You are on page 1of 9

Int. J, Fatigue Vol, 17, No. 5, pp.

365-373, 1995
Copyright 1995 Elsevier Science Limited
Printed in Great Britain. All rights reserved
0142-1123/95/$10.00

TTERWORTH
I

tq

~"

Detailed evaluation of methods for estimation


of fatigue properties
Jun-Hyub Park and Ji-Ho Song
Department of Automation and Design Engineering, Korea Advanced Institute of
Science and Technology, 207-43, Cheongry angri-dong, Dongdaemun-gu, Seoul,
Korea
(Received 7 November 1994)
Using extensive experimental strain-life curve data on 116 steels, 16 aluminium alloys and six titanium
alloys, nearly all methods currently available for estimation of fatigue properties from simple tensile
data are discussed in detail. The four-point correlation method, the universal slopes method, Mitchell's
method, the modified universal slopes method, the uniform material law by Bgumel and Seeger. and
the modified four-point correlation method by Ong are evaluated in a quantitative manner by using
new criteria proposed in this work, along with conventional error criterion. The modified universal
slopes, Seeger's and Ong's methods give good life predictions. Among them, the modified universal
slopes method provides best results.
(Keywords: fatigue properties; fatigue life prediction; estimation method; simple tensile data; evaluation criteria)

Fatigue analysis is very important in the design


of mechanical structures and components. Fatigue
properties of materials are essential for fatigue analysis.
Fatigue properties such as stress-life ( S - N ) or
strain-life (e-N) curves are usually obtained by
performing fatigue tests. However, as fatigue testing
requires a lot of time and effort, there have been
many attempts to estimate fatigue properties from
simple tensile data.
Manson ~ first proposed two methods, the four-point
correlation method and universal slopes method, to
estimate the strain-life curve composed of the plastic
and elastic lines on a log-log scale, using only
tensile data. Mitchell-~ has proposed another method,
particularly suitable for steels. Recently, B~iumel and
Seeger. 3 have proposed a new method, the uniform
material law. In contrast, Ong 4 has proposed a modified
four-point correlation method.
The universal slopes method, in which the slopes of
the plastic and elastic lines are universalized as -0.6
and -0.12 respectively for all materials, has been
widely used for its simplicity and ease of application.
However, the method tends to produce overconservative estimates of life in the very high cycle life range.
To improve the original universal slopes method,
Muralidharan and Manson 5 have proposed a new,
modified universal slopes method.
The assumption that the slope of the plastic
strain-life relationship (the plastic line) on a log-log
plot is almost constant for all materials is relatively
well accepted. However, the slope of the elastic
strain-life relationship (the elastic line) has been
frequently observed to vary over a relatively wide
range and to be material-dependent. Considering this

material dependence, Mitchell proposed a method to


estimate the slope of the elastic line from the tensile
strength of material.
The uniform material law proposed by Baumel and
Seeger. has been derived from a large amount of
fatigue data collected by them. This method may be
said to be a kind of universal slopes method, which
assigns different slopes to unalloyed and low-alloy steels
and to aluminium and titanium alloys respectively. One
advantage of the method is that only the tensile
strength of the material is needed for estimation of
the strain-life curve, in contrast to other methods,
which also require the data of the reduction in area
or the fracture ductility of the material. Using the
fatigue data collected by them, they checked the
prediction capability of the uniform material law and
the modified universal slopes method. They found
that both methods showed larger deviations between
the predicted and experimental results for aluminium
and titanium alloys and for high-alloy steels, compared
with unalloyed and low-alloy steels.
Quite recently, Ong 6 has evaluated three commonly
used methods, the four-point correlation method, the
original universal slopes method and Mitchell's method
(or the method proposed by Socie et al.), and found
that the four-point correlation method and the original
universal slopes method give satisfactory results, while
Mitchell's method gives inferior predictions. Moreover,
he proposed a new, modified four-point correlation
method to improve the original four-point correlation
method. However, the data used for the evaluation
were not actual experimental data but calculated values
from the fatigue properties of materials provided in
the A S M Metals Handbook 7.

