Safety, Reliability and Risk Management – Steenbergen & van Gelder (eds)

© 2014 Taylor & Francis Group, London, ISBN 978-1-138-00123-7

Reliability based structural design

A.C.W.M. Ton Vrouwenvelder

TU-Delft, TNO, Delft, The Netherlands

ABSTRACT:  According to ISO 2394, structures shall be designed, constructed and maintained in such
a way that they are suited for their use during the design working life in an economic way. To fulfil this
requirement one needs insight into the risk and reliability under expected and non-expected actions. A key
role in this respect is played by the structural reliability analysis (SRA). In this paper the present state of
the art will be summarised, including the simplifications to semi-probabilistic calculations as being used
in daily practice. Although in principle the adopted Bayesian reliability approach should be able to take
care of all uncertainties involved, present practice still uses other safety concepts like Robustness Design
and Quality Assurance as tools for achieving the safety objectives of design and assessment.

Keywords:  Risk based decisions, structural reliability, time variant and time invariant models, structural
system reliability, uncertainty modelling, code calibration

1  Introduction 2  SRA in a Nutshell

Structural reliability analysis (SRA) comprises a The response of a structural system depends on the
set of methods and models that can be used for loading characteristics and the geometrical and mate-
the probability and risk based decision making rial properties of the structure itself. Two main cat-
with respect to design and assessment of structural egories of structural responses may be distinguished:
systems. The widest application may be found in the desired state and the adverse or undesired state.
building, civil and offshore engineering, sometimes The boundary between the two is referred to as the
explicit but in most cases in the form of so called limit state. A structure being in the undesired state
semi probabilistic procedures. is considered as having failed. Given the scatter and
The core business of SRA is the estimation of uncertainties in the various loading and structural
the lifetime (or annual) failure probability for a parameters as well as in the models, we may derive
given structure. As input the calculation pro- the probability of failure according to:
cedure requires structural behaviour models as
well as a probabilistic description of all relevant Pf = P ( g (X ) ≤ 0 ) = ∫ f ( x ) dx (1)
actions, material properties and geometrical g ( x )≤ 0 X
parameters. The establishment of these models
themselves is in fact already an essential part of where X is the vector of basic random variables,
the SRA. Failure may be defined with respect to g(x) is the limit state function for the failure mode
all kinds of structural performances, but usually it considered and fX(x) is the joint probability density
is related to collapse or to the violence of service- function of the random variables X. The limit state
ability limits. function g(x) is defined in such a way that negative
One should realise that the calculation of the values correspond to failure and positive values to
failure probability is not a goal is itself. The final non-failure.
goal is to make decisions with respect to the design Instead of the failure probability Pf the reliabil-
of new structures and inspection or repair pro- ity of the structure may also be expressed by means
grams for existing ones. This means that we also of the so called reliability index β, which may be
need to have insight into the consequences of obtained from:
failure, the corresponding risks and cost optimisa-
tion. This final goal also determines the choice of β = − Φ −1(Pf) (2)
Bayesian probability theory as the basis of SRA.
We will return to this statement later on in this where Φ (..) is the standard normal cumulative dis-
paper. tribution function. For β in the range from 1 to 4

