Professional Documents
Culture Documents
A B S T R A C T Based on the assumption that the fatigue life and endurance limit is mainly controlled by
the materials particle distribution, different possible statistical distributions to describe
the fatigue have been assessed. Some distributions lead to high statistical uncertainty
or physically non-realistic size effects. It is shown that, if the maximum particle size
distribution is described by a Frechet distribution, both the fatigue life, as well as the
endurance limit are described by a Weibull distribution.
Keywords endurance limit; extreme value distribution; fatigue life; frechet; gumbel;
statistics; weakest link; weibull.
c 2010 Blackwell Publishing Ltd. Fatigue Fract Engng Mater Struct 33, 333–344 333
334 K. WALLIN
FATIGUE LIFE
c 2010 Blackwell Publishing Ltd. Fatigue Fract Engng Mater Struct 33, 333–344
STATISTICAL ASPECTS OF FATIGUE LIFE AND ENDURANCE LIMIT 335
⎝ rlimit ⎠ 20
Equation (4), with a slight approximation, consisting of
0.4 deleting the constant 1, leads to a Type-II asymptotic
distribution for the maximum of the form of Eq. (9). This
0.2 is also known as the Frechet distribution.
m
r0
P (rmax ≤ rc ) = exp − . (9)
0.0 rc
0 1 2 3 4 5
rc/rave The physically promising feature of the Frechet distri-
bution is that it cannot have negative particle sizes. Both
Fig. 2 Some possible inclusion size distributions. the reversed Weibull as well as the Gumbel distribution
allow mathematically for negative particle sizes, which is
physically not possible. For an engineering application
distribution types that can be used to describe particle size this may not be significant, but it does undermine the
distributions. A comparison of them is shown in Fig. 2. theoretical justification for their use.
Equation (4) represents a simple form of an inverse power The three extreme value distributions can be trans-
law function, Eq. (5) is a simple form of a decaying power formed into one general form called the General Extreme
law function and Eq. (6) represents a simple form of a de- Value distribution (GEV) in the form of Eq. (10).5 If the
caying exponential function. Often, more complex forms parameter ξ is close to 0, it indicates a Gumbel distribu-
of the basic functions are used. For example Eq. (6) is in tion, if ξ is positive it indicates a Frechet distribution and
form similar to the log-normal distribution, which also is if ξ is negative it implies a reversed Weibull distribution.
often used to describe the particle size distribution.
rc − μ −1/ξ
−m P (rmax ≤ rc ) = exp − 1 + ξ · . (10)
rc σ
P (r ≥ rc ) = +1 , (4)
rref
The mean and standard deviation expressions for the
different distributions are as given in Table 1. The scat-
rc m
P (r ≥ rc ) = 1 − , (5) ter of the reversed Weibull distribution is proportional
rlimit
to the difference between the limiting size and the mean
rc size (Table 1). The scatter for the Frechet distribution
P (r ≥ rc ) = exp − . (6)
rref on the other hand is directly proportional to the mean
size (Table 1). The scatter for the Gumbel distribution is
The value 1 in Eq. (4), corrects basically for the resolu- proportional to the difference between the location pa-
tion of the particle size measurements. rameter and the mean size (Table 1). These differences
Equations (5) and (6) provide very similar results, with are important factors, when considering the effect of size
the exception that Eq. (5) contains an upper limiting par- on the distribution.
ticle size (r limit ). This may actually be physically more As an example, fits of the three distributions are com-
reasonable than the assumption of a possible infinite par- pared in Fig. 3 for three different steels. The figures show
ticle size as predicted by Eq. (6). Of course, some limit- the distributions of the maximum particle sizes, found in
ing particle size criterions can be introduced in the other a 5 × 5 mm2 area.6
functions as well. The data seem overall to be best described by the Gum-
Normally, functions similar to Eqs (5) and (6) are used bel distribution, but the reversed Weibull distributions
to estimate the maximum particle size distribution, since provide practically identical fits and also the Frechet dis-
they lead to well known extreme value distributions. tribution provides a comparatively good description of the
Eq. (5) leads to a Type-III asymptotic distribution for the data. In essence, any of the three distributions could be
maximum of the form of Eq. (7). This is also known as used on equal grounds.
