Professional Documents
Culture Documents
The
university of
Michigan
Liabraries
1 01 7
A ft T E S 8C1ENTIA VB RIT AS
A GUIDE FOR
FATIGUE TESTING AND THE
STATISTICAL ANALYSIS OF
FATIGUE DATA
Prepar'ed by
COMMITTEE £-9 ON FATIGUE
AMERICAN SOCIETY FOR TESTING AND MATERIALS
1 963
Published by the
AMERICAN SOCIETY FOR TESTING AND MATERIALS
1916 Race Street, Philadelphia 3, Pa.
©BY AMERICAN SOCIETY FOR TESTING AND MATERIALS 1963
Library of Congress Catalog Card Number: 63-16331
The First Edition of this Guide was the composite work of many
people who contributed a great deal of time to the discussion and writ-
ing of the text under the guidance of Task Group Leader, F. B. Stulen.
A major portion of the statistical section was written by Miss Mary N.
Torrey. George R. Gohn not only contributed to the discussion and plan-
ning, but also edited and arranged for the printing of the advance copies
of the text. The coordination of contributions and discussions was done by
H. N. Cummings. Appreciable contributions to the statistical parts of the
Guide were also made by D. H. Shaffer. In addition to the above, R. E.
Peterson, H. F. Dodge, D. P. Gaver, R. Hooke, W. T. Lankford, R. B.
Murphy, W. C. Schulte, P. R. Toolin, and M. B. Wilk contributed to the
discussions at various conferences.
The original Task Group was organized under the leadership of J. T.
Ransom, and a first rough draft was prepared in 1954 and revised in 1955.
Other contributors to these drafts were E. W. Ellis, W. T. Lankford, F. A.
McClintock, R. E. Peterson, E. H. Schuette, F. B. Stulen, and E. J. Ward.
In 1956, F. B. Stulen became Leader of the Task Group and the Guide was
completed under his direction.
Upon the formation of Subcommittee VI on the Statistical Aspects of
Fatigue, this subcommittee was asked to review the First Edition and to
make any revisions necessary to bring the Guide up to date. As a result
of this study, extensive revisions have been made in various sections as
printed in this Second Edition. They include: (1) revisions in the definitions
(Section II) and their separate publication as ASTM Tentative Definitions
E 206,1 (2) an expansion of Section IV on the number of test specimens,
(3) changes in Section V on tests of significance, and (4) the preparation
of a new section, Appendix IV, on the use of the Weibull distribution func-
tion for fatigue Me.
This work was carried out by four Task Groups headed by S. M. Marco,
H. E. Frankel, Miss M. N. Torrey, and C. A. Moyer, respectively. Others
who assisted in the preparation of the Second Edition were W. N. Findley,
R. A. Heller, J. H. K. Kao, H. N. Cummings, W. S. Hyler, B. Ruley, and
G. R. Gohn, Chairman of Subcommittee VI.
1
Definitions of Terms Relating to Fatigue g and the Statistical Analysis of Fatigue Data
(E 206), 1962 Supplement to Book of ASTM Standards, Part 3.
iii
NOTE.—The Society is not responsible, as a body, for the statements
and opinions advanced in this publication.
CONTENTS
PAGE
I. Purposes of Fatigue Testing 1
II. Definitions, Symbols, and Abbreviations 2
III. Test Procedures 8
IV. Minimum Number of Test Specimens and Their Selection 16
V. Analysis of Fatigue Data 22
Appendices
Miscellaneous Reference Tables 55
Additional Technique for Distribution Shape Not Assumed 68
Analysis of Correlation Between Two Variables 69
The Weibull Distribution Function for Fatigue Life 71
References 78
Index 81
LIS T O F T A B L E S
TABLE PAGE
1.—Allocation of Test Specimens for "Probit" Method of Test 11
2.—Minimum Number of Specimens Needed for Determining 95 Per Cent Con-
fidence Intervals of Stated Width for a Population Mean, p 19
3.—Minimum Number of Specimens Needed for Determining 95 Per Cent Con-
fidence Intervals of Stated Width for a Population Standard Deviation, a 19
4.—Minimum Number of Specimens Needed to Detect if the Standard "Deviation
of a Population Is a Stated Percentage of a Fixed Value 20
5.—Minimum Number of Specimens Needed in Each Sample to Detect if a Stand-
ard Deviation of One Population Is a Stated Multiple of the Standard Devia-
tion of Another Population 20
6.—Minimum Number of Specimens Needed to Detect a Stated Difference Between
a Mean and a Fixed Value 21
7.—Minimum Number of Specimens Needed to Detect a Stated Difference Between
the Means of Two Populations 21
8.—Median Percentage of Survivors for the Population 24
9.—Confidence Intervals for the Median 26
10.—Approximate Confidence Intervals for the Mean 27
11.—Confidence Intervals for Percentages 28
12.—Fatigue Test Data 29
13.—Fatigue Test Data 31
14.—Percentages Surviving 108 Cycles 32
15.—-"Probit" Test Data 33
16.—Computations for Fitting a Response Curve by Method of Least Squares 35
17.—Computation of Standard Deviation, s 36
18.—Method of Computing 95 Per Cent Confidence Limits for Per Cent Survival
Values 37
19.—Method of Computing 95 Per Cent Confidence Limits for Fatigue Strength
Values 38
20.—Computations for Significance Tests 46
21.—R. R. Moore Rotating Beam; Step Tests of 42 Specimens 50
22.—Analysis of Data in Table 21 51
23.—Prot Test Computations 52
v
vi CONTENTS
TABLE PAGE
24.—Prot Test Computations 53
25.—Minimum Per Cent of Population Exceeding Median of Low Ranking Points.. 56
26.—Unpaired Rank Test2 58
27.—Percentiles of the x Distribution 60
28.—Areas of the "Normal" Curve 61
29.—Values of t 62
30.—Percentiles of the x2 /d-f- Distribution 63
31.—Mo.25 and uo.yis for Runs Among Elements in Samples of Sizes Ni and Nz 64
32.—F Distribution 65
33.—k Factors for S-N Curves (Normal Distribution Assumed) 67
34.—Working Significance Levels for Quadrant Sum 69
35.—Ordinate Locations Corresponding to Per Cent Failed Values 72
36.—Mean-Rank Estimates of the Per Cent Population Failed Corresponding to
Failure Order in Sample 74
37.—Typical Fatigue Test Data 75
38.—Typical Fatigue Test Data, Without Runouts 75
39.—Typical Fatigue Test Data, with Runouts 77
LIS T O F F I G U R E S
FIGURE
1.—Probability-Stress-Cycle (P-S-N) Curve for Phosphor-Bronze Strip 10
2.—Response or Survival Tests 11
3.—Illustration of Staircase Method 12
4.—Representation of "Step" Testing of Single Specimen 14
5.—Graphical Illustration of Prot Data 15
6.—"Normal" or Gaussian Distribution Curve 22
7.—Response Curves for a Particular Type of Steel 34
8.—Per Cent of Specimens Having at Least the Indicated Fatigue Strength at 107
Cycles 51
9.—Prot Test: Stress as Linear Function of Stress Cycles 53
10.—Log-Log Plot of Prot Data 54
11.—Scatter Diagram 70
12.—Typical Weibull Distribution Curves 72
13.—Construction of Weibull Probability Paper from Log-Log Paper 73
14.—Estimation of Weibull Distribution Function Parameters for Data in Table 38. 75
15.—Per Cent Failed at Weibull Mean 76
16.—Estimation of Weibull Distribution Function Parameters for Data in Table 39.. 77
8. Range of Stress, Sr. —The algebraic difference between the maximum and
minimum stresses in one cycle, that is
9. Stress Amplitude (or Variable Component of Stress), Sa- —One half the
range of stress, that is
10. Stress Ratio, A or R. —The algebraic ratio of two specified stress values in a
stress cycle. Two commonly used stress ratios are:
The ratio of the stress amplitude to the mean stress, that is,
and the ratio of the minimum stress to the maximum stress, that is,
11. S-N Diagram. —A plot of stress against the number of cycles to failure. The
stress can be Smax, Smin, or Sa- The diagram indicates the S-N relationship for a
specified value of Sm, A, or R and a specified probability of survival. For N a log
scale is almost always used. For S a linear scale is used most often, but a log scale
is sometimes used.
12. Stress Cycles Endured, N. —The number of cycles of a specified character
(that produce fluctuating stress and strain) which a specimen has endured at any
time in its stress history.
13. Fatigue Strength at TV Cycles, SN. —A hypothetical value of stress for
failure at exactly N cycles as determined from an S-N diagram. The value of
SN thus determined is subject to the same conditions as those which apply to the
S-N diagram.
NOTE. —The value of SN which is commonly found in the literature is the hypothetical
value of Smai , -Smt^ror Sa , at which 50 per cent of the specimens of a given sample could
survive N stress cycles in which Sm = 0. This is also known as the median fatigue strength
at N cycles (see definition 47).
14. Fatigue Limit, S/. —The limiting value of the median fatigue strength as N
becomes very large.
NpTE.->-Certain materials and" environments preclude the attainment of a fatigue limit.
4 FATIGUE TESTING AND STATISTICAL ANALYSIS or DATA
Values tabulated as fatigue limitsun the literature are frequently (but not always) values
6f SN for 50~per cent survival at N cycles of stress in which Sm = 0.
15. Cycle Ratio, C—The ratio of the number of stress cycles, n, of a specified
character to the hypothetical fatigue life, N, obtained from the S-N diagram, for
stress cycles of the same character, that is,
51. S-N Curve for p Per Cent Survival.—A curve fitted to the fatigue life for p
per cent survival values at each of several stress levels. It is an estimate of the rela-
tionship between applied stress and the number of cycles-to-failure that p per cent
of the population would survive; p may be any number, such as 95, 90, etc.
NOTE.—Caution should be used in drawing conclusions from extrapolated portions of
the S-N curves. In general, the S-N curves should not be extrapolated beyond observed
life values.
52. Response Curve for N Cycles.—A curve fitted to observed values of per-
centage survival at N cycles for several stress levels, where N is a preassigned num-
ber such as 106, 107, etc. It is an estimate of the relationship between applied stress
and the percentage of the population that would survive N cycles.
NOTE 1.—Values of the median fatigue strength at N cycles and the fatigue strength
for p per cent survival at N cycles may be derived from the response curve for N cycles,
if p falls within the range of the per cent survival values actually observed.
NOTE 2.—Caution should be used in drawing conclusions from extrapolated portions
of the response curves. In general, the curves should not be extrapolated to other values
of p.
In "Probit" tests a group should consist of not less than five specimens
and the total tested at all stress levels should be at least 50. The distribu-
tion of the total number of available t specimens will depend upon the
purpose of the test. The relative group sizes for different stress levels are
shown in Table 1. This allocation is suggested so that the observed percent-
age survival values will have approximately equal weight, a condition neces-
sary for fitting the response curve by the usual method of least squares. This
allocation also facilitates the computation of confidence limits on the re-
sponse curves. As an alternative to the use of the relative group sizes (Ta-
ble 1), groups of other sizes can be used at each stress level, provided that
12 FATIGUE TESTING AND STATISTICAL ANALYSIS OF DATA
weighting factors7 are employed and the analysis conducted as indicated in
references (1-3) .
Figure 2 presents data that might be obtained in a "Probit" test of the
type described if the preassigned number of cycles were 107. Although not
required for the "Probit" analysis, the actual number 7of cycles-to-failure
should be recorded for each specimen that fails before 10 cycles, so that the
data may be available for other types of analysis, such as the plotting of
P-S-N Curves.
