PROBABILISTIC ASSESSMENT OF EXISTING STRUCTURES

RELIABILITY UPDATING THROUGH TESTING

MEICK J. JØRGENSEN
SIMON BAK

MASTER’S THESIS, STRUCTURAL AND CIVIL

ENGINEERING
Reference for edited cover picture [Anindita, 2017].

ii
 Department of Civil Engineering
Thomas Manns Vej 23
9220 Aalborg Ø
Phone +45 99408484
http://www.civil.aau.dk

Title: Abstract:
Probabilistic Assessment of
The purpose of this project is to study how
Existing Structures
various of testing methods can be used to
access and/or update the reliability level of
Subtitle:
existing structures. This is investigated both
Reliability Updating Through Testing on a component and system level and is
carried out considering the various methods
Project type: with respect to proposed decision models.
Master’s thesis The analyses are all carried out for a repre-
sentative linear limit state equation where
Project period: only one variable load is present at a time.
1st of February 2018 - 8th of June 2018 Relevant stochastic models are established,
where especially the stochastic models related
Project participants: to the resistance are of interest in regards to
Meick J. Jørgensen the obtained results.
Simon Bak
The first part of the project mainly revolves
Supervisors: around how proof load testing of components
in a system can be used to update the reliabil-
Jannie Sønderkær Nielsen ity level based on whether prior information
John Dalsgaard Sørensen is available or not. The total expected costs
are then estimated in regards to a decision
model.
Main report: 65
Appendices: 38
The last part of the project assesses how
Submitted: 8th of June 2018
cylinder and CAPO-tests can be used to assess
the reliability level and which method is the
most cost-beneficial in regards to average total
expected costs. The total expected costs are
also in this case estimated in regards to a
decision model.

The content of the report is freely available, but publication (with source reference) may only take place in agreement
with the authors.
 Preface

This report and its appendices are the product of the undersigned master’s thesis at the 4th
semester of the Master’s programme in Structural and Civil Engineering at the Department of
Civil Engineering. The report is prepared and made in accordance and compliance to the current
curriculum at the School of Engineering and Science from the 1st of February 2018 to the 8th of
June 2018 under the supervision of John Dalsgaard Sørensen and Jannie Sønderkær Nielsen.

The theme of the project is Assessment of the structural reliability for existing structures, where
the purpose of this project is to assess an approach to reliability update a structure through different
testing methods.

In regards to the project, many appreciations goes towards above mentioned supervisors for their
supervision, expertise and flexibility throughout the project period.

Meick J. Jørgensen
Simon Bak
Aalborg, 8th of June 2018

Reading guide
The project consists of two parts: A main report and its appendices, which are found at the back
of the report. The main report refers to the appendix where the appertaining calculations and
extensional documents are to be found. The files used for software, e.g MATLAB, are attached in
a digital folder. The main report is numbered numerically, so that the first chapter is called Chapter
1 and sections here-under 1.1, 1.2 etc. The appendices are analogously referred to alphabetically
e.g. A, B etc. Figures and tables are numbered with respect to the their respective chapter,
meaning that the first figure in Chapter 2 is referred to as Figure 2.1 etc. Equations are referred to
analogously to this with the exception of having ( ) in-closing the reference, e.g. Equation (2.1).
The digital folder contains literature, experimental data, results and MATLAB scripts.

This report uses the Harvard method of bibliography with the name of the author and year of
publication into brackets after the text, e.g [Ayyub and McCuen, 2002]. If the source reference
is positioned before a punctuation it only refers to the sentence whereas if it is placed behind the
punctuation it refers back to the whole text section. A list of all the source references is given in
the bibliography list at the end of the main report.

v
 Resume

Dette kandidatspeciale er udarbejdet i perioden d. 1. februar 2018 - d. 8. juni 2018 og er produktet

af kandidatudannelsen Structural and Civil Engineering på Aalborg Universitet.

Titlen på projektet er: Probabilistic Assessment of Existing Structures med den tilhørende undertitel
Reliability Updating Through Testing.

For en eksisterende konstruktion kan et ønske om at ændre anvendelsen eller forlænge levetiden
føre til en ombygning eller renovering. Hvis en ombygning, renovering eller ændret anvendelse
indebærer udskiftning eller ændring af eksisterende, bærende konstruktionsdele, statisk virkemåde
eller lasters størrelse og omfordelingen af disse, så skal konstruktionens bæreevne naturligvis
revurderes, hvis de ikke umiddelbart erstattes med nye komponenter. Forstærkning eller
udskiftning af bærende konstruktionsdele i en eksisterende konstruktion er en dyr løsning
sammenlignet med samme tiltag for en ny konstruktion i designfasen. Herved kan det være
af stor interesse at vide, hvorvidt det er muligt gennem forsøg at bevise, om pålideligheden er
tilstrækkelig, eller hvor meget forstærkning der i så fald skal vendes for at opnå et bestemt ønsket
sikkerhedsindeks.

Formålet med dette projekt er at undersøge, hvordan forskellige testmetoder kan bruges til at
vurdere og/eller opdatere sikkerhedsniveauet for eksisterende konstruktioner. Dette undersøges
både på komponent- og systemniveau med hensyn til forskellige testmetoder i relation til foreslåede
beslutningsmodeller. Analyserne udføres alle for en repræsentativ lineær svigtfunktion, hvor kun
en variabel last er til stede af gangen. Relevante stokastiske modeller etableres, hvor især de
stokastiske modeller relateret til styrken er af interesse med hensyn til de opnåede resultater.

Den første del af projektet omhandler primært, hvordan prøvebelastninger komponenter i et system
kan anvendes til at opdatere sikkerhedsniveauet, alt efter om der foreligger forhåndsviden eller ej.
De samlede forventede omkostninger beregnes derefter i forbindelse med en beslutningsmodel.

Den sidste del af projektet vurderer, hvordan cylinder- og CAPO-test kan bruges til at
vurdere sikkerhedsniveauet af en eksisterende konstruktion, samt hvilken metode som er den
mest omkostningseffektive med hensyn til gennemsnitlige samlede forventede omkostninger.
De samlede forventede omkostninger beregnes også i dette tilfælde i forbindelse med en
beslutningsmodel.

vii
 Table of Contents

Chapter 1 Introduction 1
1.1 Safety - then and now . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Prerequisites and delimiters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Structure of the report . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Chapter 2 Assessment of Existing Structures Reliability Level 5

2.1 Reasons for assessment of existing structures . . . . . . . . . . . . . . . . . . . 5
2.2 Methods for obtaining new information . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Estimation of reliability level . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 Overall decision model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

Chapter 3 Proof Load Testing Without Prior Information 13

3.1 Reliability level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Limit state equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 Stochastic models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4 Updated reliability level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

Chapter 4 Proof Load Testing With Prior Information 19

4.1 Assessment model: Preliminary inspection and rating . . . . . . . . . . . . . . . 20
4.2 Prerequisites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.3 Reliability level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.4 Limit state equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.5 Stochastic models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.6 Estimation of system reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.7 Updated system reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.8 Analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.9 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.10 Decision model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

Chapter 5 Design Based on Sampling Through Testing 45

5.1 Cylinder test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.2 CAPO-test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

ix
 Table of Contents

Chapter 6 Cost-Benefit Analysis Between Cylinder and CAPO-tests 51

6.1 Assessment model: Preliminary inspection and rating . . . . . . . . . . . . . . . 51
6.2 Prerequisites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.3 Reliability level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.4 Limit state equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.5 Stochastic models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.6 Updating of the characteristic strength . . . . . . . . . . . . . . . . . . . . . . . 54
6.7 Simulation algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.8 Cost-benefit analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

Chapter 7 Discussion 63

Chapter 8 Conclusion 65

Bibliography 67

Appendix A Simulation Techniques 69

A.1 Simulation of stochastic variables . . . . . . . . . . . . . . . . . . . . . . . . . 69
A.2 Simulation of correlated lognormally distributed numbers . . . . . . . . . . . . . 71
A.3 Crude Monte Carlo simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
A.4 Reliability updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

Appendix B Factors for Proof Load Testing Without Prior Information 77

Appendix C Factors for Proof Load Testing With Prior Information 81

C.1 Series systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
C.2 Parallel systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

Appendix D Cost-Benefit Analysis of Sampling Methods 91

Appendix E Data by Krenchel and Bickley 106

x
Chapter

1
Introduction

An assessment or reevaluation of the reliability level of an already existing structure is in some

cases necessary due to a change in circumstances regarding the structure. For instance, this could
be caused by a change in loading or usage, discovery of damage on the structure, extension of
service life or in general doubts about the reliability level and structural safety of the given structure.

The documentation of the load-bearing capacity of components in a structure was previously based
on a different basis than today. As a result, the safety of an already existing structure based on
previous codes cannot be used as documentation in regards to the current codes. Therefore, the
structural safety of an existing structure has to be updated either by reinforcement of the load-
bearing components of concern in case the load-bearing capacity is insufficient, or by proving that
the components have sufficient strength and safety despite the change in circumstances by gathering
additional information through tests or more advanced calculations. In general, reinforcement of
components in already existing structures is expensive in comparison to new structures that are still
in the design phase. Instead a less expensive method is to conduct tests on the existing structure
in order to potentially improve the certainty of material strength proving that the reliability level
is sufficient. Different methods for testing exist, but they are only briefly accounted for in regards
to the assessment of structural safety of existing structures in the codes. In general the codes for
assessment of structural safety for existing structures are not well developed, which is problematic,
when a reevaluation of the structural safety of an existing structure has to be carried out. However,
new codes are currently under development that seek to better establish a set of guidelines or rules
for different problems to assess structural safety for existing structures. [SBI 251, 2015]

1.1 Safety - then and now

The basis for design of structures in Denmark has changed over the past years. In the following a
distinction between three periods is presented [SBI 251, 2015]:

1856-1908 - Design requirements

In the structural codes of Copenhagen from 1856 [Bygningslov, 1856], 1871 [Bygningslov, 1871]
and 1889 [Bygningslov, 1889] a detailed description of materials, dimensions and interior design
is described. The basis of these descriptions are from artisan-related traditions.

1
 1. Introduction

1908-1965 - Permissible stresses

The Danish Engineering Association’s codes for reinforced concrete structures operates with the
calculation of permissible actions in the form of permissible stresses. [Ingeniørforening, 1908]
After this period, structures are designed in regards to functional requirements, i.e. the desire for a
given load-bearing capacity instead of specified design descriptions. The permissible stresses are
expressed as a given percentage of the mean strength of the material. In 1945 the first edition of
DS 410 [1945] is published. Until then, the magnitude of the loads could be found in Norm 12 by
Ingeniørforening [1916]. Load combinations and the magnitude of the permissible stresses were
still found in the structural codes.

1965 - Partial safety method

After 1965 the structural codes began to assign partial safety factors on the load- and strength side
as well as describe methods for estimation of the characteristic values of the load and strength
properties. The value of the characteristic strength was typically expressed as the 5-10%-quantile
in a probability distribution. The value of the characteristic variable load was typically expressed
as the 98%-quantile in a distribution function corresponding to a 50 year return period.

In the beginning, the partial safety factors were based on experience such that the obtained
dimensions are similar to the ones obtained from the permissible stresses. Later on, the reliability
index method was used as basis for the calibration of the partial safety factors in the safety codes.
[NKB, 1978] This method was also used for verifying the load-bearing capacity of larger structures.

In 1982 the first edition of the Code of Practice for the Safety of Structures, DS 409, was released.
[Ingeniørforening, 1982] This code governs the rest of the structural codes by establishing a set of
rules for load combinations and rules for estimation of the partial safety factors in these codes.

In 2001 the Joint Committee of Structural Dafety publication on assessment of existing structures
was published, which recommends procedures and tools for assessment of existing structures’
safety level. [Diamantidis, 2001]

From 2009 Eurocodes are applied together with the Danish National Annexes for design of new
strutures, where the safety system is accounted for in Eurocode 0, DS/EN 1990 [2007] and loads in
Eurocode 1, DS/EN 1991-1 [2007]. In DS/INF 172 [2009] the basis for the partial safety factors in
the National Annex and Eurocodes that result in an acceptable reliability level has been described.
The partial safety factors are estimated by the use of the reliability index method and a set of
prerequisites are related to this, which involves probability distribution types and quantile-values
for load- and strength parameters.

1.2 Problem statement

This project aim is to study how to assess the reliability level of existing structures, both on a
component and system level with focus on performing tests as basis for assessment. This is to be
carried out considering various testing methods, and relevant decision models are to be proposed.

2
1.3. Prerequisites and delimiters Aalborg University

1.3 Prerequisites and delimiters

Only linear limit state equations are considered. Furthermore, only one variable load is present at
a time, which means that presented results are not valid for combination of several loads.

Analyses are not carried out for any specific materials, apart from in Chapter 5 where concrete
subjected to compression is assessed.

Due to limited simulation time, almost all results are obtained for specific defined values that in
reality can vary a lot.

1.4 Structure of the report

In this section a short description of the report’s overall structure is presented.

As a basis, reasons and methods for assessing the structural reliability of an existing structure are
described.

Some of the various of test methods are analyzed with respect to assessing and/or updating the
structural reliability of existing structures. In relation to this, assessment- and decision models are
constructed to clarify the method of approach for the case studies related to the different methods.

A cost-benefit analysis is performed between two testing methods, cylinder compression test and
CAPO-test. The stochastic models for cylinder compression tests are based on codes, while the
stochastic modelling of the CAPO-test in this project is based on experimental measurements.

Lastly, an overall assessment of the results obtained in the different analyses are evaluated and
discussed, so that a proper conclusion to the problem statement can be formulated.

The above-described is summarized in Figure 1.1, which shows a structure diagram of the content
of the report.

3
 1. Introduction

Assessment of existing
structures reliability
level

Proof Load Testing Design Based on

Proof Load Testing
Without Prior Sampling Through
With Prior Information
Information Testing

Cost-Benefit Analysis
Assessment of
Between Cylinder and
decision model
CAPO-tests

Assessment of
Discussion
decision model

Conclusion

Figure 1.1: Structure diagram of the content of the report.

4
Chapter

2
Assessment of Existing Structures
Reliability Level

In this chapter the purpose or reassessing the reliability level of an existing structures and when
to do so is described along with the various of methods to evaluate and/or update the reliability
level of an existing structure. Theory and methods described in this chapter are found in SBI 251
[2015].

2.1 Reasons for assessment of existing structures

An existing structure designed on the basis of previous codes can in most cases not be documented
in accordance to the codes currently valid for new structures. This is a result of the change of codes
from then to now, as described in Chapter 1, and the fact that the structure naturally decays over
its lifetime, meaning that the remaining lifetime will usually be shorter than what is required in
the current codes for new structures. For instance, an existing structure with a remaining lifetime
of 20 years will most likely not have an acceptable reliability level in accordance with a required
lifetime of 50 years in the current codes, if fatigue or other damage accumulation is of importance.
However, it is generally accepted that existing structures do not have to comply with the current
codes given that they comply with codes at the time of construction, unless changes has occurred
or will occur. In this case, the reliability level of the existing structure has to be assessed, which
among others might be caused by the following:

• Change of application of the structure.

• Change in static system of the load-bearing structure.
• Change in load conditions or surroundings, e.g. accumulation of snow, increase in wind
load or change in conditions of foundation.
• Change of lifetime for a structure.
• Deterioration or damage of the bearing structure, e.g. collision, overloading, settlements or
unstable cracks.
• Discovery of errors in relation to design and construction.

The assessment of an existing structure should initially be based on the current codes, since
they are assumed to be created on the basis of greater knowledge than the previous codes. The
assessment requires that information about the strength characteristics of the structure is known,
either from historical documentation or by conducting proof load tests. This information is then
used to update the reliability level of the structure in accordance to the new codes through Bayesian

5
 2. Assessment of Existing Structures Reliability Level

updating. In cases, for which the current codes do not cover the existing structures, they have to be
considered from an engineering point of view in regards to functionality and capacity. However,
circumstances related to uncertainty of strength characteristics and determination of loads still
have to be considered in accordance with the current codes.

2.2 Methods for obtaining new information

When reevaluating the structural reliability of an existing structure, it is necessary to obtain new
information regarding the resistance of the structure. This information is generally obtained
through either measurements and/or observations of the stochastic variable i.e. the resistance or
through observation of events. Methods used for measuring and observing the stochastic variables
can be either direct or indirect and examples of these are as follows:

• Direct measurements of the resistance, e.g. from bored out cylinders of a concrete structure.
• Indirect measurements of the resistance, e.g. by performing CAPO-test on a concrete
structure.

Observation of events include events caused by actively performing a test on the structure or by
natural occurrences. This includes the following examples:

• Proof load testing a structure and observe if failure occurs.

• Observation of non-failure during a natural extreme event, e.g. a storm.

Reevaluation of the structural reliability through cylinder test, CAPO-test and proof load testing
will be assessed in this project. These test methods are intended to be either destructive, non-
destructive or partly destructive and will be described in the following.

2.2.1 Proof load tests

Proof load testing is a common method for obtaining new information on a structural component
or a whole system of components. The method can be performed, when no prior information
associated to the existing structure is available or when the available information does not document
the reliability level of the structure. It is an experimental method, where a component or system is
subjected to a load close to the extreme exposures throughout the service life. This test establishes
a minimum load level that the tested specimen can sustain. Proof load testing is considered or is at
least intended to be a non-destructive testing method, meaning that a probability of not successfully
carrying out the test exists. Proof load testing is expensive and bears the risk of damage to the
tested structure and its surroundings. Thus, the decision to carry out proof load tests should depend
on a decision model that includes the cost and risks involved in comparison to alternative methods.

In regards to Bayesian statistics, the resistance of the structural component is updated by knowing
that the component survived the proof load and is therefore described as an observation of an
event. There are two types of observations in context to this, which are an equality event and an
inequality event, see Appendix A.4. If no prior information about the resistance is available when
proof load testing, then the updating of the probability of failure, should the component survive
the proof load test, is modelled by an equality event given the correct prerequisites. When no
prior information is available and the proof load is not increased to failure, then it only validates
the reliability level of the particular structural member that was tested. This means that the whole
population in a structural system has to be tested in order to have a total validation for the whole

6
2.3. Estimation of reliability level Aalborg University

structure. Proof load testing without prior information is studied in Chapter 3. If the contrary
scenario is considered e.g. having prior information about the resistance, then the updating of
the probability of failure is modelled by an inequality event, see Appendix A.4. In this case, it is
sufficient to perform proof load tests on a portion of the structural components. Likewise, if the
proof load is increased to failure, only a portion of the population has to be tested in order to assess
the reliability level. Proof load testing with prior information is studied in Chapter 4.

Proof load testing is not always a possibility when having to assess the reliability level of an
existing structure. Reasons for this could be that the proof load test is too difficult to perform, e.g.
on offshore structures, the component cannot properly be accessed or perhaps the risk of the proof
load test is too expensive. Nonetheless, there are alternative methods in relation to obtaining new
information of existing structures.

2.2.2 Resistance sampling tests

Resistance sampling tests are other methods that can be used for obtaining new information about a
structure. The principle of these testing methods is to directly measure the strength of the material,
i.e. the stochastic variable, through test samples from the given structure. The information obtained
through these type of tests can be used to update the stochastic variables, if prior information is
available, or otherwise give an estimate of the stochastic variable. This is described in Appendix
A.4. Three common methods to measure the resistance of a structural component is; cylinder
compression test, CAPO-test and LOK-test.

The cylinder compression test is carried out by boring out a cylindrical test specimen from
the component. The specimen is then compression tested in a laboratory and the resistance is
measured, which is why this is considered a destructive testing method. This method is only used
for concrete structures. For CAPO-test and LOK-test the compressive strength of the material
is directly measured on-site by the use of a pull-out test. The pull-out force (tensile force) is
converted to a strength resistance. These testing methods are considered partly destructive. A
more detailed description of the various of sampling test methods is found in Chapter 5, while
Chapter 6 describes how they are used to assess the reliability level of an existing structure.

Even though these resistance sampling tests are considered either destructive or partly destructive,
it is intended that these tests are carried out so that the component is not damaged. The decision to
perform a cylinder compression test or a CAPO/LOK-test can be evaluated based on a cost-benefit
analysis of the tests, which will be studied in Chapter 6.

2.3 Estimation of reliability level

2.3.1 General
When a suitable combination of loads and load-bearing properties e.g. geometry and strengths
causes the structure to have exhausted its load-bearing capacity, then the structure will go into
failure-mode. The probability of failure is normally assessed in terms of calculation models
related to the load-bearing capacity. The calculation model has to be provided with specific values
representing the loads and the load-bearing properties. These values are uncertain (stochastic) and
are expressed by the use of distribution functions. By using these distribution functions together
with the calculation model it is possible to estimate the probability of obtaining a combination

7
 2. Assessment of Existing Structures Reliability Level

of loads and load-bearing capacities that yields failure. The stochastic values are included in the
calculation model for the probability of failure with different weights. This is why it is appropriate
to be able to describe the most influential stochastic variables as accurate as possible.

2.3.2 Estimation of the probability of failure

In Figure 2.1 distribution functions are illustrated in a simple case, e.g. a single stochastic load
and a single stochastic strength property.

Figure 2.1: Illustration of a joint probability distribution function for two uncertain parameters.
[SBI 251, 2015]

The shown surface is the joint probability distribution function of the two separate distribution
functions of the load and strength property. The joint probability distribution function showcases
the frequency of the combined events. The joint probability function is standardized such that the
area below equals 1, e.g. having 100% probability of having any event. Standardizing the surface
means that the variables are transformed into standard normal distributed variables with a mean
value, µ = 0, and standard deviation, σ = 1.

Furthermore, in Figure 2.1 a limit state equation, g = R − S, is shown for which R is the resistance
and S is the load. The probability of failure is found at the most probable combination of events
that will cause failure. In case of failure, the combined events yield a resistance less than the load.

The probability of failure is equal to the volume beneath the failure surface, which is shown on
Figure 2.1. The failure surface is the area in which the strength is less than the load. The minimum

8
2.3. Estimation of reliability level Aalborg University

acceptance criteria for the probability of failure is of the magnitude 10−5 and with a reference
period of 1 year (annual probability of failure). This is why the upper tail of load-related stochastic
variables are of importance and the contrary applies for the strength-related stochastic variables.

When dealing with normal distributed stochastic variables and linear limit state equations, then the
probability of failure is estimated as Φ(−β), where β is called the reliability index. The reliability
index is defined as the shortest distance between the origin of the standardized stochastic variables
and the failure surface in the standardized space, while Φ(•) is the standardized normal distribution
function. This explanation is illustrated in Figure 2.2.

Figure 2.2: Illustration of probability distribution functions in regards to estimating the reliability
index, β, for the design point, P. SBI 251 [2015]

Non-linear limit state equations can be approximated with a tangent in the point, which has the
lowest β-value. This is an iterative process to estimate the point, P, which yields the smallest
value for β. This point is defined as the design point and represents the combination of multiple
stochastic events that will most likely cause failure. The components in the vector from the origin
to P express the reliability index’s sensitivity towards changes in the various of stochastic variables’
distribution functions.

