Engineering Fracture Mechanics Vol. 53, No. 5, pp.

687-706, 1996
Copyright © 1996 Elsevier Science Ltd.
Pergamon 0013-7944(95)00131-X Printed in Great Britain. All rights reserved
0013-7944/96 $15.00 + 0.00

A STOCHASTIC SYSTEMS APPROACH TO FATIGUE

R E L I A B I L I T Y - - A N APPLICATION TO Ti-6A1-4V
KHALIL F A R H A N G D O O S T
Department of Mechanical Engineering, McGill University, Montr6aL Qu6bec, Canada
JAMES W. PROVAN
Department of Mechanical Engineering, University of Victoria, Victoria, British Columbia, Canada

Abstract--Reliability analyses are proposed for the life prediction of fatigue sensitive components. In
essence, the models are based upon the recently developed McGill-Markov and closure-lognormal
stochastic processes. Not only do these models have the capacity of predicting the statistical dispersion
of crack growth, they also, by incorporating the concept of crack closure, have the capability of
transferring stochastic fatigue properties, measured under ideal laboratory conditions, to situations of
industrial significance, such as those occurring under adverse loading and/or environmental conditions.
The primary data required for an estimation of the pertinent parameters governing these stochastic models
are obtained from a statistically significant number of replicate tests. Both the theory and experimental
technique are illustrated using a Ti-6AI-4V alloy of considerable importance to the aerospace industry.
Finally, the application of these models to the assessment of structural integrity and component reliability
is detailed in this paper.

1. INTRODUCTION
WHILEMOSTindustrial failures involve fatigue, the assessment of the fatigue reliability of industrial
components being subjected to various dynamic loading situations is one of the most difficult
engineering problems that remains. This is because material degradation processes due to fatigue
depend upon material characteristics, component geometry, loading history and environmental
conditions. As a result, the concepts of Reliability, Availability and Maintainability (RAM) have
been brought into play to control, within specified performance limits, failures of all types of
components and systems.
The seminal papers on structural reliability of Freudenthal [1] and Weibull [2] regarded scatter
observed in fatigue failure as stochastic phenomena. As both fracture and fatigue failures became
more critical in higher performance structures other engineers adopted statistical models of
increasing complexity. These models utilized known statistical distributions such as the exponential,
normal, lognormal, gamma, Weibull and Gumbel. In the past few years, more research effort has
been devoted to the derivation of reliability models based on probabilistic interpretations of the
fatigue process. Thus, substantiated by several sets of fatigue data, Birnbaum and Saunders [3], and
Freudenthal and Shinozuka [4] proposed similar life distributions to characterize fatigue crack
extension failures. Subsequently, Payne [5] introduced a statistical reliability model for assessing
the fatigue strength of aircraft structures by evaluating the random variability in crack propagation
rates and residual strengths of cracked structures at any stage during their evolution.
Provan et al. [6-8] have derived and experimentally investigated new reliability distributions
based on probabilistic micromechanic concepts applied to the fatigue failure of polycrystalline
metals. Furthermore, Bogdanoff and Kozin [9], and Yang et al. [10] have investigated the stochastic
fatigue crack growth process based, respectively, on Markov and lognormal processes,
interpretations that are incorporated into the current investigation.
Turning attention to the fatigue process itself, Elber [11] observed that fatigue-crack surfaces
contact each other even during tension-tension cyclic loading. This simple observation of the
crack-closure phenomenon explained many crack growth characteristics that had been troubling
researchers during the 1960s. Since then, several closure mechanisms have been identified, among
them, "plasticity" induced closure. These new closure mechanisms and the influence of the plastic
wake on the local crack-tip strain field have greatly advanced an understanding of the fatigue-crack
growth and fracture behaviour of metallic materials, under, for example, variable amplitude
loading. Furthermore, the occurrence of crack-closure significantly affects the local crack driving
~ 5~/s--B 687
688 K. FARHANGDOOSTand J. W. PROVAN

force stress intensity factor and plays a crucial role in quantifying the fatigue crack growth or arrest
characteristics of the material in question. Thus, while the qualitative interpretation of the closure
phenomena justifies a large number of fatigue crack anomalies, a quantitative knowledge of the
crack closure stress intensity level is required in order to incorporate fatigue crack growth data
into the design of engineering components.
Most fatigue design requirements based on damage tolerance concepts assume the existence
of flaws in a component. Such flaws may either exist in the initially delivered component or develop
during its service life. To assure a high level of reliability and before these flaws grow to critical
lengths, it is necessary to either repair or replace the component. Thus, in-service inspections are
required to detect the various sizes and shapes of cracks and other defects. For a reliability analysis
of an in-service airframe, for example, fatigue, environmental and accidental damage are three
sources that must be taken into consideration.
Hence, the important motivational aspects of fatigue, the unifying nature attributed to the
closure phenomenon, the need for a stochastic process interpretation of the whole fatigue process
and the need for in-depthreliability analyses are clear. The remainder of this paper systematically
elaborates on these topics. Indeed, while crack closure is briefly detailed in the next section and
probabilistic approaches to fatigue are presented in the following, this paper investigates new
methods of assessing the fatigue reliability of components based upon probabilistic approaches.
Its objectives may be summarized as:

(1) the experimental determination of a statistically significant number of crack growth rates
for a Ti-6A1-4V titanium alloy of importance to the aerospace industry (Section 4);
(2) the determination of the closure-lognormal, c, m and ~2, ~ parameters for Ti-6A1-4V
(Section 5);
(3) the simulation of the crack propagation based upon these closure-lognormal parameters
(Section 6);
(4) the determination of the McGill-Markov 2 and x parameters for the Ti-6A1-4V alloy in
question (Section 7); and
(5) the RAM assessment of Ti-6A1-4V using the McGill-Markov model (Section 8).

2. FATIGUE CRACK CLOSURE

The fatigue crack closure concept, definition, mechanisms and measurement techniques are
briefly described and reviewed in this section.
An understanding of crack closure under cyclic tension was developed by Elber [11]. He
showed that the occurrence of premature contact between the opposing crack faces during
unloading was due to residual plastic stretch in the crack wake. During the loading portion of
a cycle, the elastic constraints acting on the residual material in the wake of an advancing crack,
keep the crack tip closed until these constraints are overcome by the externally applied load.
Figure 1 shows one period of a typical cyclic load-displacement response of Ti-6A1-4V which
illustrates closure. While an ideal crack exhibits a linear elastic response, passing through the
origin, the actual response is nonlinear. The opening load level, Pop, lies between linear and
nonlinear parts of the load-displacement curve. Hence, the nonlinear part of this curve is due to
the closure effect.
As discussed in ref. [12], the stress intensity factor, Kop, associated with a quantitative
knowledge of the crack opening load, Pop, is required for the determination of the effective stress
intensity range factor, AKc~ defined by:

AK,~ = Kmax Kop.

