Derivation of Endurance Curves From Fatigue Test Data, Including Run-Outs

ESDU 97024
TM
Engineering Sciences Data Unit Issued June 2003

An IHS GROUP Company
Derivation of endurance curves from

ESDU Copyright material. For current status contact ESDU.
fatigue test data, including run-outs
Associated software: ESDUpac A9724

VIEWpac W9724
Endorsed by
The Royal Aeronautical Society
The Institution of Mechanical Engineers
ESDU
TM
Engineering Sciences Data Unit 97024

ESDU DATA ITEMS
Data Items provide validated information in engineering design and analysis for use by, or under the supervision of,
professionally qualified engineers. The data are founded on an evaluation of all the relevant information, both published and
unpublished, and are invariably supported by original work of ESDU staff engineers or consultants. The whole process is
subject to independent review for which crucial support is provided by industrial companies, government research
laboratories, universities and others from around the world through the participation of some of their leading experts on ESDU
Technical Committees. This process ensures that the results of much valuable work (theoretical, experimental and
operational), which may not be widely available or in a readily usable form, can be communicated concisely and accurately
to the engineering community.
We are constantly striving to develop new work and review data already issued. Any comments arising out of your use of our
data, or any suggestions for new topics or information that might lead to improvements, will help us to provide a better service.
THE PREPARATION OF THIS DATA ITEM

The work on this particular Data Item was monitored and guided by the Fatigue Committee and the Stress Analysis and
Strength of Components Committee. The Fatigue Committee first met in 1955 and now has the following membership:
Chairman
Mr A.R. Simpson British Aerospace Airbus Ltd
Members
Dr A. Burgess Messier-Dowty Ltd
Mr K.E. Cheverton Independent
Mr D. Crouch MATRA BAe Dynamics
Mr R.G. Eastin* Federal Aviation Administration, USA
Ms C. Escobedo* Airbus, Germany
Dr M. Miller* Boeing Commercial Airplane Group, USA
Mr D. Painter Westland Helicopters Ltd
Dr P.M. Powell QinetiQ
Dr S.H. Spence BAE SYSTEMS, Warton
Dr C.H. Wang* Aeronautical and Maritime Research Laboratory, Australia
Mr R.T. Watanabe* Boeing Commercial Airplane Group, USA
Dr R.N. Wilson Independent.
*
Corresponding Member

Member Emeritus
ESDU
TM

DERIVATION OF ENDURANCE CURVES FROM FATIGUE TEST DATA, INCLUDING
RUN-OUTS
CONTENTS
Page
1. INTRODUCTION, THEORY AND APPLICATION 1

1.1 INTRODUCTION 1
1.2 NOTATION 1
1.2.1 Suffixes 2
1.3 FATIGUE TEST DESIGN 2

1.3.1 Definition of Population 2
1.3.2 Sample Size and Stress/Strain Levels 2
1.3.3 Censored Data 3
1.4 THEORY AND APPLICATION 3
1.4.1 Form of SN Curve 3
1.4.2 Parameter Estimation 4
1.5 CURVE-FITTING PROCEDURE 5
1.5.1 Running Strategy 5
1.5.2 User Input 5
1.5.3 Program Output 6
1.5.4 Outliers 7
1.6 DERIVATION AND REFERENCES 8
1.6.1 Derivation 8
1.6.2 References 8
2. ESDUpac A9724 AND VIEWpac W9724 9

2.1 INTRODUCTION TO THE PROGRAMS 9
2.2 SCOPE AND LIMITATIONS OF THE PROGRAMS 9
2.3 DATA INPUT REQUIREMENTS FOR THE FORTRAN PROGRAMS 10
2.3.1 Running the Programs 11
2.3.1.1 Program A9724 11
2.3.1.2 Program W9724 11
2.4 OUTPUT FROM THE PROGRAMS 13
2.4.1 Output of Input Data 13
2.4.2 Results and Output 13
2.4.2.1 Failure to provide a solution owing to errors in the input data 13
2.4.2.2 Failure to provide a solution from valid input data 14
i
ESDU
TM

2.5 EXAMPLES 15
2.5.1 Example 1, Data for PH13-8Mo Forged Steel Bar taken from MIL-HDBK-5G 15
2.5.1.1 Two parameter curve 15
2.5.1.2 Three parameter curve 19
2.5.2 Example 2, Data from 2025-T6 Forged Aluminium Alloy Slab 21
2.5.2.1 Automatic selection of curve 21
2.5.3 Example 3, Data for a Metal Matrix Composite 25
2.5.3.1 Two parameter curve 25
2.5.3.2 First data point removed 27
2.5.3.3 q predefined 28
APPENDIX A THE METHOD OF MAXIMUM LIKELIHOOD APPLIED TO THE ANALYSIS OF

CENSORED FATIGUE DATA
A1. INTRODUCTION 31
A1.1 NOTATION 31
A1.1.1 Suffixes 31
A1.1.2 Operators 31
A2. THE METHOD 32

A2.1 Uncensored Observations 32
A2.2 Right-censored Observations (Run-outs) 32
A2.3 Left-censored Observations 32
A2.4 Application of Maximum Likelihood Method 32
ii
ESDU
TM

DERIVATION OF ENDURANCE CURVES FROM FATIGUE TEST DATA, INCLUDING
RUN-OUTS
1. INTRODUCTION, THEORY AND APPLICATION
1.1 INTRODUCTION
Section 1 describes the theory and application of a method for consistent derivation of an SN curve from
a set of fatigue data. The method is applicable to data for both metallic and non-metallic materials as well
as to data from tests on coupons or specimens representing structural features. The method encompasses
the low-cycle, high-cycle and endurance limit regimes and also takes account of censored data (for example,
run-outs) in a rigorous manner.
Data relating to static failures, corresponding to a single load application, are not catered for. In addition
to this, it is assumed that the curve of logS versus logN is concave upward; this is believed, however, to
encompass the vast majority of published SN data.
S may be in terms of a maximum, an alternating or an equivalent stress, strain or loading with the life, N,
expressed in terms of number of cycles, time or some other convenient unit (for example, repeated program
blocks). The cycles may be of either constant or variable amplitude. The derived curve is the median curve,
that is, it represents a fifty per cent probability of failure.
Ideally, the process of deriving an SN curve should begin at the test planning stage, the aim being to obtain
the maximum amount of useful information from a limited number of tests. Section 1.3 therefore gives
guidance on the design of fatigue tests with a view to optimising the testing procedure.
Section 1.4 outlines the theory on which the derivation of the SN curve is based. Section 1.5 describes
the curve-fitting procedure that is implemented by a Fortran computer program, the use of which is
described in Section 2.
Definitions of the statistical terms used in this Data Item can be found in Data Item No. 92040*. A definition
of the normal distribution is given in Data Item No. 91041.
1.2 NOTATION
The units for S and N are totally independent and can, therefore, be chosen as required. However, the values
of P 1, P 2 and P 3 will only apply to the chosen units for S and N (see Section 1.4).
C number of parameters by which fitted SN curve is defined (see Table 2.1)
D number of values of S for which values of N are required to be calculated by the program
(see Table 2.1)
l likelihood (see Section 1.4.2 and Appendix A)
M number of pairs of SN values in a particular data set
N life
*
Data Item No. 92040, An introduction to the statistical analysis of engineering data.

