Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

STRUCTURAL

SAFETY

Structural Safety 29 (2007) 6676


www.elsevier.com/locate/strusafe

Inherent variability of an experimental crack growth curve


F. Casciati

a,*

, P. Colombi b, L. Faravelli

Department of Structural Mechanics, University of Pavia, Via Ferrata, 1, 27100 Pavia, Italy
Department of Structural Engineering, Technical University of Milan, P.zza L. da Vinci, 32, 20133 Milan, Italy
Received 5 March 2005; received in revised form 1 February 2006; accepted 8 February 2006
Available online 31 March 2006

Abstract
A statistical data analysis method is applied to a sample of crack growth curves collected by the authors. Details on the
experimental environment are provided and the resulting sample is illustrated. The statistical estimates of the model
parameters are compared with the analytical assessment made possible by an approximate analytical framework.
2006 Elsevier Ltd. All rights reserved.
Keywords: Crack growth; Fracture; Fatigue tests; Random fatigue; Material non-homogeneities

1. Introduction
Two main approaches have been proposed in the literature [1,2] to model the random eect in fatigue crack
growth as observed in laboratory tests:
1. a non-linear mathematical model is used to estimate the average crack propagation path. Realisations of a
stochastic process are used to move from the average path to the actual one. A stochastic dierential equation is written for the fatigue crack growth rate [1]. Numerical or analytical techniques are used to solve the
stochastic crack growth equation and to compute the lifetime distribution [36];
2. a suitable mathematical model is tted to each experimental crack growth curve. For a sucient large sample one can estimate the probability distribution model and hence the distribution itself for the model
parameters [710]. The relevant residuals represent a stochastic process able to reproduce the irregularity
details of the experimental tests. First order (FORM) and second order (SORM) reliability model
(FORM/SORM) [2,11] and simulation techniques are used to evaluate the lifetime distribution.
This paper reports the results of the experimental activity carried out by the authors and shows how the
parameters of these models can be estimated to quantify the randomness in the fatigue resistance. Uncertainties in the applied loading can be treated as shown by, among others, [1214].
*

Corresponding author. Fax: +39 0382 528422.


E-mail address: fabio@dipmec.unipv.it (F. Casciati).

0167-4730/$ - see front matter 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.strusafe.2006.02.001

F. Casciati et al. / Structural Safety 29 (2007) 6676

67

Nomenclature
a
B
C, m
C1
C 2(a)
E
R
R2
Rf(s)
V
Y(a)
x(N)
w(N)
W
DK
DS
q, r
rC 2
f()

crack length
compact tension thickness
Paris law constants
positive random variable describing the behaviour deviation in dierent specimens
random process which models the deviation of the crack growth rate within each specimen
Youngs modulus
stress ratio
multiple correlation coecient
residuals autocorrelation function
crack mouth opening displacement
correction factor, depending on the specimen and crack shape, in the evaluation of the stress
intensity factor
nonnegative random process which represents the scatter in fatigue crack growth rate
Wiener process
compact tension width
stress intensity factor range
applied stress range
material parameters of the dierential equation modelling the crack growth
standard deviation of the process C 2(a)
residuals of the regression analyses

A summary of the governing relations is provided in the next section in order to summarize the identication process and to explain how the parameters are actually estimated from the experimental data.
2. Governing relations
2.1. Dierential equation models
Both the eects of random loading and material fatigue resistance have been introduced in such models
[1,2,15]. This paper considers only the second source of uncertainty.
Most of the empirical models for fatigue crack growth have the form of a dierential equation and, hence,
randomisation leads to a stochastic dierential equation for fatigue crack growth. The empirical dierential
equation:
daN
F DK; S; R; a
1
dN
shows an evident non-linear character. In the above expression a(N) is the crack size after a given number of
cycles N, F is a non negative function, DK is the stress intensity range [16], S is the applied stress and R = Smin/
Smax the stress ratio. The empirical fatigue crack growth equation is usually randomised in the form:
daN
F DK; S; R; a  xN
2
dN
where x(N) is a non negative random process. It turns out that the crack size a(N) is also a random process
(and this is emphasised by underlining the symbol). The probability distribution of the crack size, as required
for a reliability prediction, can be obtained analytically under very restrictive hypotheses or, in the general
case, by adopting a suitable numerical algorithm.
Most of the proposed models [16] for fatigue crack growth have been developed from the randomisation of
the simple Paris-type equation. In this case one has F(DK,S,R,a) = C DKm and then:
daN
C  DK m  xN
3
dN

