Engineering Structures 201 (2019) 109774

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Strain-life fatigue assessment of scanned weld geometries considering notch T

effects
⁎
A. Niederwanger , M. Ladinek, G. Lener
Department of Structural Engineering and Material Sciences, Unit of Steel Construction and Mixed Building Technology, University of Innsbruck, 6020 Innsbruck, Austria

A R T I C LE I N FO A B S T R A C T

Keywords: Fatigue failure is one of the most encountered problems with dynamically loaded engineering structures. A
Fatigue crack initiation sophisticated assessment approach to tackle fatigue is the strain-life concept, which enables the consideration of
Life prediction many different aspects, such as mean stress effect, size effect, surface roughness, or net section yielding.
Strain-life Originally developed for base material, the present work aims to enable its reasonable application to welded
Cyclic loading
components. However, there are two significant problems. On the one hand, the strain-life approach does not
Laser scan technology
provide any recommendations on weld modeling, which means that the geometry of each assessed weld has to
be known. On the other hand, the majority of the available material properties correspond to base material.
Especially conventional approximation formulae, which relate strain-life fatigue parameters to more accessible
data, such as static properties or hardness measurements, predict poor results if applied to weld material. The
first issue is tackled by incorporating the real weld geometry into the assessment through 3D laser-scanning. The
latter is addressed by proposing fatigue properties for welded components. After showing that conventional
approximation methods do not accurately describe the fatigue behavior of weldments, thoughts on adjusting the
slope of the strain-life curve, based on comparison to other fatigue approaches and least-squares fitting using
experimental data, are presented. Additionally, recommendations are provided on the statistical size effect,
namely the highly stressed area and the Weibull exponent. By using the proposed parameters for the strain-life
concept, significant improvement in terms of prediction accuracy compared to conventional parameter esti-
mation methods was observed. The proposed model is able to explain more scatter of experimentally determined
fatigue lives than state-of-the-art stress-life fatigue models for welded components.

1. Introduction stresses is not possible or other essential prerequisites for the applica-
tion of standard methods are not fulfilled. Due to recent developments
Welding is a frequently used method to join different components of [2–4], the strain-life approach (also referred to as the local concept)
engineering structures. Especially in steel construction, welding is of may become more accepted by practitioners. Using this approach, ob-
high importance. In such assemblies or whole structures, proofing fa- served characteristics, such as plastic yielding and mean stress effects,
tigue strength is often design relevant. The issue of fatigue is based on a can be taken into account. In contrast to the stress-life concepts, where
complex procedure of damage evolving in the material due to the cyclic fatigue failure is defined as through-section cracking, the estimated
nature of the loading. This process of crack initiation on a micro scale cycles in the strain-life concept usually refer to crack initiation. Even
happens long before the crack becomes detectable by non-destructive though the local concept is not new, its application may seem chal-
testing methods or visual inspection. Due to this complicated damage lenging due to several different possibilities to consider mean stress
process, engineering approaches to ensure safe fatigue design depend effects, or to allow for plasticity. Besides, several material properties are
heavily on phenomenological assumptions and experimentally obtained required. For this very reason, it is appreciated that an enhanced ver-
data as well as safety considerations. sion of the strain-life concept is considered in a recent design guideline
More advanced design approaches exist, yet struggle to become [3], enabling design against fatigue for base material.
equally accepted compared to established concepts, such as the nominal The application of this fatigue approach to welded joints remains
stress, the hot-spot, or the notch stress concept [1]. Nevertheless, so- more challenging as this case is not covered in the recommendations
phisticated concepts are necessary if either the definition of nominal provided in [3]. In the area of a weld, material heterogeneity exists

⁎
Corresponding author.
E-mail address: [email protected] (A. Niederwanger).

https://doi.org/10.1016/j.engstruct.2019.109774
Received 24 July 2019; Received in revised form 6 October 2019; Accepted 7 October 2019
Available online 25 October 2019
0141-0296/ © 2019 Elsevier Ltd. All rights reserved.
A. Niederwanger, et al. Engineering Structures 201 (2019) 109774

resulting from the welding process. The material in the heat affected other words, while estimations in low cycle fatigue lead to significant
zone and the weld material itself show individual microstructures and, underprediction, overprediction for high cycle fatigue results was ob-
therefore, diverging material responses [5,6]. Besides their different served. This is an indicator that the slope of the strain-life curve is not
material properties, the spatial distribution of these zones may not be accurately represented by currently available material data sets.
known, and is, as a consequence, difficult to consider in computations. Therefore, a parameter fitting with experimental results of welded
An engineering approach to this issue is the fitting of free parameters components seems promising to improve the prediction accuracy of the
based on experimental data. concept. Consequently, the contribution at hand investigates the
Regarding the application of the strain-life approach, several adaptation of the strain-life approach to welded joints, especially em-
methods are available to determine the necessary cyclic material phasizing the determination of the fatigue strength exponent, the fa-
properties. Besides the experimental determination, other approaches tigue ductility exponent, and the Weibull exponent from fatigue test of
suggest the approximation of these values from more available prop- welded joints manufactured in S355 mild steel. While the use of the
erties or hardness measurements. Such relations have been fitted to local notch strain approach for the assessment of welded joints is not
existing data, and their validity and accuracy have been critically as- new, its application to scanned weld geometries in combination with
sessed [7,8]. In most of these approaches, e.g., [7,9–11], parameters incorporating support effects and deriving adjusted material parameters
describing the slope of the Wöhler-curve are set to constant values for a is a novelty to the best of the authors’ knowledge.
variety of different steel grades. Roessle and Fatemi [10] as well as After a brief discussion of the theory behind the strain-life concept,
Lopez and Fatemi [11] reported that no sufficient correlation was found the methods used in the present work are explained in Section 3. Sec-
between these parameters and measured hardness values. Un- tion 4 focuses on describing the proposed model, whereas Section 5
fortunately, these two parameters influence the estimated fatigue lives provides information about the different experiments used for the fit-
the most [12]. ting and evaluation process. Afterward, the results are presented in
The local notch strain approach is frequently used to investigate the Section 6, before conclusions are drawn and prospect for further re-
fatigue behavior of base material, for instance [9,13–18]. A modifica- search is given.
tion of the model, proposed by Jesus et al. [19], incorporating failure
probabilities has been recently applied to engineering problems 2. Theoretical background
[20–22].
Despite its widespread use for base material, the strain-life concept’s The strain-life equation consists of two parts. The first one, initially
application to welded components is much less common. One potential proposed by Basquin [36], describes the fatigue behavior of a material
reason for this is that it is challenging to incorporate the weld geometry in high cycle fatigue, where linear elastic material behavior is assumed.
since, unlike with other concepts, no recommendations regarding the Therefore, the relation links the applied elastic strain amplitude εae to
modeling of the weld exist. For the local notch stress approach, usually, the reversal to fatigue crack initiation 2N .
a fictitious notch radius is defined to enable the calculation of stresses
σ ′f
at the notch. By doing so, Neuber’s micro support theory [23] is auto- εae = (2N )b
E (1)
matically taken into account and, therefore, the effects of the stress
gradient considered [24]. For an in-depth discussion of the fictitious The parameters σ ′f and b are material dependent properties and la-
notch stress concept, the interested reader is referred to the literature beled the fatigue strength coefficient and the fatigue strength exponent,
[25], where an overview of recent developments in the field is pro- respectively. E is the modulus of elasticity. The second part of the ap-
vided. proach relates the reversals to crack initiation to the occurring plastic
As sophisticated as well as convenient the definition of a fictitious strain amplitude εpe and is known as the Coffin-Manson equation [37].
notch radius might be, the concept is not quite applicable to the local εap = ε′f (2N )c (2)
notch strain approach since the latter is based on the idea that the fa-
tigue behavior of an arbitrary notched specimen is identical to that of a The fatigue ductility coefficient ε′f and the fatigue ductility exponent
test specimen loaded under the same cyclic deformation [26]. If a fic- c are material parameters as well. By combining Eqs. (1) and (2), the
titious notch radius is applied, its value obviously influences the mag- strain-life relation in its basic form is obtained.
nitude of the stresses and strains at the notch. In the effective notch σ ′f
stress concept, the fatigue strength is simply adjusted depending on the εa = (2N )b + ε′f (2N )c
E (3)
radius used [27]. With the strain-life concept, however, this is not
possible as its reference is the fatigue behavior of a test specimen, at Another assumption of the strain-life approach is that the material
least theoretically. Therefore, the chosen radius has a significant impact behavior can be described by the stabilized cyclic stress-strain curve.
on the fatigue behavior. After initial hardening or softening, the stress-strain hysteresis loops are
Nevertheless, some works covering the application of the strain-life said to maintain steadily after some cycles [38]. If the peaks of these
approach to weldments exist. The origins can be traced back to Mattos stabilized hysteresis loops are connected for different load amplitudes,
and Lawrence [28]. Lawrence et al. [29] as well as Testin et al. [30] the cyclic stress-strain curve is obtained and is often expressed through
also assessed weld toes using the strain-life concept, assuming a worst- the Ramberg-Osgood [39] relation.
case notch stress concentration factor. Skorupa [31] predicted fatigue 1/ n′
σ σ
lives using stress concentration factors based on measured weld radii ε = εe + ε p = +⎛ ⎞
E ⎝ K′ ⎠ (4)
and misalignments. Years later, Saiprasertkit et al. [32,33] proposed to
determine effective notch strains at the weld using a notch radius of Masing behavior and memory effect are assumed to be valid. As
1 mm in combination with modeling the different material zones of the they are not detailly covered in the present work, the interested reader
weld and applying an elastic-plastic material law in the FE calculation. is, therefore, referred to the literature, e.g., [26,37].
They also derived approximation factors, which can be used to obtain Before solving Eq. (3), one needs to determine the elastic and plastic
the effective strain at the notch alternatively. strains and their amplitudes. This is either accomplished by using an
Previous research showed that the application of the strain-life elastic-plastic material law in the finite element (FE) model, or by
concept to welded components is feasible based on its theoretical calculating stresses according to the theory of linear elasticity followed
background. However, depending on the chosen material parameters, by a plasticity correction. The present work incorporates the latter
accurate predictions were achieved only for a specific range of cycles, option by utilizing Neuber’s rule [40], which assumes that the product
and substantial deviations were observed for other regions [34,35]. In of stresses and strains remains constant. The basic idea is expressed via

