APPENDIX D

STRUCTURAL RELIABILITY ANALYSIS

Development of FORM model in using spreadsheet

In this study, FORM model was built using Microsoft Excel and Visual Basic for
Application. A practical procedure presented by Zhao and Ono (1999), Low and Tang
(2004) and Low and Tang (2007) was used to develop FORM model in Excel. The
reliability index was obtained by calling Excel’s built-in optimization program,
Solver, with the objective function to minimize the reliability index subject to the
constraint that the limit-state function, g(x) = 0. The objective function is

Minimize:

where is a vector representing the set of random variables, is the mean value,
is the correlation matrix and is the failure domain. The probability of failure is then
calculated. Figure D1 represents FORM framework for assessing the failure
probability.

185
 Input variables : Distribution type,
parameter for each distribution

No Normal
distribution?

Transform non-normal to normal

distribution using Rackwitz-Fiessler
equivalent normal transformation Specify
Yes

Specify correlation
matrix, R
Normal distribution Calculate nx=(x*-mx)/ x

Calculate the inverse

Multiply the R-1 and nx (A) of R (R-1)

Multiple the transpose of

Transpose nx nx with A (B)

Calculate square root of B

Specify the limit-state

Calculate g(X)
function g(X)

Is g(X) = 0? No

Yes

Calculate the reliability index β

Calculate the probability of failure

PoF = (-β)

Model validation
PoF = (-β)

Figure D1: FORM Model Framework in Excel

186
Estimation of probability distribution and its parameter

It is evident that the limit state function used in this study is a function of five
basic variables (i.e. material strength , diameter , material thickness , operating
pressure and corrosion rate ). To conduct FORM analysis, the variability of the
random variables has to be estimated in terms of their probability distributions and
their parameters:
1. The variability in material property (i.e. material strength )
Material strength is often used to represent the limit state, in structural
reliability problems of corroded systems and hence, is crucial information in
structural reliability assessments. It is also subjected to variations and is
considered as a basic random variable that can be statistically represented.
The variation in material property is typically assumed normally distributed
for it is a reasonable model for many natural processes or physical properties
(Melchers, 1999). Other distributions that have been suggested for
representing the distribution of material strength are lognormal and Extreme
Value Type I distributions. The mean value of the material strength is assumed
to be the value typically specified for specific pipe material under study. COV
will be based on the Swedish yield stress data in Melchers (1999, pg. 278) as
shown in Table D.1.

Table D.1: Swedish yield stress data (Melchers, 1999)

Nominal yield stress (MPa) COV Number of samples
220 0.103 19,857
260 0.099 19,217
360 0.057 11,170
400 0.054 2447
Note: Number of samples is the number of data points to estimate coefficient of variation

2. The variability in pipe properties (i.e. diameter , material thickness )

Relatively the initial pipe diameter and thickness are not available. Therefore,
the variation in pipe diameter and thickness is assumed to follow a normal
distribution since many processes are assumed to follow this distribution
(Melchers, 1999). The mean value of pipe diameter and thickness is assumed
to be the value of the specified diameter and the nominal pipe thickness,

187
 respectively. For COV, Ellingwood et al. (1999) suggested an average COV of
0.05.
According to ASME SA 530 or ASTM A530 specification, the requirement
for pipe wall thickness and diameter tolerance is 12.5%. Assume the pipe wall
thickness is 5.54 mm. The tolerance for the pipe is mm which means
the thickness has to be manufactured between 5.19 to 5.89 mm. Table D.2
shows how the variation in the manufacturing process can be translated into
COV. Figure D2 shows the effect of having different values of COV on the
pipe thickness distribution. Therefore, the COV value of 0.05 suggested by
Ellingwood et al. (1999) is considered in this study since it is within the
tolerance specified by the standard requirement.

Table D.2: Variation in manufacturing process

Manufacturing process COV
= 0.35 0.35 0.06
2 = 0.35 0.175 0.03
3 = 0.35 0.117 0.02
6 = 0.35 0.058 0.01

Figure D.2: The effect of having different values of COV on the pipe thickness
distribution

188
3. The variability in operating parameters (i.e. operating pressure )
To estimate the probability distribution for operating pressure and its
parameters, the actual historical operating pressure data is taken for respective
pipe under study. Then, the data is analyzed to establish its probability
distribution and parameters using the distribution fitting method.
4. The variability in corrosion rate
In structural reliability assessment, the defect growth rate is the primary input
to allow future prediction of the defect size, which in turn allows prediction of
the failure probability. The traditional corrosion analysis assumes a
deterministic defect growth function, even though in reality, the defect growth
rates also exhibit a large stochastic behavior. In this study, the probability
distribution of CUI corrosion rate is to be established using corrosion rate data
collected during inspection period. The method to estimate the probability
distribution for corrosion rate and its parameters has been discussed in Chapter
3 and 5 using the bootstrapping method.

Model Validation

Data from Teixeira et al. (2008) was used to validate the FORM spreadsheet model as
shown in Table D.3. The limit-state function for intact pipes is given by

1 .1 y 2t
g (X) Po
D

where y is the material yield stress, D is the outside diameter of the pipe, t is the

wall thickness of the pipe and Po is the internal operating pressure of the pipe.

