Structural Safety 66 (2017) 27–37

Contents lists available at ScienceDirect

Structural Safety
journal homepage: www.elsevier.com/locate/strusafe

Spatial reliability analysis of a wind turbine blade cross section subjected

to multi-axial extreme loading
N. Dimitrov ⇑, R.D. Bitsche 1, J.P. Blasques 2
Technical University of Denmark, Department of Wind Energy, Frederiksborgvej 399, Building 115, 4000 Roskilde, Denmark

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a methodology for structural reliability analysis of wind turbine blades. The study
Received 15 September 2015 introduces several novel elements by taking into account loading direction using a multiaxial probabilis-
Received in revised form 30 January 2017 tic load model, considering random material strength, spatial correlation between material properties,
Accepted 31 January 2017
progressive material failure, and system reliability effects. An example analysis of reliability against
material failure is demonstrated for a blade cross section. Based on the study we discuss the implications
of using a system reliability approach, the effect of spatial correlation length, type of material degradation
Keywords:
algorithm, and reliability methods on the system failure probability, as well as the main factors that have
Probabilistic
Wind turbine
an influence on the reliability.
Composite Ó 2017 Elsevier Ltd. All rights reserved.
Blade
Cross section
Reliability
System
Spatial

1. Introduction different loading directions. The most critical loading direction

does not necessarily coincide with the direction where the highest
Wind turbine blade design loads are usually estimated with stress is observed. The random loading and the directional depen-
aeroelastic simulation tools. Here the turbine structure, repre- dence of the material strength imply that for different load direc-
sented by a beam finite element model [1], is subjected to random, tions the likelihood of exceeding the ultimate strength will differ.
dynamically varying environmental conditions. The outcome is the Structural reliability analysis provides the opportunity to take
random loads obtained in terms of beam section forces and into account the inherent variability in both loads and material
moments. The loading direction at a given time is defined by the properties, and finding the most likely combination of factors that
orientation of the resultant of the section moments. Since the will result in failure. Previous studies on structural reliability of
airfoil-shaped regions of wind turbine blades are not symmetric wind turbine blades exist in the literature, for example [4,5], which
the blade strength will be dependent on the loading direction. It demonstrate a reliability analysis of a blade section, and [6], where
is therefore important to identify the most critical loading direc- the reliability along a line on the main spar of the blade is analyzed.
tions and base the blade design criteria on the structural perfor- In the aforementioned studies, the blade is subjected to pure bend-
mance in these directions. While sufficient for predicting the ing moments in the flapwise direction, and system reliability
global behaviour of the structure, the beam fintie element mod- effects are not taken into account. In the current article, we present
elling level does not provide information about the stress field at a methodology for carrying out structural reliability analysis of
the material level. A more advanced approach is to combine the wind turbine blade cross sections which, compared to previous
aeroelastic simulation tools with a finite element based cross- studies, adds several novel elements: (1) the effect of loading direc-
section analysis tool [2,3], where local stress design limits can be tion by means of a probabilistic multiaxial extreme load model is
evaluated at multiple locations within a cross section and under considered; (2) the blade section is considered as a reliability sys-
tem and it is shown how the failure probability for the entire sec-
tion can be estimated by means of a system reliability analysis; (3)
⇑ Corresponding author. we demonstrate how spatial correlation of material properties can
E-mail address: nkdi@dtu.dk (N. Dimitrov). be taken into account and investigate the effect of different corre-
1
SCALE GmbH, Germany.
2
lation lengths on the system reliability estimates.
Suzlon Energy A/S, Denmark.

http://dx.doi.org/10.1016/j.strusafe.2017.01.007
0167-4730/Ó 2017 Elsevier Ltd. All rights reserved.
28 N. Dimitrov et al. / Structural Safety 66 (2017) 27–37

2. Problem definition (2) A separate limit-state function is defined for each failure
mode of interest. The overall failure probability is computed
A reliability analysis with the features defined in the previous as the joint failure probability of all limit states by treating
section requires a number of input models and variables: each of the failure modes as a component in a reliability sys-
tem. For a blade cross section, the section area is divided into
– Structural response model (defining the stress within the sec- a set of sub-divisions (e.g., individual finite elements in a
tion as function of section geometry, composite layup, material finite element model), and each sub-division is treated as a
properties and loads) system component in a reliability system.
– Limit state function (definition of failure and a function indicat-
ing the state of the structure) Both approaches given above represent a spatial-reliability
– Stochastic load distribution model and corresponding random problem solved in a system reliability context. In the following sec-
variables describing the load distribution tions, we discuss the calculation of the system failure probability
– Material strength model and stochastic variables describing for a blade cross section using both a Monte Carlo-based technique
material property distributions as well as a component-reliability based approach.
– Uncertainty variables (accounting for inaccuracies in the mod-
els and inputs)
– A definition for the dependencies between input variables 2.2. Section model