The described method is called First Order Reliability Method (FORM). The error related to the
probability of failure, Φ(−β), is directly related to the failure surface’s deviation from the linear
surface, and hence the part of the failure surface that is not included. In these cases another
method can be used instead. This method is called Second Order Reliability Method (SORM),
which approximates the failure surface with 2. order derivatives and thus gives more precise
results. Using FORM yields sufficient results as stated in DS/EN 1990 [2007]. The tangent line

9
 2. Assessment of Existing Structures Reliability Level

develops into a tangent plane if three stochastic variables are introduced and into a hyper plane if
further are added.

The stochastic variables are not always given in terms of normal distributions. When dealing
with various of distribution functions it is essential to transform these into normal distributed
uncertainties instead. This is carried out by substituting the stochastic variables in the limit state
equation with normal distributed stochastic variables, see Sørensen [2011] for further explanation.

Alternatively, Monte-Carlo simulation of a large amount of combined events can be carried out
and the amount of obtained failures compared to the non-failure events is counted, see Appendix
A. Hereby, an estimate of the probability of failure is found. The amount of samples that has to be
simulated has to be sufficiently large, so that an even larger amount of samples will yield the same
results. In this project, it is chosen to use Monte-Carlo simulation for all analyses for estimating
the probability of failure.

2.3.3 Distribution functions

The distribution functions that form a basis for the safety assessment includes both physical
uncertainties, statistical uncertainties and model uncertainties. The physical uncertainties on the
load side is related to the variations that are expected to happen, e.g. wind load during a storm.
On the resistance side the physical uncertainty is related to e.g. the construction, geometry and
natural variations in the construction materials.

The statistical uncertainty is governed by the extent of observations and test results, which the
physical uncertainty is based on.

A model uncertainty is associated to the load-bearing model. This model uncertainty indicates
how good a relation there is between the resistance estimated by a model and the strength estimated
through tests, see DS/EN 1990 [2007], Annex D.

For the load model the uncertainty expresses the uncertainty related to the conversion between the
load and the load effect.

2.3.4 Reliability level

When the involved distribution functions related to the load- and resistance side describe the
probability of having a maximum load within a given reference period, then this means that the
reliability level is the probability of having failure within this reference period. The reference
period that is most commonly used is an annual one, i.e. one year.

For structures under normal circumstances this annual probability of failure has to have a magnitude
less than 10−5 . This corresponds to a reliability index with a value of approximately β = 4.3.
Since the required probability of failure is annual, it means that the remaining life time on the
structure is of no importance in the assessment of the structures safety given that no degradation
is present, e.g. fatigue or corrosion.

As a minimum the desired reliability level has to be obtained. In a specific existing structure some
circumstances might cause this to be unreasonably expensive or impossible to carry out. In these
cases the government governs deviation from the codes if necessary.

10
2.4. Overall decision model Aalborg University

The requirement for the reliability level depends on the structure’s consequence class and failure
type, see Table 2.1. A description related to the failure types are found in Table 2.2. Values for
failure type II are found from DS/EN 1990 [2007]. Values for the other failure types I and III are
found by respectively increasing and reducing the annual probability of failure with a factor 10, as
indicated in DS/INF 172 [2009].

Table 2.1: Tentative target values for the annual reliability indices for different reliability classes
and failure types cf. DS/INF 172 [2009].

Reliability Class Failure Type I Failure Type II Failure Type III

CC1 (less serious) 3.2 3.8 4.3
CC2 (serious) 3.8 4.3 4.8
CC3 (very serious) 4.2 4.7 5.2

Table 2.2: Classification and description of failure types cf. DS/INF 172 [2009]

Failure type Description

Type I Alerted failure with extra load-bearing capacity reserves from e.g. hardening of steel
Type II Alerted failure without extra-load bearing capacity
Type III Failure caused by e.g. loss of stability

Alternative tentative reliability indices have also been proposed by the Joint Committee of
Structural Safety, which are shown in Figure 2.3. These values are based on a cost-benefit
analysis and the choice of target reliability index depends on an assessment of consequence class
and cost of safety measures.

Table 2.3: Tentative target values for the annual reliability indices for different reliability classes
and cost of failures cf. Joint Committee on Structural Safety [2001].

Large cost of safety Normal cost of safety Low cost of safety

Consequence Class
measures (A) measures (B) measures (C)
CC1 (less serious) 3.1 3.7 4.2
CC2 (serious) 3.3 4.2 4.4
CC3 (very serious) 3.7 4.4 4.7

2.4 Overall decision model

An overall decision model over the different test methods that can be used for reevaluating the
reliability level of existing structures will be established in this section. The decision model takes
basis in a realization that one of the circumstances listed in Section 2.1 leads to the conclusion that
the reliability level of an existing structure has to be reevaluated. The possible test methods used
in the reevaluation are illustrated in the decision model in Figure 2.3.

11
 2. Assessment of Existing Structures Reliability Level

Figure 2.3: Overall decision model.

An initial decision is made between performing proof load tests or test sampling methods like
cylinder compression tests or CAPO-tests. This decision depends on parameters such as the type
of material used in the structure and total expected costs. If proof load tests are performed,
then it has to be assessed whether prior information is available of not. The case without prior
information in regards to reevaluating the reliability level of the structure is evaluated in Chapter
3. The case with prior information is studied in Chapter 4 in which a more detailed decision model
is established for the different decisions and actions that can be taken in this case. If test sampling
methods are decided as choice of action, then it has to be decided whether to perform cylinder or
CAPO-tests as well as assess whether prior information is available of not. The cases for which
prior information is available are not investigated in this report. The other cases for which no
prior information is available are evaluated in Chapter 5 and 6. The decision between performing
cylinder or CAPO-tests is based on a cost-benefit analysis and is used as part of a more detailed
decision model in Chapter 6.

12
Chapter

3
Proof Load Testing Without Prior
Information

This chapter will describe how proof load testing can be used to assess the reliability level of an
existing structure, when no prior information about the resistance is available.

When no prior information associated to the existing structure is available, then the reliability level
has to be assessed solely from the information obtained through proof load testing. In this case
information about a system can be obtained by either proof load testing a portion of the components
to failure or by proof load testing all components in a system with a proof load corresponding to
a target reliability level. This chapter will only study the latter in which an analysis of how proof
load testing can be used to update the reliability level of an existing structure. In relation to this,
the chapter will study how different stochastic models for the variable load influence the reliability
level, when performing proof load tests.

3.1 Reliability level

The analysis for assessing how the reliability level can be updated by performing proof load testing
takes basis in the reliability levels for new structures. The target reliability level is chosen for
different cases listed in Table 2.1. This includes the following component reliability indices with
a reference period of one year in accordance to Table 3.1:

Table 3.1: The target reliability indices that are used for assessing how the reliability level can be
updated by performing proof load testing without prior information.

Target reliability index, βcomp

t

β = 3.2 β = 3.8 β = 4.3 β = 4.7 β = 5.2

3.2 Limit state equation

A limit state equation has to be established that includes both the proof load and the permanent and
variable loads multiplied with model uncertainties related to the loads. It is assumed that only a
single variable load is present at a time, e.g. wind or snow load. This is assessed to be a reasonable
assumption, since e.g. no snow load is present when wind is assumed to be the dominating load.

13
 3. Proof Load Testing Without Prior Information

Thus, a generic representative linear limit state equation can be expressed as following:

g = Ppl − XG,i G(1 − α) + XQ,i Qα

(3.1)

where

Ppl Proof load.

G Permanent load.
Q Variable load.
XG,i Model uncertainties for the permanent load.
XQ,i Model uncertainties for the variable load.
α α = 0 corresponds to full permanent load and α = 1 corresponds to full variable load.

The ratio for the permanent and variable load can be varied in the interval between 0 and 1, i.e.
the interval between full permanent and full variable load respectively for the characteristic load
requirement. The loads and model uncertainties depend on the chosen stochastic models listed in
the following. The proof load is a deterministic load that is determined by:

Ppl = Ptest · k test (3.2)

where

ktest Multiplication factor dependent on the percentage permanent load of the characteristic
load and the target reliability index, βt .
Ptest Characteristic load requirement.

The characteristic load requirement is defined as the additional load that the structure has to carry
in addition to the permanent load already applied to the structure before subjecting it to the proof
load. The purpose is to estimate the k test -factors that corresponds to the target reliability indices.
This can be done by assuming that the resistance of the tested component in the structure is emptied
at a value just above the proof load, since no knowledge of the resistance above the proof load is
available, given that the conditions for the proof load is equivalent to the real load size and duration.
Thus, the assessment of the reliability level is based on an equality event, which is described in
Appendix A.4, i.e. the resistance is assumed equal to the proof load. The probability of failure
can be estimated as the probability that a load greater than the proof load subjects the structure.

3.3 Stochastic models

Stochastic models are established for all stochastic variables in the limit state equation. The
models include distribution type, mean value,µ, coefficient of variation, V, and quantiles, p, for
the characteristic values. The stochastic models of interest are listed in Table 3.2 for which model
2 and 3 for the wind and snow load respectively are determined from Equation (3.3) - (3.6).

Q w,2 = Cpe Cr Cg Qr e f ,2 for wind model 2 (3.3)

Q w,3 = XQr e f Cpe Cr Cg Qr e f ,3 for wind model 3 (3.4)
Q s,2 = XQs g,2 Q sg,2 for snow model 2 (3.5)
Q s,3 = XQs,3 Q sg,3 for snow model 3 (3.6)

14
3.3. Stochastic models Aalborg University

Table 3.2: Stochastic models for loads and model uncertainties.

Load type Model Random variable Symbol Dist. type Mean V Quantile
Permanent Model 1 Permanent load GP Normal 1.00 0.10 0.50
Model 1 Wind load Q w,1 Gumbel 1.00 0.40 0.98
Pressure coeff. Cpe Gumbel 1.00 0.15 0.78
Roughness coeff. Cr Lognorm. 1.00 0.15 1.00
Model 2
Gust coeff. Cg Lognorm. 1.00 0.10 1.00
Wind Wind pressure1 Qr e f ,2 Gumbel 1.00 0.25 0.98
Model uncertainty XQr e f Lognorm. 0.80 0.20 1.00
Pressure coeff. Cpe Gumbel 1.00 0.10 0.78
Model 3 Roughness coeff. Cr Lognorm. 0.80 0.10 1.00
Gust coeff. Cg Lognorm. 1.00 0.10 1.00
Wind pressure1 Qr e f ,3 Gumbel 1.00 0.25 0.98
Model 1 Snow load Q s,1 Gumbel 1.00 0.40 0.98
Model uncertainty XQs,2 Lognorm. 1.00 0.30 (µ + σ)
Model 2
Snow Snow load2 Q sg,2 Gumbel 1.00 0.40 0.98
Model uncertainty XQs,3 Lognorm. 1.00 0.35 0.50
Model 3
Snow load2 Q sg,3 Gumbel 1.00 0.40 0.98
1 Annual reference wind pressure.
2 Snow load at ground level.

Currently, there is no clear-cut stochastic model for the variable loads that should be used. Thus,
it is of interest to assess the difference between the models in regards to assessing the reliability
level of an existing structure through proof load testing.

The stochastic models for the permanent load as well as model 1 for wind and snow load are used
as basis for the loads in DS/EN 1991-1 [2007] and often used for design. Physical and model
uncertainties are already included in model 1 for the wind and snow load. Model 1 for wind and
snow load are identical and normally referred to under the same category of ‘variable load’ in
the current codes. However, it might be fair to assume that wind and snow loads do not follow
the same stochastic model, which is why it is interesting to also evaluate alternative proposed
stochastic models for the wind and snow load. This includes model 2 and 3 for the wind and
snow load respectively in Table 3.2. Model 3 for the wind load and model 2 for the snow load
are described in JCSS [2001], while model 2 for the wind load and model 3 for the snow load are
described in S A K O [1999] and Sanpolesi [1997] respectively.

For model 2 and 3 of the wind load in comparison to model 1, it can be seen that the coefficient
of variation is lower for the annual reference wind pressure compared to the wind load, which is
caused by the models not yet accounting for a number of different time-invariant factors related
to pressure, roughness and gust. An uncertainty and thus a stochastic model is related to each of
these factors as listed in Table 3.2. Furthermore, a model uncertainty is added in model 3.

The stochastic model for the snow load at ground level in model 2 and 3 is identical to model 1 of
the snow load. However, model 2 and 3 have been added a model uncertainty with a rather high
coefficient of variation to take the snow time-invariant part into account. The high uncertainty
is caused by the snow load being dependent on the shape of the structure and shelter from the
surroundings.

15
 3. Proof Load Testing Without Prior Information

3.4 Updated reliability level

The reliability level of a system can be updated by proof load testing components in the system.
Since no prior information is available, all components have to be proof load tested, given they are
not tested to failure. Firstly, the stochastic models for model 1 described in Section 3.3 will be used
for estimating the necessary k test -factors for obtaining the target reliability levels listed in Section
3.1. These ktest -factors are estimated for a varying weight factor, α. This study has previously
been carried out in SBI 251 [2015], and thus the initial purpose is to recreate these results before
different alternative proposed stochastic models are used in the analysis. The results are listed in
Table 3.3.
Table 3.3: k test -factors for the 1. stochastic model of permanent and wind/snow load in SBI 251
[2015].

Characteristic permanent
Factor: k test
load percentage of
characteristic load
β = 3.2 β = 3.8 β = 4.3 β = 4.8 β = 5.2
requirement [%]
0 1.52 1.85 2.19 2.48 2.87
25 1.39 1.65 1.89 2.11 2.40
50 1.28 1.45 1.61 1.75 1.95
60 1.24 1.38 1.51 1.62 1.78
65 1.22 1.34 1.46 1.56 1.69
70 1.20 1.31 1.41 1.50 1.61
75 1.20 1.29 1.37 1.44 1.54
80 1.20 1.27 1.34 1.40 1.48
85 1.21 1.28 1.33 1.37 1.43
90 1.24 1.30 1.34 1.38 1.43
95 1.28 1.34 1.38 1.42 1.47
100 1.32 1.38 1.43 1.47 1.52

Instead of estimating the ktest -factors in the entire interval of α, an average value is usually suggested
to be used. The weighting between the permanent and variable load usually lies in the interval
α = [0.3; 0.8] and thus an average value of the k test -factors can be estimated for this interval. The
results are listed in Table 3.4.
Table 3.4: Average k test -factors in the interval α = [0.3; 0.8] for model 1.

k test -factor
β = 3.2 β = 3.8 β = 4.3 β = 4.7 β = 5.2
1.39 1.60 1.79 1.96 2.19

Similarly, the k test -factors for the four alternative stochastic models of the variable loads can be
estimated. If the target reliability indices listed in 3.1 are used again, the estimated k test -factors will
become unreasonable large due to all the additional variables in the models that add up to a larger
coefficient of variation. Thus, the probability of the components actually surviving the proof load
is really low and therefore not feasible. Instead, it will be assessed how the reliability level differs
between the different stochastic models of the variable load when using the same magnitude of

16
3.4. Updated reliability level Aalborg University

proof load. These reliability indices can then be used as new target reliability indices for which
k test -factors can be estimated. These target reliability indices can be estimated by inserting k test -
factors in the interval α = [0.3; 0.8] from model 1 into the limit state equation in Equation (3.1)
for the stochastic models 2 and 3 and calculating an average reliability index for these stochastic
models. This is done for a reliability index of β = 4.3. The following values are found for the
different stochastic models:

Table 3.5: Average annual targeted reliability indices for the alternative stochastic models in the
interval α = [0.3; 0.8] in comparison to β = 4.3 for model 1.

Target reliability index, βcomp

t

Wind model 2 Wind model 3 Snow model 2 Snow model 3

2.6 3.5 1.9 1.8

The results clearly shows that the reliability level decreases significantly when using the alternative
stochastic models of the variable load, especially for the snow load. Average values of the k test -
factor in the interval α = [0.3; 0.8] can now be estimated for the new target reliability indices
and compared in order to assess the sensitivity of the k test -factors, when using different stochastic
models. The k test -factors are estimated for the target reliability indices in Table 3.5 as well as
βcomp
t ± 0.5 and 1.0 that are assessed to represent two reliability classes above and below the
reliability indices in Table 3.5 similar to the case for model 1. The results are seen in Tables 3.6
- 3.9, while more detailed tables for varying α-values can be found in Appendix B, if accurate
information about the percentage permanent load of the characteristic load requirement is known.

Table 3.6: Average k test -factors in the interval α = [0.3; 0.8] for wind model 2.

k test -factor
β = 1.6 β = 2.1 β = 2.6 β = 3.1 β = 3.6
1.32 1.48 1.68 1.92 2.21

Table 3.7: Average k test -factors in the interval α = [0.3; 0.8] for wind model 3.

k test -factor
β = 2.5 β = 3.0 β = 3.5 β = 4.0 β = 4.5
1.28 1.46 1.68 1.94 2.27

Table 3.8: Average k test -factors in the interval α = [0.3; 0.8] for snow model 2.

k test -factor
β = 0.9 β = 1.4 β = 1.9 β = 2.4 β = 2.9
1.21 1.42 1.68 2.01 2.42

17
 3. Proof Load Testing Without Prior Information

Table 3.9: Average ktest -factors in the interval α = [0.3; 0.8] for snow model 3.

k test -factor
β = 0.8 β = 1.3 β = 1.8 β = 2.3 β = 2.8
1.19 1.40 1.68 2.03 2.48

By comparing the average k test -factors between the models, it can be seen that the results only
differ slightly from each other. This indicates that the k test -factors have a low sensitivity in regards
to which model that is used, given that the resulting reliability indices are accepted.

18
Chapter

4
Proof Load Testing With Prior
Information

This chapter will make an assessment of how the reliability level of an existing structure can be
updated based on proof load testing of components in a system, when prior information about the
resistance is available.

When prior information is available for a structural system consisting of a number of components
with resistances that can be considered as realizations from a homogeneous population in the
existing structure, then the reliability level can be updated by conducting proof load tests on a
portion of the components. The prior information includes knowledge about the mean value, µR ,
and coefficient of variation, VR , of the resistance, R, of the components in a system. An example
of such a system is illustrated in Figure 4.1 for a reinforced concrete flooring structure, where the
steel reinforcement bars in the upper and lower part respectively are considered components of a
homogeneous population in a system. If changes to the structure lead to an increase in the loads,
which are carried by the reinforcement bars, then proof load testing can be performed to ensure an
acceptable reliability level of the structure.

G, Q

R, XR
Figure 4.1: A concrete component consisting of a number of identical reinforcement bars with
the resistance parameters, R and XR , and subjected to a permanent and variable load,
G and Q.

For an existing structure, the stochastic models for the loads and model uncertainties can be
considered known, but there will always be an uncertainty related to the resistance of the
components, despite prior information being available. Based on the prior information available

19
 4. Proof Load Testing With Prior Information

and an assessment of the uncertainty related to this information, different actions can be taken to
ensure that the structure has an acceptable reliability level, and these are described by a decision
model that is based on a preliminary inspection and rating of the resistance.

4.1 Assessment model: Preliminary inspection and rating

An assessment model that takes basis in a preliminary inspection and rating of the resistance will
be established. This assessment model constitutes a scenario in which the reliability level of an
existing structure has to be reevaluated. For instance, an office building has to be changed into
a fitness center, meaning that the loads are increased on the load-bearing structures such as a
concrete flooring structure as previously illustrated in Figure 4.1. The flooring structure can be
considered a system consisting of a number of reinforcement bars from a homogeneous population.
In this case, the reinforcement bars in the lower part of the flooring structure will be subjected to
increased tension forces, and thus the reliability level has to be reevaluated. The assessment model
in Figure 4.2 has been established to illustrate the process behind a reevaluation.

Figure 4.2: Assessment model that takes basis in a preliminary inspection and rating of the
resistance and describes the different assessment that have to be made when the
reliability level of an existing structure has to be reevaluated.

Assessment (A) comprises an assessment of the prior information, Pi , available for the resistance of
the existing structure. The prior information might be believed to still hold true or having changed
due to time having passed. For example, cracks might be discovered, which leads to a belief that
the resistance is less than documented. Likewise, the resistance of e.g. a concrete structure might
be believed to be greater than documented, since the compressive strength of concrete increases
with time.

In assessment (B), the level of uncertainty, V(µR ) and (VR ), related to the prior information, µR
and VR , have to be assessed, since the resistance is an uncertain variable. This is modelled by the
symbol Ui in the figure. The third assessment (C) is to identify what type of system that has to be
reevaluated and the number of components in the system, denoted Si in the figure. This includes
whether the system is a series or parallel system as well as the static mode of action of the system,
i.e. is the failure alerted or instantaneous.

Finally, decision (A) is whether to perform proof load testing on a number of components in
the system or to reinforce the components in the system to ensure an acceptable reliability level.
This decision is based on an analysis of cost, Ci , for performing the respective actions. This cost
depends on the number of proof load tested components and the probability of failure occurring
during the proof load testing. The optimal decision can then be made based on the action with the
least cost.

20
4.2. Prerequisites Aalborg University

4.2 Prerequisites
The aim of this chapter is to assess the influence of proof load testing on the reliability level of
a structural system consisting of a number of components and in order to do so, it is necessary
to establish some prerequisites for the resistance. This includes both the stochastic model of
the resistance, µR and VR , as well as the uncertainties, V(µR ) and V(VR ), related to µR and VR .
The prerequisites used for µR and VR are described in Section 4.5.2 for the stochastic model. The
uncertainty on the statistical parameters, µR and VR , is a result of the structure having already existed
in a period of time, which generates an uncertainty related to the resistance of the components.
Therefore, when assessing the reliability of an existing structure, the level of uncertainty, V(µR )
and V(VR ), is important to take into consideration. In this analysis, the uncertainty is considered
to be classified in the following levels assuming that all parameters are lognormal distributed:
[Sørensen, 2016]

• Small for V(µR ) ≤ 15% and V(VR ) ≤ 7.5%

• Large for 15% < V(µR ) ≤ 20% and 7.5% < V(µR ) ≤ 10%
• Very large for 20% < V(µR ) ≤ 30% and 10% < V(µR ) ≤ 15%

The influence of proof load testing will be assessed for when the uncertainty for estimating µR and
VR is either small, large or very large respectively. This analysis will be carried out for both series
and parallel systems.

4.3 Reliability level

The analysis for assessing the influence of proof load testing on the system reliability will take
its basis in a target reliability level for new structures. The target reliability level is chosen for a
reliability class CC2 and failure type II, which corresponds to the following component reliability
index with a reference period of one year in accordance to Table 2.1:

• βcomp
t = 4.3

However, it should be noted that the reliability level for existing structures is normally accepted
to be lower than that of new structures, since the cost of safety measure is larger for an existing
structure compared to a new structure.

4.4 Limit state equation

In order to estimate the reliability level of a system, the following generic, representative, linear
limit state equation is used for each of the components for calibrating the mean value of the
resistance, µR , to result in a targeted reliability index of βcomp
t = 4.3:

g = RXR − (G(1 − α) + Qα) (4.1)

where

G Normalized permanent load.

