International Journal of Fatigue 27 (2005) 1031–1039

www.elsevier.com/locate/ijfatigue

Statistical distribution functions and fatigue of structures

J. Schijve*
Faculty of Aerospace Engineering, Delft University of Technology, Kluyverweg 1, P.O. Box 5058, 2600 GB Delft, The Netherlands
Received 5 August 2004; received in revised form 9 February 2005; accepted 22 March 2005
Available online 13 May 2005

Abstract
Three statistical distribution functions are compared. The log(N) normal distribution, the 3-parameter Weibull distribution and the
3-parameter log(NKN0) normal distribution. The latter function is relatively new. Attention is paid to very low probabilities of failure.
Various sources of scatter of fatigue test data and fatigue of structures in service are recapitulated. Different practical problems for which
statistics are important are defined. Limitations of statistical predictions are briefly discussed. The significance of fatigue acceptance tests and
realistic full scale service-simulation tests are emphasized.
q 2005 Elsevier Ltd. All rights reserved.

Keywords: Fatigue; Statistical distribution functions; Structures; Service life

1. Introduction In general, the statistical analysis of fatigue problems

requires a statistical distribution function of the relevant
The evaluation of statistical fatigue problems of variables. Unfortunately, statistical distributions for fatigue
structures and materials is addressed in numerous publi- problems cannot be derived from a physical description of
cations in the open literature. The variety of objectives the fatigue phenomenon, in spite of the fact that the
includes essentially different purposes. The most-simple knowledge about fatigue damage of a material is fairly well
case is to characterize the variation of fatigue properties, developed in a qualitative way [1]. As a consequence, a
such as the standard deviation of a fatigue limit. A practical statistical distribution function must be assumed. The most
case occurs in comparative test programs if the question is well-known function is the normal distribution (also
whether option A is statistically superior to option B. As an Gaussian distribution). In this paper three distribution
example, is the fatigue life of a joint with a new type of functions are considered; (i) the log(N)-normal distribution
fasteners statistically superior to the life obtained with function, (ii) the 3-parameter Weibull distribution function,
fasteners previously used? In the industry comparative and (iii) the log(NKN0)-normal distribution function. The
fatigue tests are also carried out to check the quality of a latter function appears to be a new one although it was
material or a product procured from suppliers. proposed long ago. The three functions are used for data
Another category of publications is dealing ‘with fitting to results of a two fatigue test series. Attention is paid
designing against fatigue’, predictions on fatigue, and to very low probabilities of failure. The paper continues with
safety issues. Statistical aspects must then be recognized. general observations about scatter of fatigue results caused
Safety factors to avoid fatigue failures of industrial products by different sources. The significance of scatter for practical
are to be used for reasons of economy and liability. This problems of industrial production and the structure in
implies that very low probabilities of failure have to be service is summarized and different types of planning
considered. Sometimes the growth of fatigue cracks is part fatigue tests are stipulated. The paper is completed with a
of a statistical scenario. Inspections should then prevent number of conclusions.
catastrophic failures.

* Tel.: C15 3695 194; fax: C15 278 1151. 2. Three statistical distribution functions
E-mail address: [email protected]

0142-1123/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. The most well known and classical distribution function
doi:10.1016/j.ijfatigue.2005.03.001 is the normal distribution function, also called the Gaussian
1032 J. Schijve / International Journal of Fatigue 27 (2005) 1031–1039