In the following sections, we introduce the necessary models For the purpose of the study we use the Technical University of
and explain the procedures for defining the probability distribu- Denmark’s (DTU) 10 MW reference wind turbine [10] with rotor
tions of the input variables. Due to possible ambiguity in the terms diameter of 178.3 m and hub height of 119.0 m3. The analysis is
used (where e.g. components can both refer to components in a based on a finite element model of the cross section located at
reliability system or vector components), throughout the paper 62:4m radial position on the 86.4 m long blade. The finite element
we denote terms with possible double meaning by using a two- model is implemented in the specialized cross section modelling
word description, as follows: software BECAS [2], using 9004 finite elements with approximately
44,000 nodes. The choice of radial location for analysis is dictated
– finite element for the elements in a finite-element structural by the distribution of the ratio between material capacity and load-
model; ing under operation, which is almost constant from 0 m to 65 m
– system component for the components in a reliability system; radial distance. Thus the 62:4m radial position is towards the end
– component reliability refers to the reliability of a component in a of this plateau, and also coincides with the blade region which is
reliability system subjected to the highest aerodynamic pressure.
– load component for vector components of structural loads The geometry of the cross section is shown on Fig. 1, and a more
defined in a Cartesian coordinate system; detailed view of the region where the spar cap joins the shear web
– vector component for the components of a vector which does not is shown on Fig. 2. The analysis focuses on the composite materials
represent loads, e.g. importance factors from reliability analysis; in the section, namely uni-, bi- and tri-axial glass-epoxy laminates
– structural component for a major part in a wind turbine struc- referred to as Uniax, Biax and Triax, respectively. An epoxy-based
ture, e.g., a wind turbine blade or a part of a blade. adhesive and balsa wood sandwich core material are also present
in the section, however they are not considered in the present
2.1. Limit state function study where ultimate failure of fiber composites is chosen as the
main focus. The sandwich core will carry some transverse shear
The standard approach to reliability analysis requires defining stresses, however for typical blades these stresses are relatively
the state of the structure in terms of a vector of input random vari- small unless a very large non-linear deformation of the cross sec-
ables X and a scalar indicator function, the so-called limit state tion occurs [11]. In cases where the failure of sandwich core is a
function gðXÞ, where gðXÞ 6 0 indicates structural failure. This relevant problem as in e.g. [12], the same procedure as discussed
broad definition of the limit state function allows representing here can be applied. A detailed description of the cross section
complex system problems with multiple failure mechanisms (e.g. design is available in [10]. We assume typical elastic properties
through a finite-element model). However, the scalar definition for the laminates under consideration. The values of the elastic
of the function means that a single output is produced for any com- properties for all materials in the cross section are available in
bination of input variables. As a result, the limit state function [10]. Table 1 lists the elastic properties for the uni-axial and bi-
value can provide information for only one failure mode at a time. axial materials which are the only materials analyzed in details
Setting up a reliability model with multiple failure mechanisms in the remainder of this paper. The stiffness and strength values
can be done with two system reliability approaches: in Table 1 are given in the local material coordinate system, which
as illustrated on Fig. 3 is individual for each material layer. The
(1) A ‘‘global” limit-state function can be designed such that it material is assumed to be linear elastic and effects from geometri-
incorporates all failure modes. The output of the global limit cal and material nonlinearities are not taken into account.
state function will reflect the failure mode which results in Nonetheless, the progressive failure analysis approach presented
the lowest safety margin. Probability integration over the here is also applicable when working with geometrical and mate-
failure domain defined by the ‘‘global” limit-state function rial nonlinear models as well (see [13,11] for example applica-
will directly compute the system failure probability. How- tions). The material property distributions used in the present
ever, such a limit state function may be very nonlinear, dis- study are based on coupon tests, and as a result they only represent
continuous or non-differentiable, which prevents the use of the uncertainty due to small-scale variations, and do not take
simple gradient-based reliability methods. Instead, the sys- large-scale defects into account. Large-scale defects invoke differ-
tem failure probability can be computed using simulation- ent failure mechanisms and in general need to be treated as sepa-
based approach (e.g., Crude Monte Carlo [7], Asymptotic
Sampling [8], Adaptive Importance Sampling [9]). 3
Detailed information on this turbine is available online: http://dtu-10mw-rwt.
vindenergi.dtu.dk.
 N. Dimitrov et al. / Structural Safety 66 (2017) 27–37 29

trailing edge

spar cap

0.4 suction side

reference center
shear webs
0.2 elastic center
y [m]

spar caps shear center

0 mass center
elastic axis 1
-0.2
elastic axis 2
-0.4 pressure side

-1.5 -1 -0.5 0 0.5 1 1.5

x [m]
Fig. 1. Cross section of the DTU 10 MW reference turbine blade at 62.4 m radial position.

the failure mechanism and is defined by the stress component

Uniaxial GFRE which exceeds the ultimate stress envelope; and the limit state
provides the definition of failure – e.g. the number of finite ele-
Triax GFRE
ments exceeding the max stress criterion which are required to
Biaxial GFRE consider the structure as failed. For all limit states, we consider a
Core finite element in the cross section has failed if any stress vector
Adhesive component exceeds the material strength limit for that stress com-
ponent. This is the so-called maximum stress failure criterion,
commonly used for fiber-reinforced composite materials. Unlike
Fig. 2. Material distribution at detail where shear web meets the cap of cross
the Tsai-Wu [15] criterion, the maximum stress criterion distin-
section of the DTU 10 MW reference turbine blade at 62.4 m radial position.
guishes between different failure modes, which has relevance in
rate failure modes. An example stochastic model for the distribu- reliability analysis as each failure mode can be considered a system
tion of defects in wind turbine blades is given in [14]. component in a reliability system. With respect to the approach
The load input to the BECAS finite element model is given in taken for the failure of the cross section, we consider two ultimate
terms of six cross section loads as obtained from a 3-D beam finite limit states:
element model (i.e. an aeroelastic simulation tool): two shear
forces (F x ; F y ), a longitudinal force (F z ), two bending moments – Single-ply failure: in such analyses, the material failure criterion
(Mx ; M y ) and a torsional moment (M z ). BECAS then returns stresses is evaluated for one finite element at a time. If the material
strength is exceeded in any of the finite elements in the cross
and strains in the material coordinate system at each finite ele-
section, it is considered that the entire section has failed (ulti-
ment center. The stresses are used for failure criteria evaluation
mate failure has been reached). When the limit state is applied
and thus for assessment of the failure states of individual elements.
to all finite elements in the cross section simultaneously, the
Furthermore, the program computes K s , a 6  6 constitutive stiff-
first-ply failure load can be found as the load level at which
ness matrix for the entire cross section. The 4th and 5th terms in
the weakest finite element in the cross section fails. If instead
the diagonal of K s represent the cross sectional bending stiffness
the single-ply limit state is evaluated for a particularly chosen
in flapwise and edgewise loading directions, respectively. This is
finite element, the result is a system component failure
later used for defining the ultimate failure condition in a progres-
analysis.
sive failure analysis (see Fig. 4).
– Progressive failure: this analysis assumes progressive material
degradation, with ultimate failure condition based on the
2.3. Ultimate limit states decrease in the bending stiffness of the cross section. For a spec-
ified load level, the section stiffness and the element stresses
We characterize failures in the structure according to two cate- are evaluated in a fully intact cross section. The stiffness of
gories: limit state and failure mode. The failure mode represents failed finite elements is eliminated, and the cross section

Table 1
Effective elastic properties of the glass–epoxy uni-axial and bi-axial plies (from [10]). The properties are given according to a coordinate system where the 1-axis points in the
span-wise direction and the 3-axis points in the local normal direction of the composite ply.

