You are on page 1of 24

Fatigue Reliability Analysis Using Dynamic Bayesian Network

Dooyoul Lee
Korea National Defense Univerisity, 1040 Hwangsanbeol-ro, Yangchon-myeon, Nonsan, 33021, Republic of Korea

Abstract

A comprehensive approach is developed to integrate crack growth model, nondestructive test-


ing (NDT), and repairs for fatigue reliability evaluation using Dynamic Bayesian network. All
three cases of inspection results are considered in the model. Repair is The application of the
proposed methods is illustrated through an in-service fatigue problem and inspection techniques
with uncertain POD characteristics are compared.
1. Introduction

Risk is usually interpreted as the chance or the probability of a failure event occurring within a
population details over a period of time [1]. Alternatively, risk is interpreted as the instantaneous
chance or the instantaneous probability of a failure event occurring within a population of details.
For aircraft failure, the instantaneous rate is often defined in terms of per hour or per flight. It is
important that the risk of failure is kept below the allowable level of risk acceptable for continuous
safe operations.
Structural risk and reliability analysis is the probabilistic engineering approach, also known
as the “reliability method”, for supporting the structural integrity analysis of structures and
components under service loads [1]. In recent times, a new challenge has arisen in the aircraft
structural integrity program with the need to maintain aircraft longer in an environment of
reduced funding levels. The probabilistic approach is capable of identifying the source of variables
affecting the fatigue life and fatigue strength of the structure in terms of risk, while eliminating
the over-conservatism that maintains safety [1]. Especially in aging aircraft problems, inspection
timing and cost effectiveness is important for life extension programs in allowing the fleet to
continue operation exceeding by their design life in the most economical manner. For these
reasons, the probabilistic approach is beginning to gain new interests in civil and military aircraft
operators.
In many instances, the interest is on the reliability of a deteriorating structure conditional on
inspection and repair data (e.g. [Straub2020 ref. 14, 16]). What is the novelty of this paper?
Since full-scale testing may not be feasible in many engineering applications, experimental
data on smaller or intermediate modules must be used to update the prior reliability prediction
distribution. If the entire system of equations that are involved in the formulation of the limit
state can be represented by a network showing the relations among the basic random variables,
intermediate quantities, and the overall output, the updating process would be much easier. The
Bayes network approach is well suited for this purpose. The attractive feature of the Bayes
network is the ability to update the statistical information for all the nodes, given an observation
for one node [Mahadevan2005]. This feature is very valuable in the context of model calibration
and validation, when in-service inspection data may not be available on one or more intermediate

2
quantities.
DBN models are able to represent the impacts on the operational conditions (e.g. mainte-
nance actions, production levels, environmental conditions ...) on system reliability by means of
exogenous variables [Weber2012 ref. Weber et al., 2004]. One main contribution is a DBN rep-
resentation of non-homogeneous Markovian processes when using chageable parameters though
time [Weber2012 ref. Ben Salem et al., 2006]. When exploiting a DBN model, there are sev-
eral inference algorithms that are appropriated to different situations [Weber2012]. With the
exact inference algorithm [Weber2012 ref. Jensen, 1996], the 2TBN model is similar to a Marko-
vian model with dynamic independent variables. Several inference algorithms exist are still in
development. Their efficiency depends on the model complexity [Weber2012 ref. Murphy, 2002].
Dynamic Bayesian network is an attractive method for the risk analysis. DBN parses prob-
lem into nodes and edges. Nodes are random variables and edges indicate their probabilistic
relationship. The problem is that existing DBN models do not follow established framework
developed by Lincoln. Yang’s model consider repair [Yang2018]. As mentioned previously, repair
herein indicates the replacement of the considered fatigue-critical detail, which resets the crack
size distribution in the next time slot (Node Ai+1 ) back to the initial crack distribution.
Therefore, a practical and mathematically consistent approach is developed in this paper
to combine the information about various NDT performance measures to update the failure
probability. The proposed methodology is comprehensive in that it includes the possibility of
false calls and non-detection of large flaws [2]. It uses commonly available practical measures of
NDT performance and does not require the definition of any new measure. Section 2 of the paper
presents theoretical background of structural risk analysis and nondestructive testing (NDT).
Section 3 describes the proposed models. In section 4, the model is applied to the computation
of the fatigue reliability of high-strength aircraft bolt.