- - (1)

AK,n is, in turn, an important parameter for estimating the crack growth rate through, for example,
a suitably modified Paris-Erdogan relationship [13]:

da(t) = c(AK¢~)m' (2)

dN
 Stochastic systems approach to fatigue reliability 689

where da(t)/dN represents the rate at which a crack of dimension a(t) grows, and c and m are
material parameters which characterize the material response.
AK~, is an appropriate field parameter for relating crack growth rates under
constant-amplitude loading conditions to the influence of a number of variables that are
known to influence the rate at which a crack grows. Specifically, knowing the crack growth
rate in terms of AKe~ for a specific material, such as Ti-6A1-4V, the influence of an actual
load spectrum or load ratio for a specific environment may be incorporated into an estimate
of the fatigue reliability of a specific component. As an example, fatigue crack closure effects
at different load ratios have been extensively investigated by Ritchie et al. [14], who initially
showed a significant difference in the (da/dN)-AK behaviour due to different load ratios over
a wide range of AKs, but when plotted as a function of AKe, were able to consolidate the
curves into a narrow band. Similar results over a broad range of AK and crack growth rates
have been observed in Ti-6A1-4V under different air environments[15] and different heat
treatments [16].
It is as a result of the unifying properties of crack closure that this concept is incorporated
into the lognormal stochastic description of the crack growth rate that is developed and explained
in Section 5. As will be shown, the combination of crack closure with a stochastic interpretation
of fatigue crack growth may contribute significantly to reliable design procedures.
Finally, several experimental methods of characterizing crack closure have recently been
reviewed [17-20]. The proposed techniques are primarily based on:

• direct observations of crack closure,

• Crack Mouth Opening Displacement (CMOD) compliance measurements,
• indirect observations based on fatigue crack growth, and
• the Rigid-Insert Crack Closure (RICC) Model.

As is detailed in Section 4 of the current investigation, closure is established on the basis of

a CMOD technique. The closure load is found on the basis of the minimum value of the comparison
between a fitted curve and a linear regression of the linear portion of the load-CMOD response.
Figure 2 depicts typical experimental load-CMOD behaviours of Ti-6AI-4V for two different crack
lengths, with the upper and lower limits indicating the range of the linear regression analysis
obtained during the test.

Pmax 8 ¸ ................................ ?.......

Pop1
P.~. .,~.oO" ",°oo,o,O* F)n~/1

I , I , I * I , I ~ I , 0 i I , I , I , I
O0 100 t10 110 410 140 110 O.ml ~ ~ ~I0o

TIME (mlll4eoond) CMOD (ram)

Fig, I. A schematic of the fatigue crack closure response of Ti-6AI-4V.
 690 K. FARHANGDOOST and J. W. PROVAN

I c~cKsm" 1~4"~ I /
CLOeUMLO~- I=0~ I /
0

g5% .........i ............................/ ...........

4 upperlimit 4 upperlimit/
~=
O, / Iow]rlimit / lowerlimit
2

Pop
1 1

0 o
0.01 0.0~ 0.03 0.04 O.OG 0.0e 0.07 O.no 0.04 0.1~ O.¢le 0.1
CMOD (turn) CMOD (mm)
Fig. 2. A typical load-CMOD plot for Ti-6AI-4V.

3. S T O C H A S T I C P R O C E S S E S AND RELIABILITY
Stochastic processes describe random phenomena with the property that under a particular
set of conditions their observation leads to a multitude of possible outcomes in such a way that
their statistical regularity may be determined.
3.1. The lognormal process
The aim of the present study of the lognormal process is to consider fatigue crack growth rates
as random phenomena. In reliability investigations, common distribution functions are often used
to fit experimental data. Two are briefly introduced here because of their importance to the further
development of this paper.
To formulate mathematical models for random phenomena with reasonable precision, it is
helpful to start with an understanding of the normal distribution. A random variable X is said to
have a normal distribution with mean/~ and variance #2 if its probability density function is given
by:
1
f(x) - #x/~ e -I(-'-")-~/2e21, -- oo < x < ~ . (3)

A normal distribution with p = 0 and 6 2 = 1 is called a standard normal distribution and may be
derived from f ( x ) by replacing X by Z = ( X - bt)/6. The cumulative distribution function of the
standardized normal probability distribution is usually denoted by ~(z), which is defined for any
real z by:

• (z)- 1 fl e - .-2/2dz. (4)

The normal process is often assumed to play a significant role in the stochastic analysis of
random phenomena, such as the crack growth rate vs stress intensity factor. Hence, a stochastic
process x(t) is said to be a normal (or Gaussian) process if for any t,, t2. . . . . t,, the random
variables x(tO, x(t2) . . . . . x(t,) are jointly normally distributed. For a stationary normal process
the mean function is constant and the covariance function, Cov[x(t), x(t + z)], depends only on
the time difference z.
 Stochastic systemsapproach to fatigue reliability 691

A random variable X has a lognormal distribution if In X is normal and its probability density
function is given by:
1 (Inx - #)2
f(x)- 62x//~e- 2,-' , 0 < x < m . (5)

The lognormal distribution has the two parameters, # and #2, and has a variety of properties which
are useful in the analysis of reliability.
3.1.1. Applications of the lognormal process. The validity of lognormal crack growth rate
models, including the lognormal random process, white noise and random variable models along
with the general lognormal random process model, have been investigated using extensive fatigue
crack growth data gathered from fastener hole specimens [10]. For any lognormal random process
model, an additional parameter appearing in the autocorrelation function is determined from the
experimental test results. The estimation of such a correlation parameter indicates the stochastic
property of the crack growth rate.
Deterministic models for fatigue crack propagation, which are based upon the principles of
fracture mechanics, are generally of the form [21]:

da(t) _ F(AK,Km~,R,S,a), (6)

dt
where a(t) is the deterministic crack size, t is time or number of cycles, F is a non-negative function,
AK is the stress intensity factor range, Km,x is the maximum stress intensity factor, S is the stress
amplitude and R is the stress ratio. Some commonly used crack growth rate functions are the
Paris-Erdogan [13] [see eq. (2)], the Forman [22] and the Larsen models [23].
In order to account for the random nature of crack growth rate, however, several researchers
(Lin and Yang [24], L i n e t al. [25], Sobcyzk [26], Yang et al. [10]) have suggested the following
model for fatigue loading situations:
da(t)
dt - X(t)F(AK, Km~x,R,S,a), (7)

where a(t) is the random crack size and X(t) is a non-negative random process. Based on
experimental results, Yang et al. [10] proposed X(t) to be a non-negative, stationary lognormal
random process with median unity. The lognormal random process, X(t), is defined by its logarithm
being a normal random process, i.e. Z(t) is a normal random process, where:
Z(t) = logX(t). (8)
The stationary normal random process Z(t) is defined by the mean value p- and the autocorrelation
function R:-(T) between Z(t) and Z(t + ~) which is given by:
R:-(z) = E[Z(t)Z(t + z)]. (9)
The autocorrelation function plays a significant role in random process analysis and specifies the
statistical behaviour of the random process. In general, it is referred to as an ensemble of a
stochastic process and is given by:
Rx,(t,,t2) = E[x(h)x(t2)]

- - [kx(t,)kx(t2)l, (10)
n k=l

where k is the report number and n is the total number of reports. For a stationary normal random
process, the autocorrelation function Rx.~(t~t2) becomes a function of the time difference only, i.e.
R.~x(O = E[x(t)x(t + ~)]

1 ~ [kx(t)kx(t + T)], n ~ . (11)