45
the approximation Pf  = 10−β is adequate and serves In the case of n variables every calculation step
the purpose of getting a first impression of the requires (n + 1) limit state function evaluations (one
relation between these two equivalent measures of at the design point and one for finding all deriva-
reliability. tives). The number of iterations depend on the
In many SRA-techniques all random variables type of problem and the calculation method. In the
X are transformed into the so called U-space, that course of time a large variety of possible calculation
is a set of independent normal variables with zero procedures have been developed and can be found
mean and unit standard deviation. If the variables in the literature (see Rackwitz 2001). When using
Xk are independent, the transformation may be FORM, three types of problems may however be
performed by: encountered: (1) no convergence is achieved, (2)
convergence is achieved but at a local (instead of
Φ (Uk) = FXk (Xk) (3) a global) minimum and (3) convergence is achieved
but there is loss of accuracy due to high non-linear-
for every variable Xk. For a Gaussian distributed ity. The first problem at least has the advantage that
variable, the relation between Xk and Uk is simply it is recognised, the other two may go unnoticed.
given by Xk = µk + Uk σk. In case of dependent varia- In some cases it may be advantageous to use a
bles the transformation is more complex and usually mix of methods, including analytical results. As an
referred to as the Rosenblatt transformation (Rosen- example, consider a case where one or two variables
blatt, 1952, Hohenbichler and Rackwitz 1981). are known to be responsible for a heavy non-linear
The integral (1), of course, can be solved by aspect. In that case we may combine the method
straightforward numerical integration. This, how- of numerical integration and FORM by the using
ever, requires a large number of so called “limit Total Probability Theorem:
state function evaluations”. If the limit state func-
tion g( ) is a simple ‘one liner’, we may still be able Pf  = ∫P(g(U) < 0 | un)) ϕ (un) dun (5)
to handle problems with say a maximum of 7–10
random variables. For more complex limit state In this approach we may derive the conditional
functions, requiring the call of large computer probabilities P(g(U)  <  0  |  un) using FORM for a
codes, the method may hardly be feasible. An alter- series of un-values.
native is Monte Carlo simulation. This method at
least is not sensitive to the number of random vari-
ables, but also requires a large number of limit state 3  Time dependency
function evaluations in case of small probabilities.
Some special Monte Carlo techniques exist to The parameter time plays an important role in reli-
reduce this problem (Rubinstein, 1981). Examples ability engineering. Practically all statements on
are Directional Sampling, Importance Sampling, the reliability of a structure are meaningless with-
Latin Hypercube sampling, possibly in combina- out making reference to a certain period of time.
tion with response surface techniques. However, It is helpful to state this explicitly in the problem
for large systems this still may be cumbersome. statement and to reformulate (1) as:
Another way to reduce the calculation time is
provided by first and second order approximations Pf  = P(min{g(X, t)}< 0) (6)
(FORM, SORM) that have become very popular
in SRA (Hasofer and Lind, 1974; Hohenbichler The elements of X may refer to random vari-
and Rackwitz, 1983). Standard FORM is based ables as well as to time dependent random proc-
on a linearization of the limit state function in esses. The minimisation is performed over the
the U-domain and SORM provides a local second period of consideration, say 0  <  t  <  T. The time
order correction (Breitung, 1994). The method dependency may be the result of variation in the
works very well for linear or almost linear failure loads, but also of various degradation processes in
surfaces in the U-domain. FORM usually is for- the structure.
mulated as an optimisation problem. The reliabil- Loads on a structure may have short and long
ity index β, found by the method can be interpreted scale fluctuations. Slowly varying, long term fluc-
as the minimum distance in the U-space from the tuations may represent changes over hours of
origin to the limit state equation g(u) = 0: even years. In SRA the slow component is often
modelled as a time series of constant blocks (FBC
β = min || u ||,  sub g(u) = 0 (4) processes). Fast and short scale mechanical load
fluctuations are often modelled as Poisson pulse
The point in the U-space where the minimum processes or continuous Gaussian processes. The
occurs is called beta point or design point. It has short term fluctuations may cause dynamic effects
to be found using an iterative calculation scheme. in the structure. In that case spectral analysis

46
techniques are often used to deal with them. In that interval. In terms of the discrete time intervals
some load models both types of fluctuations may and limit state functions this comes down to:
be present. Figure  1  shows a schematized wind
load as the sum of a slowly changing hourly mean h(ti) = lim P(g(ti) < 0 | g(tj) > 0 for 0 < tj < ti)/∆ (9)
vale and a fast fluctuating Gaussian gust process.
Also the resistance is usually not constant in time. with ∆ = ti – ti − 1. The function h is referred to as
In most cases structural properties may deteriorate the conditional failure rate; in many engineering
under the influence of (random) mechanical, chem- mechanics applications it is often taken as a con-
ical, physical or biological influences. Typical exam- stant value: h = λ.
ples are fatigue (mechanical), corrosion (chemical) Cramer and Leadbetter (1967) showed that if
and freezing-thawing (physical) mechanisms. In the random vector process X(t) is a sufficiently
all these cases it is preferred to have physics based mixing process, the above expressions may be
models for the deterioration processes to feed the approximated by replacing the conditional failure
limit state functions in the reliability analysis. Many h( ) rate by the outcrossing rate ν( ):
deterioration processes, however, are so complex
that no physics based models exists. In that case a
classification system of structural states as indicated
in Fig. 1 could be useful. Markov models may be
{ 0
T
Pf (0,T ) = 1 − exp − ∫ v(τ )dτ } (10)
v(ti ) = lim P ( g (ti ) < 0 | g (ti −1 ) > 0 ) / ∆
applied to describe the transition probabilities from
one state to another in an empirical way. Another
Note that (10) usually is obtained by assum-
option is to formulate an empirical function for the
ing that individual outcrossings are independent
conditional failure rate h (see next section).
events. If failure at the start of the period (t = 0) is
One popular option to elaborate (6) is by dis-
also taken into account we arrive at:
cretizing the time axis and looking to the survival
probability:
Pf (0,T ) = Pf (0 )
  (11)
+ [1 − Pf (0 )] 1 − exp  − ∫ v(τ )dτ 
T
1-Pf  = P(g(t1) > 0 ∩ g(t2) > 0 ∩ ….)
= P(g(t1) > 0) Π{P{g(ti) > 0 | ∩ g(tj) > 0} (7)   0 