68

F. Casciati et al. / Structural Safety 29 (2007) 6676

where C and m are material constants. In this scheme the Paris equation represents a mean-behaviour equation (resulting from averaging over possible scattered results). Actually, the linear regression on crack growth
data, when organized in logarithmic format, estimates the median. When the variability of the noise is low, the
mean and median are close to one other.
Dierent solutions have been proposed in the literature for the stochastic process x(N). It should characterise the random interactions aecting the crack growth, but it should also have features that allow an eective treatment of the fatigue crack growth equation. A Gaussian white noise was proposed for x(N) [1]. Its
Gaussian nature allows, theoretically, negative values with non-zero probability for the crack growth rate.
A ltering technique [1,3,5] can be introduced in order to exclude the even small possibility of negative
crack growth rates and to take into account the necessary correlation information. But care should be used
in preserving the advantages that the Markov process theory oers in the analysis of fatigue crack growth.
As an example, the following lter was adopted in [3]:
8
>
< xN lx yN
yN q  cos/N
4
>
: d/N r  dwN
where lx is the mean value of x(N), w(N) is a Wiener process [11] and q and r are material parameters.
2.2. Reliability theory models
In these approaches, the randomness in the fatigue crack growth rate is split in two parts: (a) the specimento-specimen variation and (b) the variation within each specimen. The rst one is modelled by random variables, while the second one is represented by a random process [710].
Consider again the Paris equation. The material parameter C is modelled as:
Ca

C1
C 2 a

where C1 is a positive random variable describing the deviation between the behaviour of dierent specimens,
and C 2(a) is a random process modelling the deviation of the crack growth rate within each specimen. The
presence of the stochastic term in the denominator is consistent with an earlier proposal by Madsen [10]: additive or multiplicative solutions could result more suitable from a stochastic point of view but the algebra
the
pof

following formulas makes Eq. (5) preferable. The Paris equation is then rewritten as DK DS  Y a  p  a:
p m
da
C1

 DS  Y a  p  a
dN C 2 a

where Y(a) is a factor which takes into account the shape of the crack and of the specimen and DS is again the stress
range. In this scheme the Paris equation represents the crack growth in each specimen for a given (xed) realisation
of C1 and m. For the special case of constant amplitude loading, the above equation can be integrated as:
Wa C 1  DS m  N
where W(a) is a random variable. It is given by the following stochastic integral:
Z a
C 2 z
p m dz
Wa
Y
z
 p  z
a0

The probability distribution of W(a) is not easily computed, but the mean value and the variance are given
by [10]:
Z a
Z a
EC 2 z
1
p m dz
p m dz
EWa
9
p  a
p  a
a0 Y a 
a0 Y a 
Z aZ a
CovC 2 z1  C 2 z2 
10
VarWa
p m
p dz1 dz2
p  z1  Y z2  p  z2 m
a0
a0 Y z1 

F. Casciati et al. / Structural Safety 29 (2007) 6676

69

If the correlation length of C 2(a) is short compared to the crack increment of interest, the following approximation [10] can be used for Var[W(a)]:
Z a
1
11
VarWa qC 2  VarC 2  
p 2m dz
p  z
a0 Y z 
where qC2 is the correlation length of C 2(a). Moreover, the random variable W(a), under the assumption of
short correlation length for C 2(a), is the sum of a large number of independent random variables with approximately the same variance. Its probability distribution is then well approximated by the Gaussian distribution.
The failure of the specimen is usually dened as the point at which the crack size reaches the critical size ac.
Then the following limit state function can be dened:
Wac  C 1  DS m  N 6 0