O k= I
692 K. F A R H A N G D O O S T and J. W. PROVAN

The mean value, #:, of Z(t) is equal to the logarithm of the median value of X(t). Since the median
value of X(t) is equal to unity, the mean value #: of Z(t) is equal to zero, i.e. #: = log(/zx) =0.0.
Hence, Z(t) is a stationary normal random process with zero mean and is completely defined by
the autocorrelation function R::(O.
According to Yang et al. [10], the autocorrelation function R=(z) of the normal random fatigue
crack growth process Z(t) is an exponentially decaying function of time difference z, as follows:
R::(z) = ffZe- ¢ITI (12)
where ¢ is the correlation parameter estimated from test results. This correlation parameter plays
a significant role in describing the statistical fatigue crack growth rate and propagation behaviour
of materials.
Within the class of random processes Z(t) or X(t), there are two extreme cases. At one extreme
when ¢ ~ oo, the autocorrelation function becomes a Dirac delta function as follows:
R=(z) = 626(z), (13)
indicating that the random process Z(t) is totally uncorrelated at any two distinct time instants.
Such a random process is referred to as a white noise process which agrees very well with the mean
crack growth behaviour. As a result, the Gaussian white noise model gives minimum scatter for
crack growth propagation prediction. For the other extreme case in which ~ ~ 0, the lognormal
random process X(t), or the normal random process Z(t), has a correlation of unity between any
two successive time instants since, according to eq. (12), R:=(z)= 02 is independent of time
difference z. Under this condition X(t) becomes the independent lognormal random variable
investigated by Yang et al. [10, 27]. It follows that the distribution functions of the crack size at
any given number of flight hours require only the crack growth rate parameters as well as the
variance 8. It also yields a maximum scatter prediction. Upon acceptance of Yang's formulation,
the stochastic behaviour of propagating cracks has been shown to lie between the two extreme cases
described above. In short, the stochastic nature of a material may be described by the correlation
parameter ~ under Yang's assumptions.
The lognormal model, investigated by Yang et al. [10, 27], has been demonstrated to be very
flexible and it fits adequately the available experimental data. Among the advantages of this model
for practical analysis and design applications are its mathematical simplicity, and the fact that the
correlation parameter and variance are the only two required material characteristics, implying that
a small number of replicate specimens is adequate.
To illustrate these points and referring again to both the Paris-Erdogan model and the
lognormal model of Yang and co-workers, a special case of the general stochastic model may be
taken as: da/dt = X(t)c[a(t)]'. The parameters c and m may be interpreted as a deterministic
description of the inherent material property, while the variance 82 and autocorrelation parameter
describe the statistical nature of the normal process. As examples, the results presented in Table 1
have been estimated from a linear regression analysis of the test results (i.e. the crack growth rate
vs the crack size) presented by Yang et al. [10]. The WPB and XWPB for A17475-T7351 data sets
and the CWPF for AI7075-T7651 in 3.5% NaCi solution data set are presented.
3.2. The McGill-Markov process
To further study the stochastic characteristics of fatigue crack propagation processes, it is
necessary to consider the concept of Markov processes. In this subsection, discussion will centre
on only one Markov process, namely, the McGill-Markov process.
A Markov process is based upon the assumption that the future of the process is influenced
only by the current state and not by the history that led to its present state. This statement explains

Table 1. Linear regression estimate of m, c, 0= and autocorrelation parameter

Data set rn c (10 -3) 0.- ~ (10 -4) Ao (in.) Ar (in.)
WPB 0.941 0.116 0.070 1.5 0.004 0.04
XWPB 1.014 0.284 0.109 1.0 0.004 0.04
CWPF 1.372 2.128 0.202 3.5 0.010 0.35
 Stochastic systemsapproach to fatiguereliability 693

why Markov models have the potential of being useful in engineering applications• If a thorough
inspection of a structure is carried out, the current state of damage is known• Using a Markov
interpretation, the future damage may be predicted without any knowledge of how the structure
arrived at the current damage state• The reader is referred to refs [28-30] for further details•
According to assumptions made in ref. [8] and the nature of fatigue crack growth, the crack
propagation process is here assumed to be a discrete-state, continuous-parameter and
nonhomogeneous Markov process• By doing so, the transition probability density becomes a
variable which depends only on the time difference z. Suppose that for any time, t,, the initial
distribution, f{x(t,)}, is known• This distribution, in addition to a knowledge of the transition
density function, gives a procedure by which any future time of interest may be obtained by simply
repeating the process as many times as necessary. Although it may appear that the crack formation
process should be thought of as a continuous process it is reasonable to treat crack size as a discrete
quantity due to the built-in limitations of crack detection and measuring systems. As Bogdanoff
and Kozin [9] point out, computational advantages are found from treating crack sizes as discrete.
The crack size variable, a(t), can only be measured to within equipment and operator
limitations• By considering the observable zones, i, the crack size may be written as:
ai < a(t) < ai + Aa~ ; l<_i < NR <_ NF, (14)
where i is the state number, Aa~ is the width of a state, NR indicates a state where a repair is to
be carried out and N F designates the state at which failure will occur. A discrete-state and
continuous-parameter stochastic process, {a(t), t ~ T}, has the Markov property if for n = 1,2,3,
• . . and for t, ~ t (n --0, 1, 2 . . . . . n), where tl < < t,, and for arbitrary integers i and.j the
• • •

equality:
P{a(t,+ ,) =jla(t,,) = i,a(t,_ i) = in_ ,,'",a(to) = i0}
= P{a(t,+,)=j[a(t,,)= i}, (15)
is satisfied for any integers i,_ 1. . . . . to. The probability on the right-hand-side of this equation
may be taken in the general form:
P{a(t) =j[a(z) = i} =po(r,t), 0 <_ z < t, (16)
where i and j are integer states, and z and t are times. This probability is called the transition
probability and is defined as the probability of a transition from state i to state j during the time
interval z to t.
In order to solve the (Chapman-)Kolmogorov differential equations which govern pi~, an
infinitesimal transition scheme must be specified [31]. In 1982, Provan and his co-workers, through
examination of a variety of processes, developed a new form of intensity function which adequately
describes the time evolution of material property degradation processes [32]. The intensity functions
for this "McGill-Markov" process are:

qj(t)=j2(t),qkj(t)=~{(j-1)2(t) fork=j-i, for t > 0 ; j = 1,2,..., (17)

(0 otherwise,
with the two degree of freedom 2(0 for the McGill-Markov process being given by:

2( 0 _ 2(1 + 20
1 + 2r " (18)

2 and x are positive empirical system parameters which are determined by a fit to experimental
data. When these parameters are found, the material degradation response of components
operating in a specified fatigue situation may be modelled. If the fatigue situation is changed, new
system parameters must be determined.
The solution of the governing (Chapman-)Kolmogorov forward differential equation is well
known [31]. For this linear, non-homogeneous birth process it is given by:

p~(z,t) = (~.-- li) q~(1--q) j- ', (19)

694 K. FARHANGDOOSTand J. W. PROVAN

0.4

Ir ''=I.IOJRS 1
O.g
0.11

=m==:=
i O.R
~
!
0,7

0,1
0,6

0.§ t I I
o 6 e 10 12 14
lO 40
CRACK SIZE (Stm)
70 TIME, (xlO s FUGHT HOURS)
a) Reliability at a given time b) Rellabil~ ~ • function of tin'~

Fig. 3. Determination of reliabilityat a given time.

where for the McGiil-Markov process:

q = e-a"'", A(T,t) =
f, 2(t') dt'.

