Paper 03M-161

Probabilistic Approach to Component Condition Assessment, Remaining

Life Prediction and Maintenance Engineering

Dan M. Ghiocel
STI Technologies

Copyright © 2003 Society of Automotive Engineers, Inc.

ABSTRACT maintenance costs are very sensitive to the to the engine

component failure risk estimates.
The paper illustrates how probabilistic physics-based
models can be used for risk-based condition assessment The overall scope of the research effort was to develop a
and life prediction of aircraft jet engines, including the prototype engineering computational tool for predicting
uncertainties in maintenance activities. Although this the unscheduled maintenance event rates by developing
paper focuses on engines, the proposed approach can physics-based stochastic models that are based on
be extended elsewhere. Probabilistic modeling includes actual flight and component design data for current
all significant uncertainties that affect engine reliability, engines. Based on this physics-based stochastic models
such as flight conditions, loading history, manufacturing approach the key engine component reliability measures
deviations, material properties and behavior under can be computed. These computed reliability estimates
random loading and maintenance activities. Maintenance are used to perform a risk-based optimal-cost
uncertainties include those related to NDI techniques and maintenance analysis
operator’s skills. The paper shows the uncertainty effects
of different NDI techniques, maintenance intervals, TECHNICAL APPROACH
operator skills, etc. on the engine reliability. Unscheduled
maintenance rates are computed for given a The prototype engineering tool developed by STI under
maintenance schedule. this project, called ProMACOR (Probabilistic
Maintenance for Cost Reduction) is based on a physics-
INTRODUCTION based reliability engineering approach that bridges and
integrates intimately the structural reliability concepts and
In contrast to the current engine maintenance tools with the classical reliability engineering concepts
engineering based on field data and empirical FMECA and tools.
studies, STI, in collaboration with GEAE, proposed a
physics-based stochastic approach to maintenance To increase ProMACOR’s practicality, its computational
engineering that is based on the detailed integration of all reliability results are also converted to equivalent
design and post-design aspects and uncertainties reliability (failure) models that are currently used in
including significant manufacturing variabilities, practice, such as the Weibull failure models and Uniform
maintenance activities and cost aspects. Hazard Rate (Exponential) reliability models.