12

The corresponding failure probability can be computed by a standard reliability method (FORM/SORM) [8
11].
2.3. Parameter identication
Standard statistical analysis was used in the literature to t to experimental data a given random fatigue
model. Assume one has n curves aji N i available. The jth (j = 1, n) curve is the record of n1 couples (ai, Ni)
(i = 1, n1).
2.3.1. Reliability theory model
Consider rst the reliability theory model [79] given by Eq. (6). The average behaviour of each single crack
growth curve was tted [10] according to the Paris law by computing the material parameters outcomes mj and
C1j and the consequent realisation C2j(a) of the stochastic process C 2(a) (see Eq. (6)). For this purpose, consider the logarithmic form of the crack growth rate:
 
da
log
log C 1j mj  log DK fj a
13
dN j
where the residuals fj(a) are related with C2j(a);
 
da
fj a  logC 2j a log
 logC 1j  mj  logDK
dN j

14

The coecients C1j and mj are computed by a least squares regression analysis which minimises the residuals.
In order to measure the goodness for the tted line, the multiple correlation coecient, R2j , was computed [7].
In general, 1  R2j quanties the portion of the residual sum of squares SST that is not explained by the model,
with:
SST

n1
X
2
fj aji 

15

i1

In the previous equation aji is the i-th value of the curve aj. The variance of the residual, r2fj , is computed as:
r2fj

n1
1 X
fj aji 2
n1  2 i1

16

The process f(a) of which fj(a) is a realization is usually assumed in the literature to be a stationary Gaussian
process [7]. Then, the standard normal to log-normal transformation formula for the second order statistics
provides:
r2C2j e

r2f

 e

r2f

 1

17

70

F. Casciati et al. / Structural Safety 29 (2007) 6676

The autocorrelation function, Rfj s, of the residuals fj(a) is computed as the inverse Fourier transform of the
power spectral density function of the residuals [17]. The estimate of the model parameters is nally obtained
by averaging the n values computed for j = 1, n.
In order to capture the non-linear behaviour of fatigue crack growth data, an alternative cubic fatigue crack
growth relationship was proposed in the literature [5]:
2
3
da
1

 ec0 c1 log DKc2 log DK c3 log DK 


dN C 2 a

The residuals are now dened as:


 
da
fj a log
 c0j c1j  log DK c2j  log DK2 c3j  log DK3 
dN j

18

19

Clearly, the cubic model reduces the correlation length of the random process C 2(a). This can be explained
[1] by noting that if C 2(a) is uncorrelated (i.e. a white noise) the smallest statistical dispersion is achieved, while
the largest statistical dispersion is obtained if it is perfectly correlated (i.e. a random variable).
2.3.2. Dierential equation model
Coming to the dierential equation model given by Eq. (2), the Paris law is tted to the whole experimental
data set by linear least squares regression analysis. One has for the logarithmic form of the fatigue crack
growth law:
 
da
log
20
logC 1 m  logDK fN
dN
where:


da
fN logxN log
dN


 logC 1  m  logDK

21

are the residuals. The variance of the residual is now given by [1]:
r2f

v
1 X
fN i 2
v  2 i1

22

where v = n n1 is the sample dimension. Again, the lognormal and stationary assumption is adopted in the
literature for the stochastic process x(N) of Eq.(3) [5]. Then, making use of Eq. (22) and the associated equivalent of Eq. (17), one has the standard deviation, rx, of the random process x(N).
In order to improve the goodness-of-t, again a cubic fatigue crack growth relationship similar to Eq. (20)
could be adopted as suggested in the literature [5]:
2
3
da
ec0 c1 log DKc2 log DK c3 log DK   xN
dN

The residuals are now dened as:


 
da
2
3
fN log
 c0 c1  log DK c2  log DK c3  log DK 
dN

23

24

3. Experimental program
When the fatigue process is characterised by the growth of a dominant crack, couples of values crack length
number of cycles represent the experimental basis required to build any of the models described in the previous section.