Finally, in order to determine the entire history of the crack propagation distribution, the
(20)

total probability may be continuously monitored via the fundamental absolute probability
relation:
/

Pj(t) = ~ po.(~,t)Pi(z), (21)

i=l

where P~(t) is the probability of being in the j-th (later) state at time t; Pi(T) is the probability of
being in the i-th (earlier) state at time ~; and p~(~,t) are the transition probabilities of a Markov
process. Hence, for the McGill-Markov process, it is necessary to specify the initial state and
transition probabilities in terms of ). and ~: in order to describe the evolution of the entire process.
3.3. Reliability analysis
One of the main objectives of this paper is to introduce reliability maintenance and
inspection/correction procedures via the McGill-Markov interpretation of material property
degradation. Inspection/correction failure control systems play a significant role in a reliable repair
policy which adheres to well-scheduled inspection programs. Reliability is defined as the probability
of satisfactory performance of a component for a specified period in a specified environment.
Hence, the reliability can be found if the critical crack size is known, it is the probability that the
crack does not exceed this critical size. This quantity can be obtained by summing up the
probability of a crack being any sub-critical size. This is shown in Fig. 3(a) where the reliability
is equal to the sum of the areas to the left of the critical size. The reliability as a function of time
is shown in Fig. 3(b).
3.3.1. Reliability maintenance-inspection/correction process. A maintenance engineer ensuring
the safe operation of a large number or fleet of similar structures will often specify a desired level
of reliability for each component in the structure. Once this level has been determined, perhaps
by company policy, standard industry practices or other means, it becomes necessary to determine
when to schedule the maintenance procedures that correspond to this desired level of component
reliability. This may be accomplished by employing the McGill-Markov model to predict when
the probability of component failure will reach the desired limit and then calling for an
inspection/correction procedure. In this paper, inspection/correction processes are summarized a s :
stopping the degradation process for the entire fleet, identifying specific components that pose a
risk to structural integrity and carrying out the necessary maintenance procedures.
Analytically, as a result of the removal and replacement of these specific components during
 Stochastic systems approach to fatigue reliability 695

a single fleet inspection/correction procedure there results two distinct populations referred to as
population I, which consists of the remaining components from the initial group, and population
II, which is the group of replacement components. The reason for this distinction is that the
McGill-Markov process is time-dependent or non-homogeneous. This means that for an
inspection/correction at time T~,~p~, the fatigue process continues for population I, while for
population II it starts at time T = 0 and ends at time T = Tt~,,~- T,.sr~c,. As long as the fatigue
loading spectrum remains the same, the system parameters 2 and ~: for the specific component in
question, can be used for populations I and II. Following the development of Provan and
Rodriguez [32], the total probability of failure is a combination of the probabilities of failure of
populations I and II, and is given by:
PcroTaL(t) = (1 -- R))K,, + (1 -- R)')(I - g,,), (22)
where R~ and R~j are the predicted reliabilities of populations I and II at time t > t~, and K,~ is the
normalizing constant of the original population or reliability at the inspection time t~.
This process can easily be extended to include as many inspection/correction procedures as
desired.
3.2.2. Reliability maintenance-inspection optimization. Another useful form of component
reliability analysis which can be carried out with the McGill-Markov model is the optimization
of fleet inspection times. As an example, suppose it is desired to minimize the total probability of
component failure under the condition that there can be only one inspection/correction procedure
carried out during the life expectancy of the components. Hence, the question, "what is the
optimum time for this procedure?" may be posed. An inspection too early in the service life will,
on the one hand, remove few components that may subsequently fail, while a later inspection may
be too late to remove components that have failed. The optimum time for inspection will depend
on several variables such as: critical crack size, repair size, inspection process and the quality of
replacement components [8]. The quantity to be minimized, the total probability of failure at some
time t~,~t, is given by [32]:
errovAL(t,,a.) = (I -- R] ~,a.)K,+ (1 - RI'~°~,)(I - K,), (23)
where t is the time of the inspection, R, is the reliability of the initial population at the time of
inspection, K, is the normalizing constant or reliability at inspection time and Rtlinal
l and gtfinal
tJ are
the reliabilities of populations I and II at the final time.
Hence, as further illustrated in Section 8, the McGill-Markov model used in conjunction with
a failure control methodology can be a useful tool for obtaining valuable reliability information.

4. E X P E R I M E N T A L P R O C E D U R E AND BASIC RESULTS

In this section, a computer controlled, increasing stress intensity factor AK test method,
according to the American Society for Testing and Materials (ASTM) E647 [33], is applied to 18
standard C(T) specimens manufactured from a forged Ti-6AI-4V jet engine, fan disk grade,
titanium alloy. The main objectives of carrying out these experiments are to examine the stochastic
properties of crack growth for Ti-6Al-4V and to analyze the results in an effort to establish the
parameters associated with the closure-lognormal and McGill-Markov stochastic processes, and
the reliability assessment of components manufactured from this titanium alloy.
4.1. Experimental procedure
The MTS 810.14 Material Testing System was periodically calibrated in order to ensure
accurate data acquisition. Fatigue crack propagation tests were carried out under the control
of the in-house F A T I G computer program based on the crack closure compliance method
detailed in Section 2. For the compliance calculations, the C M O D measurement was determined
using a mechanical clip gauge and employed as an input to FATIG. The gauge had a full scale
displacement of 0.50 mm for a 10 volt excitation voltage monitored using a digital display unit with
the resolution of 0.005 volts. Hence, the maximum resolution of the gauge was considered to be
0.00025 mm. Employing the same procedure for the load measurements, a resolution of 0.005 kN
was obtained.
696 K. F A R H A N G D O O S T and J. W. P R O V A N

An important basis of the ASTM E647 test is that it distinguishes between crack initiation
and propagation. The test involved sinusoidally and cyclically loading the notched C(T) specimens
(W = 5.08 cm; B = 1.27 cm) with the load ratio being set at R = 0.1. Figure 4 shows the procedure
of K vs crack length determination.
The FATIG program operated according to the following six steps.

(1) The first step was carried out at a constant load.

(2) In the second, the load was increased to the maximum load Pm~x = 7 kN, at point A, that
is related to the AK value and is obtained using the FATIG program by extrapolating along a
constant load line to point B corresponding to a crack length equal to the notch length (10 mm).
Following this, the crack initiation is carried out at constant AK until the crack reaches a length
equal to the notch length.
(3) In the third step, the load is decreased exponentially until the initial crack length equals
12 mm (at point C in the figure).
(4) After detection of the initial crack length, the test is continued at a constant load (equal
to 5 kN) which represents step four of Fig. 4.
(5) During step five, the crack growth rate is high and therefore the printout rate is increased
by omitting the waiting loops 1 and 2 inside a main loop.
(6) Near the end of the test, step six is started because crack propagation is faster and a line
of data is printed when the crack length becomes more than 20 mm.

The FATIG program utilizes three main loops, corresponding to the steps mentioned. The
"calculation" part of FATIG determines:

(i) the crack length based on load vs CMOD data (compliance);

(ii) the AK calculation based upon the ASTM E647 standard formulation for C(T)
specimens [33]; and
(iii) the AKo~ calculation based on AK and closure load.

During each "data acquisition" block, the load vs CMOD curve, Fig. 2, for 200 individual data
points is obtained. The lower limit Pm~,, shown in Fig. 2, is a variable such that the nonlinearity
is distinguished by the FATIG program as the closure load. A linear fit is made through the
remaining points to obtain the normalized compliance.

