See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/267241230

RELIABILITY OF TIMBER STRUCTURES

Article

CITATIONS READS

67 6,593

1 author:

Jochen Köhler
Norwegian University of Science and Technology
112 PUBLICATIONS 1,827 CITATIONS

SEE PROFILE

All content following this page was uploaded by Jochen Köhler on 17 April 2015.

The user has requested enhancement of the downloaded file.

RELIABILITY OF TIMBER STRUCTURES

Jochen Köhler

Institute of Structural Engineering

Swiss Federal Institute of Technology

Zurich
May 2007
FOREWORD
In the last few years the interest in designing timber structures has steadily increased. The
reason for this being an increased focus in society on sustainability and environmental aspects
but also due to the positive effects on the inner climate in accommodation buildings and the
increased architectural possibilities. Furthermore, timber is technically and economically
competitive compared with steel and concrete as a building material for a broad range of
normal building structures such as e.g. accommodation buildings.
So far the basis for design of timber structures has by far and large not achieved the same
level of refinement and detail as the basis for the design of steel and concrete structures for
several reasons. First of all the variability of the material properties is much higher than for
other building materials; the raw timber material is not engineered but the result of natural
processes. The material properties cannot be designed as for other materials but must be
ensured by quality control schemes. Secondly the material properties, and therefore the
reliability, depend on the whole load and moisture history of the structure. These two effects
interact in a complicated manner for the timber materials used in timber structures, structural
timber, glued laminated timber, panel-products together with the joints between them.
The result of this being that the design of timber structures to a large degree is based on
experience, subjective engineering judgement and in many cases excessive conservative
assumptions. However, despite an in general conservative attitude the present basis for design
of timber structures does not consistently account for the uncertainties influencing the
structural performance. As a consequence hereof the presently applied basis for the design of
timber structures may lead to designs which in terms of reliability are not comparable to
equivalent structures made of steel and concrete.
The main contribution of the present thesis by PhD J. Köhler addresses the problem complex
outlined in the above and can be summarized as to establish a probabilistic model framework
(or probabilistic model code) for the design of timber structures. In this process special
emphasis is directed on the aspects of consistent modelling of the performance of timber
components subjected to different types of loading and environmental exposures.
In the thesis a rather rigorous phenomenological and hierarchical modelling of uncertainties
associated with the characteristic of timber materials in dependency of scale, applied models
and available information is developed. On this basis specific models are developed and/or
further extended with the aim to represent the performance of timber components and joints in
consistency with experimental data and engineering understanding. The introduced models
are rigorously analysed and simplified to the extent reasonable and relevant with due
consideration to their practical applicability.
The probabilistic model code developed within the present thesis has by now already been

i
adopted by the Joint Committee on Structural Safety. In this way the thesis has been ensured a
significant impact in a pre-normative context and thereby contributes to the future increased
rational, efficient and sustainable use of timber as a building material

Zürich, May 2007 Prof. Dr. M. H. Faber

ii
SUMMARY
During the last decades structural reliability methods have been further developed, refined and
adapted and are now at a stage where they are being applied in practical engineering problems
as a decision support tool in connection with design and assessment of structures. For
materials such as concrete and steel, the application of modern structural reliability methods
has led to an increasingly more consistent evaluation of the safety or reliability. Whereas
some efforts in this direction have been undertaken also for timber, the developments,
however, have been less impressive. One of the main reasons for this is that the variability of
the timber material properties is much higher than for other building materials. Furthermore,
the timber material properties depend on the entire load and moisture history of the structure.
It is important that a consistent basis for design of timber structures is established and
documented in such a way that it may be accepted for implementation by the timber
engineering and research community. The development of a consistent basis of design of
timber structures is the focal point of this thesis.
The proposed basis of design is structured into several levels of sophistication. The basic level
reflects the recent practice for reliability based code calibration. The bending strength and
stiffness and the density of timber are referred to as reference material properties and are
introduced as simple random variables. The basic limit state functions for components and
connections are given. Furthermore, proposals are made regarding the different characteristics
of timber on this simple level. Functional relationships for other material properties (based on
the reference material properties) are given and probability distribution functions for the other
material properties are proposed. Starting from this level, several possible refinements are
proposed. New information might be introduced, and it is shown how different types of new
information can be integrated by using a Bayesian updating scheme. Refinements in regard to
the modelling of damage as a consequence of time load duration are proposed. For the
bending strength, a hierarchical spatial variability model is proposed and a method is
presented for linking the properties of a cross section (which is considered as the reference
starting point for the modelling of spatial variability) with the properties of a test specimen.
The main outcomes of this thesis are related to necessary pre-codification modelling aspects
concerning the reliability of timber components in regard to strength and stiffness properties.
An achievement of this thesis is that the work performed is fully compatible with the general
probabilistic framework for establishing design basis developed by the Joint Committee on
Structural Safety (JCSS).

iii
iv
ZUSAMMENFASSUNG
Die Methoden zur Beurteilung der Bauwerkszuverlässigkeit wurden in den letzten
Jahrzehnten weiterentwickelt, verbessert und angepasst, so dass sie heute im praktischen
Ingenieurwesen ihre Anwendung finden; zum Beispiel als Entscheidungshilfe bei dem
Entwurf und dem Unterhalt von Bauwerken. Für Bauwerke aus Stahl und Beton führte die
Anwendung dieser Methoden zu einer konsistenteren Beurteilung der Sicherheit und der
Zuverlässigkeit. Obschon für Bauwerke aus Holz einige Anstrengungen unternommen
wurden, diese Methoden zur Anwendung zu bringen, sind die erreichten Entwicklungen im
Vergleich zu anderen Baumaterialien weniger fortgeschritten. Einer der Hauptgründe für
diesen Unterschied ist die hohe Komplexität des Baumaterials Holz; um das volle Potential
des Baumaterials Holz auszuschöpfen ist ein hohes Mass an Fachkenntnis erforderlich.
Es ist wichtig, dass eine konsistente Basis für die Bemessung von Holzkonstruktionen
entwickelt wird, die breite Anwendung auf dem Gebiet des Ingenieurholzbaus findet. Die
Entwicklung einer solchen Bemessungsbasis steht im Focus dieser Dissertation.
Die hier vorgeschlagene Bemessungsbasis ist in mehrere Modellierungsstufen aufgeteilt. Die
Grundstufe repräsentiert die allgemein gängige Praxis bei der Kalibrierung von Bemessungs-
richtlinien. Die Biegefestigkeit, das Elastizitätsmodul in Biegung und die Holzdichte sind als
Referenzmaterialeigenschaften definiert und werden als Zufallsvariablen eingeführt. Einige
grundlegende Grenzzustandsfunktionen für Komponenten und Verbindungen sind angegeben.
Des Weiteren werden Vorschläge zur Modellierung der speziellen Holzmaterialeigenschaften
gemacht. Andere Holzmaterialeigenschaften werden basierend auf den Referenzmaterial-
eigenschaften ermittelt und ebenfalls als Zufallsvariablen eingeführt. Ausgehend von dieser
Grundstufe werden mehrere Modellerweiterungen vorgeschlagen. Es wird aufgezeigt, wie
verschiedenartige neue Information dazu benutzt werden kann, die Parameter der
vorgeschlagenen Modelle anzupassen. Modellverbesserungen in Bezug auf die
Berücksichtigung der Schadensakkumulierung infolge von Langzeitbeanspruchung werden
vorgeschlagen. Ein hierarchisches Model, welches die räumliche Variabilität der
Biegefestigkeit berücksichtigt, wird entwickelt. Dieses Modell dient als Grundlage für eine
Methode, die die Biegefestigkeit eines Querschnitts mit der Biegefestigkeit eines
Probekörpers in Beziehung setzt.
Die vorgestellte Bemessungsbasis soll den Ausgangspunkt für die Weiterentwicklung von
Bemessungsrichtlinien wie den Eurocode 5 und die SIA 265 bilden und den effizienten
Gebrauch von Holz als Baumaterial ermöglichen. Die Ergebnisse dieser Arbeit sind
kompatibel mit dem allgemeinen wahrscheinlichkeitsbasierten Rahmenwerk für die
Entwicklung von Bemessungsrichtlinien, dem Probabilistischen Model Code, der vom Joint
Committee on Structural Safety (JCSS) entwickelt und herausgegeben wurde.

v
vi
TABLE OF CONTENTS

1 INTRODUCTION ...................................................................................... 1
1.1 Scope of Work and Limitations ......................................................................... 2
1.2 Outline and Thesis Overview............................................................................. 3

2 Aspects of Structural Reliability................................................................. 5

2.1 The Limit State Principle ................................................................................... 5
2.2 Basic Principles of Reliability Assessment........................................................ 8
2.2.1 Uncertainties in Reliability Assessment................................................. 8
2.2.2 Basic Variables and their Quantification ............................................... 8
2.3 Methods of Structural Reliability..................................................................... 15
2.3.1 First Order Reliability Method (FORM).............................................. 15
2.3.2 Monte Carlo Simulation Methods ........................................................ 18
2.4 Design Formats in Structural Engineering ....................................................... 19
2.4.1 Allowable Stress Format – the Safety Factor....................................... 19
2.4.2 Load and Resistance Factor Design Format – Partial Safety Factors .. 20
2.5 Reliability Based Code Calibration.................................................................. 20
2.5.1 Simple Code Calibration with given Target Reliability....................... 21
2.5.2 Target Reliabilities ............................................................................... 22

3 Timber as a Structural Material ................................................................ 25

3.1 The Tree ........................................................................................................... 25
3.2 Wood ................................................................................................................ 27
3.2.1 Cells and Fibres – The Ultrastructure of Wood ................................... 27
3.2.2 Growth Rings and Vessels – The Microstructure of Wood ................. 28
3.2.3 Knots and Grain Deviations – The Meso Structure of Wood .............. 30
3.2.4 Clear Wood Specimens and their Properties........................................ 31
3.3 Structural Timber ............................................................................................. 35
3.3.1 Growth Irregularities in Timber Structural Components ..................... 35
3.3.2 Material Properties of Interest of Timber Structural Components....... 36
3.4 Timber Design.................................................................................................. 41
3.4.1 Design Format ...................................................................................... 41

vii
 3.4.2 Timber Grading .................................................................................... 42
3.4.3 Strength Class Systems ........................................................................ 42

4 Probabilistic Modelling of Timber Material Properties............................ 45

4.1 Introduction ...................................................................................................... 45
4.2 Spatial Variations of Timber Material Properties ............................................ 46
4.2.1 Scales of Modelling.............................................................................. 46
4.2.2 Modelling Reference Material Properties under Reference Conditions
– Meso Level Variations .................................................................................. 48
4.2.3 Macro Level Variations – Timber Grading.......................................... 48
4.2.4 Within Member Variations – Micro Scale Variations ......................... 73
4.3 Duration of Load Effects.................................................................................. 84
4.3.1 Assessment of Duration of Load Experimental Data........................... 84
4.3.2 Modelling the Duration of Load Effect................................................ 87
4.3.3 Probabilistic Modelling of Duration of Load Effects........................... 91
4.3.4 Applications for DOL Models............................................................ 105
4.3.5 Summary and Concluding Remarks, DOL Effects ............................ 117
4.4 Moisture Effects on the Duration of Load Effect........................................... 118
4.5 Interrelation of Material Properties ................................................................ 119

5 Probabilistic Modelling of the Properties of Timber Connections......... 123

5.1 Joints with Dowel Type Fasteners ................................................................. 124
5.1.1 Single Fastener Connections – Johansen’s Yield Theory .................. 126
5.1.2 Single Fastener Connections – Jorissen’s Splitting Mode ................. 128
5.1.3 Relevant Material Properties.............................................................. 129
5.1.4 The Behaviour of Single Fastener Connections................................. 137
5.1.5 Design Framework for Single Dowel Type Fastener Connections.... 138
5.2 Effective Number of Fasteners....................................................................... 142
5.3 Probabilistic Model for the Load Bearing Capacity of Dowel Type Fastener
Connections................................................................................................................ 147
5.3.1 Model Sensitivities............................................................................. 149
5.3.2 Model Verification ............................................................................. 153
5.3.3 Evaluation of the Model Uncertainty ................................................. 160
5.3.4 Limit State Functions for a Selected Model Alternative.................... 163

viii
 5.4 Summary and Concluding Remarks, Timber Connections............................ 165

6 Probabilistic Model Code - Timber ........................................................ 167

6.1 Scope and Limitations of the Proposal........................................................... 167
6.2 Model Framework .......................................................................................... 167
6.2.1 Basic Properties.................................................................................. 168
6.2.2 Typical Ultimate Limit States ............................................................ 170
6.2.3 Typical Serviceability Limit States .................................................... 170
6.3 Simple Probabilistic Model............................................................................ 171
6.3.1 Basic Properties.................................................................................. 171
6.3.2 Correlation Matrix.............................................................................. 172
6.3.3 Strength and Stiffness Modification Functions.................................. 172
6.3.4 Model Uncertainties for Different Ultimate Limit States .................. 173
6.4 Possible Refinements ..................................................................................... 174
6.4.1 Modelling of the Spatial Variation of Timber Properties .................. 174
6.4.2 Duration of Load Effect ..................................................................... 176
6.4.3 Updating Scheme for the Basic Properties......................................... 177
6.5 Concluding Remarks ...................................................................................... 181

7 Applications ............................................................................................ 183

7.1 Assessment of Data ........................................................................................ 183
7.1.1 Non-Parametric Assessment .............................................................. 183
7.1.2 Selection of a Distribution Function and Parameters......................... 185
7.1.3 Estimation of the 5%- fractile value................................................... 187
7.2 Updating of Random Variables...................................................................... 191
7.2.1 Normal Distribution with Uncertain Mean and Known Standard
Deviation ........................................................................................................ 191
7.2.2 Normal Distribution with Uncertain Mean and Standard Deviation.. 193
7.3 Comparison of Different Bending Test Specifications .................................. 196
7.3.1 European Standards EN 408 and EN 384 .......................................... 196
7.3.2 American Society for Testing and Materials, ASTM D 4761-88 and D
1990-91........................................................................................................... 197
7.3.3 Australian/New Zealand Standard, AS/NZ 4063:1992...................... 197
7.3.4 Evaluating the Effect of different Test Standards .............................. 197

ix
8 Conclusions and Outlook........................................................................ 201
8.1 Conclusions .................................................................................................... 201
8.1.1 Retrospect and Motivation ................................................................. 201
8.1.2 Approach and Summary..................................................................... 202
8.1.3 Originality of Work............................................................................ 203
8.1.4 Limitations ......................................................................................... 204
8.2 Outlook........................................................................................................... 205

References ......................................................................................................... 207

Annex

x
1 INTRODUCTION

Timber is an efficient building material, not only in regard to its mechanical properties but
also because it is a highly sustainable material considering all phases of the life cycle of
timber structures; production, use and decommissioning. Timber is a widely available natural
resource throughout Europe; with proper management, there is a potential for a continuous
and sustainable supply of raw timber material in the future. Due to the low energy use and the
low level of pollution associated with the manufacturing of timber structures, the
environmental impact of timber structures is much smaller than for structures built using other
building materials. In addition timber, is a rather advantageous building material due to its
material properties. Timber is a light material and, compared to its weight, the strength is
high; the strength/weight ratio is even higher than for steel.
However, timber is still not utilized to its full potential in the building and construction sector
considering its beneficial properties. Many building owners, but also architects and structural
engineers, do not consider timber as a competitive building material compared to concrete,
steel or masonry. Attributes such as high performance in regard to reliability, serviceability
and durability are generally not associated with timber as a building material. One of the main
reasons for this is that timber is a highly complex material; it actually requires a significant
amount of expertise to fully appreciate the potential of timber as a structural building material.
In addition to this there are still a number of issues which need to be further researched before
timber materials can achieve the same recognition as a high quality building material such as
e.g. steel and concrete.
In daily practice the engineering codes and regulations form the premises for the use of timber
as a structural material. It is therefore of utmost importance that the codes and regulations are
based on the most relevant and exact information available in regard to the reliability of
timber structures. Traditionally, codes and regulations have been based to a very large degree
on experience. This statement is true not just for timber structures, but also for concrete and
steel structures. However, whereas the codes and regulations for the design of concrete and
steel have undergone a remarkable modernisation through the last 2-3 decades, the codes and
regulations for the design of timber structures are still falling significantly behind. The
principle for the development of the scientific basis for codes and regulations for the design
and assessment of structures is illustrated in Figure 1-1.
Whereas the various steps in the process of developing a scheme for design and assessment
are illustrated to the left in the figure, the scientific constituents required in the process are
illustrated in the right part of the figure. For identified design and assessment situations,
models for describing the performance of structures are derived. The models take basis in
physical hypotheses and experimental evidence. In general these models are associated with
uncertainties and an important task within the modelling process is to take into account these
uncertainties in a consistent manner. A set of probabilistic models, which reflects the current

1
best engineering practice, may be specified in so-called (probabilistic) model codes, which
build the basis for all further simplified design codes.

Risk Screening
(Assessment of Failures and Malfunctions)

Component System
aspects aspects
Identification of Design and
Assessment Situations Exposures
(Loads and Enviroment)

Evidence from
Model Building Experiments
& Verification
Physical
Hypothesis

Basis of Quantified Models

Design & Uncertainty

Codes

Detailed Quantification of Design and

Assessment Strategies

Figure 1-1 Schematic representation of the framework required for establishing the basis for
codes and regulations for design and assessment of structures.

Presently the existing codes and regulations for the design of timber structures are not fully
based on a framework as illustrated in Figure 1-1, but rather on a framework taking basis in
general physical understanding combined with experience achieved through hundreds of years
of use of timber as a structural material. In order to achieve the goal of optimizing the
potential benefit of timber as a building material, it is thus necessary to establish a firm
scientific basis for codes and regulations according to the process indicated in Figure 1-1.

1.1 SCOPE OF WORK AND LIMITATIONS

The main objective of the thesis is the development of a basis of design for timber structures.
It is considered that timber is a material with special properties in regard to its high variability
and its sensitivity to loading mode, load duration and changes in the surrounding climate. The
schematic layout given in the right part of Figure 1-1 is used as a working thesis, focus is

2
placed on modelling aspects of timber components exposed to different load types and
environmental conditions. This includes:
x a discussion of existing modelling approaches for the various aspects of timber
components,
x a discussion of modelling alternatives,
x the selection of the most promising models,
x the calibration of parameters,
x the presentation of selected models and parameters in a consistent format,
x example calculations.
An important issue for the design of timber structures is the behaviour of structural systems
made of a number of different components. These aspects are not considered explicitly within
this thesis. However, the discussion on spatial variability of timber material properties and the
respective modelling proposal delivers the basis for further research of the system behaviour
of timber structures.

1.2 OUTLINE AND THESIS OVERVIEW

A thesis overview is given in Figure 1-2. In chapters 2 and 3 a brief introduction into the
fields of structural reliability and timber engineering is given. Existing knowledge is reviewed
and summarized. Focus is placed on the issues which are considered as a relevant basis for the
subsequent chapters. Chapters 4 and 5 are the development parts of this thesis. In chapter 4
the focus is placed on the probabilistic modelling of timber material properties; in chapter 5
the modelling of timber connections with dowel type fasteners is discussed. These two
chapters build the basis for the development of a proposal of a probabilistic model code for
timber structures that is presented in chapter 6. In chapter 7 some further examples are given
regarding e.g. the assessment of experimental data, Bayesian updating, etc. The thesis is
concluded with chapter 8, where the main content is summarized and a discussion of proposed
further research is given.

3
Chapter 1

Introduction

Chapter 2 Chapter 3
Tools

Basic Principles of Structural Reliability Timber as a Structural Material

Limit State Principle J Uncertainty J Tree J Wood J Timber J Timber Design

Basic Variables J Methods J Code Calibration
Component

Chapter 4 Chapter 5

Modelling Timber Material Properties Modelling Timber Connections

Grading J Spatial Variability Dowel Type J Mechanical Models J

J DOL J Framework Material Properties J Code Formats J Model Proposal
JVerification J Model Uncertainty
Inference

Chapter 6 Chapter 7

Parameter Estimation
Probabilistic Model Code Applications Influence of Test Standards
Updating

Chapter 8

Conclusions and Outlook

Figure 1-2 Schematic overview of the thesis.

4
2 ASPECTS OF STRUCTURAL RELIABILITY

Typical problems in structural engineering such as design, assessment, inspection and

maintenance planning are decision problems subject to a combination of inherent, modelling
and statistical uncertainties. The structural reliability theory is concerned with the rational
treatment of these uncertainties. In the subsequent chapter some basic aspects of structural
reliability are introduced, whereas the content of the chapter is based on standard structural
reliability literature, as Madsen et al. (1986), Melchers (1999) and Faber (2003). The
interested reader is also referred to these publications for more detailed and complete
information.

2.1 THE LIMIT STATE PRINCIPLE

The performance of an engineering structure depends on the type and magnitude of the
applied load and the structural strength and stiffness. Whether the performance is considered
satisfactory depends on the requirements which must be satisfied. Among others these include
reliability of the structure against collapse, limitation of damages or of deflections, or other
criteria. In general any state, that may be associated with consequences in terms of costs, loss
of lives and impact to the environment are of interest. In the following it is not differentiated
between these different types of states but for simplicity refer to all these as being failure
events.
It is convenient to describe failure events in terms of functional relations, which if they are
fulfilled, define that the failure event F will occur:

F ^g x d 0` (2.1)

where g x is termed a limit state function. The components of the vector x are the
realisations of the so-called basic random variables X representing all relevant uncertainties
influencing the problem at hand. The failure event F is defined as the set of realisations of the
limit state function g x , which are zero or negative.
Some typical limit states are given in Table 2-1.

5
Table 2-1 Typical limit states for structures.
Limit State Type Description Examples

Rupture, plastic mechanism, instability,

Ultimate Collapse of the structure or part of it progressive collapse, fatigue, deterioration,
fire.

Excessive permanent cracking, permanent

Damage (often included in above)
irreversible deformation.

Serviceability Disruption of normal use Excessive deflection, vibration, local damage.

Due to the associated consequences, failure events linked with the most serious limit states
such as collapse and major damage should be relatively rare events. The study of structural
reliability is concerned with the assessment of the probability of failure events for engineered
structures at any stage during its service life.
The probability of occurrence of a failure event is a measure of the chance of its occurrence.
This chance may be quantified by observing the long term frequency of the event for
generally similar structures or components, or may be simply a subjective estimate of its
numerical value. Engineering structures are mostly exclusive in regard to the structural
assembly and their exposure to loads and environment which in general precludes the
assignment of relative frequencies of events within many similar structures. However, a more
generic description can be given for structural components and material. Thus in practice a
combination of subjective estimates and frequency observations about structural components
and structural assemblies is utilized to assess the probability of limit state violation of a
structure.
The probability of failure p f may be determined by the following integral:

pf P g X d0 ³ f X x dx
(2.2)
g x d0

where g X is the limit state function, X is a vector of basic random variables and f X . is
the joint probability function of the variables X .
In general the basic variables X are functions of time. E.g. the loads which are applied to a
structure are varying in time and are of uncertain magnitude at any point in time;
consequently the situation is similar for the corresponding load effects s . The resistance of a
structure or component r is also a function of time; commonly a structure is subject to some
(random) deterioration process and the resistance is decreasing with time. The situation is
illustrated in Figure 2-1. Both, the load effect and the resistance can be represented as random
processes and probability density functions for given times t ta , tb can be formulated.

6
 R, S

fR r t ta
fR r t tb
r t

s t

fS s t tb
fS s t ta
ta tb t

Figure 2-1 Schematic time-dependent reliability problem.

A situation as illustrated in Figure 2-1 is referred to as the time-dependent reliability problem.

Failure occurs when the process g x t d 0 for some time, t , during the considered time
interval, > 0; T @ . x t is a realisation of the random process ^X t ` .The probability of failure
p f in the interval > 0; T @ may be determined as:

pf T 1 P g X t ! 0, t  > 0; T @ (2.3)

The evaluation of Equation (2.3) is in general difficult, and approximations are used in
practical applications. Often an upper bound of the probability of failure in the time interval
>0; T @ is used:
T

pf T ³Q
0
t , [ dt (2.4)

where the out-crossing rate Q  t , [ is determined by a suitable application of Rice’s formula,

see e.g. Madsen et al. (1986).
In many situations it is convenient (and sufficient) to assume that the basic variables X are
not depending on the time t . Then Equation (2.2) can be used for the determination of the
probability of failure.

7
2.2 BASIC PRINCIPLES OF RELIABILITY ASSESSMENT

The aim of structural reliability assessment is to quantify the reliability of structures under the
consideration of all uncertainties associated with the formulation of the failure criteria. The
formal framework for the calculation of the failure probability is given in Equation (2.2),
whereas the term reliability is defined as the complement of the failure probability 1  p f .
A failure criterion is expressed through the limit state function g X .

2.2.1 UNCERTAINTIES IN RELIABILITY ASSESSMENT

The limit state function g X is formulated by means of models based on physical

understanding and empirical data. Due to idealizations, inherent physical uncertainties and
inadequate or insufficient data, the models themselves and the parameters entering the models
such as material properties and load characteristics are uncertain. Consequently, uncertainties
are grouped into:
x Inherent Uncertainties,
x Model Uncertainties,
x Statistical Uncertainties.
Inherent uncertainty refers to the randomness of a phenomenon. This randomness is a result
from a combination of uncontrollable fluctuations of many different factors. An example is a
strength property of a structural timber element, which is a product of several quality control
procedures during its production, but also growing conditions, sawing pattern and many other
factors in the production line of the timber structural element. Another example for the
randomness is the realization of loads on structures due to e.g. snow or wind events. The
described type of uncertainty is also referred to as physical uncertainty.
Model uncertainty is associated with the crudeness and incompleteness of mathematical
models which describe a phenomenon.
The statistical uncertainties are associated with the statistical evaluation of test results or
observations. They may result from limited numbers of tests or observations which cause
uncertainty in the estimation of statistical parameters.

2.2.2 BASIC VARIABLES AND THEIR QUANTIFICATION

In structural reliability assessment the set of basic variables of a problem in general is

constituted of both random and deterministic variables for the geometry, material properties
and load characteristics. In this section only very general remarks are made in regard to the
different physical characteristics of the different types of random basic variables. In general,
the quantification of the basic random variables can be divided into three parts:

8
x Definition of the considered populations.
x Selection of a suitable types or families of probability distribution for the basic variable.
x Estimation of suitable distribution parameters from available data and any prior
knowledge.

2.2.2.1 Population
The basic random variables should always be related to a meaningful and consistent set of
populations. The description and modelling of these random variables should correspond to
this set. A reliability analysis based on these random variables is only valid for the considered
set of populations. The basis for the definition of a population is in most cases the physical
background of the quantity. Factors which define a population are the nature and the origin of
the random quantity (e.g. strength, load or geometry), the spatial characteristics (e.g. size of
structural component, geographical origin of the considered material, regional wind speed
characteristics) and temporal conditions (e.g. duration of exposure). The choice of
specifications which define a population may depend on the objective of the analysis, the
amount and nature of the available data and the amount of resources which can be afforded. A
population with a unique set of specifications is referred to as elementary population; a
population in which specification parameters vary is referred to as a composite population.
The set of measurements associated with a certain population is referred to as an elementary
or composite sample respectively. A sampling procedure may be representative or artificial.
Representative samples or representative realisations of random variables are obtained
through random sampling. Artificial means that no direct relation exist between the statistical
properties of the sample and the statistical properties of the population. An artificial sample is
e.g. when only weak specimen are selected for testing by engineering judgment or proof
loading. Artificial samples are also termed censored samples.
Observations on a representative sample may be undertaken according to a standardised test
procedure. Hereby the test standard specifies partly the population; e.g. if a sample of timber
specimen is tested in bending and all spatial and temporary conditions are specified. Then, the
statistical properties of the sample are assumed to be the same as the statistical properties of
the population.
The statistical properties of the sample are described by a suitable probability distribution
function. The physical characteristic of the random variable determines the possible type of
distribution function.

2.2.2.2 Selecting a Suitable Probabilistic Model

Resistance Variables
The resistance of a structure is governed by so-called resistance variables. Dimensions,

9
geometrical imperfections and material properties are resistance variables. Resistance
variables can be modelled as random properties by random variables. Assuming that the
random properties belong to an elementary population, standard probability distribution
models can be utilised to represent the random variables. The following distribution models
are generally used to model resistance variables:
Normal Distribution: This is one of the most important probability distribution. It can be
shown that a sum of many independent random properties gets normal distributed. This
property is also known as the central limit theorem; see e.g. Benjamin and Cornell (1970).
The normal distribution gives finite probability for negative values. For that reason the
logarithm of the normal distribution is often preferred for modelling resistance variables.
Lognormal Distribution: The lognormal distribution arises naturally as a limiting distribution
when the random resistance is a product of a number of independent random quantities. The
lognormal distribution is frequently used to model resistance variables because it is
precluding negative values.
Weibull Distribution: For strength related material properties the Weibull distribution is used
quite frequently. It is based on the assumption that a structural body is composed of nearly an
infinite number of elements. The strength of the elements is independent and identically
distributed and the strength of the material body is assumed to be equivalent to the strength of
the weakest element. The theory behind the Weibull distribution is described in more detail in
section 4.2.4.2.
The distribution function for the Normal, Lognormal and the Weibull distribution is given in
Annex A, Table A1.
Other distributions: A number of other distributions are sometimes used to model resistance
variables; e.g. the beta-, rectangular-, t- or gamma distribution. For information about these
models it is referred to standard literature, e.g. Benjamin and Cornell (1970).
Caution is necessary if the considered sample is not homogeneous, i.e. if the source of the
considered specimens is not the same. When data from one or more sources is analysed as a
single sample, the shape of the distribution function is likely depending equally on relative
number taken from each source than on the actual, but unknown, distribution function of
every single source.

Load Variables
Loads are generally understood as forces acting on a structure which arise from external
influences, e.g. self weight, snow load, wind force, etc. Imposed deformations as dimensional
changes arising from temperature or humidity changes are also considered as loads. In a
structural analysis, so-called load effects are considered directly. They are commonly a
combination of different variables, as e.g. roof shape, snow load, wind exposition,
dimensions, etc., which make the effect of the snow load on the structure. This effect is

10
measured in the same dimension as the resistance (e.g. stress) and can be directly compared.
Load effects can be classified as follows in:
x permanent or variable (variability in magnitude with time),
x fixed or free (variability in position with time),
x static or dynamic (the nature of the induced structural response).
It should be noted that similar loads can be considered as different load effects according to
the classification made above. (E.g. a traffic load on a bridge can be seen as the effect of the
static weight of the heaviest truck or as the magnitude and number of load cycles caused by
passing vehicles, etc.).
Permanent Loads: Permanent loads on structures can be seen as a sum of dead load of many
different components of the structure and other parts. Therefore, the random character of
permanent loads is well represented by a normal distribution. Furthermore, it can be shown
that the coefficient of variation of the weight of the sum of many (independent) components is
decreasing with the number of components considered.
Variable Loads: Most loads on structures are varying in time. Examples are wind loads, snow
loads or traffic loads. Time varying loads can be modelled with random processes. Different
types of random processes are illustrated in Figure 2-2. Very often structural analysis is
simplified to be time invariant. Then only the distribution of the maximum realisations of a
process in a specified time interval is of interest. Extreme value distributions can be utilised to
model these extreme realisations. Several extreme value distributions are well described in the
literature, e.g. Benjamin and Cornell (1970).

S t S t S t

t t t

Figure 2-2 Different types of random load processes.

Load combinations
An important issue when considering load modelling is the representation of the extremes of
combinations of different load effects – the load combination problem. Different individually
acting loads may be modelled as the sum of the load effect processes X i t and the
maximum X max T is calculated by:

X max T max ^ X 1 t  X 2 t  ...  X n t ` (2.5)

T

11
A general solution to Equation (2.5) is hardly attainable but solutions exist for special cases of
continuous load processes and different types of non-continuous processes, see e.g. Thoft-
Christensen and Baker (1982) or Melchers (1999). Some approximate solutions to Equation
(2.5) exist and the most widely used approaches, the Turkstra Load Combination Rule and the
Ferry Borges – Castanheta Load Combination Rule are briefly introduced next.
The Turkstra Load Combination Rule
When n loads are combined it seems to be clear that the event of all individual loads attain
the maximum at the same point in time is highly unlikely; even if the number of different
loads n considered is high. However, it can be formulated as an upper bound for the
maximum of n combined loads:

X max T d max ^ X 1 t `  max ^ X 2 t `  ...  max ^ X n t ` (2.6)

T T T

This formulation is considered as too conservative and alternatively it is proposed in Turkstra

(1970) to evaluate the maximum load for the individual loads for the given reference period
and combining them in accordance with the scheme shown in Equation (2.7).

Z1 max ^ X 1 (t )`  X 2 (t )  X 3 (t )  ...  X n (t )
T

Z2 X 1 (t )  max ^ X 2 (t )`  X 3 (t )  ...  X n (t )
T
(2.7)

Zn X 1 (t )  X 2 (t )  X 3 (t )  ...  max ^ X n (t )`
T

and approximating the maximum combined load X max (T ) by

X max ( T ) | max^Z i ` (2.8)

i

X i (t ) is an arbitrary point in time value of X i (t ) and in general taken into account as the
mean value, provided that the process is stationary.
This approximation is called Turkstra’s rule and is commonly used as a basis for codified load
combination rules.
The Ferry Borges – Castanheta Load Combination Rule
A more sophisticated approximation to the load combination problem is based on the load
model is suggested in Ferry Borges and Castanheta (1971). A highly simplified representation
of the real load processes is utilized which facilitates a solution of the load combination
problem as defined by Equation (2.5) by use of modern reliability methods such as FORM
outlined in section 2.3.1.

12
 x1 t

0 1 t

x2 t

0 2 t

x3 t

0 3 T t

Figure 2-3 Illustration of the Ferry Borges – Castanheta Load Process.

For each process X i t it is assumed that the load changes after equal so-called elementary
intervals of time W i . Further it is assumed that the load is constant within each elementary
time interval. This is illustrated in Figure 2-3 where the reference period T has been divided
into ni intervals of equal length W i T / ni . The integer ni is called the repetition number for
the i th load. The loads in the elementary time intervals are assumed to be identically
distributed and mutually independent random variables with a point in time probability
distribution function FX i xi . The ni pulses of the process may be understood as a vector of
mutually independent random variables.
When combinations of load processes X 1 t , X 2 t , …, X n t are considered it is assumed
that the loads are stochastically independent with repetition numbers ni , where:

n1 d n2 d ... d ni d ... d nr (2.9)

and

ni ni 1  =  for i 2,3,..., j (2.10)

where =  is the set of positive natural numbers. E.g., in Figure 2-3 j 3 and n1 2 , n2 6
and n3 12 .
The distribution function of the maximum value of a period with m repetitions is calculated

13
as:

m
Fmax X i xi FX i xi (2.11)
T

According to Ferry Borges – Castanheta different combinations of individual load processes i

have to be taken into account, whereas the number of repetitions is varying and depending on
the number of elementary time intervals ni . The following combination rules are for the case
of 3 different load sequences ( j 3 ):

Table 2-2 Load combination rules for 3 different load sequences ( j 3 ).

No. of repetitions m
Combination No.
Load 1 Load 2 Load 3

1 n1 n2 n1 n3 n2

2 1 n2 n3 n2

3 n1 1 n3 n1

4 1 1 n3

I.e. for the load distributions FX1 x , FX 2 x , FX 3 x of the loads 1, 2, 3 in Figure 2-3, the
following combinations have to be considered:

2 3 2
1: Z1 FX1 x  FX 2 x  FX 3 x

1 6 2
2: Z2 FX1 x  FX 2 x  FX 3 x

2 1 6
3: Z3 FX1 x  FX 2 x  FX 3 x

1 1 12
4: Z4 FX1 x  FX 2 x  FX 3 x

The governing maximum load is X max T | max ^Z i ` .

i

2.2.2.3 Estimation of Distribution Parameters

An important task within the framework of reliability analysis and assessment is the
quantification of model parameters based on observations of the quantity to be modelled. In
general the observations are performed through measurements of the quantity in the context of
a specified test configuration or through continuous observations of the state of nature. The

14
output of these measurements is a set of numerical data, the so-called sample values. It is
assumed that the sample values are realisations of the random variable X . The statistical
properties of X can be assessed considering the sample values. Several methods can be found
in the literature, e.g. a good overview of the methods is given in Benjamin and Cornell
(1970). The most important methods are outlined in Annex A.

2.3 METHODS OF STRUCTURAL RELIABILITY

Several methods can be found in the literature to calculate the probability failure by solving
the integral in Equation (2.2). An overview and a discussion of different approaches is
presented in e.g. Ditlevsen and Madsen (1996), Melchers (1999) and Faber (2003). The most
straightforward method is that of Monte Carlo simulation, while probably the more efficient
are the so called approximate methods based on the calculation of the reliability index E
(FORM and SORM). There are also methods to increase the efficiency of the Monte Carlo
simulation like Importance Sampling or Adaptive Sampling. In this section only the main
ideas of FORM (First Order Reliability Method) and the Monte Carlo Simulation are briefly
outlined. For further information also about the other methods it is referred to standard
literature, e.g. Melchers (1999).

2.3.1 FIRST ORDER RELIABILITY METHOD (FORM)

2.3.1.1 Linear Limit State Functions and Normal Distributed Variables

At first a simple case where the limit state function g x is a linear function of the basic
variables X is considered. Then the limit state function can be written as:

n
g x a0  ¦ ai xi (2.12)
i 1

If the basic random variables are normally distributed the so-called safety margin M is
defined as:

n
M a0  ¦ ai X i (2.13)
i 1

and is normal distributed with mean value P M and standard deviation V M as:

n
PM a0  ¦ ai P X i (2.14)
i 1

and

15
 n n n
V M2 a0  ¦ aiV X2 i  ¦ ¦ Uij ai a jV X V X
i j
(2.15)
i 1 i 1 j 1, j z ì

where Uij are the correlation coefficients between the variables X i and X j . Defining the
failure event as in Equation (2.1), the probability of failure can be defined as:

pf P g X d0 P M d0 (2.16)

which in this case reduces to the evaluation of the standard normal distribution function, as:

pf ) E (2.17)

Where E is the so-called reliability index (due to Cornell (1969) and Basler (1961)) given as:

PM
E (2.18)
VM
If the two dimensional case of two independent normal distributed random variables is
considered the reliability index E has a geometrical interpretation as illustrated in Figure 2-4.

fX x
x-space u - space
x2 u2
fU u u

x1 u1
g x 0
g u 0

Figure 2-4 Illustration of the two-dimensional case of a linear limit state function and
standardized normally distributed variables U1 , U 2 . (adapted from Faber (2003)).

In Figure 2-4 the limit state function g x is transformed into the limit state function g u
by normalisation of the independent normal distributed random variables X into standardized
normally distributed random variables U as:

X i  P Xi
Ui (2.19)
VX i

such that the random variables U i have zero means and unit standard deviations.
The reliability index E has the geometrical interpretation as the smallest distance from the

16
line (or in general the hyper-plane) forming the boundary between the safe domain and the
failure domain, i.e. the line (or hyper-plane) described by g u 0 (Hasofer and Lind
(1974)). The point of the smallest distance is referred to as the design point u* .

2.3.1.2 Non-linear Limit State Functions

When the limit state function g x is non-linear in the basic random variables X , the failure
probability can only be approximated. In Hasofer and Lind (1974) it is suggested to perform a
linearization in the design point of the failure surface represented in the normalized space.
The approach is illustrated in the two dimensional space in Figure 2-5.

u - space
u2
fU u u

u1
g· u 0
u*
g u 0

Figure 2-5 Illustration of the linearization proposed in Hasofer and Lind (1974) in the standard
normal space.

In Figure 2-5 the failure surface g u 0 is linearised in the design point u* by the line
g c u 0 . The vector Į is the out ward directed normal vector to the failure surface in the
design point.
As the limit state function is in general non-linear, the design point is not known in advance
and has to be found iteratively, e.g. by solving the following optimisation problem:

n
E min
u^ g u 0`
¦u
i 1
2
i (2.20)

This problem can be solved in several different ways. Provided that the limit state function is
differentiable the following simple iteration scheme can be followed:

w
 g EĮ
wui
Di 1
, i 1, 2,..., n (2.21)
2
ª n w 2º
«¦ g EĮ »
¬ j 1 wui ¼

17
g ED1 , ED 2 ,..., ED n 0 (2.22)

First the design point is guessed with u* E Į and inserted into Equation (2.21) whereby a
new vector Į to the failure surface is achieved. Then this Į -vector is inserted into Equation
(2.22) from which a new E -value is evaluated. The iteration scheme converges normally after
a view, say 6-10, iterations and provides the design point u* as well as the reliability index E
and the outward normal vector to the failure surface in the design point Į . The components of
the vector Į may be interpreted as sensitivity factors giving the relative importance of the
individual random variables for the reliability index.
For the general case, i.e. where the basic random variables X may be not normal distributed
and correlated, the above requirements in regard to non-correlation and normality can be met
by proper transformations of the correlated non-normal variables into uncorrelated normal
ones. It is referred to Hohenbichler and Rackwitz (1981) and Der Kiureghian and Liu (1986),
or Ditlevsen and Madsen (1996) where a detailed overview upon these transformation
methods is given.

2.3.2 MONTE CARLO SIMULATION METHODS

A very simple approach to estimate probability of failure e.g. according to the general
expression in Equation (2.2) is the Monte Carlo Simulation Method. Assuming that the basic
random variables are represented through a set of independent random variables X , then
outcomes of the limit state function g X can be processed by sampling the basic random
variables at random virtually according to their distribution functions. The outcomes might be
in the failure domain ( g x d 0 ) or in the safe domain ( g x ! 0 ) and after an infinite
number of ‘tests’ the failure probability is:

nf
pf P g x d0 lim (2.23)
n of n
where n is the total number of trials and n f is the number of outcomes where g x d 0 .
Virtual random sampling or simulation is often consuming considerable computation time;
therefore the number of simulations is limited to a certain extent. With a limited number of
simulations the failure probability can only be estimated and the uncertainty of these
estimations is of interest. It can be shown that the uncertainty associated with the estimate is
proportional to 1 n f .
Since in structural reliability analysis the failure probability of interest is small, i.e. in the
order of 106 the number of simulated failures n f is scarce. To estimate the failure probability
of 106 with proper accuracy, e.g. a coefficient of variation of the estimate of 10%, the
number of simulations should be 108 .

18
2.4 DESIGN FORMATS IN STRUCTURAL ENGINEERING

In probabilistic assessments any uncertainties about a variable are taken into account
explicitly. This is not the case for more traditional ways of measuring safety, such as the
“factor of safety” and “partial safety factor” formats, which build the framework for presently
used design formats in structural engineering. Within these formats safety is expressed as a
deterministic measure since load and resistance are introduced as fixed values, whereas the
estimate of the load is considered as sufficiently high and the estimate of the resistance as
sufficiently low to guarantee an appropriate safety level.

2.4.1 ALLOWABLE STRESS FORMAT – THE SAFETY FACTOR

A customary method to define structural safety is through the factor of safety, usually
associated with elastic stress analysis and which requires that:

si \ g d s pi (2.24)

where si \ g is the i th applied stress component calculated at the generic point \ g in the
structure, and s*pi is the allowable stress for the i th stress component. The allowable stresses
s pi are in general defined in structural design codes and they are derived from material
strengths as the ultimate moment, tension or compression stress, expressed in stress terms sui
but reduced by a factor Ir :

s pi sui Ir (2.25)

where Ir is referred to as the factor of safety. The factor Ir may be selected on the basis of
experimental observations, prior experience, economic and possibly political considerations.
Usually, its selection is the responsibility of a code committee.
According to Equation (2.24) failure of a structure is assumed when the calculated i th elastic
stress si \ g component reaches the local permissible stress component s pi . Whether failure
actually does occur depends completely on how accurate si \ g represents the actual stress
in the real structure and how well s pi represents the actual material strength.
By combining Equations (2.24) and (2.25) the condition of limit state violation can be written
as:

sui \ g sui \ g
d si \ g or si \ g d 1 (2.26)
Ir Ir

When the inequality sign is replaced by an equality sign, Equations (2.26) are limit state
functions. By appropriate integration these Equations can also be expressed by stress
resultants:

19
ri \ g ri \ g
d sq ,i \ g or sq , i \ g d 1 (2.27)
Ir Ir

where ri is the i th resistance at location \ g and sq ,i is the i th stress resultant in general made
up of the effects of one or more applied loads qi .
The design formats where Equation (2.26) is utilized as a basis are referred to as ‘Allowable
Stress Design’ (ASD) formats.

2.4.2 LOAD AND RESISTANCE FACTOR DESIGN FORMAT – PARTIAL SAFETY

FACTORS

A format which is derived from the allowable stress format but allows for a more
differentiated treatment of loads and resistances is the so called load and resistance factor
design (LRFD) format. The limit state can be expressed at the level of stress resultants (i.e.
member design level) as:

zd rk
J G sG , k  J Q sQ ,k  ... (2.28)
JM
where rk is a characteristic member resistance, J M is the partial factor on rk and sG , k , sQ ,k
are the characteristic dead and live load effects respectively with associated partial factors J G ,
J Q and zd is the design variable. Characteristic values are generally given in the design codes
and correspond to fractile values of the underlying distributions of the variables.

2.5 RELIABILITY BASED CODE CALIBRATION

In the daily practice the engineering codes and regulations form the premises for the design of
safe and cost efficient structures. Code regulations in North America, Australia and Europe
are based on the limit states design (LSD) approach which is implemented via load and
resistance factor design (LRFD) formats, e.g. CIRIA (1977), CEB (1976), Eurocodes (2001),
AHSTO (1994) and OHBDC (1983). The LRFD format is outlined in Section 2.4.2.
Originally, LRFD methods where adapted as so called “soft conversions” of allowable stress
design (ASD), the design method which was prevalently used in code regulations before
LRFD was introduced and which is commonly based to a major part on experience, tradition
and judgment. Engineering traditions might be considered as the accumulation of knowledge
and experience collected over a long period of time; i.e. the in that way developed rules and
regulations imply an inherent level of safety. However, for the design of new types of
structures with new materials or subject to new environment and loading conditions, existing
design rules have to be adapted by means of careful extrapolation. Thereby it is assumed that

20
the safety and cost efficiency is still satisfactory.
In the last 3 to 4 decades the development of structural reliability methods has provided a
more rational basis for the design of structures. These methods provide a consistent basis for
the comparison between the reliability of well tested structural design and the reliability of
new types of structures. The determination of consistent design formats, i.e. the allocation of
characteristic values and partial safety factors, which provide consistent safety levels for
different types of structures based on structural reliability methods together with the choice of
desired target reliability, is commonly understood as reliability based code calibration.
Reliability based code calibration has been formulated by several researchers, see e.g.
Ravindra and Galambos (1978), Ellingwood et al. (1982) and Rosenblueth and Esteva (1972)
and has been already implemented in several codes, see e.g. OHBDC (1983), NBCC (1980),
and more recent the Eurocodes (2001). An overview about reliability based code calibration
can be found in Faber and Sørensen (2003).

2.5.1 SIMPLE CODE CALIBRATION WITH GIVEN TARGET RELIABILITY

The central task within the framework of reliability based code calibration is to evaluate
partial safety factors for already existing design concepts, which are formulated e.g. by means
of a Load and Resistance Factor Design format (LRFD). Within structural codes,
characteristic values, partial safety factor and load combination factors are utilized to achieve
target reliability for a variety of load combinations. The characteristic values are in general
determined by the code; e.g. 98%- or 50%-fractile values of the underlying probability
distribution function for loads and 5%-fractile values for resistance variables. The partial
safety and load combination factors can be calibrated so that a uniform level of reliability is
obtained for all load combinations.
As an example the calibration of partial safety factors for a set of limit states is considered.
The load is assumed to be a linear combination of a permanent load with the load effect SG
and a variable load with the load effect SQ . The material property e.g. the yield strength is
given by R and the design variable e.g. the cross-sectional area is zd ,i . D S ,i is a factor taking
into account the ratio of the effect due to the characteristic permanent load to the effect due to
the characteristic variable and permanent loads. Therefore, all possible load situations
comprising the two loads can be described by different D S ,i factors as e.g. D S ,i 0, 0.1, 0.2,
…, 1.0.
For a given set of partial safety factors the reliability index is evaluated as follows. Firstly, the
design variable zd ,i is determined in accordance with Equation (2.28) as:

JM
z d ,i D S ,i J G sG , k  1  D S ,i J Q sQ ,k (2.29)
rk

Secondly, the design variable zi is inserted into the corresponding limit state function g X :

21
gi X z d ,i R  D S , i S G  1  D S , i S Q (2.30)
T
with X R, SG , SQ is the vector of the basic random variables. The failure probability is
calculated using Equations (2.30) and (2.2) together with e.g. FORM or SORM as a solution
scheme (see Section 2.3).
The optimal set of partial safety factors is determined by the optimization formulation:

­ 2 ½ i
min ®¦ E target  E i D S ,i , zi , J M , J G , J Q ¾ with e.g. i 1, 2,3,...,10 ; D S ,i (2.31)
J M ,J G ,J Q
¯ i ¿ 10

Etarget is the target reliability.

2.5.2 TARGET RELIABILITIES

In general, the requirements to the safety of a structure are expressed in terms of the accepted
minimum reliability index or the accepted maximum failure probability. In a rational analysis
the target reliability is considered as a control parameter subject to optimisation, Faber and
Sørensen (2003).
The objective function may include in a general form cost benefit considerations in the sense
of how much has to be invested (by e.g. increasing partial safety factors) to attain the intended
benefit (e.g. the reliability of a structural system). As a consequence of this cost benefit
formulation the target reliability is not common for different types of structures; it depends on
the relative cost of a safety measure (e.g. increasing the cross section of a component) and the
expected consequences in the case of failure of the structure.
In Table 2-3 target failure probabilities and corresponding target reliability indexes are given
for ultimate limit states based on the recommendations of JCSS (2001). Note that the values
given correspond to a year reference period and the stochastic models recommended in JCSS
(2001).

Table 2-3 Tentative target reliability indices E (and associated target failure probabilities) related
to a one-year reference period and ultimate limit states (JCSS (2001)).
Relative cost of Minor consequences of Moderate consequences of Large consequences of
safety measure failure failure failure

High E=3.1 ( p f |10-3) E=3.3 ( p f |5 10-4) E=3.7 ( p f |10-4)

Normal E=3.7 ( p f |10-4) E=4.2 ( p f |10-5) E=4.4 ( p f |5 10-5)

Low E=4.2 ( p f |10-5) E=4.4 ( p f |10-5) E=4.7 ( p f |10-6)

22
The value for the most common design situation is indicated with grey shading in Table 2-3.
Guidelines for the classifications in this table can be found in the probabilistic model code,
JCSS (2001).

