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

net/publication/336141107

Bayesian Reliability Assessment with Spatially Variable Measurements: The

Spatial Averaging Approach

Conference Paper · May 2019

DOI: 10.22725/ICASP13.333

CITATIONS READS

3 152

4 authors:

Sebastian Geyer Claus Kunz

Technische Universität München Bundesanstalt für Wasserbau
15 PUBLICATIONS   66 CITATIONS    26 PUBLICATIONS   68 CITATIONS   

SEE PROFILE SEE PROFILE

Daniel Straub Iason Papaioannou

Technische Universität München Technische Universität München
256 PUBLICATIONS   4,018 CITATIONS    93 PUBLICATIONS   1,205 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

SAFEPEC - Innovative, risk-based inspection for a smarter and safer waterborne industry View project

Decision Support for Structural Health Monitoring View project

All content following this page was uploaded by Sebastian Geyer on 30 September 2019.

The user has requested enhancement of the downloaded file.

 13th International Conference on Applications of Statistics and Probability in Civil Engineering, ICASP13
Seoul, South Korea, May 26-30, 2019

Bayesian Reliability Assessment with Spatially Variable

Measurements: The Spatial Averaging Approach

Sebastian Geyer
Ph.D. student, Engineering Risk Analysis Group, Technische Universität München,
Germany
Iason Papaioannou
Senior Researcher, Engineering Risk Analysis Group, Technische Universität München,
Germany
Claus Kunz
Engineer, Federal Waterways Engineering and Research Institute, Karlsruhe, Germany
Daniel Straub
Professor, Engineering Risk Analysis Group, Technische Universität München, Germany

ABSTRACT: A large portion of the hydraulic structures in the German waterways network have been in
service for over 70 years. Verification of these structures according to current standards is not always
possible, as they were built according to the standards valid at the time of construction. The massive
shape of large hydraulic structures can lead to spatially variable properties, which should be explicitly
addressed in a reliability assessment through a random field modeling. Spatially distributed
measurements of the concrete properties are available for many structures. These measurements can be
used to update the random field model of the concrete properties through Bayesian analysis. In this
contribution, we develop and investigate a simplified approach to model non-homogeneous spatial
variability in Bayesian posterior random field models that is consistent with the common
semi-probabilistic assessment format. The proposed method reduces the random field to a small number
of random variables that correspond to local averages of the random field and, hence, generalizes the
spatial averaging method originally developed for approximating homogeneous random fields. This
should facilitate its application in engineering practice. We apply the proposed methodology to a
numerical example and compare the results to those obtained with the random field approach.

1. I NTRODUCTION tures, in particular their massive size and specific

load conditions in combination with the long ser-
An important task in infrastructure management vice life, the German Federal Waterways Engineer-
is the reliability assessment of existing structures ing and Research Institute has developed a guide-
that are part of the network, especially when these line for the structural verification of existing hy-
structures are approaching their intended service draulic structures (Federal Waterways Engineering
life. This is the case for a large portion of the and Research Institute, 2016). It allows the use of
hydraulic structures in the German waterways net- probabilistic methods for modeling the material and
work, a substantial number of which have been in geometrical properties as well as load conditions.
service for more than 70 years (Westendarp et al.,
2015). Due to the characteristics of hydraulic struc- Modeling the uncertainties of a system as ran-

1
 13th International Conference on Applications of Statistics and Probability in Civil Engineering, ICASP13
Seoul, South Korea, May 26-30, 2019