Data Item No. 91041, The statistical analysis of data from normal distributions, with particular reference to small samples.
Issued June 2003
1
ESDU
TM

P1, P2, P3 constants (see Equation (1.1))
q, r constants (see Equation (1.1))
S stress, strain or loading
v population coefficient of variation of S
random error (see Equation (1.2))
In this Data Item ln represents the natural logarithm of a number while log represents the logarithm to any
base of a number.
1.2.1 Suffixes
Suffixes 1, 2, . . . D, applied to S, denote specific values of S for which a value of N is required to be

calculated by the program.
Suffixes H and L, applied to l, denote values for higher and lower order models respectively (see Section
1.5.3).
Suffixes 1, 2, . . . M, applied to S and N, denote specific pairs of SN values.
1.3 FATIGUE TEST DESIGN
1.3.1 Definition of Population
The first step in designing a fatigue testing programme is to define clearly the population for which an SN
curve is to be obtained. When referring to fatigue testing, the population may be considered to be the set
of all possible test specimens of a particular type and all tested in a particular manner; the specimens actually
tested will constitute a sample from that theoretical population. It is important that the sample is as
representative as possible of the population and that all relevant details of the specimens, the method of
testing and the test conditions are recorded for future reference.
Fatigue data from different sources can be combined, but it should be borne in mind that the population
will then usually consist of a combination of two or more individual populations. As a result of this, the
scatter may be very large and the derived curve may cease to be representative of any of the individual
populations.
1.3.2 Sample Size and Stress/Strain Levels
For a given population, the greater the total number of tests that are performed, the more representative of
the population will be the resulting SN curve. Also, the larger the scatter, the greater the number of test
results required to obtain a median curve. Furthermore, the wider the range of endurances over which the
SN curve is to be derived, the larger the number of test results that will be needed to obtain a sufficiently
representative median curve.
A lower limit on the number of test results that is required is also imposed by the number of parameters
used to define the derived SN curve (see Section 1.5.2).
2
ESDU
TM

The minimum number of specimens to be tested will also depend on the amount of information about the
form and parameters of the curve that already exists, since this will determine the number of unknown
constants that must be evaluated from the test data.
The optimum allocation of a given total number of tests to different stress or strain levels will also depend
on the degree of scatter. The larger the scatter, the more specimens should be tested at each level.
Conversely, if the degree of scatter is low it would be advantageous to perform tests at an increased number
of stress or strain levels.
In general, it is advisable to choose test levels that result in a reasonably even coverage of the range of
endurances of interest. In cases where it is not necessary for all the specimens to be tested simultaneously,
good coverage of the range of endurances can sometimes be best achieved by adjusting the stress or strain
levels as the test programme proceeds, on the basis of the information gained up to that point. Exceptions
to this are cases where the relationship between logS and logN is known from experience to be linear over
the range of interest. In such cases a good strategy may be to focus testing activity at the lower and upper
extremes of the range of endurances of interest.
It is impossible to prescribe the minimum total number of test results that will be needed to obtain a
sufficiently representative curve because of the number of separate factors involved. However, two tests
in which the specimen fails, performed at each of three different stress or strain levels, is considered an
absolute minimum but even this would be unlikely to give a reliable median curve (with the exception of
cases where the relationship between logS and logN is known to be linear). Section 1.5.2 gives guidance
on checking retrospectively whether or not the sample size is adequate.
1.3.3 Censored Data
Although unfailed specimens (run-outs) are taken into account in a mathematically rigorous manner by the
computer program that accompanies this Data Item, the aim of a testing programme should be to test to
failure as many of the specimens as possible over the range of interest, since more information is obtained
from failed specimens than from run-outs. The only exception to this rule is when the user is particularly
interested in determining the endurance limit, in which case a number of run-outs is desirable.
Left-censored data, that is data where the number of cycles to failure has not been recorded but is known
to be less than a certain value, can also be used in the analysis and should not therefore be discarded.
1.4 THEORY AND APPLICATION
1.4.1 Form of SN Curve
The program that accompanies this Data Item derives from the fatigue data a curve of the form:
q r
S = P1 N + P2 N + P3 , (1.1)
where P 1, P 2, P 3, q and r are constants with P 1, P 2, P 3 0 and q < r 0 .
This general equation is capable of representing a wide range of shapes of SN curve where S may be in
terms of a maximum, an alternating or an equivalent stress, strain or loading with the life, N, expressed in
terms of number of cycles, time or some other convenient unit. The equation has a sound theoretical basis
that is applicable to many materials.
3
ESDU
TM

q
S = P1N
q r
S = P1N + P2N + P3
r
S S = P2N
S = P3
N
Sketch 1.1 Components of SN curve plotted on a log-log scale
The first term on the right-hand side of Equation (1.1), when taken in isolation, represents a straight line
on a log-log scale, the slope of the line being q. Similarly, the second term on the right-hand side represents
a straight line with shallower slope r on a log-log scale. The third term represents a horizontal straight line.
Since different terms are dominant over different ranges of N, the curve tends to be asymptotic to each
straight line in turn, as illustrated in Sketch 1.1.
The model described by Equation (1.1) is consistent with the results of research on metallic materials which
have shown that the low-cycle fatigue region, which is characterised by predominantly plastic strain, can
be described on a log-log scale by a straight line. These results have also shown that the high-cycle fatigue
region, which is characterised by predominantly elastic strain, can be described by a straight line with a
fairly shallow slope, with a gradual transition taking place between the low-cycle and high-cycle regions.
Further information on the origins of Equation (1.1) and details of other SN relationships that are
commonly used elsewhere are given in Derivation 3. The third term on the right-hand side of Equation (1.1)
enables an endurance limit to be modelled, since the first two terms tend to zero for high values of N. The
value of the constant term fitted by the program should not, however, be taken as an estimate of the
endurance limit unless there is a significant number of data points in that region.
In practice, an appropriate model is often described by just one or two terms on the right-hand side,
especially when the available data may cover only a restricted range of endurances.
1.4.2 Parameter Estimation
The Fortran program performs a non-linear constrained optimisation procedure using a maximum
likelihood criterion. This means that the values chosen for the unknown parameters in Equation (1.1) are
those that maximise the probability of the observed data being obtained. The method of maximum
likelihood is covered in most text books on statistics. The simplified treatment given in Reference 7 and
the more rigorous treatment given in Reference 6 are recommended. A brief description of the use of the
method in the derivation of an SN curve from fatigue test data is given in Appendix A. The method is
particularly appropriate for application to a mixture of censored and uncensored data. In practice, it is
convenient to maximise the natural logarithm of the likelihood, lnl, rather than the likelihood, l, itself, and
4
ESDU
TM

the value of lnl corresponding to the fitted relationship provides a useful measure of how good a fit has
been obtained. (The same optimal solution is obtained whether it is l or lnl that is maximised.)
The regression relationship that is fitted is:
q r
log S = log ( P 1 N + P2 N + P3 ) + , (1.2)
where is a normally distributed error with a mean of zero and a constant standard deviation, independent
of N.
This approach was chosen, in preference to the derivation of a regression curve of N on S, because the
standard deviation of logN is not constant but increases with increasing N and tends to infinity as the
endurance limit is approached. This makes it impossible to fit a regression curve of N on S that extends to
high endurances where the slope of the curve levels out. Analysis of typical fatigue data from a variety of
materials has shown that the standard deviation of logS is, on the other hand, approximately constant; this
is equivalent to stating that S has a constant coefficient of variation. The regression relationship for S on
N can therefore be fitted over the whole applicable range of endurances.
The optimisation procedure is a globally-iterated version of a simplex algorithm constructed by Nelder and
Mead (Derivations 1 and 5). Constraints in the algorithm mean that negative values of the parameters
P 1, P 2 and P 3 , and positive values of the parameters q and r are not permitted. This ensures that, even
for very small samples where the true shape of the curve tends to be obscured by scatter, the resulting SN
curve will have an appropriate shape.
1.5 CURVE-FITTING PROCEDURE
1.5.1 Running Strategy
When using the program the limitations on its use that are set out in Section 2.2 should be given careful
consideration. It is envisaged that the program will probably be used in the following way. Initially a plot
of the data will be obtained that will enable a preliminary assessment of the data to be made. Depending
upon this and the amount of information about the population already available, the number of parameters
used to define the curve can be decided upon and, if desired, values assigned to some or all of them. A plot
showing the data and the derived curve can then be viewed together with the corresponding value of the
natural logarithm of the likelihood, lnl. A satisfactory curve having been derived, values of N for specific
values of S can be obtained.
1.5.2 User Input
The basic input to the program consists of pairs of S and N values, plus an indication of which values of N
are censored; that is, cases where the value of N is known to be either greater or less than the value input.
The program provides a facility for the user to display a scatter plot of the input data at this stage which
might assist in deciding what order of model is appropriate.
Equation (1.1), with all the constant parameters being given values other than zero, represents the highest
order (that is the most complex) model that can be fitted. The options available to the user are as follows.
1 parameter model S = P1 ,
q
2 parameter model S = P1 N ,
q
3 parameter model S = P1 N + P2 ,
5
ESDU
TM