Q Normalized variable load.
XR Model uncertainty of the resistance, R.

21
 4. Proof Load Testing With Prior Information

In Equation (4.1), g < 0 will lead to failure. It is assumed that only a single variable load is
applied to the structure at a time, meaning Equation (4.1) is not valid, when multiple variable loads
are present simultaneously, e.g. both wind and snow load. Furthermore, the weighting factor is
initially assumed to be α = 0.5 for all analyses, since it is assessed to be a representative average
value, but will in the later analyses be varied to assess the sensitivity of the weighting factor.

4.5 Stochastic models

The stochastic models used for the variables in Equation (4.1) will be established in the following,
including the mean value, µ, coefficient of variation, V, and characteristic value for the stochastic
variables, Rk . Firstly, the load models will be described, since the loads are the known stochastic
variables, which can then be used to give an estimate of the resistance that is the unknown variable.

4.5.1 Loads
The following assumptions are chosen for the normalized characteristic value of the permanent
and variable load applied to the components in the system:

• Gk = 1
• Qk = 1

The permanent and variable load are assumed to be distributed in accordance to the stochastic
models described in DS/INF 172 [2009] and are listed in Table 4.1. The statistical parameters, i.e.
µ and V are estimated directly from the characteristic values of the 50%-quantile for the permanent
load and 98%-quantile for the variable load. The statistical parameters for the permanent load are
estimated by:

µG = G k (4.2)
σG
VG = = 0.10 (4.3)
µG
Similarly to Equation (4.2) and (4.3), the statistical parameters for the variable load are estimated
by the following, assuming V = 40% and the characteristic value is equal to 1:

µQ = 0.490 Q k (4.4)
σQ
VQ = = 0.40 (4.5)
µQ

The distribution of the variable load is an expression for the probability that a particular load is
the maximum load to occur in a reference period of one year, i.e the variable load models are the
maximum annual load.

Table 4.1: Statistical parameters for the stochastic models of the permanent and variable loads.

Parameter Symbol Distribution type µ V Characteristic value

Permanent load G Normal 1.00 0.10 1.00
Variable load Q Gumbel 0.49 0.40 1.00

22
4.5. Stochastic models Aalborg University

4.5.2 Resistance and model uncertainties

The resistance is the uncertain and unknown variable, meaning that it is necessary to set up
prerequisites for establishing a stochastic model that can be used for the analysis. The resistance
will be evaluated in two cases for which the following assumptions listed in Table 4.2 are made for
the distribution type, V, and quantile for Rk corresponding to the total resistance, RXR .

Table 4.2: Stochastic models of the resistances and model uncertainty.

Parameter Symbol Distribution type V Rk

Resistance 1 R1 Lognormal 0.10 5 %-quantile
Resistance 2 R2 Lognormal 0.20 5 %-quantile
Model uncertainty XR Lognormal 0.05

The first stochastic model of the resistance corresponds to a material in between steel and concrete in
regards to VR = 10% while the second stochastic model represents laminated wood with VR = 20%.
The same model uncertainty related to the resistance has been chosen for both stochastic models
with VXR . The mean, µR , and characteristic value, Rk , for the two stochastic models will be
estimated corresponding to a component reliability index, βcomp t = 4.3. The mean values can
be estimated through iteration by using Monte-Carlo simulation for which the inverse method,
x = Fx−1 (FU (U)), is used, where FU (U) is the distribution function for a uniform distributed
stochastic variable and FU (U) = U for 0 ≤ U < 1 and FU (U) = 1 for u > 1, see Appendix A. The
following simulation algorithm is used for estimating these mean values and is also illustrated in
the flowchart in Figure 4.3 to better visualize the procedure.

1. A mean value of the resistance, µR , is guessed.

2. N = 107 random realizations of the resistance, Ri , model uncertainty, XR,i , and the loads Gi
and Qi using the appropriate stochastic parameters and distribution functions are simulated.
3. The simulated variables in step 2 are placed into Equation 4.1 and g is calculated. g < 0
leads to failure.
4. The probability of failure, P f , for a component is estimated as the ratio between the amount
of failed components and number of total simulations, N.
5. The reliability index, β, is calculated by βcomp = −Φ−1 (P f ).
6. If | βt − β| ≤ 0.1, the correct mean value, µR , has been guessed, otherwise step 1-6 is
repeated until convergence is met.

23
 4. Proof Load Testing With Prior Information

Guess a mean value, 𝜇 .

1

Simulate 𝑁 realizations of the

2 Resistances, 𝑅 , model uncertainties,
𝑋 , , and loads, 𝐺 and 𝑄

Insert 𝑧, 𝑅 , 𝑋 , , 𝐺 and 𝑄 into the limit

3 state function

Estimated the probability of failure, 𝑃 ,

4 for a component

Estimate the reliability index, 𝛽, for a

5 component

If |𝛽 − 𝛽| ≤ 0.1, the correct mean

6 value, 𝜇 , has been guessed, otherwise
repeat step 1-6 with 𝜇 = 𝜇 ± 0.01

Figure 4.3: Algorithm for estimating µR corresponding to βcomp

t = 4.3 for a component, when
using different µR,i and VR,i .

The results for the mean values are listed in Table 4.3.

Table 4.3: Mean values of the resistances.

Parameter Symbol µR,i

Resistance 1 R1 1.82
Resistance 2 R2 2.34

When µR,i for the stochastic models have been estimated, then Rk,i can be determined as the
5%-quantile of the total resistance, RXR . Both R and XR are lognormal distributed, meaning that
their product is likewise lognormal distributed.

Firstly, the lognormal statistical parameters, σL,XR and µ L,XR , of XR can be determined by:
q
σL,XR = ln(VX2R + 1) (4.6)
1 2
µ L,XR = ln(µXR ) − σL,X (4.7)
2 R

24
4.6. Estimation of system reliability Aalborg University

The only parameters in Equation (4.6) and (4.7) are VXR and XR . The assumed value of XR and
VXR can be found in Table 4.2. Analogously, the lognormal statistical parameters, σL, R and µ L, R ,
of R are determined from:
q
σL, R = ln(VR2 + 1) (4.8)
1 2
µ L,R = ln(µR ) − σL, (4.9)
2 R
The characteristic value, Rk , of the 5%-quantile of the resistance can be determined by the
following:

ln(x0.05 ) = µ L − 1.645σL (4.10)

The value −1.645 corresponds to the 5% quantile of the standard normal distribution. The mean
value and standard deviation of the product of R and XR are determined by:

µ L = µ L,XR + µ L, R (4.11)
q
σL = σL,X 2
R
+ σL,R2 (4.12)

Through Equation (4.6) - (4.12) it is possible to determine Rk if µ and V of R and XR are known.
The results for Rk as well as the final stochastic models for R and XR are shown in Table 4.4.

Table 4.4: Statistical parameters for the stochastic models of the resistances and model uncertainty.

Parameter Symbol Distribution type µ V Rk

Resistance 1 R1 Lognormal 1.82 0.10 1.49
Resistance 2 R2 Lognormal 2.34 0.20 1.55
Model uncertainty XR Lognormal 1.00 0.05

4.6 Estimation of system reliability

Now that the appropriate stochastic models and limit state equation have been described, the
procedure used to simulate and estimate the reliability index of a system, βsys , will be explained.

As previously mentioned βt is set to 4.3 and relates to a single component’s safety. If more than one
component engages with other components, it can be considered either a series, parallel system or a
combination of both. For a series system, the corresponding βsys s will be less than the components’

βcomp
t , since the typical prerequisite of a series system is that if a single component fails, then
failure will occur in the entire system. The opposite would most likely be the case for a parallel
p
system, where the system reliability index, βsys , usually is higher than that of the components due
to mechanical behaviour and the requirement that all components fail.

The first step in this analysis is to get an idea of the current reliability level of the systems, both
series and parallel system, before they are proof load tested. This is done with a varying amount of
components and with varying uncertainties on the stochastic models related to the resistance, see
the beginning of Chapter 4. A system consisting of 10, 50 and 100 components will be evaluated
for the following cases:

25
 4. Proof Load Testing With Prior Information

Table 4.5: The cases that will be evaluated including the statistical parameters and uncertainties.

Cases µR VR V(µR ) V(VR )

1a 1.82 10 % 0% 0%
2a 1.82 10 % 15 % 7.5 %
3a 1.82 10 % 20 % 10 %
4a 1.82 10 % 30 % 15 %
1b 2.34 20 % 0% 0%
2b 2.34 20 % 15 % 7.5 %
3b 2.34 20 % 20 % 10 %
4b 2.34 20 % 30 % 15 %

These cases are denoted 1a-4b, where a indicates the first stochastic model for the resistance and b
the other stochastic model, while 1-4 indicates the different uncertainties related to the statistical
parameters.

The established stochastic models are used to estimate the corresponding system reliability level
for the eight cases. For this analysis, the following prerequisites are used in regards to correlation:

• The statistical parameters, µR,i and µR,i , are fully correlated, i.e. all components in a system
have the same statistical parameters.
• Three different cases of correlation between resistances, Ri , of components in a system
will be assessed. This includes an uncorrelated case and a case for which the correlation
coefficient, ρ, is set equal to ρ = 0.5 and ρ = 0.8 respectively.
• The loads, Gi and Qi , are assumed to be fully correlated in the systems, i.e. the load is the
same for all components in the system.

The procedure for estimating the system reliability levels is described in the following three
subsections for series systems and parallel systems respectively.

4.6.1 Series system

Firstly, series systems are considered. The prerequisite for this procedure is that a series system
fails if any of its components fails. The simulation algorithm used to estimate βsys
s is similar to the

procedure above, but with a few modifications.

1. N = 107 set of the stochastic parameters, µR,i and VR,i , related to a batch of components are
Monte-Carlo simulated, including the uncertainties V(µR ) and V(VR ) for each case.
2. N · n random realizations of the resistance, Ri , model uncertainty, XR,i , and loads, Gi and
Qi , are simulated depending on the assumption for the correlation. n is the number of
components in each batch/system. The theory for adding correlation between component
resistances in a system is described in Appendix A.
3. The random simulated variables are inserted into Equation 4.1 and g is calculated for n
components in N batches/systems. Failure in any component leads to failure of the system.
sys
4. The probability of failure, P f , is estimated as the ratio between the amount of failed
batches/systems and the number of simulations, N.  
s , is calculated by βs = −Φ−1 P sys .
5. The reliability index, βsys sys f

26
4.6. Estimation of system reliability Aalborg University

By following the above-mentioned procedure, βsys comp = 4.3 for no uncertainty

s corresponding to β t

can then be estimated. The results are listed in Tables 4.6 - 4.8 for a correlation coefficient of
ρ = 0, ρ = 0.5 and ρ = 0.8 respectively.

Table 4.6: The reliability indices of the series system for the various of cases with varying n and
ρ = 0.

Cases βsys
n=1 βsys
n=10 βsys
n=50 βsys
n=100

1a 4.30 3.83 3.55 3.43

2a 3.57 3.10 2.82 2.71
3a 3.17 2.70 2.44 2.33
4a 2.45 2.05 1.81 1.72
1b 4.30 3.75 3.36 3.18
2b 3.72 3.13 2.72 2.55
3b 3.40 2.79 2.38 2.21
4b 2.78 2.17 1.80 1.65

Table 4.7: The reliability indices of the series system for the various of cases with varying n and
ρ = 0.5.

Cases βsys
n=1 βsys
n=10 βsys
n=50 βsys
n=100

1a 4.30 3.94 3.72 3.63

2a 3.57 3.20 3.00 2.93
3a 3.17 2.81 2.63 2.55
4a 2.45 2.14 1.98 1.92
1b 4.30 3.81 3.50 3.38
2b 3.72 3.23 2.92 2.80
3b 3.40 2.91 2.61 2.50
4b 2.78 2.31 2.06 1.96

Table 4.8: The reliability indices of the series system for the various of cases with varying n and
ρ = 0.8.

Cases βsys
n=1 βsys
n=10 βsys
n=50 βsys
n=100

1a 4.30 4.05 3.90 3.83

2a 3.57 3.32 3.19 3.14
3a 3.17 2.92 2.81 2.76
4a 2.45 2.24 2.15 2.11
1b 4.30 3.94 3.72 3.63
2b 3.72 3.36 3.16 3.09
3b 3.40 3.05 2.86 2.79
4b 2.78 2.47 2.31 2.24

The tables show a significant decrease in βsyss when adding an uncertainty, V(µ ) and V(V ). This
R R
is caused by an increase in the deviation of the resistances, meaning that more values in the lower
tail of the distribution are obtained in the system. Because of the prerequisite for the series system
that failure in a single component leads to total failure of the system, this means that there is a

27
 4. Proof Load Testing With Prior Information

higher probability of failure. Moreover, this prerequisite is also a cause for the decrease in βsys s

for an increasing n in the system. Furthermore, the results show that an increase in correlation
between the resistance of components in a system consisting of more than one component result in
larger β sys , especially for n = 100. This agrees well with what is to be expected of a series system.

The influence of the correlation is also illustrated in Figure 4.4 for case 1a and 1b with V(µR ) = 0
and V(VR ) = 0. In the figure, it should be noted that case 1a for ρ = 0.5 and case 1b for ρ = 0.8
have the same curve by chance.

5
Case 1a for =0
4.8 Case 1b for =0
Case 1a for = 0.5
4.6 Case 1b for = 0.5
Case 1a for = 0.8
4.4 Case 1b for = 0.8
Reliability index,

4.2

3.8

3.6

3.4

3.2

3
0 10 20 30 40 50 60 70 80 90 100
Number of components, n
Figure 4.4: Illustration of βsys
s for V(µ ) and V(V ) as function of n components in a series system.
R R

4.6.2 Parallel system - brittle

Secondly, parallel systems are considered. For this case, the parallel system is assumed brittle,
meaning that each component in the system does not have any extra bearing capacity and fails when
the load reaches the load-bearing capacity, i.e. resistance. This is illustrated by the stress-strain
curve in Figure 4.5.

σy

ε
Figure 4.5: Stress-strain curve for a brittle material.

28
4.6. Estimation of system reliability Aalborg University

When a component fails, the load applied to the failed component is assumed to be uniformly
p
distributed to the rest of the system. The simulation algorithm used for estimating βsys for a
parallel brittle system is as follows:

1. N = 107 set of the stochastic parameters, µR,i and VR,i , related to a batch of components are
Monte-Carlo simulated, including the uncertainties V(µR ) and V(VR ) for each case.
2. N · n random realizations of the resistance, Ri , model uncertainty, XR,i , and loads, Gi and
Qi , are simulated depending on the assumption for the correlation. n is the number of
components in each batch/system. Correlation between the resistance of components in
each system can be varied.
3. The random simulated variables are inserted into Equation 4.1 and g is calculated for n
components in N batches/systems. If any component fails in a system, the applied load is
uniformly distributed to the rest of the components. This process is repeated until no failure
occurs in a component or the system fails.
sys
4. The probability of failure, P f , is estimated as the ratio between the amount of failed
batches/systems and the number of simulations, N.  
p p sys
5. The reliability index, βsys , is calculated by βsys = −Φ−1 P f .
p
The results for βsys for the same cases as the series system are listed in Tables 4.9 - 4.11

Table 4.9: The reliability indices of the brittle parallel system for the various of cases with varying
amount of components in the system and ρ = 0.

Cases βsys
n=1 βsys
n=10 βsys
n=50 βsys
n=100

1a 4.30 3.93 3.85 3.81

2a 3.57 3.17 3.06 3.03
3a 3.17 2.76 2.64 2.61
4a 2.45 2.06 1.95 1.93
1b 4.30 4.37 4.32 4.33
2b 3.72 3.56 3.48 3.45
3b 3.40 3.14 3.04 3.00
4b 2.78 2.40 2.27 2.25

Table 4.10: The reliability indices of the brittle parallel system for the various of cases with
varying amount of components in the system and ρ = 0.5.

Cases βsys
n=1 βsys
n=10 βsys
n=50 βsys
n=100

1a 4.30 3.96 3.86 3.80

2a 3.57 3.22 3.12 3.09
3a 3.17 2.83 2.72 2.70
4a 2.45 2.15 2.05 2.03
1b 4.30 3.99 3.90 3.87
2b 3.72 3.37 3.26 3.24
3b 3.40 3.02 2.92 2.89
4b 2.78 2.49 2.40 2.38

29
 4. Proof Load Testing With Prior Information

Table 4.11: The reliability indices of the brittle parallel system for the various of cases with
varying amount of components in the system and ρ = 0.8.

Cases βsys
n=1 βsys
n=10 βsys
n=50 βsys
n=100

1a 4.30 4.04 3.93 3.91

2a 3.57 3.33 3.23 3.21
3a 3.17 2.93 2.85 2.82
4a 2.45 2.25 2.17 2.16
1b 4.30 3.95 3.83 3.84
2b 3.72 3.39 3.29 3.27
3b 3.40 3.08 2.98 2.96
4b 2.78 2.49 2.40 2.38

p
The results for βsys are slightly higher than for the series system, but the same tendencies are seen,
when varying V(µR ), V(VR ) and changing n in a system. The similarities are due to the parallel
system being considered brittle, which means that even though all components in the system have
to fail for total failure of the system, the redistribution of the full load from the failed components
often leads to a chain reaction of component failures. Thus, system failure might have a high
probability to occur anyway due to an initial failure in a component. The opposite will most likely
be the case for a ductile parallel system, which is assessed in the following. The influence ρ is also
illustrated in Figure 4.6 for the cases with V(µR ) = 0 and V(VR ) = 0.

4.8

4.6

4.4
Reliability index,

4.2

3.8

3.6 Case 1a for =0

Case 1b for =0
3.4 Case 1a for = 0.5
Case 1b for = 0.5
3.2 Case 1a for = 0.8
Case 1b for = 0.8
3
0 10 20 30 40 50 60 70 80 90 100
Number of components, n
p
Figure 4.6: Illustration of βsys for V(µR ) = 0 and V(VR ) = 0 as function of n in a brittle parallel
system.

p
The deviation between the curves for βsys is small compared to the series system. Some of the

30
4.6. Estimation of system reliability Aalborg University

curves crosses each other, which might indicate that there is a minor instability in regards to the
number of simulations. Furthermore, it is seen that case 1b for ρ = 0 deviates from the tendency
p
of the other cases with βsys being almost constant regardless of n. This indicates that a high VR
is beneficial for a brittle parallel system, when there is no correlation between the resistance of
p
components. In general, it is seen that βsys increases for an increasing ρ for VR = 10%, while the
opposite is the case for VR = 20%.

4.6.3 Parallel system - ductile

The second type of parallel system that is considered is a ductile system. It is assumed that the
components in this ductile parallel system behave as ideal plastic materials as illustrated by the
stress-strain curve in Figure 4.7.

σy

ε
Figure 4.7: Stress-strain curve for an ideal plastic material.

When a component in the ductile parallel system is subjected to a load larger than the resistance,
then the component will not fail. Instead the component will carry the load up to the resistance,
while the residual load will be uniformly redistributed to the rest of the components in the parallel
system. This means that the system only fails, if the load surpasses the resistance of all the
components in the system. The simulation algorithm is similar to the one for brittle material
with the exception of step 3 for which only the load surpassing the resistance of a component is
redistributed to the remaining components. This analysis is only carried out for the cases with
no uncertainty, i.e. 1a and 1b for the various ρ. The reason for this is that the ductile system
p
is expected to have a very high βsys , thus it is only necessary to confirm this hypothesis. The
p
results are listed in Table 4.12, 4.13 and 4.14. It should be noted that βsys > 5.20 means that no
components fail in N = 107 simulations.

Table 4.12: The reliability indices of the ductile parallel system for the various of cases with
varying n in the system and ρ = 0.
sys sys sys sys
Cases βn=1 βn=10 βn=50 βn=100
1a 4.30 4.83 >5.20 4.72
1b 4.30 >5.20 >5.20 >5.20

31
 4. Proof Load Testing With Prior Information

Table 4.13: The reliability indices of the ductile parallel system for the various of cases with
varying n in the system and ρ = 0.5.
sys sys sys sys
Cases βn=1 βn=10 βn=50 βn=100
1a 4.30 4.55 4.65 4.58
1b 4.30 4.99 4.86 4.94

Table 4.14: The reliability indices of the ductile parallel system for the various of cases with
varying n in the system and ρ = 0.8.
sys sys sys sys
Cases βn=1 βn=10 βn=50 βn=100
1a 4.30 4.43 4.37 4.46
1b 4.30 4.53 4.56 4.51

The results show that the ductile parallel system in general has a very high reliability level,
especially when having no correlation between the resistance of the components as would have
been expected. The basis reliability levels indicate that a ductile parallel system will have a
satisfying safety and it is likely that a change of function or loading on an existing structure will
not lead to failure. Thus, it will only be necessary to proof load test a single component in this
ductile parallel system with a magnitude equal to the characteristic load requirement to validate
the safety level. This means that the ductile parallel system will not be analyzed in the following
Section 4.7.

4.7 Updated system reliability

The reliability level of a system can be updated by using information gained through proof load
tests on a number of components in a system, see Appendix A.4. The general idea behind proof
load testing is to apply a proof load on e.g. 5 components in a system of 25 identical components
and observe if failure occurs in any component in the system. An observation of no failure means
that βsys can then be updated depending on the magnitude of the proof load. The limit state function
used for proof load testing is expressed as the following:

gtest = RXR − Ppl where Ppl = k test Pk (4.13)

where

Ppl Proof load.

ktest Multiplication factor to the characteristic load requirement.
Pk Characteristic load requirement.

The characteristic load requirement is determined by a weighting between the permanent and
variable load as expressed by:

Pk = G k (1 − α) + Q k α (4.14)

The proof load is applied to a portion of the components in each system. The portion sizes that
are used are 10%, 20%, 30%, 40% and 50% of the total numbers of components rounded to the

32
4.7. Updated system reliability Aalborg University

following whole number respectively. The procedure for assessing the influence of proof load
testing is as following and can be seen visualized in the flowchart in Figure 4.8.

Simulate 𝑁 random realizations of the

1 stochastic parameters, 𝜇 , and 𝑉 ,

Simulate 𝑁 ⋅ 𝑛 realizations of the

2 resistance, 𝑅 , model uncertainty, 𝑋 , ,
and loads, 𝐺 and 𝑄

Insert 𝑅 , 𝑋 , and 𝑃 into the limit state

3 function related to the proof load and
disregard the systems that fail

Insert 𝑅 , 𝑋 , , 𝐺 and 𝑄 into the limit

4 state function related to the real loads for
the successful systems from step 3

Estimated the probability of failure,

5 𝑃 , for the system

Estimated the reliability index, 𝛽 , for

6 the system

If 𝛽 ≤ 𝛽 , step 3-7 is repeated for a

7 𝑘 -factor of 𝑘 =𝑘 + 0.01 until
convergence is met

Step 1-7 is repeated 10 times due to

8 computational limitations and 𝑘 is
found as an average of the 10 results

Figure 4.8: Algorithm for estimating the k test corresponding to a targeted system reliability index,
βsys
t , when using different V(µ ) and V(V ).
R R

1. N = 106 random realizations of the stochastic parameters, µR,i and VR,i related to a system
are Monte-Carlo simulated, including the uncertainties V(µR ) and V(VR ) for each case.
2. N · n random realizations of the resistance, Ri , model uncertainty, XR,i , and loads, Gi and Qi ,
are simulated for which n is the number of components in each batch/system. Correlation
between the resistance of components in each system can be varied.
3. The simulated variables related to the resistance in step 2 and the proof load, Ppl , are placed
into Equation 4.13 and gtest is calculated. gtest < 0 leads to failure and the failed systems are
disregarded for the remaining procedure.

