Probabilistic Methods

for Durability Design

The European Union – Brite EuRam III

Probabilistic Methods
for Durability Design

DuraCrete
Probabilistic Performance based Durability Design
of Concrete Structures

Contract BRPR-CT95-0132, Project BE95-1347

Document BE95-1347/R0, January 1999
Although the project consortium does its best to ensure that any information given is accurate,
no liability or responsibility of any kind (including liability for negligence) is accepted in this re-
spect by the project consortium, the authors/editors and those who contributed to the report.

Acknowledgement
The present report is an additional deliverable of Task 1 ‘Design Framework’.
At the start of the project it was felt valuable to organise a short course on probabilistic methods
for durability design as an additional activity. This report contains the lecture notes of the
course.
Michael Havbro Faber (COWI Consulting Engineers and Planners), Ton Siemes and Ton Vrou-
wenvelder (both TNO Building and Construction Research) have written the report.

Information
Information regarding the report:
Ton Siemes, TNO Building and Construction Research, PO Box 49, NL-2600 AA Delft, The
Netherlands. Tel +31 15 2842281, Email A.Siemes@bouw.tno.nl
Information regarding the project in general:
Jan P.G. Mijnsbergen, CUR, PO Box 420, NL-2800 AK Gouda, The Netherlands
Tel +31 182 540620, Email jan.mijnsbergen@cur.nl
The European Union – Brite EuRam III

Probabilistic Methods
for Durability Design

DuraCrete
Probabilistic Performance based Durability Design
of Concrete Structures

Contract BRPR-CT95-0132, Project BE95-1347

Document BE95-1347/R0, January 1999

CUR, Centre for Civil Engineering Research and Codes, NL

Hollandsche Beton Groep N.V., NL
COWI Consulting Engineers and Planners AS, DK
Taywood Engineering Limited, GB
E. Schwenk Zementwerke KG, DE
Geotecnia y Cimientos S.A., ES
ibac, Institute for Building Research, Technical University of Aachen, DE
Institute Éduardo Torroja of Construction Science of the CSIC of Spain, ES
RWS, Directorate-General for Public Works and Water Management, NL
TNO Building and Construction research, Netherlands Organisation for Applied
Scientific Research, NL
Intron, the Quality Assessment Institute for the Building Industry, NL
Chalmers University of Technology, SE
Probabilistic methods for Durability Design 1

Table of Contents
1 Probabilities and random variables 3
1.1 Definition of probability 3
1.2 Sources of uncertainties 6
1.3 Conditional probability and (in)dependence 7
1.4 Important rules in probability theory 10
1.5 Random variable 12
1.6 Reliability calculations 18

2 Limit States 20
2.1 Background of the structural limit state
design process 20
2.2 Limit state format in the present codes 24
2.3 Limit state approach for durability 28
2.4 Durability design in some existing building
codes 31
2.5 Possible bases for durability limit states 33

3 Probabilistic calculations 37
3.1 Introduction to reliability analysis 37
3.2 Exact approach 37
3.3 First Order Reliability Methods (FORM) 42
3.4 Mean value approach 43
3.5 Advanced level II approach 45
3.6 Level I approach 49
3.7 Practical proposal for level I Codes 52

4 Principles of probabilistic modelling 54

4.1 Introduction 54
4.2 General principles of modelling 55
4.3 Formal Bayesian Analysis 56
4.4 Informative and non-informative priors 59
4.5 The selection and updating of a distribution 61

BE95-1347 DuraCrete
2 Probabilistic Methods for Durability design

5 Spatial an temporal fluctuations 62

5.1 Time variability 62
5.2 Spatial variability 66

6 Decision making subject to uncertainty 69

6.1 Philosophical basis for uncertainty
modelling 69
6.2 The decision tree 70
6.3 Decisions based on expected values 71
6.4 Utility assignment 72
6.5 Assignment of probability 72

7 Decision making subject to uncertainty 75

7.1 Decision analysis with given information -
prior analysis 75
7.2 Decision analysis with additional information
- posterior analysis 76
7.3 Decision analysis with ‘unknown’
information - pre-posterior analysis 79
7.4 Postscript 81

8 References 82

9 Table for the standard normal

distribution 84

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 3

1 Probabilities and random variables

1.1 Definition of probability

Introduction
Everybody has some notion what is meant by the term probability. Answers
like ‘probably yes’ or ‘probably no’ are part of the daily language. The word
‘probably’ indicates that we are not completely certain about the present status
or future development of the item under discussion: two or more options are
still open and in the best case we have only some idea that one option is more
likely than the other. In every day speech we may even attach numbers to the
probability, for instance by saying that the probability of something to happen
is about ‘fifty-fifty’.

The purpose of the theory of probability is to enable calculations in this field. In

such a theory words like probability and likelihood have received a more for-
mal meaning. However the real meaning and interpretation of probabilities and
probabilistic calculations as such is not a part of the theory: two people may
have completely different interpretations of the probability concept, but still use
the same calculus.

Mathematical definition
To start the discussion, let us begin at the purely mathematical side. The prob-
ability of occurrence of some event 'A' is denoted as P(A). This probability has
(loosely formulated) the following two properties:

• It is always between zero and unity:

0 < P(A) < 1 (1.1)

where '0' corresponds to the certainty that A will not occur and '1' to the cer-
tainty that A will occur. In intermediate cases A 'might occur' with increasing
probability.

• For two events A and B the following addition rule holds:

P(A or B) = P(A) + P(B) (1.2)

provided that A and B are exclusive, that is 'A and B' is an impossible event.

BE95-1347 DuraCrete
4 Probabilistic Methods for Durability design

These two properties together are sufficient to build up a complete theory. In

this short course we will show some of the further definitions and statements.

Frequentistic definition
The simplest interpretation of a probability in the real world is the frequentistic
one. In this interpretation the probability of an event A is defined as the number
of times that A occurs divided by the number of experiments that is carried out,
provided that the number of experiments is large enough:

N( A )
P(A) = lim for Ntot → ∞ (1.3)
Ntot

N(A) number of experiments where A occurred

Ntot total number of experiments.

For instance, if we have counted 3650 rainy days in the past 100 years, we may
say that for an arbitrary day, somewhere next year, the probability of having
rain is 10 percent.

Classical definition
The classical definition refers to the days that Pascal and Fermat founded the
probability calculus. The inspiration for this theory was found in the games of
cards and dices. The classical definition of the probability of event A can be
formulated as:

n(A)
P(A) = (1.4)
ntot

n(A) number of equally likely ways an experiment may lead to A

ntot total number of equally likely ways in the experiment

Example: the probability of a dice falling on '4' is 1/6 as there are 6 equally
likely results in the experiment of throwing a dice and '4' is just one of them. In
fact there is no real contradiction to the frequentistic definition: the differences
are:

1. that the experiment need not to be carried out as the answer is known in
advance

2. that the classical theory gives no solution if the problem is too difficult to
know all possible equally possible ways.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 5

Bayesian definition
In the Bayesian probabilistic concept probability is formulated as a degree of
belief that A will occur:

P(A) = degree of belief that A will occur (1.5)

The reasons why a person has a certain degree of belief are not stated explicitly.
It may partly follow from statistical data (as in the frequentistic interpretation),
it may be because of symmetry conditions (like in the classical case) or it might
be a matter of judgement based on experiences in comparable situations. This
definition opens the possibility that two persons assign different probabilities to
the same event: the probability is therefor called 'subjective'. It tells something
about the 'state of mind' of the person who assigns the possibility. This is op-
posed to the frequentistic point of view, where probabilities are supposed to tell
something objectively about the 'state of matter'.

The degree of freedom, of course, is not so wide as sketched above. In a formal

Bayesian framework, the subjective element should be formulated before the
relevant data are observed. Arguments of objective symmetrical reasoning and
physical constraints, of course, should be taken into account. Next there is a
formal procedure to deal with statistical data. We will discuss this in Chapter 4.

Coin Example
The following example may help to get a better understanding of the various
concepts:

First consider an ordinary coin. The probability that head comes up is 0.5 in
all cases: Frequentistically because the experiment will show it, Classical be-
cause of symmetry and Bayesian because of a lack of any preference.

Next, consider a coin with two heads or two tails. Assume that it is not known
whether the coin is 'double headed' or 'double tailed'. What is the probability
to get ‘head'?
• In the frequentistic framework the probability to get head is either 0 or 1,
but it is unknown which of the two: we need an experiment to find out.
• The classical definition has no way to find out the 'equally possible ways
to come to such a type of coin'. Is the production of double head and dou-
ble tail a symmetric one?
• Bayesian there is no difficulty: there still is no preference, so P(head) =
0.5.

Closure
The frequentistic interpretation of probability is popular in those fields of appli-
cation where lots of data exists to estimate probabilities and other statistical pa-
rameters. In those circles the Bayesian concept is often considered as 'unscien-
tific' as it contains subjective assessments.

The Bayesian interpretation is popular in those cases where the number of data
is scarce or incomplete. The basic ideas are that lack of knowledge is an uncer-

BE95-1347 DuraCrete
6 Probabilistic Methods for Durability design

tainty that should be treated by probabilistic reasoning in a similar way to other

types of uncertainty. Furthermore, decisions have to be made anyway and sta-
tistical tools are a great help in that process. One may better use it and supple-
ment missing data by judgement. Note that the judgement needs not to be
purely subjective, but that the judgements of various experts in the same pro-
fession can be merged.

1.2 Sources of uncertainties

Probabilities are intended to quantify uncertainties. In structural analysis the
following types of uncertainty are generally distinguished:

• inherent random variability or uncertainty

• uncertainty due to inadequate knowledge (model uncertainty)
• statistical uncertainty.

The inherent random variabilities and uncertainties may be split up into two
categories: uncertainties which can, and uncertainties which cannot, be affected
by human activities. Many kinds of action parameters, for example, wind speed
and earthquake ground motion intensity belong to the second category. The
first category concerns, for example, the uncertainties of strength values and
dimensions of concrete. These uncertainties can be decreased by the use of
more advanced production and quality control methods which, on the other
hand, may cause an increase of production costs.

The uncertainties due to inadequate knowledge can also be subdivided into two
categories. To one of these categories belong, for example, the uncertainties of
physical models or for which knowledge can be increased - and thus uncer-
tainty can be decreased - by research or other similar activities. Also measure-
ment errors belong to this category of uncertainties. To the other category be-
long, for example, uncertainties which depend on future development. One ex-
ample may be the future development of the traffic loads on road bridges. The
possibility to decrease these uncertainties by research or similar activities is
very limited.

• The statistical uncertainties are associated with the statistical evaluation of

results of tests or observations. The may result from:
• a limited number of test results which cause uncertainties in the estimation
of statistical parameters, e.g. mean and standard deviation
• neglecting systematic variations of the observed variables e.g. of climate
variables
• neglecting possible correlations.
• The statistical uncertainties can normally be decreased by increasing test
and observational efforts.

In structural engineering all three types of uncertainty need to be considered.

This also indicates the need for a Bayesian interpretation of probability.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 7

The above descriptions of uncertainties make clear that some uncertainties are
unavoidable (the wind velocity for the next year) while some others might be
removed at the cost of some expense. In other words: uncertainty is in some
cases voluntary. On might even consider the possibility to treat completely de-
terministic variations by the using the theory of probability. Such a procedure is
called 'randomising'.

1.3 Conditional probability and (in)dependence

The events A and B
Consider two events A and B with probabilities P(A) and P(B). We then may
put forward the question: what is the probability that A and B will occur simul-
taneously. To analyse this problem we will adopt the classical definition of
probability. Let an experiment have n equally likely outcomes as indicated in
figure 1.1. A set of n(A) outcomes is associated with event A and a number of
n(B) outcomes are associated with event B. From the classical definition we
have:

P(A) = n(A)/n and P(B) = n(B)/n

In figure 1.1 we have n(A) = 6, n(B) = 9, and n = 20. So P(A) = 0.2 and P(B) =
0.45. In a similar way we may count the number of outcomes that can be asso-
ciated with both A and B. This number is denoted as n(AB). The probability of
(A and B) then is defined as:

P(A and B) = n(AB) / n

Numerically in figure 1.1 we have P(A and B) = 2/20 = 0.10. It is clear n(AB)
must be greater than zero but will always be smaller than n(A) or n(B), which-
ever is the smallest, and so we have:

0 < P(A and B) < min (P(A), P(B))

The actual value of P(A and B) depends on the nature of the problem. If the
events A and B exclude each other the probability P(A and B) is zero. If event
B implies automatically the occurrence of A we have simply P(A and B) =
P(B).

BE95-1347 DuraCrete
8 Probabilistic Methods for Durability design

x x
Bx x x

x x x x x

x x x x x

x x x x x
A
Figure 1.1: Venn diagram for events A and B; every cross represents an
equally likely outcome of the experiment

As an example, consider a box with 4 balls, called Q, R, S, and T. Somebody is

invited to pick up randomly two balls from the box (without returning the first
one before choosing the second). Now, what is the probability to pick up balls
Q and R in that order?
Let event A be : 'the first ball is Q'
Let event B be : 'the second ball is R'

As we have 4 balls the experiment has n = 4 * 3 = 12 equally likely outcomes:

QR QS QT
RQ RS RT
SQ SR ST
TQ TR TS

As only the outcome 'QR' corresponds to (A and B) we may conclude that:

P(A and B) = P(first Q, then R) = 1/12 = 0.083.

Let us now consider the same problem form a different angle. We may find P(A
and B) also from:

P(A and B ) = n(AB)/n = {n(A)/n} * {n(AB)/n(A)} = P(A) * {n(AB)/n(A)}

This looks trivial, but the advantage is that the two factors are relatively easy
to calculate. For instance to calculate P(A) we only have to know that there
are four possible outcomes for the first ball (Q, R , S and T) and that event A
occurs only in one case, so:

P(A) = 1/4
In order to find (n(AB)/n(A)) we only have to look at the n(A) = 3 outcomes at

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 9

the upper line in the above matrix:

QS QT

The other lines simply will not contribute and we may derive directly:

n(AB)/n(A) = 1/3

For P(A and B) we have to calculate the product of 1/4 and 1/3 which is 1/12
as before.

The ratio [n(AB)/n(A)] can be interpreted as the probability that B occurs, as-
suming that A already has occurred. That is: if the first ball is Q, the probabil-
ity that the second ball is R equals 1/3. This is derived simply form the fact that
three balls are left and only one of them is the right one.

Conditional probability
From the above it can be inferred that we may introduce the so called condi-
tional probability:

P(A|B ) = n(AB)/n(A) = {n(AB)/n}/(n(A)/n} = P(A and B )/P(A) (1.6)

A technical example of conditional probability is e.g. the probability of failure,

given that corrosion of the reinforcement has been observed:

P(failure | corrosion of the reinforcement)

It is clear that in most cases the observation of corrosion may lead to a substan-
tial increase in the failure probability.

Independence
Two events A and B are called 'independent' if the occurrence of B does not
affect the probability of A:

P(A|B) = P(A) (1.7)

So 'failure' and 'corrosion' cannot be called independent.

Consider next a non-technical example. Assume that we pick up randomly two

cards out of a package of 52 Bridge playing cards, not putting the first one
back. The probability to find the second card an ace, given the first card to be
not being an ace is:

P(card2 = ace | card1 = no ace) = 4/51

BE95-1347 DuraCrete
10 Probabilistic Methods for Durability design

However, if we return the first card to the package before picking the second,
we have:
P(card2 = ace | card1 = no ace) = 4/52

As the probability to have an ace is 4/52 we see that the two events are de-
pendent in the first case and independent in the second case.