23
24
3 TIMBER AS A STRUCTURAL MATERIAL

It is the intention of this chapter to give an overview of timber as a structural material. This
includes the description of wood as a fibre composite material on a micro scale 1 and the
specification of irregularities like knots and fissures on a meso scale 2 . Furthermore it is
described how timber material is usually used in construction and the relevant material
properties are identified and defined.
In order to gain a better understanding of the reason for the special behaviour of wood and
timber material it is helpful to start thinking about where the wood and the timber are
‘produced’; in the stem of a tree.

3.1 THE TREE

Action

Mechanical Biological

Leaves

Wind Photo-
Resistance synthesis

Branch

Support Transport

Stem

Support Transport

Root

Anchoring Uptake of Minerals

and Water

Figure 3-1 The tree.

A tree can be seen as a structure that faces an ongoing optimisation process over millenniums.

1
In this context micro scale specifies a spatial reference most conveniently measured in P m .
2
In this context meso scale specifies a spatial reference most conveniently measured in cm .

25
Its function is to expose leafs to sunlight and due to the competitive situation with other plants
leafs have to be lifted up far above the ground. As illustrated in Figure 3-1 every part of the
tree has different mechanical and biological functions. The uptake of water and minerals in
the roots, the transport of these nutritional agents, the sap, through the stem and the branches
and the photosynthesis in the leafs exposed to the sunlight describes briefly the biological
function of the different parts. Mechanically the leafs have to resist the direct exposure to
wind, rain and snow whereas the branches and the stem have to support the crown and
transfer the corresponding load effects to the roots where the tree is anchored into the ground.
As evident from Figure 3-1, the design of the tree must necessarily be a compromise. Even
though it is a good photoreceptor, a large crown is heavy and more susceptible to wind loads;
consequently many supporting members (stem and branches) are required and the stem must
be more resistant to bending. Furthermore, the load has to be transferred through a large and
efficient root system. A large crown also looses much water by transpiration. The water
evaporated must be replaced by sufficient quantities of fresh supplies transported to the shoots
by an equivalent root system. The available space for the root system is limited, since the
ground has to be shared with other plants.
All these counteracting requirements result in an ideal compromise tree characterised by the
largest possible crown but the smallest number of supporting members, all of which are of
sufficient strength, but not so large as to add weight and increase energy consumption. Over
millenniums this optimal design set up has been achieved, moreover trees have developed the
ability to react continuously on every specific boundary condition; i.e. during growth, trees
are able to optimise their shape and their mechanical properties. Regarding this aspect, the so-
called secondary growth in the thickness direction of the tree components is of high
importance. The thickness of branches and stems is determining the strength and the stiffness;
the supporting qualities of the tree. Secondary growth takes place just below the outer surface
of the stem or branch in the cambium, where phloem and finally bark are produced outwards,
and wood is grown inwards, Niemz (2004). The inwardly grown wood is an annual tree ring
consisting mostly of more porous early wood, and denser and stronger late wood1. Aiming at
an even stress distribution in the supporting part of the tree the mechanical performance is
continuously reassessed and more and stronger material is accumulated by secondary growth
on the locations where the stresses are the highest. Simultaneously, the primary growth is
aspiring to a tree design which is balanced between light exposure and affordable moment
forces in the supporting parts. A more complete description of a tree as a highly optimised
structure can be found e.g. in Mattheck (1998).
Coming back to the issue of interest – to describe and understand the composition of wood
and structural timber, the briefly described self-optimisation property of trees should be kept
in mind. On every scale it can be recognised that wood is a material with highly optimised
properties as long as the original purpose is aimed at – the mechanical support and the

1
The terms ‘early wood’ and ‘late wood’ are further explained in section 3.2.2.1.

26
transportation of sap in a tree. However, the material properties become sub-optimal when
wood material is reused for human purpose, e.g. as timber for building structures.

3.2 WOOD

3.2.1 CELLS AND FIBRES – THE ULTRASTRUCTURE OF WOOD

Wood is a natural, organic cellular solid. Wood cells are also called tacheids which are long (2
– 5 mm) and slender (0.01 – 0.05 mm) and which have tapered or flattened ends. The
mechanical design of wood cells is found to be common among many wood species.

}
Inner Layer (S3)

Middle Layer (S2) Secondary Wall

Outer Layer (S1)

Primary Wall

Middle Lamella

Figure 3-2 The composition of a wood cell and its bonding in a lignin matrix (middle lamella).

The basic skeletal substance of the wood cell wall is cellulose. Cellulose is made up of sugar
molecules that fit together in a regular manner with their long-chain direction parallel to each
other. The cellulose molecules form bundles called micro fibrils. As illustrated in Figure 3-2
the cell wall consists of 2 main layers; the primary wall in which the micro fibrils are arranged
in a random, irregular network and a secondary wall which normally determines the static
behaviour of the wood cell. The secondary wall itself consists of three fairly distinct layers.
The outermost so-called S1 layer is very thin (around 0.1 ȝm) and exhibits an average micro
fibril angle of about 60° relative to the longitudinal axis of the cell. The main part of the
secondary wall is the S2 layer, which is typically several micro meters thick. The micro fibrils
are oriented in a very small angle to the fibre axis. Within the S3 layer the micro fibrils are
arranged with a gentle slope but not in a strict order, Hoffmeyer (1995).
From an engineering point of view, the cell wall structure is a very effective construction. The
predominant S2 layer with the mainly axially oriented micro fibrils takes up tension forces. In
compression the long and slender micro fibrils of the S2 layer are prevented from buckling by

27
the outer and inner S1 and S3 layer.
The wood cells are orientated and glued together by a matrix of lignin, the middle lamella
(Figure 3-2), forming the characteristic wood tissue (see also Figure 3-3) whose mechanical
properties are tuned to fulfil the special supporting functions of the stem and branches. Along
the fibre orientation the strength and stiffness properties for compression and tension are high
compared to the strength and stiffness perpendicular to the fibre orientation.

3.2.2 GROWTH RINGS AND VESSELS – THE MICROSTRUCTURE OF WOOD

Wood is obtained from two main categories of trees known in colloquial terms as hardwoods
(deciduous trees) and softwoods (coniferous trees or conifers). The observable difference
between these two categories is that in general deciduous trees have leafs and covered seeds
and coniferous trees evergreen needles and uncovered seeds. But there are also differences in
the microstructure of the wood, namely the specific assembly, function and production
strategy of wood cells.
Coniferous wood consists of 90% - 95% tracheids, which are prearranged in radial arrays, and
their longitudinal direction is oriented along the axis of the stem of the tree. In evolving from
early wood to latewood the cell walls become thicker, while the cell diameters become
smaller. This difference in growth may result in a ratio between latewood density and early
wood density as high as 3:1. The storage and transportation of sap takes place within
parenchyma cells which in conifers are mostly arranged in radial rays. In Figure 3-3a and
Figure 3-3b, two microscopic exposures of a southern pine specimen are shown. The different
proportions of the early- and late wood tracheids can be observed. The parenchyma cells can
be seen in Figure 3-3b. The interchange of sap between the cells is facilitated by small
openings in the fibre wall, the so-called pits.

a 0.3 mm b 0.3 mm c 0.3 mm d 0.3 mm

Figure 3-3 The microstructure of coniferous wood – southern pine (a and b) and deciduous wood
– white ash (c and d). (Pictures adapted from Niemz (2004))

The anatomy of deciduous wood is more varied and complicated than that of coniferous
wood, but most of the structural characteristics are similar. The difference between early

28
wood and late wood is minor and the tracheids have mainly strengthening function. The
transport of sap is assured through special vessels in between the tracheids; see Figure 3-3c
and Figure 3-3d. Deciduous wood fibres have thicker cell walls and smaller Lumina than
those of fibres from coniferous wood (compare Figure 3-3b, coniferous with Figure 3-3d,
deciduous), Kollmann et al. (1968).

3.2.2.1 Variations of the Properties of Wood on a Micro Scale

Growth Rings
Within a very small volume in the stem of a tree the properties of the wood are varying
systematically. The different properties of wood cells build in spring and early summer (early
wood) and the cells produced from summer to fall (late wood) are well pronounced for many
wood species grown in a temperate climate. Conifers tend to produce high density late wood
rows of a relatively constant thickness and the variation of the thickness of the entire annual
rings is governed by low density bands of early wood. In Figure 3-4 late and early wood can
be distinguished as dark and bright coloured ring pattern. For some deciduous wood species,
the so-called ring-porous species such as oak and ash, there is a high concentration of open
vessels in the early wood which forms also the typical annual ring pattern (compare Figure
3-3). The width of these rings is relatively constant and the variation of growth ring width is
mainly caused by the variation in the thickness of the high density rings of latewood
tracheids. This can not be observed for so-called diffuse porous deciduous wood as poplar and
beech, Hoffmeyer (1995).

Sapwood and Heartwood

The young outer part of a tree stem conducts the upward flow of sap from the root to the
crown. This part of the stem is known appropriately as sapwood. As the cells grow old, they
stop functioning physiologically; this inner part of the stem is known as heartwood. In most
species heartwood is darker in colour due to the incrustation with organic extractives
(compare with Figure 3-4). These chemicals provide heartwood with a better resistance to
decay and wood boring insects. Heartwood formation normally results in a significant
reduction in moisture content; in conifers the pits in the wood cell wall constrict, in deciduous
wood the vessels become plugged. This results in a marked reduction of permeability,
Kollmann et al. (1968).

Juvenile Wood
The wood of the first 5 to 20 growth rings (depending on species) of any stem cross section is
called juvenile wood and exhibits properties different from those of the outer part of the stem
(mature wood). This is particularly significant for coniferous wood. In juvenile wood, the
wood cells are relatively short and thin walled with a remarkable slope of the micro fibrils or

29
the S2 layer of the secondary wall. Therefore, juvenile wood typically exhibits lower strength
and stiffness particularly in tension. The longitudinal shrinkage is much greater than in non-
juvenile wood. Often heartwood holds all the juvenile wood, which possesses inferior quality
with respect to the mechanical properties. The boundary between juvenile and mature wood is
gradual and not visible, Thörnquist (1990).

anual growth ring pattern

sap wood

pith

juvenile wood

heart wood

bark and cambium

Figure 3-4 Cross section of a stem of a tree, Douglas-fir, annual growth ring pattern, juvenile
wood, heart wood, bark and cambium.

3.2.3 KNOTS AND GRAIN DEVIATIONS – THE MESO STRUCTURE OF WOOD

As seen in the preceding section wood is a material optimised to fulfil its biological and
mechanical function in the supporting parts of a tree. Wood as a mechanical body is highly
anisotropic mainly due to the elongated shapes of the wood cells and the oriented structure of
the cell walls; i.e. the strength and stiffness along the grain direction is much larger than
perpendicular to the grain direction. In a living tree this property is used in such a way that the
fibre direction is following the direction of the main stresses in the supporting components;
mainly along the axis of the stem and the branches. However, where branches are connected
to the stem or where specific load circumstances cause torsion forces in the supporting
components the fibre direction deviates from the direction of the axis of the components.

Knots
Knots are parts of branches that are embedded and anchored in the main stem of a tree. The
lateral branch is connected to the pith of the main stem. As the perimeter of the stem
increases, successive growth rings form continuously over the stem and branches and a cone

30
of branch wood (the inter-grown knots) develops within the stem. The result is a notch-free
transition from stem to branch with minor stress concentrations where the fibre direction is
following the predominant main stress direction in the ‘joint’. Sometimes the branch may die
or brake off. The succeeding growth rings added to the main stem simply encircle the dead
limb stub and the dead part of the stub becomes a so-called encased or entrapped knot.
Softwoods are characterised by having a dominant stem from which clusters of lateral
branches occur at regular intervals, Isaksson (1999).

Grain Deviations
Some specific but permanent loading conditions on a tree are inducing cell orientation
forming a helix around the stem axis (spiral grain). This is the case when torsion forces
mainly in one direction due to e.g. asymmetric wind loading, are present in the stem.

Reaction Wood
A tree reacts on systematic external forces on the supporting components by forming so-called
reaction wood. Conifers develop compression wood in areas with subdominant high
compression, whereas deciduous species develop tension wood in regions where tension
forces are prevailing. Compression wood has the appearance of wider growth rings and a
higher late wood proportion than normal coniferous wood. Furthermore, the structural
arrangement of the wood cell walls is different for compression wood, i.e. the micro fibrils of
the S2 layer are arranged with a 45° slope which results in excessive longitudinal shrinkage
property, similar to juvenile wood. The compression strength and stiffness of compression
wood is higher than of normal wood whereas the tension strength and stiffness is
approximately the same, i.e. the 45° angle of the S2 layer and their increased thickness are
effects compensating each other. For deciduous species tension wood differ from normal
wood for a number of biochemical, anatomical and mechanical characteristics. Mechanically,
the frequency of vessels and their porosity is significantly lower in tension wood, whereas
fibre and vessel lengths are significantly longer, Jourez et al (2001). However, in a number of
tree species such as poplar the most striking differences are found in the fibres of tension
wood. In these fibres, named G-fibres, one layer of the secondary wall (generally the S3
layer) is replaced by a very thick layer in which the micro fibrils are almost parallel to the axis
of the cell which contributes, to the specific mechanical properties of tension wood, Timell
(1969).

3.2.4 CLEAR WOOD SPECIMENS AND THEIR PROPERTIES

The characteristics of wood are usually obtained from tests of small pieces of wood termed
“clear” or “straight grained” because they do not contain characteristics such as knots, cross

31
grain and splits1. These test pieces have anatomical characteristics such as growth rings that
occur in consistent patterns within each piece. Clear wood specimens are usually considered
“homogeneous” in wood mechanics, Niemz (2004).

Moisture Content
The moisture content is defined as the ratio of the mass of removable water to the dry mass of
the wood. The dry mass is obtained by oven drying. The moisture content may be expressed
as a fraction or in percentage terms. In so-called green wood, i.e. wood of a freshly cut tree,
the moisture content could be higher than 100%; the so-called free water in the cell lumen
together with the water in the saturated wood cell walls have a larger mass fraction than the
oven dry wood. At a moisture content of around 28%, the so-called fibre saturation point, the
wood cell walls are still saturated, whereas there is no free water in the cell lumen. Below the
fibre saturation point the cell wall is loosing water which has a mayor influence on the
mechanical properties of the wood. Above the fibre saturation point the mechanical properties
are nearly independent of moisture content, Hoffmeyer (1995).

Shrinkage and Swelling

Figure 3-5 Distortions of wood prisms cut out at different spots in the stem due to shrinkage.
(Niemz (2004))

Wood is dimensionally stable when the moisture content is greater than the fibre saturation
point. Wood changes dimension as it gains or looses moisture below that point. It shrinks
when loosing moisture from the cell walls and swells when gaining moisture in the cell walls.
Shrinkage and swelling are also termed movements. The magnitude of these movements is
depending on the direction to the wood cells and is mainly governed by the substantial S2
layer of the wood cell wall, compare Figure 3-2. The micro fibrils in the S2 layer are nearly
parallel oriented to the longitudinal axis of the cell and the water is absorbed between the

1
Splits are cracks along the fibre direction due to shrinkage of the wood

32
micro fibrils. This means that movements in the transverse direction are much more
pronounced than in the longitudinal direction; the ratio is in the order of 20:1. Juvenile wood
and compression wood exhibit micro fibril angles much larger than non juvenile wood, which
result in much larger longitudinal movements.
In the transverse direction a differentiation between the radial and the tangential direction to
the growth rings can be made. The movements in the tangential direction may, for practical
purposes, be taken as twice as radial movements. The consequence of this property can be
observed in Figure 3-5. Wood prisms are deformed due to shrinkage in different manner,
depending on the orientation of the annual rings in the prism; these deformations are in
general referred to as distortions, Hoffmeyer (1995).

Density
The wood density is the most important physical characteristic of wood. The density U den is
defined as:

mg
U den (3.1)
vVol

where mg is the mass in kg and vVol is the volume in m3 . In wood science and engineering
the dry density U den ,0 and the density U den ,12 at 12% moisture content are most frequently
used. While the density of the cell wall material is relatively constant among wood species,
which is about 1500 kg m3 (Kollmann et al., 1968), the density of wood depends mainly on
the ratio between cell lumen and cell wall. The density ranges from 200 kg m3 to 650 kg m3
for coniferous wood and from 300 kg m3 to 1100 kg m3 for deciduous wood.

Strength and Stiffness Properties

The strength and stiffness properties of wood depend on which direction the fibres are
stressed. For small clear wood specimens it is in general distinguished between properties
parallel and perpendicular to the grain.

Table 3-1 Strength properties and density of some structural materials (Thelandersson and Larsen
(2004)).

Material Density Strength Strength / Density

kg m3 MPa 103 MPa ˜ m3 kg
Structural Steel 7800 400-1000 50-30
Aluminium 2700 100-300 40-110
Concrete, compression 2300 30-120 13-50
Clear coniferous wood, tension parallel to the grain 400-600 40-200 100-300
Clear coniferous wood, compression parallel to the grain 400-600 30-90 70-150
Clear coniferous wood, tension perpendicular to the grain 400-600 2-8 5-10

33
Due to the fact that wood in a tree is a highly optimised structural material it is no surprise
that its performance is impressive if it is stressed as in nature. E.g. clear wood exhibits high
strength and stiffness in tension parallel to the grain; the ratio between strength and density is
even higher than for steel (compare Table 3-1). On the other hand, if clear wood is loaded
perpendicular to the grain the load carrying capacity is very low, Thelandersson and Larsen
(2004).

Moisture Content and Mechanical Properties

The mechanical properties of wood depend on the moisture content. An increase in moisture
leads to lower values for strength and stiffness properties. Water, when penetrating the cell
wall, weakens the coherence of the cell wall. Moisture variations above the fibre saturation
point of around 28% have no effect on the mechanical properties, since such variations are
related to free water in the cell lumen. Not only the moisture content is important for the
strength properties of wood specimen; fast changes cause moisture gradients in wood which
induce stresses perpendicular to the grain direction. Wood is a hygroscopic material that takes
water from the surrounding air, and the moisture content of wood tends to attain equilibrium
with the air humidity.
The effect of moisture on the mechanical properties of wood is given in Table 3-2. For
moisture contents between 8% and 20% a linear relationship between moisture content and
strength can be assumed, Hoffmeyer (1995).

Table 3-2 Approximate change (%) of clear wood properties for a one percentage change of
moisture content. (Hoffmeyer(1995)).

Material Property Change [%]

Compression strength parallel to the grain 5
Compression strength perpendicular to the grain 5
Bending strength parallel to the grain 4
Tension strength perpendicular to the grain 2.5
Tension strength parallel to the grain 2
Modulus of elasticity parallel to the grain 1.5

Duration of Load
Wood experiences a significant loss of strength over a period of time. The failure mechanism
of wood under long term load is referred to as creep rupture. The first major investigation on
the duration of load effect on small clear wood specimen is published in Wood (1947). On the
basis of bending tests of duration of up to 7 years, a stress – lifetime relationship is
established, which predicts the 10-years strength to be slightly less than 60% of the
approximated short term strength. Similar effects are observed in the literature for loading
modes different from bending.

34
3.3 STRUCTURAL TIMBER

“Timber is as different from wood as concrete is different from cement.” - Borg Madsen

Structural timber components are sawn from logs1. Apart from some exceptions they have a
prismatic shape with a rectangular cross section. In contrast to small clear wood specimens
timber components have structural dimensions. The maximum possible dimension of the
timber components is determined by the size of the trees in the forest. E.g. one hundred years
ago timber components with a length of 20 m and a rectangular cross section of about 150 x
450 mm were commonly available. Nowadays, the forestry strategy has changed and in most
countries timber with a cross section over 75 x 225 mm and a length more than 5 m attracts a
cost premium due to scarcity, Steer (1995). Timber components are applied for load carrying
functions in structures; therefore there are several provisions in the production line to obtain
appropriate structural elements. For example it is aimed for that the longitudinal axis of the
timber structural component coincides with the grain direction of the wood cells. Due to the
dimension of the timber components this is mostly not strictly possible. So called growth
irregularities as knots and grain deviations are affecting the uniform and parallel grain
direction. It should be remembered that these ‘growth irregularities’ are just part of the
excellent property of a tree to react on stress peaks and special load conditions constantly
during its growth. However, for a sawn piece of timber these growth irregularities are sub-
optimal.

3.3.1 GROWTH IRREGULARITIES IN TIMBER STRUCTURAL COMPONENTS

3.3.1.1 Knots
Being an optimal stress reducing connection of the stem with the branches in a tree, knots in
sawn structural timber are by far the most important defects affecting the mechanical
properties. It can be differentiated between inter grown knots and loose knots (see Section
3.2.3).

3.3.1.2 Cross Grain

The phenomenon, when the mean grain direction within a large section is not coinciding with
the longitudinal axis of the timber component is referred to as cross grain. This is quite
common, but only recognised at a certain magnitude. The reason for cross grain might be
technical, i.e. due to sawing configurations, or due to growth particularities, e.g. helix-growth,
or a cone shaped stem of the tree.

1
A log is a cut proportion of the stem of a tree, in general already prepared for transportation and further
processing; i.e. the log is cylindrical shaped and free of branches.

35
3.3.1.3 Distortion
As for clear wood prisms, distortion due to inhomogeneous shrinkage and swelling properties
is also present for timber structural elements. (compare shrinkage and swelling in section
3.2.4)

3.3.1.4 Wane
Timber structural elements, in general, have a rectangular shaped cross section. Due to sawing
practice it is possible that the edge of the timber log becomes visible after cutting the
rectangular shaped elements from the conical shaped log. This is in general referred to as
wane.

3.3.1.5 Permanent Compressive Yield

Permanent compressive yield is a local phenomenon due to an instant compression in the stem
of the living tree. E.g. during a storm the compression strength can be reached on a spot
where local buckling of the fibres occurs. This happens without further consequences for the
tree. In a sawn timber component these spots can be very disadvantageous due to very low
strength characteristics. Permanent compressive yield is very hard to detect by visual quality
control, Arnold and Steiger (2005).

3.3.2 MATERIAL PROPERTIES OF INTEREST OF TIMBER STRUCTURAL

COMPONENTS

Structural timber is a non-homogeneous material. It’s also known as orthotropic, whereas in

structural timber the major irregularities in a component (knots or cross grain) weaken this
statement. However, material properties of timber components are considered as the property
of an entire structural component, i.e. tension parallel to the grain means tension parallel to
the axis where a coincidence with the main fibre direction is assumed. Consequently, both,
stresses perpendicular to the grain and parallel to the grain exist in such a component, whereas
the latter case is the dominant.
Timber components are mainly used in structures in form of beams, columns and bars which
are loaded in transversal bending and parallel compression and tension correspondingly. In
these elements the stress bearing capacity of the wood material is used to its full potential.
The wood fibres and the stress directions are mainly orientated along the longitudinal axis of
these elements. In structural design, timber components are also subjected to other loading
modes, namely tension and compression perpendicular to the component axis or shear. These
properties are comparatively weak and should be treated with care during design.
In timber engineering, material properties are defined on an element level, i.e. material
properties are defined as the load bearing capacity of timber material specimens of defined

36
size and conditioning1 and assessed in accordance with an agreed testing procedure. E.g. the
bending strength of timber is not an ultimate stress property of the timber material; it is rather
the moment capacity of the test specimen, divided by the elastic section modulus. In this
section timber material properties according to test standards are introduced. The principle
test configurations are illustrated to show how material properties are defined at an element
level (based on ISO 8375). Regional specifications for the test configuration of bending tests
are given in section 7.3 according to the Australian/New Zealand Code, the US-American
Code, the Canadian Code and the European Code.

3.3.2.1 Bending Strength, Bending Modulus of Elasticity

The bending strength according to test standards is determined with a test configuration
illustrated in Figure 3-6. The load q is applied with constant rate until the collapse of the
beam. The maximum load qmax is observed. The projected test time to failure t f is also
specified in the standards. The bending strength rm is expressed in stress, assuming
homogeneous ideal elastic material and pure bending as:

3qa
rm (3.2)
h 2b

q q
2 ew 2

h
w'
b
a|l 3 a|l 3 a|l 3

Figure 3-6 Typical bending test configuration; a, b, ew, h are the specified dimensions, q and w'
are the measurements.

The bending stiffness is assessed with the same test configuration by observing the deflection
w' and the load at two different times during the test, i.e. w' ,1 , q1 and w' ,2 , q2 . It has to be
assured that the two measurements are made within the proportionality limit2 of the beam. The
bending stiffness, i.e. the bending modulus of elasticity moem is calculated as:

1
Conditioning in terms of a predefined constant temperature and relative humidity surrounding.
2
The proportionality limit confines the elastic domain of the stress strain relation ship of a material. In the
elastic domain stress is proportional to strain.

37
 2
3 a q2  q1 ew
moem (3.3)
4 h3b w' ,2  w' ,1

3.3.2.2 Tension Strength Parallel to the Grain, Tension Modulus of Elasticity

The tension strength according to test standards is determined with a test configuration as
illustrated in Figure 3-7. The load q is applied with constant rate until failure of the specimen
and the maximum load qmax is recorded. The projected test time to failure t f is also specified
in the test standards. Assuming ideal elastic and homogenous material behaviour the tension
strength parallel to the grain rt ,0 is calculated as:

qmax
rt ,0 (3.4)
bd
where the index 0 specifies the angle to the grain.

ew
q
d
w'
b
l

Figure 3-7 Typical tension test configuration; l, b, d, ew are the specified dimensions, q and w'
are the measurements.

The tension stiffness is measured with the same test configuration by recording two
simultaneous observations of the deflection w' and the load at different times during the test.
It has to be assured that the two measurements are made within the proportional limit of the
specimen. The tension stiffness, i.e. the tension modulus of elasticity moet ,0 is calculated
with:

ew q2  q1
moet ,0 (3.5)
bd w' ,2  w' ,1

3.3.2.3 Compression Strength Parallel to the Grain, Compression Modulus of Elasticity

The compression strength according to test standards is determined with a test configuration
as illustrated in Figure 3-8. The load q is applied with constant rate until the collapse of the
specimen and the maximum load qmax is recorded. The projected time to failure t f in the test
is also specified in the standards. Assuming ideal elastic and homogenous material behaviour

38
the compression strength rc ,0 is calculated according to:

qmax
rc ,0 (3.6)
bd
The compression stiffness is measured with the same test setup by recording two
simultaneous observations of the deflection w' and the load at different times during the test.
It has to be assured that the two measurements are made within the proportional limit of the
specimen. The compression stiffness, i.e. the compression modulus of elasticity moec ,0 is
calculated with:

ew q2  q1
moec ,0 (3.7)
bd w' ,2  w' ,1

l w' ew

Figure 3-8 Typical compression test configuration; l, b, d, ew are the specified dimensions, q and
w' are the measurements.

3.3.2.4 Tension and Compression Strength and Stiffness Perpendicular to the Grain
The compression and tension strength perpendicular to the grain according to test standards is
determined with a test configuration as illustrated in Figure 3-9. The load q is applied with
constant rate until the collapse of the specimen and the maximum load qmax is recorded. (For
compression perpendicular to the grain qmax has to be estimated with proper iteration
techniques). The projected time to failure t f is also specified in the standards. Assuming ideal
elastic and homogenous material behaviour the compression or tension strength perpendicular
to the grain rt ,90 , rc ,90 is calculated according to:

39
 qmax
rt ,90 rc ,90 (3.8)
bl

q
l

w'
h ew

Figure 3-9 Typical compression and tension test configuration for perpendicular to the grain; l, b,
h, ew are the specified dimensions, q and w' are the measurements.

The tension or compression stiffness perpendicular to the grain is measured with the same test
configuration by recording two simultaneous observations of the deflection w' and the load at
different times during the test. It has to be assured that the two measurements are made within
the proportional limit of the specimen. The tension or compression stiffness perpendicular to
the grain, i.e. the tension or compression modulus of elasticity perpendicular to the grain
moet ,90 , moec ,90 is calculated with:

ew q2  q1
moet ,90 moec ,90 (3.9)
bl w' ,2  w' ,1

3.3.2.5 Shear Strength, Shear Modulus

For evaluating the shear strength and stiffness different test configurations exist between
different test standards, i.e. in contrast to the other material properties the differences are not
only in dimension. For further particularities it is referred to the standards. Shear strength is in
general understood as the resistance of a timber specimen against a shear load on the surface
of the specimen in direction parallel to the grain combined with a compression load
perpendicular to the shear load. The shear modulus is in general evaluated with a 3-point
bending arrangement.

40
3.3.2.6 Density
A small volume is cut from the test specimen, whereas the removal shall be made over the
entire cross section of the specimen. The volume has to be free of defects. After destructive
tests the volume should be removed near the spot of the failure. The density is defined as the
ratio between the mass and the volume (see also section 3.2.4). The density is highly
dependent on the moisture content of the timber. Specifications are given in the test standards.

3.3.2.7 Test Standards

The test standards referred to are:
European Standards: EN 408 and EN 384
American Society for Testing and Materials: D 4761-88 and D 1990-91
Australian/New Zealand Standard: AS/NZ 4063:1992

3.4 TIMBER DESIGN

3.4.1 DESIGN FORMAT

The general layout of the most recent timber design codes is based on the load and resistance
factor design (LRFD) format (see section 2.4.2), which can be written as in Equation (2.28).
Several limit state equations are given in the codes, ultimate and serviceability, to cover the
most typical situations in practical design. Characteristic values for the material properties are
also given in these design codes which correspond to fractile values of the underlying
distribution functions. As described above the material properties are defined as the properties
of test specimen tested under specified conditions. In real structures these conditions may
deviate in terms of the size of the elements, loading modes and duration, climate variations,
etc. As seen before, timber material properties are rather sensitive to these deviations. In
modern timber design codes modification factors on the strength and stiffness related material
parameters are introduced to satisfy this circumstance. The calibration of these modification
factors is the matter of ongoing discussions in the timber research community; some
phenomena seem to be too complex to cope just with a multiplication of a factor, for other
problems the experimental evidence is rather poor. These issues will be discussed in more
detail in chapter 4.
Another particularity of timber is how its material properties are ensured to stay within some
predictable limits.

41
3.4.2 TIMBER GRADING

In comparison to building materials such as steel and concrete, the properties of structural
timber are not designed or produced by means of some recipe but may be ensured to fulfil
given requirements only by quality control procedures – also referred to as grading. For this
reason quality control and selection schemes are implemented in the production line, typically
already at the sawmill where the construction timber is produced from the timber logs.
Various schemes for grading have been developed using different principles, however, the
basic idea behind them all is that the material properties of interest are assessed indirectly by
means of other properties, measured non-destructively such as e.g. the density, the modulus of
elasticity or the visual appearance of the timber see Madsen (1992), Walker et al. (1993) or
Green and Kretschmann (1997).
As a result of timber grading, timber is represented at the market as a graded material. The
grades imply that the material properties lie within desirable and predictable limits. However,
the material properties of timber grades have to be considered as random variables and the
properties of timber grades are characterised (and communicated) through specific fractile
values of the assumed probability distribution functions of the material properties of interest.

3.4.3 STRENGTH CLASS SYSTEMS

In general, structural timber is assigned to a specific strength class. Several strength class
systems exist on an international scale, e.g. in Europe it is the EN 338 which constitutes the
classification of timber based on the prescription of characteristic values or the mean values
for the material properties; i.e. for every timber strength class a characteristic value or a mean
value for every relevant material property is given. Timber that is assigned to a certain
strength class is also referred to as a timber grade. In Table 3-3 an example for some strength
classes according to EN 338 is given. The characteristic values for the strength properties and
the density are defined as the 5% fractile values of the underlying distribution functions. The
modulus of elasticity (MOE) and shear modulus are specified by mean values. For the
qualification of a timber population to a certain grade, values for three material properties are
mandatory; the 5% fractile value of the bending moment capacity, the 5% fractile value of the
density and the mean value of the bending MOE. These material properties are subsequently
also referred to as the reference material properties. It is in general assumed that the indicated
values for other properties are representative for the specified grades. These values are based
on empirical relationships with the reference material properties evaluated based on several
tests on European softwoods and given in Glos (1995) as:

rt ,0,k 0.6rm ,k (3.10)

42
rc ,0,k 5rm0.45
,k (3.11)

rv ,k 0.2rm0.8,k (3.12)

rc ,90,k 0.015U den , k (3.13)

rt ,90,k 0.001U den ,k (3.14)

moe0,mean moem,mean (3.15)

moe90,mean moe90,mean moem ,mean 30 (3.16)

mog v , mean moem ,mean 16 (3.17)

For the nomenclature compare Table 3-3, the subscript k indicates the characteristic value of
the material property; the subscript mean indicates the mean value of the material property.
0 and 90 indicate the angle of the main stress direction to the fibre direction.

Table 3-3 The EN 338 strength classes C24, C30, C35 and C40; characteristic values for strength
and stiffness properties and density. Mandatory properties are indicated with grey
shading.
C24 C30 C35 C40
> MPa@
Bending moment capacity rm ,k 24 30 35 40

Tension strength parallel to the grain rt ,0,k 14 18 21 24

Tension strength perp. to the grain rt ,90,k 0.4 0.4 0.4 0.4

Compression strength parallel to the grain rc ,0,k 21 23 25 26

Compression strength per. to the grain rc ,90,k 5.3 5.7 6 6.3

Shear strength rv ,k 2.5 3.0 3.4 3.8

>GPa @
Bending MOE in bending moem ,mean 11 12 13 14

MOE - tension parallel to the grain moe0,mean 11 12 13 14

MOE - tension perp. to the grain moe90,mean 0.37 0.4 0.43 0.47

Shear modulus mog v , mean 0.69 0.75 0.81 0.88

3
ª¬ kg m º¼

Density U den , k 350 380 400 420

43
44
4 PROBABILISTIC MODELLING OF TIMBER MATERIAL
PROPERTIES

4.1 INTRODUCTION

Timber is a rather complex building material. Its properties are highly variable, spatially and
in time. In structural engineering, material properties of timber are in general understood as
the stress and stiffness related properties of standard test specimen under given (standard)
loading and climate conditions and the timber density. Test configurations according to ISO
8375 are defined in section 3.3.2.
As introduced in section 3.4.2 timber is a graded material. Due to the grading process, the
material properties are associated with some control scheme, whereas only the so-called
reference material properties are considered explicitly. The so-called other material properties
are only assessed implicitly. Following the European standard EN 338 the distinction between
reference properties and other properties is made as illustrated in Figure 4-1. The bending
moment capacity, the bending modulus of elasticity and the timber density are referred to as
the reference material properties.
When modelling timber material properties in a structure, i.e. at any generic point, in time and
in space, several issues have to be taken into account, see also chapter 3. As illustrated in
Figure 4-2 the cornerstone of the modelling of timber material properties are the reference
material properties under reference conditions. The material property of interest at any generic
point may deviate in terms of type (‘other material properties’), of dimensions (‘scale’) and of
specific loading and climate conditions (‘time (load/moisture)’). In general, these aspects are
treated separately, i.e. research and modelling scheme proposals are focused on one dimension
in Figure 4-2. This scheme is also followed in this chapter. In Section 4.2 focus is directed to
the spatial variability of timber material properties, i.e. material properties are considered at
different scales. The time dependency of timber material properties is discussed in section 4.3
(load dependency) and section 4.4 (moisture dependency). A framework for the interrelation
of reference material properties and other material properties is given in section 4.5.

45
 Reference Material Other Material
Properties Properties

rm = bending moment capacity rt ,0 / 90 tension strength

moem = bending MOE moet ,0 / 90 tension MOE

den density rc ,0 / 90 compression strength

moec ,0 / 90 compression MOE
rv shear strength
mog v shear modulus
rh ,0 / 90 embedding strength

Figure 4-1 Reference material properties and other material properties.

s ial
tie er
er at
op r m
pr the
o

time
reference
(load/moisture)
material
properties
material property at
a generic point of a structure
graded
scale

timber

Figure 4-2 Outline of the modeling of timber material properties.

4.2 SPATIAL VARIATIONS OF TIMBER MATERIAL PROPERTIES

4.2.1 SCALES OF MODELLING

Timber material properties vary locally in space. A material property in one point of a
structure might not be the same as in another point in the structure or in a point in a different
structure. In general, three hierarchical levels of variation may be considered: micro, local
(meso) and global (macro).
Timber from different origins is graded by applying different grading schemes by different

46
producers to one common grade. Every single sub-population of the timber grade might have
differently distributed material properties. The variability of the distribution parameters
represents a typical global parameter variation. Parameter variations may also be due to
statistical uncertainties.
The variations between timber test specimens or components of one specific sub-population
are modelled at the meso level. Information about the specific sub-population may be
obtainable from tests with the purpose of reducing the parameter variation. The geometrical
scale of the test specimens is in the same order as the size of a structural element and probably
most conveniently measured in meters.
At the micro-level the irregularities in the timber material itself are represented. These are
basically uncontrollable as they originate from natural variability such as the random
distribution of knots fissures and grain deviations. The geometrical scale of these irregularities
could be measured in P m at wood cell level or in cm if knots and weak sections are
considered.
probability density
MACRO

r
Sequence of Lots
probability density
MESO

Lot of Structural Elements / Test Specimen r

ln(bending strength)
MICRO

Knot cluster [dm]

{
{

Structural Element / Test Specimen [m] ln(r ij )

ij
{
=> sequence of weak zones

longitudinal direction of the beam

Figure 4-3 Different scales of modelling of timber material properties.

47
4.2.2 MODELLING REFERENCE MATERIAL PROPERTIES UNDER REFERENCE
CONDITIONS – MESO LEVEL VARIATIONS

Material properties are in general defined on the scale of test specimen capacity and the
loading history and climate conditions until failure are also specified in test standards, see
section 3.3.2. Deriving a probabilistic model for the material properties on this level, in
general, is a straightforward task. Material properties can be represented by random variables
X i ,... , the statistical characteristics of these variables can be described by distribution models
which parameters are calibrated according to data taken from standard tests. In section 2.2.2 a
general outline is given, some basic approaches and an example how this should be performed
is given in Annex A and in Chapter 7.

4.2.3 MACRO LEVEL VARIATIONS – TIMBER GRADING

In practical design the situation is different in regard to the amount and quality of available
information. Typically it is known what timber grade will be utilized in the structure. Various
grading schemes are calibrated to identify timber grades, i.e. sub-populations of timber
elements for which the strength class requirements (see section 3.4.3) are fulfilled. The
statistical characteristics of the material properties of these sub-populations are strongly
dependent on the properties of the timber supply which is used for grading and the particular
grading procedure which is applied (Faber et al. (2004) and section 4.2.3.2). This means that
for different sub-populations all assigned to the same timber grade, but by applying different
grading schemes and/or using different ungraded material, the statistical properties of the
material properties might be different, although the target fractile values are similar. These
macro variations have to be taken into account.
A possible method to quantify macro material variations is demonstrated in Rackwitz and
Müller (1977), where the macro variability of concrete is analysed. Concrete from a particular
grade but from several different producers is analysed. The sample moments of each sub-
population are quantified and functional relationships between the sample moments are
derived. A similar scheme can be realized for graded timber by taking timber from different
regions and graded to the same grade but identified by different grading schemes.
Alternatively the macro variability can be explicitly assessed if the applied grading scheme
can be formalized to a probabilistic framework which takes into account all uncertainties
involved into the grading procedure.
Coming back to this issue with a proposal in this direction in the second half of this sub-
section, first existing grading practice is introduced and discussed.

4.2.3.1 Existing Grading Practice

It is in general differentiated between two different strategies of timber grading: visual

48
grading and machine grading.

Visual Grading
Visual grading is based on visual inspection of timber structural elements. Visible defects as
knots, fissures and cross grain 1 are assessed and according to the appearance of timber
structural elements in regard to these defects they are sorted to a certain grade. Some more or
less rudimentary forms of visual strength grading have been used since timber was utilised as
a construction material. The first formal visual grading rules, the USA ASTM Standard D245
were published in 1927 (Madsen (1992)). Since the 1930s formalized rules for visual grading
were introduced in the European countries (Glos (1995)). These rules are further developed
until today, however they differ widely with respect to grading criteria, number of grades and
grade limits. Recent efforts to harmonize these visual grading rules at least throughout Europe
have not been successful because no single set of grading rules would cover the different
species, timber dimensions and uses in a satisfactory manner (Glos (1995)). However, for
Europe some general requirements for regional visual grading codes are prescribed (prEN
14081-1); the following has to be taken into account:
x limitations for visible strength and stiffness reducing characteristics: knots, slope to the
grain, rate of growth (annual ring width), fissures,
x limitations for geometrical characteristics: wane, distortion,
x limitations for biological characteristics: insect and fungal damage.
It has to be considered that the visual grading of a timber structural element in the usual
production line takes place within 2-4 seconds (Glos (1995)); the before mentioned character-
istics are not measured by some device, they are subjectively estimated and it is decided
within seconds whether the structural element belongs to a certain grade or not. Visual
grading has proven as an efficient tool to reduce the variability of timber material properties,
however, the grading effect strongly depends on the person who is performing the visual
grading. The statistical characteristics of the material properties of visual graded timber are
therefore difficult to assess based on information about the applied visual grading rules.

Machine Grading
The above mentioned disadvantages of visual strength grading may be overcome by machine
grading, where a more formal assessment of the grading process can be performed. In contrast
to visual grading, machine grading is in general based on indicative characteristics of a timber
structural element which can be measured non-destructively by some device. The indicative
characteristics have to be related to the basic material properties of interest. Typical indicative
properties are:

1
See Section 3.3.1, Page 35.

49
x Directly related to the MOE:
- flat wise bending stiffness,
- ultrasonic pulse measurement,
- frequency response measurement.
x Directly related to the density:
- measurements on weight and dimensions,
- J - ray detection.
x Directly related to visible defects:
- microwave response,
- optical detection and subsequent image processing.
A good overview about the different measuring schemes can be found in Thelandersson and
Larsen (2003). It is the result from several research projects, e.g. Johansson et al. (1992),
Boström (1994) that measurements related to the MOE are also highly related to the bending
strength. Several grading machines operate with a single MOE related indicator as the flat
wise bending stiffness and deliver comparable results compared to more complex machines
measuring several indicators (Thelandersson and Larsen (2003)).
Several grading machine systems can be found at the market, however they operate according
to similar principles; one or more indicative properties of the timber to be graded are
measured by the machine and based on these measurements a population of the un-graded
timber is subdivided into sub-populations of graded timber material. The grading acceptance
criteria are formulated in form of boundary values for indicative properties which have to be
matched to qualify a piece of timber to a certain grade. These boundaries are also termed
grading machine settings. The performance, i.e. the statistical characteristics of the output of
grading machines strongly depends on these settings, and in general very much attention is
kept on how to control these machine settings.

Control of Grading Machine Settings

Grading machines can operate either machine controlled or output controlled. The output
controlled grading system was developed in North America. Control is based on frequent
destructive strength testing or proof loading of control samples of the machine graded timber.
This system is relatively costly but it permits a modification of the machine settings in order
to optimize the yield, i.e. the predictability of the properties of the graded timber material.
This method requires large quantities of timber of similar dimension and origin, so that it can
be assumed that the characteristics of the timber are stationary. These conditions rarely exist
in Europe, where a variety of sizes, species and grades in small quantities are typical. For
these conditions the machine controlled systems are developed. Machine control means that

50
the settings are derived within a substantial assessment procedure prior to the operation phase
of the machine. The settings are optimized to a representative un-graded timber population
which might be typical for the daily use of the grading machine. In general these assessments
are done for entire geographical regions, e.g. assessments for the gross supply in France or
Scandinavia suggest common settings for certain grading machines used in these countries or
regions.
Rules for machine grading are described in building standards such as prEN 14081 in Europe,
AS 4490, AS 3519 in Australia, ASTM D6570 in the US and NLGA SPS2 in Canada. Except
of the US where in general exclusively the output control system is used, both, the output
control and machine control system is used. For the output control in general the so called
CUSUM methods are used for the continuous control of the grading machine output. In the
prEN 14081 an explicit description of a machine control system is given. Both methods are
briefly described in the following.

CUSUM – An Output Control Scheme

The most common used schemes for continuous output control machine grading systems are
the cumulative sum ( CUSUM ) methods. CUSUM methods for quality control problems are
described e.g. in Ewan and Kemp (1960) and linked to the timber grading problem in Warren
(1978). To date CUSUM methods are proposed by building standards as the NLGA SPS2
(Canada) and prEN 14081 (Europe).
The principle feature of CUSUM techniques is that successive values of a variable, e.g. mean
values of control samples, are compared with a predetermined target or reference value, and
the cumulative sum of deviations from this value is plotted on a chart or recorded in
tabulation. If the accumulation reaches or exceeds a pre-determined decision interval, this is
taken to indicate that a change has occurred in the mean level of the variable. The decision
interval may represent either an ordinal distance on a chart having time or sample number as
its abscissa, or a constant for tabulation. To illustrate the main aspects of CUSUM methods a
possible layout for output controlled machine timber grading is described, mainly following
an example given in Leicester and Breitinger (1994).
It is differentiated between the attribute chart and the variable chart procedure. The attribute
chart procedure is based on measuring the number of times that a required attribute, catt ,i is
missing. Specifically the attribute is taken as the event e.g. a strength value is in excess of the
value of the proof load. The variable chart procedure is based on measuring a mean value,
msc ,i of a control sample taken frequently from the output. The usual sample size for these
samples is n 5 . The charting procedure is based on three control parameters kc , yc and zc ,
where yc  zc . The values for the control parameters depend on the size n of the control
sample, the coefficient of variation of the variable of interest and are different for the attribute
and variable chart procedure. For example, if the size of the control sample is n 5 and the
coefficient of variation ( cov ) of the property to be controlled is 25% the control parameters

51
kc , yc and zc can be specified as:
x Attribute Chart (Independent from cov ): kc , yc , zc 1,1, 6
x Variable Chart : kc , yc , zc 0.9625P , 0.475P , 0.672 P
with P , the mean value of the property to be controlled.
For the i  th sample, a sum can be computed as:

SUM i CUSUM i 1  catt ,i  kc (4.1)

for an attributes chart and

SUM i CUSUM i 1  kc  msc ,i (4.2)

for a variable chart. CUSUM i 1 denotes the value of CUSUM after the previous sample is
tested. E.g. the actual CUSUM i for the variable chart is derived according to Table 4-1.
When the process switches to ‘out-of-control’, in general, some check on the stress grading
process must be performed. If no processing errors are detected, an intensive sampling is
performed, e.g. more samples are collected. If the process does not return to ‘in-control’ the
production is halted. This procedure is illustrated in Figure 4-4.

Table 4-1 Rules for computing CUSUM (Leicester and Breitinger (1994)).
Actual CUSUM i
Previous CUSUM i 1
SUM i d 0 0  SUM i  yc SUM i d yc yc  SUM i  zc SUM i t zc

CUSUM i 1 0 0 SUM i zc zc zc

0 d CUSUM i 1  yc 0 SUM i zc zc zc
yc d CUSUM i 1  zc 0 0 0 SUM i zc
CUSUM i 1 zc 0 0 0 SUM i zc
CUSUM i 1 d yc : Process is in-control
CUSUM i 1 ! yc : Process is out-of-control

52
SUM i out-of-control
in-control

yc
t

clear run
production run
START STOP

Figure 4-4 Principle of CUSUM control procedure.

Discussion of the CUSUM method

Being a common control procedure in the general field of quality control the CUSUM method
has proven to be an operational scheme for the output control of machine grading systems.
However, the control mechanism concentrates on mean values of variables of interest and is
not sensitive to possible changes of the variance of the variables which are assumed to be
known and constant. This could be seen as a reasonable assumption, but on the other hand it
could be interpreted as a shortcoming of the method, since changes in variance may be
present and lower fractile values are sensitive to these changes.
Alternatively to the CUSUM method, a Bayesian updating scheme (see Annex A) could be
employed to integrate the gathered information more efficiently, i.e. continuously gain more
information about the variable and its statistical properties.

The prEN 14081 approach for machine control systems

The new European grading code prEN 14081 prescribes an approach for the derivation of
grading machine settings according to the machine control system. The basis of this approach
is first introduced in Rouger (1996). The machine’s grading performance is compared with
that of a perfect machine capable of grading each piece of timber to its optimum grade. The
comparison is made by assigning utilities for wrongly graded timber. In the following the
main aspects of this approach are introduced and discussed.
1) Optimum Grade:
Timber grades are defined by requirements which have to be fulfilled in terms of the
statistical properties of some material properties of the graded sub-population. In Europe these
requirements are defined through a strength class system given in EN 338 (see section 3.4.3)
and are referred to the:
x 5%-fractile value of the bending moment capacity,
x 50%-fractile value of the bending stiffness,
x 5%-fractile value of the timber density,

53
all properties measured in tests according to EN 408.
The optimal set of different timber grades for a given population is defined as a set where
every single component is assigned to its highest possible grade. To obtain this set the
following steps are examined (following Rouger (1996)):
x For a specific geographic region a large ( n t 900 ) and representative sample is assessed
in regard to measurements on the bending moment capacity rm ,i , the bending stiffness
moem,i , the density U den ,i and the indicative properties.
x The values of the measurements are ranked in ascending order that:
rm rm ,1 , rm ,2 ,..., rm ,n with rm ,1 d rm,2 d ... d rm,n ;
moe m moem ,1 , moem,2 ,..., moem ,n with moem,1 d moem ,2 d ... d moem ,n ;
ȡ den U den ,1 , U den ,2 ,..., U den ,n with U den ,1 d U den ,2 d ... d U den ,n .
Note that following this approach rm ,i , moem ,i , U den ,i do not essentially correspond to the
same specimen. The ranked data is plotted in quantile plot (see Annex A1).
x For the highest grade the sample is cut at a quantile level such, that the sub-sample above
the cut level fulfils the requirements of this grade. If the cutting levels are at different
quantiles for the different properties the cutting level with the greatest rank has to be
taken as being the relevant.
x The sub-sample below the cutting level is taken to be cut again to obtain a sub-sample
corresponding to the next timber grade below and so on (see Figure 4-5).
x The optimal grading is defined as the proportions of different grades assigned according
to the segmentation procedure as described above. Typically the proportion of the
highest grade is large. A possible optimal grading could be as given in Table 4-2.

54
 Fulfills the bending strength
1 requirement for C 40
0.8

PDF
0.6

0.4

1 0.2
Fulfills the bending strength
requirement for C 35
0.8 0 1
0 20 40 60 80 100
0.6 Bending Strength [MPa] 0.8
PDF

PDF
0.4 0.6
Cut Level j
0.2 0.4
1

0 0.2
0 20 40 60 80 100 0.8
0
Bending Strength [MPa]
PDF

0.6
Cut Level k 0 20 40 60 80 100
Bending Strength [MPa]
0.4

0.2

0 1
0 20 40 60 80 100
Bending Strength [MPa] 0.8

0.6

PDF
Cut Level l
0.4

0.2

0
0 20 40 60 80 100
Bending Strength [MPa]

Figure 4-5 The principle of the segmentation method exemplified on the bending strength
(arbitrary but typical example).