where i = 2 … n, j runs from 1 to (i - 1) and n is the The outcrossing approach is quite accurate in
number of time intervals. Within one time interval, case of ergodic processes. So when substantial non
one may consider the g-functions as being constant ergodic elements are present (e.g. the resistance)
in time. By taking load parameters on their maxi- one can better take care of them separately. The
mum and resistance parameters on their minimum same holds for slowly varying processes. So, for
an approximation on the safe side is obtained. The instance, in the case of windloading we may have:
result of (7) can be further developed into:
Pf  = ER{Pf(0) + (1 - exp n EQ(1 - exp(- ∫νRQ) dτ))}
{ T
Pf (0,T ) = 1 − exp − ∫ h(τ )dτ
0 } (8) (12)

where ER and EQ indicate expectations over the

where h(τ) dτ is the probability to fail in the inter- constant variables R and the intensities Q of the
val (τ, τ + dτ ), conditional upon no failure before slowly varying processes respectively and νRQ is the

Figure 1. Example of the sum of an FBC Block process Figure  2.  Relative frequency of structural conditions
and a Gaussian process. states as function of time (Kallen and Noortwijk, 2005).

47
outcrossing rate of the fast fluctuating gust proc- In the linear elastic case the focus is on the deriva-
ess conditional upon R and Q. The integration tion of mean values and the covariance matrix of
runs over the duration Ts of an individual FBC the response (displacements, strains and stresses).
block (single storm, sea state..) and n = T/Ts is the In fact this is comparable to the mean value approx-
number of FBC-blocks. imation of the FORM family. Suitable methods
based on perturbation techniques are outlined in
(Sudret and Der Kiureghian, 2000, Chakraborty
4  System Failure and Bhattacharyya, 2002). In the Neumann series
expansion (Haldar and Mahadevan, 2000) the ran-
In reliability theory a system failure is a failure dom stiffness matrix is split up into its mean and
defined as a combination of events and condi- deviatory parts as:
tions, using logic operators like AND and OR.
Structural system behaviour is often represented K = K0 + Kθ (14)
using elementary parallel and series models, fault
trees, event trees, failure trees, Bayesian Networks, where K0 is the stiffness matrix associated with the
and so on. mean value of the structural properties and Kθ is
The distinction between a single mode failure the stiffness matrix which contains deviatory part
and a system failure however is often very subtle. of the random properties. The resulting displace-
Consider for instance a series system of two ele- ment vector w may be found from a recursive solu-
ments 1 and 2, for which we have: tion of:

Failure ≡ {g1 < 0 ∪ g2 < 0} (13) K0 w(i) = F - Kθ w(i-1) (15)