In the case of independence we have further:

P(A and B) = P(A|B) P(B) = P(A)P(B) (1.8)

So in the case of independence we may find the probability of (A and B) sim-

ply by multiplying the individual probabilities P(A) and P(B). Whether or not
two events are dependent may follow from calculations, from observations or
from judgement.

1.4 Important rules in probability theory

The event 'A or B'
In this section we will introduce some important rules in probability theory:
The first one is the expression for the calculation of the event (A or B):

P(A or B) = P(A) + P(B) - P(A and B) (1.9)

The proof is relatively easy if we consider once more figure 1.1:

n(A or B) = n(A) + n(B) - n(AB)

In numbers: n(A or B) = 6 + 9 - 2 = 13. Then the subtraction of n(AB) is neces-

sary to avoid the 'double counting' of the points in (A and B). Dividing by n and
applying the classical definition of probability completes the proof.

Some special cases (see figure 1.2):

Case 1: A and B are independent:

P(A and B) = P(A|B) P(B) = P(A) P(B)

P(A or B) = P(A) + P(B) - P(A) P(B)

Example: the probability to throw at least one 6 with two dices is given by:
1/6 + 1/6 -1/36 = 11/36

Case 2: A and B exclude each other

P(A and B) = P(A|B) P(B) = 0 * P(B) = 0

P(A or B) = P(A) + P(B)

Example: the probability to throw 1 or 2 with a dice is equal to 1/6 + 1/6 =

1/3.
Case 3: B implies A, that is A always occurs if B occurs:

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 11

P(A and B) = P(A|B)P(B) = 1.0 * P(B) = P(B)

P(A or B) = P(A) + P(B) - P(B) = P(A)

Example: the probability to throw '3' is 1/6 and the probability to throw 'odd'
is 1/2; as throwing '3' implies throwing 'odd' the probability to have 'odd and
3' is simply 1/6.

Case 4: In the case of three events A, B and C:

P(A or B or C) =
P(A) + P(B) + P(C) -P(A and B) - P(A and C) - P(B and C) +
+ P(A and B and C)

The proof can be performed in a similar way as in the case of two events. Note
that in the case of mutual exclusive events we simply have:

P(A or B or C or..) = P(A) + P(B) + P(C) + ..

Independence A and B exclusive B implies A.

A B B
B
A
A
(in)dependent exclusive B implies A

Figure 1.2: Events A and B, special cases

Total probability theorem

The second important theorem is the so-called total probability theorem. In the
simplest form the presentation is as follows:

P(A) = P(A|B) P(B) + P(A|Bnot) P(Bnot) (1.10)

To proof the statement we may again refer to figure 1.1:

n(A) = n(AB) + n(ABnot) = {n(AB)/n(B)} n(B) + {n(AB)/n(Bnot)} n(Bnot)

The 'numbers' in this case are: n(A) = 6, n(B) = 9, n(Bnot) = 11, n(AB) = 4,
n(ABnot) = 4.
A technical example is the probability of failure of a beam in an aggressive
environment:

BE95-1347 DuraCrete
12 Probabilistic Methods for Durability design

P(failure) = P(failure given corrosion) P(corrosion) +

P(failure given no corrosion) P(no corrosion)

The above statement can easily be extended to a set of events Bi:

P(A) = Σ P(A|Bi)P(Bi) (1.11)

For the proof it is necessary that the events Bi and Bj are mutually exclusive
and that the sum of P(Bi) = 1.0.

Bayes theorem
Bayes theorem is a simple theorem. It starts with the definition of the condi-
tional probability (1.6):

P(A|B) = P(A and B) / P(A)

In a similar way we might define:

P(B|A) = P(A and B) / P(B)

So, combining these two expressions:

P(A|B) = P(B|A) P(A) / P(B) (1.12)

This theorem will be used later in Chapter 4.

1.5 Random variable

Distribution function
The output of many experiments can be expressed as a numerical value, for
instance, the compressive strength of concrete is 50 MPa or the carbonation
depth is 20 mm. If such a numerical value cannot be predicted precisely, we
call the variable a random variable. For random variables we can define the
event that the variable X is smaller than some value x and assign a probability
to that event:

P( X<x )

The probability P( X<x ) is called the probability distribution function (see fig-
ure 1.3):

FX(x) = P( X<x ) (1.13)

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 13

As FX(x) is defined as probability it is always between 0 and 1. In fact, the

function always starts at 0 for x = -∞ and increases monotonically to 1.0 for x
= +∞.

Probability density function

If we differentiate Fx(x) with respect to x we obtain the probability density
function (see figure 1.3):

fX(x) = dFX(x) / dx (1.14)

This function gives the probability that X is in the interval from x to x+dx:

P( x < X < x+dx ) = fX(x) dx (1.15)

FX
1

0
µ X
fX

µ X

Figure 1.3: Distribution function FX and probability density function fX of X

The most commonly used probability models in reliability calculations are:

• the uniform distribution
• the beta distribution
• the normal distribution
• the lognormal distribution
• the extreme value distributions

The beta and uniform distributions are defined for a finite interval of non-zero
probability.

BE95-1347 DuraCrete
14 Probabilistic Methods for Durability design

The normal distribution is used if the stochastic variable is the sum of great
number of individual contributions (Central limit theorem). A lognormal dis-
tribution follows if log(X) is normally distributed. As a result a lognormally
distributed variable is always non-negative, which makes it appropriate for
non-negative variables.

The extreme distributions form a complete family for cases when a variable is
the maximum or minimum of a great number of variables. Typical examples
are the Gumbel, the exponential, the Rayleigh, the Weibull, and the Frechet
distribution.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 15

Distribution type Para- Moments

meters
Rectangular a+b
m=
a≤x≤b A 2
1 b b−a
f X ( x) = s=
b−a 12
Normal
1  1  x − µ 2 µ m=µ
f X ( x) = exp −    σ>0 s=σ
σ 2π  2 σ  
Lognormal  ζ2 
m = exp λ + 
x>0 λ  2
1  1  ln x − λ  2  ζ>0  ζ2 
f X ( x) = exp −    s = exp λ +  exp(ζ 2 ) − 1
Xζ 2 π  2 ζ    2
Shifted Lognormal  ζ2 
m = ε + exp λ + 
x>ε λ  2
1  1  ln( x − ε ) − λ  2  ζ>0  ζ2 
f X ( x) = exp −    ε s = exp λ +  exp(ζ 2 ) − 1
( x − ε )ζ 2 π  2 ζ    2
Shifted Exponential 1
m=ε+
x≥ε ε λ
f X ( x) = λ exp( − λ ( x − ε )) λ>0 1
s=
λ
Gamma p
m=
x≥0 p>0 b
b p
b>0 p
f X ( x) = exp( − bx) x p −1 s=
Γ ( p) b
Beta A r
m = a + ( b − a)
a ≤ x ≤ b, r , t ≥ 1 b r+t
( x − a ) r − 1 ( b − x ) t −1 r>1 b−a rt
f X ( x) = s=
( b − a ) r + t −1 B( r , t ) t>1 r+t r + t +1
Gumbel (L) 0.577216
m= u+
−∞ < x < +∞ U α
f X ( x) = α exp( −α ( x − u) − exp( −α ( x − u))) α>0 π
s=
α 6
Frechet (L)  1
m = ε + ( u − ε )Γ  1 − 
ε ≤ x < +∞, u, k > 0 u>0  k
k  x − ε
− k −1
  x − ε −k
 k>0
f X ( x) = exp −   2 2 1
   ε s = ( u − ε) Γ1 −  − Γ 1 − 
u − ε  u − ε   u − ε   k  k
Weibull (S)  1
m = ε + ( u − ε )Γ  1 + 
ε ≤ x < +∞, u, k > 0 u>0  k
k  x − ε
− k −1
  x − ε  −k  k>0
exp −  2 2 1
f X ( x) =     ε s = ( u − ε) Γ1 +  − Γ 1 + 
u − ε  u − ε   u − ε   k   k

Table 1.1: Probability Distributions (from ‘Sicherheit und Zuverlässigkeit

im Bauwesen by Jörg Schneider’)

BE95-1347 DuraCrete
16 Probabilistic Methods for Durability design

Table 1.1 gives an overview of the mathematical expressions, the shapes and
the expressions for mean and standard deviation, defined as:
∞

µ( X ) = ∫
−∞
x f x ( x ) dx (1.16)

σ (X)= ∫ ( x − µ x ) 2 f x ( x )dx
2
(1.17)
−∞
The mean is the quick characterisation for the order of magnitude of a quantity.
The standard deviation is the quick characterisation for the scatter. As an alter-
native one often uses the coefficient of variation:
V = σ/µ (1.18)
Of course, this measure is valid only if µ is a positive quantity.

A special case of a normal distribution is the standard normal distribution.

This is a normal distribution with a mean value µ = 0 and a standard deviation
σ = 1. For this distribution tables are available for derivating the probability
and the density. In Chapter 9 such a table is presented.

Two random variables X and Y

In most reliability problems there may be several random variables. Let us first
consider the case of two random variables. Formally these two random vari-
ables have a two dimensional distribution function and a two-dimensional den-
sity function:
FXY (xy ) = P {X < x and Y < y} (1.19)
∂ ∂ (1.20)
f XY ( xy ) = FXY ( xy )
∂x ∂y
An example of a two dimensional density function is presented in figure 1.4. In
most cases such a function is represented by drawing the lines of equal height
(i.e. probability density).

f XY

Figure 1.4: Two dimensional probability density function

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 17

The distribution of the two variables can also be characterised by their respec-
tive means and standard deviations. An important additional parameter in the
case of two variables, however, is the covariance:

covariancexy = ρxy σxσy (1.21)

where ρxy = coefficient of correlation, given by:

 x − µ  y − µ y 
 x   f ( xy )dxdy
ρ
xy
= ∫∫  σ  σ  xy
(1.22)
 x  y 

Without proof it is stated that the following statements hold:

• if X,Y are independent: ρxy = 0
• if X,Y are fully linear dependent: ρxy = + 1 or ρxy = -1
• otherwise: 0 < | ρxy | < 1.
• if (X,Y) are independent and if Z = X + Y we have: σz2 = σx2 + σy2(1.23)

Figure 1.5 gives some possible interpretations for ρ.

Y Y Y
*
* * *
* *
* *
*
* * * * *
*
* * *
* ρ = 1.0 ρ = 0.7 ρ = 0.0
* * *
*
* * X
X X
Y Y
* **
* **
*
*
* * *
*
*
ρ = - 0.7 ρ =*- 1.0
*
*
X X

Figure 1.5: Correlation between X and Y

BE95-1347 DuraCrete
18 Probabilistic Methods for Durability design

1.6 Reliability calculations

Consider a function Z of a number of variables Xi:

Z = g(X1, X2, X3, ....., Xn)

Assume that the density functions fxi (xi) for all random variables Xi are known.
In that case it is in principle possible to calculate the density function fZ(z). The
following methods may be used:
• Monte Carlo
• analytical integration
• numerical integration
• numerical approximation (FORM)

As an example consider Z = X + Y

Monte Carlo
The procedure is as follows:
1. sample u from a uniform distribution [0,1]
2. sample v from a uniform distribution [0,1]
3. calculate X via F(X) = u
4. calculate Y via F(Y) = v
5. calculate Z = X + Y
6. repeat this many times, for example 1 000 000 times
7. make a histogram of Z

Analytical integration
Apply the total probability theorem (1.11) with A = 'Z < z' and Bi = xi < X <
xi+1. Using xi+1 = xi + dx we may write:

P(Z<z) = ∫ P(Z<z | x<X<x+dx) P(x<X<x+dx)

At the left-hand side we recognise the distribution function of Z and at the right
hand side the density function of X:

FZ(z) = ∫ P(Z<z | x<X<x+dx) fx(x) dx

Now, as Z = X+Y we have:

P(Z<z | x<X<x+dx) = P(X+Y<z | x<X<x+dx)

Let us now 'use' the condition 'x<X<x+dx'. This condition means that X is very
close to x, and so we may replace the X by x in the conditional event X+Y<z :

P(Z<z | x<X<x+dx) = P(x+Y<z)

And so we have finally:

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 19

FZ(z) = ∫ P(Z<z | x<X<x+dx) fx(x) dx = ∫ P(x+Y<z) fX(x) dx =

∫ FY(z-x) fX(x) dx
In some cases it is possible to find analytical solutions for such an integral.
This will not further be explored. More general is the use of this integral as a
basis for numerical integration. In Chapter 3 this approach will be explained
more in detail.

Numerical integration
Based on the previous result:

FZ = ∑ Fy(z-x) fx(x) ∆x

The practical problem is to find the correct values of upper and lower integra-
tion bounds and the value for ∆x. Numerical integration is feasible for a limited
number of random variables (say maximum 8).

FORM
FORM stands for First Order Reliability Method. The basic steps in FORM
are:
1. linearise Z
2. assume Z to be normal
3. calculate µ(Z) and σ(Z) using analytical expressions

In this example problem the first step can be omitted, as Z = X+Y is already
linear. Further we have:

µ(Z) = µ(X) + µ(Y)

σ2(Z) = σ2(X) + σ2(Y).

In non-linear problems and problems with non-normal variables the method is

an approximation. In Chapter 3 this method will be elaborated more in detail.

BE95-1347 DuraCrete
20 Probabilistic Methods for Durability design

2 Limit States

2.1 Background of the structural limit state design

process
The process of designing, construction and use of buildings is complex. In the
following this is simplified overview of the successive steps in the process are
presented.

REQUIRED PERFORMANCES

DEFENCE STRATEGY

MAINTENANCE STRATEGY

DESIGN

CONSTRUCTION

MAINTENANCE PLAN

USE AND MAINTENANCE

Required performances

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 21

The first step in the building process is of course the initiative. After that step a
definition should be given with respect to the functions and the quality level of
these functions. This must be translated into required performances. Perform-
ance means that the fulfilment of the function must be expressed in quantifiable
terms.

Example
One of the functions of a floor in a parking garage is to carry the loads of the
cars. In terms of performances this can be expressed as: the floor must have a
safe load bearing capacity capable to carry a uniform load over the whole
floor of 4 kN/m2. For a design engineer this means in general that he has to
take care for bending and shear capacity.

A structure that exactly fulfils the performance requirements has reached its
limit state. Exceeding this limit state means respectively that the structure is
functioning or disfunctioning.

A basic definition of a limit state is: A limit state is the border that separates
desired states, that is situations acceptable to the owner of the structure, in
which a structure may be during its lifetime from the undesired or adverse
states. Acceptable means in this case that the desired states should satisfy the
primary functional demands.

The owner does not directly set some of the primary demands, but he is legally
forced to commit himself to them. In most countries the authorities set the con-
straints on safety, for example in the building codes. But many primary de-
mands may be set by the owner and usually the architect or the structural de-
signer translates them to functional demands. Besides safety one can think
about demands with respect to serviceability and aesthetics.

The basis of a limit state consists of the fact that crossing into the unacceptable
region, directly measurable consequences follow: the structure does no longer
meets its primary demands. In principle there will be costs involved to improve
(repair) the structure to get it back in the desired state. Crossing into the unac-
ceptable region can be a gradual phenomenon (like staining and soiling) or a
sudden consequence (like collapse).

It is also possible to define secondary limit states. If it follows for example

from designs based on primary demands that crossing from the initiation phase
of a degradation into the propagation phase will lead to excessive life cycle
costs, than this crossing may be the bases for a secondary limit state. In princi-
ple it must however be proved that the secondary limit state can logically be
derived from a primary limit state.

Using this definition of a limit state provides a sound basis for optimal design,
inspection and maintenance strategies. By economic optimisation repair actions
can be formulated in time. By defining for example failure of a concrete beam
as the limit state, economic optimisation may lead to the conclusion that the
best moment to stop carbonation is at the point when the reinforcement starts to
corrode. The corrosion of the reinforcement is not a limit state itself.