80

25

12 3 4 5

i,g

E A B

°i
i i
4
i i I
IO 8O
Crack length (ram)
Fig. 4. Procedure of K vs crack length for Ti-6AI-4V utilizing the F A T I G program.
 Stochastic systems approach to fatigue reliability 697

0.8

0.8

Ig
0.4

0.2

0
0 0.05 0.1 0.15 0,2 0.25 0,3
tJ
X

Fig. 5. Comparison between eqs (25) and (26), using eq. (24).

The relationship between compliance and crack length has been analytically derived for a
number of standard specimens [33]. Such relationships are usually expressed in terms of the
dimensionless compliance, E v B / P , as follows:

ct - W - Co + c,u~ + c2u~ + c3u 3 + c4ux + csu~ ; u~= + 1 , (24)

where a / W is the normalized crack length, E = 117 000 MPa [34] is the elastic modulus, v is the
displacement between measurement points, B is specimen thickness, P is load, W is the specimen
width, and Co, cj, c2, c3, c4, c5 are specimen dependent coefficients. Figure 5 shows a comparison
between:
EBv 3(~ + 1)(~ + 2.3)
p = ~ 2 _ 2~ + 1 + 8(1 + v)~, (25)

proposed by Mirzaei [17] and:

-ff -- 1+ ~ {1.61369 + 12.6778~-- 14.2311cc 2

-- 16.6102~ 3 + 35.0499~ 4 -- 14.4943~5}, (26)

proposed by Saxena and H u d a k [35]. There is good agreement between these two curves above a~ W
= 0.2. However, comparison with experimental crack propagation data has shown eq. (25) to be
in good agreement below the 0.2 value while eq. (26) is not. The coefficients used in F A T I G are
presented in the Table 2.
The stress intensity factor range, AK, for C(T) specimens is calculated using a formula

Table 2. Comparison between coefficients of eqs (25) and (26), using eq. (24)
Eq. no. co cn c2 c3 c4 c5
(25) 1.0003 --4.49 5.0629 --47.1328 231.0345 -309.889
(26) 1.001 --4.6695 18.460 -236.82 1214.9 -2143.6
698 K. FARHANGDOOST and J. W. PROVAN

24

I C(T) 8peclmen8 of
Ti-SAI-4V I
j 0.0~6 ..,

-~ I II III """
/

20 E =, 2.9
Z ~'
"0
•s

~" .o
n-,=-= "'~-"~ 69
10
O~.~ . r.: '

14 1E,.~

t2 =' I ~ I I I
0 800 Io000 1,500 2,000 1o 11 =o s
Number of Cycles, xlO a Effective Stress Intensity Range Factor (MPaVm)
a) Crack propagation b) Crack growth rate
Fig. 6. Crack propagation and fatigue crack growth rate for C(T) specimens of Ti-6AI-4V.

developed by Srawley [36]. This formulation has been adopted by ASTM E674 and is given in the
following expression:

AK- AP (2 + ~ ) (0.886 + 4.64~ - 13.32~ 2 + 14.72~ a - 5.6~4), (27)

B x / ~ (1 - (~)3/2
where AP is the load range. The expression for the effective stress intensity range factor, A / ~ , is
the same as eq. (27), the only difference being that Ap is defined by:
A P = Pmax - Pop, (28)
where Pop is the closure load.
When a specific test was completed, the resulting data were stored in another data file and
were analyzed with the second in-house FADA program to obtain a da/dN vs AKo~curve. Several
methods of data processing exist in the literature. For example, the methods proposed by ASTM
E647 are the secant and incremental polynomial methods. It has been shown by Virkler et al. [37],
and Wei and Shim [38], however, that the incremental polynomial method is more precise. Hence,
the FADA program utilizes this scheme for n = 3, that is, 7 successive data points. The FADA
program also provides the data output for plotting crack length, a, vs the number of cycles, N,
and the crack growth rate, da/dN, vs AKe~.
4.2. Experimental test results for Ti-6A1-4V
Test results were obtained from 18 standard C(T) Ti-6AI-4V specimens prepared according
to conventional procedures. The fatigue crack propagation for all 18 specimens is shown in
Fig. 6(a). The F A T I G program produced approximately 150 data points for each specimen. The
test was run at room temperature in a random order while the MTS unit operated under load
control. A sinusoidal signal with a frequency of 17 Hz was used as the input for the alternating
portion of the load. The crack tip region was continually monitored by a microscope and a cross
hair mounted in the microscope was used as a reference line during crack propagation. Visual crack
lengths were measured simultaneously with the printout of the crack length data from the computer
and a comparison was immediately made. In all cases, an excellent agreement was observed during
the running of all tests.
 Stochastic systemsapproach to fatiguereliability 699

From Fig. 6(a) and AKin, the fatigue crack growth rates vs the effective stress intensity range
factor were determined for Ti-6AI-4V and are presented in Fig. 6(b). Examination of the data
revealed a few surprises, the most unexpected being the change in the growth rate which occurred
in almost every test. A close observation of the crack growth rates shows two transition points,
namely, at AKo~ - 8-9 M P a x / ~ and at AK~,r ~- 13-14 MPax//-m. Hence, the fatigue crack growth
process is highly complex, especially in the Ergodan regime. In spite of being smooth da/dN vs
AKon curves, it is not a uniform linear curve in the logarithmic scale Fig. 6(b). Accordingly, it is
reasonable to have different Paris-Erdogan regimes correlated with these experimental results.
However, from the viewpoint of reliability analyses the first part of the fatigue crack growth rate
curve, i.e. prior to the second transition point, is more important than the second.
The raw data of crack length vs number of cycles for each of the 18 experiments are the
information required by the McGill-Markov interpretation of stochastic fatigue crack growth
(Section 7). The reduction of this information to obtain the crack propagation rates (da/dN) which,
along with the determination of the effective stress intensity range factor for the 18 experiments,
are the data required for an implementation of the closure-lognormal process (Section 5).

5. THE C L O S U R E - L O G N O R M A L CHARACTERISTICS OF Ti-6AI-4V

According to the general form of fatigue crack propagation, eq. (6), the crack growth rate is
a function of the stress intensity factor, maximum stress intensity factor, stress amplitude, load ratio
and so on. Some commonly used crack growth rate functions, such as the Paris-Erdogan model
of eq. (2), are such that the coefficients c and m are functions of load condition and environment.
A comprehensive assessment of the closure phenomena necessitates the evaluation of the state
of residual stress and strain in the neighbourhood of the crack tip and the extent of crack closure.
Crack closure effects are most pronounced at low AK levels. This can be shown in Fig. 7 which
resulted from the 18 Ti-6A1-4V specimens tested as described in the previous section. Hence, Pop
for Ti-6AI-4V as a function of crack length (a) is found by a polynomial curve fit to experimental
results of closure load vs crack length depicted in Fig. 7. The result of this curve fitting for the
titanium alloy in question is:
Pop = 2.90203 - 0.140572a + 0.0021723a2; (kN;mm). (29)
As detailed in Section 2, the crack closure concept may be used to describe the influence of
load spectra, load ratio and/or environmental parameters. Hence, using a quantitative knowledge

M~*• •a•
•4 ~..*.~ .':'.:...
,'..~'..,~
r~-
.-~?-..
* o f , ,. • ~v •.

,;~,: .... ~ ,..:. ,,.....