Property Description Unit Fabric type

Uniax Biax
E1 Axial modulus GPa 41.6 13.9
E2 Transverse modulus GPa 14.9 13.9
E3 Out of plane modulus GPa 13.4 12.1
G12 In-plane shear modulus GPa 5.05 11.5
G13 Out-of plane shear modulus GPa 5.05 4.54
G23 Out-of plane shear modulus GPa 5.05 4.54
m12 Poisson ratio – 0.241 0.533
m13 Poisson ratio – 0.268 0.275
m23 Poisson ratio – 0.330 0.333
30 N. Dimitrov et al. / Structural Safety 66 (2017) 27–37

although with bigger uncertainty than in the dependence between

z
M x and F y . There is also a conditional dependence between M x and
y'
My due to variations of the load distribution as function of rotor
y x' 2 3
azimuth and pitch angle, as well as fluctuations due to turbulence
z' which affect both load components. The joint probability distribu-
x tion of M x and M y is modelled using a multivariate model for the
1 distribution of contemporaneous extreme loads on wind turbines
which is described in detail in [16]. The model considers the load
Fig. 3. Blade (x, y, z) and local material (1, 2, 3) coordinate systems for DTU 10 MW components as a combination of a cyclic, gravity-driven compo-
reference turbine blade. The local material coordinate system is individual for each nent and a randomly distributed perturbation. The subtraction of
material layer.
the sinusoidal signal from the random data results in better possi-
bilities for assessing the correlations between loads in different
directions, and thus leads to an improved model for the joint dis-
tribution of loads.
The load distribution used in the present study is based on data
from load simulations carried out on the DTU 10 MW reference
turbine mentioned earlier, using the Hawc2 aeroelastic code [1].
The load case chosen is the extreme turbulence load case DLC1.3
(as defined by the IEC61400-1 standard [17]), which is representa-
tive for the current problem because DLC1.3 is very often the
design-driving case for ultimate loads on wind turbine blades,
and it is also the design-driving case for the DTU 10 MW reference
wind turbine [10]. The DLC1.3 load case simulates a scenario where
the wind turbulence (the standard deviation of the wind field)
attains an extreme value which corresponds to a 50-year return
period for the joint distribution of mean wind speed and turbu-
lence. According to [17] this distribution is defined in the following
way:

– The 10-min average wind speed, U, is described in terms of a

Fig. 4. Diagram of the progressive failure analysis algorithm used to determine the
Weibull distribution (A ¼ 11:28 and k ¼ 2 for IEC class I
section failure state. conditions).
– The turbulence rU (the standard deviation of the wind speed)
stiffness and stress variation are evaluated again. The procedure for 10-min reference periods is defined as log-normal, condi-
continues until either a ‘quasi-static equilibrium’ is reached tional on the wind speed with the relation
where no new element failures are observed, or until any of EðrU Þ ¼ Iref  ð0:75U þ 3:8Þ
the two bending stiffness of the cross section is reduced below
VarðrU Þ ¼ 1:4  Iref
98% of its original value, which indicates failure. The progressive
failure algorithm is outlined graphically in Fig. 4. In the context where EðÞ and VarðÞ are the expected value and variance oper-
of reliability analysis, the progressive failure procedure is only ators, and Iref ¼ ½EðrU j ðU ¼ 15ÞÞ=15 is the mean turbulence
suited for determining the system reliability, since by definition intensity at 15 m/s mean wind speed.
with this approach it is not possible to consider individual finite – For any wind speed, the ETM turbulence is found as the 10-min
elements. turbulence intensity corresponding to a joint return period of
50 years, i.e., a joint probability of 3:8  107 . This is estimated
using the inverse-FORM method [18].
2.4. Loads model

As an example to be used in this study, we choose wind speed of

For the purpose of detailed reliability analysis or structural
11 m/s, which within the operating wind speed range (4–25 m/s)
design of a wind turbine blade section in general, the probability
can be considered as most critical as this is the wind speed at
of observing a given combination of load components acting on
which the turbine reaches rated power at the lowest pitch angle,
the cross section has to be determined. The main loads acting on
resulting in the maximum thrust force and blade bending
a wind turbine blade are distributed loads caused by the aerody-
moments. The following results are based on 1000 load simula-
namic pressure on the airfoils and the own weight of the blade.
tions of DLC1.3 with IEC Class 1A turbulence at 11 m/s mean wind
For a given cross section, the action of these loads is described in
speed. Fig. 5 shows an example of the joint probability density for
terms of six degrees of freedom – the moments and forces acting
M x and M y obtained from load simulations at 11 m/s wind speed,
in three axes of a Cartesian coordinate system. Since both the
as well as 10-min return period contours estimated by the load
bending moments and shear forces are predominantly a function
model.
of the same distributed load, there is strong correlation between
Since the loads and environmental conditions are defined over
pairs of forces and moments, i.e., the flapwise moment M x and flap-
10-min reference periods, using the above definition in a reliability
wise shear force F y , and the edgewise moment M y with the edge-
analysis will result in failure probability estimates per 10-min ref-
wise shear force F x . Given the long and slender structure of the
erence periods, however conditional on a specific, design-relevant
wind turbine blade, the bending moments M x and M y are large in
extreme combination of wind speed and turbulence with 50-year
magnitude and as a result they have the highest impact on axial
return period:
stresses and subsequently on blade design for ultimate strength.

The torsion moment M z is also strongly dependent on aerodynamic pf ðU; rU Þ ¼ pf X j ðU; rU Þ50y ð1Þ
forces and therefore it can also be considered a function of M x ,
 N. Dimitrov et al. / Structural Safety 66 (2017) 27–37 31

Fig. 5. Joint pdf and 10-min return period contours of simulated flapwise bending moment Mx extremes and edgewise bending moment M y extremes. The loads are acting on
a cross section located at 62.4 m radial position for a 86.4 m long wind turbine blade. The data are based on aeroelastic load simulations for a 10 MW turbine with 1000
independent realizations of turbulent wind fields with 10 min duration and 11 m/s mean wind speed. The 10-min return period contour shown in red is determined
empirically from the sample statistics, while the black contour is determined from the model using i-FORM [18].

where pf denotes failure probability, ðU; rU Þ50y represents a given

4
combination of wind and turbulence with 50-year return period, Fy from simulations
and X are the model random variables including the loads. Comput- 3 Fy from correlation dependence
ing a failure probability without the conditional dependence will Mx from simulations
require taking the distribution of U and rU into account, either by
2
including them as random variables in the reliability analysis, or
Load (Normalized)

by integrating the conditional probability of failure:

1
Z Z
pf ;10min ¼ pf ðX j ðU; rU ÞÞf ðUÞf ðrU j UÞdUdrU ð2Þ 0
U rU

In the present study, we use a single combination of U and rU , -1

meaning that the end result will correspond to (1).
-2
2.4.1. Deriving loads in additional degrees of freedom
The model outlined above can be used to define the joint distri- -3
bution of the bending moments Mx and M y , which are typically the
loads most relevant for blade design. However, for some more 360 365 370 375
detailed analyses it may be necessary to include the other load
Time [s]
components, for example the shear forces F x and F y . Based on
observations of the correlations between the simulated time series Fig. 6. Example blade load time series showing the correlation between the
of different load components, for most practical applications the flapwise bending moment M x and flapwise shear force F y .
values of shear forces F x ; F y and the torsion moment M z can be esti-
mated through their dependence on the flapwise and edgewise rF j
bending moments M x and M y . Table 2 shows the correlation coef- F j ¼ lF j þ qF i F j ðF  lF i Þ ð3Þ
rF i i
ficients between section load components observed from aeroelas-
tic simulations. The values shown in bold are used to establish the
dependence relations between the load components. Given two
where lFi ; lF j are the mean values of load components i and j
load components designated F i and F j , and knowing the mean respectively, rF i and rF j are the standard deviations, and qF i F j is
and variance for each component as well as the correlation the correlation coefficient between the two load components.
between them, component j can be expressed as function of com- Fig. 6 shows an example comparing the actual time series of the
ponent i with the following relation [19]: blade flapwise shear force F y to the time series derived using Mx

Table 2
Correlations between different cross section loads estimated from aeroelastic simulations for a section located 62.4 m from the root of a 86.4 m wind turbine blade. The
coefficients given in bold are significant and used in (3) to estimate the values of additional load components.

Fx Fy Fz Mx My Mz

Fx 1 0.37 0.27 0.37 0.98 0.40

Fy 0.37 1 0.40 0.97 0.41 0.83
Fz 0.27 0.40 1 0.45 0.32 0.54
Mx 0.37 0.97 0.45 1 0.43 0.82
My 0.98 0.41 0.32 0.43 1 0.40
Mz 0.40 0.83 0.54 0.82 0.40 1
32 N. Dimitrov et al. / Structural Safety 66 (2017) 27–37

Table 3
List of stochastic variables needed for calculating the component reliability index for finite element number n with Uniax material. Since M x and M y are represented by their joint
distribution, their marginal distributions are listed as conditional on each other. The joint distribution of M x and M y defines the joint probability of the extremes of M x and M y for a
10-min reference period. The material strength distributions are estimated based on the OptiDat public database [20].

Description Notation Range Total count Distribution Mean c.o.v

Flapwise bending moment Mx – 1 Joint with My – –
Edgewise bending moment My – 1 Joint with Mx – –
Longitudinal tensile strength rmax
1t 1...N N Lognormal 909 MPa 0.05
Longitudinal compressive strength rmax
1c 1...N N Lognormal 673 MPa 0.09
Transverse tensile strength rmax
2t 1...N N Lognormal 57 MPa 0.05
Transverse compressive strength rmax
2c 1...N N Lognormal 232 MPa 0.05
In-plane shear strength rmax
12 1...N N Lognormal 80 MPa 0.07

and (3). As visible from Table 2, the axial force component F z is tor components described by the maximum stress failure criterion.
poorly correlated to other load components. This is because F z is As a result, the total number of system components in the full reli-
predominantly caused by the cyclic gravity forces and as such it is ability system will be equal to five times the number of finite ele-
an almost deterministic function of the rotor azimuth. Modelling ments, i.e., M ¼ 5N. However, depending on the geometry and
F z would thus require introducing the rotor azimuth as a random loading conditions, typically no more than one or two failure
variable, which would also require modifications of the model of modes per system component will have a non-zero failure
the joint distribution for M x and M y , where the rotor azimuth is probability.
implicitly included. For the present analysis of a blade section rela- The elastic properties are considered deterministic in the anal-
tively close to the tip, the effect of axial force due to gravity is very ysis and their influence on the reliability is considered negligible.
small and for simplicity we choose to model F z as deterministic and We define the limit state for a given finite element in a cross sec-
replace it with its mean value, which is F z ¼ 0. The model in (3) is tion in terms of a failure criterion evaluated for a specific failure
only applicable for cases where there is a strong linear dependence mode and as a function of the element stress vector rE and mate-
between the variables. For q ¼ 1 this is equivalent to an exact linear rial stiffness and strength properties vector C,
dependence, while for q < 1 the unexplained fraction of the model
C ¼ ½rmax
1t ; r1c ; r2t ; r2c ; r12 ; E1 ; E2 ; E3 ; G12 ; G13 ; G23 ; m12 ; m13 ; m23 :
max max max max
variance (the uncertainty) will be equal to 1  q2 . For example, for
q ¼ 0:7; 1  q2 ¼ 0:51, i.e., less than half of the variance will be Considering one failure mode at a time, and deterministic elas-
explained by the linear model.
tic properties, the failure criterion will be a function only of the
stress vector component and the ultimate strength in the material
3. Reliability analysis setup direction corresponding to the failure mode under consideration.
For a given failure direction i; i ¼ 1; 2; 12, the limit state equation
3.1. Definition of limit state and random variables reads

riE ðXÞ
In the presence of spatial correlation, a random change in the gðXÞ ¼ 1  ð4Þ
properties of one finite element will influence the properties of rmax;i
E ðXÞ
other finite elements. This requires correct treatment of dependen-
where riE is the stress for material direction i, and rEmax;i is the mate-
cies between different system component limit states, and the
rial strength in material direction i. Note that, in Eq. (4), both the
component reliability analysis needs to include the properties of
element stresses rE and the material strengths rmax E are given as
all finite elements in the model. This leads to an increase in the
functions of the stochastic input variables X, and that for axial
number of stochastic variables in comparison to a single-
and transverse directions (i ¼ 1 or i ¼ 2) there are two failure
component reliability analysis. Based on the definition of the struc-
modes (compressive and tensile) per direction, thus the ultimate
ture, material properties and the loads acting on the cross section,
we can summarize the stochastic variables necessary for calculat- stress rmax;i
E will depend on the sign of riE .
ing the reliability for system component (or finite element) num-
ber n out of a total of N as given in Table 3. Note that Table 3 3.2. Model uncertainties
only lists the stochastic strength variables characterizing a single
material type – in this case the Uniax material which is used in In order to take into account the epistemic uncertainties pre-
the reliability analyses in Section 4. With this setup however, using sent in the analysis, we introduce some additional model uncer-
multiple materials will not change the number of variables as long tainty variables (Table 4). The variable X load accounts for possible
as each finite element consists of a single material type. uncertainty in the load simulations due to non-representative
The reliability system considers a single blade cross section and input conditions or inaccuracy of the loads model, X mat accounts
so the spatial correlation of material properties is assumed only for uncertainties in the actual capacity of the material (e.g. devia-
within the cross section. Nonetheless, the approach presented here tions in the strength due to defects, or due to inaccurate estimate
could be easily extended to consider correlation between different of the material strength distribution), and X st – statistical uncer-
sections as well. Assuming spatially correlated material properties tainty due to the limited number of simulations used.
in a given cross section with N finite elements, a single component All model uncertainty variables are considered to be log-
reliability analysis based on the set up presented in Table 3 will normally distributed (denoted by ‘LN’ in Table 4) with unit mean.
require the use of 5N þ 2 variables. Furthermore, as composites This choice is often used in uncertainty modelling [21,22]. It fol-
have non-isotropic behaviour and multiple failure modes, the fail- lows that the uncertainties are assumed not to change the sign of
ure at a particular location cannot be sufficiently well character- the model outputs (as the log-normal distribution is only defined
ized by a single safety margin and a single reliability component for non-negative arguments), and a log-normal distribution with
[13]. Instead, five failure modes per finite element are possible, unit mean and a relatively small variance is similar to a normal dis-
corresponding to the five different material in-plane strength vec- tribution with the same mean. The uncertainty model in the pre-
 N. Dimitrov et al. / Structural Safety 66 (2017) 27–37 33