2
A. Niederwanger, et al. Engineering Structures 201 (2019) 109774

1/ κ
Eq. (5). A
nst = ⎛ ref ⎞
⎜ ⎟ .
A
⎝ 90% ⎠ (8)
σ2K 2
σε = n t
E (5)
The reference area Aref is defined as 500 mm2, and the Weibull
The left side of the equation above represents the local (corrected) exponent κ is set to 30 for steel [2]. The parameter κ can also be esti-
notch stress and strain, while the right side indicates the elastically mated through the expected scatter Tσ according to Eq. (9), [49].
calculated ones. If an FE-calculation is used, the product of the nominal 1.34
net section stress σn and the elastic stress concentration factor Kt is κ=
log10 Tσ (9)
replaced by the stress value obtained from the finite element post
processing. The link between σ and ε is achieved through the Ramberg- The stress distribution in the nearfield of a notch also affects the
Osgood relation, Eq. (4). By doing so, the equation reduces to one un- fatigue life of a component. Whereas a slight stress gradient has a ne-
known variable. Solving Eq. (5) is an iterative task as the local stresses gative impact on fatigue behavior, steep ones are less harmful. Eq. (10)
cannot be expressed explicitly. attempts to describe these effects mathematically.

5 + Gσ
2.1. Mean stress correction nbm =
kbm (10)
The effects of mean stress on the fatigue behavior of a component The stress gradient at the notch Gσ is expressed in relative terms,
should not be neglected [41,42]. For welded constructions, mean stress which means that it represents the percentage stress decrease per length
independence is often assumed, yet the influence of mean stress is ob- unit. Eqs. (10) and (11) demand a dimension of (1/mm). The parameter
served as well, and, therefore, taken into account in the present work kbm is determined as follows.
[1]. Different methods to incorporate mean stress into the strain-life
approach exist. The contribution at hand, uses the energy-based ap- Su 7.5 + Gσ
proach by Smith, Watson, and Topper [43]. The authors of the concept kbm = 5 nst +
Su, bm 1 + 0.2 Gσ (11)
proposed that the product of the maximum stress σmax and the strain
amplitude εa governs fatigue crack initiation. Based on this assumption, The reference tensile strength Su, bm for steel is defined as 680 MPa.
Eq. (3) is, after some rearrangements, modified to According to [3], nbm cannot be <1, which means that the stress gra-
dient at the notch can only increase the estimated fatigue life, but not
σmax εa E = (σ ′f )2 (2N )2b + σf ′εf ′E (2N )b + c . (6) decrease it. The last part of Eq. (7) refers to the macro support theory,
but is implicitly considered in the local concept; hence, set to 1.
The left part of the equation above is labeled the Smith-Watson-
The support factor is incorporated into the assessment approach
Topper damage parameter (PSWT ).
through a shift of the PSWT -life curve. Mathematically, this is expressed
via Eq. (12).
2.2. Size effects
PSWT = np (σ ′f )2 (2N )2b + σ ′f ε′f E (2N )b + c (12)
A possible way to incorporate size effects into the assessment of
welded joints is the use of a highly stressed volume [44] or a highly
stressed weld length [45]. Muniz-Calvente et al. [46] recently showed
2.3. Material parameters
an alternative method, also based on Weibull’s weakest link model
[47], to incorporate size effects. In the present contribution, a different
One drawback of the strain-life approach might be the need of
approach, originally proposed by Liu and Zenner [48,49], incorporating
several material parameters, as opposed to most of the stress-life con-
the highly stressed area, is used since it is recommended in the guide-
cepts, which only requires two (three if the knee point of the Wöhler
lines [2,3].
curve at the fatigue limit is considered). These material properties can
According to [2], the support factor np is split into three parts that
either be determined through experimental testing or by using an es-
allow for different effects.
timation method. The literature provides various of these approxima-
np = nst nbm n vm (7) tion methods [7–9,56,57]. Park and Song [58] provide a comprehensive
overview and discusses prediction errors. Most of these approaches try
This relation was orginially developed for base material. As size
to link the cyclic material properties to static ones, as they are more
effects are reported to influence the fatigue lives of weldments too
easy to obtain. Others, e.g. [10,11,59], try to relate them to hardness
[50–52], it appears reasonable to include these effects into the present
measurements. The present work focuses on deriving material proper-
work.
ties according to the Uniform Material Law (UML) [8,9] and the FKM-
The statistical support factor nst incorporates the size of the highly
method [56], as they are reported to have superior prediction ac-
stressed area around the notch. The larger the region of high stresses,
curacies compared to other concepts and only require static material
the greater the probability of crack initiation. The literature provides
parameters [2]. The approximation formulae of both methods are
different suggestions on how to measure the highly stressed region of a
shown in Table 1. Since the FKM-method does not provide re-
component [45,52–55]. We opted for the 90%-area as it is a tangible
commendations for K ′ and n′, they are determined using the following
measure, which is straightforward to interpret. Additionally, by setting
compatibility relations [26].
a percentage contribution limit instead of weighting the areas with
respect to the Weibull exponent κ , the determination of the highly
Table 1
stressed area was independent from any material properties. Thus, the Recommendations for strain-life parameter estimations. *for Su⩽633 MPa.
subsequent parameter optimization process could be decoupled from
σ ′f b ε′f c K′ n′
the FE-computation, which saved a significant amount of time and
(MPa) (–) (–) (–) (MPa) (–)
enabled a more detailed post-processing and parameter fitting. Ac-
cording to the definition, it involves all parts of the surface that are UML [9] 1.5Su −0.087 0.59 −0.58 1.65Su 0.15
within the 90% stress range of the maximum stress value [54]. The FKM [56] 3.1148(Su )0.987 −0.097 0.338* −0.52 – –
statistical part of the support factor follows to

3
A. Niederwanger, et al. Engineering Structures 201 (2019) 109774

b σ ′f 3. Methodology
n′ = and K′ =
c (ε′f )n′ (13)
This section briefly describes some of the methods and measure-
ments used. For an in-depth discussion of these methods, literature
2.4. Alternative method recommendations are given in the respective subsections.

In addition to the described version of the strain-life approach, a 3.1. Root mean square error
modification of the concept, which recently established its way into the
design guideline [3], is investigated. The approach is based on the The sum of squared error (SSE) loss function is a commonly used
parameter PRAM , which is an enhancement of the Smith-Watson-Topper measure to compare the accuracy of different prediction models, or to
damage parameter that tries to incorporate mean stress effects more calculate residuals in parameter optimization methods [60]. The square
accurately [3]. The indices relate to the Research Association for Me- in Eq. (21) makes the measurement sensitive to outliers as the effect of
chanical Engineering. The standard strain-life equation is modified as large errors is over-proportional compared to small errors.
follows. n
SSE = ∑ (yi ̂ − yi )2
1/ d1 (21)
⎧103
⎪
⎪
( P RAM
P RAM,Z ) if PRAM ⩾ PRAM,Z i=1

In the equation above, yi ̂ are the predicted values and yi the real
N= 1/ d2
⎨103
⎪ ( P RAM
P RAM,Z ) if PRAM,Z ⩾ PRAM ⩾ PRAM,D (observed) ones. The root-mean-square error (RMSE) is obtained by
dividing the SSE by the number of predictions n and applying the
⎪∞ otherwise (14)
⎩ square-root [61]. The RMSE is more natural to interpret because it has
The fatigue strength is defined by the value of the damage para- the same dimension as the underlying data and represents an average
meter at 103 cycles. error.
n
PRAM,Z = np aP,Z (Su )bP,Z , (15) 1
RMSE = ∑ (yi ̂ − yi )2
n i=1 (22)
whereas no fatigue damage is expected beyond the fatigue limit

