vector of the limit state at the minimum distance point.

Figure 5 describes the problem for two random variables indicating that the problem
consists of a simply rotation of coordinates. This rotation can be carried out by an
orthogonal transformation:

Y′ = RY (42)

where R is the rotation matrix. For instance, in the case of two random variables:

⎡ cosθ sinθ ⎤
R = ⎢⎢ ⎥
⎥ (43)
⎣ −sinθ cosθ ⎦

where θ is the counterclockwise angle of rotation of the axes (Figure 5). For n ran-
dom variables, the reader may refer to [Hal00] for the determination of the matrix R.

Y2#
Y’1# Failure'region'

α##
Design#point,#
(y1*,#y2*)#

β# g(y1,#y2)#=#0#
θ#
# Safety'region'
Y1#
#

Y’2#
Figure 5: Rotation of axis in the standardized space

The matrix R is after used to estimate a matrix A, whose elements are denoted aij, as
follows:

(RDR )
T
ij
aij = , i, j = 1,2,...,n −1 (44)
∇G(y* )

where D is the n × n second-derivative matrix of the limit state function in the

standard normal space evaluated at the design point and |∇G(y*)| is the length of the
gradient vector in the standard normal space.
The last variable Yn coincides with the β vector computed with the FORM approach
in the rotated normal space. In the next step, the last row and last column in the A
matrix and the last row in the Y’ vector are dropped to take this factor into account.
The limit state can be rewritten in terms of a second-order approximation in this
rotated standard normal space Y’ as:  

1 T
yn′ = β + y′ Ay′ (45)
2

where the matrix A has now the size (n – 1) × (n – 1). Afterwards, the curvatures κi
of eq. (41) are computed as the eigenvalues of the matrix A, to estimate the proba-
bility of failure.

Breitung’s SORM method uses a parabolic approximation (it does not consider a
general second order approximation) by ignoring the mixed terms and their deriva-
tives in the Taylor series approximation in eq. (40). This approach uses the theory of
asymptotic approximation to derive the probability estimate. This approximation is
accurate for large values of β (which is the case for engineering purposes). However,
the assessment is less accurate for smaller β.

Example 3 Assessment of the reliability index by using SORM

(Adapted from [San10])
Suppose that the performance function of a problem is defined by

g(X1, X2) = X1X2 – 80

where X1 follows a normal distribution with mean µX1 = 20 and standard deviation
σX1 = 2. X2 follows a lognormal distribution with mean µX2 = 7 and standard devia-
tion σX2 = 1.4.

The results of the FORM approach provide a reliability index βFORM =2.402 and the
two following vectors x* and α:

x * = (17.612, 4.542)
α = (0.497,0.868)

where x* is the final design point in the original space and α is the vector of direc-
tion cosines.

Question 1: Estimate the probability of failure by using SORM and compare the
results with the probability of failure estimated from FORM.
We already have the results of the FORM approach. We will transform X1 and X2
from the original to the standard normal space. Since X2 is lognormal, we use the
Rackwitz and Fiessler procedure to estimate the equivalent parameters:

φ ⎡⎢Φ−1 ⎡⎢ FX2 ( x2* )⎤⎥ ⎤⎥ φ ⎡⎢Φ−1 ⎡⎢ FX2 ( 4.542)⎤⎥ ⎤⎥

σ = ⎣
e ⎣ ⎦⎦ = ⎣ ⎣ ⎦ ⎦ = 0.9
f X2 ( x2* )
X2
f X2 ( 4.542)

µXe 2 = x2* −Φ−1 ⎡⎢ FX2 ( x2* )⎤⎥ σ Xe 2 = 4.542 −Φ−1 ⎡⎣⎢ FX2 ( 4.542)⎤⎦⎥ 0.9 = 6.418
⎣ ⎦

Transforming X1 and X2 from the original to the reduced space gives:

Y1 = (X1 −µX1 ) / σ X1
Y2 = (X2 −µXe 2 ) / σ Xe 2

Then the values of the design point for each random variable in the reduced space
are:

y1* = (17.612−20) / 2 = −1.194

y2* = (4.542−6.418) / 0.9 = −2.085

The first step in SORM is to determine the matrix R from eq. (43). In the rotated
coordinates, the second coordinate need to coincide with the unit gradient vector α.
Figure 6 presents the example in the reduced space including the performance func-
tion and the design point. It also includes the geometrical description of the assess-
ment of θ. Thus, it can be noted that θ = 270º + tan–1(–2.085/–1.194) = 330.194º, and
R becomes