2. The Staircase Method:
The staircase (or "up-and-down") method of testing is.a variation of the
"Probit" method. It may require fewer specimens than the latter but is
likely to be useful only when the primary interest is in the mean fatigue
strength corresponding to a preassigned cycle life, N* The advantage gained
where E = one half the desired width of a 95 per cent confidence interval
(see ASTM Recommended Practice E 122)9.
It is more difficult to determine the minimum number of specimens needed
for a confidence interval of a given width for fatigue life corresponding to a
stated value of per cent survival other than 50 per cent. See footnote 15,
page 29, for equations for setting up tables similar to Table 9 (in Section
V A2) for other percentage points.
In general, the sample sizes would be larger than for medians.
3. Tests of Significance:
The minimum number of specimens needed depends upon the desired
magnitude of the difference that should be detected and the size of the risks
that can be tolerated.
When the rank test is used to test the differences of group medians, it is
difficult to relate the desired values and the criteria for the significance test
given in Table 26 for two groups and in Table 27 for more than two groups.
At least five specimens should be included in each group.
For differences of two or more s (other than 50 per cent) no
precise estimate of the minimum number of specimens needed is possible
unless prior estimates of the percentages are available. At least 15 speci-
mens should be included in each group.
4. Response Curves:
A discussion of the minimum number of specimens and their allocation
to stress levels is given in Section III Bl.
NUMBER or SPECIMENS AND THEIR SELECTION 19
B. LITE DISTRIBUTION SHAPE ASSUMED
1. Normal Distribution:
It is assumed here, as well as in Section V B that the fatigue data can be
transformed so that they will be approximately Normally distributed. A
Normal distribution is assumed in all cases. Each sample is assumed to be
drawn at random from its population.
2. S-N Curves:
Table 33 gives k factors for computing points on 75, 90, 95, 99, and 99.9
per cent survival curves for four values of confidence level, including 50 per
cent, and for n = 3 to 25. The minimum number of specimens should in-
crease as the per cent survival increases, but there is no definite criterion
for choosing a particular group size except for the relative magnitudes of
the k values. (Note that the rate of decrease is less as,« increases.) The num-
ber of specimens tested at each stress level can be smaller than the group
sizes needed when the life distribution is not assumed (Section V A).
TABLE 2.—MINIMUM NUMBER TABLE 3.—MINIMUM NUMBER
OF SPECIMENS 0 NEEDED FOR DE- OF SPECIMENS" NEEDED FOR DE-
TERMINING 95 PER CENT CONFI- TERMINING 95 PER CENT CONFI-
DENCE INTERVALS OF STATED DENCE INTERVALS OF STATED
WIDTH FOR A POPULATION WIDTH FOR A POPULATION STAND-
MEAN, /*. ARD DEVIATION, a.
Standard Deviation, a, Assumed Known. Some Estimate of <r Available.
Width of Confidence Number of
Specimens, n
Number of
Width of Interval
Interval Limits Specimens, »
Where:
._ n=
E= width of interval
2 0
Based on Fig. 1 of Greenwood and
x = sample mean. Sandomire (13).
"ASTM Designation E 122 (see foot- * The values of re given in
note 9). Table 30.
3. Confidence Intervals:
For the Mean. —If a good estimate of the population standard deviation,
0, is available, Table 2 gives the minimum number of specimens needed
for confidence intervals of stated width for the mean, p, of the population.
If the sample is used to estimate a as well as /*, the sample sizes should be
20 FATIGUE TESTING AND STATISTICAL ANALYSIS or DATA
larger, since ^0.975 values from Table 29 should be used instead of the 1.96
in the equation for n (Table 2).
For the Standard Demotion. —In order to find the minimum number of
specimens needed for determining confidence intervals of stated width for
the standard deviation, <r, of a population, some estimate of a must be avail-
able, since the width of the interval is measured in units of a. However,
Table 3 can be used as a guide even if no good estimate of a is available.
For example, if n = 8, the sample-standard deviation, used to estimate the
population standard deviation, may be above or below a by 0.5 <r, whereas an
estimate based on n = 30 will not be expected to deviate from the true value
by more than 0.25<r.
TABLE 4.—MINIMUM NUMBER TABLE 5.—MINIMUM NUMBER
OF SPECIMENS" NEEDED TO DE- OF SPECIMENS 0 NEEDED IN EACH
TECT IF THE STANDARD DEVIA- SAMPLE TO DETECT IF A STAND-
TION OF A POPULATION IS A ARD DEVIATION OF ONE POPULA-
STATED PERCENTAGE OF A FIXED TION IS A STATED MULTIPLE OF
VALUE. THE STANDARD DEVIATION OF
ANOTHER POPULATION.
Number of Specimens, n
Percentage of Fixed Number of Specimens, «
Value 80% 90%
Chance Chance
of of Multiple 80% 90%
Detection Detection
Chance Chance
of of
40 7 8 Detection Detection
45 8 9
50 9 12 1 .5. 39 52
55 12 14 2.0 15 20
60 15 19 2.5 9 13
65 20 26 3.0 7 9
70 28 38 3.5 6 8
75 42 55 4.0 5 7
0
a
Based on Fig. 2 from Ferris et al (14). Based on Fig. 3 from Ferris et al »(14).
4. Tests of Significance:
Difference Between Two Standard Deviations. —The sample sizes for testing
the difference between two means are given in Tables 6 and 7. In some cases,
the principal interest is in the difference between standard deviations.
1. One Standard Deviation a Fixed Value. —If one standard deviation is a
fixed value—for example, the long-time standard deviation of data based
upon an old procedure—and if the other standard deviation is to be com-
puted from data based upon a new procedure that may reduce the varia-
bility, Table 4 gives the minimum number of specimens needed to detect a
reduction of a stated amount. These e sizes apply when the observed
standard deviation, s, for the new procedure is indeed smaller than the fixed
value, and the ratio s2/(fixed value) 2 is compared with 1/F 0.95, corresponding
to °o and n — 1 degrees of freedom for numerator and denominator re-
spectively. (See Section V B4(a) and Table 32.)
2. Two Sample Standard Deviations. —If the problem is to test whether
the variability of procedure 1, say, is greater than the variability of pro-
cedure 2 (the numbers having been assigned prior to taking the data), Ta-
NUMBER or SPECIMENS AND THEIR SELECTION 21
ble 5 gives the minimum number of specimens needed in each sample to
detect that si is a stated multiple of s%. If the observed value of si is indeed
larger than the observed value of sz, compare s?/s<? with ^0.95 corresponding
to (HI — 1) degrees of freedom for numerator and denominator (since
n\ = HZ). (See Section V B4(a) and Table 32.) In this case it is not correct
to make the test if s22 is greater than Si2.
Difference Between Two Means:
1. One Mean a Fixed Valise. —If one mean is a fixed value—for example,
the long-time mean of data based on an old procedure or a commonly used
material—and the other mean is to be computed from data based upon a new
TABLE 6.—MINIMUM NUMBER OF TABLE 7.—MINIMUM NUMBER OF
SPECIMENS 0 NEEDED TO DETECT SPECIMENS 0 NEEDED TO DETECT
A STATED DIFFERENCE BETWEEN A STATED DIFFERENCE BETWEEN
A MEAN AND A FIXED VALUE. THE MEANS OF TWO POPULATIONS.
a = Unknown Standard Deviation of the <7 = Unknown Standard Deviation of
Population Being Estimated. Each Population; <n = az •
Number of Specimens, » Number of Specimens, n
Difference Difference
80% 90% 80% Chance 90% Chance
Chance Chance of Detection of Detection
of of
Detection Detection
0 50<r 64 86
0 50<r 34 44 0 75 29 39
0 75 16 21 1 00 17 23
1 00 10 13 1 25 12 15
1 25 7 9 1 50 9 11
1 50 6 7 1 75 7 9
1 75 5 6 2 00 6 7
2 00 4 5 2 50 4 5
2 50 3 4 0
0
Taken from Table E.I of reference
Taken from Table E of reference (11). (11) •
procedure that may shift the mean, Table 6 gives the minimum number of
specimens needed to detect a shift in either direction, measured in terms
of the population standard deviation of the new procedure. These sample
sizes apply when the computed value of
The constants in the formula are the population mean, /*, and <r, the popula-
ANALYSIS or FATIGUE DATA 23
tion standard deviation (a measure of the dispersion). 11 It should be empha-
sized that values of the parameters of the population can only be estimated
from tests on the specimens in the sample; to obtain exact values would re-
quire that the total population be tested.
While some fatigue tests, particularly those made in the finite life range
of an S-N curve, may yield approximately Normal distributions of cycle
life, generally a transformation to log cycle life is required. Others do not
yield Normal distributions, even after various transformations are performed
on the data. This is particularly true in the case of tests made at applied
stresses near the fatigue limit where runouts are observed. Hence, other
distributions, such as the Weibull distribution, 12 the "extreme value" dis-
tribution with and without lower limits, as used by Freudenthal and Gumbel
(IS), and other distributions, that are just as normal in the usual sense, as the
Normal or Gaussian distribution, have been applied to the analysis of fatigue
data. While references to some ,of these distributions are included in this
Guide, analysis of the fatigue data has been confined mostly to methods that
require no assumptions of distribution shape or to the methods based upon
the assumption that the raw data or the transformed data have a Normal
distribution.
As stated previously, however, any set of observations to which these
statistical methods are applied is assumed to come from a random sample
from the population of interest. If a series of samples is drawn, procedures
for testing for statistical control are given in the ASTM Manual on Quality
Control of Materials (see footnote 5). Lack of statistical control in data in-
dicates that the series of samples does not come from the same population.
A. LIFE DISTRIBUTION SHAPE NOT ASSUMED
1. S-N Curves:
These techniques should be used when the actual shape of the distribution
of fatigue life values for a given material is unknown or sketchy and the
number of specimens tested at each applied stress level is too small, say less
than 50, to estimate the shape of the distribution. In such cases, these tech-
niques give conservative results.
(a) One Group at Each Stress Level. —Usually the first step in the analysis
of fatigue data is to draw the S-N curve for 50 per cent survival; it is the
curve fitted to the medians of the groups at the several applied stress levels.
The median, an "order statistic," is the middlemost value when the observed
values are arranged in order of magnitude, or the average of the two middle-
most values if the group size is even.
Other S-N curves, those for p per t survival (where p is not 50), may
be fitted to other order statistics if the group size is greater than 1. If the
group values are arranged in order of magnitude, NI is the minimum cycle
life value, or the first order statistic, A7 2 is the second observed value, or the
second order statistic, and so forth.
The estimated percentage of survivors for the population at cycle life
11
In the Normal distribution, the median and the mean are equal.
12
See Appendix IV, p. 71 .
24 FATIGUE TESTING AND STATISTICAL ANALYSIS or DATA
values of Ni, or 7V 2 , depends upon the group size.13Table 8 gives the median
percentages at Ni and Nz for several group sizes. Some of these percent-
ages also are given in Table 25 for one group, m = 1, and the 50 per cent
confidence level.
The median percentage of survivors at the maximum value of the sample,
Nn , is 100 — (per cent for N\), etc. Examples of how Table 8 may be used
follow:
1. The 50 per cent survival curve may be estimated from the median of
any sample size.