PRAM,D = np aP,D (Su )bP,D. The accuracies of different models predicting values of the same
(16)
dataset can be compared using RMSE. However, the comparison be-
Support effects are incorporated equivalent to the Smith-Watson- tween different sets of data is not possible as the measure is scale-de-
Topper approach using the parameter np according to Eq. (7). pendent [62].
The damage parameter PRAM is calculated using Eq. (17). Since fatigue assessment (and, therefore, the evaluation of the
model) ranges from a few cycles to a very high number of cycles, errors
(σa + kσm ) εa E if (σa + kσm) ⩾ 0
PRAM = ⎧ of equal magnitude (in terms of percent) in the latter region have a
⎨
⎩ 0 otherwise (17) significant impact on the RMSE, while those in the low cycle regime
only contribute marginally to its value. To address this issue, and also
The parameter k accounts for the material’s sensitivity to mean
because most fatigue material parameters are defined on a log10-log10
stress σm .
scale, we adapted Eq. (22) by applying the logarithm of base 10 to both
⎧ Mσ (Mσ + 2) if σm ⩾ 0 the estimated and the observed cycles.
k=
⎩
M
( M
)
⎨ 3σ 3σ + 2 otherwise,
(18) 1
n
1
n 2
⎡log ⎛ yi ⎞̂ ⎤
RMSE = ∑ ( log10 yi ̂ − log10 yi )2 = ∑ ⎢ 10 ⎜ y ⎟ ⎥
n n ⎣ ⎝ i ⎠⎦ (23)
with i=1 i=1

a
Mσ = M Su + bM
1000 (19) 3.2. Mean absolute percentage error
Since this new method incorporates other (even more) material-
The mean absolute percentage error (MAPE) is another measure to
specific parameters than the original strain-life equation does, the
evaluate the accuracy of a model and is expressed via Eq. (24).
previously described approximation formulae (UML and FKM) are no
n
longer applicable. Therefore, the authors of the concept give re- 1 yi ̂ − yi
commendations based on heuristic determinations from a large number MAPE =
n
∑ yi
i=1 (24)
of experiments [3]. Table 2 presents the suggested values for steel.
The plasticity correction according to Neuber remains the same, but The division by the actual value yi scales the error measurement,
different values for the cyclic stress-strain curve are recommended. The allowing application and comparison across different data sets. Owing
respective suggestions for steel are shown in Table 2 and Eq. (20). to the missing power of two (compared to the RMSE), the MAPE is less
Values for the ultimate strength Su in Eqs. (19) and (20) are inserted in sensitive to outliers [62]. Furthermore, no subsequent modification is
MPa. required to apply it to fatigue models because the normalization to the
actual value equally weighs the percentage errors in low cycle and high
aσ (Su )bσ
K′ = cycle fatigue. However, this error measurement favors somewhat un-
(min(εlim; aε (Su )bε ))n ′ (20) derprediction as the maximum possible error is 100%, whereas the
maximum error for overprediction is not limited.
Table 2
Material parameters of the PRAM -method for steel, according to [3]. 3.2.1. Scattering bandwidth
aP,Z bP,Z aP,D bP,D d1 d2 aM The bandwidth (BW) of the scatter measures the difference between
20 MPa 0.587 0.82 MPa 0.92 −0.302 −0.197 0.35 the highest overprediction and the lowest underprediction in terms of
aσ aε ε lim
cycles to fatigue failure. If all values are over- or underpredicted, it
n′ bσ bε bM
0.187 3.1148 MPa 1033 0.897 −1.235 0.338 −0.1 represents the difference between the best and the worst approxima-
tion. Eq. (25) defines the bandwidth for the scattered values.

4
A. Niederwanger, et al. Engineering Structures 201 (2019) 109774

entry of the data set. The Euclidian norm in Eq. (26) refers to least-
squares fitting. This is equivalent to the minimization of the RMSE as
the value n in Eq. (22) only scales the loss function and with regard to
the square root, the following applies to positive values.

∀ a, b ∈ +: a ⩽ b ⇔ a ⩽ b (27)

A feasible method for solving such optimization problems (finding x

that minimizes the error residual) is the trust-region-reflective algorithm.
For a detailed discussion regarding its underlying theory, the interested
reader is referred to the literature [63].

4. Proposed model

Other fatigue assessment methods for welded components, such as

the nominal stress concept, the hot-spot stress concept, and the effective
notch stress concept, have slopes of the Wöhler curves (or S-N curves)
Fig. 1. Determination of different scatter measurements for a given set of data. which differ from the one of the strain-life concept. The methods
mentioned are covered in various standards and guidelines, and are,
therefore, considered state of the art. For welded structures, most of
⎛ ŷ ŷ ⎞
BW = exp ⎜max ⎧log ⎛⎜ i ⎞⎟ ⎫ − min ⎧log ⎛⎜ i ⎞⎟ ⎫ ⎟ them recommend a slope of m = 3 (on a logarithmic scale) up to a
i ⎨ y ⎬ ⎨ ⎬
⎩ ⎝ yi ⎠ ⎭ ⎠
i
⎝ ⎩ ⎝ i ⎠⎭ (25) certain number of cycles. Regarding the constant amplitude fatigue
limit, different suggestions exist. Whereas the FKM guideline [64]
The inner term of Eq. (25) can be interpreted as the distance be-
predicts no fatigue failure after 5e6 cycles, other references, such as the
tween the highest and lowest deviation in the log10-log10 plot. By ap-
IIW guidelines [65] and the Eurocode [66], suggest a knee point of the
plying the exponential function, one obtains a more tangible measure as
Wöhler curve at 1e7 and 5e6 cycles, respectively. Beyond this point, the
it expresses the bandwidth in terms of a factor. Fig. 1 provides an il-
Eurocode predicts no fatigue damage if a component is subjected to
lustration.
constant amplitude loading. For variable amplitude loading, the curve
It should be noted that because this measure is basically derived
continues with a slope of m2 = 5 up to 1e8 cycles, after which no fatigue
from only two extreme values, it is very sensitive to outliers. Also, no
failure is assumed for any loading. In contrast, the IIW guidelines adjust
information regarding the distribution of the predictions inside the
the slope beyond the knee point to m2 = 22 for both constant and
scatter band is obtained. However, we still consider it a good proxy as it
variable amplitude loading and do not define an ultimate cut-off limit
is straightforward to interpret and a rigorous measurement that favors
since recent research indicates that a constant amplitude fatigue limit
models of constant prediction accuracy.
does not exist for steel [65].
A slight modification, the expected scatter is less sensitive to outliers
According to [66], the fatigue resistance is calculated using Eq. (28).
and compares the stress values (Tσ ) or fatigue lives (TN ) at 10% and 90%
probability of survival, calculated with a log-normal distribution. Δσ mN = Δσcm 2e6 (28)

where Δσc is the fatigue resistance at 2e6 cycles (FAT class), Δσ is the
3.3. Nonlinear least-squares parameter fitting
applied stress range, and N represents the estimated cycles up to fatigue
crack initiation. Concerning the strain-life approach, the Basquin-rela-
The determination of material properties can be achieved using
tion is primarily determining the fatigue resistance in high cycle fatigue
parameter fitting algorithms and experimental data. In the case of the
and is expressed via Eq. (1).
strain-life concept, which represents a nonlinear relation between the
The slopes of the curves, which are illustrated in Fig. 2, are ex-
input and output parameters, we used an iterative method that focuses
pressed differently for the strain-life and the stress-life approach. By
on minimizing an error residual.
applying the logarithm to Eqs. (1) and (28), and taking the derivatives,
min ||f (x )||22 with fi (x ) = yi ̂ − yi the slopes are comparable.
x (26)
∂ log10 (Δσ ) 1
In the equation above, x represents the vector of input parameters =−
∂ log10 (N ) m (29)
and f (·) the errors between predicted and observed values for each

Fig. 2. Different definitions of the slopes of the stress-life curve (a) and strain-life curve (b).

