1

Automated Health Management for Gas Turbine

Engine Accessory System Components
Carl S. Byington
Matthew J. Watson
Sudarshan P. Bharadwaj
Impact Technologies, LLC
2571 Park Center Blvd., Suite 1
State College, PA 16801
814-861-6273
[email protected]
AbstractTraditional engine health management
development has focused on major gas turbine engine
turbomachinery components, such as disks, blades, and
main bearings, because these components are expensive to
maintain and their failures frequently have safety
implications. However, the majority of the events that
compromise mission success and equipment availability in
military aircraft arise from the degradation or failures of
engine accessory system components, such as valves,
pumps, and actuators. Failure or statistical-based
maintenance of these components fails to account for
unanticipated and extreme operating scenarios, which are a
major cause of unscheduled maintenance events. U.S.
military systems are thus moving toward condition-based
maintenance (CBM), wherein maintenance is performed as
and when required, thus improving asset availability and
contributing significantly to mission success.
1 2

The authors have developed low-overhead diagnostics and
prognostics techniques, which would enable a shift toward
CBM of engine accessory components . The current work
focused on aircraft fuel and lubrication systems. Model-
based and data-driven techniques were developed to provide
reliable health assessments of hydraulic pumps and valves,
which are essential components on these systems.

TABLE OF CONTENTS
1. INTRODUCTION..................................................... 1
2. TECHNICAL APPROACH...................................... 2
3. DATA COLLECTION.............................................. 5
4. RESULTS AND DISCUSSION................................. 6
5. CONCLUSIONS .....................................................11
6. FUTURE WORK...................................................11
7. ACKNOWLEDGMENTS........................................11
8. REFERENCES .......................................................11
9. BIOGRAPHY.........................................................12


1
1-4244-1488-1/08/$25.00 2008 IEEE.
2
IEEEAC paper#1313, Version 2, Updated October 23, 2007
1. INTRODUCTION
Prognostic Health Management (PHM) technologies are key
enablers of Condition-based Maintenance (CBM)[1],[2],
CBM Plus (CBM+) [3], and Autonomous Logistics (AL).
True CBM enabled by PHM allows operational availability
to be balanced with total operating cost using a risk-based
decision making process. Such algorithms would find ready
application on board the current fleet of military aircraft, as
well as on emerging platforms such as the F-35 Lightning II
(a.k.a, Joint Strike Fighter). Not only does the current work
help address safety and life cycle cost concerns, but it also
addresses the risk that unforeseen failures can have on
system readiness. Effective military operations require
accurate knowledge of the health of operational assets,
which can be obtained by using equipment monitoring and
assessment technologies. This knowledge, when combined
with the anticipated wear-and-tear effects of the planned
mission on each component, is vital for optimal planning
and mission execution. Automated condition assessments
and prognostics thus achieve the goal of putting the right
asset on the right mission at the right time.
Low overhead techniques are needed that could provide
reliable health assessment, diagnostics and prognostics of
gas turbine accessories. Such techniques would enable a
shift in maintenance strategies towards condition-based
maintenance. Solutions to this problem would rely on
applying novel algorithms to the existing signal inputs and
outputs of the accessory components, rather than on the
induction of additional sensors, which would add
significantly to cost and complexity.
The authors have therefore developed embeddable software
modules for autonomous health management of gas turbine
engine fluid system components. These modules allow
incipient detection of impending failure and pre-emptive
action at opportune times for maintenance. An experimental
setup was chosen to represent aircraft fuel and lubrication
systems. Faults were then seeded on the setup, and the
developed routines were trained and validated with
experimental data.
2
2. TECHNICAL APPROACH
Figure 1 illustrates the developed approach to gas turbine
engine fluid system health management. The combination of
model-based and analytical approaches, as shown in the
figure, captures the physical characteristics of the
component, along with diagnostic information extracted
from analyzing available data from the system. This ensures
a more confident and reliable prediction. The model-based
approach ensures that faults are traced back to physically
meaningful parameters, while the data-driven approach
captures higher order dynamics, which may be difficult to
model.
Model-based and Data-driven PHM
The model-based approach to PHM applies physical
modeling and advanced parametric identification
techniques. As an advantage over black-box or purely
data-driven health-monitoring schemes, faults and failure
modes are traced back to physically meaningful system
parameters. The approach employs a mathematical dynamic
model of the system that is directly tied to the physical
processes that drive the health of the component. The
difference between the simulated response (obtained by
feeding the control command to the system model) and
actual response (obtained from the system running real-
time) is used to perform an estimation of system parameters
(e.g., efficiency, friction factors, etc.). The estimated
parameters are then compared with the baseline parameters
to identify, isolate, and quantify system faults.
An additional benefit of developing a model for model-
based PHM is that the model can also be used to simulate
fault cases or fault levels that are not cost effective (or
possible) to execute on a full -system, full-scale
experimental test platform. These faults may be quickly and
inexpensively seeded into the model for simulation of the
system response. Use of available experimental data is,
however, still crucial to this effort for the purpose of tuning
and validation.
In addition to the model-based approach, the authors have
implemented data-driven techniques to capture information
that might be difficult to capture in a physical model, owing
to model limitations, measurement uncertainties, higher
order effects, or other dynamic confounds. Data-driven
approaches continuously monitor sensor signals and identify
differences from baseline signals. Faulted conditions may
cause certain signals to increase or decrease, while having
negligible effect on the remaining signals. By identifying
the signals and features that increase or decrease as
compared to baseline operation, the presence of various
faults may be inferred, while the degree of deviation from
baseline responses is indicative of the severity of the fault.
A core concept within the data-driven PHM approach
(especially when applicable to high bandwidth data) is the
extraction of features, a process that is common within the
condition monitoring and automated health management
community. Fundamentally, features represent a reduced set
of data, or information that can be closely tied to the health
of the system. The need for feature extraction arises
primarily due to a recognized inability to store raw data over
long periods of time. Second, and perhaps more
importantly, most of the raw data does not contain insightful
information. Proven techniques for generating features
include signal processing and neural network (black-box)
modeling. Within the overall PHM architecture, these
features provide collaborative, quantitative evidence of
degradation in the system.