Table 4-2 Optimal Grading (arbitrary but typical example).

Grade Proportion

C40 720 (72%)

C30 130 (13%)

C24 89 (8.9%)

reject 61 (6.1%)

2) The establishment of a model which relates bending strength with the indicating properties
of the machine. Beside linear regression models also more complex models can be used.
3) The determination of a set of machine settings which results in timber grades by using the
model derived in 2.
4) The entire population is (virtually) graded according to these settings, into the so-called
assigned grades.
5) The assigned grades are compared with the optimal grades in a so-called size matrix. A
possible size matrix could be given as in Table 4-3.

55
Table 4-3 A possible size matrix for the grade combination C40-C30-C24-reject (arbitrary but
typical example).
Assigned Grades
Optimum Grade
C40 C30 C24 reject

C40 254 412 51 3

C30 11 28 87 4

C24 1 20 45 23

eject 0 3 48 10

6) The determination of an elementary cost matrix.

According to Rouger (1997) it is considered as unprofitable if timber is assigned to a grade
greater than its optimal grade (upgrading) or if timber is assigned to a grade less than its
optimal grade (downgrading). Both up- and downgrading is assumed to be associated with
costs derived according to the scheme described next.
a) Upgrading:
The mean value of the optimum grade P X ,opt is calculated based on its characteristic value,
i.e. the 5%-fractile value x0.05,opt and assuming that the property is lognormal distributed with
a coefficient of variation ( cov ) of 30%:

P X ,opt exp ln x0.05,opt  0.5cov 2  1.65cov (4.3)

Similar, the mean value of the assigned grade P X ,ass is estimated.

By assuming a target reliability index of E target 3.0 , an applicable stress to the timber of the
assigned grade sass is calculated as:

sass P X ,ass  Etarget P X ,ass cov (4.4)

sass is then applied to the timber sub-population corresponding to the optimum grade and a
reliability index E is calculated as:

P X ,opt  sass
E (4.5)
P X ,opt cov

A cost of upgrading is then defined as the difference between this reliability index and the
target reliability index:

Cost upgrading Etarget  E (4.6)

56
b) Downgrading:
The cost of downgrading is assigned to be related to the over dimensioning of bending
elements in deflection (it is considered that deflection is the most common design criteria):

moem ,opt
Cost downgrading 3 1 (4.7)
moem ,ass

where moem,opt and moem,ass are the MOE’s of the optimum and the assigned grade.
Based on the above calculations a so-called cost matrix can be derived. A possible cost matrix
is illustrated in Table 4-4.

Table 4-4 A possible elementary Cost Matrix (upgrading, dark grey; downgrading, light grey),
(arbitrary but typical example).
Assigned Grades
Optimum Grade
C40 C30 C24 eject

C40 0 0.053 0.084 0.326

C30 0.111 0 0.029 0.26

C24 0.222 0.083 0 0.224

eject 0.778 0.5 0.333 0

7) The determination of the global Cost Matrix.

Each value globalij in the global cost matrix is obtained by multiplying the corresponding
number in each cell of the size matrix sizeij by the corresponding value of the elementary cost
matrix elementaryij and then by dividing with the total number of pieces in the assigned
grade, as:

elementaryij sizeij
globalij (4.8)
¦ size
i
ij

A possible global cost matrix is given in Table 4-5.

57
Table 4-5 A possible global Cost Matrix (arbitrary but typical example).

Optimum Assigned Grades

Grade C40 C30 C24 eject

C40 0 0.047162 0.018545 0.02445

C30 0.00459 0 0.010922 0.026

C24 0.000835 0.003585 0 0.1288

eject 0 0.00324 0.069195 0

8) Assessment of the global cost matrix.

The grading machine settings are considered as appropriate when all global upgrading costs
are lower than 0.02. In Table 4-5 this requirement is violated.

Discussion on the prEN 14081 approach

The principles of the procedure for the derivation of grading machine settings according to
prEN 14081 part II are briefly summarized above. In the opinion of the author the procedure
exhibits several shortcomings in both, the practical and the theoretical point of view:
Practically, it is considered as a disadvantage that the procedure does not allow for a
probabilistic assessment of the material properties based on the derived grading settings. I.e.
the gathered information cannot be directly used for the estimation of distribution parameters
of the graded timber.
From the theoretical viewpoint several inadequacies can be identified, which, in the opinion
of the author, are too serve to be considered as simplifications or assumptions. The general set
up of the method, the assignment of so-called optimum grades by the segmentation technique
(compare Figure 4-5) and taking these segments as a reference is debatable. Furthermore, it
seems incorrect to assign Lognormal distributions with a coefficient of variation ( cov ) of
30% to these segments and perform a reliability calculation (see Equations (4.3) and (4.5)).
By inspecting the right part of Figure 4-5 it is clear that the segments are not log-normally
distributed and the coefficient of variation is not 30%. In addition, the assumption is not
consistent, i.e. for some of the optimum grades the assumption is less inadequate than for
others. Another feature which appears inappropriate is the assumption that the cost for
upgrading is proportional to the reliability index. Possible cost, however, would better be
related to the probability of failure.
Even more features of the method could be questioned; however, the discussion is not
continued here.
Quite recently the prEN 14081 reached its approval phase. A future task for the research
community is to reflect upon the proposed method for the derivation of grading machine
settings. Real costs are involved in the structural timber production line and one major issue

58
of competitiveness of structural timber is the consistent representation of the uncertainties
according to the timber material properties. It is expected that new more efficient methods
will come up soon.
An example for an alternative procedure is presented in the following. Referring to Faber et
al. (2004) and Köhler and Faber (2003) a procedure for the probabilistic modelling of graded
timber material properties is derived. In section 4.2.3.2 the procedure is introduced with the
example of one relevant (or grade determining) material property. In section 4.2.3.4 the
method is followed further and a framework for the optimal assignment of grading machine
settings is presented.

4.2.3.2 Probabilistic Modelling of Grading Schemes

In structural reliability applications it is necessary to be able to assess the probability
distribution function of the relevant material properties. Assuming that a sufficiently large
number of experiments have been performed regarding the relevant material property X of
the ungraded timber, it is in principle a straightforward task to select a probability density
function and to estimate the parameters of this correspondingly. The resulting density function
might be considered as a prior density function f Xc x , i.e. the probability density function,
which might be assumed if no grading schemes are invoked. However, when a grading
scheme has been applied the prior density function is no longer representative for the graded
timber specimens. In order to assess the representative probability density function use may
be made of Bayes’s rule, see e.g. Benjamin and Cornell (1970), yielding the posterior
probability density function f Xcc x , i.e. the probability density function, which can be
assumed for the material properties, categorised into a particular grade by application of the
grading acceptance criteria AC :

1
f Xcc x P X x AC f Xc x P AC X x (4.9)
c
where c P AC .
The grading acceptance criteria AC is the rule applied for the categorisation of timber into
different grades and also refered to as the grading machine settings. The grading acceptance
criteria AC may be formulated in terms of the values of the indicators. Typically the criteria
have the following appearance:

AC ^bL d I d bU ` (4.10)

where bL and bU are lower and upper bounds for the indicator for a particular grade.
It is seen from Equation (4.9) that it is necessary to estimate the likelihood of the
implementation of the grading acceptance criteria, i.e. P AC X x as a function of the
specific value of the material property X . This likelihood may be assessed if test results are

59
available from timber specimens tested both, in regard to the indirect characteristic, e.g. the
flatwise bending stiffness, and the relevant material property, e.g. the edgewise bending
strength. Assuming that such test results are available a regression analysis can be performed
based on which the statistical characteristics of the indicator, e.g. the flatwise bending
stiffness can be assessed for a given value of the relevant material property. The prior
probability distribution function for the un-graded timber material properties together with the
likelihood of implementing a given selection criteria for a given sample thus forms the basis
for assessing the posterior probability distribution function using Bayes’s rule.
The regression analysis takes basis in n simultaneous observations of the relevant material
T T
property x x1 , x2 ,..., xn and the indicator Ț L1 ,L2 ,...,Ln . Assuming that at least locally a
linear relationship between x and L exist the regression may be performed on the basis of:

L a0  a1 ˜ x  H (4.11)

where a0 and a1 are the regression coefficients and where H is an error term. Assuming that
the error term H is normal distributed with zero mean and unknown standard deviation V H the
maximum likelihood method, see e.g. Lindley (1965) may be used to estimate the mean
values and covariance matrix for the parameters a0 , a1 , V H .
The likelihood is then given as:

n
1 § 1 § L  a  a ˜ x · 2 ·
L a0 , a1 , V H – exp ¨  ¨ i
¨ 2©
0 1 i
¸ ¸ (4.12)
i 1 2SV H ©
VH ¹ ¸¹

The parameters are estimated by the solution p to the optimisation problem:

max L p (4.13)
p

T
where p a0 , a1 , V H .
Having performed the regression analysis, it is possible to assess the acceptance probability
i.e. P AC X x by:

P AC X x P bL d a0  a1 X  H d bU X x P bL d a0  a1 x  H d bU (4.14)

which is straightforward to assess recognising that the indicator I is normal distributed.

For structural reliability applications the probability distribution function for the relevant
material characteristic is often more directly useful than the probability density function.
Considering the probabilistic modelling of the relevant material characteristics corresponding
to the different grades it may, e.g. in the context of First Order and Second Order Reliability
(FORM/SORM) analysis be more appropriate to consider the probability distribution function
given in Equation (4.15) instead of the probability density function from Equation (4.9):

60
 P X d x P AC X d x
P X d x AC (4.15)
P X d x P AC X d x  P X ! x P AC X ! x

However, the probability distribution defined by Equation (4.15) is also not always
straightforward to work with because it involves the calculation of the terms P AC X d x
and P AC X ! x which in general are difficult to express analytically.
If the only aim is to assess the probability distribution function FORM/SORM (Madsen et al.
(1986)) as well as Monte Carlo based simulation techniques (Melchers (1987)) may, however,
also be directly applied by use of the following formulation:

P X d x  AC
P X d x AC (4.16)
P AC

Example
As an example the case where the grading of the timber material is performed using the
Computermatic grading machine is presented. A data set from Johansson (1992) is considered
concerning the bending strength of 239 timber specimens of European spruce.

F = constant force
deflection measurement
=> stiffness indication
infeed speed
flatwise

l = 914 mm

Figure 4-6 Principles of the Computermatic Grading Machine.

The Computermatic grading machine is a widely used grading machine which considers the
flatwise bending stiffness as an indicator of the bending strength. As illustrated in Figure 4-6,
the stiffness characteristic is obtained through a measurement of the deflection under a fixed
load applied at the centre of a span of 914 mm. The bending strength is obtained in four-point
bending tests according to the European standard EN 408, compare section 7.3.1. Further
information regarding the Computermatic grading machine can be found in the literature, e.g.
Computermatic user manual (2000). The entire population of the 239 timber specimen is
assumed to be un-graded. The probability distribution function of the bending strength, i.e. the
prior probability distribution function, is assumed to be 2-parameter Weibull min distributed.
The 2-parameter Weibull min distribution function is given in Annex A. It is observed e.g. in

61
Glos (1978) that the 2-parameter Weibull distribution gives too conservative estimates for the
lower tail region of the distribution. The parameters of the Weibull distribution are therefore
also estimated with explicit consideration of the lower tail region of the data.
In Figure 4-7 the probability distribution function representing all data is compared to the
probability distribution function estimated by using the censored maximum likelihood
estimation according to section 7.1 with a threshold value corresponding to the lower 30 %
quantile. For the purpose of comparison the sample probability distribution function using all
observations is also illustrated.
In Figure 4-7 it is seen that a significant refinement of the representation of the strength data
in the lower tail domain can be achieved, by using the distribution model fitted to the lower
data set domain.
In Johansson (1992) not only the bending strength of the 239 specimens but also the
Computermatic based indicator of the bending strength for each specimen has been observed
and recorded prior to the bending strength tests. By regression analysis of simultaneously
measured bending strengths and the observed Computermatic based indicators the regression
coefficients are estimated according to Equations (4.11)-(4.13). Using the Method of the
Maximum Likelihood both the parameters of the prior probability distribution functions and
the regression parameters are estimated as normal distributed random variables with mean
values, standard deviations and correlations as given in Table 4-6. Ignoring the statistical
uncertainty associated with the estimated parameters, which for the present case is
insignificant, the acceptance probability may thus be assessed directly from Equation (4.14).

1.0
fitted to data
0.8 lower 30%
fitted to all
Probability

0.6 data

0.4

0.2 Sample PDF

0
0 20 40 60
Bending Strength [MPa]

Figure 4-7 2-parameter Weibull min probability distribution functions with parameters estimated
using all data and data below 30%, compared with the sample probability distribution
function (Sample PDF).

62
Table 4-6 Mean values, standard deviations and correlations of the parameters of the prior
probability distribution function (w, k see also the definition in Annex A) together with
the estimated regression parameters.
Prior Distribution Parameters (bending
Regression Parameters (Computermatic)
mod. of rupture)

Weibull parameters as
as normal distributed random
normal distributed correlation correlation
variables
random variables

U a0 a1 [Comp/ VH U
w [MPa] k [MPa]
[Comp] MPa] [Comp]

All
P = 48.0 P = 4.32
w, k = 0.33 a0 , a1 = -0.02
data V = 0.76 V = 0.21 P = 4673 P = 98.6 P = 1073
V H , a1 =0
Data P = 43.9 P = 6.44 V = 291 V = 6.45 V = 40.08
lower w, k = -0.06 V H , a0 =0
30% V = 0.02 V = 0.49

1.0 E-Modulus
Comp
0.8

Density
Probability

0.6

0.4

0.2

0
0 15 30 45 60 75
Bending Strength [MPa]

Figure 4-8 Acceptance probability of different indicators.

In Figure 4-8 the acceptance probability (Equation (4.14) for three different types of
indicators is illustrated. The Computermatic based indicator is compared with the case when
the timber density and edgewise modulus of elasticity are used as an indicator. The grading
acceptance criterion for each probability curve is chosen such that the mean value of the
expected bending strength is the same in all three cases. It is not surprising that the ‘better’ the
linear regression is between the indicator and the relevant material property, the steeper is also
the corresponding acceptance probability curve. The steepness of the acceptance probability
curves thus forms a basis for comparison of the efficiency of different grading methods.
Assuming that the timber specimens are categorised in three grades of same volume according

63
to the observed Computermatic indicators, i.e. 1) 4000 – 8100 MPa, 2) 8100 – 9650 MPa and
3) 9650 – 15000 MPa, the probability density function for the graded timber can be obtained
directly from Equation (4.9).
In Figure 4-9 the prior probability density function of the bending strength for the un-graded
timber (2-parameter Weibull min distributed with mean value 43.7 MPa and standard
deviation 11.4 MPa) is shown together with the probability density functions for the three
timber grades in accordance with the abovementioned grading acceptance criteria. Note that
the grading is performed using the prior probability density function with parameters
estimated based on all data.

0.05 Grade 2
Grade 3
Grade 1
0.04

0.03
Density

0.02
All
0.01

0
0 15 30 45 60 75
Bending Strength [MPa]

Figure 4-9 Illustration of the probability density functions for the bending strength for un-graded
and graded timber specimens.

In Figure 4-10 the probability distribution functions for the different grades are shown in
comparison with the sample probability distribution functions (sample PDF’s) established
from the experimental data fulfilling the respective grading acceptance criteria.

64
 Sample PDF Grade 1
1.0 Sample PDF Grade 2
Sample PDF Grade 3

0.8
Probability

0.6
Grade 1
0.4 Grade 3

0.2
Grade 2

0
0 15 30 45 60 75
Bending Strength [MPa]

Figure 4-10 Probability distribution functions for the bending strength for the three different
grades together with the corresponding sample probability distribution functions.

In Figure 4-10 the lower tail domain of the probability distribution functions are shown in
larger scale together with the lower 5% fractile values.

Sample PDF Grade 1

0.5 Sample PDF Grade 2
Sample PDF Grade 3

0.4
Grade 2
Probability

0.3
Grade 1

0.2
Grade 3

0.1
0.05
0
18.2 28.8 37.75
Bending Strength [MPa]

Figure 4-11 Lower tail domain of the graded probability distribution functions. Estimates of the
lower 5% fractiles (characteristic values) are shown.

Note, that especially the representation of the data by the probability distribution function for
grade 1 is not satisfactory. For the weaker timber, e.g., grade 1 and grade 2, special emphasis
on the lower tail domain in the representation of the prior probability distribution function is
thus required. Figure 4-12 and Figure 4-13 shows the three probability distribution functions
compared with the sample probability distribution functions. For identifying the probability

65
density function of grade 1 and 2, now, the prior probability density function is fitted with
special emphasis on the lower tail domain using the censored maximum likelihood method
according to section 7.1. The data which fulfils the grade 3 acceptance criteria are mostly
taken from the middle and upper part of the prior distribution. For that part the prior
probability distribution function modelled under consideration of all data can be used. The
sample cumulative probability distribution functions corresponding to the graded observations
are also shown in Figure 4-12.

1.0 Sample PDF Grade 1

Sample PDF Grade 2
Sample PDF Grade 3
0.8
Probability

0.6

0.4 Grade 3
Grade 1

0.2
Grade 2
0
0 15 30 45 60 75
Bending Strength [MPa]

Figure 4-12 Probability distribution functions for the bending strength for the three grades. The
prior probability distribution function for grade 1 and 2 is fitted to the lower tail.

0.5 Sample PDF Grade 1

Sample PDF Grade 2
Sample PDF Grade 3
0.4
Probability

0.3

0.2
Grade 1
Grade 2
0.1 Grade 3
0.05
0
23.4 30.5 37.75
Bending Strength [MPa]

Figure 4-13 Lower tail domain of the probability distribution functions for the bending strength for
the three grades. Estimates of the lower 5% characteristic values are shown.

66
In Figure 4-13 the lower tail domain is shown in larger scale. It is seen by comparison
between the theoretically derived probability distribution functions and the sample probability
distribution functions (corresponding to the experimental observations in the considered
intervals) that a rather satisfactorily agreement in the lower tail domain is achieved.
In Table 4-7 the lower 5 % fractile values are given for the different grades with and without
estimating the parameters of the prior probability distribution function in the lower tail
domain using the censored maximum likelihood method compared with the 5% fractile values
resulting directly from the data.
From Figure 4-13 it is seen that a great benefit may be obtained by estimating the prior
probability distribution with special consideration of the tail domain before the probability
distributions of the graded specimens are evaluated.
Whether a prior fitted with special emphasis to the goodness of fit in the lower tail domain
has to be preferred or not, depends on the domain for which the posterior is of special interest.
It is thus difficult to provide any generally applicable rules for this selection. However, in
most normal cases the selection would follow naturally from the available data on the
simultaneous observations of an indicator and the bending strength together with previous
experience to judge which approach is appropriate.

Table 4-7 Lower 5% fractile values for the graded timber specimens.
char. value [MPa] char. value [MPa] char. value [MPa]
acceptance interval f Rcm fitted to lower tail f Rcm fitted to all sample values

4000 - 8100 23.4 18.2 25.0

8100 - 9650 30.5 28.8 30.5
9650 - 15000 - 37.71 36.3
All data (ungraded) 24.1 28.4

4.2.3.3 Probabilistic Model Format for Timber Materials in consideration of grading

As a prerequisite to the consistent probabilistic design of timber structures and
correspondingly, the calibration of e.g. load and resistance factor design formats, it is
necessary to establish a standardised format for the representation of the relevant timber
material properties. Such a format must contain both a description of the prior probability
distribution of the relevant timber material properties as well as a description of the quality of
the different quality control and selection for grading procedures.
In Table 4-8 a possible format for probabilistic models of relevant timber material properties
of interest is suggested. First the distinction of different prior timber populations is proposed
e.g. in respect to the geographical origin of the timber, the species, the dimension of the
considered timber elements and other possible characterising properties. Representative prior
probability distribution models of the relevant material property may be specified in terms of

67
probability distribution functions, their parameters and the uncertainties associated with their
parameters. Furthermore the different schemes for quality control and grading may be
represented in terms of their linear regressions and associated uncertainties.
To establish a probabilistic description of timber material properties in a code format such as
e.g. the Probabilistic Model Code by the Joint Committee on Structural Safety, the parameters
and their uncertainties should be represented in the same way. The procedure described in this
chapter is a proposal in this direction and should be taken as a basis for further discussions.

Table 4-8 Suggested Format for Probabilistic Models for relevant Material Properties of Timber
Materials.
Prior distibution Parameters Regression Parameters
characteristics of the population
(relevant material property) (for a given grading method)
origin timber species size ....... w k ȡ a0 a1 ıİ ȡ
(a 1 ,a 0 ) =
Specie or μ= μ= μ= μ= μ=
Region or Country 1

(w,k) = (a 1 ,ı İ ) =
speciegroup 1 ı= ı= ı= ı= ı=
(a 0 ,ı İ ) =
(a 1 ,a 0 ) =
Specie or μ= μ= μ= μ= μ=
(w,k) = (a 1 ,ı İ ) =
speciegroup 2 ı= ı= ı= ı= ı=
(a 0 ,ı İ ) =
......

......

......

......

......

......

......

......

......

......
(a 1 ,a 0 ) =
Country 2
Region or

Specie or μ= μ= μ= μ= μ=
speciegroup 1 ı= ı=
(w,k) = ı= ı= ı=
(a 1 ,ı İ ) =
(a 0 ,ı İ ) =
...... ......

...... ......

...... ......

...... ......

...... ......

...... ......

...... ......

...... ......

...... ......

...... ......
......

User specified populations

(a 1 ,a 0 ) =
μ= μ= μ= μ= μ=
Pop 1
ı= ı=
(w,k) = ı= ı= ı=
(a 1 ,ı İ ) =
(a 0 ,ı İ ) =
Pop 2

The statistical properties of the graded timber materials depend as discussed previously also
on the grading acceptance criteria (grading machine settings) implemented for the different
grades. It is hardly possible to prescribe how these should be selected as this should be a
matter of economical optimisation for the individual saw mills. The selection of a quality
control and grading scheme must be made on the basis of cost benefit considerations. These
aspects are considered in the next section.

4.2.3.4 Assessment of the Optimal Grading Procedure

Based on the proposed statistical modelling of timber properties as a function of the type and
efficiency of the grading procedures, a cost optimisation scheme may be formulated for the
identification of the optimal grading procedure. Therefore, the cost of the control and the
benefit of fulfilling the requirements set for the material characteristics belonging to different
grades have to be given.

68
The benefit BT of a set of timber grades identified by the grading procedure GP may be
written as:

BT ( f Xc ( x), AC , GP) V T ˜ CGrade (4.17)

where V is a vector of volumes of particular grades, which can be identified depending on

the prior probability distribution of the relevant material property, the set of grading
acceptance criteria and the grading procedure. CGrade is a vector of the (monetary) benefit of
the timber grades per unit volume. If the prior probability distribution of the relevant material
property, the grading procedure, the set of timber grades and their monetary benefits are
known, the optimal set of grading acceptance criteria AC may be found by solving the
following optimisation problem:

max BT ( f Xc ( x), AC , GP)

Ac (4.18)
subject to: N req , CGrade

subject to normative requirements which have to be fulfilled by the grades, N req , and the cost
vector CGrade . Involving the investment, maintenance over lifetime costs, costs for personnel,
etc. of a particular grading procedure by the function CG (GP ) the optimal grading procedure
may be identified by solving the following optimisation problem:

ª§ · º
max max BT ( f Xc ( x), AC , GP)  CG (GP) »
« ¨ ¸ (4.19)
GP «¨ Ac ¸ »
«¬© subject to: N req , CGrade ¹ »¼

Example - continuation
Considering the data and the inference from the example in section 4.2.3.2, it can now be
demonstrated how an optimal set of grading acceptance criteria for the considered grading
procedure of the Computermatic grading machine can be found by applying Equation (4.18).
The intention is to set up three different grading acceptance criteria to select three different
timber grades. The grades have to fulfil the requirements for the strength classes according to
EN 338 in terms of the 5% fractile value of the bending strength and the mean of bending
modulus of elasticity and the 5% fractile value of the density. The prior population is assumed
to be checked according to visual defects of the timber as warp, insect damage etc. Five
different grades according to EN 338 are selected to be possible grades and a reject domain is
defined for timber, which not has to fulfil any requirements. They are shown together with
example values for the monetary benefit and the material requirements in Table 4-9, therefore
the vector C grade (units: monetary unit per volume) becomes:
C grade (0.5, 1.0, 1.4, 1.6, 1.8, 2.0)T .
To identify three different grades, three different grading acceptance criteria have to be

69
defined:

AC1 ^bL ,1 d I  bL ,2 ` ; AC2 ^b

L ,2 d I  bL ,3 ` ; AC3 ^b
L ,3 d I d bU ,3 ` (4.20)

Table 4-9 Monetary benefit (example values) and requirements for grades according to EN 338.
‘reject’ C16 C24 C30 C35 C40

Monetary Benefit 0.5 1.0 1.4 1.6 1.8 2.0

Bending strength, 5%-frac [MPa] - >16 >24 >30 >35 >40

Modulus of elasticity, mean value

- >8000 >11000 >12000 >13000 >14000
[MPa]

Density, mean value [kg/m3] - >310 >350 >380 >400 >420

In addition to the posterior distributions of the ultimate bending strength for these three
grading acceptance criteria P Rm d rm AC , the posterior distributions of the bending
i

modulus of elasticity P MOEm d moem AC and of the density P 5 den d U den AC have to be
i i

assessed, according to the prior distribution parameters and the regression parameters given in
Table 4-6, Table 4-10 and Table 4-11. These posterior distributions are representing timber
sub-populations with the property I  ACi and can be checked in regard to the requirements in
Table 4-9.
The three grading acceptance criteria are defined through the values of bL ,1 , bL ,2 , bL ,3 and bU ,3
in Equation (4.20). The value bU ,3 is fixed to an upper threshold value, namely
bU ,3 16000 MPa . The optimal values of bL ,1 , bL ,2 and bL ,3 are then found by solving
Equation (4.18) using the simplex algorithm, see e.g. Nelder and Mead (1965), for which
differentiability of the objective function and constraints is not required. The applied
algorithm does not allow for the definition of the constraints explicitly. Therefore the
constraints have been included directly into the formulation of the objective
function BT ( f Xc ( x), AC , GP) in line with the requirements given in EN 338. Due to the discrete
character of the objective function the simplex algorithm is applied in conjunction with a
random search to facilitate the identification of the optimal solution over the set of discrete
optimisation variables. In this search it has been found that 100 simulations provided stable
results.

70
Table 4-10 Mean values, standard deviations and correlations of the parameters of the prior
probability distribution function together with the estimated parameters for the
regression between bending modulus of elasticity and the Computermatic indicator.
Prior Distribution Parameters (bending
Regression Parameters (Computermatic)
modulus of elasticity)

Log-normal parameters as
normal distributed random correlation as normal distributed random variables correlation
variables

a 1 [Comp
[ [MPa] G [MPa] U a0 [Comp] VH[Comp] U
/MPa]

a0 , a1 = -0.03
P = 12749.1 P = 0.1949
P = 2892 P = 0.469 P = 961.7
[ ,G =0 V H , a1 =0
V= 332.1 V= 0.025 V= 35.91
V= 160.7 V= 0.0089 V H , a0 =0

Table 4-11 Mean values, standard deviations and correlations of the parameters of the prior
probability distribution function together with the estimated parameters for the
regression between density and the Computermatic indicator.
Prior Distribution Parameters (density) Regression Parameters (Computermatic)

Normal parameters as
normal distributed random Correlation as normal distributed random variables correlation
variables

U a0 [Comp/ U
P[kg/m3] G [kg/m3] a0 [Comp] VH[Comp]
kg/m3]

P = 406.3 P = 35.47 (a0, a1) = -0.02

P = 1063.6 P = 19.51 P = 1341.0
[ ,G =0 (VH, a1) = 0
V= 997.24 V= 2.44 V= 50.1
V= 2.295 V= 1.622 (VH, a0) = 0

Table 4-12 Optimal set of timber grades, limiting properties are shaded and framed.
bending mod.
bend. mod. of density 5%-
of rupture 5%- Grade
I [MPa] Vi elasticity mean fractile value
fractile value EN 338
value [MPa] [kg/m3]
[MPa]
AC r [0,5255) 0.5% reject
AC1 [5255,8094) 29.5% 26.0 11000.4 339* C24
AC2 [8094,10161) 49% 30.3 12952.34 351.6* C30
AC3 [10161,16000] 21% 40.1 15738.9 366* C40

*)density is not taken into account as a requirement.

Subject to the given conditions summarised in Table 4-9, the optimal set of grading

71
acceptance criteria is given in Table 4-12. From the assumed set of data the grades C24, C30
and C40 have been identified, where for the grade C24 the mean value of the bending
modulus of elasticity is the limiting property. For the grades C30 and C40 the bending
strength is limiting. The relationship between the Computermatic indicator and the timber
density is too weak to make any quantitative prediction about the density based on a
Computermatic indication. Therefore, in this example the density is not taken into account as
a requirement for assigning grades.
As the Computermatic grading procedure apart from values of the flatwise bending stiffness
(Indicator) also provides values of density, the parameters of Table 4-11 can be updated
continuously. By doing spot test, e.g. for every 100th timber specimen, of the bending
strength and the bending modulus of elasticity the parameters of Table 4-6 and Table 4-10 can
also be updated continuously.

4.2.3.5 Summary and Concluding Remarks on Grading

A brief overview about timber grading is given, with focus on the control of grading
machines. Grading machines can be either output controlled or machine controlled. A
common output controlled procedure is the so-called CUSUM method and the main features
of this method are briefly explained and discussed. A procedure for machine control is
introduced following the method which is prescribed in the European standard prEN 14081. It
is noted that both control methods do not allow for a probabilistic assessment of the graded
timber material properties based on the corresponding formalism of the control methods.
Furthermore, several shortcomings within the machine control method according to the
European standard prEN 14081 are outlined and discussed.
Alternatively to the described methods, the statistical assessment of timber material properties
has been considered with special emphasis on the modelling of the effect of different schemes
for quality control and grading of timber. It is shown that typically applied indicators for the
relevant material properties such as measurements of the flatwise bending stiffness, the
measurement of densities and others may be utilised for the statistical differentiation of
different populations of the strength characteristics. The suggested approach not only forms a
very strong tool for the statistical quantification of the material characteristics of timber but
furthermore provides a consistent basis for quantifying the efficiency of different quality
control and grading procedures. The probabilistic models for the graded timber material
properties have been formulated such that they readily may be applied in structural reliability
analysis.
It has been found that the assessment of the statistical characteristics of the graded timber
should be performed on the basis of a probabilistic modelling of the un-graded timber where
special emphasis is made on the representation of the probability distribution function in the
lower tail domain.
It is of utmost importance that the statistical characteristics of timber material properties are

72
assessed and treated in consistency with the implemented quality control and grading
procedures. Only then a consistent basis may be established for the quantification of the
reliability of timber structures - the basis for codification of design and assessment. The
suggested probabilistic modelling seems to provide the required framework for establishing
such a basis by means of quantifying the efficiency of the different quality control and
grading procedures. It is envisaged that different quality control grading procedures may be
described by means of their regression characteristics and acceptance probability curves
corresponding to different grading criteria. A format for the standardisation of the
probabilistic modelling of timber materials subject to different quality control and grading
procedures is suggested. It is important that the appropriateness of such a format is discussed
and that a consensus is achieved in this respect in the near future.
In section 4.2.3.4 it has been demonstrated how an optimal (in terms of monetary benefit) set
of timber grades can be identified through the solution of an optimisation problem. The
objective function of the optimisation problem is defined based on the findings of section
4.2.3.2. The identified timber grades can be described by means of the probabilistic
characteristics of the relevant material properties as e.g. the bending strength, the bending
modulus of elasticity and the density of the timber. The simplex algorithm for the optimisation
of non-differentiable objective functions in conjunction with a simulation procedure has been
applied for the identification of the optimal grading procedure. The constraints to the
optimisation problem, in terms of the requirements for timber grades according to EN 338
have been incorporated directly into the objective function.
An example has been presented illustrating the suggested approach to cost optimal timber
grading. The assignment of monetary benefit to the different grades of timber has, however,
been based on judgement rather than true values. In practice the benefit associated with timber
of a particular grade would depend on a number of factors such as the size of the individual
timber specimen, the total amount of available timber for a given grading, the production
capacity of a given sawmill, the available grading machines and not least the market price for
the different timber grades. The implementation of the proposed approach in practice would
have to incorporate these and other factors more accurately into the formulation of the benefit
function. Further studies in close collaboration with the timber industry should be undertaken
and discussed to clarify these aspects and to set up a rational basis for their assessment.
However, according to the preferences of a sawmill owner the proposed approach facilitates
the identification and the calibration of a grading procedure and thus an increase in the overall
production benefit.

4.2.4 WITHIN MEMBER VARIATIONS – MICRO SCALE VARIATIONS

For modelling the variations of strength related timber material properties within a
component, the variability on a micro scale has to be taken into account. In general,
assumptions about the micro system behaviour of the material are providing guidelines for the

73
development of stochastic material models. Two classical probabilistic strength models can be
found in the literature: the ideal brittle material model and the ideal plastic material model.

4.2.4.1 Classical Probabilistic Strength Models

Ideal Brittle Material

An ideal brittle material is defined as a material that fails if a single particle fails, see e.g.
Bolotin (1969). The strength of the material is thus governed by the strength of the ‘weakest’
particle; therefore the model for ideal brittle materials is also called the weakest link theory
proposed by Weibull (1939). If the strength X of the individual particles are assumed to be
mutually independent and identically distributed with distribution function FX ( x) , the
distribution function of the strength R of a body with nvVol identical loaded particles1 is:

nvVol
FR x 1  1  FX x 1  exp nvvol ln 1  FX x (4.21)

The behaviour of ln 1  FX x for small arguments can be represented as:

k
ln 1  FX x # c x  x0 ; x t x0 (4.22)

where x0 is the smallest possible value of x , c and k are positive constants. Equation (4.21)
now can be written as:

k
FR x 1  exp cnvvol x  x0 (4.23)

Considering an isotropic solid under a homogeneous stress distribution and by introducing a

reference volume vVol ,0 , e.g. the volume of standard test specimen substituting by
cn 1 vVol ,0 wk in (4.23) gives:

§ v § x  x0 · k ·
FR x 1  exp ¨  Vol ¨ ¸ (4.24)
¨ vVol ,0 © w ¸¹ ¸
© ¹
which is the expression of a 3-parameter Weibull distribution. w is called the shape parameter
and k the scale parameter. The expected value E > R @ and the variance Var > R @ of R can be
given as:

1
k
§ 1 ·§ v ·
E > R@ x0  w* ¨1  ¸ ¨ Vol ¸ (4.25)
© k ¹ ¨© vVol ,0 ¸¹

1
n particles of finite volume vVol .

74
 2
k
§ § 2· § 1 ··§ v ·
Var > R @ w2 ¨ * ¨1  ¸  * 2 ¨ 1  ¸ ¸ ¨ Vol ¸ (4.26)
© © k¹ © k ¹ ¹ ©¨ vVol ,0 ¹¸

where * . is the gamma distribution.

The expected value function decreases with the considered volume vVol and for x0 0 the
coefficient of variation is independent of vVol . The size effect is observed for many brittle
materials such as hardened steel, concrete and stiff clay. The coefficient of variation and the
correlation of the strength of the reference volumes vVol ,0 determine the importance of the
volume effect.
Assuming unique failure probability for two material bodies with different volumes vVol ,1 and
vVol ,2 gives the following expression:

§ v § x  x ·k · § v k
§x x · ·
FR x1 , vVol ,1 1  exp ¨  Vol ,1 ¨ 1 0 ¸ ¸ 1  exp ¨  Vol ,2 ¨ 2 0 ¸ ¸ FR x2 , vVol ,2 (4.27)
¨ vVol ,0 © w ¹ ¸ ¨ vVol ,0 © w ¹ ¸
© ¹ © ¹
which results in:

1 1
k k
x2  x0 § V1 · x
x0 0 § V1 ·
¨ ¸ o 2
 ¨ ¸ (4.28)
x1  x0 © V2 ¹ x1 © V2 ¹
Equation (4.28) can directly be used to compare the load bearing capacity of different
volumes for loading modes resulting in constant stress fields and brittle failure modes.
If an inhomogeneous stress distribution can be considered by introducing the stress
s [1 , [ 2 , [3 as a product of a reference stress s , (e.g. the maximum stress in the body) and a
dimension free function h [1 , [ 2 , [3 , the probability distribution function of the strength r of
the body can be written as:

k
­° 1 § xh [1 , [ 2 , [3  x0 · ½°
FR x 1  exp ® ³ ¨ dv
¸ Vol ¾ (4.29)
°¯ vVol ,0
xh [1 ,[ 2 ,[3 ! x0 w
© ¹ °¿

Equation (4.29) is utilised in Johnson (1953) to compare the mean strength of beams with
various support and loading conditions and standard deviation of reference strength. A
graphical representation of the result of these calculations is presented in Figure 4-14. The
ordinate in the figure is the mean strength of the beam normalised to the strength of a beam
loaded by a constant bending moment. It is seen that the mean strength of beams exposed to
other load configurations is larger and also depending on the coefficient of variation of the
material strength.

75
 2.00
mean beam strength [-]

1.75

1.50

1.25

1.00
0 5 10 20 50
coeff. of variation [%]

Figure 4-14 Comparison of mean strength of rectangular beams with various loading conditions
and standard deviation of reference volume strength, from Johnson (1953).

Ideal Plastic Material

In an ideal plastic material a particle yields when reaching its maximum capacity (yield load),
i.e. a particle is still capable of carrying the yield load but further load increments are
transferred to other particles in the loaded material body. The maximum load is reached when
yielding takes place in all particles in a cross section. The strength R of an ideal plastic
material is thus equal to the sum of the strength Ri of the yielded particles. The cross section
strength is:

R ¦R
i
i (4.30)

When the number of particles is large and the dependency between the particle strength is
sufficiently low, the distribution of R converges to a normal distribution due to the central
limit theorem.
The sum can be approximated by the integral:

R ³ su* [1 , [ 2 daare (4.31)

aare

where su* [1 , [ 2 is the yield stress at [1 , [ 2 .

4.2.4.2 Timber Material as an ideal brittle material

Over the last 40 years, the ideal brittle material model has been widely applied to study the
tension, bending, bending shear and compression strength variability of various wood
products including sawn and glued laminated timber. However, whenever the theory is

76
utilised to explain some observed phenomena, inconsistencies to the theory have been
noticed. In Bohannan (1966) one of the first studies is published, showing that size effects and
load configuration effects on the bending strength could be explained by the weakest link
theory for brittle materials. Clear wood specimens of different size and loading configurations
are analysed. It is shown that increasing length or depth causes a decrease in bending strength
but the bending strength is independent of the width of the specimen. Inductively, it is
concluded that bending members do not exhibit a perfectly brittle material behaviour upon
which the weakest link theory is based. In Madsen and Stinson (1982), the width effect on the
bending strength of structural timber is studied. Their results suggest that bending strength
increases as the member width increase, which is contrary to the concept of a weakest link
fracture process. It is suggested in Madsen and Buchanan (1986) that this inconsistency could
result from visual grading rules which limit knot sizes on the wide and narrow faces of the
member. These rules tend to limit the maximum size of knots for a fixed member depth
independent of member width. In the same study it is suggested to consider size effects on the
strength of timber separately depending on every single dimension of the member. According
to Equation (4.32) it is distinguished between length, width and depth effect:

1 1 1 ml mb md
kl kb kd
rm,2 § l1 · §b · §d · § l1 · § b1 · § d1 ·
¨ ¸ ¨ 1 ¸ ¨ 1 ¸ ¨ ¸ ¨ ¸ ¨ ¸ (4.32)
rm ,1 © l2 ¹ © b2 ¹ © d2 ¹ © l2 ¹ © b2 ¹ © d2 ¹
where l is the length, b is the width and d is the depth of the specimen; rm is the bending
strength property of the specimen.
A similar distinction is made in a study of the size effect on the tension strength, Madsen
(1992). In that comprehensive research project, specimen of different sizes, grades and
species are analysed. Equation (4.32) is utilised to quantify m -values for two different fractile
values of the underlying probability distribution functions of the tension strength. The
following values are suggested, Madsen (1992):
Length effect: 50%-Fractile ml 0, 20
10%-Fractile ml 0,15

Width1 effect: 50%-Fractile mb 0,15

10%-Fractile mb 0,10

For tension perpendicular to the grain a nearly perfect agreement with the weakest link theory
can be observed. In Barrett (1974) the work of several authors is reviewed and summarised.
Data from clear wood specimens, structural timber and glued laminated timber is brought in

1
For tension specimen the width refers to the wide face of the cross section.

77
one context.
Furthermore, in Barrett et al. (1975) and subsequently in Colling (1986) the perfectly brittle
material model is successfully applied to model tension perpendicular to the grain failure
modes in curved and pitched tapered beams and connections. These studies confirm that the
Weibull model has wide application in modelling of wood fracture especially for tension
perpendicular to the grain. In Foschi and Barrett (1975) and in Colling (1986) it is confirmed
that shear strength of beams varies with member size and the effects are quantified by using
the ideal brittle material model.
In Madsen (1992) it is shown that the bending strength of sawn timber of constant thickness
varies with member length and loading condition in a manner consistent with the ideal brittle
material model (Equation (4.28) and (4.29)). Further, tension and compression tests with
different lengths are analysed. The length dependencies are quantified with the shape
parameter k (compare Equation (4.28)); for tension and bending similar values are derived,
ktension kbending 5 , and for compression kcompression 10 is established. In Barrett and Fewell
(1990) Canadian, US and European species data is analysed and length effect factors for
bending and tension are derived by also using the ideal brittle material model. Similar values
for bending and tension are found; ktension kbending 5.9 . An important result of these studies
is the observation that length effects in tension and bending are very similar; i.e. inductively it
could be concluded that the bending and tension strength of structural timber is both governed
by the ultimate tension strength. On the other hand in Rouger and Fewell (1994) it is found
that 60-90% of the bending specimens considered in this study showed ductile failure mode
governed by the compression side.
In Larsen (1986) and Colling (1986) a formulation based on Equation (4.29) is utilised to
investigate the stress distribution effect on the shear strength and the tension perpendicular
strength of curved and tapered beams.
There is an ongoing list of further literature and many applications of the ideal brittle material
model for timber can be found. However, the results of these investigations are delivering
partly contrary results. Contrary in regard to the value of quantified parameters as e.g. the
scale parameter k , but also in regard to the observed phenomena. This can be explained by
the variability of the timber material at a macro scale, i.e. different species and grades exhibit
different size effects (Rouger and Fewell (1994)); but also by the fact that the assumptions
underlying the theory of brittle fracture are not strictly fulfilled when considering timber.
Timber is referred to as an orthotropic material; strength and stiffness properties are
depending on the stress direction in a timber solid. Therefore the theory can only be used for
individual loading modes for which the stress direction can be assumed to be constant, i.e.
timber components loaded to tension perpendicular or parallel to the grain, shear along the
grain or pure bending of beam shaped timber specimen. The latter loading mode includes
tension and compression and has to be considered with caution because the failure mode in
compression is not brittle. The material irregularities are in the scale of P m if the fibre

78
configuration is considered or in the scale of dm when considering major defects as knots
and grain deviations1. The latter ones are almost in the same scale as structural components,
which is in conflict of the assumption of ‘a big number of defects, identical distributed and
independent’ – the material is not statistically homogeneous.
In summary it can be concluded that the ideal brittle material model can be used in some cases
as an empirical model basis which describes the major phenomenon, but where the
parameters, i.e. scale and shape parameters do not have any physical meaning; but can be
derived by fitting to experimental data. Especially for tension perpendicular to the grain the
physical deviations to the initial theory seem to be small.

4.2.4.3 Weak section models for special load cases for timber structural elements
A typical timber structural element is of beam shape which means that its longitudinal
extension is much larger than its transversal extensions and the main fibre direction is
orientated along the longitudinal (main) axis of the element. These elements are mainly
loaded in tension, compression and/or bending along the main axis. According to the
inhomogeneous structure of these elements due to major defects such as knots and grain
deviations, the direct application of the ideal brittle material model is questionable. Therefore,
an alternative model for the variation of strength and stiffness related material properties is
desirable. In regard to the variation of the modulus of elasticity ( MOE ) several references
can be found in the literature, see e.g. in Corder (1965), Kass (1975), Suddarth and Woeste
(1977) and Foschi and Barrett (1980). In Kline et al. (1986) a probabilistic approach to
describe the lengthwise variation of bending stiffness is introduced. As illustrated in Figure
4-15 timber beams are divided in segments of identical length and the stiffness of segment 2
to segment n  1 is measured with a 4-point bending test. The correlation between the
MOEi ’s of the different segments is investigated and quantified by the so-called Lag-k serial
correlation. The Lag-k serial correlation is the correlation of an observation from segment i
and an observation from segment i  k ; e.g. if k 2 , Lag-k means the correlation of the
MOE of segments 2 and 4 or of the segments 3 and 5 etc. For realisations of the MOE X ;
x1 , x2 , x3 ,..., xn the Lag-k serial correlation U k is defined as:

nk

¦ xi  x xi  k  x
Uk i 1
n (4.33)
2
¦
i 1
xi  x

with x being the sample mean value of X . It is found that the serial correlation decreases
with increasing k . The lengthwise variation of X is modelled by a second order Markov
model as:

1
Grain deviation is defined as in section 3.3.1.2.

79
X i 1 E 0 X i  E1 X i 1  H i 1 (4.34)

where E i , i 1, 2 are regression parameters and İ is the vector of random errors, the
components are assumed to be normal distributed with mean value 0 and unknown standard
deviation.
Similar approaches are utilised to describe the lengthwise variation of tension and bending
strength. Obviously only every second section can be tested destructively (compare Figure
4-15). In Showalter et al. (1987), Lam and Varoglu (1991) and Taylor and Bender (1991) the
tension strength is considered, in Czmoch (1991) the bending strength is considered; all
studies find decreasing serial correlation with increasing k .

Load

Segment 1 Segment 2 Segment n

Figure 4-15 Separation in segments and test arrangement.

As seen in Figure 4-15 the regular segmentation does not facilitate the explicit consideration
of observable irregularities in the beams. Examples for such observable irregularities are
knots and knot clusters. In Riberholt and Madsen (1979) it is observed that low bending
strength and bending stiffness coincides with the presence of knots and knot clusters. In this
study it is assumed that failure can only occur at such weak sections and due to the discrete
distribution of knots and knot clusters an idealised model is proposed in terms of discrete
weak sections separated by strong sections – sections of clear wood, see Figure 4-16.
Furthermore equicorrelation of the strength of weak sections is assumed, which means that
the correlation between the strength of weak sections is independent on their distance over the
length of the beam.

beam defect
no defect

bending strength - reality

bending strength - model

Figure 4-16 Bending strength of a timber beam; implied reality and as in the proposed model
(Riberholt, Madsen (1979)).

80
Failure of one single weak section is determining the strength of the entire component and
consequently the component can be modelled as a series system of weak sections. The
strength of the weak section is described by a random variable whereas the location of the
weak sections is modelled as an arrival ‘time’ of a Poisson process. The parameters of the
Poisson process are estimated by direct measurements of the distances of knots and knot
clusters. The parameters of the distribution function of the weak section strength is estimated
indirectly; through measurements of the bending stiffness as an indicator for bending strength
or by existing bending strength tests according to EN 408.
The proposed model for the variation of bending strength properties is investigated by
Czmoch et al (1991). Here the length and load configuration effect of beams is studied and it
is found that the available experimental information to verify the parameters of the proposed
model is insufficient. This problem is considered in Isaksson (1999).
Following Isaksson (1999) the variability in bending strength within and between members
can be modelled as:

ln rm ,ij Q  Y i  F ij (4.35)

where,
rm ,ij is the strength of weak section j in component i .
Q is the logarithm of the mean strength of all weak sections of all components.
Yi is the realisation of the difference between the logarithm of the mean of the strength of
the sections within a specific component i and Q . Y is modelled by a normal
distributed random variable with zero mean and standard deviation equal to V Y .
F ij is the realisation of the difference between the strength of weak section j in
component i and Q  Y i , i.e. the variability within one particular component in one
particular population. F is modelled by a normal distributed random variable with
zero mean and standard deviation equal to V F .

Based on experiments with Norwegian spruce, the following random variables of the model
are quantified:

Distance between weak sections

It is assumed that the weak sections (and correspondingly the number of weak sections) are
distributed spatially according to events of a Poisson process X i along the longitudinal axis
of the component. Following this assumption the distances between the weak sections are
exponentially distributed. However, due to the experimental evidence it is found that the
Gamma distribution fits better to the data. The parameters of the Gamma distribution are
given as: * 2.55,194 .

81
Length of weak sections
The length of the weak sections, i.e. the length of a knot cluster, is assumed to be constant
with 150mm .

Bending strength of weak sections

Based on 673 observations the random variables from Equation (4.35) are quantified with the
parameters as given in Table 4-13 (Isaksson (1999)).

Table 4-13 Parameters for the bending strength of weak sections according to Equation (4.35).
Q Y [Normal] F [Normal]

P 4.03038 0 0
V 0.24773 0.18747 0.16194

The model is illustrated in Figure 4-17.

ln(bending strength)

ln( ri,max)

ij
ln(r ij )

X1 X2 X3 Xj longitudinal direction
of the beam

Figure 4-17 Modelling of the longitudinal variation of bending strength of a timber beam, Isaksson
(1999).

Bending strength between weak sections

In Isaksson (1999) it is assumed that the strength of the strong sections is equal to the strength
of the strongest weak section in the component.
A similar study on the lengthwise variation of bending strength is reported in Källsner and
Ditlevsen (1994, 1997), Ditlevsen and Källsner (2004). Out of 26 beams 197 weak sections
are identified and tested in regard to their bending strength by cutting them out, finger joining
them between two pieces of stronger wood and testing them in a four point bending
arrangement. A particular feature of this study is that unintentionally a large fraction of the
test specimen did not fail in the section of interest, i.e. failure took place in the finger joints or

82
in the stronger wood. Due to the censored data special methods are introduced to consider the
observations properly. However, the results are in general consistent with the results presented
in Isaksson (1999). A hierarchical model is introduced with two levels to represent the
variation of bending strength within and between members. The Lag  k correlation is found
to be constant, i.e. equicorrelation is assumed. The model proposed by Isaksson in Equation
(4.35) can be interpreted as a two level hierarchical model if v is assumed to be constant. A
closed form analytical expression for the distribution function of the bending strength with
any given number of weak section is presented and the derivation is summarised in the
following.
The ultimate strength of a timber component containing j weak sections and loaded with a
constant load effect is defined by:

rm ,i ,min exp Zi  min ^ F i ,1 , F i ,2 ,..., F i , j ` (4.36)

with Zi Q  Y i , compare Equation (4.35).