q r
4 parameter model S = P1 N + P 2 N
and
q r
5 parameter model S = P1 N + P2 N + P3 .
No general rules on choosing a model can be given but, as well as examining a scatter plot of the data, it
might be useful, especially when analysing very small samples, to draw upon experience with similar but
larger data sets. For small data sets the absolute maximum number of unknown parameters that can be
evaluated by the program is equal to one less than the number of pairs of SN values corresponding to failed
specimens but, because of the high degree of scatter inherent in fatigue data, a lower-order model will
usually be found to be more satisfactory. When doubt exists as to whether or not there are sufficient data,
it is recommended that each data point be removed in turn and the analysis repeated; if the fitted model
varies greatly between the analyses then a lower order model should be tried, or more data collected. If,
however, the inclusion of a particular data point affects the model significantly, it is possible that this point
is an outlier that is not representative of the population (see Section 1.5.4).
The full 5 parameter model is rarely needed, and a satisfactory fit to the vast majority of fatigue data can
be obtained using 2, 3 or 4 parameters. The simplest model that gives an adequate fit is to be preferred,
since this is likely to give a better representation of the population SN curve than a higher order model
that gives a slightly better fit. In the absence of any prior knowledge, all five orders of model can be fitted
automatically and the most appropriate model selected; this facility is described in Section 1.5.3.
A facility is also provided for the user to specify the values of any of the non-zero parameters that can be
estimated from experience; in practice this is likely to apply only to the indices q and r. Again this might
be of value when dealing with very small data sets, for which the shape of the population SN curve might
be completely obscured by scatter. It might be useful to analyse old data sets using this Data Item in order
to obtain this information. This facility is not available in conjunction with automatic selection of the most
appropriate model (see preceding paragraph).
An additional option, which is only available when censored data are involved, is that of specifying the
value of v, the coefficient of variation of S, if it is known.
Equation (1.1) gives a convenient means of estimating S from N, but not vice versa. The program will,
however, if required, calculate values of N corresponding to values of S specified by the user; these values
of S form an additional input into the program if it is decided to make use of this facility.
1.5.3 Program Output
The program outputs values of the parameters that define the derived SN curve, a maximum likelihood
estimate of v, the population coefficient of variation of S, a list of any outliers identified, and values of N
for any values of S that were specified in the input. The program also outputs the value of the natural
logarithm of the likelihood corresponding to the fitted curve. If program W9724 is used, a plot of the data
and the fitted SN curve is output.
The curve obtained is that which gives the best fit to the data within the constraints imposed by the basic
model of Equation (1.1) and by the user. The purpose of these constraints is to ensure that all available
information, including that obtained from relevant experience, is taken into account to obtain the best
possible estimate of the population median SN curve.
The graphical output covers the same range of endurances as the input data. Extrapolation beyond this
range is not recommended and a facility to do this is not provided in the program. Similarly, if the program
is used to calculate values of N corresponding to given values of S, results will only be provided within the
range of the fitted curve.
6
ESDU
TM

It should be emphasised that the maximum likelihood estimate of the coefficient of variation of S given by
the program should not be taken to give an accurate indication of the population value. This is because
there will be some error in the estimated median curve and also the degree of scatter will vary from sample
to sample. The user should therefore exercise great caution when using this value for design purposes.
The value of the natural logarithm of the likelihood corresponding to the fitted curve provides a useful
means of comparing and choosing between different orders of model, using Akaikes criterion
(Derivations 2 and 4). Akaikes criterion states that, when comparing the natural logarithm of the likelihood
of a lower order model, lnlL, with that of a higher order model, lnlH, involving an extra unknown parameter,
the higher order model should only be preferred if lnlH is greater than or equal to lnlL + 1. This comparison
can be carried out by the user, or, if automatic selection has been requested, the most appropriate order of
model will have been selected according to Akaikes criterion. However, if knowledge of the expected
form of the curve is available, this knowledge may be used to override Akaikes criterion.
1.5.4 Outliers
Outliers are points that appear to be inconsistent with the majority of the data. They may, of course, be
genuinely extreme values, but, on the other hand, they may result from defective specimens or improper
test procedures, or they may simply be recording errors. Another possibility is that an inappropriate SN
model has been fitted. The program identifies all points corresponding to failures that lie at a distance of
more than two standard deviations of logS from the fitted curve. The purpose of this is to alert the user
who can then investigate the possible causes of the discrepancy and, if necessary, take appropriate action.
Outliers should only be removed from the data set if there is positive evidence that they do not represent
genuine population values.
The fact that no outliers have been identified does not necessarily mean that all of the data points are genuine,
particularly in small samples where a single erratic value could modify the fitted curve to such an extent
that it does not appear to be an outlier.
7
ESDU
TM

1.6 DERIVATION AND REFERENCES
1.6.1 Derivation
This section lists selected sources that have assisted in the preparation of this Data Item.
1. NELDER, J.A. A simplex method for function minimization. Computer Journal, Vol. 7,
MEAD, R. pp. 308-313, 1965.
2. AKAIKE, H. A new look at the statistical model identification. IEEE Transactions on

Automatic Control, Vol. AC-19, No. 6, pp. 716-723, 1974.
3. MANSON, S.S. Predictive analysis of metal fatigue in the high cyclic life range. Methods
for predicting material life in fatigue. Proceedings of the winter annual

meeting, American Society of Mechanical Engineers, pp.145-183,
December 1979.
4. PRIESTLY, M.B. Spectral analysis and time series. Academic Press, London, 1981.
5. PRESS, W.H. Numerical recipes in FORTRAN. Second Edition, Cambridge University

et al. Press, Cambridge, 1992.
1.6.2 References
The references given are recommended sources of information supplementary to that in this Data Item.
6. SILVEY, S.D. Statistical inference. Chapman and Hall, London, 1975.
7. CHATFIELD, C. Statistics for technology. Third edition, Chapman and Hall, London,
1984.
8
ESDU
TM

