materials

Article
Fatigue Assessment Strategy Using
Bayesian Techniques
Enrique Castillo 1,2 , Miguel Muniz-Calvente 3, * , Alfonso Fernández-Canteli 3
and Sergio Blasón 3
1 Royal Academy of Engineering, Don Pedro 10, 28005 Madrid, Spain; [email protected]
2 Royal Academy of Sciences, Valverde 22, 28004 Madrid, Spain
3 Department of Construction and Engineering Manufacturing, Polytechnic School of Gijón,
University of Oviedo, Campus de Viesques, 33206 Asturias, Spain; [email protected] (A.F.-C.);
[email protected] (S.B.)
* Correspondence: [email protected]




Received: 13 September 2019; Accepted: 30 September 2019; Published: 3 October 2019


Abstract: Different empirical models have been proposed in the literature to determine the fatigue
strength as a function of lifetime, according to linear, parabolic, hyperbolic, exponential, and other
shaped solutions. However, most of them imply a deterministic definition of the S-N field, despite the
inherent scatter exhibited by the fatigue results issuing from experimental campaigns. In this work,
the Bayesian theory is presented as a suitable way not only to convert deterministic into probabilistic
models, but to enhance probabilistic fatigue models with the statistical distribution of the percentile
curves of failure probability interpreted as their confidence bands. After a short introduction about
the application of the Bayesian methodology, its advantageous implementation on an OpenSource
software named OpenBUGS is presented. As a practical example, this methodology has been applied
to the statistical analysis of the Maennig fatigue S-N field data using the Weibull regression model
proposed by Castillo and Canteli, which allows the confidence bands of the S-N field to be determined
as a function of the already available test results. Finally, a question of general interest is discussed as
that concerned to the recommendable number of tests to carry out in an experimental S-N fatigue
program for achieving “reliable or confident” results to be subsequently used in component design,
which, generally, is not adequately and practically addressed by researchers.

Keywords: fatigue; bayesian model; openbugs software; density function; confidence bands

1. Introduction
Repeated application of variable loads over time may lead to fatigue failure of real structures and
components. Besides, this type of failure occurs often unexpectedly because the magnitude of the
stresses acting on these components usually remains far below the static material strength. Therefore,
accurate estimation of the fatigue strength of materials is crucial to ensure safe design and maintenance
of structures and components.
In the literature, there is an extensive list of models devoted to the study of material fatigue
strength, which are focused on predicting the service life (N) in terms of a particular generalized
parameter (GP), such as the equivalent range of stresses (∆σ), strains (∆ε) or combinations of both
(Smith–Watson–Topper, etc.). However, most of the models implied are deterministic, despite the
fact that one of the main characteristics of the fatigue problem is its associated uncertainty. The high
variability is inherent to the fatigue phenomenon forces fatigue failure to be considered as a random
phenomenon, so that only models, including random variables should be considered valid for predicting
fatigue life reliably. For that reason, there are different standards and guidelines, such as ISO [1],

Materials 2019, 12, 3239; doi:10.3390/ma12193239 www.mdpi.com/journal/materials

Materials 2019, 12, 3239 2 of 17

ASTM [2] or ASM [3], that include specific sections to suggest some recommendations for the statistical
analysis of the fatigue data. Furthermore, among the different probabilistic fatigue modes proposed in
the literature, the Weibull probabilistic regression model proposed by Castillo–Canteli [4], the same as
those proposed by Freudenthal, Gumbel and Bolotin [5–8], defines the S-N or ε-N fields as hyperbolic
percentile curves representing the same probability of failure.
Despite the satisfactory probabilistic definition of the S-N field and the robust parameter estimation
provided by the model of Castillo-Canteli, as extensively proven after being applied under consideration
of different driving forces to different materials [9–12], a question still remains open and repeatedly
unanswered in academic forums: “How many experimental tests should be performed to achieve
a suitable probabilistic definition of the S-N field?”. Alternatively, even more precisely, “[h]ow many
experimental tests should be performed to define the percentiles curves and fatigue limit for a given
precision and confidence level”? The question is perhaps motivated by the fact that confidence
intervals of the percentiles shrink according to the increasing number of available results as in the
case of an extensive experimental campaign. Nevertheless, according to the concept of confidence
bands [13,14] it can be concluded that both questions are incorrectly posed, since they cannot be
answered “a priori” because of the complex relation implied among the number of tests and parameter
values, or even because the influence of the suitability or unsuitability of the test strategy is applied.
Only after testing and data assessment, it means “a posteriori”, the problem can be formulated in
the following terms: The fatigue results for a certain probability of failure and precision can be
statistically provided at any time of the experimental campaign irrespective of the number of tests
available, but we have to count on the results being consequently penalized as a function of the
number of tests: The smaller the number of test the higher the penalty applied in the derivation of the
value of the particular fatigue property estimated for a given probability of failure according to the
precision required.
In the past, other procedures, as for instance the bootstrap method [15–18], were applied to face the
reliability question of the S-N field, but their high computational cost together with their questionable
applicability when applied to small samples have contributed them not to be extensively widespread
and consequently applied in the practical current fatigue characterization. Furthermore, different
attempts have been made to incorporate the Bayesian methodology to the evaluation of fatigue results
and fatigue lifetime prediction [19] under low and high cycle fatigue conditions from deterministic
fatigue models, although using a mathematical-statistical formulation less accessible for laboratory
and practitioner engineers. In those contributions, the advantages of applying Bayesian procedures are
emphasized as a way of enhancing fatigue design reliability by incorporating technological knowledge
from theoretical studies and previous experimental experience in particular in the case of a scarce
number of tests available [19,20].
As will be highlighted in this work, a satisfactory solution for the calculation of the confidence
bands of the problem will be found by applying the Bayesian theory, the response of the approach
becoming now simple, unambiguous and rigorous from a statistical viewpoint particularly when
applied to a probabilistic model of assessment.
The practical application of the Bayes methodology is facilitated by means of the application
of a powerful OpenSource and free software named OpenBUGS, which despite its currently limited
dissemination in the engineering area, might completely transform the way in which Bayesian methods
will be applied in the future. The implementation of the OpenBUGS software into the Bayesian theory
is basically explained, and thereafter applied to the assessment of the fatigue S-N data provided by
Maennig [21–23] according to the Weibull regression model proposed by Castillo and Canteli [4],
which allows the confidence bands of the S-N field to be determined as a function of the already
available test results. In the assessment of this practical case, not only the variability of the model
parameters is calculated, but also the confidence bands for each percentile failure curves are determined.
In this way, the fatigue strength or lifetime can be estimated for any probability of failure and given
precision level. Finally, the main conclusions of the work are summarized.
Materials 2019, 12, 3239 3 of 17

Furthermore, a question of general interest is discussed as that concerned to the recommendable

number of tests to carry out in an experimental S-N fatigue program for achieving “reliable or confident”
results to be subsequently used in component design, which, generally, is not adequately and practically
addressed by researchers.

2. Fundamentals Bases of the Model

In the following, a general description of the main features of the Weibull regression model
proposed by Castillo-Canteli for analyzing the S-N field is made and the implementation of the
OpenBugs software into the Bayesian methods are explained in order to facilitate the comprehension
of the methodology proposed and its practical application to the case of fatigue data assessment.