function. The equation for the probability of values v%x is A second 3-parameter distribution function with a lower
given in Eq. (1). limit for the fatigue life will now be discussed. In his book
 y K m 2 Statistical Theory with Engineering Applications [3], Hald
ðx discussed skew distributions. According to Hald, distri-
1 K12
PðxÞ Z pffiffiffiffiffiffi e s dv (1) butions influenced by economics, psychological and
s 2p KN
biological factors are generally skew. Hald also states that
In this equation m is the mean value, s is the standard theoretically it is always possible to determine a function
deviation, and v the integration variable. A linear function g(x) which will transform a skew distribution of x into a
between P(x) and the variable x is obtained if this function is normal distribution. As a noticeable example, Hald
plotted on so-called normal probability paper, see Fig. 1. discusses the logarithmic normal distribution with the
The slope of the straight line depends on s. Usually, log(N) transformation function g(x)Zlog(x). Actually, this applies
is adopted as the variable x for the evaluation of test results to the log(N)-normal distribution function discussed above.
of similar fatigue tests. Hald also mentions the transformation function g(x)Z
Another popular distribution function is the 3-parameter log(xKa) with ‘a’ as an additional parameter of the
Weibull distribution function [2], presented as Eq. (2). logarithmic normal distribution. This function was already
discussed in 1916 by Kapteyn and Van Uven [4]. They
 x K x b
K
0 showed that statistical data of a skew distribution could be
PðxÞ Z 1 K e a (2) fitted by a linear relation between P(x) and log(xKa) if
plotted on normal probability paper. With the present
There are three parameters in this equation. The function notation of N for the fatigue life, the statistical variable x is:
has a lower limit x0 for which P(x)Z0. It implies that
fatigue lives smaller than N0 (corresponding to x0Zlog(N0)) x Z logðN K N0 Þ (3)
cannot occur. This appears to be realistic from a physical The probability function is still the same as in Eq. (1) but
point of view. The parameter a is called the scale parameter, now xZlog (NKN0), and m and s are the mean value and
and b the shape parameter in view of their effects on standard deviation of log (NKN0), respectively. The
the probability density curve. In several publications x0 distribution function is still a normal distribution, but now
is assumed to be zero and Eq. (2) then reduces to the three parameters are involved. The attractive feature is that a
2-parameter Weibull distribution function. If a Weibull lower limit of the fatigue life N0 is introduced as a location
distribution is fitted to test results, three parameters parameter, although mathematically in a different way as in
obviously allow a better correlation with the data than two the Weibull 3-parameter equation. It may be appreciated
parameters. The 2-parameter Weibull distribution function that the normality of the distribution function is maintained.
will not be considered any further in this paper. The option of the log(NKN0)-normal distribution
function apparently escaped the attention of papers on
99
statistical evaluations of fatigue test results. De Jonge in
1983 [5] compared this distribution function with the
3-parameter Weibull distribution function. He showed that
90
approximately similar shapes of the probability density
function, depending on the 3 parameters, can be obtained by
both distribution functions. It then may be expected that
50 the log(NKN0)-normal distribution and the 3-parameter
Weibull distribution can both agree in a similar way with
P (%)

test data. This is explored in the section below.

10

1 3. Fitting distribution functions to test data

0.1 Results of similar fatigue tests can be plotted in a normal

probability plot at an estimated probability of failure
0.01
equal to:
100 200 500 1000 2000
Fatigue life (1000 cycles)
i K 0:5
Pi Z (4)
n
Fig. 1. Test results of 15 similar fatigue tests on welded specimens plotted
in a normal probability graph [6]. Data fitting with the log(N)-normal with n as the number of test results and the rank number iZ1
distribution function (straight line) and the 3-parameter Weibull distri- to n for the fatigue life data in an increasing order of
bution function (non-linear curve, N0Z168 kc). magnitude. Other approximation equations for Pi are
 J. Schijve / International Journal of Fatigue 27 (2005) 1031–1039 1033

presented in the literature (e.g. PiZ(3iK1)/(3nC1)) but the 99

differences with Eq. (4) are small if n is not too small. The
differences are insignificant for the discussion below.
90
3.1. First test series

The results of 15 fatigue tests on a welded joint specimen 50

P (%)
are presented in Fig. 1 [6]. The polished specimens with a
transverse welding joint were made of an 8 mm thick plate
of a low carbon steel. The fatigue tests were carried out at
SaZ242 MPa, RZK1. The straight line in Fig. 1 applies to 10
the log normal distribution function plotted with the
calculated values of m and s (slog(N)Z0.17). Data fitting
1
was also done with the 3-parameter Weibull distribution
function. The values of the three parameters x0, a and b were
obtained with a regression analysis previously used and 0.1
described in [7]. It implies that the root mean square (rms) 300 500 1000 2000 4000
value of the deviations between the test data and the Fatigue life (1000 cycles)
corresponding values according to the distribution function
Fig. 2. Results of 30 similar tests on a notched 2024-T3 Al-alloy specimens,
at the same Pi is minimized. plotted in a normal probability graph [8]. Data fitting with the log(N)-
The test data in Fig. 1 do not show a tendency towards a normal distribution function (straight line) and the 3-parameter Weibull
skew distribution, and both distribution function are distribution function, non-linear curve with N0Z631 kc.
apparently in good agreement with the test results. The
question which of the two functions gives the better and a lower limit might be relevant. The log(N)-normal
agreement with the test result can obviously not be distribution function, i.e. the straight line in this figure,
answered by Fig. 1. The rms deviation between the test does not agree with the test results, but the 3-parameter
results and the distribution function was 0.0260 for the Weibull distribution function gives a good agreement with
log(N)-normal distribution function and 0.0250 for the data. The lower limit x0 corresponds to NZ631 kc,
3-parameter Weibull distribution function. The two values which is fairly close to the lowest test results of 650 kc.
are small (good fit) and hardly different. However, the value The log(NKN0)-normal distribution function for the
of x0 for the 3-parameter Weibull distribution function same test data is shown in Fig. 3. The value of N0 is again
corresponds to a minimum fatigue life of 168 kc (dotted calculated with the rms deviation method. The value now is
vertical line in Fig. 1) whereas a lower limit for the log- N0Z561 kc. It turns out that the log(NKN0)-normal
normal distribution function is theoretically equal to zero. distribution function gives a similarly good fit to the test
The lower tail of the Weibull distribution function in Fig. 1
clearly suggests a deviation from the normal distribution 99
function. As said earlier, a lower limit of statistically
possible fatigue lives appears to be realistic from a physical
point of view. Unfortunately, it cannot be concluded from 90
Fig. 1 that the predicted lower limit of 168 kc according to
the 3-parameter Weibull distribution function is realistic
information. 50
P (%)