2. ESDUpac A9724 AND VIEWpac W9724
2.1 INTRODUCTION TO THE PROGRAMS
ESDUpac A9724 and VIEWpac W9724 are Fortran programs that accept fatigue SN test data. The aim
of the programs is to derive a median curve through a set of SN data. Both programs accept the same
input data in the same format and perform identical computations*. The differences between the programs
are basically cosmetic: Program A9724 is a file-input/file-output program which accepts input data in the
form of an ASCII text file and produces output in a similar form; Program W9724 makes use of Winteracter
(hence W9724) subroutines to provide a Windows-style interface, including context-sensitive help, and
graphical output in the form of a plot of the SN data and the curve derived by the program, together with
output in the form of an ASCII text file.
This Data Item relates to Version 1.0 of ESDUpac A9724V10 and VIEWpac W9724V10. The ESDUpac
and VIEWpac numbers contain the version number encoded in the three characters V10. Throughout this
Data Item the ESDUpac and VIEWpac are referred to by the numbers A9724 and W9724 and the current
version number is assumed to apply. The Contents page of the Series Software Volume gives the current
version number.
Program A9724 is provided in the form of Fortran source code file accompanied by the corresponding
executable file; Program W9724 is provided only as an executable file. Both programs can be found in
ESDUscope (the electronic version of the Series), on the internet version of the Series, and on disc in the
Series Software Volume. Example input and output files are also provided.
Section 2.2 gives the scope and limitations of the programs, Section 2.3 gives details of the inputs required
by the Fortran programs and Section 2.4 gives an explanation of the output, including all messages that
may be encountered when using the program. Section 2.5 provides a selection of examples that illustrate
the use of the programs.
Every reasonable effort has been made to ensure that the programs perform the intended calculations
satisfactorily. If, however, any problems are encountered in their use, please inform ESDU and every effort
will be made to overcome the difficulties. In common with other software producers, ESDU makes no
representation as to the suitability or fitness of the programs for any particular purpose and accepts no
liability for any loss occasioned by any persons as a direct or indirect result of use of the programs whether
arising from negligence or otherwise. In no event shall ESDU or any individuals associated with the
development of the programs be liable for any damages, including loss of profit or consequential loss,
arising out of or in connection with the programs.
2.2 SCOPE AND LIMITATIONS OF THE PROGRAMS
The purpose of the programs is to derive a median SN curve through a set of SN data, that is, the curve
that corresponds to a fifty per cent probability of failure. The programs analyse the SN data and derive
values for the number of parameters specified by the user. These parameters define a median curve through
the data points. The programs take account of data for unfailed specimens (run-outs) and also of
left-censored data using the maximum likelihood method, details of which are given in Section 1.4.2,
Appendix A and References 6 and 7. (Left-censored data are data for which the number of cycles to failure
has not been recorded but is known to be less than a certain value.) The curve is constrained by the program
to be concave upwards.
*
Owing to the differences in the manner in which data are input into the programs, small differences in the values output by the programs
for the same input data might be encountered; in all tests on the programs these differences have been negligible.
9
ESDU
TM

Using the data, the program calculates the values of up to five parameters that define a curve through the
data. The program outputs values of the parameters requested, an estimate of v, the coefficient of variation
of S, the value of lnl, the natural logarithm of the likelihood and a list of any data points that fall a significant
distance from the curve. The user can also request values of N for specified values of S. These values are
calculated using the equation of the derived SN curve. Graphical output showing the SN data with or
without the fitted curve is also available if program W9724 is used. The user can predefine in the input to
the program any of the parameters as well as, when censored data are being analysed, the population
coefficient of variation of S. Note that the order of fit less the number of predefined parameters should be
less than the number of failed data points.
The accuracy of the derived curve will depend upon a number of factors including the accuracy of the data
on which the curve is based, the number of data points available, in particular the number of failed
specimens, and the nature of the set of data, that is its scatter and its distribution over the range of stress,
strain or endurance under consideration. Section 1.3 gives notes on the design of a fatigue testing
programme with the aim of optimising the usefulness of the results obtained from the programme.
A further factor that will affect the accuracy of the curve is the amount of information concerning the form
of the curve that is already available. This information can be taken from other comparable data sets or
from other sources. Care should be taken when using such knowledge that the source from which it is taken
is genuinely comparable to the data set under consideration.
2.3 DATA INPUT REQUIREMENTS FOR THE FORTRAN PROGRAMS
This section gives notes on data input for the Fortran programs. Table 2.1 gives a formal list of the data
required and their order and format. This section and Table 2.1 need not be consulted if Program W9724
is used as that program includes a user-friendly interface which hides the formalities of the preparation
of the input data; comprehensive, context-sensitive, help is also included. Files of the format defined by
Table 2.1 can be loaded into Program W9724 and input data saved from that program will also be in the
format defined by Table 2.1. Section 2.5 provides examples of input files and Tables 2.1 and 2.2 give the
acceptable ranges for the input values.
The first line is reserved for the program identifier, which must be ESDU9724, see the end of Section
2.3.1.1 for details of why this is necessary. The next three lines are provided for run identification details.
These lines may be left empty or they can carry up to 72 alphanumeric characters each. The contents of
these lines appear in the output.
Immediately after these, two further lines are provided for the specification of units. These lines may be
left empty or they can be used to record the units of S and N. These lines, which can carry up to 12
alphanumeric characters each and are echoed to the output, are provided solely to give the user the
opportunity to record the units in which the data are entered and may be left blank if the user so wishes.
Whatever is entered in these two lines, the results will be in the same units as the input data and will be
correct provided that the input data are all in a coherent set of units.
Following the two lines for the units, the name of the output file, to which the results produced by the
program will be written, is required. If the output file is to be placed in a location different to that of the
program itself, the path of the file should be included. 72 alphanumeric characters are allotted.
Following this, the numerical data required by the program are entered, in the format described in Table 2.1.
10
ESDU
TM

2.3.1 Running the Programs
2.3.1.1 Program A9724
Program A9724 may be run in two ways. For both ways, it is first necessary to construct and save an input
file as described above and in Table 2.1. Then follow the instructions given in (a) or (b) below.
(a) Double-click on the executable file A9724v10.exe and type the name of the previously-prepared
input file as prompted, followed by a carriage return. (Note that if the input file is in a different
location to the program, the path of the input file should also be included.)
(b) Open a command (DOS) prompt and change directory so that the prompt is at the location of the
program executable file A9724v10. Then type
A9724v10 <[input file name]

followed by a carriage return, where [input file name] represents the name of the
previously-prepared input file, which should include the path of the input file if it is in a different
location to the program.
In both cases, the results produced by the program will be written to the file specified on the seventh line
of the input file.
It is the specification of the program identifier, ESDU9724, on the first line of the input file that allows
the program to be run in these two distinct ways and it is therefore important that it always be included.
2.3.1.2 Program W9724
To run Program W9724, simply double-click on the executable file W9724v10.exe. The program includes
comprehensive, context-sensitive, help which gives guidance both on using the program and on technical
aspects of the data required by it.
11
ESDU
TM