2.1. The Probabilistic Fatigue Model

The S-N field solution for lifetime prediction proposed by Castillo and Canteli [4] is a probabilistic
model, based on physical and statistical conditions to be necessarily fulfilled by any valid fatigue model,
in particular on the necessary compatibility condition existing between the driving force (generally
identified with the stress range, ∆σ) distribution for given number of cycles, and the lifetime distribution
for given driving force. As shown graphically in Figure 1 for the two discretional intersecting straight
lines parallel to both axes, each S-N curve pertaining to the S-N field that crosses the left part of the
horizontal line must necessarily cross the lower segment of the vertical line as well, so that the areas
of both shaded zones pertaining to the probability density distribution must be identical. Due to
statistical considerations, see References [4,7,8], Weibull distributions are the most suitable ones for
being proposed to model fitting. As a result, the following equality arises:

E(N∗ , ∆σ∗ ) = F(∆σ∗ , N∗ ), (1)

where E(N∗ , ∆σ∗ ) and F(∆σ∗ , N∗ ) are the cdf of the lifetime N∗ given a certain stress range ∆σ∗ and the
cdf of the stress range ∆σ∗ given a certain fatigue lifetime N∗ , respectively. Equation (1) represents
a functional equation [24–26], which provides the only two possible solutions for the probabilistic
S-N field represented by failure hyperbolic shaped percentiles curves, i.e., the curves representing
the number of cycles to failure for a certain probability of failure [4]. Due to physical conditions,
only the solution represented by Equation (2) is considered to be acceptable, where N0 and ∆σ0 are the
characteristic asymptotes (limit number of cycles and fatigue endurance limit) and λ, δ and β are the
location, scale and shape Weibull model parameters, respectively:
β 
  log N log ∆σ − λ  
     
N0 ∆σ0
p f ailure = 1 − exp−
   
(2)
δ
 
   
Materials 2019, 12, x FOR PEER REVIEW 3 of 17

1. Illustration
FigureFigure ofofthe
1. Illustration the compatibility condition
compatibility condition in thein the
S-N S-N field.
field.

2. Fundamentals Bases of the Model

In the following, a general description of the main features of the Weibull regression model
proposed by Castillo-Canteli for analyzing the S-N field is made and the implementation of the
OpenBugs software into the Bayesian methods are explained in order to facilitate the comprehension
of the methodology proposed and its practical application to the case of fatigue data assessment.
Materials 2019, 12, 3239 4 of 17

Equation (2) provides the probabilistic definition of the S-N field, i.e., the p-S-N field, which can be
advantageously assessed by assuming the normalized variable V = log(N/N0 )log(∆σ/∆σ0 ). In this
way, all the S-N field results are evaluated as pertaining to a single Weibull cumulative distribution
function, thus, enhancing reliability in the parameter estimation. The satisfactory application of the
model to practical S-N fatigue programs is confirmed for different materials and driving forces. The free
and easy-to-use ProFatigue software is available to facilitate parameter estimation [27]. According to
this classical concept of probabilistic approach, the five parameters of the model, i.e., λ, δ, β, N0 and
∆σ0 , are estimated as if they were fixed values. Nevertheless, the reliability of the parameters obtained
during model fitting depends, obviously, on the total number of tests performed and the suitability
of the test strategy applied in the experimental program. As a result, a sounder evaluation requires
determining the densities of the estimated parameters or the confidence bands of the percentiles curves.
Until now, the calculation of the confidence intervals was achieved by applying the Bootstrap
method following a procedure consisting in a previous estimation of the model parameters for the
test data available and successive application of the Monte-Carlo simulation based on the model
parameters just obtained following a faithful replication of the fatigue program planning adopted.
The model parameters for the newly resulting simulated S-N field are estimated by applying the same
probabilistic model, so that, the remaining outgoing simulated S-N fields provide the variability field
of the model parameters (confidence bands) or, alternatively, the confidence bans for the particular
percentile curve of interest. However, this method is usually computationally expensive and does
not always report satisfactory results. As an alternative, Bayes’ methodology and its efficiency are
investigated here, when applied to an advanced model, as that represented by the probabilistic S-N
field solution proposed by Castillo and Canteli corresponding to Equation (2).

2.2. Bayesian Methods

First of all, it should be noted that the main difference between conventional and Bayesian statistical
models is that the formers ones accept the model parameter estimates as fixed or deterministic values,
whereas the second ones consider them as random variables. This applies to both deterministic and
probabilistic models. In the application of Bayesian methods, an initial family of parametric models
is assumed in which the parameters are considered to be random variables rather than constants.
This means that deterministic models are transformed automatically in probabilistic ones, so they are
improved. The probability distribution for each of the parameters involved, i.e., λ, δ, β, N0 and ∆σ0 is
the goal to be achieved [5] when Bayesian methods are applied to the probabilistic S-N fatigue model
considered in Section 2.1, and therefore, not their point estimations. In the following, the Bayesian
methodology is shortly introduced for highlighting its application to fatigue model, including a brief
description of the four fundamental steps to be applied, namely, (a) definition of the prior distribution,
(b) estimation of the prior predictive distribution, (c) derivation of the posterior distribution and
(d) calculation of the posterior predictive distribution.

2.2.1. Prior Distribution

Bayesian methods are initiated assuming a prior distribution of the parameter, which may
reflect the initial engineer’s knowledge about the parameters of the model and their uncertainty.
This distribution can be “non-informative”, when there is no information about the parameters,
or “informative”, when it contains some knowledge based on previous experience.
For example, a parametric model with a parameter vector θ is defined to represent a variable X
taking x values, such as:
X ∼ p(x θ). (3)

The parameters θ of this function are random and are associated with a prior distribution function:

θ ∼ p(θ α), (4)

Materials 2019, 12, 3239 5 of 17

where α is a vector of the parameter distributions, called hyper-parameter to be distinguished from the
initial model parameters (θ). The prior distribution may be discrete, continuous or mixed, depending
on the type of parameters used.

2.2.2. The Prior Predictive Distribution

Based on the description of the model and the distribution functions of the prior parameters
already defined, it is possible to use the model to make predictions. Based on the description of the
model and the distribution functions of the prior parameters already defined, it is possible to use the
model to make predictions e x:
Z  
p(ex θ) = pXe θ p(θ|α)dθ (5)
θ

It is worth mentioning that this model only contains the information provided by the scientist or
engineer knowledge on the variability of the model parameters without including any information
related to evidence shown in experimental results.