Figure 1 Technical approach
3
Demonstration Target
In order to demonstrate the model-based and data-driven
routines described previously, experimental data was
collected on Impacts Fluid System Test Bench (FSTB).
This system is rated to pressures of up to 10 bar (150 psi)
and temperatures greater than 120
?
C (250
?
F), and is
capable of pumping up to 120 LPM (30 GPM) of oil. The
system is representative of aircraft lubrication systems,
which typically operate at pressures of 60 to 100 psi. While
aircraft fuel systems operate at pressures of 100 bar (1500
psi), the techniques are being developed to be independent
of operating pressures or flow rates, and would thus also be
applicable to aircraft systems that operate at higher
pressures and flow rates.

Figure 2 Fluid System Test Bench (FSTB)

Figure 3 Line Schematic of the FSTB
The FSTB system consists of an internal gear pump that is
driven by an induction motor. The setup is instrumented
with temperature, pressure, and flow sensors. A valve
controls the oil pressure, while a heating element regulates
its temperature. A Variable Frequency Drive (VFD) controls
the pump, and control software is used to control the valve
and heating elements. Figure 2 shows a snapshot of the
Fluid System Test Bench, while Figure 3 presents a line
schematic depicting the test section and the controls and
sensors with which the test stand is equipped.
Modeling
A dynamic model of the FSTB was developed to facilitate
the development of model-based PHM routines. A block
diagram of this model is shown in Figure 4. As shown in the
figure, the model consists of various blocks representing the
variable frequency drive (VFD) controller, induction motor,
gear pump, PID pressure controller, and the flow passages
within the test stand. Fault blocks may be incorporated into
any of these structure blocks, so as to simulate faults in the
corresponding test stand components. The important pump
parameters are the pressure head developed and the flow
rate delivered by the pump.