Table D.3: Probabilistic models of the basic variables of the intact pipe
Basic variables Distribution Mean Std. Dev.
Yield stress, y Lognormal 410.7 MPa 32.86 MPa
Diameter, D Normal 406.4 mm 0.41 mm
Thickness, t Normal 12.7 mm 0.13 mm
Operating pressure, Po Gumbel 17.28 Pa 1.20 mm

189
Figure D3 and D4 show the Excel template developed and how to run Solver in order
to get the solutions.

Figure D3: Excel template for FORM model

Figure D4: Solver to solve the limit state function

Figure D5 shows the results generated by FORM model developed and it showed that
the result is almost identical as the result published in the paper by Teixeira et al.
(2008). Therefore, the developed model is validated.

190
 Figure D5: Generated results by FORM model

Excel VBA Coding

Below are parts of the VBA coding used in the FORM model. This coding is used to
transform any distributions to normal distribution.

Function EqvN(DistributionName, para1, para2, para3, para4, x, code)

'This function is used to transform any types of distributions to normal distribution

'del is the value assign to x if x < del (to avoid division of 0)

del = 0.0001

'Select the following case for specified distribution.

'UCase function returns a Variant (String) containing the specified string,
'converted to uppercase.

Select Case UCase(Trim(DistributionName))

Case "NORMAL":

If code = 1 Then EqvN = para1

If code = 2 Then EqvN = para2

Case "UNIFORM":

a1 = para1: a2 = para2: min = para3: max = para4

191
 If x <= min Then x = min + del
If x >= max Then x = max - del

cdf = Application.BetaDist(x, a1, a2, min, max)

pdf = Betapdf1(x, a1, a2, min, max)

EqvN = EqvTransform(x, cdf, pdf, code)

Case "LOGNORMAL":

para3 = 0
para4 = 0

If x < del Then x = del

lambda = Log(para1) - 0.5 * Log(1 + (para2 / para1) ^ 2)

If code = 1 Then
EqvN = x * (1 - Log(x) + lambda)
ElseIf code = 2 Then
EqvN = x * Sqr(Log(1 + (para2 / para1) ^ 2))
End If

Case "EXTVALUE1":

para3 = 0
para4 = 0

alfa = 1.2825498302 / para2

u = para1 - 0.5772 / alfa

cdf = Exp(-Exp(-alfa * (x - u)))

pdf = alfa * Exp(-alfa * (x - u)) * cdf

EqvN = EqvTransform(x, cdf, pdf, code)

Case "EXPONENTIAL":

para2 = 0
para3 = 0
para4 = 0

b = para1

If x < del Then x = del

cdf = 1 - Exp(-x / b)
pdf = 1 / b * Exp(-x / b)
192
 EqvN = EqvTransform(x, cdf, pdf, code)

Case "GAMMA":

para3 = 0
para4 = 0

If x < del Then x = del

cdf = Application.GammaDist(x, para1, para2, True)
pdf = Application.GammaDist(x, para1, para2, False)
EqvN = EqvTransform(x, cdf, pdf, code)

Case "WEIBULL":

para3 = 0
para4 = 0
If x < del Then x = del
cdf = Application.Weibull(x, para1, para2, True)
pdf = Application.Weibull(x, para1, para2, False)
EqvN = EqvTransform(x, cdf, pdf, code)

Case "TRIANGULAR":

para4 = 0

a = para1: Mode = para2: c = para3

If x <= a Then x = a + del

If x >= c Then x = c - del
If x < Mode Then cdf = (x - a) ^ 2 / (Mode - a) / (c - a)
If x < Mode Then pdf = 2 * (x - a) / (Mode - a) / (c - a)
If x >= Mode Then cdf = 1 - (c - x) ^ 2 / (c - a) / (c - Mode)
If x >= Mode Then pdf = 2 * (c - x) / (c - a) / (c - Mode)
EqvN = EqvTransform(x, cdf, pdf, code)

Case "BETADIST":

a1 = para1: a2 = para2: min = para3: max = para4

If x <= min Then x = min + del

If x >= max Then x = max - del
cdf = Application.BetaDist(x, a1, a2, min, max)
pdf = Betapdf1(x, a1, a2, min, max)

EqvN = EqvTransform(x, cdf, pdf, code)

193
 Case "PERTDIST":

para4 = 0

a = para1: Mode = para2: c = para3

If x <= a Then x = a + del

If x >= c Then x = c - del

mean = (a + 4 * Mode + c) / 6

If Mode = mean Then

f=6
Else
f = (2 * Mode - a - c) / (Mode - mean)

End If

a1 = (mean - a) * f / (c - a)
a2 = a1 * (c - mean) / (mean - a)

cdf = Application.BetaDist(x, a1, a2, a, c)

pdf = Betapdf1(x, a1, a2, a, c)

EqvN = EqvTransform(x, cdf, pdf, code)

End Select

End Function

Function EqvTransform(x, cdf, pdf, code)

epsi = 10 ^ (-16)

If cdf < epsi Then cdf = epsi

If cdf > 1 - epsi Then cdf = 1 - epsi

EqvSigma = Application.NormDist(Application.NormSInv(cdf), 0, 1, False) / pdf

If code = 1 Then EqvTransform = x - EqvSigma * (Application.NormSInv(cdf))

If code = 2 Then EqvTransform = EqvSigma

End Function

Function betapdf(x, a1, a2, min, max)

With aplication.WorksheetFunction
194
BetaFunc = Exp(.GammaLn(a1) + .GammaLn(a2) - .GammaLn(a1 + a2))

End With

betapdf = 1 / BetaFunc * (x - min) ^ (a1 - 1) * (max - x) ^ (a2 - 1) / (max - min) ^

(a1 + a2 - 1)

End Function

Public Function TEST(x, Y)

TEST = x + Y

End Function

Public Function SQRT(a1)

With aplication.WorksheetFunction

a = .SQRT(a1)

End With

End Function

195