BE95-1347 DuraCrete
22 Probabilistic Methods for Durability design

Defence strategy
A building structure can provide its performances in many different ways. The
design engineer must make therefor many decisions with respect to the choice
of the material, the material composition, the material quality, the structural
system and the defence actions. Some examples of these numerous decisions
are shown in table 2.1.

Maintenance strategy
The defence decisions are directly related to the maintenance that will be
needed by the structure. The design engineer must take decisions with respect
to maintenance free (in practice this means maintenance with a low probability)
or planned maintenance. Planned maintenance can be subdivided in periodic
maintenance, used based maintenance or maintenance on complaints.

Design
In the design phase the prove is given that with the chosen defence and mainte-
nance strategy the required performances can be realised by the structure. The
result of the design is a detailed construction concept.

Construction
The actual realisation of the structure takes place during the construction phase.
The starting points of the design must be realised during construction. In case
of deviations from the original design plan a redesign must be made. An impor-
tant aspect in the construction phase is the workmanship.

Maintenance plan
After the structure has been built a maintenance plan must be made. This plan
is derived from the maintenance strategy, the design and the actual construc-
tion.

Use and maintenance

During use and operation of the structure the planned maintenance must be exe-
cuted. In case of deviation from the starting points of the design repair or redes-
ign will be necessary. The latter is also needed if the use of the structure is
changed.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 23

Table 2.1: Examples of decisions to be taken by the design engineer

Decision example
choice of the building material concrete
steel
masonry
timber
aluminium
composition of the material plain concrete
special type of concrete (rumble aggregate,
high strength concrete)
fibre reinforced concrete
reinforced concrete
prestressed concrete with unbonded tendons
prestressed concrete with bonded tendons
material quality p.m.
structural system parallel or series system
static dependent or independent
framework or single system
type of defence avoiding mechanism
single or multi-defence system
robustness
redundancy
second bearing system

BE95-1347 DuraCrete
24 Probabilistic Methods for Durability design

2.2 Limit state format in the present codes

If a certain performance B of a structure depends on only one parameter X1 the
performance requirement is simple. This one parameter may not exceed a de-
fined limit L.

B{X1} > L or B{X1} < L (2.1)

Where:
B{ } function that describes the performance
X1 basic parameter for the performance function
L limit for the performance.
Depending on the type of performance the limit may not be exceeded from the
underside or the upper side. These options are expressed in relationship 2.1.

For example: the width w of a door may not exceed the clear width = 0.8 m of
the doorframe.

The performance function will in general be described with more than one, say
n, parameters. The performance requirement is than more complicated:

B{X1, X2, .., Xi, .., Xn} > L or B{X1, X2, .., Xi, .., Xn} < L (2.2)

where
B{ } performance function
Xi i-th basic parameter
n total number of basic parameters, n >1
L limit

Example
A simply supported beam with a length l is carrying a uniform load q over the
total length of the beam. The requirement for this beam is that it must be able
to carry this load. If the bending capacity of the beam is equal to Mu =50 kNm
and the shear capacity is equal to Tu = 10 kN. Then the performance require-
ment is:

B{ } = Mu - 1/8 ql2 = 50 - 1/8 ql2 > 0

B{ } = Tu - 1/2 ql = 10 - 1/2 ql > 0

In Figure 2.1 two diagrams have been given for all values of q and l that fulfil
the performance requirement. In the diagram two areas can be distinguished.
An area with combinations of q and l that lead to a safe structure (B > 0) and
an other area with combinations that are not safe (B < 0). All combinations on
the curve between the desired and the adversed area form the limit state
(B = 0).

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 25

q q
8 16

6 12

limit state
limit state
4 8
collapse
collapse

2 4
equilibrium
equilibrium

0
l 0
l
5 10 15 20 5 10 15

shear bending

Figure 2.1: Two examples of limit states

In all modern building codes the structural performance requirements are ex-
pressed in terms of limit states. Two main types of limit states can be distin-
guished:
• ultimate limit state (ULS)
• serviceability limit state (SLS).

The ultimate limit states refer to collapse, fracture, overturning, lifting or slid-
ing and other events where the safety of the structure is of direct importance. In
principle this refers to the loss of static equilibrium. The serviceability limit
states refer to comfort for the user, functionality (fitness for purpose), aesthetics
and durability. In table 2.2 some examples of serviceability problems are given.
In these table the problems related to user comfort are the conventional service-
ability problems. They are treated in structural building codes. Some of the
other problems are also dealt of in other building codes. It was not usual to call
them serviceability problems. Nowadays the term serviceability is used in the
broader sense, as given in table 2.2.

Many building codes relate durability to serviceability. This is especially true

for concrete codes. Crack width limitations and water/cement ratio limits are in
the present codes defined as serviceability limit states. In table 2.2. some exam-
ples from existing codes have been presented. It is however becoming more and
more common opinion that durability is an aspect of ultimate and serviceability
limit states where degradation’s can play a role. A special durability limit is
therefor not necessary.
Table 2.2: Examples of problems related to serviceability

BE95-1347 DuraCrete
26 Probabilistic Methods for Durability design

Aspect Example
User comfort static deflection
Vibrations
limited local structural damage
unacceptable damage in non-structural parts
settlements of foundations
Functionality unevenness of floors
drainage of roofs
jamming of doors and windows
out of plumbness
functioning of services
Aesthetics out of plumbness
large deformations
cracks and fissures
gaps between components
gravel pockets
soiling and staining
Graffiti
Durability crack width in concrete
minimum crack width in concrete
maximum water/cement ratio in concrete
fully prestressing concrete
galvanising of steel
application of an organic coating
Prescription of type of material (for example
hardwood or stainless steel)

We define this as durable ultimate and durable serviceability limit states. The
limit states have therefor to be extended to the time domain. Within each limit
state that is considered attention must therefor be paid to possible durability
problems. As a consequence the enumeration of durability aspects in table 2.2
is not correct according to the modern opinion on limit states.
For structural engineers it is more or less usual to write performance functions
like (2.1) and (2.2) in two parts, representing respectively the strength (bearing

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 27

capacity) and the effect of the loading (action):

R(X1, X2, ..., Xi, ..., Xj) - S(Xj+1, Xj+2, ..., Xn ) > 0 (2.3)

Where:
R the function (model) that describes the bearing capacity or strength of the
structure
S the function (model) that describes the influence of the load on the
structure
Xi basic variable.
In general this function is called the limit state function.

The strength model, the load model and most of the basic variables are not de-
terministic but stochastic. Therefor it is inevitable to express relationship (2.3)
in terms of probability. In the design the probability of failure (i.e. exceeding an
ultimate or a serviceability limit state) must be restricted [1]. The design for-
mula becomes therefor:

P{R(X1, X2, ..., Xi) - S(Xi+1, Xi+2, ..., Xn ) < 0} < Pacc = Φ(-β) (2.4)

or more simplified:

P{R - S < 0} < Pacc = Φ(-β) (2.5)

Where:
P{..} probability that the failure event {..} will occur
Pacc accepted value of the probability of failure.
Φ distribution function of the standard normal distribution
β parameter for the standard normal distribution.

The parameter β is called the reliability index. The accepted failure probability
is usually related to the standard normal distribution Φ through the parameter β.
In Chapter 9 a table for the standard normal distribution is given, where β is
related to the failure probability.

Relationship (2.5) may be approximated for 1 < β < 4 with:

P{R - S < 0} < Pacc = Φ(-β) ≈ 10-β

The failure probability of relationship (2.4) res. (2.5) can be calculated with the
aid of probabilistic techniques. As these calculations can be complicated and
time consuming the actual design formula is, for the practical design, rewritten
as:

Rc /γR - γS Sc > 0 (2.6)

BE95-1347 DuraCrete
28 Probabilistic Methods for Durability design

Where:
Rc characteristic value of the load bearing capacity, this is a value with a
distinct probability of exceedence, for example 5 %
Sc characteristic value of the load effect, this is also a value with a distinct
probability of exceedence, for example 1 %
γR material factor
γS load factor.

The values of Rc, γR , γS, and Sc have been defined in the building codes in such
a way that it may be presumed that the failure probability will not be exceeded.
In a number of codes these values have actually been calibrated to the failure
probability.

Relationship (2.6) can be simplified:

Rd - Sd >0 (2.7)

Where:
Rd design value of the load bearing capacity
Sd design value of the load effect.

2.3 Limit state approach for durability

In the performance based structural design, relationships (2.3) to (2.7) both the
resistance R and the load S are considered to be time independent. In many
loading situations this is not realistic. The limit state function should then be
rewritten as a time dependent limit state function [2]:

R(t) - S(t) > 0 (2.8)

Where:
R(t) time dependent function (model) that describes the bearing capacity or
strength of the structure
S(t) time dependent function (model) that describes the load effect on the
structure.

A special case for this limit state function occurs if either R or S is not time de-
pendent. Relationship (2.8) applies for all values of t in the time interval (0,T).
T is the reference period (intended service life or design working life). Even if
the loading or the load bearing capacity of a structure is time dependent, the
limit state functions for designing structures are rarely formulated in this way
(with an exception for fatigue). They are often simplified to time independent
quantities. Well-known simplifications are:
• assuming that the material strength during the service life period is either
equal to the short term strength or to the long term strength
• for (semi) static loads one characteristic maximum value, related to the ref-
erence period, is defined

• for fluctuating loads (like wind, traffic or waves) one characteristic maxi-

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 29

mum fluctuation in the reference period is defined.

Relationship 2.8 can also be used as the bases for durable ultimate and service-
ability limit state design.

The durability design can be presented in two different, but theoretically

equivalent, formats for the limit state functions. These are the 'intended service
period design' and the 'lifetime design'. In the intended service period design
the condition is that the limit state may, with a certain reliability, not be reached
within the intended service period. The format for the intended service life de-
sign is largely comparable with the format for the conventional structural de-
sign. In the lifetime design the reliability of the structure is related to the prob-
ability that the design working life will be exceeded. The working life ends at
the moment that the limit state is exceeded.

The concept of the intended service period design can be expressed in a design
formula:

Pf,T = P{R(t) - S(t) < 0}T < Pacc = Φ(-β) (2.9)

Where:
Pf,T the probability of failure of the structure within the period T
T intended service period.
Pacc the accepted maximum value of the probability of failure
Φ standard normal distribution function
β reliability index

Probably it will be possible to simplify this relationship in a later stage to a de-

sign format equivalent to the conventional design formulas.

Example :
A steel rod with a diameter of ø = 20 mm and a tensile strength fa = 500 N/mm2
is loaded with an axial tensile force F =105 kN. Due to corrosion the diameter
reduces with a speed s = 0.1 mm/year. In this case the design formula for the
period t = 0 to T years is:

Pf,T = P{R(t) - S(t) > O }T = P{¼ p (ø - s t)2 fa - F > 0}T

= P{ [¼ p (20 - 0.1 t)2 - 150.103 > 0] }T < Pacc = Φ(-β)

In some structural codes indications have been given for values of T and Pacc.
For example in Eurocode 1 for the ultimate limit state (collapse): T = 50 years
and β = 3.8 (corresponding with Pacc = 7.10-5 ≈ 10-4).

For the lifetime design relationship (2.8) must be transformed to a lifetime func-
tion. This can be done by writing it as an explicit function of the time:

L = t{R,S} (2.10)

BE95-1347 DuraCrete
30 Probabilistic Methods for Durability design

where:
L the lifetime of the structure.

The reliability of the structure can be introduced by limiting the probability of

exceeding the accepted value:

Pf = P{L < T} < Pacc = Φ (-β) (2.11)

Where:
Pf failure probability.

A possible, more practical form for a design formula is:

td > ttarget (2.12)

Where:
td design value of the lifetime
ttarget target working life.

Example:
For the same steel rod that was presented in example 1 the life time function is:

Pf = P{ L > T } = P{ø - √{[ F / ¼ π fa ] / s} < T}

= P{ [20 - √{[105000 / ¼ π 500 ] / 0.1 } < Ptarget = Φ (-β)

In Figure 2.2 the durability design procedure is illustrated. It shows the simi-
larities and differences between the service period design and the lifetime de-
sign. The illustration makes clear that for both approaches the same information
is used. Consequently they will lead to exactly the same result. Moreover the
lifetime distribution in Figure 7.2 shows that the margin between the mean ser-
vice life m(L) and the target value T can be very large. This margin depends on
the type of the distribution, the scatter and the target failure probability (As the
target failure probability is very small; it is almost impossible to draw the figure
in the right scale!).

The performance based durability design provides the basis for an objective
assessment. It is based on realistic and quantified environmental and material
models capable to predict the future behaviour of the structure. In this way it
describes the performance in relation to time. This offers the following bene-
fits:
• durability design is based on the same principles as structural design (safety,
serviceability, limit states and reliability)
• objective designs based on the total life cycle costs will be possible
• a reduction of the consumption of materials and energy by optimal use of the
materials

• basis for design engineers, contractors and maintenance engineers to make

tailor made designs, especially if the design is based on specific material
properties

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 31

• realistic performance test procedures to establish the building material prop-

erties and the building component behaviour
• methodology can be extended from concrete to other existing structural ma-
terials, new materials and new applications; nowadays much attention is
paid to a sustainable development and as a consequence waste materials or
secondary materials will be used more and more.
Figure 2.2: Similarity between the service period design and the lifetime design

service period design

R,S R(t)
distribution of R(t)

S(t)
distribution of S(t)

µ(L) time
service mean lifetime
Ptarget
period
lifetime distribution

lifetime design

2.4 Durability design in some existing building codes

Eurocode
In Eurocode 1: 'Basis of design and actions on structures' (ENV 1991-1, 1994)
an attempt has been made to give fundamental requirements with respect to the
durability of structures. It is stated that structures should be designed with ap-
propriate levels of reliability for safety and serviceability, including durability.
For this purpose a 'design working life' is defined, being the period for which
the structure has to be used for its intended purpose with anticipated mainte-
nance but without major repairs being necessary. For the design working life a
classification is given varying from 1 to 5 years for temporary structures to 100
years for monumental buildings, bridges, and other civil engineering structures.

It is assumed in design that the durability of a structure or part of it in its envi-

ronment is such that it remains fit for use during the design working life. The
structure should be designed so that the deterioration does not impair the dura-

BE95-1347 DuraCrete
32 Probabilistic Methods for Durability design

bility and the performances of the structure. The environmental conditions have
to be appraised at the design stage to assess their significance in relation to du-
rability and to enable adequate provisions to be made for protection. The degree
of deterioration may be estimated from calculations, experimental investiga-
tions, experience from earlier structures, or a combination of these considera-
tions.

Although it is important that in Eurocode 1 attention has been paid to durability

the practical significance is still restricted. The main reason for that is the lack
of a strict design format and of objective requirements. In some material related
Eurocodes more elaborated requirements are given. These are however implicit
and not related to performances or to the fundamental durability requirements
in Eurocode 1.

All necessary ingredients for a performance based durability design procedure

are nevertheless present in Eurocode 1. These ingredients are the performance
based structural design method, levels of reliability, and reference periods.

CEB-FIP Model Code - Design Code 1990

The CEB-FIP Model Code is based on limit state design with partial safety fac-
tors and characteristic values. The design refers to reliability levels. Levels of
reliability are however not defined. The reference period is in principle 50
years. For practical reasons this period is equal to the intended working life.