The log(NKN0)-normal distribution function was also

used to fit the test data in Fig. 1. The least square method
was adopted again to determine the three parameters. It
10
turned out that N0 was as small as 1100 cycles which implies
that the log(NKN0)-normal distribution function practically
coincides with the log(N)-normal distribution function.
1

3.2. Second test series

0.1
300 500 1000 2000 4000
Results of a larger test series with 30 similar fatigue tests
on edge notched specimens (KtZ2.85) of an aluminium Fatigue life (1000 cycles)
alloy (2024-T3 Alclad, thickness 5 mm) [8] are presented in Fig. 3. The same data of Fig. 2 but now data fitted with the log(NKN0)-
Fig. 2, again in a normal probability graph with log(N) as normal distribution function. Slightly different non-linear curve with
the variable. The results clearly show a skew distribution N0Z561 kc.
1034 J. Schijve / International Journal of Fatigue 27 (2005) 1031–1039

results. The rms deviation for the log(NKN0)-normal covers two periods: the crack initiation period and the crack
distribution function and the 3-parameter Weibull distri- growth period. The fatigue mechanism during the crack
bution function are 0.0189 and 0.0193, respectively. Again initiation period is a surface phenomenon. Crack growth in
the two rms deviations are small and hardly different. The the second period is a material bulk phenomenon associated
rms deviation for the log(N)-normal distribution function is with the fatigue crack growth resistance of the material. The
0.0459, which is a significantly larger value. It confirms that crack initiation is much more sensitive to various circum-
the log(N)-normal distribution function does not agree with stances at the material surface, with local surface roughness
the test results. as a predominating factor. At a high Sa (low-cycle fatigue)
The somewhat lower N0-value of 561 kc for the the crack can occur relatively easy and scatter is low. At a
log(NKN0)-normal distribution function, as compared to low Sa (high-cycle fatigue) the initiation of the first
631 kc of the 3-parameter Weibull distribution function, microcrack meets with more difficulties depending on the
appears to be a bit more realistic when comparing the local circumstances. As a result scatter is larger. For
curves for the lower test data of the skew distributions in amplitudes around the fatigue limit it is even a kind of a
Figs. 2 and 3. threshold problem whether a microcrack will be initiated
In conclusion, the analysis of the two test series indicate and can pass the very first micro-barriers for further growth.
that a good fit of test results in the range of the test data is The previous discussion on three statistical distribution
very well possible with the log(NKN0)-normal distribution functions was related to the finite life region of the S–N
function and the 3-parameter Weibull distribution function. curve. For scatter of the fatigue limit, the statistical fatigue
This is not true for the log(N)-normal distribution function if life distribution functions are not applicable. It is generally
the test data indicate a skew distribution. Secondly, assumed that the normal distribution applies to the fatigue
extrapolation of the distribution function to very low limit, a common assumption for strength properties.
probabilities of failure remains an uncertain procedure. It Although this may be considered to be reasonable, it should
means that the distribution function is actually unknown. be recognized that the distribution function is not really
known. Statistical trends about the fatigue limit can be
obtained from fatigue experiments. The stair-case method is
4. Some general trends of scatter in fatigue tests the most well known procedure to obtain an estimate of the
mean value and the standard deviation [11]. Unfortunately,
4.1. Scatter and the S–N curve it is an expensive procedure because to obtain a reasonable
confidence level, a large number of long lasting tests is
Observations on scatter of fatigue were usually made in required to obtain these properties with a reasonable
laboratory experiments under constant-amplitude (CA) confidence level.
loading. In general scatter was larger at low stress The fatigue limit is supposed to be a relevant material
amplitudes and lower at high amplitudes as shown by property for components subjected to very high numbers
Fig. 4. This trend is discussed in [9,10]. The fatigue life of cycles in service while failure in service is not allowed.