365

366

Jun-Hyub Park and Ji-Ho Song

Though many useful estimation methods have been


proposed as described above, there is little work to
evaluate them extensively, particularly using actual
experimentally obtained data on a variety of materials.
The importance of fatigue data has been emphasized
recently in several remarkable fatigue data books,
containing a large amount of experimental fatigue data
for a broad range of materials, published in Japan
and Germany.
In this work, using the actual fatigue data available
in fatigue data books, the predictive accuracy of all
the estimation methods described above is discussed.
In general, the accuracy of the estimation methods
has been evaluated in terms of the error or deviation
of the predicted values from the actual ones. However,
it seems that the error criterion on its own is not
always sufficient for detailed discussion of estimation
methods. Some additional criteria are proposed in this
work to examine in more detail the predictive accuracy
of estimation methods. As the fatigue data used
here for discussion of estimation methods are almost
independent from the data from which the estimation
methods have been derived, the conclusions obtained
in this work may be unbiased and highly reliable.

value of elastic strain at Nf = 1 0 4 cycles on the elastic


line.
As the true fracture stress ~rf is not always given in
the literature, Manson ~ recommended the following
approximation suggested by J. O'Brien:
trf = ~r~(1 + El)

(2)

Original universal slopes method by Manson ~


In this method, the coefficients C~, b, Cp, c of
Equation (1) are as given in the following equation:
Ae = Ae~ + Aep = 3.5 E N ( .'2 + o-f.~N~,~'

Method by MitchelF
In this method, the strain-life relation is expressed
using the strain amplitude Ae/2 and the number of
reversals 2Nf as
A~ = Aee + Aep
2
2
2
!

: tyf (2Nf)h + e~ (2Nf) c


E
M E T H O D S F O R E S T I M A T I O N OF F A T I G U E
P R O P E R T I E S F R O M TENSILE D A T A
The estimation methods to be discussed in this work
are briefly explained below.

Original four-point correlation method by Manson 1


The strain-life relation is usually expressed as the
sum of the elastic and plastic lines on a log-log scale
as follows:
A ~ = A~. e "~ A~.p =

C~N } + CpN'f"

--

(4)

where ~} is the fatigue strength coefficient, b is the


fatigue strength exponent, e} is the fatigue ductility
coefficient and c is the fatigue ductility exponent. The
equation is represented graphically in Figure 2. 2N, in
the figure is the transition fatigue life at which the
elastic and plastic lines intersect. Mitchell suggested
that for steels with hardness below 500 BHN,
~ ~ ~rf

(1)

where Ae, A~ and A% are the total, elastic and plastic


strains respectively, and Nf is the number of cycles to
failure. In the original four-point correlation method,
the coefficients C~, b, Cp, c of Equation (1) are
calculated using the four points, P1-P4, located on
the elastic and plastic lines as shown in Figure 1.
Every point is determined from tensile data: the true
fracture stress ~rf, the tensile strength 0% and the
fracture ductility el. Ae* in the figure indicates the

(3)

au

50 (ksi)

(5)

= ~u + 345 (MPa)
!o
b =

1o
(6)

log(lO o) - log(lO 6)

i log 2(o'~ + 345)


= - ~ log
= -a
~'
\ trB /
[
O-B
J

~a

pl

~5

a%}cp.~

"0

Of

~'

e'~

,
Ae~ =

A
I~1

100

Figure 1

..L
4

101

10 2

~
p4~,

103

104

Number of cycles to failure, Nf

Four-point correlation method by Manson

I ~

l0 0

t
2Nt

I
106

107

Number of reversals to failure,~d',lf

10 5

Representation of strain amplitude versus number of


reversals to failure

Figure 2

367

Detailed evaluation of methods for estimation of fatigue properties

6 ~ ef = In 10

RA

where RA is the percentage reduction in area.