2. Theoretical Background

2.1. Aircraft structural risk and reliability analysis

In the following, theoretical background for structural risk analysis is provided.

3
The disadvantage of the probability of failure (POF) as a risk definition was that it made no
indication on the risk of failure per hour or per flight [1]. Traditionally in the safe life philosophy,
to protect airframe structures and engine components against fatigue fracture, a scatter factor
is used to limit the operational life of components and structures to a safe life. An acceptable
cumulative POF from fatigue is 10−3 per aircraft per safe life. It (helicopter) simply relies on its
conservative assumptions regarding loads and material properties to achieve the one in a million
POF over the retirement period [Tong2001 ref. 8].
Methods used to set safe lives or inspection intervals are generally based on one or more of the
internationally recognized airworthiness standards such as DEFSTAN 970 or the Joint Services
Specification Guide (JSSG) which superseded the standards MIL-A-8860B used by U.S. Navy
(USN) and MIL-A-87221 used by the U.S. Air Force (USAF) [3]. DEFSTAN 970 specifies safety
limits in terms of a total probability of failure of 1/1000 for an aircraft, whereas the JSSG specifies
acceptable levels of 1/107 per flight hour.
The important random variables in aircraft structural fatigue problems are summarized into
four groups [1]:

• The variability in material properties

• The initial fatigue quality of the component

• The variability in the crack growth rate and

• The reliability of the inspection

Lincoln establishes general framework for structural risk analysis. Uncertainties mentioned
above are considered in this framework as shown in Fig. 00.
Model-based reliability analysis generally uses a demand versus capacity format, correspond-
ing to a desired performance criterion [Mahadevan2005]. At time t, suppose R (t) is the capacity
and S (t) is the demand, then a corresponding safety margin is constructed as

M (t) = R (t) − S (t) . (1)

More generally, the system can be modeled by a limit state function g (X, t). Therein, X is the
vector of input random variables [Straub2020 ref. 1, 2, 17]. Failure is defined to occur when the

4
Material Environmental
Variablity Variablity

Crack Tip
Loading
Shape

Crack Concentration of
Tortuosity Chemical Species

Crack Growth Rate


Variability

Meta Model
Measurement Uncertainty
Error
FEA
Discretization Error

Sparse Data Crack Growth


Law

Data Model
Uncertainty Uncertainty

Figure 1. Fishbone diagram shows the sources of variability for fatigue growth.

5
limit state function is less than zero, and a point-in-time failure event is defined as

F * (t) = {R (t) ≤ S (t)} , (2)

knowing the statistics of R and S, or, more generally, as

F (t) = {g (X, t) ≤ 0} . (3)

Computation of the corresponding probability Pr [F (t)] is typically straightforward. Monte Carlo


simulation methods as well as analytical first-order and second-order approximations have been
developed for this computation, based on either actual computational models (e.g. finite element
models) or response surface approximations, for a large variety engineering problems [Mahade-
van2005 ref. 16].
The function Pr [F (t)] and the derived quantities summarized above describe the reliability of
a single structural system [4]. When the system fails, it might be systematically replaced. Such
a process can be described by renewal theory [Straub2020 ref. 23, 24], which enables calculation
of the cost and risk of the system considering replacements. Such studies have been performed,
e.g. for optimizing risk acceptance criteria [Straub2020 ref. 25, 26].

2.2. Reliability characterization of NDT methods

The objective of an NDT system is to enhance the reliability of structural components by


detecting flaws that can degrade the strength of the component [2]. The structural integrity of
aircraft is largely dependent on NDT. NDT methods include visual, ultrasonic testing (UT), eddy
current testing (ET), magnetic particle testing (MT), penetrant testing (PT), and radiographic
testing (RT). The result of inspection can be classified into two categories. Inspections in the first
category returns hit and miss data. There’s no quantitative signal amplitude but the detection or
non-detection is reported. MT, PT, and some technique of RT belong to this category. Inspections
in the second category returns quantitative signal amplitude. UT, ET and some technique of RT
belong to this category.
Due to many uncertain factors, such as the condition of the structure under inspection and its
service environment, the sensitivity of inspection equipment, material imperfections, and operator
skills, no in-service inspection can be perfect. The probability of detection [POD (a)] expresses