TABLE 2.1 PROGRAM INPUT DATA
Entry Entry Enter the numerical values for the quantities indicated in free format.
order The entries must be in the order shown below and must start on a
separate line with no blank line in between (except Entries 1 and 2).
The values given for each entry on a single line should be in the order
shown and must have a space or a comma between them.
1 ESDU9724 Enter the program identifier, ESDU9724.
2 Text Enter run identification data, input file title or notes. Three lines 72
characters wide are allotted. Lines may be left blank.
3 Unit of S Enter details of units used for the input data. Two lines 72 characters
Unit of N wide are allotted. Lines may be left blank.
4 Output file name Enter the name of the output file, to which the results produced by the
program will be written. If the output file is to be placed in a location

different to that of the program itself, the path of the file should be
included. One line 72 characters wide is allotted.
5 M Enter the number of points, M, in the SN data set.
6 N1 S1 1, 0 or 1 For each point enter the value of N followed, on the same line, by the
corresponding value of S. Finally, again on the same line, enter a
N2 S2 1, 0 or 1 number, either 1 to indicate that the specimen is left-censored (see
and so on up to Section 1.3.3), 0 to indicate that the specimen failed at a specific value
of N, or 1 to indicate that the specimen did not fail. The data pairs may
NM SM 1, 0 or 1 be in any order.
7 C Enter the number of parameters, C, by which the fitted SN curve is to
be defined (see Section 1.5.2). If automatic selection of the number of
parameters is required (see Section 1.5.3), enter 0.
8 1 P1 for C = 1 to 5 If the option for automatic selection was chosen in Entry 7, this entry
1 q for C = 2 to 5 must be omitted. Otherwise, for all values of C, make entries as
1 P2 for C = 3 to 5 follows.
1 r for C = 4 or 5 If the parameter is to be predefined, enter 1 followed, on the same
1 P3 for C = 5 line, by the value of the parameter.
If a parameter is not to be predefined, or is not included in the order
or, as appropriate, of fit specified in Entry 7, enter 0 followed, on the same line, by 0.
0 0
9 0 0 If v, the coefficient of variation of S, is not to be predefined, enter 0
or followed, on the same line, by 0. If v is to be predefined, enter 1
1 v followed, on the same line, by the value of v. The option to predefine v
is limited to data sets containing censored data.
10 D Enter D, the number of values of S for which values of N are required.
If no values of N are required, enter 0; in this case the input is complete
and Entry 11 may be left blank.
11 S1 Enter the values of S for which values of N are required. If no values of
S2 N are required, this entry may be left blank.
and so on up to The input is then complete.
SD
12
ESDU
TM

2.4 OUTPUT FROM THE PROGRAMS
2.4.1 Output of Input Data
For all runs the program outputs a standard header (see the examples in Section 2.5) followed by the three
lines of run identification details that were input.
All the data that were input are included in the output. After the three lines of run identification details,
the two input lines giving details of the units in which the data were entered are given. If no units were
input none will appear in the output.
Program W9724 includes graphical output in the form of a plot of the SN data, with or without the fitted
curve.
2.4.2 Results and Output

In successful computations the output provides the following for the set of data that was input:
(i) a maximum likelihood estimate of v, the coefficient of variation of S,
(ii) values of up to five parameters defining the derived SN curve,
(iii) a list of outliers identified,
(iv) a plot of the data and the derived curve (when Program W9724 is used),
(v) lnl, the value of the natural logarithm of the likelihood corresponding to the fitted model
and
(vi) estimated values of N for specified values of S if requested.
Values for the coefficient of variation and the parameters defining the curve may be predefined by the user
in which case it is these values that are output.
Successful runs are terminated with the statement Computation completed, solution obtained.
2.4.2.1 Failure to provide a solution owing to errors in the input data
When the programs are run, the input values in Table 2.2 are checked and if they are found to be out of
range an appropriate message is issued.
13
ESDU
TM

TABLE 2.2 INPUT DATA CHECKS AND ERROR MESSAGES
Check Message if check is violated
2 M 500 The number of data points must be between 2 and 500 inclusive.
1C5 The number of parameters defining the curve must be an integer between
1 and 5 inclusive. C minus the number of predefined parameters must be less
than the number of data points corresponding to failed specimens.
v 0.001 If the coefficient of variation, v, is to be predefined, it must be given a value
greater than or equal to 0.001.
P1 > 0 If any of the parameters P1, P2 or P3 are to be predefined, they must be given
P2 > 0 values greater than zero.
P3 > 0
q<r<0 If the parameters q and/or r are to be predefined, the value given to q must be
algebraically less than that given to r which must be less than zero.
0 D 20 The number of values of S for which the program will calculate values of N is
limited to a maximum of 20.
2.4.2.2 Failure to provide a solution from valid input data
In tests on the programs with fatigue test data no failures of this kind have been encountered. The following
failures may be encountered in extreme cases. In each case the input data are reproduced, followed by the
message, and the program is terminated.
Calculation failed. Attempted division by near zero in solution of equations. Check input data.
Calculation failed. Exponentiation attempted which would exceed machine limit. Check the input data.
Failed to converge. Check the input data and/or try a different number of parameters.
14
ESDU
TM

2.5 EXAMPLES
2.5.1 Example 1, Data for PH13-8Mo Forged Steel Bar taken from MIL-HDBK-5G
The specimens were tested under axial loading with a stress ratio of 0.1 and the data were recorded in terms
of maximum stress versus the number of cycles to failure. Tests were performed on specimens from a
number of separate casts. A scatter plot of the data, produced using Program W9724, is shown in Sketch 2.1
230
220
210
Left-censored
200
190
Failure
Run-out
180
170
160
150
140
4 5 6 7 8
10 10 10 10 10
Sketch 2.1 Scatter plot of data
The scatter of the data suggests that either a two or three parameter curve might provide a good fit to the data.
2.5.1.1 Two parameter curve
The input file specifying a two parameter curve is shown below. The text in square brackets should not
form part of the input file.
ESDU9724 [ Program identifier ]

Example 1 [ Three
June 2003 lines
Data from PH13-8Mo Forged Steel Bar of text ]
[ Units not
specified ]
Ex1_2par.out [ Output file name ]
51 [M]
14397 215 0 [ N1 S1 type of data ]
20670 215 0 [ N2 S2 type of data ]
25912 210 0
31757 210 0
31757 205 0
15
ESDU
TM

34762 205 0 [ and so on ]
52215 200 0
60478 200 0
50360 185 0
63993 195 0
78429 190 0
70048 175 0
82056 175 0
101480 200 0
100567 195 0
103800 185 0
115173 190 0
121593 190 0
128952 175 0
144381 185 0
9450616 172.5 0
239545 190 0
262211 185 0
325749 180 0
333196 175 0
271867 165 0
421488 155 0
446999 160 0
623623 180 0
656414 180 0
630245 170 0
586271 155 0
718525 170 0
768935 170 0
841692 180 0
2296051 165 0
3052549 180 0
4207742 180 0
4284517 175 0
5371077 150 0
7044542 160 0
8067676 175 0
7370291 150 0
8672804 150 0
172996 180 1
10677418 175 1
9887626 170 1
15054582 175 1
10000000 145 1
1986825 175 1
4689922 175 1 [ NM SM type of data ]
2 [C]
0 0 [
0 0
0 0 No predefined parameters
0 0
0 0 ]
0 0 [ Coefficient of variation of S is not predefined ]
0 [D]
When a two parameter curve is specified, the program gives the following output.
16
ESDU
TM

********************************************************************************
ESDU International plc
ESDUpac Number: A9724
ESDUpac Title: Derivation of endurance curves

from fatigue test data,
including run-outs
Data Item Number: 97024
Data Item Title: Derivation of endurance curves

including run-outs
ESDUpac Version: 1.0 Issued June 2003

(See Data Item for full input/output specification and interpretation)
********************************************************************************
Example 1, Two parameter curve
June 2003
Data from PH13-8Mo Forged Steel Bar
********************************************************************************
INPUT DATA
==========
Number of pairs of S-N values = 51
N values S values Type of data (see Table 2.1)

14397.000 215.00000 0
20670.000 215.00000 0
25912.000 210.00000 0
31757.000 210.00000 0
31757.000 205.00000 0
34762.000 205.00000 0
52215.000 200.00000 0
60478.000 200.00000 0
50360.000 185.00000 0
63993.000 195.00000 0
78429.000 190.00000 0
70048.000 175.00000 0
82056.000 175.00000 0
101480.00 200.00000 0
100567.00 195.00000 0
103800.00 185.00000 0
115173.00 190.00000 0
121593.00 190.00000 0
128952.00 175.00000 0
144381.00 185.00000 0
9450616.0 172.50000 0
239545.00 190.00000 0
262211.00 185.00000 0
325749.00 180.00000 0
333196.00 175.00000 0
271867.00 165.00000 0
421488.00 155.00000 0
446999.00 160.00000 0
623623.00 180.00000 0
17
ESDU
TM

