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

net/publication/257940733

Fatigue reliability of structural components

Article  in  International Journal of Materials and Product Technology · January 2006

DOI: 10.1504/IJMPT.2006.008274

CITATIONS READS

11 668

3 authors, including:

G. Glinka
University of Waterloo
178 PUBLICATIONS   5,218 CITATIONS   

SEE PROFILE

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

Spanish Network on Notch Mechanics (Red Española de Mecánica de Entallas) View project

International Symposium on Fatigue Crack Growth – Experimental, Theoretical and Numerical Approach (ISFCG2017) View project

All content following this page was uploaded by G. Glinka on 24 April 2019.

The user has requested enhancement of the downloaded file.

64 Int. J. Materials and Product Technology, Vol. 25, Nos. 1/2/3, 2006

Fatigue reliability of structural components

Hieronim Jakubczak* and

Wojciech Sobczykiewicz
Department of Construction Machinery Engineering, Warsaw
University of Technology, Narbutta 84, Warsaw 02 -524, Poland
E-mail: Hieronim.Jakubczak@simr.pw.edu.pl
E-mail: wso@simr.pw.edu.pl
*Corresponding author

Grzegorz Glinka
Department of Mechanical Engineering, University of Waterloo,
Waterloo, Ontario N2L 3G1, Canada
E-mail: gggreg@mecheng1.uwaterloo.ca

Abstract: The paper is concerned with reliability assessment of structural and

mechanical components subjected to cyclic variable amplitude service load.
Due to relatively easy simulation and simple computational procedure,
computer simulations based on the Monte Carlo method are often used for
reliability assessments of engineering objects. The only disadvantage of such a
method is a large number of repetitions of the fatigue-life assessment procedure
required for attaining good accuracy in the range of high levels of reliability.
Useful tips enabling the decrease of the number of samplings in the simulation
procedure and preserving required reliability levels are the subjects of the
discussion presented below. The rules for dimensioning structures and
mechanical components using the deterministic and probabilistic input data
have been formulated, and the relation between the safety factor and the
fatigue-failure probability of structural components has been analysed as well.

Keywords: structural components; fatigue; safety factor; reliability;

Monte Carlo simulation; fast simulation algorithm.

Reference to this paper should be made as follows: Jakubczak, H.,

Sobczykiewicz, W. and Glinka, G. (2006) ‘Fatigue reliability of structural
components’, Int. J. Materials and Product Technology, Vol. 25, Nos. 1/2/3,
pp.64–83.

Biographical notes: Professor Hieronim Jakubczak has a PhD and DSc from
the Warsaw University of Technology, where he is now with the Institute of
Construction Machinery Engineering. Dr. Jakubczak is a specialist in Fatigue
Design and Reliability of Structures. His research interest includes Fatigue,
Fracture and Reliability of Welded Steel Structures, Low Cycle Fatigue and
Fracture Mechanics. He has gained his experience through design, analysis and
testing of load carrying structures for earth moving machines. Dr. Jakubczak
and Dr. Glinka have also jointly developed a software package for the
Deterministic and Probabilistic Fatigue Life Assessment of Engineering
Components and Structures.

Copyright © 2006 Inderscience Enterprises Ltd.

 Fatigue reliability of structural components 65

Professor Wojciech Sobczykiewicz, has a PhD and DSc from the Warsaw
University of Technology, where he is now Managing Director of the Institute
of Construction Machinery Engineering. He is also a consultant to the
Industrial Institute of Construction Equipment. His professional field of interest
is Fatigue Design and Reliability of Load Carrying Mechanical Structures.
Most of his experience was gained through studying cranes and earth moving
machines. In recent years he has also explored the problems associated with the
operational safety of cranes.

Professor Grzegorz Glinka has been with the University of Waterloo since
1989. He has also lectured at the University of Metz, France and at the
University College London, England. He holds a PhD and DSc from the
Warsaw University of Technology. Dr. Glinka is a specialist in fracture and
fatigue of mechanical components and welded steel structures. He has also
acted as a UN expert. His research interests include Fracture of Materials,
Fatigue of Structures, Low Cycle Fatigue, Fatigue of Welded Structures,
Multiaxial Fatigue and Creep of Engineering Materials, Computer Aided
Design, and FEM-Elastic-Plastic Stress-Strain Analysis.

