You are on page 1of 12

Engineering Fracture Mechanics 71 (2004) 725736

www.elsevier.com/locate/engfracmech

Comparison of fatigue data using the maximum


likelihood method
Luca Goglio, Massimo Rossetto

Dipartimento di Meccanica, Politecnico di Torino, Corso Duca degli Abruzzi, 24, Torino 10129, Italy
Received 30 October 2002; accepted 31 October 2002

Abstract
The paper describes procedures to compare the fatigue behaviour of two populations corresponding to batches that
differ for a factor on the basis of fatigue data of two samples. Fatigue data are elaborated by means of the maximum
likelihood method to obtain the SNP curves and their confidence intervals. The likelihood ratio test is used to identify
the most adequate distribution and model. To analyse the influence of the factor, two integrated procedures are
adopted: the first one is based on the LR test, the second is an empirical procedure, based on the comparison of the
confidence intervals of the curves.
! 2003 Elsevier Ltd. All rights reserved.
Keywords: Fatigue data; Analysis; Comparison; Maximum likelihood; Likelihood ratio test

1. Introduction
In this paper we describe a statistical procedure to compare the SNP curves of two or more samples to
assess the influence of a factor and to state if the samples are extracted from the same population (at least
with respect to the fatigue behaviour).
In literature the comparison is usually performed visually, by plotting the mean curves (B50 ), with or
without the data points, on the same graph (for example, [17]); usually the data are fitted by means of a
least square method that has some limitations because it is assumed that the scatter, as a function of the
stress level, is uniform and runouts are not considered. In some cases also the scatter bands of the curves,
that can help in decisions, are plotted [8,9], and data are elaborated using the maximum likelihood method
[8].
The visual comparison of the curves can be sufficient only if there is a large difference in the obtained
curves and original data, but in some case some doubts can arise. In these cases it is desirable to have at
disposal statistical tests allowing for a convincing comparison of the data.

Corresponding author. Fax: +39-11-5646999.


E-mail address: massimo.rossetto@polito.it (M. Rossetto).

0013-7944/$ - see front matter ! 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/S0013-7944(03)00009-2

726

L. Goglio, M. Rossetto / Engineering Fracture Mechanics 71 (2004) 725736

A more sophisticated method consists in comparing the fatigue data for each stress level, or selected
ones, using the classical statistical tests [10,11], but this method requires to have a large number of data for
each level. Moreover, all the samples must be obtained at the same levels, and the comparison is not made
on overall data.
In [12] the comparison is performed using also the analysis of variance and of covariance, but this
statistical methods requires the homogeneity of variance (scatter constant as a function of stress level) and
the data cannot contain runouts.
In [13] the likelihood ratio test is proposed to compare K samples fitted with the same model by means
of the maximum likelihood method.
In this paper also a graphical method based on the comparison of the SNP curves with their confidence
interval, more convincing, in our opinion, in respect to the usual graphical methods, is presented.
For a correct comparison of the data, and in any case to have reliable SNP curves, it is also important to
state the most adequate model to fit the data, considering the type of diagram (semilog or loglog), the
theoretical distribution that has to be considered and the type of model as a function of the stress level
(linear or nonlinear scale parameters of distribution, constant or non-constant shape parameter, etc.).
Even if a lot of theoretical distribution have been proposed [1418], the most used are two parameters
log-normal and Weibull distributions. In some cases three-parameter distributions are proposed [11,14], but
usually there are some difficulties in the determination of the location parameter (corresponding to the B!0
curve) due to the low number of fatigue data in the samples.
To state the theoretical distribution that best fit the data it is possible to apply the usual statistical test to
the data obtained at a single stress levels, based essentially on the linear regression of the (transformed) data
on probability plots [14]. Also in this case the method considers the single stress level and not on overall
data, and runouts cannot be considered.
In [19] Little describes a statistical test to discriminate between the log-normal and Weibull distribution
considering all data simultaneously (all stress levels, failures and runouts); this test is based on the ratio
between the maximum likelihood obtained using the two distributions. A similar test that uses the general
likelihood ratio test is proposed by Nelson in [13] on the basis of a theoretical work of Farewell and
Prentice [20].
For the choice of the most adequate model in [13,21] a test, again based on likelihood ratio test, is
presented. Also the use of graphs for the raw and normalised residuals, which are evaluated as distance of
the data from the mean value divided by the standard deviation, can help in the decision.
In this paper a procedure to fit and to compare the fatigue data of two samples that integrate the
likelihood ratio tests proposed by Nelson in [13] and graphical methods is presented. The procedure is
developed in two steps: determination of models that best fits the data of the two sample and comparison of
the data. Some consideration about some possible difficulties are reported.
Furthermore an example concerning the influence of defects on the fatigue behaviour of M12 welded
studs with mean stress of 400 MPa is reported.