the Weibull distribution for the maximum or the reversed The previously mentioned analysis was based on the
Weibull distribution. asymptotic extreme value distributions, corresponding to
rlimit − rc m a large number of volumes. The statistical nature of fa-
P (rmax ≤ rc ) = exp − . (7)
rlimit − r0 tigue life can also easily be derived based on the weakest
c 2010 Blackwell Publishing Ltd. Fatigue Fract Engng Mater Struct 33, 333–344
336 K. WALLIN
1.0 1.0
(a) (b)
Bearing steel A Bearing steel B
0.8 all inclusions 0.8 all inclusions
P{rmax≤√area}
P{rmax≤√area}
0.6 0.6
Maximum inclusion size Maximum inclusion size
2 2
within 5x5 mm area within 5x5 mm area
0.4 0.4
0.0 0.0
0 5 10 15 20 25 30 35 40 45 0 5 10 15 20 25
√area [μm] √area [μm]
1.0
(c)
Q&T steel C
0.8 all inclusions
P{rmax≤√area}
0.6
Maximum inclusion size
2
within 5x5 mm area
0.4
0.0
5 10 15 20 25
√area [μm]
Fig. 3 Example fits of the extreme value distributions for three steels, containing several different inclusion types.6
link theory, which is also applicable for smaller volumes. For small samples, the number of particles is not de-
This is presented next. fined by a single value. There is a specific probability of
finding a certain number of particles in a sample. For
a macroscopically uniform sample, this probability fol-
ESTIMATE BASED ON THE WEAKEST LINK lows the Poisson distribution and can be expressed as
Eq. (12), where n̄ is the average number of particles in the
If the sample contains n particles, the probability that the
sample.
fatigue life (Nf ) is longer than Nc is expressed in the form
of Eq. (11). The form of Eq. (11) comes from the fact
n̄n · exp(−n̄)
that if one or more particles in the sample are larger than Pn = . (12)
rc , Nc is not reached. Thus it is required that none of the n!
particles are larger than rc .
Equation (11) can thus be re-written in the form of
P (N f ≥ Nc ) = [1 − P (r ≥ rc )]n . (11) Eq. (13), which becomes as Eq. (14).
c 2010 Blackwell Publishing Ltd. Fatigue Fract Engng Mater Struct 33, 333–344
STATISTICAL ASPECTS OF FATIGUE LIFE AND ENDURANCE LIMIT 337
∞
and the distribution function P(x), then the probability
P (N f ≥ Nc ) = [1 − P (r ≥ rc )]n · Pn (13) function of Yr is given by Eq. (18).
n=0
n!
∞
{n̄ · [1 − P (r ≥ rc )]}n · e −n̄ p Yr (x) = · [P (x)]r−1
P (N f ≥ Nc ) = . (14) (r − 1)! · (n − r)!
n!
n=0 ·[1 − P (x)]n−r · p(x), (18)
Equation (14) looks complicated, but it can be sim-
smallest value ⇒ r = 1 p(xmin )
plified by making use of the exponential equation. By
definition, the exponential equation can be expressed as = n · [1 − P (x)]n−1 · p(x), (19)
Eq. (15).
∞
largest value ⇒ r = n p(xmax ) = n · P (x)n−1 · p(x).
xn (20)
ex = . (15)
n!
n=0 If x corresponds to probability, it represents a uniform
Inserting Eq. (15) into Eq. (14) yields the simple form of random variable. In this case P(x) = x (between 0 and
Eq. (16). 1) and p(x) = dx. The distribution for the largest value
becomes in this case simply.
P (N f ≥ Nc ) = exp{−n̄ · P (r ≥ rc )}. (16)
p(xmax ) = n · x n−1 · d x ⇒
Different volumes can be accounted for simply by adding 1
a volume term in the form of Eq. (17) P (x > xmax ) = p(xmax ) = 1 − xmax
n
x max
V (21)
P N f ≥ Nc = exp −n̄ Vo · · P (r ≥ rc ) . (17) P (x ≤ xmax ) = n
xmax .
V0
The result can also be expressed as Eq. (22), where
Depending on the particle size distribution specifics,
P rmax = probability that the largest r is smaller or equal
again the results will be Frechet, Eq. (4); reversed Weibull,
to r max .