Fig. 4. S–N curve with narrow scatter band at high Sa (low cycle fatigue) and wide scatter band at low Sa (high-cycle fatigue). Three probability density
functions are indicated [10].
 J. Schijve / International Journal of Fatigue 27 (2005) 1031–1039 1035

This applies to several engine components. It is not always all specimens from a single sheet, plate or bar, and
realized that the fatigue limit of unnotched specimens does manufacturing of all specimens with great care in one
not indicate whether a material is notch sensitive. Moreover, machine shop. From a fatigue point of view, materials are not
different sources of scatter apply to laboratory experiments fully characterized by the standardized composition of the
and the production of components as discussed in [9,10]. major alloying elements. In several materials crack nuclea-
tion starts from microscopic inclusions at or near the material
4.2. Scatter under variable-amplitude loading surface. It implies that the contents of impurities, the material
production and the heat treatment can affect the fatigue
In the last century, fatigue tests with a variable-amplitude properties. It has been observed that the fatigue properties of
(VA) load sequence were often carried out to check damage aircraft aluminium alloys can be different for the same
accumulation models, noteworthy the Pålmgren-Miner rule. nominally alloy obtained from different producers, and even
Scatter in such tests was occasionally analyzed. It was for different batches of the same producer [10]. Similarly, the
concluded in [9] that the standard deviation in such tests is effect of impurities on the fatigue limit of low-alloy high-
of the same order of magnitude as obtained in CA tests at the strength steels (alloying elements Cr, Mo, Ni, V) heat treated
highest stress amplitudes applied in the VA tests, even if to high stress levels can also be significant [10]. The cleaner
high fatigue lives are obtained in the VA tests. It implies that steels are superior from a fatigue point of view. The effect of
the scatter can be relatively low under VA loading. The different sizes and shapes of inclusions on crack initiation
trend was explained by recognizing the prevailing influence and initial crack growth has extensively been investigated by
of the most severe cycles of the VA load sequence on the Murakami [12].
crack initiation mechanism. In view of this observation, it The production of components and structures in the
must be concluded that it is not correct to predict the scatter industry cannot be associated with an exactly constant
under VA loading by pooling scatter data obtained in CA quality of the material procured from year to year, in spite of
tests for the stress amplitudes used in the VA tests. ordering the material in agreement with the same material
specification.
4.3. Scatter and size effects
4.5. Scatter due to production variables
Effects of the specimen size on the fatigue limit of both
unnotched and notched specimens have drawn attention in The industry has to take care of the production quality of
the past. It was thought that a weakest link theory could a structure or a component to be maintained at a good
account for these effects. It is indeed an intriguing but standard. This problem is easily recognized for mass
difficult question. Size effects should then be related to the production of various vehicles with motorcars as out-
probability of favourable locations for microcrack nuclea- standing examples. It also applies to aircraft, which are
tion in unnotched specimens and at the root of a notch in produced during many years. The quality of the product can
notched specimens. In the latter case, stress gradients must change over the years for several reasons, partly because of
be considered. A fundamental question is whether the variations of the production techniques and the materials
probability of microcrack nucleation will be a volume effect used, but also because of introducing a new and more
(three-dimensional), a surface effect (two-dimensional), or economic production process.
a corner edge effect (one-dimensional) [10, p. 144]. The Walter Schütz carried out an investigation on different
probability of having favourable sites for microcrack types of riveted lap joint used in the aircraft industry [13].
initiation can also depend on the type of material including Some years later he was asked to repeat a series of these
inclusions and impurities. Although the occurrence of a size tests on riveted joint specimens manufactured according to
effect can be understood from a weakest link point of view, the original drawing. He then found that the fatigue life was
this has not led to a rational theory to account for the effect. about three times larger then obtained in the first test
The same is true for the correlation between the fatigue limit program. Schütz showed that the deviating results were due
of notched and unnotched specimens. In practice, empirical to more intensive squeezing of the rivets. Later this riveting
rules are still used for the prediction of notch effects. variable was extensively investigated in Delft by Müller
[14]. It confirmed that the squeezing force had a large effect
4.4. Scatter due to variations of the material on the fatigue life and thus could be a source of scatter due
to the variability of the production.
It is a good practice of laboratory investigations to plan In the motor car industry, but also in other industries,
fatigue experiments in such a way that scatter will be various components are procured from small companies
minimal. This is done to avoid that scatter of the test results specialized in the production of those components. It then
will obscure quantitative conclusions about the phenomenon should be ascertained that the product meets the standard of
to be studied. Useful information can be obtained with a sufficient quality for each supply. Partly this can be done by
lower number of tests if scatter can be avoided or minimized. simple means such as hardness testing, visual inspection
Well-known precautionary measures are the production of and tolerance measurements. However, with respect to
1036 J. Schijve / International Journal of Fatigue 27 (2005) 1031–1039