The probability distribution function Fln( Rm ) z j of the logarithm of the ultimate bending
strength ln(rm ) z conditional on j , the number of weak sections, becomes:

j
1
f ª § z  x · º § x  PZ ·
Fln( Rm ) z j P Z  min ^ F1 , F 2 ,..., F j ` d z ³ ««1  ) ¨¨  V F ¸¸ »» M ¨© V Z ¸ dx (4.37)
V Z f © ¹ ¼ ¹
¬
If j is considered as a discrete random variable with probability p j , the probability
distribution function becomes:

f
1 ª § x  z ·J º § x  P ·
f
Fln( Rm ) z ¦F ln( Rm ) z j pj 1 ³ E ««) ¨¨ V F ¸¸ »» M ¨© V Z Z ¸¹ dx
V Z f
(4.38)
j 0
¬ © ¹ ¼

ª § x  z ·J º f § xz·
j

with E «) ¨ ¸ »
« ¨© V F ¸¹ »
¦ )¨ p
¨ V ¸¸ j
¬ ¼ j 0 © F ¹
which can be expressed by the so-called probability generating function \ y E ª¬ y J º¼ of the
integer random variable J . For J following a Poisson distribution with parameter O l the
probability generating function is \ y exp O l y  1 in which case Equation (4.38)
becomes:

1
f ª § z  x · º § x  PZ ·
Fln( Rm ) z 1
VZ ³ exp ««Ol) ¨¨ V
M
¸¸ »» ¨ V ¸ dx (4.39)
f ¬ © F ¹¼ © Z ¹

where l is the length of the considered component. Equation (4.39) is the probability
distribution function of the logarithm of the bending strength of a component of length l .

83
4.3 DURATION OF LOAD EFFECTS

One of the distinctive characteristics of timber is that its strength is influenced by the intensity
and the duration of the applied stresses. Although this phenomenon is similar to that of fatigue
in metals, strength degradation in timber is observed even under static (permanent) loading.
This effect is referred to as the duration of load (DOL) effect. Numerous experimental
programs have focused on the investigation of the DOL effects in clear wood specimen and
later on also in full size timber components. A variety of models have been proposed to
describe the phenomenon. Hereby, it has been mainly focused on the duration of load effect of
bending specimen. Some of the proposed models have a physical hypothesis of the
phenomena as a basis; however, they all consist of variable model parameters which can be
calibrated to observed experimental data.

4.3.1 ASSESSMENT OF DURATION OF LOAD EXPERIMENTAL DATA

The assessment of the DOL phenomenon by experimental data introduces a variety of

challenging research problems. First of all by the nature of the problem, experiments involve
long periods of time, are therefore costly and consequently rare. To reduce costs, the test
configurations are rather simple, i.e. in general a constant (dead) load or a ramp-load is
applied and time until failure is measured. A typical bending duration of load test
configuration is the three- or four-point bending test as illustrated in Figure 4-18. Also
illustrated in Figure 4-18 are the typical stress histories for short-term tests, long-term tests
and impact load test.
In a short-term test, the stress is applied with a constant rate k R , which in general is kept
constant within the entire test program. k R is calibrated so that the time to failure stays within
some pre-defined limits; e.g. according to the test standard ISO 8375 failure should take place
after ts 300 seconds +/- 120 seconds. If, for example, k R is chosen such that in average the
short-term bending test duration is 300 seconds and the coefficient of variation of the short-
term bending strength is 25%, the test duration for a specimen with a low bending strength,
say a value similar to the 5%-fractile value is only 176 seconds (300 - 124 seconds); i.e. the
time limit from ISO 8375 would be violated.
In a long-term test the specimens are loaded with a constant stress ŝ over a period of time and
the time until failure t f is measured. Therefore, the specimens have to be loaded at the
beginning of the test. In general, this can be done by applying any specified load rate; often
the same constant rate k R as in the short-test (as illustrated in Figure 4-18) is applied. Or the
stress is applied instantly, i.e. with a very large rate which in general is not specified.
The third type of duration of load test is the impact loading test. In this type of test the loading
rate is greater than for shot- term tests, k I ! k R and the time to failure t I is measured.

84
 4-point bending test 3-point bending test

s s s
rI
r0 r0 r0

kR kR kI
ts tc ts tf t tI ts
Short-term test Long-term test Impact loading test

Figure 4-18 Stress history used in test set up; short-term and long-term tests and impact loading
test.

The short-term strength r0 is in general the reference property, i.e. the focus of duration of
load experiments is for how long a specimen can sustain a constant fraction of its short-term
strength. The ratio sˆ r0,i is therefore of importance. Similarly, impact loading tests focus on
measuring rI ,i which is then expressed as the ratio rI ,i r0,i .
Obviously, rI ,i , ŝ and r0,i cannot be measured for the same specimen and since the applied
constant stress ŝ respectively the impact load capacity rI ,i are the targets, the short-term
strength r0,i must be estimated indirectly. Several methods have been applied to overcome this
problem.
One method does not rely on the comparison of rI ,i , sˆi and r0,i of individual specimen, but in
the comparison of average values of rI ,i , sˆi and r0,i at average time to failures. This method is
applied in Wood (1949) for the analysis of data from short-term, long-term and impact load
tests and in Hanhijärvi et al. (1998) for the analysis of data from short-term and long-term
tests.
Another method is pair matching, which is undertaken by sampling specimens in pairs cut out
of the tree next to each other and therefore implying similar properties. One specimen is then
tested in a short-term test while the other is long-term loaded. The method is mainly applied
for clear wood specimen free of knots and other irregularities, but still has to allow for
substantial short-term strength differences between the two specimens (Hoffmeyer (1990)).
For structural timber, the pair matching method is assessed in Norén (1986). Pairs of
specimens are obtained by very carefully placing the saw cut not only through the pith of a
piece of timber, but even divide the critical knots into equally sized halves. However, short-
term test on the paired specimen confirm that it is not possible to find a method for pair

85
matching timber with sufficient accuracy by applying this technique.
The predominant approach for estimating the short-term strength of duration of load test
specimen is based on the so-called equal rank assumption. This method is e.g. applied in
Madsen (1992). An adequate number n of specimen is assessed in regard to a strength related
property which has to be measured non-destructively, e.g. machine grading indication or knot
size measures. The measurements Li are ranked in ascending order, so that L1 d L2 d L3 d ...Ln .
The specimens are now sub-divided into two groups with similar distribution of the strength
related properties, e.g. one group with L1 ,L3 .L5 ,...,Ln 1 and the other group containing
L2 ,L4 ,L6 ,...,Ln given that n 2  IN. Now it is assumed that both groups have also similar
bending strength distributions. One group – the short-term group – is tested in a short-term
bending strength test, while the other group – the long-term group – is long-term loaded to a
specific percentile of the short term strength distribution. Some of the specimens fail before
the constant load level is reached. The specimens which survive the duration of the long-term
test are taken to be tested in a short-term strength test. The strength measurements of the
short-term group r0,i are ranked in ascending order, so that r0,1 d r0,2 d r0,3 d ...r0,n 2 . The time to
failure measurements t f ,i of the long-term group are also ranked in ascending order, so that
t f ,1 d t f ,2 d t f ,3 d ...t f ,n 2 (note that in general, the first time to failure measurements already
take place in the ramp-load phase of the long duration test). In Figure 4-19 the strength
measurements of the specimen of both groups are plotted against their rank. The basic idea of
the equal rank assumption is that a specimen from the long-term group of rank i (failed at
time t f ,i ) has the same short term strength as the specimen of the short-term group with rank
i , r0,i . Following this assumption it is possible to express the so-called stress level sli as the
ratio of the applied constant stress ŝ and the expected short-term strength r0,i :

sˆ
sli (4.40)
r0,i

In Figure 4-20 the stress levels for the specimen of the long-term group are estimated
according to Equation (4.40) and plotted against the logarithm of the time to failure
log10 t f .
In Figure 4-20 the common representation of duration of load data is illustrated. Many DOL
models are represented in form of the stress level as the function of time to failure. Typically
the data points are arranged on a string, which is the consequence of the equal rank
assumption. The shape of the data string does not necessarily reflect the DOL behaviour of the
single specimens; it may only reflect the DOL behaviour of the particular distribution of the
considered sample, i.e. specimen within the sample which have a low quality and therefore
have a low short-term strength always sustain a high stress level and fail after relatively short
time, specimen with high quality have to sustain a low stress level and fail after relatively
long time (Hoffmeyer (1990)).

86
 60
bending strength [MPa]

50

40

30
n = 102
20
Short-term-test
10
Long-term test
0
0 20 40 60 80 100
rank i

Figure 4-19 Strength measurements from short-term and from long-term tests plotted against their
rank (the data for this figure is taken from Hoffmeyer (1990)).

sl >% @
100.0

80.0

60.0

40.0
Failed during DOL test
20.0 Survived DOL test

0.0
-3.00 0.00 3.00 log10 t f

Figure 4-20 Stress level in [%] against logarithmic time to failure in [log(hours)] (the data for this
figure is taken from Hoffmeyer (1990)).

4.3.2 MODELLING THE DURATION OF LOAD EFFECT

Several models for describing the duration of load effect for timber and timber materials can
be found in the literature. Some of them have a conceptual framework from some physical
mechanism leading to failure over time as a background; others are just empirically derived to
represent DOL test data. The models can be divided into the following groups:
x Empirical formulations for time to failure. The parameters of simple mathematical
formulations are fitted to DOL test results.
x Cumulative damage theories. A damage state variable without any precise physical

87
 definition is introduced. Empirical parameters are fitted to test results. The models and
their parameters are conditional on loading mode (e.g. bending, compression).
x Fracture mechanics. Relative damage is defined as the growth of cracks. Model
parameters have direct physical meaning.
x Deformation kinetics. Rupture is completely determined by the magnitude and nature of
deformation preceding rupture.
x Energy based models. Failure is defined as the excess of critical strain energy.
In the following sections all the different modelling approaches are briefly reviewed and the
most relevant references are given.

4.3.2.1 Empirical Representation of time to failure – the Madison Curve

The so-called Madison curve (Wood (1951)) is formulated on the basis of results from
laboratory tests on clear wood bending specimens, i.e. specimens, with no visible defects,
exposed to loads of constant intensity. The aim of this model is to cover the effect of relatively
long load durations and the effect of very short load durations. Therefore, data bases from
several resources are pooled together to obtain a model capable to cover a duration of load
ranging from 0.015 seconds to 156 days. The Madison curve is an empirical curve describing
the stress level sl (see Equation (4.40)) as a function of the time to failure t f . The curve is
calibrated to three pre-selected points representing average values derived from several test
programs:
x The short-term strength is defined as the failure load of a specimen loaded in a ramp load
test with a corresponding average time to failure of 7.5 minutes. Correspondingly, the
stress level is 100% at 7.5 minutes.
x From impact loading tests it is found that the average stress level sl of specimen loaded
in a very fast ramp-load test of the duration of 0.015 seconds is 150%.
x From long-term loading tests it is found that the average stress level sl of specimen
loaded in a constant load test of the duration of 156 days is 69%.
The resulting curve has the equation:

sl 18.3  108.4 t f 0.0464 (4.41)

where sl is the stress level and t f is the time to failure in seconds. The Madison curve is
often taken as a reference for the duration of load effect for clear wood but also for structural
timber.
In Pearson (1972) data from several investigations is reviewed, whereas it is focused solely on
long-term data. Based on the analysis of the data an alternative formulation is found:

88
sl 91.5  7 log t f (4.42)

where t f is the time to failure in hours.

In Figure 4-21 Equations (4.41) and (4.42) are compared. The models differ especially for
very long durations of loading time. The up bend of the Madison curve can not be confirmed
by explicitly considering long-duration data. Therefore the Pearson curve should be taken as a
reference for the long-term DOL effect in clear wood.

sl >% @
100
80
60
Madison Curve
40
Pearson
20
0
-2 -1 0 1 2 3 4 5 log10 t f

Figure 4-21 Models for the DOL behavior of clear wood; the so-called Madison curve (Wood
(1951), Equation (4.41)) and a model following Pearson (1972), Equation (4.42)
represented in a stress level – log(time) diagram. (time in hours).

4.3.2.2 Damage Accumulation Models

A major part of the models for describing the duration of load effect for timber involve a
damage state variable to assess damage accumulation in timber structural members subject to
their loading histories. These models are referred to as damage accumulation models. It
should be noted that damage in the context of the damage accumulation models does not refer
to a quantity with a direct physical interpretation; damage is simply deducted from the time to
failure recordings of long-term loading experiments under a given loading history. However,
possible interpretations could range from the creation and the propagation of micro cracks in
the material, the increase of material porosity, local fibre buckling, the destruction of
molecular bonds, some other irrecoverable process or a combination of these. (Hoffmeyer in
Thelandersson and Larsen (2003)).
In general, damage is expressed by a damage state variable D D , where D D 0 is equivalent
with no damage and D D 1 is associated with full damage or failure. Damage accumulation
models are usually given as rate equations of damage over time dD D dt , as:

dD D
h s t , r0 , ș for 0 d DD d 1 (4.43)
dt
where h . is a function of the applied stress s t , the short term strength r0 and model

89
parameters ș ; or:

dD D
h s t , r0 , D D , ș for 0 d DD d 1 (4.44)
dt
where h . is a function of the applied stress s t , the short term strength r0 , the actual
damage state D D and model parameters ș .
The first damage accumulation model is proposed in Gerhards (1979), where the rate of
damage accumulation is expressed as a function of the applied stress (equivalent to Equation
(4.43)). Initially, the model is developed and calibrated to long-term test data on small clear
wood specimen. In Gerhards and Link (1987) the same model is calibrated to long-term test
on full- size timber specimen. In Barrett and Foschi (1978) a damage accumulation model is
introduced, where the rate of damage is a function of the applied stress, the actual damage and
a threshold value for the stress below which no damage accumulation occurs (equivalent to
Equation (4.44)). In Foschi and Yao (1986) a model is presented, taking basis in the first two
terms of the extension of a power series. The model is also presented as the damage rate as a
function of the applied stress, the actual damage and a threshold value for the stress below
which no damage accumulation occurs, i.e. is corresponding to Equation (4.44).

4.3.2.3 Models based on Fracture Mechanics

In contrast to the damage accumulation models, where damage is introduced as an abstract
quantity without any direct physical background, the models based on fracture mechanics
introduce damage as the increase of crack size in a viscoelastic solid.
An approach based on viscoelastic fracture mechanics is introduced in Nielsen (1979) with
the so called Damaged Viscoelastic Material (DVM) theory. The main idea behind the DVM -
model is that structural timber may be seen as an initially damaged material, where the
damage is represented by cracks. The time dependent behaviour of timber under load is
modelled by a single crack under stress perpendicular to the crack plane. The crack is
modelled as a Dugdale crack with a time dependent modulus of elasticity; see e.g. Dugdale
(1960).
Another approach using the Dugdale crack model applied to a viscoelastic material is
presented in Mindess (1976), where it is focused on delayed crack propagation under constant
or increasing load:

dacrack b2
b1k stress (4.45)
dt
where acrack is the crack length, t is the time, k stress is the stress intensity factor and b1 , b2 are
model parameters depending on the material. Equation (4.45) can be integrated for several
load histories. In Sharpely (1975) Equation (4.45) is utilized and the model parameters b1 , b2
are assigned based on theoretical considerations taking basis in the so called Barenblatt

90
theory, see Barenblatt (1962).

4.3.2.4 Models based on Chemical Kinetics

In Caultfield (1985) and van der Put (1986) two DOL models based on chemical kinetics are
proposed that relate fracture times to bond breaking processes in the wood. These models are
based on physical parameters, however, the governing equations are rather complex and the
application to structural timber with its inherent defects and high material variability may be
difficult (Morlier et al. (1996)).

4.3.2.5 Models based on Energy considerations

A strain energy model to predict load duration effects in timber is proposed in Fridley et al.
(1991) and the model is developed further in Philpot et al. (1994). A critical strain energy
density function is established from experimental observations. Failure is defined as the
excess of a critical strain energy density, gu ,cr , which is assumed to correspond to the
initiation of non-linear material behaviour. For instance, gu ,cr in a standard short-term test
would correspond to the proportional limit; in a long-term test gu ,cr would correspond to the
initiation of tertiary creep. The strain energy definition of failure is found to be invariant with
respect to load history, environmental conditions and strength and stiffness related material
property. Another proposal for an energy-based DOL model can be found in Bach (1973).

4.3.3 PROBABILISTIC MODELLING OF DURATION OF LOAD EFFECTS

The probabilistic modelling of duration of load effects is a prerequisite for reliability analysis
and reliability based code calibration for timber structures. The analysis takes basis in the time
dependent reliability problem as discussed in section 2.1, where Equation (2.3) specifies the
time dependent failure criterion as:

R t S t d0 (4.46)

at any time t ; i.e. normally the considered time interval is limited to a projected service
lifetime of a structure; t  > 0, Tsl @ .

4.3.3.1 Modelling Scheme - Uncertainty

In principle, the probabilistic modelling of the life time of a timber structural element follows
a scheme as illustrated in Figure 4-22. On the load side (the left hand side in Figure 4-22),
relevant scenarios have to be identified which might be idealized and represented by proper
random processes. The parameters of these processes are in general estimated based on
observations. On the resistance side (the right hand side in Figure 4-22) the procedure starts
with the identification of relevant failure mechanisms. In section 4.3.2 it is already discussed

91
how diverse the assumptions regarding the relevant mechanisms are. However, a
mathematical model is formulated based on these assumptions and the parameters of the
models are calibrated to experimental data. It is an important task within the probabilistic
modelling that all uncertainties involved are consistently taken into account.

Load Resistance
Identification of relevant Identification of relevant
load scenarios failure mechanism

Idealisation of load scenarios Idealisation of failure

- random process mechanism - DOL model

Parameter estimation Parameter estimation -

Observations - data - random process probabilistic DOL model Observations - data

Probabilistic
Modeling
of
DOL
Effects

Figure 4-22 Scheme for the probabilistic modeling of duration of load effects.

The uncertainties involved in the long term resistance model are briefly discussed in the
following:
x Failure mechanism assumptions/model formulations: The model can be formulated based
on different assumptions about the failure mechanism. The basic assumptions and the
possibly simplified mathematical representations are associated with uncertainties.
x Parameter estimation: the model parameters are calibrated with (limited) experimental
data.
x Extrapolation to other load scenarios: due to the limited availability of experimental data,
also the load scenarios applied within the experiments are limited. It has to be assumed
that models can be applied to load scenarios different to the scenarios used in the
experiments.
x Sampling techniques such as the equal rank assumption: as discussed in section 4.3.1 the
stress ratio sli sˆ r0,i is a central measure for the analysis of long-term strength data.
Assumptions such as the equal rank assumption induce additional uncertainty to the
problem.

92
The uncertainties discussed above can be partly quantified by proper calibration methods as
the maximum likelihood method. However, the remaining uncertainty e.g. due to the equal
rank assumption or due to extrapolation has to be accounted for separately.

4.3.3.2 DOL model representations

In general, damage accumulation models are used for the probabilistic modelling of the DOL
effect; see e.g. Ellingwood and Rosowsky (1991), Foschi et al. (1989), Sørensen et al. (2002).
The basic characteristic of damage accumulation models is that the performance of the
element is described by one single damage state variable D D , which is in general a function of
the short-term strength R0 and the applied stress history S t . Taking this into account the
limit state function in Equation (4.46) can be simplified to:

1  D D S t , R0 , ș d 0 (4.47)

where ș is a vector of model parameters.

The model parameters of the damage accumulation models are mostly calibrated to data from
standard long-term experiments, i.e. where a ramp load phase is followed by a period of
constant load until failure and the time to failure is measured (compare section 4.3.1). For
reliability analysis and code calibration these models are used for any realistic and relevant
load process. This includes the assumption that the failure mechanism is solely due to creep
rupture, i.e. failure is assumed to be a consequence of accumulated time under load. This
means that the effect of repeated load cycles is assumed to be irrelevant, a fact which
provokes some scepticism, especially for load processes with a high number of load
repetitions. Furthermore, the parameters of the damage accumulation models have no direct
physical meaning and cannot be adapted to other situations in regard to climate or loading
mode.
The model proposed by Nielsen (2000) (see section 4.3.2.3) seems to be promising in order to
overcome these problems. For constant loads a relatively simple representation for this model
can be given and parameters can be calibrated by using standard long-term experiment data as
for the damage accumulation models. However, the parameters are linked to some physical
meaning and an algorithm for taking into account repeated load cycles exist.
In the following, three damage theories are investigated further. The Gerhards model
(Gerhards (1979) and Gerhards and Link (1987)) and the Foschi and Yao model (Foschi and
Yao (1986)) are typical representatives for damage accumulation models. The Nielsen model
(Nielsen (1979) and Madsen (1992)) with its promising properties is also discussed and
compared with the two other models.

Gerhards Model
The Gerhards model has the form:

93
dD D § s t ·
exp ¨  aD  bD ¸ for 0 d DD d 1 (4.48)
dt © r0 ¹

aD and bD are model parameters, r0 is the short-term capacity and s t is the applied stress.
The Gerhards model corresponds to the general formulation given in Equation (4.43).
Equation (4.48) can be integrated for any stress history s t to determine the accumulated
damage D D . The time to failure t f is defined at D D t 1 .

t
§ s t ·
DD t ³ exp ¨© a D  bD ¸dt (4.49)
0
r0 ¹

For a constant load effect and assuming that the effect of the initial ramp loading can be
neglected the time to failure according to the Gerhards model can be estimated as:

§ sˆ ·
tf exp ¨ aD  bD ¸ (4.50)
© r0 ¹

where ŝ is the constant stress level (Note that this format is equivalent to the one proposed by
Pearson (1972), Equation (4.42)).
The Gerhards model allows for the consideration of the characteristics of the short-term test
configuration. As illustrated in Figure 4-18 the short-term test involves a certain time of ramp
loading with rate k R . The integral given in Equation (4.49) can be solved for any ramp
loading s t k R ,a t , as:

r0 § § kR ,at · ·
DD t exp  aD ¨¨ exp ¨ bD ¸  1¸ (4.51)
bD k R ,a © © r0 ¹ ¸¹

For the case of the short-term test, where k R ,a kR , D D 1 , k R t r0 1 the parameter aD can
be represented by:

§ r0 exp bD  1 ·
aD ln ¨ ¸ (4.52)
¨ bD k R ¸
© ¹

This means that whenever the short-term strength is linked to a certain short-term test loading
rate k R , aD can be eliminated through Equation (4.52).

Considering this elimination, Equation (4.51) can be simplified to:

94
 § k t·
exp ¨ bD R ,a ¸  1
kR © r0 ¹ (4.53)
D D (t )
k R ,a exp bD  1

The integral given in Equation (4.49) can be solved for the stress history of a long-term test as
illustrated in Figure 4-18. A ramp loading with rate k R is applied until the constant stress level
ŝ is reached after the time tC . Damage accumulation can be assessed through D D (t ) as:

§ § kRt · ·
¨¨ exp ¨ bD ¸  1¸
© © r0 ¹ ¸¹
DD t for 0 d t  tC
exp bD  1
(4.54)
§ sˆ · § § sˆ · ·
exp ¨ bD ¸ ¨¨ exp ¨ bD ¸  1¸¸
bD t  tC © r0 ¹  © © r0 ¹ ¹
DD t for tC d t
tC exp bD  1 exp bD  1

When ŝ  r0 , the time to failure can be evaluated as

sˆ r § § § sˆ · · ·
tf  0 ¨ exp ¨¨ bD ¨1  ¸ ¸¸  1¸ (4.55)
k R k R bD ¨ ¸
© © © r0 ¹ ¹ ¹
Equation (4.55) is used for calibrating model parameters with test results.
The residual strength rr corresponding to the damage D D can be derived such that
D D 0 Ÿ rr r0 1 and D D 1Ÿ rr r0 0 . For Gerhards model the residual strength can be
given as:

rr 1
1  1  D D exp bD  1 (4.56)
r0 bD

Foschi and Yao Model

In Foschi and Yao (1986) a power series expansion is used to express damage accumulation
according to the general form as given in Equation (4.44), as:

dD D
h s t , r0 , D D , ș h1 s t , r0 , ș  h2 s t , r0 , ș D D  ...  hn s t , r0 , ș D D n 1 (4.57)
dt
As an approximation, Yao (1987) considered only the first two terms of the expansion.
Accordingly the so called Foschi and Yao model is expressed as:

95
dD D bD dD
aD s t  K D r0  cD s t  K D r0 DD t for s t t K D r0
dt
(4.58)
dD D
0 for s t  K D r0
dt
The accumulated damage is also depending on the actual value of the damage state variable
D D t . aD , bD , cD , d D and K D are model parameters. K D defines the stress level threshold
below which no damage is assumed to occur. The model can be rewritten as (Köhler and
Svensson 2002):

bD dD
dD D § s (t ) · § s (t ) · s t
aD ¨  K D ¸  cD ¨  K D ¸ D D (t ) for t KD
dt © r0 ¹ © r0 ¹ r0
(4.59)
dD D s t
0 for  KD
dt r0

Equation (4.59) has the advantage to present stress as a dimensionless ratio of the applied
stress and the assumed short-term strength.
By solving Equation (4.59) for the case of standard short-term tests, one will ascertain also for
Foschi and Yaos’ model one model parameter, aD , with:

k R (bD  1)
aD (4.60)
r0 (1  K D )( bD 1)

where k R is the rate of loading applied for evaluating r0 .

The time to failure, t f , for a prior-to-testing undamaged component loaded with a ramp load,
with the same rate of loading, k R , as the standard test, until the constant load level, ŝ , is
reached and held until failure occurs has the following expression (see Köhler and Svensson
(2002)):

sˆ 1 § 1  OD ·
tf  ln ¨ ¸¸ sˆ t K D R0 , sˆ  r0
kR § sˆ ·
dD ¨D O
© D ,0 D ¹
cD ¨  K D ¸ (4.61)
© r0 ¹
tf f sˆ  K D R0

with

bD  d D bD 1
k R bD  1 § sˆ · § sˆ r0  K D ·
OD ( bD 1) ¨
KD ¸ , D D ,0 ¨ ¸ (4.62)
cD r0 1  K D © r0 ¹ © 1 KD ¹

96
where D D ,0 is the degree of damage prior the studied period of constant load.
Equation (4.61) is used when calibrating the model against test results.
Residual strength rr corresponding to the damage D D for the Foschi and Yao model can be
expressed as:

rr 1 1 bD
KD  1 KD 1  D D (4.63)
r0

Nielsen Model
In case of constant load intensity, a damage accumulation law can be formulated from the
DVM-theory as:

2
dD N S ˜ fl DN ˜ slN2
dt 8qcW c 2
1
1
bc (4.64)
DN ˜ slN 1

where slN is the stress level (or load intensity) defined by the ratio sˆ rcr between applied
stress ŝ (load intensity) and the strength rcr measured in a very fast ramp-load test. fl is the
strength level defined as the ratio rcr rl between the strength rcr (as defined above) and the
intrinsic strength of the (hypothetical) non-cracked material rl . DN l l0 is the damage,
defined as the ratio between the actual crack length l and the initial crack length l0 . DN 1
corresponds to no damage and DN slN2 to full damage. W c and bc are (creep) material
parameters depending e.g. on loading mode and moisture history. The parameter qc is given
as a function of the creep exponent bc as:

1
bc
qc 0.5 bc  1 bc  2 (4.65)

and considers a parabolic increasing crack opening progression.

Based on Equation (4.64) time to failure t f under constant stress can be derived as:

slN2 1

8qcW c M 1 bc

tf 2 ³ dM (4.66)
S ˜ fl ˜ slN 1
M

where M is a damage state variable. Nielsen’s definition of the stress level slN thus differs
from the usual definition, which relates applied stress to short-term strength as it is defined in
Figure 4-18, i.e. a ramp load test with a specified average time to failure of e.g. 5 minutes
according to ISO 8375. The creep parameters W c and bc are related to the viscoelastic
behaviour around the crack tip. These are very local phenomena and cannot be consistently
reflected in macro-scale creep tests. qc is a function of the creep parameter bc . The parameter
fl indicates the state of initial damage of the material as it is defined by the ratio of the

97
material strength rcr and the (hypothetical) material strength without any damage rl which
can only be assessed theoretically. This means that, although the model parameters have some
physical meaning, they must be obtained by calibration, as it is the case for other damage
accumulation models. Equation (4.66) can be used for calibrating the parameters of the
Nielsen model to data from standard long-term tests. Therefore it is assumed that the Nielsen
stress level slN can be replaced by the conventionally defined stress level sl . Further, it is
supposed that the initial ramp load phase of a standard long-term experiment (Figure 4-18)
can be neglected.
The integral, included in Equation (4.66) can be solved for discrete realisations of the
parameter bc ; in general values for 1 bc 3, 4,5 are used and the solutions can be given as
(Bronstein and Semendjajew (1981)):

bc 1 3.1W c § x 3 x 2 ·
3 tf 2 2 ¨
  x  log x  1 ¸
slN fl © 3 2 ¹

bc 1 3.2W c § x 4 x 3 x 2 ·
4 tf 2 2 ¨
   x  log x  1 ¸ (4.67)
slN fl © 4 3 2 ¹

bc 1 3.2W c § x 5 x 4 x3 x 2 ·
5 tf 2 2 ¨
    x  log x  1 ¸
slN fl © 5 4 3 2 ¹
with x slN2  1 .
The parameters fl and W c are not independent. The model therefore includes only one free
parameter to calibrate.
For the Nielsen model the residual strength rr is given by:

1
rr 
DN 2
(4.68)
r0

Fatigue modelling with the Nielsen model

Theoretical investigations as well as physical evidence from test results indicate that the
duration of load failure mechanism in timber is not entirely due to creep rupture but also due
to accumulated fatigue damage. Frequency dependent fatigue effects in timber have been
considered theoretically in Nielsen (2000) and experimentally e.g. in Clorius (2001). Based on
the solution for constant load intensity the DVM-model is developed further in Nielsen (2000)
to consider rectangular pulse loading by introducing crack closure and contact effects due to
unloading.
The load process which can be considered by the Nielsen model is a square wave loading with

98
constant amplitude and constant frequency. The load process is determined by frequency
f 1 / T , maximum load intensity slmax , load ratio psl slmin slmax and fractional time under
maximum load E sl . The parameters are illustrated in Figure 4-23.

sl

sl max

sl min

0
T T 1 T 2T t
0 sl sl

Figure 4-23 Basic load intensity variation considered: Square wave loading.

The algorithm proposed in Nielsen (2000) is illustrated in Figure 4-24 and starts with the
input of material and load parameters; į is a vector of fatigue parameters. The other
parameters are defined in the foregoing. After setting the start-values, a x which satisfies the
Equation in step C3 is found by iteration. With the Equation in step C4 and the constant
damage increment 'DN the number of load cycles 'n leading to that damage increment is
calculated. The damage is accumulated until the failure criterion formulated in the Equation
in step L1 is fulfilled. Load conditions may only be changed when the algorithm returns to
step C3 again, i.e. after an unknown number of load cycles. Not that this circumstance
restricts the model to consider only harmonic load pulse processes.
By setting 'n instead of 'DN to a constant value, e.g. 'n 1 , the load intensity can be
changed after every load pulse and the corresponding damage increment is calculated. This
facilitates the consideration of realizations of a random process with rectangular load pulses
(see Köhler and Faber (2003)).
The model as it is presented in Figure 4-24 is referred to as the general Nielsen model. The
model as it is presented in Equation (4.64) is also capable to model the duration of load effect;
damage is then accumulated as a consequence of time under constant load. The model
presented in Equation (4.64) is referred to as the simple Nielsen model.

99
 INPUT:

Material: fl, c , bc , C1
Load: sl, psl , f sl

START:
t 0; n 0; 1;
'
C2
const (Nielsen);
'n const (this study)

configure
load parameters

f1 , sl , f , sl , c , bc x bc  f 2 , sl , x
C3
 f3 , sl 0 Ÿx

§ ' ·
x ¨ fl 2 ¸ Ÿ 'n or ' C4
© 'n ¹

ACCUMULATION:
n n  'n; ' ;t t  'n f C5

true
 sl 2
L1

false

END:
n cat n ; t cat t C6

Figure 4-24 Algorithm for predicting the time till failure.

4.3.3.3 Calibrating Models to experiment observations

Long-term experiments are illustrated in Figure 4-18. The experiments indicate a time to
failure t f after a certain stress history, where a period of constant stress ŝ follows a period of
loading with the rate k R .

100
For the considered DOL models an expression for the time to failure considering the
experiment stress history can be derived. These expressions can be written in the general form
as:

tf t f sl,ș ,k R (4.69)

The estimation of the model parameters is performed considering n simultaneous

T
observations of the time to failure t f t f ,1 , t f ,2 ,..., t f ,n and the stress level
T
sl sl1 , sl2 ,..., sln . Assuming that (at least) locally the relationship between t f and sl can
be described with the models presented above, the parameter assessment may be performed
by introducing an error term H which takes account for the difference between observed time
to failure t f and modelled time to failure t f ,m :

tf t f ,m sl,ș ,k R  H (4.70)

Assuming that the error term is normally distributed with zero mean and unknown standard
deviation, V H , the maximum likelihood method, see e.g. Lindley (1965), may be used for
estimating the mean values and covariance matrix for the parameters ș and V H .
The likelihood is then given as

n
1 § 1 § t f ,i  t f ,m ,i ( sli , ș, k R ) · 2 ·
L T1 ,T 2 ,...,T n , V H – 2SV H
exp ¨  ¨
¨ 2© VH
¸ ¸¸ (4.71)
i 1
© ¹ ¹

The parameters are estimated by the solution of the optimization problem

max L(p) (4.72)

p

T
where p T1 ,T 2 ,...,T n , V H .
The parameters of the damage models are calibrated to results from long-term tests
(Hoffmeyer (1990)):
The considered tests are performed on structural timber in four point bending. Two different
test climate conditions are considered: constant climate corresponding to 20% and 11%
moisture content of the timber. From the 20% moisture content timber, 306 specimens of
graded Norway spruce are tested. One third is tested in standard short-term test with ramp
load until failure (EN 408). The remaining specimens are tested with constant load
corresponding to the 5 percentile of the short term strength for one half the remaining sample
and the 15 percentile for the other half. From the 11% moisture content timber the approach is
similar ( n 204 ) but for long term loading solely the 5th percentile of the short term strength
is used as a load level. The matching of the short-term capacity for the long-term test is based
on the assumption of equal rank (see section 4.3.1).

101
For the Gerhards damage model and the Foschi and Yao damage model the ramp load rate k R ,
is considered and is set to 500 MPa/hour; for the Nielsen model the initial ramp load phase is
neglected. The threshold ratio K D in the Foschi and Yao model is set to a constant value of
0.5. For the Nielsen damage model the parameters bc and fl are held constant with values 0.2
and 0.25, respectively.

Table 4-14 Model parameters for the Foschi and Yao damage model, the Gerhards model and the
Nielsen model calibrated on results form DOL tests (Hoffmeyer, 1990) by the ML-
method. The model parameters are represented as normal distributed random variables.
Average Timber Moisture Content: Average Timber Moisture Content:
11% 20%

Model Standard
Expected value Expected value Standard deviation
parameter deviation

bD : BD 30.03 0.92 20.16 0.610

Foschi and Yao

cD : C D 17.05 12.30 12.06 7.29

d D : DD 5.69 0.45 4.37 0.31

V eps : 6eps 0.30 0.03 0.37 0.02

bD : BD 51.41 0.01 43.35 0.004

Gerharts

V eps : 6eps 0.55 0.05 0.476 0.03

W c : 7c 875 113 75.74 5.41

Nielsen

V eps : 6eps 0.51 0.04 0.38 0.02

The correlations for the estimated parameters of the Foschi and Yao model can be given as:

Table 4-15 Correlations for the parameters of the Foschi and Yao model.
Average Timber Moisture Content: Average Timber Moisture Content:
11% 20%

BD CD DD 6 eps BD CD DD 6 eps

BD 1 0.41 0.32 0 BD 1 0.62 0.49 0

CD 1 0.98 0 CD 1 0.97 0

DD 1 0 DD 1 0

6 eps 1 6 eps 1

102
For the Nielsen and the Gerhards model as for the Foschi and Yao model (see Table 4-15)
6 eps is not correlated with the model parameters.
The test results corresponding to a timber moisture content of 20% are plotted together with
the expected realizations of the calibrated models in Figure 4-25. The stress level sl is plotted
together with the logarithm of the time until failure in hours log10 t f . It can be observed that
especially for the centre part, for time to failure between 1 hour and 1 year the differences
between the models are quite minor.

sl >% @

100

90

80

70
Data 5%
Data15%
60 Nielsen
Gerhards
50 Foschi

40
-2 -1 0 1 2 3 4 5 log10 t f

Figure 4-25 Duration of load data (Hoffmeyer, 1990) and damage models based on the mean value
of the estimated parameters (time in hours).

4.3.3.4 Modelling Stress Histories

The DOL models presented above can be used for time variant stress histories. Therefore the
stress histories have to be separated into incremental periods of constant load. Damage is
accumulated according to the following scheme:

D D ,i D D ,i 1  'D D ,i (4.73)

where 'D D ,i can be calculated as:

'D D ,i hmodel,i sli 'ti (4.74)

where h . is a function of the assumed constant stress level sl . For the Gerhards model and
the Foschi and Yao model h . can be written as (based on Equations (4.48) and (4.59)):

103
hGerhards ,i sli exp aD  bD sli (4.75)

bD dD
hF Y ,i sli aD sli  K D  cD sli  K D D D ,i 1 for sli  K D ! 0
(4.76)
hF Y ,i sli 0 for sli  K D d 0

The time increments 'ti have to be chosen small enough to map the time-variant load process
and the non-linear effect of the Foschi and Yao model. The starting condition of the
calculation is D D 0 for an undamaged material. The failure condition is D D ,cr 1 .
For using the general Nielsen model, i.e. modelling damage as a consequence from both, time
under load and number of load cycles (with a scheme as illustrated in Figure 4-24), the
random load sequence has to be transformed into a square wave shaped load process. A
harmonic square wave shaped load process is illustrated in Figure 4-23. As shown in Köhler
and Faber (2003) the parameters from the harmonic square wave process can be modified
after every load cycle as illustrated in Figure 4-24, i.e. random load variations can be
reasonably idealized. Superior stress cycles can be modelled by standard rain flow counting
algorithm, see e.g. in Downing and Socie (1982).
For moderately fluctuating load processes a similar damage accumulation scheme as
presented above in Equation (4.74) can be applied by using the simple Nielsen formulation for
the case of constant load, h . can be written as (compare with Equation (4.64)):

2
S fl D D sli2
hNielsen ,i sli
8qcW c 1
1
bc
(4.77)
D D sli2 1

where the starting condition of the calculation is here D D 1 , failure condition is D D ,cr sl 2 .
As shown in Köhler and Faber (2003) for some load cases it is a reasonable simplification of
the Nielsen model to consider only the time under load as the cause of damage accumulation
instead of applying the model in it’s fully developed general format as illustrated in Figure
4-24. According to the Nielsen model fatigue damage is due to fast unloading and it can be
shown that this damage can be neglected if the unloading is slow enough (order of magnitude
6 hours) or seldom enough (the order of magnitude is around 60 unloadings per lifetime).
The limit state function as given for the damage accumulation models in Equation (4.47) is
not valid for the Nielsen model. For the Nielsen model the limit state function becomes:

D D ,cr  D D S t , R0 , ș d 0 (4.78)

where D D ,cr SL2 .

104
4.3.4 APPLICATIONS FOR DOL MODELS

To gain a better understanding about the predictive behaviour of the three DOL models
introduced above some example calculations are performed within this section. The models
are used with the parameters as presented in Table 4-14.

4.3.4.1 Lifetime Prediction and Damage Accumulation for Constant Load

When comparing the DOL-model estimations of the time to failure for a constant load only
minor differences can be found, as it can be seen in Figure 4-25 for e.g. a stress level sl 0.7 .
When investigating the damage accumulation for a case with constant load, however, the
models show very different behaviour. In Figure 4-26 the damage accumulation until failure
for the case of constant load with load ratio sl 0.7 is shown. The ordinate of the diagram in
Figure 4-26 is the difference between damage, D D , and damage at failure, D D ,cr , on the
abscissa is time. The damage accumulation according to the Gerhards model is for this load
case linear with time. The Nielsen model shows for the same case high non-linearity as does
the Foschi and Yao model. The latter is, however, smoother than the first. The model
parameters for constructing Figure 4-26 are taken from Table 4-14, with a moisture content of
the timber equal to 20% (Köhler and Svensson (2002)).

D -Dcr
1

0.8

0.6
Gerhards
0.4
Foschi and Yao
0.2 Nielsen

0
time
40 days

Figure 4-26 Damage accumulation for the case of constant load ratio of 0.7 according to the
Gerhards model, the Foschi and Yao model and the Nielsen model.

4.3.4.2 Reliability based code calibration

In timber design codes, the effect of load duration is taken into account by a strength
modification factor kmod , which depends on the type of load acting on the structure. The loads
are classified in terms of their expected duration time and intensity. The modification factor
kmod is calibrated by using a DOL model in a probabilistic analyses. kmod calibration has been
carried out and reported in e.g. Foschi et al. (1989), Ellingwood et al. (1991), Svensson et al.

105
(1999), Sørensen et al. (2002) and Köhler and Faber (2003). In the following an approach for
determining kmod is presented.

Procedure for calibrating the duration of load modification factor kmod

1) Design Format:
The design codes specify by means of design equations that the design strength of structural
components rd is larger than or equal to the design load effect sd . In modern LRFD code
format, the design values are defined in terms of characteristic values and partial safety
factors J as (compare Equation (2.28)):

r0,k
rd zd t sk ˜ J S sd (4.79)
JM
where the index k denotes characteristic value and zd is the design parameter. Irrespective of
the duration of load, the load description (right hand-side of Equation (4.79)) is based on the
annual maximum load intensity, i.e. the description of the load effect will not change to
account for the load duration. For materials being affected by the duration of the loading such
as timber a modification factor, kmod , is introduced on the short term capacity as:

kmod ˜ r0,k
rd zd (4.80)
JM
The effect of the duration of the load is accounted for entirely by the kmod factor. The partial
safety factor for the timber material resistance J M , is thus independent of the load duration.
2) Limit State Functions:
Two failure modes are considered in the following: A) failure is defined as the load effect
intensity exceeding the short-term strength and B) where failure is defined as the consequence
of damage accumulation during long term load application. Hence, when conducting the
reliability analyses each failure mode is associated with a specific limit state function.
For failure mode A) i.e. the failure mode where load effect intensity, S , is exceeding the short
term strength R0 the short-term limit state function is given as (compare Equation (4.46)):

g X zd R0 X M  S (4.81)

where X M is the model uncertainty and zd is a design variable.

For failure mode B) where failure occurs as a consequence of the accumulated damage
exceeding the critical damage level D D ,cr the long-term limit state function is given as
(compare Equation (4.78)):

106
g X D D ,cr  D D S t , R0 , zd , p (4.82)
T
where D is the damage as a function of the model parameters p T1 ,T 2 ,...,T n , V H , S t is
the load process, and zd is the design variable.
3) Procedure for determining kmod :
By utilizing Monte Carlo simulation for generating random variables the modification factor
kmod is determined according to the following procedure. First the partial safety factor for the
short-term strength J M , s is calibrated to a target reliability index, E target , using the limit state
function given in Equation (4.81) and the design Equation (4.79) for constant J S . The
probabilistic model for the load intensity is based on the maximum load for a reference period
of 50 years. The partial safety factor for the long-term strength J M ,l is calibrated using the
limit state function in Equation (4.82), and the design equation given in Equation (4.79) for
constant J S . The realization of the load process S t entering the limit state function given in
Equation (4.82), is generated based on some specified load model. For the generated load
process realization, the damage models calculate the accumulation of damage. After a
sufficient number of simulations, the ratio of the number of failures n f and the total number
of realizations n estimates the probability of failure p f n f n . The corresponding reliability
index E is determined from E ) 1 p f . The long-term partial safety factor, J M ,l , is re-
calculated until the target reliability index, E target , is reached. kmod is then determined as:

J M , s E target
kmod (4.83)
J M ,l E target

The calculations may be based on several different load models. The short-term strength R0 is
modeled by a log-normally distributed random variable with expected value equal to one and
coefficient of variation equal to 25%. Referring to the Joint Committee on Structural Safety
(JCSS, 2001) and Faber and Sørensen (2003) a target reliability index E target 3.2 for a
reference period of 50 years is used. When not specifically specified, the model uncertainty in
the short-term limit state function (Equation (4.81)) is assumed log-normal distributed with
mean value equal to 1 and standard deviation equal to 0.05.

4.3.4.3 Calibration Example – Live Load

Live Load Model

The live load model used in this example is equivalent to the model proposed in (CIB 1989)
and recommended by the Joint Committee on Structural Safety (JCSS (2001)). The model
contains two parts, sustained live load and intermittent live load. The sustained live load
covers ordinary live load such as furniture, average utilization by persons, etc. The
intermittent live load describes the exceptional load peaks, e.g. furniture assembly while re-

107
modelling, people gathering for special occasions, etc. The live load model is based on the
following definitions:
The occurrences of changes in magnitude of the sustained load are modelled as a Poisson
process. The duration of load is then exponential distributed with expected value Osus . The
magnitude of the sustained load is modelled by a Gamma distribution with expected value
P sus and standard deviation V sus :

2 2 aare ,0
V sus V sus , r  V sus , sp NP (4.84)
aare

where aare ,0 is the reference area, aare is the considered area ( aare ,0  aare ), N P is the peak
2 2
factor, V sus ,J specifies the variation between structures, V sus , sp is the small scale variation
related to the area aare ,0 .
The occurrences of intermittent loads are also modeled as a Poisson process. The duration
between intermittent loads are thus exponentially distributed with expected value Q int . The
magnitude of the intermittent load is modeled by a Gamma distribution with expected value
Pint and standard deviation V int :

2 aare ,0
V int V int, sp NP (4.85)
aare
2
where V int, sp is the small scale variation related to the area aare ,0 .

The duration of the intermittent load tint is considered exponentially distributed with
parameter Oint . The parameters used for the live load model are given in Table 4-16.

Table 4-16: Parameters used for model live load in office space.

Type of
aare ,0 P sus V sus ,J V sus , sp 1 Osus Pint V int, sp Q int 1 Oint
building ª¬ m 2 º¼
ª kN º ª kN º ª kN º ª kN º ª kN º
«¬ m 2 »¼ «¬ m 2 »¼ «¬ m 2 »¼ > year @ «¬ m 2 »¼ «¬ m 2 »¼ > year @ >days @

Office 2 0.5 0.3 0.6 5.0 0.2 0.4 0.3 2.0

In this example, a floor structure based on joists with a span of 4.5 m and a spacing of 0.6 m
is considered. The area aare is assumed to cover the span and twice the spacing of the joists
aare 5 m 2 . The peak factor, N P is calculated for load influence on the joist for maximum
mid-span bending moment, therefore N P 1.778 (CIB 1989).
By using Monte Carlo simulation and the load model parameters from Table 4-16 the mean
value and the coefficient of variation of the annual maximum load SQ ,1 and the 50 year
maximum load SQ ,50 can be derived. In Table 4-17 the results of the simulation are given. A
typical realization of a 50 years live load history is shown in Figure 4-27 (Köhler and

108
Svensson (2002)).

Table 4-17: Results from simulations of live load in office space (non-parametric).
Annual max SQ ,1 50 year max SQ ,50

mean COV 98% quantile mean COV

2 2 2
ª¬ kN m º¼ [-] ª¬ kN m º¼ ª¬ kN m º¼ [-]

0.96 0.78 3.10 3.05 0.29

Load
3

0
0 10 20 30 40 Time [years]

Figure 4-27: A typical live load realization acting as midspan bending moment on a joist beam in
an office space.

Result
The three DOL models discussed in section 4.3.3.2 are utilized for comparison. The
parameters for timber with 20% moisture content as specified in Table 4-14 are used. The
Nielsen model is applied with its simple format as given in Equation (4.64), i.e. only time
under load is considered. The results from the calibration of kmod factor for the case with live
load in office space are shown in Table 4-18. The results show that the damage models give
almost the same results (Köhler and Svensson (2002)).

Table 4-18: Calibrated values of kmod for the different damage models.

Foschi and Yao Nielsen Gerhards

0.77 0.76 0.75

4.3.4.4 Calibration Example – Snow Load

Snow Load Model

Snow on roofs of buildings is considered and a stochastic load model is formulated in
accordance with the load model presented in Sørensen et al. (2002). Snow load on the ground

109
is modeled by triangularly shaped load pulses (snow packs). This is illustrated in Figure 4-28.

PSG

Pm ,1
Pm ,2

X1 X 1  T1 X2 X 2  T2 t

Figure 4-28: Snowload at the ground modeled by triangular pulses.

Based on Danish snow data over 32 years from five locations the following load process
parameters have been estimated (see in Sørensen et al. (2002)). The occurrence of a snow
package at times X 1 , X 2 ,... is modelled by a Poisson-process. The duration between snow
packages is therefore exponential distributed with expected value 1 O , where O is the
expected number of annually snow packages. The magnitude of each snow package Pm is
assumed to be Gumbel distributed. The duration of a snow package TSP is assumed to be
related to the snow package magnitude with TSP X T Pm where the factor X T is exponential
distributed. The load history on the ground P t can be simulated and transformed to the roof
load Q t by introducing a Gumbel distributed shape factor C with Q t CP t . The
parameters for the stochastic models are summarized in Table 4-19.

110
Table 4-19: Overview of the stochastic Models used in this example.
Distribution Mean Standard
Stochastic Variable
Type Value Deviation

Short-term strength R0 Log-normal 1 0.2

Short term LS

Snow load (50 y max) QS Gumbel max 1 0.21

Model uncertainty XM Log-normal 1 0.05

Duration between snow pack [y] 1O Exponential 1/1.175 -

Load model

Max. per snow pack [kN/m2] Pm Gumbel max 0.33 0.21

Duration pack factor [d/kN/m2] XT Exponential 75 -

Roof shape factor [-] C Gumbel max 1 0.35

The long term simulations are performed by using the simple Nielsen model with the
parameters given in Table 4-14, timber with 11% moisture content.

Result
The modification factor kmod is calibrated as previously described. A value of kmod 0.8 is
found. This result is consistent with the results presented in Sørensen et al. (2002), where the
same snow load model but different damage accumulation models are used in the calibration
procedure. To illustrate the effect of damage DN to the strength of the timber material, the
residual strength ratio rr t r0 can be evaluated according to Equation (4.68). In Figure 4-29,
one realisation of the load process and the corresponding residual strength ratio is given.
Failure occurs when the residual strength ratio drops below the load effect as consequence of
an extreme snow load, i.e. when rr t r0  s t r0 (Köhler (2002)). In the case illustrated in
Figure 4-29 failure occurs after a time of approximately 32 years. The load history prior to
failure seems not to be of significance to the damage and the residual strength respectively. A
similar observation is made in Rosowsky and Bulleit (2002) where other damage accumu-
lation models are investigated.