Figure 4- Dynamic model of the FSTB
A number of assumptions have been made for this model,
which neglect dynamic effects in the pump system. The
assumptions that were made are:
(1) Incompressible flow since the pressures
encountered in the FSTB are less than 100 psi, this
is a good assumption for the hydraulic fluid in the
setup.
(2) Fluid capacitance is discounted since the flow is
assumed to be incompressible, there is no
volumetric storage.
(3) Fluid inertance is discounted since the flow is
steady state, there is no time variation of the flow
rate.
(4) Leakage terms are modeled with a linear hydraulic
resistance.[6]
(5) Fluid viscosity is assumed to vary with temperature
in Arrhenius fashion.[7]
4
(6) Fluid viscosity is assumed to be invariant with
pressure this is a good approximation, for the
small range of pressures encountered in the FSTB.
A dynamic model of this type might require more resources
than available on legacy on-board computing platforms.
Since the end goal of this research is to develop PHM
systems that operate real-time on aircraft platforms, a
reduced-order model was also developed and validated
against the full, dynamic model. This reduced-order model
was developed based on the governing equations of fluid
flow in positive displacement pumps and makes the same
assumptions as the dynamic model above. In addition, the
reduced-order model also assumes steady state flow. The
mathematical model is described below.
For positive displacement pumps, the amount of fluid
transported per shaft revolution is a fixed quantity, which
means that at steady state, the flow rate delivered is ideally
only a function of the pump RPM, and is independent of the
pressure head developed or the temperature of the fluid.
However, in practice, there is a slight dependence of the
flow rate on the fluid pressure and temperature even for
incompressible fluids, owing to the fact that the seals
between the stationary and rotating elements within the
pump are not perfect. Thus, there is always a small flow loss
due to leakage, and this leakage increases with the pressure
difference between the outlet and the suction sides. In
addition, leakage flows also vary inversely with the
viscosity of the fluid, which in turn decreases with the fluid
temperature. These relationships are summarized in the
equation below.
( )
i
L
P P P
T R
P Nq
Q
=

=
60
(1)
where:
(K) emperature t luid F T
(psi) ressure p Inlet P
(psi) ressure p Outlet P
(GPM) rate Flow Q
(psi/GPM) resistance Leakage R
RPM Shaft N
rev GPM volution re per Flow q
i
L
=
=
=
=
=
=
= ) / (

In Equation (1) above, the leakage resistance, R
L
, represents
the hydraulic resistance offered by the seals to the fluid
leakage flow. This resistance term is a function of the fluid
viscosity:

L
R

(2)
where ? is the viscosity of the fluid. As the above equation
shows, the resistance to fluid flow increases with increasing
fluid viscosity. The viscosity of hydraulic fluids may be
assumed to vary with temperature in Arrhenius fashion, as
below:

=
T T
b
1 1
exp
0
0


(3)
where ?
0
is the viscosity at the reference temperature, T
0
,
and b is the temperature coefficient of viscosity. As the
above equation indicates, viscosity of most liquids decreases
with increasing temperature. Thus, the leakage resistance
may be assumed to decrease with temperature as:

=
T T
b R R
L L
1 1
exp
0
0

(4)
where R
L0
is the leakage resistance at the reference
temperature. The leakage flow, Q
L
, may thus be assumed to
vary with pressure and temperature as below:
( )
( )
0
0
0
0 0
0
0 0 0
0
0 0
1 1
exp
1 1
exp
1 1
exp ,
L
i
L
i
i
L
i
i
L
i
L
i
L
L
R
P P
Q
T T
b
P P
P P
Q
T T
b
P P
P P
R
P P
T T
b
R
P P
T R
P
T P Q

=

(5)
where Q
L0
is the leakage flow at the reference conditions,
(P
0
,T
0
). Equation (1) may then be rewritten as:

=
0 0
0
1 1
exp
60 T T
b
P P
P P
Q
Nq
Q
i
i
L

(6)
Recasting this equation in terms of the pump RPM gives:
q T T
b
P P
P P
Q Q N
i
a
L
60 1 1
exp
0 0
0