dom variables may not suffice in the case where brief introduction to spatial variability and the mod-
an uncertain property is spatially variable, as is the eling of random fields with the SA method is given.
case for material properties in large hydraulic struc- This is followed by a short discussion of Bayesian
tures. To correctly account for this spatial vari- analysis and the conjugate prior concept that is used
ability, uncertainties have to be modeled as ran- to update the random field parameters. The combi-
dom fields instead of random variables (Vanmar- nation of Bayesian analysis and SA is then applied
cke, 2010). to a numerical example to investigate the effects
The structural verification of existing structures and efficiency of the proposed approach.
differs from the verification of newly built struc-
tures as often data is available from inspection 2. M ODELING SPATIAL VARIABILITY
and repair actions that have taken place throughout The traditional approach of modeling uncertainties
the service life. A powerful tool to include these in material properties with random variables is suf-
data into an existing (prior) probabilistic model is ficient for cases where the spatial variability of the
Bayesian analysis (Gelman et al., 2013). modeled quantity is negligible or not of impor-
tance to the performance of the investigated sys-
By combining random field modeling and tem. In large hydraulic structures, some material
Bayesian analysis to incorporate available data, it properties may vary significantly in space. Model-
is possible to define an accurate probabilistic model ing these properties with random variables would
of the material properties for reliability analysis of assume perfect correlation of all spatial locations
a structure. However, this requires to perform all and hence underestimate the uncertainty of the sys-
computational steps with a full random field model, tem. By modeling these uncertainties with random
which is computationally expensive and not easy to fields instead of random variables, this spatial vari-
couple with most commercial finite element soft- ability can be quantified and accounted for in the
ware. The complexity of the random field model reliability analysis.
can be reduced by employing a discretization tech-
nique which expresses the random field by a finite 2.1. Random fields
number of random variables. A variety of such A random field is defined as an indexed collec-
methods exist, including point discretization meth- tion of random variables X (t) , t ∈ Ω, where Ω
ods, averaging discretization methods and series describes the one-, two- or three-dimensional do-
expansion methods (Sudret and Der Kiureghian, main on which the random field is defined (Vanmar-
2000). In this contribution we use the spatial av- cke, 2010). The index t denotes the spatial location
eraging method (SA) (Vanmarcke, 2010) for the within Ω. The random field follows the probabil-
discretization of random fields conditional on spa- ity density function (PDF) f (x (t)) at each point
tially variable measurements. First, we employ di- t ∈ Ω. The correlation of different spatial loca-
rect measurements of the parameter to update the tions is described by the autocorrelation function
random field model. The reduction to a set of ran- ρ (t , t ). The auto-covariance function is then ex-
1 2
dom variables by SA is then performed on the (non- pressed as
homogeneous) posterior random field. This gener-
alizes the application of the SA method, which was Cov [X (t1 ) , X (t2 )] = σX (t1 ) · σX (t2 ) · ρ (t1 , t2 ) ,
originally developed for the discretization of homo- (1)
geneous random fields. The proposed approach ac- where σX (t) is the standard deviation function of
counts for the spatial correlation of the measure- the random field. If µX (t) is constant in space, i.e.
ment locations in the Bayesian analysis, while the µX (t) = µX ∀ t ∈ Ω and Cov [X (t1 ) , X (t2 )] is a
SA discretization allows direct application of com- function of the spatial distance of X (t1 ) and X (t2 )
mercial finite element software in the reliability only, X (t) is said to be (weakly) homogeneous.
analysis process. In this contribution we model spatial variability
The outline of the paper is as follows: First, a with Gaussian random fields, which means that

2
 13th International Conference on Applications of Statistics and Probability in Civil Engineering, ICASP13
Seoul, South Korea, May 26-30, 2019

f (x (t)) is a Gaussian density for any t ∈ Ω. That X(t)

is, the random field is fully described by the spatial
function for the mean µX (t) and the autocovariance
function Cov [X (t1 ) , X (t2 )]. However, the method-
ology could be extended to a class of non-Gaussian
random fields that can be expressed by a nonlinear xTi
transformation of an underlying Gaussian random
field - the so-called translation fields (Grigoriu, t
2009). t1 Ti t2

Figure 1: Approximation of the random field X (t) via

2.2. Spatial average method the local spatial average XTi in the interval [t1 ,t2 ].
The exact representation of a random field at any
point t ∈ Ω typically requires an infinite number of In contrast, the variance of the random field is
random variables. With discretization methods, a always affected by the averaging process. For a
random field is approximated in terms of a finite set homogeneous random field it can be calculated by
of random variables. We choose the SA method to multiplying the variance of the random field by a
approximate a random field (Vanmarcke, 2010) be- reduction factor found by an integration over the
cause it lends itself to easy implementation in prac- autocorrelation function, which is a function of the
tice. It represents the random field as a set of local spatial distance only (Vanmarcke, 2010). This func-
integrals that correspond to the average of the ran- tion is often termed variance reduction function. In
dom field over small portions of the domain Ω. For the non-homogeneous case, however, the correla-
the sake of lucidity, the following derivations re- tion of two points may depend not only on their
strict to the one-dimensional case but the concept spatial distance and the variance of the random field
applies to the multidimensional case as well. is not constant in space. Therefore, the variance of
The spatial average XTi of the random field X (t) the averaging integral has to be calculated by the
over the interval [t1 ,t2 ] is defined as the following following double integral over the autocovariance
integral (Vanmarcke, 2010): function:
Z t2 Z t2
1
Z t2 1
XTi = X (t) dt, (2) Var [XT ] = Cov [X (a) , X (b)] dadb (4)
Ti t1
T2 t1 t1