He further recommended constructing the plastic
line using an empirical representation of the hardness
versus transition fatigue life 2Nt, instead of using a
specific value of c. However, as the hardness-transition
life data are not always readily available, a constant
value of c = - 0 . 6 , which is regarded as an average
value of c for almost all materials, may be assigned
to the slope of the plastic line, for convenience.
The resulting equation will be
AtE __ A e c

9~-

The slopes of the elastic and plastic lines for steels,


-0.087 and -0.58, are very close to the values -0.09
and -0.56 in the modified universal slopes method.
However, the slope of the plastic line for aluminium
and titanium alloys, -0.69, is rather steep.

Modified four-point correlation method by Ong 4


To improve the original four-point correlation
method, this method uses the four points P I ' - P 4 '
shown in Figure 3, instead of the four points P1-P4
in Figure 1. The strain-life relation is given as
Ae
-2

Ae~
+ Aep
- __
2
2

-~ A ~ p

= uf (2Nf)b + ~f(2Nf)c
E

2
[-2( o-B + 3 4 5 )-]

(orB + 345)
-Ugl
=
E
(2Nf) 6 L

where
b

+ ~,(2N,) "~

(12)

(8)

= ~{log[ 016tt(r ) \'"8' ] -

This will be hereafter referred to as Mitchell's equation.

Modified universal slopes method by Muralidharan


and Manson 5
As was briefly mentioned in the introduction, this
method has been proposed to improve the original
universal slopes method, and expresses the strain-life
relation as
[ O.B \ 0 . 8 3 2

A e = 1.17[ /~)

U~-'9
/O"

+ 0.0266 ~

155

\-(I.53

I - EB )

N?56

(9)

This method provides lower universalized slopes for


the elastic and plastic lines than the original method.
The crB/E term is newly introduced into the coefficient
of the plastic line, indicating that the tensile strength
of the material also has a marked effect on the life in
the low-cycle fatigue.

Uniform material law by Biiumel and Seeger 3


In this method, two different expressions for strain
versus life are used for unalloyed and low-alloy steels
and for aluminium and titanium alloys as follows.
For unalloyed and low-alloy steels:
Ae

--

A~ e

log

2~074

- log(e,)
trl,-

crB 0.81

The value of uf is needed also in this method as in


the original four-point correlation method. Ong has
noted that the value of (re can be approximated by
Equation (2).
F A T I G U E D A T A F O R DISCUSSION OF
ESTIMATION M E T H O D S
The fatigue data used are strain-life data obtained
through fatigue testing, which are contained mainly in
the databooks, the JSME Data Book: Fatigue of Metals
IV 8 and Materials Data for Cyclic Loading by Seeger
and others 3'9. In particular, we have used the data
obtained under the so-called normal testing condition:
that is, on polished smooth or hourglass-type specimens
tested in air at room temperature under axial loading

-{- A E p

2
....

c =

pi

2
-E-\~,,f]

-b

0.59t~2Nf) -58

(lO)

where
O"B

--<~
E

.@

0.003,0= 1

O'BE> 0.003, 0 = 1.375 - 1 2 5 . 0 ~


For aluminium and titanium alloys:
Ae
2

Ae~
2

A%
2

= 11.ut~-Iz.,zvf!gc'lO'B['~--(I.095
31~!
q'- 0.35(2Nr) -'69

10

(11)

104

106

Number of reversals to failure,2Nf


Figure 3 Modifiedfour-point correlation method by Ong

368

Jun-Hyub Park and Ji-Ho Song

at a stress ratio of R,~ ( = O ' m a x / O ' m i n ) = --i or a strain


ratio of R~ ( = emaJer, i , ) = - - 1 . Run-out data are
excluded. Further, only the data on materials with
both data of tensile strength o-B and percent reduction
in area RA (or fracture ductility el) are utilized.
The data used are listed in Table 1, classified by
data sources and materials. The total number of e-N
curves is 315 and the total number of experimental
strain-life data points is 2717 obtained on 138 materials.
For convenience, each individual strain-life datum is
referred to a e - N data point, and a set of ~-N data
points forming an E-N curve is hereafter referred to
a (E-N) data set. To apply successfully the evaluation
criteria to be proposed later, data sets that consist of
less than four data points are not considered in
the discussion. When needed, the value of ~rf is
approximated by Equation (2).