2. If three specimens are tested at each applied stress level, the 79, 50, and
the 21 per cent survival curves may be estimated from the entries in Table
8 and their complements. The value 79 per cent is found opposite sample
size 3 in the second column, the value 50 per cent is taken from the median
TABLE 8.—MEDIAN PERCENTAGE OF
SURVIVORS FOR THE POPULATION.
cSample
™~i«, Size,
c; n
M - At^the Lowest At the Next^Lowest
Vajue ValuCj
1 . 50
2 70 30
3 79 50
4 84 61
5 87 69
6 89 73
7 90 77
8.". 91 80
:' 9. 92 82
10 93 84
" :• 11.-. 94 85
12 94 86
13 95 87
14 95 88
15.' 95 89
16, 96 90
S-N curve, and the value 21 per cent is obtained by subtracting the value
in the second column from 100 per cent.
3. If 7 specimens are tested at each applied stress level, the 90, 77, 23, and
10 per cent survival curves may be estimated from the entries in Table 8 and
their complements. The 50 per cent survival curve may be estimated from
the median.
At least 13 specimens must be tested at each applied stress level to esti-
mate the 95 per cent survival curve.
In practice, values of per cent l less than? 50 usually are not wanted.
Hence, if all of the specimens in a sample are tested simultaneously, the tests
may be stopped as soon as the specimen having the median value of fatigue
life for the sample has failed, unless the data are required for other purposes.
13
These are called "median percentages" because, half of the time, the true*percentage
will be larger, and for the other half of the time, smaller. They are close to, but usually
not equal to, the "expected" percentage of survivors, which is equal to 1 — i/(n + 1), where
i is the number of the order statistic and n is the sample size. The confidence level associ-
ated with expected percentages varies with the sample size, whereas it is constant for
median percentages.
ANALYSIS or FATIGUE DATA 25
As mentioned previously, the percentage survival values given in Table 8
are median values; they are based on a "confidence level" of 50 per cent. 14
Percentage survival values corresponding to higher confidence levels, such
as 95 or 99 per cent, are given in Table 25 for a single sample when m = 1.
For example, if three specimens are tested at an applied stress level, 79
per cent of the population are expected to survive N\ cycles (50 per cent
confidence level), but the statement that at least 37 per cent of the popula-
tion will survive N\ cycles may be made with greater confidence (confidence
level = 95 per cent). If estimates of the population percentage are made
from a series of samples tested at one applied stress level and the statement
is made that at least 79 per cent of the population will survive N\ cycles,
50 per cent of such statements are expected to be incorrect. If the statement
is made each time that at least 37 per cent will survive N\ cycles, only 5
per cent of such statements are expected to be incorrect. However, S-N
curves corresponding to a 50 per cent confidence level are usually shown.
The effect of fitting a curve to the same order statistics at several stress
levels probably increases the confidence level; how much is not known. If
S-N curves are based on other confidence levels, the fact should be plainly in-
dicated on the chart.
(Z>) Several Samples, or Groups, at Each Stress Level. —If it is not possible
to test all the specimens in a sample simultaneously and if stopping the tests
before all the specimens have failed is desirable to save time, the required
sample may be divided, at random, into two or more groups (see references
17 and 18). Then the median of the particular order statistics (the first,
second, and so forth) for the several groups may be used for constructing
the S-N curve. Table 25 gives values of percentage survival for several num-
bers of groups and several confidence levels.
EXAMPLE.—With five testing machines available, 15 specimens were tested at a
constant applied stress level in three groups of 5 each. For each group, all machines
were assumed to be stopped after the second failure. (Actually, all machines were
allowed to run until fracture occurred or until 10 million cycles of fatigue stressing
had been applied, so that the time saved could be estimated for this particular set
of tests.)
The test data are:
Group Life, kilocycles
1 162, 229, (261, 668, 2 281)
2 105, 131, (140, 245, 10 000+)
3 275, 373, (5 503, 8 695, 10 000+)
From these data we have:
Lowest Ranking Points Median
162, 105, 275 162
229, 131, 373 229
2Vi .
#2
Entering Table 25, under "Lowest Ranking Points," in the column for m = 3
groups, opposite n = 5 in each group, and at a confidence level of 50 per cent,
14 Technically speaking, the S-N curves based on order statistics are "nonparametric
tolerance limits," which are described by Murphy (16). The probability that at least p
per cent of the population lies above Ni cycles, where Ni is the ith order statistic of the
sample, is properly called a "tolerance level"; but the term confidence level appears to
have been used more frequently.
26 FATIGUE TESTING AND STATISTICAL ANALYSIS or DATA
read 87.05 per cent. This value is an estimate of the percentage of the population
from which the original 15 specimens were selected that will survive 162 kilocycles.
Similarly, at a confidence level of 95 per cent, 67.03 per cent or more of the popula-
tion are estimated to survive the 162 kilocycles. Again, for the "Second Ranking
Points," at a confidence level of 50 per cent, 68.61 per cent of the population are
estimated to survive 229 kilocycles and, at a 95 per cent confidence level, 45.40 per
cent or more of the population are estimated to survive 229 kilocycles.
Additional information can be obtained from the preceding test data by con-
sidering all 15 specimens as one "group" and determining the percentage of the
population expected to survive 105 kilocycles, which is the lowest ranking point
for m = 1 and n = 15 in Table 25. For a 95 per cent confidence level, straight-line
interpolation between 74.11 per cent for n = 10 and 86.09 per cent for n — 20
TABLE 9.—CONFIDENCE INTERVALS
FOR THE MEDIAN."
Confidence Level =£ 0.95.
Confidence Limits
Sample Size, n
Lower Upper
6 # #6
7 #saa #7
8 # #8
9 #2 N8
10 #2 #9
11 #2 N10
12 N2 #W
13 #3 #11
14 #3 #12
15 N4 N12
20 #
N6 N15
25 # #W
30 #1 »
35 #1 1 #24
40. #1 . #27
45 #1 S #30
50 N13 » #3
" Based on a table in Nair (19).
gives about 80 per cent. From this, it is estimated that at a 95 per cent confidence
level about 80 per cent of the population will survive 105 kilocycles.
2. Estimates of Parameters—Single Stress Level:
(a) Median Fatigue Life:
1. Point Estimate. —A point estimate of the population median is the
sample median, described above in n V Al(a).
2. Confidence Interval Estimate. —A confidence interval for the median
that does not assume a particular frequency distribution for the population
may be computed if the sample size is larger than five.
The n observed values offatigue life, N, are arranged in order of magnitude
as follows:
Ni£N*gNf- £ N.
ANALYSIS OF FATIGUE DATA 27
The confidence limits corresponding to a confidence level of at least 0.95 are
given by the order statistics designated in Table 9, p. 26.
EXAMPLE. —Assume that ten specimens are tested at a particular stress level and
the observed values of fatigue life in kilocycles are 201, 224, 226, 230, 232, 238, 24
244, 245, and 248. The point estimate of median fatigue life is the average of the
two middlemost values, namely 235 kilocycles. The interval estimate is defined by
A/2 and Ng (see Table 9), which are 224 and 245 kilocycles, respectively.
The population median may be above or below the sample median—235
kilocycles —but the chances are at least 95 in 100 that the statement, "the
median lies between 224 and 245 kilocycles," is correct if the sample came from
one population.
(b} Mean Fatigue Life:
1. Point Estimate. —A point estimate of the population mean is the
sample average.
3 ...... add the range of the observed values to the largest 3 X range
value and subtract it from the smallest value:
that is, Ni - (N3 - JVi) and N3 + (N3 - Ni).
4 ...... add (range) /4 to the largest value and subtract it lj£ X range
from the smallest value:
»=s n = 10 » = 20 » = 40
outlined in the Section on S-N curves: "One Group at Each Stress Level"
(Section V A). A particular value is e median, corresponding to 50 per
cent survival.
Another point estimate may be derived from the. cumulative frequency
distribution of the observed values. In general, the two point estimates would
not be exactly equal.
2. Confidence Interval Estimate. —Interval estimates for medians '(50 per
cent survival) are described in Section V A2(a). Interval estimates for fatigue
ANALYSIS or FATIGUE DATA 29
life values corresponding to other percentage points may be computed by
using reference (21) ,15
3. Tests of Significance:
(a) Differences of Group Medians—Single Stress Level.—-If two or more
groups of specimens are tested, the question of whether the observed differ-
ences in the values are due to chance or to some differences in the popula-
tions from which the groups were drawn often arises. The observed differ-
ences, for example, could arise because of differences in material lots or
differences in the characteristics of the testing machines.
The rank tests given in this section assume that the several groups are in-
dependently and randomly drawn from populations that are of the same
shape but may differ with respect to their medians. All the observed values
in one group are assumed to come from one population. Since the populations
are assumed to be of the same (though unknown) shape, only those groups
that are tested at the same stress level should be compared, since the form
of the distribution tends to change with change in stress level.
1. Rank Test for Two Groups. —In the rank test for two groups the rank
of each observation in the two groups combined is determined. The lowest
value is given the rank of 1, the next higher observed value is given the rank
of 2, and so forth. If one value appears several times, that is, there is a tie,
the average of the ranks for those numbers is assigned to each one. For ex-
ample, if the llth, 12th, 13th, and 14th values are all equal, they are each
given the rank of (11 + 12 + 13 + 14)/4 = 12.5. The ranks for the two
groups are totaled separately and the total for one of the groups (the one
with the smaller number of observations if the group sizes are unequal) is
compared with the critical values given in Table 26 for sample sizes equal to
the group sizes.
If the observed value falls within the range of values given in Table 26 for
the chosen significance level (5 or 1 per cent), the groups may be considered
to have come from one population. If the observed value falls outside the
range of values given in the table, the two groups are said to be significantly
different, that is, to have come from two populations with different medians.
The use of the 1 per cent significance level gives a smaller risk of calling the
15 The interval Nk to N may be computed as follows:
m
(1) k is chosen so that
where
« = 1 — (confidence level), and
p = (stated value of per cent survival)/100.
30 FATIGUE TESTING AND STATISTICAL ANALYSIS or DATA
groups significantly different when they are actually drawn from one popu-
lation and the observed difference is due to chance.
EXAMPLE.—To compare two machines, the rank test was applied to the data
from 27 specimens randomly assigned to two testing machines. (See Table 12.)
According to Table 26, the rank total for Machine A in Table 12, which has "the
smaller number of measurements," should be between 101 and 179 (Ni = 10,
NZ = 17) for the 5 per cent level of significance, and between 89 and 191 for the
1 per cent level of significance. This means that the actual total, 87, would not be
expected to occur as often as once in a hundred samples due to chance alone, if the
two machines were completely interchangeable. Thus, on the average, the machines
give significantly different fatigue life values.
TABLE 12.—FATIGUE TEST DATA.
Machine A Machine B
2. Rank Test for More than Two Groups. —The method of assigning ranks
is the same as for the two-group test, ranking the observations for all the
groups combined. The ranks are totaled separately for each group and the
following test-statistic, H, is computed from the rank totals (22):
where:
k =
number of groups,
Hi =
number of observations in the ith group,
N =
y^, ni, the number of observations in all groups combined, and
Ri =
sum of the ranks in the ith group.
The test-statistic H is distributed approximately as x2 with k — 1 de-
grees of freedom if each Ui is at least five. (For a discussion of x2, see refer-
ence (9)). Thus, the value of H calculated from the observed data may be
ANALYSIS or FATIGUE DATA 31
compared with the values of x2 given in Table 27 to determine whether
there may be a significant difference among the populations from which
the groups were drawn or not. If H is greater than the x2 value for k — 1
degrees of freedom and the chosen significance level, the populations are
said to be different; that is, the groups may be said to have been drawn
from two or more populations. Inspection of the rank totals will usually
show which groups are different from the others if the difference is signifi-
cant.