By developing physics-based stochastic models for PROMACOR APPROACH

idealizing the operating environment, aero-thermal
loading, structural behavior and material progressive The integration between the structural reliability tools and
damage under variable loading, the maintenance the reliability engineering tools is the essence of the
engineering and cost analysis is approached from an ProMACOR approach. This integration brings with it the
advanced physical understanding and modeling of the capability of stochastic modeling of the physics of
engine behavior. Only by using such a physics-based failures within the classical reliability engineering
stochastic approach, can the engine life-cycle cost framework (and further to cost modeling).
process be adequately understood and controlled from
the design stage. In the classical reliability engineering approach the
designer or maintenance engineer has no ability to
The actual approaches to risk-based maintenance cost understand and control the effects of any environment or
analysis, that are based solely on simple Weibull life component changes on risk and maintenance costs. In
models developed for only very limited tests or field data, contrast, in the physics–based reliability engineering
suffer severely from having a weak foundation by approach the engineer understands the stochastic
neglecting the stochastic physics of failures. And the physics of failure and can control the effects of changes
on risk and costs. The physics-based reliability mortality” failure risk due to a poor component design,
engineering approach links engineering decision to material defects, or due to mishandling, etc. (this is the
reliability and costs. Of great importance is that this responsibility of OEM engine designer and in
approach also reduces the needed amount of testing ProMACOR is considered to input by the analyst).
data. (v) stochastic modeling of component stress and
strain histories at critical locations; this step includes the
ProMACOR brings a new dimension to engine design construction of equivalent stress random histories based
and maintenance by the fact that the comparison can be on the stochastic mission profile inputs and component
made directly in terms of risks and costs. This type of stress computations.
what-if cost analysis forms the basis for an optimized (vi) probabilistic condition assessment and life
design-maintenance approach for fleets of aircraft prediction based on stochastic cumulative damage
engines. By examining how the variation of different mechanics models for crack nucleation stage and
parameters affects component risk prediction accuracy stochastic fracture mechanics-based models for crack
and life-cycle cost estimates, it would be possible to propagation stage (effect of maintenance is not
perform trade studies that maximize the value of the included). This step includes the maintenance
funding used in research for new and existing aircraft uncertainties.
designs.
ProMACOR predicts probabilistic life of an engine
The ProMACOR prototype software has basically two component, including both the crack initiation and crack
important functional options: (i) For an accepted reliability propagation stages, using the following cumulative
it computes the required maintenance intervals to inspect damage models:
the investigated component (the results depend also on
uncertainties in operational environment, component Crack Initiation: Stochastic Cumulative Damage Models
behavior, selected inspection techniques) and (ii) For an (fatigue and creep)
anticipated maintenance strategy it computes component 1) Linear Damage Rule (Miner’s Rule)
reliability and associated maintenance costs for obtaining 2) Damage Curve Approach
an optimum engine design over a family of alternate 3) Double Damage Curve Approach
designs. 4) Lemaitre-Caboche CDM Model
5) Larson-Miller Model (pure creep rupture)
Using ProMACOR a designer or maintenance engineer
can quickly perform what-if analyses to see how different Crack Growth: Fracture Mechanics-based Models
design modifications affect a component’s risk of failure, (fatigue and creep)
the predicted life and/or the induced maintenance costs. 1) Forman Model
For a designer, what-if analyses represent a key aspect 2) Sine Hyperbolic Model
for obtaining robust, affordable and durable cost-effective 3) Modified Sigmoidal Equation Model
designs. Through what-if analyses, the designer
understands, in much more detail, the behavior of his PHYSICS-BASED RELIABILITY ENGINEERING
design, so that he can make the optimal technical
decision. The current approaches to engine maintenance cost
analysis are based solely on FMECA and simple Weibull
The ProMACOR probabilistic condition assessment and life (failure) models that are developed for very limited
life prediction analysis of an engine component include: tests or field data. These practical approaches suffer
(i) stochastic modeling of flight profiles severely from having a weak foundation by neglecting
(ii) stochastic modeling of component loading, the stochastic physics of failures.
environmental surface conditions, material and structural
properties; this step may include modeling of the ProMACOR is developed to handle all situations
component surface boundary conditions, such as including the “break-in” and “wear-out” failure mode
pressure and temperature variations, contact surface periods. However, ProMACOR is developed mainly to
constraint effects on stiffness and damping, material handle the “wear-out” failure types that are generated by
property variations, manufacturing deviations from the progressive damage mechanisms. In the actual version,
baseline geometry, etc. the “break-in” period failures can be included, but it is the
(iii) probabilistic component stress/strain analysis analyst’s responsibility to define the initial failure
to compute multiaxial stress/strain state in the probabilities due to material defects or poor
component for given steady and dynamic boundary workmanship. Figures 1 and 2 show the failure
conditions that are time dependent, such as variable probability curves and hazard rate curves computed for
blade stresses induced by speed variations or variable three critical crack limits with ProMACOR assuming an
thermal stresses in a hot component due to the transient initial reliability condition defined by (i) a zero initial failure
thermal effects, etc. probability (at time zero) that corresponds to a situation
(iv) probabilistic component reliability analysis or with no “break-in” period, Figure 1, and (ii) a non-zero
risk analysis for initial no-usage conditions (no failure probability that corresponds to a situation that
deterioration due to progressive failure mechanisms is includes a “break-in” period, Figure 2.
included). This risk analysis is to identify the “infant
The basic relationship that links the physics-based unplanned failures is greater than the cost of planned
reliability engineering approach with the classical replacements, there is an optimal replacement interval. If
reliability engineering approach is the interval is too short, the replacement costs are too
high. If it is too long, the unplanned failures drive the total
P[Tf £ t + Dt Tf > t ] = cost too high. An optimal replacement strategy finds the
most cost effective interval for replacing the hardware.