2. Determination of the model for the SNP curves


To determine the SNP curves and their confidence limits the maximum likelihood method [13,21] has
been used. Unlike the methods based on least squares, this allows for considering also interrupted tests
(runouts) and life data affected by scatter variable as a function of the stress level. For computational
convenience usually the logarithm of the likelihood is used ( maximum of the log likelihood).
In our implementation of the method, developed in the MATLAB environment, we have considered
both loglog and semilog diagrams, two distributions (two parameters log-normal and Weibull), and four
type of the models of life as a function of the stress level or log of stress: linear and nonlinear mean,

L. Goglio, M. Rossetto / Engineering Fracture Mechanics 71 (2004) 725736

727

constant and non-constant standard deviation in the case of log-normal distribution; linear and nonlinear
log of characteristic life, constant and non-constant shape factor in case of Weibull distribution.
In the case of loglog diagram a Weibull model [22] (also referred as Stromeyer model [23]) is considered.
For the log-normal distribution the models are
adl
bdl
cdl
ddl

l C1 C2 log S $ log S
l C1 C2 log S $ log S
l C1 C2 logS $ S0 $ log S
l C1 C2 logS $ S0 $ log S

r expD
! 1
"
r exp D1 D2 log S $ log S
r expD1
r expD1 D2 logS $ S0 $ log S'

where l is the mean of log N and r is the standard deviation of log N , Ci ; Di ; S0 are the coefficients of the
models that maximise the log likelihood of the sample () and log S is the mean of log of stress; this
normalisation of the stress allows a better numerical convergence and narrower confidence intervals for
coefficient C1 .
It is worth of note that using the maximum log likelihood method the coefficient S0 , sometimes referred
to as fatigue limit [23,24], is only a mathematical coefficient used to have a nonlinear model. In fact, for
computational reasons, its value must be lower of the lowest stress level of the sample, and this value
usually differs from the 50% fatigue limit.
Considering a Weibull distribution the models are
adw
bdw
cdw
ddw

ln h C1 C2 log S $ log S
ln h C1 C2 log S $ log S
ln h C1 C2 logS $ S0 $ log S
ln h C1 C2 logS $ S0 $ log S

b expD
! 1
"
b exp D1 D2 log S $ log S
b expD
! 1
"
b exp D1 D2 logS $ S0 $ log S

where h is characteristic life and b is the shape factor.


In the case of semilog diagram the nonlinear behaviour of the mean or of the log of the characteristic life
is obtained using a hyperbolic model [25], thus the considered models for semilog diagram and log-normal
distribution are
asl
bsl
csl
dsl

l C1 C2 S
l C1 C2 S
l C1 C2 S C3 =S
l C1 C2 S C3 =S

r expD1
r expD1 D2 S
r expD1
r expD1 D2 S

and considering a Weibull distribution:


asw
bsw
csw
dsw

ln h C1 C2 S
ln h C1 Cs S
ln h C1 C2 S C3 =S
ln h C1 C2 S C3 =S

b expD1
b expD1 D2 S
b expD1
b expD1 D2 S

In this case it is not possible to normalise the stress because in the computational procedure some data
become infinite.
The analysis of the data requires some initial assumptions, namely the type of diagram, the theoretical
distribution, the type of model for the distribution parameters. Therefore it is desirable to have at disposal
statistical tests and procedures allowing the best choices for the representation of the data.
2.1. Choice of the type of diagram
To the best of our knowledge, there is no statistical or empirical test to choose the type of diagram
(semilog or loglog); therefore, this choice is due to experience and convenience in using the data.