EXAMPLE.—To compare five machines, the rank test was applied to the data
from 25 specimens, randomly assigned to the five machines (see Table 13).
TABLE 13.—FATIGUE TEST DATA.
Machines A B c D E Total
(5)596 (6)599 (3)539 (2)530 (1)477
(10)640 (13)661 (12)651 (8)624 (4)568
(11)646 (21)760 (14)662 (9)638 (7)607
(18)733 (22)774 (15)675 (16)684 (17)719
(24)807 (23)781 (19)744 (25)889 (20)757
Sum of ranks, Ri. . . 68 85 63 60 49
R* 4624 7225 3969 3600 2401
Bf 924.8 793.8 720.0 480.2 4363.8
m 1445.0
H = 80.56 - 78 = 2.56
Entering Table 27 with 4 degrees of freedom, one less than the number of groups,
gives x2 = °.49, corresponding to a 5 per cent significance level or a percentile of
95. Since the computed value of H, 2.56, is very much smaller than 9.49, the ob-
served values of fatigue life may be considered to be from one population; the ma-
chines may be considered to be interchangeable.
(ft) Differences of Two or More Percentages (for example, per cent survival
values). —The test-statistic used to test the significance of the differences
among percentage values computed from observed data is x2- The formula
for x2 may be written in two ways; the second one is usually better for
computation purposes.
1. When the sample sizes are unequal:
where:
k = number of samples;
sum over k samples;
32 FATIGUE TESTING AND STATISTICAL ANALYSIS OF DATA
»,• = size of ith sample (i = 1, 2, • • • k);
Xi = observed number of "events" in the ith sample; an event may be a
failure, a survival, etc. ;
pi = observed fraction for the ith sample: pi = #*/«;; and
P = Z_, Xi/£_, ni = average fraction for all samples combined.
where n = sample size and x = ^ Xi/k (23, pp. 175-178). The other terms
were defined previously.
The computed value of x2 may be compared with the tabular values given
in Table 27 for k — 1 degrees of freedom (d.f.). If the computed value of
X2 is larger than the tabular value corresponding to: percentile =
100 — (chosen significance level), the percentages are said to be signifi-
cantly different; that is, the samples were drawn from different populations.
If the computed value of x2 is smaller than the tabular value, the samples
may be considered to have come from one population.
Another use of the x2 test is to test whether or not the observed per-
centage values are significantly different from an arbitrary value, such as
50 per cent. The method of computation is the same as that given previously,
except that: (1) the first way of writing the formula for x2 is used for the
computations, (2) the arbitrary value, which may be called p', replaces
p, and (3) d.f. = k.
EXAMPLE.—To compare six lots of phosphor-bronze 16strip, the x2 test was applied
to the data given in Table 14, using a significance level of 10 per cent.
TABLE 14.—PERCENTAGES SURVIVING 108 CYCLES.
Stress = ±25,000 psi.
Lot Sample Size, nt- Per Cent Number Sur- Xi*
Surviving, V)0pi viving, Xi »i
1 15 60.0 9 5.40
2 20 40.0 8 3.20
3 17 58.8 10 5.88
4 25 48.0 12 5.76
5 19 57.9 11 6.37
6 14 50.0 7 3.50
Total 110 = 57 = 3 0.1 1 =
16
Significance levels commonly used are 10 and 5 per cent.
ANALYSIS OF FATIGUE DATA 33
The tabular value of x2 corresponding to degrees of freedom = 5 and percentile
= 90 per cent is 9.24. Since the computed value, 2.28, is much smaller than 9.24,
no significant difference among samples is indicated by these data and the samples
may be considered to have come from one population. If the computed value of x2
were larger than 9.24, the lots from which the samples were drawn would be con-
sidered to be significantly different.
4. Response Curves:
Instead of drawing S-N curves when specimens are tested at several
levels of applied stress, response curves may be constructed from the data
if the applied stress levels have been properly chosen. A response curve is
constructed for a stated value of the number of cycles-to-failure, such as
107, 108, and so forth. It is based on the values of per cent survival at N
cycles that are observed for several (at least 3) values of applied stress.
This method of analyzing fatigue data is especially useful when some of
the specimens "run out," that is, survive the duration of the test.
While a response curve may be drawn on any type of graph paper, ex-
perience has shown that the per cent survival values tend to lie along a
straight line when the data are plotted on Normal probability paper. 17
(a) Fitting a Response Curve. — If the observed per cent survival values lie
along a straight line when plotted on Normal probability paper, a straight
line may be fitted to the points by eye or by the method of least squares.
The latter method is, of course, more precise and not subject to the biases
that may be introduced by a person fitting a line by eye.
The equations for the slope and intercept of a line fitted by the method
of least squares to per cent survival values having equal weight are: 18
where :
X = applied stress value (usually coded to reduce the size of the number) ,
p = per cent survival in each group of specimens tested,
Y = z, a value of the Normal deviate obtained from Table 28. The value
of 0 corresponding to p is obtained by entering the column headed
"Area" with 1 — (/>/100) and reading the corresponding value of z.
Interpolation between tabular values may be required. Y is called
the transformed value of p,
X — ^L, X/k = average of . X values,
Y = ^^ Y/k = average of Y values,
k = total number of groups tested, or total number of observed
p values, and
2^ = sum from 1 to k.
Intercept, a = Y
17
No. 3127, Codex Book Co., Inc., Norwood, Mass; or No. 358-23, Keuffel & Esser
Co.,18 New York, N. Y.
If the specimens have been allocated as indicated in Table 1, the per cent survival
values will have approximately equal weight. If this has not been done, references (1),
(2), or (3) should be consulted for a method of analyzing the data.
34 FATIGUE TESTING AND STATISTICAL ANALYSIS OP DATA
The equation for the fitted line isf Y= a + b (X — X} JThe subscript, /,
is used to denote the fitted line values.
TABLE 15.— "PROBIT" TEST DATA.
AppliedStress, Ksi Applied Stress, ksi Number^ Specimens Numberof Specimens Per Cent Survival, ,
40.0. 15 14 93. 33
41. 5. 8 6 75. 00
43.0. 5 3 60.00
44.5. 8 2 25. 00
46.0. 15 1 6.67
F/ Pf
Intercept, a = 5
-0.05 + 0.730T - X)
Yf =
pf = 100 (1 - "Area" value)
where F = the transformed value of the stated per cent survival value. This
derived estimate is called the fatigue strength for p per cent survival at N
cycles.
TABLE 17.—COMPUTATION OF STANDARD DEVIATION, s.
40.0 — 1 50 — 1 51 - 0 0001
41. 5 — 0 67 — 0 78 0 0121
43.0 -0. 25 — 0.05 0.0400
44.5 + 0.67 + 0 68 0.0001
46 . 0 + 1 . 50 + 1 41 0.0081
Sum 0.0604
served Y values about the fitted line is needed. Provided that all Y values
have equal weight (see Section III Al), the following equation may be used:
1. Confidence Limits for Per Cent Survival Values at N Cycles. —In com-
puting such confidence limits for a stated value of applied stress, the trans-
TABLE 18.—METHOD OF COMPUTING 95 PER CENT CONFIDENCE
LIMITS FOR PER CENT SURVIVAL VALUES.
Confidence Limits,
Applied _ Coded
Stress, ksi Value, Y Values p Values
X
i
Lower Upper Lower Upper
formed values of per cent survival are used. The formula for symmetrical
confidence limits is:
where:
X' = the stated value of applied stress (coded),
X = average of X values for the data used in fitting the line,
^(X — X) 2 is computed for the X values for the data, and
FZ , k -2 = a factor obtained from Table 32 for 2 and k — 2 degrees
of freedom for numerator and for denominator, respec-
tively. 21
The other symbols have been defined previously.
The confidence limits should be computed for all the applied stress levels
21
Hald (24) uses t = (Fi, k - 2)1/2 instead of (2F2, k - 2)1/2. The wider limits given here
ensure the desired confidence level when several confidence statements are made using
the same fitted line.
38 FATIGUE TESTING AND STATISTICAL ANALYSIS OF DATA
used in the experiment and may be computed for any other value within the
range of the applied stress values. The method of computation is given in the
following example.
EXAMPLE.—Table 18 shows the method of computing 95 per cent confidence
limits for per cent survival, at 107 cycles, using the value of 5 computed in Table
17 and the Yf values from Table 16, where:
n = 5, the number of observed p values,
^2,3 = 9.55, the value of F0.95 in Table 32 for 2 d.f. for numerator and 3
d.f. for denominator,
(2F2, 3 )1/2 = (19.10)1 /2 = 4.37.
s(2F2, s) 1/2 = 0.14 (4.37) = 0.61, and
p = 100 (1 — Area value); an Area value corresponds to a z = Y value in
Table 28.
The confidence limits computed in Table 18 are plotted in Fig. 7(a) and
connected by dotted lines to form a confidence band. The band applies to
TABLE 19.—METHOD OF COMPUTING 95 PER CENT CONFIDENCE
LIMITS FOR FATIGUE STRENGTH VALUES.
Confidence Limit! •CJ- Es(2F*c i)"t
Coded
p F X(Y)
Values Coded Values Fatigue
Strength, ksi°
Lower Upper Lower Upper
where:
F = the transformed value of the stated per cent survival value and
The confidence limits computed in Table 19 are plotted in Fig. 7(6) and
connected by dotted lines to form a confidence band. The band applies to
any value of per cent survival within the range of observed values, that is,
the extent of the fitted line. It will be noted that this confidence band differs
from the one in Fig. 7(a) only at the ends and that, numerically, the bands
are equivalent. Thus it is recommended that confidence bands be computed
according to the simpler method illustrated in Table 18 and the confidence
limits for fatigue strength be read from the intersections of the confidence
band and the stated value of per cent survival. For example, for 50 per cent
survival, the point estimate of the population value of fatigue strength is
43.1 ksi. The confidence interval estimate, corresponding to a confidence
level of 0.95, is 42.5 to 43.7 ksi. Likewise, for 80 per cent survival, the point
estimate is 41.4 ksi and the confidence interval estimate is 40.5 to 42.0 ksi.
The method illustrated in Table 19 should be used when a confidence
band is not presented or when the s are needed to more decimal
places than can be read from a chart.
B. LITE DISTRIBUTION SHAPE ASSUMED
1. Normal Distribution of Fatigue Life:
The preceding analyses, which made no specific assumptions about the
shape of the fatigue life distributions, are always applicable but, because of
their generality, may not give results that are precise enough for all pur-
+40 FATIGUE TESTING AND STATISTICAL ANALYSIS or DATA
poses. In order to sharpen the inferences which may be made from the data,
particular functional forms can be assumed for these distributions and the
statistical analyses carried out relative to these forms. It should be pointed
out, however, that the precision so gained is real only when the particular
distribution chosen honestly reflects the random process being studied.
(a) Choice of Distribution Shape. —In the literature are several specific
fatigue life distributions which have some basis in theory. However, em-
pirical evidence gathered to date has not allowed a clear-cut choice to be
made among them. In this position, it is sensible to choose a form for the
fatigue life distribution according to the following two criteria:
1. There should be satisfactory agreement between the data gathered and
the results predicted by the "theoretical" distribution.
2. The distribution form should be mathematically tractable or, at worst,
one on which extensive studies have already been made so that laborious
computations can be kept at a minimum.
These two conditions can be fulfilled by finding a transformation which,
when applied to a set of fatigue life data, results in new data which closely
follow the Normal or Gaussian distribution. Once this has been done, all
the results of Normal distribution theory are available to analyze the trans-
formed data.