In order to obtain a realization of the random field

where Ti is the length of the integration domain, i.e.
in terms of local spatial averaging elements, it is
Ti = t2 − t1 . Figure 1 shows a single realization of
necessary to correctly account for the covariances
a homogeneous Gaussian random field and the spa-
and hence the correlation of these local averages.
tial average in the interval Ti determined by applica-
Consider two intervals, T1 and T2 , for which the
tion of Equation 2. It can be seen that the realization
covariance is to be calculated as shown in Fig-
of X (t) is approximated by a constant value in the
ure 2. The covariance of the two spatial averages
interval [t1 ,t2 ].
Cov [XT2 , XT1 ] can then be calculated from the local
Accordingly, the mean of the spatial average can
averaging variances as follows (Vanmarcke, 2010):
be determined:
Z t2 Cov [XT2 , XT1 ] =
1
µXTi = µX (t) dt (3) 1
Ti t1 = (∆ (L1 ) − ∆ (L2 ) + ∆ (L3 ) − ∆ (L4 )) , (5)
2T1 T2
If µX (t) is constant in space (e.g. in a homogeneous where
random field), this integration has no effect on the
mean. Then µXTi = µX ∀ t1 ,t2 ∈ Ω. ∆ (Li ) = Li2 · Var [XLi ] , i = 1, . . . 4 (6)

3
 13th International Conference on Applications of Statistics and Probability in Civil Engineering, ICASP13
Seoul, South Korea, May 26-30, 2019

L1 , L2 , L3 and L4 are the distances between the error measures can be employed, e.g. the error vari-
endpoints of T1 and T2 in Figure 2. The local av- ance, the bias or the mean square error (e.g. Su-
eraging variances of Equation (6) are calculated by dret and Der Kiureghian (2000); Betz et al. (2014)).
means of Equation (4). It is noted that Equation (5) As these quantities are usually defined point-wise,
is also valid for overlapping, containing or coincid- global error measures can be defined by integra-
ing intervals T1 and T2 . tion over Ω or by taking the supremum norm of
the point-wise error (Sudret and Der Kiureghian,
T1 T2 2000). The required number of elements for the SA
t method can be determined by setting a target global
t11 t12 t21 t22
L1 discretization error.

L2 3. BAYESIAN UPDATING OF RANDOM FIELDS

L3 The incorporation of available information in the
analysis can be done via Bayesian analysis, which
L4
allows to update the probabilistic model based on
the information at hand.
Figure 2: Intervals for the calculation of the covari-
ance of the local averaging intervals T1 and T2 . 3.1. Bayesian analysis
Bayesian analysis with available information M is
Based on Equations (3) to (6), all necessary pa- based on the application of Bayes’ rule (Gelman
rameters of the local averaging random variables et al., 2013):
used to represent the Gaussian random field can be fX00 (x) ∝ L (x) · fX0 (x) , (7)
calculated. An approximation of the realization of
the random field of Figure 1 with the spatial average where fX0 (x) is the prior joint PDF of some random
method with five elements is illustrated in Figure 3. vector X, fX00 (x) is the posterior joint PDF of the
The random field is approximated with a set of five random vector X after the update and L (x) denotes
random variables, that follow a multivariate Gaus- the likelihood function, which is proportional to the
sian distribution with mean vector and covariance probability of the information M conditional on x.
matrix calculated following Equations (3) to (6). M can be available as measurements of a continu-
ous quantity Q. The measurements are often sub-
X(t) ject to a measurement error. In the case of an addi-
tive measurement error, the measurement outcome
xT xT
1 xT xT 5 qm is expressed as
2 4

qm = q + εm . (8)
xT The mathematical expression for the likelihood
3
function L (x) is then obtained in terms of the PDF
t of the measurement error:
T1 T2 T3 T4 T5
L (x) = fεm (qm − q) (9)
Figure 3: Approximation of the random field X (t) with
Under the relatively mild assumption of indepen-
SA with nSA = 5.
dence between multiple measurement errors, the
joint likelihood function of a set of nm measure-
The number of averaging elements determines ments is defined as the product of the marginal like-
the accuracy of the approximation. To quantify the lihoods:
nm
approximation error, i.e. the difference between the
L (x) = ∏ fεm (qm,i − qi ) (10)
true random field and the approximation, different i=1