111
 1.0
0.8
rr t r0 and sl

0.6

0.4
0.2

0
0 5 10 15 20 25 30 time [years]

Figure 4-29 Maximum magnitude of each snow package and the residual strength of the timber
specimen is plotted over lifetime.

In Figure 4-30, a snow load process is used where the standard deviation of the maximum
load level of each snow package is lower compared to the example shown in Figure 4-29,
V Pm 0.05 compared to V Pm 0.21 . It is seen that the residual strength drops as a
consequence of major snow loads several times until failure occurs. The damage accumulates
with the number of major snow events.

1.0
rr t r0 and sl

0.8

0.6

0.4
0.2

0
0 5 10 15 20 25 time [years]

Figure 4-30 Maximum magnitude of each snow package with lower standard deviation and the
residual strength of the timber specimen is plotted over lifetime.

Depending on the load level as well as the temporal characteristics of the underlying random
load process failure may be considered as time independent. Rather than simulating complete
load histories and incrementally accumulating damage, a time independent limit state function
can be formulated. Only the load pulse that cause significant damage, may be considered for
reliability analysis with the considered snow load characteristics. If the standard deviation of
the maximum load level of a snow package is getting smaller, damage accumulates several
times (Figure 4-30) and the time to failure has to be evaluated by simulation as described in
this paper. The DVM-model does not include a threshold value K D (compare the Foschi and
Yao model, Equation (4.59)) below which no damage accumulates as some other damage
accumulation models. However, through the highly non-linear properties of the model an
effective threshold can be observed by analyzing the results of the simulations.

112
4.3.4.5 The Nielsen Model and Fatigue - Parametric Variable Load.
For timber structures as for structures made of any other material, in principle all types of
loads must be anticipated. For the load types considered above, live load in buildings and
snowload, it seems to be a reasonable assumption to regard damage solely a consequence of
accumulated time under load. For some types of loads such as traffic load on bridges, wind
loads on roofs and facades or the dynamic load on a rotor blade of a wind turbine it is
intuitively understood that beside damage accumulation as a consequence of accumulated
time under load, a significant damage contribution is due to the number of applied load cycles
of given amplitude, i.e. due to fatigue effects. In Bonfield and Ansell (1991), the DOL-effect
in wind turbine rotor blades is investigated considering timber as an elastic material, i.e. no
damage caused by creep is considered.
In the calibration examples above (sections 4.3.4.3 and 4.3.4.4) and e.g. in Sørensen et. al.
(2003), damage accumulation laws accounting for the visco-elastic damage accumulation are
calibrated to long term test with structural timber specimen subjected to constant intensity
loads. In the assessment of the kmod factors for different load scenarios thus only the
accumulated duration of load is taken into account, and fatigue effects disregarded.
In the following, two damage accumulation models are compared. The general Nielsen model
for which the computational algorithm is illustrated in Figure 4-24 is considered. As
previously mentioned this model is able to predict the lifetime of timber components subject
to random time variant loads considering both creep and fatigue effects. Also the simple
Nielsen model given in Equation (4.64) is considered. This model which also rests on the
DVM-theory is compared with other damage accumulation models in section 4.3.4.3, where it
is applied specifically for the purpose of calibration of kmod .
Both models are used for the estimation of the time to failure of one timber specimen with the
short term resistance r0 . A rectangular pulse process with constant amplitude and constant
frequency as defined in Figure 4-23 is applied. The following load parameters are used:
psl 0.5 , E sl 0.5 , slmax 0.6 , f 0.001 Hz ( T 100s). In Figure 4-31 the residual
strength ratio rr r0 defined as rr r0 DN0.5 and the maximum and minimum load intensities,
slmax and slmin over the logarithm of time in hours are illustrated. A significant difference is
observed in the predicted lifetimes, i.e. 14 and 60 years using the general and the simple
model, respectively see also Figure 4-31.
It can be shown that a timber specimen is able to survive a reference lifetime period of 50
years applying a load as described above for slmax  0.606 for the simple model and
slmax  0.548 for the general model.
In the following the general model is investigated in more detail. It is first examined whether
the general model deviates significantly from the currently applied simpler damage
accumulation model (without consideration of fatigue effect) for different types of random
load processes.

113
A simple load process model is applied in the following, see Figure 4-32, where the
rectangular load pulses are random in both their intensity and duration. The occurrence of
load events at times X 1 , X 2 ,... is modeled with the constant rate OS , i.e. the constant number
of load events per year. The intensity of the load at a given load event Pm is assumed to be
Gumbel distributed. The duration of a load event TLE is assumed to be related to the load
event intensity with TLE X T Pm , where the factor X T is assumed exponential distributed. The
load model is illustrated in Figure 4-32.

1.0

0.8
rr t r0 and sl

sl max
0.6

0.4 sl min

0.2 Simple
General
0
2 3 4 5 6 log10 t , >hours @

Figure 4-31 Predicted lifetime in accordance with using the simple model and the general model
respectively.

To investigate the effect of the characteristics of the random load process on the model
predictions, the variation of the intensity Pm and the number of load events per year OS are
varied.

114
Pm t

Pm ,0
Pm ,1

T0 X1 X 1  T1 t

Figure 4-32 Load model applied in the example.

The parameters for the probabilistic models for the load, the long term strength and the short
term strength are given in Table 4-20.
According to the presented procedure the modification factor kmod is calibrated for the
following cases. OS = 1, 4, 8, 16 (in 1/years) and V Pm 0.64 according to Table 4-20;
different numbers of load events per year are applied. The duration of the load events is
decreasing by increasing number of load events per year so that the expected value of the time
under load is the same for the considered load cases. In Figure 4-33 the results as obtained by
the general model are illustrated and compared with the results obtained by using the simple
model, i.e. the model in which fatigue effects are disregarded.

115
Table 4-20 Overview of the probabilistic model used in the example.
Distribution Mean Standard
Stochastic Variable
Type Value Deviation

Short-term strength R0 Log-normal 1 0.2

Short term LS

Load, 50 years max QS Gumbel max f V Pm , C 0.21

Model uncertainty XM Log-normal 1 0.05

Load events per year OS Exponential C -

Intensity of the load event

Load model

[kN/m2]
Pm Gumbel max 1 VP m

Load duration factor [d/kN/m2] XT Exponential 75/C -

1.065
1
0.925
0.9
0.815 Simple
Simple
kmod

0.8
General
General
0.74
0.7

0.6

0.5
Os = 1 Os = 4 Os = 8 Os = 16
Load Case

Figure 4-31 Calibrated kmod factors for different load frequencies.

From Figure 4-31 it can be observed that the estimated values for the modification factor are
virtually identical and independent on the choice of the damage accumulation model.
In Figure 4-32 the calibrated kmod factors for different load cases are illustrated as a function
of the coefficient of variation of the load intensity. The number of load events per year is
constant with OS = 16, the expected value of the load duration factor is 4.7 d/kN/m2 according
to Table 4-20. The effect of varying the coefficient of variation in the range between 0.64 and
0.00001 is investigated, see Figure 4-32.

116
 1.2
1.065 1.04
1 1.065 1.03 0.9
0.8 0.87
0.622
kmod

General
0.6
0.588 Simple
0.4
0.2
0
0.64 0.1 0.001 0.00001
coefficient of variation

Figure 4-34 Calibrated kmod factors as function of the coefficient of variation of the intensity of
load events.

From Figure 4-34 it is seen that the smaller the coefficient of variation of the intensity of the
load events, the larger the difference achieved by using the two different models.

4.3.5 SUMMARY AND CONCLUDING REMARKS, DOL EFFECTS

In section 4.3 an approach for the probabilistic modelling of the effect of load duration is
presented. The method is exemplified for calibrating the design code short term strength
modification factor kmod for different characteristics of the applied loading process. The
considered damage models are introduced. The Nielsen model takes basis in fracture
mechanical considerations and facilitates the consideration both creep and fatigue effects. The
two other models, the Gerhards model and the Foschi and Yao model take solely creep effects
into account. The damage models are calibrated against duration of load tests using the
maximum likelihood method. Furthermore, the damage models have been investigated and
compared by applying several different load scenarios.
In section 4.3.4.3 the so called simple Nielsen model is compared with the other two models
by calibrating kmod for a live load scenario, using the three different DOL models. It is shown
that the result is not sensitive to the particular DOL model which is used.
In section 4.3.4.4 the modification factor kmod is also calibrated for a snow load scenario, this
time only the simple Nielsen model is used. The result is in line with results for snow load
calibration results found in the literature, where the Gerhards and the Foschi and Yao model is
used. From section 4.3.4.3 and 4.3.4.4 it can be concluded, that the simple Nielsen model
delivers similar results compared to other damage accumulation models.
In section 4.3.4.4 it is further observed that for a typical snowload realisation the damage
accumulation process is highly non-linear, i.e. the undamaged state is directly leading into the
full damage (failure) state (Figure 4-29). In Figure 4-30 it is illustrated that this is not the case
when the variation of the amplitudes of each load cycle is decreased. Then, damage

117
accumulates as a consequence of several major load events.
In section 4.3.4.5 the general Nielsen model is compared with the simple Nielsen model. A
significant difference in predicted lifetimes using the two different models can be observed in
Figure 4-31. Under the given load conditions, a square wave process with constant load
intensities, lifetime predictions obtained using the simple model are more than four times
longer than those achieved using the general model. From this observation it can be concluded
that, for the given harmonic square wave load process a significant fatigue effect might exist
following the fracture mechanical modelling proposed by Nielsen (2000). This conclusion is
also consistent with experimental evidence e.g. in Clorius (2001).
By applying random load processes with rectangular load pulses and random intensity and
duration no significant fatigue effect is observed between the models in the frequency range
considered. The estimated load duration modification factors kmod are almost the same for
both using the general and the simple models for the calibration.
By decreasing the coefficient of variation of the intensity of the load pulses Pm to nearly 0,
i.e. Pm is nearly deterministic, a difference of the value of kmod of 6% can be observed in the
considered example.
The considered frequency range between 1 and 16 load cycles per year is, however, not
representative for load scenarios likely to induce fatigue damages. In the present example this
limited frequency interval is due only to the numerically cumbersome and time demanding
calculations. It still remains to investigate higher frequency loads.

4.4 MOISTURE EFFECTS ON THE DURATION OF LOAD EFFECT

The DOL effect is dependent on the moisture content ( mc ) of the timber, i.e. the higher the
moisture content, the shorter the time to failure. The DOL effect is also subject to variations in
moisture content. Such variations increase creep significantly and shorten time to creep
rupture (see e.g. Toratti (1992)). This effect is in general referred to as the mechanosorptive
effect.
In line with the presented damage accumulation models for the duration of load effect,
proposals for modelling the combined (damage) effect of moisture and duration of load exist.
E.g. the models proposed in Fridley (1992) or in Toratti (1992) consider the moisture effect as
an extra term in damage accumulation formulations similar to Equation (4.43). The
parameters of these models are calibrated to rather limited data from experiments, i.e. can not
be verified by a reasonable set of climate/load scenarios (Morlier (1994)). However, the
combination of moisture and duration of load effects in a damage accumulation model still
seems promising; current research projects focus into that direction (Sørensen and Svensson
(2005), Nielsen (2005)).
In practical design, as in the Eurocode 5, the effect of moisture on the duration of load effect

118
is considered with the modification factor kmod which is given for different climate exposures
in design codes. Values for this factor are prescribed in a matrix for three different so-called
service classes, i.e. different climate scenarios, and five different load classes, i.e. load
scenarios. With a similar framework the stiffness degradation aspect is accounted for with the
factor kdef , which is also prescribed in a matrix for different service and load classes. The
different service and load classes according to the EC 5 are given in Table 4-21. The service
classes sc are defined in regard to the average moisture content mc of the timber.

Table 4-21 Load and service classes according to the EC 5.

mc Permanent Long term Medium term Short term
sc Instantaneous
[%] ( t ! 10 years) ( 0.5  t  10 years) ( 0.25  t  6 month) ( t  1 week)

1 <12

2 <20

3 >20

4.5 INTERRELATION OF MATERIAL PROPERTIES

As already discussed at the beginning of this chapter, other material properties are estimated
based on the information about the reference material properties. In Equations (3.10)-(3.17)
relationships based on experiments on European softwoods are given. The estimates derived
with these equations are associated with uncertainties which are not quantified in the
literature. Furthermore, only characteristic values or mean values are estimated with these
formulas.
To get estimates for the probability distribution functions of other material properties the
above equations are utilized as a basis. Based on several discussions within the COST action
E 24 (2005) it is decided, as a simplification, to take the same interrelations as given for
strength characteristic values also for the expected values for the strength related material
properties. Furthermore, coefficients of variation of the other material properties and a
correlation matrix is identified, both based on judgment and a result of several discussions
within the COST action E24 (2005).
The bending moment capacity and the bending modulus of elasticity are assumed to be
lognormal distributed. The density is assumed to be normal distributed.

119
Table 4-22 Probabilistic models of other timber material properties. (quantities are tentative and
basis for further discussion)
Coefficient of variation
Property Distribution Expected Values E > X @
COV > X @

Tension strength
parallel to the lognormal E ª¬ Rt ,0 º¼ 0.6 E ª¬ Rm º¼ COV ª¬ Rt ,0 º¼ 1.2 COV ª¬ Rm º¼
grain:

Tension strength E ª¬ Rt ,90 º¼ 0.015E > 5 den @ COV ª¬ Rt ,90 º¼ 2.5COV > 5 den @
weibull
perp. to the grain:

MOE - tension COV ª¬ MOEt ,0 º¼ COV ª¬ MOEm º¼

parallel to the lognormal E ª¬ MOEt ,0 º¼ E ª¬ MOEm º¼
grain:

MOE - tension E ª¬ MOEm º¼ COV ª¬ MOEt ,90 º¼ COV ª¬ MOEm º¼

lognormal E ª¬ MOEt ,90 º¼
perp. to the grain: 30

Compression
0,45
strength parallel to lognormal E ª¬ Rc ,0 º¼ 5 E ª¬ Rm º¼ COV ª¬ Rc º¼ 0.8 COV ª¬ Rm º¼
the grain:

Compression
strength per. to the lognormal E ª¬ Rc ,90 º¼ 0.008 E > 5 den @ COV ª¬ Rc ,90 º¼ COV > 5 den @
grain:
E > MOGm @ COV > MOGv @ COV > MOEm @
Shear modulus: lognormal E > MOEv @
16
0.8
Shear strength: lognormal E ª¬ Rv º¼ 0.2 E ª¬ Rm º¼ COV ª¬ Rv º¼ COV ª¬ Rm º¼

120
Table 4-23 Correlation coefficient matrix. The values in this table are quantified by judgment; 0.8
ļ high correlation, 0.6 ļ medium correlation, 0.4 ļ low correlation, 0.2 ļ very low
correlation.
MOEm 5 den Rt ,0 Rt ,90 MOEt ,0 MOEt ,90 Rc ,0 Rc ,90 MOGv Rv

rm
0.8 0.6 0.8 0.4 0.6 0.6 0.8 0.6 0.4 0.4

MOEm
0.6 0.6 0.4 0.8 0.4 0.6 0.4 0.6 0.4
5 den
0.4 0.4 0.6 0.6 0.8 0.8 0.6 0.6
Rt ,0
0.2 0.8 0.2 0.5 0.4 0.4 0.6
Rt ,90
0.4 0.4 0.2 0.4 0.4 0.6
MOEt ,0
0.4 0.4 0.4 0.6 0.4
MOEt ,90
0.6 0.2 0.6 0.6
Rc ,0
0.6 0.4 0.4
Rc ,90
0.4 0.4
MOGv
0.6

It should be underlined that the information given in Table 4-22 and Table 4-23 is rather
vague and associated with uncertainties. However, if no more information is available it can
be taken as a reference. The information can be also utilized to quantify a prior distribution
function as a basis for probability updating if more information becomes available (see Annex
A).

121
122
5 PROBABILISTIC MODELLING OF THE PROPERTIES OF
TIMBER CONNECTIONS

For timber structures, the structural performance depends to a considerable part on the
connections between different timber structural members; connections can govern the overall
strength, serviceability and fire resistance. Assessments of timber structures damaged after
extreme events as storms and earthquakes often point to inadequate connections as the
primary cause of damage (Foliente (1998)). Despite their importance timber connection
design frameworks are not based on a consistent basis compared to the design regulations of
timber structural components.
Explanations for this difference in progress of design provisions for members and connections
can be found in the relative simplicity of characterising mechanical behaviour of members, as
compared to connections. A diversity of connections types is used in practice and these types
have infinite variety in arrangement. This usually precludes the option of testing large
numbers of replicas for a reliable quantification and verification of statistical and mechanical
models.
For commonly used connections, a distinction is made between carpentry joints and
mechanical joints that can be made from several types of fasteners. An overview of timber
connections can be found in the literature, e.g. in Thelandersson and Larsen (2003). The
mechanical joints are divided into two groups depending on how they transfer the forces
between the connected members. The main group corresponds to the joints with dowel type
fasteners. Connections with dowels, nails, screws and staples belong to this group. The second
type includes connections with fasteners such as split-rings, shear-plates and punched metal
plates in which the load transmission is primarily achieved by a large bearing area at the
surface of the members. In addition to connections with mechanical fasteners, also glued
joints should be mentioned. This technique is mainly carried out using glued-in bolts for beam
connections or large finger joints for frame corners, (Blass et al. (1995)).
In this chapter it is focused on connections with dowel type fasteners. On the example of
timber to timber dowel type fasteners it is demonstrated how a probabilistic framework can be
introduced. Therefore it is first focused on the physical modelling of this type of connections,
while single fastener connections are at first considered. Beside a common physical model
framework – the so called Johansen model – a model refinement considering timber splitting
based on the work presented in Jorissen (1998) is introduced and discussed. The ingredients
of the physical models, the material properties and their probabilistic modelling, are
discussed. Different implementations of the considered models in present design formats are
illustrated before it is examined how the effective number of fasteners for multiple fastener
joints is commonly evaluated.
A large data base of load carrying capacity data of dowel type fastener connections described
in Jorissen (1998) is utilized to illustrate the possible model verifications implied by

123
comparing model predictions with observations from the load bearing capacity tests. In the
same line model uncertainties are assessed based on the similar database.
Different formats for possible probabilistic models are presented and the sensitivity of the
model to the input parameters is assessed. The chapter is concluded with a discussion of the
findings, the shortcomings in the presented approach and a perspective for further research is
given.

5.1 JOINTS WITH DOWEL TYPE FASTENERS

Joints with dowel type fasteners are the most common joints in timber structures. Dowel type
fasteners include bolts, dowels, nails and staples (see Figure 5-1). The main characteristic of
this type of connections is that the fasteners are mainly laterally loaded.

a) b) c) d)

Figure 5-1 Different dowel type fasteners: a) Bolt, b) Dowel, c) Nail, d) Staple.

Laterally loaded joints with dowel type fasteners are illustrated in Figure 5-2. The load can be
transferred through one or more shear planes per fastener.
By using dowel type fastener joints the load can be transferred in pure tension or
compression, but also at an angle in e.g. truss joints. In Figure 5-3 two examples for timber to
timber double shear fastener connections are given; a tension or compression joint (left) and a
typical truss joint (right). Three timber components are connected, two so-called side
members with the side member thickness d s and one so-called middle member with thickness
d m . The fastener diameter is specified by d‡ and furthermore by the so-called end- and in-
between distance for the side members, a3 and a1 , are given in the left part of Figure 5-3.

124
Figure 5-2 Laterally loaded timber to timber joints with dowel type fasteners. Dowels in double
shear (i.e. two shear planes per dowel (left)). Dowels in single shear (middle and
right).

The load carrying capacity of dowel type fasteners is governed by four main characteristics:
x The embedding strength of the timber. The embedding strength is the property of a
timber solid to resist the lateral penetration of a stiff fastener.
x The bending moment capacity of the dowel. The bending moment capacity is mainly
influenced by the dowel diameter and the yield strength of the dowel material. A plastic
deformation capacity is necessary to provide bending moment capacity even after
considerable deformation of the dowel.
x The pulling out resistance of the dowel. Under special circumstances the so called pulling
out resistance of dowel type fasteners can be mobilised even in lateral loading. A large
bending deformation of the fastener is required. This effect is also referred to as the rope
effect.
x The resistance against splitting. This resistance is mainly governed by a fracture
mechanical phenomena.

Joints with dowel type fasteners usually contain more than one fastener. Modelling of the load
bearing capacity of multiple fastener connections, however, is always based on the
calculations made considering one fastener. This might be for traditional and practical
reasons; since the physical behaviour of single fastener connections is rather complex, the
behaviour is even more complicated for multiple fastener connections, not least due to the
multitudinous configurations which could be considered.

125
 d ds d ds
dm dm

ds ds
a3 a1

h h

Figure 5-3 Tension or compression joint (left) and a typical truss joint (right), where the load is
transferred under the angle D .

5.1.1 SINGLE FASTENER CONNECTIONS – JOHANSEN’S YIELD THEORY

The load bearing capacity of laterally loaded single dowel type fastener connections has been
studied for several decades. In most timber design codes the so-called Johansen’s Yield Model
has been implemented for this type of connection. The embedding strength of timber and the
bending capacity of the fastener are considered in this model. However, the model in its very
first formulation (Johansen (1949)) is subject to rather strict assumptions. These are:
x The embedding behaviour of timber is idealised to be an ideal rigid-plastic type.
x All timber members of the connection have identical embedding strength properties.
x For single shear connections the involved members have the same thickness; for double
shear connections the assembly is symmetric.
x Only the elastic bending capacity of the fastener is considered.
In Meyer (1957) these restrictions have been released; different material properties and
member thickness are considered and the plastic bending capacity for the dowel type fastener
can be taken into account. However, the modelling of double shear connections is restricted to
symmetry in regard to its geometry and material properties. Despite of this historical
background the formulations from Equation (5.1)-(5.4) are generally referred to as the
Johansen Equations.
For double shear connections 4 different failure modes are differentiated; three different for
the side members and one common for the middle member as illustrated in Figure 5-4. Note
that for failure modes II and III two respectively four plastic hinges are developed.

126
 rJoh , I , s rJoh , Ia ,m rJoh , II , s rJoh , III , s

I Ia II III

d ds 1
2 dm ds 1
dm
d 2

2 rJoh , I , s 2 rJoh , Ia ,m 2 rJoh , II , s 2 rJoh , III , s

Figure 5-4 Johansen failure modes for double shear connections.

rJoh , I , s d s d‡ rh, s (5.1)

rJoh , Ia ,m d m d‡ rh , m (5.2)

2
d s d‡ rh , s rh ,m 1 § 2d‡ d s rh , s rh ,m · 4 d s2 d‡ rh , s  4m y d‡ rh , s rh , m
rJoh , II , s   ¨¨ ¸¸  (5.3)
2rh , s  rh ,m 2 © 2rh , s  rh , m ¹ 2rh , s  rh , m

4my d‡ rh , s rh ,m
rJoh , III , s (5.4)
rh , s  rh ,m

The parameters used in Equations (5.1) - (5.4) are:

ds / dm the side/middle member thickness > mm @ ,
d‡ the diameter of the fastener > mm @ ,
rh , s / rh , m the embedding strength of side/middle member > MPa @ ,
my the bending moment capacity of fastener > Nmm @ .
rJoh , I , s , rJoh, II , s , rJoh , III ,s is the load bearing capacity in > N @ of each side member according to
failure modes I , II and III (see Figure 5-4) respectively. rJoh , Ia ,m is the load bearing capacity
in > N @ of the middle member according to failure mode Ia .
In the Johansen equations forces other than the force normal to the axis of the fastener are
neglected. The load bearing capacity per shear plane is determined for any failure mode by
static equilibrium or from the principle of virtual work. The governing failure mode is
determined by the minimum capacity per shear plane, as:

127
 § 1 ·
rJoh min ¨ rJoh , I , s , rJoh , Ia , m , rJoh , II , s , rJoh , III , s ¸ (5.5)
© 2 ¹
As the fastener is assumed to exhibit ideal rigid-plastic behaviour, it can only translate and/or
rotate as a rigid body or segments of it can behave as rigid bodies for failure modes, where
one or more plastic hinges form in the fastener (compare Figure 5-4). The embedding of the
fastener into the timber interface is also assumed to behave ideal rigid-plastic. The
specification of the bending moment capacity of the fastener and the embedding strength of
the fastener plays an important role in the Johansen equations.

5.1.2 SINGLE FASTENER CONNECTIONS – JORISSEN’S SPLITTING MODE

Beside the Johansen failure mechanisms, a failure mechanism associated with timber splitting
is possible (Jorissen (1998)). Whether connections fail in timber splitting or according to one
of the Johansen failure modes mainly depend on the positions of the fasteners; the distances
from the fastener to the end of the timber member and the distance in between fasteners, both
in direction parallel to the grain. In Figure 5-3 the end distance a3 and the in between distance
a1 are specified. In general, design regulations provide guidelines for identifying connection
geometries providing that the Johansen mechanisms are obtained and very brittle failure
modes as timber splitting are avoided. These guidelines are in regard to minimum end
distance and timber thickness both in relation with dowel diameters. However, tests by
Jorissen (1998) show that connections with very rigid dowel type fasteners fail by timber
splitting, even if the members are loaded parallel to the grain and designed according to the
guidelines referred to above. In Jorissen (1998) a model based on fracture mechanical
considerations is proposed to cover the splitting failure mode. According to Jorissen (1998)
the splitting load bearing capacity of a member can be estimated as:

g c ,i MOE0,i d‡ 2h  d‡
rsplitt ,i di md 1
h
(5.6)
g c ,i MOE0,i d‡ h  d ‡
rsplitt ,i 2d i md 2
h
with
g c ,i the mixed mode fracture energy1 > N mm @ ,
d‡ the diameter of the dowel type fastener > mm @ ,
MOE0,i the modulus of elasticity parallel to the grain > MPa @ ,

1
The mixed mode fracture energy is defined in section 5.1.3.3 and specifies a mixture between fracture
opening failure mode and fracture sliding failure mode.

128
h the member width > mm @ ,
di the member thickness > mm @ (see Figure 5-3),
md the number of rows (perpendicular to the grain),
r splitt ,i the load bearing capacity > N @ .
Note that in Equation (5.6) it is differentiated between the case of one single row of fasteners1
( md 1 ) and two parallel rows ( md 2 ).

5.1.3 RELEVANT MATERIAL PROPERTIES

5.1.3.1 Embedding Strength

Several definitions of embedding strength exist for timber materials and when comparing test
results it is of importance to be aware of these definitions. A quite common definition of the
embedding strength is the definition according to the European test standard EN 383, which
reads: the embedding strength is “an average compressive stress at maximum load in a
specimen of timber (…) under the action of a stiff linear fastener”. The test configuration for
parallel and perpendicular to grain tests is given in Figure 5-5. This standard defines the
embedment strength as the highest embedment stress within 5 mm displacement for both
parallel and perpendicular to grain tests. The load displacement diagrams for the embedding
strength parallel and perpendicular to the grain are illustrated in Figure 5-6. For the parallel to
grain test the maximum load is usually reached within 2 or 3 mm displacement and the load-
displacement curves show a typical linear and full plastic branch. For tests perpendicular to
the grain this is not the case. The load is increasing with increasing displacement and the
measured embedding strength strongly depends on at which displacement it is defined. It
should be underlined that in Johansen’s yield model ideal rigid-plastic behaviour is assumed
for the embedding strength. This might be reasonable for the case parallel to the grain, but it is
a crude approximation for the case perpendicular to the grain (compare Figure 5-6).
In practice the embedding strength must be estimated based on indirect information, such as
geometry or material properties which are related to the embedding strength. In most design
codes the embedding strength is estimated based on information about the dowel diameter, the
timber density and the force direction relative to the grain. These estimates are based on the
analysis of experiment data.

1
Rows of fasteners are defined in Figure 5-15.

129
Figure 5-5 Embedment test according to EN 383. (Taken from Sawata and Yasumura (2002)).

Figure 5-6 Load-displacement behaviour parallel (left) and perpendicular (right) to grain test.
(Taken from Sawata and Yasumura (2002))

The embedment strength expressions in EC 5 (EN 1995-1-1:2004) are based on a

comprehensive study by Whale and Smith (1986b) and Ehlbeck and Werner (1992). The
influence of the timber density and the fastener diameter is derived by regression analyses.
The following equations build the basis for the formulation found in the EC 5.

rh ,0 0.082 1  0.01d‡ U den for dowels, coniferous and deciduous (5.7)

rh ,0 0.082 U den d ‡ 0.3 for nails (not pre-drilled), coniferous and deciduous (5.8)

When the embedding strength perpendicular to the grain is estimated according to EC 5 the
following formulations are utilised:

130
 0.082 1  0.01d‡ U den
rh ,90 for dowels, coniferous (5.9)
1.35  0.015d‡

0.082 1  0.01d‡ U den

rh ,90 for dowels, deciduous (5.10)
0.9  0.015d‡

f h ,90 0.082 U den d‡0.3 for nails, coniferous and deciduous (5.11)

with
rh ,0 the embedding strength parallel to the grain [MPa],
rh ,90 the embedding strength perpendicular to the grain [MPa],
U den the timber density [kg/m3],
d‡ the diameter of the fastener [mm].
Equations (5.7) - (5.11) specify an estimate of the embedding strength by given fastener
diameter and density of the timber. In EC 5 the same expressions are used for the estimation
of the 5%-fractile value. Therefore, it is assumed that a sufficiently good estimate for the 5%-
fractile value of the embedding strength can be obtained by exchanging the expected value of
the density by the 5%-fractile value. This is only true if it can be assumed that the uncertainty
of the derived regression models is equal to zero.
In Leijten et al. (2004) embedding strength data is analysed. Data from research projects
associated with the following references is used:
x Whale and Smith (1986b) and Ehlbeck and Werner (1992) (coinciding with data used for
calibration of code formats, e.g. see the EC 5 formulations above),
x Vreeswijk (2003), Sawata and Yasamura (2002), Mischler (not published) (not yet
considered in code drafting).
As in the EC 5 formulation it is found that the embedding strength mainly depends on the
timber density and the diameter of the fastener. Eight cases are differentiated, namely
coniferous and deciduous wood loaded parallel or perpendicular to the grain with pre-drilled
(bolts and dowels) or non pre-drilled (staples and nails) fastener holes. For all cases it is found
that the following model fits the data best:

bh
rh ah U den d‡ch H (5.12)

where,
rh is the embedding strength > MPa @ ,
ah , bh , ch are the model parameters >  @ ,

131
U den is the timber density ª¬ kg m3 º¼ ,
d‡ is the diameter of the fastener > mm @ ,
H is the error term > MPa @ .
Rewriting Equation (5.12) as:

ln rh ah*  bh ln U den  ch ln d‡  ln H with ah* ln ah (5.13)

and assuming that ln H is normal distributed with zero mean and unknown standard
deviation V ln H the parameters ah* , bh , ch and the standard deviation V ln H can be estimated
with the maximum likelihood method, see e.g. Annex A. Therefore, the parameters and the
standard deviation of the logarithm of the error term are normal distributed random variables.
According to Leijten et al. (2004) the parameters of Equation (5.12) are:

Table 5-1 Parameters for the estimation of the embedding strength based on Equation (5.12)
according to Leijten et al. (2004).
Nails Dowels
parallel perpendicular parallel perpendicular
Coniferous Deciduous Coniferous Deciduous Coniferous Deciduous Coniferous Deciduous
n = 397 n = 120 n = 319 n = 80 n = 448 n = 285 n = 506 n = 37

Ah* -4.563 -5.533 -3.086 -7.905 -2.334 -2.441 -2.547 -2.245

Ah* 0.44 0.513 0.391 0.533 0.232 0.296 0.309 0.655
Bh 1.345 1.507 1.148 1.887 1.066 1.091 1.099 1.128
Bh 0.072 0.077 0.064 0.079 0.038 0.044 0.052 0.097
Ch -0.273 -0.181 -0.420 -0.418 -0.253 -0.253 -0.432 -0.455
Ch 0.03 0.037 0.027 0.039 0.012 0.018 0.021 0.038
6 0.175 0.119 0.141 0.101 0.107 0.129 0.129 0.112
6 0.004 0.005 0.004 0.006 0.003 0.004 0.003 0.010
Ah* Bh -0.995 -0.994 -0.995 -0.994 -0.991 -0.986 -0.984 -0.987
Bh Ch 0.085 0.006 0.039 0.004 0.105 -0.115 -0.126 -0.129
Ah*Ch -0.182 -0.115 -0.138 -0.113 -0.235 -0.049 -0.055 -0.030

P and V are the mean values and the standard deviations of the parameters Ah* , Bh , Ch and
6ln H . U˜˜ denotes the correlation between the parameters. With the presented information it is
possible to express the probability distribution function of the embedding strength and to
estimate values of the embedding strength based on this. The probability distribution function
of the embedding strength FRh x can be written as:

FRh x P exp Ah*  Bh ln 5 den  Ch ln d‡  ln H dx (5.14)

where Ah* , Bh , Ch and ln H are introduced as normal distributed random variables.

However, it is important to note that the parameters of the model and consequently all
estimates derived from this model are conditional on the set of data based on which the model
is calibrated. It has to be assumed that the data set considered for deriving the model and its

132
parameters are representative.
Based on the above listed parameters the embedding strength parallel to the grain can be
estimated with given density and dowel diameter and disregarding the uncertainty of the
models as follows:

1 1.345 0.273
rh ,0 U den d‡ for nails, coniferous (5.15)
96

1 1.507 0.181
rh ,0 U den d‡ for nails, deciduous (5.16)
253

1 1.066 0.253
rh ,0 U den d‡ for dowels, coniferous (5.17)
10.3

1 1.091 0.253
rh ,0 U den d‡ for dowels, deciduous (5.18)
11.5

nails_deciduous
dowels_
100
embedding strength [MPa]

deciduous

80
dowels
_coniferous
60 dowels_EC5
nails_EC5
40

20
nails_coniferous
0
0 250 500 750 1000

density [kg/m3]

Figure 5-7 Comparison of the embedding strength parallel to the grain predictions according to
Leijten et al. (2004) and the EC 5.

In Figure 5-7 Equations (5.7) and (5.8) (EC 5) and Equations (5.15) - (5.18) (Leijten et
al.(2004)) are plotted as functions of the density U den and given diameter (nails d‡ 4mm
and dowels d‡ 12mm ) and compared. Except for nails coniferous, the EC 5 formulation
gives lower estimates for the embedding strength.

5.1.3.2 Bending Moment Capacity

The bending moment capacity of a steel dowel type fastener is characterised through its yield
moment. Johansen considered the elastic bending moment capacity m y .el , which can be
derived for circular cross sections as:

133
 S
m y ,el f y d‡3 (5.19)
32
where f y is the yield stress of the material of the fastener.
The plastic bending moment capacity for a circular cross section my , pl is derived as:

1
m y , pl f y d‡3 (5.20)
6
which requires large strains and therefore large bending rotations, of up to 45° in the
fasteners, Blass et al. (2001). For mild steel the yield stress f y in tension can in general be
approximated by f y 0.6 f u , where fu is the ultimate tension stress of the steel material. Due
to strain hardening at large bending rotations, the yield stress in bending is estimated by
f y 0.8 fu . In Figure 5-8 the stress – strain relation of mild steel is illustrated.
stress

f u ,t

f y ,m | 0.8 fu ,t

f y ,t | 0.6 fu ,t

strain

Figure 5-8 Stress – strain relation for mild steel. f y ,t yield strength in tension, f y , m yield
strength in bending, fu ,t ultimate strength in tension.

In Europe the load carrying capacity and the deformation behaviour of connections with
dowel type fasteners is assessed by tests according to EN 26891. Here the connection strength
is defined as the maximum load before a deformation of 15 mm parallel to the load direction
is observed. In Blass et al. (2001) it is shown that, considering the strength criteria of 15 mm
relative displacement in the connection, an angle of 45° is rarely reached. Therefore the full
plastic moment capacity cannot be mobilised. To quantify this effect an algorithm is derived
to relate the plastification reached at 15 mm displacement for different dowel diameters and
different joint geometries. As a result of this an alternative equation for the estimation of the
bending yield moment of bolts and dowels in dowel type fastener connections is proposed in
Blass et al. (2001), as:

134
my 0.3 fu d ‡2.6 (5.21)

The parameters 0.3 and 2.6 are identified by regression analysis. Note that Equation (5.21) is
used in the present Eurocode 5 (EN 1995-1-1:2004).
In many tests recorded e.g. in Jorissen (1998) the load bearing capacity is reached at even
lower displacements in the connection, when compared with the 15mm requirement in EN
26891. This is especially the case for multiple fastener joints. Therefore it may be assumed
that even Equation (5.21) in some cases gives too optimistic estimates of the bending yield
capacity.

5.1.3.3 Mixed Mode Fracture Energy

A fundamental material property in fracture mechanics is the (fracture) energy which is
required to open a crack in a solid material. In fracture mechanics three different fracture
modes are differentiated; the opening mode, the sliding mode and the tearing mode, compare
Figure 5-9.

Mode I Mode II Mode III

Opening Mode Sliding Mode Tear Mode

Figure 5-9 Three opening modes in fracture mechanics.

Generally, in dowel type fastener timber connections timber cracks are due to a combination
of stresses perpendicular to the grain (opening Mode I) and shear (slide Mode II). The result is
the so called mixed mode fracture. According to Petersson (1995) and Gustafsson (1992) the
fracture energy in mixed mode I and II can be estimated as:

1 § N § 4N ··
gc ¨ 1  g ,2 ¨1  1  g ,1 ¸¸ (5.22)
N g ,1 ¨ 2N g ,1 ¨© N g ,2 ¸¸
¹¹
©
where

135
 1  N g ,3
N g ,1 (a)
g IIc

N g ,3
N g ,2 (b)
g Ic
(5.23)
2
§ rt ,90 ·
¨ ¸
N g ,3 © rv ¹
2 (c)
§ rt ,90 · MOE90
¨ ¸ 
© rv ¹ MOE0

and

g Ic 162  1.07 U den ª Nm º (a)

¬« m 2 ¼»
(5.24)
g IIc 3.5 g Ic ª Nm º (b)
«¬ m 2 »¼
where
g Ic is the fracture energy required for opening mode I > N mm @ ,
g IIc is the fracture energy required for opening mode II > N mm @ ,
U den is the timber density ª¬ kg m3 º¼ ,
rt ,90
is the ratio of tension strength perpendicular to the grain and shear strength,
rv
MOE90 is the MOE perpendicular to the grain > MPa @ ,
MOE0 is the MOE parallel to the grain > MPa @ .

Simplification of the Mixed Mode Fracture Energy Model

In practice the input parameters of the above presented framework for deriving the mixed
mode fracture energy may be not specified with high accuracy. E.g. in Jorissen (1998) the
ratio of tension strength perpendicular to the grain and shear strength rt ,90 rv is assumed to be
constant and approximated by 0.6. The ratio between the MOE perpendicular to the grain and
the MOE parallel to the grain MOE90 MOE0 is assumed to be equal to 1/30. Based on these
assumptions the above presented model can be simplified; thus the mixed mode fracture
energy g c can be modelled as a function of the density (Equation (5.25) and Figure 5-10):

136
gc 0.0013U den - 0.1918 ª Nmm º (5.25)
«¬ mm 2 »¼

g c ª« Nmm º
¬ mm 2 »¼
0.8

0.6

0.4

0.2

0
0 200 400 600 ª kg º
3
den
¬« m ¼»

Figure 5-10 The mixed mode fracture energy as a function of the density.

5.1.4 THE BEHAVIOUR OF SINGLE FASTENER CONNECTIONS

The load-slip behaviour of single fastener connections is illustrated schematically in Figure

5-11. It is dependent on the slenderness of the fastener which can be expressed roughly as the
ratio between the thickness of the middle member d m and the diameter of the fastener d‡ . In
Figure 5-11 rigid corresponds to a ratio d m d‡ 2 , rigid/slender to d m d‡ 6 and slender to
d m d‡ 12 .

q [kN]
slender

rigid/slender

rigid

3 6 9 12 15 w [mm]

Figure 5-11 Influence of the slenderness of the fastener on the load-slip behavior of a timber to
timber joint parallel to the grain, adapted from Jorissen (1998).

The slenderness ratio is also important for the assessment of the failure mode which can be
expected in a joint with a specific configuration. In Jorissen (1998) slenderness ratios are
introduced which divide expected rigid behaviour, rigid/slender behaviour and slender
behaviour for double shear connections; corresponding to the Johansen failure modes I, II and

137
III specified in Equations (5.1)-(5.4). In Figure 5-12 the dependency between slenderness -
ratio and failure mode is illustrated. The threshold values Ogr ;1 and Ogr ;2 can be quantified as:

ds fy
Ogr ;1 0.41 ; d m t 2d s (5.26)
d‡ rh

ds fy
Ogr 1.39 ; rm t 0.82rs (5.27)
d‡ rh

Here, d s is the side member thickness, d‡ is the dowel diameter, f y the yield strength for
bending, rh the embedding strength and d m the middle member thickness.

gr ;1 gr

Mode I Mode II Mode III

Figure 5-12 Graphical representation of the dependency between slenderness - ratio and
(Johansen) failure mode. Characteristic capacity over slenderness - ratio (fixed
diameter).

5.1.5 DESIGN FRAMEWORK FOR SINGLE DOWEL TYPE FASTENER

CONNECTIONS

5.1.5.1 The Eurocode Design Format

According to EC 5 (EN 1995-1-1:2004) the design resistance rd is calculated as (compare
sections 2.4.2 and 4.3.4.2):

rk
rd kmod zd (5.28)
JM
where:
rk is the characteristic value for the load bearing capacity,
zd is the design variable,

138
JM is the partial factor for material property,
kmod is a modification factor taking into account the effect of the duration of load and
moisture.
The characteristic value rk is equivalent to the 5%-fractile value of the underlying probability
distribution function of the load bearing capacity.
The characteristic value of the resistance for dowel type fastener joints is derived from the
Johansen equations (Equation (5.5)). In Equation (5.5) the index m specifies the middle
member and s the side members which are assumed to have identical properties. The
embedding strength and the fastener yield capacity are introduced as characteristic values
(5%- fractile values). The characteristic value of the embedding strength is approximated by
Equations (5.7) - (5.11) using the 5%- fractile value of the density; the characteristic value of
the fastener bending moment capacity is approximated by Equation (5.21) using the 5%-
fractile value of the tension yield strength of the fastener material.
The composite modification factor kmod is taking into account the combined effect of load
duration and moisture environment. kmod values are tabulated in the EC 5 for different timber
materials and it is distinguished between three ‘service classes’ (moisture environments) and
five ‘load duration classes’ (compare section 4.4). The values given in the EC 5 are based on
judgement and therefore mainly reflect tradition and experience. It is interesting to note that in
the EC 5 design format for timber connections kmod factors are directly applied on
characteristic values of the resistance of connections (and not on a timber material property).
The partial safety factor for structural timber J M is 1.3 and in the EC 5 it is suggested to use
the same value for the resistance of timber connections. In the present version of the EC 5
possible axial resistances of the fasteners are accounted for by multiplying factors on the
equations for the failure modes II and III (Equations (5.3) and (5.4)). For these failure modes
the fastener deformation is pronounced and the additional resistance can be explained by the
so called rope effect, i.e. axial forces in the fastener due to transferred friction between the
fastener shaft and the timber and on a compression force component due to possible washers.
This effect, however, is hard to quantify and the factors are exclusively a result of judgement.
In the present considerations these factors are thus not taken into account.

5.1.5.2 Simplified format according to DIN 1052:2004-08

Design formats based on the Johansen equations require the calculation of at least four
equations for each failure mode. In design practice this is often considered as being too
complicated and seen as a disadvantage of the Johansen equations.
In Blass et al. (1999) an alternative and simplified design framework is presented. It focuses
on the case of slender fasteners and therefore ductile failure behaviour. It is suggested to aim
for Johansen failure mode III and a minimum member thickness is introduced to induce
slender failure mechanisms. For example for double shear timber to timber connections the

139
following minimum member thicknesses for side members d s ,req and for middle members
d m ,req are proposed:

§ 9f · my
d s ,req ¨2  2¸ (5.29)
¨ 1 9 f ¸ rh , s d‡
© ¹

§ 4 · my
d m ,req ¨ ¸ (5.30)
¨ 1 9 f ¸ f h ,m d‡
© ¹
9 f is the ratio between the embedding strength of the middle member rh ,m and the embedding
strength if the side member rh , s .
In the simplified design rules it is proposed to use the minimum member thickness and use the
Johansen equation for failure mode III exclusively. If the minimum member thickness cannot
be reached it is proposed to decrease the calculated joint capacity in a linear manner by
multiplying the ratio between the member thickness and the minimal member thickness (see
Equation (5.32)).
The resistance rs of one fastener assuming that the minimum member thickness (Equation
(5.29) and (5.30)) is used is given as:

4my d‡ rh , s rh ,m
rs rJoh , III , s (5.31)
rh , s  rh ,m

Here the index m specifies the middle member and s the side members which are assumed
to have identical properties.
If the member thickness is smaller, the resistance rs is decreased as:

rs Kreq rJoh , III , s with

§ d d · (5.32)
Kreq min ¨ s ; m ¸¸
¨d
© s , req d m ,req ¹
As in EC 5 the material properties for estimating the load bearing capacity rs are introduced
as 5%-fractile values. The 5%-fractile value of the embedding strength is estimated according
to Equations (5.7) - (5.11) using the 5%-fractile value of the timber density; the characteristic
value of the fastener bending moment capacity is approximated by Equation (5.21) by using
the 5%- fractile value of the tension strength of the fastener material.
The effect of using the simplified design format can be illustrated by considering Figure 5-13.

140
r

simplified

dreq d

Figure 5-13 Effect of using the simplified design format on the predicted strength of the connection.
Characteristic capacity over member thickness and fixed diameter.

In Figure 5-13 it can be observed that the strength predictions according to the simplified
equations are always lower than predicted by using the complete Johansen equations.

5.1.5.3 Refined Design Format Considering Timber Splitting

As observed in a comprehensive test program documented in Jorissen (1998) the predictions
of the load bearing capacity for cases when the Johansen failure mode I is governing are not
properly reflecting the real load bearing capacity of connections with rigid fasteners. In this
case Equation (5.6) (splitting) may be considered as more appropriate than Equations (5.1)
and (5.2) (Johansen failure mode I).
According to the here presented refined design format the load bearing capacity is estimated
based on the Johansen Equations (Equations (5.1)-(5.4)) and the splitting mode capacity
according to Jorissen (Equation (5.6)). The relevant failure mode is equivalent to the failure
mode corresponding to the minimum value of the load bearing capacity estimated according
to Equations (5.1)-(5.4) and (5.6):

§ 1 ·
rs min ¨ rJoh , I , s , rJoh , Ia ,m , rJoh , II , s , rJoh , III , s , rsplitt , s , rsplitt , m ¸ (5.33)
© 2 ¹
The index m specifies the middle member and s the side members which are assumed to
have identical properties.
The model predictions according to the refined design method are compared with the model
predictions of the design framework according to EC 5 and the simplified design format
according to DIN 1052:2004-08 presented in Section 5.1.5.2. A parameter study is carried out,
considering different connection geometries and different material properties for timber and
steel. (end distance: a3 >5d‡ ;8d‡ @ ; timber thickness: d ª¬ d‡ ; d req º¼ ; characteristic value of
the timber density: U den ,k >300; 450@ ª¬ kg m3 º¼ ; characteristic value of the bending moment
capacity of the fastener material: m y ,k >300; 450@ > MPa @ ). Material properties are

141
introduced as 5%-fractile values of the underlying distribution functions. The result shows
one characteristic pattern for all configurations considered (Figure 5-14): For the considered
configurations the predictions according to Equation (5.6) never fall below the predictions
following the simplified approach and never exceed the predictions according to Johansen I.

simplified

splitting

dreq d

Figure 5-14 Predictions considering the splitting mode, compared to the simplified approach.
Characteristic capacity over member thickness and fixed diameter.

5.2 EFFECTIVE NUMBER OF FASTENERS

In timber engineering practice, single dowel type fasteners are rarely used. So called multiple
fastener connections are utilised; i.e. connections containing more than one fastener. In Figure
5-15 a multiple fastener connection is outlined. The connection is loaded in tension. Twelve
fasteners are arranged in two rows.

ds

dm d

a3,s a1 a1 a3,m ds

nd= 6
md= 2

q q
h

Figure 5-15 Multiple fastener connection loaded parallel to the grain in tension with nd 6
fasteners in md 2 rows.

142
A multiple fastener connection in general fails at a load level lower than the load carrying
capacity of a single fastener connection multiplied by the number of fasteners. In most
national design codes this reduction in load bearing capacity is taken into account by
introducing an effective number of fasteners nd ,ef which is smaller than the real number of
fasteners nd . In general the load bearing capacity of multiple fastener connections rmultiple is
estimated as:

rmultiple nd ,ef rs (5.34)

Different formulations for nd ,ef can be found in the codes partly based on physical modelling
and on experimental evidence. The present formulation in the EC 5 involves the in between
distance a1 , the number of fasteners in a row nd and the diameter of the fasteners d‡ and is
based on Jorissen (1998):

0.25
0.9 a1 §a ·
0.9
nd ,ef nd 4 0.53 n ¨ 1 ¸
d (5.35)
13d‡ © d‡ ¹
In the Canadian code (CSA O86-1:2001) the slenderness d d‡ of the fastener is also
considered:

0.2 0.5
§a · § d ·
0.7
nd ,ef 0.33 n ¨ 1 ¸ ¨
d ¸ (5.36)
© d‡ ¹ © d‡ ¹
where d is the thickness of the component.
The common practice in the USA is described in the NDS (1997). The effective number of
fasteners is based on an analytical study described in Lantos (1969). The multiple fastener
connection is abstracted by one-dimensional elements. The elastic elongation stiffness EA of
these elements and the foundation modulus1 are considered for modelling the load bearing
capacity of multiple fastener connections. The results of this study are tabulated in the US
American code (NDS (1997)).
In Figure 5-16 the nd ,ef - nd relationships according to the different code formats are
illustrated. It is clear that no agreement exists on the design values for nd ,ef . Although the
design formulations in Canada or Europe are both experimentally based, for large nd the
difference reaches 90%.

1
The foundation modulus specifies the gradient of the embedding stress - displacement diagram. (e.g. Figure
5-6)

143
 nd,ef
8 SIA 265
Input: EC 5_new
7
ds 36 mm 6
h 72 mm 5 Lantos
d‡ 12 mm 4 Canada

a1 84 mm 3
a3 84 mm 2
1

1 2 3 4 5 6 7 8 9 10 nd

Figure 5-16 nd ,ef according to different design codes.

