Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Structural Health Monitoring and Damage

Prognosis in Fatigue

S. S. Kulkarni and J. D. Achenbach*

Center for Quality Engineering and Failure Prevention, Northwestern University


Evanston, IL 60208, USA

A conceptual structural health monitoring (SHM) system to monitor fatigue damage is discussed in this
study. The first part of the study is concerned with damage prognosis. A methodology to use the data
from a pre-crack damage monitoring system to predict the number of cycles to macrocrack-initiation in
a probabilistic sense, is presented. Issues related to quantifying damage, a damage evolution law, and
numerical evaluation of the probability of macrocrack initiation are discussed. The second part of the
study deals with quantifying the effects of imperfect inspections during the growth of a macrocrack.
The probability that there exists an undetected critical crack is the quantity of interest. An analytical
expression for this quantity is derived for the case of a two-dimensional crack. A Monte Carlo
simulation method to quantify the effects of imperfect inspections is also presented. Numerical results
are presented for two examples of surface-breaking cracks with different geometries.

Keywords fatigue crack  initiation  growth  inspections

1 Introduction life (see [1]). The underlying concept is based on


detecting and characterizing current damage
The purpose of a structural health monitoring level as determined from the SHM system and
(SHM) system is to provide continuous (or on- evaluating it in terms of failure mechanics and a
demand) information about the state of a struc- damage growth law. The aim is to determine
ture so that assessment of the structural integrity whether damage in a structure is sufficiently
can be made at any time, and timely remedial small that failure can be precluded with a high
actions may be taken as necessary. A SHM degree of certainty within a preset interval. If so,
system typically consists of an array of sensors the structural system can be allowed to function
for periodic inspections of the structure. Features for that interval; otherwise, the structure must be
of interests are first extracted from the data taken off-line for further inspection, repair, or
provided by the sensors and then analyzed to replacement.
determine the current state of the structure. The current study discusses two aspects of a
Optimum use of the data acquired from the conceptual local SHM system for monitoring
SHM can be made by using it for the purpose of damage evolution during fatigue, namely,
damage prognosis. Damage prognosis is defined pre-macrocrack damage (Stage I) followed by
as an estimate of a structure’s remaining useful subsequent crack growth (Stage II). In the first

Copyright ß SAGE Publications 2008


Los Angeles, London, New Delhi and Singapore
*Author to whom correspondence should be addressed. Vol 7(1): 0037–13
E-mail: achenbach@northwestern.edu [1475-9217 (200803) 7:1;37–13 10.1177/1475921707081973]

37

Downloaded from shm.sagepub.com at PENNSYLVANIA STATE UNIV on March 5, 2016


38 Structural Health Monitoring 7(1)

part, a damage prognosis procedure to predict nucleated as the damage accumulation further
macrocrack initiation is briefly described. proceeds (see [7,8]). In the next stage, the
The procedure is based on a local ultrasonic microcracks grow and eventually coalesce to
measurement system which monitors pre-crack form a macrocrack. Once a macrocrack of
damage. This approach has been discussed in detectable size has been formed, further crack
considerable detail in a recent study by Kulkarni growth under continuing cyclic loading to a
et al. [2]. In the second part, the growth of the critical size, may take a number of loading cycles
macrocrack is considered whose initiation was that is relatively small as compared to the
predicted in the first part. It specifically addresses number of cycles leading to the formation of a
the issue of quantifying the effects of inspections detectable macrocrack. For high strength
on the probability of detecting the growing crack materials it is, therefore, important that a
when it becomes greater than a critical size. probabilistic prediction of the time to macrocrack
Since the growth of a crack is a stochastic formation can be established, since this prediction
phenomenon, various approaches have been may well be the most realistic estimate of fatigue
proposed in literature to model this behavior. safety in a damage-tolerant approach of such
Bogdanoff and Kozin [3] have accounted for the materials.
random nature of fatigue crack using the Markov To quantify the damage accumulation in a
process while Yang and Manning [4] and Yang component undergoing fatigue, it is first neces-
et al. [5] have modeled the stochastic nature of sary to relate the accumulated damage to an
the crack growth using a lognormal process. observable variable. This can be achieved as
A comprehensive review of the different follows. The accumulated damage which is
approaches appears in Sobczyk and Spencer [6]. caused by changes in the microstructure of the
Like Sobczyk and Spencer, the authors assume component in turn introduces material nonlinear-
that the growth rate of the detected macrocrack ity in the specimen. To characterize the material
can be represented by Paris law and introduce nonlinearity, a single frequency ultrasonic
randomness by considering the size of the initia- wave generated by a transducer (generator) is
ted macrocrack to be a random variable with transmitted through the specimen and the signal
known probability distribution. Two cases are con- received by the receiving transducer is analyzed.
sidered: (i) a two-dimensional surface-breaking The material nonlinearity distorts the single
crack in a tensile field and (ii) a semi-elliptical frequency wave and leads to the generation of a
surface-breaking crack in a tensile field. A wave which not only consists of a component
methodology is described to account for imper- at the fundamental frequency but also
fect inspections at different cycle numbers and is components at higher frequencies (integer multi-
followed by representative numerical results. ples of the fundamental frequency) (see e.g., [9]).
The components at the higher frequencies are
collectively referred to as higher harmonics and
2 Part I: Monitoring of Fatigue the individual components at twice the funda-
Damage Leading to Macrocrack mental frequency, thrice the fundamental fre-
Formation quency, etc., are referred to second harmonic,
third harmonic, and so on. Owing to the material
The nucleation and growth of a fatigue crack nonlinearity, the signal received at the receiver
in a metal is generally preceded by a process of not only consists of a component at the transmit-
damage accumulation on the microstructure level. ting (fundamental) frequency, but also second
As the number of fatigue cycles increases, the and higher harmonics whose amplitudes increase
damage accumulation involves the formation of as the material nonlinearity increases. Hence, the
dislocation monopoles, followed by dislocation material nonlinearity in a specimen can be
loops and dipoles, and subsequent dislocation quantified in terms of the amplitude of the
veins and persistent slip bands. Microcracks are fundamental, A1, and the second harmonic, A2,