2.2.3. The Posterior Distribution

The knowledge about the parameters is complemented by random samples, which lead to
a posterior distribution of the parameters, including the two sources of knowledge. Thus, once a sample
of size n is obtained from the population to be modelled:

X = (x1 , x2 , . . . , xn ). (6)

It is possible to improve the predictions made by applying the maximum likelihood method by
combining the prior distribution and the information provided by the sample:

p(X|θ) ≡ L(X|θ). (7)

At this point, the posterior distribution of the parameters may be derived by means of the Bayes
theorem expressed by the following formula:

p(X|θ)p(θ|α)
p(X|θ, α) = . (8)
p(X|α)

2.2.4. The Posterior Predictive Distribution

Finally, the posterior predictive distribution is obtained using the posterior distribution from the
previous step, which includes both the prior provided information and the experimental evidence
collected in the posterior distribution of the model parameters, thus, allowing predictions (e x) to
be made: Z  
x X, α) =
p(e pXe θ p(θ|X, α)dθ. (9)
θ

Note that Bayesian models do not work with the initial family of distributions, but with the
predictive distribution, i.e., a convex combination, possibly infinite, of a set of models from this family,
whose weights or coefficients arise from the posterior distribution. In this way, an extended family of
the initially selected family of distributions is found. This allows the sample to achieve a better model
fit. Note that the resulting models are extensions of the initially assumed family of models because by
considering Dirac distributions as posterior distributions, the models of the initial family are obtained.
This extension plays a significant role in the real engineering practice because it provides high
flexibility in building models allowing better correlation between theoretical models and the reality to
be attained as pursued in modelling fatigue problems.
 Materials 2019, 12, 3239 6 of 17

2.2.5. Application Example

With the aim of illustrating the different distributions, i.e., prior, prior predictive, posterior and
posterior predictive distributions presented in the previous subsections, an example of the application
Materials 2019, 12, x FOR PEER REVIEW 6 of 17
of Bayesian Methods is presented, in which a very simplified model that follows a normal random
Materials 2019, 12, x FOR PEER REVIEW 6 of 17
distribution
application N(θ,1) isMethods
of Bayesian assumed. is presented, in which a very simplified model that follows a normal
random Indistribution
applicationthis example,N(θ,1)
of Bayesian theisnormal
assumed.
Methods israndom
presented,model only aimplies
in which one internal
very simplified model vector
that follows a normalθ,
of parameters,
which
In this
random is distribution
assumed
example, theto N(θ,1)
be random
normal as well
is random
assumed. following
model a uniform
only implies one internal vectorθ~U(4,8)
distribution (see dot
of parameters, θ, line in
Figure
which 2). Asexample,
isInassumed
this previouslythedescribed,
to be random normalas well the initial
following
random model random
a uniform distribution
distribution
only implies ofθ~U(4,8)
the parameters
one internal (seeofdot
vector represents
line in
parameters, the
θ,
Figure
prior 2). As previously
distribution, and described,
its parameters the initial
are random
called distribution
hyperparameters. of the parameters
Therefore, the
which is assumed to be random as well following a uniform distribution θ~U(4,8) (see dot line in represents
prior the
distribution in this
prior distribution,
example
Figure 2). As and
would beits parameters
the
previously uniform are called
distribution,
described, hyperparameters.
the initialand Therefore, the
the hyperparameters
random distribution prior
will
of the distribution
be in this the
4 and 8.represents
parameters
example would be the
prior distribution, and uniform distribution,
its parameters and the
are called hyperparameters
hyperparameters. will be 4the
Therefore, and 8. distribution in this
prior
example would be the uniform distribution, and the hyperparameters will be 4 and 8.

Figure 2. An example of a normal N(θ,1) model with a uniform U(4,8) prior distribution of θ.
Figure 2. An example of a normal N(θ,1) model with a uniform U(4,8) prior distribution of θ.
Figure 2. An example
Once the model and theofprior
a normal N(θ,1) model
distribution havewith
beena defined
uniform (θ~U(4,8)),
U(4,8) prior itdistribution
is possibleoftoθ.obtain
Once the model and the prior distribution have been defined (θ~U(4,8)), it is possible to obtain
the prior predictive distribution (see blue line in Figure 2):
the prior
Once predictive
the model distribution (seedistribution
and the prior blue line inhave Figure been 2): defined (θ~U(4,8)), it is possible to obtain
the prior predictive distribution
p x| (see (
blue ) 
line
( )
in
( )
 4,8 = fZU( 4 ,8) θ f N(θ ,1) x dθ .
Figure 2):
(10)
fU(4,8) (θ) fN(θ,1) (e
( ) 
p(ex 4, 8) θ= x)dθ. (10)
p x| ( ) ( )
 4,8 =θ fU( 4 ,8) θ f N(θ ,1) x dθ . (10)
After that, a sample of size n = 10 is obtained (see points on the horizontal axis in Figure 3) and
θ
the prior distribution
After improved,
that, a sample of size n = 10 in
resulting is aobtained
posterior(see distribution
points onofthe thehorizontal
parametersaxis
(seeindot line 3) and
Figure
in the After
Figure 3), that, a sample of size n = 10 is obtained (see points on the
prior distribution improved, resulting in a posterior distribution of the parameters (see dot horizontal axis in Figure 3) and
line
the prior distribution improved, resulting in a posterior distribution of the parameters (see dot line
in Figure 3),
in Figure 3), (
p θ|X ,U ( 4 ,8) ∝
p(θ X, U (4, 8)) ∝
) ∏
Y f N (θ ,1) ( xi ) fU ( 4 ,8) (θ ) .
fN(θ,1) (xi ) fU(4,8) (θ).
(11)
(11)
i =1,10

p (θ|X ,U ( 4 ,8) ) ∝ ∏f
i=1,10
N (θ ,1) ( xi ) fU ( 4 ,8) (θ ) . (11)
i =1,10

Figure 3. Illustration of obtaining the posterior distribution by updating the prior distribution using
a sample of data.

Figure3.3. Illustration
Figure Illustration of obtaining the posterior
posterior distribution
distributionby
byupdating
updatingthe
theprior
priordistribution
distributionusing
using
Finally, the posterior predictive distribution is obtained (see violet line in Figure3), which can be
aasample
sampleof
ofdata.
data.
considered the final predictive model:

Finally, the posterior predictive

p x|X (
 ,U ( 4 ,8distribution )  ) p(θ |X ,U ((see
) = f N (θ ,1)is( xobtained 4 ,8) )violet
dθ . line in Figure3), which
(12)
can be
considered the final predictive model: θ

x|X (
 ,U ( 4parameters
As noted, Bayesian methodspconsider ,8) = f N (θ (θ)
obtain convex combinations of the initial model family.
) 
,1) ( x ) p(θ |X ,U ( 4 ,8) ) dθ
as random variables . and subsequently
In fact, the distribution, actually assumed at
(12)
θ

As noted, Bayesian methods consider parameters (θ) as random variables and subsequently
Materials 2019, 12, 3239 7 of 17

Finally, the posterior predictive distribution is obtained (see violet line in Figure 3), which can be
considered the final predictive model:
Z
p(e
x X, U (4, 8)) = x)p(θ X, U (4, 8))dθ.
fN(θ,1) (e (12)
θ

As noted, Bayesian methods consider parameters (θ) as random variables and subsequently
obtain convex combinations of the initial model family. In fact, the distribution, actually assumed at
the end of the method, is a weighted or mixed distribution, which results after combining the prior or
posterior distribution with the initial distributions family.

