You are on page 1of 13

A Bayesian analysis of fatigue data

Maurizio Guida
a,
*
, Francesco Penta
b
a
Department of Information and Electrical Engineering, University of Salerno, 84084 Fisciano (SA), Italy
b
Department of Mechanics and Energetics, University of Naples Federico II, Naples, Italy
a r t i c l e i n f o
Article history:
Received 4 August 2008
Received in revised form 5 June 2009
Accepted 7 August 2009
Available online 13 September 2009
Keywords:
Fatigue testing
PSN curves
Statistical analysis
Bayesian inference
Fatigue design curves
a b s t r a c t
The aim of the present paper is to bring arguments in favour of Bayesian inference in the context of fati-
gue testing. In fact, life tests play a central role in the design of mechanical systems, as their structural
reliability depends in part on the fatigue strength of material, which need to be determined by experi-
ments. The classical statistical analysis, however, can lead to results of limited practical usefulness when
the number of specimens on test is small. Instead, despite the little attention paid to it in this context,
Bayes approach can potentially give more accurate estimates by combining test data with technological
knowledge available from theoretical studies and/or previous experimental results, thus contributing to
save time and money. Hence, for the case of steel alloys, a discussion about the usually available techno-
logical knowledge is presented and methods to properly formalize it in the form of prior credibility den-
sity functions are proposed. Further, the performances of the proposed Bayesian procedures are analysed
on the basis of simulation studies, showing that they can largely outperform the conventional ones at the
expense of a moderate increase of the computational effort.
2009 Elsevier Ltd. All rights reserved.
1. Introduction
The development of systems with a high structural reliability
together with low manufacturing and maintenance costs requires
the mechanical strength of materials be accurately known. This le-
vel of knowledge, however, is difcult to obtain as far as the fatigue
phenomenon is concerned, for several reasons: the inherent com-
plexity of the fatigue mechanism, the variety of inuential factors
and the laboriousness of fatigue testing.
Even if the laboratory fatigue tests are carried out under con-
trolled operating conditions, fatigue data are characterized by a
high variability and require a statistical analysis. This analysis
essentially consists in: assuming a probabilistic model for the life-
times, estimating its unknown parameters and constructing a fa-
tigue design curve which should take into account both the
inherent randomness of fatigue phenomenon and the uncertainty
on the true values of model parameters due to the nite sample
size. The classical (frequentist) statistical analysis, however, can
lead to results of limited technical usefulness when the sample size
is small. This circumstance often happens at the beginning of the
development phase of new systems or components, when limited
experimental data are usually available for both cost and time
reasons.
A lot of theoretical studies and experimental life testing results
have been produced by engineers and manufacturers in order to
evaluate the properties of different materials. When new constant
amplitude life tests are carried out, however, this body of techno-
logical knowledge is completely ignored by the conventional esti-
mation procedures which are usually adopted. In principle, the
available technological information, when properly formalized as
a prior credibility on model parameters, might be combined via
Bayes theorem with test data to make more accurate inference
on the quantities of interest (for a comprehensive reference on
Bayes inferential approach see, as an example, [1]).
Although a number of papers have appeared which exploited
Bayesian inference in the analysis of the propagation of fatigue
cracks, very few attempts have been made in the past to use the
Bayes approach in the context of SN fatigue tests. In fact, as far
as the authors are aware, only two papers have specically ad-
dressed this topic. For the case of offshore structural joints, Madsen
[2] considered a Bayesian linear regression analysis where the
prior information was derived from the behaviour of similar joints.
The posterior distribution of model parameters was then used to
predict the fatigue life. Edwards and Pacheco [3] in their paper crit-
icised two limitations of the conventional procedures for establish-
ing design SN curves from laboratory fatigue tests, namely the use
of design curves based on point estimates of regression model
parameters and the impossibility of accounting for fatigue run-
outs in a rational manner. Then, they proposed the use of Bayesian
method (in its noninformative setting) in order to handle both
these limitations.
The little attention paid to Bayesian methods in the context of
SN fatigue testing is probably due to two main reasons: (a) the
0167-4730/$ - see front matter 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.strusafe.2009.08.001
* Corresponding author. Tel.: +39 089 964282; fax: +39 089 964218.
E-mail address: mguida@unisa.it (M. Guida).
Structural Safety 32 (2010) 6476
Contents lists available at ScienceDirect
Structural Safety
j our nal homepage: www. el sevi er. com/ l ocat e/ st rusaf e
greater computational burden usually associated to Bayesian
methods when compared to the classical methods available for
SN tests, and (b) the suspiciousness still existing in many people
towards subjective probability which is the basis of Bayesian
inference.
The aim of this paper is to bring arguments in favour of Bayes-
ian inference in the context of constant amplitude fatigue testing,
by taking into account that the nowadays computational capabili-
ties make the use of Bayesian viewpoint feasible and by showing,
on the basis of simulation studies, that the performances of Bayes-
ian procedures can sensibly outperform the conventional ones,
when the available technical information is properly formalized.
The paper is organized as follows: the next section reviews the
basic concepts of SN tests and related classical estimation proce-
dures. In Section 3 the Bayes point of view is briey outlined and
suitable prior and posterior probability density functions (pdf)
are introduced. Section 4 reviews technical knowledge on the cor-
relations among static and fatigue properties of steels, which is
commonly available. Section 5 is, then, devoted to the mathemat-
ical formalization of this technical knowledge in the form of prior
pdfs on material parameters. Finally, Sections 6 and 7 present cri-
teria to compare Bayesian and classical estimators along with
numerical results obtained by Monte Carlo simulation.
2. SN tests and the related classical inferences
The lifetime of mechanical systems depends in part on the fati-
gue strength of the material, whose properties need to be deter-
mined by experiments. In constant amplitude fatigue tests, test
specimens are subjected to alternating load until failure. The mag-
nitude of the load amplitude is the controlled (independent) vari-
able and the number of cycles to failure is the response
(dependent) variable. In the nite life region, the number of cycles
to failure N
i
is usually assumed to depend on the load amplitude S
i
by the Basquins relation [4]
N
i
= AS
B
i
(1)
where A(> 0) and B(> 0) are (usually unknown) material
parameters.
Because of inherent microstructural inhomogeneity in the
materials properties, differences in the surface and in test condi-
tions of each specimen and other factors, the number of cycles to
failure exhibits a random behaviour. To model this experimental
variability, a multiplicative randomerror is introduced in (1) giving
N
i
= AS
B
i
e
i
(2)
where e
i
is usually assumed to be distributed as a log-normal ran-
dom variable (rv) with (unknown) variance independent of S
i
. In re-
peated testing, the errors e
i
are also assumed to be stochastically
independent rvs.
For statistical analysis it is convenient to rewrite model (2) by
taking the logarithm (either the log base-10 or the natural log) of
both sides of (2) and subtracting the sample mean of log S
i
times
B, arriving to Y
i
= a b(u
i
u) re
i
, where Y
i
= log N
i
, u
i
= log S
i
,
u = (1=n)

n
i=1
log S
i
, b = B, a = logA B u, and e
i
- N(0; 1).
By letting x
i
= (u
i
u) with

n
i=1
x
i
= 0, the simple linear regres-
sion model is obtained
Y
i
= a bx
i
re
i
(3)
Under the above assumptions, the Y
i
are independently distrib-
uted Normal rvs with mean l
x
i
= a bx
i
and constant variance r
2
.
After having observed the couples (x
i
; y
i
)(i = 1; :::; n), point esti-
mates of the unknown parameters a and b are obtained by the
Least Squares method or, equivalently, by the maximum likelihood
method, which both give
^
a =
1
n

n
i=1
y
i
^
b =

n
i=1
x
i
y
i

n
i=1
x
2
i
(4)
Thus, ^ a and
^
b are linear estimators of a and b, i.e. they are linear
combinations of the observed data y
i
. Since Y
i
are Normal rvs, in
repeated sampling ^ a and
^
b also are Normal rvs. They are unbiased
minimum variance estimators of a and b.
The error variance r
2
is usually estimated by
s
2
=

n
i=1
[y
i
^ l
x
i
[
2
n 2
(5)
where ^ l
x
i
= ^ a
^
bx
i
. This is a minimum variance unbiased estimator
of r
2
.
From the frequentist (classical) point of view, the uncertainty
about an unknown quantity (model parameters or functions there-
of, or future values of observable random variables) is measured
through a random interval generated by a rule which has in re-
peated sampling a known probability (1 c) of generating intervals
that include the unknown quantity of interest. Depending on the
particular quantity to be estimated, random intervals are given dif-
ferent names: condence intervals, when population parameters
such as mean or variance are involved; tolerance intervals, when
percentages of a population are involved; and prediction intervals,
when future values of observable rvs are involved.
For the linear regression model (3), the following exact results
hold (see, for example, [5,6]).
Condence interval on the median l
x
= a bx: The rv
(^ l
x
l
x
)=s^ lx
, where ^ l
x
= ^ a
^
bx and s^ lx
= s