Table 4
Model uncertainties in a cross section strength analysis: definitions and distribution properties.

Name Description Distribution Mean C.O.V

X mat Material strength LN l¼1 0.25
X st Statistical uncertainty LN l¼1 0.05
X load Load uncertainty LN l¼1 0.15

sent paper is not meant to be fully accurate and in principle cannot bility of failure, and b is the Hasofer-Lind reliability index [27]. The
be made to be so, because there is no information for some of the gradient a of the limit state function gðUÞ at the design point u
key factors introducing uncertainty in the design. As the DTU also has a specific meaning: its vector components ak are called
10 MW reference wind turbine is not a physically existing system importance factors, and they indicate the relative influence of each
and some design and manufacturing aspects are not considered in of the stochastic variables uk to the reliability.
details, there are a number of uncertainties which can not be
directly estimated. Some examples are the uncertainty in blade 3.4. System reliability
profile aerodynamic coefficients, uncertainties related to the blade
manufacturing process, and other. Therefore, as a simplification, Considering a single-ply limit state (i.e., a series reliability sys-
we choose to use only a few, aggregate uncertainty variables which tem), the system reliability can be estimated from the component
altogether have an effect on the reliability problem roughly equiv- reliability indexes by
alent to the effect achieved by the use of partial safety factors for
loads and material strength. Clearly, the choice of uncertainty pf ;system ¼ 1  Uk ðb; RÞ ð6Þ
modelling will affect the resulting reliability, but using the same
uncertainty model and changing other factors allows for compar- where b is the vector of component reliability indexes, R is the
ison studies, which is the main scope of the present analysis. The matrix containing the correlation coefficients between system com-
variances of the uncertainty variables X mat and X load were chosen ponent reliabilities, and Uk ðÞ is the k-dimensional multivariate
such that the variable values corresponding to a 10-min probabil- standard normal distribution. The correlation coefficient between
ity of exceedance of 104 are equivalent to applying a partial safety system components i and j can be estimated from the importance
factor to the characteristic load and strength values in a determin- factors resulting from the component reliability analysis,
istic design. The coefficient of variation for the material strength Rij ¼ aTi aj . There is no closed-form expression for the multivariate
uncertainty was selected so that the 5th percentile of the material normal density, and it has to be evaluated either by direct numeri-
strength divided by a partial safety factor of cm ¼ 2:205, equals cal integration (e.g., [28]), or by some approximate method as for
example in [29,30]. In the present study the efficient asymptotic
the mean material strength divided by the 99:99th percentile of
approximation described in [30] is used.
X mat (where 99.99th percentile corresponds to probability of excee-
Using a damage progression algorithm in conjunction with ran-
dance of 104 ). The coefficient of variation for the load uncertainty dom material properties on multiple finite elements in a model
was selected so that the 98th percentile of the flapwise bending means that the results of the analysis will be path-dependent:
moment multiplied by a partial safety factor of cf ¼ 1:35, equals the global failure load will depend on the sequence in which indi-
the mean of the flapwise load extremes multiplied by the vidual element failures occur, and it will not be smoothly varying
99.99th percentile of X load . The partial safety factors cm ; cf are the with respect to changing initial conditions. Such behaviour cannot
ones used in the design of the DTU 10 MW reference turbine be accounted for by the FORM method where limit state functions
[10], and are based on IEC61400-1 (cf ), and on the GL blade design with defined gradients are required. The problem can neither be
guidelines [23](cm ). The variance of the statistical uncertainty vari- represented as a parallel reliability system, because each element
able X st is taken as recommended by Tarp-Johansen et al. [21]. The failure leads to stiffness loss and load redistribution, a behaviour
uncertainty model defined above is intended to mimic the reliabil- which is not modelled by a parallel system, but is similar to a so-
ity which should be achievable by designing according to the called Daniel’s system [31,32]. Therefore, the reliability model with
IEC61400-1 design standard, which according to earlier studies progressive material degradation requires a simulation-based
ensures reliability levels corresponding to annual failure probabil- analysis, where the problem is treated as a single global limit state
ities in the order of 5  103 to 1  104 [24,21,6]. representing the entire system. In the present paper, we use the
The addition of model uncertainties results in an updated limit Importance Sampling (IS) method [26], which is an efficient
state equation: Monte-Carlo-based technique. With the IS method, the sampling
density is shifted closer to the failure domain in order to obtain a
X st  X load  rE ðXÞ larger fraction of failure events and subsequently faster
gðXÞ ¼ 1  X mat  ð5Þ
rmax
E ðXÞ convergence.
Finding an appropriate location for the IS sampling center point
requires knowledge about the location of the failure domain. This
3.3. Component reliability analysis using FORM information can be obtained either by an iterative search tech-
nique (i.e. Search-based Importance Sampling [9]) or by a determi-
Component reliability with respect to a single safety margin can nate choice based on other analyses, for example the design point
be calculated using FORM (First-Order Reliability Method [25,26]). from a FORM-analysis [33,13]. However, in the case of a reliability
The FORM method estimates the probability of failure by assuming system, multiple safety margins and design points will be present.
that the limit state surface is approximately linear in the close In such cases, an Importance Sampling analysis centered at a
vicinity of the design point u , where u is defined as the most system-component design point will be efficient only if the design
likely point of failure, i.e., the point in the failure domain with points for all other components in the system are in the near vicin-
the highest probability density. With this assumption, the failure ity of the sampling center point.
probability pf can be defined as pf  pFORM ¼ UðbÞ ¼ 1  UðbÞ, In the context of spatial reliability analysis, the reliability index
where UðÞ denotes the standard normal distribution, pf the proba- can be defined as a spatial field which, under the assumption of no
34 N. Dimitrov et al. / Structural Safety 66 (2017) 27–37

Fig. 7. Schematic example of the difference between the upper bound of reliability calculated by finite element analysis, and the continuous spatial reliability index.