5
A. Niederwanger, et al. Engineering Structures 201 (2019) 109774

failure criteria, such as crack initiation, a previously defined crack

length, or the complete rupture of the specimen, were used in the ex-
perimental series. In fact, the majority of the experiments on which the
recommendations are based on used the latter criterion to define fatigue
failure. Therefore, the stress-life concept for welded components refers
to through-section cracking [65]. The crack propagation up to the
failure of the test specimen affects the resulting slope of the Wöhler
curve. Contrarily, fatigue failure in the strain-life concept usually refers
to crack initiation, yet other criteria also exist. In this contribution, the
fatigue failure is defined in terms of technical crack initiation as this
corresponds to the experiments used in the verification process. The
fatigue strength exponent b is still chosen based on a comparison to the
stress-life concept because, on the one hand, since the S-N curves are
based on a larger number of experiments, the fitted slope values do not
necessarily accurately reflect each of the underlying test series, but
adequately describe the fatigue behavior of the aggregated data. On the
Fig. 3. Derivative of the Wöhler curve with respect to cycles N for the strain-life
other hand, although the experimental S-N curves of the experiments
approach, the stress-life concept according to Eurocode 3 [66], and the pro-
posed model. The black dashed line corresponds to the region beyond the
used in the present contribution correspond to fatigue crack initiation,
constant amplitude fatigue limit (CAFL) and only applies to variable amplitude they do have slope values around m = 3 as well.
loading. The fatigue ductility exponent c can be determined through para-
meter fitting with experimental data. Concerning the fatigue strength
and ductility coefficients, the limited available literature [1,6,28,67] on
∂ log10 (σa )
=b welded components provides significantly varying suggestions. The
∂ log10 (2N ) (30) reported values for ε′f range from 0.218 to 1.224. The majority of the
collected values appear to be above the UML recommendation of 0.59.
The mismatch between stress range (Δσ ) and stress amplitude (σa ) However, the amount of data was too small to make a reasonable
as well as cycles (N) and reversals (2N ) only leads to a shift of the curve prediction, which is why ε′f remained unchanged from the UML in the
(on a logarithmic scale) but does not affect its slope. present work. The fatigue strength coefficient σ ′f was also chosen ac-
As far as the whole strain-life equation is concerned, the slope of the cording to the UML, as the correlation to the ultimate strength was
curve depends on the estimated reversals to failure. In low cycle fatigue, apparent in the data surveyed, but no other significant trend was
it is represented by the fatigue ductility exponent c, whereas in high visible.
cycle fatigue, the slope converges to the fatigue strength exponent b. An In addition to the strain-life curve, we also suggest changes re-
illustration of this behavior is given in Fig. 3. garding the statistical size effect. The applied model in [49] was ori-
The parameters σ ′f and ε′f influence the shape of the transition curve ginally developed for base material. Its application to other materials is
in Fig. 3. For this example, material properties according to the Uniform generally possible as long as the notch radius ρ is of greater magnitude
Material Law (UML) for an ultimate strength of Su = 550 MPa (related than the grain diameter d of the material [49], which is given for the
to the investigated specimens, Section 5) are used. If the fatigue specimens used. S355 steel is reported to have grain sizes ranging from
strength coefficient increases, the curve in Fig. 3 shifts to the left, which 9 to 17 μm for base material [68,69] and a value of approximately 3 μm
means that the value of the slope progresses faster to the fatigue for weld material [69]. The authors in [49] also state that the value of
strength exponent. Increasing the fatigue ductility coefficient has the the highly stressed area must be greater than the square of the grain size
opposite effect. In addition, Fig. 3 also illustrates the stress-life relation d 2 . This criterion is also fulfilled if A90% is used, yet we decided to ad-
as recommended in [66]. The conversion of the stress-life curve to ditionally investigate the influence on the prediction accuracy of the
strain amplitudes was based on the assumption of linear-elastic material model if the contribution limit is set to 80%, as suggested by
behavior (stresses divided by the modulus of elasticity). [45,53,55], for instance. Due to sharp notches resulting from the
It shall be mentioned that comparing the strain-life relation to the welding process, stresses close to the notch show steep gradients,
stress-life equation is only feasible for high cycle fatigue, where no (or leading to very small highly stressed areas, especially if scanned geo-
little) plastic strains occur, and the Basquin equation primarily governs metries (described in the following section) are used. Therefore, the
the estimated cycles. Since determining fatigue life using the strain-life A80% area might be a reasonable adjustment as it is less sensitive to the
equation is preceded by Neuber’s plasticity correction, the strain am- stress gradient at the very vicinity of the notch. Fig. 6b and c illustrate
plitude consists of an elastic and a plastic part. For higher loads, the the difference in the highly loaded area between a stress limit of 90%
proportion of the plastic strain becomes more significant, which does and 80% at the example of specimen S03-3.
not allow the comparison to stress-life curves that are based on elastic Regarding the Weibull exponent, the literature, such as
material behavior. Local plastifications are implicitly considered in the [3,49,51,45], provides different recommendations depending on the
FAT classes of the respective approaches [65], but elastically calculated material and the chosen reference area. While the authors in [49] state
stresses are used for fatigue assessment. that the values for steel lie between 10 and 30, the guideline [3] re-
Accordingly, we propose to adjust only the fatigue strength ex- commends κ = 30 for steel (base material). However, the question
ponent b based on comparisons to the stress-life concept. As seen in arises whether this value is also applicable to weldments. Since the
Fig. 3, the actual slope of the strain-life curve approaches the value of b welding process causes additional heterogeneity, the scatter in terms of
only for very high cycles. Therefore, b was not chosen in relation to the material quality is expected to increase, which leads to a different value
slope of the stress-life curve of m = 3 but set to b = −0.2 as it replicates of κ . The calculation according to Eq. (9) requires the expected scatter
the stress-life curve more accurately. A value of b = −0.33, which cor- Tσ , which is depending on the underlying Wöhler curve. As the Wöhler
responds to m = 3, would result in an actual slope of the strain-life curve is not a priori known, we chose to determine the value of κ also
curve in the region of interest way below that value, as one can imagine through the parameter optimization process.
by shifting the green horizontal line in Fig. 3 downwards.
Note that the recommendations of the referenced concept used for
the comparison are based on a large number of experiments. Different

6
A. Niederwanger, et al. Engineering Structures 201 (2019) 109774

Fig. 4. Overview of the three test series, including specimen geometry and experimental set-up. Red circles indicate the fatigue critical areas. (For interpretation of
the references to colour in this figure legend, the reader is referred to the web version of this article.)

5. Experimental and numerical investigations manufactured in S355 mild steel.

The first series was part of a master thesis [74]. Eight specimens,
5.1. Specimens which consisted of two steel plates that were welded to a continuous
main plate, were tested for fatigue. The experiments were stopped at
Available data (scanned geometries, static material properties, and the first visible crack initiation at the weld toes of any of the four welds.
fatigue results) of three different experimental series were gathered to Liquid penetration inspection was used, which resulted in detecting
use in the present work. An overview of the assessed specimens is given cracks of a visible length at the surface of approximately 5 mm. The
in Fig. 4. Applied loads, observed cycles to fatigue crack initiation, and same procedure was also used for the other two test series.
static material properties can be found in Table 3. In the test series 1, 2, The experiments of the second series were also conducted in the
and 3, the nominal stress ranges refer to the continuous main plate, the course of a master thesis [75]. In each of the three specimens, four
section of the stiffener right above the weld, and the flange of the main stiffeners were welded to two other plates, as seen in Fig. 4. Occurring
girder, respectively. Young’s modulus and the ultimate strength were cracks were recorded and repaired by additional welding in order to
determined through tensile tests of the base material. For the last series, prevent stiffness loss and an resulting redistribution of loads. The ex-
the available data did not include the ultimate strength. Therefore, it periments were stopped when cracks in each of the four welds were
was estimated through measuring Brinell hardness of the base material observed.
using the formula presented by Roessle and Fatemi [10]. The final test series consisted of three specimens with four fatigue
It is noted that the hardnesses between the individual zones of a critical butt welds each. The experiments were carried out until crack
weld seam, namely weld material (WM), heat-affected zone (HAZ), and initiation in each of the welds was visible, which lead to a total of
base material (BM), can vary significantly, with the latter two usually twelve fatigue results. Due to the large specimen size, the loss of stiff-
having higher values than base material, leading to a higher strength ness caused by the cracks was not significant. Unfortunately, 3D scans
but less ductility [5,70–72]. These variations are due to the different in sufficient quality were only available for nine of the twelve welds.
microstructures of the three zones, which also affects both the fatigue Detailed information regarding the specimens and the experimental
strength and the crack growth behavior. Since hardness measurements setups is available in the respective literature (R-x [74], Sx-x [75], and
of the individual zones are often not available during the design pro- Vx-x [76]).
cess, the weld is considered as one part in the present paper and the The weld geometries of the specimens were measured using a mo-
fatigue strength coefficient, therefore, related to the ultimate strength bile scanning device by the company FARO [77]. The resolution was set
of the base material. For further discussion about incorporating the to 50 μm with an accuracy of 25 μm (statistical significance of twice the
different material zones in the fatigue assessment approach, the inter- standard deviation). After the scanning process, possible scan defects
ested reader is referred to the literature [5,32,33,70,73]. were cleared, and the obtained point cloud was triangulated into me-
All experiments were conducted in the laboratory at the University shes and subsequently processed into NURBS (Non-Uniform Rational B-
of Innsbruck under constant amplitude loading, and all specimens were Spline) surfaces of C1-continuity. The C1-continuity ensured

7
A. Niederwanger, et al. Engineering Structures 201 (2019) 109774

Table 3
Details of experiments and specimens. *Loads were applied as forces in the FE
models; these values represent nominal stresses, calculated assuming an idea-
lized specimen geometry. †Values estimated from Brinell Hardness measure-
ments of the base material according to [10].
Specimen Stress Range* Stress Ratio Cycles E Su
No. (MPa) (–) (–) (GPa) (MPa)

R-01 348 0 (0.01) 65,069 220 567

R-02 345 0 (0.01) 54,055 220 567
R-03 346 0 (0.01) 67,425 220 567
R-04 347 0 (0.01) 54,298 220 567
R-05 163 0 (0.01) 611,703 220 567
R-06 163 0 (0.01) 963,948 220 567
R-07 163 0 (0.01) 609,118 220 567
R-08 163 0 (0.01) 678,034 220 567