1=n x
2
=

n
i=1
x
2
i
_
, has
a Student distribution with m = n 2 degrees of freedom (df), so
that the (1 c) two-sided equal tailed condence interval is
^ l
x
st
1c=2;n2

1=n x
2
=

n
i=1
x
2
i

_
: (6)
Condence interval on the 100p-quantile y
p
x
= a bx rz
p
(toler-
ance interval): The 100p-quantile of the distribution of a rv Y is de-
ned as the value y
p
such that Pr(Y 6 y
p
) = p. It can be readily
shown that the rv (^ l
x
y
p
x
)=s^ lx
has a non-central Student distribu-
tion with m = n 2 df and non-centrality parameter
d = z
p
=

1=n x
2
=

n
i=1
x
2
i
_
. Thus, the (1 c) two-sided equal
tailed condence interval for the 100p-quantile y
p
x
is given by
^ l
x
st
d;1c=2;(n2)

1=n x
2

n
i=1
x
2
i
_

_
;
^ l
x
st
d;c=2;(n2)

1=n x
2

n
i=1
x
2
i
_

_
: (7)
In particular, a lower condence limit of the 100p-quantile y
p
x
is
usually of interest (a one-sided c size condence limit). Relative to
such a limit, we may say that we are (1 c) condent that the
true 100p-quantile y
p
x
is greater than this limit. This is equivalent
to say that we are (1 c) condent that (at the most) a fraction p
of all future observations of the sampled population will not ex-
ceed this limit or equivalently that (at least) a fraction (1 p)
of all future observations of the sampled population will exceed
this limit. When the latter formulation is used, this limit is often
called a lower tolerance limit.
Although (under the model hypothesis) the only correct way to
dene a tolerance interval is via Eq. (7), in the practice different or
approximate intervals are often used. As an example, practitioners
often refer to a design curve obtained shifting the estimated
median SN curve in the logarithm coordinates to the left by two
or three times the estimated standard deviation. This approach,
however, being based on point estimates, suffers of the obvious
M. Guida, F. Penta / Structural Safety 32 (2010) 6476 65
limitation that it does not take into account the sample variability
which, in case of small or moderate samples, may even be the pre-
dominant part of the experimental uncertainty. Also, the approxi-
mate Owen one-side tolerance limit [7] has been proposed to
account for the uncertainty in regression analysis.
Prediction interval on a future value of the rv Y: The rv
(Y ^ l
x
)=s
Y^ lx
, where s
Y^ lx
= s

1 1=n x
2
=

n
i=1
x
2
i
_
, has a Student
distribution with m = n 2 df, so that the (1 c) two-sided equal
tailed prediction interval is
^ l
x
st
1c=2;n2

1 1=n x
2

n
i=1
x
2
i
_

_
(8)
Fig. 1 depicts the aforementioned classical lower limits. Note
that the usual representation of SN tests assigns the horizontal axis
to the rv Y. Coherently with this setting, the proper regression anal-
ysis is to minimize horizontal distances.
We recall for future use, that it is common practice among prac-
titioners to write the deterministic link equation between the load
amplitude and the number of cycles to failure in the following
parametric form
S = S
/
f
(2N)
b
or S=S
u
= (S
/
f
=S
u
)(2N)
b
(9)
where 2N is the number of reversals, S
u
is the ultimate tensile
strength and S
/
f
is a material parameter, known as the fatigue
strength coefcient, which represents the hypothetical fatigue
strength value which one would obtain by extrapolating the SN line
at one reversal. Then, material fatigue properties can also be set in
terms of the parameters a = log(S
/
f
=S
u
) and b which are related to
the parameters a and b in (3) by
a = C
2
b
1
(a C
1
)
b = b
1
_
(10)
where C
1
= log2 and C
2
= (1=n)

n
i=1
log(S
i
=S
u
) is the mean of the
logarithm of the relative load levels that will be used in the fatigue
testing.
Finally, we recall that it is often of interest to explore the behav-
iour of steel alloys at load levels close to the fatigue limit, that is the
cut-off point at which the SN curve changes to a nearly horizontal
line. In fact, for many components the fatigue limit is a design cri-
terion. In such a case, the straight line may be a poor approxima-
tion and S-shaped SN curves might be considered. Also, the
assumption of constant standard deviation of the random log-life-
time at all stress levels is untrue. Hence, statistical approaches dif-
ferent from the one described above should be considered. It is to
be stressed, however, that data samples with a relatively limited
number of specimens, which are often common at the beginning
of the development phase of new systems or components, cannot
be used for the purpose of determining the whole shape of a SN
curve. In fact, in those cases it is hardly possible to estimate even
the slope of the high cycle fatigue straight line portion of the fati-
gue curve, with an acceptable accuracy [8].
3. Bayesian measures of uncertainty
Let Y be a rv with pdf p(y[H), indexed by a vector of parameters
H. Given a random sample of observations y = y
1
; :::; y
n
, Bayes-
ian inference on H is obtained via Bayes rule
p(H[y) =
p(y[H)p(H)
_
H
p(y[H)p(H)dH
; (11)
where p(H) is the personal degree of belief that a coherent person
has about H before observing the sample y, p(y[H) is the likelihood
function of the observed data, and p(H[y) is the updated personal
degree of belief about H after having observed the sample y.
Since a coherent persons degree of belief satises the Kolmogo-
rovs axioms of probability theory, we may think of p(H) and
p(H[y) as (subjective) probability density functions. In particular,
p(H) is the prior pdf and p(H[y) is the posterior pdf of H. The
denominator in (11) is simply a normalizing factor which ensures
that, over the support of H, the posterior pdf integrates to one.
Inference on H can be made on the basis of posterior credibility
regions, that is regions over the support of H, with a given posterior
probability content. Inference on the scalar components of vector
H can be obtained via marginal posterior pdfs. For such compo-
nents both (1 c) two-sided equal tailed probability intervals or
(1 c) highest posterior density (HPD) intervals can be derived.
Point estimates of scalar components of H are often taken to be
the expected value of the corresponding marginal posterior pdf.
This is because the posterior mean
^
h = E(h[y) minimizes the mean
quadratic loss function E[(h
^
h)
2
[.
3.1. Noninformative Bayesian inference for the simple linear model
Different kind of prior functions can arise depending on the de-
gree of initial personal knowledge about model parameters. In par-
ticular a noninformative (or reference) prior assumes that very
little is known about H . Although there have been various points
of view about how to express the notion of knowing little, it is
commonly recognised that a noninformative prior must at least en-
sure a consistent inference under simple parameters transforma-
tions. Noninformative priors are often improper, that is they do
not integrate to one. This is usually accepted, provided that the
posterior is a proper density. The regression model (3) is indexed
by the parameters vector H = [abr[. For this model a and b are loca-
tion parameters, while r is a scale parameter. Assuming indepen-
dence among parameters information, the noninformative
Jeffreys prior (which is consistent under parameters transforma-
tions) is [9]
p(a; b; r) 1=r a; bR; rR

(12)
which is an improper prior.
When using prior (12), it can be shown (see, as an example, [9])
that the (1 c) two-sided noninformative credibility intervals for
l
x
, y
p
x
and for a future observation Y, numerically coincide with
the corresponding classical intervals (6)(8), respectively. Recall,
however, that credibility intervals and condence intervals are
quite different inferential objects, in that a (1 c) credibility level
has to be thought of as a nal measure of precision (the precision
1 2 3
log
10
(N)
l
o
g
1
0
(
S
)
4
1 - Estimated median life;
2 - 90% LCL for median life;
3 - 90% LPL for future life;
4 - 90% LCL for the 5%-quantile
of lifetimes distribution;
Fig. 1. Data points, estimated SN line and classical lower limits
66 M. Guida, F. Penta / Structural Safety 32 (2009) 6476
after the experiment is run), whereas a (1 c) condence level has
to be thought of as an initial measure of precision (the precision
of the generating rule before the experiment is even run). This
might appear as just a philosophical difference with no conse-
quence from a practical point of view, thus considering not so
appealing the use of noninformative Bayesian inferential frame-
work. It is to be noted, however, that, even from a practical point
of view, noninformative inference should be considered as an use-
ful tool when censored sampling is involved. Indeed, in presence of
Type I censoring, classical inferential procedures fail to be applied
whereas Bayesian approach does not suffer from the experimental
context.
3.2. Informative Bayesian inference for the simple linear model
Informative Bayesian inference assumes that a subject is able to
express his/her personal knowledge about an unknown quantity H
in a quantitative manner. The available information is formalized
in terms of a prior pdf p(H), which encapsulates all relevant
knowledge one has about H. For model (3), p(H) = p(a; b; r). Such
a pdf could be based upon some or all of the following:
(1) engineering, physics, etc. information,
(2) mathematical or physical models,
(3) experts judgements,
(4) corporate memory, commercial databases, historical data.
After having observed a random sample of (transformed) num-
ber of cycles to failure y = y
1
; :::; y
n
at (transformed) loads
x = x
1
; :::; x
n
, one can combine the prior p(a; b; r) with the
likelihood
p(y; x[a; b; r) =
1
(2p)
n=2
r
n
exp
1
2r
2