656414.00 180.00000 0
630245.00 170.00000 0
586271.00 155.00000 0
718525.00 170.00000 0
768935.00 170.00000 0
841692.00 180.00000 0
2296051.0 165.00000 0
3052549.0 180.00000 0
4207742.0 180.00000 0
4284517.0 175.00000 0
5371077.0 150.00000 0
7044542.0 160.00000 0
8067676.0 175.00000 0
7370291.0 150.00000 0
8672804.0 150.00000 0
172996.00 180.00000 1
10677420. 175.00000 1
9887626.0 170.00000 1
15054580. 175.00000 1
10000000. 145.00000 1
1986825.0 175.00000 1
4689922.0 175.00000 1
Number of parameters defining curve = 2
No predefined parameters
Number of values of S for which

values of N are required = 0
********************************************************************************
RESULTS
=======
P1 = 270.1
q = -.3099E-01
Equation of fitted curve is:
S = 270.1 N^ -.3099E-01
Natural logarithm of likelihood

corresponding to fitted curve = 127.6
Estimate of coefficient of variation of S = .6789E-01
Number of outliers identified = 1
at N = .4215E+06 , S = 155.0 at 2.103 std devs from curve
No values of N were calculated for specified values of S
********************************************************************************
Computation completed, solution obtained
A plot of the data and the derived two parameter curve, produced using Program W9724, is shown in
Sketch 2.2.
18
ESDU
TM
230
220
210
Left-censored
200
190
Failure
Run-out
180
170
160
150
140
4 5 6 7 8
10 10 10 10 10
Sketch 2.2 Plot of two parameter curve and data
2.5.1.2 Three parameter curve
The input file specifying a three parameter curve is the same as that specifying a two parameter curve except
that the 2 on the eighth to last line is replaced by a 3.
When a three parameter curve is specified, the program gives the following output.
S = 14220N0.576 + 167.1
lnl = 134.2
v = 0.0588
No outliers were identified by the program.
A plot of the data and the derived three parameter curve, produced using Program W9724, is shown in
Sketch 2.3.
19
ESDU
TM
230
220
210
Left-censored
200
190
Failure
Run-out
180
170
160
150
140
4 5 6 7 8
10 10 10 10 10
Sketch 2.3 Plot of three parameter curve and data
Requesting a curve defined by four parameters gives almost the same curve as that defined by three
parameters and with the same value of lnl. Thus, using Akaikes criterion, the three parameter curve is
likely to be most representative of the population.
20
ESDU
TM

2.5.2 Example 2, Data from 2025-T6 Forged Aluminium Alloy Slab
The data are from Figure 7 of Data Item No. 87026*. The specimens were tested under rotating bending
loading and the data were recorded in terms of alternating stress versus the number of cycles to failure. A
scatter plot of the data, produced using Program W9724, is shown in Sketch 2.4.
300
280
260
240
Left-censored
220
Failure
Run-out
200
180
160
140
120
100
10 4 10 5 10 6 10 7 10 8 10 9
The scatter of the data suggests that a four parameter curve might provide a good fit to the data, but the
option to fit automatically all five orders of model and select the most appropriate according to Akaikes
criterion is chosen. It is also desired to obtain values of N for seven values of S from the selected curve for
use in cumulative damage calculations. The seven values of S are therefore input and the corresponding
values of N obtained.
2.5.2.1 Automatic selection of curve
The input file specifying automatic curve selection is shown below. The text in square brackets should not

Example 2 [ Three
June 2003 lines
Data for 2025-T6 Forged Al Alloy Slab of text ]
[ Units not
specified ]
Example2.out [ Output file name ]
17 [M]
48000 275 0 [ N1 S1 type of data ]
*
Data Item No. 87026, Fatigue strength of anodized aluminium alloy.
21
ESDU
TM

100000 240 0 [ N2 S2 type of data ]
140000 223 0
210000 207 0
230000 207 0
280000 172 0 [ and so on ]
10000000 151 0
25000000 137 0
40000000 137 0
62000000 132 0
70000000 117 1
100000000 125 1
100000000 110 1
100000000 110 1
100000000 110 1
100000000 103 1
100000000 103 1 [ NM SM type of data ]
0 [ 0 indicating automatic curve selection ]

7 [D]
140 [ S1 ]
160 [ S2 ]
180 [ S3 ] [ Values of S for which
200 [ S4 ] values of N are required ]
220 [ S5 ]
240 [ S6 ]
260 [ S7 ]
The program selected a three parameter curve as being most representative of the population. The program
output is given below.
********************************************************************************

including run-outs

including run-outs
********************************************************************************
Example 2
June 2003
Data for 2025-T6 Forged Al Alloy Slab
********************************************************************************
22
ESDU
TM

INPUT DATA
==========

48000.000 275.00000 0
100000.00 240.00000 0
140000.00 223.00000 0
210000.00 207.00000 0
230000.00 207.00000 0
280000.00 172.00000 0
10000000. 151.00000 0
25000000. 137.00000 0
40000000. 137.00000 0
62000000. 132.00000 0
70000000. 117.00000 1
.10000000E+09 125.00000 1
.10000000E+09 110.00000 1
.10000000E+09 110.00000 1
.10000000E+09 110.00000 1
.10000000E+09 103.00000 1
.10000000E+09 103.00000 1
Automatic selection of number of parameters defining curve
No predefined parameters

S values 140.00000 =
160.00000
180.00000
200.00000
220.00000
240.00000
260.00000
********************************************************************************
RESULTS
=======
P1 = .3376E+05
q = -.5043
P2 = 132.3
S = .3376E+05 N^ -.5043 + 132.3

Number of outliers identified = 1
at N = .2800E+06 , S = 172.0 at 2.356 std devs from curve
S values Calculated N values

140.0 .1660E+08
160.0 .1314E+07
23
ESDU
TM

180.0 .4474E+06
200.0 .2235E+06
220.0 .1338E+06
240.0 .8901E+05
260.0 .6350E+05
********************************************************************************
A plot of the data and the derived three parameter curve, produced using Program W9724, is shown in
Sketch 2.5.
300
280
260
240
Left-censored
220
Failure
Run-out
200
180
160
140
120
100
10 4 10 5 10 6 10 7 10 8 10 9
Sketch 2.5 Plot of three parameter curve and data
24
ESDU
TM

2.5.3 Example 3, Data for a Metal Matrix Composite
The specimens were tested at an elevated temperature and the data were recorded in terms of percentage
strain versus the number of cycles to failure. A scatter plot of the data, produced using Program W9724,
is shown in Sketch 2.6.
.55
.50
.45
Left-censored
.40
Failure
Run-out
.35
.30
.25
.20
.15
3 4 5 6 7
10 10 10 10 10
Since there are only three data points corresponding to failed specimens, only a two parameter curve can
be fitted (see Section 1.5.2 and Tables 2.1 and 2.2).
2.5.3.1 Two parameter curve
The input file specifying a two parameter curve is shown below. The text in square brackets should not

Example 3 [ Three
Data for a Metal Matrix Composite lines
Two parameter curve of text ]
[ Units not
specified ]
Ex3_2par.out [ Output file name ]
4 [M]
1100 0.5 0 [ N1 S1 type of data ]
12000 0.4 0 [ N2 S2 type of data ]
68000 0.25 0 [ N3 S3 type of data ]
1000000 0.15 1 [ NM SM type of data ]
2 [C]
0 0 [
0 0
25
ESDU
TM