If the various loads on the structure are defined, the loads

inside the structure depend on the dynamic response of the
structure. A simple case is a pressure vessel, which is
periodically loaded to a certain pressure level. However,
complex problems can arise for e.g. a motorcar, a coach, a
truck, or an aircraft wing where the dynamic response may
be rather complex. Different loads in various directions and
not necessarily in phase can occur simultaneously.
Fig. 5. A forged control lever of a motorcar tested under a service- By the end of the previous century, techniques and
simulation fatigue load history to check the variability of the quality of apparatus for measuring load-time histories were developed
the product. In two years 130 levers were tested. Scatter was large, to a high level of perfection. Load-time measurements can
slog(N)Z0.29 [15]. even be used as input signals for simulation tests. In the
automotive industry, this is attractive, because it opens the
the fatigue quality, this can only be done in a realistic way possibility for a virtual reality simulation in the laboratory.
by some kind of a ‘fatigue acceptance test’. Several car Unknown load spectra can be measured under service
industries have developed testing procedures for this conditions if a structure is available for that purpose.
purpose. A few items of a large number of components of Continuous measurements are sometimes made on military
a supply are randomly selected for fatigue testing under a aircraft in order to know the scatter of the load spectra in a
cyclic load history, which should discriminate between fleet of aircraft. In summery, load spectra, if they cannot be
sufficient and insufficient quality. The load sequence is calculated, can now be measured without great difficulties.
conservatively chosen for the loads on the component in
service. Depending of the component, the sequence may be
either CA or VA loading. 5. Scatter of the fatigue endurance in service
In an older investigation of the German motorcar
industry 130 control levers shown in Fig. 5 were tested Case histories of scatter of fatigue lives of a structure in
under a service-simulation load history, [15]. The levers service are rarely published. If some similar failures occur,
were taken from different deliveries during a period of and if these failures are unacceptable, remedial actions are
2 years. The material was a 0.4C steel (40Mn4) with carried out to prevent more failures in the future. Just one or
SUz850 MPa. The life to failure showed much scatter with two data are then available. Secondly, fatigue failures are
slog(N)Z0.29. Analysis of the 130 levers indicated that generally associated with poor publicity for either the
scatter was mainly caused by changing dimensions due to industry or the operator of the structure. As a result
wear out of the forging dies and to somewhat different interesting case histories are documented in restricted
machining shapes near the fatigue critical cross section. company reports, but not in the open literature.
The situation is different for civil aircraft structures
4.6. Scatter and load spectra because accident reports are not always confidential. If a
fatigue failure with serious consequences has occurred,
Scatter of load spectra applied to a structure in service relevant information is spread amongst the aircraft oper-
is an entirely different problem that the other sources ators. It must be explored whether similar cracks are present
summarized before. Those sources could be manipulated in the same type of aircraft. If this is true, it is a case of a
by the manufacturer of the structure. However, the loads symptomatic fatigue critical component. Such cases have
on the structure are depending on the operator of the been reported [9]. In 1976, a Hawker Sideley 748 lost a
structure. Nevertheless, the load spectrum is an essential wing due to a large fatigue crack in the tension skin. Cracks
input for any prediction on the fatigue performance of a were then observed in 19 similar aircraft. In 1977, a Boeing
structure. In general, a structure will not be used in the 707 lost a horizontal tail plane. Cracks were found in 12
same way by all owners. Severe load spectra and mild load similar aircraft.
spectra may be applicable to the same structure. An Sometimes it can be shown that a failure in service is due
overloaded truck driving along roads in a poor condition to an incidental cause, such as damage of the structure
will see a relatively severe load spectrum. An example of which in general should not occur. The problem then is how
an opposite case occurred to a Fokker aircraft, which was such damage can be prevented. A well-known example of
used by a firm as an executive aircraft. Since the an incidental case is the crash of an F-111 aircraft due to a
executives did not fly when stormy weather was fore- relatively large flaw already present in the wing skin plate
casted, the counting accelerometer data of the aircraft material as delivered to the aircraft industry. The flaw was
indicated a low severity spectrum [10, p. 326]. not detected during inspection of the material. The flaw
Scenarios about how a structure will be used and can be initiated fatigue crack growth followed by wing failure [16].
used must lead to a decision about the design load spectrum It has led to various precautionary measures to eliminate
for the analysis of the fatigue behaviour of a structure. similar accidents in the future.
 J. Schijve / International Journal of Fatigue 27 (2005) 1031–1039 1037