Downloaded from shm.sagepub.com at PENNSYLVANIA STATE UNIV on March 5, 2016


Kulkarni & Achenbach SHM and Damage Prognosis in Fatigue 39

generated by propagating a single frequency damage with the number of cycles. A suitable
ultrasonic wave. This phenomenon of generation model that has been proposed is:
of higher harmonics due to material nonlinearity 8  
is generally referred to as acoustic nonlinearity. >
<  m
1 =2  rc ðÞ 1 
dD 
n if 2 > rc ðÞ
The acoustic nonlinearity is usually characterized ¼ N c =2 ð1  DÞ
by a nonlinearity parameter, which quantifies the dN > :
0 if  
2 < rc ðÞ
amount by which an ultrasonic wave is distorted
as it travels through the body. Measurements of ð1Þ
the second harmonic amplitude for a commercial
steel containing 0.25 and 0.35 mass percent Here Nc is a normalizing constant,  is the
carbon were presented by Ogi et al. [10].  is the endurance limit
stress range in a cycle, rc ðÞ
Their data are laboratory measurements on when the mean stress in a cycle is ,  and m > 0
fatigue samples, which indicate an essentially and n > 0 are parameters which depend on the
monotonic increase of the measured second material and the loading conditions. For a
harmonic amplitude with number of fatigue constant loading amplitude, Equation (1) can be
cycles up to the point of macrocrack initiation. solved to obtain
Subsequent to the formation of the macrocrack, 
there is a rapid decrease in the nonlinearity, N  N0
DðNÞ ¼ 1  ð1  D0 Þnþ1 
possibly caused by escape of dislocations as a Nc
macrocrack is formed. Based on these observa-  m 1=ðnþ1Þ

=2  rc ðÞ
tions, the second harmonic amplitude has been  ðn þ 1Þ ð2Þ
=2
chosen to monitor the pre-macrocrack damage
accumulation process during fatigue.
Here D0 is the damage at cycle N0. The cycle at
The sensors that are part of an SHM system
which a macrocrack appears, Nini follows by
can only provide information about the current
substituting D ¼ 1 in Equation (1) and solving for
state of a structure. If damage develops in the
N. The result is
structure either as discrete flaws, or as
distributed material degradation, it is necessary  m
Nc nþ1 =2
to incorporate damage growth models to assess Nini ¼ N0 þ ð1  D 0 Þ
the structural integrity of the structure. nþ1 
=2  rc ðÞ
For discrete macrocracks, Paris law type models ð3Þ
may be applicable where the size of the flaw,
the load level, and material properties are used as In a purely deterministic setting, all the quantities
input to gauge the growth rate of the flaw at appearing in Equation (2) are assumed to be
any given point in the life cycle of the known exactly and hence one can calculate the
component. These models are well established exact number of cycles required for crack
and will not be described here. For distributed initiation using Equation (3). But it is well known
damage heuristic approaches can be used as that the crack initiation process is stochastic in
discussed in [2]. nature and to account for this, some of the
Consider the case of the pre-macrocrack quantities appearing in Equation (2) are taken as
initiation phase of fatigue damage in a metal. random variables with known probability
Let the state of damage in the specimen at a densities. The probability of macrocrack initiation
particular cycle during fatigue be represented by Pma, which is defined as the probability that the
a scalar damage function D(N), where N is the number of cycles to macrocrack initiation, will be
number of cycles. The state D ¼ 0 corresponds to less than a specified number of cycles Ns, is then
no damage, and D ¼ 1 corresponds to the appear- given by
ance of the first macrocrack. A phenomenological
model is needed to represent the evolution of the Pma ¼ PrðNini < Ns Þ ð4Þ