Eq. (5) or Gumbel, Eq. (6). The value with the form of
Eq. (17) is that it provides a direct link between the actual Pr max,n = Prnmax . (22)
particle distribution and the distribution for the maxi-
mum values. It also allows the use of more complicated, The size effect can thus for the Gumbel distribution be
physically perhaps more sound size distributions than the expressed as Eq. (23),
simple ones used here. Mathematically, the difference be-
tween the extreme value distributions and Eq. (17) is that α − rmax
Pr max = exp − exp = (Pr max,n )1/n , (23)
the extreme value distributions give the probability that β
the maximum particle is of a certain size or less, whereas which becomes Eq. (24),
Eq. (17) gives the probability that all particles are of a
certain size or less. In practice the equations are equiva- 1 1 α − rmax
· ln = exp , (24)
lent. However, the extreme value distributions should be n Pr max,n β
better for the description of maximum particle size data,
whereas Eq. (17) should be better for the description of which finally becomes Eq. (25).
fatigue data. 1 rmax α
ln ln =− + + ln n. (25)
Pr max,n β β
EFFECT OF ‘SIZE’ The size effect affects only the location parameter α, not
the scatter parameter β. This means that for the Gumbel
The following derivation is for the special case where distribution the absolute scatter is unaffected by size. If
the stress is constant, so that only the stressed volume is the scatter is expressed in proportional terms (e.g. in %),
affected by size. The case where also the stress varies with the scatter will be affected by size.
size is more complicated and is addressed later. The Gumbel size effect can be expressed in the reduced
The effect of size can simply be studied using order form, Eq. (26).
statistics.7 If a sample has n values x1 , . . . , xn , they can be
ordered by rank so that Y 1 <Y 2 <· · ·<Yn . In this case Y i is βn = β & αn = α + ln n. (26)
called the ith order statistic. Special cases include the min-
imum value Y 1 = xmin and maximum the maximum value The size effect can for the Frechet distribution be ex-
YN = xmax . If x has the probability density function p(x) pressed as Eq. (27),
c 2010 Blackwell Publishing Ltd. Fatigue Fract Engng Mater Struct 33, 333–344
338 K. WALLIN
r0 m The different size effects on scatter for the Gumbel on
Pr max = exp − = (Pr max,n )1/n , (27)
rmax one hand and the Reversed Weibull and Frechet on the
other hand are important for the selection of proper size
which becomes Eq. (28).
distribution.
1 Figure 4 compares two staircase tests8 on cast iron 1563-
ln ln = m · ln r0 + ln n − m · ln rmax . (28)
Pr max,n GJS-500–7. The scatter in the plain and notched speci-
mens seems clearly proportional to the mean. The size
In this case the size effect affects the parameter r 0 , which
effect on the scatter in the case of the Gumbel distribu-
controls both the location as well as the scatter. The ab-
tion is not proportional to the mean. This indicates thus
solute scatter is therefore affected by size, but expressed
that a Gumbel distribution may not suitable to describe
in proportional terms (e.g. in %), the scatter will not be
the fatigue threshold data.
affected by size.
A comparison of the different size effects is visualised in
The Frechet size effect can also be expressed in the re-
Fig. 5. The Frechet size effect is solely controlled by m,
duced form of Eq. (29).
whereas the Gumbel size effect is solely controlled by α.
mn = m & r0n = r0 · n1/m . (29) The size effect for the reversed Weibull is controlled both
by m and r limit . Since the parameter r limit is practically im-
The size effect can for the reversed Weibull distribution possible to estimate with any reliability, the size effect of
be expressed as Eq. (30), the reversed Weibull distribution becomes unpredictable.
rlimit − rmax m Depending on chosen distribution, the predicted size ef-
Pr max = exp − = (Pr max,n )1/n , (30) fect can thus vary strongly.