Table 5
Correlation coefficients qm between material strength variables.

rmax
1t rmax
1c rmax
2t rmax
2c rmax
12

rmax
1t 1 0.8 0 0.2 0.2
rmax
1c 0.8 1 0 0.2 0.2
rmax
2t 0 0 1 0.8 0.8
rmax
2c 0.2 0.2 0.8 1 0.8
rmax
12 0.2 0.2 0.8 0.8 1

major material defects, can be considered as continuous. We face properties, and that properties in different faces are not corre-
attempt to characterize this field using a discretized grid of points, lated either.
e.g., at the nodes of a finite element model. Such an analysis can
only resolve phenomena down to element size, and some local 3.6. Choice of section components for analysis
minima of the reliability index may be omitted. Therefore, the
analysis represents an upper bound of the reliability of the system, Applying component reliability analysis on all approximately
i.e., a lower bound on the probability of failure, see Fig. 7. [34] sug- 9000 finite elements in the blade cross section is impractical.
gests a method where the upper bound of system failure probabil- Therefore, we first carry out a simpler deterministic analysis which
ity can also be approximately determined. For most problems, is aimed at identifying parts of the section which would be of inter-
increasing the spatial resolution of the model will lead to reduced est for further analysis. We carry out a series of stress calculations
difference between the estimated lower and upper bounds of the while sweeping the loads in M x and M y along the 10-min return
failure probability. period contour drawn on Fig. 5. The flapwise and edgewise bending
moments have the dominating contribution to the stresses within
3.5. Dependence between variables the cross section, and as it was shown in Section 2.4.1, other load
components can be expressed as a function of M x and M y . The load-
The material strength at an individual location in a composite ing direction is then defined as
structure is subject to uncertainties related to, among others, the
presence of micro-scale and material-scale defects, variations in My
h ¼ arctan ð8Þ
fiber orientation, fiber quality, curing process, etc. We assume that Mx
this will result in spatial dependence between properties of differ- Having a series of analyses with values of M x and M y sweeping
ent finite elements expressed in terms of a correlation coefficient along the entire length of their joint probability contour will corre-
ql , and that the degree of dependence will diminish with the dis- spond to analysing loading directions from 0 to 360 degrees. For
tance between finite elements. Such behaviour can be described any combination of values of M x and M y , the values of three addi-
as an exponential function of the distance between two points, r, tional load components (F x ; F y and M z ) are determined using the
h i
and a characteristic correlation length Lc , as ql ðrÞ ¼ exp  Lrc . In relation in Eq. 3 and the correlation values in Table 2.
addition to the correlation between values of the same variable For each section load combination defined using the procedure
at close spatial locations, the values of different anisotropic above, the stresses in the material coordinate system are evaluated
strength parameters at the same point will also be correlated. for all finite elements in the cross section using BECAS. Then, the
We denote this correlation as qm . It is very difficult or virtually failure state of each finite element is assessed using the maximum
impossible to experimentally determine qm ; therefore an estima- stress failure criterion, with three failure modes being considered:
tion has to be done theoretically. In [35] the correlation between tensile failure in axial direction, compressive failure in axial direc-
mechanical properties is estimated based on micromechanics laws, tion, and in-plane shear failure. The failure modes corresponding to
and the results are shown in Table 5. the transverse material direction (matrix failure due to transverse
For any two variables numbered i and j, the overall correlation compression or tension) were not included in the analysis as trans-
coefficient qij is defined as the product of ql and qm . This leads to verse stresses in a blade cross section are typically very small
unless a strong nonlinear deformation of the cross section has
the equation
taken place. The material strengths used to evaluate the failure
qij ¼ ql ðrEðiÞ; EðjÞÞqm ðMðiÞ; MðjÞÞ ð7Þ state are taken as the design strengths that would be used in a
standard deterministic design: they are computed by dividing
where i; j refers to the variable number; rEðiÞ; EðjÞ is the distance the 5th percentile of the material strengths listed in Table 3 to a
between the centers of finite elements EðiÞ and EðjÞ, and MðiÞ; MðjÞ material safety factor cm ¼ 2:205. From the cross section areas
are the two material properties which the variables i and j refer shown on Figs. 1 and 2, the highest values of the maximum stress
to, respectively. For a cross section with sandwich structure, it is failure criterion are observed in finite elements representing Uniax
considered that the core strength properties are not correlated with material in the spar cap on the suction side, and Biax material in
 N. Dimitrov et al. / Structural Safety 66 (2017) 27–37 35

Fig. 8. Polar plot of the maximum-stress failure criterion for the spar cap on the suction side and the shear web closer to the trailing edge of a blade cross section at 62.4 m
radial position, as function of load direction and failure mode. The x and y axes represent the principal loading directions. For any point on the plotted curves, the length of the
line between this point and the origin represents the max-stress failure criterion, while the angle between the line and the x-axis defines the loading direction. The joint
distribution of the load tensor at any loading direction corresponds to the same probability.

the shear web closer to the trailing edge. As a result these locations Table 6
are considered being of highest interest for further analysis. Fig. 8 Importance factors of variables in component reliability analysis of finite elements in
shows failure criterion variation as function of loading direction for a blade cross section.
the most loaded finite elements in the two locations mentioned Variable Notation Importance factor a
above. The loading direction corresponding to highest failure crite-
Flapwise moment Mx 0.21
rion values is h ¼ 174 for both compression stress in the spar cap Edgewise moment My 0.15
and shear stress in the shear web, showing that for this particular Material uncertainty X mat 0.79
blade design and loading conditions, the most critical loading Statistical uncertainty X st 0.16
direction is close to pure flapwise bending. In the following section, Load uncertainty X load 0.47
we present results from the reliability analysis of the spar cap on Material compressive strength rmax
c 0.2

the suction side.