The particular transformation which will do the job mentioned above is
not necessarily the same for all cases. It has been found that replacing the
observed fatigue life, N, by log N will often give a distribution which is
approximately Normal. Some situations may arise, however, where this
transformation is inappropriate. The joint efforts of the test engineer and
experimental statistician will be needed to make a specific choice which is
suitable for the problem being studied. If usable inferences are to be made,
the data transformation should not be so variable as to be a function of tjje
individual sets of data, but the problem of what transformation to use is
not a cut-and-dried statistical one.
In Section B, all random variables, unless otherwise noted, are assumed
to have been transformed so that they are Normally distributed.
(b) Estimates of Parameters. —On the assumption that the variable being
studied is Normally distributed, the information in the sample can be used
to obtain estimates of the parameters for the particular population. In
many situations it is sufficient to obtain point estimates of n and a. More
meaningful estimates, however, may be gotten by the use of confidence
intervals, since these utilize not only the point estimates, but also depend on
the amount of variation expected in these estimates from experiment to
experiment. A 95 per cent confidence interval on /*, for example, can be
described in the following way:
Once the data from a particular experiment have been collected, an inter-
val which is presumed to include the true parameter value, p, can be calcu-
lated from the sample values in a prespecified way. Since this interval is a
function of the observations, it is itself a random quantity, both as to posi-
tion and length, so that it will not be certain to include /x in any single
experiment. The process for constructing the interval from the sample can
be determined, however, in such a way that the chance of \L being covered
ANALYSIS OF FATIGUE DATA 41
by the interval in a given experiment is exactly 0.95. It is then said,, "With
a confidence level of 0.95, the true value, >, is included in the interval so
computed."
The mean, /*, of the Normal population, which is equal to the median of
the population, may be estimated either by the sample mean or by the sample
median. The sample median is the less efficient of the two, in the sense that
it experiences a larger variation from experiment to experiment but has the
advantage of being simpler to compute. The standard deviation, a, is
estimated by the sample standard deviation. The sample quantities—the
mean, x, and the standard deviation, s—are summarized below, where
Xi is the tth sample value from a sample of size n:
or
4.9175, 5.0147
Similarly, the desired confidence interval for <r has the limits:
or
0.0512, 0.1304
Note that the point estimates and confidence interval estimates are log-
arithms and may be converted to cycle life values for presentation.
Sample median = antilog X = 92 thousand cycles. (This value may be
44 FATIGUE TESTING AND STATISTICAL ANALYSIS or DATA
found from the original, imtransformed observations.) Confidence limits
for the population median are derived directly from the confidence limits
for JLI: antilog 4.9175 = 82,700 cycles and antilog2 5.0147 = 103,400 cycles.
Sample average, N = antilog (X + 1.1513<r ).23 Since o-2 is unknown,
an approximate value of the sample average may be found by replacing
er2 by s2. Then:
sample average, N = antilog (X ± 1.1513 s2) = 93,792 cycles
where = means "approximately equal to".
Sample standard deviation:
where N' is the population mean in terms of cycle life. Replacing N' by
N gives:
Sample standard deviation
4. Tests of Significance:
If cycle life data are collected at several stress levels for the purpose of
estimating the life distributions of these several levels, the techniques of
analysis just described are applicable. If, however, the principal interest is
a comparison of distributions in the sense of deciding whether their paranr-
eters are the same, then a different set of analyses is needed in order to use
the information in the data more effectively. A summary of these techniques,
classified according to the questions they are designed to investigate, is
given in this section. Except for the t on "Differences Among k Means,"
this section is concerned with a comparison between two sets of data.
(a) Difference Between Two Standard Deviations. —Before testing whether
or not the means of two samples are significantly different, it is important
to investigate whether the standard deviations are significantly different or
not. Let $i be the standard deviation for a sample of size n\ from the first
population and let s 2 be similarly defined for the second. Compute the ratio
23 See Hald (24), p. 161, and Soyero and Olds (25), p! 671.
ANALYSIS OF FATIGUE DATA 45
of the squares of the standard deviations, V, putting the larger standard
deviation in the numerator; if
then
where x\ and £ 2 are the sample means for the first and second samples,
respectively.
With a preassigned significance , a, compute 0 = 1 — (a/2) and
find tp , from Table 29, corresponding to d.f. = n\ , + n z — 2. If \t \ > t$,
conclude that the populations are different in mean or in variance or both.
(On the average, identical populations will be erroneously judged different
about 100 a per cent of the time.) If the samples are large enough so that
the test for the difference between two standard deviations would have
probably detected any important difference in the variances, a value
\ 1 1 > fy can be attributed to different population means. With the same
46 FATIGUE TESTING AND STATISTICAL ANALYSIS or DATA
reservation about sample size, if I / 1 < tp, one may conclude that the
population means are not significantly different.
2. Sample Standard Deviations Significantly Different. —Ifthe sample
standard deviations are significantly different, the hypothesis that the24
population means are equal may be tested by the following procedures.
Compute the numbers:
and
fatigue lives at this stress level is roughly log-normal; therefore, the computations
are in terms of log N rather than N. (See Table 20.)
We now have the mean, x, and the variance, s2, of each sample. To test them
for significant differences, use the F-ratio test and then the /-test, choosing a
significance level of 0.05.
From Table 29, t0. 976 (d.f. = 16) = 2.12. Since t = 10.8 is larger than t0.y, 5 =
2.12, the mean of the first sample is judged to be significantly larger than the mean
of the second sample. In the matter of fatigue life at the test stress level, the
second surface finish appears to be inferior to the first.
(c) Difference Among k Means. :
If k (k > 1) sets of data have been obtained, each of which is a random
sample from a Normal population, these populations being taken as having
a common standard deviation, then one can test the hypothesis that these
populations have a common mean. Let:% be the/th observation from the iih
group, Xi 2be the mean of the ith group, and Ui be the size of the tth group.
Define sw , average variance within samples, and sb*, variance among sam-
ples, by
and
Sz 2 N2 2N2
Si 1 Ni N!
So 0 No 0
Sum: A
where:
i = 0 is assigned to the lowest stress, S0, on which the less frequent
event occurs;
i = 1 is, assigned to the stress level for i0 + d, etc.;
Ni = number of the less frequent events at the corresponding levels;
iNj = product of i and Ni at each level; and
A = "£iNi.
3. Estimate the mean fatigue strength by
where:
N = total number of the less frequent events,
d = preselected stress increment, an
S0 = first stress level.
And use +1 if the less frequent event is a run out, or — ^, if the less frequent
event is a failure.
6. The Modified Staircase Method:
(a) For Small Sample Sizes. —For sample sizes less than 15, use the fol-
lowing simplified procedure:
ANALYS h
1. Start the staircase procedure as before, but use a step size, d, equal to
twice the preselected stress increment until the first pair of opposite results
is obtained. Subsequently, use a step size, d, equal to the preselected stress
increment.
2. In the analysis, disregard the specimens used up to the first pair
which give opposite results. Let n = total number of specimens tested,
starting with the first pair of opposite results. Let C = sum of the stresses
used on the last n — 1 tests plus the stress which would have been used
on the next test had it been run. Then,
C
m ——
n
or merely the average stress used on the last n — 1 tests plus that which
would have been used on the (n + l) th test.
EXAMPLE.—The following data were obtained in a modified staircase fatigue
test:
Stress, ksi » = 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
62 h h h
h h h h
61 h h h h h h h
60 h h h
59
58 h
Where x = specimen has failed and o = specimen has not failed (as in Fig. 3).
The n is taken as 1 with the first specimen of the pair which gave opposite
results. The (n + l) th, or 17th, test would have been run at 61.0 ksi. The last
n — 1 tests started at n = 2. Therefore, the total C is composed of the stresses
from n = 2 (62.0), to n = 17 (61.0).
C 980
m — — = -— = 61.25 ksi
n 16
(V) For Reducing Testing Time—For staircase runs involving more than
30 specimens, the time required to complete the staircase method is long
because of the sequential nature of the procedure. This long time can be
reduced by dividing one long staircase program into several shorter, inde-
pendent staircases and conducting these simultaneously.
1. Proceed by dividing the total r of specimens, T, into r groups
of n each, so that rn = T. Test each group as a separate small-sample stair-
case program, as described previously. Test as many groups simultaneously
as desired.
2. The best estimate of m for the whole set of T specimens is then:
s iMii
m = h
= arithmetic average of r values of m.
50 FATIGUE TESTING AND STATISTICAL ANALYSIS or DATA
EXAMPLE.—The step size is 5 ksi; and the data arose from testing five groups
of 15 specimens each.
Group m
No. 1 101
No. 2 102
No. 3 97
No. 4 99
No. 5 99
TABLE 21.—R. R. MOORE ROTATING BEAM; STEP0 TESTS OF
42 SPECIMENS.
SAE 4330 Steel: Tensile Strength, 140 ksi (Nominal).
Failure Stress, Kilocycles
Failure at
Stress Failure Stress, Kilocycles
Failure Stressat
ksi ksi
928 950
585 692
2 082 1 077
665 65 1 177
969 5 184
60 < 1 567 7 661
1 522
490 997
1 085 1 524
1 285 566
10 390 630
108
289 70 815
595 3 850
384 1 166
479 681
1 039 967
65 < 973 606
329
1 489 680
1 265 h 886
2 076 735
(. Continued, nex ! column) 1 479
0 Starting stress = 55 ksi. Steps = 5 ksi. Run = 107 cycles or to failure.
The estimate of the mean fatigue strength of the population is given by:
stress survived and the failure stress is the estimated fatigue strength for
each specimen. The data are then tabulated as indicated in Table 22. The
data in the last column can be plotted on arithmetic probability paper, as
shown in Fig. 8.
FIG. 8.—Per Cent of Specimens Having At Least the Indicated Fatigue Strength at
107 Cycles.
The stress levels of 100 per cent and 0 per cent survivals are indicated
on Fig. 8 by short arrows. It is interesting to note that the dash-line ex-
trapolations of the "reasonably straight line" suggest that, if a very much
larger number of specimens had been tested, some of them would probably
52 FATIGUE TESTING AND STATISTICAL ANALYSIS or DATA
have a fatigue strength higher than 72.5 ksi and some might not have
survived a stress of 55.0 ksi. Extrapolation on probability paper of the type
used for Fig. 8, however, depends critically on the assumption of a Normal
population; this assumption is almost certainly not safe. G. W. Snedecor
(26) says that extrapolation is mostly a guessing game.
If a straight line does not reasonably represent the plotted points, the
population distribution is probably not nearly Normal enough to warrant
analysis by logarithmic or arithmetic probability paper. In such cases, a
statistician should be consulted to determine whether some transformation
will Normalize the data. If the straight line is acceptably close to the plotted
points, the mean fatigue strength is approximately the value of the ab-
scissa at P = 0.50. Otherwise, that value is the median, that is, the value
above which half of the specimens survived, but it is not necessarily the
mean. If the straight line is acceptably close, the approximate fatigue
strength corresponding to any given mortality ratio (per cent failure) can
be estimated by reading directly from the line, but only between the first
and last plotted points. Extrapolation to very low or very high probabilities
is entirely unjustified and may be very misleading.
TABLE 23.—PROT TEST COMPUTATIONS.