tions, as shown in Figure 4, where Np and Ne are the


predicted and experimental lives respectively. The
values calculated by Equation (13) are all the same
for the three methods. However, as can be easily
found, method C can hardly explain the experimental
results and method B may be expected to give superior
predictions to method A for a wider range of life.
This drawback in the error criterion may be overcome
by considering the linear relationship between the
predicted and experimental lives in the following
manner.
If the relationship between the predicted and experimental lives is approximated by a least-squares line
as

log(2Np) = a + N o g ( 2 N , . )

10~

CRITERIA FOR EVALUATION OF


ESTIMATION METHODS

Method!

A )

A criterion most frequently used is the error criterion,


which evaluates the predictive accuracy of an estimation
method in terms of the fraction of data falling within
a scatterband of a specified factor s as follows:

(14)

MethodA /
/

/ ,/
McthodB/

.ili
Method C

//~/' NP = I'Nf

Number of data falling within


The experimental value ~< s
Total number of data

(13)

There is a problem inherent in the criterion,


as illustrated below. Assume that three different
estimation methods provide three different life predic-

Table 1

10

Figure 4

Experimental life, Nf

109

Illustration of drawback inherent in the error criterion

Data used for comparison of estimation methods


Data source

Materials

unit

Unalloyed
steels

Number of
E - N curves
Number of data
points

Low-alloy
steels

Number of
e - N curves
Number of
data points

High-alloy
steels

Number of
~ - N curves
Number of
data points

Aluminium
alloys

Number of
E-N curves
Number of
data points

Titanium
alloys

Total

Materials Data ]br


Cyclic Loading Vol
JSME Data Book."
Fatigue o f Metals IV ~ 1-4"

Materials' Data for


Cyclic Loading Suppl.
1'
S A E data"'

Total

35

28

72

51

317

234

6(12

13

8O

42

137

87

903

299

27

1316

23

36

68

133

3(13

69

505

12

1(1

29

100

76

64

(I

24(I

Number of
E - N curves
Number of
data points

54

54

Number of
E-N curves
Number of
data points

57

17(I

86

315

371

1653

666

27

2717

369

Detailed evaluation of methods for estimation of fatigue properties


it can be said that the closer the values of intercept a
and slope/3 are to 0 and 1 respectively, the better the
estimation method is. And if a is rather greater than
zero, the case when/3 is moderately smaller than unity
gives better predictions than the case when/3 is greater
than unity. For a rather less than zero, the case
when /3 is moderately greater than unity gives better
predictions than the case when/3 is smaller than unity.
Thus the values of a,/3 and (a + /3) can be utilized
to evaluate the predictive accuracy of an estimation
method from the viewpoint of agreement between the
predicted and experimental lives. The drawback above
described can be overcome by employing the values
of a, /3 and (a + /3) as additional criteria.
In using the additional criteria, the following possibility should be considered. Assume that Figure 5
shows the prediction results obtained by applying two
estimation methods, D and E, to two (e-N) data sets
I and II. That is, method D gives different leastsquares lines for (e-N) data sets [ and II, while
providing a least-squares line of a ~ 0 and/3 ~ 1 for
the combined data of data sets I and II, as shown in
Figure 5a. In contrast, method E gives nearly the
same least-squares lines of a ~ 0 and /3 ~- 1 for both
data sets I and II, and also for the combined data of
data sets I and II, as shown in Figure 5b. Limiting
attention to only the results obtained for the combined
data, methods D and E appear to be equally excellent
methods. However, it is evident from the results
obtained for individual (E-N) data sets that method
E is preferable.
The correlation coefficient between the predicted
and experimental values, r, will also be one other
additional criterion. When two different methods
happen to provide identical least-squares lines, an
estimation method that gives a higher r value is
preferable.

In this work, the predictive accuracy of estimation


methods will be evaluated in a quantitative manner
by using the evaluation measure defined below.