728

L. Goglio, M. Rossetto / Engineering Fracture Mechanics 71 (2004) 725736

2.2. Choice of the distribution


The choice between log-normal and Weibull distribution can be performed, as proposed by Nelson
in [13], using the likelihood ratio. The test requires to calculate the statistics:
Twl 2w l

where w and l are respectively the maximum log likelihood obtained using the Weibull and log-normal
distributions considering consistent types of model (with the same structure, for instance csl and csw ). Twl is
then compared to the v2 distribution for the desired confidence level (1 $ a) and for one degree of freedom;
three cases are possible:
ii(i) Twl < 0: log-normal distribution fits better the data;
i(ii) 0 < Twl < v2 1 $ a; 1: the Weibull distribution fits better but not significantly better;
(iii) Twl > v2 1 $ a; 1: the Weibull distribution fits significantly better the data.
Therefore, to chose the distribution, the data must be elaborated using both Weibull and log-normal
distribution adopting all the type of models for the chosen type of diagram.
If cases (i) and (ii) occur for all type of models the log-normal distribution can be assumed, whereas cases
(iii) and (ii) occur the Weibull distribution can be assumed. If cases (i) and (iii) occur for the different types
of model the choice must be carried out contextually to the choice of the type of model.
In the following the likelihood ratio test used to state the most appropriate distribution will be referred
to as LR(D) test.
2.3. Choice of the model
The most adequate model for the available sample, after that the diagram and distribution have been
chosen, can be selected by applying both graphical and numerical methods.
The graphical method requires the construction of graphs for the raw and normalised residuals, which
are evaluated as distance of the data from the mean value divided by the standard deviation [13]. No
difficulties arise in constructing such diagrams, but their interpretation requires experience, the larger the
number of considered alternatives the higher the difficulty.
The numerical method uses again the likelihood ratio test [21,13]; in this case the test can be used when
the models belong to the same set, i.e. when a model can be derived from a more general one by setting to
zero the values of some coefficients.
For instance in our case models asl can be derived from model dsl by setting to zero coefficients C3 and
D2 , the same applies to other models.
The likelihood ratio test allows for checking whether the coefficients added to obtain a higher degree
model are statistically significant, i.e. if the addition significantly improves data representation [21].
The test requires to calculate the statistics:
Tpq 2p $ q

where p is the logarithm of the maximum likelihood obtained using the higher order model and q is the
logarithm of the maximum likelihood obtained using the lower order model. The value Tpq is then compared
to the v2 distribution for the desired confidence level (1 $ a) and a number of degrees of freedom equal to
the number of coefficients being added (G). If Tpq > v2 1 $ a; G the higher order model fits significantly
better the data.

L. Goglio, M. Rossetto / Engineering Fracture Mechanics 71 (2004) 725736

729

If different response of the test for the distribution for different types of model occurred the data must
be elaborated using both the distributions and the best combination of type of model and distribution must
be chosen.
In the following the likelihood ratio test used to state the most appropriate type of model will be referred
to as LR(M) test.
Another possible procedure to assess the significance of a higher order model, consists in checking that
the confidence interval of the added coefficient does not enclose zero [21]. The (approximate) confidence
intervals are easy to obtain from the covariance matrix that can be evaluated considering the Fisher matrix,
i.e. the matrix of negative second partial derivatives of with respect to coefficients Ci and Di and S0 .
However, this procedure may lead to incorrect conclusions. In fact, whether the LR(M) test states that a
coefficient is statistically significant, its confidence interval does not enclose zero; but, whether the LR(M)
test states that a coefficient is not statistically significant, not necessarily the confidence interval encloses
zero.