Dealing with scatter of fatigue properties of a structure is and if inspections are not carried out a sufficient crack
not just a single well-defined problem. The requirements initiation fatigue life must be provided. The statistical
applicable as a fatigue design scenario of a structure can be problem is then associated with scatter of the crack initiation
quite different. Some major categories are: life, and of course again with various sources of scatter.

(1) Fatigue cracks are unacceptable in service. 5.2. Fatigue cracks in service are acceptable, but failure
(2) Fatigue cracks in service are acceptable, but failure should not occur
should not occur.
(3) Fatigue cracks in service are acceptable, even if failure This problem is well known as the damage tolerance
can occur. philosophy. It is documented in airworthiness requirements
for aircraft, both for civil and military aircraft. The same
Brief comments on these conditions are made below. approach is also adopted in various other technical
disciplines, noteworthy in large welded structures.
5.1. Fatigue cracks are unacceptable in service It is obvious that the application of the damage tolerance
procedure requires information about several aspects. It
Two different cases should be mentioned. First, the includes the crack initiation lifetime, possibilities for crack
structure during its life time will be loaded by large numbers detection, crack growth, inspection periods, and final
of cycles (millions) with relatively high stress amplitudes, failure. All aspects are subjected to scatter. With respect
which was defined in [10] as a flat load spectrum, see Fig. 6. to crack detection the concept of the probability of detection
It can apply to high speed rotating systems, such as engines, (POD) has been introduced for aircraft operations [17].
pumps and helicopters. It general crack initiation will not be Another source of scatter is again the load spectrum of the
allowed because crack growth can lead to failure relatively structure.
soon. Frequent inspections might prevent final failure
(damage tolerant in aerospace engineering), but in many 5.3. Fatigue cracks in service are acceptable, even if failure
cases inspections are not a feasible option. Theoretically the can occur
criterion is to design for an infinite life of the structure, or in
more practical terms, designing with allowable stress levels Failures in service can be acceptable if no liability
below the fatigue limit. The statistical problem then is how consequences are involved and if the economy of the
to account for possible scatter of the fatigue limit, and also industry is not adversely affected. The latter condition can
for other possible sources of scatter (e.g. component still be significant in a way that sufficient life until failure
production, load spectra). should be available. In that case scatter of the fatigue life
A second case of cracks being not acceptable is associated until failure is still involved.
with ‘finite life structures’. The load spectrum with load
cycles above the fatigue limit of the structure implies that
crack initiation can occur. This applies to load spectra with 6. Discussion
load cycles with amplitudes above the fatigue limit, such as
for the steep load spectrum in Fig. 6. But it can also apply to Various sources of scatter have been revealed in the past
structures, which meet with low numbers of high-amplitude as a result of extensive laboratory investigations. Moreover,
cycles, e.g. pressure vessels. Crack initiation is possible it is well understood why and how these sources are
affecting fatigue properties, but the understanding is
qualitative in nature. Because of the practical design and
100 production circumstances of the industry and the exploita-
load level (% max. load)