4
 13th International Conference on Applications of Statistics and Probability in Civil Engineering, ICASP13
Seoul, South Korea, May 26-30, 2019

3.2. Conjugate prior for Gaussian random fields structural example of a one-dimensional pinned-
Equation (7) often has to be evaluated numerically, fixed beam subject to a uniformly distributed ver-
as a closed-form solution for the posterior distribu- tical load illustrated in Figure 4. The length of the
tion is typically not available. However, it is possi- beam and the applied load are chosen deterministi-
ble to choose the prior PDF and the likelihood func- cally as L = 10 m and q = 1.2 kN m−1 .
tion such that the posterior PDF can be obtained in
q
functional form. In these cases, the prior PDF is
termed a conjugate prior for the likelihood function
(Gelman et al., 2013).
This concept can be employed for updating the
parameters of a Gaussian random field. To this
end, we model the additive measurement error with
L
a normal distribution with zero mean and standard
deviation σεm . In such case, the conjugate prior for Figure 4: Pinned-fixed beam under uniformly dis-
the likelihood function is the Gaussian distribution. tributed vertical load.
The posterior random field will also be Gaussian
with mean and auto-covariance functions as follows The prior flexibility of the beam is defined by
(Papaioannou and Straub, 2017): a homogeneous Gaussian random field with mean
µX00 (t) =µX0 + ΣX (t) · Σ−1
m,ε · δm (11) value µF0 = 1 × 10−4 N−1 mm−2 and standard devi-
ation σF0 = 2 × 10−5 N−1 mm−2 . The spatial corre-
Σ00X (t1 ,t2 ) =Σ0X (t1 ,t2 ) − ΣX (t1 ) · Σ−1 T lation is modeled with the exponential autocorrela-
m,ε · ΣX (t2 )
(12) tion function with a correlation length of lC = 5 m:
 
µX0 and Σ0X (t1 ,t2 ) are the prior mean and prior co- |t2 − t1 |
ρ (t1 ,t2 ) = exp − (13)
variance function of the Gaussian random field. lC
ΣX (t) is a 1×nm row vector function with element i The system response is evaluated with the linear fi-
equal to Σ0X (t,ti ), where ti , i = 1, . . . , nm denotes the nite element method based on the Euler-Bernoulli
measurement locations. Σm,ε is defined as Σm + Σε , beam theory with a finite element size of lFE =
where Σm is a n
m × nm matrix with element (i, j) 0.01 m.
equal to Σ0X ti ,t j and Σε is a nm × nm diagonal ma- The accuracy of SA is measured by means of
trix with the error variance σε2m on the diagonal. The the statistics of the output of the numerical model.
off-diagonal terms of Σε are 0 due to the assump- To this end we compare the bending moment and
tion of statistical independence between the mea- the displacement of the beam computed with SA
surements. δm is a nm × 1 column vector containing to a reference solution obtained with the midpoint
the difference of the measured data from the prior method with a very fine discretization. We employ
mean with element i equal to (xm (ti ) − µX0 ). the supremum of the point-wise root mean square
The resulting non-homogeneous random field error normalized by the reference mean value as er-
can then be approximated with any discretization ror measure:
method, e.g. with the spatial average method pre- r h i
sented in Section 2.2. sup E (Q (t) − QT (t))2
t∈Ω
4. N UMERICAL EXAMPLE εQ = , (14)
E [Q (t)]
This section applies the SA method to represent
the spatial variability of a material property and as- where Q refers to the quantity of interest, i.e. the
sesses the accuracy of the obtained approximation bending moment or the displacement. We calculate
prior and posterior to the inclusion of spatially vari- the error of Equation (14) based on Monte Carlo
able measurements. We apply the method to the Simulation, i.e. for each realization of the random