• ' * *:,~.¶*" ..'~, . ~ • ,~'.." t~ .'. * . .
• "' ".~" "" " "<"~'*'""" • " "'" t'"

i 1

e , I a I a I n I i L n I n
18 II 14 lO 11 II It 14

Crack length (mm)

Fig. 7. Closure load vs crack length for Ti-6AI-4V.
700 K. F A R H A N G D O O S T and J. W. PROVAN

Table 3. Closure-lognormal characteristics of Ti-6AI-4V

Closure-lognormal
parameters 62 10 4 ¢10 4 m cl0 -9
Ti-6AI-4V 47 1.4 3.25 5.4

of the crack closure stress intensity level in AKe,, the following equation for the description of
fatigue crack propagation is suggested:

da(t) = X(t)c(AKe~)", (30)

dt
where a(t) is now the random crack size and X(t) is a non-negative random process. In this case
AKc, is a function of crack length and closure effects and X(t) is a lognormal random process. By
taking the logarithm of both sides of eq. (30), it follows that:

(31)

By substitution

(da)
7J(t) = log -d~ ' Z ( t ) = logX(t), C = logc and F = log(AKo~),

eq. (31) may be rearranged into:

~(t) = mF + C + Z(t). (32)
Z(t), fully specified as detailed in Section 3 by its variance 02 and autocorrelation parameter, ¢,
describes the scatter inherent in crack growth rates due to the microstructural inhomogeneities of
a specific material. These parameters, along with the deterministic c and m, are obtained from the
Ti-6A1-4V test results of the crack growth rate vs AKo~. Hence, the four closure-lognormal
parameters, namely, m, c and ~, 02 are material specific parameters. Table 3 lists these
closure-lognormal parameters for Ti-6A1-4V.
In this way, the statistical scatter of actual test results, or as fully discussed in the next section,
simulated fatigue crack growth data, may be presented. By using these parameters and the
simulation procedure, definitive reliability, availability and maintainability procedures (see
Section 8) may now be carried out using the appropriate McGill-Markov parameters, 2 and x (see
Section 7). Hence, under these assumptions, a determination of a material's stochastic crack growth
characteristics based upon AKe~ and the closure-lognormal interpretation of scatter leads to an
estimation of crack growth rates.

6. SIMULATION OF CRACK PROPAGATION--Ti-6AI-4V

A simulation procedure is required in order to utilize both the closure-lognormal and
McGill-Markov procedures of describing the statistics associated with fatigue crack growth
processes, and to transfer the standardized closure-lognormal information into an estimation of
crack propagation as it occurs in actual components and under specified loading situations.
The stationary Gaussian random process Z(t) = logX(t), as described in Section 3, may be
simulated using the well-known Fast Fourier Transform (FFT) technique. With a specification of
and 02, Z(t) may be simulated using the FFT method with the following expression [39, 40]:

ZfJ'AT) = 2x/2x/2x/2x/2x/2x/2~Re ~==(kAw)e~*qe

~;~2~/~ , (33)
k 1

where ¢~_-..(w)is the power spectral density. Equation (33), for w _> 0, is evaluated at equally spaced
intervals Aw with kAw > 0. Re{ } represents the real part of the complex quantity in the bracket
 Stochastic systems approach to fatigue reliability 701

and q~, (k = 1, 2 , . . . , M) are statistically independent and identically distributed random variables
with the uniform distribution in [0,2hi. The stationary normal random process, Z(t), is computed
at equally spaced discrete time points tj =jAt for j = 1, 2. . . . . M. Furthermore, in eq. (33) the
power spectral density is given by:

~=(w) = ~ R=(z)e -'"~ dr, (34)

in which R_-:(r) is the autocorrelation function described by eq. (12). By substitution of R--(z) from
eq. (12) into eq. (34), it follows that:

• ::(w) = ~2 2¢
+ w-------762, (35)

where ¢ and 62 are the closure-lognormal parameters.

By taking the integral of both sides of eq. (30), the crack propagation equation becomes:

f (A-~)" - c~X(t) dt, (36)

where AKe~ is determined by a straightforward combination of eqs (27)-(29).

Equation (36) may be incrementally written as:

N Aa M
[AK~(tAa)] - c / =~1 X(/At)At,
• m (37)
i=

where N is the number associated with the crack size state and M is the number of time increments
(cycles). The results of these analyses are shown in Fig. 8.
AKin, incorporating a quantitative knowledge of the crack opening stress level, now appears
as an appropriate field parameter for incorporating a knowledge of constant-amplitude crack
growth rates into an estimation of practical crack growth rates. Specifically, knowing in terms of
m, c, # and ~ the crack growth rate as a function of AKo~for a specific material, such as Ti-6A1-4V,
the influence on the fatigue reliability of a specific component of a specific fatigue loading may
be estimated during the design process. In this way, the simulation process plays an important role

II4

[ /t/H
l D:-~o I

]-
14

1%
800 1,000 l,S00 2,000 12 0 800 1 000 1 I;00 2.000
-. Num.ber of .Cy~u, x l 0 ~ Nurnberof CyoiN, X l 0 s
a) ~lmulatea crack propagation b) Real crack propagation
Fig. 8. Simulated and real crack propagation of C(T) specimens of Ti-6AI-4V.
702 K. F A R H A N G D O O S T and J. W. P R O V A N

Table 4. McGilI-Markov parameters for

C(T) specimens of Ti-6AI-4V--experimental
and simulated
System parameters 2 10 -5 h
Ti-6AI-4V 0.14 0.98

in removing the dependency on actual experimental results obtained under specific loading and
environmental conditions. Hence, the closure behaviour, either experimentally determined or
predicted by a model [17], expressed in terms of the effective stress intensity factor range (AKe~)
expression found in handbooks and the closure--lognormal parameters constitute the only
information required for analyzing a particular situation.

7. THE M C G I L L - M A R K O V P A R A M E T E R S FOR Ti-6AI-4V

This section presents the McGill-Markov stochastic fatigue crack growth properties of
Ti-6AI-4V. As described in Section 3, the two constant parameters 2 and K in the intensity functions
q~(t) and q~j(t), defined in eqs (l 7) and (18), are related to the stochastic properties of the material
in question. Accordingly, the probability of the crack tip being in state i is increased, state by state,
as the process progresses. Failure of a specified component in a large number of similar components
in a finite time interval implies that there is a probability of the state going past a prescribed limit.
Hence, 2 and K are the two system parameters which control this component's failure and play
a significant role in both reliability assessment and inspection/correction procedures. Applying the
McGill-Markov model to the Ti-6AI-4V data is an appropriate way to illustrate the capabilities
of this approach.
These parameters are determined by a fit to the experimental or simulated data. Several
steps, however, must be taken before this iterative process is undertaken. The first step is
the normalization of data to an initial crack length of a0 at time t = 0. This is done to eliminate
the crack initiation stage. In the next step, the data is discretized into states of width Aa. Using
the 2 and x, the probability histograms at future times are generated, and the mean and variance
calculated from /~i(t) = E/pj(t) and ~ ( t ) = E/(/"- I~j(t))2pj(t). 2 and x, as system parameters, are
thus determined by an iterative process of fitting the model predictions to the experimental results
from the 18 identical C(T) specimens of Ti-6AI-4V. For the Ti-6AI-4V data set, a state size of
0.4 mm was chosen and a total of 70 states was used since this was judged to be a sufficient
number for the interval of time t = 0 cycles through t = 2E + 06 cycles. Hence, the input data were
as follows:

• initial crack size distribution at 7.20E + 05 cycles;

• number of states = 70;
• initial time = 7.20E + 05;
• increment time = 0.20E + 05;
• final time = 20.00E + 05;
• failure state = 40.