33
 4. Proof Load Testing With Prior Information

4. The real loads, Gi and Qi , are now applied to the remaining systems from step 3 and Equation
4.1 is evaluated. The failure criterion is g < 0.
sys
5. The probability of failure, P f , for a system is estimated as the ratio between the amount of
failed systems and number of total systems that survived the proof load test.
sys
6. The reliability index, β, is calculated by βsys = −Φ−1 (P f ).
7. If βsys < βsys
t , step 3-7 is repeated for a k -factor of k
test test = k test + 0.01 until convergence
is met.
8. Step 1-7 is repeat 10 times due to computational limitations in regards to memory when
saving matrices in step 2 of more than 106 · n in size. The k test -factor is then found as an
average of the 10 simulations.

4.8 Analyses
In this section, a number of analyses will be carried out for assessing the influence of proof load
testing on the reliability level. This includes assessment of k test -factors for the series system and
brittle parallel system as well as sensitivity analyses on selected results.

4.8.1 Proof load test on series system

For the first analysis, the series system will be evaluated in order to determine the k test -factors
for the respective cases. The factors will be estimated for a varying amount of components and
varying portion size of proof load tested components in the system. The number of components
and test portions are described in Section 4.6 and 4.7 respectively.

The analysis will be carried out for all cases with an uncertainty related to the statistical parameters,
µR and VR , i.e. case 2a - 4b listed in Table 4.5. The reason for this is that proof load testing a
portion of components in a system, when V(µR ) = 0 and V(VR ) = 0, does not provide information
about the rest of the components. The k test -factors will be calibrated to have a βsyst corresponding

to V(µR ) = 0 and V(VR ) = 0, i.e. case 1a and 1b in Tables 4.6, 4.7 and 4.8 depending on the
correlation between components in the system. For instance, βsys t = 3.83 is used for a series system

consisting of 10 components in case 2a, 3a and 4a, when estimating the k test -factors.

The procedure for estimating the k test -factors are explained in Section 4.7 and the stochastic models
are shown in Section 4.5. The results are listed in Table 4.15, 4.16 and 4.17 for ρ = 0, ρ = 0.5 and
ρ = 0.8 respectively. The tables are generalized on the basis of the results obtained in Appendix
C for the sake of clarity. Because of the generalization, the results have been made conservative,
but more detailed results can be found in Tables C.1 - C.18.

34
4.8. Analyses Aalborg University

Table 4.15: Test portions and k test -factors for a various number of components, n, for a correlation
of ρ = 0. Values inside the parenthesis () indicate the amount of series systems that
succeeded the proof load testing.
Uncertainty
related to VR ≤ 10% (a-cases) 10% < VR ≤ 20% (b-cases)
µR and VR
10% and k test = 1.9 if n ≥ 10 (≥ 43%) 10% and k test = 2.9 if n ≥ 10 (≥ 18%)
10% and k test = 1.5 if n ≥ 100 (≥ 58%) 10% and k test = 1.7 if n ≥ 100 (≥ 48%)
Small
50% and k test = 1.5 if n ≥ 10 (≥ 67%) 50% and k test = 1.7 if n ≥ 10 (≥ 59%)
50% and k test = 1.3 if n ≥ 100 (≥ 67%) 50% and k test = 1.3 if n ≥ 100 (≥ 71%)
10% and k test = 1.9 if n ≥ 10 (≥ 45%) 10% and k test = 3.0 if n ≥ 10 (≥ 16%)
10% and k test = 1.5 if n ≥ 100 (≥ 57%) 10% and k test = 1.7 if n ≥ 100 (≥ 48%)
Large
50% and k test = 1.5 if n ≥ 10 (≥ 60%) 50% and k test = 1.8 if n ≥ 10 (≥ 51%)
50% and k test = 1.3 if n ≥ 100 (≥ 62%) 50% and k test = 1.3 if n ≥ 100 (≥ 65%)
10% and k test = 1.9 if n ≥ 10 (≥ 38%) 10% and k test = 3.3 if n ≥ 10 (≥ 13%)
10% and k test = 1.5 if n ≥ 100 (≥ 55%) 10% and k test = 1.7 if n ≥ 100 (≥ 46%)
Very large
50% and k test = 1.5 if n ≥ 10 (≥ 60%) 50% and k test = 1.8 if n ≥ 10 (≥ 47%)
50% and k test = 1.3 if n ≥ 100 (≥ 56%) 50% and k test = 1.3 if n ≥ 100 (≥ 58%)

Table 4.16: Test portions and k test -factors for a various number of components, n, for a correlation
of ρ = 0.5. Values inside the parenthesis () indicate the amount of series systems that
succeeded the proof load testing.
Uncertainty
related to VR ≤ 10% (a-cases) 10% < VR ≤ 20% (b-cases)
µR and VR
10% and k test = 1.6 if n ≥ 10 (≥ 71%) 10% and k test = 1.9 if n ≥ 10 (≥ 81%)
10% and k test = 1.4 if n ≥ 100 (≥ 76%) 10% and k test = 1.4 if n ≥ 100 (≥ 88%)
Small
50% and k test = 1.4 if n ≥ 10 (≥ 79%) 50% and k test = 1.4 if n ≥ 10 (≥ 92%)
50% and k test = 1.3 if n ≥ 100 (≥ 79%) 50% and k test = 1.2 if n ≥ 100 (≥ 92%)
10% and k test = 1.7 if n ≥ 10 (≥ 64%) 10% and k test = 1.9 if n ≥ 10 (≥ 72%)
10% and k test = 1.4 if n ≥ 100 (≥ 71%) 10% and k test = 1.4 if n ≥ 100 (≥ 83%)
Large
50% and k test = 1.5 if n ≥ 10 (≥ 72%) 50% and k test = 1.5 if n ≥ 10 (≥ 86%)
50% and k test = 1.3 if n ≥ 100 (≥ 73%) 50% and k test = 1.2 if n ≥ 100 (≥ 88%)
10% and k test = 1.7 if n ≥ 10 (≥ 58%) 10% and k test = 2.1 if n ≥ 10 (≥ 57%)
10% and k test = 1.4 if n ≥ 100 (≥ 63%) 10% and k test = 1.5 if n ≥ 100 (≥ 72%)
Very large
50% and k test = 1.4 if n ≥ 10 (≥ 64%) 50% and k test = 1.5 if n ≥ 10 (≥ 75%)
50% and k test = 1.3 if n ≥ 100 (≥ 65%) 50% and k test = 1.3 if n ≥ 100 (≥ 76%)

35
 4. Proof Load Testing With Prior Information

Table 4.17: Test portions and k test -factors for a various number of components, n, for a correlation
of ρ = 0.8. Values inside the parenthesis () indicate the amount of series systems that
succeeded the proof load testing.
Uncertainty
related to VR ≤ 10% (a-cases) 10% < VR ≤ 20% (b-cases)
µR and VR
10% and k test = 1.5 if n ≥ 10 (≥ 82%) 10% and k test = 1.5 if n ≥ 10 (≥ 95%)
10% and k test = 1.4 if n ≥ 100 (≥ 85%) 10% and k test = 1.3 if n ≥ 100 (≥ 96%)
Small
50% and k test = 1.4 if n ≥ 10 (≥ 85%) 50% and k test = 1.3 if n ≥ 10 (≥ 97%)
50% and k test = 1.4 if n ≥ 100 (≥ 85%) 50% and k test = 1.2 if n ≥ 100 (≥ 97%)
10% and k test = 1.5 if n ≥ 10 (≥ 76%) 10% and k test = 1.6 if n ≥ 10 (≥ 90%)
10% and k test = 1.4 if n ≥ 100 (≥ 79%) 10% and k test = 1.3 if n ≥ 100 (≥ 93%)
Large
50% and k test = 1.4 if n ≥ 10 (≥ 78%) 50% and k test = 1.4 if n ≥ 10 (≥ 94%)
50% and k test = 1.4 if n ≥ 100 (≥ 79%) 50% and k test = 1.2 if n ≥ 100 (≥ 94%)
10% and k test = 1.5 if n ≥ 10 (≥ 67%) 10% and k test = 1.6 if n ≥ 10 (≥ 81%)
10% and k test = 1.4 if n ≥ 100 (≥ 69%) 10% and k test = 1.4 if n ≥ 100 (≥ 84%)
Very large
50% and k test = 1.5 if n ≥ 10 (≥ 68%) 50% and k test = 1.4 if n ≥ 10 (≥ 85%)
50% and k test = 1.4 if n ≥ 100 (≥ 69%) 50% and k test = 1.3 if n ≥ 100 (≥ 86%)

From the results in Table 4.15, 4.16 and 4.17 it can be seen that the k test -factor does not differ much
for different V(µR ) and V(VR ), i.e. from case 2a to 4a and from case 2b to 4b respectively. It could
have been expected that the k test -factors should increase with an increasing V(µR ) and V(VR ), since
Table 4.6 shows a decrease in βsys s pre-proof load testing with an increasing uncertainty. However,

this indicates that more information is gained through proof load testing of a system with a larger
uncertainty.

Furthermore, by comparing a-cases and b-cases it can be seen that a larger VR generally results in
larger k test -factors, especially for a low value of ρ when proof load testing a smaller test portion of
components in a system or when having a system consisting of few components. Nevertheless, the
tendencies between a-cases and b-cases are similar. A large test portion and/or a large amount of
components in a system results in a lower k test -factor. The survival rate of the proof load testing
has the opposite tendencies in comparison to the k test -factors, i.e. increasing survival rate for
larger test portions and amount of components in a system. This is the opposite of what would
be expected, if the same proof load is applied, thus indicating that the decrease in k test -factors
out-weights the negative effect on the survival rate when increasing test portion and number of
components. Furthermore, by comparing the results for different ρ, it can also be seen that an
increase in ρ leads to a decrease in k test -factors, which is to be expected for a series system.

The survival rate of proof load testing is an important factor to take into consideration, when
deciding whether to perform the proof load testing or just reinforce the components instead. This
aspect will be treated in relation to a decision model based on a cost-benefit analysis in Section
4.10.

4.8.2 Proof load test on brittle parallel system

The brittle parallel system is now considered in regards to estimating the k test -factors for the
respective cases. The procedure is similar to that for the series system with the exception of how
failure of the system is modelled. Failure in a brittle parallel system occurs, if all the components

36
4.8. Analyses Aalborg University

fail, under the assumption that the full load applied to a component that fails is evenly redistributed
to the remaining components.

The analysis is similarly to the series system carried out for all cases with an uncertainty V(µR ) , 0
and V(VR ) , 0. The k test -factors are calibrated to have βsys
t for case 1a and 1b listed in Table 4.9,

4.10 and 4.11 depending on ρ. The results are seen in Table 4.18, 4.19 and 4.20 for ρ = 0, ρ = 0.5
and ρ = 0.8 respectively.

Table 4.18: Test portions and ktest -factors for a varying n and ρ = 0. Values inside the parenthesis
() indicate the amount of brittle parallel systems that succeeded the proof load testing.
Uncertainty
related to VR ≤ 10% (a-cases) 10% < VR ≤ 20% (b-cases)
µR and VR
10% and k test = 1.9 if n ≥ 10 (≥ 44%) 10% and k test = 2.8 if n ≥ 10 (≥ 24%)
10% and k test = 1.5 if n ≥ 100 (≥ 59%) 10% and k test = 1.8 if n ≥ 100 (≥ 43%)
Small
50% and k test = 1.5 if n ≥ 10 (≥ 66%) 50% and k test = 1.7 if n ≥ 10 (≥ 64%)
50% and k test = 1.4 if n ≥ 100 (≥ 61%) 50% and k test = 1.4 if n ≥ 100 (≥ 56%)
10% and k test = 1.9 if n ≥ 10 (≥ 45%) 10% and k test = 3.0 if n ≥ 10 (≥ 17%)
10% and k test = 1.5 if n ≥ 100 (≥ 57%) 10% and k test = 1.7 if n ≥ 100 (≥ 45%)
Large
50% and k test = 1.5 if n ≥ 10 (≥ 63%) 50% and k test = 1.7 if n ≥ 10 (≥ 56%)
50% and k test = 1.4 if n ≥ 100 (≥ 59%) 50% and k test = 1.5 if n ≥ 100 (≥ 49%)
10% and k test = 1.9 if n ≥ 10 (≥ 41%) 10% and k test = 3.0 if n ≥ 10 (≥ 18%)
10% and k test = 1.5 if n ≥ 100 (≥ 54%) 10% and k test = 1.7 if n ≥ 100 (≥ 46%)
Very large
50% and k test = 1.5 if n ≥ 10 (≥ 56%) 50% and k test = 1.7 if n ≥ 10 (≥ 54%)
50% and k test = 1.3 if n ≥ 100 (≥ 55%) 50% and k test = 1.4 if n ≥ 100 (≥ 49%)

Table 4.19: Test portions and k test -factors for a varying n and ρ = 0.5. Values inside the parenthesis
() indicate the amount of brittle parallel systems that succeeded the proof load testing.
Uncertainty
related to VR ≤ 10% (a-cases) 10% < VR ≤ 20% (b-cases)
µR and VR
10% and k test = 1.7 if n ≥ 10 (≥ 70%) 10% and k test = 1.8 if n ≥ 10 (≥ 83%)
10% and k test = 1.4 if n ≥ 100 (≥ 78%) 10% and k test = 1.4 if n ≥ 100 (≥ 90%)
Small
50% and k test = 1.5 if n ≥ 10 (≥ 79%) 50% and k test = 1.4 if n ≥ 10 (≥ 93%)
50% and k test = 1.3 if n ≥ 100 (≥ 79%) 50% and k test = 1.2 if n ≥ 100 (≥ 92%)
10% and k test = 1.7 if n ≥ 10 (≥ 65%) 10% and k test = 1.9 if n ≥ 10 (≥ 75%)
10% and k test = 1.4 if n ≥ 100 (≥ 72%) 10% and k test = 1.4 if n ≥ 100 (≥ 85%)
Large
50% and k test = 1.5 if n ≥ 10 (≥ 72%) 50% and k test = 1.4 if n ≥ 10 (≥ 87%)
50% and k test = 1.3 if n ≥ 100 (≥ 73%) 50% and k test = 1.3 if n ≥ 100 (≥ 87%)
10% and k test = 1.6 if n ≥ 10 (≥ 59%) 10% and k test = 2.0 if n ≥ 10 (≥ 61%)
10% and k test = 1.3 if n ≥ 100 (≥ 70%) 10% and k test = 1.4 if n ≥ 100 (≥ 74%)
Very large
50% and k test = 1.4 if n ≥ 10 (≥ 65%) 50% and k test = 1.5 if n ≥ 10 (≥ 78%)
50% and k test = 1.3 if n ≥ 100 (≥ 69%) 50% and k test = 1.3 if n ≥ 100 (≥ 76%)

37
 4. Proof Load Testing With Prior Information

Table 4.20: Test portions and k test -factors for a varying n and ρ = 0.8. Values inside the parenthesis
() indicate the amount of brittle parallel systems that succeeded the proof load testing.
Uncertainty
related to VR ≤ 10% (a-cases) 10% < VR ≤ 20% (b-cases)
µR and VR
10% and k test = 1.5 if n ≥ 10 (≥ 82%) 10% and k test = 1.5 if n ≥ 10 (≥ 96%)
10% and k test = 1.4 if n ≥ 100 (≥ 88%) 10% and k test = 1.2 if n ≥ 100 (≥ 99%)
Small
50% and k test = 1.4 if n ≥ 10 (≥ 84%) 50% and k test = 1.3 if n ≥ 10 (≥ 97%)
50% and k test = 1.3 if n ≥ 100 (≥ 88%) 50% and k test = 1.1 if n ≥ 100 (≥ 99%)
10% and k test = 1.6 if n ≥ 10 (≥ 74%) 10% and k test = 1.6 if n ≥ 10 (≥ 91%)
10% and k test = 1.4 if n ≥ 100 (≥ 80%) 10% and k test = 1.2 if n ≥ 100 (≥ 96%)
Large
50% and k test = 1.4 if n ≥ 10 (≥ 78%) 50% and k test = 1.4 if n ≥ 10 (≥ 94%)
50% and k test = 1.3 if n ≥ 100 (≥ 81%) 50% and k test = 1.2 if n ≥ 100 (≥ 96%)
10% and k test = 1.6 if n ≥ 10 (≥ 66%) 10% and k test = 1.6 if n ≥ 10 (≥ 82%)
10% and k test = 1.4 if n ≥ 100 (≥ 72%) 10% and k test = 1.3 if n ≥ 100 (≥ 88%)
Very large
50% and k test = 1.4 if n ≥ 10 (≥ 69%) 50% and k test = 1.4 if n ≥ 10 (≥ 86%)
50% and k test = 1.3 if n ≥ 100 (≥ 71%) 50% and k test = 1.2 if n ≥ 100 (≥ 90%)

The same tendencies as for the series system are seen for the brittle parallel system. However, the
brittle parallel system results generally in slightly lower k test -factors compared to the series system
due to the assumption that all components in a system have to fail for total failure to occur. The
survival rate is almost identical. Thus, it can be concluded that there is only a small difference
between the brittle parallel system and the series system. The reason for this could be that the
failure in a component in the parallel system might start a chain-reaction of components that fail
due to redistribution of the load.

4.9 Sensitivity analysis

This section will evaluate the sensitivity of various important parameters used for obtaining the
results in Section 4.8. This includes a sensitivity analysis of the mean value, µR , and the weighting
factor, α. The sensitivity analyses will only be performed for case 2a and 2b in the uncorrelated
cases, ρ = 0, since analyzing a selection of the results in regards to sensitivity is sufficient for
assessing the influence of the parameters.

4.9.1 Influence of µ R
Firstly, the influence of µR is evaluated. The reason for this is that µR is the uncertain parameter,
when dealing with an existing structure, and thus a parameter of importance. The results in Section
4.8 are obtained on the basis of the assumption that µR in the stochastic models can be estimated
by having it correspond to βcomp t = 4.3. This implies that the structure initially has a satisfying
resistance, which might not have been the case. Therefore, it is of interest to evaluate the influence
of the results for k test -factors, when decreasing and increasing µR . This sensitivity analysis will be
performed for a 10% decrease and increase of µR . The results are shown in Table 4.21 for which
the k test -factors corresponds to βsys in Table 4.6 for case 1a and 1b respectively, i.e. the k test -factors
for 0.9µ, 1.0µ and 1.1µ correspond to the same βsys .

38
4.9. Sensitivity analysis Aalborg University

Table 4.21: Comparison between k test -factors and survival rate when having different mean values,
µR .
Scaling of
the mean VR ≤ 10% (a-cases) 10% < VR ≤ 20% (b-cases)
value, µR
10% and k test = 2.1 if n ≥ 10 (≥ 11%) 10% and k test = 3.5 if n ≥ 10 (≥ 2%)
10% and k test = 1.6 if n ≥ 100 (≥ 25%) 10% and k test = 1.8 if n ≥ 100 (≥ 16%)
0.9 · µR
50% and k test = 1.6 if n ≥ 10 (≥ 31%) 50% and k test = 2.0 if n ≥ 10 (≥ 19%)
50% and k test = 1.4 if n ≥ 100 (≥ 32%) 50% and k test = 1.4 if n ≥ 100 (≥ 33%)
10% and k test = 1.9 if n ≥ 10 (≥ 43%) 10 % and k test = 2.9 if n ≥ 10 (≥ 18%)
10% and k test = 1.5 if n ≥ 100 (≥ 58%) 10 % and k test = 1.7 if n ≥ 100 (≥ 48%)
1.0 · µR
50% and k test = 1.5 if n ≥ 10 (≥ 67%) 50% and k test = 1.7 if n ≥ 10 (≥ 59%)
50% and k test = 1.3 if n ≥ 100 (≥ 67%) 50% and k test = 1.3 if n ≥ 100 (≥ 71%)
10% and k test = 1.7 if n ≥ 10 (≥ 85%) 10% and k test = 2.2 if n ≥ 10 (≥ 72%)
10% and k test = 1.4 if n ≥ 100 (≥ 89%) 10% and k test = 1.5 if n ≥ 100 (≥ 86%)
1.1 · µR
50% and k test = 1.4 if n ≥ 10 (≥ 93%) 50% and k test = 1.5 if n ≥ 10 (≥ 92%)
50% and k test = 1.2 if n ≥ 100 (≥ 94%) 50% and k test = 1.2 if n ≥ 100 (≥ 95%)

The sensitivity analysis shows a small decrease in k test -factors for 0.9µR and opposite a small
increase k test -factors for 1.1µR for obtaining the same reliability level, when having VR ≤ 10%.
For 10% < VR ≤ 20% a larger decrease and increase is seen when testing a small portion of
components in a system consisting of few components. As a result, the survival rate when having
0.9µ is very low, especially for a large VR . The opposite is seen for 1.1µ, where the survival rate
is very high and satisfying. Thus, the results are sensitive to a variation of µR .

Instead of estimating the k test -factors corresponding to the same reliability level when having
different µR it might also be interesting to evaluate a case in which βsys
s are estimated for different

µR when using the k test -factors for 1.0µR . This corresponds to a situation, where µR is thought
to be 1.0µR , but in reality is either lower or higher, meaning that the system will have a different
reliability level than βsys
t . These βs for different µ are shown in Table 4.22 and 4.24.
sys R

Table 4.22: Reliability indices for 0.9µR .

Cases βsys
n=10 βsys
n=50 βsys
n=100

2a 3.70 3.48 3.38

2b 3.53 3.21 3.05

Table 4.23: Reliability indices for 1.0µR .

Cases βsys
n=10 βsys
n=50 βsys
n=100

2a 3.83 3.55 3.43

2b 3.75 3.36 3.18

39
 4. Proof Load Testing With Prior Information

Table 4.24: Reliability indices for 1.1µR .

Cases βsys
n=10 βsys
n=50 βsys
n=100

2a 4.03 3.77 3.65

2b 3.97 3.57 3.40

The results show that the reliability level becomes lower for a smaller µR and opposite, higher for
a larger µR , which is what would have been expected. Furthermore, it can be seen that an increase
in µR has a bigger influence on the reliability level than a decrease in µR , especially for case 2a
with the smaller VR .