rlmit − r0
which becomes Eq. (31). 2.2
1 r0 = α = 10
2.0
ln ln = m · ln(rlimit − rmax ) m = 10
Pr max,n 1.8
−m · ln(rlimit − r0 ) + ln n. (31)
1.6
r0N/r0
Also in this case the size effect affects only the parameter 1.4
(r limit − r 0 ), which controls the location as well as the
1.2
scatter. Since the limiting size r limit cannot be affected by Frechet
size, the size effect acts on the location parameter r 0 . The 1.0 Gumbel
absolute scatter is therefore affected by size, but expressed Rev. Weibull r limit/r0 = 2
0.8
in proportional terms (e.g. in %), the scatter will not be Rev. Weibull r limit/r0 = 5
99.5 99.5 95 %
95 %
95
95
70
P [%]
P [%]
70 5% 235.7 MPa
195.9 MPa 40 5%
40
10
10 σ ≈ 23.8 MPa
σ ≈ 18.5 MPa
1
1 0.1
170 180 190 200 210 220 230 240 190 200 210 220 230 240 250 260
Local stress amplitude [MPa] Local stress amplitude [MPa]
Fig. 4 Distribution of fatigue strength, based on staircase test results for plain and notched specimens of EN 1563-GJS-500– 7.8
c 2010 Blackwell Publishing Ltd. Fatigue Fract Engng Mater Struct 33, 333–344
STATISTICAL ASPECTS OF FATIGUE LIFE AND ENDURANCE LIMIT 339
150 μ
da ⎛ ΔK ⎞
Δσ [MPa]
Approximate ≈⎜ ⎟ ⋅1 mm
solution 2 dN ⎝ K0 ⎠
EFFECT OF STRESS
Stress controls the critical fractured particle size lead- 100 K0 = 30 MPa√m
μ=5
ing to fatigue life Tf , but it may also limit the fac- Δa = 10 μm
ture of particles to begin with. A simplified expression Nc = 10000
for the relation between fractured particle size rc and
σ can be derived from basic fracture mechanics. The 10 20 40 60 80 200
relation between stress intensity factor, stress and flaw
rc [μm]
size, for a small embedded flaw is given by the classical
Eq. (33). Fig. 6 Simple schematic example of the effect of stress on the
critical particle size.
K ≈ σ · (rc + a) · π. (33)
c 2010 Blackwell Publishing Ltd. Fatigue Fract Engng Mater Struct 33, 333–344
340 K. WALLIN
By combining Eqs (35) and (41), it is possible to express for the case where C 2 = 0, Eq. (36).
the threshold in terms of an effective stress amplitude.
P (σ f ≤ σ )
Using realistic values for rc (<100 μm) and making a
⎛ ⎛ 2 ⎞m ⎞
small with respect to rc (threshold indicates minimal crack 1/μ
N0 σlb
growth), leads to a very simple relation in the form of = 1 − exp⎝−σ0 · ⎝1 − · ⎠ ⎠.
Nf σ − σth
Eq. (42). This effective stress can be used together with
Eq. (39)
(45)
σth If Eq. (39) is combined with Eqs (6) and (17), which
σe f f ≈ σ · 1 − = σ − σth . (42)
σ for the particle size distribution results in a Gumbel dis-
tribution, the resulting distribution can, adjusted for the
If Eq. (39) is combined with Eqs (4) and (17), which for
fatigue crack growth threshold, be expressed in the form
the particle size distribution result in a Frechet distribu-
of Eq. (46). Also this result is of the same form for the
tion, the resulting distribution for the stress is represented
case where C 2 = 0, Eq. (36).
by a distribution that is close to a Weibull distribution.
The power m equals the power in the Frechet distribu- P (σ f ≤ σ )
tion. The distribution can be expressed in the form of 1/μ
Eq. (43). N0 σβ
= 1 − exp − exp σ0 − · 2
.
Nf σ − σth
P (σ f ≤ σ )
⎛ ⎛ ⎞m ⎞ (46)
1/μ 2
Nf σ The previously mentioned analysis can be summarised
⎜ ⎜ · ⎟ ⎟
⎜ ⎜ σ0 ⎟ ⎟ as such that if the maximum particle size distribution fol-
⎜ ⎜ N0 ⎟ ⎟
⎜ ⎜
= 1 − exp⎜− ⎜ ⎛ ⎟ ⎟.