6
the probability of detecting a flaw of a given size a and is the common measure to evaluate
the capability of an NDT technique [2]. The POD of inspections yields binary output can be
evaluated by getting the fraction of detection over total number of inspection. Inspection with
quantitative signal amplitude can be evaluated using the linear relation between the actual flaw
size and the measured flaw size.
Measurement errors can be either systematic (i.e., the measured size is consistently larger or
smaller than the actual size) or random in nature, and they are typically made up a combination
of systematic and random errors [2]. The relationship between the actual and measured flaw size
is described with a linear function determined by regression analysis [5].

â = β0 + β1 a +  (4)

where a is the actual flaw size, â is the measured flaw size, β0 and β1 is the regression coefficients,
and  is the residual random error with zero mean value. The smaller the σepsilon is the better
the NDT performs [2]. The extreme case that β0 = 0, β1 = 1, and σ = 0 represents the ideal
NDT system, which detects a flaw of any size with exact measurement. Usually, a high capability
of flaw detection is connected to a high probability of false call, so there is a trade-off between
detection of flaws and rejection of good components [2]. Detection threshold is determined by the
level of noise amplitude with acceptable false call probability (FCP) because repair or replacement
due to a false call would cause unnecessary economic cost and may result in potential damage.
At detection threshold POD is 50 %.
A convenient expression for the POD curve is
αaγ
POD (a) = . (5)
1 + αaγ
where α and γ are the fitting parameters [6]. The probability that a crack is not detected follows
from Eq. (5) as
αaγ 1
PND (a) = 1 − POD = 1 − γ
= (6)
1 + αa 1 + αaγ
PND is shown in Figure 2. Crack length under PND represent the fraction of undetected cracks
after inspection. The variance of posterior distribution is reduced after inspection. Therefore, the
inspection update our belief on the structural integrity. Handling the detected fraction is related
to the quality of repair and the replaced parts.

7
1 1 1 1
f0(aa) f0(aa)

Likelihood of Not Detecting A

Likelihood of Not Detecting A


Normalized Density of A

Normalized Density of A
0.8 f1(aa) 0.8 0.8 f1(aa) 0.8
L(aa) L(aa)

0.6 0.6 0.6 0.6

0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0
0 2 4 6 8 10 0 2 4 6 8 10
Crack Size (Not to Scale) Crack Size (Not to Scale)

(a) (b)

Figure 2. Effect of inspection with (a) pass-fail and (b) crack size measured cases

3. Models

3.1. Fatigue crack growth model

Fatigue crack growth rate is represented by Paris’ law.

da
= C (∆K)m (7)
dN

where a is the crack length, t is the time of operation, ∆K is the stress intensity factor range,
and C and m are experimentally determined model parameters [7].
Fig. 3 shows a Two-Timeslice BN (2TBN) for fatigue crack growth. Each column represents
the evolution of the crack length distribution for t. The nodes are connected by edges, which
corresponds to a conditional dependency among the variables of the fatigue-crack growth model.
Nodes with incoming edges are conditioned by their parents.
According to Paris’ law (Eq. (7)), the crack length distribution at timeslice t (at ) is determined
by four variables: the two Paris’ law parameters (Ct and mt ), stress range (∆σt ), and crack length
at timeslice t − 1 (at−1 ). Since material parameters C and m are closely correlated [8], only m
is represented in the model. Stress intensity factor is a function of applied stress and crack
geometry. Therefore, the range of applied stress is represented in the model. Crack length at
current timeslice is updated using crack length previous timeslice at−1 .

8
mt−1 mt

∆σt−1 ∆σt

a0t−1 at−1 a0t at

Figure 3. Basic DBN for fatigue crack growth

Yn at Bt

Yt apt

Figure 4. Nodes and edges considering inspections in the DBN

3.2. Modeling inspection

Fig. 4 shows a BN for inspection. Node Yt is the signal amplitude node representing signal
response from a discontinuity in the material. Node Yn depicts the noise of the NDT system
including equipment, wiring, specimen, etc. In the node apt , crack length is updated with a
binary inspection result provided by node Bt .
There are two cases in the result of inspection. The first case is the inspection returns binary
indications, i.e., detected and non-detected. The detection or non-detection are determined by
the POD in Eq. (5). In the DBN, the detection state is determined by comparing the POD with
random number from zero to one.