The design for durability is based on the principle of slowing down the process
of deterioration of the parts of the structure that are critical in that respect. This
normally implies a multistage strategy, which may often be based on successive
protective barriers against corrosion. This concept leads to an integral treatment
of the planning and design phase, the execution phase and the period of use.
The elaboration of this is done on basis of environmental classes. The concrete
structure must be designed, constructed and operated in such a way, that it
maintains its safety, serviceability and appearance during its intended lifetime.
In the Model Code and extensive chapter on durability is present. The back-
ground for the durability design is worked out. The requirements are however
presented in terms of deem-to-satisfy rules.

Dutch Building Decree

The Dutch Building Decree and the structural building codes derived from it,
are based on the principles from the reliability theory. The ultimate and ser-
viceability limit states are defined in terms of:
• performances
• reliability index
• reference period

In table 2.3 an overview is given of the reliability indexes according to the

Dutch Building Decree. These indexes must be combined to reference periods
of either 15 years or 50 years. The period of 15 years refers to buildings where
a short life span is usual like factories and caravans. In special situations longer
reference periods must be used. This applies for example for important infra-

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 33

structural structures.

Table 2.3: Overview of reliability indexes β in the Dutch Building Codes

Ultimate limit state serviceability
Safety class wind decisive other loads limit state
1 2,3 3,2 1,8
2 2,4 3,4 1,8
3 2,6 3,6 1,8

A practical elaboration of the durability limit states has however not been given
in the codes. For concrete structures the durability is arranged by means of
deem-to-satisfy rules. In practice the Dutch Building Decree is directed to limit
state design for mechanical loadings.

In some other building codes an attempt has been made to define performance
based durability design. Despite this situation the practical elaboration is still
missing.

2.5 Possible bases for durability limit states

It is important to realise that both the intended service period design and the
lifetime design are not necessarily restricted to the conventional ultimate and
serviceability limit states. The processes (or mechanisms) involved can be of
mechanical, physical, chemical, elector-chemical or biological nature. The limit
states can refer to aspects such as structural safety, serviceability, functionality,
comfort, aesthetics and so on. Already in Table 2.1 an overview of possible
limit state aspects has been presented. The accessory target reliability indexes
and reference periods depend on the type and the amount of damage.

In the limit state approach, as presented so far, the first exceedence of the limit
state determines failure. Other types of criteria are however possible. Examples
are: the number of exceedences of a limit state is restricted. Or: the duration of
the exceedence of a limit state is restricted. The latter two examples may apply
in reversible limit states. The two examples may be true for vibrations that
cause discomfort.

Some worked out examples of both concepts were presented some years ago [3,
4, and 5]. These examples made clear that the concepts of the intended service
period design and the lifetime design can be used for various building materi-
als, showing the generic character of these concepts. Rilem committee 130 CSL
'Calculation of the Service Life of Concrete' has prepared a report [6] that
shows calculation examples for this approach for various degradation’s of con-
crete structures.

In defining limit states and the target reliability indexes related to durability the
following possibilities can be distinguished:

BE95-1347 DuraCrete
34 Probabilistic Methods for Durability design

• lack of durability leading to an ultimate limit state

• lack of durability leading to a conventional serviceability limit state
• lack of durability leading to other serviceability limit states
• durability limit states intending to reduce the amount of repair and annoy-
ance.
The target reliability is strongly related to the costs due to failure of the struc-
ture. If people can get killed or injured social considerations have also much
influence on the level.

In [7] a basis has been presented for introducing the gradual loss of function in
a limit state design. The specification of a target probability for durability re-
lated limit states is not an easy task. In fact, the fitness-for-purpose of a struc-
ture is not explicitly defined. A degradation slightly above a predefined limit
does often not lead to dramatic effects. The same holds if the exceedence of a
certain level has a longer duration then some specified admissible time period.

Cost of damage Cost of damage

state parameter state parameter

a b

Figure 2.3 Step and gradual loss functions

In terms of decision theory it is said that the 'loss function' for the decision is a
step function or a gradual one (see Figure 2.3). The loss function describes the
costs of damage as a function of a state parameter as for instance stress, deflec-
tion, and corrosion. A further distinction that has to be made is that a limit state
can be irreversible (like fracture) or reversible (like vibrations, deformation due
to shrinkage).

The conclusion is that, at least for reversible serviceability problems, it does not
make much sense to talk about a limit state, a limit state function or a service-
ability limit state target reliability, although some attempts in such directions
have been made. Mathematically such a procedure might lead to an operational
procedure; but the basis for such a procedure may be questioned. An alternative
and fundamentally better approach is to start with the gradual loss function
from the beginning.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 35

The basis equation for an ultimate limit state or a reversible serviceability limit
state is to optimise the life cycle cost function. In a simple form and neglecting
interest this function is:

C(tot) = C(b) + C(f) * P{R< S}T (2.13)

Where:
C(b) costs of building the structure
C(f) costs in case of failure
P{R<S}T probability of failure within the intended working life T.

Optimising (2.13) leads (at least formally) to an optimum for the failure prob-
ability P{S>R}. Based on such a target safety standard procedures lead to de-
sign values for load and resistance that can be used in every day design.

In the case of a gradual loss function problem, the life cycle cost function is:

C(tot) = C(b) + ∫ C(υ) fu(υ) dυ (2.14)

Where:
C(υ) loss related to the state parameter u = n
fu(υ) probability density function of the state parameter u

Using partial integration this can be rewritten as:

C(tot) = C(b) + ∫ C’(υ) * P{u > υ} dυ (2.15)

Where C’(υ) denotes the first derivative of C(υ) with respect to υ. The advan-
tage of (2.15) is that we may use standard reliability programs to calculate the
probabilities.

Next consider the case of a time dependent reversible limit state. One simple
model for the gradual loss function is to assume that the damage costs are di-
rectly related to measure of exceedence of the limit. Once this function is
known, the designer may choose an optimal design for the building owner by
minimising building costs and maximising rent income.

Neglecting interest effects, the object function for this optimisation problem
can be formulated as:
C(tot) = C(b) + ∫ c(a) dt (2.16)

Where:
C(b) costs of building a structures with known properties
c(a)dt the damage costs during the period dt
a the time fluctuating effective value of the state parameter

Solving this minimisation problem gives the optimal structure. However, func-

BE95-1347 DuraCrete
36 Probabilistic Methods for Durability design

tions like C(b) and c(a) are difficult to establish. It would be an interesting re-
search project to define them, but for the time being it seems more promising to
start with some kind of calibration procedure.

This means that functions for C(..) and c(..) are postulated. On the basis of these
postulates the optimum is calculated and compared with present day practice.
These comparisons may be used to improve the postulations. he elaboration of
this technique will be given in Chapter 4.

Another point of reflection is that it should be possible to derive on the basis of

the equation (2.14) design values for the everyday design practice, in a similar
way as is done for conventional structural designs.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 37

3 Probabilistic calculations

3.1 Introduction to reliability analysis

Nowadays it is commonly accepted that the safety of structures should be ex-
pressed in terms of reliability. This means as the probability of failure. In litera-
ture [8, 9, 10, and 11] the bases have been given for the calculation of the fail-
ure probability. Making probabilistic calculations can be done on different lev-
els, ranging from theoretically correct to practically simplified. The Joint
Committee on Structural Safety (CEB, ECCS, CIB, FIP, RILEM, CECM and
IABSE) has introduced three different levels for the reliability analysis of struc-
tures [9]:
• exact approach
• first order reliability approach (FORM)
• partial factor method (PFM).

Exact approach, the calculations are fully based on the theory of probability.
All parameters that have influence on the load bearing capacity of a structure
are taken in account as stochastic quantities. Between these parameters a
mathematical relationship must be formulated, describing the limit state under
consideration. After that the probability of failure can be calculated. Such a cal-
culation is not practical, especially if many parameters are considered.

First order reliability methods (FORM) are probabilistic approaches with spe-
cific simplifications to make the calculations manageable. Nevertheless these
calculations can be complicated and time-consuming. For this reason they are
not be applied for ordinary calculations.

Partial factor method (PFM) is a semi-probabilistic approach with characteris-

tic values for the loadings and the strengths in combination with partial safety
factors. This method is usually based on the results of FORM calculations. This
PFM is very appropriate for modern building regulations and it is more and
more accepted. In older building regulations an approach with fixed values for
the loadings was used.

3.2 Exact approach

The completely probabilistic approach will be illustrated for a simple case in
which the limit state can be described with only two parameters (see also Chap-

BE95-1347 DuraCrete
38 Probabilistic Methods for Durability design

ter 1). These are the effect of the loading S and the capacity of the structure R
(or in short the load S and the capacity or strength R). If the loading is equal to
the capacity: R = S the limit state is reached and failure will occur. For the as-
sessment of the reliability a limit state function Z is introduced:

Z=R-S (3.1)

With this function it is possible to distinguish the relevant situations for the
limit state:
Z > 0 the structure is in the desired state
Z = 0 the structure has reached the limit state
Z < 0 the structure is in a adverse state.

The structure will fail if Z < 0 or R < S. The probability P that the load S is
equal to s1 can be calculated from the probability density function fS:

P{s1 < S < s1+ds} = fS(s1)ds (3.2)

The probability that the capacity R is smaller than s1 follows from:

s1

P{R < s1} = ∫

0
fR(r)dr (3.3)

The combination of these two probabilities is consequently:

s1 s1
P{s1 < S < s1+ds and R < s1 } = ∫ ∫
0 0
fR(r)dr . fS(s1)ds (3.4)

The probability P{Z < 0) = P{R < S} follows from integrating fS(s) over all val-
ues s:
s1 ∞ ∞
P{R < S} = ∫ ∫
0 0
fR(r)dr . fS(s)ds = ∫
0
FR(r) fS(s)ds (3.5)

Calculating this probability can be very complicated. Especially if Z is a non-

linear function or a function with more than two parameters. If Z is however a
linear function of R and S the solution of relationship (3.5) is rather simple. The
mean value mm(Z) can be calculated as:

µ (Z) = µ (R) -µ (S) (3.6)

and the standard deviation σ(Z) as:

σ(Z) =√ [σ2(R) + σ2(S)] (3.7)

If R and S have a normal (Gaussian) distribution, Z will also have a normal dis-
tribution. The probability of failure PF can be calculated by using the standard
normal distribution Φ:

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 39

µ ( Z) µ ( R ) − µ (S )
Pf = P{R < S} = Φ { - } = Φ {- } = Φ [-β] (3.8)
σ ( Z) [σ 2 ( R ) + σ 2 (S )]

The value Φ[-β] can be found in tables for the standard normal distribution in
handbooks of statistics. In table 3.1 the relationship between -β and the prob-
ability of failure Pf given. In Chapter 9 a more elaborated table is given.

Table 3.1: Relationship between β and the probability of failure Pf

-β 1.3 2.3 3.1 3.7 4.2 4.7 5.2 5.6
Pf 10-1 10-2 10-3 10-4 10-5 10-6 10-7 10-8

The reliability of a structure can be expressed as the failure probability Φ[-β],

or directly as β. In that case β is called the reliability index. Note that from
(3.8) follows:
µ ( R ) − µ (S )
β= (3.9)
[σ 2 ( R ) + σ 2 (S )]
This special case is important for the FORM approach, because all the simplifi-
cations on that level aim to express Z as a linear function of the loading and the
capacity parameters. And to transform the distributions of these parameters to
normal distributions.

S fR,S(r,s) = constant

µ(S)
Z<0

Z=0

Z>0
R
µ(R)
Figure 3.1: Two-dimensional probability density function fR,S and the
limit state Z

In figure 3.1 a graph is given that can explain the probabilistic calculation. The
figure gives the two-dimensional probability density function
fR,S(r,s) = fR(r).fS(s). The straight line Z = R -S = 0 gives the limit state. The
volume of the probability density function that is cut off by this line is equal to
the probability of failure according to relationship 3.5

BE95-1347 DuraCrete
40 Probabilistic Methods for Durability design

Figure 3.1 can be simplified by:

• transforming the R- and S-axes to normed axes, according to R` = R/σ(R)
and S` = S/σ(S)
• shifting the origin to the point [µ(R)/σ(R), µ(S)/σ(S)], this is the transfor-
med mean of the probability density function fR,S(r,s).
The new co-ordinates after both transformations (figure 3.2) are:
R − µ ( R)
R* =
σ( R)
S − µ (S)
S* =
σ(S)
S*

R*

Z=0

Figure 3.2 Two-dimensional probability density function with transformed axes

The limit state Z = 0 has now the relationship:

Z = µ(R) - µ(S) - R* {σ(R) - S*} σ(S) = 0 (3.10)

In figure 3.2 it can be seen, that the probability of failure is equal to the volume
of the body that is cut off by the limit state from the probability density func-
tion. It is however more important to see, that the distance from the new origin
to the limit state line is equal to β.

In figure 3.3 a new parameter r = µ(R)/µ(S), the ratio between the mean values
of the strength R and the loading S is introduced. With r relationship (3.5) can
be written as:

r−1
β = (3.11)
[ r 2V 2 ( R ) + V 2 ( S )]

In figure 3.3 this relationship is represented for various combinations of V(R)

and V(S). It is obvious that the reliability index β is dependent on the ratio r
and the variation coefficients.

β
BE95-1347 DuraCrete
Probabilistic methods for Durability Design 41

5.0

4.0

V(R) = 0.1
3.0 V(R) = 0.2

V(R) = 0.3
2.0

1.0
V(S) = 0.3

0 1.0 1.5 2.0 2.5 3.0 3.5 4.0 r

Fig. 3.3: Relationship between β and r for various values of V(R) and V(S)

Example
if a safety factor γ = 1.7 is applied between characteristic values (5 % normal
fractiles) for the strength and the loading, than this means:

µ ( R ) − 164
. σ(R)
γ =
µ ( S + 164
. σ(S )

For V(R) = V(S) = 0.1 this implies:

0 .836 µ ( R )
γ = 1.7 =
. µ( S )
1164
or:

r = 2.37

From relationship 3.11 it follows that β = 5.3. Table 3.1 indicates that the
probability of failure is a little less then 10-7. For higher values of the variation
coefficients the ratio r has to be bigger to give the same reliability.

The examples given till now are all based on limit state functions with two pa-
rameters. For realistic structures this will in general not be enough. Sometimes
it can be necessary to introduce more parameters like the geometry, material
properties, loads, loading combinations, shape factors etc. In the general case

BE95-1347 DuraCrete
42 Probabilistic Methods for Durability design

the limit state should be described as a function with n stochastic parameters:

Z = G(X) = G(X1, X2,.. , ......., Xn) (3.12)

By analogy with relationship (3.5) the failure probability is given by:

Pf = P{Z < 0} = ∫∫ . . . . ∫ fX1(x1)dx1 . fX2(x2)dx2......fXn(xn)dxn

= ∫∫ . . . . ∫ fX(xi) dx1dx2....dxn (3.13)

This integration should be done for the area where Z < 0. If the number of pa-
rameters n is high, this calculation will be very time consuming. This applies
especially if the reliability function is non-linear. An other complication can be,
that many probability density functions of the individual parameters are not
known. In that case it is not possible to calculate fX(x).

Because of these types of problems the exact approach is not used for most of
the practical problems. This has given rise to the development of the FORM
and the PFM approaches with justified simplifications.

3.3 First Order Reliability Methods (FORM)

In the level II approach the n-dimensional integral (3.13) is calculated after two
approximations:
• the reliability function Z = g (X) is linearised;
• the distributions of the basic parameters (X) are replaced by normal distribu-
tions.

Both approximations are characteristic for the level II approach. They are
called the 'First-Order-Second-Moment Approximations'. The term 'First-Order'
relates to the linearisation. The scatter of the basic parameters is calculated on
basis of the second moments, i.e. the variance.