Downloaded from shm.sagepub.com at PENNSYLVANIA STATE UNIV on March 5, 2016


40 Structural Health Monitoring 7(1)

The technique which had been employed in is evaluated by sampling from the density hV ðvÞ
[2] to compute Pma is now summarized. First a centered at u*. In this case, hV ðvÞ is the joint
limit state surface gðXÞ ¼ Ns  Nini ðXÞ is defined normal density with mean u* and standard
where X denotes the random variables in deviations equal to one. Results for a specific
Equation (3). Observe that the region gðXÞ  0 example are given in [2].
corresponds to the region of macrocrack forma-
tion. The probability of macrocrack initiation,
which is the integral of the joint probability 3 Part 2: Fatigue Crack Growth and
density function fX ðxÞ of the random variables the Effect of Multiple Inspections
over the region gðXÞ  0, can therefore be on Fatigue Crack Detection
written as
Z The previous section has now provided the
Pma ¼ PrðNini < Ns Þ ¼ fX ðxÞdx ð5Þ probability of the formation of a macrocrack.
gðXÞ0
The size of the macrocrack has not been speci-
In general, it is difficult to evaluate the integral fied, but it may be assumed that the size of the
in Equation (5) analytically and hence it is macrocrack that appears at the conclusion of the
usually estimated by various different techniques pre-crack damage phase may be described by a
like Monte Carlo integration, Monte Carlo inte- probability density function. Of course it is also
gration with importance sampling, and the first conceivable that such a density may be deduced
order reliability method (see for e.g. [11–13]). from other considerations than presented in the
Equation (5) is chosen to be evaluated using previous section. Attention is now focused on the
Monte Carlo integration with importance sam- growth of a macrocrack with a known density
pling. In this method, the sampling density is function, and particularly on quantifying the
centered in the region of interest so as to improve effects of inspections on the detection of a crack
the efficiency of the simple Monte Carlo method. that has grown under continued cyclic loading.
In this case the sampling density has been centered The authors are especially interested in finding
at the most probable point (MPP) in the trans- the probability that a crack with a critical
formed (u) space. This choice of the sampling dimension greater than a predefined magnitude
center ensures that approximately half the samples exists in the component in the presence of
will be in the failure region and also that they are successive inspections at increasing cycle
concentrated about the point which has the highest numbers.
probability of failure. This lets one use a smaller To proceed, one needs to assume a suitable
number of simulation points when compared to model for the crack growth which accounts for
the direct Monte Carlo integration for which the the variabilities observed in fatigue crack growth.
sampling density is centered at the origin in the Various such models are available in literature
transformed (u) space. Following the procedure (see e.g., [6]). In the present case, for simplicity,
needed for FORM, the random variables X are the authors have represented the crack growth
first transformed into uncorrelated standard by a nonlinear differential equation with fixed
normal variables u. The modified Hasofer-Lind, parameters and have introduced the variabilities
Rackwitz-Fiessler (HL–RF) algorithm described only through the random size of the macrocrack
in Kiureghian and Liu [14] is then used to obtain which appears at the end of the initiation stage.
the point closest to the origin u* on the surface The capability of a specified monitoring techni-
g(u) ¼ 0. Finally the integral (see e.g., [11]) que has been modeled using the probability of
detection (POD) concept. A POD is a statistical
Z
representation of the probability that a given
Pma ¼ PrðNini < Ns Þ ¼ fX ðxÞ dx monitoring technique is able to detect a specific
gðXÞ0
Z Z flaw in a given material or structure. Note that
fU ðvÞ
¼ fU ðuÞ du ¼ hV ðvÞdv the general approach described to account for the
gðUÞ0 gðVÞ0 hV ðvÞ
inspections is similar to the one presented in [15].

Downloaded from shm.sagepub.com at PENNSYLVANIA STATE UNIV on March 5, 2016


Kulkarni & Achenbach SHM and Damage Prognosis in Fatigue 41

In [15], the authors presented expressions for the a


σ σ
crack size distribution after inspections
assuming that a crack detected in an inspection is
repaired. In contrast to the present case where
the parameters in the crack growth rate can be Figure 1 Surface-breaking crack in a tensile field.
randomized, they started with a power crack
growth rate which is randomized by a stochastic
process with a lognormal distribution. from the damage evolution and macrocrack
initiation process described in the preceding
section.
3.1 Two-dimensional Surface- The crack length after N cycles, aN, is given
Breaking Crack by Equation (8) which can be rewritten as:
First the case of a two-dimensional  1=
surface-breaking crack with depth a in a layer of aN ¼ a0 þ NC ð9Þ
thickness much larger than the crack depth,
subject to a uniform tensile field (see Figure 1)
where
is considered. This case is equivalent to a
surface-breaking crack in a half-space whose  pffiffiffi m
C ¼ A 1:12  ,
stress-intensity factor may be written as:
1
pffiffiffiffiffiffi  ¼ ð2  mÞ
K ¼ 1:12 a ð6Þ 2

From Equation (9), one derives


For cyclic loading, the commonly used crack
growth law is Paris law (see [16]), which is  1=
written as: a0 ¼ aN  NC  hðaN Þ ð10Þ

da The probability density function of aN is given by


¼ AðKÞm ð7Þ
dN (see [17])

where N is the number of cycles, da/dN is the



da0

fN ðaN Þ ¼ f0 ½hðaN Þ

ð11Þ
rate of crack growth, A and m are material daN

parameters, and K is the amplitude of the


stress-intensity factor. For constant amplitude which may be written as:
loading and substitution of Equation (6), one can
integrate Equation (7) (m 6¼ 2) to get  1= 

 ð1Þ=

fN ðaN Þ ¼ f0 aN  NC
a1 a  NC


N N

m pffiffiffi m
a1m=2
N ¼ a1m=2
0 þ NA 1  1:12  ð8Þ ð12Þ
2

where a0 and aN are the crack lengths at cycles Note that the probability that there exists a crack
N ¼ 0 and N, respectively. To account for the with length a > acr is given by:
inherent uncertainty in the fatigue behavior, some Z 1
of the quantities appearing in Equation (8) can Prða > acr ; NÞ ¼ fN ðaN ÞdaN ð13Þ
be taken to be random with known probability acr
density functions. For simplicity, the study is
restricted to the case in which the only random The effect of inspections on the probability that
quantity is the initial crack length a0 with density there exists an undetected crack with a > acr is
given by f0 ða0 Þ. This follows naturally since the now discussed. It is assumed that the POD
initial crack length is assumed to have come forth curve of the inspection technique is known.