Fracture Change in
Group Stress, S « (Mean),
psi per cycle Log a Line'i Log a
(Mean), psi (A Log a)
C 92 700 0.208 9, .3181 - 3
B 80 600 0.0592 1 ,7723 - 3 C -B 0.5458
A 71 800 0.00705 0 ,8482 - 3 B—A 0.9241
C 92 700 0.208 9, ,3181 - 3 C -A 1.4699
0 See Fig. 11.
8. The Prot Method:
The Prot method is based upon the assumption that the curve relating
the failure stresses and some power of the rate of increase of stress per
cycle is a straight line, provided that (1) the power is properly selected and
(2) this straight line will intersect the axis of failure stresses at the fatigue
limit. This may be expressed by the equation:
where:
,5 = fracture stress (final breaking stress), psi,
E = fatigue limit, psi,
K = a constant,
a = rate of increase of stress, psi per cycle, and
n — exponent of a that makes S a linear function of an.
A mathematical solution for the value of the fatigue strength and the
optimum value of n, the power of a that will place the plotted test values
on a straight line, and for the standard deviation of the fatigue strength is
ANALYSIS or FATIGUE DATA 53
proposed by Corten et al (5). The following "cut-and-try" method provides
an alternative to the mathematical solution.
The arithmetic mean of the individual values of a in each of the three
groups is determined, as well as the arithmetic mean of the individual values
of S—the fracture stress—in each group. (Strictly, the geometric means
should be used instead of arithmetic means, but the difference would usually
be unimportant.) This calculation gives three pairs of values, such as those
included for illustration in the second and third columns of Table 23. They
represent the entire set of test data fairly well. The test data may also be
plotted as shown in Fig. 9.
Stress Cycles
FIG. 9.—Prot Test: Stress as Linear Function of Stress Cycles.
TABLE 24.—PROT TEST COMPUTATIONS.
Assumed Log Change in Slope, n =
E, psi Group S-E (S-E) Line Log (S — E) A log (S - E)
[A log (S - E)] -5- A log a
The cut-and-try method depends upon the fact that Prot's equation can
be written thus:
log (S-E) = log K + n log a
This plots as a straight line through points C, B, and A, for the correct
value of E, with slope n. The procedure is to assume values of E, as shown
in Table 24. The slopes of lines C-B, B-A, and C-A are computed. Since
54 FATIGUE TESTING AND STATISTICAL ANALYSIS or DATA
C, B, and A, for the correct value of E in the logarithmic equation, all lie
on a straight line, the slopes of the three lines must come out the same
when the correct value of E has been determined. 25 Table 24 is a suggested
form for systematizing the trial computations.
A value for E close to but a little less than the value of the mean failure
stress for A must first be assumed. The quantities indicated in Table 24
are then computed, using values of S and A log a given in Table 23. If the
slopes are not within 1 or 2 per cent of one another (see Fig. 10), another
value of E must be assumed and the calculation repeated. The value of E
that makes the slopes sufficiently close to one another so that the three
points lie on the same straight line is the estimate of the fatigue limit as*
given by the Prot method.
25
Although theoretically only three groups of specimens need be used, in some cases
more groups with fewer specimens per group may be advisable. The "cut-and-try" pro-
cedure described for three groups can easily be adapted to four or more groups. Some ma-
terials have been found—for example, some case-hardened steels—for which the Prot
method gave no solution. Whether or not this means that they have no true fatigue limit
is not known.
STP91 A-EB/Feb. 1 964
APPENDIX I
MISCELLANEOUS REFERENCE TABLES
1 3 5 7 9 15 25 1 3 5 7 9 15 25
57
58 FATIGUE TESTING AND STATISTICAL ANALYSIS OF DATA
TABLE 26.—UNPAIRED
Critical lower and upper rank totals for the 5 per cent and 1 per cent levels of significance. Values
h
4 5 6 7 8 9 10 11 12 13
Ni
4 0.05 11 25
0.01
5 0.05 12 28 18 37
0.01 15 40
6 0.05 13 31 19 41 26 52
0.01 10 34 16 44 23 55
7. 0.05 13 35 20 45 28 56 37 68
0.01 11 37 17 48 24 60 33 72
8. 0.05 14 38 21 49 29 61 39 73 49 87
0.01 11 41 18 52 25 65 34 78 44 92
9. 0.05 15 41 22 53 31 65 41 78 51 93 63 108
0.01 12 44 19 56 27 69 36 83 46 98 57 114
10 0.05 16 44 24 56 33 69 43 83 54 98 66 114 79 131
0.01 12 48 19 61 28 74 37 89 47 105 59 121 71 189
11. 0.05 16 48 25 60 34 74 44 89 56 104 68 121 82 138 96 157
0.01 13 51 20 65 29 79 38 95 49 111 61 128 74 146 88 165
12. 0.05 17 51 26 64 36 78 46 94 58 110 71 127 85 145 100 164 116 184
0.01 13 55 21 69 30 84 40 100 51 117 63 135 76 154 90 174 106 194
13. 0.05 18 54 27 68 37 83 48 99 61 115 74 133 88 152 104 171 120 192 137 214
0.01 14 58 22 73 31 89 41 106 53 123 65 142 79 161 93 182 109 203 126 226
14. 0.05 19 57 28 72 39 87 50 104 63 121 77 139 91 159 107 179 124 200 142 222
0.01 14 62 23 77 32 94 43 111 55 129 67 149 81 169 96 190 112 212 129 235
15 0.05 30 75 41 91 52 109 65 127 79 146 95 165 111 186 128 208 146 231
0.01 23 82 33 99 44 117 56 136 70 155 84 176 99 198 116 220 133 244
16 0.05 42 96 54 114 68 132 82 152 98 172 114 194 132 216 150 240
0.01 35 103 46 122 58 142 72 162 87 183 102 206 119 229 137 253
17. 0.05 56 119 70 138 85 158 101 179 118 201 136 224 155 248
0.01 47 128 60 148 74 169 89 191 105 214 122 238 141 262
18 0.05 73 143 88 164 104 186 121 209 140 232 159 257
0.01 62 154 76 179 92 198 108 222 126 246 144 272
19 0.05 91 170 107 193 125 216 144 240 163 266
0.01 79 182 94 206 111 230 129 255 147 282
20 0.05 111 199 129 223 148 248 168 274
0.01 97 213 114 238 133 263 151 291
21. 0.05 132 231 152 256 172 283
0.01 117 246 136 272 155 300
22. 0.05 156 264 177 291
0.01 139 281 158 310
23. 0.05 181 300
0.01 162 319
24. 0.05
0.01
25 0.05
0.01
26 0.05
0.01
27 0.05
0.01
28. 0.05
0.01
29. 0.05
0.01
30. 0.05
0.01
160 246
147 259
165 255 185 280
151 269 171 294
170 264 190 290 212 316
156 278 175 305 196 332
175 273 196 299 218 326 241 354
159 289 179 316 200 344 223 372
179 283 201 309 223 337 247 365 271 395
163 299 184 326 205 355 228 384 251 415
184 292 206 319 229 347 253 376 278 407 303 438
167 309 188 337 210 366 233 396 257 427 282 459
189 301 211 329 234 358 259 387 284 418 310 450 338 482
171 319 193 347 215 377 238 408 263 439 288 472 315 505
194 310 216 339 240 368 265 398 290 430 317 462 345 495 374 529
175 329 197 358 220 388 244 419 268 452 294 485 321 519 349 554
199 319 222 348 246 378 271 409 297 441 324 474 352 508 381 543 412 579
179 339 202 368 225 399 249 431 274 464 300 498 328 532 356 568 385 605
203 329 227 358 251 389 277 420 303 453 331 486 360 521 389 556 420 592 451 630
184 348 206 379 230 410 254 443 280 476 306 511 334 546 363 582 392 620 423 658
208 338 232 368 257 399 283 431 310 464 338 498 367 533 397 569 428 606 460 644 493 683
188 358 211 389 235 421 260 455 286 489 313 524 341 559 370 596 400 634 431 673 463 713
237 378 263 409 289 442 316 476 345 510 374 546 405 582 436 620 469 658 502 698 536 739
215 400 239 433 265 466 291 501 319 536 347 573 377 611 407 649 438 689 471 729 505 771
268 420 295 453 323 487 352 522 382 558 412 596 444 634 477 673 511 713 546 754
244 444 270 478 297 513 325 549 354 586 383 625 414 664 446 704 479 745 513 787
301 464 329 499 359 534 389 571 420 609 452 648 486 687 520 728 555 770
275 490 303 525 331 562 360 600 390 639 422 678 454 719 487 761 522 803
336 510 366 546 396 584 428 622 461 661 494 702 529 743 565 785
308 538 337 575 367 613 397 653 429 693 462 734 495 777 530 820
373 558 404 596 436 635 469 675 503 716 538 758 575 800
343 588 373 627 404 667 436 708 469 750 504 792 539 836
411 609 444 648 477 689 512 730 547 773 584 816
380 640 411 681 444 722 477 765 512 808 547 852
60 FATIGUE TESTING AND STATISTICAL ANALYSIS OF DATA
TABLE 29.—VALUES 0 - 5 OF t.
When the table is read from the foot, the tabled values are to be prefixed with a
negative sign. Interpolation should be performed using the reciprocals of the degrees of
freedom.
0 From Dixon and Massey (9) by n of the publisher.
6 The values in this table were computed from percentiles of the distribution. See
F
Pearson and Hartley (2).
MISCELLANEOUS REFERENCE TABLES 63
Nl 2 3 4 5 6 7 8 9 10 11 12 13 14 IS 16 17 18 19 20 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
tf
2 4
3 5 6
4 5 7 8
5 2 2 5 7 g 9
6 2 2 3 3 S 7 g 9 10
7 2 2 3 3 3 s 7 9 10 11 1?
g 2 3 3 3 4 4 s 7 9 10 11 1? 13
9 2 3 3 4 4 5 5 5 7 9 11 12 13 13 14
10 2 3 3 4 s 5 5 6 5 7 9 11 12 13 14 15 15
11 2 3 4 4 5 5 6 6 7 5 7 9 11 12 H 14 15 16 16
12 2 2 3 4 4 s 6 6 7 7 7 S 7 9 11 12 1S 15 15 16 17 18
13 2 2 3 4 5 5 6 6 7 7 g g S 7 9 11 H 14 IS 16 17 18 18 19
14 2 2 3 4 s 5 6 7 7 g g 9 9 5 7 9 11 13 14 15 16 17 18 19 19 20
15 2 3 3 4 5 6 6 7 7 g g 9 9 10 S 7 9 11 13 14 15 17 17 18 19 20 21 21
16 2 3 4 4 5 6 6 7 8 8 9 9 10 10 11 S 7 9 11 13 15 16 17 18 19 20 20 21 22 22
17 2 3 4 4 5 6 7 7 g 9 9 10 10 11 11 ii s 7 9 11 13 IS 16 17 18 19 20 21 22 22 23 74
18 2 3 4 5 5 6 7 g g 9 9 10 10 11 11 12 12 s 7 9 11 13 15 16 17 18 19 20 21 22 23 24 24 25
19 2 3 4 s 6 6 7 g g 9 10 10 11 11 12 12 13 H 5 7 9 11 13 15 16 17 19 20 21 22 22 23 24 25 25 76
20 2 3 4 s 6 6 7 8 9 9 10 10 11 12 12 n n n 14 S 7 9 11 H IS 16 17 19 20 71 22 23 24 ?4 7S 26 76 77
The values listed are such that a number less than or equal to the Wo.o26
Nl = #2 «0 .025 MO .976 Ni = Nz Mo .025 «0. 975
value will occur not more than 2.5 per cent of the time and a number greater
20 14 31 50 than Mo.975 will occur not more than 2.5 per cent of the time.