is subjected to the highest stress under flapwise loading. The load-

4. Reliability analysis results and discussion ing condition at the design point is almost pure flapwise loading,
showing that for this particular blade design and load case, the
4.1. Component reliabilities in single-ply failure mode design-driving loading condition is close to pure flapwise bending.
The design points in all component reliability analyses are located
Component failure probabilities for all 630 finite elements of very close to each other - with almost identical design values for
the spar cap located on the suction-side of the blade cross section the material properties, and with some variations in the load val-
are evaluated using the First-Order Reliability Method and the ues at failure due to the varying stress field. As a result, the corre-
limit state equation given in 5. The resulting component reliability lations between system components, evaluated by the inner
indexes with respect to a 10-min reference period vary between product of the respective importance factor vectors, Rij ¼ aTi aj , have
5.2 and 5.8, and their distribution is shown on Fig. 9. The variation high values in the range of 0.9–0.95. Table 6 shows the importance
of the reliability index is relatively smooth and uniform, and it factors of the variables which have the most influence on the reli-
resembles the variation of the stress field in the finite elements. ability index. The factors shown are for the finite element with
The highest probability of failure (lowest reliability index) is highest failure probability, however the results are almost identical
observed in a finite element which is located in the outermost with those for other finite elements. The factor with highest influ-
layer, farthest from the cross section elastic center, and as a result ence on the reliability is found to be the material uncertainty X mat ,
which can be expected since the variance of X mat was chosen so it is
representative of an uncertainty mitigated in design calculations
by a very large partial safety factor of 2.205. This indicates that
reducing uncertainties in the material strength has a high potential
for obtaining a more optimal blade design.

4.2. System reliability

Having determined component reliability indexes for all com-

ponents in the reliability system (i.e. the finite elements in the
Fig. 9. Variation of component reliability indexes over the spar cap on the suction cross section), the system reliability for a series system may be cal-
side of a wind turbine blade cross section. culated using (6). However, the asymptotic approximation to the
36 N. Dimitrov et al. / Structural Safety 66 (2017) 27–37

multivariate normal distribution [30] has insufficient accuracy in 10-min failure probability conditional on a 50-year extreme turbu-
determining the probability integral for large systems with high lence event is not directly comparable to the annual failure proba-
correlation between the system components as present in the cur- bility considered in [21,24] and other; second, the DTU 10 MW
rent problem. As a result it was not possible to determine the series wind turbine blade design may not be detailed enough to capture
system failure probability based on the multivariate normal inte- all uncertainties; and third, the ultimate material failure may not
gral. The high correlation and close spacing of the design points be the dominating failure mode, e.g., a lower reliability may be
from component reliability analysis however means that the sys- associated with a different failure mode. Nevertheless, the reliabil-
tem reliability could be calculated efficiently by carrying out an ities shown in the present work are realistic for the given model
Importance Sampling simulation with sample center at a chosen assumptions and choice of structural component. The absolute
design point. Six such Importance Sampling runs were carried reliability levels do not influence the conclusions in the study as
out, using six assumed spatial correlation lengths: they are based solely on comparison. However, if the reliability
Lq ¼ ½0; 1; 10; 100; 1000; 10000 mm. Two analysis sets are carried of an entire blade has to be estimated for design purposes, all rel-
out, the first using a single-ply failure algorithm with resulting evant failure modes have to be considered for multiple sections on
estimate for the reliability of the blade cross section as a series sys- the blade, while taking all known uncertainties into account.
tem, and a second calculation using the progressive failure analysis The use of a cross section analysis tool (BECAS) proved to be an
algorithm outlined in Section 2.3, where the result represents the efficient way of carrying out system reliability analysis due to its
system reliability as a general system. In Fig. 10, the system relia- high computational efficiency, allowing a fine discretization of
bility results are shown as functions of the spatial correlation the cross section area. An alternative analysis with a full-scale
length. For comparison, Fig. 10 also shows the theoretical reliabil- finite element model would require significantly more computa-
ity for a series system of uncorrelated system components (i.e., the tional efforts or the use of a simplified surrogate model (e.g.
sum of all component failure probabilities), and the reliability of [13]) which adds to the complexity of the problem. However, the
the weakest single component. In all cases the series system shows presently used approach also has some limitations since the cross
higher reliability than what corresponds to summing all compo- section analysis tool assumes linear elastic structural response,
nent failure probabilities, a behaviour typical for a series system which does not allow taking into account geometric nonlinearities
where the presence of correlation between component failure or solution path dependency, which may be relevant in a progres-
modes results in increased reliability. Expectedly, the progressive sive failure analysis [11].
failure approach results in higher reliability than the single-ply
failure analysis. Under the single-ply failure assumption, the fail-
5. Conclusions
ure of a single or few isolated finite elements with very low
strength will be considered as a total system failure even if the
In the present paper we demonstrated the assessment of the
overall structural properties have not deteriorated significantly.
reliability against material failure of a blade cross section while
In a progressive-failure approach, the same event will be analysed
taking random multiaxial loading, random material properties,
further and if there are no additional failures, it will be considered
spatial correlation in material properties, progressive material fail-
as non-failure, which in a system analysis perspective decreases
ure, and system reliability effects into account. For the structure
the number of observed failure events. Increasing the spatial corre-
under consideration, the spar cap and the shear web close to the
lation to 100mm or higher also leads to increased reliability.
trailing edge were found to be the most critically loaded parts of
Although the spatial correlation length in composite materials is
the section. The results indicate that in these regions the most crit-
not known in reality and it is difficult to determine experimentally,
ical loading direction is very close to pure negative flapwise bend-
the results of this study are an indication that efforts in improving
ing moment. It was shown how spatial reliability analysis of a
materials as for example achieving more uniform production pro-
structure defined by a large number of finite elements subjected
cess that reduces the amount of small-scale defects in the compos-
to a relatively uniform stress field can be carried out using Impor-
ite materials, can lead to improved blade reliability.
tance Sampling simulation, while the standard approach with the
The general reliability levels obtained in the present study are
Multivariate Normal integral fails due to insufficient accuracy of
higher than the ones estimated in earlier studies, e.g., [21,24].
the asymptotic approximations. The work described the system
There are several factors which may contribute to this: first, the
reliability analysis of the section applying two different failure
approaches, a single-ply failure approach which results in a series
reliability system, and a progressive failure approach correspond-
ing to a general system. A progressive failure approach results in
higher reliability estimates as it disregards some failures of iso-
lated finite elements with very low strength, which do not result
in significant deterioration of the cross section properties. The
uncertainty variables had a large influence on the probability of
failure, with the largest contribution from the uncertainty related
to the material strength. This is an indication that reducing uncer-
tainties in the material strength has a high potential for obtaining a
more optimal blade design. The analysis showed also that spatial
correlation length of  100 mm or more results in increased sys-
tem reliability. This also is considered as an indication that achiev-
ing more uniform material strength properties will result in
improved blade reliability.