After linearisation g (X) has the following form:

Z ≈ g (X) = a0 + a1X1 + a2X2 + ........... + anXn

The mean value µ(Z) and the standard deviation σ(Z) of this linear function
follow from:

µ(Z) ≈ g[µ(X)] = a1µ(X1) + a2µ(X2) + ...... + anµ(Xn) (3.14)

σ2 (Z) ≈ σ2 [g(X)] = [a1σ(X1)]2 + [a2σ(X2)]2 + .... + [anσ(X2)]2 (3.15)

In the special case that all the basic parameters Xi have a normal distribution,
the linearised reliability function g will also have a normal distribution. The
limit state function Z will approximately have a normal distribution.
Based on the level II approximations several approaches are possible. In the

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 43

next part the 'mean value approach' and the 'advanced approach' will be ex-
plained. The mean value method can be done by hand. The calculated reliability
can differ sometimes substantially from intrinsic one. For more precise results
the improved 'advanced approach' is necessary. For simple reliability functions
the calculations can be done by hand. In general these have to be done with a
computer.

3.4 Mean value approach

The first step in both the mean value and the advanced approach is the li-
nearisation of the reliability function. In the point X* the non-linear function is
developed in a Taylor progression, that is cut of after the linear terms:

Z ≈ g (X) = g(X*1, X*2, ....X*n) + ∂g/∂x1[(X* )] (X 1- X*1) +

∂g/∂x2[(X*)] (X 2 - X*2) + ..... + ∂g/∂xn[(X*)] (Xn - X*n) =
n
*
= g(X ) + ∑
i= 1
(Xi-X*i) [∂g /∂xi (3.16)

The linearisation point X* in the mean value approach is taken as the mean
value of all basic parameters µ(X) = (µ(X1), µ(X2), ...., µ(Xn)). The distribu-
tions of the basic parameters are replaced by normal distributions with the
original mean values and standard deviations. This results in an approximated
reliability function with the following properties:
• Z has a normal distribution
• µ(Z) = g[µ(Xi), µ(X2),....,µ(Xn)]
n
∂g *
• σ (Z) =
2
∑ ∂x
i =1 i
σ2(Xi)

The reliability index β = µ(Z)/σ(Z) can be calculated just like the failure prob-
ability Pf = Φ(-β).

Example
The probability of failure of a cantilevered steel pipe loaded with a force F has
to be calculated. It is assumed that failure corresponds with exceeding the
bending capacity:

BE95-1347 DuraCrete
44 Probabilistic Methods for Durability design

t
l
D
D

The load is S = F l
The capacity is R = Wp Sp = t D2 Sp
The reliability function is Z = R - S = t D2 Sp - F l

In table 3.4 the properties of the basic parameters for this example are ex-
plained and summarised.

Table 3.4: Basic variables in the calculation example

Xi Description distribution µ(Xi) σ(Xi)

t Wall thickness deterministic 0.03 m -
D diameter normal 1.00 m 0.05 m
Sp yield stress normal 280 N/mm2 19.6 N/mm2
F load deterministic 0.6 MN -
l pipe length deterministic 10.0 m -

The mean value of Z follows from:

µ(Z) = g{µ(X)} = t µ2(D)µ(Sp) - Fl =

0.03*1.002*280.106 - 6.106*10.0 = 2.4 106 N/m2

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 45

The standard deviation follows from:

σ2(Z) = [∂Z/∂D.σ(D)]2 + [∂Z/∂Sp.σ(Sp)]2 =

[ 2t µ(D) µ(Sp) s(D) ]2 + [ t µ2(D)σ(Sp) ]2 = 1.03 106 N/m2

As a consequence β = µ /σ = 2.33 and the failure probability is 0.011.

The reliability function Z in this example could also have been written as:

Z = Sp - Fl / tD2

This results however in a reliability index β = 2.86, or a probability of failure

PF = 0.002. The result of the mean value is obviously influenced by the form of
the reliability function. This is a known disadvantage of the mean value ap-
proach. Without further proof it will be clear that replacing distributions that
deviate strongly from normal ones also give errors.

3.5 Advanced level II approach

In the advanced approach two modifications are made:
• The linearisation is taken place in a special point, the design-point X*, and
no longer in the mean point µ( Xi ). The design-point is the contact point of a
circle with the mean point as centre and the linearised failure boundary g(X)
as tangent. Because of the iterative steps the calculations can in general not
be done by hand. This is a disadvantage. As the design-point lies on the fail-
ure boundary, in the point with the maximum probability density of Z
(Venezeano [10] has shown this effect), it is comprehensible that the result
of the calculation is better then with the mean value approach.
• The replacement of the non-normal distributions is improved. These will be
replaced by normal distributions, according to:

FXi(X*i) = Φ (X*i)
fXi(X*i) = φ (X*i)

Where:
Φ normal distribution function
φ probability density function for a normal distribution

Or in other words the replacing normal distribution Φ and probability density

function φ of the basic parameter Xi have for X*i the same values as the original
distribution and probability density function (Figure 3.3). This improvement is
called the 'Approximate Full Distribution Approach' (AFDA). Rackwitz [11]
has given more information about its accuracy, that is close to level III.

BE95-1347 DuraCrete
46 Probabilistic Methods for Durability design

FX

representative
original

fX
representative
original

X* X
i

Figure 3.3: Approximate Full Distribution Approach (AFDA)

Relationship (3.12) applies also for the advanced approach. As in the design-
point g(X*) = 0, this gives:

n
Z = ∑
i =1
(Xi - X*i) [∂g/∂x]*

So:
n
µ(Z) = ∑ i=1
[µ(X i) - X*i] [∂g/∂xi]*

n
σ2(Z) = ∑ [∂g/∂xi]*2 σ2(Xi)
i =1

If:
*2
 ∂g  2
  σ ( Xi )
αi =  ∂x i  (3.17)
n
∂g *2 2
∑ [
i = 1 ∂x i
] σ ( Xi )

n
∂g 2
σ(Z) = ∑
i= 1
αi [
∂xi
] σ (Xi) (3.18)

From µ(Z) and σ(Z) the reliability index β can be found:

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 47

n
∂g ∗
µ( Z )
∑ [ µ( X
i=1
i ) − X i, ][
∂x i
]
β= = (3.19)
σ( Z ) n
∂g ∗
∑
i=1
αi[
∂x i
] σ ( Xi )

After some arrangements this can also be formulated as:

n
∂g *
∑
i= 1
[
∂x i
] [ µ(Xi) - X*i - αi β σ(Xi)] = 0 (3.20)

Or:

X*i = µ(Xi) - αI β σ(Xi) for i = 1 → n (3.21)

In figure 3.4 the flow chart for the iteration process is given. First a value is
chosen for b and the first approximation for the design-point is taken as the
mean point, like in the mean value approach. Then two internal loops are made.
The first one to calculate a better value for the design point. The second loop
tries to find a design-point on the failure boundary by changing the value of β.

Example
Recalculation of the example of the steel pipe in paragraph 3.1 with the ad-
vanced approach and AFDA gives as result:

µ(Z) = 2.18 106 Nm

σ(Z) = 0.82 106 Nm
β = 2.66
Pf = 0.004
D contributes with 67 % to the variance of Z
Sp contributes with 33 % to the variance of Z.

In this case the basic parameter D with the relatively low scatter has the big-
gest contribution to the variance of Z.

The results of the mean value approach based on reliability functions for bend-
ing moments or for stresses differ from each other. The advanced approach in
combination with AFDA gives also another outcome. The three approaches
resulted in:
• Pf = 0.011 for the mean value approach based on bending moments
• Pf = 0.002 for the mean value approach based on stresses
• Pf = 0.004 for the advanced approach with AFDA.

BE95-1347 DuraCrete
48 Probabilistic Methods for Durability design

START

choose β

x* = µ(X)
β = β + ∆β

replace the distributions of X

(only with AFDA)

calculate αi

new value for Xi*

X *i = µ(Xi) - αi β σ(Xi)

no
Is X *i stable?

yes

no Z(X*) = 0?

yes

P{failure} = P{Z<0} = Φ(-β)

output:

µ(Z), σ(Z), αi, β, P{failure}

STOP

Figure 3.4: Flow chart of the advanced approach with AFDA

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 49

3.6 Level I approach

Even the level II approach is still too complicated for standard design calcula-
tions. For that reason most of the designs are nowadays based on a semi-
probabilistic level I approach. It is evident that the simplifications on which the
level I approach is based will tend to bigger errors in the calculation of the reli-
ability. One of the starting points for this approach is therefor to minimise the
difference between the aimed and attained reliability.

The semi-probabilistic approach is based on characteristic values Xi,k and par-

tial safety factors gXi for the most relevant basic parameters. Usually there are
safety factors for the load and for the capacity. To account for this the reliabil-
ity function Z = g(X) = 0 is split in two partial functions R = gR(X) and S =
gS(X). Each of them containing only parameters related respectively to loads
and to strength:

Z = R - S ≈ gR(X) - gS(X) = gR(X1, X2, ..., Xj) - gS(X j+1, Xj+2, ..., Xn) = 0

This means that the partial safety factors are:

γR = Rc/R* or R* = Rc/γR
γS = S* /Sc or S * = γ S Sc

With relationship (3.5):

R* = µ(R) - αR β σ(R) = Rc/γR

S* = µ(S) - αS β σ(S) = γS Sc

Or with k as a parameter indicating the fractile for the characteristic value (usu-
ally k = 1.64 for the 5 % fractile of normal parameters):

µ ( R ) − kσ ( R ) 1 − kV ( R )
γR = = (3.19)
µ ( R ) + αRβσ ( R ) 1 + αRβV ( R )

µ ( S ) − αSβσ ( S ) µ ( S ) − αSβV ( S )
γS = = (3.20)
µ ( S ) + kσ ( S ) 1 + kV ( S )

With variation coefficients V(S) = σ(S)/µ(S) and V(R) =σ(R)/µ(R) applies ac-
cording to (3.2):

σ( R)
αR = (3.21)
[ σ ( R ) + σ 2 ( S )]
2

σ( S )
αS = (3.22)
[ σ ( R ) + σ 2 ( S )]
2

This means that:

BE95-1347 DuraCrete
50 Probabilistic Methods for Durability design

αR2 + αS2 = 1

Example
For a number of combinations of V(R) and V(S) the partial safety factors γR and
γS are given in table 3.3. They have been calculated on basis of relationships
(3.19) and (3.20) for:

R and S having normal distributions

µ(S) = 1.0
β = 4.0
V(R) = 0.1
K = 1.64

Table 3.3: Example of calculated partial safety factors

V(S) µ(R) αR αS γS γR
0.05 1.71 0.96 - 0.28 0.98 1.36
0.10 1.83 0.88 - 0.48 1.02 1.29
0.15 2.00 0.80 - 0.60 1.09 1.23
0.20 2.18 0.74 - 0.68 1.16 1.19
0.25 2.38 0.69 - 0.72 1.22 1.15

This table indicates, just like the relationships (3.19) and (3.20) that the partial
safety factors are dependent respectively on the factor αR or αS and conse-
quently on both σ(R) and σ(S). This means that the load factor γS and the mate-
rial factor γR are always mutual dependent. An other property is the fact that the
partial safety factors have for each limit state other values.

A simple method to overcome the first problem is to establish constant values

for αR and αS. Because of relationships (3.21) and (3.22) this restricts also the
values of σ(R) and σ(S). Introducing constant factors αR and αS means however
that the reliability function is no longer linearised in the design point. The fac-
tors αR and αS must be chosen in such a way that the error in the aimed reliabil-
ity index β is as small as possible and on the safe side.
In figure 3.5 the ideal situation is reproduced. The limit state in the design point
can be described as:

[µ(R) - αR β σ(R)] - [µ(S) + αS β σ(S)] = 0

If for instance the linearisation factors are taken as αR = 1.0 and αS = 1.0 the
limit state function becomes:
[µ(R) -β σ(R)] - [µ(S) +β σ(S)] = 0
This can be seen in figure 3.6. In this case the minimum and the maximum val-
ues of the reliability index differ a factor √2. The minimum value βmin = β, be-

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 51

cause it is on the circle. The maximum value βmax differs by a factor √2 from
the aimed reliability index β. This difference is caused by the fact that the lin-
earisation point lies too far away from the circle.

R - µ(R)
σ(R) Z > 0

Z =0

αS
S - µ(S)
σ(S)
β

αR
design point

Z < 0

Figure 3.5: Linearisation in the design point

R - µ(R)
σ(R)

αS = 1.0 S - µ(S)
βmin σ(S)
αR=1.0
βmax

Figure 3.7: Linearisation in the point (αR = 1.0, αS = 1.0)

A better choice is for instance the linearisation point (αR = 0.8 ; αS = 0.7). In
figure 3.8 it can be seen that this point lies much closer to the circle. From the
slope of both tangent lines it can be calculated that this situation applies for

BE95-1347 DuraCrete
52 Probabilistic Methods for Durability design

0.40 < σ(S)/σ(R) < 1.78. Most of the ratios between σ(S) /σ(R) will fall in this
region.

R - µ(R)
σ(R)

arctan max

αS = 0.7 S - µ(S)
σ(S)
^
β
^
αS = 0.8 β

arctan min

Figure 3.8: Linearisation in the point (αR = 0.8; αS = 0.7) (see 5.3)

If R and S are functions of more than one basic parameter the relationships
(3.21) and (3.22) chance to:

µ ( Si ) − αS βσ ( Si ) 1 − αS βV ( Si )
i
γSi =
i
= (3.23)
µ ( S i ) + kσ ( S i ) 1 + kV ( Si )

µ ( Ri ) − kσ ( Ri ) 1 − kV ( Ri )
γRi= = (3.24)
µ ( Ri ) + αRβσ ( Ri ) 1 + αRβV ( Ri )

The values of αSi and αRi can be calculated with relationships (3.21 and 3.22).
But analogue to the two-dimensional case it will be possible to establish con-
stant values for them.

3.7 Practical proposal for level I Codes

International situation
As an indication is in table 3.4 an overview given of the reliability indices β
and the linearisation factors αR and αS in a number of countries.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 53

Table 3.5: Overview of β, αR and αS for the design codes in various countries

Country SLS ULS αR αS

β β β
1 year 1 year 50 year
Netherlands 1.8 2.6 3.6 .. ..
Switzerland [13] 4.0 0.9* 0.8**
Germany [14] .. 3.0 4.7 0.8 0.7
Scandinavia [15] .. 2.0 4.2 √(a2+6.25/β2) 2.5/β
a=1+2[β-2.5/(100V(R)]
USA [16] - 3.0
* for two parameters αR = 0.7
** for the remaining parameters αS = 0.5

The value of the reliability index β depends of the time interval for which the
analysis is conducted. In the literature reference periods of in general 1 year
and 50 years are mentioned. The reliability index β can be converted to another
reference period with:

βT2 = Φ-1[ Φ(βT1) ]T2/T1

If or instance β1 year = 4.0 this gives:

β50 = Φ-1[ Φ(4.0) ]50/1 = 3.0

This example shows that the reliability index in the USA has the same magni-
tude as the European values.

Simple application
Normal distribution:

1 + αSβV(S
γS =
1 + 164
. V (S )

1 − 164
. V( R )
γR =
1 − αRβV( R)

Lognormal distribution:

γS = exp[ (αS β - 1.64) V(S) ]

γR = exp[ (αR β - 1.64) V(R)]

BE95-1347 DuraCrete
54 Probabilistic Methods for Durability design

4 Principles of probabilistic modelling

4.1 Introduction
In the majority of cases a random variable can be characterised by:
• a mean value
• a standard deviation (or coefficient of variation)
• a distribution type.

In Chapter 5 we shall see that we may have some additional parameters ex-
pressing the spatial and temporal variability. For these parameters, of course,
similar methods as discussed here hold.