F. Casciati et al. / Structural Safety 29 (2007) 6676

71

3.1. Testing equipment and specimens conguration


Fatigue crack growth tests have been conducted by a MTS testing machine making use of a commercial
software package (Fatigue Crack Growth Test Application) by the MTS Systems Corporation [18]. Three
methods are available to measure the crack length: visual measurement, compliance measurement or external
measurement. The compliance measurement option was adopted. The crack length is estimated using a polynomial equation that, for a given specimen geometry correlates the crack length to the specimen stiness.
There is also an inverse correlation between specimen stiness and crack length. By measuring the specimen
stiness, one can calculate the crack length.
As an example for the compact tension (CT) specimen, as dened in appendix A2 of ASTM E647, Standards [19] provides the following polynomial form:
a
c0 c1  u c2  u2 c3  u3 c4  u4 c5  u5
25
W
where a is the crack length, W the specimen width and u is the dimensionless compliance. For the CT specimens u is given by:
1
26
u q
BEV
1
P
where P is the acting loading, V is the crack mouth opening displacement as measured with the crack opening
displacement (COD) gage, and the ratio V/P is the compliance. This quantity is made dimensionless by multiplying it by B, the specimen thickness, and E, the modulus of elasticity.
To get a better measurement of the compliance, the software package excludes the top and the bottom portions of the COD versus load curve, and ts just the interior points. This eliminates the non-linearity due to
phenomena such as crack closure from aecting the slope of the curve. The crack length is recorded at a crack
step selected by the user. This option allows one to avoid any interpolation of the records.
As stated above, the results are sensitive to the modulus of elasticity, E, of the material. The software package provides the further option of having a preliminary assessment of the value of E in order to reduce this
source of uncertainty.
The experiments were conducted on ASTM E647 standard Fe E 235 CT specimens (see Fig. 1), having a
thickness B = 12.7 mm and a width W = 55 mm. Specimens were extracted from a slab of Fe360 steel having
321.5 MPa yield strength. A straight notch, h = 3 mm, was prepared by electro-discharge machining (EMD)
technique. The initial notch, an, to width, W, ratio was 0.2 and then an = 11 mm.
Before starting the fatigue crack growth test, the specimens were pre-cracked (which leads the crack from 11
to 14 mm) by the same software package used to perform the crack growth tests. For this purpose, the

thickness 12.7 mm
Fig. 1. The CT specimen.

72

F. Casciati et al. / Structural Safety 29 (2007) 6676

software applies variable amplitude loading in order to reduce the number of cycles required to achieve the
selected initial crack length. The load history is automatically selected in order to meet the ASTM E647
requirement and to avoid crack growth retardation.
3.2. The collected fatigue crack growth data
All the tests were conducted in air environment, with sinusoidal constant amplitude loading waveform at
30 Hz frequency. The maximum applied loading was Pmax = 13 KN and the load ratio R = Pmin/Pmax was 0.3.
The initial crack length after the pre-cracking operation was a = 14 mm. Crack length was measured by the
0.028

0.026

0.024

a [m]

0.022

0.02

0.018

0.016

0.014

0.5

1.5

2.5

N [-]

3.5
5

x 10

Fig. 2. Fatigue crack growth for 9 replicate tests of the CT data set.
-15.5

-16

log (da/dN) [1/m]

-16.5

-17

-17.5

-18

-18.5
2.7

2.8

2.9

3
3.1
log(K) [MPa m)]

3.2

3.3

Fig. 3. Fatigue crack growth rate for 9 replicate tests of the CT data set.