1 Introduction

The inherent feature of fatigue-behaviour of structural components in service is the

scatter of fatigue-lives resulting from the variability of parameters such as the loading
magnitude, material properties, geometrical features of a component etc. The scatter of
fatigue-lives of structural components is often the reason for differences between
deterministically predicted fatigue-lives and those observed in service (Buxbaum, 1982;
Hudak, 1990). The differences are sometimes very large irrespective of the life prediction
method and are often attributed to imperfections of models used for fatigue-life
calculations. However, the disagreement may also be caused by the uncertainty of the
input data used for fatigue-life predictions. The apparent feature of the uncertainty of the
input data is its scatter, characterised often by the coefficient of variation (COV),
expressed as the ratio of the standard deviation, σx, to the mean value, µx, of the parameter
(random variable) X, i.e., Vx = σx/µx. More information on the uncertainty of the parameter
being analysed is given by the probability distribution. Therefore, in order to account for
the uncertainty of the input data in the fatigue-life assessment procedure, a probabilistic
approach is needed.
An algorithm for fatigue dimensioning of a structural component is shown in
Figure 1. Such an algorithm requires specific input data such as the geometry of the
‘potential weak link’ (PWL), i.e., the critical component or critical cross section, the
service loading and the material fatigue. To avoid fatigue-failure, a design/acceptability
criterion is formulated either in terms of stress (R > S) or in terms of life (T > Tr).
Parameters S and R denote the loading and fatigue resistance respectively, which may
have different meanings, depending on the method used in the fatigue-life assessment
procedure. The second condition, used when designing for a finite fatigue-life Tr and T,
denotes the required and calculated fatigue-life respectively. In the traditional design
procedures all the parameters, R, S and Tr above are considered as deterministic
quantities.
66 H. Jakubczak, W. Sobczykiewicz and G. Glinka

Figure 1 An algorithm for fatigue dimensioning

The ratio of the fatigue-resistance level R, to load S, or of the calculated fatigue-life

T, to the required one Tr, is often called as a safety factor, SF. The larger the magnitude
of the safety factor the higher is the fatigue-safety of a machine or structural component
in service. However, the same value of the safety factor of several components does not
necessarily mean the same fatigue-safety of each of them. This is due to different scatters
of fatigue properties associated with individual components and, as a consequence, to
different scatters of their fatigue-lives.

2 Fatigue life assessment

2.1 Deterministic methods for fatigue life evaluations

There are three fatigue-life assessment methods frequently used in engineering practice,
i.e., the nominal stress method called often as the S–N approach, the strain-life method
called as the local strain, ε–N, approach and the fracture mechanics based approach
denoted often as the da/dN – ∆K method.
The nominal stress or simple engineering reference stress, S, is used in the
S–N method as the load parameter (Figure 2) and therefore the fatigue properties must be
determined in terms of identically defined nominal stress in specimens used for the
determination of the basic S–N curves. The main task and the main difficulty of an
analyst is to assure that the same reference stress is used for the determination of the
basic fatigue S–N curves and for the quantification of fatigue damages in a machine part
represented in Figure 2 by the notched component.
The local strain-life (ε–N) method is based on the analysis of the actual elastic-plastic
strains and stresses at the critical point (such as the notch tip) and the local strain
ε, represents the load parameter (Figure 3).
 Fatigue reliability of structural components 67

Figure 2 The (S–N) nominal stress approach

Figure 3 The (ε–N) strain-life approach

The actual local strains and stresses are usually determined using simplified methods of
elastic-plastic, stress-strain analysis such as the Neuber rule (Neuber, 1961) or the
Equivalent Strain Energy Density (ESED) method (Molski and Glinka, 1981).
The fracture-mechanics based method (Figure 4) requires the analysis of the
fatigue-crack growth from its initial dimension ao, to the critical size af. The load
parameter is represented by the Stress Intensity Factor K, and the material fatigue
properties are characterised by the relationship correlating the fatigue-crack growth rate
da/dN, and the stress intensity range, ∆K.
68 H. Jakubczak, W. Sobczykiewicz and G. Glinka

Figure 4 The (da/dN – ∆K) fracture mechanics approach

Most of the input data parameters used in the approaches presented above are scattered
(uncertain). In order to predict in deterministic analyses, sufficiently safe (conservative)
fatigue lives, the uncertainty of the input fatigue data is accounted for by selecting
parameters of the basic S–N, ε–N or da/dN – ∆K curve that correspond to the required
survival probability (e.g., 97%).
However, only the probabilistic analysis makes it possible to account for the
uncertainty of the input data and to rationally assess its effect on the resultant fatigue-life
of analysed structures or machine components.

3 Uncertainty of data

The uncertainty of the data in a fatigue analysis concerns three main groups of data,
i.e., service loadings, fatigue properties of the analysed material or component, and the
local geometry of a component. The input data depends on the methodology and might be
expressed by one set curves (or mathematical expressions), representing jointly, both
material properties and the geometry. Geometrical features are sometimes included into
fatigue properties of a component (e.g., S–N curves for welded joints) and the scatter of
the geometrical parameters of a component can also influence the resultant
fatigue-resistance of the component.