P OD (a) >A, (Detected)


(8)
A ∼ U (0, 1)

The second case is the inspection returns quantitative signal amplitude. In this case, the crack
length is estimated by Eq. (4) in node Yt . If the signal amplitude is lower than noise, the signal
amplitude cannot be distinguished from the noise (Fig. 5). Nodes Yt and Bt are the observable
nodes, which update current crack length distribution.

9
1
0.46 mm
0.72 mm
0.8 1.13 mm
1.78 mm

Probability
0.6 Noise Amplitude = 6%

0.4

0.2

0
0 20 40 60 80 100
Signal Amplitude (%)

Figure 5. Signal response for different crack lengths

In Fig. 4, each state of the node Bt represents each of the above mentioned cases including
no inspection. When Bt is instantiated as 1, a crack is detected but not measured, and only
detected portion is passed to the following timeslice.

f (ap | B = 1) ∝ L (a | B = 1) f (a; t) (9)

Node Bt instantiated as 2 indicates a crack is not detected, and only non-detected portion is
passed to the following timeslice.

f (ap | B = 2) ∝ L (a | B = 2) f (a; t) (10)

Bt = 3 means that a crack is detected and measured, and the whole detected crack length
distribution is used in the next timeslice.

f (ap | B = 3, Y = y) ∝ L (a | B = 3) L (a | Y = y) f (a; t) (11)

The last state (Bt = 4) is used when no inspection is conducted. In this case, nodes at and apt
have the same distribution.

3.3. Modeling repair


Fig. 6 shows nodes and edges related to repairs after inspection. Node ar represents the
repair crack length node. Node at is the crack length node calculated at timeslice t. As described

10
ar at B

apt

a0t

Figure 6. Nodes and edges considering repairs in the DBN

in the previous section, node Bt indicates the types of inspection result with repair actions: (i)
detected and repaired, (ii) not detected, (iii) detected and measured, and (iv) no inspection.
Repair decision for the case (iii) depends on the size of crack. Using the model shown in Fig. 6,
all four cases of inspection results can be modeled.
The first case (Bt = 1) is a crack is detected, and repair is conducted at timeslice t. In this
case, the crack length distribution is replaced by repair crack length probability.

Pr (apt = i | ar = i, at = j, B = 1) = 1 (12)

The second case (Bt = 2) is a crack is not detected during inspection conducted at timeslice t.
The crack length distribution of the second case is the mixture of repair crack length distribution
and crack length distribution calculated at timeslice t. The detected portion is removed from
the crack length distribution. Thus, each state of crack at timeslice t is likely to transit to the
same state with the same probability as PND of given state. Thus, the probabilities of diagonal
elements coincide with the PND.

Pr (apt = j | ar = i, at = j, B = 2) = PNDi (13)

Provided that crack length distribution after repair is available, the detected portion will be
replaced by the repair distribution. In this case, the replaced portion is the same as POD of
given state.
Pr (apt = i | ar = i, at = j, B = 2) = PODi (14)

11
Figure 7. Transition matrix for crack detected case considering repair crack length distribution

As a result, crack length distribution after inspection is the mixture of the repair and calculated
crack length distributions at timeslice t. Fig. 7 shows the example of transition matrix for the
second case given repair crack length state. In Fig. 7, crack states will be transit to the given
repair crack length state which is thirtieth state (0.72 mm).
The third case (Bt = 3) is the crack is detected and the signal amplitude is measured. As
mentioned above, repair decision for the case (iii) depends on the size of crack. If no repair is
decided, a crack length distribution is not updated at node apt , and the transition probability is
given as:
Pr (apt = j | ar = i, at = j, B = 3) = 1 (15)

Otherwise, the transition probability is the same as the case (i). The last case (B = 4) has the
same transition matrix because no inspection is conducted.