2 ⎞ ⎟ ⎟
lows a Frechet distribution, the fatigue life will follow
⎜ ⎜ Nf 1/μ
σ ⎟ ⎟ a Weibull distribution, but if the maximum particle size
⎝ ⎝ ⎝1 − · ⎠⎠ ⎠
N0 σub distribution follows either a reversed Weibull distribu-
tion or a Gumbel distribution, the fatigue life will follow
(43) more complicated, non-standard distributions. Thus, if
Physically, the parameter σ ub represents the situation the Frechet distribution provides comparable fits to the
where the maximum particle size is less than the resolu- maximum particle size distributions, it may be advisable
tion of the particle size measurements. If σ ub
σ , to describe the fatigue life with a Weibull distribution.
which also corresponds to C 2 = 0, described by Eq. (36), In order to examine the realism of different distribution
the result is simply a two-parameter Weibull distribution options, the maximum defect distribution, arising from
with the power 2·m. Actually, applying Eq. (4) directly a number of fatigue specimens, was fitted by non-linear
with Eq. (17), without omission of the constant 1, in the regression, with the three possible size distributions. The
derivation of the stress effect causes C 2 to become 0 and data corresponded to a cast iron (EN 1563-GJS-500–7)8
the result is a simple two-parameter Weibull distribu- and the results are shown in Fig. 7. Additionally, the
tion, without approximations. The Weibull distribution data was also fitted by the GEV distribution given by
is thus equally possible as the other resulting distributions.
Applying the expression for the effective stress, given by
1.0
Eq. (42), adjusting for the fatigue crack growth threshold,
the result is a three-parameter Weibull distribution in the EN 1563-GJS-500-7
P (σ f ≤ σ ) 0.6
m Distribution of maximum
N f 1/μ σ − σth (44) defect size on fatigue specimen
= 1 − exp − · .
N0 σ0 0.4 fracture surface
c 2010 Blackwell Publishing Ltd. Fatigue Fract Engng Mater Struct 33, 333–344
STATISTICAL ASPECTS OF FATIGUE LIFE AND ENDURANCE LIMIT 341
Eq. (10). The Frechet distribution provides for this data though a slight over-conservatism is evident for low prob-
set the best overall description of the data with respect to abilities. This is likely due to the omission of the initiation
the regression coefficient r 2 and it provides an identical life Ni . The right hand figure shows the fit in terms of re-
fit as GEV. This indicates that the simple Weibull dis- lation between fatigue life and stress amplitude. The most
tribution, given by Eq. (44), should in this case provide a striking feature of the data is that no threshold stress is
good description of fatigue life data. predicted (likely caused by an environmental effect). This
The reversed Weibull and Gumbel distribution provides is in a strong contrast to fatigue data in air for the same
practically identical fits to the data. This is interesting material,9 shown in Fig. 9. Again, the left hand figure
since their size effects differ considerably. shows the fit of Eq. (44) to fatigue life data corresponding
The previously mentioned equations, lack one feature. to three different stress amplitudes. The fit to the cumu-
Before crack growth can start, some fatigue cycles are lative rank probability data is satisfactory, even though,
usually required to damage the fatigue initiating particles. also here, a slight over conservatism is evident for low
Sometimes inclusions, acting as stress concentrations, probabilities. The right hand figure shows the fit in terms
may also initiate cracks in the matrix. This leads to an of relation between fatigue life and stress amplitude. This
effective fatigue life in the form of Neff = Nf – Ni . The time a clear threshold stress is predicted. Eq. (44) predicts
initiation life (Ni ) will be a function of stress, material only a single threshold stress value. However, since also
strength and particle type and geometry. For high stresses the true endurance stress will show a scatter, the threshold
and sometimes weak particles, Ni will be small, whereas stress σ th has been varied to provide realistic fits close
for low stresses and sometimes strong particles Ni may to the threshold region.
constitute a large part of the fatigue life. However, a de- Figure 10 shows fatigue data for a 0.21% carbon steel
tailed investigation of Ni is outside the scope of the present tested in air.2 Again, the left hand figure shows the fit of
study. It is sufficient to note that Ni may explain some con- Eq. (44) to fatigue life data corresponding to three dif-
servatism in the estimates based on the previously men- ferent stress amplitudes. The fit to the cumulative rank
tioned equations, corresponding to low probabilities. probability data is satisfactory, even though, also here, a
slight over conservatism is evident for low probabilities.