S01-1 227 0.1 144,000 195 524

S01-2 227 0.1 155,000 195 524
S01-3 227 0.1 101,600 195 524
S01-4 227 0.1 98,500 195 524
S02-1 134 0.1 runout 195 524
S02-2 134 0.1 870,500 195 524
S02-3 134 0.1 677,600 195 524
S02-4 134 0.1 898,000 195 524 Fig. 5. Influence of two meshing options (element size and curvature angle) on
S03-1 182 0.1 581,000 195 524
the von Mises stress (blue = averaged, green = non-averaged) and the highly
S03-2 182 0.1 505,000 195 524
stressed area A90% . Results are scaled to the values obtained with the finest
S03-3 182 0.1 393,000 195 524
S03-4 182 0.1 571,000 195 524 mesh (=100%). (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)
V1-1 99 0 (0.02) 304,000 222 541†
V1-6 99 0 (0.02) 401,000 222 541†
V2-7 139 0 (0.01) 165,000 222 541† submodel in addition to curvature sizing at the weld surface. This
V2-9 139 0 (0.01) 85,000 222 541† means that the element size is related to the curvature of the surface,
V2-11 139 0 (0.01) 85,000 222 541†
automatically meshing critical fatigue areas (notches) in great detail.
V2-12 139 0 (0.01) 235,000 222 541†
V3-13 119 0 (0.01) 153,000 222 541† Fig. 5c and d present another mesh study that covers the determination
V3-15 119 0 (0.01) 294,000 222 541† of an appropriate curvature angle. In contrast to the previous study,
V3-18 119 0 (0.01) 163,000 222 541† decent results are already achieved with coarse mesh values. The
chosen meshing parameters are given in Table 4. The whole submodel
was initially meshed with and element size of 1 mm. At the scanned
convergent stresses in the later FE-calculation. A thorough demonstra- surface, curvature sizing was applied, which relates the element size to
tion of this workflow is given in [78]. the notch radius. A curvature angle of 3° indicates that 120 elements are
used along the edge of a full circle. The transition between the fine
5.2. Numerical calculations mesh at the weld surface and the base mesh is defined by the element
growth rate. The value of 1.25 implies that the size of an adjacent
Due to the realistic representation of the weld through the scanning element is at most 1.25 times larger than that of the current one. Fig. 6a
process, stresses and strains at the weld can no longer be determined uses the example of specimen S03-3 to illustrate the mesh resulting
analytically, but numerically, for instance, using finite element analysis. from parameters according to Table 4.
Therefore, the scanned geometries, available in the form of NURBS- The maximum von Mises stress (non-averaged) of both load steps
surfaces were processed into solids using Boolean operations. These (upper and lower peak of the sine-shaped loading curve), and the highly
solids were imported into ANSYS 19.1 [79], a commercial finite ele- stressed area, which was determined using a post-processing extension
ment software code. After aligning the imported specimen with a pre- in ANSYS, were exported to MATLAB [80]. The strain-life equation,
defined coordinate system, and applying boundary conditions, the including the preceded plasticity correction, was solved afterward. This
meshing of the model followed. The choice of an appropriate element was possible because changing the material properties of the strain-life
size can have a significant influence on the estimated fatigue life, relation does not affect the elastically calculated stresses or the highly
especially if the weld geometry is modeled in such great detail. stressed area. The outsourcing of the fatigue post-processing to MA-
Therefore, a mesh study preceded the main calculations in order to TLAB enabled numerous parameter variations without solving the FE-
obtain consistent results. We used 10-node tetrahedral volume elements model for every parameter set, saving a considerable amount of
with quadratic shape functions for the entire model.
Fig. 5a demonstrates the influence of the element size on the max-
imum von Mises stress. The two lines represent averaged nodal values Table 4
and the maximum overall stress value (unaveraged). Additionally, the Meshing options used in the present study.
change of the highly stressed area with respect to the element size is Global model
illustrated (Fig. 5b). Whereas sufficiently accurate stresses are obtained
from element sizes smaller 0.05 mm (stresses within a 2% range of the Element size 3 mm
Element shape Tet
final value), obtaining accurate highly stressed areas represents a more
strict criterion, leading to element sizes of around 0.02 mm. Besides, it Submodel
is seen that the highly stressed area fluctuates considerably if the mesh
is insufficient, even changing from under- to overprediction from one Base element size 1 mm
Element shape Tet
element size to the next one.
Growth rate 1.25
Meshing the whole model with such small elements is not yet fea- Curvature angle 3°
sible in terms of computing power. Therefore, we opted for using a

8
A. Niederwanger, et al. Engineering Structures 201 (2019) 109774

Fig. 6. Mesh and stress distribution (a) at the example of specimen S03-3. Comparison of highly loaded areas with 90% (b) and 80% (c) stress contribution limit.
Images show the fatigue critical area of the submodel used in the FE calculation.

computing time.
Fatigue lives were calculated using the formulae presented in
Section 2. Mean stress effects were taken into account through the
Smith-Watson-Topper damage parameter, and support effects were in-
corporated by determining both the highly stressed area and the stress
gradient in order to calculate the support factor np . Material properties
were chosen according to the Uniform Material Law (UML), the FKM-
method, the PRAM approach, and the model proposed in the present
work. The numerical results, including the comparison to experiments
and a statistical evaluation according to the metrics discussed in Section
3, are presented in Section 6.

5.3. Parameter fitting

As we only derived the fatigue strength exponent b from the stress-

life equation, the fatigue ductility exponent c was determined through
least-squares parameter fitting. Therefore, the trust-region-algorithm,
which is already implemented in MATLAB, was used. Start values were
chosen according to the Uniform Material Law. The parameters of the
stabilized cyclic stress-strain curve (K ′ and n′) were maintained at their
original values according to Table 1. By doing so, the compatibility
relations in Eq. (13) might be violated, which means that the con-
Fig. 7. Comparison of fatigue lives calculated using the strain-life approach and
gruency of the ratio between the elastic and plastic parts in the cyclic
material properties according to different recommendations with experimen-
stress-strain curve and the strain-life curve is no longer given. However, tally determined fatigue lives.
this relation is lost anyways if the strain-life (or PSWT -life) curve is
shifted by the support factor np . Besides, the literature also reports
substantial deviations concerning the compatibility relations for a great Table 5
Evaluation of different models (Uniform Material Law, FKM-method, PRAM
amount of investigated data [11].
damage parameter, and the proposed model) based on the fatigue experiments
presented in Section 5.
6. Results and discussion
Method RMSElg MAPE BW

In this section, fatigue estimations, according to the previously de- UML 0.658 65.6% 22.0
scribed procedure, are compared with experimentally determined cy- FKM 0.716 68.8% 18.9
cles. Fig. 7 and Table 5 show the results for three sets of material PRAM 0.517 52.8% 20.8
parameters. All three yield to similar estimation errors, with the PRAM Proposed* 1.896 98.6% 5.1
method providing the best results. It is also seen that the error (distance Proposed, A90% 0.205 45.0% 6.2
to the 45°-line) varies significantly depending on the number of cycles. Proposed, A80% 0.195 40.5% 5.8

Regarding the proposed model, so far, only the fatigue strength ex-
ponent b has been modified. Its scatter bandwidth of 5.1 opposed to the
previous values. The least-squares optimization algorithm yielded va-
other values of 22.0, 18.9, and 20.8 for UML, FKM, and PRAM, re-
lues of c = −0.344 and κ = 18.85 for the 90% highly stressed area. If a
spectively, is promising for fatigue results. However, the first stage of
stress contribution limit of 80% was used ( A80% ), the values for c and κ
the proposed model (just adapting b) leads to significant under-
resulted in −0.338 and 15.43, respectively. The determined value of κ
prediction for all the investigated specimens as seen in Fig. 7.
is in accordance with estimations from statistical analyses of the ex-
The goal is, therefore, to take advantage of the model’s narrow
periments. The Weibull exponents, derived separately from experi-
scatter band and shift it upwards in order to reach an overall good
mental one-parameter Wöhler curves for each of the investigated series,
prediction accuracy. The approach chosen was to optimize the fatigue
are 24.4, 12.1, and 9.6.
ductility exponent c, as well as the Weibull exponent κ to find a
In order to avoid the risk of finding local minima, the optimization
minimum of the RMSE. The rest of the parameters were kept at their

9
A. Niederwanger, et al. Engineering Structures 201 (2019) 109774

Fig. 8. Comparison of fatigue lives estimated by different versions of the pro-

posed model with experimentally determined ones.

process was performed with different starting values for c and κ , all
leading to the same final values. Compared to the first stage of the
proposed model, where only the fatigue strength exponent b was ad-
justed, the scatter bandwidth slightly increased to values of 6.2 ( A90% )
and 5.8 ( A80% ), but both the RMSE and the MAPE improved sig-
nificantly as seen in Table 5. The prediction accuracy of the proposed
model is additionally illustrated in Fig. 8. Regarding the highly stressed
Fig. 9. Influence of the amount of experimental data used in the least-squares
area, a contribution limit of 80% leads to superior prediction accuracy
parameter optimization on the resulting fatigue ductility exponent (a) and the
in terms of RMSE, MAPE, and scatter bandwidth, compared to a value Weibull exponent (b). Boxplots illustrate the range of values, the 25%- and
of 90%. 75%-quantiles, and the Median.
To test the sensitivity of the parameter fitting process, we split the
available data into training sets containing different numbers of ex-
Table 6
periments. For each training set, the fatigue ductility exponent and the
Proposed strain-life material properties for welded components.
Weibull exponent were determined. The fluctuations and deviations in
the resulting values of c and κ are presented in Fig. 9a in the form of σ ′f b ε′f c K′ n′ κ
boxplots. A complete variation, including all possible combinations of (MPa) (–) (–) (–) (MPa) (–) (–)

experimental data, would have led to millions of different training sets.