n
i=1
(y
i
a bx
i
)
2
_ _
; (13)
thus obtaining the posterior density
p(a; b; r[y; x) =
r
n
D
exp
1
2r
2

n
i=1
(y
i
a bx
i
)
2
_ _
p(a; b; r)
(14)
where D is the normalizing factor given by
D =
_ _ _
r
n
exp
1
2r
2

n
i1
(y
i
a bx
i
)
2
_ _
p(a; b; r)da db dr:
The marginal pdfs of a, b and r can be obtained integrating over
the other variables. Moreover, by changing variables in (14), the
posterior pdf of l
x
and y
p
x
can be derived. Also, by combining
(14) with the likelihood of a future observation Y and integrating
over the parameters space, the predictive pdf of Y can be obtained.
In general a triple integration will be required to obtain a 100p-
quantile of the above quantities and the use of simulation methods
(such as Monte Carlo Markov Chain) could be considered.
In this paper we will assume that estimates of the material
parameters of steel alloys belonging to the same steel family of
the one to be tested are available and empirical relationships can
possibly be established between these parameters or functions
thereof (see the next section for details). Thus, when formulating
a prior pdf based on this information, model parameters a and b
need usually to be viewed as correlated rvs, while parameter r
can usually be considered as a rv a priori independent of a and b.
Then, the joint prior can be factored as p(a; b; r) p(a; b)p(r),
p(a; b) being the joint prior of a and b, and p(r) the prior for r.
Moreover, the available information about r is often more vague
than that about a and b. Then, we will assume
p(r) 1=r r
L
< r < r
U
(15)
since this pdf appears to be a very exible one, while keeping sim-
ple mathematics. In fact, when r
L
0 and r
U
this pdf con-
verges to the noninformative Jeffreys prior, whereas it can
conveniently describe less vague knowledge about r by simply
adjusting the interval (r
L
; r
U
) in a suitable manner.Then, the prior
pdf results in
p(a; b; r) p(a; b)=r a A; b B; r R (16)
where A, B and R are suitable subsets of R
By combining the prior (16) with the likelihood (13), the poster-
ior (14) results in
p(a; b; r[y; x) =
r
(n1)
D
exp
1
2r
2

n
i=1
(y
i
a bx
i
)
2
_ _
p(a; b)
(17)
The normalizing constant D is given by
D = 2
(n2)=2
_
A
_
B
g(a; b)
n=2
G[g(a; b); R[p(a; b)dadb
where g(a; b) =

n
i=1
(y
i
a bx
i
)
2
and
G[g(a; b); R[ =
C(n=2) when R = (0; )
IG
n
2
;
g(a;b)
2r
2
L
_ _
IG
n
2
;
g(a;b)
2r
2
U
_ _
when R = (r
L
; r
U
)
_
_
_
with IG(m; f) =
_
f
0
z
m1
e
z
dz (the incomplete gamma function).
On the basis of this prior pdf, the following posterior pdfs are
obtained:
Joint posterior pdf of a and b: Integrating (17) over r, the joint
posterior pdf of a and b is obtained
p(a; b[y; x) = 2
(n2)=2
D
1
g(a; b)
n=2
G[g(a; b); R[p(a; b): (18)
Posterior pdf of l
x
: By making the change of variable a = l
x
bx
in (17) and integrating over b and r, the posterior pdf of l
x
is
obtained
p(l
x
[y; x) = 2
(n2)=2
D
1
_
B
g(l
x
; b)
n=2
G[g(l
x
; b); R[p(l
x
; b)db (19)
where g(l
x
; b) =

n
i=1
[y
i
l
x
b(x
i
x)[
2
and
G[g(l
x
; b); R[ =
C(n=2) when R = (0; )
IG
n
2
;
g(l
x
;b)
2r
2
L
_ _
IG
n
2
;
g(l
x
;b)
2r
2
U
_ _
when R = (r
L
; r
U
)
_
_
_
Posterior pdf of y
p
x
: By making the change of variable
r = (y
p
x
a bx)=z
p
in (17) and integrating over a and b, the pos-
terior pdf of y
p
x
is obtained
p(y
p
x
[y; x) = D
1
[z
p
[
1
_
A
_
B
g(y
p
x
; a; b)p(a; b)dadb (20)
where g(y
p
x
; a; b) = [z
p
=(y
p
x
abx)[
n1
exp
1
2
[z
p
= (y
p
x
abx)[
2

n
i=1
(y
i
abx
i
)
2
. The integration domain in (20) is
A =
(y
p
x
bx r
U
z
p
; y
p
x
bx r
L
z
p
) when z
p
>0
(y
p
x
bx r
L
z
p
; y
p
x
bx r
U
z
p
) when z
p
<0
_
B =(; )
Predictive pdf of a future value of Y: By combining (17) with the
likelihood of a future observation and integrating over the param-
eters space, the predictive pdf of Y is derived
p(y[y; x) =2
(n1)=2
D
1
_
A
_
B
[g(y; a; b)[
(n1)=2
G[g(y; a; b); R[p(a; b)dadb
(21)
where g(y; a; b) = (y a bx)
2

n
i=1
(y
i
a bx
i
)
2
and
M. Guida, F. Penta / Structural Safety 32 (2010) 6476 67
G[g(y; a; b); R[ =
C[(n1)=2[ when R=(0; )
IG
n1
2
;
g(y;a;b)
2r
2
L
_ _
IG
n1
2
;
g(y;a;b)
2r
2
U
_ _
when R=(r
L
; r
U
)
_
_
_
4. Empirical knowledge on material parameters
Many studies have been addressed in the past to establishing
empirical relations between static and fatigue properties of steel,
aluminium and other alloys. Most of them dealt with data gener-
ated under completely reversed strain cycling. Indeed, besides
the traditional method based on nominal stresses and Basquins
equation, a more recent approach for the prediction of crack initi-
ation or nucleation life is now commonly used, which is based on
data obtained by strain controlled fatigue tests. The fundamental
relation, often referred to as the MansonCofn equation, relates
the reversals to failure, 2N
f
, to the total strain amplitude De=2:
De
2
=
De
e
2