4.9.2 Influence of α
The weighting factor, α, is another parameter that might have a significant influence on the k test -
factors for obtaining βsys
t in Table 4.6. Therefore it is interesting to analyze the sensitivity of α, since

it describes the weight between the two load variables that have very different stochastic models.
All previous analyses have been performed based on the assumption that α = 0.5. However, this is
a rough assumption that is not always the case as α = 0.5 was initially used as an average assessed
value. This sensitivity analysis will be performed for α = 0.3 and α = 0.8, i.e. a high weighting
on the permanent and variable load respectively. The results are listed in Table 4.25.

Table 4.25: Comparison between k test -factors and survival rate when having different weighting
factors, α.
Weighting
VR ≤ 10% (a-cases) 10% < VR ≤ 20% (b-cases)
factor, α
10% and k test = 1.8 if n ≥ 10 (≥ 56%) 10% and k test = 3.2 if n ≥ 10 (≥ 10%)
10% and k test = 1.4 if n ≥ 100 (≥ 67%) 10% and k test = 1.8 if n ≥ 100 (≥ 39%)
0.3
50% and k test = 1.4 if n ≥ 10 (≥ 79%) 50% and k test = 1.9 if n ≥ 10 (≥ 47%)
50% and k test = 1.3 if n ≥ 100 (≥ 75%) 50% and k test = 1.4 if n ≥ 100 (≥ 62%)
10% and k test = 1.9 if n ≥ 10 (≥ 43%) 10% and k test = 2.9 if n ≥ 10 (≥ 18%)
10% and k test = 1.5 if n ≥ 100 (≥ 58%) 10% and k test = 1.7 if n ≥ 100 (≥ 48%)
0.5
50% and k test = 1.5 if n ≥ 10 (≥ 67%) 50% and k test = 1.7 if n ≥ 10 (≥ 59%)
50% and k test = 1.3 if n ≥ 100 (≥ 67%) 50% and k test = 1.3 if n ≥ 100 (≥ 71%)
10% and k test = 2.4 if n ≥ 10 (≥ 8%) 10% and k test = 3.4 if n ≥ 10 (≥ 6%)
10% and k test = 1.7 if n ≥ 100 (≥ 25%) 10% and k test = 1.9 if n ≥ 100 (≥ 30%)
0.8
50% and k test = 1.8 if n ≥ 10 (≥ 22%) 50% and k test = 2.0 if n ≥ 10 (≥ 32%)
50% and k test = 1.5 if n ≥ 100 (≥ 32%) 50% and k test = 1.5 if n ≥ 100 (≥ 49%)

Table 4.25 show that for the k test -factors becomes smaller for α = 0.3 in the a-cases and opposite
larger for α = 0.8. A different tendency is seen for the b-cases for which both α = 0.3 and
α = 0.8 result in larger k test -factors. This indicates that the relation between α and the k test -factors
is non-linear with an optimum/minimum point between the two α-values. Though, in general it
is seen that α = 0.8 results in larger k test -factors, which means that a large weight on the variable
load is more critical. This is caused by the large V on the variable load despite the permanent load
having a higher µ in the stochastic models described in Section 4.5. Thus, VR of the stochastic
load has a significant influence on the results.

40
4.10. Decision model Aalborg University

4.10 Decision model

The decision for whether to perform proof load testing or not usually comes down to an economical
point of view, although the practical aspect also is of importance, since it might not always be
possible to perform proof load tests. A decision model can be established to provide an overview
of different decisions and actions that can be taken. The decision model is illustrated in Figure
4.9 and is based on the prerequisite that the endpoint of each branch in the model corresponds to
a fixed target reliability level for a component. In the analyses in this report, βcomp
t = 4.3 has been
used. Furthermore, the cost related to collapse of the structure for βcomp = 4.3 is assumed to be
t

very small and thus negligible, and the proof load tests are assumed to be carried out such that
failure in a component does not lead to collapse of structure.

The decision model constitutes of an initial decision to replace all components in the system with
new components that corresponds to a desired βt or proof load test a number of components in the
system. Replacement of components is usually an expensive solution, why the decision to proof
load test might be the preferable choice of action. The decision to perform proof load testing leads
to an event, where failure is observed or not. If not, then the system has the required βcomp
t given
the prerequisites about the resistance hold true. If failure is observed, then the structure does not
have the required βcomp
t . The question then becomes whether to replace the components or to
continue performing tests. If it is decided to continue testing, then the failed component firstly
has to be replaced. An option is to repeat the proof load testing, but this requires that the analysis
in this chapter is repeated with the prerequisite that the simulation in Section 4.7 only disregards
system for which more than one component fail. If failure is observed again, this process can be
repeated again and so on. This is not investigated in this report and therefore it is assumed that
failure in one component results in all components having to be replaced. This is a strict rule, but
the analyses in this report can not account for these situations. Another option is to proof load test
a few components to failure in order to obtain information about the resistance. This can be used
in addition to the information already obtained from the previously failed component and hereby
estimate how much the components must be reinforced. If the components are made of concrete,
other test methods like cylinder and CAPO-tests can be used for this estimation. The analysis

Figure 4.9: Decision model for reassessment of the reliability level for an existing structure through
proof load testing with prior information. The endpoint of a branch corresponds to a
fixed target reliability level, βcomp
t = 4.3.

41
 4. Proof Load Testing With Prior Information

then becomes similar to that described in Chapter 6 that studies how the reliability level can be
updated by observation of the stochastic variable that is the resistance. However, this is not further
investigated in this report.

An analysis of the cost related to each decision will be performed to provide a basis for choosing
between the different solutions for updating the reliability level of an existing structure. The
comparison will be between the decision to replace the components in a system immediately or to
proof load test a portion of the components based on the average cost of these. The analysis in this
report will not take basis in the correct costs of the different options for updating the reliability
level, since they are difficult to obtain given that they depend on a lot of parameters such as the
present market values, the man hours used for performing the solutions etc. Instead a generic
analysis will be performed for which a ratio is specified between the cost of proof load testing and
replacement of components in order to illustrate the principle behind the decision problem. The
cost of collapse for the fixed βcomp
t = 4.3 is assumed to be negligible and thus not listed. The
following costs are chosen for the analysis, where E indicates a unit cost and each cost is for testing
and replacing of a single component.

• Proof load testing, Cpl : 1 E per component

• Replacement of components, Cre : 10 E per component

Proof load testing a single component in a system corresponds to a unit cost, E, while replacement
of components is 10 times the unit cost per component. The replacement of components has been
chosen to have a significantly larger cost than proof load testing, since that is to be expected in
reality. This assumption will probably result in the decision to perform proof load test usually
being less expensive.

Example
The specified cost and results for the survival rate of systems obtained in Section 4.8 can be used in
the decision model in Figure 4.9 to determine the optimal. An example of this is shown in Figure
4.10 for a series system consisting of 100 components for which 10% of the components are proof
load tested. The survival rates of the test is taken for a case with a small uncertainty related to µR
and VR for VR = 10% and ρ = 0 as seen in Table 4.15, which is 58%. The last part of the decision
model is not included in the analysis of cost, since the analysis required for the continuation of
testing has not been investigated in this report. Thus, observation of failure during the initial proof
load testing results in all the components having to be replaced.

42
4.10. Decision model Aalborg University

Figure 4.10: Example of a decision tree for reevaluating the reliability level of an existing structure.

The decision between replacing the components in the system or performing proof load testing
from an economical point of view depends on the average cost of the options. These are calculated
in the following:

Replacing the components: 1000 E

Proof load testing: 0.58 · 10 E + 0.42 · 1010 E = 426 E

For this particular case, proof load testing has a significantly less average cost for updating the
reliability level. Based on the survival rates, the ratio between the cost of the two options for when
the average cost is equal to each other can be estimated to Cpl /Cre = 5.8. This means that the proof
load tests have to be significantly more expensive than replacement of the components.
This analysis should be repeated for different test portions of component and compared to determine
the optimal solution. The survival rates for a test portion of 20%, 30%, 40% and 50% are listed in
Table C.1 in Appendix C for case 2a with ρ = 0. By repeating the above procedure, the following
costs in Table 4.26 are estimated.
Table 4.26: Comparison between cost of option A and B for different test portions for case 2a with
a coefficient correlation of ρ = 0.

Replacing Proof load testing

Test portion k test Survival rate
components (replacement)
10% 1.46 58% 1000 E 426 E
20% 1.39 61% 1000 E 402 E
30% 1.35 64% 1000 E 379 E
40% 1.32 65% 1000 E 376 E
50% 1.30 67% 1000 E 364 E

In this case, the optimal solution in regards to the cost is to proof load test 50% of the 100
components. The results show that performing proof load tests in general becomes less expensive,
if more components are tested, which is due to the survival rate increasing with the size of test
portion and the fact that the specified cost for proof load tests in comparison to replacement of
components is small. The decreasing cost could also mean that it is beneficial to proof load test
more than 50% of the components, but this has not been investigated. The increase in survival

43
 4. Proof Load Testing With Prior Information

rate is caused by the decrease in k test -factors that are used for the proof load. If ρ is increased,
the survival rate will tend to become more constant for all test portions, meaning that a lower test
portion will probably be the optimal solution. Thus, it is important to assess different test portions
of components.

In reality, an assessment of the resistance has to be made through a preliminary inspection and
rating, and if this assessment fits the prerequisites that are assumed about the resistance in this
chapter, then the optimal solution for updating the reliability level can be estimated by using the
results obtained in this chapter and Appendix C, given that the correct costs are known.

44
Chapter

5
Design Based on Sampling Through
Testing

This chapter describes how to perform cylinder compression tests and CAPO-tests. Furthermore,
the model uncertainty related to CAPO-tests is assessed in regards to various of calculation models,
which will be used in a cost-benefit analysis between the two test methods in Chapter 6.

When having to reevaluate the reliability level of existing structures, it may be very important
to obtain new or additional information about the resistance of the structure. This information
can be obtained through testing of components in regards to measuring the strength and updating
the stochastic variables through Bayesian updating of variables, see Appendix A.4. In general,
two types of tests are distinguished between, which are destructive and non-destructive tests. The
cylinder compression test is a destructive test method, while CAPO and LOK-tests are examples of
non-destructive test methods, despite being partly destructive methods. A more detailed description
of the test methods will be described in the following sections. LOK-testing, which is very similar
to CAPO-testing, will not be investigated in this study.

5.1 Cylinder test

A cylinder test is a commonly used destructive method for concrete structures and lightweight
concrete structures for directly measuring the compressive strength of the structure. This test
method is used for both existing structures and new structures still in the design phase. When
performing destructive tests, they have to be performed in accordance with the relevant codes
or other relevant documented instructions. Moreover, when extracting test samples, it has to be
ensured that it does not reduce the functionality of the structure or if so that the functionality
can be reestablished. Similarly, the consequences of potential damage on the structure as well
as possible structural failures during testing have to be assessed, before performing a destructive
test. [SBI 251, 2015] For existing structures, cores are bored out directly from the structure with
a typical dimension of 150 × 300 mm for diameter and height respectively. It is recommended to
bore a large number of core samples in different locations on the structure in order to evaluate the
compressive strength of the concrete with a high precision. When boring out the cores as well as
transporting them to the laboratory, it has to be ensured that the structures and cores receive minor
damage. [FPrimeC Solutions Inc., 2016] The extracted cylinder of the structure is then tested in a
laboratory for estimating the characteristic values of the resistance in accordance with Annex D7.2
in DS/EN 1990 [2007].

45
 5. Design Based on Sampling Through Testing

5.2 CAPO-test
The CAPO-test is a partly destructive testing method that can be performed on existing structures
without having pre-installed inserts in the specimen. The test is regarded as a non-destructive test,
since the test is intended to be carried out with almost no consequence in terms of damage and
cost. The test makes use of a pullout system as illustrated in Figure 5.1 for indirectly estimating
the compressive strength of concrete. It is a fast and cheap alternative to core testing and does not
require much space, but yields a larger uncertainty.

18.4 mm 25 mm
25 mm

55 mm
Figure 5.1: Cross-sectional view of a CAPO-test. [Germann Instruments, 2018]

The procedure for performing a CAPO-test is as follows [Germann Instruments, 2018]:

• The pullout system is placed on the specimen with an inner diameter of 55 mm between the
reinforcement bars at the desired location. When doing this, it has to be ensured that the
reinforcement bars are not within the failure region of the concrete.
• A 65 mm deep hole is cored with a diameter of 18.4 mm perpendicular to the surface of the
specimen.
• The surface is levelled to a depth of approximately 3 mm, since an irregular surface will
provide invalid results.
• A recess with a diameter of 25 mm is drilled at a depth of 25 mm.
• A split ring is expanded and placed in the recess, while a counter pressure ring is placed
between the reinforcement bars on the surface.
• The expanded ring is pulled out by applying a tension force. When performing the test, the
concrete in between the expanded ring and counter pressure ring will be in compression.
Thus, the pullout force is directly related to the compressive strength of the concrete.

5.2.1 Theoretical models

It is important to note that the CAPO-test result does not equal the compressive strength of concrete,
meaning that a correlation between the pullout force and compressive strength of the concrete has
to be established for the CAPO-test to be applicable. Various of models for correlating these
have been developed during recent decades through a number of studies. These models will be
presented in the following with the purpose of carrying out a statistical assessment of the models
in relation to experimental data.

46
5.2. CAPO-test Aalborg University

A commonly used method takes its basis in Equation (5.1), which relates the pullout force to the
compressive strength of 150 × 300 mm cylinders. [Germann Instruments, 2018]

fcyl = 0.69F 1.12 (5.1)

where

fcyl Compressive strength of cylinder [MPa]

F Maximum pullout force [kN]

Another common model is the one found in BS EN 13791 [2007] where the compressive strength,
fc,E N , is expressed as following:

fc,E N = 1.33(F − 10) 10kN ≤ F ≤ 60kN (5.2)

The last model that will be presented is the best linear fit to the experimental data, which will have
the following appearance:

fc = aF + b (5.3)

where

a Parameter estimated by linear least squares method, a = 1.18 [MPa/kN]

b Parameter estimated by linear least squares method, b = −3.40 [MPa]

5.2.2 Experimental data

The experimental data which will be used in regards to the statistical assessment of the theoretical
models is based on the results in Appendix E and is shown in Figure 5.2. Each point in the data is
an average of 2-9 cylinders and 4-24 CAPO-test performed on the vertical faces of accompanying
200 mm cubes.

Figure 5.2: Experimental data

47
 5. Design Based on Sampling Through Testing

5.2.3 Statistical assessment

The procedure used to perform the statistical assessment is based on the procedure found in DS/EN
1990 [2007] Annex D - Design assisted by testing to estimate the mean, µ, and the coefficient of
variation, V, of the model uncertainty. The procedure will be presented in the following.

The model is rewritten as:

Y = f (X)  b∆h(X) (5.4)

where

b Bias corresponding to the mean value of Fc

∆ ∆ modelled by a lognormal distributed stochastic variable with mean 1 and standard deviation σ∆
h(X Theoretical model, see Equations (5.1) - (5.3)

It is assumed that the parameters Fi, ..., Fn are included directly in the model h(X).

Using the experimental results in Appendix E and assuming that the test results are statistically
independent, then the bias, b, is estimated using a linear least squares method:
ÍN
yi h(xi )
b = Íi=1
N
(5.5)
i=1 h(xi )

where

yi Experimental value
h(xi ) Theoretical value

Realizations of the lognormally distributed variable Fc for each test is obtained from:
 
yi
∆i = ln (5.6)
bh(xi )

The standard deviation, σ∆ , is estimated as:

v
u
t N
1 Õ
s∆ = ∆i − ∆¯


(5.7)
N − 1 i=1

where ∆¯ is the mean:

N
¯∆ = 1
Õ
∆i (5.8)
N i=1

The corresponding coefficient of variation of the model uncertainty is found from:

q
V∆ = exp s∆2 − 1


(5.9)

By following this procedure the model uncertainty related to the various of theoretical models can
be estimated. In Figure 5.3 the different calculation models are plotted with respect to the data.

48
5.2. CAPO-test Aalborg University

80

Compressive strength [MPa]

60

40

Measured
20
Best Linear fit
Power function
EN13791
0
0 10 20 30 40 50 60 70
Pull-force [kN]
Figure 5.3: Calculation models versus experimental data

5.2.4 Results
By using the procedure described above for each of the three theoretical models, see Equation
5.1-5.3, the results shown in Table 5.1 are obtained.

Table 5.1: Statistical parameters of the model uncertainty for the various of theoretical models.

Model Bias (Mean value), µ Coefficient of variation, V

Best linear fit 1.00 0.29
Power function 1.01 0.08
EN13791:2007 1.07 0.26

What can be said about the bias in the table above for the various of calculation models, except
the best linear fit, is that they generally produce a conservative result, which can also be seen on
Figure 5.3. The best linear fit’s bias is expected to be 1, since the model is calibrated from a least
squares method of the data. In relation to the cost-benefit analysis that will be carried out in the
next section, the model with the smallest V is chosen, which in this case is the power function.

49
Chapter

6
Cost-Benefit Analysis Between
Cylinder and CAPO-tests

In this chapter the purpose is to perform a cost-benefit analysis of cylinder and CAPO-tests and
compare these in different situations. The cost-benefit is part of a decision model for choosing
the optimal solution in regards to cost when having to assess the reliability level of an existing
structure. This chapter makes use of results obtained in Chapter 5.

For existing structures, cylinder and CAPO tests can be performed to obtain information about the
resistance of the components in the structure. This information can be used in combination with
prior or without prior information to estimate the stochastic variables. Based on the information
obtained, the reliability level of the structure can then be assessed and updated if necessary by
reinforcing the component to an acceptable reliability level. The decision whether to perform one
test or the other can be chosen based on an economical point of view, where a cost-benefit analysis
is carried out for the different tests. A cost-benefit analysis is a tool used for calculating the total
expected cost of carrying out a chosen strategy. The cost-benefit analysis will depend on the type
of test, amount of performed tests and the information gained from performing the individual test.
The results can then be used in a decision model to determine the optimal solution in regards to
cost.

The planning of tests should follow the general guidelines/requirements described in Annex D,
DS/EN 1990 [2007]. This includes prerequisites, that requires the data to be representative,
statistically homogeneously and statistically independent.

6.1 Assessment model: Preliminary inspection and rating

An assessment model related to a preliminary inspection and rating of the resistance of an existing
structure is established. The reason behind this assessment model is to give a qualitative guess
about the resistance, since the decision whether to perform cylinder or CAPO tests based on a
cost-benefit analysis will depend on the actual resistance of the structure. When prior information
is available, this rating of the resistance will be less uncertain than a case for which no prior
information is available. Depending on the rating, different actions can be taken and these are
illustrated in the assessment model in Figure 6.1 similar to that in Chapter 4.

51
 6. Cost-Benefit Analysis Between Cylinder and CAPO-tests

Figure 6.1: Decision model that takes basis in a preliminary inspection and rating of the resistance
and describes the different decisions that have to be made, when the reliability level
of an existing structure has to be reevaluated.

Assessment (A) is a rating of the resistance of the structure, either based on prior information or no
prior information. Cylinder tests might be better for a given resistance and CAPO tests for another,
but without giving a qualitative rating of the resistance, the decision then becomes completely
random. However, the cost-benefit analysis that will be carried out in this chapter might prove that
one of the test types in general is less expensive, since the analysis is made for several different
cases. Assessment (B) is an assessment of an uncertainty related to the rating of the resistance.
Based on the rating of resistance and uncertainty, decision (A) is whether to perform cylinder or
CAPO tests based on a cost-benefit analysis.

The prerequisites about the resistance and uncertainty that are used for the cost-benefit analysis in
this chapter are described in Section 6.2.

Case study
The following case study can be considered for when cylinder and CAPO tests can be used for
assessment of the reliability level of an existing structure. The reliability level of a concrete structure
consisting of a number of components from a homogeneous population has to be reevaluated due
to an increase in loads. This could be due to a change of function of the structure, for instance an
office building has to be changed into a fitness center, meaning new, heavier interior and perhaps
more screed is added.

There are two different scenarios to consider if it is not immediately chosen to replace the
load-bearing components, which is an expensive solution and can be difficult to perform, if
the components act as a series system. Either the tests are performed and it is discovered that the
characteristic strength of the concrete, Rk is satisfying or it is discovered that the components have
to be reinforced or replaced. If Rk is satisfying, it means no further action has to be taken.

6.2 Prerequisites
The aim of this chapter is to perform a cost-benefit analysis considering the two testing methods
that can be used in relation to a decision model, when assuming no prior information is available,
hence having a diffuse prior. In regards to the uncertainty related to the resistance, the uncertainty is
assumed classified in the following levels assuming that all parameters are lognormally distributed:

• Small for V(µR ) ≤ 15% and V(VR ) ≤ 7.5%

• Large for 15% < V(µR ) ≤ 20% and 7.5% < V(µR ) ≤ 10%
• Very large for 20% < V(µR ) ≤ 30% and 10% < V(µR ) ≤ 15%

Furthermore, it is assumed that there cannot be performed more tests than the amount of
components in a system. The mechanical function of the system (series or parallel system) is

52
6.3. Reliability level Aalborg University

not included in this study. A limit state equation with concrete compression strength is considered,
and the stochastic models are described in Section 6.5.

6.3 Reliability level

The target reliability level used for this analysis is the same as described in Section 4.3, i.e.
βt = 4.3.

6.4 Limit state equation

The limit state equation used for this analysis is the same as the one described in Section 4.4.

6.5 Stochastic models

Loads
The stochastic load model is the same as the one described in Section 4.5.

Resistance and model uncertainties

The stochastic resistance model is specifically for in-situ concrete in compression and is shown in
Table 6.1 Sørensen [2009]. The characteristic value, Rk is estimated for the 5%-quantile of the
total resistance, RXR Sørensen [2009]. The model uncertainty related to the CAPO-test is the one
chosen from Section 5.2.4 for the power function because of its small V. This model uncertainty
is assumed representative for all CAPO tests in this analysis. The additional model uncertainty
related to the conversion to concrete strength is taken into account when simulating realizations
from a CAPO-test.

Table 6.1: Stochastic models of the resistances and model uncertainties for in-situ concrete in
compression.

Parameter Symbol Distribution type V Rk

Resistance R Lognormal 0.14 5 %-quantile
Model uncertainty, cylinder XR Lognormal 0.11
Model uncertainty, CAPO XR,CAPO Lognormal 0.08

The mean, µR , and Rk of R will be estimated corresponding to βcomp

t = 4.3. The mean values
are estimated by a Monte-Carlo simulation for which the theory can be found in Appendix A. The
result for µR is listed in Table 6.2.

Table 6.2: Mean values of the resistance.

Parameter Symbol Mean value, µR

Resistance R 2.15

53
 6. Cost-Benefit Analysis Between Cylinder and CAPO-tests

When µR for the stochastic model have been estimated, Rk can then be estimated. The resistance and
model uncertainty are lognormally distributed, meaning that their product is likewise lognormally
distributed. Therefore, the same procedure as described in Section 4.5.2 is applied.

The results for Rk as well as the final stochastic model for the resistance and model uncertainties
are shown in Table 6.3.