3.1 Service loading

The knowledge of service loading is crucial for the fatigue-life predictions of structural
components. In most practical cases, the load is usually a variable amplitude process
(history), dependent on the operating conditions, character of the service task and
operator’s (user’s) skills and habits. The service loading is a random variable and not one
unique quantity, because it is difficult to account for all possible operating conditions of a
machine or structure. The available data concerning loads is always limited and the
variability that may occur in service needs to be accounted for by using the probabilistic
life prediction methods.
 Fatigue reliability of structural components 69

Service loadings can be formally defined in certain cases and recommended by

special design guidelines or standards. They are usually given in the form of
loading/stress spectra or stress histories (Aicher, 1976). The standard loading must be
scaled for the actual object by using a scaling parameter such as the maximum stress
range, ∆Smax. Such an approach requires the stress analysis to be carried out for
prescribed representative loads.
A simple way of accounting for the potential uncertainty of service loading is to
assume the scaling parameter as a random variable. In the case of stress history acquired
in service, the mean value of the scaling parameter can be assumed as equal to one.
The load scaling parameter, usually described by normal distribution and its scatter,
measured by the COV, is in the range of VS = 0.05 – 0.10 (Szala and Zawiślak, 1990).

3.2 Fatigue data

As far as the fatigue properties are concerned, three types of randomness can
be distinguished (Bolotin, 1996), i.e., within-specimen, specimen-to-specimen, and
batch-to-batch randomness. The within-specimen randomness can be observed in the
crack growth rate used in the fracture mechanics approach, whereas most of other fatigue
properties show the randomness of the last two types, observed in laboratory tests under
the same loading and environment conditions. This is usually the result of imperfections
and differences occurring within acceptable tolerances of the manufacturing process.
Statistical parameters characterising fatigue properties, which can be considered as
random variables, are obtained in a natural way from the statistical analysis of the test
results (see Figures 5(a)–(c)).
Parameters such as the fatigue-limit, Se, the strain-life fatigue coefficients, σ’f and ε’f,
and the crack growth constant, C, are usually log-normally distributed random variables
and logarithmic coordinates are used for their graphical presentation. The scatter of the
fatigue-limit, Se, depends on the specimen (component) geometry, and its COV
varies usually from VR = 0.05 – 0.1 for plain specimens to VR = 0.1 – 0.2 for welded
joints (Committee on Fatigue and Fracture Reliability of the Committee on Structural
Safety and Reliability of the Structural Division, 1982; Wirshing, 1995).
The scatter of parameters, σ’f and ε’f, used in the strain-life approach is usually
smaller than that appearing in the S–N data. This is due to a more strictly controlled
manufacturing process of small laboratory specimens used for the determination of
fatigue-resistance curves. The coefficients of variation of the strain-life parameters can be
approximately assumed as Vσ = 0.05 and Vε = 0.1. The third parameter, which can be
considered in the notch-strain approach as a random variable is the cyclic strength
coefficient, K’. It is assumed that K’ is log-normally distributed with an average value of
the coefficient of variation of VK’ = 0.1. The question then arises, whether parameters σ’f ,
ε’f and K’, due to the known relationship between them, should be considered as
uncorrelated random variables. An analysis carried out on data sets obtained for four
carbon and low-alloy steels has revealed that the assumption of no correlation results in
conservative life predictions.
Scatter of the fatigue-crack growth parameter C is small if the entire da/dN – ∆K
curve is obtained from one specimen (VC = 0.1); however if the da/dN vs. ∆K curve is
based on data obtained from several specimens, the scatter of parameter C may increase
to VC = 0.2. In such a case, parameter C can also be treated as dependent on the exponent
70 H. Jakubczak, W. Sobczykiewicz and G. Glinka

n (Bertini and Marmorini, 1989) which in turn is characterised by normal distribution

with the COV value of Vn = 0.05 – 0.1.

Figure 5 Scatter of fatigue properties used in the (a) nominal stress, (b) strain-life and (c) fracture
mechanics approach respectively
 Fatigue reliability of structural components 71

3.3 Component geometry

Geometrical features of structural components are the result of both the design and
manufacturing processes. The overall geometrical shape (Figure 6) and dimensions
(diameters, cross-section, plate thickness, etc.) as well as parameters describing the local
geometry (notch tip radii, angles, etc.) are prescribed (defined) at various stages of the
design process. These parameters are obtained with certain accuracy and repeatability
levels during the manufacturing process, regardless of whether they were considered as
deterministic single-valued quantities or they were assigned certain tolerance ranges.
In both cases they are, in reality, random variables. Hence, the scatter of the global and
local geometrical parameters of structural components reflects their quality of
manufacturing.

Figure 6 (a) The global t and e and (b) local ρ, θ and a, geometrical parameters of a structural
component