Fraction of data within a factor of s, E j s )


As noted above, the error criterion can be expressed
by Equation (13), i.e. fraction of data within a factor
of s, Ef(s). A value of s of 3 is employed for life
prediction. The evaluation value Er(s = 3) is given as

Ef(s

3)

(15)

Np
Number of data falling within 3_ <- Nr -< 3
1

Number of total data


The closer El(S) is to 1, the better the prediction.
Accordingly, all the evaluation measures to be here
employed will be formulated to be unity for ideally
good prediction.

Goodness of fit between the predicted and


experimental values, E.~
Using the additional criteria presented above, the
goodness of fit is defined for the combined data of all
(e-N) data sets and for individual (E-N) data sets.
separately as
(E.),,,,,,1 =

(l-

IOftotal [) -c (1 - -

l1 - -

/-~total I )

(1--[1-

O~total-

~[~tota I

-1

[)+(1--11--

r,,,,,[) (16)

4
/V

(E.)D~c, = N ,=~ (E~,),

N i=1
z.., [

~+ (1 -- 1] -- ~ ' - - ~'1)
4 + (1 -- [I - - ' " l ) ]
--t----m--

fore Ncurvel I
fore N c u r v e l l l
for curves (1+ Ill I

I1 ff
T,
1+I1
i//

/ /

(17)

where the subscripts, total and i, refer Io the combined


data of all (e-N) data sets and the ith (e-N) data set
respectively, and N is the number of (e-N) data sets.
(E,0Dset of Equation (17) represents the goodness of
fit for individual (e-N) data sets.
C O M P A R I S O N OF ESTIMATION M E T H O D S

i "2'
/

z~

a) Method D

If/I+II

e~

/
./ "A"

N,=j4"
/

b) Method E
Experimental life, Nf

Figure 5 Application of least-squares analysis to the predicted data


by two different estimation methods

Life predictions were performed with the aforementioned six estimation methods on the five material
groups listed in Table 1. Figure 6 shows a typical
result of life predictions obtained on low-alloy steels,
for which the largest amount of data is obtained among
the five material groups. In the figure, the perfect
correlation line and a factor of 3 scatter band are
expressed by the dashed and solid lines respectively.
The equation and the value of r in the figure represent
the result of least-square analysis and the correlation
coefficient for the combined data respectively.
As can be found from Figure 6a, the predicted data
by the four-point correlation method tend to level off
in the long-life range, resulting in over-conservative
life predictions. Such a trend is also discernible in the
result for the original universal slopes method shown
in Figure 6b. In contrast, the modified universal slopes
method, the uniform material law by B~iumel and

370

Jun-Hyub Park and Ji-Ho Song


109
a)

l0 s

109

Four-PointCorrelati,

10 s

107

Zt"q

10n

t"q

105

>

>

104
h3

107

0 0

106

o o~8

103

EF

o
;.~

103

10 101 102 103 104 105 106 I07 10s 109

Experimentalreversals,2N
109

Experimentalreversals,2Nf
10 9

c)ModifiedUniversal
SlopesMethod

10 8

106

t'-,I

105

103

102

d)Mitchell'sMethod

10 s

107

;~

r = 0.936
~ ' ' ~ ~ l )=0.858+0.781
g ( 2 Nlog(2Nr)

101
10

10 101 102 103 104 105 106 107 10s 109

104

102
0.893
Iog(2Np)=1.375+0.5241og(2Nr )

10

oJ 1

r =

10 ~

>

105
104

[]

10 2

Z
t"q

b) OriginalUniversalSlopes

ee m
engB ~no ~9

107
106

~ 1
~

104

103

102

I)

F
[]

101
I0 ~

r = 0.936
( )=0.100+0.9531og(2Nr)
2
N

I01 [,_ ~_) fro//

m ii

dim

r = 0.922
log(2Nv)=0.642+0.9271og(2Nr )

10
I0 l01 102 103 l04 105 106 107 108 109

]0

l01

Experimentalreversals,2Nf

10 3

10 4

10 5

10 6

10 7

l0 8

10 9

Experimentalreversals,2Nf

109

109

108

e) Seeger'sMethd ~~ ~ ~

107

108

If)Ong'sMethd

,
z~

105

~
. An

'@

10 o

~1~

1o4

ai

~>

105 !~-

~'