Parameters (mean, standard deviation) as well as distribution type will depend

on the specific population under consideration . As an example, see figure 4.1,
the C50 concrete production from one particular production unit might be char-
acterised by a mean strength of 65 N/mm2 and a variation coefficient of 10 %.
For the production in the European Community the numbers might be 70
N/mm2 and 20 % (these figures indicate that the intended quality has not been
achieved in a number of cases).

One should always keep in mind the definition of the population under consid-
eration, both in the case of modelling on the basis of judgement, as well as in
the case of estimating statistical properties from data. The basis question al-
ways is: what does the data represent? The existence of various levels of popu-
lations is often referred to as a 'hierarchy of populations'. In other words: within
a population one might distinguish a number of sub-populations.

Especially in probabilistic calculations, having many parameters, one should be

careful to use consistent models for all random variables.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 55

one production unit

many productionunits

2
50 N/mm
f'
c

Figure 4.1: Distributions of various populations C50 concrete

4.2 General principles of modelling

To establish distribution type and parameters one has to rely on a combination
of:
• data
• theoretical background
• engineering intuition.
One may recognise the frequentistic, the classical and the Bayesian approach.
We shall discuss them one by one.

Data
If one has a lot of statistical data, there is a minimum need for the other meth-
ods of modelling. One may simply use the data directly or find a probability
density model that fits best with the data. In the latter case one may find the
parameters of a distribution by methods like:
• graphical curve fitting: by eye or using least square methods
• the method of moments: calculate the mean and standard deviation of the
distribution from the mean and standard deviation of the data
• maximum likelihood: find those parameters that make the data most likely.

We will not discuss these methods in detail. Note that in the case of limited data
these methods also provide confidence limits.

Theory
Theoretical arguments may follow from physics and or probabilistic arguments
itself. Examples are:
• a variable is always positive; this indicates a lognormal distribution
• a variable is bounded on two sides: this indicates a β distribution
• a variable is the sum of great number of other variables; this indicates a
normal distribution

BE95-1347 DuraCrete
56 Probabilistic Methods for Durability design

• a variable is the maximum or minimum of a great number of distributions;

this indicates an extreme value distribution
• a variable may or may not be independent with others
• a variable is related to other variables through models
• probabilities can be estimated on the basis of symmetry (as in the case of a
die)

Judgement
Judgement is necessary if the number of theoretical arguments is not sufficient
and directly applicable data is limited. The judgement is normally based on the
opinion of experts who try to translate experiences in similar problems to the
problem at hand. If possible, this judgement should not be the judgement of one
person, but of a group of persons.

Of course, the judgement should never be in conflict with the available theo-
retical arguments. If we have the combination of judgement and data we need
to use the formal Bayesian analysis.

4.3 Formal Bayesian Analysis

The basic principle of a formal Bayesian approach has been presented in figure
4.2. The idea is that we start to formulate a so-called 'prior-distribution' on the
basis of a combination of theoretical arguments and judgement. Next the prior
is combined with the data. This process is called 'updating' and results in the
'posterior distribution'.

Theory Judgement

Prior Data

Updating

Posterior

Figure 4.2: Schematic presentation of a formal Bayesian procedure

In order to update the prior use is made of Bayes Theorem (1.12):

P(B| A) P(A)
P(A| B) = (4.1)
P(B)

In this application of Bayes Theorem A represents some hypothesis (e.g. with respect

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 57

to the unknown mean value µx of a distribution) and B represents the data relevant
for the hypothesis (e.g. the mean x of a sample). If we substitute this in Bayes Theo-
rem we get:

P(hypothesis | data) = C P(data | hypothesis) P(hypothesis) (4.2)

For example:
P{µ(x) = m | x } = C P{ x | µ(x) = m ) P( µ(x) = m }

where:
P{µ(x) = m} prior distribution for µx
P{µ(x) = m | x } posterior distribution for µx
P{ x | µ(x) = m} likelihood
C constant.

In words this means that the probability that the mean value µx equals m , given
some sample average x , is equal to the product of a the prior probability, the
likelihood of the x if µx = m and a constant.

Numerical example
Let x be the carbonation depth for some structural element. Assume that x is
normal with unknown mean and a standard deviation of 3 mm. As a prior it is
guessed that µx is 10 or 15 mm, both with 50 % probability. Let there be two
measurements with an average value of 11 mm. We then have for the posterior
probabilities:

P(µx = 10 | x =11) = C P( x =11 | µx=10) P(µx=10)

P(µx = 15 | x =11 ) = C P( x =11 | µx =15) P(µx=15)

The units mm have been omitted for clarity. As x is continuous, the probability of x
being exactly 11 is simply zero and so we should actually write:

P(µx = 10 | x =11) = C P(11< x <11+dx | µx=10) P(µx=10)

P(µx = 15 | x =11 ) = C P(11< x <11+dx | µx=15) P(µx=15)

We will elaborate the first line. The prior probability P(µx = 10) = 0.5. To evaluate
the likelihood one should note that if µx = 10 mm the sum of two observations (x1 +
x2) has a mean equal to 20 mm and a standard deviation equal to √(9+9) = 4.24; so
the mean and standard deviation of the sample average x = (x1 + x2)/2 are 10 and
2.12 respectively. Given the normal distribution, the likelihood is given by (see Table
1.1):

− 1 1 11 − 10 2
P (11 < x < 11 + dx|µ x = 10) = exp{− ( ) }dx
2.12 2π 2 2.12
And so we find for the posterior probability that µx = 10:

BE95-1347 DuraCrete
58 Probabilistic Methods for Durability design

− 1 1 11 − 10 2
P( µ x = 10| x = 11 ) = C exp{ − ( ) }dx * 0 .5 = 0 .17 Cdx
2.12 2π 2 2.12

Similar we may have:

−
P( µ x = 15| x = 11 ) = 0 .03 Cdx

We now may find C from the requirement that the sum of these two probabilities
should be equal to one:

C = 5/dx

So:
−
P(µx = 10| x =11) = 0.85
−
P(µx = 14| x =11) = 0.15

So the probability of µx = 10 has gone up from 0.50 to 0.85.

General formulas
If the parameter µx is not a discrete variable, but a continuous variable, or even
a vector of parameters θ, we may rewrite (4.1) as:

fθ ( θ | D ) = C f X ( x ,θ ) fθ ( θ ) (4.3)

Where:
fθ(θ|D) posterior distribution for the parameters θ
fX(x|θ) probability density function for X, given θ (likelihood function)
fθ(θ) prior distribution for the parameters θ
C normalising constant

Graphically the Bayesian analysis is presented in figure 4.3. Once the distribution
for the parameter θ has been established, two theoretically equivalent ways of intro-
duction into the reliability analysis can be followed. One way is to consider θ as an
additional set of random variables. This is straightforward but increases the number
of random variables. The other way is to incorporate the statistical uncertainty into
the distribution of the basic variables X:

fx(x|D) = ∫ fx(x;θ) fθ(θ|D) dθ (4.4)

Where:
fx(x|D) predictive distribution

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 59

P(µ) P(µ)

PRIO R

µ µ

P(x|µ) P(x|µ)

LIK ELY H O O D

µ µ
P(u|x) P(u|x)

PO STERIO R

µ µ

Figure 4.3: Graphical presentation of a Bayesian updating procedure

4.4 Informative and non-informative priors

Non-informative or vague priors
As the choice of a prior may very have great influence on the final result, very often
a so-called 'vague' or non-informative prior is used. Graphically this means that the
priors in figure 4.3 are constant, flat and almost equal to zero. A very common ex-
ample is the following situation:

BE95-1347 DuraCrete
60 Probabilistic Methods for Durability design

Let x be normal with unknown mean and unknown standard deviation. Assume that
we have a set of n observations. It can be proven that on the basis of non-informative
priors the posterior fractile for x can be found from:

xp = m ± kn,p s (4.5)

Where:
m sample mean
s sample standard deviation
k a coefficient following from:

n=3 n=6 n = 10 n = 30 n =∞
p = 0.1 2.18 1.62 1.44 1.33 1.28
p = 0.05 3.27 2.21 1.89 1.72 1.64
p = 0.01 8.00 3.69 2.90 2.50 2.08

For instance: the 5 percent fractile of a concrete cube strength fc (the character-
istic value), where a mean value of 30 MPa and a standard deviation of 4 MPa
have been measured out of 6 cubes is equal to:

f c = m - 2.21 s = 30 - 2.21*4 = 21.2 MPa.

Note that if we have the same mean and standard deviation, but resulting from
more than 100 observations we have 30 - 1.64*4 = 23.4 MPa. The difference
between 23.4 and 21.2 is caused by the so-called statistical uncertainty in the
mean and the standard deviation.

Note: in standard frequentistic analysis this uncertainty is dealt with via the so-
called confidence intervals. In Bayesian analysis this confidence level is al-
ready incorporated in the outcome.

Informative priors
In some cases one might want to use an informative prior, based on general
data (statistical as well as non-statistical) and circumstantial evidence. A useful
strategy in those cases is often to specify an interval { Θ iL , Θ iU } for which one
reasonably can assume that there is a 95 % probability that the unknown quan-
tity will be inside. Such values can be established as the aggregated opinion of
a number of experts Assuming a normal distribution, the mean and standard
deviation of Θ i can then be estimated from:

µ( Θ i )={ Θ iL + Θ iU }/2 (4.6)

σ( Θ i )={ Θ iL + Θ iU }/4 (4.7)

Similar expressions exist if a lognormal distribution for the parameter is as-

sumed.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 61

4.5 The selection and updating of a distribution

In most cases theoretical considerations will not be sufficient to establish the
distribution type. Also in those cases Bayesian updating procedures may help to
find the distribution. A practical solution is to model X as some polynomial
function of for instance the standard normal variable U (mean 0, standard de-
viation 1.0):

X = a0 + a1u + a2u2+ .... + anun (4.8)

The constants a in the polynomial can be found in the same way as the parame-
ters θ of a distribution (Section 2.2).

Another approach, which sometimes might be required, is to use mixed distri-

bution:

fX(x) = P(C1) fX |C1 (x) + P(C2) fX |C2(x) + ... (4.9)

In here fX|Ci(x) is the probability density given some condition Ci and P(Ci) is
the probability that this condition occurs.

Whether or not theoretical arguments and supporting data are available, the
choice of a distribution is a difficult task. The results of reliability analyses
based on low values for the probability’s are very sensitive to the tail of the
probability distribution, and observations to give information about the tail
characteristics are in most cases not available.

BE95-1347 DuraCrete
62 Probabilistic Methods for Durability design

5 Spatial and temporal fluctuations

5.1 Time variability

Introduction
Many physical quantities vary in time. If the physical quantity is random this is
called a time dependent random variable or a random process.
In general we may be interested in the probability density function for:
• Values at specific or arbitrary points in time
• Extreme values for some given period of time

In some cases only the extremes are interesting for the structural behaviour. This
is for instance often the case for loads: only extreme loads may lead to collapse of
the structure. Especially for durability aspects, however, the general fluctuations
may be more important than a single extreme.

Stochastic processes are very often subdivided in two categories:

• stationary processes
• non-stationary processes

For a stationary process the distribution for the 'arbitrary point in time value' (and
some other characteristics) do not depend on time.

To describe a time varying process completely it is not sufficient to have either

the distribution of the arbitrary point value or the extreme value distribution or
even both. The full description of a time varying process requires some additional
modelling with respect to its correlation structure in time. The correlation struc-
ture primarily describes whether processes vary slow or fast in time.

Some models for describing time varying processes are:

• parametric processes
• Gaussian processes
• rectangular wave processes
• Poisson processes.

Parametric processes
In this case we have a simple deterministic function of the time. Some of the vari-
ables in this function may be random. As the random variables have different val-
ues in every realisation, every realisation of the process is different. A simple ex-

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 63

ample of a non-stationary parametric process is x(t) = b t. Where b is a random

variable.

Another and very special case is: x(t) = a, where a is random. This of course is a
stationary process.

Gaussian processes
Here the density function is Gaussian for all points in time. In order to describe
such a process we need:
• the mean as a function of time
• the standard deviation as a function of time
• the correlation for each pair of points

If the mean and the standard deviation do not depend on time (and the correlation
is only a function of the time gap) the process is called stationary. We will not
further develop this process.

Rectangular wave processes

This type of process is relatively easy to deal with in a calculation. We may dis-
tinguish between a number of varieties:
• the time intervals may be random or deterministic
• values in subsequent intervals may be dependent or independent
• the process may be stationary or non-stationary

The simplest case is characterised by:

• deterministic and equal time intervals ∆t
• independent values of the process in the various intervals
• the same distribution function Fx(X) for the value in every interval (stationary
process)

In that simple case we have the following relation between the arbitrary point in
time value and the extreme for some period T:

Fxmax ( x ) = P { X max < x } = P { X 1 < x and X 2 < x .. } =

= P { X i < x } n = F xi n ( x ) (5.1)

Where:
Xmax maximum value in period T
Xi arbitrary point in time value
n T/∆t

Poisson processes
This is in fact also a rectangular wave process, but the probability to have a non-
zero value for an interval is very small. The modelling of variable depends on the
nature of the variable and the type of limit state under consideration. One might
of course also use a combination of the models.

BE95-1347 DuraCrete
64 Probabilistic Methods for Durability design

Example: carbonation depth

Consider as an example the carbonation process in a concrete beam. The carbona-
tion depth after some period of time is given by:

dt = m A B C √t (5.2)

Where:
m model factor
A factor determined by the water cement ratio
B factor determined by the local temperature and humidity
C factor determined by the cement type
t time

For a concrete to be made the water cement ratio may be unknown and so A is a
random variable (in the design stage). The cement type of course is known and B
is deterministic. Local temperature and humidity may be unknown and add ran-
domness. The model factor m is intended to cover all unknown and unprecise
aspects of the model

For the random variables A, C and m we need probability density functions, that
is distribution type and parameters as mean values and standard deviations. But
what about the time variability?

It is interesting to find out what happens if all variables are time independent. In
such a model it is sufficient to have one measurement. After one measurement,
the future has become completely deterministic. This of course is not very realis-
tic. Even if we know that the carbonation depth after 5 years is 5 mm, we have no
absolute certainty that after 20 years this will be 10 mm. So let us have a look at
the variables in this model.

Parameter A is typically a time independent variable: the water cement ratio is

determined during the technological process of production and remains constant
afterwards. The local climate, of course may vary in time: we may have dry
summers and humid summers, hot and cold winters and so on. In the model the
value of C reflects the averaging effect of the climate in a the period form 0 to t,
so we might model C a random process with a standard deviation slowing down
in time. We will not further develop this.

Finally we have m: this variable reflects the model uncertainty. Of course there is
a large time independent part: the whole model for the carbonation depth d might
be too high or too low. However, the model may also be wrong in the basic con-
ception of a homogeneous diffusion process and there may be physical reasons
that the process is not t but for instance 3 t . In those cases the model over-
predicts in the first period and under-predicts in the second period. Finally there
may be inhomogenities which cause fluctuations in the diffusion process.

As a possible model for m we could take a rectangular wave process with periods
of 5 years and independent or dependent realisations in subsequent periods. A
possible realisation of m and d is then as depicted in figure 5.2. In some cases the

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 65

changes are too abrupt or not logic, especially if short-term decisions may have to
be taken. In those cases one might refine the mesh, comparable to a finite element
programme.

As working with dependent variables is always difficult one might try to split up
the m in a time invariant part and a time variant block process with independent
blocks:

m = m1 + m2 (5.3)

Where:
m1 time independent with µ(m1) = 1.0 and σ(m1) = 0.10
m2 rectangular wave process with µ(m2) = 0 and σ(m2) = 0.10 and
independent blocks.