The EC 5 design equation for nd ,ef given in Equation (5.35) is based on Jorissen (1998). The
parameters are estimated based on a comprehensive data base of 875 test observations
supported by the results of a numerical model for the load bearing capacity of multiple
fastener joints. All in all two different equations are proposed in Jorissen (1998). Besides the
equation used in the EC 5 design format the other equation is equivalent to the one used in the
Canadian code; the equations are rewritten as in Equation (5.37) and (5.38) respectively.

cn
§a ·
bn
nd ,ef an ¨ 1 ¸
n d (5.37)
© d‡ ¹

cn dn
§a · § d ·
bn
nd ,ef an ¨ 1 ¸ ¨
n d ¸ (5.38)
© d‡ ¹ © d‡ ¹
an , bn , cn , d n are model parameters,
a1 the spacing parallel to the grain > mm @ ,
nd the number of fasteners in a row in the grain direction,
d the member thickness > mm @ .
The quantification of the parameters of Equation (5.37) and (5.38) depends on the basis of the
load bearing capacity derivation of the single fastener joint. Two situations are differentiated:
a) Rs is based on Johansen Equations only (Equations (5.1)-(5.4)).

­
°min ª¬ RJoh , II , s1 , RJoh , III , s1 º¼  min ª¬ RJoh , II , s 2 , RJoh , III , s 2 º¼
Rs min ® (5.39)
°¯ min ª¬ RJoh , Ia ,m , 2 RJoh , I , s1 , 2 RJoh , I , s 2 º¼

See Johansen failure modes in Equations (5.1)-(5.4). Note that the Johansen failure mode I is
assumed to be brittle, failure modes II and III are assumed to exhibit ductile behaviour.

144
 R R

Brittle
Three Johansen modes Element
for the two side members
Johansen mode
middle member

Ductile
R R Element

Figure 5-17 Schematic interaction of failure modes; a) Rs is based on Johansen Equations only
(Equations (5.1)-(5.4)).

b) Rs is based on Johansen Equations and splitting (Equations (5.1)-(5.4) an Equation (5.6)).

­min ª¬ RJoh , II , s1 , RJoh , III , s1 º¼  min ª¬ RJoh , II , s 2 , RJoh , III , s 2 º¼

°
°
Rs min ® min ª¬ RJoh , Ia ,m , 2 RJoh , I , s1 , 2 RJoh , I , s 3 º¼ (5.40)
°
°¯ min ª¬ Rsplitt ,m , 2 Rsplitt , s1 , 2 Rsplitt , s 2 º¼

In Equation (5.39) and (5.40) the index m specifies the middle member, s1 and s 2 the side
members.

R R

Three Johansen modes Splitting mode

for the two side members side members
Johansen mode Splitting mode
middle member middle member

R R

Ductile Brittle
Element Element

Figure 5-18 Schematic interaction of failure modes; b) Rs is based on Johansen Equations and
splitting (Equations (5.1)-(5.4) and Equation (5.6)).

The following regression parameters are presented in Jorissen (1998):

145
Table 5-2 Regression parameters according to Jorissen (1998).
1 2 3 4

4-parameters 4-parameters 3-parameters 3-parameters

Eq.(5.38) a) Eq. (5.38) b) Eq. (5.37) a) Eq. (5.37) b)

an 0.42 0.52 0.53 0.55

bn 0.91 0.9 0.92 0.92

cn 0.28 0.28 0.28 0.29

dn 0.19 0.07 - -

These parameters are derived by regression analysis, and therefore represent the parameters of
the mean regression curve. In Jorissen (1998) it is found that these values are not appropriate
for a design equation; they are rounded off and shifted in a way such that 5% of the observed
and simulated data is larger than the predicted values.

Table 5-3 Design parameters according to Jorissen (1998).

1 2 3 4

4-parameters 4-parameters 3-parameters 3-parameters

Eq.(5.38) a) Eq. (5.38) b) Eq. (5.37) a) Eq. (5.37) b)

an 0.37 0.43 0.53 0.56

bn 0.9 0.9 0.9 0.9

cn 0.3 0.3 0.25 0.25

dn 0.2 0.1 - -

EC 5

To demonstrate the consequence of the parameter differences in Table 5-2 and Table 5-3 nd ,ef -
nd relationships for the case a) are illustrated in Figure 5-19. Four cases are compared; design
parameters (Table 5-3) or regression parameters (Table 5-2) for the formulations with 3 or 4
parameters respectively. The curves are compared with the Canadian Design Equation (5.36).
It can be observed that the differences within the different parameter sets suggested in
Jorissen (1998) (Table 5-2 and Table 5-3) are rather small compared with the rather large
difference compared to the formulation in the Canadian design code.

146
 nd,ef
“regression” 3 par
8
Input: “regression” 4 par
7
“design” 4 & 3 par
ds 36 mm 6
h 72 mm 5
d‡ 12 mm 4 Canada
a1 84 mm 3
a3 84 mm 2
1

1 2 3 4 5 6 7 8 9 10 nd

Figure 5-19 nd ,ef according to different suggestions in Jorissen (1998) compared to the
formulation in the Canadian Design Code (CSA O86.1:2001).

5.3 PROBABILISTIC MODEL FOR THE LOAD BEARING CAPACITY

OF DOWEL TYPE FASTENER CONNECTIONS

In this section a model for the estimation of the load bearing capacity of dowel type fastener
connections is derived. Therefore, at first two alternative models are presented and discussed.
In this context when referring to a model two parts are considered separately; the physical and
the probabilistic part. The physical part is based on proper mechanical hypothesises; i.e. here,
the distinction of failure modes and the corresponding interaction between material and
geometry parameters. The probabilistic part is concerned with the identification and the
quantification of the parameters as random or deterministic variables and the quantification of
model uncertainties.
As mentioned above two models are proposed;
a) A model where the physical part is closely related to the framework presented in the EC 5.
b) A model which is based on best possible knowledge, i.e. which accounts for the splitting
mode, a refined embedding strength model and also a more complex equation for
assessing nd ,ef .
The physical part of the models is summarised in Table 5-4:

147
Table 5-4 Two alternative models for the load bearing capacity: a) traditional (Following EC 5), b)
refinements according to failure mode and embedding strength evaluation.
a) b)

RJoh,.,. according to Equations (5.1)-(5.4);

Failure RJoh,.,. according to Equations (5.1)-(5.4).
modes RSplitt ,. according to Equation (5.6).

Double shear ­ min ª RJoh,C , s1 º¼  min ª¬ RJoh,C , s 2 º¼

­ min ª¬ RJoh,C , s1 º¼  min ª¬ RJoh,C , s 2 º¼ °C II , III ¬ C II , III
capacity of ° °
Rs min ®C II , III C II , III
Rs min ® min ª¬ RJoh , Ia ,m , 2 RJoh , I ,s1 , 2 RJoh , I , s 2 º¼
single °¯ min ¬ RJoh , Ia ,m , 2 RJoh , I , s1 , 2 RJoh , I , s 2 º¼
ª °
fasteners °¯ min ª¬ Rsplitt ,m , 2 Rsplitt , s1 , 2 Rsplitt , s 2 º¼

Based on density and diameter, according to Based on density and diameter, according to
Embedding Equations(5.7) - (5.11); e.g. for dowels Equation (5.12), parameters according to
strength parallel: Table 5-1:
bh
rh ,0 0.082 1  0.01d‡ U den rh ah U den d ‡ch H

Based on the ultimate tension capacity of the Based on the ultimate tension capacity of the
Yield fastener material and the diameter; fastener material and the diameter;
Capacity according to Equation (5.21): according to Equation (5.21):
my 0.3 f u d‡2.6 my 0.3 f u d ‡2.6

Based on the density; according to Equation

Mixed mode (5.25):
-
frac. energy ª Nmm º
gc 0.0013U den - 0.1918
¬« mm 2 ¼»

Based on density, according to Wood

MOE - Handbook (1987): moe0 48Uden 0.91

Main Ultimate tension capacity of the fastener Ultimate tension capacity of the fastener
material: fu . material: fu .
material
properties Density of the timber material: U den . Density of the timber material: U den .

Parameters modelled by deterministic Parameters modelled by deterministic

Geometry
values. values.

Distance Minimal end- and in-between distance Minimal end- and in-between distance
Requirements according to EC 5. a1 t 4d‡ ; a3 t 7 d‡ . according to EC 5. a1 t 4d ‡ ; a3 t 7 d‡ .

Based on single fastener connections as a Based on single fastener connections as a

Multiple reference (Equation (5.34); nef according to reference (Equation (5.34); nef according to
fasteners Equation (5.37), parameters according to Equation (5.38), parameters according to
Table 5-3 col. 3. Table 5-2 col. 2.

In the probabilistic part of the model input parameters have to be identified. The geometry
parameters are assumed to be deterministic and the timber density U den and the ultimate
tension strength of the fastener material fu are introduced as random variables. As well
important is the identification and quantification of model uncertainties.

148
5.3.1 MODEL SENSITIVITIES

The models a) and b) (Table 5-4) are representing the load bearing capacity of single dowel
type fasteners as a system of different failure modes. In order to assess the sensitivities of the
different failure modes and of the input parameters a tentative reliability calculation using
model a) is performed. According to this model, it is assumed that the Johansen failure modes
I and Ia are brittle and the failure modes II and III are ductile. Furthermore, both side
members are considered separately. The interaction of the different failure modes as
illustrated in Figure 5-20 can be obtained (compare Figure 5-17).

FJoh , III , s1 FJoh , II , s1 FJoh , I , s1

R FJoh , Ia ,m R
FJoh , III , s 2 FJoh , II , s 2 FJoh , I , s 2

Figure 5-20 Interaction of failure modes.

The component failure events F associated with the different failure modes are indicated in
Figure 5-20. The index m specifies the middle member, s1 and s 2 the side members. System
failure SF is related to the following combination of component failure events:

SF FJoh , I , s1 ‰ FJoh , I , s 2 ‰ FJoh , Ia , m ‰ FJoh , II , s1 ˆ FJoh , II , s 2 ‰ FJoh , III , s1 ˆ FJoh , III , s 2

(5.41)
‰ FJoh , II , s1 ˆ FJoh , III , s 2 ‰ FJoh , III , s1 ˆ FJoh , II , s 2

According to Equation (5.41) system failure occurs when one of the side members follows the
failure mode I , the middle member follows the failure mode Ia or the side members are
following both a failure mode II or III .
The governing combination of failure modes for the determination of the single fastener
connection strength Rs can be quantified according to a set of limit state functions as e.g.
presented here for one permanent and one variable load effect, SG and SQ :

g1 2 zd RJoh , I , s1 X M  SG  SQ 0
g2 2 zd RJoh , I , s 2 X M  SG  SQ 0
g3 2 zd RJoh , Ia , m X M  SG  SQ 0
g4 zd RJoh , II , s1  RJoh , II , s 2 X M  SG  SQ 0 (5.42)
g5 zd RJoh , III , s1  RJoh , III , s 2 X M  SG  SQ 0
g6 zd RJoh , II , s1  RJoh , III , s 2 X M  SG  SQ 0
g7 zd RJoh , III , s1  RJoh , II , s 2 X M  SG  SQ 0

with

149
zd is a design variable,
R is the resistance according to a Johansen failure mode,
SG is the permanent load,
SQ is the variable load,
XM is the model uncertainty.
The failure probability p f can be estimated as:

§ 7 ·
pf P ¨  gi d 0 ¸ (5.43)
©i 1 ¹
The failure probability can be evaluated according to e.g. first or second order reliability
method (FORM/SORM see e.g. Ditlevsen and Madsen (1996)). The equivalent reliability
index E E is defined as:

EE ) p f (5.44)

The reliability analysis is performed by using the following input variables:

Table 5-5 Input variables for the tentative reliability calculation.

timber yield permanent model
variable load
density capacity load uncertainty

5 den ª¬ kg m3 º¼ Fu > MPa @ SG > N @ SQ > N @ X M >@

distribution Normal Lognormal Normal Gumbel Lognormal

mean value 450 427 1000 1200 1

st. dev. 45 17 100 480 0.1

COV 0.1 0.04 0.1 0.4 0.1

fractile 5% 5% 50% 98% -

char. value 376 400 1000 2444 -

par. safety fac. J m 1.3 J G 1.35 J Q 1.5 -

geometry member thickness fastener diameter fastener placing

dm 2d s d‡ 12mm a3 7 d ‡ ; a4 3d‡

150
The design variable zd is derived according to the EC 5 design format, as:

J G SG , k  J Q SQ , k
zd JM (5.45)
rk

The load characteristic values SG ,k and SQ ,k and the partial safety factors J M , J G and J Q are
given in Table 5-5. The characteristic value for the resistance rk is derived according to
Equation (5.5) using the characteristic value for the embedding strength and the fastener
bending moment capacity as given in Table 5-5.
To investigate different cases, i.e. different dominant failure scenarios a parameter study upon
d m > 24mm,160mm @ is carried out.

5.3.1.1 Results
The equivalent reliability index E E (Equation (5.44)) for different values for the middle
member thickness d m is calculated. In order to investigate the sensitivity of the results on the
system model assumptions E E -values for two alternative system model assumptions are also
calculated; assuming that all Johansen failure modes are brittle and assuming that all Johansen
failure modes are ductile (both in contrast to the assumptions in Table 5-4, Model a)
brittleness for Johansen I and Ia and ductile behaviour for Johansen II and III ).

proposed system interaction all failure modes brittle all failure modes ductile

6.60 6.53
6.5 6.46 6.46
6.0
5.5
5.34

24 42 56 104 160 d m >mm @

Figure 5-21 The equivalent reliability index (Equation (5.44)) for different middle member
thickness. The results of the proposed system are compared with the results of two
alternative system model assumptions.

The E E -values range from 5.34 to 6.6 which is equivalent to a failure probability of 4.8 ˜108
and 2 ˜1011 respectively. The results according to the alternative system model assumptions
are in the same order of magnitude.
In Figure 5-22 the component (failure mode) sensitivities W comp , j are illustrated. W comp , j can be
seen as a measure of the importance of a particular failure mode. E.g. for d m 24mm the
brittle failure modes I and Ia are relevant for the reliability calculation; for d m 56mm
exclusively failure mode II is relevant.

151
 Ia 0.6 0.3 0.0 0.0 0.0
I 2x0.6 2x0.3 0.0 0.0 0.0
II 0.0 0.8 1.0 0.5 0.0
II/III 0.0 0.0 0.0 0.7 0.0
III 0.0 0.0 0.0 0.6 1.0

comp,j

0.6
0.4
0
24 42 56 104 160 d m >mm @
Figure 5-22 The componental (failure mode) sensitivities W i for different middle member thickness.

In Figure 5-23 the D - values, or sensitivity factors, of the basic variables are illustrated. The
sensitivity factors are a measure for the relative importance of the uncertainty in the
(stochastic) basic variables on the reliability index, (Madsen et al. (1986)). It can be observed
that the uncertainty associated with the model uncertainty X M and the variable load SQ is
dominating the result of the reliability calculation.

SQ
0.69 0.69 0.69 XM
i 0.68 0.68 0.68 0.65 0.65
0.52 0.59 5 den,m
SG
0.26 0.19
0.3 0.29 0.23 0.15 0.16
0.14
0.14 Fu
0.14 0.14 0.14 0.13
0.10 0.09 5 den,s
0.00 0.06 0.08
0
24 42 56 104 160 d >mm @
m

Figure 5-23 The sensitivity factors D i for the basic variables for different middle member thickness.

5.3.1.2 Intermediate Conclusions and Discussion

Cases with different relevant failure mode combinations are considered. For all combinations
it can be observed that the variables model uncertainty X M and variable load SQ are
dominating the result of the reliability calculation. The evaluated E E -values are significantly
larger than usual target reliability indices ( Etarget 4.2 according to the recommendations in
JCSS (2001)). A possible reason for this could be the inappropriate quantification of the
model uncertainty. Considering all the model assumptions integrated into the model a large
variation of the model uncertainty can be expected. The possible sources of model uncertainty
are briefly discussed next:
x Resistance is modelled as a consequence of the embedding strength of the timber and the
plastic bending capacity of the fastener. Other possible effects on the resistance of dowel
type connections are not considered.
x The embedding strength is defined as e.g. in EN 383. This standard defines the
embedment strength as the highest embedment stress within 5 mm displacement for both

152
 parallel and perpendicular to grain tests. Regression rules (associated with uncertainties)
are evaluated based on these test data (Leijten et al. (2004), Whale and Smith (1986)).
Further, the embedding stress strain relationship according to the Johansen equations is
idealized with an ideal rigid-plastic model, which obviously does not reflect the real
behaviour.
x The plastic bending capacity of the fastener is estimated by Equation (5.21), which is
derived in Blass et al. (2001). The parameters 0.3 and 2.6 are found by applying an
iterative procedure, estimating the activated plastic capacity for different geometrical
configurations and for an assumed strength criterion of 15 mm relative displacement in
the connection. In fact and as shown in Jorissen (1998) the ultimate relative displacement
can be less than 15 mm.
x The system assumptions (Table 5-4) are idealisations and associated with uncertainties.
x The statistical modelling of the material properties, the timber density and the ultimate
fastener capacity in tension is associated with model uncertainties.
Because of the high importance of the model uncertainty in reliability evaluation (Figure
5-23) and the several sources of uncertainty the model uncertainty is assessed in section 5.3.3
based on observations from experiments published in Jorissen (1998). But first the same data
set is utilized to assess the proposed models empirically and to discuss some possible model
verifications.

5.3.2 MODEL VERIFICATION

The presented models given in Table 5-4 are assessed and verified under consideration of a
comprehensive data base, Jorissen (1998). Double shear timber to timber connections are
loaded parallel to the grain in tension and compression. The fasteners are bolts without nuts.
Teflon layers between the timber members are minimising the friction between the members.
The motivation for these tests is to investigate the Johansen failure modes under additional
consideration of a splitting mode. Undesired side effects as an axial tension force components
in the fastener (rope effect) and friction between the timber members are minimised by having
no nuts on the bolts and by introducing the Teflon layers respectively.

153
Table 5-6 Data base description:

n 564 nd >1;3;5;9@ md >1; 2@

ds >12, 60@ mm dm > 24,80@ mm d‡ >10.65,19.75@ mm h >6,9@ d‡

a1 >3,12@ d‡ a2 3d‡ a3 >5,12@ d‡

Material: North European timber, visual graded according to the Dutch class B (NEN 5466),
which corresponds approximately to the European strength class C 24. The sample
characteristics for density U den are: mean value mUden 450.3 and the standard deviation
sUden 44.04 both in ª¬ kg / m3 º¼ ; n 2700 .
Measurements: Within the experimental program, the following quantities are measured:
x the density of each side and middle member: U den , s1 , U den , s 2 and U den , m ,
x the load bearing capacity: r ,
x the end-displacement, i.e. the relative displacement at failure: w' .
Based on the models presented in Table 5-4 assessments for the load bearing capacity of the
test specimen are performed. Therefore, the measured timber densities are used together with
the assumed ultimate tension strength of the fastener material and the specified geometry. As
in Jorissen (1998), the ultimate tension strength fu of the fastener material is approximated
by:

fu 400 MPa (5.46)

The model assessments are compared with the observations from the experiments.

5.3.2.1 Single Fastener Connections

The first comparison is made for single dowel type fastener connections.
In the left part of Figure 5-24 model a) from Table 5-4 is utilised to assess the load bearing
capacity r . The different failure modes according to Johansen are indicated in the Figure.
Predictions of the load bearing capacity according to the Johansen equation for failure mode I
(Equations (5.1) and (5.2)) are always larger than the load bearing capacity observed from the
experiments and are therefore on the unsafe side. Mode II and III model calculations are
smaller than the measured values. In the right part of Figure 5-24 model b) from Table 5-4 is
utilised to calculate the load bearing capacity. Johansen failure mode I is entirely replaced by
the splitting mode according to Jorissen (1998), i.e. for the configurations considered in this
test programme Johansen failure mode I is never relevant. The predictions of the splitting

154
model are in good agreement with the measured values from the experiments. The Johansen
modes II and III are also in better agreement with the measurements compared to model a).
This might be due to the different model for the embedding strength.

25 25
Mode III

20 Mode III 20

Measurement [kN]
Measurement [kN]

15 15 Mode II
Mode II

10 10 Splitting
Mode I

5 5

0 0
0 5 10 15 20 25 0 5 10 15 20 25
Model Prediction [kN] Model Prediction [kN]

Figure 5-24 Comparison of model assessments and observations; predictions are made based on
model a) (left) and model b) (right) from Table 5-4, (data from Jorissen (1998)).

For both models the results for semi rigid and slender fasteners (Johansen Modes II and III)
are consistently lower than the experimentally assessed capacity of these connections.
According to the Johansen model the load bearing capacity of a dowel type connection r is
exclusively governed by the embedding strength of the timber rh and the plastic bending
moment capacity of the fastener my , pl . Other effects than friction between the timber side and
middle members are not considered by the model. The reason for the special test
configuration in Jorissen (1998) is to exclude these side effects as much as possible and to
focus on the Johansen mechanisms. However, the consistent underestimation of the load
bearing capacity by the model provokes a deeper thought in regard to what is modelled and
what is really observed.
One reason for the deviation might be friction which occurs between bolt shaft and timber and
this effect might increase the load bearing capacity. This effect is assumed to be proportional
to the normal force between shaft and timber and different for the case of two plastic hinges
(Mode II) and four hinges (Mode III) in the fastener (Figure 5-4). The model could be refined
by introducing the two factors k f , II and k f , III , as:

ª 2
4 d s2 d‡ rh , s  4m y d‡ rh , s rh ,m º
rJoh , II , s k f , II «  d s d ‡ rh , s rh , m  1 §¨ 2d‡ d s rh , s rh ,m ·¸  » (5.47)
« 2rh , s  rh ,m 2 ¨© 2rh , s  rh ,m ¸¹ 2rh , s  rh ,m »
¬« ¼»

155
 4my d‡ rh , s rh ,m
rJoh , III , s k f , III (5.48)
rh , s  rh ,m

The factors are quantified by simple least squares technique minimising the difference
between model calculation and observation, to k f , II 1.19 k f , III 1.29 for model b).

25
three plastic
hinges
20
Measurement [kN]

15
one plastic
hinge

10

splitting
5

0
0 5 10 15 20 25
Model Prediction [kN]

Figure 5-25 Comparison of model calculations (original, refined) with measurements, (data from
Jorissen (1998)).

In Figure 5-25 the model calculations of the refined model b) are plotted and compared with
the measurements. Due to the assumed friction effect the dots in the middle and upper part of
the graph are shifted to the right.
The above results seem to be promising to capture common effects which might occur for any
connection with dowel type fasteners. This would inherently include the assumption that the
considered range of connections is representative for all dowel type connections, which
obviously is not the case: Only few different diameters are considered, the bolts fit loose into
the holes (in contrast to dowels, where the fit is tight); all tests are parallel to the grain,
friction between the members is suppressed by the Teflon layers, etc.. The result has to be
interpreted as an example and cannot be extrapolated to real situations. Therefore the idea of a
design format with a factor taking into account a friction effect k f as calibrated above is not
followed further.

5.3.2.2 Multiple Fastener Connections

In Figure 5-26 model calculations following model a) and b) (Table 5-4) are compared with
observations from tests. When comparing the two graphs a slight difference between the two
models can be observed by visual inspection. Model b) seems to deliver more accurate
predictions. However, the considerable differences in model formulations are not reflected in
Figure 5-26.

156
 200 200

150 150
Measurement [kN]

Measurement [kN]
100 100

50 50

0 0
0 50 100 150 200 0 50 100 150 200
Model Prediction [kN] Model Prediction [kN]

Figure 5-26 Comparison of model calculations and measurements for multiple dowel type
fasteners (N = 474) for model a) (left) and b) (right), (data from Jorissen (1998)).

5.3.2.3 Estimation of Model Parameters for the nd ,ef - Equations

In Table 5-4, the nd ,ef -equations are specified according to Equations (5.37) (Model a)) and
(5.38) (Model b)) with the corresponding parameters according to Jorissen (1998). These
parameters are derived based on the data base used herein with support of numerical
calculations. The same test observations as utilised by Jorissen and the calculated load bearing
capacity according to models a) and b) are utilized to reassess the parameters of the models.
Therefore the following regression equations based on Equations (5.37) and (5.38) can be
formulated:

cn
§a ·
bn
robs md nd ,ef rs md a n ¨ 1 ¸ rs H
n d
© d‡ ¹
(5.49)
§ r · §a ·
œ ln ¨ obs ¸ ln nd ,ef ln an  bn ln nd  cn ln ¨ 1 ¸  ln H
© md rs ¹ © d‡ ¹

cn dn
§a · § d ·
bn
robs md nd ,ef rs md a n ¨ 1 ¸ ¨
n d ¸ rs H
© d‡ ¹ © d‡ ¹
(5.50)
§ r · §a · § d ·
œ ln ¨ obs ¸ ln nd ,ef ln an  bn ln nd  cn ln ¨ 1 ¸  d n ln ¨ ¸  ln H
© md rs ¹ © d‡ ¹ © d‡ ¹
Assuming ln H is following a normal distribution with zero mean and standard deviation V H
the maximum likelihood method can be used to estimate the parameters of the model. Only
test data with nd ! 1 is used for this consideration. The number of observations is then
n 474. For the estimation of the load bearing capacity of a single fastener rs the models

157
presented in Table 5-4 are used with the input data specified in section 5.3.2.1.
The parameters of Equation (5.37) and (5.38) are estimated as follows (expected values):

Table 5-7 Parameters for estimating nd ,ef according to Equations(5.37) and (5.38) (mean values).

Model a) Model b)

3-parametric 4-parametric
Equation (5.37) Equation (5.38)

E > An @ 0.604 0.419

E > Bn @ 0.991 0.957

E > Cn @ 0.211 0.183

E > Dn @ - 0.363

E > 6H @ 0.195 0.122

The evaluated parameters are utilised for further comparisons. The same models a) and b) as
for Figure 5-26 is utilised with the difference that the parameters from above are used for the
nd ,ef - equation.

200 200

150 150
Measurement [kN]

Measurement [kN]

100 100

50 50

0 0
0 50 100 150 200 0 50 100 150 200
Model Prediction [kN] Model Prediction [kN]

Figure 5-27 Comparison of model results. Reassessed parameters for Equations (5.37) and (5.38)
and measurements for multiple dowel type fasteners (n = 474) , (data from Jorissen
(1998)).

As illustrated in Figure 5-27 the model calculations are in good agreement with the test
results. However, the models are mainly based on empirical considerations (the model
parameters are calibrated to test observations) rather than on physical understanding (the strict

158
differentiation of the failure modes for single fasteners etc.). It is important to note that
therefore the appearance of the models is mainly conditional to the tests considered for its
verification. The physical differentiations made by modelling one single fastener and the
involved calculation efforts are not consistent with the fact that the multiple fastener
connection model is mainly governed by the nd ,ef - equation. This fact is illustrated in the
following.
As a simple example one could think of a very simple model. By setting up the mathematical
framework of the model one could consider some physical hypothesis. The model of the
capacity could be for instance be:

cn dn
§a · § d ·
bn en
rm md a n ¨ 1 ¸ ¨
n d ¸ d m d‡ rh H (5.51)
© d‡ ¹ © d‡ ¹
where f h is the embedding strength evaluated according to model b). The parameters of such
a model could be estimated similarly to the approach used in Equation (5.50) to
an ; bn ;...; en ; V H 0.43;0.95;0.19; 0.16;1.087;0.11 . The result is illustrated in Figure 5-28,
left.
Alternatively, the embedding strength can be replaced by the timber density.

cn dn
§a · § d ·
bn en
rm md a n ¨ 1 ¸ ¨
n d ¸ d m d‡ U den H (5.52)
© d‡ ¹ © d‡ ¹
The parameters of this model are estimated similar to the approach used in Equation (5.50) to
an ; bn ;...; en ; V H 0.095;0.95;0.20; 0.07;0.92;0.11 . The result is illustrated in Figure 5-28,
right.

159
 200 200

150 150
Measurement [kN]

Measurement [kN]
100 100

50 50

0 0
0 50 100 150 200 0 50 100 150 200
Model Prediction [kN] Model Prediction [kN]

Figure 5-28 Comparison of model results according to Equation (5.51) and (5.52) with
measurements for multiple dowel type fasteners nd  1 (n = 474), (data from Jorissen
(1998)).

The models presented in Equation (5.51) and (5.52) do not include any steel material property,
nor any differentiation or detailed description of failure modes. The result, however, is rather
good. It is underlined that with Equation (5.51) and (5.52) no alternative models for multiple
fastener connections are proposed. It is demonstrated, however, that models for multiple
fastener connections are governed rather by the empirical nd ,ef -equation than by physical
consideration as the Johansen equations or the splitting mode.

5.3.3 EVALUATION OF THE MODEL UNCERTAINTY

Probabilistic models for uncertain load and resistance characteristics may in principle be
formulated at any level of approximation within the range of a purely scientific mathematical
description of the physical phenomena governing the problem at hand and a purely empirical
description based on observations and tests. In engineering analysis the physical modelling is,
however, normally performed at an intermediate level.
The proposed models for the load bearing capacity of double shear timber to timber
connections with single and multiple dowel type fasteners is based on a strong hypothesis of
the physical load bearing behaviour. However, some of the input variables of these physical
models are uncertain and have to be modelled probabilistically. E.g. the formulation for the
effective number of fasteners nef is found semi empirically, i.e. assuming that nef is a
function of the number of fasteners nd , the ratio between the fastener distance and the
diameter a1 d ‡ and the ratio between the member thickness and the diameter d d ‡ , a
multiple linear formulation of the logarithms of the assumed model parameters is obtained
(compare Equation (5.50)). Hereby both the identification of model parameters and the
mathematical formulation of the model constitute the model assumptions. The appropriateness

160
of these model assumptions can be quantified by comparing model predictions and
measurements from experiments. In the JCSS Probabilistic Model Code three possible
representations of the model uncertainty are given:

Y X M f X 1... X n (5.53)

Y X M  f X 1... X n (5.54)

Y f X M ,1 X 1 , X M ,2 X 2 ,..., X M ,n X n (5.55)

with
Y the structural performance as a random variable,
f . the model function,

XM the model uncertainty as a random variable,

Xi the basic random variables.
For the model of the strength for single fastener connections the formulation in Equation
(5.55) seems to be most appropriate since it differentiates between different model
uncertainties for different variables within the model. For instance for the model presented in
Table 5-4b) the mixed mode fracture energy is estimated based on the timber density, the
embedding strength of the timber is estimated based on the timber density together with the
diameter of the fastener, etc. These estimations are associated with uncertainties. However,
the test configuration which is considered here does not facilitate observations of the mixed
mode fracture energy or the embedding strength directly. The geometry of the connection and
the timber density is measured first and thereafter the load bearing capacity of the connection
is measured by destructive test. Based on these the model uncertainty can only be quantified
as a factor relating the entire model with the density and the geometry as the input with the
load bearing capacity as the output. Consequently the formulation of Equation (5.53) is
utilised in this case and can be written as:

Y X M , s f s z, X (5.56)

Here the model is represented as a function of a vector of random variables X and a vector of
deterministic parameters z .
In this case the model uncertainty can be quantified as log-normal distributed random variable
X M with the realisations:

161
 y
xm (5.57)
f z, x

If the mean value of X M is not equal to 1 the model is biased.

For the three models discussed in section 5.3.2.1 the model uncertainty can be quantified.

Table 5-8 Model uncertainties for the strength estimation of single dowel type fastener connections.
Model Uncertainty Mean St.dev

1 Model a) 1.117 0.240

2 Model b) 1.105 0.128

3 Model b) + Friction 0.995 0.087

Model a) tends to underestimate the load bearing capacity and is therefore biased. This holds
also for Model b) but the standard deviation of the model uncertainty is much less than for
Model a). Model b) + Friction is not biased and more accurate since the standard deviation of
the model uncertainty is less. Referring to Figure 5-24 and Figure 5-25 it has to be noted that
the model uncertainty is depending on the failure mode which is considered. For example, for
Johansen failure mode I in Figure 5-24 the model is overestimating the obtained load bearing
capacity, i.e. the model is biased and the mean value of the model uncertainty is less than one.
This effect is not reflected by the overall model uncertainty which is assumed in the tentative
reliability assessment in section 5.3.1. In particular, when considering the individual failure
modes the model uncertainties are significantly different as seen in Table 5-9.

Table 5-9 Model uncertainties for the strength estimation of single dowel type fastener connections.
(failure modes are differentiated)
Mean St.dev

Joh I 0.802 0.061

Model a)
Joh II 1.202 0.080

Joh III 1.278 0.131

Splitt 0.974 0.077

Model b)
Joh II 1.112 0.066

Joh III 1.205 0.122

Splitt 0.974 0.077

Model b) +
Joh II 0.959 0.065
Friction
Joh III 1.045 0.095

Equation (5.56) can now be extended to the case of multiple fastener joints. Considering the

162
multiplicative formulation in Equation (5.34), the model uncertainty for multiple fastener
joints can be expressed as:

Y X M ,s f s z, X X M ,nef f nef z, X (5.58)

It is not possible to assess the model uncertainty associated with the nd ,ef model individually.
It is, however, possible to quantify the model uncertainty as:

Y X M ,m f m z, X (5.59)

For the models discussed in section 5.3.2.2 the model uncertainty can be quantified as given
in Table 5-10.

Table 5-10 Model uncertainties for the strength estimation of multiple dowel type fastener
connections.
Model Uncertainty Mean St.dev.

Model a) 1.256 0.253

Model b) 1.105 0.170

Model a), nd ,ef according to Table 5-7. 1.019 0.201

Model b), nd ,ef according to Table 5-7. 1.027 0.163

Pure empirical according to Equation (5.51) 1.06 0.107

Pure empirical according to Equation (5.52) 1.06 0.108

The values Table 5-10 are evaluated for all data not differentiating between different failure
modes and different geometrical set ups.

5.3.4 LIMIT STATE FUNCTIONS FOR A SELECTED MODEL ALTERNATIVE

In the previous two sections the two alternative models are verified and model uncertainties
for different representations of the model are derived. It is shown that it is possible to reduce
model uncertainty by applying model verifications, i.e. re-calibrated model parameters and/or
changed model representations. However, none of the proposed verifications is followed
further here. The considered data set is small and not sufficiently representative. Furthermore,
the same data set is used for both, the calibration of verified models and for the assessment of
model uncertainty. Therefore, the relative low model uncertainty derived for the verified
model should be judged with caution.
The model presented in Table 5-4 a) is taken as a basis for a model proposal; the model which
is mainly following the design framework which can be found in the present version of the

163
EC 5. The timber density and the steel yield capacity in tension are introduced as random
variables. The connection is modelled as a system of the different failure modes according to
Johansen. The interaction of the different failure modes as illustrated in Figure 5-20 can be
obtained.

5.3.4.1 Single Dowel Type Fastener Connections

As in Equation (5.42) the governing combination of failure modes for the determination of the
single fastener connection strength Rs can be quantified according to a set of limit state
functions as here for one permanent and one variable load effect, SG and SQ :

g1 2 zd RJoh , I , s1 X M , I  SG  SQ 0
g2 2 zd RJoh , I , s 2 X M , I  SG  SQ 0
g3 2 zd RJoh , Ia ,m X M , I  SG  SQ 0
g4 zd RJoh , II , s1  RJoh , II , s 2 X M , II  SG  SQ 0 (5.60)
g5 zd RJoh , III , s1  RJoh , III , s 2 X M , III  SG  SQ 0
g6 zd X M , II RJoh , II , s1  X M , III RJoh , III , s 2  SG  SQ 0
g7 zd X M , III RJoh , III , s1  X M , II RJoh , II , s 2  SG  SQ 0

with
zd is a design variable,
R is the resistance according to a Johansen failure mode,
SG is the permanent load,
SQ is the variable load,
XM is the model uncertainty.
In Equation (5.60) the index m specifies the middle member, s1 and s 2 the side members.
For the quantification of the model uncertainty X M it is differentiated between the different
failure modes. The model uncertainties for the failure modes are taken from Table 5-9, Model
a).
The failure probability p f can be estimated as in Equation (5.43) and can be evaluated
according to e.g. first or second order reliability method (FORM/SORM see e.g. Ditlevsen
and Madsen (1996)). The equivalent reliability index E E is defined as in Equation (5.44).

5.3.4.2 Multiple Dowel Type Fastener Connections

In engineering design, the load bearing capacity of multiple fastener connections is assessed
based on the estimated load bearing capacity of a single fastener multiplied by the effective
number of fasteners nd ,ef , compare Equation (5.34). Accordingly, the proposed probabilistic

164
model for multiple fastener connections is based on the assessment of the single fastener load
bearing capacity. The set of limit state functions, here for one permanent and one variable
load effect, SG and SQ is given as:

g1 2 zd nd ,ef RJoh , I , s1 X M  SG  SQ 0
g2 2 zd nd ,ef RJoh , I , s 2 X M  SG  SQ 0
g3 2 zd nd ,ef RJoh , Ia , m X M  SG  SQ 0
g4 zd nd ,ef RJoh , II , s1  RJoh , II , s 2 X M  SG  SQ 0 (5.61)
g5 zd nd ,ef RJoh , III , s1  RJoh , III , s 2 X M  SG  SQ 0
g6 zd nd ,ef RJoh , II , s1  RJoh , III , s 2 X M  SG  SQ 0
g7 zd nd ,ef RJoh , III , s1  RJoh , II , s 2 X M  SG  SQ 0

with
zd is a design variable,
nd ,ef the effective number of fasteners, according to Equation (5.35),
R is the resistance according to a Johansen failure mode,
SG is the permanent load,
SQ is the variable load,
XM is the model uncertainty.
The model uncertainty is specified as in Table 5-10, EC a). The failure probability p f can be
estimated as given in Equation (5.43); the equivalent reliability index E E is defined as given
in Equation (5.44).

5.4 SUMMARY AND CONCLUDING REMARKS, TIMBER

CONNECTIONS

Timber connections with dowel type fastener are addressed in this chapter. Calculation
models for single dowel type fasteners are presented; beside the common Johansen equations
a splitting mode according to Jorissen is introduced. The relevant material properties for
dowel type fasteners are discussed and a new empirical equation for the estimation of the
embedding strength is presented, which is based on a substantial embedding strength database
consisting of data from North America, Europe and Japan.
Different design frameworks are discussed; the framework according to EC 5, a simplified
one according to DIN 1052:2004-08 and a refined format also considering the splitting mode
according to Jorissen (1998) are compared. Several equations for the estimation of the
effective number of fasteners are introduced and compared.

165
Based on these findings two alternative probabilistic model frameworks for dowel type
fastener connections are derived. One model, the traditional model, is based on the present
design rules found in Eurocode 5, and therefore rests on a broad consensus in the research
community. The other model, the refined model, features more recent research findings as
failure mode of timber splitting and an alternative formulation for the embedding strength.
Both models are assessed by considering test data from dowel type fastener connections.
Some possible model verifications are identified and model parameters are reassessed. The
model uncertainty is analysed for the two different models; it is observed that the model
uncertainty is different for different failure modes and might be reduced by using verified
parameters. The model predictions by using the refined model are closer to observed test data
than the predictions made by using the traditional model.
However, the model improvements are not considered further when all the findings are
summarised into a proposal for a probabilistic model for dowel type fasteners. The proposed
model is based on the traditional model. The model uncertainty as evaluated under
consideration of the test data is given.
The several promising achievements in regard to possible model refinements are not taken
into account. The data set used in this study, although the number of observations is rather
large, is not considered as representative for the entire domain of dowel type fasteners. In this
context it should be mentioned that this statement also has to be considered when utilizing the
model uncertainties quantified based on this test observations. However, the presented scheme
for the verification of the model for the load capacity of connections with dowel type
fasteners should be followed further under consideration of more experimental data.

166
6 PROBABILISTIC MODEL CODE - TIMBER

In chapter 4 and 5 existing knowledge about some aspects of the modelling of timber
structures is reviewed, discussed and extended. However, the different aspects are treated
rather independently. In this chapter it is the aim to envelope the above discussed aspects into
a consistent modelling framework.

6.1 SCOPE AND LIMITATIONS OF THE PROPOSAL

The proposal contains models and limit state formulations which are found to be most
relevant for the probabilistic analysis and code calibration of timber structures. The content of
this proposal should serve as a general guideline and common reference, but also as a basis
for further discussions in the research community; discussions on how to describe the rather
complex strength and stiffness related behaviour of timber and timber related materials
sufficiently precise and operational for engineering purposes.
The proposed probabilistic model code PMC for timber structures mainly concerns the
modelling of material properties of solid structural timber and the modelling of dowel type
fastener connections. The proposed models are predominantly based on test programs and
investigations considering European and North American softwoods. For some other
softwoods and especially for hardwood the underlying assumptions are less appropriate. It
should also be noted that part of given numbers in this proposal should be considered as
indicative values.

6.2 MODEL FRAMEWORK

In Figure 6-1 the general principle of the proposed modelling framework is illustrated. The
framework takes basis in the reference material properties represented by random variables in
the vector X . As introduced in chapter 4 the reference material properties are the bending
moment capacity Rm , the bending modulus of elasticity MOEm and the timber density 5 den .
Other timber material properties are represented by random variables in the vector Y . The
material properties are seen in three different contexts. In the upper part of Figure 6-1 the
material properties of test volumes measured under reference conditions1 are specified with
X s and Y s . Test standards prescribe dimensions, duration and climate conditions for these
tests. Nearly all available experimental evidence is related to tests performed according to
these test standards, i.e. to the corresponding test volumes and conditions. In the middle part

1
Reference test condition is defined in regard to the climate in which the volumes are conditioned prior to the
test and in regard to the type of loading.

167
of Figure 6-1 the indicated material properties are related to reference volumes considered
under reference conditions, X0 and Y 0 . A reference volume is defined as the volume for
which the material properties are assumed to be constant. In the bottom part of Figure 6-1 the
material properties under any conditions and any size of a component are represented with
XD and YD .

Volume
Xs

Test
Rm , s , MOEm , s , 5 m , s Ys

Condition
Test
Reference
Volume
X0 Rm ,0 , MOEm ,0 , 5 m ,0 Y0

Condition
+Volume

Any
Xα Rm ,α , MOEm. α , 5 m , α Yα

Figure 6-1 Model Framework. It is differentiated between different volumes and different
conditions. The arrows indicate where models are proposed.

6.2.1 BASIC PROPERTIES

The reference properties of structural timber are:

the bending strength rm , s in > MPa @ and the
bending modulus of elasticity (MOE) moem , s in > MPa @ ,
both measured on short-term standard test specimens evaluated according to ISO 83751 and
the
timber density U den , s in ª¬ kg m3 º¼ ,
measured according to ISO 31312.
The subscript s refers to material property according to a standardised test procedure.
For bending components it is assumed that the material properties of a cross section are

1
symmetrical 4-point bending test, span 18h ( 3 ˜ 6 ˜ h ) with h | 150mm , ramp load test duration 300 r 120s ,
specimen conditioned at nominal climate, 20 r 2 °C, 65 r 5 % relative humidity.
2
from a disc of full cross section, free of knots and resin pockets.

168
equivalent with the material properties of the reference volume; rm ,0 , moem ,0 , U den ,0 .
The reference material properties are sensitive to the deviations from the standard test
conditions. The reference material properties of a cross section in situ (i.e. under any
condition) can be estimated as:

Bending moment capacity in situ, rm ,0,D : rm ,0,D D Ex s, Z ,W , t rm ,0 (6.1)

MOEm ,0
Bending MOE in bending in situ, moem ,0,G : moem ,0,G (6.2)
1  G Ex s, Z ,W , t

Density in situ, U den ,0 :

U den ,0 U den , s (6.3)

where Ex s, Z ,W , t is the exposure of the structure to loads, humidity and temperature, over
time t ; D Ex . is a strength modification function, in general defined for a particular set of
exposures; G Ex . is a stiffness modification function, in general defined for a particular
set of exposures.
Other material properties are estimated based on the reference material properties. Referring
to section 4.5 expressions for the expected values E >.@ and the coefficient of variation
COV >.@ are given in Table 6-1.

Table 6-1: Relation reference properties – other properties (based on section 4.5).

Property Expected Values E > X @ Coef. of variation COV > X @

Tension strength par. to the grain, rt ,0 : E ª¬ Rt ,0 º¼ 0.6 E ª¬ Rm º¼ COV ª¬ Rt ,0 º¼ 1.2 COV ª¬ Rm º¼

Tension strength perp. to the grain, rt ,90 : E ª¬ Rt ,90 º¼ 0.015E > 5 den @ COV ª¬ Rt ,90 º¼ 2.5COV > 5 den @

MOE - tension par. to the grain, moet ,0 : E ª¬ MOEt ,0 º¼ E ª¬ MOEm º¼ COV ª¬ MOEt ,0 º¼ COV ª¬ MOEm º¼

E > MOEm @ COV ª¬ MOEt ,90 º¼ COV ª¬ MOEm º¼

MOE - tension perp. to the grain, moet ,90 : E ª¬ MOEt ,90 º¼
30
0,45
Compression strength par. to the grain, rc ,0 : E ª¬ Rc ,0 º¼ 5 E ª¬ Rm º¼ COV ª¬ Rc º¼ 0.8 COV ª¬ Rm º¼

Compression strength per. to the grain, rc ,90 : E ª¬ Rc ,90 º¼ 0.008 E > 5 den @ COV ª¬ Rc ,90 º¼ COV > 5 den @

mog v : E > MOEm @ COV > MOGv @ COV > MOEm @

Shear modulus, E > MOGv @
16
0.8
Shear strength, rv : E ª¬ Rv º¼ 0.2 E ª¬ Rm º¼ COV ª¬ Rv º¼ COV ª¬ Rm º¼

The relations are derived for standard test specimen properties tested under reference
conditions. However, it is assumed that the relations can be used at any level, i.e. for reference
volumes and for other climate and load conditions.

169
6.2.2 TYPICAL ULTIMATE LIMIT STATES

6.2.2.1 Components
The ultimate limit state equation for a cross section subjected to stress in one particular
direction is given as:

g zd R X M  ¦ Si (6.4)
i

where zd is a design variable, e.g. cross-sectional area, R is the resistance, e.g. tension
strength, 6Si is the sum of all possible load effects, e.g. axial stresses, X M is the model
uncertainty.
The ultimate limit state equation for cross sections subjected to combined bending and tension
stress parallel to grain is given as:

1 ¦ ¦S
§ S t ,i m ,i ·
¨ i 1 i ¸X
g 1  (6.5)
¨z ¸ M
¨ d , A Rt ,0 zd , M Rm ¸
© ¹
where zd , A , zd , M are design variables, e.g. the cross sectional area and the section modulus,
Rt ,0 , Rm are the tension strength and the bending moment capacity, 6St ,i , 6S m,i are the sum of
all possible load effects, e.g. axial stresses and bending stresses and X M is the model
uncertainty.
Ultimate limit state equations for cross sections subjected to other combined stresses can be
formulated similarly.

6.2.2.2 Connections
The proposed limit state functions for dowel type fastener connections are given in section
5.3.4.

6.2.3 TYPICAL SERVICEABILITY LIMIT STATES

E.g. when a deflection exceeds an allowable deflection limit:

g t G L  W' 6Si , MOEm , t X M (6.6)

where G L is an allowable deflection limit and W' 6Si , MOEm , t is the deflection at time t ,
depending on load effects 6Si and modulus of elasticity and X M is the model uncertainty.

170
6.3 SIMPLE PROBABILISTIC MODEL

A rather simple but operational proposal for the quantification of the input parameters is
presented in this section. The specifications may be seen as a common reference for, e.g. code
calibration procedures.

6.3.1 BASIC PROPERTIES

The distribution type and the recommended coefficient of variation ( cov ) of the basic
material properties for European softwood are given in Table 6-2. It is assumed that the
properties of test specimen rm , s , moem , s , U den , s are equivalent to the reference properties of
cross sections in a structure rm ,0 , moem ,0 , U den ,0 .

Table 6-2: Probabilistic models for reference properties.

Distribution cov
Bending strength Rm Rm , s Lognormal 0.25

Bending MOE MOEm MOEm, s Lognormal 0.13

Density 5 den 5 den , s Normal 0.1

The proposed distribution functions for the other material properties are indicated in Table
6-3.

Table 6-3 Distribution functions for other material properties.

Property Distribution Function

Tension strength par. to the grain, Rt ,0 : Lognormal

Tension strength perp. to the grain, Rt ,90 : 2-p Weibull

MOE - tension par. to the grain, MOEt ,0 : Lognormal

MOE - tension perp. to the grain, MOEt ,90 : Lognormal

Compression strength par. to the grain, Rc ,0 : Lognormal

Compression strength per. to the grain, Rc ,90 : Normal

Shear modulus, MOGv : Lognormal

Shear strength, Rv : Lognormal

171
6.3.2 CORRELATION MATRIX

The relations to other material properties are given in Table 6-1. Indicative values of the
correlation coefficients are given in Table 6-4.

Table 6-4 Correlation coefficient matrix – indicative values.

MOEm 5 den Rt ,0 Rt ,90 MOEt ,0 MOEt ,90 Rc ,0 Rc ,90 MOGv Rv

rm 0.8 0.6 0.8 0.4 0.6 0.6 0.8 0.6 0.4 0.4

MOEm 0.6 0.6 0.4 0.8 0.4 0.6 0.4 0.6 0.4

5 den 0.4 0.4 0.6 0.6 0.8 0.8 0.6 0.6

Rt ,0 0.2 0.8 0.2 0.5 0.4 0.4 0.6

Rt ,90 0.4 0.4 0.2 0.4 0.4 0.6

MOEt ,0 0.4 0.4 0.4 0.6 0.4

MOEt ,90 0.6 0.2 0.6 0.6

Rc ,0 0.6 0.4 0.4

Rc ,90 0.4 0.4

MOGv 0.6

The values in Table 6-4 are quantified by judgment (COST E24 (2005)), such that 0.8 l
high correlation, 0.6 l medium correlation, 0.4 l low correlation, 0.2 l very low
correlation.

6.3.3 STRENGTH AND STIFFNESS MODIFICATION FUNCTIONS

Values for the strength modification function D . are quantified for discrete exposures
Ex s, Z ,W , t as specified in Table 6-5. The particular sets of exposures are defined as in EC 5
(ENV 1995-1-1:2004); different load duration classes and different service classes (sc)
depending on the expected moisture content of the timber. The values for D . are taken from
EC 5.

172
Table 6-5: Strength modification function table.
Permanent Long term Medium term Short term
sc Instantaneous
( t ! 10 years) ( 0.5  t  10 years) ( 0.25  t  6 month) ( t  1 week)

1/2 D 0.6 D 0.70 D 0.80 D 0.9 D 1.1

3 D 0.5 D 0.55 D 0.65 D 0.7 D 0.9

Values for the stiffness modification function G . are quantified for discrete exposures
Ex s, Z ,W , t as specified in Table 6-7. The particular sets of exposures are defined as in the
EC 5 (ENV 1995-1-1:2004); classified as in Table 6-5. The values for G . are taken from the
EC 5.