1.5Su −0.20 0.59 −0.34 1.65Su 0.15 15
Therefore, 10,000 combinations of data were chosen randomly for each
size of the training set. In other words, each of the boxplots in Fig. 9
consists of 10,000 values of c and κ (excluding the last four entries, investigated (base material versus welds), on the other hand, strain-life
where a full combination led to less than 10,000 sets). material properties reported by different publications show a distinct
Fig. 9a shows that the values of the determined fatigue ductility scatter [7], even for the same steel grade [34].
exponent fluctuate only slightly based on the underlying data set. Some Concerning the statistical size effect, we recommend to use the
outliers stretch the range of data, but the quartiles are located within highly stressed surface area with a stress contribution limit of 80%,
± 3% around the Median. The scatter obviously decreases the more since it led to superior prediction accuracy compared to a value of 90%,
experiments are incorporated in the training set. Concerning the Wei- on the one hand, and is further away from the model’s limits of ap-
bull exponent, a greater scatter compared to the parameter c is ob- plication [49], on the other.
served. The values of the quantiles are approximately within ± 10% of Table 7 shows how well the proposed model is capable of explaining
the Median. Overall, the sensitivity analysis confirms the robustness of the observed scatter in the experiments. Therefore, loading over cycles
the parameter fitting process as the value of c does not change con- was plotted for both experimental and estimated fatigue results, and the
siderably based on which experimental data is used in the optimization scatter in both dimensions (Tσ and TN ) was determined as described in
process. The Weibull exponent κ experiences slightly more scatter. Section 3. We additionally fitted experimental one-parameter Wöhler
Based on the conducted calculations and analyses, we propose fa- curves (constant slope) for each of the series individually using the
tigue material parameters for welded components according to Table 6. applied load range ΔF and linear regression analysis. A combination of
In comparison with recently published strain-life material properties for all the series was not possible as the geometries and set-ups varied,
S355 base material of b = −0.098 and c = −0.664 [17,18], it is seen that making a comparison of applied loads impossible. Furthermore, we also
the fatigue strength exponent proposed in the present contribution is used the stress ranges Δσ obtained from FE-calculations of the scanned
lower by approximately a factor of two whereas the fatigue ductility is weld geometries. The slope of the Wöhler curve was, again, determined
higher by about the same factor. However, this comparison should be for each series individually, and once for all specimens combined. To
taken with caution. On the one hand, different material zones were evaluate the proposed version of the strain-life approach, we chose the

10
A. Niederwanger, et al. Engineering Structures 201 (2019) 109774

Table 7
Scatter measurements in terms of cycles and applied load (expressed as applied
• Therefore, the present work focused on determining suitable mate-
rial properties that lead to an increase in prediction accuracy re-
force range ΔF , notch stress range Δσ , or the damage parameter corrected by garding the assessment of weldments. The fatigue strength exponent
the support factor PSWT / np ) of different models and experimental series. *not was derived from stress-life concept considerations, whereas the
applicable due to different specimens and set-ups. fatigue ductility exponent resulted from least-squares optimization
Series Experiment Notch stress Proposed with experimental results. The proposed values are b = −0.20 and c
y-axis – ΔF Δσ PSWT/ np = −0.34.

Tσ R-x, [74] 1.13 1.17 1.14

• Concerning the statistical size effect, it seems reasonable to use the
highly loaded surface area with a contribution limit of 80% of the
Sx-x, [75] 1.29 1.53 1.23
Vx-x, [76] 1.38 1.65 1.20 maximum stress for welded components. This appears more suitable
all * 1.75 1.37 for welds as it led to better prediction accuracy for the investigated
experiments compared to setting the limit to 90%. Additionally, the
TN R-x, [74] 1.51 1.70 1.64 Weibull exponent was also derived through the parameter fitting
Sx-x, [75] 2.50 3.29 2.20
Vx-x, [76] 2.55 2.37 1.98
process and resulted in κ = 15. This value is within the range of the
all * 4.50 3.22 ones obtained from scattering estimations of the individual experi-
mental series that were used in this study (24.4, 12.1, and 9.6).
• Although the experimental Wöhler curves were determined in-
calculated Smith-Watson-Topper damage parameter divided by the dividually for each series, the proposed model was capable of ex-
support factor, to make a comparison to the PSWT -life curve possible. plaining more scatter in the fatigue behavior overall. This implies
Table 7 demonstrates that the proposed model describes the fatigue that the application of the strain-life approach to welded compo-
behavior of the investigated specimens more accurately than the ex- nents is possible and that it has the potential to predict fatigue
perimental Wöhler curves for both Tσ and TN (except for the first series, failure more accurately than conventional stress-life concepts. It was
where the empirical fitting showed slightly less scatter). This is espe- also shown that including scanned geometries in the assessment
cially notable since the experimental Wöhler curves were determined approach on its own does not necessarily lead to more reliable fa-
separately for each of the series, whereas only one set of material tigue results. Only by subsequently taking into account plasticity
parameters was used to evaluate the proposed model for all series. The and support effects, an increase in the quality of the predictions was
fact that the stress values used by standard stress-life approaches observed.
(nominal stress, hot-spot stress, or notch stress concept) are perfectly • The proposed parameters represent an estimate, and further data is
correlated to the applied loads (linear relation), and the slopes of their required to validate or enhance the model. Particularly, experiments
curves are constant too, implies that the proposed model is superior to in low cycle fatigue are aspired to determine the fatigue ductility
these approaches as well. The column referring to the notch stress range exponent more accurately, as the sensitivity analysis showed some
shows that solely using scanned geometries in the finite element ana- variation based on which data was used; yet the trend was apparent.
lysis leads to worse predictions. However, by incorporating plasticity Furthermore, σ ′f and ε′f were kept at their values proposed by the
correction as well as accounting for support effects, a considerable UML in the current study. If more data were available, these values
amount of scattering is explained in the present study. could also be determined more precisely, which would presumably
One should mention that most of the available experiments are in further increase the accuracy of the model. However, since the
the regime of high cycle fatigue (the lowest observed fatigue life in the strain-life concept does not provide recommendations regarding the
present study was around 55,000 cycles), where Basquin’s relation and, modeling of welds, complete datasets must include information
thus, the fatigue strength exponent b is primarily governing the strain- concerning the weld geometry, preferably obtained through laser
life equation. Based on this circumstance, we recommend to validate scanning, and are, therefore, of limited availability.
the proposed model with further experiments and also plan to do so.
Especially experimental results in LCF are beneficial as plastic strains Declaration of Competing Interest
are more prevalent in this region, and, therefore, the Coffin-Manson
equation, driven by the fatigue ductility exponent c, is mainly influ- There are no conflict of interest in this work.
encing the estimated fatigue lives. By incorporating more experimental
data, the proposed fatigue exponents for welded components can be References
verified, and adjusted if needed.
[1] Radaj D, Sonsino CM, Fricke W. Fatigue assessment of welded joints by local ap-
7. Conclusion proaches. 2nd ed. Woodhead, Cambridge, England: Woodhead Publishing in ma-
terials; 2011http://www.sciencedirect.com/science/book/9781855739482.
[2] Fiedler M, Varfolomeev I, Wächter M. Richtlinie Nichtlinear: Rechnerischer
The present work investigated the application of the strain-life ap- Bauteilfestigkeitsnachweis unter expliziter Erfassung nichtlinearen Werkstoff-
proach to welded components. The following can be concluded based Verformungsverhaltens, 1st ed., FKM-Vorhaben Nr. 301, VDMA; 2016.
[3] Fiedler M, Wächter M, Varfolomeev I, Vormwald M, Esderts A. Richtlinie
on the experimental and numerical investigations presented. Nichtlinear: Rechnerischer Festigkeitsnachweis unter expliziter Erfassung nichtli-
nearen Werkstoffverformungsverhaltens; für Bauteile aus Stahl, Stahlguss und
• Conventional strain-life material parameter estimation methods, Aluminiumknetlegierungen. 1st ed. FKM-Richtlinie: VDMA Verlag GmbH, Frankfurt
am Main; 2019.
such as the UML or the FKM-method, yield to inaccurate results if [4] Austrian Standards Institute, ÖNORM prEN 13001: Cranes - General design - Part 3-
applied to welded structures. While significantly too low fatigue 8: Limit states and proof competence of machinery - Shafts; 2018.
lifetimes were estimated for LCF, overprediction occurred in HCF for [5] Cheng G, Kuang ZB, Lou ZW, Li H. Experimental investigation of fatigue behaviour
for welded joint with mechanical heterogeneity. Int J Press Vessels Pip
some experiments. Scatter bandwidths of factor 22.0 and 18.9 were 1996;67(3):229–42. https://doi.org/10.1016/0308-0161(94)00020-4.
observed. The PRAM-approach lead to only slightly better results in [6] Higashida Y, Burk JD, Lawrence FV. Strain-controlled fatigue behavior of ASTM
terms of RMSE, and showed a bandwidth of 20.8. The fact that the A36 and A514 Grade F Steels and 5083–0 Aluminum Weld Materials. Weld Res –
Suppl Weld J 1978;57:334–44.
estimation error varies considerably depending on the number of
[7] Meggiolaro MA, Castro JTP. Statistical evaluation of strain-life fatigue crack in-
cycles (also seen in the broad scatter bands) is an indicator that the itiation predictions. Int J Fatigue 2004;26(5):463–76. https://doi.org/10.1016/j.
slope of the strain-life curve, described by the fatigue strength ex- ijfatigue.2003.10.003.
ponent b and the fatigue ductility exponent c, does not accurately [8] Haibach E. Betriebsfestigkeit: Verfahren und Daten zur Bauteilberechnung, 2nd
Edition, VDI-Buch. Berlin u.a.: Springer; 2002.
represent the fatigue behavior of welded components.