.~

103

lOj
10

r = 0.934
log(2Np)=0.330+0.9161og(2Nf)

~
l0

101

102

103

104

105

106

107

l08

ll010

109

10

D,

103

~
mull

~k

106

>

10 2

r = 0.935
Ig(2Np)=0"563+0"8621g(2NI)

~
lO 1

Experimentalreversals,2Nf
Figure 6 Comparison of the predicted and experimental fatigue lives for low-alloy steels

102

103

104

105

106

107

Experimentalreversals,2Nf

l0 x

109

Detailed evaluation of methods for estimation of fatigue properties


Seeger. and the modified four-point correlation method
by Ong provide relatively good life predictions. However, Mitchell's method tends to give considerably
non-conservative predictions in the longer-life range.
A close inspection of Figure 6 reveals that the
modified universal slopes method tends to give slightly
conservative predictions at shorter lives, but nonconservative predictions at long lives. A similar tendency can be found for Seeger's method. The predicted
data due to Ong's method are found on average to
fall slightly above the perfect correlation line.
The trend observed on low-alloy steels was observed
consistently on other materials. However, particularly
on aluminium alloys, a few excessively non-conservative predictions were observed at short lives for all
the methods except Mitchell's.
Unless otherwise stated, Figure 6 is hereafter
employed as representative of the results for all
materials, for convenience.
Table 2 shows comparisons of the predictive accuracy
of estimation methods on the basis of the evaluation
valu__es defined in the previous section. The values of
E, E and E in the table are given as
3

= k_7:t -

= Ef(s

= 3) + (Ea)total + (Ea)Dset

3
(18)

(Ek)iUi
Ek

-- -

j = l~

---

(19)

EEk
:

k=l

E l ( s = 3) +
=

371

+
total

_ Oset

3
(20)

where the subscripts j and k refer to the jth material


and kth evaluation value respectively, and Nj is the
number of (e-N) data sets of the jth material.
The bold-faced figures in the table represent the
best prediction for each evaluation item. First, giving
attention to the result on low-alloy steels along with
Figure 6, the evaluation value of fraction of data
E f ( s = 3 ) is nearly the same for the original and
modified universal slopes methods, Seeger's and Ong's
methods, among which Ong's method gives the best
predictions. With respect to the goodness of fit, the
modified universal slopes method is the best for the
combined data, while Seeger's is the best for individual
(E-N) data sets. It is interesting that the method giving
the best result differs depending on the evaluation item.
Considering the evaluation value averaged over three
items, E, the modified universal slopes method may
be said to be the best for low-alloy steels.
Based on the value of E, the modified universal
slopes method is also the best for unalloyed and highalloy steels. For unalloyed and high-alloy steels,
Seeger's method also gives good results, comparable
to the modified universal slopes method. On the other
hand, Mitchell's method is the best for aluminium
alloy, while Ong's is the best for titanium.

ENj

J=l

Table 2

Comparison of estimation methods in terms of evaluation values

Material
group
Unalloyed
steels

E value
Ef(s = 3)

(Ea)total
(E~)D~
E

Low-alloy
steels

Ef(s = 3)

High-alloy
steels

Er(s = 3)