De
p
2
=
S
/
f
E
(2N
f
)
b
e
/
f
(2N
f
)
c
(22)
where De
e
=2 and De
p
=2 are the elastic and plastic strain amplitude,
respectively, and E is the Young modulus. In Eq. (22), two additional
parameters with respect to Eq. (9) are needed, which dene the
plastic strain amplitude, namely the fatigue ductility coefcient e
/
f
and the fatigue ductility exponent c.
Recall that, in the region of high cycles fatigue, i.e. N
f
P10
5
,
De
p
De De
e
. Moreover, the way the test is controlled has little
effect on the lifetimes [10], so that S
N
EDe=2 EDe
e
=2, and the
estimates obtained by strain controlled tests are also valid for load
controlled tests.
The main empirical formulae that have been proposed for esti-
mating the parameters S
/
f
and b [1118] are listed in Table 1, where
r
f
is the true fracture stress, e
f
is the ductility or fracture elonga-
tion, n
/
is the hardening exponent and HB is the Brinnel hardness.
Even if the various proposed relations are not always in a close
agreement, most of them assume that a correlation exists between
the fatigue coefcient S
/
f
and the ultimate tensile strength S
u
.
Experimental results also seem to show that, for several types of
steel alloys, the quantities b and log(S
/
f
=S
u
) are correlated, too.
In particular, it is worth mentioning the recent paper [18],
where a sample of 845 different structural alloys is analysed. The
authors conclude that, on the average, steels present signicantly
higher b exponent than aluminium and titanium alloys, therefore
different estimates of this parameter should be considered for each
alloy family. Furthermore, correlation between b and tensile prop-
erties resulted poor, while the fatigue strength coefcient S
/
f
pre-
sented a fair correlation with the ultimate tensile strength. Also
in [15], some years before, it was recognised the importance of sep-
arating the estimation method by alloy families. These results,
along with the fact that tensile characteristics of many kinds of
steels show a correlation with several metallurgical factors which
affect the fatigue behaviour, motivated us to analyse the fatigue
properties by steel families, dened on the basis of common pri-
mary alloy elements and thermal treatment.
Some empirical data from the literature regarding steel alloys
[19] are reported and analysed in the following for an illustrative
purpose. In particular, data referring to some hot-rolled (HR) and
quenched and tempered (QT) carbon steels are presented in
Fig. 2. From this gure it can be inferred that, for HR and QT steel
alloys, log S
/
f
and log S
u
as well as b and log(S
/
f
=S
u
) appear to be
highly correlated quantities. Moreover, the point estimates of
material parameters of HR and QT carbon steels appear to form
well separated clusters on the graphs in Fig. 2. Data referring to
some low alloys steel (AISI 41xx) are also given in Fig. 3 and corre-
lations among material parameters are analysed. Even if, for these
alloys, logS
/
f
and logS
u
appear to be highly correlated, nevertheless
b and log(S
/
f
=S
u
) appear to be not. Finally, data referring to some
stainless steels SUS 3xx-B are given in Fig. 4: in this case it appears
that log S
/
f
and log S
u
are uncorrelated, whereas b and log(S
/
f
=S
u
)
seem to be correlated.
These empirical observations suggest to consider HR, QT, AISI
41xx and SUS 3xx-B steels as different steel families, in order
to increase the estimation accuracy when empirical regression
lines based on a steel population are used to predict the values
of S
/
f
and b, once the ultimate tensile strength of the material is gi-
ven. The observed variability in the material parameters is poten-
tially due to (at least) two non deterministic factors: (a) the
randomness of their estimates due to nite sampling, and (b) an
inherent random variation of material parameters over the popula-
tion of steels, which would determine a departure from a mean
trend even in absence of estimation error. In order to detect the
nature of the underlying variability, specic tests should be carried
out in which the estimation error is made negligible by using large
sample sizes. It is to be stressed, however, that the present paper is
not focused on proposing or establishing relationships between
material parameters. Data in Figs. 24 are simply reported with
the aim of suggesting that a lot of information about material
parameters is often available, which is totally ignored by conven-
tional estimation procedures. In the framework of the Bayes para-
digm instead, this information, when properly formalized, could be
used in conjunction with test data in order to render more efcient
the inference on quantities of interest. To this end, some Bayesian
procedures will be developed and discussed in the next section.
5. Formalizing the empirical knowledge on material parameters
as a prior pdf
Assume that fatigue tests have to be run in order to estimate the
material parameters b and S
/
f
of a certain steel alloy with a known
ultimate tensile strength S
u
. Let h denote the data set of the esti-
mates of these material parameters observed in a population of
steel alloys characterized by the same primary alloy elements
and thermal treatment as the steel to be tested, which we call
hereinafter the reference population for this steel. In the frame-
work of Bayes inference, the quantities b and a = log(S
/
f
=S
u
) are rvs
and their joint pdf, given h and S
u
, say p(a; b[h; S
u
), can be factored
as p(a; b[h; S
u
) = f (b[a; h; S
u
)f (a[h; S
u
).
Without loss of generality, assume that data pertaining to the
reference population, which the steel under investigation belongs
to, support the hypothesis that a linear relationship exists between
b and a. Then, it is reasonable to model the prior uncertainty
about the parameter b of the steel to be tested (given a and S
u
)
Table 1
Empirical estimation methods for Basquins formula parameters.
Estimation method S
/
f
b
Morrow [11] n
/
=(1 5n
/
)
Mansons universal
slopes [12]
1:9Su 0.12
Mansons four points
[12]
1:25r
f
2
b
r
f
Su(1 e
f
)
log(0:36r
f
=Su)
1=5:6
Mitchell et al. [13] Su 345 (MPa)
log[(Su 345)=(0:5Su)[
1=6
Muralidharan and
Manson [14]
0:623E(Su=E)
0:832
0.09
Baumel and Seeger
[15]
1:5Su 0.087
Ong [16] Su(1 e
f
)
log[(6:25r
f
=E)=(Su=E)
0:81
[
1=6
Roessle and Fatemi
[17]
4:25HB 225 (MPa) 0.09
Meggiolaro and Castro
[18]
1:5Su 0.09
68 M. Guida, F. Penta / Structural Safety 32 (2009) 6476
Fig. 2. Carbon steels: (a) plot of logS
/
f
vs log Su; (b) plot of b vs log(S
/
f
=Su).
2.9 3.0 3.1 3.2 3.3 3.4 3.5
2.6
2.8
3.0
3.2
3.4
3.6
regression line ( ):
y = 0.7803 x + 0.7808 (R
2
=0.773)
l
o
g
(
S
'f
)
log(S
u
)
-0.05 0.00 0.05 0.10 0.15 0.20
0.00
0.04
0.08
0.12
0.16
0.20

log(S
'
f
/ S
u
)
regression line ( ):
y = 0.1423 x + 0.0729 (R
2
= 0.408)
b a
Fig. 3. AISI 41xx steels: (a) plot of log S
/
f
vs log Su; (b) plot of b vs log(S
/
f
=Su).
b a
0.10
0.15
0.20
0.25
0.30

log(S
'
f
/ S
u
)
regression line ( ):
y = 0.2328 x + 0.0666 (R
2
=0.617)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
2.65 2.70 2.75 2.80 2.85 2.90
2.6
2.8
3.0
3.2
3.4
3.6
l
o
g
(
S
'f
)
log(S
u
)
regression line ( ):
y = - 0.1764 x +3.7489 (R
2
=0.0036)
Fig. 4. SUS 3xx-B steels: (a) plot of log S
/
f
vs logSu; (b) plot of b vs log(S
/
f
=Su).
M. Guida, F. Penta / Structural Safety 32 (2010) 6476 69
by assuming that f (b[a; h; S
u
) is the Bayes predictive pdf for the
simple (Normal) linear regression model b
i
= A
1
A
2
a
i
r
b
e
i
(i = 1; :::; m), under the noninformative prior (12) on the parame-
ters A
1
, A
2
and r
b
. This predictive pdf implies that the rv
(b ^ l
b[a
)=[s

c(a)
_
[ is Student with m = m2 df, where m is the
number of steel alloys in the reference population,
^ l
b[a
=
^
A
1

^
A
2
a,
^
A
1
and
^
A
2
being the estimated coefcients of the
regression line, and
s
b
=

m
i=1
(b
i

^
A
1

^
A
2
a
i
)
2
m2

;
c(a) = 1
1
m

(a

a)
2

m
i=1
(a
i

a)
2
with

a =

m
i=1
a
i
=m:
Thus, the proposed conditional prior pdf of b, given a and S
u
, is
f (b[a; h; S
u
) =
C

m2
_
^ r
b[a
1
b ^ l
b[a

m2
_
^ r
b[a
_ _
2
_
_
_
_
_
_
(m1)=2
(23)
where C = C[(m1)=2[=C(1=2)C[(m2)=2[ and ^ r
b[a
= s
b

c(a)
_
.
Now, analyse logS
/
f
and log S
u
and assume that data in the refer-
ence population support the hypothesis that a linear relationship
exists between these variables, too. Then, following the same
above-mentioned arguments, it can be assumed that, given the
ultimate tensile strength of the material to be tested, the prior
knowledge on the rv c = log S
/
f
can be expressed by the Bayes pre-
dictive pdf for the simple (Normal) linear regression model
c
i
= B
1
B
2
log S
i
u
r
c
e
i
(i = 1; :::; m), under the noninformative
prior on the parameters B
1
, B
2
and r
c
.
Thus, the conditional prior pdf of c, given S
u
, is
f (c[h; S
u
) =
C

m2
_
^ r
c[Su
1
c ^ l
c[Su

m2
_
^ r
c[Su
_ _
2
_
_
_
_
_
_
(m1)=2
where ^ l
c[Su
=
^
B
1

^
B
2
logS
u
,
^
B
1
and
^
B
2
being the estimated coef-
cients of the regression line,
^ r
c[Su
= s
c

c(S
u
)
_
; s
c
=

m
i=1
(c
i

^
B
1

^
B
2
log S
i
u
)
2
m2

;
and
c(S
u
) = 1
1
m

(log S
u
log S
u
)
2

n
i=1
(log S
i
u
log S
u
)
2
; with log S
u
=

m
i=1
log S
i
u
=m
Then, given S
u
, it readily follows that the conditional pdf of the
rv a = log(S
/
f
=S
u
) = c log(S
u
) is
f (a[h; S
u
) =
C

m2
_
^ r
a[Su
1
a ^ l
a[Su

m2
_
^ r
a[Su
_ _
2
_
_
_
_
_
_
(m1)=2
(24)
where ^ l
a[Su
= ^ l
c
logS
u
=
^
B
1
(1
^
B
2
) logS
u
, and ^ r
a[Su
= ^ r
c[Su
=
s
c

c(S
u
)
_
.
By combining (23) and (24), the joint prior pdf of a and b, given
h and S
u
, is obtained
p(a; b[h; S
u
) = f (b[a; h; S
u
)f (a[h; S
u
)
=
C
2
(m2)^ r
a[Su
^ r
b[a
1
b ^ l
b[a