3. Comparison of two or more samples


To compare two, or more, samples two methods are here reported: the first one is an empirical graphical
method, valid in the case of two samples, the second one is a numerical method based on likelihood ratio
test.
3.1. Graphical method
The data obtained from the two samples are treated separately by means of the maximum likelihood
method, using the same model chosen by the method described previously.
Then the characteristic life hS for the Weibull distribution, or the mean life lS for the log-normal
distribution, with their related confidence interval at the chosen level (e.g. 95%) are plotted and superposed
on the same graph. If the curve hS (lS) of one batch is included in the confidence interval of the curve of
the other batch, and vice versa, from the statistical viewpoint the two curves may not be distinguished, thus
the characteristic life (mean life) of the two batches is statistically the same. If this is not the case, it may be
concluded that the behaviour of the batches is different.
If the characteristic lives (mean lives) are the same, the curves corresponding to the highest interesting
survival percentile (for instance 90%) are plotted together with their confidence intervals. If the curve of the
survival percentile of one batch is included in the confidence interval of the curve of the other batch, and
vice versa, from the statistical viewpoint the two curves may not be distinguished. In this case it may be
assumed that the shape parameters bS (rS) are statistically equal. Thus the data from the two batches
represent two samples extracted from the same population (at least with respect to the fatigue behaviour).
The latter step is the most critical, because the functions of the percentile and the related confidence
interval also depend on hS (or lS), that are numerically different for the two samples. This makes the
procedure weak from the theoretical viewpoint, but we think that as far as application is concerned having
curves not distinguishable at the highest percentile of interest justifies the assumption of unique population.
3.2. Likelihood ratio test method
This method allows to compare K samples elaborated using the same model with Q coefficients and it is
based on the simultaneous comparison of the Q coefficients [13].
The method requires to compute the maximum likelihood, for each sample (i ) and for the pooled data
(0 ). The test requires to calculate the statistics:

730

L. Goglio, M. Rossetto / Engineering Fracture Mechanics 71 (2004) 725736

T 2

#X

i $ 0

and its value is compared to the v2 distribution for the desired confidence level (1 $ a) and a number of
degrees of freedom equal to Q.
If T 6 v2 1 $ a; Q the Q coefficients of the samples not differ significantly (at the chosen confidence level)
and it is possible to state that the samples are extracted from the same population (at least with respect to
the fatigue behaviour).
If T > v2 1 $ a; Q at least one pairs of coefficients differ convincingly and it is possible to state that the
samples are extracted from different population (with respect to the fatigue behaviour).
In the latter case, it is possible to compare the confidence intervals of the coefficients to state why the
samples differ. In fact it is possible, for instance, that the two samples differ only for the dispersion
parameter, but have the same mean; also the superposed plots can help in this comparison.
Obviously, after that it has been concluded that the data are all from the same population the fatigue
curves may be obtained considering the pooled data.
In the following the likelihood ratio test used to compare samples will be referred to as LR(C) test.

4. Example: defective or non-defective welded studs


The example here reported concerns the influence on the fatigue behaviour of porosity in the weld zone
of M12 studs welded, by means of a welding gun, on circular plates 30 mm diameter and 5 mm thick. All
specimens have been controlled by ultrasonics adopting the immersion technique, and were classified as
defective or non-defective.
Fatigue tests have been carried out on a resonant machine (AMSLER 10HFP 422) applying a mean
stress values of 400 MPa (approximately corresponding to 70% of the yield stress). The testing frequency
was 190 Hz.
For both the samples the fatigue limit (alternate component) has been evaluated by means of the staircase procedure [26,27], assuming as threshold for infinite life 5 ( 106 cycles and a step d 10 MPa; for each
sample 15 specimens have been used.
For both samples the estimate of the fatigue limit was SD50% 26:4 MPa and its 95% confidence interval
{21.131.7 MPa}, the standard deviation r 5:3 MPa, so it is obvious that defects have no influence on
fatigue limit. (The precision adopted in reporting the results (10$1 MPa) is consistent with that achieved
when setting the load on the testing machine. Obviously, from the technical viewpoint it would be enough
to round values to the closest integer or even five. However, data have been kept in this form to allow for
further comparison.)
This result may not be directly applied to whole Whler curve. For this reason tests at stress higher than
the fatigue limits have been carried out separately for defective and non-defective specimens (at 1108060
MPa).
Considering also the data from staircase the non-defective sample has 28 data (seven runouts) and
defective sample 30 data (seven runouts).
These data were elaborated using the developed !Maximum Likelihood" program considering semilog
diagrams and compared using the methods described above. Resulting log likelihood for the two samples
are reported in Tables 1 and 2.
Applying the LR(D) test for both samples and for all models result Twl < 0, except for model as of the
defective sample. In this case, considering a confidence level of 95%, result Twl 2:48 < 3:84 v2 0:95; 1.
Therefore for all cases log-normal distribution has been chosen.
To state the most adequate model the LR(M) test was applied (Table 3).