Acknowledgements

Fig. 10. System reliability indexes for the spar cap on the suction side of a wind The work presented herein is a part of the IRPWind project
turbine blade cross section, shown as function of spatial correlation length. funded by the European Union’s seventh programme for research,
 N. Dimitrov et al. / Structural Safety 66 (2017) 27–37 37

technological development and demonstration under grant agree- [17] IEC, International Standard IEC61400-1: Wind Turbines - Part 1: Design
Guidelines; 2005.
ment No. 64011-0352. The financial support is greatly appreciated.
[18] Winterstein SR, Ude TC, Cornell CA, Bjerager P, Haver S. Environmental
parameters for extreme response: inverse form with omission factors. In:
References ICOSSAR-93 Conference, Innsbruck, Austria. p. 8.
[19] Gardoni P, DerKiureghian A, Mosalam KM. Probabilistic capacity models and
[1] Larsen TJ, Hansen AM. How to HAWC2, the user’s manual, Tech. Rep. R-1597, fragility estimates for reinforced concrete columns based on experimental
Technical University of Denmark, Department of Wind Energy; 2012. observations. J Eng Mech 2002;128(10):1024–38.
[2] Blasques JP, Stolpe M. Multi-material topology optimization of laminated [20] Nijssen R. Optidat Reference Document, WMC knolwedge centre, The
composite beam cross sections. J Compos Struct 2012;94:3278–89. Netherlands, http://www.wmc.eu; 2006.
[3] Pollayi H, Yu W. Modelling matrix cracking in composite rotor blades within [21] Tarp-Johansen N, Madsen H, Frandsen ST. Partial safety factors for extreme
VABS framework. J Compos Struct 2014;110:62–76. load effects, Tech. Rep. Ris-R-1319, Ris national laboratory, Roskilde, Denmark;
[4] Lekou DJ, Philippidis TP. Pre- and post-thin: a tool for the probabilistic design 2002.
and analysis of composite rotor blade strength. J Wind Energy 2009;12 [22] Srensen JD, Frandsen S, Tarp-Johansen NJ. Effective turbulence models and
(7):676–91. fatigue reliability in wind farms. Probab Eng Mech 2008;23:531–8.
[5] Bacharoudis KC, Philippidis TP. Estimating design reliability of composite rotor [23] Lloyd G. Guideline for the certification of wind turbines, 2010.
blades under ultimate loading, Journal of Wind Energy, doi: 10.1002/we.1728. [24] Veldkamp HF. Chances in wind energy: a probabilistic approach to wind
[6] Toft HS, Srensen JD. Reliability-based design of wind turbine blades. J Struct Saf turbine fatigue design. Delft University, Wind Energy Research Institute; 2006.
2011;33:333–42. Ph.D. thesis.
[7] Rubinstein RY. Simulation and the Monte Carlo method. Wiley & Sons; 1981. [25] Ditlevsen O, Madsen HO. Structural reliability methods. Wiley; 1996.
[8] Bucher C. Asymptotic sampling for high-dimensional reliability analysis. [26] Madsen HO, Krenk S, Lind NC. Methods of structural safety. Prentice-Hall;
Probab Eng Mech 2009;24:504–10. 1985.
[9] Melchers RE. Search-based importance sampling. J Struct Saf 1990;9:117–28. [27] Hasofer AM, Lind NC. Exact and invariant second-moment code format. J Eng
[10] Bak C, Zahle F, Bitsche R, Kim T, Yde A, Henriksen LC, Natarajan A, Hansen M. Mech Div, ASCE 1974;100(EM1):111–21.
Description of the DTU 10MW reference wind turbine, Tech. Rep. I-0092, [28] Drezner Z. Computation of the multivariate normal integral. ACM Trans Math
Technical University of Denmark, Department of Wind Energy; 2013. Software 1992;18(4):470–80.
[11] Dimitrov N. Structural reliability of wind turbine blades: Design methods and [29] Gollwitzer S, Rackwitz R. An efficient solution to the multinormal integral.
evaluation, Ph.D. thesis, Technical University of Denmark, Department of Probab Eng Mech 1988;3(2):98–101.
Mechanical Engineering; 2013. [30] Pandey MD. An effective approximation to evaluate multinormal integrals. J
[12] Dimitrov N, Berggreen C. Probabilistic fatigue life of balsa cored sandwich Struct Saf 1998;20:51–67.
composites subjected to transverse shear. Sandwich Struct Mater 2015;17 [31] Daniels HE. The statistical theory of strength of bundles of threads. Proc R Soc
(5):562–77. London, Ser A 1945;183(995):405–35.
[13] Dimitrov N, Friis-Hansen P, Berggreen C. Reliability analysis of a composite [32] Gollwitzer S, Rackwitz R. On the reliability of daniels systems. J Struct Saf
wind turbine blade section using the Model Correction Factor method: 1990;7:229–43.
numerical study and validation. J Appl Compos 2013;20(1):17–39. http://dx. [33] Bourgund U, Bucher CG. Importance sampling procedure using design points,
doi.org/10.1007/s10443-011-9246-3. Tech. Rep. 8–86, Institute of Engineering Mechanics, University of Innsbruck,
[14] Toft HS, Branner K, Berring P, Srensen JD. Defect distribution and reliability Austria; 1986.
assessment of wind turbine blades. Eng Struct 2011;33(1):171–80. [34] Kiureghian AD, Zhang Y. Space-variant finite element reliability analysis.
[15] Tsai S, Wu E. A general theory for strength of anisotropic materials. J Compos Comput Methods Appl Mech Eng 1999;168:173–83.
Mater 1971;5:55–80. [35] Toft HS, Branner K, Mishnaevsky Jr LM, Sørensen JD. Uncertainty modelling
[16] Dimitrov N. Probabilistic model for multi-axial dynamic load combinations for and code calibration for composite materials. J Compos Mater 2013;47
wind turbines. J Eng Struct 2016;117:239–49. (14):1729–47.