m2
_
^ r
b[a
_ _
2
_
_
_
_
_
_
_
_
1
a ^ l
a[Su

m2
_
^ r
a[Su
_ _
2
_
_
_
_
_
_
_
_
(m1)=2
(25)
whose parameters can be estimated from data pertaining to the
reference population of steel alloys the one to be tested belongs
to.
As an example, the prior knowledge (25) on the material
parameters a and b of a new HR steel alloy with ultimate tensile
strength S
u
= 350 MPa or S
u
= 1000 MPa, is depicted in Fig. 5a.
The joint prior pdf (25) was derived under the assumption that
both the couples of variables (b; a) and (log S
/
f
; log S
u
) are linearly
dependent, as historical data suggest that it happens for HR and
QT steel alloys. This prior pdf, however, can be easily adapted to
the situation where one (or even both) of the above-mentioned
couples of variables are uncorrelated.
In particular, when b and a are uncorrelated, as for example it
seems to happen for AISI 41xx steel alloys (see Fig. 3), the
predictive pdf of b implies that the rv (b ^ l
b
)=^ r
b
is Student with
m = m1 df, where ^ l
b
=

m
i=1
b
i
=m =

b, and ^ r
b
=

1 1=m
_

m
i=1
(b
i

b)
2
=(m1)
_
.
0.10
0.30
0.60
0.90
0.10
0.30
0.60
0.90
0.04
0.08
0.12
0.16
0.20 a b

S
u
=1000 MPa
S
u
=350 MPa
0.1
0.3
0.6
0.1
0.3
0.5
0.0 0.1 0.2 0.3 0.4 0.5 3 4 5 6 7 8 9
5.0
7.5
10.0
12.5
15.0
17.5
0
.9
0
.9 b
a
0.7
S
u
=1000 MPa
S
u
=350 MPa
Fig. 5. Prior joint degree of belief on a and b (a) and on a and b (b) for HR steels with Su = 350 MPa and Su = 1000 MPa.
70 M. Guida, F. Penta / Structural Safety 32 (2009) 6476
Hence
f (b[h; S
u
) =
C[m=2[
C(1=2)C[(m1)=2[

m1
_
^ r
b
1
b ^ l
b

m1
_
^ r
b
_ _
2
_
_
_
_
_
_
m=2
(26)
Instead, when log S
/
f
and log S
u
are uncorrelated, as for example
it seems to happen for SUS 3xx-B steel alloys (see Fig. 4), the prior
pdf of a is
f (a[h; S
u
) =
C[m=2[
C(1=2)C[(m1)=2[

m1
_
^ r
a
1
b ^ l
a[Su

m1
_
^ r
a
_ _
2
_ _
m=2
(27)
where ^ l
a[Su
=

m
i=1
c
i
=mlogS
u
= c log S
u
, c
i
being the log fatigue
strength coefcient of the ith steel in the reference population, and
^ r
a
=

1 1=m
_

m
i=1
(c
i
c)
2
=(m1)
_
.
The parameters a and b are related to the parameters a and b of
model (3) by Eq. (10). For xed values of a and b, the equations sys-
tem (10) has a unique solution. Moreover, the Jacobian of the
transformation is b
3
. Then, by the transformation theorem, the
pdf of a and b can be converted into the pdf of a and b. It results
p(a; b[h; S
u
) =
C
2
b
3
(m2)^ r
a[Su
^ r
b[a
1
b
1
^ l
b[a

m2
_
^ r
b[a
_ _
2
_
_
_
_
_
_
_
_
1
b
1
(a C
1
) C
2
^ l
a[Su

m2
_
^ r
a[Su
_ _
2
_
_
_
_
_
_
_
_
(m1)=2
(28)
where ^ l
a[Su
= ^ l
a[Su
, ^ r
a[Su
= ^ r
a[Su
, ^ l
b[a
=
^
A
1

^
A
2
[b
1
(a C
1
) C
2
[ and
^ r
b[a
= s
b

1
1
m

[b
1
(a C
1
) C
2
a[
2

m
i=1
(a
i
a)
2

_
:
With reference to HR steels with ultimate tensile strengths
S
u
= 350 MPa and S
u
= 1000 MPa, respectively, in Fig. 5b the prior
knowledge about a and b, derived from the reference population, is
presented, when the k = 3 relative loads S
i
=S
u
= 0:35; 0:375; 0:40
are used in the fatigue testing.
Expressions for the prior pdf analogous to (27) or (28) can be
derived by using one of (or even both) Eqs (28) and (27) instead
of (23) and (24), when it is the case.
6. Comparison of classical and Bayesian estimation methods
When choosing among different inferential procedures, opti-
mality criteria are usually adopted. This is possible, however, only
within the same inferential framework. As already observed, in-
stead, classical and Bayesian inference represent quite different
points of view about knowing from experience. Thus, in principle
it makes no sense trying to compare these two approaches. Never-
theless, it is a common practice to carry out simulation studies in
order to analyse the behaviour of Bayesian procedures in repeated
sampling. Although philosophically questionable, this approach ap-
pears to be the only available tool to compare classical and Bayes-
ian methods.
To this end, for a given set of (transformed) load amplitudes x
i
(i = 1; :::; k), constant amplitude fatigue tests are simulated where,
for each x
i
, m (transformed) number of cycles to failure Y
i
are ob-
tained by sampling m pseudo-random values from a Normal distri-
bution with mean l
x
i
= a bx
i
and variance r
2
both known. Thus, a
sample of pseudo-random observations of size n = k m is ob-
tained and Bayesian point and interval estimates of the regression
model parameters and functions thereof are calculated, under a gi-
ven prior pdf. By repeating this procedure a large number N
S
of
times (keeping xed the simulation context and the prior pdf),
the sampling properties of Bayesian estimators (i.e., bias, mean-
squared error, covering percentage) can be empirically evaluated
and compared with the homogeneous precision measures of the
corresponding classical estimators.
In particular, we recall that for regression model (3), exact values
of bias and mean-squared error of classical point estimators of a, b
and y
p
x
(for any value of p) are available from statistical theory. In
fact, ^ a,
^
b and s
2
are unbiased minimum variance estimators with
known sampling distributions. They also are stochastically inde-
pendent rvs. These theoretical results are also useful for calibrating
the size N
S
of simulated tests to be carried out for the comparative
analysis. In the present case, it was found that N
S
= 2000 simulated
tests ensure a good agreement between exact and empirical results
both in terms of mean-squared errors and covering percentages.
As to the estimation of the material parameters b and S
/
f
in the
classical framework, we recall that they are known functions of the
model parameters a and b. Hence, maximum likelihood estimators,
say
^
b and
^
S
/
f
, of these quantities are easily obtained from ^ a and
^
b,
due to the invariance property of the maximum likelihood estima-
tion principle. The sampling distributions of
^
b and
^
S
/
f
are unknown,
however, and bias and variance of these estimators have to be
empirically calculated by simulation.
The following precision measures are used when comparing
point estimators: the normalized bias (NB), dened as the ratio of
the bias of the estimator to the true value of the parameter; and
the relative efciency (RE), dened as the ratio of the root mean-
squared error (rmse) of the classical estimator to the rmse of the
Bayesian estimator.
Comparison between the 90% Bayesian lower credibility limit
and the 90% lower condence limit on the 100p-quantile is given
in terms of: the relative mean distance (RMD) dened as the ratio
of the mean distance of the classical lower limit from the true
quantile to the corresponding mean distance of the Bayesian lower
limit; the relative standard deviation (RSDV) dened as the ratio of
the standard deviation of the classical lower limit to the standard
deviation of the Bayesian limit; and the covering percentage (CP),
dened as the fraction of times the Bayesian limit is less than or
equal to the true 100p-quantile, in repeated sampling.
7. Analysis of the simulation study
A large Monte Carlo study was carried out to assess Bayes esti-
mators performances with respect to classical ones, when chang-
ing: number of specimens on test, true values of regression
model parameters and prior information.
In particular, fatigue tests at k = 3 different constant amplitude
load levels were simulated. The values 0.35, 0.375 and 0.40 were
selected for the ratio S
i
=S
u
(i = 1; 2; 3). Then, for each load level,
m pseudo-random determinations of the number of cycles to fail-
ure were generated by assuming known values of the material
properties (S
u
; S
/
f
; b) which in turn determine known values of
the regression model parameters a and b and a known value of
the regression model paramerer r. Letting m vary from 2 to 5, fa-
tigue tests with sample sizes n = 6, 9, 12 and 15 were considered.
The pseudo-random lifetimes were generated from two hypo-
thetical steel alloys, namely: a HR steel with S
u
= 1000 MPa,
S
/
f
= 1300 MPa and b = 0:085 (HR1000, for short) and a SUS steel
with S
u
= 535 MPa, S
/
f
= 1870 MPa and b = 0:195 (SUS535). Both
HR1000 and SUS535 have material properties very close to that
of some real steels included in the respective reference population.
Also, the material properties are close to the corresponding esti-
mated values from the respective reference population.
M. Guida, F. Penta / Structural Safety 32 (2010) 6476 71
As to the regression model parameter r, a consolidated empir-
ical evidence of fatigue testing on steels (see, for example, [20])
indicates that the variation coefcient of the lifetimes ranges from
0.3 to 0.5, which implies the standard deviation of the log-lifetimes
to range from 0.13 to 0.20. Thus, the central value of this interval
was assumed for r in the simulation study.
The prior pdf dened by Eq. (28) was assumed for modelling the
uncertainty about the model parameters a and b for HR1000. For
SUS535, instead, the prior pdf was obtained by combining Eq.
(23) with Eq. (27), to take into account that for SUS 3xx-B steel
log S
/
f
and logS
u
appear to be uncorrelated. Given the reference
population of steels, both pdfs are completely dened by the S
u
va-
lue pertaining to the steel alloy on test, only. No indications were
found in the literature of a correlation between r and material
parameters, so that an independent prior was used to model the
uncertainty about the parameter r. In particular, two different
prior were used, namely: the prior p(r) 1=r over R