3. The Proposed Model

In this section, the application of the Bayesian method to the probabilistic S-N field model of
Castillo–Canteli, see Equation (2), is presented with the five parameters of the model being considered
as random variables. To this end, the OpenBUGS software package has been used, which, given
prior distributions and a sample of experimental results, generates large samples of the posterior
model based on Markov Chain Monte Carlo (MCMC) techniques. This means that a large sample
of any statistic can be accomplished and, consequently, an excellent approximation of its probability
distribution is achieved.
On the following, the implementation of the Bayesian Weibull fatigue Model in OpenBUGS is
introduced in its two variants: Code text in flat format and visual programming by the Doodle module.

3.1. Model Implementation Through Openbugs Code

The only extreme value distribution implemented in OpenBUGS is the generalized extreme value
for maxima; so that the Castillo-Canteli model, which is based on a Weibull extreme value for minima,
must be implemented by this function (See Appendix A).
Once the model is defined, the implied parameters (N0 , ∆σ0 , β, λ, δ) are defined in the prior as
uniform random ones, ranging between a minimum and a maximum value:
 
N0 ∼ U N0min , N0max ,
 
∆σ0 ∼ U ∆σ0min , ∆σ0max ,
β ∼ U (βmin , βmax ), (13)
λ ∼ U (λmin , λmax ),
δ ∼ U (δmin , δmax ).

Equation (13) are the prior distributions associated with the five parameters of the Castillo-Canteli
model. The values of these hyperparameters have been established in a limited range of the
neighbourhoods of N0cc , ∆σ0cc , βcc , λcc and δcc parameters as follows:

N0min = 0.7 N0cc ; N0max = 1.5 N0cc ,

∆σ0min = 0.8 ∆σ0cc ; ∆σ0max = 1.2 ∆σ0cc ,
β
βmin = 1.5
cc
; βmax = 1.5 βcc ; (14)
λcc
λmin = 3 ; λmax = 2 λcc ,
δmin = δcc2 ; δmax = 1.5 δcc ,

where the values of ∆σ0cc , βcc , λcc and δcc are those obtained by implementing the original version
Castillo–Canteli model using, for example, the current version of the ProFatigue software [27].
Alternatively to the implementation of the Castillo–Canteli model by the generalized extreme function,
the ones/zeros Poisson model trick can be used to generate the model in an easier way (see Appendix A).
 Materials 2019, 12, 3239 8 of 17

3.2. Model Implementation Using Graphic Doodle in OpenBUGS

Materials 2019, 12, x FOR PEER REVIEW
OpenBUGS has a graphical programming environment called Graphic Doodle, which can be used 8 of 17
to program the previous code easily. Figure 4 shows the acyclic graph of the Bayesian network of the
of the model.model.
As can Asbe
canseen, the
be seen, theprior distributions
prior distributions of theof the
five five parameters
parameters haveoutside
have been defined been defined
the outside
model loop as in the previous case.
the model loop as in the previous case.

Materials 2019, 12, x FOR PEER REVIEW 8 of 17

of the model. As can be seen, the prior distributions of the five parameters have been defined outside
the model loop as in the previous case.

Figure 4. Block diagram of the Castillo and Canteli model introduced with Graphic Doodle in
Figure 4. Block diagram of the Castillo and Canteli model introduced with Graphic Doodle in OpenBUGS.
OpenBUGS.
Figure 4. Block diagram of the Castillo and Canteli model introduced with Graphic Doodle in
3.3. Execution of the Code in OpenBUGS and Analysis of the Results Provided by the Program
OpenBUGS.3.3. Execution of the Code in OpenBUGS and Analysis of the Results Provided by the Program.
Once the model has been implemented into the program, it is necessary to define the initial values
of theOnce the model
variables has been
(Equation implemented
(14)) and into the data
the experimental program,
of theitfatigue
is necessary
life thattotodefine the by
be fitted initial
the
3.3. Executionmodel,
of theof
values Code
the in OpenBUGS
variables (Equation and
(14)) Analysis
and the of the
experimental
giving rise to the posterior distributions (Figure 5). Results
data of Provided
the fatigue by
life the
that to Program.
be fitted by
the model, giving rise to the posterior distributions (Figure 5).
Once the model has been implemented into the program, it is necessary to define the initial
values of the variables (Equation (14)) and the experimental data of the fatigue life that to be fitted by
the model, giving rise to the posterior distributions (Figure 5).

Figure
Figure 5. Example of
5. Example of definition
definition of
of the
the initial
initial values
values of
of the
the variables
variables and
and the
the experimental
experimental fatigue
fatigue data
data
for a hypothetic fatigue case assessment.
for a hypothetic fatigue case assessment.

Finally, the number of simulations is defined for the initial process (burn-in process) and for the
Finally, the number of simulations is defined for the initial process (burn-in process) and for the
final sample, and the model is executed. In this case, taking into account the computational time of
final sample, and the model is executed. In this case, taking into account the computational time of
OpenBugs and the scatter associated with the fatigue problems, it is proposed to use 1,000 and 20,000
OpenBugs and the scatter associated with the fatigue problems, it is proposed to use 1,000 and 20,000
simulations for the initial and final samples, respectively.
simulations for the initial and final samples, respectively.
Once the
Figure 5. Example code is executed,
of definition of thetheinitial
program provides
values both
of the a convergence
variables and theanalysis and the model
experimental fatigue data
parameters of the posterior distributions, which are introduced into Equation (12) to derive the
for a hypothetic
posteriorfatigue case
predictive assessment.
distribution. In this way, the percentiles of the percentiles the fatigue failure
probability are obtained; in other words, the confidence intervals of each percentile as obtained by
Materials 2019, 12, 3239 9 of 17

Once the code is executed, the program provides both a convergence analysis and the model
parameters of the posterior distributions, which are introduced into Equation (12) to derive the posterior
predictive distribution. In this way, the percentiles of the percentiles the fatigue failure probability are
obtained; in other words, the confidence intervals of each percentile as obtained by the original version
of the Castillo–Canteli model.

4. Practical Example
In this section, the Maennig fatigue data [21–23], (see on Table 1), is used to illustrate the method
proposed in the previous sections and its implementation in OpenBUGS. This extensive fatigue program
comprising 360 results, i.e., an exceptionally high amount of data, that is not impaired, or questions the
potential application of Bayesian methodology to assess fatigue programs with a limited amount of
data. Its selection only obeys to a unique opportunity of observing and discussing the evolution of the
confidence limits as a function of the number of results available.

Table 1. Maennig fatigue data.