731

L. Goglio, M. Rossetto / Engineering Fracture Mechanics 71 (2004) 725736


Table 1
Log likelihood () for non-defective sample
Model

Log-normal

Weibull

Twl

as
bs
cs
ds

)290.74
)285.00
)280.88
)277.99

)291.31
)287.44
)283.63
)281.56

)1.14
)4.88
)5.50
)7.14

Table 2
Log likelihood () for defective sample
Model

Log-normal

Weibull

Twl

as
bs
cs
ds

)315.38
)312.52
)298.99
)297.90

)314.14
)313.75
)301.43
)300.33

2.48
)2.46
)4.88
)4.86

Table 3
LR test to chose models
Sample

Tba 1 (D2 )

Tca 1 (C3 )

Tdc 1 (D2 )

Tdb 1 (C3 )

Tda 2 (C3 ; D2 )

Non-defective
Defective

11.46
5.72

19.72
32.78

5.78
2.18

14.02
29.24

25.50
34.96

Bold if Tpq < v2 1 $ a; G (no influence of coef.): 1 limit value v2 0:95; 1 3:84, 2 limit value v2 0:95; 2 3:84.

Observing these data the most adequate model for the non-defective sample is dsl , whereas that for the
defective sample is csl . It is worth of note that the difference between the two models is the constant or
non-constant standard deviation.
In addition, it can be noted that, in this case, the observation of the confidence intervals of the coefficients (reported in Tables 4 and 5) leads to the same conclusions.
The difference in the model is not sufficient to conclude that the populations present different fatigue
behaviour, in fact the differences could be due to the low number of points.
The graphical comparison can help to decide about the difference between the models with engineering
judgement.
In our case the comparison of the confidence interval of the B10 curves obtained using the two models
(Fig. 1) leads to state that it is not possible, from a statistical point of view, to distinguish between the two
models, so it is possible to continue in the comparison procedure using the higher order model dsl for both
samples (this can be convenient but it is not compulsory).
The fatigue strengths at 5 ( 106 cycles calculated using model dsl is 24.6 MPa for non-defective sample
and 27.1 MPa for defective sample, both in the 95% confidence interval of the fatigue limit at 5 ( 106 cycles
obtained from staircases (SD50% 26:4 MPa {21.131.7 MPa}). It is worth of note that models asl and bsl
(for non-defective sample) give fatigue strengths out of this confidence interval. This fact confirm that the
use of a wrong model gives non-reliable results.
To compare graphically the two samples B50 and B10 curves and their 95% confidence interval are superposed (Figs. 2 and 3 respectively); it is possible to note only a small difference at lower stress levels,
where the staircases indicate an equal fatigue behaviour. This fact can be due to the presence at these levels
of all the runouts [13].

732

L. Goglio, M. Rossetto / Engineering Fracture Mechanics 71 (2004) 725736

Table 4
Non-defective sample: 95% interval confidence of the models coefficients
M12Sm 400 MPa non-defective

Lower

Upper

Model asl

C1
C2
D1

6.0553
)0.0199
)2.0458

6.0043
)0.0216
)2.3625

6.1064
)0.0183
)1.7291

Model bsl

C1
C2
D1
D2

6.0172
)0.0180
)2.0619
)0.0279

5.9545
)0.0192
)2.4044
)0.0397

6.0798
)0.0168
)1.7194
)0.0160

Model csl

C1
C2
C3
D1

6.1959
)0.0124
19.8400
)2.3322

5.7495
)0.0161
8.9864
)2.6352

6.6422
)0.0087
30.6940
)2.0292

Model dsl

C1
C2
C3
D1
D2

6.3549
)0.0138
16.7770
)1.5306
)0.0157

6.0172
)0.0164
8.0588
)2.3333
)0.0282

6.6926
)0.0112
25.4950
)0.7278
)0.0032

Lower

Upper

Table 5
Defective sample: 95% interval confidence of the models coefficients
M12Sm 400 MPa defective
Model asl

C1
C2
D1

6.0308
)0.0205
)2.0723

5.9823
)0.0220
)2.3665

6.0792
)0.0190
)1.7781

Model bsl

C1
C2
D1
D2

5.9384
)0.0170
)1.9411
)0.0324

5.8609
)0.0185
)2.2668
)0.0442

6.0159
)0.0154
)1.6154
)0.0206

Model csl

C1
C2
C3
D1

5.8674
)0.0101
30.2940
)2.6517

5.5186
)0.0129
21.7240
)2.9444

6.2162
)0.0073
38.8640
)2.3590

Model dsl

C1
C2
C3
D1
D2

5.9953
)0.0111
27.2460
)2.0161
$0:0104

5.6394
)0.0138
17.8690
)2.9923
$0:0245

6.3511
)0.0085
36.6220
)1.0400
0:0038

Bold if the 95% confidence interval includes zero.