(i.e., the
noninformative Jeffreys prior) and the same prior p(r) 1=r over
the nite interval 0.130.20, which is an informative prior.
7.1. Point estimation of model and material parameters
In Table 2, classical and Bayes (under the noninformative prior
on r) point estimation of parameters a, b, S
/
f
and b is analysed,
when simulating from HR1000 and SUS535 steel. It is immediately
evident that Bayes estimators of parameters related to the slope
of the regression line are exceedingly better than classical ones,
achieving rmses which are up to hundreds times smaller. The most
critical situation for classical estimation particularly occurs in case
of steels with a high b value. In fact, when simulating from SUS535
(b = 0:195), near to zero, possibly negative, estimates of b were ob-
served in a fraction of over 20% of the simulated tests, and it was
not even possible to evaluate the empirical mean of the classical
point estimator of S
/
f
. Instead, even in case of samples where clas-
sical estimation fails, Bayes method provides accurate estimates.
Both methods show the same efciency in the estimation of a,
however.
This can be explained as it follows. Recall that under the fully
noninformative Bayes approach, i.e. under the Jeffreys prior
p(a; b; r) 1=r, Bayes point and interval estimates of the regres-
sion model parameters numerically coincide with the correspond-
ing classical estimates. This is because the quantities

n
_
(a ^ a)=s
and

n
i=1
x
i
_
(b
^
b)=s are distributed as Student rvs with (n 2)
df under both Bayes and classical framework, although with a quite
different meaning of the quantities involved. Thus, the joint nonin-
formative posterior is a bivariate Student pdf with mode at (^ a,
^
b),
and (for the cases being treated here) with an uncertainty on a
much smaller than that on b. At the same time, this pdf, when
centered on the true a and b values, also describes the sam-
pling uctuation of ^ a and
^
b estimators. When using the prior
p(a; b; r) p(a; b)=r, it can be readily shown that the joint poster-
ior (18) can be factored as the product of the joint noninformative
posterior times the prior p(a; b). In particular, the priors for
HR1000 and SUS535 steels, derived from their reference popula-
tions in Figs. 2 and 4 and used for the simulation study, are approx-
imately centred on the true material parameters, with an
uncertainty on a (on b) greater (smaller) than that conveyed by
the noninformative posterior. In repeated sampling, the noninfor-
mative posterior mode presents small uctuations around the
prior mean of a and dominates the prior in terms of this variable
so that the informative posterior tends to have the mode close to
^ a. Instead, the noninformative posterior presents large uctuations
around the prior mean of b and it is dominated by the prior in
terms of this variable so that the informative posterior tends to
have the mode close to the prior mean of b. This explains why, in
repeated sampling, the Bayes point estimate of a practically has
the same properties of ^ a, whereas the Bayes point estimate of b
has dramatically better properties than
^
b, in terms of mean-
squared error.
In Table 3 classical and Bayes point estimation of the 50% (med-
ian) and the 5%-quantile of the log-lifetime distribution (usually
known as the R50 and the R95 quantile, respectively) is analysed,
for HR1000 and SUS535. For both steels, Bayes point estimates
show very similar properties. In term of bias, Bayes estimators be-
have as classical ones. Moreover, under the noninformative prior
on r, the Bayes point estimator appears to be more efcient than
the classical one in correspondence of low and high load levels,
showing root-mean-squared errors from about 45% to 75% lower
for the R50 quantile, and from about 10% to 20% for the R95
quantile.
For both steels, under the informative prior p(r) 1=r over the
interval 0.130.20, only a slight improvement was observed in
Bayes point estimation of a; b; S
/
f
; b and of the median log-life,
whereas the point estimator of the 5%-quantile of the log-life dis-
tribution showed to be exceedingly more efcient than classical
one in correspondence of all the load levels, with rmses from
about 50% to 75% lower in correspondence of the intermediate
load, and from about 75% to 125% in correspondence of the two ex-
treme loads.
7.2. Lower tolerance limit estimation
In Table 4, for HR1000 and SUS535 steels, the 90% lower toler-
ance limit and the Bayes 90% lower credibility limit on the 50% and
on the 5%-quantile of the log-lifetimes distribution (usually known
as R50C90 and R95C90, respectively) are analysed.
It appears that, for both steels, Bayes limits greatly outperform
classical ones. In fact, under the noninformative prior on r, in cor-
respondence of the low and the high load level, the Bayes
R50C90 limit is (in the mean) closer to the population median from
38% to 56% than the classical one for HR1000 (47% to 50% for
SUS535); and the Bayes R95C90 limit is (in the mean) closer to
the 5%-quantile from about 21% to 31% for HR1000 (25% to 38%
for SUS535), depending on the sample size. Further improvements
Table 2
HR1000 and SUS535 noninformative prior on r: comparison of point estimates of regression model and material parameters (subscript c for classical, b for Bayes).
Steel Sample size Parameter estimates
a b S
/
f
b
NB
c
NB
b
RE NB
c
NB
b
RE NB
c
NB
b
RE NB
c
NB
b
RE
HR1000 6 0 0.000 1.02 0 0.003 9.30 0.534 0.015 234.53 0.066 0.009 30.13
9 0 0.000 1.01 0 0.002 6.99 0.126 0.014 16.98 0.044 0.008 8.71
12 0 0.000 1.01 0 0.002 5.68 0.071 0.014 8.64 0.029 0.008 6.46
15 0 0.000 1.01 0 0.002 4.95 0.048 0.014 6.18 0.020 0.008 5.32
SUS535 6 0 0.000 1.01 0 0.014 10.91 0.073 0.009
15 0 0.000 1.00 0 0.011 5.90 0.073 0.194 0.009 33.46
72 M. Guida, F. Penta / Structural Safety 32 (2009) 6476
are obtained under the informative prior on r, especially for the
R95C90 Bayes limit which is (in the mean) closer to the 5%-quan-
tile than the classical one from about 93% to 111% when n = 15 and
from about 188% to 214% when n = 6 for HR1000, and from about
101% to 108% when n = 15 and from about 202% to 214% when
n = 6 for SUS535. It is worth noting that these results are obtained
with a covering percentage that is even greater than 90%. The
above-mentioned results can be explained as it follows. Recall that
under the fully noninformative approach, Bayes and classical limits
numerically coincide. Hence, when we compare the two inferential
methods in repeated sampling, the properties of the two estima-
tors coincide, too. If one looks at Eq. (7), he/she immediately recog-
nises that the sampling uctuation of the classical lower limit
depends on the sampling uctuation of ^ l
x
(which, in turn, depends
on the uctuation of ^ a and
^
b) and on the sampling uctuation of s.
Roughly speaking, we may say that, when using an unbiased infor-
mative prior pdf on a and b it is as if, in a sense, we contrasted -
through this prior - the sample variability of the estimates of a
Table 3
HR1000 and SUS535 comparison of point estimation of 50%- and 5%-quantile (subscript c for classical, b for Bayes).
Steel Quantile Prior on r Sample size Quantile estimates
Low load Intermediate load High load
NB
c
NB
b
RE NB
c
NB
b
RE NB
c
NB
b
RE
HR1000 R50 Noninform 6 0 0.001 1.55 0 0.001 1.01 0 0.000 1.74
9 0 0.000 1.53 0 0.000 1.01 0 0.000 1.69
12 0 0.000 1.48 0 0.000 1.01 0 0.000 1.63
15 0 0.000 1.46 0 0.000 1.00 0 0.000 1.59
Informative 6 0 0.000 1.55 0 0.000 1.01 0 0.000 1.74
9 0 0.000 1.53 0 0.000 1.01 0 0.000 1.69
12 0 0.000 1.49 0 0.000 1.01 0 0.000 1.64