5
 13th International Conference on Applications of Statistics and Probability in Civil Engineering, ICASP13
Seoul, South Korea, May 26-30, 2019

field, QT (t) is obtained by averaging the realization ×10-5

of the reference solution Q (t). It is noted that the 2

σF(t) [N-1mm-2]
error in Equation (14) lumps both the error in the
σ′F(t)
representation of the mean as well as the error in 1.5 σ″F(t)
the variance of the random field.
1
4.1. Parameter update with measurements 0.5
In this study, it is assumed that two direct measure-
ments of the beam flexibility are available: 0
0 2 4 6 8 10
" # t [m]
  −4
3 m 0.8 × 10 N mm−1 −2
tm Fm = (15) Figure 6: Prior (blue) and posterior (red) point-wise
7 m 1.2 × 10−4 N−1 mm−2
standard deviation of the flexibility of the beam.
This data can be used to update the random field pa-
rameters by means of Equations (11) and (12). The the posterior parameters, the approximation of both
additive measurement error is modeled with the parameters differs from the true point-wise value.
normal distribution with zero mean and standard
deviation σεm = 0.5 × 10−5 N−1 mm−2 . The result- ×10-4
1.2
ing posterior parameters compared to the prior pa-
μF (t) [N-1mm-2]

rameters are illustrated in Figures 5 and 6. Obvi- 1.1 μ′F (t)

ously, the posterior random field is not homoge- T μ″F (t)
T
neous anymore, the influence of the measurements 1
is clearly visible.
T

0.9
×10
-4
1.2 0.8
μ″F(t) 0 2 4 6 8 10
t [m]
μF(t) [N-1mm-2]

1.1
μ′F(t)
Figure 7: Prior (blue, dashed) and posterior (red,
1 dashed) point-wise mean with nSA = 5.
0.9
×10-5
0.8
0 2 4 6 8 10 2
t [m]
σF (t) [N-1mm-2]

Figure 5: Prior (blue) and posterior (red) point-wise 1.5 σ′F (t)
T
mean of the flexibility of the beam. 1
T

0.5 σ″F (t)

T
4.2. Representation with spatial averages
The number of averaging elements determines the 0
0 2 4 6 8 10
accuracy and the computational effort for the rep- t [m]
resentation with SA as one random variable is re-
Figure 8: Prior (blue, dashed) and posterior (red,
quired for the realization of each element. Figures 7
dashed) point-wise standard deviation with nSA = 5.
and 8 show the parameters of an approximation
with five elements. While the prior mean is rep-
resented exactly with SA, the approximation of the 4.3. Evaluation of the finite element model
prior standard deviation underestimates the point- To determine the accuracy of SA in terms of the
wise standard deviation of the random field. For response of the finite element model, a parameter

6
 13th International Conference on Applications of Statistics and Probability in Civil Engineering, ICASP13
Seoul, South Korea, May 26-30, 2019

study on nSA is conducted. The error in the sys- random field than it is for the prior random field if
tem response according to Equation (14) is evalu- nSA is small and it decreases rapidly for increasing
ated with Monte Carlo Simulation based on 1 × 105 nSA . However, it requires more elements to achieve
repeated simulations. Figures 9 and 10 show the an error of less than 1 %, namely nSA ≥ 8 in this ex-
error in the bending moment and the displacement ample. Contrary to the error in the moment, the lo-
respectively as function of nSA . cation of the maximum displacement error depends
on the random field parameters and nSA .
0.08 4.4. Reliability analysis
ε″M
The application of a discretization method to ap-
0.06 ε′M
proximate random fields in the reliability analysis
εM

0.04
of a structural system requires good approximation
in the tails of the distribution for the system re-
0.02 sponse. To illustrate this behavior, a reliability anal-
ysis is conducted for the simply supported beam
0 with the following limit state function:
0 4 8 12 16 20
nSA
g (X(t)) = Mcrit − M (X(t)) , (16)
Figure 9: Prior (blue) and posterior (red) error in the
bending moment. where Mcrit is chosen 15 % higher than the bending
moment at the fixed end of a beam with a constant
flexibility, i.e. Mcrit = 17.25 kN m.
The reliability analysis is performed with Monte
ε″w Carlo Simulation with 1 × 105 independent simula-
0.06 tion runs. The results are illustrated in Figure 11. It
appears that a small nSA gives non-conservative es-
0.04 timates of the system behavior as the failure prob-
εw