Table 6.3: Statistical parameters for the stochastic model of the resistance and model uncertainties.

Parameter Symbol Distribution type µ V Characteristic value

Resistance R Lognormal 2.15 0.14 1.52
Model uncertainty, cylinder XR Lognormal 1.00 0.11
Model uncertainty, CAPO XR,C APO Lognormal 1.01 0.08

It is important to keep in mind that this obtained characteristic strength is the minimum value
needed to obtain a satisfying reliability level of the component of βt = 4.3.

6.6 Updating of the characteristic strength

When no prior information is available and the characteristic strength, Rk , of a component has to
be estimated through tests, then Rk can be estimated from Equation (6.1), which corresponds to
having an unknown coefficient of variation of the resistance, VR . [DS/EN 1990, 2007]

Rk = X̄ − k n S (6.1)

where

X̄ Mean value of test results

kn Factor that accounts for statistical uncertainty.
S Standard deviation of test results

The mean value and standard deviation of the resistance are estimated from realizations, x̂ on a
number of components by Equation (6.2) and (6.3):
n
1Õ
X̄ = x̂i (6.2)
n i=1
n
1 Õ
S2 = ( x̂ − X̄)2 (6.3)
n − 1 i=1

The realization, x̂, are from the total distribution RXR or RXR XCAPO depending on the test type
that is being evaluated. The k n -factors are found in Table 6.4 and takes into account the statistical
uncertainty due to a limited number of tests.

Table 6.4: Values of k n for the 5% characteristic value. [DS/EN 1990, 2007]

n 1 2 3 4 5 6 8 10 20 30 ∞
VX unknown - - 3.37 2.63 2.33 2.18 2.00 1.92 1.76 1.73 1.64

54
6.7. Simulation algorithm Aalborg University

6.7 Simulation algorithm

A simulation algorithm that takes basis in the stochastic models and equation for estimating the
characteristic value from test results will now be explained. The simulation algorithm is used to
produce results of the characteristic value of components in a system based on a number of cylinder
or CAPO-tests. These results will then be used in a cost-benefit analysis to determine the optimal
number of tests to perform in regards to total cost in different cases. The procedure is as following:

1. A sensitivity analysis with respect to the assumed value of µR is investigated, why the
following values of µR will be considered: µR = 0.6µR , µR = 0.8µR and µR = 1.0µR
respectively.
2. Simulate N random realizations of the stochastic parameters, µR,i and VRi , for each value of
µR,i depending on the chosen level of uncertainty, V(µR ) and V(VR ).
3. Simulate N · n realizations from the total distribution RXR or RXR XCAPO depending on the
test type that is being evaluated. The realizations in each system are correlated by having
the same distribution function.
4. Estimate the mean, X̄ and standard deviation, S, of each test series
5. Each of the test series characteristic strengths are calculated from Equation (6.1).
6. The estimated characteristic strengths are compared to the necessary one estimated earlier
to be Rk,necessary = 1.52
7. The possibility of obtaining an characteristic value smaller than the necessary is estimated
for every value of µR,i for various amount of tests and chosen uncertainty levels. This
possibility is denoted PR .
8. The average amount of reinforcement needed is expressed as the average, ratio between
Rk,necessary / R¯k . This is the value, z, used in regards to the cost-benefit analysis.
9. The average amount of reinforcement needed when having to reinforce is expressed as the
average ratio between Rk,necessary / R¯k for Rk,necessary > Rk . This is the value, zT , used in
regards to calculating the expected cost for the alternatives in a decision model, see Figure
6.2. The ’T’ indicates that the values are truncated.

The results for the characteristic values and average amount of reinforcement needed to obtained
the βcomp
t is used in a cost-benefit analysis. This cost-benefit analysis is part of the decision model
illustrated in Figure 6.2. The decision model is based on the prerequisites that the endpoint of
each branch corresponds to a fixed βcompt and that cylinder and CAPO-tests are always performed
without damaging the structure, meaning that collapse is not considered a risk when performing
the tests.

55
 6. Cost-Benefit Analysis Between Cylinder and CAPO-tests

Figure 6.2: Decision model for reassessment of the reliability level for an existing structure by
performing cylinder and CAPO-tests. The branches corresponds to a fixed βcompt .

The decision model constitutes of an initial decision to immediately replace components in a

system with new components that corresponds to βcomp t or perform cylinder or CAPO-tests with
the objective to update the reliability level to βcomp if necessary. Based on a cost-benefit analysis
t

between cylinder and CAPO-tests, the optimal method can be used for assessing the current
reliability level. If βcomp ≥ βcomp
t , then no further action has to be performed. If βcomp < βcomp
t ,
then the components must be reinforced to correspond to βcomp , given that the reinforcement is less
t

expensive than replacing the components. A cost-benefit analysis will be carried out for cylinder
and CAPO-tests in the following.

6.8 Cost-benefit analysis

A cost-benefit analysis is performed as part of the decision model for the cylinder and CAPO-tests
in order to assess the optimal method in regards to total expected cost. The expected cost that is
used for each solution in the decision model will depend on the choice of µR , level of uncertainty,
test method, amount of tests and amount of components in the system considered. The correct
costs have not been obtained for the analysis in this report. Instead the analysis is carried out for
the following unit costs, E:

• CAPO-test, CCAPO : 1 E per test

• Cylinder test, Ccyl : 5 E per test
• Reinforcement, Cz : 50 E per component for z > 1
• Replacement, Cre : 500 E per component

Based on the expected costs listed above, the total expected cost for cylinder and CAPO-tests can
be estimated by Equation (6.4) and (6.5) respectively.

Ctot,cyl = n(Ccyl + (z − 1)Cz ) (6.4)

Ctot,CAPO = n(CCAPO + (z − 1)Cz ) (6.5)

56
6.8. Cost-benefit analysis Aalborg University

This analysis is based on average values for the expected cost of reinforcement needed, meaning
that the probability of having to reinforce the components is not considered in the cost-benefit
analysis. However, this will later be taken into account, when giving an example of expected cost
in the decision model.

The cost-benefit analysis will be carried out for a system with varying mean value, µR , uncertainties,
V(µR ) and V(VR ), number of components, n, and correlation coefficient, ρ. The following cases
are analyzed:

• Mean value: µR = 0.6µR , µR = 0.8µR and µR = 1.0µR

• Uncertainty: Small, large and very large, see Section 6.2
• Number of components: n = 10, n = 50 and n = 100
• Coefficient correlation: ρ = 0, ρ = 0.5 and ρ = 0.8

All results obtained for the cost-benefit analysis is graphically shown in Appendix D. In the
following, a few selected cases are presented to show the influence of the different parameters. The
case for non-correlated simulated realizations, a small level of uncertainty and a system consisting
of 100 components will be used as reference for assessing the influence of the different parameters.
Relevant results obtained for the analysis of this case are shown in Figure 6.3 and Table 6.5 and
6.6.

0.6µR,cylinder
6000 0.8µR,cylinder
1.0µR,cylinder
5000 0.6µR,capo
Total cost in units, E

0.8µR,capo
1.0µR,capo
4000

3000

2000

1000

0
10 20 30 40 50 60 70 80 90 100
Number of tests, ntest
Figure 6.3: Cost-benefit analysis of cylinder vs CAPO-tests for total average expected cost.

Table 6.5: Optimal amount of tests and minimal average expected cost for the two test types.

Optimal amount of tests Average minimum expected cost

Test type
Cylinder CAPO Cylinder CAPO
0.6µR 20 50 3205 3277
0.8µR 20 50 1193 1228
1.0µR 10 20 50 20

57
 6. Cost-Benefit Analysis Between Cylinder and CAPO-tests

Table 6.6: Corresponding average reinforcement values and probability of having to reinforce.

z zT PR
Type type Cylinder CAPO Cylinder CAPO Cylinder CAPO
0.6µR 0.63 0.66 0.63 0.66 99.9 99.8
0.8µR 0.22 0.24 0.26 0.27 90.8 92.5
1.0µR 0 0 0.16 0.14 50.6 54

It is seen from Figure 6.3 and Table 6.5 that the cylinder test is the most cost-beneficial for small
values of µR . For an increasing resistance the CAPO test tends towards becoming more cost-
beneficial than cylinder tests based on the specified costs. The curves tend to become linear after a
certain amount of tests performed, which is due to the cost of performing the tests starting to heavily
out-weight the benefits from performing additional tests. Furthermore, the results show that the
cylinder test reaches its optimum at a lower amount of performed tests than CAPO-tests, which
indicates that more information is obtained per test as would be expected due to the additional
model uncertainty for CAPO-tests. Additionally, Figure 6.3 indicates that a lot of information
in comparison to cost is obtained when increasing the amount of components tested from a low
amount. The difference in expected cost between 0.6µR and 1.0µR is significant, but this is caused
by no reinforcement being needed for 1.0µR on average and reinforcement of existing structure
being very costly.

6.8.1 Variation of components in a system

In this analysis, the same analysis is carried out with the exception of the amount of components
in a system being changed to 10 components, i.e. small uncertainty, n = 10 and ρ = 0. Results are
shown in Figure 6.4 and Table 6.7 and 6.8.

800
0.6µR,cylinder
700 0.8µR,cylinder
1.0µR,cylinder
600 0.6µR,capo
Total cost in units, E

0.8µR,capo
500 1.0µR,capo

400

300

200

100

0
3 4 5 6 7 8 9 10
Number of tests, ntest
Figure 6.4: Cost-benefit analysis of cylinder vs CAPO-tests for total average expected cost.

58
6.8. Cost-benefit analysis Aalborg University

Table 6.7: Optimal amount of tests and minimal average expected cost for the two test types.

Optimal amount of tests Average minimum expected cost

Test type
Cylinder CAPO Cylinder CAPO
0.6µR 8 10 374 359
0.8µR 8 10 166 146
1.0µR 8 10 42 19.5

Table 6.8: Corresponding average reinforcement values and probability of having to reinforce.

z zT PR
Type type Cylinder CAPO Cylinder CAPO Cylinder CAPO
0.6µR 0.70 0.70 0.70 0.71 99.8 99.9
0.8µR 0.24 0.27 0.32 0.29 90.57 91.8
1.0µR 0 0 0.17 0.17 54.0 57.1

Figure 6.4 and Table 6.8 show that CAPO-tests tend to become more cost beneficial than cylinder
tests for the all three cases, when the amount of components in the system is decreased from 100 to
10. The cylinder test reaches its optimum for less tests than the amount of components, while the
minimum cost for CAPO-tests is obtained by testing all components. This indicates that there is a
possibility for the real optimum to be located for ntest > 10. Due to the assumption made in this
analysis that only one test per component can be performed, this optimum will not be estimated.

6.8.2 Variation of uncertainty

This analysis is carried out for a very large uncertainty, n = 100 and ρ = 0. The results are shown
in Figure 6.5 and Table 6.9 and 6.10.

0.6µR,cylinder
6000 0.8µR,cylinder
1.0µR,cylinder
5000 0.6µR,capo
Total cost in units, E

0.8µR,capo
1.0µR,capo
4000

3000

2000

1000

0
10 20 30 40 50 60 70 80 90 100
Number of tests, ntest
Figure 6.5: Cost-benefit analysis of cylinder vs CAPO-tests for total average expected cost.

59
 6. Cost-Benefit Analysis Between Cylinder and CAPO-tests

Table 6.9: Optimal amount of tests and minimal average expected cost for the two test types.

Optimal amount of tests Average minimum expected cost

Test type
Cylinder CAPO Cylinder CAPO
0.6µR 20 50 3208 3298
0.8µR 20 50 1194 1202
1.0µR 10 20 50 20

Table 6.10: Corresponding average reinforcement values and probability of having to reinforce.

z zT PR
Type type Cylinder CAPO Cylinder CAPO Cylinder CAPO
0.6µR 0.65 0.66 0.72 0.72 96.0 96.6
0.8µR 0.19 0.23 0.36 0.40 78.8 80.6
1.0µR 0 0 0.29 0.26 55.9 55.2

By increasing the level of uncertainty from small to very large uncertainty, the tendencies are
similar to that of small uncertainty for which the cylinder test is now more cost-effective for a small
µR , while CAPO-tests is less expensive for 1.0µR . This indicates that the results for the cost is not
very sensitive towards changes in the uncertainty. Though, it should be noted that the probability
of having to reinforce the components is lower for a very large uncertainty for 0.6µR and 0.8µR .
The reason for this is that the very large V(µR ) and V(VR ) results in more characteristic values in
the upper-tail of the distribution and thus being beneficial for low values of µR .

6.8.3 Variation of correlation coefficient in a system

The following analysis is now carried out for n = 100, small uncertainty and a correlation coefficient
of ρ = 0.8. The results are illustrated in Figure 6.6 and listed in Table 6.11 and 6.12.

Table 6.11: Optimal amount of tests and minimal average expected cost for the two test types.

Optimal amount of tests Average minimum expected cost

Test type
Cylinder CAPO Cylinder CAPO
0.6µR 10 30 1867 1803
0.8µR 15 30 163 119
1.0µR 3 3 15 3

Table 6.12: Corresponding average reinforcement values and probability of having to reinforce.

z zT PR
Type type Cylinder CAPO Cylinder CAPO Cylinder CAPO
0.6µR 0.37 0.38 0.41 0.42 93.3 92.4
0.8µR 0.01 0.01 0.20 0.20 57.3 57.6
1.0µR 0 0 0.19 0.20 37.9 39.7

60
6.8. Cost-benefit analysis Aalborg University

0.6µR,cylinder
4000
0.8µR,cylinder
3500 1.0µR,cylinder
0.6µR,capo
Total cost in units, E 3000 0.8µR,capo
1.0µR,capo
2500

2000

1500

1000

500

0
10 20 30 40 50 60 70 80 90 100
Number of tests, ntest
Figure 6.6: Cost-benefit analysis of cylinder vs CAPO-tests for total average expected cost.

By comparing the results above with those obtained in the first analysis for ρ = 0 it is clearly
seen that the cost significantly decreases by introducing a large ρ. As a result, the optimums are
located for a smaller amount of tests and CAPO-tests are the most cost-beneficial solution for
all the evaluated values of µR . In reality, it is fair to assume that their is a correlation between
components in a system, if they are from the batch. Hence, a lot of cost is saved by using a
correlation coefficient in the cost-benefit analysis.

6.8.4 Cost related to decision model

The cost-benefit analysis can be used in addition to the specified cost in the decision model in
Figure 6.2 to decide whether to replace components in a system or perform cylinder or CAPO-tests
based on the average cost. An example of this will now be evaluated, which is shown in Figure 6.7
for 0.8µR , small uncertainty, n = 100 and ρ = 0. The cost related to collapse of the structure for
βcomp
t = 4.3 is assumed to be negligible.

61
 6. Cost-Benefit Analysis Between Cylinder and CAPO-tests

Figure 6.7: Cost related to decision model for reassessment of the reliability level for an existing
structure by performing cylinder and CAPO-tests. The branches corresponds to a
fixed βcomp
t .

The cost for cylinder and CAPO-tests at the end of the branches are estimated for a truncated
distribution of the reinforcement needed, zT , expressed as an average value. The average cost of
the different options in Figure 6.7 is calculated by:

C = Cre n (6.6)
C = PR (Cz zT n + Ccyl/CAPO ntest ) + (1 − PR ) Ccyl/CAPO ntest (6.7)

The values for zT and the probability of having to reinforce, PR , is found in Table 6.6. The results
for the average cost of replacing the components in a system or performing cylinder or CAPO-tests
are as following:

Replace comp.: C = 500 E · 100 = 50000 E

Cylinder test: C = 0.908 · (50 E · 0.26 · 100 + 5 E · 20) + 0.092 · 5 E · 20 = 1280 E
CAPO-test: C = 0.925 · (50 E · 0.27 · 100 + 1 E · 50) + 0.075 · 1 E · 100 = 1299 E

The results show that performing cylinder test on average is the optimal solution in terms of cost
in this case for the specified costs. Similarly, this analysis can be carried out for the other case
studies evaluated in this chapter, but this will not be the case, since the purpose of this study is to
describe the process behind choosing the optimal solution based on cost in a decision model.

In reality, the cost-benefit analysis in addition to a decision model has to be carried out for the
correct costs based on an assessment of the resistance through a preliminary inspection and rating.
The optimal decision can then be chosen based on the average cost.

62
Chapter

7
Discussion

The purpose of this chapter is to discuss the importance and choice of different prerequisites that
has been used for obtaining the results in this report. Furthermore, a perspectivation will be
carried out discussing different analyses that can be investigated for future studies in addition to
this project.

The analyses in this report, except for some analyses in Chapter 3, use a target reliability level
of βcomp
t = 4.3 corresponding to a probability of failure, PF , of a magnitude 10−5 . This is
not necessarily an unrealistic target reliability level, as it is commonly used for new structures,
cf. Table 2.1. However, for existing structures it is sometimes proposed and accepted to use a
reliability level lower than that of new structures, see Sørensen [2016]. The reason for accepting
a lower reliability level of existing structures is that the cost of safety measures is very large,
e.g. reinforcement or replacement of components, in comparison to new structures still in the
design phase. This means that if the analyses in this report are carried out again for βcomp t = 3.8,
then smaller k test -factors are necessary for obtaining βcomp and thus having an increased survival
t

rate, given the same prerequisites are kept. This is beneficial in regards to the decision model
for choosing to perform proof load tests instead of replacing the components, still neglecting the
cost for collapse of the structure. In the cost-benefit analysis for cylinder and CAPO-tests, the
lower reliability level will result in less total expected costs, if the same resistance models are kept.
This is due to the necessary characteristic value of the resistance, Rk,necessary , being smaller, hence
less average reinforcement is needed, if any. Therefore, the choice of the target reliability level is
important when deciding how to update the reliability level of an existing structure.

Another aspect to be discussed is the prerequisites used for the stochastic models of the resistance
in relation to the reevaluation of the reliability level by performing proof load tests with prior
information. The coefficient of variation for the resistance, VR , in the evaluated cases is assumed
to be 10% and 20%, which does not correspond to commonly used materials. In addition to this,
the model uncertainty, XR , related to the resistance has a fixed value for both cases, despite XR
normally being calibrated for specific materials. Thus, these prerequisites make it difficult to relate
the results to a real situation, but by keeping the same model uncertainty, it is possible to evaluate
the influence of VR . Instead, recommended stochastic resistance models from the Joint Committee
on Structural Safety [2001] for e.g. concrete in compression or steel could have been specifically
evaluated.

63
 7. Discussion

In addition to not using recommended stochastic resistance models, a fixed value of the weighting
factor, α, between permanent and variable load has been used. The weighting factor normally
depends on the material, e.g. high weighting on permanent load for concrete structures as described
in Sørensen [2011]. A value of α = 0.5 has been used for all analyses, meaning that the results
solely corresponds to this weighting. The sensitivity of α was investigated, see Table 4.25, for
which the results for a high weighting on the variable load showed significantly larger k test -factors
due to the large V. The large k test -factors indicates that the reliability level pre-proof load testing is
lower than for α = 0.5. Instead of using a fixed value for α, the average of a representative interval
for α could have been used similar to Chapter 3, so that the results on an average satisfies βcomp t .

The decision model associated to updating of the reliability level of an existing structure through
proof load testing with prior information, see Figure 4.2, has been given a strict rule that all
components have to be replaced, if failure is observed during proof load testing. This is most
likely not the optimal choice of action, but is what have been analyzed within the scope of this
project. The reason for ending the decision model a this point is that there exist a large number
of alternative actions that require in-depth analyses analogously to the the ones evaluated in this
project. For instance, new k test -factors can be estimated for which the simulation in Section 4.7
is performed such that the systems with failure in more than one component is disregarded in the
analysis. Another example of a choice of action is to proof load test a few components to failure and
use the information about the resistance to update the stochastic variables similar to test sampling
methods evaluated in Chapter 6. More choice of actions are in other words available and these can
be included in future extensional case studies.

64
Chapter

8
Conclusion

An assessment of how to reevaluate the reliability level of existing structures on component and
system level has been carried out for proof load testing with and without prior information as well
as cylinder and CAPO-tests without prior information.

The reliability level can be reevaluated through proof load testing without prior information by using
multiplication factors found in Table 3.4 on the characteristic load requirement. Multiplication
factors have also been estimated by using alternative proposed stochastic models of the variable
load, but the corresponding reliability levels are low, meaning that the results are only valid, if the
these reliability levels are accepted.

If prior information is available, then the reliability level can be reevaluated through proof load
testing of a number of components in a series or parallel system from a statistical homogeneous
population. This is done by using the multiplication factors on the characteristic load requirement
presented in Chapter 4. The results are based on a preliminary assessment of the uncertainty
related to the resistance. A decision model including this analysis has been established to estimate
the optimal choice of action based on total expected costs.

Test sampling methods like cylinder compression tests and CAPO-tests can be used to reevaluate
the reliability level of an existing structure. A cost-benefit analysis between the two tests has been
performed. The analyses show that CAPO-tests tend to become more cost-beneficial on average for
larger values of the resistance, smaller amount of components in a system and a larger correlation
coefficient. The cost-benefit analysis is used as part in a decision model to estimate the optimal
decision in regards to total expected cost.

65
 Bibliography

Anindita, 2017. Anindita. The Challenge Of Converting Existing Buildings Into Energy-Efficient
Structures, 2017. URL https://www.makaan.com/iq/living/
the-challenge-of-converting-existing-buildings-into-energy-efficient-structures.

BS EN 13791, 2007. Eurocode BS EN 13791. Assessment of in-situ compressive strength in

structures and pre-cast concrete components (BS EN 13791:2007). Dansk Standard, 2007.

Bygningslov, 1856. Bygningslov. Bygningslov for Staden Kjøbenhavn og dens forstæder. 1856.

Bygningslov, 1871. Bygningslov. Bygningslov for Staden Kjøbenhavn, Lovtidende nr. 53, lov nr.
164. 1871.

Bygningslov, 1889. Bygningslov. Bygningslov for Staden Kjøbenhavn, Lovtidende nr. 15, 1889.

Diamantidis, 2001. D. Diamantidis. Probabilistic Assessment of Existing Structures - JCSS

Report. JCSS, 2001.

DS 410, 1945. Dansk Standard DS 410. Normer for bygningskonstruktioner, 1.

Belastningsforskrifter. Dansk Standard, 1945.

DS/EN 1990, 2007. Dansk Standard DS/EN 1990. Projekteringsgrundlag for bærende
konstruktioner (DS/EN 1990:2007). Dansk Standard, 2007.

DS/EN 1991-1, 2007. Dansk Standard DS/EN 1991-1. Last på bærende konstruktioner (DS/EN
1991-1:2007). Dansk Standard, 2007.

DS/INF 172, 2009. Dansk Standard DS/INF 172. Baggrundsundersøgelser i forbindelse med
udarbejdelse af Nationale Annekser til EN 1990 og EN 1991 - Sikkerhedsformat,
lastkombinationer, partialkoefficienter, udmattelse, snelast, vindlast, m.m. (DS/INF172).
Dansk Standard, 2009.