Changes in the global geometry resulting in changes of the cross-section area or

introducing additional local bending (see misalignment in Figure 6(a) of a component
may lead to changes in the magnitude of the service loading and subsequently in the
magnitude of nominal and local stresses. This may increase the scatter of the service
loading parameters used in the analysis. The local geometry of a component (Figure 6(b))
affects significantly, the fatigue process, especially in the early fatigue-crack-initiation
stage or the fatigue-crack growth. As a result, the scatter of geometrical parameters will
significantly affect the scatter of the fatigue-life. Depending on the approach used for the
fatigue-life analysis and the fatigue-life calculation procedure, the local geometry can be
described using the notch factor Kf, (nominal stress approach), or the stress concentration
factor Kt, in the case of the strain-life approach. In the case of the fracture mechanics
based method the initial crack size ao, plays as important role as the stress concentration
factor Kt.
The measured values of the weld toe radius ρ, and the weld angle Θ, shown in
Figure 6, can be used as a base for the calculation of the stress concentration Kt, and
the fatigue-notch-factor Kf. The experimental data (Engesvik and Moan, 1983;
Jakubczak, 1998) indicates that both factors can be sufficiently accurate described by the
Weibull probability distribution with the COV in the range of VK = 0.05 – 0.2.
As far as the initial crack size is concerned, several different statistical probability
distributions are used (Jakubczak, 2002). The log-normal or exponential probability
distributions seem to fit best to a variety of experimental data. The mean value of the
initial crack depth ao, depends on the quality of the material and the component and
varies from 0.005 mm for turbine blades (Brückner-Foit, 1993) to 5 mm for welded
bridge structures (Zhao and Halder, 1996). The scatter of the initial crack size is
relatively large, i.e., Va = 0.2 – 1.0 (Jakubczak, 2002).
72 H. Jakubczak, W. Sobczykiewicz and G. Glinka

4 Probabilistic fatigue life assessment

There are many methods of reliability estimation suitable for probabilistic assessment of
fatigue-lives of structural components (Zhou and Shen, 1998). Among them the Monte
Carlo based computer simulation is often used as a tool. The Monte Carlo simulation
procedure enables the statistics of the fatigue-life to be determined from the distributions
of the input variables. It requires the repetition of the deterministic fatigue life assessment
procedure for a number of input data sets sampled according to their probability density
distributions (Figure 7). The probability of failure is calculated as the ratio of the number
of failures Lf, (T < Tr) to the number of simulation runs, L.
The main advantage of the simulation approach is the possibility of using even very
complex models for fatigue-life predictions. It is also important to note that the
simulation procedure is not limited to any specific type of probability density distribution
or the maximum number of probability distributions or the number of variables used in
the simulation or the magnitude of scatter (Bertini and Marmorini, 1989; Cruse, 1997) of
these variables.

Figure 7 The algorithm for the simulation based fatigue life assessment

The simulation based approach enables, also, adjustment of some limits concerning the
magnitude of random variables by truncating a portion of the probability distribution.
The truncation of the probability distribution of fatigue-strength from the left hand side,
in the region of low values of Se, or the limitation of the probability distribution
of the initial crack size ao, from the right hand side (large ao) is equivalent to the
rejection of badly manufactured objects that would be normally rejected during the final
quality-control tests.
The actual values of input parameters to be considered as random variables can be
selected, with some approximation, by perturbation analysis (Zhou and Shen, 1998).
However, for a more accurate assessment of the importance of particular random
variables, the sensitivity analysis needs to be carried out (Zhou and Shen, 1998).
 Fatigue reliability of structural components 73

Sensitivity analysis involves the determination of the rate of variation of the resultant
variable (fatigue-life or probability of failure) with respect to a small variation of an input
random variable.
The sensitivity can be obtained by finding the derivative of the output variable with
respect to the input parameter (Zhou and Shen, 1998). For fatigue lives resulting from
simulation analyses, the following formula for sensitivity wTv has been proposed
(Jakubczak, 2002):
∆Tr µ x
wTv = . (1)
Tr ∆σ x

The sensitivity wTv shows the relative change of the fatigue-life Tr, at given probability of
failure Pf, caused by the change of the scatter, ∆Vx, (∆Vx = ∆σx/µx) of random variable X.
In the same way, the sensitivity wTm can be calculated, showing the relative change
of the fatigue-life Tr, due to the relative change of the mean value ∆µx/µx, of the random
variable X:
∆Tr µ x
wTm = . (2)
Tr ∆µ x

The change of the scatter ∆Vx, and the change of the mean value ∆µx/µx, of the random
variable X should be very small, e.g., 1%.
The formulae defined above make it possible to compare directly, the sensitivities
wTv and wTm corresponding to all random variables. As a result it is possible to assess the
importance of random variables in relation to both their scatter and the mean value.
Equations (1) and (2) can be rewritten for sensitivities wPv and wPm showing the relative
change of the probability of failure Pf due to the relative change of the scatter and the
mean value of the random variable respectively.