∆σ(MPa) Lifetime (Thousands of Cycles) (N)

385 51,57,60,67,68,69,75,76,82,83,87,95,106,109,111,119,122,128,132,140
380 59,66,69,80,87,90,97,98,99,100,107,109,117,118,125,128,132,158,177,186
375 65,71,78,84,89,93,98,103,105,109,113,118,124,131,147,156,171,182,199,220
370 83,98,100,104,110,111,122,125,132,136,141,143,146,155,165,194,200,201,251,318
365 89,105,108,118,119,121,130,133,152,164,170,181,182,192,199,211,238,273,324,398
360 117,127,141,151,162,173,181,186,192,198,203,209,218,255,262,288,295,309,394,585
355 112,125,133,156,166,168,173,202,227,247,253,261,285,286,309,365,442,559,702,852
350 115,129,143,169,177,178,218,230,271,280,285,305,326,342,381,431,493,568,734,1101
345 140,155,169,174,218,248,265,293,321,326,348,350,364,374,397,426,461,504,738,1063
340 146,159,168,224,246,253,291,326,358,385,397,425,449,498,532,610,714,763,987,1585
335 154,180,210,254,305,332,363,415,457,482,528,559,593,611,678,767,835,957,1274,1854
330 166,184,241,251,273,312,371,418,493,562,683,760,830,981,1306,1463,1842,1867,2220,2978
325 196,227,250,271,308,347,393,475,548,669,799,879,975,1154,1388,1705,2073,2211,2925,4257
206,231,283,370,413,474,523,597,605,619,727,815,935,1056,1144,1336,1580,1786,1826,1943,2214,
320
3107,4510,6297
226,257,307,370,457,549,570,590,672,781,850,974,1093,1460,1477,1936,2662,2731,3487,4396,5803,
315
7215
310 206,317,393,446,502,570,627,809,956,1022,1327,1745,2001,2139,2314,3425,4576,5453,7868,8297
305 253,311,329,370,726,845,935,954,1139,1456,1792,2578,3776,5161,8131
300 411,606,700,707,919,1587,1595,2295,4628,6280
295 503,1191,1282,1609,4070,6337
290 1055,1369
285 1220

Firstly, the data has been fitted following the standard procedure of the Castillo–Canteli model
using the ProFatigue software, in order to obtain an initial approximation of the model parameters
values (see Equation (15)) and an initial estimation of the p-S-N field (see Figure 6).

N0cc = 14958 cycles,∆σ0cc = 257.881 MPa,βcc = 2.97,λcc = 0.34,δcc = 0.56. (15)

Those values have been used to define the uniform prior distributions (Equation (13)) according
to Equation (14).
Materials 2019, 12, 3239 10 of 17
Materials
Materials 2019,
2019, 12,
12, xx FOR
FOR PEER
PEER REVIEW
REVIEW 10
10 of
of 17
17

Figure
Figure 6.
Figure 6. p-S-N field
6. p-S-N
p-S-N field obtained
field obtained by
obtained by the
by the standard
the standard approach
standard approach of
approach of the
of the Castillo–Canteli
the Castillo–Canteli Model.
Castillo–Canteli Model.
Model.

Thereafter, data
Thereafter, has been
been entered into OpenBUGS, and
and the simulations
simulations have been
been carried out
out to
Thereafter, data
data has
has been entered
entered into
into OpenBUGS,
OpenBUGS, and the
the simulations have
have been carried
carried out to
to
calculate
calculate the posterior distributions of the model parameters (see Figure 7).
calculate the
the posterior
posterior distributions
distributions of
of the
the model
model parameters
parameters (see
(see Figure
Figure 7).
7).

7. Posterior distributions
Figure 7.
Figure distributions of the
the model parameters
parameters fitted to
to the Maennig
Maennig data.
Figure 7. Posterior
Posterior distributions of
of the model
model parameters fitted
fitted to the
the Maennig data.
data.
Finally, the
Finally, the posterior distributions
distributions of the
the parameters have have been used
used to obtain
obtain the posterior
posterior
Finally, the posterior
posterior distributions of of the parameters
parameters have been been used toto obtain the
the posterior
predictive
predictive distribution of the model as the 0.01, 0.10, 0.50, 0.90 and 0.99 quantiles of the S-N curves
predictive distribution
distribution of
of the
the model
model as
as the
the 0.01,
0.01, 0.10,
0.10, 0.50,
0.50, 0.90
0.90 and
and 0.99
0.99 quantiles
quantiles of
of the
the S-N
S-N curves
curves
(see lines
(see lines in
in Figure
Figure8),
8),and
andthethecorresponding
corresponding0.01–0.99
0.01–0.99confidence
confidence intervals (see shaded regions
(see lines in Figure 8), and the corresponding 0.01–0.99 confidence intervals (see shaded regions in
intervals (see shaded regions in
in Figure
Figure 8).
Figure 8).
8).
Materials 2019, 12, 3239 11 of 17
Materials 2019, 12, x FOR PEER REVIEW 11 of 17

Figure 8.
Figure The 0.01,
8. The 0.01, 0.10,
0.10, 0.50,
0.50, 0.90,
0.90, and
and 0.99
0.99 quantile
quantile S-N
S-N curves
curves and
and the
the 0.01–0.99
0.01–0.99 confidence
confidence intervals
intervals
for the Maennig data.
for the Maennig data.
Influence of the Number of Fatigue Tests Performed
4.1. Influence of the Number of Fatigue Tests Performed.
As mentioned above, the number of tests has significant influence on the results of the fitting
As mentioned above, the number of tests has significant influence on the results of the fitting
process of a p-S-N curve. For that reason, the experimental program developed by Maennig, as an almost
process of a p-S-N curve. For that reason, the experimental program developed by Maennig, as an
ideal case of fatigue data provided in the literature, may not be representative to check the suitability
almost ideal case of fatigue data provided in the literature, may not be representative to check the
of the S-N assessment using the new probabilistic model. Nevertheless, the utility of the Bayes model
suitability of the S-N assessment using the new probabilistic model. Nevertheless, the utility of the
and that of the fatigue model proposed can be checked, assuming that the virtual planning of the test
Bayes model and that of the fatigue model proposed can be checked, assuming that the virtual
program consists of limited samples, which will be randomly select from the original Maennig’s test
planning of the test program consists of limited samples, which will be randomly select from the
program. The number test results considered for the assessment will be gradually increased, in order
original Maennig’s test program. The number test results considered for the assessment will be
to observe their influence on the final S-N field. After their fitting, the evolution in the assessment of
gradually increased, in order to observe their influence on the final S-N field. . After their fitting, the
the fatigue program concerning the failure percentiles and fatigue limit distribution is investigated.
evolution in the assessment of the fatigue program concerning the failure percentiles and fatigue limit
To this aim, a random permutation of the 360 results contained in the vector is generated,
distribution is investigated.
providing a simulation of the experimental test sequence as virtually performed by Maennig. Thereafter,
To this aim, a random permutation of the 360 results contained in the vector is generated,
the Bayesian Weibull Fatigue Model is applied to fit initially only the first ten experimental results
providing a simulation of the experimental test sequence as virtually performed by Maennig.
furnished by the random permutation vector (see Figure 9). The procedure is continued to define the
Thereafter, the Bayesian Weibull Fatigue Model is applied to fit initially only the first ten
evolution of the 0.01, 0.10, 0.50, 0.90 and 0.99 percentiles and the corresponding confidence intervals
experimental results furnished by the random permutation vector (see Figure 9). The procedure is
when the initial sample with 10 results is enlarged stepwise up to 20, 50, 100, 150, 200, 250 and 300 tests,
continued to define the evolution of the 0.01, 0.10, 0.50, 0.90 and 0.99 percentiles and the
respectively, as represented in Figure 9 and Video S1. As expected, the confidence intervals shrink
corresponding confidence intervals when the initial sample with 10 results is enlarged stepwise up
with the number of tests, but in a non-linear way (see Figure 9).
to 20, 50, 100, 150, 200, 250 and 300 tests, respectively, as represented in Figure 9 and Video S1. As
expected, the confidence intervals shrink with the number of tests, but in a non-linear way (see Figure
9).
Materials 2019, 12, 3239 12 of 17
Materials 2019, 12, x FOR PEER REVIEW 12 of 17

Figure 9.
Figure 9. Evolution
Evolution of
of the
the whole
whole p-S-N
p-S-N field
field (S-N
(S-N percentile
percentile curves
curves and
and confidence intervals) with
confidence intervals) with the
the
number of
number of tests
tests implied
implied in
in the
the fitting
fitting process.
process.