3.4. DBN model

Fig. 8 shows the DBN model for the analysis. Both inspections and repair were considered.
The circular node shown in Fig. 8 is a stochastic node, which implies that the variable is stochastic
for the specified values of the parent nodes. Therefore, the arrow pointing toward it represents the
conditional probability table (CPT). A rectangular node is a functional node, where the variable

12
mt−1 mt

∆σt−1 ∆σt
a0t−1

at−1 a0t at

Yn

Yn
Yt−1 Yt
Bt−1

Bt
apt−1 apt

ar ar
Rt−1 Rt

Figure 8. DBN model with repairs

is obtained from the deterministic calculation based on the specified values of the parent nodes.
The arrow pointing toward a functional node represents a deterministic function. Circular nodes
with dashed lines represent static variables, meaning that the CPTs are identical to the initial
node (circular nodes with solid lines) for all slice.
Reliability is calculated at node R. Simple limit state function is used to calculate the prob-
ability of failure (POF).
g = ac − a (t) (16)

where ac is the critical crack length and a (N ) is the crack length at timeslice t. The limit state
is defined as g = 0, which is the boundary between the failure domain and the no failure domain
in the fatigue crack length space. The POF is the area of the crack length distribution over the
critical crack length.

13
(a) (b)

Figure 9. (a) View of the compressor rotor blade, and (b) magnified view of blade tangs that are held by a pin.
The dashed circle indicates the location of the fatigue crack.

4. Application

4.1. Background

The target of the current analysis is a J85 engine compressor first stage rotor blade (Fig. 9).
The blade is attached to the disc using a tang-splice joint. The joint fails when a fatigue crack
reaches the critical crack length. In depth failure analysis was conducted by Lee et al. [9]. The
reliability of NDT system was evaluated by the authors [10]. Based on the findings from the
failure analysis and the reliability of NDT system, the simple DBN model was constructed and
used to evaluate the risk corresponding to the fatigue crack growth [11]. In the following sections,
we re-evaluate the risk using the proposed DBN model.

4.2. Model parameters and discretization

The maximum revolution per minute (RPM) of J85-GE-21C engine is 16,600. The number of
the first stage stator vanes are 15. Assuming that the average RPM is 70 %, then the revolution
per 100 h is 1.048 × 109 . Each slice of Fig. 8 represents 100 h of operation.

14
Priors of stochastic nodes are determined from experimental data. Initial crack length dis-
tribution is estimated using the equivalent initial flaw size (EIFS) method proposed by Liu and
Mahadevan [12]. EIFS is determined by matching the infinite fatigue life of a component with
and without a presumed initial crack. The fatigue limit (∆σf ) can be expressed using the fatigue
threshold (∆Kth ) and fictitious crack length (aEIF S ) as follows:

∆Kth = ∆σf πaEIF S Y (17)

where Y is a geometric correction factor that depends on the crack configuration. The aEIF S
term in Eq. (17) can be written as
 2
1 ∆Kth
aEIF S = (18)
π ∆σf Y
The EIFS represents the size of crack lengths with an infinite fatigue life. Cracks longer than the
EIFS developed further, and the growth rate can be estimated using a long crack growth law.
aEIF S is a random variable because ∆Kth and ∆σf are random variables. Experimental data
[13] with extremely high cumulative cycles exceeding 1 × 109 cycles were used (Fig. 10(a)) as
well as the fatigue threshold shown in Fig. 10(b) [14].
Since we have no crack length distribution after repair, the same distribution as the initial
crack length was used instead. The reliability of NDT is largely affected by human factors. Fig.
11 shows all inspectors conducting NDT of blade and shows variation of capability from inspector
to inspector. Thus, the single POD curve cannot represent overall inspection reliability. Thus,
the prediction should be conducted per inspector and per base. Noise amplitude also varies per
inspector and per base as shown in Fig. 11(b). Fig. 11(c) illustrates a versus â with a linear model,
showing 95 % confidence and prediction bounds. Both x and y axes are log-transformed. In the
figure, the two horizontal lines represent the noise amplitude and decision threshold, respectively.
The signal amplitude below the noise amplitude is left-censored because it is indistinguishable
from the background noise. The response was normalized to the maximum signal amplitude.
Thus, no right censoring is considered. A Tobit model was utilized as the censored regression
model. Both axes were shows the POD curve used for the analysis. The POD curve estimated
using the calibrated linear model is displayed in Fig. 11(d). We obtained a90/95 = 1.422 mm
(0.056 in) as the detectable crack size, corresponding to a 90 % probability with 95 % confidence.