5 Fast fatigue life prediction based on the simulation technique

5.1 Relative fatigue life distribution

It has been found advantageous to carry out the simulation-based fatigue-life
analyses using the concept (index) of relative fatigue-life. This conclusion was drawn
from observations of the basic formula in the nominal stress (S–N) approach used for
fatigue-life calculations. The simple formula given below accounts for the service loading
in the form of the stress spectrum, (Figure 2, where si = ∆Si/∆Smax) and it ignores the
existence of the fatigue-limit, i.e., all stress spectrum levels including those below the
fatigue-limit are accounted for in the damage analysis.
m
 Se  N0
T =  (3)
 ∆S K
 max f  ∑ si )m ni
(
n

If parameters Se, ∆Smax and Kf are assumed to be random variables and xR, xS and xK are
the corresponding relative random variables, equation (3) can be written in the form of
equation (4).
74 H. Jakubczak, W. Sobczykiewicz and G. Glinka

m
 x 
T = T0  R  (4)
 xS x K 
A relative random variable xX, is defined as the ratio of the actual value of the random
variable X, to its mean value µx. The quantity To, is the reference life corresponding to the
appropriate mean values of all random variables:
m
 µ  N0
T = T (µ )  R  . (5)
 µS µK  ∑ ( si )m ni
n

The relative fatigue-life is defined as Tw = T/To. It is apparent that the relative life does
not depend on the form of the service-loading spectrum and therefore both the probability
density distribution of the relative fatigue-life Tw, and the relative fatigue-life Trw at given
probability of failure Pf, depend only on the distributions of the random variables.
Hence, in order to obtain a probability distribution of the relative fatigue-life, one can use
in probabilistic fatigue-life assessments, a more computationally efficient stress spectrum
or constant amplitude (C.A.) loading instead of the actual long reversal-by-reversal
stress history.
Due to the more complicated procedure of fatigue-life calculation based on the
strain-life approach, similar independence of the distribution of the relative fatigue-life,
Tw, in the form of the service loading can be only assumed by analogy. The results
presented in Figure 8, which were generated using the strain-life approach, indicate that
the same results, in terms of the probability of failure, are obtained for the actual long
stress history (H), the corresponding cumulative stress spectrum (S), and the constant
amplitude loading (C). The uncertainty of the input parameters in the example was
assumed as follows:
• notch geometry (K): Weibull distribution (W), VK = 0.1, (KW10)
• service loading (S): normal distribution (N), VS = 0.1, (SN10)
• material (M1): log-normal distribution (L), Ve = 0.05, Vs = 0.05, VK’ = 0.1
(ML1–Figure 5(b)).

Figure 8 Comparison of fatigue-lives calculated for a stress history (lines), stress spectrum
(open symbols) and C.A. loading (solid symbols)
 Fatigue reliability of structural components 75

Thus, by using the stress spectrum, the computing time needed for the simulation can be
reduced by a factor equal to the ratio of the number of cycles in the stress history to the
number of steps (stress levels) in the stress spectrum.
The computing time reduction factor can be increased a few times more if the
simulation is carried out using a constant amplitude loading history. However, for the
notch-strain approach care must be taken while selecting the mean value, µS, of the stress
range ∆S, since the loading level may affect predicted relative fatigue-lives Trw
(Figure 9). It is suggested to use the mean value of the stress range ∆S for the C.A.
loading equal to or less (10%) than the mean value of the maximum stress range ∆Smax
occurring in the entire stress history.

Figure 9 Effect of the loading level on the relative fatigue-life Trw at probability of failure
Pf = 0.001 (four different carbon and low alloy steels)

Due to the complex calculation algorithm of the crack growth analysis, the accelerated
procedure described above has negligible effect on the simulation time concerning
fatigue-life assessments based on the fracture-mechanics approach. Shortening of
the simulation time is possible in this case if the relative life Tw, is calculated through the
direct integration of the crack-growth law for C.A. loadings. This requires neglecting
the dependency of the stress-intensity-factor correction function on the crack
length, which, in the deterministic approach would be in most cases inappropriate and
non-conservative.

5.2 Fitting a theoretical distribution into a data set

The probability of failure in the direct Monte Carlo simulation method is calculated as the
ratio of the number of failures to the number of simulations (repetitions). Such an
approach is not very efficient since a large number of repetitions are necessary if a high
level of reliability is required. Several methods (variance reduction techniques) are
available (Zhou and Shen, 1998), which enable reduction of the number of repetitions
without increasing of the error of estimated quantities. These procedures are useful when
the computing process is very long and time-consuming (Neves, 2001). However, thanks
76 H. Jakubczak, W. Sobczykiewicz and G. Glinka

to the steady increase of computers computational power even a very large number of
simulations (repetitions) of the same calculation procedure do not, at present, pose any
serious problem.
The failure probability can be also calculated from the distribution function fitted into
the simulation results. However, if the fitting is carried out for the entire set of simulated
results, the failure probabilities calculated in the low probability range may significantly
differ from those obtained later from the fitted distribution. This is usually due to the
inaccurate fitting of the selected probability distribution function into the region spanning
the left tail-part of the simulated set of results (Figure 10). Therefore, it is advantageous if
the distribution function is fitted into the truncated distribution.