⎡ cosθ sin θ ⎤ ⎡ 0.868 –0.497 ⎤

R = ⎢⎢ ⎥=⎢
⎥ ⎢ 0.497
⎥
⎥
⎣ −sin θ cosθ ⎦ ⎣ 0.868 ⎦

Notice that the elements of R are also easily available from the direction cosines, i.e.
the components of the unit gradient vector α.
 5
Performance function Y’2
4
Design point 3
2
θ1
!
Y2

0
-5 -4 -3 -2 -1 -1 !0 1 2 3 4 5
-2
-3
-4
-5
Y1 Y’1
Figure 6: Rotation of the axis in the reduced space

The next step is to construct the matrix D that contains the second derivatives of the
performance function with respect to each variable in the standard normal space. We
use the chain rule of differentiation for determining the derivatives from the limit
state function in the original space

∂2 g ∂ ⎡⎢ ∂g ∂ X1 ⎤⎥ ∂ X1 ∂ ⎡
= = X2σ X1 ⎤⎥⎦ σ X1 = 0
∂Y12
∂ X1 ⎢⎣ ∂ X1 ∂Y1 ⎥⎦ ∂Y1 ∂ X1 ⎢⎣
∂2 g ∂ ⎡ ∂g ∂ X ⎤ ∂ X ∂ ⎡ e ⎤
⎢ 2⎥ 2
=
∂Y22 ∂ X2 ⎢ ∂ X ∂Y ⎥ ∂Y
=
∂ X ⎢⎣ X1σ X2 ⎥⎦ σ X2 = 0
⎣ 2 2 ⎦ 2 2
2
∂ g ⎡ ⎤
∂ ⎢ ∂g ∂ X2 ⎥ ∂ X1 ∂ ⎡
= = ⎢
⎣ X1σ Xe 2 ⎤⎥⎦ σ X1 = σ Xe 2 σ X1
⎢ ⎥
∂Y1Y2 ∂ X1 ⎣ ∂ X2 ∂Y2 ⎦ ∂Y1 ∂ X1

Consequently

⎡ ⎤
⎢ ∂2 g ∂2 g ⎥
⎢ ⎥ ⎡ ⎤
σ Xe 2 σ X1 ⎥ ⎡ 0
⎢ ∂Y12 ∂Y1Y2 ⎥ ⎢ 0
⎥=⎢ 1.799 ⎤⎥
D=⎢ ⎥=⎢ e
⎢ ∂2 g ∂ g ⎥ ⎢⎢ σ X2 σ X1
2
0 ⎥ ⎢ 1.799 0 ⎥⎦
⎢ ⎥ ⎣ ⎥⎦ ⎣
⎢ ∂Y1Y2 ∂Y22 ⎥⎥
⎢⎣ ⎦

We estimate the following partial derivatives for the assessment of A

 ∂g ∂g ∂ X1
= = X2σ X1
∂Y1 ∂ X1 ∂Y1
∂g ∂g ∂ X2
= = X1σ Xe 2
∂Y2 ∂ X2 ∂Y2

These derivatives evaluated at the design point are

⎡ * ⎤ ⎡
⎢ x2σ X1 ⎥ ⎢ (17.612)(2) ⎤⎥ ⎡ 35.223 ⎤
*
∇G(y ) = ⎢ * e ⎥=⎢ =⎢ ⎥
⎢ xσ ⎥ ⎢ (4.542)(0.9) ⎥⎥ ⎢ 4.086 ⎥
⎢⎣ 1 X2 ⎥⎦ ⎣ ⎦ ⎣ ⎦

and the length of the vector is

∇G(y* ) = 35.2232 + 4.086 2 = 35.46

Therefore by applying eq. (44), the matrix A is

⎡ −0.044 0.026 ⎤
A = ⎢⎢ ⎥
⎥
⎣ 0.026 0.044 ⎦

Afterwards, we modify the matrix A by deleting the last row and column. In this
case, we have only one element a11 = –0.044. Then, the eigenvalue of this one ele-
ment matrix is simply the same element: κ1 = a11 = –0.044.