This process is equivalent to saying that m is a rectangular wave process with

µ(m) = 1.0, σ(m) = 0.14 and ρ(m(t), m(t+∆t)) = 0.5.

We call such a process a hierarchical process as it has been build up from pref-
erably independent variables, each representing the variability at another level.
We will see the same model for the spatial variability.

t
d

t
Figure 5.1: Carbonation process with a rectangular wave process for m

BE95-1347 DuraCrete
66 Probabilistic Methods for Durability design

5.2 Spatial variability

Many physical quantities vary in space. If the physical quantity is random this is
called a random field. The types of modelling are similar to those for time vari-
ability. Also here we might combine the various models. Consider as an example
the concrete strength f in the building of figure 5.3:
fi = fs +∆i (5.4)
Here the term fs describes the variability of concrete form one building to an-
other. The second term ∆i describes the variability between the various ele-
ments within a structure.

f i = fs + ∆ i

Figure 5.3: Structure with spatial variability

Example
The following table presents some numerical values:

µ σ
fs = mean concrete strength of structure 30 MPa 3 MPa
∆i = deviation for member I 0 MPa 3 MPa

σ ( f i ) = σ 2 ( f s ) + σ 2 ( ∆ i ) = 4 .2 MPa

µ ( f i ) = µ ( f s ) + µ ( ∆ i ) = 30 + 0 = 30 M Pa

σ 2( fs ) 9
ρ( fi f j ) = = = 0 .5
σ 2( fs ) + σ 2(∆i ) 9 + 9

A modelling as presented in (5.4) is again an example of a hierarchical model.

The idea, of course, can be extended. We might include space independent un-
certainties, or differences between countries, or variability’s within elements.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 67

For variability’s within an element we have the same possibilities as described

for the modelling of time domain variations. (Gaussian processes, wave proc-
esses, see also figure 5.3).

Whether or not the explicit inclusion of some variability in the model makes
sense, of course, depends on the problem. Consider again the problem of car-
bonation depth. If we neglect the spatial variation within one element, it would
be sufficient or have one measurement of the depth. Given this single meas-
urement the carbonation depth of all other points of the beam or plate would be
known. If, for physical reasons, no systematic differences are present, this
would also mean that the carbonation depth is exactly the same over the entire
element. Such a model would only accept a total repair or replacement of the
element: local repair seems to be ruled out as a reasonable option. It may be
clear that such a model is not suited in all cases.

A probabilistic model consists, as already defined in Chapter 4, in the first place

of:
• type of distribution
• parameters as mean and standard deviation

A probabilistic model, however, needs to take account of the random variations in

time space. The most convenient way to do so is by the construction of hierarchi-
cal models for time and space. A model is called hierarchical if it can be written
as the sum of independent parts, each one representing a certain variability.

In the time domain we may have the sum of:

• a time invariant part
• time variant independent blocks

In the space domain we may have the sum of:

• a space independent part
• a part depending on the country
• a part depending on the building
• a part depending on the element (column, wall, floor)
• independent blocks reflecting the variability within an element.

The subject of spatial correlation is closely related to the choice of the population
and the existence of sub-populations as discussed in Chapter 4.

In principle for all parameters one needs also to consider the statistical uncer-
tainty. In fact this means that we have a very large model which can be described
only with 20 or 30 parameters. This, of course, is not very practical and a conse-
quent treatment of all variables in this way will not contribute to popularity of the
probabilistic methods. The point is that the expert will have to make a rational
choice. This choice depends on the availability of data and the intention of the
model. Some aspects may be completely irrelevant for some design situations or
maintenance decisions. The main point, however, is that one should be aware of
all these possible fluctuations that might be important. If it is estimated that it is
important, it should be modelled that way to avoid the wrong decisions to be

BE95-1347 DuraCrete
68 Probabilistic Methods for Durability design

taken from the model.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 69

6 Decision making subject to uncertainty

The ultimate task for the structural engineer is to establish a consistent decision
basis for the design and management of structures such that the overall life cy-
cle costs of the operation of the structures are minimised and such that the per-
sonal safety hazards are kept within the limits specified by legislation or soci-
ety.

As the available information regarding e.g. loading, material properties and de-
terioration processes in general is incomplete or uncertain the decision problem
is a decision problem subject to uncertain information.

The present report presents fundamental issues of decision making subject to

uncertain information. The notes are highly inspired by the work of Benjamin
and Cornell [18] on Bayesian decision making to which the reader is referred
for a more complete presentation.

6.1 Philosophical basis for uncertainty modelling

When dealing with structural systems and components it is important to realise
that these are more or less unique in the sense that completely identical struc-
tures made of identical materials and exposed to identical loading are extremely
rare. A sufficient amount of data for classical frequentistic statistical analysis of
such structures is therefore in general not available.

By use of sound engineering judgement it is, however, possible to establish

models which describe the structural response characteristics in terms of basic
variables such as structural geometry, material properties and loading charac-
teristics. Furthermore the statistical characteristics of the basic variables can be
estimated not only on the basis of frequentistic material (observations) but also
on the basis of physical understanding, engineering experience and judgement,
thus facilitating a probabilistic analysis of the structural response.

This type of statistical analysis is not frequentistic but can still be considered a
consistent way of representing uncertain information. A statistical framework
suitable for this type of statistical analysis is the Bayesian statistics, see e.g.
Lindley [19].

Due to the ability of Bayesian statistics to incorporate engineering judgement

and experience into the statistical modelling, different engineers may reach dif-

BE95-1347 DuraCrete
70 Probabilistic Methods for Durability design

ferent decisions due to the use of different statistical models. This may obvi-
ously be a serious obstacle for the use of such methods. In order to avoid this it
is necessary to define and agree on a common basis for such analysis thus en-
suring that the decision making is performed on a consistent and comparable
basis.

6.2 The decision tree

In practical decision problems such as reassessment, inspection and mainte-
nance planning for structures the number of alternative actions can be ex-
tremely large and a framework for the systematic analysis of the corresponding
consequences is therefore expedient.

An action/event tree as illustrated in figure 6.1 may conveniently represent the

decision problems.

A ction /ch oise E ven t C on seq u en ce

d ep th = 4 0 ft 0

4 0 ft p ile
-4 0 0
d ep th = 5 0 ft

d ep th = 4 0 ft
-1 0 0

5 0 ft p ile 0
d ep th = 5 0 ft

Figure 6.1 Action/event tree.

The action/event tree illustrated in figure 6.1 considers a pile driving decision
problem. The engineer is faced with the problem that the depth to solid rock is
unknown, either 40 ft or 50 ft. The engineer has two piles available with differ-
ent length of 40 ft and 50 ft respectively. If he chooses a pile which turns out to
be too short the pile must be spliced and welded and crew must be kept idle in
the meanwhile. This situation will induce costs corresponding to 400 units. If
on the other hand he chooses a pile which turns out to be too long, the pile must
be shortened. This situation will induce costs corresponding to 100 units. How-
ever there is also the possibility that the engineer will choose a pile with a cor-
rect length and there will be no costs.

The task is now to analyse such decision problems in a way making consistent
use of all the information available to the engineer, including his degree of be-
lief in the possible states, his subsequent observed data and his preferences
among the various possible action/state pairs.

To this end, use will be made of the fact that decisions shall be based on ex-
pected values of the corresponding consequences. This issue is addressed fur-
ther in the following.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 71

6.3 Decisions based on expected values

Consider the simple case where the engineer must choose between actions a1
and a2. The consequence of action a1 is with certainty B whereas the conse-
quence of action a2 is uncertain. The state of nature may be θ1 in which case the
consequence is A and the state of nature may be θ2 in which case the conse-
quence is C. The decision tree is illustrated in figure 6.2.

Action State of Nature Consequence

B
a1

A
a2 θ1

θ2
C

Figure 6.2 Decision tree illustrating a basic decision problem.

Before the true state is known the optimal decision depends upon the likelihood
of the various states of the nature θ and the seriousness of the consequences A,
B and C.

A further analysis of the decision problem requires the numerical assessment of

the preferences of the decision-maker. It is assumed that the decision maker
prefers A to B and B to C. This statement of preferences may be expressed by
any function u such that :

u( A) > u( B ) > u(C ) (6.1)

The task is to find a particular function u namely the utility function such that it
is logically consistent to decide between a1 and a2 by comparing u(B) with the
expected value of the utility of the action a2, namely

p u( A) + (1 − p ) u( C ) (6.2)

where p is the probability that the state of nature is θ1.

Assuming that u(A) and u(C) have been given appropriate values the question
is - what value should u(B) have in order to make the expected value a valid
decision criteria? If the probability of θ1 being the state of nature p is equal to 1
the decision maker would choose a2 over a1 because he prefers A to B. On the
other hand if the probability of θ1 being the state of nature is equal to 0 he
would choose a1 over a2. For a value of p somewhere between 0 and 1 the deci-
sion-maker will be indifferent to choosing a1 over a2. This value p* may be de-
termined and u(B) assigned as

BE95-1347 DuraCrete
72 Probabilistic Methods for Durability design

u( B ) = p * u( A) + (1 − p * )u( C ) (6.3)

From equation (6.3) it is seen that u(B) will lie between u(A) and u(C) for all
choices of p* and therefore the utility function is consistent with the stated pref-
erences. Furthermore it is seen that the decision maker should choose the action
a1 to a2 only if the expected utility given this action E[ua1] is greater than
E[ua2]. This is realised by noting that for all p smaller than p* and with u(B)
given by equation (6.3) we have

u( B ) > pu( A) + (1 − p )u( C )

c
p * u( A) + (1 − p * )u( C ) > pu( A) + (1 − p )u( C ) (6.4)
c
u( C ) + ( u( A) − u( C )) p * > u( C ) + ( u( A) − u( C )) p

This means that if u(B) is properly assigned in consistency with the decision
makers stated preferences i.e. A preferred to B preferred to C and the indiffer-
ence probability p* the ranking of the expected values of the utility determines
the ranking of actions.

6.4 Utility assignment

The utility function may be and may not be associated with monetary value,
however, if the utility function is a linear function of a monetary unit (e.g. dol-
lars) then the expected monetary value is a valid criterion for choosing among
alternative actions.

The phenomenon that some decision makers are risk averse i.e. accept less
probability of loss the larger the loss, is an issue which has been devoted a tre-
mendous interest by especially the engineers and decision makers applying de-
cision analysis to problems where the loss of human lives is a possible conse-
quence. The consequence of risk aversion is that the utility function becomes
non-linear. Many unclosed discussions have been initiated on which shape the
non-linear utility function should have in order to reflect the preferences of the
decision-makers. It may, however, be realised that there would be no need for
non-linear utility functions if all consequences of the state of nature would be
included in the utility function. Sometimes it is difficult to identify all conse-
quences and decision-makers are tempted to guess at the shape of the utility
function instead.

6.5 Assignment of probability

The assignment of probability is a central issue in decision analysis as this to-
gether with the assignment of a utility function and the consequences of the
possible states of nature determines the expected utility. Therefore a brief
treatment of uncertainty and assignment of probability is given in the follow-
ing.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 73

Several sources of uncertainty influence the probability of the state of nature.

The uncertainties which usually must be considered in structural engineering
are the physical uncertainties, statistical uncertainties and model uncertainties.
The physical uncertainties are typically uncertainties associated with the load-
ing environment, the geometry of the structure, the material properties and the
repair qualities. The statistical uncertainties arise from incomplete statistical
information e.g. due to a small number of materials tests. Finally the model un-
certainties must be considered to take into account the uncertainty associated
with the idealised mathematical descriptions used to approximate the actual
physical behaviour of the structures.

For some uncertainties, physical arguments can be used to identify the appro-
priate family of distribution function, however, in general there is some free-
dom in the choice of distribution function.

The basis for the uncertainty model can be any mixture of frequentistic and
subjective information. It should, however, be emphasised that on the one hand
the model should aim for simplicity in the formulation of the uncertainty mod-
elling and on the other hand the model should be close enough to reality to al-
low for including important information collected during the lifetime of the
structure, thus allowing for updating the uncertainty structure in the considered
problem. In this way uncertainty models which initially are based entirely on
subjective information will as new information is collected eventually be based
on frequentistic information.

In principle, it does not matter whether the information is quantitative or quali-

tative. Formal reliability assessment, however, requires a quantitative type of
statement as a starting point for further processing.

Qualitative statements like ‘the structure looks fine’ should therefore be trans-
lated into quantitative statements like: no cracking of concrete, no sign of cor-
rosion and so on. If one also knows, from other experiments, what the threshold
values are for observing visual cracks and corrosion, these statements can be
used in the formal procedure.

The probabilities of the possible states of nature which are used in the decision
analysis may conveniently be estimated using modern reliability methods.
There the uncertain variables X are often denoted basic variables with realisa-
tions x. The probability of the event F that x ∈ ωF, where the domain ωF is de-
fined by g(x) ≤ 0, can be evaluated from

PF = ∫f
g ( x )≤ 0
X ( x )dx (6.5)

where f X ( x ) is the joint density function of X and g(x) is a so-called limit state
function defining the boundary of the set F .

The integral in equation (6.5) plays a central role in the reliability analysis and
has therefore been devoted special attention over the last decades. As the inte-

BE95-1347 DuraCrete
74 Probabilistic Methods for Durability design

gral in general has no analytical solution it is easily realised that its solution or
numerical approximation becomes a major task for integral dimension larger
than say 6 and for small probabilities. Sufficiently accurate approximations
have been developed which are based on asymptotic integral expansions. These
methods which are called FORM/SORM have become standard in reliability
analysis and commercial software with these methods are available.
FORM/SORM methods and commercial software are described in Ditlevsen
[20].

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 75

7 Decision making subject to uncertainty

Having formulated the decision problem in terms of an action state of nature
tree, with properly assignment of utility and probability structure the numerical
evaluation of the decision alternatives may be performed.

Depending on the state of information at the time of the decision analysis tree
different analysis types are distinguished, namely prior analysis, posterior
analysis and pre-posterior analysis. Each of these are important in practical ap-
plications of decision analysis and are therefore discussed briefly in the follow-
ing.

7.1 Decision analysis with given information - prior

analysis
When the utility function has been defined and the probabilities of the various
state of nature corresponding to different consequences have been estimated the
analysis reduces to the calculation of the expected utilities corresponding to the
different action alternatives.

At this stage the probabilistic description P[θ] of the state of nature θ is usually
called a prior description and denoted P’[θ].

To illustrate the prior decision analysis the pile driving decision problem from
section 6.2 is considered again. The decision problem is stated as follows. The
engineer has a choice between two actions:
a0 : Select a 40 ft pile to drive
a1 : Select a 50 ft pile to drive
The possible states of nature are the following two
θ0 : The depth of the stratum is 40 ft
θ1 : The depth of the stratum is 50 ft

The prior assessment of probabilities is based on experience from pile driving

several hundred feet away from the present site together with large-scale geo-
logical maps. The prior probabilities are
P’[θ0] = 0.70
P’[θ1] = 0.30

BE95-1347 DuraCrete
76 Probabilistic Methods for Durability design

Based on the prior information alone it is easily seen that the expected utility
E ' [u] amounts to:
E' [u] = min{ P' [θ 0 ]* 0 + P' [θ 1 ]* 400 , P' [θ 0 ]* 100 + P' [θ 1 ]* 0 } =
min{ 0 .7 * 0 + 0 .3 * 400 , 0 .7 * 100 + 0 .3 * 0 } =
0 .7 * 100 + 0 .3 * 0 = 70

The decision tree is illustrated in figure 7.1 together with the utilities.

p=0.70 u = 0
θ0
120
p=0.30
a0 θ1
u = 400 (splice needed)
p=0.70
a1 70 θ u = 100 (pile shortened)
0

p=0.30
θ1
u=0

Figure 7.1 Simple decision problem with assigned prior probabilities and util-
ity.