Pf ( t ) + [1 - Pf ( t )] expéê- h ( x )dx ùú
t + Dt

ò
ë t û
(1) The optimal replacement interval is the time with the
minimum ratio of the mean cost to the mean time to
failure, MainCost( t ) . The maintenance cost per unit can
The above equation relates the failure probability within a
time interval that in fact defines probabilistic distribution be expressed as follows:
of component life, to the instantaneous failure probability
computed at the starting time of the interval and to the C p [1 - Pf ( t )] + C up Pf ( t )
hazard failure rate variation in the interval. Then, the MainCost ( t ) = t
= min (6)
MTBF can be computed integrating the reliability function
that translates in the equation ò [1 - P (t )]dt
0
f

ò
¥
MTBF = [1 - Pf ( t )]dt (2) where C up = Cost of an unplanned on-line replacement,
0
where C up > C p and C p = Cost of a planned off-line
Equivalent Weibull failure models replacement before failure. If b is greater that one and
the cost ratio is greater that one, the maintenance cost
For practicality purposes, equivalent Weibull component per unit time has a minimum.
life models are determined based on the computational
results of the physics-based reliability analysis. These Non-Destructive Inspection (NDI) Techniques
equivalent “physics-based Weibull life models” have the
advantage that they can be easily compared with the Inspection routines are adopted to detect and remove
existing Weibull models developed based on component cracks with size larger than a rejection limit, resulting in
field failure data. the improvement of reliability towards an acceptable
level. The advantages taken from regular inspections
To compute the two parameters of the Weibull can be reduced or even completely lost if the inspection
distribution a least-square error minimization technique is technique is not appropriately selected in view of the
used to fit the random sample life data. It can be shown given component or the inspection is not able to detect
by doing some simple mathematical manipulations that damage indications that would lead to failure before any
the Weibull distribution assumes a linear relationship repair action could be taken.
between ln(t) and ln(ln(1/1-F(t))) as follows:
The rejectable crack size can be used to evaluate the
ln(ln(1 /(1 - F( t ))) = b ln( t ) - b ln(q) (3) following probabilities, where independence between
additive sizing error and detection is assumed:
By choosing ln(t) as x, the scale on the abscissa, and 1. The probability PR(a) of rejecting a crack with
ln(ln(1/(1-F(t))) as y, the scale on the ordinate, the size a, calculated as the product of the detection
Weibull CFD is represented as a straight line in this probability and the probability of sizing the detected
transformed space (these are the coordinates of the crack larger than aR:
Weibull paper). Thus, by a simple linear regression in the
transformed space, the two Weibull parameters are
easily computed. For equivalent Weibull model the PR (a ) = PD (a )[1 - FE (a R - a )] (7)
instantaneous failure probability is computed by
2. The probability PA(a) of accepting a crack
t
- ( )b with size a , calculated as the product of the detection
Pf ( t ) = 1 - e q
(4) probability and the probability of sizing the detected
crack smaller than aR, added to non-detection
Also, the hazard failure rate at time t is given by probability:

1 PA (a ) = PD (a )FE (a R - a ) + [1 - PD (a )] = 1 - PR (8)
h (t ) = (5)
t
qb( ) b -1 For a given crack size a, the sum of these two
q
probabilities equals unity, since a crack must always be
either rejected or accepted. For a particular case where
Computation of Optimal-Cost Replacement Time: If a
a > aR the function PR(a) is called the probability of
part has a Weibull wear-out distribution and the cost of
correct rejection, while for a < aR the function PA(a) is
called the probability of correct acceptance. It should be 1. Crack Length Statistics Evolution with No or
observed that PR(a) and PA(a), both depend on the Multiple Inspection Intervals
reliability of the inspection technique and on the specified
rejection limit aR. 2. Failure Risk Evolution with No or Multiple
Inspection Intervals
ProMACOR can consider the uncertainties related to
both the crack POD curve and the crack length sizing. 3. Reliability Index Evolution with No or Multiple
Inspection Intervals
Reliability-Based Maintenance Analysis
4. Hazard Failure Rate Evolution with No or
The effect of NDE inspections on the crack propagation Multiple Inspection Intervals
process is illustrated in Figure 3. The plots show the time
evolution of the PDF of crack length in an axonometric 5. Average Hazard Failure Rates per Inspection
view and using contour plots (time flows is from right to Intervals
left). The left-side plot corresponds to no NDE
inspection, while the right-side plot corresponds to four 6. Number of Failures (Removals) per Inspection
NDE inspections at 4000 EFH each. It can be visualized Intervals
from Figure 3 that crack propagation process is a
diffusion stochastic process with no boundary for no 7. PDF of the Parent Crack Length Population after
NDE inspection and with intermittent “mutation” (or Each Inspection
reflective “jump-back”) boundaries that have also a
stochastic nature for repeated NDE inspections. For 8. Equivalent Weibull Failure (Life) Models
each NDE inspection a stochastic “mutation” boundary is
placed at a random crack length crossing level. The 9. Posterior PDF of Life Via Bayesian Updating to
“mutation” boundary has two functions: (i) accept a Incorporate Failure Data
parent crack or (ii) reject a parent crack and produce a
new kid crack (the crack of the replaced or repaired 10. Posterior PDF of Crack Size Via Bayesian
component). Updating to Include Inspection Data
Probabilistic Crack Growth Process Including Inspections For performing reliability-based maintenance analysis,
ProMACOR has two functional options: (i) For
Thus, after each inspection new cracks are born due to anticipated maintenance strategy it computes component
the rejection of the old cracks, i.e. component reliability (Figure 4) (ii) For given reliability level it
replacement or local repair. The new crack populations computes the required maintenance intervals (Figure 5).
are kids produced by the rejection of cracks from the
previous populations. For example after two NDE Based on reliability analysis results associated
inspections, the new kid cracks can have as parents maintenance costs can be computed. A risk-based
(produced them by rejection-mutation) the rejected optimal-cost maintenance strategy can be identified
cracks form the original crack population that was born (minimum cost with risk constraints).
at the starting time or from the next generation of crack
population produced by the rejected cracks at the first The three curve plots in Figure 4 and 5 correspond to
inspection. Thus, an accurate stochastic modeling of the three different values of the critical crack lengths, namely
crack growth process including inspections has to 0.20 in, 0.50 in and 0.80 in. The figures compare the
include the presence of evolutionary multiple statistical computed reliability estimates for two basic cases: (i) No
populations. Inspection and (ii) Multiple Inspections at 4000, 9000 and
11,000 FHs. The rationale behind selecting these
ProMACOR uses the non-normal probabilistic mixture inspection times was not to produce an optimum
model of populations for predicting component reliability. maintenance strategy, but to highlight some key aspects
For each crack length population a lognormal probability of problem. An average operator-skill ultrasonic POD
distribution is assumed. This assumption appears to be curve was considered for this numerical investigation.
reasonable and slightly on the conservative side as
shown by repeated simulation studies performed. The By exploring these types of results, important questions
use of Weibull distribution for crack length population is of the maintenance engineer can be answered. These
less accurate. questions include: What is the overall effect of
maintenance strategy on failure risk evolution and overall
Component reliability is expressed in terms of the associated costs? Is the selected maintenance strategy
instantaneous failure probabilities and reliability indices. efficient from reliability and cost point of view? How
Based on physics-based reliability analysis, for each much does the NDE technique influence the component
critical location of a component ProMACOR computes: reliability and the overall maintenance costs? How much
does the operator skills influence the component
reliability? How much does the crack rejection criteria
influence reliability?