FPrimeC Solutions Inc., 2016. FPrimeC Solutions Inc. How to Evaluate Compressive Strength
of Concrete, 2016. URL http:
//www.fprimec.com/how-to-evaluate-compressive-strength-of-concrete/.

Germann Instruments, 2018. Germann Instruments. CAPO-TEST catalog, 2018. URL

http://germann.org/products-by-application/category-1/capo-test.

67
 Bibliography

Ingeniørforening, 1908. Dansk Ingeniørforening. Normer for jærnbeton-konstruktioner (Norm

7). Dansk Ingeniørforening, 1908.

Ingeniørforening, 1916. Dansk Ingeniørforening. Norm for beregning af

husbygningskonstruktioner (Norm 12). Dansk Ingeniørforening, 1916.

Ingeniørforening, 1982. Dansk Ingeniørforening. Sikkerhedsbestemmelser for konstruktioner

(DS 409). Dansk Ingeniørforening, 1982.

JCSS, 2001. JCSS. JCSS probabilistic model code part 2: Load models. Joint Committee on
Structural Safety, 2001.

Joint Committee on Structural Safety, 2001. Joint Committee on Structural Safety.

Probabilistic Assessment of Existing Structures. RILEM Publications S.A.R.L., 2001.

NKB, 1978. NKB. Recommendations for loading and safety regulations for structural design,
Report 36. NKB, 1978.

S A K O, 1999. S A K O. Basis of design of structures - Proposals for Modification of Partial

Safety Factors in Eurocodes. Nordic Committee on Building Regulations, 1999.

Sanpolesi, 1997. Luca Sanpolesi. Scientific support activity in the field of structural stability of
civil engineering works snow loads. 1997.

SBI 251, 2015. Erik Steen Pedersen SBI 251. Vurdering af eksisterende konstruktioners
bæreevne. Statens Byggeforskningsinstitut, 2015.

Sørensen, 2016. John Dalsgaard Sørensen. Bæreevnevurdering af eksisterende konstruktioner.

2016.

Sørensen, 2011. John Dalsgaard Sørensen. Structural Reliability Theory And Risk Analysis.
John Dalsgaard Sørensen, 2011.

Sørensen, 2009. John Dalsgaard Sørensen. Baggrundsundersøgelse ifm. udarbejdelse af

brospecifikke Nationale Annekser til Eurocodes EN1990 og EN1991, Aalborg Universitet,
2009.

68
Appendix

A
Simulation Techniques

In this appendix the necessary tools used to estimate the probability of failure will be described.
This includes simulation of outcomes of stochastic variables with an arbitrary distribution and
with/ without correlation between the variables. It also includes the simulation technique used to
estimate the probability of failure and the error related to this. The theory accounted for in this
appendix is based on Sørensen [2011].

A.1 Simulation of stochastic variables

Consider a stochastic variable, X, with a corresponding distribution function, FX (x). In the
following, the inverse method is used to simulate random realizations from the distribution function.
There are two steps that are needed to generate an outcome x̂ of X:

1. Generate an outcome v̂ of V which is uniformly distributed between 0 and 1.

2. Determine the outcome of x̂ from Equation (A.1).

x̂ = FX−1 (FV (v̂)) = FX−1 (v̂) (A.1)

The method is illustrated in Figure A.1 and it is seen that the distribution function for X̂ with
outcomes simulated by this procedure is:
   
FX̂ (x) = P X̂ ≤ x = P FX−1 (V) ≤ x = P (V ≤ FX (x)) = Fx (x) (A.2)

F v ( v) F x( x)

1 1

v x
v^ x^
Figure A.1: The concept of the inverse method

69
 A. Simulation Techniques

In the following, the derived equations, which will be obtained by applying the inverse method to
a normal distribution, lognormal distribution and gumbel distribution will be presented. These are
the various of distribution functions that have been used throughout this project.

Simulation of random realization of normally distributed random variable

The distribution function for a normally distributed variable with expected value, µ, and standard
deviation, σ is defined by:
 x − µ
FX (x) = Φ (A.3)
σ
where Φ (u) is the standardized distribution function for a normal distributed stochastic variable
with µ = 0 and σ = 1.

By applying Equation (A.1) following expression to simulate realizations of a normal distributed

random variable is obtained:

x̂ = µ + Φ (v̂) σ (A.4)

Simulation of random realization of lognormally distributed random variable

The distribution function for a lognormally distributed stochastic random variable with expected
value, µ, and standard deviation, σ is defined by:

ln x − µY
 
FX (x) = Φ (A.5)
σY

where

µY Lognormal parameter
σY Lognormal parameter

The lognormal parameters µY and σY is calculated from Equations (A.6) and (A.7), respectively.
v
u ! q
t  2
σ
σY = ln + 1 = ln V 2 + 1


(A.6)
µ
1
µY = ln µ − σY2 (A.7)
2
The lognormally distributed random variable, Y , has the following relationship to the normal
distributed random variable:

Y = ln X (A.8)

The Equation that is obtained by applying Equation (A.1) and (A.8) is:

x̂ = exp ŷ = exp (µY + Φ (v̂) σY ) (A.9)

70
A.2. Simulation of correlated lognormally distributed numbers Aalborg University

Simulation of random realization of Gumbel distributed random variable

The distribution function for a stochastic variable with expected value, µ, and standard deviation,
σ is defined by:

FX (x) = exp (− exp (−α (x − β))) (A.10)

where

α Shape parameter
β Scale parameter

The shape and scale parameter is related to the mean and standard deviation:
0.5772
µ= β+ (A.11)
α
π
σ= √ (A.12)
α 6

By applying Equation (A.1) following expression is obtained:

1
x̂ = v̂ − ln (− ln (v̂)) (A.13)
α

A.2 Simulation of correlated lognormally distributed numbers

In the previous section the inverse method used to simulate stochastic variables was explained.
In this section a technique to simulate correlated log- and normally distributed variables will be
described.

Let the stochastic variables Xi, i = 1, ..., n be normally distributed with expected value
µX1 , ..., µXn , standard deviations σX1 , ..., σXn and correlation coefficients ρi j , i, j = 1, ..., n.
Firstly, a transformation from correlated to uncorrelated stochastic variables is necessary. This
transformation can be carried out in various of ways, but here the Choleski triangulation is used.
The procedure described in the following requires the correlation matrix ρ to be positive definite.

The first step is to determine the normalized variables Yi, i = 1, ..., n with µ = 0 and σ = 1:
Xi − µXi
Yi = , i = 1, ..., n (A.14)
σXi

It should be noted that Y will have a covariance matrix and correlation coefficient matrix equal to
ρ. Now a transformation from Y to uncorrelated and normalized variables U is defiend as:

Y = TU (A.15)

where T is a lower triangular matrix meaning. The covariance matrix CY for Y can be expressd
as:

CY = E YYT = E[TUUT TT ] = TE UUT TT = TTT = ρ

   
(A.16)

71
 A. Simulation Techniques

The transformation from X to U is written as:

X̂ = µx + DTÛ (A.17)

where D is a diagonal matrix with standard deviations in the diagonal.

Equation (A.17) is used to simulate random correlated normal distributed variables. If lognormally
distributed correlated random variables are to be simulated, then the following equation is used:
   
X̂ = exp Ŷ = exp µx,Y + DTÛ (A.18)

where D is a diagonal matrix with the lognormal standard deviations in the diagonal.

In the following section a simulation method to estimate the probability of failure is described:

P f = P (g (X)) ≤ 0 (A.19)

Here the presented failure function is assumed to be modelled in the physical space.

A.3 Crude Monte Carlo simulation

A crude Monte Carlo simulation is a method for estimating the probability of failure, P f , which is
estimated by:
N
1 Õ 
P̂ f =

I g(x̂ j ) (A.20)
N j=1

where

N Number of simulations
x̂ j Sample no. j of a randomly sampled stochastic vector U.
I [g(u)] The indicator function

The indicator function is defined by the following:

(
0 if g(x) > 0 (safe)
I [g(u)] = (A.21)
1 if g(x) < 0 (failure)

Thus, if evaluation of the limit state function results in a value above 0, the component in the
sample is denoted 0 corresponding to no failure in the component. Likewise the component is
denoted 1, if the limit state function results in a value below 0.
The standard error of the estimated probability of failure, P̂ f , is estimated by the following:
s
P̂ f (1 − P̂ f )
(A.22)
N

Confidence intervals can also be established for the estimate of the probability of failure by using
that P̂ f becomes normally distributed for the number of samples, N, going towards infinite, i.e.
N −→ ∞.

72
A.4. Reliability updating Aalborg University

A.4 Reliability updating

Reliability updating is a method for which new available information is used to update the stochastic
models and probability of failure for a given component or structure. Thus, reliability updating is
useful for assessing the reliability of existing structures. When performing the updating, it depends
on the type of information available. In general, two types of new information are distinguished
between, which are:

• Observation of events that are described by a single or multiple stochastic variables. If a

non-failure event is observed, it is modelled by an event margin, while a failure event is
modelled by a safety margin.
• Test samples or measurements of a stochastic variable, e.g. concrete resistance for which
the updating can be performed by using Bayesian statistics.

The methods for performing the updating depending on the new available information are described
in the following part of the section.

A.4.1 Bayesian updating of failure events

When modelling the observed events, firstly an event function has to be established for the stochastic
variable:

H = h(X) (A.23)

The event function, h, corresponds to the limit state function, while the observations are considered
as samples of the stochastic variable H. An example of an observed event can be the damage level
of an existing structure that is subjected to a well defined proof load. Usually no damage is
observed in this case. These observations are assumed to be modelled by either:

• Inequality events {H ≤ 0}, i.e. the observed quantity is less than or equal to a limit.
• Equality events {H = 0}, i.e. the observed quantity is equal to a limit.

When modelling the observations by inequality events, the updated probability of failure is
estimated by the following, cf. Sørensen [2011]:
P (g(X) ≤ 0 ∩ h(X) ≤ 0)
f = P (g(X) ≤ 0|h(X) ≤ 0) =
PU (A.24)
P (h(X) ≤ 0)
In equation (A.24), M = g(X) is the safety margin related to the limit state function, g(X), while X
represents the stochastic variables. The equation takes into account that the probability of an event
P(A ∩ B)
A given an event B, normally denoted by P(A|B), can be expressed as . Furthermore,
P(B)
it shoud be noted that the expression P (g(X) ≤ 0 ∩ h(X) ≤ 0) corresponds to a parallel system
consisting of two elements. When evaluating equation (A.24), FORM or SORM methods can be
used.
If the observations are modelled by equality events instead for which the observed quantity is equal
to a limit, then the probability of failure is estimated by, cf. Sørensen [2011]:
∂
P (g(X) ≤ 0 ∩ h(X) ≤ z)
P (g(X) ≤ 0 ∩ h(X) = 0) ∂z
P f = P (g(X) ≤ 0|h(X) = 0) =
U
=
P (h(X) = 0) ∂
P (h(X) ≤ z)
∂z
(A.25)

73
 A. Simulation Techniques

Equation (A.25) can similarly to equation (A.24) be evaluated by FORM or SORM methods and
the equation is simple to generalize in case that more than one event is observed.

A.4.2 Bayesian updating of stochastic variables

For parameter estimations, it is recommended to use Bayesian techniques whenever possible,
since the statistical uncertainty related to the estimated parameters can be determined through
Bayesian estimation. Furthermore, Bayesian techniques makes it easy to update a model, when
new information becomes available.
When one or more observations are available of the stochastic variables, X, it is possible to update
the probabilistic model related to the stochastic variables and thus the probability of failure. The
density function for a stochastic variable, X, is now considered:

fX (x, q) (A.26)

In equation (A.26), q denotes a vector of parameters governing the distribution of the stochastic
variable, X, e.g. mean value and standard deviation of X, if X is normally distributed.
If there is an uncertainty related to the parameters, q, then the density function, fX (x, q), can instead
be considered as a conditional density function, fX (x|Q), for which q denotes a realization of Q.
The prior density function of Q is denoted fQ0 (q) and contains information from e.g. historical
knowledge or previous documentation about the parameter, Q.
For updating the density function, it is assumed that n observations/realizations of the stochastic
variable, X, are available, i.e. x̂ = ( xˆ1, xˆ2, ..., xˆn ), and these realizations are independent. The
posterior density function of the uncertain parameters, Q, given the realizations can then be
expressed as:
fX ( x̂| q) fQ0 (q)
fQ00(q| x̂) =∫ (A.27)
fX ( x̂| q) fQ0 (q)d q
N
In equation (A.27), fX ( x̂|q) =
Î
fX ( x̂i |q) is the probability density for the observations based on
i=1
the assumption that the distribution parameter are q. The distribution type of the prior distribution
for a stochastic variable, X, is typically chosen identical to the posterior distribution.
Finally the predictive density function of the stochastic variable, X, given the realization, x̂, can
be expressed by:
∫
fX (x| x̂) = fX (x| q) fQ00(q| x̂)d q (A.28)

It is possible to quantify both the physical and statistical uncertainty related to the given variable
and the model parameters respectively by using the Bayesian methods described above. However,
it should be noted that the measurement and model uncertainties have to be taken into account
when formulating the probabilistic model.

Example When updating stochastic variables through Bayesian methods, the distribution type of
the density function as well as the parameters of interest have to be identified, since the updating
depends on these. As an example, consider a normally distributed stochastic variable, X, with a
known variance, σ 2 . The normal distribution is expressed as the following for which the parameter
of interest is the unknown mean value of the stochastic variable:
1  x − µ 2
 
1
fX (x| µ, σ) = √ exp − (A.29)
σ 2π 2 σ

74
A.4. Reliability updating Aalborg University

The prior distribution of the unknown mean value, µ, is normally distributed with a mean value,
µ0, and variance, σ 02 . New objective information has been obtained by performing tests, resulting
in a test sample, n, with a mean value, X. A likelihood function can be computed based on the
normal distribution in equation (A.29) and inserted in equation (A.27). The posterior mean value,
µ00, and variance, σ 00, is then determined by:

nXσ 02 + µ0 σ 2
µ00 = (A.30)
nσ 02 + σ 2

σ 02 σ 2
σ 002 = (A.31)
nσ 02 + σ 2
Similarly, the mean value, µ00, and standard deviation, σ 000, for the predictive distribution can be
expressed, which also becomes normal distributed:

nXσ 02 + µ0 σ 2
µ00 = (A.32)
nσ 02 + σ 2

p
σ 000 = σ 002 + σ 2 (A.33)

When the stochastic variable, X, has a different distribution type, the same procedure is followed,
i.e. the likelihood function of the given distribution type can be inserted in equation (A.27) for
which the parameters of interest are determined.

75
Appendix

B
Factors for Proof Load Testing
Without Prior Information

Table B.1: k test -factors for the 2. stochastic model of wind load.

Characteristic permanent
Factor: k test
load percentage of
characteristic load
β = 1.6 β = 2.1 β = 2.6 β = 3.1 β = 3.6
requirement [%]
0 1.56 1.85 2.21 2.65 3.18
25 1.42 1.64 1.91 2.24 2.64
50 1.29 1.44 1.62 1.83 2.10
60 1.24 1.36 1.50 1.68 1.89
65 1.22 1.32 1.45 1.60 1.78
70 1.19 1.29 1.39 1.52 1.68
75 1.18 1.25 1.34 1.45 1.58
80 1.16 1.22 1.30 1.39 1.49
85 1.15 1.21 1.26 1.33 1.41
90 1.15 1.20 1.25 1.30 1.35
95 1.15 1.20 1.25 1.30 1.35
100 1.16 1.21 1.26 1.31 1.36

77
 B. Factors for Proof Load Testing Without Prior Information

Table B.2: k test -factors for the 3. stochastic model of wind load.

Characteristic permanent
Factor: k test
load percentage of
characteristic load
β = 2.5 β = 3.0 β = 3.5 β = 4.0 β = 4.5
requirement [%]
0 1.49 1.81 2.21 2.70 3.30
25 1.37 1.61 1.91 2.27 2.72
50 1.26 1.42 1.62 1.86 2.16
60 1.22 1.35 1.50 1.70 1.94
65 1.20 1.31 1.45 1.62 1.83
70 1.18 1.28 1.40 1.54 1.72
75 1.17 1.25 1.35 1.47 1.62
80 1.17 1.23 1.31 1.40 1.52
85 1.18 1.23 1.29 1.35 1.43
90 1.20 1.24 1.29 1.34 1.39
95 1.22 1.27 1.32 1.37 1.42
100 1.26 1.31 1.36 1.41 1.46

Table B.3: k test -factors for the 2. stochastic model of snow load.

Characteristic permanent
Factor: k test
load percentage of
characteristic load
β = 0.9 β = 1.4 β = 1.9 β = 2.4 β = 2.9
requirement [%]
0 1.38 1.75 2.22 2.82 3.57
25 1.28 1.56 1.92 2.36 2.93
50 1.19 1.38 1.62 1.92 2.29
60 1.16 1.31 1.50 1.74 2.04
65 1.14 1.27 1.44 1.65 1.91
70 1.13 1.24 1.38 1.56 1.79
75 1.11 1.21 1.33 1.48 1.67
80 1.10 1.18 1.28 1.40 1.55
85 1.09 1.15 1.23 1.32 1.43
90 1.08 1.14 1.20 1.26 1.33
95 1.09 1.14 1.18 1.23 1.28
100 1.09 1.14 1.19 1.24 1.29

78
 Aalborg University

Table B.4: k test -factors for the 3. stochastic model of snow load.

Characteristic permanent
Factor: k test
load percentage of
characteristic load
β = 0.8 β = 1.3 β = 1.8 β = 2.3 β = 2.8
requirement [%]
0 1.34 1.72 2.22 2.86 3.69
25 1.25 1.54 1.92 2.40 3.02
50 1.17 1.37 1.62 1.94 2.35
60 1.14 1.30 1.50 1.75 2.08
65 1.13 1.26 1.44 1.66 1.95
70 1.11 1.23 1.38 1.58 1.82
75 1.10 1.20 1.33 1.49 1.69
80 1.09 1.17 1.27 1.40 1.57
85 1.08 1.15 1.23 1.32 1.44
90 1.08 1.13 1.19 1.26 1.34
95 1.08 1.13 1.18 1.23 1.28
100 1.09 1.14 1.19 1.24 1.29

79
Appendix

C
Factors for Proof Load Testing
With Prior Information

C.1 Series systems

C.1.1 Results for correlation ρ = 0
Table C.1: k test -factors for ρ = 0 in case 2a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.85 (43%) 1.53 (57%) 1.46 (58%)
20% 1.66 (53%) 1.45 (60%) 1.39 (61%)
30% 1.56 (60%) 1.40 (63%) 1.35 (64%)
40% 1.51 (63%) 1.37 (65%) 1.32 (65%)
50% 1.48 (67%) 1.35 (65%) 1.30 (67%)

Table C.2: k test -factors for ρ = 0 in case 3a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.83 (45%) 1.52 (55%) 1.44 (57%)
20% 1.65 (52%) 1.44 (58%) 1.36 (61%)
30% 1.56 (56%) 1.39 (60%) 1.35 (59%)
40% 1.52 (57%) 1.36 (62%) 1.30 (63%)
50% 1.48 (60%) 1.33 (63%) 1.30 (62%)

Table C.3: k test -factors for ρ = 0 in case 4a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.90 (38%) 1.50 (52%) 1.40 (55%)
20% 1.65 (48%) 1.40 (55%) 1.35 (55%)
30% 1.50 (56%) 1.35 (57%) 1.30 (58%)
40% 1.50 (54%) 1.35 (55%) 1.30 (56%)
50% 1.45 (60%) 1.30 (59%) 1.25 (56%)

81
 C. Factors for Proof Load Testing With Prior Information

Table C.4: k test -factors for ρ = 0 in case 2b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 2.88 (18%) 1.87 (42%) 1.68 (48%)
20% 2.24 (33%) 1.66 (50%) 1.51 (57%)
30% 1.98 (44%) 1.55 (56%) 1.43 (61%)
40% 1.83 (51%) 1.47 (62%) 1.37 (65%)
50% 1.71 (59%) 1.41 (67%) 1.30 (71%)

Table C.5: k test -factors for ρ = 0 in case 3b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 2.98 (16%) 1.87 (42%) 1.66 (48%)
20% 2.28 (31%) 1.66 (49%) 1.50 (54%)
30% 2.03 (40%) 1.55 (53%) 1.42 (58%)
40% 1.89 (45%) 1.47 (58%) 1.35 (62%)
50% 1.77 (51%) 1.40 (63%) 1.31 (65%)

Table C.6: k test -factors for ρ = 0 in case 4b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 3.30 (13%) 1.90 (39%) 1.66 (46%)
20% 2.41 (27%) 1.65 (47%) 1.50 (50%)
30% 2.10 (35%) 1.55 (50%) 1.40 (54%)
40% 1.90 (43%) 1.46 (54%) 1.35 (56%)
50% 1.78 (47%) 1.40 (56%) 1.30 (58%)

C.1.2 Results for correlation ρ = 0.5

Table C.7: k test -factors for ρ = 0.5 in case 2a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.61 (71%) 1.45 (75%) 1.40 (76%)
20% 1.51 (74%) 1.39 (77%) 1.35 (78%)
30% 1.48 (75%) 1.38 (76%) 1.34 (77%)
40% 1.45 (77%) 1.34 (79%) 1.32 (78%)
50% 1.41 (79%) 1.32 (80%) 1.30 (79%)

82
C.1. Series systems Aalborg University

Table C.8: k test -factors for ρ = 0.5 in case 3a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.64 (64%) 1.45 (70%) 1.40 (71%)
20% 1.53 (68%) 1.40 (71%) 1.35 (72%)
30% 1.49 (69%) 1.36 (72%) 1.33 (73%)
40% 1.44 (72%) 1.35 (72%) 1.31 (73%)
50% 1.42 (72%) 1.34 (72%) 1.30 (73%)

Table C.9: k test -factors for ρ = 0.5 in case 4a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.63 (58%) 1.44 (62%) 1.39 (63%)
20% 1.52 (61%) 1.38 (63%) 1.34 (64%)
30% 1.46 (63%) 1.35 (64%) 1.32 (64%)
40% 1.43 (64%) 1.35 (63%) 1.30 (65%)
50% 1.41 (64%) 1.31 (66%) 1.29 (65%)

Table C.10: k test -factors for ρ = 0.5 in case 2b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.82 (81%) 1.46 (88%) 1.40 (88%)
20% 1.60 (87%) 1.36 (90%) 1.30 (90%)
30% 1.50 (89%) 1.30 (91%) 1.25 (91%)
40% 1.43 (91%) 1.26 (92%) 1.23 (91%)
50% 1.39 (92%) 1.25 (92%) 1.20 (92%)

Table C.11: k test -factors for ρ = 0.5 in case 3b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.91 (72%) 1.50 (81%) 1.40 (83%)
20% 1.67 (79%) 1.40 (83%) 1.34 (83%)
30% 1.55 (82%) 1.35 (84%) 1.28 (85%)
40% 1.48 (84%) 1.30 (86%) 1.25 (86%)
50% 1.43 (86%) 1.27 (87%) 1.20 (88%)