To apply the LR(C) test the maximum log likelihood 0 for the pooled data have to be computed: the
results are reported in Table 6. Applying the LR(D) test to the pooled data and considering a limit value for
Twl v2 0:95; 1 3:84 the most suitable distribution for the pooled data is again the log-normal.
Comparing the sample applying the LR(C) test, considering the model dsl (number of coefficients Q 5)
and a 95% confidence interval results:
T 21 2 $ 0 2$277:99 $ 297:90 577:82 3:86 6 11:07 v2 0:95; 5

L. Goglio, M. Rossetto / Engineering Fracture Mechanics 71 (2004) 725736

733

Fig. 1. Defective sample: comparison of B10 curves obtained using csl and dsl models.

Fig. 2. Comparison of defective and non-defective samples: B50 curves.

so, the two Q pairs of coefficients of the samples do not differ significantly and it is possible to state that
the samples are extracted from the same population (at least with respect to the fatigue behaviour).
It is worth of note that the same holds if the model csl is considered, in fact in this case it is
T 21 2 $ 0 2$280:88 $ 298:99 583:31 6:88 6 11:07 v2 0:95; 5
On the basis of the above considerations and the result of the LR(C) test it is possible to state that the
two samples are extracted from the same population (with respect to the fatigue behaviour). The LR(M)
tests for the pooled data (Table 7) confirm the choice of the dsl model. The fatigue strength at 5 ( 106 cycles
computed from this model for the pooled data is 25.8 within the 95% confidence interval of the fatigue limit
at 5 ( 106 cycles obtained from staircases. Fig. 4 reports the B10 , B50 and B90 curves obtained with this
model for the pooled data.

734

L. Goglio, M. Rossetto / Engineering Fracture Mechanics 71 (2004) 725736

Fig. 3. Comparison of defective and non-defective samples: B10 curves.

Table 6
Log likelihood 0 for pooled data
Model

Log-normal

Weibull

Twl

as
bs
cs
ds

)606.67
)599.06
)583.31
)577.82

)605.94
)602.37
)586.60
)583.22

1.46
)6.62
)6.58
)10.80

Table 7
LR test to chose models
Pooled data

Tba 1 (D2 )

Tca 1 (C3 )

Tdc 1 (D2 )

Tdb 1 (C3 )

Tda 2 (C3 ; D2 )

15.22

46.72

10.98

42.48

57.70

Bold if Tpq < v2 1 $ a; G (no influence of coef.) 1 limit value v2 0:95; 1 3:84, 2 limit value v2 0:95; 2 3:84.

Fig. 4. SNP curves of the pooled data.

L. Goglio, M. Rossetto / Engineering Fracture Mechanics 71 (2004) 725736

735

5. Conclusions
The two procedures described in the paper allow to compare the fatigue behaviour of two population
corresponding to batches that differ for a factor on the basis of fatigue data of two samples; the first is based
on the likelihood ratio test, the second one is an empirical graphical procedure.
In both cases the fatigue data have to be elaborated using the maximum likelihood method, and procedures to identify the most adequate distribution and model to fit the data and to perform correctly the
comparison are stated.
In the reported example is pointed out that the two procedures can be used complementary to perform
a correct comparison with engineering judgement.

Acknowledgements
This research was supported by the Italian Research Council (CNR Grants. 96.02437.CT07 and
97.03106.CT07).