One may now simply introduce the function In deriving the solution one should keep track
g = min (g1 < 0, g2 < 0) and claim that this is a sin- of the correlation between the (random) load vec-
gle mode limit state function. It would not help tor F and displacement vector w.
to “forbid” the use of “min” of “max” operators, For the nonlinear structural behaviour a variety
because we can easily produce analytical functions, of options exist. (Morutso, (1983), Gierlinski et al
continuous and infinitely many times continuously (1991), Guiterez (1999)). Very popular is the use of
differentiable, but showing effectively the same the Response surface method (Bucher and Bour-
behaviour as (13). gund, 1990): the idea is that the limit state func-
Monte Carlo and Numerical Integration can tion is generated for a number of selected points
easily take system complications on board, but in the u- or x-space. One option is to set up the
FORM-procedures are not giving appropriate response surface first and do the reliability calcu-
results in the case of strong system effects, hid- lations afterwards. Another option is to develop
den or not. In those cases one needs to address the surface as a part of the reliability procedure
the system effects explicitly. The general proce- and to adapt the set of selected points in order to
dure is that as a start for all individual modes get better and more accurate results. An advanced
a FORM analysis is performed, resulting in a and interesting calculation scheme by mixing
vector of reliability indices βi and a matrix of Directional Sampling and an Adaptive Response
influence coefficients αij. Based on these results Surface Technique (DARS) has been proposed by
a treatment of the system effects is possible. Waarts (2000). He proved that the total number
Methods for this type of System Analysis are of actually needed limit state function evaluations
widely available in the literature (Stevenson and may be as low as 7n, where n is the number of ran-
Moses 1973, Ditlevsen 1979, Hohenbichler and dom variables.
Rackwitz 1983). One of the advantages of such Clearly, at the operational edge of SRA is the
an explicit system procedure compared to direct combined nonlinear dynamic analysis. The spec-
calculation is that valuable intermediate results tral approach is usually restricted to linear sys-
become available. tems but linearization techniques for nonlinear
In the case of continuum structural systems frequency domain analysis are available (Robert
(beams, plates, slopes) the notion of one or more and Spanos, 1990, Schuëller et  al, 1991). Also
dimensional random fields enters the game. For time domain analysis is an option, again in com-
relative simple problems the outcrossing approach bination with FORM or Monte Carlo. In the lat-
mentioned before, but then in space domain, may ter case an interesting reduction in calculation
be of value. For more complex systems the stochas- time may be obtained if use is made of so called
tic FEM (Finite Element Method) may be applied. Constrained Simulation (Harland et  al, 1999),
Many publications exist, below we give a short which may be conceived as a kind of importance
overview on the basis of (Karadeniz et  al, 2004). sampling.

48
5  Inspection and maintenance

Once a structure has been built, it can be inspected.

In order to combine the data from the measure-
ments with the original data, Bayesian updating
procedures can be used. In principle two proce-
dures for updating are available (JCSS, 2000). In Figure 4.  Decision–event tree for inspection planning.
the first procedure one updates the probability
density distributions of all random variables and updated reliability index after inspecting a crack
then recalculates the probability of failure. One in a steel structure, both as functions of time. It
may also directly update the probability of failure. is interesting to see that the reliability index first
The first procedure is more informative, but some- increases, because of reduced uncertainty, but
times more complex. later on (in this example) becomes lower than the
An essential point in the updating procedure is original one, because of a somewhat disappointing
the reliability of the inspection method itself. First measurement result.
of all there is the Probability of Detection (PoD), The ultimate test is the proof load. Although it
for instance in the case of inspecting fatigue cracks looks an easy concept, it is recommended to con-
in steel structures. For many techniques PoD sider carefully the uncertainties still present.
curves are available but not for all. At present this Performing inspections and processing the
is a typical omission in the reliability assessment results is a costly matter. So one should find out
of existing structures. A second important piece which inspections are worth the effort and which
of information is the accuracy of the inspection are not. The theoretical tool to support those deci-
method: if we measure a fatigue crack of 3  mm, sions is the so called Preposterior Bayesian Analy-
could in reality this also be 2  mm or 4  mm? The sis (JCSS, 2000). Preposterior means that we try
same holds for most non-destructive test methods to find out beforehand whether the result will be
estimating the strength of materials. profitable or not. As the outcome of the inspection
Apart from explicit inspections, all kinds of beforehand is random, we can only optimise on the
observed behaviour of the structure should be basis of expectations. The idea is often presented
included in the updating analysis like permanent in the form of an event-decision tree as in Figure 4.
deformations, settlements, cracks, loose elements, We first have to choose an inspection plan or strat-
corrosion and so on. It should be realized that these egy “s”. Once we have chosen our strategy, nature
phenomena may require a double treatment in the will come with an inspection result “ω”. Given this
SRA: (1) it directly affects the structural model as inspection result we have to choose an action “a”
such (e.g. a cracked beam instead of a non-cracked (do nothing, repair option A, B, C.., reduce the
beam as a starting model) and (2) a change in the use, demolish the structure, etc.) and finally this
statistical models of possibly all random variables: may result in costs (or utilities) “u” depending
the presence of a crack might for instance may be on the state of nature “θ”. For instance, we may
an indication that the load is higher than originally choose to check some strength parameter, decide
expected. (in case of a positive result) to do nothing, but have
Given the extra data we may update the prob- a failed structure in the end after all. Taking the
ability of failure or the reliability index beta. expectations over all random outcomes (ω,θ) and
Figure  3  gives an example of the original and the optimum over all decisions (s,a) we arrive at
the optimal strategy. In practice this might prove
to be a tedious procedure, but even when using a
more pragmatic procedure, it always pays to keep
the principles in mind. A simple example is: try in
advance to find out which outcome is necessary to
change a decision and estimate how likely such an
outcome is. The profit then should outweigh the
product of inspection costs and the probability of
a good outcome.