11
A. Niederwanger, et al. Engineering Structures 201 (2019) 109774

[9] Bäumel A, Seeger T. Materials data for cyclic loading. Mater Sci Monogr Vol. 61. University Press; 1998. https://doi.org/10.1017/CBO9780511806575.
Amsterdam: Elsevier; 1990. [38] Stephens RI, Fatemi A, Stephens RR, Fuchs HO. Metal fatigue in engineering. 2nd
[10] Roessle M, Fatemi A. Strain-controlled fatigue properties of steels and some simple ed. New York: A Wiley-interscience publication, Wiley; 2001.
approximations. Int J Fatigue 2000;22(6):495–511. https://doi.org/10.1016/ [39] Ramberg W, Osgood WR. Description of stress-strain curves by three parameters,
S0142-1123(00)00026-8. National Advisory Committee for Aeronautics 1943 (Technical Note 902); 1943.
[11] Lopez Z, Fatemi A. A method of predicting cyclic stress-strain curve from tensile [40] Neuber H. Theory of stress concentration for shear-strained prismatical bodies with
properties for steels. Mater Sci Eng: A 2012;556:540–50. https://doi.org/10.1016/j. arbitrary nonlinear stress-strain law. J Appl Mech 1961;28(4):544. https://doi.org/
msea.2012.07.024. 10.1115/1.3641780.
[12] Niederwanger A, Ladinek M, Lang R, Timmers R, Lener G. On the stability and [41] Pook LP. Metal fatigue: what it is, why it matters. Solid mechanics and its appli-
sensitivity of the strain-life approach using the example of mild steel. J Constr Steel cations Vol. 145. Dordrecht: Springer; 2007.
Res 2019;153:483–94. https://doi.org/10.1016/j.jcsr.2018.10.030. [42] Bannantine JA, Comer JJ, Handrock JL. Fundamentals of metal fatigue analysis.
[13] Seeger T, Heuler P. Generalized application of Neuber’s rule. J Test Eval Englewood Cliffs, New Jersey: Prentice Hall; 1990.
1980;8(4):199. https://doi.org/10.1520/JTE11613J. [43] Smith KN, Watson P, Topper TH. A stress-strain function for the fatigue of metals. J
[14] Boller C, Seeger T, Laird C. Materials data for cyclic loading: low-alloy steels. Mater 1970(4):767–78.
Burlington: Elsevier Science; 1987http://gbv.eblib.com/patron/FullRecord.aspx? [44] Kuguel R. A relation between theoretical stress concentration factor and fatigue
p=1839036. notch factor deduced from the concept of highly stressed volume. ASTM Proc
[15] Langlais TE, Vogel JH. Overcoming limitations of the conventional strain-life fa- 1961;61:732–48.
tigue damage model. J Eng Mater Technol 1996;118(1):103. https://doi.org/10. [45] Baumgartner J, Bruder T, Hanselka H. Fatigue strength of laser beam welded au-
1115/1.2805921. tomotive components made of thin steel sheets considering size effects. Int J Fatigue
[16] Joadder B, Shit J, Acharyya S, Dhar S. Fatigue failure of notched specimen—a 2012;34(1):65–75. https://doi.org/10.1016/j.ijfatigue.2011.01.022.
strain-life approach. Mater Sci Appl 2011;02(12):1730–40. https://doi.org/10. [46] Muniz-Calvente M, de Jesus A, Correia J, Fernandez-Canteli A. A methodology for
4236/msa.2011.212231. probabilistic prediction of fatigue crack initiation taking into account the scale
[17] de Jesus AM, Matos R, Fontoura BF, Rebelo C, Simões da Silva L, Veljkovic M. A effect. Eng Fract Mech 2017;185:101–13. https://doi.org/10.1016/j.engfracmech.
comparison of the fatigue behavior between s355 and s690 steel grades. J Constr 2017.04.014http://www.sciencedirect.com/science/article/pii/
Steel Res 2012;79:140–50. https://doi.org/10.1016/j.jcsr.2012.07.021. S0013794417300929.
[18] Correia JAFO, de Jesus AMP, Fernandez-Canteli A. Modelling probabilistic fatigue [47] Weibull W. A statistical theory of the strength of materials. Royal Swedish Institute
crack propagation rates for a mild structural steel. Fract Struct Integr 2015;9(31): of Engineering Research; 1939.
80–96. https://doi.org/10.3221/IGF-ESIS.31.07. [48] Liu J, Zenner H. Berechnung der Dauerschwingfestigkeit unter Berücksichtigung
[19] de Jesus AM, Pinto H, Fernández-Canteli A, Castillo E, Correia JA. Fatigue assess- der spannungsmechanischen und statistischen Stützziffer. Materialwiss.
ment of a riveted shear splice based on a probabilistic model. Int J Fatigue Werkstofftech. 1991;22(6):187–96. https://doi.org/10.1002/mawe.19910220602.
2010;32(2):453–62. https://doi.org/10.1016/j.ijfatigue.2009.09.004http://www. [49] Liu J, Zenner H. Berechnung von Bauteilwöhlerlinien unter Berücksichtigung der
sciencedirect.com/science/article/pii/S0142112309002771. statistischen und spannungsmechanischen Stützziffer. Materialwiss. Werkstofftech.
[20] Correia JA, de Jesus AM, Fernández-Canteli A. Local unified probabilistic model for 1995;26(1):14–21. https://doi.org/10.1002/mawe.19950260106.
fatigue crack initiation and propagation: application to a notched geometry. Eng [50] Feng L, Qian X. Size effect and life estimation for welded plate joints under low
Struct 2013;52:394–407. https://doi.org/10.1016/j.engstruct.2013.03.009http:// cycle actions at room and low ambient temperatures. Thin-Wall Struct
www.sciencedirect.com/science/article/pii/S0141029613001260. 2018;132:195–207. https://doi.org/10.1016/j.tws.2018.08.017.
[21] de Jesus AM, da Silva AL, Correia JA. Fatigue of riveted and bolted joints made of [51] Kaffenberger M, Vormwald M. Considering size effects in the notch stress concept
puddle iron—a numerical approach. J Constr Steel Res 2014;102:164–77. https:// for fatigue assessment of welded joints. Comput Mater Sci 2012;64:71–8. https://
doi.org/10.1016/j.jcsr.2014.06.012http://www.sciencedirect.com/science/ doi.org/10.1016/j.commatsci.2012.02.047.
article/pii/S0143974X14001837. [52] Fricke W, Gao L, Paetzold H. Fatigue assessment of local stresses at fillet welds
[22] Sanches RF, de Jesus AM, Correia JA, da Silva A, Fernandes AA. A probabilistic around plate corners. Int J Fatigue 2017;101:169–76. https://doi.org/10.1016/j.
fatigue approach for riveted joints using monte carlo simulation. J Constr Steel Res ijfatigue.2017.01.011.
2015;110:149–62. https://doi.org/10.1016/j.jcsr.2015.02.019http://www. [53] Bruder T, Störzel K, Baumgartner J. Fatigue assessment of seam welds of auto-
sciencedirect.com/science/article/pii/S0143974X15000826. motive components by local stress approaches. Materialwiss. Werkstofftech.
[23] Neuber H. Kerbspannungslehre: Theorie der Spannungskonzentration Genaue 2008;39(10):726–33. https://doi.org/10.1002/mawe.200800354.
Berechnung der Festigkeit, 4th Edition, Klassiker der Technik, Springer, Berlin and [54] Wächter M, Müller C, Esderts A. Angewandter Festigkeitsnachweis nach FKM-
Heidelberg, 2001. https://doi.org/10.1007/978-3-642-56793-3. URL <https:// Richtlinie: Kurz und bündig. Wiesbaden: Springer Vieweg; 2017. https://doi.org/
doi.org/10.1007/978-3-642-56793-3>. 10.1007/978-3-658-17459-0.
[24] Radaj D, Lazzarin P, Berto F. Generalised Neuber concept of fictitious notch [55] Ai Y, Zhu S-P, Liao D, Correia JA, de Jesus AM, Keshtegar B. Probabilistic modelling
rounding. Int J Fatigue 2013;51:105–15. https://doi.org/10.1016/j.ijfatigue.2013. of notch fatigue and size effect of components using highly stressed volume ap-
01.005. proach. Int J Fatigue 2019;127:110–9. https://doi.org/10.1016/j.ijfatigue.2019.06.
[25] Radaj D, Sonsino CM, Fricke W. Recent developments in local concepts of fatigue 002http://www.sciencedirect.com/science/article/pii/S0142112319302324.
assessment of welded joints. Int J Fatigue 2009;31(1):2–11. https://doi.org/10. [56] Wächter M, Esderts A. On the estimation of cyclic material properties – part 2:
1016/j.ijfatigue.2008.05.019. introduction of a new estimation method. Mater Test 2018;60(10):953–9. https://
[26] Lee Y-L, Barkey M, Kang H-T. Metal fatigue analysis handbook: practical problem- doi.org/10.3139/120.111237.
solving techniques for computer-aided engineering. Oxford: Butterworth- [57] Muralidharan U, Manson SS. A modified universal slopes equation for estimation of
Heinemann; 2011. fatigue characteristics of metals. J Eng Mater Technol 1988;110(1):55. https://doi.
[27] Sonsino CM, Fricke W, de Bruyne F, Hoppe A, Ahmadi A, Zhang G. Notch stress org/10.1115/1.3226010.
concepts for the fatigue assessment of welded joints – background and applications. [58] Park J-H, Song J-H. Detailed evaluation of methods for estimation of fatigue
Int J Fatigue 2012;34(1):2–16. https://doi.org/10.1016/j.ijfatigue.2010.04.011. properties. Int J Fatigue 1995;17(5):365–73. https://doi.org/10.1016/0142-
[28] Mattos RJ, Lawrence FV. Estimation of the fatigue crack initiation life in welds 1123(95)99737-Uhttp://www.sciencedirect.com/science/article/pii/
using low cycle fatigue concepts. SAE SP-424; 1977. 014211239599737U.
[29] Lawrence FV, Ho NJ, Mazumdar PK. Predicting the fatigue resistance of welds. [59] Lee K-S, Song J-H. Estimation methods for strain-life fatigue properties from
Annu Rev Mater Sci 1981;11(1):401–25. https://doi.org/10.1146/annurev.ms.11. hardness. Int J Fatigue 2006;28(4):386–400. https://doi.org/10.1016/j.ijfatigue.
080181.002153. 2005.07.037http://www.sciencedirect.com/science/article/pii/
[30] Testin R, Yung J, Lawrence F, Rice R. Predicting the fatigue resistance of steel S0142112305002112.
weldements. Weld Res – Suppl Weld J 1987;66(4):93–8. [60] Wackerly DD, Mendenhall W, Scheaffer RL. Mathematical statistics with applica-
[31] Skorupa M. Fatigue life prediction of cruciform joints failing at the weld toe. Weld tions. 7th ed. Belmont, CA: Thomson Higher Education; 2008.
Res - Suppl Weld J 1992;71:269–75. [61] Barnston AG. Correspondence among the Correlation, RMSE, and Heidke Forecast
[32] Saiprasertkit K, Hanji T, Miki C. Fatigue strength assessment of load-carrying cru- Verification Measures. Refine Heidke Score, Weather Forecast 1992;7(4):699–709.
ciform joints with material mismatching in low- and high-cycle fatigue regions https://doi.org/10.1175/1520-0434(1992) 007<0699:CATCRA>2.0.CO;2.
based on the effective notch concept. Int J Fatigue 2012;40:120–8. https://doi.org/ [62] Hyndman RJ, Koehler AB. Another look at measures of forecast accuracy. Int J
10.1016/j.ijfatigue.2011.12.016. Forecast 2006;22(4):679–88. https://doi.org/10.1016/j.ijforecast.2006.03.001.
[33] Saiprasertkit K, Hanji T, Miki C. Local strain estimation method for low- and high- [63] Conn AR, Gould NIM, Toint PL. Trust-region methods, Vol. 1 of MPS-SIAM series on
cycle fatigue strength evaluation. Int J Fatigue 2012;40:1–6. https://doi.org/10. optimization, Society for Industrial and applied mathematics, Philadelphia, PA.;
1016/j.ijfatigue.2012.01.021http://www.sciencedirect.com/science/article/pii/ 2000.
S0142112312000382. [64] Rennert R, Kullig E, Vormwald M, Esderts A, Siegele D. Rechnerischer
[34] Ladinek M, Niederwanger A, Lang R, Schmid J, Timmers R, Lener G. The strain-life Festigkeitsnachweis für Maschinenbauteile: Aus Stahl, Eisenguss- und
approach applied to welded joints: considering the real weld geometry. J Constr Aluminiumwerkstoffen. 6th ed. VDMA: FKM-Richtlinie; 2012.
Steel Res 2018;148:180–8. https://doi.org/10.1016/j.jcsr.2018.04.024. [65] Hobbacher AF. Recommendations for Fatigue Design of Welded Joints and
[35] Niederwanger A, Ladinek M, Lener G. Consideration of imperfections and support Components. 2nd Edition, Cham: IIW Collection, Springer International Publishing;
effects in the fatigue assessment of welded cruciform joints. Mater Test 2020;62(1). 2016. https://doi.org/10.1007/978-3-319-23757-2.
https://doi.org/10.3139/120.111448. [66] Austrian Standards Institute, ÖNORM EN 1993-1-9: Design of steel structures. Part
[36] Basquin OH. The exponential law of endurance tests. Proc Am Soc Test Mater 1–9: Fatigue; 2013.
1910;10:625–30. [67] Bohlmann B. Zur Schwingfestigkeit von Schweißverbindungen: das örtliche
[37] Suresh S. Fatigue of materials. 2nd ed. Cambridge and New York: Cambridge Konzept am Beispiel einer Rohrverzweigung. https://doi.org/10.15480/882.981.