Aluminium
alloys

Titanium
alloys

All
materials

Total

(E,)~o,,
(E~)Ds~,

(E,),,,~
(E~)D~e,
Ef(s = 3)
(E~)~,,~
(E,)Ds~,

Original
universal
slopes

Modified
universal
slopes

0.570
0.431
0.234
0.412

0.769
0.669
0.462
0.634

0.824

0.548
0.229
0.224
0.334
0.663
0.313
0.4549
0.477

Four-point
correlation

Mitchell's
method

Seeger's
method

Ong's
method

0.925
0.542
0.764

0.738
0.774
0.330
0.614

0.805
0.885
0.569
0.753

0.811
0.868
0.49l
0.723

0.750
0.444
0.503
0.566

0.764
0.747
0.621
0.711

0.532
0.528
0.514
0.524

0.768
0.654
0.632
0.684

0.793
0.562
0.614
0.656

0.865
0.648
0.645
0.719

0.867
0.956
0.656
0.826

0.60/)
0.808
0.605
0.671

0.846
0.970
0.662
0.826

0.865
0.801
0.681
0.782

0.800

0.713
0.723
0.712
0.716

0.567
-0.043
0.437
0.320

0.586
0.666
0.684

0.796
0.720
0.704
0.740

0.721
0.907
O.734
0.787

0.738
0.877

Ef(s = 3)
( E,),,,~
( E,)D~,

0.519
0.331
0.462
0.437

0.759
0.558
0.500
0.606

0.852
0.892
0.448
0.731

0.648
0.847
0.580
0.692

0.778
0.746
0.472
0.665

0.870
0.921

Ef(s = 3)
=
(E~)~o,~J
( E~ )1. . . .
~

0.576
0.271
0.303

0.780
0.556
0.539

0.801
0.834
0.613

0.609
0.688
0.514

0.788
0.798
0.629

0.805

0.384

0.627

0.751

0.606

0.739

(/.709

0.735

0.783

0.592
0.794
0.708
0.609

372

Jun-Hyub Park and Ji-Ho Song

In the evaluation averaged o v e r all material groups


and expressed by the value of E, almost the same
results are obtained as described above on low-alloy
steels,
In the total evaluation represented by the value of
~, the modified universal slopes method is best and
Seeger's method is next best, followed by Ong's.
DISCUSSION
From the results in the preceding section, the features
of each estimation method can be summarized as
follows.
The predictive accuracy of both the four-point
correlation method and the original universal slopes
method is considered inferior to the other methods,
because the prediction results from both methods tend
to level off in the long-life range, as noted in Figure
6. As can be found from the value of Ef(s = 3) in
Table 2, only about 58% of the data predicted by the
original four-point correlation method fall within a
factor of 3 scatterband, indicating that the method
may not be recommended.
The original universal slopes method may be satisfactory in the respect that about 80% of the data fall
within a factor of 3 scatterband, However, the goodness
o f fit due to th_e method is inferior, as the values of
(E~)tot~l and (E~)Dset in Table 2 indicate.
It is worth noting that the above two methods tend
to provide slightly non-conservative predictions at
short lives, but are excessively overconservative at
long lives. This fact has been already pointed out by
Manson and Muralidharan 5, particularly for the original
universal slopes method.
In contrast, the modified universal slopes method,
developed to overcome this drawback of the original
universal slopes method, tends to give slightly conservative predictions at shorter lives, but is non-conservative
at long lives, as already noted in Figure 6c. While the
fraction of data within a factor of 3 scatterband is
80%, as can be found in Table 2, the goodness of fit
is remarkably improved in the modified universal
slopes method. As the equation in Figure 6c shows,
the values of intercept (~ = 0.100) and slope
(/3 = 0.953) are very close to 0 and 1 respectively.
This result implies that the modified universal slopes
method gives accurate predictions.
As can be seen in Figure 6d, Mitchell's method also
provides a value of slope, /3 = 0.927, relatively close
to 1, indicating excellent direct correspondence
between the predicted and experimental lives. However, the value of the intercept, a = 0,64, is relatively
high, leading to non-conservative predictions over the
entire life range. Consequently, only about 60% of
the data fall within a factor of 3 scatterband as listed
in Table 2.
Good direct correspondence (/3 = 0.914) between
the predicted and experimental lives can be observed
also in Figure 6e for Seeger's method. In particular,
the method is the best in goodness of fit for individual
data sets, as can be found from the value of (E~)D~ct
in Table 2. The fraction of data falling within a factor
of 3 scatterband is 79%.
As can be seen in Figure 6f, Ong's method gives
the value of slope, /3 = 0.862, less than those for the
above three methods, and a relatively high intercept