+ =

(7)
Equation (7) above provides a basis for applying model-
based PHM on the gear pump system.
5
This mathematical model uses the control command signal
to predict the system response, and the differences between
this predicted response and the actual response of the system
yield system parameters. The estimated parameters are then
compared with the baseline health level parameters to
identify and isolate system faults and provide a measure of
fault severity. The pump parameters may be computed real-
time on-board aircraft platforms.
Data Driven PHM
The authors have also implemented data-driven techniques
for fault detection and isolation. These techniques utilize
real-time parameter observation, and monitor signal
behavior for increased levels of deviation from the expected
baseline profile. These techniques do not rely on detailed
system models, and are thus well suited to detection of
faults such as cavitation, which are not easily represented by
mathematical modeling. However, understanding the
physics of the system and critical failure modes is vital for
the selection of parameters to monitor and for the
development of data features, which best capture deviations
from baseline behavior.
Fault Pattern Classification and Evolutionary Prognostics
A Statistical Fault Pattern Classification approach was also
developed and used to map changes in model parameters
(identified using parameter optimization) to levels of
degradation or failure. This approach relies on gauging the
proximity and rate of change of the current component
condition (i.e., features) to known fault conditions within N-
dimensional feature space. As part of this approach, an
estimation of the uncertainty is required and statistical
approaches were implemented. Figure 5 illustrates this
approach in two-dimensional parameter space. Starting at
the origin, which represents the initial, normal operation,
measured parameter distributions begin to shift as some type
of degradation begins to occur. In the figure, the points
labeled 2% Fault and 4% Fault represent the parameter
space at known fault conditions. Over time, the measured
parameter joint distribution moves to other points in the
space (Time1 and Time2) and the path of this movement
can be projected to determine the future health state of the
system. The developed probabilistic approach calculates the
dis tance between the current measured condition and known
fault conditions in parameter space (termed the Euclidean
distance). The fault regions having the shortest Euclidean
distance are used to predict impending failures.
Evolutionary Prognostics were used to model fault-to-
failure progression and determine which failure mode will
result in system failure, as well as the time remaining before
the current health state progresses to functional failure. In
the case of Figure 5, the Euclidean distance between the
current state and the fault region representing functional
failure becomes gradually smaller as the system
degrades.[8]

Figure 5 Fault pattern classification and evolutionary
Prognostics
The time to failure is determined by tracking and projecting
the path of each feature within feature space using a
statistical trending method. These projections can then be
fused in feature space to determine the amount of time until
the current condition reaches the known functional failure
region in feature space. A double exponential smoothing
(DXS) algorithm was implemented to track and project the
features. DXS employs an exponentially weighted averaging
function that can forecast future values of a vector based on
past observations. These past observations are weighted
using an exponential function, which ensures that the
current observation receives the maximum weightage, and
that progressively lower weights are assigned to past
observations, with the earliest observations receiving the
least weightage. DXS forecasts were made one time unit
into the future and smoothing statistics were updated and
used to make the next prediction. This process was repeated
until the predicted feature space cluster approached the
functional failure region for any failure mode. The
advantage of representing fault regions within feature space
is that multiple, competitive failure modes can be mapped
within the same N-dimensional feature space. As a result,
the current health and time-to-failure can be assessed for
each failure mode concurrently.
3. DATA COLLECTION
As described in an earlier section, experimental data is
needed for training and validation of the developed routines.
In addition to baseline data, the developed routines require
data collected on the system under faulted conditions. These
data enable the derivation of faulted features, i.e., features
that were obtained from a faulted system. The full set of
baseline and faulted features then provide a means to map
the features obtained from an actual operating system, to the
health of the system. The authors therefore developed and
executed a test matrix for collecting data from the FSTB
system.
6
Fault Simulation
The collection of faulted data requires that known faults
be seeded into the system. In most cases, it would be
desirable not to permanently damage the system by
introducing the fault, and this requirement dictates a careful
choice of faults to be simulated, as well as their severity
levels. The faults that were simulated on the setup include:
(1) Flow loss due to leakage
(2) Stuck valve
(3) Damaged gear teeth
(4) Cavitation
These faults were seeded on the FSTB setup as follows.
Pump seal leakage was simulated by diverting some of the
flow from the pump back to the sump, bypassing the test
section. This is illustrated in Figure 6. As a result , the pump
has to operate at a higher RPM to deliver the same flow to
the test section.
The stuck valve condition was simulated using a gain block,
which was added to the valve command signal that controls
the fractional opening of the valve. This simulated valve
blockage, since a reduction in the effective valve area would
cause a higher pressure in the test section. Thus, the valve
controller would compensate by opening the valve further,
so as to maintain the pressure setpoint.

Figure 6 Simulating leakage
Gear faults were simulated by reducing the flow command
to the pump for a fraction of each revolution. For example,
if the gear within the pump (which drives fluid flow) had
five teeth, a damaged tooth would cause a reduction in the
flow for one fifth of each shaft revolution. The extent of this
flow reduction would depend upon the severity of the
damage.
Pump cavitation was induced by reducing the supply side
pressure to the pump, by draining the sump, as shown in
Figure 7. This also introduces air bubbles into the fluid
flow, which is another cause of cavitation.