12
A. Niederwanger, et al. Engineering Structures 201 (2019) 109774

[68] González R, García JO, Barbés MA, Quintana MJ, Verdeja LF, Verdeja JI. Ultrafine [73] Madia M, Zerbst U, Beier HT, Schork B. The IBESS model – elements, realisation and
grained HSLA steels for cold forming. J Iron Steel Res Int 2010;17(10):50–6. validation. Eng Fract Mech 2018;198:171–208. https://doi.org/10.1016/j.
https://doi.org/10.1016/S1006-706X(10)60183-5. engfracmech.2017.08.033http://www.sciencedirect.com/science/article/pii/
[69] Lehto P, Romanoff J, Remes H, Sarikka T. Characterisation of local grain size var- S0013794417304885.
iation of welded structural steel. Weld World 2016;60(4):673–88. https://doi.org/ [74] Hörmann L. Experimentelle und numerische Untersuchung zur Ermüdungsfestigkeit
10.1007/s40194-016-0318-8. gekerbter Bauteile Master thesis University of Innsbruck, Innsbruck; 2019.
[70] Ladinek M, Niederwanger A, Lang R. An individual fatigue assessment approach [75] Schmid J. Geometrieoptimierung von Freischnitten ermüdungsbeanspruchter
considering real notch strains and local hardness applied to welded joints. J Constr Bauteile Master thesis University of Innsbruck, Innsbruck; 2015.
Steel Res 2018;148:314–25. https://doi.org/10.1016/j.jcsr.2018.06.005. [76] Lang R. Ein Beitrag zur Bestimmung der Anrisslebensdauer geschweißter Bauteile,
[71] Bosch A, Lang E, Vormwald M. Low cycle fatigue of seam welds – numerical si- Dissertation, Universität Innsbruck, Innsbruck; 2015.
mulation under consideration of material inhomogeneities. Proc Eng [77] FARO Technologies Inc., Laser Line Probe; 2019. URL <https://www.faro.com/>.
2014;74:218–27. https://doi.org/10.1016/j.proeng.2014.06.252http://www. [78] Lang R, Lener G. Application and comparison of deterministic and stochastic
sciencedirect.com/science/article/pii/S1877705814008224. methods for the evaluation of welded components’ fatigue lifetime based on real
[72] Zerbst U, Madia M, Schork B, Hensel J, Kucharczyk P, Ngoula D, et al. Fatigue and notch stresses. Int J Fatigue 2016;93:184–93. https://doi.org/10.1016/j.ijfatigue.
fracture of weldments: the IBESS approach for the determination of the fatigue life 2016.08.023.
and strength of weldments by fracture mechanics analysis. Springer International [79] ANSYS Inc., ANSYS 19.1; 2018. <https://www.ansys.com/>.
Publishing, Cham, Switzerland; 2019. https://doi.org/10.1007/978-3-030- [80] The MathWorks Inc., MATLAB R2019a; 2019. <https://www.mathworks.com/>.
04073-4.

13