flat spectrum
tion of structures in service, the statistical properties of a
80 steep spectrum structure in service cannot be predicted with a clearly
defined and rational theory. The relevant distribution
60
functions of the production variables and load spectra are
40 quantitatively unknown. Moreover, assumed distribution
functions for very low probabilities of failure offer
20 additional uncertainties as discussed earlier. Under such
conditions, clever engineering judgement should be guiding
to some practical procedures for the industry as well as for
1 10 102 103 104 105 106
the operator of structures in service.
number of exceedings in 1000 hrs As pointed out before, ‘fatigue acceptance tests’ have
Fig. 6. A flat load spectrum with many high-amplitude cycles and a steep been introduced for mass production. Such tests should be
load spectrum with a relatively low number of high amplitude cycles and as realistic as possible. The tests should be carried out on
many low-amplitude cycles [10]. components as produced or delivered by sub-contracted
1038 J. Schijve / International Journal of Fatigue 27 (2005) 1031–1039

suppliers. The load history in the tests should be component specimen, including small components, welded
representative and conservative for the expected service and mechanical joints and various technical metallic
load histories. Knowledge about fatigue damage accumu- materials. Average values of slog(N) of the test series were
lation associated with load cycle interactions is essential for in the range of 0.10–0.20. It depends on the uniformity of
planning these tests. Employing the Miner rule for the quality of the component if a low or a high value in this
simplification of the test load history must be discouraged range must be selected.
because the Miner rule is unreliable for the assessment of Instead of the log(N)-normal distribution function, a
equal damage. Results of periodic fatigue acceptance tests 3-parameter distribution can also be used for the above
are a kind of a data bank which can provide information question. However, in addition to the mean value, two other
about the uniformity of the production population. parameters have to be estimated. It then becomes difficult to
If the number of structures to be manufactured is not very arrive at credible choices. For the 3-parameter Weibull
large, a realistic service-simulation fatigue test on a distribution function the shape parameter could be selected
prototype structure is advisable. This procedure is fre- at a value bZ3.5 because then the distribution becomes
quently adopted in aerospace engineering. The test article somewhat similar to a normal distribution. For the log(NK
should preferably include the full structure, but an N0)-normal distribution function, the standard deviation of
alternative is to test the most fatigue critical part of the log(NKN)0 could again be chosen in the range 0.10–0.20.
structure selected by FE analysis of the stress distribution in However, for both 3-parameter distributions a value for the
the structure. Such tests are more informative if they are run location parameter N0 is needed. As discussed earlier, there
until failure occurs. If a failure does not occur, then the test is no rational criterion for this question. Anyway, it will be
can be continued at an increased severity level of the load clear that the 2-parameter log(N)-normal distribution
history. function will be on the safe side (see Fig. 1).
Actually, a major argument for a full-scale test on a new In the literature, the calculation of statistical confidence
aircraft structure is also to find fatigue critical areas. If levels are proposed in order to be more safe about statistical
cracks are found in the test, then a local redesign is probabilities. However, a basic requirement for such
introduced. As a result, the aircraft structure at the end of the calculations is that all structures can be considered to be a
flight-simulation fatigue test may still be free from visible sample of one and the same statistical population with a
fatigue cracks. It is common to continue such tests until known distribution function. In general this requirement is
twice the anticipated operational life in service. The not satisfied. As a consequence, assumptions must be made
structure then is torn down in order to see whether there based on available experience and understanding. A special
are small cracks in the joints and other hidden areas. problem is how to account for environmental conditions and
Prototype testing is also becoming more usual for motor- structural maintenance. It remains a delicate subject to
cars and trucks. A fully instrumented prototype (strain gages balance the significance of the various relevant conditions.
and accelerometers) is subjected to severe driving circum- Furthermore, aspects associated with safety, economy and
stances to see whether the dynamic behaviour of the liability may come into play in order to see whether
structure agrees with FE calculations, and to indicate conservative assumptions can be made. Profound judgement
locations with severe stress concentrations. It is more or is required.
less a function test to find local areas where the structure can
still be improved before it goes into production.
The results obtained in a service-simulation fatigue test
provide information about the fatigue performance of a 7. Summarizing conclusions
single structure. Statistical information for a large number
of structures is not obtained. Again based on engineering 1. A comparison has been made between three statistical
judgement, assumptions have to be made about scatter for a distribution functions: (1) the log(N)-normal distribution
large number of structures. The uniformity (or the lack of function, (2) the 3-parameter Weibull distribution
uniformity) of the industrial production of the structure, function, and (3) the log(NKN0)-normal distribution
including the uniformity of the material, must be considered function. The third one is a normal distribution of
for this purpose. Data about scatter observed in test series on log(NKN0), a function which was largely overlooked in
components can be instructive. Let it be assumed that an the literature. The second and the third function both
estimate of the mean value of the distribution is available. contain a statistical minimum of the fatigue life (N0) as a
The estimate can be the test result obtained in a service- 3rd parameter, although mathematically in a different
simulation fatigue test, or a predicted value, which is way. Both functions gave a good data fit of the results of
considered to have a reasonable credibility. The most simple 30 similar tests with a skew distribution, but it still has to
approach then is to assume that the log(N)-normal be recognized that the distribution function is actually
distribution function is applicable. A standard deviation of unknown. The physical significance of the estimated
log(N) must then be chosen. Adenstedt [18] analyzed scatter N0-value remains unclear. A major problem of fatigue
data of numerous test series covering CA and VA tests of life distribution functions is the extrapolation to low
 J. Schijve / International Journal of Fatigue 27 (2005) 1031–1039 1039