Figure 7 Seeding pump cavitation
In addition to the seeded faults just described, baseline data
were also collected. The data consisted of the RPM,
pressure, temperature, and flow signals at various operating
conditions. These data were then used in the model-based
and data driven approaches described above to estimate the
health of the major components in the system.
4. RESULTS AND DISCUSSION
The collected data were used to define model parameters
that minimize the error between the predicted and measured
signals. Shifts in these parameters were then evaluated to
identify features that indicate the presence of faults on the
FSTB setup. Some of these features pertained to the pump,
while others were related to the valve. One such feature was
a parameter derived from the pump signals . This parameter
was found to be a good indicator of pump leakage, being
remarkably invariant over a wide range of operating
pressures. This model parameter was obtained by feeding
the pump RPM signal to the model of the system, and
predicting the output flow. Least-squares minimization of
the error between the model-predicted flow and the actual
flow read by the flow meter on the setup yielded a best-fit
value for this parameter.
Figure 8 shows the value of this model parameter for the
various fault cases simulated on the FSTB. The top half of
the figure shows the mean and standard deviation (error bars
represent 1 standard deviation) of this feature for each fault
case (including the baseline), while the bottom half shows
the same parameter over each data run for each of the fault
cases.
7
Baseline Leakage Gear Fault Gear Fault+Leakage Valve Fault Valve Fault+Leakage
0.015
0.0175
0.02
Case
M
e
a
n

V
a
l
u
e
Mean and Standard Deviation of Pump Parameter A: 100 F
0 5 10 15 20 25 30
0.015
0.0175
0.02
P
u
m
p

P
a
r
a
m
e
t
e
r

A
Data Serial Number

Baseline
Leakage
Gear Fault
Gear Fault+Leakage
Valve Fault
Valve Fault+Leakage

Figure 8 Leakage fault results
As Figure 8 shows, the value of the feature is significantly
lower for the cases with leakage, and is unaffected by the
presence of other faults. Thus, the parameter is a reliable
indicator of the presence of leakage in the pump, and the
magnitude of the deviation in this parameter from its
baseline value indicates the severity of the leakage fault.
Figure 9 shows the Probability Density Functions (PDFs) of
the inverse of the pump model parameter for the baseline
and leakage fault cases. These PDFs were computed using
the means and standard deviations of the inverse of the
values displayed in the bottom half of Figure 8, for the
baseline and leakage cases (similar to the first two
datapoints in the top half of Figure 8). As Figure 9 shows,
there is good separability between the baseline and faulted
cases.
50 55 60 65
0
0.4
0.8
1.2
1.6
Inverse of Pump Parameter A
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y

F
u
n
c
t
i
o
n

Baseline
Leakage Fault

Figure 9 PDF of pump parameter A
Valve faults were detected using another developed
diagnostic feature derived from the valve signals . Figure 10
displays the variation of this parameter for each fault case
for all the data runs (bottom half of the figure). The top half
of Figure 10 shows the mean and standard deviation of this
parameter for each of the fault cases simulated on the FSTB.
As seen, the feature value is significantly lower for the valve
fault cases, and is relatively constant for the other cases.
Thus, this parameter provides a reliable means of detecting
valve faults in the system.
Baseline Leakage Gear Fault Gear Fault+Leakage Valve Fault Valve Fault+Leakage
0.015
0.02
0.025
0.03
Case
M
e
a
n

V
a
l
u
e
Mean and Standard Deviation of Valve Parameter A: 120 F
0 5 10 15 20 25
0.015
0.02
0.025
0.03
Data Serial Number
V
a
l
v
e

P
a
r
a
m
e
t
e
r

A


Baseline
Leakage
Gear Fault
Gear Fault+Leakage
Valve Fault
Valve Fault+Leakage

Figure 10 Valve fault results
The probability density functions of the inverse of this valve
parameter for the baseline and valve fault cases are
displayed in Figure 11. Since the valve fault seeded on the
setup was an incipient fault (approximately 10% change in
the valve control signal), there is some overlap between the
two distributions. Even so, Figure 11 shows reasonable
separability between the baseline and fault ed cases (6.2%
missed detection for a 5% false alarm rate).
30 35 40 45 50 55 60 65 70 75 80
0
0.04
0.08
0.12
0.16
Inverse of Valve Parameter A
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y