15
6
Linear Model Linear Model
500 95% Confidence Bounds 95% Confidence Bounds
95% Prediction Bounds 5 95% Prediction Bounds
400
(MPa)

4
300
3
200
= 387.661 = 4.613
0
0 2 = -2.713
100 = -354.678 1
1
= 0.259
= 10.983
0 1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
R R

(a) (b)

0.8 Al2024-T351
Probability

0.6

0.4

0.2

0
0 0.02 0.04 0.06 0.08 0.1 0.12
aEIFS (mm)

(c)

Figure 10. Ti-6Al-4V (a) fatigue limit, (b) fatigue threshold, and (c) cumulative distribution function of EIFS

16
1 0.25

Base A: mean = 2.96, stdev = 2.58


Probability of Detection

0.8 0.2 Base B: mean = 1.74, stdev = 1.63


Base C: mean = 2.07, stdev = 2.56
Base D: mean = 3.29, stdev = 2.86
0.6 0.15

Density
0.4 0.1
Base A
0.2 0.05 Base B
Base C
Base D
0 0
0 0.5 1 1.5 2 2.5 3 0 5 10 15
Crack Length, 2a (mm) Noise Amplitude (%)

(a) (b)

1
10 2 Mean POD
95% Lower Bound
Probability of Detection

0.8 Log-odds log-scale


Amplitude (dB)

1 Decision Threshold = 14.00


10 a = 1.078
0.6 50/50
Noise Amplitude = 6.00
a 90/50 = 1.343
b = 2.51 a = 1.422
0 90/95
b = 1.77
1 0.4
0 = 0.319
10 RMSE = 0.30
= 9.632
Linear Model 0.2
95% Confidence Bounds
95% Confidence Bounds
-1
10 0
0.2 0.5 1 2 3 4 0 0.5 1 1.5 2
Crack Length, 2a (mm) Crack Length, 2a (mm)

(c) (d)

Figure 11. (a) POD curves of all inspectors, (b) noise distributions of all bases, (c) linear relationship between
crack size and ET signal, and (d) estimated POD curve

17
Table 1. Prior distributions for each node

Variable Distribution Mean Standard deviation Cor. Ref.

m, ln(C) Bi-Normal (2.5, log(5.2 × 10−12 ) (0.3, 0.47) −0.9 [14]

∆σ (MPa) Normal 40 5 - -

a0 (mm) Lognormal −7.28 −0.288 - [12, 13, 14]

ar (mm) Lognormal −7.28 −0.288 - Same as a0

Yn (%) Normal 3.29 2.86 - [10]

The prior distribution for each node are summarized in Table 1.


Discretization of Bayesian networks enables the use of exact inference algorithms that exist
for general discrete Bayesian networks [15]. These include the variable elimination algorithm [16]
and the junction tree algorithm [17]. Since the size of discrete Bayesian networks, and associated
computational effort, increases approximately exponentially with the number of discrete states
of its nodes [15], both static and dynamic discretization schemes were developed [18, 15]. Static
discretization has the advantage of being computationally faster and simple to implement. The
discretization of continuous random variables must be done properly to handle reliability problems
in BNs accurately. Discreteization of the distribution tails should be done with care in applications
where extreme events are important. It has been demonstrated that this discretization scheme
can generally preserve the accuracy of the posterior distributions and significantly improve the
computational efficiency compared to MCMC sampling [19].
In this work, the conditional probability distribution was represented by a CPT because
discrete conditional probability was used as a parent. Discretization schemes with non-uniform
spacing were used to minimize the expected approximation error with respect to possible future
observations (evidence). As the number of discrete states improves accuracy. In the general case,
the computational effort for solving the BN is a direct function of the CPT size of the largest
node. In Fig. 8, both at and apt have large CPT. The size of these CPT is 2 ni=1 , where n is
Q

the number of random variables in X, and ni is the number of states used for discretizing Xi .
As shown in Fig. 12(a), result with 80 states for a is more accurate than with 40 states. It is
because coarse discretization speed up crack growth (Fig. 12(b)). In this study, however, we used