The description of the lowest stress amplitude could be
APPLICATION TO DATA
better, but probably, the fit would be improved if the
Since it is physically justifiable to use Eq. (44) to estimate parameter Ni would be accounted for. The right hand
the effect of stress on fatigue life data, a number of liter- figure shows the fit in terms of relation between fatigue
ature data sets were analysed with it, to check its ability life and stress amplitude. Also here a clear threshold stress
to fit actual data. Figure 8 shows fatigue data for a S45C is predicted and the threshold stress σ th has been var-
carbon steel tested in a 3% NaCl solution.9 The left hand ied slightly to provide realistic fits close to the threshold
figure shows the fit of Eq. (44) to fatigue life data cor- region.
responding to five different stress amplitudes. The fit to Figure 11 shows two more examples of the use of Eq. (44)
the cumulative rank probability data is satisfactory, even to describe fatigue life data. The data sets did not allow
10 400
3 % NaCl solution 95 %
300 3 % NaCl solution
S45C carbon steel Weibull S45C carbon steel
σYl = 340 MPa 5% with Δσeff σYl = 340 MPa
200
1 σU = 605 MPa σU = 605 MPa
ln(1/(1-Pf))
Δσa [MPa]
Rotary bending
Weibull
with Δσeff 100
Δσa MPa
μ = 3.2
0.1 280 μ = 3.2
m = 9.0 190 m = 9.0
Δσth = 0 MPa 100
Δσth = 0 MPa
80
50
4 5 6 7 8 5 6 7 8
10 10 10 10 10 10 10 10 10
Nf Nf
Fig. 8 Weibull analysis of fatigue life data of S45C steel in 3% NaCl solution showing lack of fatigue threshold. Data taken from.9
c 2010 Blackwell Publishing Ltd. Fatigue Fract Engng Mater Struct 33, 333–344
342 K. WALLIN
10 400
S45C carbon steel S45C carbon steel
Air Air
σYl = 340 MPa 95 % σYl = 340 MPa
σU = 605 MPa
350
σU = 605 MPa
Weibull
Rotary bending
1 with Δσeff
ln(1/(1-Pf))
5% Rotary bending
Δσa [MPa]
300
Weibull
with Δσeff
200
4 5 6 5 6 7
10 10 10 10 10 10
Nf Nf
Fig. 9 Weibull analysis of fatigue life data of S45C steel in air showing fatigue threshold. Data taken from Ref. 9.
10
270 95 % 0.21% carbon steel
0.21% carbon steel
σYl = 263 MPa
σYl = 263 MPa
260 Weibull σU = 467 MPa
σU = 467 MPa
5% with Δσeff
1 Rotary bending
ln(1/(1-Pf))
Δσa [MPa]
Rotary bending
250
Δσa MPa
240
255 Weibull
270 with Δσeff 240
0.1 μ = 1.35
μ = 1.35
m = 10 230 m = 10
Δσth = 220 MPa
Δσth = 220 ± 3 MPa
220
5 6 7 6 7
10 10 10 10 10
Nf Nf
Fig. 10 Weibull analysis of fatigue life data of carbon steel in air showing fatigue threshold. Data taken from Ref. 2.
300 200
7075-T651 aluminium 180 95 %
95 % Weibull 1018 Steel
250 Reverse bending, R = -1 with Δσeff
160 Notched R = -1
Weibull 140
Δσa [MPa]
with Δσeff
σa [MPa]
200 5% 5%
120
Δσ
100
μ = 1.28 μ = 1.41
150 m = 10
m = 13
Δσth = 135 ± 8 MPa Δσth = 100 ± 15 MPa
80
No cracks
5 6 7 6 7
10 10 10 10 10
Nf Nf
Fig. 11 Weibull analysis of fatigue life data of 7075-T651 aluminium alloy and 1018 steel in air showing fatigue threshold. Data taken from
Refs 10 and11.
c 2010 Blackwell Publishing Ltd. Fatigue Fract Engng Mater Struct 33, 333–344
STATISTICAL ASPECTS OF FATIGUE LIFE AND ENDURANCE LIMIT 343
the assessment of the fatigue life variation at different Weibull distribution (based on a Frechet distribution for
stress amplitudes, only the relation between fatigue life the maximum particle size) appears justified.
and stress amplitude. Even so, they confirm the capability
of Eq. (44) to describe the fatigue life data wit a satisfactory DISCUSSION
accuracy.