3.4

F. Casciati et al. / Structural Safety 29 (2007) 6676

73

compliance method up to a nal crack length equal to 25 mm with a constant crack step equal to 0.1 mm. A
sample of nine fatigue crack growth curves was obtained as showed in Fig. 2.
The crack growth rate, as required by Eqs. (3) and (6), was computed by the secant method following the
ASTM E647 recommendation. Its values are reported in Fig. 3 as a function of the stress intensity factor range
DK.
4. Scatter of the fatigue crack growth curve: analyses and discussion
The scatter in fatigue crack growth of the CT data set collected in this paper is analysed in this section
according to both the reliability theory and the dierential equation model. The parameter identication is
performed according to the technique previously summarised. The CT fatigue crack growth curves were
obtained, as discussed previously, by a closed-loop fatigue machine with data acquisition tools and control
features that make the experimental error negligible.
First, the reliability theory model was considered. For each single fatigue crack propagation curve the Paris
model (see Eq. (6)) was then tted by linear least squares regression analysis, as depicted in Fig. 4. The evaluated samples of C1 and m are reported in Table 1. One has E[C1] = 1.6560E-12 (MPa units) and
E[m] = 3.3945.

-16
-16.2
-16.4

log (da/dN) [1/m]

-16.6
-16.8
-17
-17.2
-17.4
-17.6
-17.8
-18
2.7

2.8

2.9

3
3.1
log(K) [MPa m)]

3.2

3.3

3.4

Fig. 4. Crack growth rate vs. the stress intensity factor range for a given record of the CT data set.

Table 1
Samples of random variables C1 and m
Specimen

C1 (MPa units)

m ()

R2

1
2
3
4
5
6
7
8
9

2.823E-12
1.810E-12
1.757E-12
1.439E-12
1.153E-12
1.473E-12
1.502E-12
1.923E-12
1.024E-12

3.2271
3.3747
3.3677
3.4254
3.4845
3.4126
3.3910
3.3256
3.5415

0.9618
0.9605
0.9858
0.9977
0.9907
0.9951
0.9969
0.9731
0.9973

74

F. Casciati et al. / Structural Safety 29 (2007) 6676


1.2

Autocorrelation function

0.8

0.6

0.4

0.2

-0.2
-0.015

-0.01

-0.005

0
Spatial lag [m]

0.005

0.01

0.015

Fig. 5. Autocorrelation function for the CT data tests.

A multiple correlation coecient E[R]2 = 0.984 (see Eq. (15) with n1 = 110) was obtained indicating an
excellent t of the data. The mean value of C 2(a) was equal to 1.001 (MPa units) while the standard deviation
of C 2(a) was computed by Eqs. (15) and (16) and is E[rC 2]=0.059 (MPa units). Since the goodness of the t
was excellent, the cubic fatigue crack propagation model was not implemented. The autocorrelation function
of the residuals E[Rf(s)] was computed as described before and it is shown in Fig. 5.
Since the scatter of the fatigue data is small, the process C 2(a) has a very short correlation length and can
adequately be represented by a white noise.
Consider now the dierential equation model given by Eq. (3). A plot of the logarithmic crack growth rate
versus the logarithmic stress intensity factor range is depicted in Fig. 3. Again, the Paris law is tted to the
whole experimental data set by linear least squares regression analysis as showed in Fig. 6. The computed
-15.5

-16

log (da/dN) [1/m]

-16.5

-17

-17.5

-18

-18.5
2.7

2.8

2.9

3
3.1
log(K) [MPa m)]

3.2

3.3

3.4

Fig. 6. Crack growth rate vs. the stress intensity factor range for the whole CT data set.