18
Table 2. Discretization scheme

Variable Number of states Interval boundaries

m 20 0, log(exp(1) : (exp(3.9) − exp(1))/19 : exp(3.9)), inf

∆σ 20 1 : 60/19 : 60, inf

a0 40 0, exp(log(0.01) : (log(3) − log(0.01))/38 : log(3)), inf

ar 40 0, exp(log(0.01) : (log(3) − log(0.01))/38 : log(3)), inf

Yn , Y 50 0 : 2 : 100

7 4
DBN Model (# state = 40) DBN Model (# state = 40)
6 DBN Model (# state = 80) DBN Model (# state = 80)
MCS MCS
3
Reliability Index

Density
After inspection with no crack detected
4 2 t = 1000 h

3
1
2

1 0
0 500 1000 1500 2000 0 0.5 1 1.5 2 2.5 3
Time of Operation (h) Crack Length, 2a (mm)

(a) (b)

Figure 12. Comparison of DBN model with MCS showing (a) reliability index β, and (b) crack length distribution
after the first inspection at 1000 h.

40 states to reduce computational demand.

4.3. Calibration and validation

The model is calibrated using field inspection data included in the previous analysis [11]. In
this study, the parameters of the initial variables (m0 , ∆σ0 , and a0 ) were estimated using the
EM algorithm. Fig. 13(a) shows the comparison of reliability indices between prior and posterior
models inspected at 1000 h with no crack detected (case 2). The validity of the calibrated model
was evaluated using k-fold cross validation. Entire data was divided into five mutually exclusive
and collectively exhaustive data sets. Fig. 13(b) shows k-fold cross validation results showing
confidence calculated using Bayes factor [20].

19
1

0.998

Confidence
0.996

0.994

0.992

0.99
1 2 3 4 5
5-fold Cross-validation

(a) (b)

Figure 13. (a) Comparison of reliability indices between prior and posterior models inspected at 1000 h with no
crack detected (case 2) and (b) k-fold cross validation result

4.4. Discussion

Among all inspection cases, case 2 is the mostly affected when considering repair crack length
distribution. Case 2 is also the most important case because in most cases the crack is not
detected. Without considering repair crack length distribution, the resultant reliability is either
underestimated or overestimated. Initially, the risk is overestimated but overestimated eventually
as shown in Fig. 14(a). It is because inspections without detection remove “detected” portions.
The detected portion is “repaired” and moved back to the left to the repair initial flaw size
distribution. In Fig. 14(b), posterior crack length distribution without considering repair has
thicker right tail. As the number of inspection increases, the variance is decreased. When the
variance on the state of crack becomes too narrow, the underestimated risk results. This process
is widely used in commercial risk analysis software such as PROF developed by UDRI (University
of Dayton Research Institute) with support from USAF.
Fig. 15 shows reliability indices by all three cases of inspection results. The first case is the
crack is detected and replaced (Fig. 15(a)). The second case is the crack is not detected (Fig.
15(b)). The third case is the crack is detected and measured (Figs. 15(c) and 15(d)). When
the crack is detected and measured, the posterior crack length depends on the measured signal
amplitude. Figs. 15(c) and 15(d) show reliability indices with inspection where 10 % and 30 %

20
7 4
Repair Reapir
6 No Repair No Repair
3
Reliability Index

5 After inspection with no crack detected


t = 1000 h

Density
4 2

3
1
2

1 0
0 500 1000 1500 2000 0 0.5 1 1.5 2 2.5 3
Time of Operation (h) Crack Length (mm)

(a) (b)

Figure 14. Comparison of DBN model with or without considering repair showing (a) reliability index β, and (b)
crack length distribution after the first inspection at 1000 h.

signal amplitudes are measured.

5. Conclusion

[1] Tong, Y. C., 2001. Literature Review on Aircraft Structural Risk and Reliability Analysis.
Tech. rep., Defence Science and Technology Organisation, Melbourne, Victoria.

[2] Zhang, R., and Mahadevan, S., 2001. “Fatigue Reliability Analysis Using Nondestructive
Inspection”. Journal of Structural Engineering, 127(8), pp. 957–965.

[3] White, P., 2006. “Review of Methods and Approaches for the Structural Risk Assessment
of Aircraft”. In DSTO-TR-1916, Defense Science and Technology Organisation.