probabilities of failure which are not supported by [3] Hald A. Statistical theory with engineering applications. New York:
Wiley; 1957.
relevant experimental data. [4] Kapteyn JC, Van Uven MJ. Skew frequency curves in biology and
2. A summary is presented of different sources for scatter in statistics. Groningen: Hoitsema Brothers; 1916.
laboratory test series and scatter of fatigue lives in [5] De Jonge JB. Some properties of the Weibull-distribution (in Dutch).
service under practical circumstances. Although labora- Nat Aerospace Laboratory, NLR, Amsterdam. Memorandum SB-83-
tory test data can be informative, it remains a difficult 008; 1983.
[6] Results of a test series carried out at the DAF industry, Eindhoven.
issue to account for scatter to be expected in service. Data provided by HTM van Lipzig.
Theories with a physical and rational basis for this [7] Schijve J. A normal distribution or a Weibull distribution for fatigue
problem do not exist. lives. Fatigue Fract Eng Mater Struct 1993;16:851–9.
3. If the industry is forced to consider the statistical [8] Schijve J, Jacobs FA. Fatigue tests on notched and unnotched clad 24
variability of mass production for reasons of economy, S-T sheet specimens to verify the cumulative damage hypothesis. Nat
Aerospace Laboratory, NLR, Amsterdam. Report M.1982; 1955.
safety and liability, so-called fatigue acceptance tests on [9] Schijve J. Fatigue predictions and scatter. Fatigue Fract Eng Mater
components should be carried out under a conservative Struct 1994;17:381–96.
and realistic fatigue load history. A realistic service- [10] Schijve J. Fatigue of structures and materials. Dordrecht: Kluwer
simulation fatigue test on a prototype structure is also Academic Publishers; 2001.
advised although it provides little information on scatter [11] A guide for fatigue testing and the statistical analysis of fatigue data.
American Society for Testing and Materials, Philadelphia, ASTM
due to the variability of the production. However, such a STP No.91-A; 1963.
test can also show the dynamic response on external [12] Murakami Y. Metal fatigue: effects of small defects and nonmetallic
loads. Moreover, design improvements can then still be inclusions. Oxford: Elsevier; 2002.
made at an early stage. [13] Schütz W. The endurance of single shear light-alloy riveted joits (in
4. In order to reduce scatter of the fatigue properties of a German). Laboratorium fúr Betriebsfestigkeit LBF, Darmstadt,
Bericht Nr.F-47; 1963.
structure, the uniformity of the production process must [14] Müller RPG. An experimental and analytical investigation on the
be carefully controlled. fatigue behaviour of fuselage riveted lap joints. The significance of the
5. If the risk of fatigue failures of a structure is a crucial rivet squeeze force, and a comparison of 2024-T3 and Glare 3. Doctor
issue, load-history measurements in service should be thesis, Delft University of Technology; Oct. 1995.
advised. [15] Lipp W. Statistical analysis of fatigue life scatter of a large number of
forged components (in German). Laboratorium für Betriebsfestigkeit,
LBF, Darmstadt, TM Nr.56/70; 1970.
[16] Mar JW. Structural integrity of aging airplanes: a perspective. In:
Structural integrity of aging airplanes. Berlin: Springer; 1990.
References p. 241–262.
[17] POD assessments of NDI procedures using a round robin test.
[1] Schijve J. Fatigue of structures and materials in the 20th century and AGARD report 809. Advisory Group for Aerospace Research and
the state of the art. Int J Fatigue 2003;25(8):679–702. Development. NATO, Paris; 1995.
[2] Weibull W. A statistical distribution function of wide applicability. [18] Adenstedt R. Scatter of the fatigue strength (in German). Doctor
J Appl Mech 1951;18(3):293–7. thesis, Technical University Claustahl, Germany; 2001.