6  Types of Uncertainty
Figure 3. Effect of (fatigue) inspection at year 10 on the
annual reliability index beta (example with disappointing In the previous sections we discussed the elabora-
inspection result leading to a lower reliability index after tion of equation (1). The structural models and
some time). probability distributions were assumed to be given

49
and correct. But, how sure can we be about these If we move to Column 4 of Table  1, we see a
models themselves. We might easily forget some list of human influences, actions that are not acci-
important action or mechanism and we may have dental but deliberate. These actions are extremely
very limited technical and statistical information. difficult to model as in general the aim is to get
For the sake of discussion we will split the haz- a load higher than the resistance. We will not dis-
ards (actions or mechanisms) into the following cuss them here. Finally the last column shows the
three categories: various types of human errors. Methods of human
reliability analysis (HRA) are under progress but
- Foreseeable and dealt with.
a meaningful interaction between SRA and HRA
- Known in principle, but unrecognized or ignored.
is still far away. In structural design procedures
- Unknown or unforeseeable.
Quality Assurance and SRA, until now, live next
Table 1 gives an overview of the category of fore- to each other a separate life.
seeable hazards in structural engineering. The list In addition to the uncertainties and errors in the
is not claimed to be complete, if such completeness treatment of the known and recognised hazards
would ever be possible. In principle similar lists for we must face the fact that certain phenomena may
degradation and collapse mechanisms exist on the be completely overlooked or, until now, objectively
resistance side. Information on the most relevant unknown (the so called black swans). Let us first
(deterministic) load and resistance models can be observe that the completely black swans are very
found in codes and text books. Information on the seldom. More important, but essentially not very
statistical modelling of the variables in those regu- different, are the forgotten, neglected or under-
lar types of loads and resistances may be found in estimated hazards or mechanisms. To deal with
the literature or data bases and in particular in the those issues the notion of robustness has been
JCSS Probabilistic Model Code (JCSS, 2001). developed.
However, all models are approximations of Technically speaking, robustness is related to
reality for which reason we have to introduce the scenarios where due to unintentional or unforeseen
notion of so called model uncertainties and statis- exposures the resistance of the structural system
tical uncertainties Given the usually present sub- has been reduced. An illustration is presented in
stantial lack of substantial information to come Figure 5 (from Eurocode EN 1991-1-7). Due to an
to objective quantifications, a frequentistic prob- exposure of any nature (a), local damage (b) may
ability in SRA interpretation is not meaningful. occur. Given the direct or local damage the struc-
Therefore, in SRA the Bayesian intuitive degree of ture may survive or (partly) collapse. Robustness
belief approach is generally adopted as the basis to requirements are especially related to the step from
step forward (Ditlevsen, 1988). The advantage of (b) to (c), i.e. to avoid that a local damage, regard-
this approach (above other options) is that treating less its origin, develops to total collapse (Faber,
aleatory and epistemic uncertainties in the same 2011).
way opens the possibilities to combine them in a Estimating some reasonable number for
coherent probabilistic decision making processes. the pattern and probability of the initial dam-
Further-more, when new or additional informa- age may give the opportunity to bring robust-
tion becomes available, it can easily be incorpo- ness design again within the scope of structural
rated using the concepts of Bayesian updating. optimisation.