Figure 10 Fitting the standard probability distribution function into the simulation results

If the left-hand-side-tail region of the probability distribution is properly selected, the

fatigue-lives, calculated from the fitted distribution, should agree fairly well with the
simulated values.
This phenomenon can be seen in Figure 11, where the relative fatigue-lives, Trw,
calculated from the distributions fitted into a set of 2,000 simulation results, (F), and
obtained using the nominal stress approach are compared to those determined from the
theoretical (exact) pdfs of the relative fatigue-life Tw. Since the relative fatigue-life for
one random variable can, according to equation (4), be written as Tw = xXk, the
corresponding theoretical pdf can be written as f(Tw) = f(xXk) and determined according to
the formula: fT(t)dt = fX(x)dx (Benjamin and Cornell, 1970). The parameter k in the
 Fatigue reliability of structural components 77

formula is equal to the slope m of the S–N curve if the random variable xX represents the
fatigue-limit, Se, (xX = xR) and/or if it represents the fatigue-notch-factor, Kf, (xX = xK).
The exponent k = –m if the random variable xX represents the maximum stress range,
∆Smax, (xX = xS).
The range of spread of the left-hand-side region of the probability distribution of the
simulated set of results can be selected by minimising the statistics based on the χ2 test.
The examples discussed above indicate that adequate fit of the probability distribution
function into the simulated set of results may potentially decrease the number of
necessary repetitions of the Monte Carlo samplings.

Figure 11 Comparison of theoretical lives with those obtained from the distribution function fitted
into the set of simulated results

5.3 Neglecting the fatigue limit

The fatigue-limit is an inherent feature of fatigue properties of the material and
components tested under constant amplitude loading. It is defined as the endurance limit
Se on the S–N curve analogously to the threshold stress range ∆σth on the stress-life curve
in the strain-life approach, and as threshold stress intensity range ∆Kth on the da/dN curve
in the fracture mechanics approach. The calculated fatigue-life is finite if the parameter
corresponding to the maximum stress range in the loading block is greater than the
fatigue limit (e.g., ∆Smax > Se/Kf in the nominal stress approach), otherwise the fatigue-life
would be infinite. Therefore the results obtained in the simulation may consist of two
sets: one consisting of results concerning finite fatigue-lives and the other containing
results with infinite fatigue-lives. The number of simulations resulting in the infinite life
prediction depends on the loading level and it also affects the scatter of fatigue lives
(Figure 12) caused by reduced number of predicted finite life results.
78 H. Jakubczak, W. Sobczykiewicz and G. Glinka

Figure 12 Effect of the fatigue-limit on the life scatter (dashed lines correspond to fatigue-limit
equal to zero)

It is worth noting that the loading level does not affect the left-hand-side tail of the
probability distribution. Therefore, the loading level does not affect the fatigue-lives
calculated for specified value of failure probability. Such a finding suggests that the
fatigue-limit can be neglected in the fatigue-damage calculation procedures used for the
Monte Carlo simulation, provided that the finite fatigue-lives are of primary concern.
This is particularly important when the loading levels are small and the maximum stress
range does not significantly exceed the fatigue-limit.

5.4 Fast Simulation Algorithm (FSA) for the fatigue life assessment procedure
Let us carry out the probabilistic fatigue-life analysis for a long service loading block,
e.g., a long stress history.
The Fast Simulation Algorithm for the probabilistic fatigue-life assessment, based on
the Monte Carlo simulation that assures required accuracy and reduces the computing
time consists of the following steps:
• Calculate the reference fatigue-lives To and T0* , for the real and the simplified
(i.e., C.A. loading with the min and max load values from the stress history) loading
blocks respectively, based on the mean values of all random variables.
• Carry out the Monte Carlo simulation and calculate fatigue-lives, T*, using the
simplified loading block. Neglect the fatigue-limit. Calculate fatigue-lives T*, using
the direct integration of the crack-growth law in the case of the fracture mechanics
approach.
• Translate already calculated fatigue-lives, T*, into fatigue-lives, T, corresponding to
the real loading by multiplying the fatigue-lives obtained in the simulation by the
ratio of the reference lives, To, and To*, i.e., T = T*To/ T0* .

• Calculate fatigue-life Tr, for specified probability of failure Pf, or the probability of
failure Pf, for the required fatigue-life Tr (use possibly the probability distribution
function fitted into the set of results obtained from the simulation).
 Fatigue reliability of structural components 79

The results obtained from the Fast Simulation Algorithm differ from the standard
simulation by less than 5%. In addition, reduction of the simulation time allows one to
increase the number of simulations, resulting in better prediction accuracy.

6 Fatigue safety of structural components

A better than the safety factor, SF, measure of the component’s fatigue-safety is the
probability of its premature failure, Pf. The deterministic fatigue-design requirements
specified in Figure 1 can be replaced now with the following one: Pf < Pfr, where Pfr and
Pf are the required and predicted failure probability of a component respectively.
In the case of stress criterion, the predicted component’s failure probability Pf, is
defined according to the notations specified in Figure 13, as Pf = P(S < R), i.e., as the
probability of the fatigue-resistance R, being smaller than the loading parameter S, and it
can be calculated from equation (6).
∞
Pf = ∫ f S ( s ) FR ( s )ds, (6)
0

where fS and FR are the probability density and the cumulative probability density
functions of the loading S, and fatigue-resistance R, respectively.