The study compared three basic distributions for the max-
imum particle size as candidates for the description of
THE ENDURANCE LIMIT scatter in fatigue life and endurance limit. The size effect
connected to the reversed Weibull distribution is strongly
From Eq. (41), the relation between endurance limit dependent on the upper limiting particle size, a value that
(σ th ) and particle size becomes simply: contains a large statistical uncertainty. This makes the re-
K th versed Weibull distribution poorly applicable for the de-
σth = √ . (47) scription of fatigue data. A similar problem is connected
π · rc
to the Gumbel distribution, where the absolute scatter is
For small particles, local yielding starts to affect the independent of specimen size. This leads to the physically
threshold. For this, a simple modification originally pro- impossible possibility of obtaining negative particle sizes,
posed by El-Haddad et al.12 is usually used. The modifi- with decreasing specimen size. The Frechet distribution
cation consists of replacing rc by the term rc + r 0 . The (and its modification) appears best suited for the task of
modification leads to the relation given by Eq. (48). This describing fatigue data. If the particle size distribution is
is equivalent to Eq. (39) for the fatigue life. described by Eq. (4), the fatigue life is predicted to fol-
low a three-parameter Weibull distribution with respect
1 K th 2 C1 2
rc ≈ · − r0 = − C2 . (48) to stress and the endurance limit is predicted to follow a
π σth σth
two parameter Weibull distribution with respect to stress,
When Eq. (48) is combined with Eqs (4) and (17) the both results being backed by actual data. It is recognised
result is, analogous to the fatigue life, simply a two- that the Gumbel distribution has been used successfully to
parameter Weibull distribution in the form of Eq. (49). describe fatigue data, e.g.13 but the implications of the size
effect on the scatter has not previously been considered
V σ 2·m
P (σth > σ ) = exp − · . (49) in detail.
V0 σre f
The three-parameter Weibull distribution, Eq. (44), was
Figure 12 shows the data from Fig. 4, plotted in Weibull shown to provide satisfactory, but not excellent descrip-
‘coordinates’, fitted by Eq. (49). The data corresponding tions of the fatigue life data. Also, the values for the Paris
to both plain and notched specimens, show a clear size law exponent μ, predicted by the fits are less than normally
effect and the best description of the combined data sets seen for metals in air. The reason for the discrepancy is
is obtained by a straight line when plotted in Weibull attributed to the omission of an initiation life Ni , which
‘coordinates’. Neither of the estimates based on reversed describes the time spent prior to crack growth. Inclusion
Weibull or Gumbel for the maximum particle size result of Ni would lead to Eq. (50). Application of Eq. (50) to
in a straight line in these coordinates. Thus, the use of the the fatigue life data in Fig. 9, provides an excellent fit,
with a realistic value for μ (Fig. 13). As predicted, the
initiation life Ni decreases with increasing stress. Since
1.5
the initiation life is not directly related to the particle size
1.0 2-par Weibull
m = 5.9 Δσref = 206 MPa
distribution, the assessment of its value in a consistent
0.5 manner is outside the scope of the present study.
lnln[1/(P{Δσth>Δσ})]
0.0
P (N f ≤ N)
-0.5
2 1/μ m
.1
-1.0 =
1
=
0 V N − Ni σ − σth
V0 V0 = 1 − exp − · · .
-1.5 V/ V/ V0 N0 − Ni σ0
-2.0 (50)
Plain
-2.5
Notched The endurance limit, σ th , is not constant. If the parti-
-3.0 cle size distribution leads to Eq. (50), the endurance limit
180 200 220 240 260 follows Eq. (49). The Weibull exponent in both Eq. (49)
Δσth [MPa] and (50) should be the same, being equal to the Frechet
exponent in the description of the maximum particle size.
Fig. 12 Data from Fig. 4, analysed by Eq. (49). The maximum particle size distribution shown in Fig. 7,
c 2010 Blackwell Publishing Ltd. Fatigue Fract Engng Mater Struct 33, 333–344
344 K. WALLIN
24
1
74
ln(1/(1-Pf))
13
Δσa MPa impossible possibility of obtaining negative particle sizes.
=
N
i
340
310 If the maximum particle size distribution is described by
280
1 a modified Frechet distribution in the form of Eq. (51), the
4
Weibull
9040
Ni = 4772
0.1 with Δσeff fatigue life will follow a double three-parameter Weibull
N=
m = 4.9
Δσth = 210 MPa
life Ni is a function of stress. Corollary the endurance limit
will follow a simple two-parameter Weibull distribution.
5 6
10 10
Nf Acknowledgements
c 2010 Blackwell Publishing Ltd. Fatigue Fract Engng Mater Struct 33, 333–344