References
[1] Kihl DP, Sarkani S. Thickness effect on the fatigue strength of welded steel cruciforms. Int J Fatigue 1997;19(Suppl 1):S3116.
[2] Ninh Nguyen T, Wahab MA. The effect of weld geometry and residual stresses on the fatigue of welded joints under combined
loading. J Mater Process Technol 1998;77(13):2018.
[3] Scholz R, Mueller R. The effect of hold-times on the fatigue behavior of type AISI 316L stainless steel under deuteron irradiation.
J Nucl Mater 1998;258263:16005.
[4] Lin C-K, Lee W-J. Effects of highly stressed volume on fatigue strength of austempered ductile irons. Int J Fatigue 1998;20(4):301
7.
[5] Lin C-K, Yang S-T. Corrosion fatigue behavior of 7050 aluminum alloys in different tempers. Engng Fract Mech 1998;59(6):779
95.
[6] Hassan T, Liu Z. On the difference of fatigue strengths from rotating bending, four-point bending, and cantilever bending tests. Int
J Pressure Vessel and Piping 2001;78(1):1930.
[7] Ochi Y, Masaki K, Matsumura T, Sekino T. Effect of shot-peening treatment on high cycle fatigue property of ductile cast iron.
Int J Fatigue 2001;23(5):4418.
[8] Beretta S, Boniardi M. Microstructure and fatigue properties of a welded duplex stainless stell. Fatigue Fract Engng Mater Struct
1996;19(6):64754.
[9] Lazzarin P, Milani V, Quaresimin M. Scatter bands summarizing the fatigue strength of aluminium alloy bolted joints. Int J
Fatigue 1997;19(5):4017.
[10] Audenino A, Goglio L, Rossetto M. Fatigue resistance of welded threaded studs. Proc Eurojoin 1994;2:75766.
[11] Goto M. Statistical investigation of the effect of laboratory air on the fatigue behaviour of a carbon steel. Fatigue Fract Engng
Mater Struct 1998;21(6):70515.
[12] Haibach E, Olivier R, Rinaldi F. Statistical design and analysis of an interlaboratory program on the fatigue properties of welded
joints in structural steel. In: Little RE, Ekuvall JC, editors. ASTM STP 744 statistical analysis of fatigue data. 1981. p. 2454.
[13] Nelson W. Accelerated testing. New York: John Wiley; 1990.
[14] Zhao Y-X, Gao Q, Sun X-F. A statistical investigation of the fatigue lives of Q235 steel-welded joints. Fatigue Fract Engng Mater
Struct 1998;21(7):78190.
[15] Beretta S, Boniardi M. Fatigue strength and surface quality of eutectoid steel wires. Int J Fatigue 1999;21(4):32935.
[16] Gong Yi, Norton MP. Materials fatigue life distribution: a maximum entropy approach. J Test Eval 1998;26(1):5363.
[17] Castillo E, Fern!
andez-Canteli A, Hadi AS. On fitting a fatigue model to data. Int J Fatigue 1999;21(1):97106.
[18] Owen WJ, Padgett WJ. A BirnbaumSaunders accelerated life model. IEEE Trans Reliability 2000;49(2):2249.
[19] Little RE. Review of statistical analyses of fatigue life data using one-side lower statistical tolerance limits. In: Little RE, Ekuvall
JC, editors. ASTM STP 744 statistical analysis of fatigue data. 1981. p. 323.
[20] Farewell VT, Prentice RL. A study of distributional shape in life testing. Technometrics 1977;19(1):6975.
[21] Nelson W. Fitting of fatigue curves with nonconstant standard deviation to data with runouts. J Test Eval 1984;12(1):6977.

736

L. Goglio, M. Rossetto / Engineering Fracture Mechanics 71 (2004) 725736

[22] Sors L. Fatigue design of machine components. Oxford: Pergamon; 1971.


[23] Spindel JE, Haibach E. Some considerations in the statistical determination of the shape of SN curves. In: Little RE, Ekuvall JC,
editors. ASTM STP 744 statistical analysis of fatigue data. 1981. p. 89113.
[24] Pascual FG, Meeker WQ. Analysis of fatigue data with runouts based on a model with noncostant standard deviation and fatigue
limit parameter. J Test Eval 1997;25(3):292301.
[25] Conway JB, Sjodahl LH. Analysis and representation of fatigue data. ASM International; 1991.
[26] Dixon WJ, Massey FJ. Introduction to statistical analysis. International Student Edition. McGraw-Hill; 1985.
[27] Lin S-K, Lee Y-L, Lu M-W. Evaluation of the staircase and the accelerated test methods for fatigue limit distributions. Int
J Fatigue 2001;23(1):7583.

You might also like