Downloaded from shm.sagepub.com at PENNSYLVANIA STATE UNIV on March 5, 2016


42 Structural Health Monitoring 7(1)

1
It should be noted that Pr(N1) is the probability
0.9 that there exists an undetected crack after N1
0.8 cycles.
Probability of detection --->

A B C
0.7 To determine the crack length density at
0.6 N > N1, but before the second inspection at N2
cycles, i.e., N > N2, the same steps are followed
0.5
as the ones leading to Equation (15). From
0.4
Equations (8) and (9), one has
0.3
0.2 a ag h i1=
POD(a) = aN ¼ aN1 þ ðN  N1 ÞC ð18Þ
0.1 1+a ag
0
0 2 4 6 8 10
Crack length (a) in mm --->
which can be solved for aN1 as:

 1=
Figure 2 POD curves. A:  ¼ 1.00 mm ,  ¼ 3.0; B:
aN1 ¼ aN  ðN  N1 ÞC  h1 ðaN Þ ð19Þ
 ¼ 0.05 mm ,  ¼ 3.0; C:  ¼ 0.005 mm ,  ¼ 3.00.

The crack length density of an undetected crack


Typical POD curves are shown in Figure 2 where at cycle N > N1 with N < N2 is then given by:
the letters ‘A’, ‘B’, and ‘C’ denote three different
inspection techniques. A convenient expression
for the POD curves is shown in the insert of fN ðaN ; N > N1 Þ
fN ðaN ; N > N1 Þ ¼ ð20Þ
Figure 2 (see [15]). The probability that the crack PrðN1 Þ
is not detected is then given by:
where,
 a 1
PNDðaÞ ¼ 1  ¼ ð14Þ

1 þ  a 1 þ  a
daN1

fN ðaN ; N > N1 Þ ¼ fþ ½ h ða Þ; N 
1



ð21Þ
Now the case of a first inspection at cycle N1 is
N1 1 N
daN

considered. Just prior to the inspection,


 the crack
length density is given by fN1 aN1 , which follows Using
from Equation (12) by replacing all Ns with N1s.
Just after the inspection, the crack length density daN1 h  ið1Þ=
1
of the undetected crack is ¼ aN  ðN  N1 ÞC aN
daN ð22Þ
 ¼ ½h1 ðaN Þ1 a1
 fþ
N1 aN1 ; N1
N

N1 aN1 ; N1 ¼ ð15Þ
PrðN1 Þ
Equation (21) is simplified to
where
   fN ðaN ;N > N1 Þ ¼ PND½h1 ðaN Þ

N1 aN1 ; N1 ¼ PND aN1 fN1 aN1 ð16Þ n 1= o
 f0 h1 ðaN Þ N1 C

In Equation (16), PND aN1 follows from


ð1Þ=

Equation (14), and 


h1 ðaN Þ1 h1 ðaN Þ N1 C

1  

PrðN1 Þ ¼ PND aN1 fN1 aN1 daN1 ð17Þ 


h1 ðaN Þ1 a1
N
ð23Þ
0

Downloaded from shm.sagepub.com at PENNSYLVANIA STATE UNIV on March 5, 2016


Kulkarni & Achenbach SHM and Damage Prognosis in Fatigue 43

which can be further simplified to Note that PrðN2 Þ is the probability that there
exists an undetected crack after the inspection at
fN ðaN ; N > N1 Þ ¼ PND½h1 ðaN Þ cycle N2 given that it was not detected during
 1=  the inspection at cycle N1. Substitution of
 f0 aN  NC Equation (24) yields

 ð1Þ=



  


aN  NC aN

1
ð24Þ fþ
N2 aN2 ; N2 ¼ PND aN2 PND h1 aN2
 1= 
 f0 aN2  N2 C
The probability that there exists a crack with
aN > acr at cycle N, which was undetected at the
 ð1Þ=

inspection at cycle N1 can then be written as:


aN2  N2 C 1

aN2
ð30Þ

PrðaN > acr , N  N1 ; NDÞ


Z1 For N > N2 , one needs to go once more through
¼ PrðN1 Þ fN ðaN ; N  N1 ÞdaN the same steps. Then
acr
Z 1  1= 
h i1=
¼ PND½h1 ðaN Þf0 aN  NC
acr aN ¼ aN2 þ ðN  N2 ÞC ð31Þ