value, c~ = 0,563. Consequently, Ong's method is


inferior to both the modified universal slopes method
and Seeger's method in the goodness of fit, and
tends to provide non-conservative predictions more
frequently, compared with the modified universal
slopes method. The fraction of data within a factor of
3 scatterband is slightly more than 80%.
It can be noted from the above discussion that the
modified universal slopes method, Seeger's method
and Ong's method all give reasonably good life
predictions, with about 80% of the data falling within
a factor of 3 scatterband. Based on the goodness of
fit, the modified universal slopes method is the best
and Seeger's method is the next best. While not shown
here explicitly, the modified universal slopes method
has the fewest non-conservative data falling outside
the upper line of a factor of 3 scatterband, in
comparison with other two methods.
In conclusion, the modified universal slopes method
can be recommended as the best estimation method
at present. This method essentially requires the elastic
modulus E, tensile strength o-B and fracture ductility
el, of the material. When the fracture ductility is not
available, Seeger's method, which requires only the
elastic modulus and tensile strength, may be utilized
as an alternative to obtain satisfactory results.
On the other hand, Ong's method, while being a
good estimation method, requires the true fracture
stress o-t,, which is not always given, in addition to the
above three material properties.
Although not shown here, the fraction of data falling
within a factor of 2 scatterband is found to be about
60% for all the above three methods. The major
conclusions obtained above for a factor of 3 scatterband
are hardly affected by employing a factor of 2.
It has already been noted in the introduction that,
on its own, the error criterion expressed by Equation
(15) is not always sufficient for detailed evaluation of
estimation methods, Based on the error criterion
alone, the original universal slopes method may be
said to be one of the best estimation methods, as the
fraction of data falling within a factor of 3 (or 2)
scatterband amounts to 78% (or 60%), comparable
to the results due to the above three methods,
However, the original universal slopes method is in
practice inferior to the modified universal slopes
method as reported by Manson and Muralidharan ~.
Such a fact can be easily clarified in a quantitative
manner by using the additional criteria proposed here,
as Table 2 shows.
CONCLUSION
The predictive accuracy of various methods developed
to estimate fatigue properties from simple tensile data
is quantitatively evaluated by using the additional
criteria proposed in this work. with the conventional
error criterion, The conclusions obtained are summarized as follows:
1. New criteria based on a least-squares analysis
between the predicted and experimental fatigue
lives are proposed to evaluate estimation methods.
2. The three methods - the modified universal slopes,
Seeger's and Ong's methods - give reasonably
good life predictions. Among them, the modified
universal slopes method provides the best results.

Detailed evaluation of methods for estimation of fatigue properties


3. Consequently, the modified universal slopes method
can be recommended as the best estimation method
at present. When the fracture ductility required by
the method is not available, Seeger's method may
be utilized as an alternative to obtain satisfactory
results.
4. The fraction of data falling within a factor of 3 (or
2) scatterband is found to be about 80% (or 60%)
for all the above three methods.
REFERENCES
1
2
3

Manson, S.S, Exp. Mech. 1965, 5, 193


Mitchell, M.R. "Fatigue and Microstructure', American
Society for Metals, Metals Pack, OH, 1979, p. 385
B~iumel, A. Jr and Seeger, T. ~Materials Data for Cyclic

4
5
6
7
8

10

373

Loading, Supplement 1' Elsevier Science Publishers, Amsterdam, 1990


Ong, J.H. Int. J. Fatigue 1993, 15, 213
Muralidharan, U. and Manson, S.S.J. Eng. Mater. Technol.
1988, 110, 55
Ong, 3.H. Int. J. Fatigue 1993, 15, 13
'ASM Metals Handbook', 9th edn, American Society for
Metals, Metals Park, OH, 1978, 680
"JSME Data Book: Fatigue of Metals', IV Low Cycle Fatigue
Strength, Japan Society of Mechanical Engineers, Tokyo,
1983
BOller, Chr, Jr and Seeger, T. 'Materials Data for Cyclic
Loading', Vol. l - 4 , Elsevier Science Publishers, Amsterdam,
1987
Wetzel, R.M. (ed.) 'Fatigue Under Complex Loading:
Analyses and Experiments', Society of Automotive Engineers, Inc., 1975 p. 41

You might also like