Table 6-6: Stiffness modification function table.

Permanent Long term Medium term Short term
sc Instantaneous
( t ! 10 years) ( 0.5  t  10 years) ( 0.25  t  6 month) ( t  1 week)

1 G 0.6 G 0.5 G 0.25 G 0.0 G 0.0

2 G 0.8 G 0.5 G 0.25 G 0.0 G 0.0

3 G 2.0 G 1.5 G 0.75 G 0.3 G 0.0

6.3.4 MODEL UNCERTAINTIES FOR DIFFERENT ULTIMATE LIMIT STATES

The model uncertainties cover deviations and simplifications related to the probabilistic
modelling and the limit state equations. The reference properties are determined by
standardized tests. Therefore, model uncertainties related to estimation of other material
parameters (e.g. tension and compression strengths) have to be accounted for. Geometrical
deviations from specified dimensions, durations of load and moisture effects (damage
accumulation) also contribute to model uncertainties if not explicitly accounted for in the
stochastic modelling. Furthermore, the idealized and simplified limit state equations introduce
model uncertainties. In Table 6-7 indicative values for model uncertainties related to
components and connections are shown. The model uncertainty for components depends on
the limit state (bending or e.g. combined stress effects) and how much the actual condition
deviates from the standard test conditions. The model uncertainty for connections is based on
statistical analysis of test data when compared to the calculation models in the limit state
equations in section 5.3.3. It is noted that the expected value of X M can be considered as a
bias on the calculation models.

173
Table 6-7: Model uncertainties X M for different limit states.

mean st.dev. Distribution

Component, e.g. Eq. (6.4) Short term 1 0.05 – 0.10 Lognormal

Component, e.g. Eq. (6.13) Long term 1 0.10 Lognormal

Dowel Type Connection, section 5.3.3.

­ XM I 0.8 0.12 Lognormal

°
Single Fastener °
® X M II 1.2 0.15 Lognormal
(Failure Mode I, II, III) °
°¯ X M III 1.3 0.20 Lognormal

Multiple Fastener XM 1.25 0.3 Lognormal

6.4 POSSIBLE REFINEMENTS

6.4.1 MODELLING OF THE SPATIAL VARIATION OF TIMBER PROPERTIES

6.4.1.1 Bending moment capacity

Following the model proposed by Isaksson (1999) which is described in more detail in section
4.2.4.3, the bending moment capacity rm ,0 at a particular point j in component i of a
structure/batch is given as:

rm ,0 rm,ij exp Q  Y i  F ij (6.7)

where Q is the unknown logarithm of the mean of all sections in all components, Y i is the
difference between the logarithm of the mean of the sections within a component i and Q . Y
is normal distributed with mean value equal to zero and standard deviation V Y , Fij is the
difference between the strength of weak section j in the beam i and the value Q  Y i . F is
normal distributed with mean value equal to zero and standard deviation V F . Y and F are
statistically independent.
Accordingly Rm ,0 is a lognormal distributed random variable.

174
 ln(bending strength)

ij
ln(r ij )
xj

longitudinal direction
of the beam

Figure 6-2 Section model for the longitudinal variation of bending strength.

It is assumed that the bending moment capacity of a cross section is related with the bending
moment capacity of a test specimen Rm , s as:

Rm ,0 Rm- , s (6.8)

where - is a constant depending on the applied bending test standard and the characteristics
of the timber.
The bending moment capacity Rm ,0 is assumed to be constant within one segment (compare
Isaksson (1999)). The discrete section transition is assumed to be Poisson distributed, thus the
section length follows an exponential distribution.
The exponential distribution is given as:

FX x 1  exp O x (6.9)

For Nordic spruce the following information basis can be given (Isaksson (1999)):
The variation of the logarithm of the bending capacity ln Rm ,0 is related by 40% to the
variable Y and by 60% to the variable F . The expected length of a section is 1 480mm .
O

Table 6-8: - -values for the estimation of the strength of weak sections.
See Equation (6.8) EN US AUS
- 1.05 1.03 1.02
The different values for - given in Table 6-8 are due to the different definitions of bending
strength of test specimen. The values are derived by simulation, see section 7.3.

6.4.1.2 Bending Modulus of Elasticity

The within component variation of the bending modulus of elasticity (MOE) is based on the
within component variation model for bending moment capacity. The realizations of the MOE

175
are directly related to the realizations of the bending moment capacity following the model
proposed for bending in the foregoing. The section MOE and the bending moment capacity
are related as follows:

E ª¬ MOEm,0 º¼ 4000  150 E ª¬ Rm ,0 º¼ (6.10)

where E >.@ is the expected value of the corresponding property. The regression coefficients
are determined by judgement (based on the EN 338 strength class system, European
softwoods).

COV ª¬ MOEm ,0 º¼ 0.7COV ª¬ Rm,0 º¼ (6.11)

where COV >.@ is the coefficient of variation of the corresponding property.

The discrete section transition is assumed to be Poisson distributed; for European softwoods
the parameters of the section transition model can be taken from above.
It is noted that the empirical basis for the presented MOE model is rather scarce. However, if
no better information is available and the within component variation of the MOE is of
interest the proposed model is considered as a reasonable modelling basis for European
softwoods.

6.4.2 DURATION OF LOAD EFFECT

The mechanism leading to strength reduction of a timber member under sustained load is
referred to as creep rupture and is modelled by so-called cumulative damage models with the
general form (compare section 4.3.2.2):

dD D dD D
h s t , r0 , ș or h s t , r0 , D D , ș for 0 d DD d 1 (6.12)
dt dt
where t is time, D D is the damage state variable which commonly ranges from 0 (no damage)
to 1 (failure), the function h . contains parameters ș that must be determined from
experiment observations, s t is the applied stress and r0 the failure stress under short term
ramp loading.
The following long term limit state function is utilized:

g 1 X MDD S t (6.13)

where D D is the damage state variable after the intended service life of a structure and X M is
the model uncertainty.
Three different models are proposed:

176
1.) The model referred to as the Gerhards model (Gerhards (1979)).

dD D § s t ·
exp ¨  aD  bD ¸ for 0 d DD d 1 (6.14)
dt © r0 ¹

2.) The model referred to as the Foschi and Yao model (Foschi and Yao (1986)).

bD dD
dD D § s (t ) · § s (t ) · s t
aD ¨  K D ¸  cD ¨  K D ¸ D D (t ) for t KD
dt © r0 ¹ © r0 ¹ r0
(6.15)
dD D s t
0 for  KD
dt r0

3.) The model referred to as the Nielsen model (Nielsen 2000).

1

2 § 2 1 · bD 2
dD N § s (t ) · ¨ § § s (t ) · · § s (t ) ·
aDD N ¨ ¨
¸ ¨ DN ¨ ¸ ¸  1¸ for 1 d DN d ¨ ¸ (6.16)
dt r ¨
© 0 ¹ ¨© © 0 ¹ ¹ r ¸ ¸¸ © r0 ¹
© ¹
The parameters in Equations (6.14)-(6.16) are:
D D , DN is the damage state variable,

t is the time,

s t is the load effect,

r0 is the timber initial capacity,

aD , bD , cD , d D ,K D are model parameters fitted to experimental results e.g. as in Sørensen et al.

2005 and section 4.3.3.
For model 3.) the limit state function given in Equation (6.13) is not valid. According to
model 3.) a time variant limit state function can be given as:

2
g t SL t  X M DN SL t (6.17)

where DN is the damage state variable after the intended service life of a structure, X M is the
model uncertainty and SL t S t R0 is the load level.

6.4.3 UPDATING SCHEME FOR THE BASIC PROPERTIES

When information has been collected about the basic material properties the new knowledge
implicit in that information might be applied to improve any previous (prior) estimate of the

177
material property. For the type of information e.g. it can be differentiated between direct and
indirect information; i.e. direct measurements of the material property and the measurement
of some indicator of the property respectively.
In the proposed framework new information in principle can be introduced at any stage of
modelling. In the following the principle of considering new direct information e.g. in form of
test results is presented. The utilisation of new indirect information, e.g. inform of grading
indications, is illustrated with two different methods.

6.4.3.1 Updating - Direct Information

The bending strength Rm and the bending modulus of elasticity MOEm are modelled by
lognormal distributed random variables which can be represented through the normal
distributed random variables Rm* ln Rm and MOEm* ln MOEm . All basic properties may
be represented with the uncertain mean value 0 and standard deviation 6 as illustrated in
Equations (6.18) - (6.20).

Rm exp Rm* o Rm* ~ N 0 R , 6 R (6.18)

MOEm exp MOEm* o MOEm* ~ N 0 MOE , 6 MOE (6.19)

5 den ~ N 0 5 , 65 (6.20)

The parameters 0 and 6 of Rm* , MOEm* and 5 den are quantified with a Normal-Inverse-
Gamma-2 distribution with the parameters m, s, n, v which is equivalent to the natural
conjugate prior of a normal distribution with unknown mean and standard deviation. Given
the parameters m, s, n, v the predictive distribution of Rm* , MOEm* and 5 den can be derived as:

§ xm n ·
FX xˆ x m, n, s, v Tv ¨¨ ¸ (6.21)
© s n  1 ¸¹

where TQ . is the student-t-distribution with Q degrees of freedom.

The prior predictive distribution can be quantified with parameters m, s, n, v mc, sc, nc, vc .
See e.g. the JCSS Probabilistic Model Code, JCSS (2001) and Annex A.
New measurements on the material properties can be used for updating the parameters given
above. For a sample of n observations xˆ1 , xˆ2 ,..., xˆn , the sample characteristics can be
quantified as:

178
 1
m
n
¦ xˆi (6.22)

1 2
s2
n 1
¦ m  xˆi (6.23)

v n 1 (6.24)

(note that for observations on the bending moment capacity rm ,i and the bending MOE
moem,i , the natural logarithm of the observations, i.e. ln rm ,i and ln moem ,i has to be used).
The parameters corresponding to the prior information mc, sc, nc, vc and the sample
characteristics m, s, n, v can be combined as (see e.g. Raiffa and Schlaifa (1960)) :

ncmc  nm
mcc (6.25)
nc  n

ncc nc  n (6.26)

vcsc2  ncmc2  vs 2  nm 2  nccmcc2

scc2 (6.27)
vc  G nc  v  G n  G ncc

vcc vc  G nc  v  G n  G ncc (6.28)

with:

­0 for xd0
G x {® (6.29)
¯1 for x!0

and the posterior predictive distribution can be derived as

§ x  mcc ncc ·
FX xˆ x mcc, ncc, scc, vcc Tvcc ¨¨ ¸ (6.30)
© scc ncc  1 ¹¸

6.4.3.2 Updating - Indirect Information I

In this section a simple model for updating the statistical parameters of the Lognormal
distribution for e.g. the bending strength of a given timber grade when new information
becomes available in the form of machine grading results is described following Sørenson
(2005) and Köhler et al. (2005).

179
The Lognormal distributed strength parameter R is assumed to have a coefficient of variation
cov R . Then X ln R is Normal distributed with expected value 0 X and standard deviation
V X2 ln cov 2R  1 . V X is assumed to be known and 0 X is assumed to be Normal distributed
with expected value P0 and standard deviation V 0 .
When machine grading is based on a measured indicator L , typically related to the stiffness of
a timber test specimen, for each grading technique the following relation with the bending
strength is assumed:

L b0 r b ˜ H
1
(6.31)

where b0 and b1 are constants and H is the error term which is assumed Lognormal
distributed. ln H is then Normal distributed and is assumed to have zero mean value and
standard deviation V ln H . The parameters b0 , b1 and V ln H can be estimated using the
Maximum Likelihood method as described in section 4.2.3.2.
Given n observations of the indicator L1 ,L2 ,...,Ln for n specimens from a given timber grade,
n
the mean value of these can be estimated: L 1 n 6 L j . The updated (predictive) distribution
j 1
function for X ln R is then Normal with expected value P '' and standard deviation V ''' :

n
L  b0 V 02  P0V X2 2
b1 V 2V 2 § n b V2 · 2 (6.32)
P '' and V ''' V  20 X 2  ¨¨ 2 1 02
2
X ¸¸ V H
nV 02  V X2 nV 0  V X © nV 0  V X ¹

The updated distribution for the strength R is then Lognormal with expected value
P R'' exp P ' '0.5V ' ' '2 and standard deviation V R'''' P R'' exp V '''2  1 .

6.4.3.3 Updating - Indirect Information II (available test data – calibration of grading rules)
The probabilistic model for bending strength described in this section can be used for machine
graded timber and is based on the model described in section 4.2.3.2. The probabilistic model
can be described by the following steps:
For a given geographic region and a given type of species (e.g. Nordic Spruce) an initial
(prior) distribution function FRm x can be established for the bending strength Rm for non-
graded timber. The recommended distribution function is Lognormal. The statistical
parameters in the distribution function can be obtained using e.g. the Maximum Likelihood
method. For the identification of lower grades it is recommended to fit the initial (prior)
distribution function FRm x to the data in the lower end (e.g. 30% of the data with lowest
strengths); in order to obtain good models in the lower tail of the distribution function for the
graded timber strength. This can be done using the Maximum Likelihood method, see section
4.2.3.2.
Machine grading is based on a measured indicator L , typically related to the stiffness of a

180
timber test specimen. For each grading technique the following linear relation with the
bending strength is assumed:

L a0  a1r  H (6.33)

where a 0 and a1 are constants and H is the lack-of-fit quantity which is assumed Normal
distributed with zero mean value and standard deviation V H . The parameters a0 , a1 and V H
can be estimated using the Maximum Likelihood method which also gives the statistical
uncertainty in form of standard deviations and correlation coefficients of the parameters a0 ,
a1 and V H .
After grading the updated (predictive) distribution function for the bending strength in grade
no. j is obtained from:

FRUm , j x P Rm d x bL , j d I x d bU , j (6.34)

where bL , j and bU , j are lower and upper limits of the grading indicator L for grading no. j .
The updated distribution function FRUm , j x can then be used in reliability analyses. A detailed
description of the method can be found in section 4.2.3.2.

6.5 CONCLUDING REMARKS

A proposal is presented for probabilistic modelling of timber material properties. The basic
reference properties for timber strength parameters are described and some limit state
equations for components and connections are formulated. The recommended probabilistic
model for the basic properties is presented and indicative numerical values for the parameters
are given. Refinements related to updating of the probabilistic model given new information,
spatial variation of strength and duration of load effects are described.
The proposal can be seen as a guideline and common reference for probability based code
calibration of timber design codes. However, the parameters of the proposed models need to
be quantified on a broad and representative data base. Comprehensive experimental data
concerning the basic timber phenomena already exist, especially resulting from research
projects in North America, Europe and Australia. One major task for developing further the
presented model code is to collect and assess existing experimental data. The timber research
community is asked to contribute by making available experimental data for the quantification
of model parameters for timber predominantly used for timber design.
The presented document does not cover all aspects of the design of timber structures. Beside
solid timber other timber materials are utilized in timber engineering. Glued laminated timber
is an example of an interesting timber material, frequently used in high performance load
carrying structures. It is of utmost importance to develop consistent probabilistic models for

181
these timber materials, especially in the perspective of their potential competitiveness to other
building materials such as steel and reinforced concrete.
The further development of the probabilistic model code for timber should constitute an
important future task for the timber research community.

182
7 APPLICATIONS

7.1 ASSESSMENT OF DATA

In Appendix A a general outline for the characterisation of the statistical properties of samples
is given. In this section the methods are applied on a data set consisting of n 175
T
observations of bending moment capacity, x x1 , x2 ,..., xn . The data is arbitrary but
typical. It is assumed that the considered sample is obtained by random sampling.

7.1.1 NON-PARAMETRIC ASSESSMENT

To gain a first overview about the data set the order statistic of the sample is considered.
Therefore it is assumed that the sample values are realisations of a random variable X . The
data is ordered such that x1 d x2 d x3  d xn . Following Equations (7.1) and (7.2) the expected
value and the variation of the cumulative distribution of each bending strength observation
xm , FX xm are derived as:

m
E ª¬ FX xm º¼ (7.1)
n 1

m § m ·
VAR ¬ª FX xm ¼º ¨1  ¸ (7.2)
n  2 n 1 © n 1 ¹

E ª¬ FX xm º¼ is plotted with its scatter band in Figure 7-1.

0.5
The scatter band is defined as E ª¬ FX xm º¼ r VAR ª¬ FX xm º¼ .
A plot of E ª¬ FX xm º¼ is often compared with a probability distribution function FX x ș for
which the parameters ș are calibrated to the same set of data. This implies that the probability
specified in Equation (7.1) of an observation xm is directly compared with the outcome of the
probability distribution function FX xm . In the literature (e.g. in Foschi et al. (1989) or
Ranta-Maunus et al. (2001)) the validity of probability distribution models to represent
samples and underlying populations is often assessed based on visual judgements of these
graphical representations. However, the scatter is not taken into account.

183
 1

0.8
Probability

0.6

0.4

0.2

0
0 20 40 60 80
Bending Strength [MPa]

Figure 7-1 Plot of E ª¬ F xm º¼ (thick line) and scatter band (enveloped by thin lines).

A common non-parametric evaluation of the 5%-fractile value is based on Equation (7.3)

which specifies the probability that one further observation of X is not exceeding the value
of xm , as:

m
p n, m;1, 0 (7.3)
n 1
For the considered sample the 5%-fractile value is estimated by linear interpolation and is
equal to 21.82 MPa. The estimate is a so-called point estimate for the fractile value, only
reflecting one characteristic of a specific sample.
According to Equations (7.4) and (7.5), the sample moments of the considered sample are
evaluated as:

1 n
x ¦ xˆi
ni1
(7.4)

1 n 2
s2 ¦ xˆi  x
ni1
(7.5)

Accordingly the sample moments of the considered sample are equal to x 40.9 and
s 14.06 , both in MPa.

184
7.1.2 SELECTION OF A DISTRIBUTION FUNCTION AND PARAMETERS

A convenient approach to check the validity of a chosen distribution model is the use of
probability paper. Probability paper for a given distribution model is constructed such that the
cumulative probability density function for that distribution model will have the shape of a
straight line when plotted on the paper. A probability paper is thus constructed by a non-linear
transformation of the x and the y axis. The considered data set is plotted on Probability paper
for a Normal, a Lognormal and a Weibull distribution. The transformations are indicated in
the graph.

Normal Lognormal Weibull

ln  ln 1  F x
) 1 F x
) 1 F x

x ln x ln x

Figure 7-2 Probability paper for Normal, Lognormal and Weibull distributions.

Based on Figure 7-2 the lognormal distribution can be identified as most properly reflecting
the statistical characteristics of the considered sample, especially in the lower tail regions of
the distribution. It is well known that especially the behaviour of the probability distribution
functions in the regions of the tails is of importance in reliability assessments. For load
variables the upper tail is normally the most important whereas the lower tails are the
important for resistance parameters (see e.g. Faber (2003)).
It thus remains to model timber material properties with special emphasis on a good
representation in the lower tail region. One approach to estimate the probability distribution in
the lower tail domain is by means of a censored Maximum Likelihood estimation where only
observations in the lower tail domain i.e. below a given predefined threshold value are used
explicitly. The other observations are only utilised implicitly to the extent that it is recognised
that they exceed the threshold.
A general description of Maximum Likelihood method is given in Annex A on page 6. In this
special case of a censored Maximum Likelihood estimation two different contributions to the
likelihood are considered, i.e.:

j
L1 –f
i 1
xi ș (7.6)

and

185
 n j
L2 P X t xG ș

with: (7.7)

P X t xG ș 1  F xi ș

where L1 represents the likelihood of the j observations with values below or equal to the
threshold value xG . L 2 represents the likelihood of the observations with values exceeding
the threshold value xG . 1  F xG ș is the probability that a value exceeds the threshold
value xG given the parameters of the probability distribution function ș . If n is the total
number of observations n  j is the number of observations exceeding the threshold value
xG .
The parameters are easily estimated by the solution of the optimisation problem:

max L1 ˜ L 2 (7.8)
T

The parameters ș are estimated as normal distributed random variables with means and
covariance quantifying the statistical uncertainty due to the relative small number of
observations below the threshold.
The parameters ș of the probability distribution functions estimated on the basis of Equation
(7.8) have been found to exhibit moderate sensitivity to the choice of the threshold value xG
for moderately low values of xG . The lower 30 % percentile value has been found to be a
reasonable choice for xG (Faber et al. (2004)).
For the considered bending strength observations, the maximum likelihood method is used to
estimate the parameters of the 2-parameter Weibull, the Normal and the Lognormal
distribution (compare Annex A). For means of comparison the parameters are calibrated to the
entire data domain and by considering exclusively the data from the lower tail of the
distribution. xG is chosen to be equal to the lower 30% percentile value. In Table 7-1 the
distribution parameters are summarized as correlated normal distributed random variables. In
Figure 7-3 plots of both predictive distributions are compared with the quantile plot of the
data.

186
Table 7-1 Distribution parameters calibrated to the considered bending strength data.
Weibull Normal Lognormal

T w k P V [ G
PT 45.76 3.09 40.91 14.02 38.61 0.342

All data VT 1.19 0.17 1.04 0.93 0.99 0.018

UT 0.33 0.34 0

PT 38.66 5.29 36.29 8.56 38.77 0.345

Lower tail VT 0.02 0.48 0.70 0.24 1.014 0.019

UT 0.19 0.22 0.01

Normal Lognormal Weibull

1.0 1.0 1.0
Probability

0.8 0.8 0.8

0.6 0.6 0.6
0.4 0.4 0.4
0.0 0.0 0.0
0 40 0 40 0 40
Bending Strength [MPa] Bending Strength [MPa] Bending Strength [MPa]

Figure 7-3 Plots of the predictive distribution functions in comparison with the E ª¬ F xm º¼ plot;
compare Figure 7-1.

Investigating the plots for the Weibull and the Normal distribution it can be found that the
lower tail domain of the data (quantile plot) is significantly better represented by the
distributions which are calibrated explicitly to the lower tail. For the Lognormal distribution it
is evident that both distributions represent the entire data domain irrespectively which
calibration method is used. It should be noted that this observation is not typical for bending
strength data sets. For a comparison it is referred to Ranta-Maunus et al. (2001)) where a
considerable amount of data bases is analyzed and documented.

7.1.3 ESTIMATION OF THE 5%- FRACTILE VALUE

As discussed in chapter 2, 5%-fractile values are used as characteristic strength values in

deterministic design formats. It is of importance that the estimation of these values is fully
consistent with the probabilistic modelling of strength related material properties. In the
following the estimation of 5%-fractile values according to existing standards is discussed.

187
7.1.3.1 Standardized methods for the estimation of the 5%-fractile value
For timber materials several standardized methods for the estimation of the characteristic
value xk exist. Four different standards are considered here:
x EN 384: valid for solid timber,
x EN 14358: valid for wood based products,
x ISO 13910: valid for solid timber,
x ISO 12491: valid for all building products.
In the following these methods are briefly reviewed:

EN 384 – stratified sampling:

This method is based on stratified sampling, i.e. a sample is subdivided into at least 5 sub-
samples j of size n j ! 40 and for each sub-sample the 5%-fractile value x05, j is assessed
according to the order statistics method described in section 7.1.1. The value x05 is estimated
as the weighted mean value of the 5%-fractile values of the sub-samples as:

¦n x
j
j 05, j

x05 (7.9)
¦n j
j

The characteristic value xk is estimated as:

­° x05
xk min ® (7.10)
1.2 min x05, j
°̄ j

ISO 13910 – Weibull distribution

The sample 5%-fractile value x05 is estimated according to the order statistics method
described in section 7.1.1. The data is calibrated to a 2-parameter Weibull distribution under
explicit consideration of the lower tail data. The lower tail is defined with the threshold value
xG corresponding to the lower 15% of the data. The Weibull distribution is given in Annex A
with the parameters w and k . vtail is introduced as a approximation of the coefficient of
variation of the Weibull distribution as:

vtail k 0.92 (7.11)

The characteristic value is given as:

188
 ª vtail º
xk «1  2.7 » x05 (7.12)
¬ n¼

where n is the sample size.

ISO 12491 – Normal distribution

The estimation of the characteristic value xk is based on the sample mean x and standard
deviation s as:

xk x  ks s (7.13)

where k s is a correction factor.

Equation (7.13) corresponds to a lower bound estimate at a given double sided confidence
interval D k . Accordingly k s is given by:

kD
ks  (7.14)
n
where n is the sample size and kD is the D k -percentile of the non-central t -distribution with
n degrees of freedom and the non-centrality parameter Onc given as:

Onc ) 1 1  p n (7.15)

where p is the fractile of interest (in this case equal to 0.05) and ) . is the standard normal
distribution function.
When the population standard deviation V is assumed to be known Equation (7.13) is
rewritten as:

xk x  kV V (7.16)

where

1 ) 1 D k
kV ) 1 p  (7.17)
n

EN 14358 – Lognormal distribution

In the EN 14358 the approach of ISO 12491 is applied to a lognormal distribution.

189
7.1.3.2 Predictive 5%-fractile values
According to modern design codes as the Eurocodes (EN 1990:2002) characteristic values for
strength related material properties have to be introduced as predictive values of the 5%-
fractile. The general form of the predictive p %-fractile value can be given as:

x p , pred FX,1pred p with f X , pred x ³f x ș fĬcc ș xˆ dș (7.18)

where x̂ are the sample observations, ș are the parameters of the distribution function. The
parameters ș are realisations of the random vector Ĭ with the posterior joint probability
density function fĬcc ș xˆ (compare in Annex with Equation (A.32)). Equation (7.18) can be
generally solved by reliability methods as FORM/SORM or numerical integration, however,
analytical solutions exists, e.g. for the case where X is normally distributed. In this case the
predictive value of the p %-fractile x p , pred is given as:

x p , pred exp m  t p v s 1  1 n (7.19)

where x is the sample mean, s the sample standard deviation, n is the sample size and Q is
defined by Q n  1 . t p Q is the p %-fractile value of the t -distribution with Q degrees of
freedom.
It should be noted that this method is fully consistent with the Bayesian updating scheme
discussed later in this chapter.

7.1.3.3 Example
The described methods are used for some example calculations with the data set which is
already used in section 7.1.1. The results are presented in Table 7-2.
It can be seen in Table 7-2 that the estimations for the 5%-fractile values are very sensitive to
the applied estimation method. It can be noticed that the estimations which are based on the
assumption that the data is normal distributed are rather low (row: 3, 5, 8 (all data)). It is also
interesting to note, that the result of stratified sampling (row 1) differs significantly from the
non-empirical approach according to Equation (A.8) (row 11). Consistent results are obtained
when the parameters of a distribution function are calibrated to the lower tail domain of the
data by using the maximum likelihood method. It can be seen that the resulting predictive
fractile values are not very sensitive to the assumed distribution (compare row 8, 9, 10 (lower
tail)). The method according to Equation (7.19) is fully consistent to the Bayesian updating
scheme for normal distributed random variables. The method can also be used for lognormal
distributed random variables by transforming the data accordingly. The estimates are sensitive
to the assumed distribution (compare row 5 and 6).

190
Table 7-2 Different 5%-fractile values for the same sample.
row Method

1 EN 384, Equation (7.10) stratified sampling 22.38

2 ISO 13910, Equation (7.12) Weibull 20.86

3 ISO 12491, Equation (7.15) Normal 16.60

4 EN 14358, Equation (7.15) Lognormal 21.35

5 Predictive, Equation (7.19) Normal 17.58

6 Predictive, Equation (7.19) Lognormal 21.87

Predictive, Equation (7.18), Distribution all lower

7
and parameters as in Table 7-1: data tail

8 Normal 17.75 22.02

9 Lognormal 21.95 21.90

10 Weibull 17.43 21.90

11 Non-parametric, Equation (A.8) 21.82

7.2 UPDATING OF RANDOM VARIABLES

7.2.1 NORMAL DISTRIBUTION WITH UNCERTAIN MEAN AND KNOWN

STANDARD DEVIATION

The probability distribution of the bending strength of a sample of timber structural elements
has to be estimated. As prior information it is known that the sample is assigned to be graded
to a particular timber grade but it is neither known how the timber is graded nor where the
ungraded material is coming from. The timber grade is specified by a 5%-fractile value of the
bending strength. It is assumed that the bending strength is following a lognormal distribution
and that the standard deviation of a typical batch is constant. Between batches of different
suppliers it is implied that the location, i.e. the mean value, of the distribution is varying. Thus
the mean value is uncertain.
The lognormal distribution is equivalent to the normal distribution, if the following
transformation is applied:

z ln x and
§ ln x [ X · § ln x  ln [ X · § z  PZ ·
FX x )¨ ¸ )¨ ¸ œ FZ z )¨ ¸ for P Z ln [ X and (7.20)
© GX ¹ © GX ¹ © VZ ¹ VZ GX

Further, the coefficient of variation ( cov ) of the lognormal distribution is approximately

equivalent to its parameter G X :

191
cov exp G X2  1 # G X (7.21)

In the present example the prior information is quantified as follows: It is assumed that the
5%-fractile value of the predictive prior distribution matches the required 5%-fractile value of
the bending strength of 24 MPa (corresponding to the European strength class C24, compare
Table 3-3). Further, the coefficient of variation of the predictive distribution is assumed to be
0.3, where as 70% of the scatter is due to the constant standard deviation and 30% is due to
the standard deviation of the uncertain mean.
The prior information is:
x o ln x z : Z ~ N M Z , V Z ; M Z ~ N P c, V c

VZ 0.25
P c 3.67
V c 0.16
Based on the prior information the predictive distribution of Z is normal distributed with
P P P c 3.67 and V P V c2  V Z2 0.3 .
New information can be introduced, e.g. information from tests of the bending strength of a
number of structural elements. Here the additional information is (arbitrary chosen values):

xˆ 20,30,50, 70,80 o zˆ ln xˆ 3.0,3.4,3.9, 4.2, 4.4

n 5
z 3.87
Considering the prior information and the additional information from the tests the predictive
distribution of Z is a normal distribution with the parameters: P U P cc 3.8 and
V U V ccc 0.33 , (compare with Equation (A.33)).
Both distributions, i.e. before introducing new information and afterwards, are plotted in
Figure 7-4. It is seen that by introducing new information the predictive probability
distribution of the bending strength is shifted to the right, i.e. the expected strength values are
higher than prior to the tests. It should be noticed that the new information is quantified by the
mean value of the sample values and the number of observations within the sample, in this
case z 3.87 and n 5 . Implicitly, it is assumed that the standard deviation of the drawn
sample is already known with s 0.25 .
The predictive 5%-fractile value is calculated with (with V assumed to be known):

x0.05, pred exp P  ) 1 0.05 V (7.22)

with ) 1 0.05 | 1.64 .

192
Prior to the tests with, P P P c 3.67 and V P V c2  V Z2 0.3 o x0.05, pred 24 MPa .
Posterior to the tests with, P U P cc 3.8 and V U V ccc 0.33 o x0.05, pred 24.7 MPa .

1
Prior

0.8
Probability

0.6
Posterior

0.4

0.2

0
0 20 40 60 80
x - Bending Strength [MPa]

Figure 7-4 Predictive probability distribution functions considering only prior information and
considering both, prior and additional information (posterior).

7.2.2 NORMAL DISTRIBUTION WITH UNCERTAIN MEAN AND STANDARD

DEVIATION

Considering the same example as above, where the probability distribution of the bending
strength of a sample of timber structural elements has to be estimated. Again, as prior
information it is known that the sample is assigned to be graded to a particular timber grade
but it is neither known how the timber is graded nor where the ungraded material is coming
from. It is assumed that both, the mean value and the standard deviation are quite sensitive to
the grading procedure and the origin of the timber. Consequently the prior information is
leading to vague estimates of the parameters of the distribution; i.e. the mean and the standard
deviation are both unknown.
The prior information is quantified inform of the parameters of the Normal-Inverse-Gamma-2
distribution (Equation (A.35)) mc, sc, nc, vc . For this example the parameters are quantified by
choice, but that the resulting prior predictive distribution function is very similar to the prior
predictive distribution function from the example above:
x o ln x z : Z ~ N M Z , 6Z

193
mc 3.7
nc 5
sc 0.25
vc 6
The values of mc and sc are based on (abstract) samples of equivalent sizes of nc and vc
respectively. Note that for the given parameter set the ‘weight’ of the prior information is
relatively low.
New information can be introduced, e.g. inform of tests of the bending strength of a number
of structural elements. The same sample as for the test sample above is used:

xˆ 20,30,50, 70,80 o zˆ ln xˆ 3.0,3.4,3.9, 4.2, 4.4 with the sample characteristics:

m z 3.78
n 5
s 0.58
v n 1 4
Considering the prior information and the additional information from the tests the predictive
distribution of Z is t-distributed (Equation (A.41)) with the parameters:

mcc z cc 3.75
ncc 10
scc 0.4
vcc 11
Both distributions, i.e. before introducing new information and afterwards, are plotted In
Figure 7-5.
The predictive 5%-fractile value is calculated with

x0.05, pred exp m  t0.05 v s 1  1 n (7.23)

where t0.05 Q is the 5%-fractile value of the student-t-distribution with v degrees of

freedom.
Prior to the tests with, mc, sc, nc, vc 3.71,5, 0.25, 6 o x0.05, pred 24 MPa .
Posterior to the tests with, mcc, scc, ncc, vcc 3.75,10, 0.4,11 o x0.05, pred 20 MPa .

194
 1
Prior
0.8
Probability

Posterior
0.6

0.4

0.2

0
0 20 40 60 80
x - Bending Strength [MPa]

Figure 7-5 Predictive probability distribution functions considering only prior information and
considering both, prior and additional information (posterior).

0.8
Probability

0.6
Example b)
0.4
Example a)

0.2

0
0 20 40 60 80
x - Bending Strength [MPa]

Figure 7-6 Comparison of the posterior predictive distribution functions derived according to
example a) and b).

In Figure 7-5 it is seen that in this example the picture is quite different than in example a)
above (Figure 7-4). Here the additional information is introduced in form of the mean value of
the observed sample values, but also in form of the standard deviation of the sample
( s 0.58 ) which is significantly larger than presumed in example a). While the additional
sample is exactly the same and the predictive prior distribution is similar, the predictive
posterior is completely different (compare also with Figure 7-6).

195
When assuming known standard deviation, as for example a), the consistency of that
assumption has to be proven carefully.

7.3 COMPARISON OF DIFFERENT BENDING TEST

SPECIFICATIONS

To model the variation of strength and stiffness related material properties between
components the variation on a meso level has to be taken into account, see chapter 4.
Consider a timber producer, e.g. a saw mill, with a supply of timber raw material which can
be assumed as stationary in its properties. It is further assumed that the production and
selection schemes are constant. A typical output of the saw mill, e.g. normal construction
timber of a specific size, can be assumed to be equivalent to a population and samples can be
made to assess the characteristics of such a population. Full-size test can be performed to
assess the variability of the load bearing capacity between the sampled components directly.
Due to the sensitivity of the load bearing capacity of components not only to its size, but also
to climate and loading conditions, test standards are used when investigating the between
component variability.
Several probabilistic models where introduced to model the observed variation in strength
and/or stiffness; e.g. the Normal distribution, the log-Normal distribution and the Weibull
distribution. Large databases exist showing distribution parameters of specific populations of
timber all tested according to certain test standards. These standards are in general associated
with national guidelines or codes for testing structural timber. The most important ones are
summarised in the following; to illustrate the differences the standards for the evaluation of
bending strength are compared.

7.3.1 EUROPEAN STANDARDS EN 408 AND EN 384

The standard EN 408 specifies the laboratory methods for the determination of some physical
and mechanical properties of structural timber. It is based on ISO 8375. A symmetrical four
point bending configuration is used to evaluate the bending strength of components. The
specimens have spans of 18 times the depth and the loads are introduced vertically at the third
points; Figure 3-6. Maximum load shall be reached within 300 ± 120 secounds.
EN 384 defines the evaluation of characteristic values for the strength, stiffness and density of
structural timber. Reference conditions for bending tests are specified as follows. A critical
section of each specimen has to be identified and placed in between the loads in the centre
third of the arrangement. The tension edge shall be selected at random and the reference
moisture content shall be consistent with a surrounding climate corresponding to a
temperature of 20°C and 65% relative humidity. The depth of the specimen shall be equal to
150mm.

196
7.3.2 AMERICAN SOCIETY FOR TESTING AND MATERIALS, ASTM D 4761-88
AND D 1990-91

The North American standard for testing the mechanical properties of timber and wood based
material is ASTM D 4761-88. The span of bending specimens are also expressed as a multiple
of the depth of the specimen. The span ranges from 17 times the depth to 21 times the depth.
A four point bending configuration according to Figure 3-6 is prescribed. The time to failure
shall be approximately 60 seconds; more than 10 and less than 600 seconds. According to
ASTM D 1990-91, the North American standard for establishing allowable properties of
visually graded timber, the critical zone shall be randomly located between the supports. The
tension side shall be also selected at random. The moisture content of the timber shall be 13%.

7.3.3 AUSTRALIAN/NEW ZEALAND STANDARD, AS/NZ 4063:1992

This standard prescribes a random location of the test section. The test configuration is
identical to ISO 8375, i.e. to EN 408.

Table 7-3 An overview comparison between different bending strength test procedures.
Climate/Moisture Loading/Time
Origin/Code Geometry Bias
Content tofailure

4 point bending (By judgment)

Conditioned at Ramp load, weakest section in
Europe/EN 408 and (Figure 3-6)
Temp.: 20°C Time to failure: the middle.
EN 384 L = 18 H
Rel. Hum:: 65% 300 s ± 120 s Tension side
H = 150 mm random.

4 point bending (By judgment)

North America/ Ramp load, weakest section
(Figure 3-6) Moisture Content:
ASTM D-4761-88 Time to failure: within supports.
L = 17 H – 21 H 13%
and 1990-91 60 s ~ (10s, 600s) Tension side
H = 150 mm random.

4 point bending
Australia/New Conditioned at Ramp load, Time
(Figure 3-6)
Zealand Temp.: 20°C to failure: -
L = 18 H
AS/NZ 4063:1992 Rel. Hum:: 65% 300 s ± 120 s
H = 150 mm

7.3.4 EVALUATING THE EFFECT OF DIFFERENT TEST STANDARDS

7.3.4.1 Duration of Load

When test data from different sources are assessed it is important to be aware of the different
bending test specifications. As seen in chapter 2 timber material properties are sensitive to the

197
duration of the applied load. In the above introduced test standards average time to failure is
specified. In the European and the Australian standard it is 300s, in the US standard it is 60s.
It can be expected that the influence on the measurements is significant. The so-called
Madison curve (Equation (4.41)) can be utilised to get an estimate of the effect of these load
duration differences.
According to Wood (1949) the Madison curve is written as:

sl 18.3  108.4 t f 0.0464 (7.24)

Where t f is measured in seconds. According to the Madison curve the difference of bending
strength measurements with a test of 60s duration are in average 6.5% larger than strength
measurements with the test of 300s duration.

7.3.4.2 Weak Section Placing

The different test standards specify how to place the assumed weakest section in the bending
test. The European standard requires placing the weakest section in the middle, high stressed
region, according to the US standard the weakest section has to be placed within the supports
and following the Australian standard the weakest section shall be placed randomly. The
difference between the test standards in regard to placing the test specimen is illustrated in
Figure 7-7. In the Figure the weakest section is indicated by a knot cluster. It can easily be
imagined that the effect between the different specifications becomes larger when the length
of the entire specimen increases. In practice it is in general not clear how long the specimen
is.

Europe

United States

Australia

Figure 7-7 Illustration of the effect of different weak section placing specifications; Europe (EN
384, weak section between load application), United States (ASTM D 1990-91, weak
section between supports), Australia (AS 4063:1992, weak section at random).

To illustrate the effect of the weak section placing specification the model derived in Isaksson

198
(1999) is utilized for simulations. The model is used as it is presented in the proposal for the
probabilistic model code in chapter 6, page 174. The Monte Carlo Simulation technique (see
e.g. Melchers (2001)) is used for the simulation of the random variables in the model. It is
assumed that the beams have a length of 5 m. 1000 bending components are generated with a
weak section distribution as indicated on page 174. The components are virtually tested
according to the three different bending strength test specifications. The obtained data is
utilized to calibrate the parameters of a lognormal distribution. The corresponding distribution
functions are plotted together with the distribution function of the strength of the weak
sections in Figure 7-8. The parameters of the distribution functions are given in Table 7-4.

Table 7-4 Distribution parameters corresponding to Figure 7-8.

[ G
Weak sections 56.28 0.247

Europe 45.68 0.213

United States 50.36 0.247

Australia 51.95 0.232

0.8
US
Probability

0.6 Europe

0.4 weak section

0.2 Australia

0
0 20 40 60 80
Bending Strength [MPa]

Figure 7-8 Distribution Functions of simulated bending test specimen according to different
national test standards.

The distribution functions shown in Figure 7-8 illustrate the difference between the bending
moment capacity of test specimen and the bending moment capacity of all weak sections. As
introduced in the proposal of the probabilistic model code in chapter 6 the capacity of a weak
section corresponds to the capacity of a reference volume. An interesting question within the

199
probabilistic modelling of timber material properties is how to relate measurements on test
specimen rm , s to the properties of reference volumes rm ,0 . For the given example a simple
relation is found with the form:

rm ,0 rm-, s (7.25)

The parameter - is calibrated for different test standards by using the least squares technique.
The results are given in Table 7-5.

Table 7-5 - -values for the estimation of the strength of weak sections.
See Equation (6.8) EN US AUS
- 1.05 1.03 1.02
According to Equation (7.25) the parameters of the lognormal distribution can be related as:

-
[r m ,0
[r m ,s

(7.26)
Gr m ,0
Gr m ,s
-

It should be underlined that the given example is based on the within component bending
moment capacity variation model and the corresponding model parameters as presented in
Isaksson (1999) and chapter 6.

200
8 CONCLUSIONS AND OUTLOOK

8.1 CONCLUSIONS

8.1.1 RETROSPECT AND MOTIVATION

During the last decades structural reliability methods have been further developed, refined and
adapted and are now at a stage where they are being applied in practical engineering problems
as an decision support tool in connection with design and assessment of structures.
Furthermore, basic knowledge concerning the actions on structures and the material
characteristics has improved due to increased focus, better measuring techniques and
international research co-operation. This knowledge has now reached a level where it enables
experts to take into account uncertainties in material properties and actions when assessing the
load carrying capacity, serviceability and service life of structures. This is not least due to the
fundamental works on structural reliability methods performed within the Joint Committee on
Structural Safety (JCSS) including, among others, the basic reports on actions on structures,
basic reports on material resistances, the guideline for reliability based assessment of
structures and the almost complete JCSS Probabilistic Model Code. These documents provide
general guidelines for the use of structural reliability methods in practical applications and at
the same time constitute the basis for ensuring that such analysis are performed on a
theoretically consistent and comparative basis.
For materials such as concrete and steel, this has led to an increasingly more consistent
evaluation of the safety or reliability, i.e. the probability that a structure will fulfil its function
throughout its service life. Whereas some efforts in this direction have been undertaken also
for timber; the developments, however, have been less impressive for several reasons:
x The variability of the material properties is much higher than for other building materials.
This poses problems but also implies that the advantages of introducing reliability
concepts may be greater.
x The material properties, and therefore also the probability of failure, depend on the entire
load and moisture history of the structure.
In the last few years the interest in designing timber structures has steadily increased. The
reason for this being an increased focus in society on sustainability and environmental aspects
but also due to the positive effects of timber materials on the inner climate in residential
buildings and excellent architectural possibilities. Furthermore, timber has been found
technically and economically competitive compared with steel and concrete as a building
material for a broad range of building structures such as e.g. large span roof structures and
residential buildings.

201
It is thus now urgently important that a consistent basis for design of timber structures is
established and documented in such a way that it may be accepted for implementation by the
timber engineering and research community. Only then can a consistent basis for the design of
timber structures be insured for the benefit of the society in general and for the parties with
interest in the application of the timber as a building material in particular.
The development of a basis of design for timber structures is the main objective of this thesis.

8.1.2 APPROACH AND SUMMARY

The logical scheme of the development of a basis for the design of timber structures is
constituted by successively following chapters 2 – 7.
In chapter 2 the main issues of structural reliability are reviewed and discussed. It is focused
on the topics which are considered as most relevant for this thesis. The limit state principle is
introduced; uncertainties and their influence on engineering decision making are discussed.
Basic principles of the probabilistic modelling of loads and resistances are presented. The
most essential methods of structural reliability and different probabilistic and deterministic
design concepts are outlined and compared.
Chapter 3 addresses on timber as a structural material. The growth of a tree is described to
illustrate how wood is produced. Many particularities of structural timber can be attributed to
the growing conditions of a tree and the characteristics of the wood cell. The chapter is
concluded with the definition of timber material properties at the so-called element level and
the introduction of structural timber strength classes as a consequence of timber grading.
Chapters 4 and 5 represent the development part of the thesis. In chapter 4 a scheme for the
probabilistic modelling of the material properties of solid timber is introduced, differentiating
between three main aspects: scale dependency of timber material properties, time dependency
and the interrelation between different material properties (see also Figure 4-2). The spatial
variability is discussed in more detail, differentiating between different levels of variation.
The corresponding perceptions constitute the basis for better understanding and assessment of
the high variability of timber material properties. The properties of timber components depend
on the load- and climate history to which they are exposed during their service lives. Different
models of the strength degradation under sustained load are discussed, further developed and
applied for reliability calculations. At the end of chapter 4, functional interrelations between
different timber material properties are presented.
In chapter 5 the probabilistic modelling of timber connections is discussed, focusing on a
particular type of connection: connections with parallel loaded double shear dowel type
fasteners. Existing calculation frameworks are reviewed and discussed. A probabilistic
calculation model is developed where the different failure modes of a connection are
considered as a system reliability problem. A large database is used to calibrate the model
uncertainty and to discuss possible refinements of the calculation model. The database,

202
however, was found to not be applicable for the development of an alternative calculation
model for dowel type fastener connections.
In chapter 6 the conclusions are condensed into a proposed basis for the design of timber
structures. After several discussions 1 it was decided to structure the proposal for the
probabilistic model code for timber into several levels of sophistication. The basic level
reflects the recent practice for reliability based code calibration. The bending strength and
stiffness and the density of timber are referred to as reference material properties and are
introduced as simple random variables. The basic limit state functions for components and
connections are given. Furthermore, proposals are made regarding the different characteristics
of timber on this simple level. Scale effects are disregarded and the time, load and moisture
dependency is covered by deterministic factors for discrete time, load and moisture scenarios.
Functional relationships for other material properties (based on the reference material
properties) are given and probability distribution functions for the other material properties
are proposed. Starting from this level, several possible refinements are proposed. New
information might be introduced, and it is shown how different types of new information can
be integrated by using a Bayesian updating scheme. Refinements in regard to the modelling of
damage as a consequence of time load duration are proposed based on the models introduced
in chapter 4. For the bending strength, a hierarchical spatial variability model is proposed and
a method is presented for linking the properties of a cross section (which is considered as the
reference starting point for the modelling of spatial variability) with the properties of a test
specimen. However, some of the features explored in chapters 4 and 5 are not considered as
being sufficiently verified for the implementation in an operational model code.
Within chapter 4 and 5 several examples for the discussed models are already given. In
chapter 7 some further applications such as the assessment of experimental data, updating
with new information etc. are exemplified.

8.1.3 ORIGINALITY OF WORK

The present work has one main objective; the development and the documentation of a
proposed basis for probability based design of timber structures. The content of the proposal
is mainly based on already existing knowledge. However, the attempt to combine this
knowledge into a consistent and operational format is considered as new and original. The
main features are:
x The presentation of a very basic model as a reference for simple reliability based code
calibration which is fully consistent with possible refinements.

1
e.g. with participants of the COST action E 24 (with many representatives of the timber research community)
and members of the JCSS (as experts in the field of structural reliability)

203
x Interrelationships between different material properties based on expected values,
coefficients of variation and correlations.
x The consistent treatment of material variability on different scales (micro, meso, macro).
x The illustration of how timber material properties can be updated in regard to different
types of information.
During the development of the proposal several interesting issues are explored. The results
have not all been applied in the proposed model code, but they also deliver valuable
conclusions for the further development of the understanding of timber as a structural
material. These results are:
x The development of a consistent probabilistic framework of graded timber material; the
grading process is explicitly considered and a format of communicating the relevant
information is proposed.
x Based on these results it is demonstrated how cost optimal grading strategies can be
identified.
x A framework for the calibration of load duration modification factors is presented; a
damage accumulation model based on fracture mechanical considerations is developed
further for consideration of random load processes. It is found that the model is not
sensitive to the number of load cycle repetitions when a random load process is
considered.
x Based on experimental data existing models for the capacity of dowel type fasteners are
reassessed and possible developments are discussed. The model uncertainty is quantified
based on the data.

8.1.4 LIMITATIONS

The present work mainly concerns modelling the material properties of solid structural timber
and the load bearing properties of connections. The proposed models are predominantly based
on test programs and investigations considering North American and European softwoods.
For some other softwoods, and especially for hardwood, the underlying assumptions are less
appropriate. It should also be noted that part of the quantitative information in this thesis
should be considered as indicative values. Timber connections are considered, however, all
the concern is directed to one special type of connections; connections with parallel loaded
double shear dowel type fasteners.
System effects are not addressed explicitly within this thesis. However, the proposal for the
modelling of spatial variability can be used for performing research in this direction. Another
aspect which is nearly neglected is the complex interrelation of timber material and moisture.
In this case, the physical or biological deterioration of the timber should be considered in

204
addition to its strength and stiffness behaviour.

8.2 OUTLOOK

The main outcomes of this thesis are related to necessary pre-codification modelling aspects
concerning the reliability of timber components in regard to strength and stiffness properties.
An achievement of this thesis is that the work performed is fully compatible with the general
probabilistic framework for establishing design basis developed by the Joint Committee on
Structural Safety (JCSS).
During the course of writing this thesis, however, it has become apparent that several
fundamental issues require more research and development. In particular, it is found that the
present practice in regard to strategies to quality control of timber raw material as well as
engineered timber introduces significant uncertainties regarding the performance of timber
structural components, and the uncertainties cannot be quantified based on the present control
practice. It is necessary to develop further a framework for the probabilistic modelling of
timber grading as outlined in chapter 4 and to disseminate this within code and standard
writing bodies and timber researchers.
An attempt to model the spatial variability of timber material properties is presented within
the present thesis. More experimental work should be performed or reviewed to quantify the
parameters of the presented models. The description of size effects has earned considerable
recognition in the past research; no general consensus has been reached and future research
should be directed in the development of a consistent framework for the description of size
effects in structural timber elements.
Furthermore, it has been found that model uncertainties associated with high performance
timber structural components such as glued laminated timber beams are unsatisfactorily high.
It is of importance to direct future research in this direction, because engineered timber
elements are promising products that may compete with steel in large span high performance
load bearing structures.
More than for other building materials, the properties of timber structural components and
systems have to be seen in direct combination with the loads and environment to which they
are exposed to. Adverse effects in timber structures are not only related to critical load-
strength combinations, but rather to moisture induced strength and stiffness degradation of
structural elements. The mechanism behind this is the so called mechano-sorptive creep of
timber on the one hand, but also degradation as a consequence of physical and/or biological
decay (insects/fungi) on the other hand. Several research activities in the field of moisture
effects in buildings have taken place, but the individual studies did not yet converge to a
general consensus which could form a scientific basis for proper code formats and regulations
facilitating the engineer to cope with the problem in daily design and maintenance. Therefore,
more work is needed in this respect.