Figure 13 The scatter of service loading parameter, S, and resistance, R, and their effect on the
failure probability, Pf, of a structural component designed for an infinite life

It is worth noting that in the probabilistic approach, steady operating conditions do not
mean constant amplitude loading. They mean that the character and magnitudes of
service loading do not change from one machine to the other, i.e., it is assumed that all
characteristic parameters of a loading block (amplitudes, min and max values, etc.)
remain the same.
In the case of fatigue-life criterion, the predicted component’s failure probability, Pf,
is defined as Pf = P(T < Tr), i.e., as the probability of the fatigue-life T, being smaller than
the required one Tr, and it can be calculated from equation (7)
Tr

Pf = ∫ fT (t )dt (7)
0

where fT is the probability density function of component’s fatigue-life, T (Figure 14).

80 H. Jakubczak, W. Sobczykiewicz and G. Glinka

Figure 14 Scatter in fatigue-life of a structural component and the required life, Tr

It is obvious that the safety factor, SF, and the failure probability, Pf, are dependent
quantities. It would be advantageous to find whether it is possible to establish a
relationship between them and to determine the most important parameters used in the
fatigue-life assessment procedure.
In order to answer those questions, a reliability analysis has been carried out
following the simulation algorithm shown in Figure 7 associated with the S–N fatigue-life
assessment method. The following assumptions have been assumed:
• The safety factor is defined as SF = µR/µS for the stress safety criterion and as
SF = To/Tr for fatigue-life criterion for safety.
• The quality of the material used for the fabrication of a component is described by
the coefficient of variation of the fatigue-limit, VR. This value was assumed to be 0.1
(R10) for the standard and 0.05 (R5) for high quality material.
• The quality of manufacturing was expressed in terms of the coefficient of variation
of the notch factor, VK. The magnitudes used were assumed to be 0.1 (K10) for
standard and 0.05 (K5) for high manufacturing quality.
• The variability of the operating conditions of machines was given in terms of the
coefficient of variation of the load scaling factor, VS. The value assumed was to be
0.1 (S10). The analysis was also carried out for steady operating conditions (VS = 0).
The results of the analysis carried out using the infinite life requirement are shown
in Figure 15.

Figure 15 Failure probability vs. safety factor for infinite fatigue-life

 Fatigue reliability of structural components 81

It is worth mentioning that for SF = 1 the probability of premature failure of a component

is equal to ca. 0.5. The higher the value of the safety factor, the lower is the probability of
failure. For a material of standard quality and standard quality of manufacturing, and
unsteady operating conditions, (R10K10S10), it is sufficient to increase the magnitude of
SF to value of 1.25 in order to decrease substantially, the failure probability (Pf = 0.1).
For greater safety and more reliable components, high level reliability is required.
To achieve the probability of failure of Pf = 0.001, the safety factor must be increased to a
value of 1.65.
The high quality of material and manufacturing processes allow the designer to
reduce the traditional safety factor by ca. 15%. If the variations of operating conditions
are small, the required value of the safety factor can be further reduced.
The results of reliability analysis for components designed for finite life are shown in
Figures 16. The dashed lines correspond to value of the S–N exponent, m = 3, and the
solid lines, m = 3.5.

Figure 16 Failure probability vs. safety factor for finite fatigue-lives, m = 3

Preliminary observations indicate that the safety factor values given in terms of the ratio
of fatigue-lives must be significantly greater than those expressed in terms of stresses.
In that case the values of the safety factor required for the same level of probabilistic
safety of a structural component depend on the value of the S–N exponent. For a standard
quality of material and manufacturing process, the safety factor must reach value of 4.4
or 5.6 instead of 1.65, to assure the failure probability of Pf = 0.001. High quality of
manufacturing allows reducing those values by ca. 40%, which means greater values of
mean stress will be allowed. This phenomenon may lead to smaller cross-section of
engineering components what is important for machines such as ground vehicles, cranes,
aircrafts, etc.
The diagrams presented in Figures 15 and 16 show how much one has to ‘pay’ for the
safety of structural components quantified by the specified level of reliability, i.e., how
many times their average fatigue-lives must be greater than the required ones.
82 H. Jakubczak, W. Sobczykiewicz and G. Glinka

7 Discussion

It was mentioned earlier that the same value of the safety factor SF, of structural
components does not assure the same level of their fatigue-safety, measured by the failure
probability, Pf. This is clearly shown in Figure 17(a), where the failure probabilities of
components A, B and C are different for the same value of the safety factor, SF.
In order to assure the same fatigue-safety, the failure probabilities of analysed objects
should be the same (Figure 17(b)). Such an approach indicates that the quality of
materials and manufacturing processes of structural and machine components is very
important from the fatigue-safety point of view. It is also worth to mention, that the
probabilistic concept of equal safety makes it possible to increase the reference life
To, when high quality of the manufacturing process results in smaller scatter. It makes it
subsequently possible to decrease the load carrying cross section area.