Table 1.  Overview of foreseeable actions.

Normal loads
(including tail values) Accidental/natural Accidental/manmade Human influences Human Errors

Self-weight Earth-quake Internal explosion Vandalism Design error

Imposed loads Landslide External explosion Demonstrations Material error
Car park loads Hurricane Internal fire Terrorist attack  Construction error
Traffic Tornado External fire Misuse
Snow Avalanche Impact by vehicle etc. Lack of maintenance
Wind Rock fall Mining subsidence Miscommunication
Hydraulic High ground-water Environmental attack
Flood
Volcano eruption

50
Figure  5.  Illustration of the basic concepts in
robustness.

7  Targets and optimisation

According to ISO 2394, structures shall be designed,

operated, maintained and decommissioned such as
to support societal functionality. Mathematically
expressed one may state that one should aim at the
Figure  6.  Schematic relationship between the annual
optimisation of the total benefits, incorporating all and lifetime failure probability Pf and the design work-
relevant socio-economic cost items. In the elemen- ing life (from Bigaj, 2013).
tary case one may search for the minimum of the
sum of direct building costs and the risk:
optimal annual failure probability is high for short
Ctot = Cb + Pf Cf (16) periods and low for longer design lives. Spending
money on safety makes sense if one can have a
Such an optimisation may be performed over the longer period to profit from it.
intended lifetime of the structure, but one should When human safety is at stake, one should add
keep in mind that the design lifetime of the struc- an amount of money to the damage costs (either a
ture may be an object of optimisation in itself. formal compensation value or some real economic
Note that although the crisp limit state concept value). Techniques exist to make proper estimates
is very dominant in structural analysis (both deter- (Nathwani, J. et al, 2009). However, there may also be
ministic and probabilistic), one may also think of ethical or legal reasons to consider limits on human
applications where the distinction between desired safety. Usually they will be formulated as a maximum
and undesired states is more gradual/diffuse or acceptable value per year. So, the target curves may
where the consequences of failure depend heavily be presented as in Figure 6: for short design periods
on X. In those cases we may want to combine the the annual value form human safety is governing, for
probability and risk calculations, leading to: longer periods the economic optimum.
Eurocode EN 1990  gives a standard value of
βt  =  3,8 (Pft  =  8  ⋅  10−5) for a design working life
Ctot = Cb + ∫ C ( x ) fx ( x )dx (17) Td = 50 a. The target value of 3,8 is raised to 4.3
for high failure consequences and lowered to 3,3
In many cases the function C(x) is so complex for less consequences. The value does not depend
that FORM methods for reliability analysis can- on the structural costs involved to reach a higher
not be applied and only the more time consuming safety level as theory demands. Also no informa-
Numerical integration or Monte Carlo will work. tion is given as to which criterion (economy of
One interesting exception is if C is a monotonically human safety) is the governing one and no guid-
increasing function from zero to one. In that case ance on what to do for shorter or longer design
the integral can be interpreted as a convolution life time. This indicates that in the world of code
and a corresponding artificial limit state function makers this is still an unsolved problem.
can be formulated, enabling FORM to solve the
case (Gollwitzer, 2004).
Optimising the costs for a given fixed design 8  Codification using
lifetime leads approximately to the result that the semi-probabilistic methods
optimal life time failure probability is independ-
ent from the length of the design working life. Most codes of practice offer the reliability
(Holický, 2012). This means that the economically requirements in the form a Partial Factor (or