205
206
9 REFERENCES
AS 3519: Standards Australia: Timber – Stress graded – Procedures for monitoring structural properties.
Sydney, Australia, 1997.
AS 4490: Standards Australia: Timber – Machine proof grading. Sydney, Australia, 1997.
ASHTO (1994): LRFD Bridge Design Specifications. American Association of State Highway and
Transportation Officials, Washington, DC.
ASTM Book of Standards – Wood (Print and CD-ROM). Volume 04.10, July 2004.
ASTM D 245: Standard practice for establishing structural grades and related allowable properties for visually
graded lumber. ASTM Book of Standards.
ASTM D 6570: Standard Practice for Assigning Allowable Properties for Mechanically-Graded Lumber.
ASTM Book of Standards.
Bach L. (1973). Reiner Weisenbergs’ theory applied to time dependent fracture of wood subjected to various
modes of mechanical loading. Wood Science, 5(3), pp. 161-171.
Barenblatt G. I. (1962). The mathematical theory of equilibrium cracks in brittle fracture. Adv. Appl. Mech. 7,
pp. 55-129.
Barrett J. D. (1974). Effect of size on tension perpendicular to grain strength of douglas fir. Wood and Fiber.
6(2), pp. 126-143.
Barrett J. D. and Fewell A. R. (1990). Size factors for the bending and tension strength of structural lumber.
Proceedings of the 23th Meeting, International Council for Research and Innovation in Building and
Construction, Working Commission W18 – Timber Structures, CIB-W18, Paper No. 23-10-3, Lisbon,
Portugal, 1990.
Barrett J. D. and Foschi R. O. (1978). Duration of load and probability of failure in wood. Part I + II. Canadian
Journal of Civil Engineering, 5(4):505-532.
Barrett J. D., Foschi R. O. and Fox S. P. (1975). Perpendicular to the grain strength of douglas fir. Canadian
Journal of Civil Engineering. 2(1) pp. 50-57.
Basler E. (1961). Untersuchungen über den Sicherheitsbegriff von Bauwerken. Schweizer Archiv für
angewandte Wissenschaft und Technik, Volume 4.
Benjamin J. R. and Cornell C. A. (1970). Probability, Statistics and Decision for Civil Engineers. McGraw-Hill
Book Company, New York, USA.
Blass H.J. (editor) (1995) STEP 1: Timber Engineering, Basis of Design, Material Properties, Structural
Components and Joints. Centrum Hout, The Netherlands.
Blass H.J., Bienhaus A. and Krämer V. (2001). Effective bending capacity of dowel-type fasteners.
Proceedings PRO 22, International RILEM Symposium on Joints in Timber Structures. pp. 71-88.
Bohannan B. (1966). Effect of size on bending strength of wood members. USDA Forest Service, Research
Paper FPL 56, Forest Products Laboratory, Madison, WI.
Bolotin V. V. (1969). Statistical Methods in Structural Mechanics. Holden-Day Series in Mathematical
Physics.
Bonfield P. W., and Ansell M. P. (1991). The fatigue properties of wood in tension, compression and shear.
Journal of Material Science. 26, pp. 4765-4773.
Boström L. (1994). Machine strength grading. Comparison of four different grading systems. Swedisch
National Testing and Research Institute, SP REPORT 1994:49.
Bronstein I. N. and Semendjajew K. A. (1981). Taschenbuch der Mathematik. BSB B. G. Teubner
Vertragsgesellschaft, Leipzig.
Caultfield D. F. (1985). A chemical kinetics approach to the duration of load problem in wood. Wood and Fibre
Science, 17(4), pp. 504-521.
CEB (1976a): First Order Concepts for Design Codes. CEB Bulletin No. 112, Munich.
CIRIA (1977): Rationalization of Safety and Serviceability Factors in Structural Codes. CIRIA Report No. 63,
London.
Clorius C. O. (2001). Fatigue in wood – an investigation in tension perpendicular to the grain. PhD Thesis,
Department of Structural Mechanics and Materials, Technical University of Denmark.

207
Coder S. E. (1965). Localised deflection related to bending strength of lumber. Proceedings of the second
symposium on non-destructive testing of wood. Washington State University, Pullman, WA.
Colling F. (1986). Influence of volume and stress distribution on the shear strength and tensile strength
perpendicular to the grain. Proceedings of the 19th Meeting, International Council for Research and
Innovation in Building and Construction, Working Commission W18 – Timber Structures, CIB-W18, Paper
No. 19-12-3, Florence, Italy, 1986.
COMPUTERMATIC “Manual” (2000). MPC Measuring and Process Control Ltd., Chelmsford, Essex,
England.
Cornell C. A. (1969). A Probability Based Structural Code. ACI-Journal, 66: pp. 974-985.
COST E 24 (2005) Reliability of Timber Structures. European Research Project, Meetings and Publications
Documentation: http://www.km.fgg.uni-lj.si/coste24/coste24.htm.
Czomch I. (1991). Lengthwise variability of bending stiffness of timber beams. Proceedings of the
International Timber Engineering Conference, London, UK , Vol.2, pp. 2158-2165.
Czomch I., Thelandersson S. and Larsen H. J. (1991). Effect of within member variability on the bending
strength of structural timber. Proceedings of the 24th Meeting, International Council for Research and
Innovation in Building and Construction, Working Commission W18 – Timber Structures, CIB-W18, Paper
No. 24-6-3, Oxford, UK, 1991.
Der Kiureghian A., Lui P.-L. (1986). Structural reliability under incomplete probability information. Journal of
Engineering Mechanics, 112(1), pp. 85-104.
Ditlevsen O. and Källsner B. (2004). Statistical series system size effects on bending strength of timber beams.
Reliability and Optimization of Structural Systems – Maes & Huyse (eds). Taylor and Francis Group,
London.
Ditlevsen O. and Madsen H.O. (1996). Structural Reliability Methods. Wiley, Chichester, UK.
Downing S. D. and Socie D. F. (1982). Simple Rainflow Counting Algorithms. International Journal of
Fatigue. 4(1), pp. 31-40.
Dugdale D. S. (1960). Yielding of steel sheets containing slits. J. Mech. Phys. Solids 8, pp. 100-104.
Ellingwood B. and Rosowski D. V. (1991). Duration of load effects in LRFD for wood constructions. ASCE J.
Struct. Eng. 117(2):584-599.
Ellingwood, B., MacGregor, J.G., Galambos, T.V. & Cornell, C.A. (1982): Probability Based Load Criteria:
Load Factors and Load Combinations. ASCE, Journal of the Structural Division, Vol. 108, N0. ST5, 1982,
pp. 978-997.
EN 1995-1-1, Eurocode 5: Design of timber structures; part 1-1: general rules and rules for buildings. Comité
Européen de Normalisation, Brussels, Belgium, 2004.
EN 26891: Timber Structures – Joints made with mechanical fasteners – General principles for the
determination of strength and deformation characteristics. Comité Européen de Normalisation, Brussels,
Belgium, 1991.
EN 338: Structural Timber – Strength Classes. Comité Européen de Normalisation, Brussels, Belgium, 2003.
EN 383: Timber Structures; test methods; determination of embedding strength and foundation values for
dowel type fasteners. Comité Européen de Normalisation, Brussels, Belgium, 1993.
EN 383: Timber Structures; test methods; determination of embedding strength and foundation values for
dowel type fasteners. Comité Européen de Normalisation, Brussels, Belgium, 1993.
EN 408: European Standard: Timber structures - Structural Timber - Determination of some physical and
mechanical properties. Comité Européen de Normalisation, Brussels, Belgium, 2004.
Ewan W. D., Kemp K. W. (1960). Sampling inspection of continuous processes with no autocorrelation
between successive results. Biometrika 47, pp. 363-380.
Faber M. H. (2003). Risk and safety in Civil Engineering. Lecture Notes, Institute of Structural Engineering,
Swiss Federal Institute of Technology, Zurich.
Faber M. H. and Sørensen J. D. (2003). Reliability based code calibration – the JCSS approach. Proceedings to
the 9th International Conference on Applications of Statistics and Probability in Civil Engineering ICASP9,
Volume 2, pp. 927-935, San Francisco, USA, July 6-9, 2003.
Faber M. H., Köhler J. and Sørensen, J. D. (2004). Probabilistic modelling of graded timber material properties.
Journal of Structural Safety, 26(3), pp. 295-309.
Foliente, G.C. (1998). Design of timber structures subjected to extreme loads. Prog. Struct. Eng. Mater., 1(3),
pp. 236-244.

208
Foschi R. O. and Barrett J. D. (1975). Longitudinal shear strength of douglas fir. Canadian Journal of Civil
Engineering, 3(2), pp. 198-208.
Foschi R. O. and Barrett J. D. (1980). Glued laminated beam strength: A model. ASCE Journal of Structural
Division 106 (ST8).
Foschi R. O. and Yao Z. C. (1986). Another look at three duration of load models. Proceedings of the 19th
Meeting, International Council for Research and Innovation in Building and Construction, Working
Commission W18 – Timber Structures, CIB-W18, Florence, Italy, 1986.
Foschi R. O., Folz B. R. and Yao F. Z. (1989). Reliability-based Design of Wood Structures. Structural
Research Series, Department of Civil Engineering, University of British Columbia, Vancouver, British
Columbia, Canada. Report No. 34.
Fridley K. J. and Soltis L. A. (1992). Load-duration effects in structural lumber: strain energy approach. J. of
Struct. Eng. ASCE, 118(9), pp. 2351-2369.
Galligan W., DeVisser D. (2004). Machine Grading Procedures under American Lumber Standard.
Proceedings of the 8th World Conference of Timber Engineering, Lathi, Finnland.
Gerhards C. C. and Link C. L. (1987). A cumulative damage model to predict load duration characteristics in
lumber. Wood and Fiber Science. 19(2), pp. 147-164.
Gerhards, C. C. (1979). Time-related effects of loads of wood strength. A linear cumulative damage theory.
Wood Science. 19(2), pp. 139-144.
Glos P. (1981). Zur Modellierung des Festigkeitsverhaltens von Bauholz bei Druck-, Zug- und
Biegebeanspruchung. Berichte zur Zuverlässigkeitstheorie der Bauwerke, SFB 96, Munich, Germany.
Glos P. (1995). Strength Grading. STEP 1: Timber Engineering, Basis of Design, Material Properties,
Structural Components and Joints. Centrum Hout, The Netherlands.
Green D. W. and Kretschmann, D. E. (1997). Properties and grading of southern pine timber. Forest Products
Journal, 47(9), pp. 78-85.
Gustafsson P. J. (1992). Some test methods for fracture mechanics properties of wood and wood adhesive
joints. RILEM TC133-TF workshop, Bordeaux, France, 1992.
Hanhijärvi A., Galimard P. and Hoffmeyer P. (1998). Duration of load behaviour of different sized straight
timber beams subjected to bending in variable climate. Holz als Roh- und Werkstoff, 56(5), pp. 285-293.
Hasofer A. M. and Lind N. C. (1974). An exact and invariant first order reliability format. ASCE, J. Eng.
Mech. Div. 1974, pp. 111-121.
Hoffmeyer P. (1990). Failure of Wood as Influenced by Moisture and Duration of Load. Ph. D. Thesis, College
of Environmental Science and Forestry Syracuse, State University of New York.
Hoffmeyer P. (1995). Wood as a Building Material. STEP 1: Timber Engineering, Basis of Design, Material
Properties, Structural Components and Joints. Centrum Hout, The Netherlands.
Hohenbichler M., Rackwitz R. (1981). Non-Normal dependent vectors in structural safety. Journal of
Engineering Mechanics, 107(6), pp. 1227-1238.
Isaksson T. (1999). Modelling the Variability of Bending Strength in Structural Timber. PhD-thesis, Lund
Institute of Technology, Report TVBK-1015.
ISO 8375: Solid timber in structural sizes – determination of some physical and mechanical properties.
International Organisation for Standardisation, 1985.
Johansen K.W. (1949). Theory of timber connections. International Association of Bridge and Structural
Engineering, Publication No. 9, pp. 249-262, Bern, Switzerland.
Johansson C.-J., Brundin, J. and Gruber R. (1992). Stress grading of Swedish and German timber. A
comparison of machine stress grading. SP REPORT 1992:23.
Johnson A. I. (1953). Strength, safety and economical dimensions of structures. Swedish State Committee for
Building Research, Bulletin n. 22, p. 159.
Joint Committee of Structural Safety (JCSS, 2001). Probabilistic Model Code, Internet Publication:
www.jcss.ethz.ch.
Jorissen A. (1998). Double shear timber connections with dowel type fasteners. PhD-Thesis, Delft University
Press, Netherlands.
Jourez B, Riboux A and Leclercq A. (2001). Anatomical characteristics of tension wood and opposite wood in
young inclined stems of poplar. IAWA (International Association of Wood Anatomists) Journal, 22, pp.
133–157.

209
Källsner B. and Ditlevsen O. (1994). Lengthwise bending strength variation of structural timber. IUFRO
Sydney, Australia.
Källsner B. and Ditlevsen O. (1997). Experimental verification of a weak zone model for timber bending.
IUFRO S5.02 Timber Engineering, pp. 389-404, Copenhagen, Denmark.
Kass A. J. (1975). Middle ordinate method measures stiffness variations within pieces of lumber. Forest
Products Journal 25(3), pp. 33-41.
Kersken-Bradley M. (1981). Beanspruchbarkeit von Bauteilquerschnitten bei streuenden Kenngrössen des
Kraftverformungsverhaltens innerhalb des Querschnitts. Dissertation, TU München.
Kersken-Bradley M. and Rackwitz R. (1991). Stochastic Modeling of Material Properties and Quality Control,
JCSS Working Document, IABSE-publication, March 1991.
Kline D. E., Woeste F. E. and Bendtsen B. A. (1986). Stochastic model for the modulus of elasticity of lumber.
Wood and Fiber Science, 18(2), pp. 228-238.
Köhler J. (2002). Probabilistic modelling of duration of load effects in timber structures using a fracture
mechanics model. Proceedings 4th International Ph.D. Symposium in Civil Engineering, fib, Munich,
September 2002.
Köhler J. and Faber M. H. (2003). A probabilistic creep and fatigue model for timber materials. Proceedings to
the 9th International Conference on Applications of Statistics and Probability in Civil Engineering
ICASP9,Volume 2, pp. 1141-1148, San Francisco, USA, July 6-9, 2003.
Köhler J., Faber M. H. (2003). A probabilistic approach to cost optimal timber grading. Proceedings of the 36th
Meeting, International Council for Research and Innovation in Building and Construction, Working
Commission W18 – Timber Structures, CIB-W18, Paper No. 36-5-2, Colorado, USA, 2003.
Köhler J., Faber M. H. (2004). Proposal for a probabilistic model code for design of timber structures.
Proceedings of the 37th Meeting, International Council for Research and Innovation in Building and
Construction, Working Commission W18 – Timber Structures, CIB-W18, Paper No. 37-104-1, Edinburgh,
UK, 2004.
Köhler J., Svensson S. (2002). Probabilistic modelling of duration of load effects in timber structures.
Proceedings of the 35th Meeting, International Council for Research and Innovation in Building and
Construction, Working Commission W18 – Timber Structures, CIB-W18, Paper No. 35-17-1, Kyoto, Japan,
2002.
Kollmann F. F. P., Côté W.A. (1968). Principles of wood science. I – Solid wood, Springer- Verlag.
Lam F. and Varoglu E. (1991). Variation of tensile strength along the length of timber, Part 1 and 2. Wood
Science and Technology 25, pp. 351-360 and pp. 449-458.
Larsen H. J. (1986). Eurocode 5 and CIB structural design code. Proceedings of the 19th Meeting, International
Council for Research and Innovation in Building and Construction, Working Commission W18 – Timber
Structures, CIB-W18, Paper No. 19-102-2, Florence, Italy, 1986.
Leicester R. H. and Breitinger H. O. (1994). Statistical control of timber strength. Proceedings of the 27th
Meeting, International Council for Research and Innovation in Building and Construction, Working
Commission W18 – Timber Structures, CIB-W18, Paper No. 27-17-1, Sydney, Australia, 1994.
Leijten A., Köhler J. and Jorissen A. (2004). Review of data for timber connections with dowel-type fasteners.
Proceedings of the 37th Meeting, International Council for Research and Innovation in Building and
Construction, Working Commission W18 – Timber Structures, CIB-W18, Paper No. 37-7-13, Edinburgh,
UK, 2004.
Leijten A., Köhler J. and Jorissen A. (2004). Review of probability data for timber connections with dowel type
fasteners. Proceedings of the 37th Meeting, International Council for Research and Innovation in Building
and Construction, Working Commission W18 – Timber Structures, CIB-W18, Paper No. 37-7-6,
Edinburgh, UK, 2004.
Lindley D. V. (1965). Introduction to Probability & Statistics. Cambridge University Press.
Madsen B. (1992). Structural Behaviour of Timber. Timber Engineering Ltd., Vancouver, Canada.
Madsen B. and Buchanan A. H. (1986). Size effects in timber explained by a modified weakest link theory.
Canadian Journal of Civil Engineering 13(2), pp. 218-232.
Madsen H.O., Krenk S. and Lind N.C. (1986). Methods of Structural Safety. Prentice-Hall, Eaglewood Cliffs,
NJ, USA.
Mattheck C. (1998). Design in nature: learning from trees. Springer, Berlin.
Melchers R. (1999). Structural Reliability: Analysis and Prediction. Ellis Horwood, Chichester, UK.

210
Meyer A. (1957). Die Tragfähigkeit von Nagelverbindungen bei Statischer Belastung. Holz als Roh- und
Werkstoff. 15(2), pp. 96-109.
Mindess S., Nadeau J. S. and Barret J. D. (1975). Slow crack growth in douglas-fir. Wood Science 8(1) pp.
389-396.
Morgan J. N. and Sonquist A. (1963). Problems in the analysis of survey data and a proposal. J. Am. Statist.
Assoc. 58, pp. 415-435.
Morlier P., Valentin G. and Toratti T. (1994). Review of the theories on long-term strength and time to failure.
Proceedings COST 508 Workshop on Service Life Assessment of Wodden Structures, Espoo, Finland.
NBCC (1980): (National Building Code of Canada), National Research Council of Canada.
Nelder J. A., and Mead R. (1965). A simplex method for function minimization. Computer Journal 7: pp. 308-
313.
NEN 5466: Kwaliteitseisen voor hout (KVH 1980). Houtsoort Europees vuren (in Dutch). Quality
requirements for timber. Species European spruce. Netherlands Normalisatie Instituut, Delft, Netherlands,
1983.
Nielsen L.F. (1979). Crack failure of dead-, ramp- and combined loaded viscoelastic materials. Proceedings to
the First International Conference on Wood Fracture, Banff, Alberta, Canada, 1979.
Nielsen L.F. (2000). Lifetime and Residual Strength of wood subjected to static and variable load – Part I +II.
Holz als Roh- und Werkstoff. Vol. 58, pp. 81-90 and 141-152.
Nielsen L.F. (2005). On the Influence of Moisture and Load Variations on the Strength Behaviour of Wood.
Proceedings of the international Conference on Probabilistic Models in Timber Engineering,
Arcachon, France 2005.
Niemz P. (2004). Physik des Holzes. Lecture Notes (in German). ETH Zürich, Switzerland.
NLGA SPS2: National Lumber Grades Authority. Standard grading rules for Canadian lumber. Vancouver,
Canada, 1998.
Norén B. (1986). Results of pair matching of wood for long-term loading. International workshop on duration
of load in lumber and wood products. Forintek Canada Corperation, Special Publication, No SP27.
OHBDC (1983): (Ontario Highway Bridge Design Code), Ontario Ministry of Transportation and
Communication, Ontario.
Pearson R. G. (1972). The effect of duration of load on bending strength of wood. Holzforschung, 26(4), pp.
153-158.
Petersson H. (1995). Fracture design analysis of wooden beams with holes and notches. Proceedings of the
28th Meeting, International Council for Research and Innovation in Building and Construction, Working
Commission W18 – Timber Structures, CIB-W18, Paper 28-19-3, Copenhagen, Denmark, 1995.
Philpot T. A., Fridley K. J and Rosowsky D. V. (1994). Energy based failure criterion for wood. J. of Materials
in Civil Engineering, 6(4), pp. 578-593.
Pöhlmann S. and Rackwitz R. (1981). Zur Verteilungsfunktion der Festigkeitseigenschaften bei kontinuierlich
durchgeführter Sortierung. Materialprüfung 23, Hanser, Munich, Germany.
prEN 14081 part 1-4: Timber Structures – Strength Graded Timber with rectangular Cross Section. Comité
Européen de Normalisation, Brussels, Belgium, 2003.
Rackwitz R. and Müller K. F. (1977). Zum Qualitätsangebot von Beton, II. Beton, 27(10), pp. 391-393.
Raiffa H. and Schlaifer R. (1960). Applied statistical decision theory. John Wiley & Sons Ltd. Chichester, UK.
Ranta-Maunus A., Fonselius M., Kurkela J. and Toratti T. (2001). Reliability Analysis of Timber Structures.
VTT Research Notes, Espoo, Finland.
Ravindra, M.K. & Galambos, T.V. (1978): Load and Resistance Factor Design for Steel. ASCE, Journal of the
Structural Division, Vol. 104, N0. ST9, 1978, pp. 1337-1353.
Riberholt H. and Madsen P. H. (1979). Strength of timber structures, measured variation of the cross sectional
strength of structural lumber. Report R 114, Struct. Research Lab., Technical University of Denmark.
Rosenblueth, E. & Esteva, L. (1972): Reliability Basis for Some Mexican Codes. ACI Publication SP-31, 1972,
pp. 1- 41.
Rosowsky D. V., Bulleit W. M. (2002). Load duration effects in wood members and connections: order
statistics and critical loads. Structural Safety 2002, 24, pp. 347-362.
Rouger F. (1996). Application of a modified statistical segmentation method to timber machine strength
grading. Wood and Fibre Science. 28(4).

211
Rouger F. (1997). A new statistical method for the establishment of machine settings. Proceedings of the 30th
Meeting, International Council for Research and Innovation in Building and Construction, Working
Commission W18 – Timber Structures, CIB-W18, Paper No. 30-17-1, Vancouver, Canada, 1997.
Rouger F. and Fewell A. R. (1994). Size effects in timber: Novelty never ends. Proceedings of the 27th
Meeting, International Council for Research and Innovation in Building and Construction, Working
Commission W18 – Timber Structures, CIB-W18, Paper No. 27-6-2, Sydney, Australia, 1994.
Sawata K., Yasumura, M. (2002). Determination of embedding strength of wood for dowel-type fasteners.
Journal of Wood Science, 48, pp. 138-146.
Sharpely R. A. (1975). A theory of crack initiation and growth in viscoelastic media (3 parts). Int. J. of Fracture
11, pp. 141-159, 369-388, 549-562.
Showalter K. L., Woeste F. E. and Bendtsen B. A. (1987). Effect of length on tensile strength in structural
timber. Forest Products Laboratory, Research paper FPL-RP-482. Madison, Wisconsin.
Sørensen, J.D. and Svenson S. (2005). Reliability Based Modeling of Moisture and Load Duration Effects.
Proceedings of the international Conference on Probabilistic Models in Timber Engineering,
Arcachon, France 2005.
Sørensen, J.D., Stang, B.D., Svensson, S. (2002). Effect of load duration on timber structures in Denmark.
Proceedings 9th International Conference on Applications of Statistics and Probability in Civil
Engineering(ICASP9), San Fransisco, USA, July 6-9, 2003, pp. 345-354.
Suddarth S. K. and Woeste F. E. (1977). Influence of variability in loads and modulus of elasticity in wood
column strength. Wood Science 20(2), pp. 62-67.
Svensson S., Thelandersson S. and Larsen H. J. (1999). Reliability of timber structures under long term loads.
Materials and Structures. 32, pp. 755-760.
Thelandersson S. and Larsen H. J. (2003). Timber Engineering. John Wiley & Sons Ltd. Chichester, UK.
Thörnquist T. (1990). Juventile wood in coniferous trees. The Swedish University of Agricultural Sciences.
Department of Forest-Industry-Market Studies. Report 10.
Timell T. E. (1969). The chemical composition of tension wood. Svensk Papperstidning, 72: pp. 173–181.
Toft-Christensen P. and Baker M. J. (1982). Structural Reliability Theory and its Applications. Springer, 1982.
Van der Put T. A. C. M. (1989). Deformation and damage processes in wood. PhD thesis. Delft University
press, Netherlands.
Walker J. C. F. (1993). Primary Wood Processing. Chapman & Hall.
Warren W. G. (1978). Recent developments in statistical quality control procedures for MSR. Proceedings of
Fourth Nondestructive Testing of Wood Symposium. Vancouver, Canada.
Weibull W. (1939). A statistical theory of the strength of materials. Royal Swedish Institute for Engineering
Research, Proceedings, 141, pp. 45.
Whale, L.R.J. and Smith, I. (1986). The derivation of design clauses for nailed joints in Eurocode 5, Proc. of
the 19th CIB W18 meeting, Paper 19-7-6.
Wood Handbook: Wood as an Engineering Material. Agriculture Handbook 71. Forest Product Laboratory,
U.S. Dep. Agric., Washington D.C., 1987.
Wood L.W. (1947). Behaviour of wood under continued loading. Eng. News-Record, 139(24), pp. 108-111.
Wood L.W. (1951). Relation of strength of wood to duration of load. Report no. 1916. Madison, USDA Forest
service, Forest products laboratory. Report no. 1916.

212
ANNEX
A1
 A CHARACTERISATION OF RANDOM PHENOMENA

In general two situations may be distinguished namely, the situation where a new probabilistic
model is formulated from the very beginning; and the situation where an already existing
probabilistic model is updated based on additional information. The formulation of
probabilistic models is normally based on frequentistic information (i.e. data from tests)
mostly combined with some physical considerations, experience and judgment (subjective
information). Building a probabilistic model in general requires the following steps:
x assessment and statistical quantification (and qualification) of available data,
x selection of a distribution function,
x estimation of distribution parameters,
x model verification,
x model updating.

A1 NON-PARAMETRIC ASSESSMENT OF AVAILABLE DATA

To reach an overview and to get a crude idea about the statistical characteristics of a sample it
is common to apply some descriptive procedures. In the following, the concept of Order
Statistics and the Quantile Plot are introduced. It is also common to use the first two sample
moments to characterise the sample.

Quantile Plot
A common method for quantifying the statistical properties of a test sample x is the non-
parametric procedure. The test data is sorted (ranked) in ascending order according to its
magnitude, as

x1 d x2 d x3  d xn (A.1)

where n is the number of test specimen in the sample and the subscript is equal to the rank of
the data value.
The quantile of the data value of rank m is

m  0.5 (A.2)
qm
n
An issue that may arise here is why the quantile is given by Equation (A.2) and not for
example just by m / n . When evaluating a certain quantile the data set is split into two groups,
an upper and a lower one. The 0.5 term in the numerator of Equation (A.2) means that the
corresponding observation xm is counted as being half in the one group and half in the other.
The method is often used as a direct graphical representation of the data in a so-called

A2
quantile plot.

Order Statistics
Let X be a continuous random variable with unknown distribution function F x and the
probability density function f X x . n independent realisations of X are ordered such that
x1 d x2 d x3  d xn . The observation of rank m , xm is a realisation of the random variable X m .
X m is determined from the observation that there are m  1 observations not larger and
n  m not less than X m (considering an arbitrary number of samples of size n ). The
probability density function of X m can be expressed as (Madsen et al (1986)):

n! m 1 nm
f Xm x FX x ª¬1  FX x º¼ fX x (A.3)
m 1 ! n  m !

Since the distribution model underlying X is not known, Equation (A.3) might have no
further proposition. But if the function FX x itself is considered as a stochastic variable,
instead of the variable X , the mean and the variance of FX xm can be assessed by rewriting
Equation (A.3) as:

§n· m 1 nm
f X m xm dxm m ¨ ¸ FX xm ª¬1  FX xm º¼ dFX xm (A.4)
©m¹
The expected value of FX xm is obtained by integration over the probability mass:

f
E ª¬ FX xm º¼ ³F
f
X xm f X m xm dxm

§n·1 m nm
m ¨ ¸ ³ FX xm ¬ª1  FX xm ¼º dFX xm
© m¹ 0
(A.5)
§ n · n  m !m !
m¨ ¸
© m ¹ n 1 !
m
n 1
Accordingly the variance VAR ª¬ FX xm º¼ can be calculated as:

2 2
VAR ª¬ FX xm º¼ E ª FX xm º  E ª¬ FX xm º¼
¬ ¼
1
E ª FX xm º¼ 1  E ª¬ FX xm º¼
n2 ¬ (A.6)

m § m ·
¨1  ¸
n  2 n 1 © n 1 ¹

A3
Equation (A.5) means that the expected value of the distribution function FX xm evaluated
at the observation of order m is equal to m n  1 ; i.e. independent of the type of the
distribution function.

Quantile Predictions
For quantile predictions the viewpoint is somehow different. Of interest is now the probability
of exceeding the observation of order m from a sample of n independent observations M
times in N future trails. Following Madsen et al. (1982) it can be shown that this probability
is given as:

§ N  M  m  1· § n  m  M ·
¨ ¸¨ ¸
© m 1 ¹© M ¹
p n, m; N , M (A.7)
§n N ·
¨ ¸
© n ¹
For example the probability that one further observation of the random variable X is not
exceeding xm is

m
p n, m;1, 0 (A.8)
n 1
The probability that no of N further observations will exceed the maximum xn is

n
p n, n; N , 0 (A.9)
N n
Equation (A.7) can be utilised to estimate fractile values of a random variable without any
assumptions of the distribution model of the variable.

A2 SELECTION OF A DISTRIBUTION FUNCTION

A classical approach to select a proper distribution function for the representation of a random
quantity is to postulate a hypothesis for the distribution family, then estimate the parameters
for the selected distribution function on the basis of observations and finally perform
statistical tests to verify the hypothesis. Statistical tests are, e.g., the F 2 - goodness of fit test
the Kolmogorov-Smirnov goodness of fit test or when comparing two hypothesises e.g. the
likelihood ratio test. (see e.g. in Benjamin and Cornell (1971))
However, a quite big number of probability distribution functions exist, and it might not be
reasonable to check all in regard to their validity. Furthermore, in most cases the amount of
frequentistic information is limited and the statistical tests may thus lead to false conclusions.
In many cases there are some physical considerations paired with experience which lead to
the identification of a limited number of possible distributions. In practice the following

A4
strategy is quite common:
First consider the physical arguments why a quantity may belong to one or the other
distribution family.
Then check whether the statistical evidence (frequentistic information) is in gross
contradiction with the assumed distribution.
E.g. if the bending strength of timber is considered, the following possible distribution models
are selected for the following reason:
x Normal distribution: additive combination of random phenomena – traditionally used for
ductile material failure. Probability for negative realisations.
x Lognormal distribution: multiplicative combination of random phenomena. Only positive
realisations.
x 2-parametric Weibull distribution: extreme value distribution. Traditionally used for
brittle material failure.
The distribution functions are introduced in Table A1.
A convenient method to check the validity of a chosen distribution model is the use of
probability paper. Probability paper for a given distribution model is constructed such that the
cumulative probability density function for that distribution model will have the shape of a
straight line when plotted on the paper. A probability paper is thus constructed by a non-linear
transformation of the x and the y axis. The transformations for the Normal, the Lognormal
and the Weibull distribution are given in Table A2.

Table A2 Axis transformation for the construction of probability paper.

x’-axis y’-axis

Normal x ) 1 F x

Lognormal ln x ) 1 F x

Weibull ln x ln  ln 1  F x

A5
Table A1 Normal, Lognormal and Weibull distribution functions.
Parm.
Density / Distribution Function Range Restri Mean Standard Deviation
ct.

1 ª 1 § x  m ·2 º
fX x exp «  ¨ ¸ »
2SV «¬ 2 © V ¹ »¼
1 § xm·
Normal

M¨ ¸ f  x  f V !0 m V
V © V ¹
§ xm·
FX x )¨ ¸
© V ¹

§ ln x ·
1 ¨ [ ¸
fX x M¨ ¸ §G2 ·
xG ¨ G ¸ G !0 [ exp ¨ ¸
Lognormal

© ¹ 0 xf © 2 ¹
[ !0 P exp G 2  1
§ ln x · [ median
¨ [ ¸
FX x )¨ ¸
¨ G ¸
© ¹
k 1
k§x· ª § x ·k º
fX x ¨ ¸ exp «  ¨ ¸ »
w© w¹ «¬ © w ¹ »¼ w!0 § 1· § 2· § 1·
Weibull

0 xf w* ¨1  ¸ w * ¨1  ¸  * 2 ¨1  ¸
ª § x ·k º k !0 © k¹ © k¹ © k¹
FX x 1  exp «  ¨ ¸ »
¬« © w ¹ ¼»

A3 ESTIMATION OF DISTRIBUTION PARAMETERS

The Method of Moments

Let the density function of a random variable X be given as f X and x be every possible
realisation of X , then the m th moment of X mX( m ) is given as:

f
mX ( m ) E ª¬ X m º¼ ³x
m
f X x dx with m 2,3, 4,5... (A.10)
f

The m th central moment is given as:

f
m
E ª x  m(1) º
( m) m
m X ,c ³ x  PX f X x dx with m 2,3, 4,5... (A.11)
«¬ X »¼
f

The characteristic of a random variable X can be expressed with the moments. E.g. the mean

A6
value P X and the variance V X2 of X can be derived, based on the first and second centralized
moment respectively:

f
PX mX (1) ³xf X x dx (A.12)
f

f
2 2
V X ³ x  PX f X x dx (A.13)
f

Moments have proven to be useful numerical descriptors of the characteristics of random

variables, i.e. a random variable can be completely described through its moments. They are
also used to characterise the statistical properties of samples. The sample mean value and the
sample variance can be derived based on the first and second centralized moment
respectively. Considering a sample with n realisations of the variable X the sample mean
(first sample moment) and the sample variance (second sample moment) are given in
Equation (A.14) and (A.15) correspondingly.

1 n
x ¦ xˆi
ni1
(A.14)

1 n 2
s2 ¦ xˆi  x
ni1
(A.15)

Note that these are sample descriptors and independent of any assumption of a possible
distribution function of X .
Similar as the sample mean and the sample variance, sample moments of higher order, e.g. the
sample coefficient of skewness and the sample coefficient of kurtosis can be derived from
Equation (A.10). However, different samples drawn from the same random variable X might
have different sample moments. The real moments of X are not known and they might be
considered as random variables itself. The sample moments are just realisations of these
random variables.
To capture the uncertainty associated with the sample moments the following way of thinking
might be helpful: Prior to a test nothing is known about its outcomes, it is just known that
there will be realisations of the random variable X , i.e. X 1 , X 2 , X 3 ,..., X n , which is a general
description of the sample consisting of n observations. The sample mean and the sample
variance can be written as

1 n
X ¦ Xi
ni1
(A.16)

A7
 1 n 2
S2 ¦ Xi  X
ni1
(A.17)

With some further calculus (see, e.g. Benjamin and Cornell (1970)) the expected value and
the variance of the sample mean and the sample variance can be expressed as

1 2
E ª¬ X º¼ mX Var ª¬ X º¼ VX (A.18)
n

n 1 2 2 n 1
E ª¬ S 2 º¼ VX Var ª¬ S 2 º¼ VX2 (A.19)
n n2
The expected value of the sample mean is equivalent to the mean of the underlying random
variable X . The variation of the sample mean is given for the case where the individual
observations can be considered to be mutually independent. This is the case, e.g. if the sample
is obtained through random sampling. Also for random samples, the expected value of sample
variance is given which is not equivalent to the variance of the underlying random variable. In
other words, on the average over many different samples, S 2 will not be equal to V X2 for
which it has been introduced to serve as an estimator. Therefore S 2 is called a biased
estimator of V X2 . As the variance of the sample mean the variance of the sample variance is
also a function of the number of observations and the variance of the underlying random
variable X .
As shown the sample moments can be used to obtain estimates of the real moments of the
random variable and the uncertainty associated with this estimation can be quantified.
Obviously in practice mostly only one sample is considered, i.e. only one realisation of
sample moments is obtained. The uncertainty of this one realisation can be addressed, by
exchanging mX with x and V X with s .
Assuming the random variable X is following a probability distribution function FX the
parameters of the distribution can be estimated. However, the uncertainty associated with the
distribution type assumption cannot be assessed by the method of moments.

The Method of Maximum Likelihood

The principle of the Method of Maximum Likelihood is to find a set of parameters of an
assumed probability distribution function which most likely reflects the statistical behaviour
of the underlying set of data (sample). The format of the method can be derived by means of
T
the following considerations. Supposing that the parameters ș T1 ,...,T n of the distribution
of X are known, the joint probability distribution of a (random) sample X 1 , X 2 , X 3 ,..., X n can
be written as:

A8
 n
fX x ș f X1 , X 2 , X 3 ,..., X n x1 , x2 , x3 ,..., xn ș f X1 x1 f X 2 x2 ˜˜˜ f X n xn –f X xi ș (A.20)
i 1

In other words Equation (A.20) can be seen as a relative measure for the belief that the sample
T
X 1 , X 2 , X 3 ,..., X n belongs to the distribution function with given parameters ș T1 ,...,T n .
Obviously in general the situation is contrary. A sample xˆ xˆ1 , xˆ2 , xˆ3 ,..., xˆn is observed and the
set of distribution parameters is not known. In that context Equation (A.20) can be similarly
seen as a relative measure for the likelihood that the distribution determined by ș is reflecting
the statistical behaviour of the sample x̂ . Over the entire domain of all possible parameters ș
the likelihood L . that the parameters belong to the sample is:

n
L ș xˆ1 , xˆ2 ,..., xˆn –f
i 1
X xˆi ș (A.21)

The maximum likelihood estimators can now be defined as the set of parameters ș which are
most likely representing the set of sample values, i.e. the set of parameters which maximise
the likelihood function L . over the entire domain of ș .

ș max L ș xˆ1 , xˆ2 ,..., xˆn (A.22)

ș

To illustrate the nature of the likelihood function a set of sample values x̂ is considered and
the exponential distribution is chosen as a proper distribution function for the sample. The
probability density function for the exponential distribution is:

f xO O exp O x (A.23)

Then by given realisations of X , i.e. xˆ xˆ1 , xˆ2 ,..., xˆn the joint probability density is

n
f X xˆ O –f X xˆi O O exp O xˆ1 O exp O xˆ2 ˜˜˜ O exp O xˆn
i 1
n
(A.24)
§ ·
O exp ¨ O ¦ xˆi ¸
n
L O xˆ
© i1 ¹
which is equivalent to the likelihood function given O .
For illustration the likelihood function for a given sample is sketched in Figure A1.

A9
 L x
arbitrary scale

0.5 1.0 1.5 2.0 2.5 3.0

Figure A1 Likelihood function on O

According to Figure A1 the likelihood function is attaining its maximum at O = 1.58, the
parameter which is most likely representing the considered sample. As an intuitive
interpretation, the likelihood function might also be helpful to asses the likelihood of every
possible parameter. In Figure A1 for example 1.5 might be a more likely estimate for O than
2.5.
In general, it might be advantageous to consider the logarithm of the likelihood function. The
log likelihood is written as:

ª n º n
ln ª¬ L ș xˆ º¼ l ș xˆ ln «– f X xˆi ș » ¦ ln ª¬ f X xˆi ș º¼ (A.25)
¬i1 ¼ i 1

Then the maximum may be obtained by solving a set of m equations

n
w
¦ wTi 1
ln ¬ª f X xˆi ș º¼ 0 j 1, 2,..., m (A.26)
j

However, the solution of the system might not always be possible in a closed form. For
complex problems, computer-automated search functions like e.g. the simplex method
(Nelder and Meat (1965)) can be employed to find the values ș which maximise the
likelihood function.

Properties of the Maximum Likelihood estimators

As for the method of moments the estimators are, before the experiment, random variables
and can be studied as such. Their properties are well known and intensively studied in the
literature (Lindley (1965)). For large samples ( n ! 15 ) the estimators ș are approximately
normal distributed (due the summation form in Equation (A.25)). Their means are
asymptotically equal to the true parameter values ș ; which means, the estimators are
T
asymptotically unbiased. The covariance matrix C for the parameters ș T1 , T 2 ,...,T n may
be obtained through the inverse of the Fischer information matrix H :

A10
CĬĬ H 1 (A.27)

with components given by:

w 2l
H ij  (A.28)
wT i wT j
ș=ș

E.g. the variance of Ti is asymptotically equal to:

1
Var ªș º 
¬ ¼ ª w 2l º (A.29)
nE « 2 »
¬ wT i ¼
Contrary to the method of moments the estimations of the distribution variables are studied in
regard to their uncertainty and not the sample moments. The main properties of the maximum
likelihood estimates are:
x For large data samples the distribution parameters ș are Normal distributed.
x Maximum likelihood estimates are consistent: For large n the estimates converge to the
true value of the parameters of the distribution.
x Maximum likelihood estimates are unbiased: For all sample sizes the parameter of
interest is calculated correctly.
x Maximum likelihood estimates are efficient: The estimate has the smallest variance.
x Maximum likelihood estimate is sufficient: All the information of the observations is
utilized.
x The solution of MLM is unique.
x The probability distribution for the problem at hand has to be known.

A4 UPDATING
When information has been collected about a quantity of interest the new knowledge implicit
in that information might be applied to improve any previous (prior) estimate of the value of
the property. For the type of information it can be differentiated between so-called equality
type and inequality type information. Equality type information is corresponding to measured
variables and inequality type information denotes the information carried with a measurement
that some variable is greater than or less than some predefined limit. It can also be
differentiated between direct and indirect information; i.e. direct measurements of the quantity
of interest and the measurement of some indicator of the quantity respectively. Examples for
the different types of information are given in Figure A2.

A11
 equality direct indirect

standard grading
tests indication
inequality

proof- status
loading inspections

Figure A2 Examples for combinations of different types of information.

The framework of doing updating is Baysian statistics, which uses the Bayes theorem, see e.g.
Ang and Tang (1975).

Updating Random Variables using Direct Information

A random quantity is represented by the random variable X with the probability distribution
T
function FX x . The parameters ș T1 , T 2 ,...,T n of the distribution function are not known
with accuracy; i.e. they are product of engineering knowledge, physical understanding or
earlier observations of the quantity. The parameters ș are in general expressed as random
variables itself specified by the so-called prior density function fĬc ș . The uncertain
parameters can be updated based on new observations of realizations of the random quantity
T
X , xˆ xˆ1 , xˆ2 ,..., xˆn .
T
The general scheme for updating of the parameters ș T1 , T 2 ,...,T n is:

fĬc ș L ș xˆ
fĬcc ș xˆ (A.30)
³ f Ĭc ș L ș xˆ dș

where fĬcc ș xˆ is the posterior distribution function of the parameters ș . The likelihood
function L ș xˆ represents the knowledge gained by the observation of x̂ . It can be
interpreted as the likelihood of observing x̂ under the assumption that ș takes its current set
of values – it may be written as p xˆ ș .
The Likelihood function is proportional to the joint conditional probability of making the
observations x̂ :

n
L ș xˆ v – f X Ĭ xˆi ș (A.31)
i 1

where n is the number of observations.

A12
Based on the updated distribution of the parameters fĬcc ș xˆ , it is possible to calculate the
predictive distribution of the random property X , as:

fĬU x ³f X x ș f Ĭcc ș xˆ dș (A.32)

Closed form solutions for the predictive and the posterior distribution can be found for special
types of probability distribution functions (known as natural conjugate distributions) in e.g.
Raiffa and Schlaifer (1961). These solutions are the analytical basis for the updating of
random variables and cover a number of distribution types which are in common use in
structural engineering. In cases where no analytical solution is available FORM/SORM
techniques (Madsen et al. 1986) can be used to integrate over the possible outcomes of the
uncertain distribution parameters.

Inference: Normal Distribution with uncertain mean and known standard deviation
T
A normal distributed variable X , with the parameters ș M X , V X is considered. The
standard deviation V X is assumed to be known and the mean value M X is considered as
random variable. It can be shown, that the natural conjugate distribution for M X is the normal
distribution, i.e. M X ~ N P c, V c is the prior distribution. n new observations x̂ of X are
made and the sample characteristics x and s are quantified.
The posterior distribution function of the mean can be given as (Ditlevsen and Madsen
(1996)):

1 § 1 § P  P cc · 2 ·
f cc P X exp ¨  ¨ X ¸ ¸¸ (A.33)
¨
2SV cc © 2 © V cc ¹ ¹
where:

Pc x VX Vc

P cc n nc n nc V X2
and V cc and nc
1 1 V c2 V c2 V c2
 
n nc n nc
The predictive distribution can then be given as:

1 § 1 § x  P cc · 2 ·
f x xˆ exp ¨  ¨
2SV ccc ¨ 2 © V ccc ¸¹ ¸¸ (A.34)
© ¹
with:

V ccc V cc2  V X2
As seen in Equation (A.33) and (A.34) the posterior distribution and the predictive
distribution are both normal which is a convenient property especially for repeated updating.

A13
Inference: Normal Distribution with uncertain mean and standard deviation
T
Considered is a normal distributed variable X , with the parameters ș M X , 6 X . The
mean and the standard deviation are regarded as random variables. An analytical solution for
the updating scheme can be found by using the natural conjugate prior for the distribution of
ș , which in this case is the Normal-Inverse-Gamma-2 distribution. The Normal-Inverse-
Gamma-2 distribution is defined through (compare e.g. Rackwitz (1983)):

f M ,6 P , h mc, sc, nc, vc f M P mc, hnc f 6 h sc, vc

1
hnc § 1 § P  mc · 2 · 1 § 1
vc1
§ 1
2 · ·
2
exp ¨  ¨ ¸ ¸ ¨ vcsc h ¸ exp ¨  vcsc2 h ¸ vcsc2 (A.35)
2S ¨ ¸ c
2 © 1 hnc ¹ 2* v 2 © 2 ¹ © 2 ¹
© ¹
f  P  f; h t 0; sc, vc, nc ! 0

where mc xc is the mean of a sample of equivalent size nc and sc is the standard deviation
of a sample of equivalent size vc  1 . The uncertain variability is expressed by the precision
h 1 V 2 . Equation (A.35) is the natural conjugate prior distribution function of the
parameters ș M X , 6 X .
By observing n sample values x̂ from X the following statistics may be drawn:

1
m x
n
¦ xˆi
1 2
s2
n 1
¦ m  xi
v n 1
For the posterior distribution function the parameters mc, sc, nc, vc in Equation (A.35) are
exchanged the parameters mcc, scc, ncc, vcc given by:

ncmc  nm
mcc (A.36)
nc  n

ncc nc  n (A.37)

2
vcsc2  ncmc2  vs 2  nm 2  nccmcc2
scc (A.38)
vc  G nc  v  G n  G ncc

vcc vc  G nc  v  G n  G ncc (A.39)

with:

A14
 ­0 for xd0
G x {® (A.40)
¯1 for x!0

The predictive distribution is of student type and is given with:

§ x  mcc ncc ·
FX xˆ x mcc, ncc, scc, vcc Tvcc ¨¨ ¸ (A.41)
© scc ncc  1 ¸¹

where TQ . is the student-t-distribution with Q degrees of freedom.

Quantification of prior information

As described and illustrated before, Baysian statistics provides a proper framework for
updating random variables and failure probabilities taking into account additional
information. The basic assumption in Baysian statistics is that prior information is always
available and can be quantified. If no information is available, the state of knowledge (or in
this case the state of ignorance) can still be quantified by so-called non informative priors
(Lindley (1965)). However, for engineering problems some prior information generally exist
and the Bayesian interpretation of probability 1 enables to introduce even vague subjective
judgement as quantified prior information. According to the confidence about the subjective
judgement the prior information can be weighted. For example consider the bending strength
of a timber supply of a sawmill which is planned to be continuously tested. It is assumed that
the supply is stationary. It is planned to draw a sample of 20 test specimens from the output
every first Monday of every month. From a neighbouring sawmill it is known that the mean
and the coefficient of variation of the bending strength are around 56 MPa and 25%
respectively. If it is believed that the supply for the considered sawmill is very similar, this
information can be utilized for the quantification of the prior distribution. Depending on how
strong the belief of similarity of the outputs of the two sawmills is, the prior information can
be weighted; i.e. the parameters nc and vc can be estimated correspondingly.
Another possibility is the quantification of prior information based on data. Consider the
example above, where the prior information of the bending strength of a sawmill timber
supply has to be quantified. Distribution parameters from several neighbouring sawmills are
available and a significant scatter within these parameters can be observed. It is assumed that
the parameters are following the natural conjugate prior distribution and that the parameters of
the supply of interest are members of this distribution. In case of the assumption of normal
distributed bending strength with unknown mean and standard deviation the conjungate prior

1
Interpretations of probability.
Bayesian: probability of A , P A , is the degree of belief that the event A occurs.
Frequentistic: Probability is understood as a relative frequency; i.e. P A nA ntotal ntotal o f

A15
 is Normal-Inverse-Gamma-2 distributed with the parameters mc, sc, nc, vc . The parameters
mc, sc, nc, vc can directly estimated based on the parameter pairs from the other supplies xi
and si with mi xi and hi 1 si2 (Rackwitz (1983)):

ĥ
xc (A.42)
h

1
§  ĥ 2 ·
nc ¨ h  ¸ (A.43)
¨ h ¸¹
©

sc h 1 2 (A.44)

§ vc · § vc ·
h \ ¨ ¸  ln ¨ ¸ (A.45)
©2¹ © 2h ¹
with:

1 k 1 k 1 k  1 k
h ¦ hi ; h ¦ ln hi ; hˆ ¦ hi mi ; h ¦ hi mi2 ;
k 1 k 1 k 1 k 1

and

§ vc · § vc · 1 1 2 16
\¨ ¸ ln ¨ ¸  c  c2    ...
©2¹ © 2 ¹ v 3v 15vc 63vc6
4

The determination of vc requires the numerical solution of Equation (A.45). However, for
sufficiently large vc (order of magnitude vc ! 6 ) vc can be approximated with

1
vc | ln h  h (A.46)

A16

View publication stats