0 0
0 0 ]
0 [D]
When a two parameter curve is specified, the program gives the following output.
S = 1.608N0.160
lnl = 8.40
v = 0.0821
A plot of the data and the derived two parameter curve, produced using Program W9724, is shown in
Sketch 2.7.
.55
.50
.45
Left-censored
.40
Failure
Run-out
.35
.30
.25
.20
.15
3 4 5 6 7
10 10 10 10 10
Sketch 2.7 Plot of two parameter curve and data
Owing to the small sample size and the high degree of scatter, there is doubt as to whether or not the fitted
line is a good representation of the population SN curve. It is possible that different samples from the
same population would give rise to quite different curves. In order to test this, one point at a time is removed
from the data set and the curve refitted.
26
ESDU
TM

2.5.3.2 First data point removed
The input file with the first data point removed is shown below. The text in square brackets should not

Example 3 [ Three
First data point removed of text ]
[ Units not
specified ]
Ex3_1rem.out [ Output file name ]
3 [M]
12000 0.4 0 [ N1 S1 type of data ]
68000 0.25 0 [ N2 S2 type of data ]
1000000 0.15 1 [ NM SM type of data ]
2 [C]
0 0 [
0 0
0 0
0 0 ]
0 [D]
When the first data point is removed and a two parameter curve is specified, the program gives the following
output.
S = 2.648N0.206
lnl = 5.99
v = 0.0547
A plot of the data and the two parameter curve derived with the first data point removed, produced using
Program W9724, is shown in Sketch 2.8.
It can be seen that this curve is somewhat different from that derived from the complete data set. It is clear
that the fitted curve is too sensitive to individual data points to give a reliable representation of the population
SN curve. However, data previously obtained from testing larger samples at different temperatures show
a characteristic slope of q = 0.15, and it is reasonable to assume that this will also be applicable to this set
of data.
27
ESDU
TM

.55
.50
.45
Left-censored
.40
Failure
Run-out
.35
.30
.25
.20
.15
3 4 5 6 7
10 10 10 10 10
Sketch 2.8 Plot of two parameter curve and data first data point removed
2.5.3.3 q predefined
The input file for a two parameter curve with q predefined is shown below. The text in square brackets
should not form part of the input file.

Example 3 [ Three
Two parameter curve with q predefined of text ]
[ Units not
specified ]
Ex3_qpre.out [ Output file name ]
4 [M]
1100 0.5 0 [ N1 S1 type of data ]
12000 0.4 0 [ N2 S2 type of data ]
68000 0.25 0 [ N3 S3 type of data ]
1000000 0.15 1 [ NM SM type of data ]
2 [C]
0 0
1 -0.15 [ q predefined as 0.15 ]
0 0
0 0
0 0
0 [D]
When q is predefined and a two parameter curve is specified, the program gives the following output.
28
ESDU
TM

********************************************************************************

including run-outs

including run-outs

********************************************************************************
Example 3
Data for a Metal Matrix Composite
Two parameter curve with q predefined
********************************************************************************
INPUT DATA
==========

1100.0000 .50000000 0
12000.000 .40000000 0
68000.000 .25000000 0
1000000.0 .15000000 1
Number of parameters defining curve = 2
PREDEFINED PARAMETERS:
q = -.1500

********************************************************************************
RESULTS
=======
P1 = 1.460
q = -.1500
S = 1.460 N^ -.1500

No outliers identified
29
ESDU
TM

No values of N were calculated for specified values of S
********************************************************************************
A plot of the data and the two parameter curve derived with q predefined, produced using Program W9724,
is shown in Sketch 2.9.
.55
.50
.45
Left-censored
.40
Failure
Run-out
.35
.30
.25
.20
.15
3 4 5 6 7
10 10 10 10 10
Sketch 2.9 Plot of two parameter curve and data with q predefined
Removing individual data points now has little effect on the fitted curve. Using knowledge of the
approximate value of the index q has thus enabled a viable estimate of the population SN curve to be made,
despite the very small number of data points.
30
ESDU
TM

APPENDIX A THE METHOD OF MAXIMUM LIKELIHOOD APPLIED TO THE ANALYSIS OF
CENSORED FATIGUE DATA
A1. INTRODUCTION
The method of maximum likelihood is a well-established technique for fitting the parameters of a statistical
model. In this case it has been used to determine the values of the constants P1, P2, P3, q and r in the
relationship given in Equation (1.2) of the Data Item:
q r
log S = log ( P 1 N + P 2 N + P3 ) + , (A1.1)
where is a normally distributed error with zero mean and constant standard deviation , and logarithms
may be taken to any convenient base. An estimate of is also obtained.
The maximum likelihood estimates of the unknown parameters are those that maximise the probability of
the observed data being obtained.
A1.1 NOTATION
The notation defined in this section is supplementary to that of Section 1.2.
f probability density function
F cumulative distribution function
Z standard normal variate
standard deviation of logS
A1.1.1 Suffixes
Suffix i indicates data representing a particular uncensored observation.
Suffix j indicates data representing a particular right-censored observation.
Suffix k indicates data representing a particular left-censored observation.
A1.1.2 Operators
I
ai = a1 a2 aI
i=1
I
ai = a1 + a2 + + aI
i=1
31
ESDU
TM

A2. THE METHOD
A2.1 Uncensored Observations
The model postulates that the errors are normally distributed with standard deviation . Therefore
q r
log S i log ( P 1 N i + P 2 N i + P 3 )
Z i = ----------------------------------------------------------------------------------- (A2.1)

is a standard normal variate, and the likelihood of obtaining an observed data pair ( S = S i , N = N i ) is:
2
1 Zi
f ( log S i ) = ------------------ exp --------- . (A2.2)
2 2
A2.2 Right-censored Observations (Run-outs)
Suppose a specimen tested at a given stress or strain S j has not failed after N j cycles. This can be interpreted
as, for a given N j , the specimen would have failed at some value of S greater than S j .
The likelihood associated with this observation is thus given by the complement of the cumulative
distribution function of log S :
2
1 Z
1 F ( log S j ) = ----------- exp --------- dZ . (A2.3)
Z j 2 2
A2.3 Left-censored Observations
Similarly, if a specimen tested at a level S k failed at some (unknown) number of cycles less than N k , the
likelihood associated with this observation is:
Zk 2
1 Z
F ( log S k ) = ----------- exp --------- dZ . (A2.4)
2
2
A2.4 Application of Maximum Likelihood Method
Let the data set consist of I uncensored observations, J right-censored observations (run-outs) and K
left-censored observations. The product of the individual terms above, when regarded as a function of the
unknown parameters, is referred to as the likelihood function of the data:
I J K
l ( P 1, P 2, P 3, q, r, ) = f ( log S i ) [ 1 F ( log S j ) ] F ( log S k ) . (A2.5)
i = 1 j = 1 k = 1
In practice it is the natural logarithm of the likelihood function that is maximised. Since a logarithm is a
monotonically increasing function, the parameter values that maximise the likelihood, l, also maximise the
32
ESDU
TM

logarithm of l. This is convenient here because using the logarithm of the likelihood avoids working with
products of numbers that are often extremely small. Thus:
I J K
ln l = ln [ f ( log S i ) ] + ln [ 1 F ( log S j ) ] + ln [ F ( log S k ) ]
i = 1 j = 1 k = 1
I J K
I 1

2 2
= --- ln ( 2 ) --- ( Zi ) + ln [ 1 F ( log S j ) ] + ln [ F ( log S k ) ]. (A2.6)
2 2
i = 1 j = 1 k = 1
The program determines the values of the model parameters and the value of that collectively maximise
lnl. This maximum likelihood estimate of is used to obtain a maximum likelihood estimate of v, the
coefficient of variation of S.
33
ESDU
TM

The Stress Analysis and Strength of Components Committee first met in 1964 and now has the following membership:
Chairman
Prof. C.E. Turner Imperial College of Science, Technology and Medicine
Vice-Chairman
Prof. T.G.F. Gray University of Strathclyde
Members
Mr R. Baker Lloyds Register of Shipping
Dr I.J. Bickley Mirrlees Blackstone (Stockport) Ltd
Dr M. Heller* Aeronautical and Maritime Research Laboratory, Australia
Dr L.C. Laming Imperial College of Science, Technology and Medicine
Mr R.B. Laycock Independent
Mr J.V. Vint Independent.