Figure 6 shows the failure probability evolutions for the 2. Ghiocel, D.M, Mao, H., “Probabilistic Life
two cases, without and with maintenance. From these Prediction for Mechanical Components Including
st
plots it can be noted that if a failure probability of 0.001 is HCF/LCF/Creep Interactions”, the 1 MIT Conference of
accepted, then the component life computed for this Computational Fluid Dynamics & Computational
probability level and for the 0.50 in crack stability criteria Mechanics, MA, June 12-15, 2001
is about 8,000 FH with no inspection and about 18,000
FH with three inspections. Further, it can be seen that 3. Ghiocel, D.M.,”Stochastic Field Models for
the second inspection at 9000 FH is more efficient than Advanced Engineering Applications”, 42th
the first inspection at 4000 FH. The first inspection is too AIAAIASME/ASCE/AHS/ASC Structures, Structural
early and therefore has a more reduced effect. The Dynamics, and Materials Conference, AIAA/ASCE/ASME
second inspection reduces the failure probability by a few NDA Forum, Seattle, WA, April 16-19, 2001
orders of magnitude. The size of the downward jumps in
failure probability evolution after each inspection time is a CONTACT
measure of how efficient the inspection at that time is.
Dr. Ghiocel is the Vice President of Advanced
Figure 7 shows the computed maintenance replacement Engineering Applications at STI Technologies Inc.,
costs assuming that the unscheduled on-line located in Rochester, New York. Dr. Ghiocel has 18
replacement is ten times more expensive than scheduled years plus of extensive research lab, university research
off-line replacement. It can be observed that the effect and consulting engineering experience in advanced
of the inspections is to double the optimal-cost computational structural/mechanical analysis for systems
replacement time and to reduce the overall replacement and components, computational stochastic mechanics,
cost to half of one third. nonlinear random vibration and bifurcation analysis,
stochastic finite element analysis,damage mechanics,
Figure 8 shows the effects of the inspection NDE probabilistic risk assessment and component life
technique on the optimal-cost replacement time. It is prediction. He taught graduate courses on Structural
interesting to note that the use of Eddy Current Analysis, Dynamics, Random Vibration, Structural
inspection instead of Visual inspection changes the Reliability and supervised graduate research work and
optimal-cost replacement time from 13,000 FH to 17,000 direct final theses of M.Sc. and Ph.D. students on related
FH and reduces the overall replacement cost by almost topics.
35%.
Currently, he is performing
Figures 9 and 10 show the main screens of the collaborative work with
ProMACOR software for the “Maintenance Strategy” Carnegie Mellon, CWRU,
input options and the “Risk Analysis” output options, Cornell, University of Colorado
respectively. at Boulder, Toledo and
Vanderbilt universities.
CONCLUDING REMARKS
Since 2001, Dr. Ghiocel has
The paper presents a new, efficient risk-based received an Adjunct Professor
maintenance cost analysis tool for aircraft engines, position at Case Western
ProMACOR, based on an accurate physic-based Reserve University (CWRU),
stochastic modeling of material deterioration under Cleveland, Ohio.
random operational conditions.
He is the Chair of the Engineering Probabilistic Methods
Using ProMACOR, the maintenance and design projects and the Co-Chair of the Technology Transfer
engineers can take optimal-cost decisions based on Subcommittee of the RMSL Division of the SAE
accurate estimates of component failure risks including International. He is also the Vice-Chair of the AIAA
maintenance uncertainties Working Technical Group on Non-Deterministic
Approaches.
ACKNOWLEDGMENTS
In 1999 Dr. Ghiocel received the prestigious SAE
The author would like to acknowledge Mike Ghan, "Distinguished Probabilistic Methods Service Award" for
AFRL/PRTC, Dion Duckett and Will Richer, GEAE, for "Demonstrated Excellence in Dedication to the Growth
their support during the research project. and Success of the SAE Probabilistic Methods
Committee Through Production of Technical Documents
REFERENCES and Ceasless Assistance to Probabilistic Methods
Community".
1. Ghiocel, D. M., “Modeling of Complex
Stochasticity for Gas Turbine Applications”, the 2001 The author’s private professional email address is
International Conference on Structural Safety and dghio@aol.com. He is actively involved in the SAE G-11
Reliability, ICOSSAR, Newport Beach, June 18-22, 2001 PMC standards and projects.
Figure 1 Failure Probability and Hazard Rate Evolution for Zero Initial Condition
(“Wear-out” shape of hazard rate curve)

Figure 2 Failure Probability and Hazard Rate Evolution for Non-Zero Initial Condition
(“Bath-Tub” shape of hazard rate curve)

Figure 3. Effect of the NDE Inspection on Crack Length Probability Density

Figure 4. Reliability Analysis Results for Given Maintenance Strategy

Figure 5. Reliability Analysis Results for Given Reliability Level

 Figure 6. Failure Probability Evolution: (a) No Inspection, (b) With Inspections

Figure 7. Maintenance Cost vs. Removal Time: (a) no Inspection, (b) with Inspections

Figure 8. Maintenance Cost vs. Removal Time: (a) Visual, (b) Eddy Current
 Figure 9. Maintenance Strategy Input Screens

Figure 10. Probabilistic Risk Analysis Output Screen