15 0 0.000 1.46 0 0.000 1.00 0 0.000 1.59
R95 Noninform 6 0.003 0.006 1.10 0.003 0.007 0.90 0.003 0.005 1.13
9 0.002 0.003 1.17 0.002 0.003 0.94 0.002 0.003 1.19
12 0.001 0.002 1.18 0.001 0.002 0.97 0.001 0.002 1.21
15 0.001 0.001 1.20 0.001 0.002 0.98 0.001 0.001 1.22
Informative 6 0.003 0.001 2.04 0.003 0.001 1.73 0.003 0.001 2.25
9 0.002 0.001 1.95 0.002 0.001 1.61 0.002 0.001 2.06
12 0.001 0.001 1.76 0.001 0.001 1.48 0.001 0.001 1.87
15 0.001 0.001 1.77 0.001 0.001 1.49 0.001 0.001 1.86
SUS535 R50 Noninform 6 0 0.000 1.59 0 0.001 1.02 0 0.001 1.71
15 0 0.000 1.50 0 0.000 1.01 0 0.000 1.58
Informative 6 0 0.000 1.59 0 0.001 1.01 0 0.001 1.71
15 0 0.000 1.50 0 0.000 1.01 0 0.001 1.58
R95 Noninform 6 0.003 0.008 1.12 0.003 0.009 0.90 0.003 0.010 1.13
15 0.001 0.002 1.22 0.001 0.002 0.98 0.001 0.003 1.21
Informative 6 0.003 0.001 2.09 0.003 0.001 1.74 0.003 0.001 2.22
15 0.001 0.002 1.80 0.001 0.001 1.49 0.001 0.001 1.86
Table 4
HR1000 and SUS535 comparison of 90% lower tolerance limit estimation on the 50%- and 5%-quantile.
Steel Quantile Prior on r Sample size 90% Lower tolerance limit estimates
Low load Intermediate load High load
RMD RSDV CP RMD RSDV CP RMD RSDV CP
HR1000 R50 Noninform 6 1.42 1.60 92.6 1.00 1.04 91.2 1.56 1.73 92.0
9 1.41 1.58 92.6 1.00 1.04 90.8 1.53 1.69 92.5
12 1.38 1.54 92.6 1.01 1.02 90.3 1.49 1.65 92.6
15 1.38 1.54 92.4 1.02 1.05 90.2 1.50 1.67 92.9
Informative 6 1.61 1.76 91.5 1.15 1.17 90.4 1.77 1.89 91.3
9 1.52 1.65 92.2 1.09 1.07 90.3 1.66 1.77 92.0
12 1.44 1.57 92.7 1.06 1.05 90.7 1.58 1.70 92.1
15 1.42 1.58 92.4 1.08 1.07 90.1 1.55 1.67 92.5
R95 Noninform 6 1.27 1.29 90.8 1.15 1.15 90.1 1.31 1.34 90.1
9 1.22 1.25 90.7 1.08 1.07 90.0 1.26 1.29 90.4
12 1.21 1.24 90.4 1.06 1.04 89.3 1.26 1.27 90.2
15 1.21 1.23 90.4 1.06 1.05 89.7 1.26 1.29 90.6
Informative 6 2.88 3.27 92.4 2.77 3.00 91.4 3.14 3.55 92.2
9 2.32 2.64 92.9 2.20 2.36 91.5 2.51 2.87 92.8
12 2.08 2.36 92.8 1.94 2.08 91.2 2.24 2.53 92.8
15 1.93 2.16 92.5 1.84 1.91 91.0 2.11 2.35 92.5
SUS535 R50 Noninform 6 1.50 1.63 92.2 1.00 1.05 91.3 1.50 1.71 92.2
15 1.47 1.58 91.5 1.02 1.03 90.6 1.43 1.63 93.4
Informative 6 1.70 1.81 91.3 1.14 1.15 90.6 1.70 1.91 91.9
15 1.54 1.57 91.5 1.07 1.03 89.9 1.50 1.61 93.2
R95 Noninform 6 1.33 1.37 90.7 1.19 1.24 90.2 1.38 1.48 90.7
15 1.25 1.26 90.0 1.08 1.08 89.7 1.34 1.54 90.9
Informative 6 3.03 3.33 91.8 2.76 3.03 91.4 3.02 3.78 92.4
15 2.08 2.24 91.7 1.82 1.92 91.0 2.01 2.35 93.0
M. Guida, F. Penta / Structural Safety 32 (2010) 6476 73
and b, and when adding information on r it is as if we contrasted
the sample variability of the estimate of r. In particular, it is the
information on r which mainly contributes to reduce the uncer-
tainty on y
p
, by dramatically reducing the long left tail which char-
acterizes the two posterior pdfs obtained under a noninformative
prior and an informative prior on a and b, only. Obviously, when
the sample size increases, these effects tend to decrease, though
to an extent which is difcult to quantify in advance, however.
A different very impressive way to measure the global efciency
of the Bayes estimation method is in terms of specimens saved
with respect to the classical method. As an example, it results that
Bayes estimation of the R95C90 limit based upon samples of size
n = 6 have a precision, both in terms of mean distance from the
true quantile and in terms of standard deviation, that classical esti-
mation attains with samples of size n = 12 or n = 24, depending
upon whether prior information on r is available or not. Thus,
Bayes estimation of the R95C90 design curve results in a testing
time and cost which is 2 or 4 times smaller than in case of classical
estimation, respectively. In Fig. 6, for HR1000 steel, Bayes and clas-
sical average (90% condence) design curves for the 5%-quantile of
the lifetimes distribution are presented, for the sample sizes
n = 6; 9; 15. The average performances of the two estimation meth-
ods and the effect of using an informative prior on r against a non-
informative one are immediately evident.
7.3. Robustness analysis
It is to be noted that previous results are obtained with a prior
knowledge about model parameters which is centered on the true
values, i.e. the values used to generate the pseudo-random deter-
minations of the number of cycles to failure. This is in keeping with
the hypothesis that observed departures of material properties
from the regression lines characterizing reference population of
steels, essentially depend on estimation errors. Since, however, it
cannot be excluded that an inherent randomeffect is possibly pres-
ent, it is necessary to test the proposed Bayesian approach in case
when the prior pdf is not centered on the true values of the mate-
rial parameters. To this end, we analysed a situation where the
prior pdf is chosen through Eq. (28), but the number of cycles to
failure are generated from a hypothetical steel, whose material
parameters values are far from the values that reference popula-
tion regression lines would assign to it, on the basis of its ultimate
tensile strength.
In particular, we chose the worst case actually observed in the
populations of both HR and QT steels. This is represented by the
QT steel with (rounded) material parameters S
u
= 1300 MPa,
S
/
f
= 1800 MPa and b = 0:090 (QT 1300, for short), to which given
S
u
the regression lines instead assign parameter values
S
/
f
= 1482 MPa and b = 0:066. Then, the case where the prior is
centered on S
/
f
= 1482 MPa and b = 0:066 while fatigue tests are
simulated from S
/
f
= 1800 MPa and b = 0:090 is analysed.
For this worst case, in Tables 5 and 6, a comparison of classical
and Bayes point estimates (under both the noninformative and the
informative prior on r) is presented.
Unlike the centered prior situation, Bayes point estimators of S
/
f
and b parameters related to the slope of the SN regression line
(Table 5) appear to be moderately biased, with a normalized bias
ranging from 7.5% to 10%. It is to be noted that, except the classical
point estimator of b which is unbiased, the corresponding classical
point estimators have similar or even much greater biases. The
rmses of the Bayes point estimators, however, are much smaller
than that of the classical ones. Moreover, Bayes estimation of
parameter a is not affected by the prior, showing the same ef-
ciency as classical estimation.
2.52
2.54
2.56
2.58
2.60
2.62
5.00 5.25 5.50 5.75 6.00 6.25
n=9
n=6
n
=
1
5
n
=
9
n
=
6
n
=
6
n
=
1
5
n
=
9
n
=
1
5
R95
classical design curves
bayesian design curves (non informative prior on )
bayesian design curve (informative prior on )
log
10
(N)
l
o
g
1
0
(
S
)
Fig. 6. HR1000: classical and Bayes average R95C90 design curves in case of both noninformative and informative prior on r, for sample sizes n = 6; 9; 15.
Table 5
QT 1300 comparison of point estimates of regression model and material parameters (subscript c for classical, b for Bayes).
Prior on r Sample size Material parameters estimates
a b S
/
f
b
NB
c
NB
b
RE NB
c
NB
b
RE NB
c
NB
b
RE NB
c
NB
b
RE
Noninform 6 0 0.001 1.01 0 0.099 2.61 4.86 0.100 1485.6 0.093 0.084 6.04
15 0 0.000 1.01 0 0.087 1.71 0.059 0.090 2.95 0.023 0.075 2.19
Informative 6 0 0.000 1.01 0 0.097 2.66 4.86 0.100 1502.3 0.093 0.083 6.13
15 0 0.000 1.01 0 0.086 1.73 0.059 0.089 2.98 0.023 0.074 2.21
74 M. Guida, F. Penta / Structural Safety 32 (2009) 6476
From Table 6 it also appears that Bayes point estimators of 50%
and 5%-quantile are practically unbiased and with quite better rel-
ative efciencies, especially for the extreme loads and when the