It is seen that action alternative a1 yields the largest expected utility (smallest
expense) why this action alternative is the optimal decision.

7.2 Decision analysis with additional information -

posterior analysis
When additional information becomes available the probability structure in the
decision problem may be updated. Having updated the probability structure the
decision analysis is unchanged in comparison to the situation with given - prior
information.

Given the result of an experiment zk the updated probability structure (or just
the posterior probability) is denoted P’’[θ] and may be evaluated by use of
Bayes rule

P ''(θ ) =
[ ]
P z k θ i P ' [θ i ]
(7.1)
∑ P[ z k ] [ ]
θ j P' θ j
j

which may be explained as

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 77

 Posterior probability of θ i   Normalising  Sample likelihood   prior probability

  =      (7.2)
 with given sample outcome   constant   given θ   of θ 

The normalising factor is to ensure that P’’[θi] forms a proper probability. The
mixing of new and old information appears through the sample likelihood
P[zk|θi] and the prior probability P’[θi]. The likelihood is the probability of ob-
taining the observation zk given the true state of nature θi.

Prior

Posterior Likelih ood

P r io r P o ste r io r L ik e lih o o d

L ik e lih o o d
P r io r

P o s te r io r

Figure 7.2: Illustration of updating of probability structures.

In figure 7.2 an illustration is given of corresponding prior and posterior prob-

ability density functions together with likelihood functions. In the first case the
prior information is strong and the likelihood is weak (small sample size). In
the second case the prior information and the likelihood are of comparable
strength. In the last case the prior information is relatively weak in comparison
to the likelihood.

To illustrate the posterior decision analysis the pile driving decision problem is
considered again.

The engineer now plans to collect information about the depth by use of simple
sonic testing. Due to soil irregularities, measuring errors, etc. the depth indi-
cated by the sonic tests are not wholly reliable estimates of the true depth. As-
sume that the sample space of the sonic tests are 40, 45 and 50 ft. Two of the
possible outcomes might be said to favour a particular state of nature, namely
40 ft and 50 ft respectively whereas the indication of 45 ft is ambiguous. The
test properties are summarised in table 7.1 as sample likelihoods.

BE95-1347 DuraCrete
78 Probabilistic Methods for Durability design

True state θ0 θ1

Test result 40 ft - depth 50 ft – depth

z0 - 40 ft indication 0.6 0.1

z1 - 50 ft indication 0.1 0.7

z2 - 45 ft indication 0.3 0.2

Table 7.1 [ ]
Sample likelihoods P zi θ j for sonic depth testing. Note that
the sum of probabilities in the columns is equal to one.

The engineer performs the tests and the depth is observed as 45 ft. With this
information the posterior probability P’’[θi] is determined by

[ ] [ ]
P' ' [θ 0 ] = P θ 0 z 2 ∝ P z 2 θ 0 P[θ 0 ] = 0 .3 0.7 = 0.21

[ ] [ ]
P' ' [θ 1 ] = P θ 1 z2 ∝ P z2 θ 1 P[θ 1 ] = 0 .2 0.3 = 0.06

As the true state of nature is either θ0 or θ1 the sum of posterior probabilities

must add up to unity. Normalising the posterior probabilities with their sum
yields the probabilities

0 .21
[ ]
P' ' θ 0 z 2 =
0 .21 + 0 .06
= 0.78

0 .06
[ ]
P' ' θ 1 z 2 =
0 .21 + 0 .06
= 0.22

which are also shown in figure 7.3. Having determined the updated probabili-
[ ]
ties the posterior expected values E '' u z 2 of the utility corresponding to the
optimal action alternative is readily obtained as

[
{E '' u( a j ) z 2 } =]
min{P'' [θ0 ] * 0 + P'' [θ1 ] * 400, P'' [θ0 ] *100 + P'' [θ1 ] * 0}
min{0.78 * 0 + 0.22 * 400, 0.78 *100 + 0.22 * 0} =
0.78 *100 + 0.22 * 0 = 78

and indicated in boxes in figure 7.3.

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 79

p = 0 .7 8 u = 0
θ0
88
p = 0 .2 2
a0 θ1
u = 4 0 0 (sp lice n eed ed )
p = 0 .7 8
a1 78 θ0 u = 1 0 0 (p ile sh o rten ed )

p = 0 .2 2
θ1
u = 0

Figure 7.3 Illustration of decision tree for pile driving decision problem.

It is seen from the example that even though the engineer believes that the true
state is more likely to be θ0 his prior information and the ambiguous test results
do not give sufficient confidence to offset the relatively large costs of 400.

7.3 Decision analysis with ‘unknown’ information -

pre-posterior analysis
Often the decision-maker has the possibility to ‘buy’ additional information
through an experiment before actually making his choice of action. If the costs
of this information is small in comparison to the information on the state of na-
ture it promises the decision-maker should go ahead and perform the experi-
ment. If several different types of experiments are possible the decision-maker
must choose the experiment yielding the overall largest utility.

If we consider the example from section 7.2 again the decision problem could
be formulated as to decide whether or not to perform the ultra sonic depth
measurement.

The situation prior to performing the experiment has already been considered in
section 7.1. There it was found that the expected utility based entirely on the
prior information E ' [u] is 70 units.

What needs to be considered is the situation where the experiment is planned

and the experiment result is still unknown. In this situation the expected utility
disregarding the experiment costs can be found as

n n
E[ u] = ∑ P' [ z ]* E'' [ u z ] = ∑ P' [ z ]* min{ E'' [ u( a
i=1
i i
i=1
i
j = 1 ,m
j ]
) zi } (7.3)

where n is the number of different possible experiment findings and m is the

number of different decision alternatives. In equation (7.4) the only new term in
comparison to the previous section is P' [ zi ] which may be calculated by

BE95-1347 DuraCrete
80 Probabilistic Methods for Durability design

[ ]
P' [ z i ] = P z i θ 0 * P' [θ 0 ] + P z i θ 1 * P' [θ 1 ] [ ] (7.4)

With reference to table 7.1 the prior probabilities of obtaining the different in-
dications by the tests are P' [ z1 ] , P' [ z 2 ] , P' [ z 3 ] given by

[ ] [ ]
P' [ z 0 ] = P z 0 θ 0 * P' [θ 0 ] + P z 0 θ 1 * P' [θ 1 ] = 0 .6 * 0 .7 + 0 .1* 0 .3 = 0 .45

[ ] [ ]
P' [ z 1 ] = P z 1 θ 0 * P' [θ 0 ] + P z 1 θ 1 * P' [θ 1 ] = 0 .1* 0 .7 + 0 .7 * 0 .3 = 0 .28

[ ] [ ]
P' [ z 2 ] = P z 2 θ0 * P' [θ0 ] + P z 2 θ1 * P' [θ1 ] = 0 .3* 0 .7 + 0 .2 * 0 .3 = 0 .27

The posterior utilities in equation (7.4) are found to be

[
{ E' ' u( a j ) z0 } = ]
[ ] [ ] [
min{ P' ' θ 0 z0 * 0 + P' ' θ 1 z0 * 400 , P' ' θ 0 z0 * 100 + P' ' θ 1 z0 * 0 } ] [ ]
min{ 0 .93 * 0 + 0 .07 * 400 , 0 .93 * 100 + 0 .07 * 0 } =
0 .07 * 400 + 0 .93 * 0 = 28

[
{E '' u( a j ) z1 } = ]
[ ] [ ]
min{P'' θ0 z1 * 0 + P'' θ1 z1 * 400, P'' θ0 z1 * 00 + P'' θ1 z1 * 0} [ ] [ ]
min{0.25 * 0 + 0.75 * 400, 0.25 *100 + 0.75 * 0} =
0.25 *100 + 0.75 * 0 = 25

[ ]
where the posterior probabilities P ' ' θ i z 0 and P ' ' θ i z1 are determined as [ ]
already shown in section 7.2 for P ' ' [θ z ] . i 2

The expected utility corresponding to the situation where the experiment is

planned and disregarding the experiment costs is therefore

[ ]
E [u] = ∑ P' [ z i ]* E' ' u z i = 28 * 0 .45 + 25 * 0 .28 + 78 * 0 .27 = 40 .66
i =1

By comparison of this result with the expected utility corresponding to the prior
information it is seen that the experiment should be performed if the costs of
the experiment is less than

E' [u] − E [u] = 70 .00 − 40 .66 = 29 .34

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 81

7.4 Postscript
For the encouraged reader who has now seen how easily decision making sub-
ject to uncertain information can be supported by the Bayesian decision theory
it is now time to gain confidence in the application of the methodology. There-
fore a simple example is given below, in which ordinary common sense can
verify the Bayesian methods.

A contestant of a quiz is faced with three closed doors. He/she is told that be-
hind one of the doors there is a Mercedes and behind each of the other doors
there is a goat. The contestant is posed with the problem of choosing one of the
doors and will be rewarded with what is behind it. The contestant picks a door
but does not open it. The quizmaster then opens one of the remaining doors be-
hind which he is sure that there is a goat. Two doors remain closed. The quiz-
master asks whether the contestant would like to reconsider his/her decision
and possibly change the choice of door.

The question to you dear reader is now: would it be wise of the contestant to
open the other door? Formulate the decision basis in terms of prior posterior
probabilities.

BE95-1347 DuraCrete
82 Probabilistic Methods for Durability design

8 References
1. Vrouwenvelder A.C.W.M., Basis of Design, Theoretical Background,
Proceedings IABSE Colloquium Basis of Design and Actions on Struc-
tures - Background and Application of Eurocode 1, March 27-29, 1996,
Delft, The Netherlands, pp. 25-32

2. Siemes, T. and S. Rostam, Durable safety and serviceability - a perform-

ance based format, Proceedings IABSE Colloquium Basis of Design and
Actions on Structures - Background and Application of Eurocode 1,
March 27-29, 1996, Delft, The Netherlands, pp. 41-50

3. Sentler L., Report TVBK-3018: Service life predictions of concrete struc-

tures, 1981, Department of Structural Engineering - Lund University of
Technology, Lund, Sweden

4. Siemes A.J.M. and A.C.W.M. Vrouwenvelder, Durability of Buildings: a

reliability analysis, 1985, Heron, Volume 30, Number 3, Delft The Neth-
erlands, 1985

5. Hergenröder M., Zur statistischen Instandhaltungsplanung für

bestehende Betonbauwerke bei Karbonatisierung des Betons und
möglicher Korrosion der Bewehrung, 1992, Technische Universität
München, Germany

6. Sarja A. and E. Vesikari (ed.), Rilem report 14, Durability Design of

Concrete Structures - Calculation methods for service life design of con-
crete structures, 1996, Chapman & Hall London, England (in print)

7. Daniels, B. and O. Dijkstra and A. Vrouwenvelder, TNO Bouw report

95-CON-R1575: Serviceability Limit States, Rijswijk, January 1996, pp.
57

8. Cornell, C.A., Structural Safety Specifications Based on Second Moment

Reliability Analysis, Final report of the IABSE Symposium on Concepts
of Safety of Structures and Methods of Design, London 1969

9. Rackwitz, R., Principles and Methods for a Practical Probabilistic Ap-

proach to Structural Safety, Subcommittee for First Order Reliability
Concepts for the Design Codes of the Joint CEB-CECM-CIB-FIP-IABSE

BE95-1347 DuraCrete
Probabilistic methods for Durability Design 83

Committee on Structural Safety, CEB-report, December 1976

10. Venezeano, D., Contribution to Second Moment Reliability Theory, De-
partment of Civil Engineering, Massachusetts Institute for Technology,
Publication 289, April 1974

11. Rackwitz, R., First Order Reliability Theories and Stochastic Models,
Proceedings of the 2nd International Conference on Structural Safety and
Reliability, ICOSSAR 1977, München 1977

12. Strating, J. and C. Gouwens, Veiligheid en bouwconstructies - De reken-

kundige beoordeling van de veiligheid, Bijlage bij Bouwen met Staal

13. SIA 260: Sicherheit und Gebrauchsfähigkeit van Tragwerken,

Schweizerischer Ingenieur- und SIA Architectenverein, 5. Fassung, Mai
1980 (in German)

14. Grundlagen für die Festlegung van Sicherheitsanforderungen für

bauliche Anlagen, DIN Deutsches Institut für Normung e.v., 1e Auflage
1981, Berlin

15. Recommendations for Loading- and Safety Regulations for Structural

Design, The Nordic Committee on Building Regulations, NKB report No.
36, November 1978

16. Development of a Probability Based Load Criterion for American Na-

tional Standard A 58, Building Code Requirements for Minimum Design
Loads in Buildings and Other Structures, NBS Special Publication 577,
June 1980

17. Load and Resistance Factor Design (LRFD) Specification, AISC Com-
mittee on specifications, 5th Draft, May 1981, Chicago

18 Benjamin, J.R. & Cornell, C.A., Probability, Statistics and Decision for
Civil Engineers, McGraw-Hill, NY, 1970.

19 Lindley,D.V., Introduction to Probability and Statistics from a Bayesian

Viewpoint, Vol. 1+2, Cambridge University Press, Cambridge 1976.

20. Ditlevsen, O., Bærende konstruktioners sikkerhed, SBI-rapport 211, Stat-

ens Byggeforskningsinstitut 1990.

BE95-1347 DuraCrete
84 Probabilistic Methods for Durability design

9 Table for the standard normal distribution

β Pf β Pf
0,0 0,5 -2,0 0,022750
-0,1 0,460172 -2,1 0,017864
-0,2 0,420740 -2,2 0,013903
-0,3 0,382089 -2,3 0,010724
-0,4 0,344578 -2,4 0,008198
-0,5 0,308538 -2,5 0,006210
-0,6 0,274253 -2,6 0,004661
-0,7 0,241964 -2,7 0,003467
-0,8 0,211855 -2,8 0,002555
-0,9 0,184060 -2,9 0,001866
-1,0 0,158655 -3,0 0,001350
-1,1 0,135666 -3,1 0,000968
-1,2 0,115070 -3,2 0,000687
-1,3 0,096801 -3,3 0,000483
-1,4 0,080757 -3,4 0,000337
-1,5 0,066807 -3,5 0,000233
-1,6 0,054799 -3,6 0,000159
-1,7 0,044565 -3,7 0,000108
-1,8 0,035930 -3,8 7,24E-05
-1,9 0,028716 -3,9 4,81E-05

BE95-1347 DuraCrete
 β Pf
β Pf -6,0 9,90E-10
-4,0 3,17E-05 -6,1 5,32E-10
-4,1 2,07E-05 -6,2 2,83E-10
-4,2 1,34E-05 -6,3 1,49E-10
-4,3 8,55E-06 -6,4 7,80E-11
-4,4 5,42E-06 -6,5 4,04E-11
-4,5 3,40E-06 -6,6 2,07E-11
-4,6 2,11E-06 -6,7 1,05E-11
-4,7 1,30E-06 -6,8 5,26E-12
-4,8 7,94E-07 -6,9 2,62E-12
-4,9 4,80E-07 -7,0 1,29E-12
-5,0 2,87E-07 -7,1 6,28E-13
-5,1 1,70E-07 -7,2 3,03E-13
-5,2 9,98E-08 -7,3 1,45E-13
-5,3 5,80E-08 -7,4 6,86E-14
-5,4 3,34E-08 -7,5 3,22E-14
-5,5 1,90E-08 -7,6 1,50E-14
-5,6 1,07E-08 -7,7 6,88E-15
-5,7 6,01E-09 -7,8 3,11E-15
-5,8 3,33E-09 -7,9 1,44E-15
-5,9 1,82E-09
86 Probabilistic Methods for Durability design

BE95-1347 DuraCrete