The approximation of the failure probability is estimated from eq. (41)

n−1
Pf ≈ Φ(−βFORM )∏ (1+ βFORMκi )−1/2
i=1

1
≈ Φ(−2.402) ≈ 8.609 ⋅10−3
1+ (2.402)(−0.044)

The probability of failure estimated from FORM is Pf = Φ(–2.402) = 8.144·10–3. The

probability of failure estimate from the SORM procedure is about 5% larger. This
result is expected because of the curvature of the performance function (Figure 6). ∎

5. Conclusions
This chapter presented the First Order Reliability Method (FORM) and the Second
Order Reliability Method (SORM). The First Order Reliability Method (FORM)
makes use of the first and second moments of the random variables. This method
includes two approaches. These are First-Order Second-Moment (FOSM) and Ad-
vanced First-Order Second-Moment (AFOSM) approaches. In FOSM, the infor-
mation on the distribution of random variables is ignored; however, in AFOSM
(called also Hasofer-Lind approach), the distributional information is appropriately
used. It was shown that contrary to FOSM, the Hasofer-Lind method led to an invar-
iant reliability index regardless of the form in which the limit state equation is writ-
ten.

While AFOSM requires the transformation of the limit state surface to a standard
space of random variables, the recent ellipsoid approach led to a simple method for
computing the Hasofer-Lind reliability index in the original space of random varia-
bles using an optimization tool available in most spreadsheet software packages.

In case of analytically-unknown system response (i.e. when the system response is

computed using a finite element/finite difference method), the Response Surface
Method (RSM) can be used to calculate the Hasofer-Lind reliability index and the
corresponding design point. The basic idea of this method is to approximate the
system response by an explicit function of random variables, and to improve the
approximation via iterations.

For the computation of the failure probability, it was shown that the Hasofer-Lind
reliability index can be used to evaluate the failure probability when the limit state
function is a linear function of uncorrelated normal variables or when the nonlinear
limit state function is represented by a first-order (linear) approximation with equiv-
alent normal variables. SORM estimates the probability of failure by approximating
the nonlinear limit state function (including a linear limit state function with corre-
lated non-normal variables) by a second-order representation.

References
[Bre84] Breitung, K. Asymptotic approximations for multinormal integrals. J.
Eng. Mech., ASCE, 110(3), 357–367, 1984.

[Cor69] Cornell, C. A. Bounds on the reliability of structural systems. J. struct.

div., ASCE, 93(1), 171-200, 1967.

[Low97] Low, B. K., and Tang, W. H. Efficient reliability evaluation using spread-
sheet. J. Eng. Mech., 123(7), 749–752, 1997.

[Low04] Low, B. K., and Tang, W. H. Reliability analysis using object-oriented

constrained optimization. Struct. Safety, 26, 68–89.

[Hal00] Haldar A. and Mahadevan S. Probability, Reliability and Statistical

Methods in Engineering desing, Jonh Wiley & Sons, Inc. 2000..

[Rac76] Rackwitz, R. Practical probabilistic approach to design, Bulletin No.

112, CEB, Paris France, 1976.
[Rac78] Rackwitz, R. and Fiessler R. Structural reliability under combined ran-
dom load sequences, Structural Safety, 22(1), 27–60, 1978.

[San10] Sanchez-Silva, M. Introduction to reliability and risk assessment: theory

and applications in engineering, Ediciones Uniandes, 2010. (In Spanish).

[Tan00] Tandjiria, V., Teh, C.I. and Low, B.K. Reliability analysis of laterally
loaded piles using response surface methods, Struct Saf 22:335–355,
2000.

Appendix: Summary of Matlab® functions for the nor-

mal distribution

Function Description
normpdf(X,mu,sigma) computes the PDF at each of the values in X using the
normal distribution with mean mu and standard deviation
sigma. X, mu, and sigma can be vectors, matrices, or mul-
tidimensional arrays that all have the same size. A scalar
input is expanded to a constant array with the same dimen-
sions as the other inputs. The parameters in sigma must be
positive.
normcdf(X) returns the standard normal CDF at each value in X. The
standard normal distribution has parameters mu = 0 and
sigma = 1.
norminv(P,mu,sigma) computes the inverse of the normal CDF using the corre-
sponding mean mu and standard deviation sigma at the
corresponding probabilities in P. P, mu, and sigma can be
vectors, matrices, or multidimensional arrays that all have
the same size. A scalar input is expanded to a constant
array with the same dimensions as the other inputs. The
parameters in sigma must be positive, and the values in P
must lie in the interval [0 1].