21 15
27
28
40
42 33 52 For values of Ni and -/V2 larger than 20, a Normal approximation may be
22
23
16
16
29
31
44
46
35
37
54
56 used- The mean is
24 17 32 48 38 59
25 18 33 50 40 61
26 19 34 55 45 66
27 20 35 60 49 72
28
29
21
22
36
37
65
70
54
58
77
83
and the variance is
30 22 39 75 63 88
32 24 41 80 68 93
34 26 43 85 72 99
36 28 45 77 104
38.. 30 47
90
95 82 109 . For example, for Ni = N2 - 20, the mean is 21 and the variance is 9.74.
100 86 115
The 2.5 1and 97.5 percentiles are 21 + 1.96 (9.74)1 /2 = 27.1 and 21 -
0 With permission from Eisenhart and Swed (29). 1.96 (9.74) / = 14.9.
2
TABLE 32.— F DISTRIBUTION. 0
Degrees of Freedom for Numerator
1 2 3 4 5 6 7 8 9 10 12 IS 20 24 30 . 40 60 120 00
APPENDIX II
68
Copyright© 1 964 by ASTM International www.astm.org
STP91 A-EB/Feb. 1 964
APPENDIX in
69
Copyright© 1 964 by ASTM International www.astm.org
70 FATIGUE TESTING AND STATISTICAL ANALYSIS OF DATA
other side of the median. In this case treat the tied group as if the number of its
points before crossing the median were:
number favorable for inclusion in quadrant sum
1 + number unfavorable
APPENDIX IV
where:
N = specimen life,
No ^ 0 = minimum life parameter,
Na = characteristic life parameter occurring at the 63.2 per cent
failure point for the population [63.2 = 100(e - \/e), e = 2.718],
and
b > 0 = Weibull shape (or "slope") parameter.
This function is a simple exponential distribution function when 6 = 1; the
Rayleigh distribution function when b = 2; and a good approximation of the
Normal distribution function when b = 3.57, that is, when the mean and the
median values are equal.
The curve representing this function (Fig. 12) is usually skewed to the right,
going on to infinity, and, for b > 1, reaches zero frequency (touches the life axis)
to the left of the mode, which is the life value where the highest number of failures
occur.
The distribution is said to have a nonzero minimum life if the curve touches
the life axis at a value of life greater than 0. In other words, any specimen from
the population represented by such a distribution will have zero probability of
* This description of the Weibull distribution function, as an addition to ASTM
STP 91-A, was originally prepared by a Task Group in Subcommittee VI on Statistical
Aspects of Fatigue of ASTM Committee E-9 on Fatigue, composed of: C. A. Moyer,
chairman, Physical Laboratories, Timken Roller Bearing Co.; J. J. Bush, General Motors
Research Laboratories; and B. T. Ruley, New Departure Div., General Motors Corp. It
has been revised, prior to publication in its present form, by another Task Group in
Subcommittee VI, composed of: John K. H. Kao, chairman, New York University;
Robert A. Heller, Columbia University; B. T. Ruley, New Departures Div., General
Motors Corp.; J. M. Holt, Applied Research Laboratories, U. S. Steel Corp.; M. P.
Semenek, International Harvester Co.; and . R. Gohn and Miss M. N. Torrey, Bell
Telephone Laboratories, Inc.
There has been a demand from the roller bearing industry for the inclusion of an
additional section covering the use of the extreme-value distribution originally proposed
for the analysis of fatigue data by W. Weibull (31,32). Since Fisher and Tippett (33) are
often credited with first showing that this distribution was one of three limiting types of
the extreme-value distribution, it is sometimes referred to as "Fisher-Tippett Type III
for smallest values." As pointed out by Freudenthal and Gumbel (34), this distribution
has some theoretical basis, assuming that fatigue failures are examples of extreme values,
that is, they are smallest-strength or weakest-link values. It has also been used by others
in the analysis of life test data.
71
Copyright© 1 964 by ASTM International www.astm.org
72 FATIGUE TESTING AND STATISTICAL ANALYSIS OF DATA
which allows a simple graphical method for fitting the Weibull distribution to the
data and the subsequent graphical estimation of the parameters (b, N0, and Na)
in the formula.
95
given in Table 35. For example, in Fig. 13, the ordinate of the 90 per cent failure
value is 1.000 on the vertical logarithmic scale. Similarly, the ordinate for the 20
TABLE 36.— MEAN -RANI C ESTI MATE 3° OF THE P ER CE NT PO PULArDION i^AILEI) COR]RESPO NDINCJ TO F AILUEJE ORIDER IJf SAM!PLE.
Samp' e Size, n.
Order No., q 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
No. 1 50.00 33.33 25.00 20.00 16.67 14.29 12.50 11.11 10.00 9.09 8.33 7.69 7.14 6.67 6.25 5.88 5.56 5.26 5 00 4 76
No 2 66.67 50.00 40.00 33.33 28.57 25.00 22.22 20.00 18.18 16.67 15.38 14.29 13.33 12.50 11 76 11 11 10 53 10 00 9 52
No 3 75.00 60.00 50.00 42.86 37.50 33.33 30.00 27.27 25.00 23.08 21.43 20.00 18.75 17 65 16 67 15 79 15 00 14 29
No 4 80.00 66.67 57.14 50.00 44.44 40.00 36.36 33.33 30.77 28.57 26.67 25.00 23 53 22 22 21 05 20 00 19 05
No 5 83.33 71.43 62.50 55.56 50.00 45.45 41.67 38.46 35.71 33.33 31.25 29.41 27.78 26.32 25 00 23 81
No 6 85.72 75.00 66.67 60.00 54.55 50.00 46.15 42.86 40.00 37.50 35 29 33 33 31 58 30 00 28 57
No. 7 87.50 77.78 70.00 63.64 58.33 53.85 50.00 46.67 43.75 41.18 38.89 36.84 35 00 33 33
No 8 88.89 80.00 72.73 66.67 61.54 57.14 53.33 50.00 47.06 44.44 42.11 40 00 38 10
No 9 90.00 81.82 75.00 69 . 23 64.29 60.00 56.25 52.94 50 00 47.37 45 00 42 86
No 10 90.91 83.33 76.92 71.43 66.67 62.50 58 82 55 56 52 63 50 00 47 62
No 1 91.67 84.62 78.57 73.33 68.75 64.71 61.11 57.89 55 00 52 38
No 12 92.31 85.71 80.00 75.00 70.59 66.67 63.16 60 00 57 14
No 1 92.86 86.67 81.25 76.47 72.22 68.42 65.00 61 90
No 1 4 93.33 87.50 82.35 77.78 73.68 70 00 66 67
No 15 93.75 88.24 83.33 78.95 75 00 71 43
No 1 6 94.12 88.89 84.21 80.00 76 19
No 17 94.44 89.47 85.00 80 95
No 18 94.74 90.00 85.71
No 19 95.00 90.48
No 20 95.24
0 Mean-rank estimates = 100
WEIBULL DISTRIBUTION FUNCTION FOR FATIGUE LIFE 75
per cent failure line is 0.0969 on the logarithmic scale. On such paper, the tangent
of the angle 6 is an estimate of the Weibull "slope," b, for the population line.
The angle 0 may be measured with a protractor, or the slope of the line may be
computed.
Plotting Positions on Probability Paper:
The fatigue data for any one sample are first ordered from shortest to longest
life, each specimen being given an order number, q, from 1 through n. The hori-
zontal plotting position is its individual life value. All runouts are assumed to
have longer lives than the last ordered specimen that failed, but such data are
treated separately below under "Estimates of the Distribution Function Pa-
rameters."
The vertical plotting position of the per cent failed (Fig. 13) is the estimate of
the per cent of the population failed, F(N), based upon the specimen order num-
ber. Mean-rank estimates of the percentages of the population failed at successive
TABLE 37.—TYPICAL FATIGUE TABL,E 38.—TYPI CAL FATIGUE
TEST DATA. 1PEST DATA, WITHOUT
RUNG UTS.
Number of Revolutions to Failure Plot of N Versus 1?(N) Nonlinear
Order, q
Lot 1 Lot 2 Order, q Specimen Number of Revolu-
tions to Failure
No. 1 . . 1.1 X 106 2 0 X 106
No. 2 . . 2.3 3 7 No. 1. . . No. 4 4.0 X 105
No. 3 . . 4.0 5 0 No. 2 . . . No. 2 5.0
No. 4. . 6.5 8 0 No. 3 . . . No. 5 6.0
No. 5. . 8.6 11 5 No. 4 . . . No. 8 7.3
No. 6 13 0 No. 5. . . No. 1 8.0
No. 7 20 0 No. 6 . . . No. 7 9.0
No 8 23 5 No. 7. .. No. 6 10.6
No. 8. . . No. 3 13.0
order numbers are given in Table 36 for sample sizes ranging from 1 through 20.
Mean rank, q/(n + 1), is an unbiased estimate of F(N); such estimates are recom-
mended by Gumbel (36) and Weibull (37). Blom(38) suggests modified mean-rank
estimates. For the data given in Table 37 for the sample taken from lot 1, the
abscissa for the first specimen is plotted at its life value of N = 1.1 X 106 revo-
lutions and the ordinate at F(N) X 100 = 16.67, the plotting position for the first
of a sample of five based upon mean ranks given in Table 36.
Estimates of the Distribution Function Parameters:
1. An estimate of the population cumulative distribution that corresponds to
the data plotted in Fig. 13 can be fqund quickly by drawing a line by eye through
the failed points. More refined techniques for calculating this line can be found
by referring to Gumbel (36), Lieblein , or Kao (35). It is possible to calculate
this line by the method of least squares, as illustrated in Section V A4 of this
guide. For example:
and
76 FATIGUE TESTING AND STATISTICAL ANALYSIS or DATA
Comparisons using these methods as against the graphic method sho.w,,however,
that the latter is usually adequate for small samples.
2. An estimate of the characteristic life, Na , is obtained from Fig. 13 by reading
off the life value corresponding to the intersection of the fitted line and a hori-
zontal line corresponding to F(N) X 100 = 63.2 per cent.
3. An estimate of the median life is obtained by reading off the life value corre-
sponding to the intersection of the straight line of Fig. 13 and a horizontal line
corresponding to F(N) X 100 = 50 per cent.
FIG. 14.—Estimation of Weibull Distribution Function Parameters for Data in Table 38.
4. In Fig. 13, the minimum life, N0 , is assumed to equal zero, since the plot
of the fatigue data is approximately linear. The plotted data from Table 38
result in a line which curves downward (Fig. 14(a)); thus the existence of a finite
minimum life value greater than 0 would be suspected. To find an estimate of
minimum life, N0 : (1) note the life value which the curve approaches asymp-
totically, (2) obtain the quantity N — N0 for each point by subtracting the N0
value from each individual specimen life, and (3) plot this life difference on Weibull
paper versus the same per cent failed values as before. Thus, by trial and error,
the best estimate of N0 will be found so that the data shown in Fig. 14(o) will,
when transformed, plot as a straight line, as shown in Fig. 14(6).
5. The slope parameter, b, is equal to the tangent of the angle 6 shown in Fig.
13. Another estimate of b can be made by computing the tangent of 6 from the
logarithms of the ordinates and abscissas of two widely separate points, NI and
Nz, on the fitted line. Thus
estimate of b
6.The skewness of the Weibull distribution varies with the shape parameter,
b; and the Weibull mean, in general, may occur at various per-cent-failed values;
WEIBULL DISTRIBUTION FUNCTION FOR FATIGUE LITE 77
FIG. 16. —^Estimation of Weibull Distribution Function Parameters for Data in Table 39.