The values of 2 and x which give a good fit to the experimental or simulated data are presented
in Table 4.
A comparison between the experimental, simulated and McGill-Markov model predictions,
shown in Fig. 9, shows that the concept of describing fatigue crack growth in terms of stochastic
processes, is a flexible method of predicting the crack propagation statistics for industrially
significant materials and situations. The fact that the 2 and x parameters are applicable in reliability
prediction and inspection/correction procedure development, as detailed in the following section,
enhances the benefits of this approach. Hence, in order to calculate the crack propagation curves
in any situation, it is necessary that the initial probability, and mean and variance of the process
be known. This, as has been demonstrated, may be accomplished by using either experimental or
simulated crack propagation results.
 Stochastic systems approach to fatigue reliability 703

8. RELIABILITY ANALYSIS OF Ti-6AI-4V

The combination of the failure control system with the McGill-Markov model presented in
Sections 3 and 7 can be a very powerful tool for practical engineering reliability calculations. While
two specific uses are highlighted in this section, others may be formulated without difficulty.
Reliability, as was discussed earlier, is defined as the probability a component will perform
satisfactorily for a specified period of time. For determining reliability, the first step is to generate
the crack size probability distributions, illustrated in the form of histograms, for given future times.
The reliability can then be found if the critical crack size is known; it is the probability that the
crack does not exceed this critical length. An implementation of this procedure is illustrated in
Fig. 10.
With reference to Section 3.3.1, consider the determination of the appropriate inspection/
correction intervals as a first illustration of the power of this McGill-Markov approach to
improving reliability. By defining the replacement size (state), N R = 25 and the desired reliability
to be 0.9999 or PrrorAL(t) = 1.0E -- 04, the total probability of failure for times 0.8E + 06 to
2.0E + 06 cycles were obtained as illustrated in Fig. 1 l(a). As can be seen, the inspection times
for the appropriate inspection/correction procedures were determined as: 1 . 5 0 0 E + 0 6 ,
1.700E + 06 and 1.880E + 06 cycles.
Alternatively, the results of a change in acceptable reliability level from 0.9999 to 0.9995 for
a constant repair size, N R = 25 are presented in Fig. 1 l(b). From a comparison with Fig. l l(a)
it is apparent that not only will the first maintenance procedure be carried out at a later time, but
that one fewer procedure will be necessary.
The effect of varying the repair size while maintaining the desired level of reliability at 0,9995
is illustrated in Fig. 12. This figure shows how the inspection interval is affected by a change in
repair size (state) from N R = 20, through N R = 25, to N R = 30 prior to 2.0E + 06 cycles. There
is a total of two inspection/corrections for N R = 20, three for N R = 25 and four for N R = 30.
These figures illustrate the type of information that can be obtained from this reliability model
concerning inspection intervals.
Following developments that are similar to those described in Section 3.3.2 and as a second

81mt~ttlon
TNt R u u ~
MoGIIl~4~kov

0
~e

J, o
A
A

0 5 10 15 ~0 . . . . . . . . . *'I=

5 10 15 20
Number of Cycles, xlO ~
Number of Cycles, x10 ~
Fig. 9. Ti-6AI-4V mean and variance crack size, from McGilI-Markov, simulated and experimental crack
propagation.
704 K. FARHANGDOOSTand J W. PROVAN

0.00

0.08

.J
~ 0.07
--1
LU
n"
O.gO

0.05

0.94 I I I I I I
000 800 1,000 1,200 1,400 1 ,O00 1,800 2,000
NUMBER of CYCLES (x 10 3)
Fig. 10. Reliability vs time for Ti-6AI-4V, with NF =40, (A~r =28 mm).

illustration of the power of this McGill-Markov based approach consider a specific inspection
optimization analysis. Suppose a design engineer is asked to optimize the time for a single
inspection/correction operation during the life expectancy of an ensemble of components
being subjected to similar fatigue loading situations. By specifying N F = 40 and N R = 20 for the
loading being used throughout these illustrations, the solid line in Fig. 13 is obtained. From this
figure it is apparent that the optimum time for the single permissible inspection/correction is at
1.700E + 06 cycles and that the total probability of failure is decreased by 88% over the no
inspection/correction case. By changing the repair policy, such that N R is varied from 20 to 30,
different curves also shown in Fig. 13 are obtained. Hence, while the optimum inspection time is
increased by increasing the repair crack size from N R = 20 to 30, the overall reliability of the system
is reduced.

0.00012
REPAIR 81ZE (State): 25] I REPAIR81ZE(State):25I
0.0001 0.0005
LIJ
n-
8E-05 0,0004

~ 0.0003

i-- 4E-05

2E-Oe
i 0.00(~
E
0,0001
/
o , J j 0 , , J
1,ooo 1,=oo 1,4oo 1 , ~2 t,8oo ,ooo 8oo 1,ooo 1.=oo 1,4oo 1 , q ~,~o zooo
NUMBER of CYCLES (xlO) NUMBERof CYCLES (xlO)
a) Inspection at 0.9999 reliability level b) Inspection 8t 0.9995 rellabiUty l e v e l
Fig. 11. Inspection schedule for Ti-6AI-4Vat: (a) 0.9999 and (b) 0.9995 reliability levels.
 Stochastic systems approach to fatigue reliability 705

0.0086
REPNR SIZE (State): 20
REPAIR SIZE (State): 25 • ~ , ~ 9
~t~ 0'0005
3 ..... I ~;J ;:~ : ::!
i~.~ 0.0004
: ~,! : :~
1 0'~3

~ 0.1~2
j ~ ~::!i : ili
0,0001
: ii i ;i

o
800 1,ooo 1,2oo 1,4oo 1,6oo 1,800 2,000
NUMBER of CYCLES (xlO')
Fig. 12. Influence of repair policy on inspection time for Ti-6AI-4V.

9. CONCLUSIONS
The conversion of AK into AKo~through the inclusion of closure effects plays a significant role
in reliability analysis. By using AKin, it now appears possible to transfer the stochastic properties
of crack growth rates, measured under ideal laboratory conditions, to practical situations.
Incorporating the sense of closure into both the fatigue crack growth rate description and the
lognormal interpretation of the scatter has led to the proposed closure-lognormal model, in terms
of c, m, ~ and 0 2, which describes the statistical nature of crack growth rates.
On the other hand, the 2 and x of the McGill-Markov process, determined either directly from
experimental data or from a nontrivial simulation of the information contained in the
closure-lognormal interpretation of the basic fatigue crack growth characteristics of the material,
vary with respect to the actual loading situation. With the crack propagation characteristics being
predictable, the reliability and inspection processes may then be evaluated using these
McGill-Markov parameters.
Finally, the closure-lognormal, McGill-Markov and RAM models proposed during the
course of this study have been examined in terms of Ti-6A1-4V titanium alloy.