83
 C. Factors for Proof Load Testing With Prior Information

Table C.12: k test -factors for ρ = 0.5 in case 4b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 2.06 (57%) 1.55 (70%) 1.45 (72%)
20% 1.74 (67%) 1.41 (74%) 1.35 (74%)
30% 1.61 (71%) 1.35 (76%) 1.30 (75%)
40% 1.53 (74%) 1.31 (77%) 1.25 (77%)
50% 1.48 (75%) 1.30 (76%) 1.25 (76%)

C.1.3 Results for correlation ρ = 0.8

Table C.13: k test -factors for ρ = 0.8 in case 2a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.51 (82%) 1.42 (83%) 1.38 (85%)
20% 1.47 (82%) 1.38 (84%) 1.36 (84%)
30% 1.43 (84%) 1.36 (85%) 1.34 (85%)
40% 1.42 (84%) 1.36 (84%) 1.33 (85%)
50% 1.40 (85%) 1.35 (85%) 1.32 (85%)

Table C.14: k test -factors for ρ = 0.8 in case 3a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.51 (76%) 1.42 (77%) 1.38 (79%)
20% 1.46 (77%) 1.38 (79%) 1.35 (79%)
30% 1.42 (79%) 1.38 (77%) 1.35 (78%)
40% 1.42 (78%) 1.36 (78%) 1.34 (78%)
50% 1.41 (78%) 1.35 (78%) 1.32 (79%)

Table C.15: k test -factors for ρ = 0.8 in case 4a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.51 (67%) 1.40 (69%) 1.37 (69%)
20% 1.47 (67%) 1.38 (69%) 1.35 (69%)
30% 1.43 (68%) 1.37 (69%) 1.35 (69%)
40% 1.41 (69%) 1.35 (70%) 1.33 (69%)
50% 1.42 (68%) 1.35 (69%) 1.33 (69%)

84
C.2. Parallel systems Aalborg University

Table C.16: k test -factors for ρ = 0.8 in case 2b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.49 (95%) 1.31 (96%) 1.28 (96%)
20% 1.40 (96%) 1.27 (97%) 1.24 (96%)
30% 1.35 (96%) 1.25 (96%) 1.20 (97%)
40% 1.32 (96%) 1.23 (97%) 1.20 (97%)
50% 1.29 (97%) 1.22 (97%) 1.18 (97%)

Table C.17: k test -factors for ρ = 0.8 in case 3b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.55 (90%) 1.37 (92%) 1.30 (93%)
20% 1.45 (92%) 1.31 (93%) 1.25 (93%)
30% 1.39 (93%) 1.27 (93%) 1.25 (93%)
40% 1.36 (93%) 1.25 (93%) 1.22 (93%)
50% 1.33 (94%) 1.25 (93%) 1.20 (94%)

Table C.18: k test -factors for ρ = 0.8 in case 4b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.60 (81%) 1.40 (84%) 1.35 (84%)
20% 1.49 (83%) 1.33 (85%) 1.30 (84%)
30% 1.43 (84%) 1.30 (85%) 1.27 (85%)
40% 1.38 (85%) 1.29 (85%) 1.25 (85%)
50% 1.38 (85%) 1.26 (86%) 1.23 (86%)

C.2 Parallel systems

C.2.1 Results for correlation ρ = 0
Table C.19: k test -factors for ρ = 0 in case 2a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.84 (44%) 1.56 (53%) 1.46 (59%)
20% 1.65 (55%) 1.46 (59%) 1.40 (60%)
30% 1.59 (56%) 1.43 (58%) 1.36 (62%)
40% 1.52 (62%) 1.40 (59%) 1.35 (61%)
50% 1.47 (66%) 1.37 (62%) 1.33 (61%)

85
 C. Factors for Proof Load Testing With Prior Information

Table C.20: k test -factors for ρ = 0 in case 3a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.82 (45%) 1.52 (55%) 1.44 (57%)
20% 1.62 (55%) 1.45 (56%) 1.38 (59%)
30% 1.57 (56%) 1.40 (58%) 1.35 (59%)
40% 1.49 (61%) 1.37 (60%) 1.33 (59%)
50% 1.45 (63%) 1.36 (59%) 1.32 (59%)

Table C.21: k test -factors for ρ = 0 in case 4a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.86 (41%) 1.52 (51%) 1.42 (54%)
20% 1.64 (49%) 1.43 (53%) 1.36 (54%)
30% 1.54 (53%) 1.40 (53%) 1.33 (55%)
40% 1.49 (55%) 1.35 (55%) 1.31 (55%)
50% 1.46 (56%) 1.33 (56%) 1.30 (55%)

Table C.22: k test -factors for ρ = 0 in case 2b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 2.72 (24%) 1.90 (39%) 1.72 (43%)
20% 2.20 (37%) 1.65 (51%) 1.56 (49%)
30% 1.94 (48%) 1.56 (55%) 1.51 (48%)
40% 1.79 (56%) 1.53 (54%) 1.45 (52%)
50% 1.67 (64%) 1.49 (56%) 1.40 (56%)

Table C.23: k test -factors for ρ = 0 in case 3b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 2.96 (17%) 1.89 (40%) 1.69 (45%)
20% 2.26 (33%) 1.67 (48%) 1.54 (50%)
30% 1.96 (45%) 1.58 (50%) 1.45 (54%)
40% 1.79 (53%) 1.50 (54%) 1.46 (49%)
50% 1.71 (56%) 1.47 (55%) 1.44 (49%)

86
C.2. Parallel systems Aalborg University

Table C.24: k test -factors for ρ = 0 in case 4b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 3.00 (18%) 1.90 (39%) 1.66 (46%)
20% 2.34 (30%) 1.65 (47%) 1.53 (48%)
30% 2.00 (40%) 1.54 (51%) 1.49 (47%)
40% 1.82 (47%) 1.52 (49%) 1.42 (50%)
50% 1.68 (54%) 1.44 (53%) 1.41 (49%)

C.2.2 Results for correlation ρ = 0.5

Table C.25: k test -factors for ρ = 0.5 in case 2a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.63 (70%) 1.45 (75%) 1.38 (78%)
20% 1.53 (74%) 1.40 (77%) 1.35 (78%)
30% 1.48 (75%) 1.38 (76%) 1.32 (79%)
40% 1.44 (78%) 1.36 (77%) 1.31 (80%)
50% 1.42 (79%) 1.34 (78%) 1.31 (79%)

Table C.26: k test -factors for ρ = 0.5 in case 3a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.62 (65%) 1.45 (70%) 1.39 (72%)
20% 1.51 (70%) 1.39 (72%) 1.35 (72%)
30% 1.47 (71%) 1.36 (72%) 1.32 (74%)
40% 1.44 (72%) 1.35 (72%) 1.30 (74%)
50% 1.42 (72%) 1.35 (71%) 1.30 (73%)

Table C.27: k test -factors for ρ = 0.5 in case 4a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.61 (59%) 1.38 (67%) 1.30 (70%)
20% 1.52 (61%) 1.33 (68%) 1.27 (70%)
30% 1.46 (63%) 1.30 (69%) 1.25 (70%)
40% 1.43 (64%) 1.30 (68%) 1.24 (70%)
50% 1.41 (65%) 1.29 (68%) 1.25 (69%)

87
 C. Factors for Proof Load Testing With Prior Information

Table C.28: k test -factors for ρ = 0.5 in case 2b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.78 (83%) 1.44 (89%) 1.36 (90%)
20% 1.57 (88%) 1.37 (90%) 1.28 (91%)
30% 1.47 (91%) 1.30 (91%) 1.25 (91%)
40% 1.41 (92%) 1.27 (92%) 1.24 (91%)
50% 1.36 (93%) 1.25 (92%) 1.20 (92%)

Table C.29: k test -factors for ρ = 0.5 in case 3b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.86 (75%) 1.47 (83%) 1.38 (85%)
20% 1.64 (80%) 1.36 (86%) 1.31 (85%)
30% 1.51 (85%) 1.32 (86%) 1.25 (87%)
40% 1.45 (86%) 1.31 (85%) 1.25 (86%)
50% 1.41 (87%) 1.28 (86%) 1.22 (87%)

Table C.30: k test -factors for ρ = 0.5 in case 4b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.99 (61%) 1.52 (72%) 1.41 (74%)
20% 1.70 (69%) 1.40 (75%) 1.32 (76%)
30% 1.56 (74%) 1.35 (76%) 1.29 (76%)
40% 1.49 (76%) 1.31 (77%) 1.25 (77%)
50% 1.43 (78%) 1.29 (77%) 1.25 (76%)

C.2.3 Results for correlation ρ = 0.8

Table C.31: k test -factors for ρ = 0.8 in case 2a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.51 (82%) 1.39 (86%) 1.34 (88%)
20% 1.47 (82%) 1.36 (86%) 1.32 (88%)
30% 1.42 (85%) 1.36 (85%) 1.30 (88%)
40% 1.41 (85%) 1.34 (86%) 1.30 (88%)
50% 1.41 (84%) 1.33 (86%) 1.30 (88%)

88
C.2. Parallel systems Aalborg University

Table C.32: k test -factors for ρ = 0.8 in case 3a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.53 (74%) 1.41 (78%) 1.36 (80%)
20% 1.46 (77%) 1.38 (79%) 1.32 (82%)
30% 1.43 (78%) 1.35 (80%) 1.32 (81%)
40% 1.42 (78%) 1.35 (79%) 1.31 (81%)
50% 1.41 (78%) 1.34 (80%) 1.30 (81%)

Table C.33: k test -factors for ρ = 0.8 in case 4a with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.52 (66%) 1.39 (70%) 1.34 (72%)
20% 1.47 (67%) 1.36 (71%) 1.33 (71%)
30% 1.43 (69%) 1.35 (70%) 1.30 (73%)
40% 1.42 (69%) 1.35 (69%) 1.30 (72%)
50% 1.40 (69%) 1.34 (70%) 1.30 (71%)

Table C.34: k test -factors for ρ = 0.8 in case 2b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.45 (96%) 1.22 (98%) 1.15 (99%)
20% 1.37 (96%) 1.19 (98%) 1.11 (99%)
30% 1.33 (97%) 1.16 (98%) 1.09 (99%)
40% 1.29 (97%) 1.15 (98%) 1.09 (99%)
50% 1.27 (97%) 1.14 (98%) 1.07 (99%)

Table C.35: k test -factors for ρ = 0.8 in case 3b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.52 (91%) 1.28 (95%) 1.20 (96%)
20% 1.40 (94%) 1.22 (95%) 1.15 (96%)
30% 1.36 (94%) 1.20 (95%) 1.15 (96%)
40% 1.32 (94%) 1.19 (95%) 1.12 (96%)
50% 1.32 (94%) 1.18 (95%) 1.12 (96%)

89
 C. Factors for Proof Load Testing With Prior Information

Table C.36: k test -factors for ρ = 0.8 in case 4b with a various n and test portions. Values inside
parenthesis () indicate the amount of systems that succeeded the proof load testing.

Test portion n=10

k test n=50
k test n=100
k test
10% 1.58 (82%) 1.31 (88%) 1.25 (88%)
20% 1.46 (84%) 1.26 (88%) 1.20 (89%)
30% 1.40 (85%) 1.25 (88%) 1.19 (89%)
40% 1.36 (86%) 1.22 (88%) 1.17 (89%)
50% 1.34 (86%) 1.21 (88%) 1.15 (90%)

90
Appendix

D
Cost-Benefit Analysis of Sampling
Methods

D.0.1 Korrelation = 0
800
0.6µR,cylinder
700 0.8µR,cylinder
1.0µR,cylinder
600 0.6µR,capo
Total cost in units, E

0.8µR,capo
500 1.0µR,capo

400

300

200

100

0
3 4 5 6 7 8 9 10
Number of tests, ntest
Figure D.1: Cost-benefit analysis of cylinder tests. Total cost for performing the test and
reinforcing the component if necessary

91
 D. Cost-Benefit Analysis of Sampling Methods

0.6µR,cylinder
3000 0.8µR,cylinder
1.0µR,cylinder
2500 0.6µR,capo
Total cost in units, E

0.8µR,capo
1.0µR,capo
2000

1500

1000

500

0
5 10 15 20 25 30 35 40 45 50
Number of tests, ntest
Figure D.2: Cost-benefit analysis of CAPO tests. Total cost for performing the test and reinforcing
the component if necessary

0.6µR,cylinder
6000 0.8µR,cylinder
1.0µR,cylinder
5000 0.6µR,capo
Total cost in units, E

0.8µR,capo
1.0µR,capo
4000

3000

2000

1000

0
10 20 30 40 50 60 70 80 90 100
Number of tests, ntest
Figure D.3: Cost-benefit analysis of CAPO tests. Total cost for performing the test and reinforcing
the component if necessary

92
 Aalborg University

800
0.6µR,cylinder
700 0.8µR,cylinder
1.0µR,cylinder
600 0.6µR,capo
Total cost in units, E 0.8µR,capo
500 1.0µR,capo

400

300

200

100

0
3 4 5 6 7 8 9 10
Number of tests, ntest
Figure D.4: Cost-benefit analysis of cylinder tests. Total cost for performing the test and
reinforcing the component if necessary

0.6µR,cylinder
3000 0.8µR,cylinder
1.0µR,cylinder
2500 0.6µR,capo
Total cost in units, E

0.8µR,capo
1.0µR,capo
2000

1500

1000

500

0
5 10 15 20 25 30 35 40 45 50
Number of tests, ntest
Figure D.5: Cost-benefit analysis of CAPO tests. Total cost for performing the test and reinforcing
the component if necessary

93
 D. Cost-Benefit Analysis of Sampling Methods

0.6µR,cylinder
6000 0.8µR,cylinder
1.0µR,cylinder
5000 0.6µR,capo
Total cost in units, E

0.8µR,capo
1.0µR,capo
4000

3000

2000

1000

0
10 20 30 40 50 60 70 80 90 100
Number of tests, ntest
Figure D.6: Cost-benefit analysis of CAPO tests. Total cost for performing the test and reinforcing
the component if necessary

800
0.6µR,cylinder
700 0.8µR,cylinder
1.0µR,cylinder
600 0.6µR,capo
Total cost in units, E

0.8µR,capo
500 1.0µR,capo

400

300

200

100

0
3 4 5 6 7 8 9 10
Number of tests, ntest
Figure D.7: Cost-benefit analysis of cylinder tests. Total cost for performing the test and
reinforcing the component if necessary

94
 Aalborg University

0.6µR,cylinder
3000 0.8µR,cylinder
1.0µR,cylinder
2500 0.6µR,capo
Total cost in units, E 0.8µR,capo
1.0µR,capo
2000

1500

1000

500

0
5 10 15 20 25 30 35 40 45 50
Number of tests, ntest
Figure D.8: Cost-benefit analysis of CAPO tests. Total cost for performing the test and reinforcing
the component if necessary

0.6µR,cylinder
6000 0.8µR,cylinder
1.0µR,cylinder
5000 0.6µR,capo
Total cost in units, E

0.8µR,capo
1.0µR,capo
4000

3000

2000

1000

0
10 20 30 40 50 60 70 80 90 100
Number of tests, ntest
Figure D.9: Cost-benefit analysis of CAPO tests. Total cost for performing the test and reinforcing
the component if necessary

95
 D. Cost-Benefit Analysis of Sampling Methods

D.0.2 Korrelation = 0.5

700
0.6µR,cylinder
0.8µR,cylinder
600 1.0µR,cylinder
0.6µR,capo
Total cost in units, E

500 0.8µR,capo
1.0µR,capo
400

300

200

100

0
3 4 5 6 7 8 9 10
Number of tests, ntest
Figure D.10: Cost-benefit analysis of cylinder tests. Total cost for performing the test and
reinforcing the component if necessary

2500 0.6µR,cylinder
0.8µR,cylinder
1.0µR,cylinder
2000 0.6µR,capo
Total cost in units, E

0.8µR,capo
1.0µR,capo
1500

1000

500

0
5 10 15 20 25 30 35 40 45 50
Number of tests, ntest
Figure D.11: Cost-benefit analysis of CAPO tests. Total cost for performing the test and
reinforcing the component if necessary

96
 Aalborg University

5000 0.6µR,cylinder
0.8µR,cylinder
1.0µR,cylinder
4000 0.6µR,capo
Total cost in units, E 0.8µR,capo
1.0µR,capo
3000

2000

1000

0
10 20 30 40 50 60 70 80 90 100
Number of tests, ntest
Figure D.12: Cost-benefit analysis of CAPO tests. Total cost for performing the test and
reinforcing the component if necessary

700
0.6µR,cylinder
0.8µR,cylinder
600 1.0µR,cylinder
0.6µR,capo
Total cost in units, E

500 0.8µR,capo
1.0µR,capo
400

300

200

100

0
3 4 5 6 7 8 9 10
Number of tests, ntest
Figure D.13: Cost-benefit analysis of cylinder tests. Total cost for performing the test and
reinforcing the component if necessary

97
 D. Cost-Benefit Analysis of Sampling Methods

2500 0.6µR,cylinder
0.8µR,cylinder
1.0µR,cylinder
2000 0.6µR,capo
Total cost in units, E

0.8µR,capo
1.0µR,capo
1500

1000

500

0
5 10 15 20 25 30 35 40 45 50
Number of tests, ntest
Figure D.14: Cost-benefit analysis of CAPO tests. Total cost for performing the test and
reinforcing the component if necessary

5000 0.6µR,cylinder
0.8µR,cylinder
1.0µR,cylinder
4000 0.6µR,capo
Total cost in units, E

0.8µR,capo
1.0µR,capo
3000

2000

1000

0
10 20 30 40 50 60 70 80 90 100
Number of tests, ntest
Figure D.15: Cost-benefit analysis of CAPO tests. Total cost for performing the test and
reinforcing the component if necessary

98
 Aalborg University

700
0.6µR,cylinder
0.8µR,cylinder
600 1.0µR,cylinder
0.6µR,capo
Total cost in units, E 500 0.8µR,capo
1.0µR,capo
400

300

200

100

0
3 4 5 6 7 8 9 10
Number of tests, ntest
Figure D.16: Cost-benefit analysis of cylinder tests. Total cost for performing the test and
reinforcing the component if necessary

0.6µR,cylinder
2500
0.8µR,cylinder
1.0µR,cylinder
2000 0.6µR,capo
Total cost in units, E

0.8µR,capo
1.0µR,capo
1500

1000

500

0
5 10 15 20 25 30 35 40 45 50
Number of tests, ntest
Figure D.17: Cost-benefit analysis of CAPO tests. Total cost for performing the test and
reinforcing the component if necessary

99
 D. Cost-Benefit Analysis of Sampling Methods

0.6µR,cylinder
5000
0.8µR,cylinder
1.0µR,cylinder
4000 0.6µR,capo
Total cost in units, E

0.8µR,capo
1.0µR,capo
3000

2000

1000

0
10 20 30 40 50 60 70 80 90 100
Number of tests, ntest
Figure D.18: Cost-benefit analysis of CAPO tests. Total cost for performing the test and
reinforcing the component if necessary

D.0.3 Korrelation = 0.8

0.6µR,cylinder
600 0.8µR,cylinder
1.0µR,cylinder
500 0.6µR,capo
Total cost in units, E

0.8µR,capo

400 1.0µR,capo

300

200

100

0
3 4 5 6 7 8 9 10
Number of tests, ntest
Figure D.19: Cost-benefit analysis of cylinder tests. Total cost for performing the test and
reinforcing the component if necessary

100
 Aalborg University

0.6µR,cylinder
0.8µR,cylinder
2000
1.0µR,cylinder
0.6µR,capo
Total cost in units, E 0.8µR,capo
1500 1.0µR,capo

1000

500

0
5 10 15 20 25 30 35 40 45 50
Number of tests, ntest
Figure D.20: Cost-benefit analysis of CAPO tests. Total cost for performing the test and
reinforcing the component if necessary

0.6µR,cylinder
4000
0.8µR,cylinder
3500 1.0µR,cylinder
0.6µR,capo
Total cost in units, E

3000 0.8µR,capo
1.0µR,capo
2500

2000

1500

1000

500

0
10 20 30 40 50 60 70 80 90 100
Number of tests, ntest
Figure D.21: Cost-benefit analysis of CAPO tests. Total cost for performing the test and
reinforcing the component if necessary

101
 D. Cost-Benefit Analysis of Sampling Methods

0.6µR,cylinder
600 0.8µR,cylinder
1.0µR,cylinder
500 0.6µR,capo
Total cost in units, E

0.8µR,capo

400 1.0µR,capo

300

200

100

0
3 4 5 6 7 8 9 10
Number of tests, ntest
Figure D.22: Cost-benefit analysis of cylinder tests. Total cost for performing the test and
reinforcing the component if necessary

0.6µR,cylinder
0.8µR,cylinder
2000
1.0µR,cylinder
0.6µR,capo
Total cost in units, E

0.8µR,capo
1500 1.0µR,capo

1000

500

0
5 10 15 20 25 30 35 40 45 50
Number of tests, ntest
Figure D.23: Cost-benefit analysis of CAPO tests. Total cost for performing the test and
reinforcing the component if necessary

102
 Aalborg University

0.6µR,cylinder
4000
0.8µR,cylinder

3500 1.0µR,cylinder
0.6µR,capo
Total cost in units, E 3000 0.8µR,capo
1.0µR,capo
2500

2000

1500

1000

500

0
10 20 30 40 50 60 70 80 90 100
Number of tests, ntest
Figure D.24: Cost-benefit analysis of CAPO tests. Total cost for performing the test and
reinforcing the component if necessary

0.6µR,cylinder
600 0.8µR,cylinder
1.0µR,cylinder
500 0.6µR,capo
Total cost in units, E

0.8µR,capo
1.0µR,capo
400

300

200

100

0
3 4 5 6 7 8 9 10
Number of tests, ntest
Figure D.25: Cost-benefit analysis of cylinder tests. Total cost for performing the test and
reinforcing the component if necessary

103
 D. Cost-Benefit Analysis of Sampling Methods

0.6µR,cylinder
0.8µR,cylinder
2000
1.0µR,cylinder
0.6µR,capo
Total cost in units, E

0.8µR,capo
1500 1.0µR,capo

1000

500

0
5 10 15 20 25 30 35 40 45 50
Number of tests, ntest
Figure D.26: Cost-benefit analysis of CAPO tests. Total cost for performing the test and
reinforcing the component if necessary

4000 0.6µR,cylinder
0.8µR,cylinder
3500 1.0µR,cylinder
0.6µR,capo
Total cost in units, E

3000 0.8µR,capo
1.0µR,capo
2500

2000

1500

1000

500

0
10 20 30 40 50 60 70 80 90 100
Number of tests, ntest
Figure D.27: Cost-benefit analysis of CAPO tests. Total cost for performing the test and
reinforcing the component if necessary

104
105
 E. Data by Krenchel and Bickley

Appendix

E
Data by Krenchel and Bickley

106 Figure E.1: Data by Krenchel and Bickley from 1987 for cylinder and CAPO-tests.