*
Corresponding Member
The construction and subsequent development of the Data Item was undertaken, under contract to ESDU, by
Mrs J. Moon and Dr J.R. Moon.
The person with overall responsibility for the work in this subject area is Mr A.C. Quilter, Head of Strength Analysis Group.
KEEPING UP TO DATE
Whenever Items are revised, subscribers to the service automatically receive the material required to update the appropriate
Volumes. If you are in any doubt as to whether or not your ESDU holding is up to date, please contact us.
Please address all technical engineering enquiries and suggestions to:
ESDU International plc Tel: 020 7490 5151 (from the UK)
+44 20 7490 5151 (from outside the UK)
Fax: 020 7490 2701 (from the UK)
+44 20 7490 2701 (from outside the UK)
E-Mail: esdu@esdu.com
Website: www.esdu.com
For users in the USA, please address all Customer Service and Support enquiries and suggestions to:
IHS Engineering Products Tel: 1 800 525 7052 (toll free number)
and Fax: 1 303 397 2599
Global Engineering Documents Website: www.ihs.com
www.global.ihs.com
97024
Derivation of endurance curves from fatigue test data, including run-outs
ESDU 97024
ISBN 1 86246 026-4, ISSN 0958 0379, ISSN 0141 4089
Available as part of the ESDU Series on Fatigue Endurance Data, on Stress

and Strength and on Engineering Structures. For information on all ESDU
validated engineering data contact ESDU International plc, 27 Corsham Street,
London N1 6UA.
ESDU 97024 introduces two Fortran programs, ESDUpac A9724 and VIEWpac
W9724, for the consistent derivation of an SN curve from a set of fatigue data.
The method on which the programs are based is applicable to data for both
metallic and non-metallic materials as well as to data from tests on coupons or
specimens representing structural features. The method also takes account of
censored data, for example, run-outs, in a rigorous manner. S may be in terms
of a maximum, an alternating or an equivalent stress, strain or loading with the
life, N, expressed in terms of number of cycles, time or some other convenient
unit (for example, repeated program blocks). The cycles may be of either
constant or variable amplitude. The derived curve is the median curve, that is,
it represents a fifty per cent probability of failure. The Fortran programs perform
a non-linear constrained optimisation procedure using a maximum likelihood
criterion to derive a curve of the form S = P1Nq + P2Nr + P3. The values chosen
for the unknown parameters in the derived curve are those that maximise the
probability of the observed data being obtained. A facility is also provided for
the user to specify the values of any of the non-zero parameters that can be
estimated from experience. An additional option, which is only available when
censored data are involved, is that of specifying the value of the coefficient of
variation of S, if it is known. The equation gives a convenient means of
estimating S from N, but not vice versa. The program will, however, if required,
calculate values of N corresponding to values of S specified by the user. Details
of the input data required by the programs are provided, along with example
input and output files. ESDUpac A9724, which gives only textual output, is
provided in the form of Fortran source code and corresponding executable file.
VIEWpac W9724, which includes a user-friendly interface and gives graphical
as well as textual output, is provided as an executable file along with
comprehensive, context-sensitive, help.
ESDU International plc, 2003
All rights are reserved. No part of any Data Item may be reprinted, reproduced, or transmitted
in any form or by any means, optical, electronic or mechanical including photocopying,
recording or by any information storage and retrieval system without permission from ESDU
International plc in writing. Save for such permission all copyright and other intellectual
property rights belong to ESDU International plc.
ESDU
TM
Engineering Sciences Data Unit

An IHS GROUP Company

Derivation of Endurance Curves From Fatigue Test Data, Including Run-Outs

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Derivation of Endurance Curves From Fatigue Test Data, Including Run-Outs

Încărcat de

Drepturi de autor:

Formate disponibile

ESDU 97024

Engineering Sciences Data Unit Issued June 2003

Derivation of endurance curves from

fatigue test data, including run-outs

Associated software: ESDUpac A9724

Engineering Sciences Data Unit 97024

THE PREPARATION OF THIS DATA ITEM

Engineering Sciences Data Unit 97024

1. INTRODUCTION, THEORY AND APPLICATION 1

1.3 FATIGUE TEST DESIGN 2

2. ESDUpac A9724 AND VIEWpac W9724 9

Engineering Sciences Data Unit 97024

APPENDIX A THE METHOD OF MAXIMUM LIKELIHOOD APPLIED TO THE ANALYSIS OF

A2. THE METHOD 32

Engineering Sciences Data Unit 97024

1. INTRODUCTION, THEORY AND APPLICATION

C number of parameters by which fitted SN curve is defined (see Table 2.1)

l likelihood (see Section 1.4.2 and Appendix A)

M number of pairs of SN values in a particular data set

Engineering Sciences Data Unit 97024

q, r constants (see Equation (1.1))

S stress, strain or loading

v population coefficient of variation of S

random error (see Equation (1.2))

Suffixes 1, 2, . . . D, applied to S, denote specific values of S for which a value of N is required to be

Suffixes 1, 2, . . . M, applied to S and N, denote specific pairs of SN values.

1.3 FATIGUE TEST DESIGN

1.3.1 Definition of Population

1.3.2 Sample Size and Stress/Strain Levels

Engineering Sciences Data Unit 97024

extremes of the range of endurances of interest.

1.3.3 Censored Data

1.4 THEORY AND APPLICATION

1.4.1 Form of SN Curve

where P 1, P 2, P 3, q and r are constants with P 1, P 2, P 3 0 and q < r 0 .

Engineering Sciences Data Unit 97024

1.4.2 Parameter Estimation

Engineering Sciences Data Unit 97024

The regression relationship that is fitted is:

1.5 CURVE-FITTING PROCEDURE

1.5.1 Running Strategy

1.5.2 User Input

Engineering Sciences Data Unit 97024

1.5.3 Program Output

Engineering Sciences Data Unit 97024

Engineering Sciences Data Unit 97024

2. AKAIKE, H. A new look at the statistical model identification. IEEE Transactions on

for predicting material life in fatigue. Proceedings of the winter annual

5. PRESS, W.H. Numerical recipes in FORTRAN. Second Edition, Cambridge University

6. SILVEY, S.D. Statistical inference. Chapman and Hall, London, 1975.

Engineering Sciences Data Unit 97024

2.1 INTRODUCTION TO THE PROGRAMS

2.2 SCOPE AND LIMITATIONS OF THE PROGRAMS

Engineering Sciences Data Unit 97024

2.3 DATA INPUT REQUIREMENTS FOR THE FORTRAN PROGRAMS

Engineering Sciences Data Unit 97024

2.3.1.1 Program A9724

A9724v10 <[input file name]

2.3.1.2 Program W9724

Engineering Sciences Data Unit 97024

program will be written. If the output file is to be placed in a location

Engineering Sciences Data Unit 97024

2.4.1 Output of Input Data