ability is underestimated significantly for the prior

0.02 ε′ and posterior parameters. With increasing nSA the
w
estimate for PF converges towards the reference so-
0 lution. It is noted that the prior estimate of PF is
0 4 8 12 16 20
nSA better than the posterior estimate for the same num-
ber of averaging elements. This performance gap
Figure 10: Prior (blue) and posterior (red) error in the
decreases with increasing nSA . For nSA ≥ 10 the es-
displacement.
timates are sufficiently exact for the prior and pos-
terior random field parameters.
The error in the representation of the bending
moment is very similar for the prior and the poste- 5. C ONCLUDING REMARKS
rior random field, albeit slightly higher for the pos- The reliability assessment of massive existing
terior random field if only one averaging element is structures requires a probabilistic model that can
used. With nSA ≥ 5 both prior and posterior random account for spatially variable parameters. In gen-
field approximate the bending moment with an er- eral, such models are not easy to couple with com-
ror of less than 1 %. It is noted that the location of mercial finite element software. This study inves-
the maximum moment error is at the fixed end of tigates the suitability of the spatial average method
the beam independent of the random field parame- to reduce the complexity of the spatially variable
ters and nSA . model. Bayesian analysis with a conjugate prior ap-
The error in the representation of the displace- proach is used to include available data in the anal-
ment of the beam is also higher for the posterior ysis.

7
 13th International Conference on Applications of Statistics and Probability in Civil Engineering, ICASP13
Seoul, South Korea, May 26-30, 2019

specific part of the structure, facilitates the combi-

0.06
nation of such a probabilistic model with commer-
0.05
P″F cial finite element software. The accuracy must be
0.04 P″F,SA
studied for additional examples before general rec-
PF

0.03 ommendations can be made.

0.02 P′F
ACKNOWLEDGEMENTS
0.01
P′F,SA This work was supported by the German Federal
0 Waterways Engineering and Research Institute.
0 4 8 12 16 20
nSA
REFERENCES
Figure 11: Convergence of the prior (blue, dashed) and
Betz, W., Papaioannou, I., and Straub, D. (2014). “Nu-
posterior (red, dashed) estimate of Pf . merical methods for the discretization of random
fields by means of the karhunen–loève expansion.”
Computer Methods in Applied Mechanics and Engi-
The numerical example of a pinned-fixed beam neering, 271, 109 – 129.
under uniformly distributed load shows that by us-
ing spatial averages, the random field can be ex- Federal Waterways Engineering and Research Institute
pressed by a relatively small number of random (2016). “Bewertung der Tragfähigkeit bestehender,
variables with sufficient accuracy. The error in the massiver Wasserbauwerke.” BAW-Merkblatt, Federal
point-wise bending moment is in the order of 1 % if Waterways Engineering and Research Institute.
five or more averaging elements are used. For the Gelman, A., Carlin, J., Stern, H., Dunson, D., Vehtari,
point-wise displacement of the beam a minimum of A., and Rubin, D. (2013). Bayesian Data Analysis,
eight averaging elements is required to achieve the Third Edition. Taylor & Francis.
same level of accuracy. Remarkably, the required
number of elements is not significantly influenced Grigoriu, M. (2009). “Existence and construction of
by the Bayesian analysis although the posterior ran- translation models for stationary non-gaussian pro-
cesses.” Probabilistic Engineering Mechanics, 24(4),
dom field, unlike the prior, is not homogeneous.
545 – 551.
To investigate the accuracy of the spatial aver-
age method in the distribution tails of the system Papaioannou, I. and Straub, D. (2017). “Learning soil
response, a reliability analysis was conducted. The parameters and updating geotechnical reliability es-
failure criterion was defined by the exceedance of timates under spatial variability – theory and appli-
a critical bending moment. The failure probability cation to shallow foundations.” Georisk: Assessment
was estimated for variable number of averaging el- and Management of Risk for Engineered Systems and
ements and for the prior and posterior random field. Geohazards, 11(1), 116 – 128.
The results indicate that for an accurate estimate of Sudret, B. and Der Kiureghian, A. (2000). “Stochastic
the failure probability, the required number of av- finite element methods and reliability a state-of-the-
eraging elements is larger than the one required for art report.
estimating the mean and variance of the system re-
sponse. For this example, the estimate of the poste- Vanmarcke, E. (2010). Random fields: Analysis and syn-
rior failure probability is slightly less accurate than thesis. World Scientific Publishing Company, revised
that of the prior failure probability with the same and expanded new edition.
number of averaging elements. Westendarp, A., Becker, H., Bödefeld, J., Fleischer,
In conclusion, the spatial average method has the H., Kunz, C., Maisner, M., Müller, H., Rahimi,
potential to be applied in engineering practice to A., Reschke, T., and Spörel, F. (2015). “Erhaltung
describe material properties in the reliability anal- und Instandsetzung von massiven Verkehrswasser-
ysis of existing structures. The reduction to a set of bauwerken.” Beton-Kalender 2015, Wiley-VCH Ver-
random variables, where each of them represents a lag GmbH, 185 – 246.

View publication stats