F. Casciati et al. / Structural Safety 29 (2007) 6676

75

Paris Parameters are C = 1.5899E-12 and m = 3.3945, while lx = 1.003 rx = 0.082 (see Eq. (22) and the associated equivalent of Eq. (17)). A multiple correlation coecient R2 = 0.97 was obtained indicating also in this
case an excellent t of the data. Again, the cubic model was not implemented.
The small dispersion recorded in this sample of specimens tested gives rise to a good agreement between the
results of the two approaches.
These results do not add information on the basic question about the advantages/disadvantages of the two
approaches introduced in Section 2: the former with independent variable depending on the cycle number N
and the latter adopting the crack length a. The rst approach nds its justication in the deeply investigated
format of the consequent mathematical framework. The second one deals with Eq. (5) but allow a more physical modelling of the uncertainty directly on the crack length rather than on the number of cycles.
5. Conclusions
Experimental tests were performed on structural steel CT specimens. The sample dimension (nine replicates) was large enough to show that the fatigue scatter observed in each single experimental test is negligible.
In particular, the scatter, measured by standard deviation, was much lower than the one observed in the literature with reference to the well-known Virklers data set [20]. Of course, one cannot compare aluminium
with steel, and one should not also forget the strong dependency of the crack growth rate from the size
and number of the inclusions in the material.
Finally, one should note that the given fatigue data sets were obtained starting from an existing initial crack
and no crack initiation was considered. Surely, crack initiation signicantly contributes to the scatter of fatigue life since it strongly depends on surface conditions. Another source of scatter is given by the load spectrum
that was not considered across this paper but cannot be ignored when one considers the fatigue life distribution of operating structures.
Acknowledgements
The contributions of the rst and third authors were supported by the University of Pavia (FAR). The second author acknowledges the nancial support received by the Milan Technical University (FRA). The
authors thank the Scientic and Technologic Park of Valle Scrivia (PST) which owns the experimental equipment used within this investigation.
References
[1] Sobczyk K, Spencer Jr BF. Random fatigue: from data to theory. Academic Press; 1992.
[2] Colombi P, Faravelli L. Stochastic nite elements via response surface: fatigue crack growth problems. Probabilistic methods for
structural design. Kluwer Academic Publishers; 1997. p. 31338.
[3] Casciati F, Colombi P, Faravelli L. Fatigue crack size probability distribution via a lter technique. Fatigue Fract Eng Mater Struct
1992;15(5):46375.
[4] Lin YK, Yang JN. A stochastic theory of fatigue crack propagation. AIAA J 1985;1:11724.
[5] Tang J, Spencer Jr BF. Reliability solution for the stochastic fatigue crack growth problem. Eng Fract Mech 1992;34(2):41933.
[6] Solomos GP. First-passage solutions in fatigue crack propagation. Probabilist Eng Mech 1989;4(1):329.
[7] Ortiz K, Kiremidjian AS. Time series analysis of fatigue crack growth rate data. Eng Fract Mech 1986;24(5):65775.
[8] Ortiz K, Kiremidjian AS. Stochastic modelling of fatigue crack growth. Eng Fract Mech 1988;29(3):31734.
[9] Dolinski K. Stochastic modelling and statistical verication of crack growth under constant amplitude loading. Eng Fract Mech
1992;43(2):195216.
[10] Madsen HO. Probabilistic and deterministic models for predicting damage accumulation due to time varying loading, Dialog 5-82.
Lyngby (Denmark): Danish Engineering Academy; 1983.
[11] Casciati F, Faravelli L. Fragility analysis of complex structural system. Research Studies Press; 1991.
[12] Colombi P, Dolinski K. Fatigue lifetime of welded joints under random loading: rainow cycles vs. cycle sequence method.
Probabilist Eng Mech 2001;16(1):6171.
[13] Moreno B, Zapatero J, Dominguez J. An experimental analysis of fatigue crack growth under random loading. Int J Fatigue
2003;25(7):56897.
[14] Beck AT, Melchers RE. Overload failure of structural components under random crack propagation and loading a random process
approach. Struct Saf 2004;26(4):47188.

76

F. Casciati et al. / Structural Safety 29 (2007) 6676

[15] Casciati F, Colombi P. Fatigue crack propagation under environmental actions. Structural dynamic systems, computational
techniques and optimisation. Gordon & Breach Publishers; 1998. p. 146.
[16] Broek D. Elementary engineering fracture mechanics. Martinus Nijho Publishers; 1986.
[17] Bendat JS, Piersol AG. Random data: analysis and measurement procedures. Wiley; 1971.
[18] MTS. Fatigue crack growth. Fatigue and fracture test applications. MTS Systems Corporation; 1992.
[19] ASTM E 647 95a. Standard test method for measurement of fatigue crack growth rates, ASTM, Philadelphia; 1995. p. 55793.
[20] Virkler DA, Hillberry BM, Goel PK. The statistical nature of fatigue crack propagation. J Mater Technol ASME 1979;101:14853.

You might also like