 ð1Þ=


aN  NC 1

aN
daN ð25Þ
and

The next inspection is at N2 cycles. Just prior  1=


to the inspection at N2, the crack length density aN2 ¼ aN  ðN  N2 ÞC  h2 ðaN Þ ð32Þ
follows from Equation (20) as
 In addition,
 fN2 aN2 ; N2 > N1
fN2 aN2 ; N2 > N1 ¼ ð26Þ
PrðN1 Þ
daN2 1
 ¼ ½h2 ðaN Þ1 aN ð33Þ
where fN2 aN2 ; N2 > N1 follows from Equation daN
(24) by replacing N by N2. Just after the inspec-
tion, the crack length density of the undetected Analogous to Equations (20) one finds using
crack is Equation (30)

fNþ2 ðaN2 ; N2 Þ fN ðaN ; N > N2 Þ



N2 ðaN2 ; N2 Þ ¼ ð27Þ fN ðaN ; N > N2 Þ ¼ ð34Þ
PrðN1 Þ PrðN2 Þ PrðN1 Þ PrðN2 Þ

where
where
  

N2 aN2 ; N2 ¼ PND aN2 fN2 aN2 ; N2 > N1 ð28Þ
fN ðaN ; N > N2 Þ ¼ PND½h1 ðaN ÞPND½h2 ðaN Þ
 1= 
and 
 f0 aN  NC
Z 1  


PrðN2 Þ ¼ PND aN2 fN2 aN2 ; N2 > N1 daN2


ð1Þ=

0 

aN  NC a1
N


ð35Þ
ð29Þ

Downloaded from shm.sagepub.com at PENNSYLVANIA STATE UNIV on March 5, 2016


44 Structural Health Monitoring 7(1)

Also, 190,000 using the three monitoring techniques


with PODs labeled ‘A’, ‘B’, and ‘C’ (see
PrðaN > acr , N  N2 ; NDÞ ¼ PrðN1 Þ PrðN2 Þ Figure 2). The probability of interest is given by
Z1 Equation (37) and is evaluated using numerical
 fN ðaN ; N  N2 ÞdaN integration. The integral defined over the infinite
acr
Z1 h i interval ðacr , 1Þis first transformed to a finite
¼ PND½h1 ðaN ÞPND½h2 ðaN Þ f0 ðaN  NCÞ1= interval (0,1) using the transformation
acr


x ¼ eðaacr Þ and then the adaptive quadrature

 ð1Þ=


1
routines described in Espelid [19] are used to

aN  NC aN
daN
evaluate the integral. The results are shown in
ð36Þ Figure 3.
The solid lines in Figure 3 represent the
Note that the terms PrðN1 Þ and PrðN2 Þcancel out probability that a crack with length a > acr exists
since they appear both in the numerator and the in the component at cycle N, and is calculated
denominator. It is now straightforward to gener- using Equation (13). The dotted lines represent
alize the result given by Equation (36) to n the probability that a crack with length a > acr
inspections at cycles N1 , . . . , Nn : which is undetected in all the previous inspections
exists at cycle N. As expected, at any given cycle,
the probability is smallest for the inspection
PrðaN > acr , N  Nn ; NDÞ
Z 1  1= 

 ð1Þ=
technique with the best POD, which is labeled ‘A’

¼ f0 aN  NC
a  NC a1
in Figure 2. It can also be seen from the figure

N N

acr that immediately following an inspection, this


Y
n probability drops and then increases till the next
 PND½hi ðaN ÞdaN ð37Þ inspection is carried out. This is again to be
i¼1
expected as the crack continues to grow
in-between inspections. It is noted that poor
This expression denotes the probability that a inspection procedure with low POD, such as C,
crack with length a > acr exists at cycle N with has very little effect on the probability of an
N > Nn and was undetected by all the inspections undetected critical crack. It is important to
at Ni , i ¼ 1 . . . n. point out that inspections do not change the
An example of a surface-breaking crack and probability that there exists a crack with a > acr .
a Paris law defined by (see Moran et al. [18]) This probability will change only if a repair is
m ¼ 2:67, A ¼ 5:069  1012 ,  ¼ 280 MPa, carried out based on the inspections. The inspec-
and R ¼ 0 is now considered. It is also assumed tions only affect the probability that there exists
that the initial crack length a0 has a lognormal an undetected crack with dimension greater than
distribution with density a predefined magnitude.
  For cases where a closed form solution to
1 1 the growth law does not exists (e.g., Forman
f0 ða0 Þ ¼ pffiffiffiffiffiffi exp  2 ðlnða0 Þ  Þ2
a0  2 2 equation), or if one is dealing with a large set
of dependent random variables, it might be
Here m and  are the mean and the standard necessary to study the effect of inspections using
deviation of the random variable y ¼ lnða0 Þ. It is Monte Carlo simulation. A brief description of
assumed that the initial crack length a0 has mean the Monte Carlo simulation applied to the
0.250 mm, since this was the macrocrack size problem presented earlier follows. This starts
initiated in the experiments reported by Ogi et al. with a large sample of crack lengths which are
[10]. The standard deviation was taken to be distributed according to the density function
0.1 mm. In this example acr was chosen as f0 ða0 Þ. For 0 < N < N1 , each of these samples is
acr ¼ 3 mm. The inspections are carried out at evolved according to Paris law and the length at
cycles 50,000, 100,000, 130,000, 160,000, and cycle N is calculated. Then the probability of