78 FATIGUE TESTING AND STATISTICAL ANALYSIS OF DATA
that is, the mean does not coincide with the median. Using the estimated Weibull
slope, b, it is possible to read from Fig. 15 an estimate of the per cent failed at the
Weibull mean and then refer back to the estimated population line on Weibull
probability paper, as in Fig. 13, to read off the estimated mean life from the1
curve. Gumbel (36) and Kao (35) give methods for calculating the Weibull mean
when the characteristic life Na and the slope b are known.
7. For data containing run-out specimens (Table 39), the n' broken specimens
(6 in the example, Fig. 16), out of a total of n specimens tested, are plotted on
probability paper at the mean-rank plotting positions, corresponding to a sample
size n (8 in the example, Fig. 16(a)). The line drawn through these points will
approach a horizontal asymptote, F/racture , which is equal to the ratio of the
first plotting positions corresponding to sample sizes n and n', respectively (Fig.
16(a)).
The parameters of this distribution may be obtained graphically by plotting
only the n' broken specimens at mean-rank plotting positions, corresponding to a
sample size n' versus N — N0 , where N0 is again the estimate of the vertical
asymptote approached by the curve. The slope of the resulting straight line (Fig-
16(6)), tan 6 = b, can be obtained as described in this Section. Na , at the proba-
bility level of 63.2 per cent, is taken directly from the plotted line. The estimated
equation of the probability function for the complete sample of size n will then
become
INDEX
Abbreviations, p. 7 Estimates of parameters, single stress level, pp.
Analysis of correlation between two variables, 18, 40, 76
Appendix III (p. 69) Estimates, mean rank—Weibull distribution
Allocation of test specimens—probit method, function, Table 36 (p. 74)
Table 1 (p. 11) Estimation, p. 4, Fig. 14 (p. 76)
Analysis of fatigue data, p. 22 Estimation, Weibull distribution function pa-
Areas of the Normal curve, Table 28 (p. 61) rameters, Fig. 14 (p. 76)
Arithmetic mean, p. 5 Fatigue, p. 2
Average, sample, p. 5 Fatigue data (see analysis of), p. 22
X2/d.f. distribution, percentiles of, Table 30 (p. Fatigue life, pp. 2, 6, 27, 39
63) Fatigue life for a stated value of per cent sur-
X22 distribution, percentiles of, Table 27 (p. 60) vival, p. 28
X ,P-3 2 Fatigue life for p per cent survival, p. 6
Choice of distribution shape, p. 40 Fatigue limit, p. 3
Computations for fitting a response curve by Fatigue limit for p per cent survival, p. 6
method of least squares, p. 34, Table 16 (p. 34) Fatigue notch factor, p. 4
Computation of significance tests, Table 20 (p. Fatigue notch sensitivity, p. 4
46) Fatigue strength, p. 6, Fig. 8 (p. 51)
Computation of standard deviation of values Fatigue strength for p per cent survival at N
about fitted line, Table 17 (p. 36) cycles, p. 6
Confidence coefficient, p. 5 Fatigue test data, Table 12 (p. 30), Table 13 (p.
Confidence interval, pp. 5, 26, 27, 28, 42, Table 31), Table 37 (p. 75), Table 38 (p. 75), Table
9 (p. 26), Table 10 (p. 27), Table 11 (p. 28) 39 (p. 77)
Confidence level, p. 5 Fatigue tests, p. 1
Confidence limits (see confidence interval) F-distribution, Table 32 (p. 65)
Constant amplitude tests, pp. 1, 9-13 F-ratio test, pp. 45, 47
Constant life fatigue diagram, p. 14 Frequency distribution, p. 4
Construction of Weibull probability paper from Gaussian distribution curve, Fig. 6 (p. 22)
log-log paper, Fig. 13 (p. 73) Group, p. 4
Correlation between two variables, Appendix III Increasing amplitude tests, pp. 1, 13, Fig. 4 (p.
(p. 69) 14), Fig. 5 (p. 15)
Cycle ratio, p. 4 Interval, pp. 5, 9
Definitions, p. 2 Interval estimate, p. 5
Definitions relating to fatigue tests and test ^-factors for S-N curves, Table 33 (p. 67)
methods, p. 2 Least squares, method for fitting a response
Definitions relating to statistical analysis, p. 4 curve, Table 16 (p. 34)
Definitions relating to statistical analysis of Level, confidence, p. 5
fatigue data, p. 6 Level, significance, p. 6
Difference among k means, p. 47 Level, tolerance, p. 5
Difference between two means, pp. 21, 45 Limits, confidence, p. 5
Difference between two standard deviations, Limits, fatigue strength at N cycles, p. 38
pp. 20, 44 Limits, method of computing, Table 19 (p. 38)
Distribution, p. 4, Table 27 (p. 60), Table 30 Limits, tolerance, p. 5
(p. 63), Table 32 (p. 65), Fig. 6 (p. 22) Maximum stress, p. 3
Distribution curves, Fig. 12 (p. 72), Fig. 6 (p. 22) Mean, confidence interval for, p. 19
Distribution shape, choice of, pp. 22, 40 Mean, definition, p. 5
Estimate, p. 5 Mean, confidence limits for, Table 10 (p. 27)
, interval (see confidence interval) Mean fatigue life, p. 27
, point, p. 5 Mean, sample, p. 21
81
Copyright© 1 964 by ASTM International www.astm.org
82 FATIGUE TESTING AND STATISTICAL ANALYSIS OF DATA
Mean rank estimates: per cent of population "Quadrant sum" correlation test, Appendix III
failed corresponding to failure order in sample, (p- 69)
Table 36 (p. 74) "Quadrant sum," working significance level, Ta-
Means, confidence interval for, p. 42, Table 2 ble 34 (p. 69)
(p-1 9) Range of stress, p. 3
Means, differences between two, p. 45 Rank test, pp. 25, 30, 31, Table 25 (p. 56), Table
Means, differences among k, p. 47 26 (p. 58)
Mean stress, p. 3 References, p. 79
Mean, Weibull, Table 36 (p. 74), Fig. 15 (p. 77) Response curves, pp. 7, 18, 33, 34, 36, 38, Fig. 7
Median, confidence limits for, Table 9 (p. 26) (p. 35), Table 16 (p. 34)
Median fatigue life, pp. 6, 26 Response or survival tests, pp. 1, 10, 12, 13,
Median fatigue strength at N cycles, p. 6 Fig. 2 (p. 11)
Median percentage of survivors for the popula- Rotating beam tests, R. R. Moore, Table 21
tion, Table 8 (p. 24) (p. 50)
Median, sample, p. 43 Run test, Appendix II (p. 68)
Medians, differences of group, p. 29 Runs along elements in samples of sizes NI and
Method of least squares, pp. 34, 35, Table 16 N2, Table 31 (p. 64)
(p. 34) Sample, p. 4
Minimum per cent of population exceeding me- Sample average, p. 5
dian of low ranking points, Table 25 (p. 56) Sample means, p. 21
Minimum stress, p. 3 Sample median, p. 5
Modified staircase test method, pp. 1, 13, 48, 49 Sample standard deviation, p. 5
Moore rotating beam step test, Table 21 (p. 50), Sample percentage, p. 5
Table 22 (p. 51), Fig. 8 (p. 51) Sample variance, p. 5
Mo.025 and MO.QTS for runs among elements in sam- Scatter diagram, Fig. 11 (p. 70)
ples of sizes NI and N%, Table 31 (p. 64) Selection of test specimens, pp. 16, 17, 19
Nominal stress, p. 2 Significance level, p. 6, Table 34 (p. 69)
Normal curve, areas of, Table 28 (p. 61) Significance level, for "quadrant sum," Table 34
Normal distribution curve, Fig. 6 (p. 22) (p. 69)
Normal distribution of fatigue hie, p. 39 Significant, p. 5
Normal distribution, ^-factors for S-N curves, S-N curve for 50 per cent survival, p. 6
Table 33 (p. 67) S-N curve for p per cent survival, p. 7
Number of test specimens, minimum, pp. 16-21, S-N curves, pp. 6, 7, 17, 19, 23, 25, 40, 41, Table
Table 2 (p. 19), Table 3 (p. 19), Table 4 (p. 8 (p. 24), Table 25 (p. 56), Table 33 (p. 67)
20), Table 5 (p. 20), Table 6 (p. 21), Table 7 S-N diagrams, p. 3
(p. 21) Staircase test method, pp. 1, 48, Fig. 3 (p. 12)
Parameter, pp. 4, 18, 26, 40, Fig. 14 (p. 76), Fig. Standard deviation, pp. 19, 44-46, Table 2 (p.
16 (p. 77) 19)
Per cent of specimens having at8 least the indi- Standard deviation, confidence interval for, p. 42
cated fatigue strength at 10 cycles, Fig. 8 Standard tests, pp. 1, 9, 23, 25
(p. 51) Statistic, p. 4
Per cent survival for a stated value of fatigue Steady component of stress, p. 3
life, p. 27 Step test method, pp. 1, 13, 50, Fig. 4 (p. 14)
Per cent survival values at N cycles, confidence Step tests, R. R. Moore rotating beam specimens,
limits for, p. 37 Table 21 (p. 50), Table 22 (p. 51)
Percentiles of the x2 distribution, Table 27 (p. Stress, p. 3
60) Stress amplitude, p. 3
Percentiles of the x2/d.f. distribution, Table 30 Stress concentration factor, p. 4
(p. 63) Stress cycle, p. 2
Point estim_ate, pp. 5, 26-29 Stress cycles endured, p. 3
Population, p. 4 Stress ratio, p. 3
Probability-stress-cycle curve, Fig. 1 (p. 10) Survival tests, Fig. 2 (p. 11)
Probability paper, pp. 33, 73 Symbols, p. 7
Probit test—allocation of test specimens, Table Test of significance, pp. 6, 18, 20, 29, 30, 31, 45,
1 (p- 11) 46, 47, Table 20 (p. 46)
Probit test data, Table 15 (p. 34) Test procedures, pp. 8, 9, 10, 12, 13, 15, Fig. 3
Probit test method, p. 10 (p. 12), Fig. 4 (p. 14T, Fig. 5 (p. 15)
t test, pp. 1, 15, 52, Table 23 (p. 52), Fig. 9 Test specimens, minimum number, p. 16
(p. 53) Test specimens, selection of, p. 16
INDEX 83
Test-statistic, p. 6 Universe, p. 4
Theoretical stress concentration factor, p. 4 Unpaired rank test, p. 30, Table 26 (p. 58)
Tolerance interval, pp. 5, 41, 42 Values of t, Table 29 (p. 62)
Tolerance level, p. 5 Variable component of stress, p. 3
Tolerance limits, p. 5 Weibull distribution, p. 71
2-test, p. 45 Weibull mean, Fig. 15 (p. 77), Footnote p. 78
/-values, Table 29 (p. 62) Wohler test method, p. 1
This page intentionally left blank
T HIS PUBLICATION is one of many
issued by the American Society for Testing and Materials
in connection with its work of promoting knowledge
of the properties of materials and developing standard
specifications and tests for materials. Much of the data
result from the voluntary contributions of many of the
country's leading technical authorities from industry,
scientific agencies, and government.
Over the years the Society has published many tech-
nical symposiums, reports, and special books. These may
consist of a series of technical papers, reports by the
ASTM technical committees, or compilations of data
developed in special Society groups with many organiza-
tions cooperating. A list of ASTM publications and
information on the work of the Society will be furnished
on request.