51
similar) format, where well defined characteris- 9  Closure
tic values for loads and resistances are combined
with corresponding partial factors, combina- In the course of the past half century, Structural
tion factors, importance factors, and so on. To Reliability Analysis has developed into an impres-
some extent these factors have been calibrated sive set of computational techniques. High speed
to consequence dependent reliability targets. computers enable the use of these techniques for
This is also referred to as the semi-probabilistic quite realistic structures, in particular in a research
method. environment. Daily practice still uses primarily
The partial factor method is based on the fol- semi probabilistic methods, but good calibration
lowing assumptions: procedures, linking safety factors to reliability
requirements, are available.
- strength and loads are independent random var-
Although much progress has been made, there
iables with known distributions;
still remains important research and development
- characteristic values of strength and loads (Xk)
work to be done. To start with, an increased level
are defined as specified fractiles of the respective
of user friendliness of presently available computer
distributions.
codes is necessary in order to reach a larger group
- uncertainties are taken into account by trans-
of engineers. Reliability methods should be cou-
forming characteristic values (Xk) into design
pled to standard engineering calculation software
values (Xd), by applying partial factors γ (mul-
in the same way as is available for partial factor
tiplying with load factors, dividing by material
methods. Consider by way of example an engineer
factors) or additional elements for geometrical
who has assessed his structure using the standard
properties;
code requirements and found that some of the unity
- the assessment of safety is considered as suffi-
checks were not satisfactory. In the ideal case it
cient if the design action effects do not exceed
should then be not more than a relatively simple job
the design strengths.
to start up a reliability calculation for the same struc-
A relatively simple way (JCSS, 1996) of deriv- ture and the same limit states. The reliability analy-
ing partial factors is by first calculating the design sis should use the same structural (FEM) model and
or beta point values for each variable according to automatically transform the semi-probabilistic input
(assuming normal distributions): into input formulated in terms of means, standard
deviations and correlation patterns. The use of data
Xd = µ(1 - α βt V) (18) bases like the JCSS Probabilistic model code could
be of help. The user, of course, should select meth-
where µ is the mean value, V is the coefficient of ods and strategies of calculation, but even there an
variation, βt the target reliability index for struc- expert system could give valuable suggestions.
ture and α the averaged FORM sensitivity factor Next to such integration of semi—and full prob-
(-1  <  α  <  1); averaging is over a well-chosen set abilistic methods, the probabilistic methods them-
of representative structural elements. Given for the selves still require improvements, both from the
same variables the characteristic value Xk, theoretical as from the operational point of view.
The necessary research involves a wide spectrum
Xk = µ(1 - k V) (19) of topics, like model uncertainties, degradation
processes, fire safety, inspection and monitoring,
(usually with k  =  0 for actions and k  =  1.64 for repair, failure consequences and risk estimates. But
material properties) the partial factor for resistance also more fundamental issues like human safety
respectively actions follow from: considerations, cost optimisation, robustness
requirements and interaction between SRA and
γ = Xk/Xd  or  γ = Xd /Xk (20) Quality Assurance will ask in the years to come
our full attention.
If loads of different types are involved several
load combinations may have to be checked.
In the case of non-normal distributions the REFERENCES
formulas become more complex, but the principle
remains the same. More advanced methods for Bigaj-van Vliet, A. & Vrouwenvelder A, (2013), Reliabil-
receiving partial factors exist. One may for instance ity in the performance-based concept of the fib Model
minimize the sum of (β - βtarget)2 for a large set of Code 2010, submitted for publication in Structural
structures in the area of application (Faber and Concrete.
Breitung, K. (1994). “Asymptotic approximations for
Sorensen, 2003) or require that the averaged failure multinormal integrals”. Journal of Engineering
probability is less than the target. Mechanics, ASCE 117(3):457–477.