Despite thethe useful

usefulinformation
informationprovided
providedby bythe
thegraphs in in
graphs Figure
Figure9 showing
9 showingthethe
evolution
evolutionof the
of
confidence
the bands,
confidence andand
bands, thus, the the
thus, dependency
dependencyof the whole
of the p-S-N
whole p-S-Nfield with
field thethe
with number
number of of
cycles, a
cycles,
abetter
betterway
wayto to observe theconfidence
observe how the confidencebands
bandsareare evolving
evolving forfor an increasing
an increasing number
number of fatigue
of fatigue data
data available is to y cut the p-S-N field horizontally at fixed Δσ and record the number of cycles
obtained for the different cases. Figure 10 represents the evolving results of the confidence bands for
the 0.01, 0.1, 0.5, 0.90 and 0.99 percentiles, respectively, when such cuts are applied at Δσ = 290,320,350
Materials 2019, 12, 3239 13 of 17

available
Materials is to12,
2019, y cut
x FORthe p-S-N
PEER field horizontally at fixed ∆σ and record the number of cycles
REVIEW obtained
13 of 17
for the different cases. Figure 10 represents the evolving results of the confidence bands for the 0.01,
0.1,and
0.5,380
0.90MPa.
and As 0.99expected, all confidence
percentiles, bands
respectively, irrespective
when such cuts of are
the applied
percentile ∆σ =stress
at and ratio
290,320,350 and
considered shows a reduction for the increasing number of tests involved in the fitting
380 MPa. As expected, all confidence bands irrespective of the percentile and stress ratio consideredprocess. The
calculation of those graphs allows researchers to decide when the experimental running fatigue
shows a reduction for the increasing number of tests involved in the fitting process. The calculation of
program should be finished according to a criterion of maximization of the reliability-cost rate, i.e.,
those graphs allows researchers to decide when the experimental running fatigue program should be
when a new test does not provide a profitable increment of reliability.
finished according to a criterion of maximization of the reliability-cost rate, i.e., when a new test does
not provide a profitable increment of reliability.

Figure 10. Detailed evolution of the quantiles and confidence intervals with the number of tests
Figure Detailed
10. in
involved evolution
the fitting process forofdifferent
the quantiles and confidence intervals with the number of tests
stress levels.
involved in the fitting process for different stress levels.
In Figure 11 the evolution of the distribution function of the fatigue limit, in particular for 0.01,
Inand
0.5, Figure 11 the evolution
0.99 probability of is
of failure the distribution
shown function
as a function of theof thenumber
total fatigueoflimit, in particular
specimens tested byfor 0.01,
0.5,applying
and 0.99the probability of failure is shown as a function of the total number of
Bayesian analysis. It reminded that the fatigue limit, Δσ0, is one of the parametersspecimens tested by
applying
interveningthe Bayesian analysis. Itmodel,
in the Castillo-Canteli is reminded that the
see Equation (2),fatigue ∆σhorizontal
limit,the
representing 0 , is one of the parameters
asymptote,
i.e., the driving
intervening in the force below which no
Castillo-Canteli fatiguesee
model, damage occurs.
Equation Itsrepresenting
(2), statistical distribution, as provided
the horizontal asymptote,
by the S-N field, has great significance in the characterization of the fatigue behavior
i.e., the driving force below which no fatigue damage occurs. Its statistical distribution, as provided of materials, not by
only as a damage lower bound, but also because its role in the interpretation of non-propagating
the S-N field, has great significance in the characterization of the fatigue behavior of materials, not only
cracks, correspondence with the threshold ΔKth in the crack growth rate curve, and determination
as a damage lower bound, but also because its role in the interpretation of non-propagating cracks,
of the intrinsic crack, but also in the definition of the Kitagawa-Takahashi diagram.
correspondence with the threshold ∆Kth in the crack growth rate curve, and determination of the
intrinsic crack, but also in the definition of the Kitagawa-Takahashi diagram.
Materials 2019, 12, 3239 14 of 17
Materials 2019, 12, x FOR PEER REVIEW 14 of 17

Figure
Figure 11. 11. Evolution
Evolution of theof estimation
the estimationofofthe
the fatigue
fatigue limit and
limit the the
and 0.01–0.99 confidence
0.01–0.99 intervals.intervals.
confidence