[4] Straub, D., Schneider, R., Bismut, E., and Kim, H. J., 2020. “Reliability analysis of deteri-
orating structural systems”. Structural Safety, 82(August 2019), p. 101877.

[5] Heasler, P., Taylor, T., and Doctor, S., 1993. Statistically based reevaluation of PISC-II
round robin test data. Tech. rep., U.S. Nuclear Regulatory Commission, Washington, DC.

[6] Cohen, M. L., Kulkarni, S. S., and Achenbach, J. D., 2012. “Probabilistic Approach to
Growth and Detection of a Truncated Distribution of Initial Crack Lengths Based on Paris’
Law”. Structural Health Monitoring, 11(2), pp. 225–236.

21
(a) (b)

(c) (d)

Figure 15. Reliability when crack (a) detected and replaced (case 1), (b) not detected (case 2), (c) detected and
10 % signal amplitude measured (case 3), and (d) detected and 30 % signal amplitude measured (case3)

22
[7] Virkler, D. A., Hillberry, B. M., and Goel, P. K., 1979. “The Statistical Nature of Fatigue
Crack Propagation”. Journal of Engineering Materials and Technology, 101(2), p. 148.

[8] An, D., Choi, J. H., and Kim, N. H., 2012. “Identification of correlated damage parameters
under noise and bias using Bayesian inference”. Structural Health Monitoring, 11(3), pp. 293–
303.

[9] Lee, B. W., Suh, J., Lee, H., and Kim, T. G., 2011. “Investigations on fretting fatigue in
aircraft engine compressor blade”. Engineering Failure Analysis, 18(7), pp. 1900–1908.

[10] Lee, D., Yang, S., Park, J., Baek, S., and Kim, S., 2017. “Investigation of Detectable Crack
Length in a Bolt Hole Using Eddy Current Inspection”. Transactions of the Korean Society
of Mechanical Engineers A, 41(8), pp. 729–736.

[11] Lee, D., and Achenbach, J. D., 2016. “Analysis of the Reliability of a Jet Engine Compressor
Rotor Blade Containing a Fatigue Crack”. Journal of Applied Mechanics, 83(4), p. 041004.

[12] Liu, Y., and Mahadevan, S., 2009. “Probabilistic fatigue life prediction using an equivalent
initial flaw size distribution”. International Journal of Fatigue, 31(3), pp. 476–487.

[13] Morrissey, R. J., McDowell, D. L., and Nicholas, T., 1999. “Frequency and stress ratio effects
in high cycle fatigue of Ti-6Al-4V”. International Journal of Fatigue, 21(7), pp. 679–685.

[14] Ritchie, R. O., Boyce, B. L., Campbell, J. P., Roder, O., Thompson, a. W., and Milligan,
W. W., 1999. “Thresholds for high-cycle fatigue in a turbine engine Ti-6Al-4V alloy”.
International Journal of Fatigue, 21(7), pp. 653–662.

[15] Zwirglmaier, K., and Straub, D., 2016. “A discretization procedure for rare events in Bayesian
networks”. Reliability Engineering and System Safety, 153, pp. 96–109.

[16] Zhang, N. L., and Poole, D., 1994. “A simple approach to Bayesian network computations”.
In Proceedings of the 10th Canadian conference on artificial intelligence.

[17] Lauritzen, S. L., and Spiegelhalter, D. J., 1988. “Local Computations with Probabilities
on Graphical Structures and Their Application to Expert Systems”. Journal of the Royal
Statistical Society: Series B (Methodological), 50(2), pp. 157–194.

23
[18] Zhu, J., and Collette, M., 2015. “A dynamic discretization method for reliability inference in
Dynamic Bayesian Networks”. Reliability Engineering and System Safety, 138, jun, pp. 242–
252.

[19] Yang, D. Y., and Frangopol, D. M., 2018. “Evidence-based framework for real-time life-
cycle management of fatigue-critical details of structures”. Structure and Infrastructure
Engineering, 14(5), pp. 509–522.

[20] Sankararaman, S., Ling, Y., and Mahadevan, S., 2011. “Uncertainty Quantification and
Model Validation of Fatigue Crack Growth Prediction”. Engineering Fracture Mechanics,
78(7), pp. 1487–1504.

24

You might also like