Downloaded from shm.sagepub.com at PENNSYLVANIA STATE UNIV on March 5, 2016


Kulkarni & Achenbach SHM and Damage Prognosis in Fatigue 45

100

10−5

10−10
POD A

10−15
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
× 105
100

10−5

10−10 POD B

10−15
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
× 105
100

10−10
POD C

10−20
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
× 105
N --->

Figure 3 Probability of an undetected crack with length a > acr, as N increases.

interest, Prða > acr ; N, NDÞ, which is identical to sample size. Since the authors are interested in
Prða > acr ; NÞ up to the first inspection, is given calculating the probability of an undetected crack
by the ratio of the number of samples with exceeding the critical length, they consider only
a > acr to the total sample size. This procedure is the ‘not detected’ group for future calculations.
followed right up to the first inspection at N1. They then essentially follow the procedure
At N1, each of the samples is subjected to an described previously except that they now start
inspection. This is done as follows (see e.g., Hahn with the ‘not detected group’. Though this
and Shapiro [17]). First the probability of method of studying the effects of inspections
detecting a crack of length aN1 is found from the using Monte Carlo simulation is straightforward,
formula given in the insert in Figure 2. it does however suffer from a drawback – a large
This probability is denoted by P1. A uniform number of samples are needed to get an accurate
variate, u, is then picked over the interval (0, 1). estimate of the probability and the smaller the
The probability that u is in the interval (0, P1) is probability to be estimated, the larger the
P1 and the probability that it is in the interval required sample size.
ðP1 , 1Þ is 1  P1 . Therefore, it is considered that Table 1 displays the results obtained by
the crack is detected if u  P1 and ‘not-detected’ solving the example described earlier by numeri-
otherwise. Depending on the result of the inspec- cally integrating Equation (37) and by using
tion, it is placed either in the ‘detected’ or ‘not Monte Carlo simulation. The Monte Carlo simu-
detected’ group of samples. Prða > acr ; N, NDÞ lation was performed by using 109 samples.
immediately following the first inspection is given The () sign in the superscript denotes that the
by the ratio of the number of samples with result was calculated just prior to an inspection
a > acr in the ‘not detected group’ to the total while the (þ) sign denotes that the result was

Downloaded from shm.sagepub.com at PENNSYLVANIA STATE UNIV on March 5, 2016


46 Structural Health Monitoring 7(1)

Table 1 Comparison of the Monte Carlo simulation with numerical integration of Equation (37).

POD A POD C
N Equation (37) Monte Carlo % Error Equation (37) Monte Carlo % Error

50,000 5.8789e007 6.2100e007 5.6320 5.8789e007 6.2100e007 5.6320
50,000þ 1.6490e008 2.0000e008 21.286 4.9681e007 5.0800e007 2.2524
100,000 5.6190e005 5.6374e005 0.3275 3.1510e004 3.1561e004 0.1622
100,000þ 1.4880e006 1.4470e006 2.7554 2.5733e004 2.5766e004 0.1279
130,000 1.2071e004 1.2087e004 0.1334 3.6497e003 3.6481e003 0.0433
130,000þ 3.1930e006 3.2560e006 1.9731 2.8931e003 2.8927e003 0.0127
160,000 2.4465e004 2.4529e004 2.5996 2.1788e002 2.1786e002 0.0071
160,000þ 6.4392e006 6.3870e006 0.8107 1.6561e002 1.6558e002 0.0153
190,000 4.3600e004 4.3620e004 0.0454 7.6804e002 7.6780e002 0.0313
190,000þ 1.1364e005 1.1220e005 1.2672 5.5199e002 5.5178e002 0.0384
200,000 5.1549e005 5.1381e005 0.3259 8.7212e002 8.7189e002 0.0267

calculated just after the inspection. It can be seen developed by Newman and Raju [20] which is
from the table that for the example considered, also employed in this study.
acceptable accuracy is obtained with the Monte Consider a plate with a semi-elliptical crack
Carlo simulation when the probability to be under tension fatigue loads as shown in Figure 4.
estimated is greater than 106. The estimate of The initial depth and the half length of the
the Monte Carlo simulation can be improved by crack are denoted by a0 and c0, respectively.
increasing the sample size. The crack is assumed to growth but remain
If the mean of the initial crack size semi-elliptical with semiaxes a and c. The fatigue
increases, then for the same loading, Paris law crack growth rates at points A and B on the
parameters and POD, the same inspection cycles crack front (see Figure 4) are assumed to
and the same standard deviation, the probability independently obey the Paris relations,
that there exists an undetected crack in the
presence of inspections will decrease. If the da
¼ CA ðKA Þn
standard deviation of the initial crack size dN
increases, then for the same loading, Paris law dc
¼ CB ðKB Þn ð38Þ
parameters, POD, same inspections cycles and dN
the same mean, the probability that there exists
an undetected crack in the presence of inspec- where K is the stress-intensity range at point A
tions will increase. or B. The Paris exponent is assumed to be
the same for growth in the depth and surface
directions. The variation in fatigue resistance
3.2 Semi-elliptical Surface-
along the crack front due to variation in stress
breaking Crack
field triaxiality is accounted for by the different
The discussion can be extended in a straight coefficients CA and CB, specifically
forward manner to a semi-elliptical surface-
breaking crack to describe the effect of inspec- CA ¼ 1:1n CB
tions on the probability that there exists an
undetected critical crack. The only point of The empirical equation developed by Newman
difference is that one now has to evolve the and Raju [20] for the stress-intensity factor for a
semi-elliptical surface as opposed to the crack semi-elliptical surface crack in a finite plate
depth. In the literature (see e.g., [20,21]), the subjected to tension and bending is given by
evolution of a semi-elliptical surface-breaking rffiffiffiffiffiffiffiffi 
crack in a rectangular plate is typically modeled a a a c 
KI ¼ P  F , , , 
using the empirical stress intensity factor Q t c b