52
Bucher, C.G. & Bourgund, U. (1990). “A fast and effi- JCSS (2000), Joint Committee on Structural Safety,
cient response surface approach for structural reliabil- Assessment of Existing Structures, RILEM Publica-
ity problems”, Structural Safety, 7, 57–66. tions S.A.R.L.
Chakraborty, S. & Bhattacharyya, B. (2002). “An efficient JCSS (2001), Joint Committee on Structural Safety, “Prob-
3D stochastic finite element method.” International abilistic Model Code”, Internet [Link], ch.
Journal of Solid and Structures, 39, 2465–2475. Kallen, M.J., Noortwijk, J.M. van., Optimal maintenance
Cramér, H. & Leadbetter, M.R. (1967), Stationary and decisions under imperfect inspection. Reliability Engi-
Related Stochastic Processes—Sample Function neering and System Safety, 90(2–3):177–185.
Properties and Their Applications, Wiley. Karadeniz, H. (1989). Advanced Stochastic Analysis
Ditlevsen, O. (1979), Narrow reliability bounds for struc- Program for Offshore Structures, Report, Dept. of
tural systems, Journal of Structural Mechanics 7 (4). Civil Engineering, Delft University of Technology,
Ditlevsen, O. (1988), Uncertainty and Structural Reli- August, Delft.
ability, Hocus Pocus or Objective Modelling, ABK, Karadeniz, H. et al. (2004). “Overview reliability methods”,
DTU, Lyngby, Denmark. SAFERELNET Report: SAF-R5-1-TUD-01(6).
Faber, M.H. & Sorensen, J.D. (2003). The JCSS Recom- Murotsu, Y. (1983). “Combinatorial properties of iden-
mended Practice For Reliability Based Code Calibra- tifying dominant failure paths in structural systems.”,
tion, ICASP, San Fransisco, USA. Bulletin of the University of Osaka Prefecture, Series
Faber, M. et al. (2011 ) Robustness of Structures, Final A, Vol. 32, No. 2.
Report of COST Action TU0601, Czech Technical Nathwani, J. Lind, N. & Pandey, M. (2009) Engineering
University in Prague (Publisher), ISBN: 978-80-01- Decisions for Life Quality: How Safe Is Safe Enough?
04803-0. Springer Verlag, London, isbn = 978-84882-601-4.
Gierlinski, W.S. et al. (1991), RASOS Theoretical man- Rackwitz, R. (2001). “Reliability analysis - a review and
ual, WS Atkins. some perspectives.” Structural Safety, 23, 365–395.
Gollwitzer, S. (2004), Private Communication. Robert, J.B. & Spanos, P.D. (1990). Random vibration
Guiterrez, M.A. (1999) “Objective simulation of failure and statistical linearization, John Wiley & Sons Ltd.,
in heterogeneous softening solids”, Thesis Delft Chichester, UK.
University. Rosenblatt, M. (1952). “Remarks on a multivariate trans-
Haldar, A. and Mahadevan, S. (2000). Reliability Assess- formation.”, The Annals of Mathematical Statistics,
ment using Stochastic Finite element Analysis. John Vol. 23, pp. 470–472.
Wiley & Sons, Inc. New York. Rubinstein, R.Y. (1981), “Simulation and the Monte
Harland, L., J. Vugts, P. Jonathan, & P. Taylor (1996), Carlo Method”, Wiley, NewYork.
“Extreme response of non-linear dynamic systems Stevenson, J. & F. Moses (1970), Reliability Analy-
using constrained simulations”, Proceeding of the sis of Frame Structures, Journal of Structural
Conference of Offshore Mechanics and Artic Engi- Division, ASCE, Vol. 96, ST11, November 1970,
neering, Florence, ASME. pp. 2409–2427.
Hasofer, A.M. & Lind, N.C. (1974) “An exact and invari- Schuëller, G.I., Pradlwarter, H.J. & Bucher, C.G. (1991)
ant second-moment code format.” J. Engrg. Mech. “Efficient Computational Procedures for Reliability
Div., ASCE, 100(1), 111–121. Estimate of MDOF-systems”, Int. Journ. Non-Linear
Hohenbichler, M. & Rackwitz, R. (1981). “Non-normal Mechanics, Vol. 26, No. 6, pp. 961–974.
dependent vectors in structural safety.” J. Engrg. Sudret, B. & Der Kiureghian, A. (2000). “Stochastic
Mech. Div., ASCE, 107(6), 1227–1238. Finite Element Methods and Reliability: A state-of-
Hohenbichler, M. & R. Rackwitz. (1983), First-order the Arts Report.” Report No. UCB/SEMM-2000/08,
Concepts in System Reliability, Structural Safety (1) Department of Civil & Environmental Engineering,
Holický, M. (2012), Optimisation of the target reliabil- University of California, Berkeley.
ity for temporary structures, Civil Engineering and Thoft-Christensen, P. & Murotsu, Y. (1986). Application
Environmental Systems, 1–10, Taylor and Francis, of structural systems reliability theory, Springer-Ver-
London. lag, Berlin.
ISO 2394 (1998),”General Principles on reliability for Waarts, P. (2000) Structural reliability using Finite Ele-
structures”, ISO, Geneve, Switzerland. ment Analysis, Ph.D. thesis, Delft University Press,
JCSS, (1996), Joint Committee on Structural Safety, Delft, The Netherlands.
Background Eurocode Part 1 Basis of Design, ECCS
publication 94, Brussels, ISBN 92-9147-0022.

53