informative prior on r is used. Hence, the Bayes point estimation
approach appears to be quite robust with respect to the departure
of the material properties from the reference population regression
lines.
In Table 7, Bayes 90% lower credibility limits on the 50% and the
5%-quantile of the log-lifetimes distribution are analysed (under
both the noninformative and informative prior for r). Bayes limits
result much closer (in the mean) to the corresponding true values
than classical ones. Also, their standard deviations are quite smal-
ler than that of classical estimates. The covering percentages, how-
ever, suffer of a distorting effect, in the sense that for the high
load the covering level is sensibly greater than 90%, whereas for
the low load the covering level is about 80%. Nevertheless, in
the whole the Bayes method appears to be quite robust even in a
such unfavourable situation.
8. Conclusions
A Bayesian analysis of SN fatigue data has been presented for
estimating material properties and for establishing fatigue design
curves from small size samples.
Posterior distributions for the linear regression model parame-
ters and function thereof have been derived on the basis of prior
technological knowledge available on steel alloys with primary al-
loy elements and thermal treatment analogous to the one to be
tested. Both the case when only prior information on regression
model parameters is considered and the case when information
on the log-lifetimes variance is also available, have been analysed.
Bayes procedures have been compared against conventional
ones by using Monte Carlo simulation. It was found that the pro-
posed Bayes approach largely outperforms the classical one for
both point and interval estimation. In particular, exceedingly bet-
ter performances were observed for Bayes point estimation of
material parameters related to the slope of the regression line
and for Bayes estimation of R95C90 lower limit. In general, it
was observed that the presence of an informative prior on the
log-lifetime variance signicantly increases the performance of
Bayes estimators. For example, Bayes estimation of R95C90 curve
from a sample of size n = 6 is as efcient as classical estimation
is in case of n = 12 or n = 24 specimens on test, depending upon
whether prior information on r is available or not. Thus, efciency
being equal, Bayes estimation of a R95C90 design curve results in a
testing time and cost which is 2 or 4 times smaller, respectively.
For the application of the proposed Bayes estimation approach,
an ad hoc mathematical software is needed, however, which re-
quires a once and for all effort. Computational methods involved
essentially are multidimensional quadrature (up to the third order)
and the search for a zero of a real function. The computing time re-
quired for the complete analysis of a fatigue lifetimes data set is in
the order of a few minutes on a personal computer. Thus, when
repetitive analyses are required, computing costs appear to be
insignicant compared with the costs of obtaining the data.
It is also worth noting that the Bayes estimation procedure has
shown to be fairly robust with respect to the formulation of a
biased prior information on material parameters.
Therefore, the proposed Bayesian approach appears to be a very
interesting alternative to the conventional procedures for estimat-
ing material fatigue properties and lower bounds of design
quantiles.
References
[1] OHagan A, Forster J. Kendalls advanced theory of statistics: Bayesian
inference. London: Arnold; 2004.
[2] Madsen HO. Bayesian fatigue life prediction. In: Eddertz S, Lind NC, editors.
Probabilistic methods in the mechanics of solids and structures. Stockholm:
Proc IUTAM Symp; 1984. p. 395406.
Table 6
QT 1300 comparison of 50% and 5%-quantile point estimates (subscript c for classical, b for Bayes).
Quantile Prior on r Sample size Quantile estimates
Low load Intermediate load High load
NB
c
NB
b
RE NB
c
NB
b
RE NB
c
NB
b
RE
R50 Noninform 6 0 0.006 1.35 0 0.001 1.01 0 0.005 1.64
15 0 0.005 1.21 0 0.000 1.01 0 0.005 1.37
Informative 6 0 0.005 1.38 0 0.000 1.02 0 0.005 1.62
15 0 0.005 1.22 0 0.000 1.01 0 0.005 1.37
R95 Noninform 6 0.003 0.001 1.16 0.003 0.006 0.93 0.003 0.012 1.07
15 0.001 0.003 1.17 0.001 0.002 0.98 0.001 0.007 1.08
Informative 6 0.003 0.007 1.76 0.003 0.001 1.74 0.003 0.004 2.19
15 0.001 0.006 1.43 0.001 0.001 1.50 0.001 0.004 1.75
Table 7
QT 1300 comparison of 90% lower tolerance limit estimation on the 50% and 5%-quantile.
Quantile Prior on r Sample size 90% Lower tolerance limit estimates
Low load Intermediate load High load
RMD RSDV CP RMD RSDV CP RMD RSDV CP
R50 Noninform 6 2.20 1.62 82.5 2.80 1.09 90.3 2.37 1.85 96.6
15 2.46 1.51 79.1 2.00 1.04 90.6 1.55 1.71 98.2
Informative 6 2.48 1.73 81.2 3.11 1.15 89.7 2.56 2.01 96.5
15 2.60 1.50 78.2 2.11 1.04 90.1 1.59 1.68 98.0
R95 Noninform 6 1.56 1.35 86.7 1.19 1.20 90.4 1.17 1.39 93.6
15 1.61 1.25 83.5 1.05 1.05 89.8 0.97 1.29 95.8
Informative 6 4.41 3.19 82.1 2.90 3.04 90.6 2.43 3.71 96.8
15 3.29 2.14 79.8 1.84 1.93 91.3 1.47 2.45 98.2
M. Guida, F. Penta / Structural Safety 32 (2010) 6476 75
[3] Edwards G, Pacheco LA. A Bayesian method for establishing fatigue design
curves. Struct Saf 1984;2:2738.
[4] Basquin OH. The exponential law of endurance tests. ASTM 1910;10:62530.
[5] Draper NR, Smith H. Applied regression analysis. New York: Wiley; 1966.
[6] Owen DB. A survey of properties and applications of the noncentral t-
distribution. Technometrics 1968;10:44572.
[7] Shen CL, Wirshing PH, Cashman GT. Design curve to characterize fatigue
strength. ASME J Eng Mater Technol 1996;118:53541.
[8] Spindel JE, Haibach E. Some considerations in the statistical determination of
the shape of SN curves. In: Little RE, Ekvall JC, editors. Statistical analysis of
fatigue data, ASTM STP 744; 1981. p. 89113.
[9] Box GEP, Tiao GC. Bayesian inference in statistical analysis. Reading: Addison-
Wesley; 1973.
[10] Ruiz C, Koenigsberger F. Design for strength and production. Macmillan; 1970.
[11] Morrow JD. Cyclic plastic strain energy and fatigue of metals. Internal friction,
damping and cyclic plasticity ASTM STP 738. Philadelphia (PA): American
Society for Testing and Materials; 1964. p. 4587.
[12] Manson SS. Fatigue: a complex subject some simple approximations. Exp
Mech J Soc Exp Stress Anal 1965;5(7):193226.
[13] Mitchell MR, Socie DF, Cauleld EM. Fundamentals of modern fatigue analysis.
Fracture Control Program Report No. 26. University of Illinois; 1977, p. 385
410.
[14] Muralidharan U, Manson SS. Modied universal slopes equation for estimation
of fatigue characteristics. ASME J Eng Mater Technol 1988;110:558.
[15] Baumel Jr A, Seeger T. Materials data for cyclic loading supplement
I. Amsterdam: Elsevier; 1990.
[16] Ong JH. An improved technique for the prediction of axial fatigue life from
tensile data. Int J Fatigue 1993;15(3):2139.
[17] Roessle ML, Fatemi A. Strain-controlled fatigue properties of steels and some
simple approximations. Int J Fatigue 2000;22:495511.
[18] Meggiolaro MA, Castro JTP. Statistical evaluation of strain-life fatigue crack
initiation predictions. Int J Fatigue 2004;26:46376.
[19] Mitchell MR. Parameters for estimating fatigue life. In: Lampman S, editor.
Fatigue and fracture. ASM handbook, vol. 19. ASM International; 1996. p. 963
79.
[20] Wirsching PH. Probabilistic fatigue analysis. In: Sundararajan C, editor.
Probabilistic structural mechanics handbook (Theory and industrial
applications). Chapman & Hall ITP; 1995. p. 14665.
76 M. Guida, F. Penta / Structural Safety 32 (2009) 6476

You might also like