Downloaded from shm.sagepub.com at PENNSYLVANIA STATE UNIV on March 5, 2016


Kulkarni & Achenbach SHM and Damage Prognosis in Fatigue 47

2b
A t

φ a
2h O B
2c
2c
Z
Y

Figure 4 Semi-elliptical surface-breaking crack load under tension fatigue load.

where 0 < a=c  1:0, 0  a=t < 1:0, c=b < 0:5, and inherent uncertainty in the initial depth and
0    . Here Q is given by length, some of the quantities in Equation (38)
a1:65 are considered to be random. Again, for simpli-
Q ¼ 1 þ 1:464 a=c  1:0 city, the authors restrict to the case in which
c the only random quantities are the initial crack
The function F is taken to be depth a0 and the initial crack length c0 with the
 joint density given by f0 ða0 , c0 Þ. In the case of
a2 a4  semi-elliptical cracks one can either use the
F ¼ M1 þ M2 þM3 f g fw
t t length or the depth, or the area of the crack as
a governing criterion to calculate the desired
where, probabilities. It is more important to detect deep
a cracks during inspections and hence the depth
M1 ¼ 1:13  0:09 of the crack is used to define the governing
c criteria to calculate the probabilities. Since
0:89
M2 ¼ 0:54 þ explicit solutions to the crack depth and crack
0:2 þ ða=cÞ length are not available, the authors resort to
1:0  a24
M3 ¼ 0:5  þ 14 1:0  Monte Carlo simulation to study the effects of
0:65 þ ða=cÞ c inspections.
  1=4
a 2 2 Now a specific example of a semi-elliptical
f ¼ cos  þ sin2  surface-breaking crack with Paris law defined
c
 a2  by  ¼ 280 MPa, R ¼ 1, n ¼ 3.0, t ¼ 5 mm,
g ¼ 1 þ 0:1 þ 0:35 ð1  sin Þ2 b ¼ 20 mm, h ¼ 20 mm, CB ¼ 2:5  1011 ,
t
  rffiffiffi1=2 a0 =c0 ¼ 0:25, and acr ¼ 1 mm is considered. The
c a initial crack depth a0 is assumed to have a
fw ¼ sec
2b t lognormal distribution with mean 0.25 mm and
standard deviation 0.1 mm. The inspections are
The growth of the semi-elliptical crack surface is carried out at cycles 5000, 10,000, 15,000, 20,000,
then evolved by numerically solving the coupled 25,000, and 30,000 using the three monitoring
differential equations given by Equation (38) techniques with PODs labeled ‘A’, ‘B’, and ‘C’
subject to the initial conditions given by (see Figure 2). The results of the Monte Carlo
að0Þ ¼ a0 and cð0Þ ¼ c0 . To account for the simulation are shown in Figure 5.

Downloaded from shm.sagepub.com at PENNSYLVANIA STATE UNIV on March 5, 2016


48 Structural Health Monitoring 7(1)

0.06

POD A
0.04

0.02

0
0 0.5 1 1.5 2 2.5 3 3.5 4
× 104
0.06

POD B
0.04

0.02

0
0 0.5 1 1.5 2 2.5 3 3.5 4
× 104
0.06

POD C
0.04

0.02

0
0 0.5 1 1.5 2 2.5 3 3.5 4
N ---> × 104

Figure 5 Probability of an undetected crack with depth a > acr, as N increases.

The solid lines in Figure 5 represent the evolution law which can be suitably adapted to
probability that a crack with depth a > acr exists account for the stochastic nature of macrocrack
in the component as a function of number of initiation process is presented. An expression for
cycles. The dotted lines represent the probability the probability of macrocrack initiation along
that a crack with depth a > acr which is with a numerical procedure to evaluate it is also
undetected in all the previous inspections exists at presented.
cycle N. As expected, at any given cycle, The technique to quantify the effect of
this probability is smallest for the inspection imperfect inspections takes into account the
technique with the best POD, which is labeled ‘A’ probability of detection (POD) curve of the
in Figure 2. specified monitoring method. The probability
that there exists an undetected critical crack is
considered to be the quantity of interest. Explicit
4 Conclusions expression for this probability is obtained for
the case of a two-dimensional surface-breaking
In this study, two critical aspects of a typical crack. A method based on Monte Carlo simula-
SHM system used for monitoring a fatigue crack tion to evaluate this probability is also presented.
have been discussed: (i) a procedure to predict It has been applied to the problem of a semi-
macrocrack initiation and (ii) a technique to elliptical surface breaking crack. Crack growth
quantify the effect of imperfect inspections. laws based on the Paris law are used for both
The procedure to predict macrocrack the cases. Results obtained for sample problems
initiation is based on monitoring the acoustic have been presented. The effect of the different
nonlinearity in the material during the damage PODs on the probability of interest is also
accumulation process. A candidate damage presented.