0.08
REPAIR SIZE (State): 20
IlJ
tr REPAIR SIZE .(.State): 25
3
~ 0,06
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . -

IJ.
O

~__.0.04

ft.
i.~ 0.02
o
0 I I I I I I
600 800 1,000 1,200 1,400 1,600 1,800 2,000
NUMBER of CYCLES (xlO = )
Fig. 13. Influence of repair policy on optimum inspection time for Ti-6AI-4V.
706 K. FARHANGDOOST and J. W. PROVAN

REFERENCES
[I] A. M. Freudenthal, Safety of structures. Trans. ASCE 112, 125-180 (1947).
[2] W. Weibull, A statistical distribution of wide applicability. J. appl. Mech. 18, 293-297 (1951).
[3] Z. W. Birnbaum and S. C. Saunders, A new family of life distributions and estimation for a family of life distributions
with applications to fatigue. J. appl. Probability 6, 319-347 (1969).
[4] A. M. Freudenthal and M. Shinozuka, Structural safety under conditions of ultimate load-failure and fatigue. WADD
Technical Report, 61-77 (1961).
[5] A. O. Payne, A reliability approach to fatigue for structures. Probabilistic Aspects of Fatigue, A S T M - S T P 511,
106-155. American Society for Testing and Materials, Philadelphia, U.S.A. (1972).
[6] J. W. Provan, Probabilistic approaches to the material-related reliability of fracture-sensitive structures, in Probabilistic
Fracture Mech. Reliability (Edited by J. W. Provan), pp. 1-45. Martinus Nijhoff (1987).
[7] J. W. Provan and Y. Theriault, An experimental investigation of fatigue reliability laws, in Defects, Fracture and
Fatigue (Edited by G. C. Sih and J. W. Provan). Martinus Nijhoff (1983).
[8] J. W. Provan and S. R. Bohn, Stochastic fatigue crack growth and the reliability of deteriorating structures. Fatigue
90, IV, 2259-2265, Hawaii (1990).
[9] J. L. Bogdanoff and F. Kozin, Probabilistic Models of Cumulative Damage. John Wiley (1985).
[10] J. N. Yang, W. H. Hsi and S. D. Manning, Stochastic crack growth models for application to aircraft structures, in
Probabilistic Fracture Mechanics and Reliability (Edited by J. W. Provan), pp. 171-212. Martinus Nijhoff (1987).
[11] W. Elber, Fatigue crack closure under cyclic tension. Engng Fracture Mech. 2, 37-45 (1970).
[12] K. Farhangdoost and J. W. Provan, On fatigue reliability assessment. Fatigue 93, Vol. II, pp. 707-712, Montreal (1993).
[13] P. C. Paris and F. Erdogan, A critical analysis of crack propagation laws. J. bas. Engng, Trans. ASME, Series D,
85, 528-534 (1963).
[14] R. O. Ritchie, W. Yu, D. K. Holm and A. F. Blom, Mechanics of Fatigue Crack Closure (Edited by J. C. Newman,
Jr and W. Elber), ASTM-STP 982, pp. 300~318. American Society for Testing and Materials, Philadelphia, U.S.A.
(1988).
[15] G. T. Gray Ii| and G. Luetjering, Titanium Sei. Technol. 4, 2251-2258 (1984).
[16] J. E. Allison and J. C. Williams, Titanium Sei. Teehnol. 4, 2243 (1984).
[17] M. Mirzaei, A fatigue crack closure model and measurement technique. Ph.D. Thesis, McGill University, Montreal
(1991).
[18] N. A. Fleck and J. C. Newman, Jr, Mechanics of Fatigue Crack Closure, ASTM-STP 982, pp. 319-341. American
Society for Testing and Materials, Philadelphia, U.S.A. (1988).
[19] S. Banerjee, Review of crack closure. AFWAL-TR-84-4031, Wright-Patterson AFB, OH (1984).
[20] J. E. Allison, Fracture Mechanics: 18th Symp. A S T M - S T P 945, American Society for Testing and Materials,
Philadelphia, U.S.A. (1988).
[21] M. S. Milter and J. P. Gallagher, An analysis of several fatigue crack growth rate (FCGR) descriptions. Fatigue Crack
Growth Measurement and Data Analysis, A S T M - S T P 738, 205-251. American Society for Testing and Materials,
Philadelphia, U.S.A. (1981).
[22] R. G. Forman, V. E. Kearney and R. N. Engle, Numerical analysis of crack propagation in cyclic loaded structures,
J. has. Engng, Trans. ASME, Series D, 89, 459-465 (1967).
[23] J. M. Larsen, B. J. Schwartz and C. G. Annis, Jr, Cumulative fracture in the transition region. 12th Fracture Mechanics
Con['., A S T M STP 700, 99-111 (1980).
[24] Y. K. Lin and J. N. Yang, On statistical moments of fatigue crack propagation. Engng Fracture Mech. 18, 243-256
(1983).
[25] Y. K. Lin, W. F. Wu and J. N. Yang, Stochastic modeling of fatigue crack propagation. Proc. IUTAM Symp.
Probabilistie Methods in Mechanics of Solids and Structures, Stockholm, Sweden, 103-110 (1984).
[26] K. Sobcyzk, Stochastic modeling of fatigue crack growth. Proc. I U T A M Symp. on Probabilistic Methods in Mechanics
of Solids and Structures, Stockholm, Sweden, I 11-119 (1984).
[27] J. N. Yang and S. D. Manning, Stochastic crack growth analysis methodologies for metallic structures. Engng Fracture
Mech. 37, 1105-1124 (1990).
[28] A. T. Bharucha-Reid, Elements of the Theory of Markov Processes and Their Applications. McGraw-Hill (1960).
[29] J. L. Doob, Stochastic Processes. John Wiley (1953).
[30] E. S. Rodriguez III, On a new Markov model for the pitting corrosion process and its application to reliability. Ph.D.
Thesis, Mechanical Engineering Department, McGill University, Montreal (1986).
[31] E. Parzen, Stochastic Processes. Holden-Day (1962).
[32] J. W. Provan and E. S. Rodriguez II, Part I: development of a Markov description of pitting corrosion. Corrosion
45, 178-192 (1989).
[33] Annual Book of A S T M Standards, Section 3, Chap. 3.01, American Society for Testing and Materials, Philadelphia,
U.S.A. (1991).
[34] Metals Handbook, Metallography and Microstructures, 9th edn. Vol. 9, pp. 458-479. Metals Park, Ohio: American
Society for Metal (1985).
[35] A. Saxena and S. J. Hudak, Jr, Review and extension of compliance information for common crack growth specimens.
Int. J. Fracture 14, 463-468 (1978).
[36] J. E. Srawley, Wide range stress intensity factor expressions for ASTM method E 399 standard fracture toughness
specimens. Int. J. Fracture 12, 475-476 (1976).
[37] D. A. Virkler, B. M. Hillberry and P. K. Goel, The statistical nature of fatigue crack propagation. J. Engng Metals
Technol,, Trans A S M E 101, 148-153 (1979).
[38] R. P. Wei and G. Shim, Fracture mechanics and corrosion fatigue. Corrosion Fatigue: Mechanics, Metallurgy,
Electrochemistry and Engineering, A S T M - S T P 801, 5-25. American Society for Testing and Materials, Philadelphia,
U.S.A. (1983).
[39] J. N. Yang, Statistical modeling of fatigue-crack growth in a nickel-base superalloy. Engng Fracture Mech. 18, 257-270
(1983).
[40] J. N. Yang, Simulation of random envelope processes. J. Sound Vibration 21, 73-85 (1972).