5. Conclusions
5. Conclusions
The main conclusions drawn from this work are the following:
The main conclusions drawn from this work are the following:
a) Concerning the Bayesian model:
(a) Concerning the Bayesian model:
1. Despite the limited utility of standard Bayesian methods, which utility is generally restricted to
specific and simple cases, the Bayesian methods based on Markov Chain Monte Carlo (MCMC)
1. Despite the limited utility of standard Bayesian methods, which utility is generally restricted
techniques, implemented in the OpenBUGS, can be advantageously used in the analysis of more
tocomplex
specificand and simplemodels,
advanced cases,particularly
the Bayesian methods
probabilistic based
ones, on Markov
thus, opening Chain Monte
the applications to Carlo
(MCMC) techniques,
a broad spectrum implemented
of fracture and fatiguein the OpenBUGS,
problems to be exploredcan in be
the advantageously
future, as the one used in
the analysis here.
investigated of more complex and advanced models, particularly probabilistic ones, thus,
2. opening
The OpenBUGS software allows
the applications for the
to a broad systematicofsampling
spectrum fractureof thefatigue
and model parameters
problems to to be
be explored
integrated into posterior predictive models, once the prior information has been enriched with
in the future, as the one investigated here.
the experimental data.
2. 3. TheOpenBUGS
The new Bayesiansoftware
approach allows for large
allows very the systematic sampling
posterior samples of theof the model
model parameters
parameters to be to be
integrated
obtained and intouseposterior predictive
them to derive models, once
the approximate the prior
distribution information
of any has been
variable instead enriched with
of working
with
the closed complex
experimental formulas.
data.
4. Bayesian methods, instead of providing point estimates of the variables, provide very large
3. The new Bayesian approach allows very large posterior samples of the model parameters to
samples of them that can be interpreted as their density functions, which is much more than
beconfidence
obtainedintervals.
and use them to derive the approximate distribution of any variable instead of
5. working with closed
The application complex
of Bayesian formulas.
techniques to the probabilistic regression Weibull model proposed
4. by Castillo and Canteli for
Bayesian methods, instead of providingthe analysis of S-N data
point enables the
estimates initial
of the simple probabilistic
variables, provide very large
definition of the five parameters model to be enhanced by providing the probability distribution
samples of them that can be interpreted as their density functions, which is much more than
for any percentile failure curve of the original model (to be interpreted as confidence intervals).
confidence
Furthermore, intervals.
the approach enables the subsequent evolution of the confidence intervals to be
5. The application
defined as a of ofBayesian
function the numbertechniques toout.
of tests carried the probabilistic regression Weibull model
b)proposed bytheCastillo
Concerning and
S-N model: Canteli for the analysis of S-N data enables the initial simple
probabilistic definition
1. The Bayes methodology can be of the five parameters
applied model
at any time during theto be enhanced
testing by providing
process providing an the
probability distributiontofor
invaluable contribution any percentile
enhance failure
the confidence curve
intervals ofof
thethe original
S-N model
assessment, (to be interpreted
in particular
aswhen
confidence intervals).
using such Furthermore,
a probabilistic theofapproach
model as that enables the subsequent evolution of the
Castillo-Canteli.
2. confidence
The evolution of the confidence
intervals bandsas
to be defined fora the fatigue of
function limit
thereferred
number to of
thetests
assessment
carried of out.
the
fatigue S-N field of the Maennig example confirms the robustness of the S-N model proposed by
(b) Concerning the S-N model:
Castillo-Canteli. In fact, the confidence bands obtained by applying the Bayes methodology
proves to show a near asymptotic trend with quick diminishing for increasing number of
1. The Bayes methodology can be applied at any time during the testing process providing
an invaluable contribution to enhance the confidence intervals of the S-N assessment,
in particular when using such a probabilistic model as that of Castillo-Canteli.
2. The evolution of the confidence bands for the fatigue limit referred to the assessment of the
fatigue S-N field of the Maennig example confirms the robustness of the S-N model proposed
by Castillo-Canteli. In fact, the confidence bands obtained by applying the Bayes methodology
proves to show a near asymptotic trend with quick diminishing for increasing number of
evaluated test results evidencing a moderate variation in the ranges of the fatigue limit even for
scarce number of test data.
Materials 2019, 12, 3239 15 of 17

3. The confidence intervals in the particular case of the fatigue limit are transcendental because
the direct interpretation of the latter, whereas those for the remaining S-N parameter are of less
interest when separately interpreted. Instead, the reciprocal interdependence, i.e., the correlation,
existing among all the model parameters is better reflected through the representation of the
confidence bands (distribution of the particular probability of failure) of the percentiles S-N
curves, preferably only for low or very low probability of failure in practical design.
Materials 2019, 12, x FOR PEER REVIEW 15 of 17

Supplementary
evaluated test resultsMaterials: The following
evidencing a moderate are
variation in the available
ranges online
of the fatigue at http://www.mdpi.com/1996-1944/12/19/3239/s1,
limit even for
Video
scarceS1: Evolution
number of test data. of the whole p-S-N field, Video S2: Evolution of the quantiles and confidence intervals with
3.
theThe confidenceof
number intervals
tests.in the particular case of the fatigue limit are transcendental because the
direct interpretation of the latter, whereas those for the remaining S-N parameter are of less
interest when
Author Conceptualization:
separately interpreted.
Contributions: E.C., A.F.-C.
Instead, the reciprocal interdependence, and
i.e., the M.M.-C.; methodology: E.C.; software: E.C. and
correlation,
existing among all the model parameters is better reflected through the representation of the
M.M.-C.; validation: M.M.-C. and S.B.; writing—original draft
confidence bands (distribution of the particular probability of failure) of the percentiles S-N
preparation: All Authors; writing—review and
editing: All Authors.; supervision: E.C.; please turn to the
curves, preferably only for low or very low probability of failure in practical design. CRediT taxonomy for the term explanation.
Funding: This research received no external funding.
Supplementary Materials: The following are available online at www.mdpi.com/xxx/s1, Video S1: Evolution of
the whole p-S-N field, Video S2: Evolution of the quantiles and confidence intervals with the number of tests.
Conflicts of Interest: The authors declare no conflict of interest.
Author Contributions: Conceptualization: E.C., A.F.-C. and M.M.-C.; methodology: E.C.; software: E.C. and
M.M.-C.; validation: M.M.-C. and S.B.; writing—original draft preparation: All Authors; writing—review and
editing: All Authors.; supervision: E.C.; please turn to the CRediT taxonomy for the term explanation.
Appendix A OpenBugs Codes
Funding: This research received no external funding.

Acknowledgments:
This Appendix shows the code implemented in OpenBugs to introduce the Castillo-Canteli model
Conflicts of Interest: The authors declare no conflict of interest.
by using the generalized extreme function. To do that, and with the aim of facilitating the reading of
Appendix A. OpenBugs Codes
the code, four auxiliary variables have been defined (lines 3 to 6), which are used to define in a simpler
This Appendix shows the code implemented in OpenBugs to introduce the Castillo-Canteli
waybythe
model model
using (lineextreme
the generalized 8). function. To do that, and with the aim of facilitating the
the model
in a simpler way the model (line 8).
is defined, the implied variables (N0 , ∆σ0 , β, λ, δ) are defined as uniform random
reading of the code, four auxiliary variables have been defined (lines 3 to 6), which are used to define
Once
ones, ranging between
Once the model is defined, the a minimum
implied variables (N , Δσand
, β, λ, a
0 0 maximum
δ) are value
defined as uniform (lines 10 to 14).
random
ones, ranging between a minimum and a maximum value (lines 10 to 14).
Mode 1
Mode 1

Furtheremore, the ones/zeros Poisson model trick has been used to introduce the Castillo and
Canteli model, as can be seen in the following code:
Materials x FOR 12,
2019,
Materials 2019, 12, PEER3239
REVIEW 16 of 17 16 of 17

Furtheremore, the ones/zeros Poisson model trick has been used to introduce the Castillo and
Canteli model, as can be seen in the following code:

Mode 2
Mode 2

References
References
1. ISO. 12107:2012. Metallic Materials-Fatigue Testing-Statistical Planning and Analysis of Data, 2nd ed.; ISO:
London, UK, 2018.
1.
2. ISO. 12107:2012. Metallic Materials-Fatigue Testing-Statistical Planning and Analysis of Data, 2nd ed.; ISO:
ASTM. E739–10 Standard Practice for Statistical Analysis of Linear or Linearized Stress-Life (S-N) and Strain-Life
(ε-N) Fatigue Data; ASTM: West Conshohocken, PA, USA, 2015.
3. London, UK, 2018.
ASM International. Fatigue and Fracture, Statistical Considerations in Fatigue. In ASM Handbook; ASM
International: Cleveland, OH, USA, 1997; Volume 19, ISBN 978-0-87170-385-9.
2.
4.
ASTM. E739–10 Standard Practice for Statistical Analysis of Linear or Linearized Stress-Life (S-N) and Strain-Life
Castillo, E.; Fernández Canteli, A. A Unified Statistical Methodology for Modeling Fatigue Damage; Springer:
(ε-N) Fatigue Data; ASTM: West Conshohocken, PA, USA, 2015.
Berlin/Heidelberg, Germany, 2009.
5. Freudenthal, A.M. The statistical aspect of fatigue of materials. Proc. R. Soc. 1946, 187, 416–429.
3.
6. ASM International. Fatigue and Fracture, Statistical Considerations in Fatigue. In ASM Handbook; ASM
Freudenthal, A.M.; Gumbel, E.J. Physical and statistical aspects of fatigue. Adv. Appl. Mech. 1956, 4, 117–
158.
7.
International: Cleveland, OH, USA, 1997; Volume 19, ISBN 978-0-87170-385-9.
Bolotin, V.V. Wahrscheinlichkeitsmethoden zur Berechnung von Konstruktionen; Verlag für Bauwesen: Berlin,
4. Castillo, E.; Fernández Canteli, A. A Unified Statistical Methodology for Modeling Fatigue Damage; Springer:
Germany, 1981.
8. Bolotin, V.V. Mechanics of Fatigue; CRC Press: Boca Raton, FL, USA, 1999.
9. Berlin/Heidelberg, Germany, 2009.
Koller, R.; Ruiz-Ripoll, M.L.; García, A.; Fernández Canteli, A.; Castillo, E. Experimental validation of a
statistical model for the Wöhler field corresponding to any stress level and amplitude. Int. J. Fatigue 2009,
5. Freudenthal, A.M. The statistical aspect of fatigue of materials. Proc. R. Soc. 1946, 187, 416–429.
31, 231–241.
6.
10. Freudenthal, A.M.; Gumbel, E.J. Physical and statistical aspects of fatigue. Adv. Appl. Mech. 1956, 4, 117–158.
Pyttel, B.; Fernández-Canteli, A.; Argente-Ripoll, A. Comparison of different statistical models for
description of fatigue including very high cycle fatigue. Int. J. Fatigue 2016, 93, 435–442.
7.
11. Bolotin, V.V. Wahrscheinlichkeitsmethoden zur Berechnung von Konstruktionen; Verlag für Bauwesen:
Correia, J.A.F.O.; Apetre, N.; Arcari, A.; de Jesus, A.; Muniz-Calvente, M.; Calçada, R.; Berto, F.; Fernández
Canteli, A. Generalized probabilistic model allowing for various fatigue damage variables. Int. J. Fatigue
Berlin, Germany, 1981.
2017, 100, 187–194.
8.
12. Bolotin, V.V. Mechanics of Fatigue; CRC Press: Boca Raton, FL, USA, 1999.
Muniz-Calvente, M.; de Jesus, A.; Correia, J.A.F.O.; Fernández Canteli, A. A methodology for probabilistic
prediction of fatigue crack initiation taking into account the scale effect. Eng. Fract. Mech. 2017, 185, 101–
9. Koller, R.; Ruiz-Ripoll, M.L.; García, A.; Fernández Canteli, A.; Castillo, E. Experimental validation of
113.
13. Benjamin, J.; Cornell, C.A. Probability, Statistics and Decisions for Civil Engineers; Mc Graw-Hill Co.: New
a statistical model for the Wöhler field corresponding to any stress level and amplitude. Int. J. Fatigue 2009,
York, NY, USA, 1970.
31, 231–241. [CrossRef]
10. Pyttel, B.; Fernández-Canteli, A.; Argente-Ripoll, A. Comparison of different statistical models for description
of fatigue including very high cycle fatigue. Int. J. Fatigue 2016, 93, 435–442. [CrossRef]
11. Correia, J.A.F.O.; Apetre, N.; Arcari, A.; de Jesus, A.; Muniz-Calvente, M.; Calçada, R.; Berto, F.; Fernández
Canteli, A. Generalized probabilistic model allowing for various fatigue damage variables. Int. J. Fatigue
2017, 100, 187–194. [CrossRef]
12. Muniz-Calvente, M.; de Jesus, A.; Correia, J.A.F.O.; Fernández Canteli, A. A methodology for probabilistic
prediction of fatigue crack initiation taking into account the scale effect. Eng. Fract. Mech. 2017, 185, 101–113.
[CrossRef]
13. Benjamin, J.; Cornell, C.A. Probability, Statistics and Decisions for Civil Engineers; Mc Graw-Hill Co.: New York,
NY, USA, 1970.
14. Castillo, E. Extreme Value Theory in Engineering; Elsevier: Amsterdam, The Netherlands, 1980.
15. Efron, B.; Tibshirami, R. An Introduction to the Bootstrap; Chapman & Hall: London, UK, 1993.
16. Davidson, A.; Hinkley, D. Bootstrap Methods and Their Application; Cambridge University Press: Cambridge,
UK, 1997.
17. Chernick, M. A Practitioner’s Guide; John Wiley & Sons: Hoboken, NJ, USA, 1999.
18. Chao, M.; Fuh, C. Bootstrap methods for the up-and-down test on pyrotechnics sensitivity analysis. Stat. Sin.
2001, 1, 1–21.
19. Guida, M.; Penta, F. A Bayesian analysis of fatigue data. Struct. Saf. 2010, 32, 64–76. [CrossRef]
20. Zhu, S.P. Bayesian framework for probabilistic low cycle life prediction and uncertainty modeling of aircraft
turbine disk alloys. Probabilistic Eng. Mech. 2013, 34, 114–122. [CrossRef]
Materials 2019, 12, 3239 17 of 17

21. Maennig, W.W. Untersuchungen zur Planung und Auswertung von Dauerschwingversuchen an Stahl in den
Bereichen der Zeit-und der Dauerfestigkeit. Fortschr-Ber; VDI-ZS Reihe; VDI Verlag: Düsseldorf, Germany, 1967;
Volume 5.
22. Maennig, W.W. Bemerkungen zur Beurteilung des Dauerschwingverhaltens von Stahl und einige
Untersuchungen zur Bestimmung des Dauerfestigkeitsbereiches. Materialprüfung 1970, 12, 124–131.
23. Maennig, W.W. Vergleichende Untersuchung über die Eignung der Treppenstufenmethode zur Berechnung
der Dauerschwingfestigkeit. Materialprüfung 1971, 13, 6–11.
24. Castillo, E.; Iglesias, A.; Ruíz-Cobo, R. Functional Equations in Applied Sciences; Elsevier Science: Amsterdam,
The Netherlands, 1999.
25. Aczel, J. Lectures on Functional Equations and Applications and Their Applications; Mathematics in Science and
Engineering; Academic Press: Cambridge, MA, USA, 1966; Volume 19.
26. Aczel, J. Chapter 1: On history and applications and theory of functional equations (Introduction). In Functional
Equations: History, Applications and Theory; Aczel, J., Ed.; Reidel Publishing Company: Dordrecht, The Netherlands,
1984; pp. 3–12.
27. Fernández-Canteli, A.; Przybilla, C.; Nogal, M.; López-Aenlle, M.; Castillo, E. ProFatigue: A software
program for probabilistic assessment of experimental fatigue data sets. Procedia Eng. 2014, 74, 236–241.
[CrossRef]

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).