Figure 17 The concept of (a) constant safety factor and (b) constant reliability

8 Conclusions

Due to the increasing computational power of modern computers, the computationally

intense probabilistic fatigue-life assessment procedure for structural components seems to
be practically feasible. The simulation procedure based on the Monte Carlo method of
sampling and multiple computer simulations, becomes a very practical and useful tool for
the analysis of fatigue-durability of engineering objects. The concepts discussed in the
paper make it possible to carry out relatively accurate probabilistic analyses with
simultaneous reduction of computing time necessary for the simulation and multiple
repetitions of the fatigue calculation process.
It was also shown that instead of the classical stress based safety factor used in
engineering practice, the probability of failure should be used for assessing the
fatigue-safety of structural components. For the fatigue-life assessment based on the
nominal stress approach, the relation between the deterministic safety factor and the
failure probability has been shown. The results allow one to assess the probability of
premature fatigue-failure of a component using the deterministic safety factor.
 Fatigue reliability of structural components 83

References
Aicher, W.J. (1976) ‘Description of a fighter aircraft loading standard for fatigue evaluation
‘Falstaff’’, Common Report of F+W Emmen.LBF, NLR, IABG.
Benjamin, R.J. and Cornell, C.A. (1970) Probability, Statistics, and Decision for Civil Engineers,
McGraw Hill, Inc., New York.
Bertini, L. and Marmorini, L. (1989) ‘On the characterization of fatigue crack growth
behaviour from a statistical viewpoint via the Paris law coefficients’, in Lieurade, H.P. (Ed.):
2nd Int. Conf. on ‘Fatigue and Stress’, Pub. IITT Int., Gourmay-sur-Marne, France, pp.48–58.
Bolotin, V.V. (1996) ‘Reliability against fatigue fracture in the presence of sets of cracks’,
Engineering Fracture Mechanics, Vol. 53, No. 5, pp.753–759.
Brückner-Foit, A. (1993) ‘Prediction of the lifetime distribution of high-strength components
subjected to fatigue loading’, Fatigue Fracture of Engineering Materials and Structures,
Vol. 16, No. 8, pp.891–908.
Buxbaum, O. (1982) ‘Vergleich der Lebensdauervorhersage nach dem Kerbgrundkonzept und dem
Nennspannungskonzept’, LBF-Bericht, No. FB-162.
Committee on Fatigue and Fracture Reliability of the Committee on Structural Safety and
Reliability of the Structural Division (1982) ‘Fatigue reliability: development of criteria for
design’, Journal of the Structural Division, ASCE, Vol. 108, No. ST1, pp.71–89
Cruse, T.A. (Ed.) (1997) ‘Mechanical reliability design variables and models’, Reliability-Based
Mechanical Design, Marcel Dekker, Inc. New York.
Engesvik, K.M. and Moan, T. (1983) ‘Probabilistic analysis of the uncertainty in the fatigue
capacity of welded joints’, Engineering Fracture Mechanics, Vol. 18, No. 4, pp.743–762.
Hudak Jr., S.J. (1990) A Comparison of Single-cycle Versus Multiple Cycle Proof Testing
Strategies, Report No 4318, National Aeronautics and Space Administration, Washington DC.
Jakubczak, H. (1998) ‘Probabilistic fracture mechanics approach for reliability assessment of
welded structures of earthmoving machines’, Fatigue Design and Reliability, ESIS
Publication 23, Elsevier, pp.229–238.
Jakubczak, H. (2002) ‘Uncertainty of data in the fatigue life assessment of structural components’,
Zeszyty Naukowe Politechniki Warszawskiej, Mechanika, No. 192, in Polish.
Molski, K.L. and Glinka, G. (1981) ‘A method of elastic-plastic stress and strain calculation
at a notch root’, Material Science and Engineering, No. 50, pp.93–100.
Neuber, H. (1961) ‘Theory of stress concentration for shear strained prismatic bodies with
arbitrary non-linear stress strain law’, Journal of Applied Mechanics, Vol. 26, No. 4,
pp.544–550.
Neves, L.A.C. (2001) ‘Reliability analysis of steel components based on FEM’, Engineering
Failure Analysis, Vol. 8, pp.29–48.
Szala, J. and Zawiślak, S. (1990) ‘Application of computer simulation method for determining a
distribution type of construction parts fatigue life’, The Archive of Mechanical Engineering,
Vol. 37, No. 3, pp.145–167.
Wirshing, P.H. (1995) ‘Probabilistic fatigue analysis’, Probabilistic Structural Mechanics
Handbook. Theory and Industrial Applications, Chapman & Hall, New York, Bonn.
Zhao, Z. and Haldar, A. (1996) ‘Bridge fatigue damage evaluation and updating using
non-destructive inspections’, Engineering Fracture Mechanics, Vol. 53, No. 5, pp.775–788.
Zhou, J. and Shen, S. (1998) ‘A study on the reliability assessment methodology for pressure
piping containing circumferential defects’, International Journal of Pressure Vessels and
Piping, Vol. 75, pp.679–697.

View publication stats