Downloaded from shm.sagepub.com at PENNSYLVANIA STATE UNIV on March 5, 2016


Kulkarni & Achenbach SHM and Damage Prognosis in Fatigue 49

Work which extends the present study and 8. Suresh, S. (1991). Fatigue of Materials, Cambridge:
which deals with optimizing the inspection cycle Cambridge University Press.
numbers so as to reduce the cost function 9. Hikata, A. and Elbaum, C. (1966). Generation of
associated with a undetected critical crack is ultrasonic second and third harmonics due to
dislocations. Physical Review, 144(2), 469–477.
currently in progress.
10. Ogi, H., Hirao, M. and Aoki, S. (2001). Noncontact
monitoring of surface wave nonlinearity for predicting
the remaining life of fatigue steels. Journal of Applied
Acknowledgments Physics, 90(1), 438–442.
11. Bjerager, J. (1989). Probability computation methods in
The work presented in this study was carried out in the structural and mechanical reliability. In: Liu, W.K. and
course of research funded by the Federal Aviation Belytschko, T. (eds), Computational Mechanics of
Administration on Grant 01-C-AW-NWU/A010, moni- Probabilistic and Reliability Analysis, Lausanne:
tored by Jim White. Although the FAA has sponsored this Elempress International, Vol. 47, p. 68.
project, it neither endorses nor rejects the findings of this 12. Madsen, H.O., Krenk, S. and Lind, N.C. (1986).
research. The presentation of this information is in Methods of Structural Safety, Englewood Cliffs,
the interest of invoking technical community comment New Jersey: Prentice-Hall Inc.
on the results and conclusions of the research. 13. Melchers, R.E. (1999). Structural Reliability Analysis
and Prediction, 2nd edn, New York: John Wiley and
Sons.
14. Kiureghian, A. and Liu, P. (1989). Finite-Element
reliability methods for geometrically nonlinear
References stochastic structures. Report No. UCB/SEMM-89/05,
University of California, Berkeley: Dept. of Civil
1. Farrar, C.R., Phillip, J.C., Hunter, N.F. and Lieven, N. Engineering.
(2005). Sensing and data acquisition systems, issues for 15. Palmberg, B., Blom, A.F. and Eggwertz, S. (1987).
damage prognosis. In: Inman, D.J., Farrar, C.R., Lopes Probabilistic damage tolerance analysis of aircraft
Jr, V. and Steffen Jr, V. (eds), Damage Prognosis for structures. In: Provan, W. (ed.), Probabilistic Fracture
Aerospace, Civil and Mechanical Systems, Chichester: Mechanics and Reliability, Boston: Martinus Nijhoff
John Wiley and Sons Ltd, pp. 307–312. Publishers.
2. Kulkarni, S.S., Sun, L., Moran, B., Krishnaswamy, S. 16. Paris, P.C. and Erdogen, F. (1963). A critical analysis of
and Achenbach, J.D. (2006). Probabilistic method to crack propagation law. Journal of Basic Engineering,
predict fatigue crack initiation. International Journal of 85(Dec), 528–534.
Fracture, 137(1–4), 9–17. 17. Hahn, G.J. and Shapiro, S.S. (1967). Statistical Models
3. Bogdanoff, J.L. and Kozin, F. (1995). Probabilistic in Engineering, New York: John Wiley & Sons.
Models of Cumulative Damage, New York: John Wiley. 18. Moran, B., Xu, Y. and Achenbach, J.D. (2000).
4. Yang, J.N. and Manning, S.D. (1990). Stochastic crack Limit-state surface element method: application to
growth analysis methodologies for metallic structures. fatigue reliability with NDE inspections. Journal of
Engineering Fracture Mechanics, 37(5), 1105–1124. Engineering Mechanics, 126(7), 684–692.
5. Yang, J.N., Hsi, W.H. and Manning, S.D. (1987). 19. Espelid, T.O. (2002). Doubly adaptive quadrature
Stochastic crack growth models for application to routines based on Newton-Cotes rules. Report No.
aircraft structures, In: Provan, J.W. (ed.), Probabilistic 229, University of Bergen, Norway: Dept. of
Fracture Mechanics and Reliability, Boston: Martinus Informatics.
Nijhoff. 20. Neuman, J. Jr. and Raju, I. (1981). An empirical stress-
6. Sobczyk, K. and Spencer Jr, B.F. (1992). Random intensity factor equation for the surface crack.
Fatigue: From Data to Theory, New York: Academic Engineering Fracture Mechanics, 15(1–2), 185–192.
Press, Inc. 21. Shang-Xian, W. (1985). Shape change of surface cracks
7. Klesnil, M. and Lukáš, P. (1992). Fatigue of Metallic during fatigue growth. Engineering Fracture Mechanics,
Materials, 2nd edn, New York: Elsevier. 22(5), 897–913.

Downloaded from shm.sagepub.com at PENNSYLVANIA STATE UNIV on March 5, 2016

You might also like