F
u
n
c
t
i
o
n

Baseline
Incipient Valve Fault
P
FA
=5%
P
MD
=6.2%

Figure 11 PDF of valve parameter A
A third model parameter, also derived from the pump
signals , was found to be indicative of gear faults. Figure 12
shows the mean and standard deviation of this parameter for
the various cases (top half of the figure), as well as the
actual values of this parameter for each data run (bottom
half of the figure) at 100 F. As the figure shows, the
parameter is significantly higher for the gear fault cases
(regardless of the presence of leakage), and is unaffected by
other faults. Thus, it is a good indicator of the presence of
gear faults in the system.
8
Baseline Leakage Gear Fault GF+Leakage Valve Fault VF+Leakage
0.95
1
1.05
Case
M
e
a
n

V
a
l
u
e
Mean and Standard Deviation of Pump Parameter B: 100 F
0 5 10 15 20 25 30
0.95
1
1.05
Data Serial Number
P
u
m
p

P
a
r
a
m
e
t
e
r

B


Baseline
Leakage
Gear Fault
Gear Fault+Leakage
Valve Fault
Valve Fault+Leakage

0.96 0.98 1 1.02 1.04
0
20
40
60
80
100
120
Pump Parameter B
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y

F
u
n
c
t
i
o
n


Baseline
Gear Fault
Figure 12 Gear fault results (left) and PDF of Pump Parameter B (right)
The PDF of this second pump parameter at 100 F is shown
in Figure 12 (right plot). As the figure shows, there is good
separation between the probability density functions for the
baseline and faulted cases.
Data-driven PHM Results
As mentioned in an earlier section, data-driven techniques
were also developed to handle fault cases which are difficult
to capture through the model-based approach. In the current
work, the developed data-driven routines were used to
detect and diagnose pump cavitation. Cavitation is a classic
example of a complex fault phenomenon with multiple
causes, and which, further, is too dynamic to capture using
low-order model parameters.
Since cavitation manifests itself through unsteady pump
performance, it would be expected that variations in the
delivered pressure would be observed. The left half of
Figure 13 presents pressure transducer data for 30 seconds
of healthy baseline testing, while the right half of the figure
presents pressure data for faulted cavitation FSTB testing.
The mean value has been removed from the plots, to
eliminate the static pressure value. It is clear from the time
domain data that fluctuations in the pressure become evident
during pump cavitation.
Such clear variations in the time domain are well suited to
statistical data features. A feature was developed to indicate
an increase in low bandwidth energy, since the fluctuations
due to cavitation were noted to cycle around a frequency of
one Hertz. There were also several possibilities for features
to be derived from the pump pressure signal, and these
features were evaluated based on their fault response. The
fault response for four of these features is illustrated in the
bar plot in Figure 14. As seen, three of the four features in
the plot respond very well to the development of pressure
value spikes in the time domain, with the last feature
responding the best.

Figure 13 Baseline (left) and cavitation (right) mean
removed pressure data

Figure 14 Cavitation feature response
Feature Results Summary
The model-based routines yielded three features, two for the
pump, and one for the valve. It was found that each of these
features uniquely correlated to one particular pump or valve
fault. In addition, a feature was derived from the data-driven
routines, which indicated pump cavitation faults. Table 1
summarizes the results of the PHM routine development.
9
Table 1 Summary of PHM results
Parameter System Fault Fault Severity
False Alarm
Rate (%)
Missed Detection
Rate (%)
Pump Parameter A Pump Leakage Moderate 5 Negligible
Pump Parameter B Pump Gear Fault Moderate 5 Negligible
Valve Parameter A Valve Blockage Incipient 5 6.2%
Data-driven Feature 4 Pump Cavitation Moderate to severe 5 Negligible

Fault Classification and Failure Prognostics
The authors then implemented the developed Fault Pattern
Classifiers to separate the various classes of faults on the
Fluid System Test Bench (FSTB). These classifiers operate
on a set of features derived from real-time data from the
FSTB. This feature set is plotted in N-dimensional feature
space and then compared to known fault regions. The
Euclidean distance between the set of current features and
each fault region is then computed, and the fault region with
the minimum Euclidean distance from the current feature set
indicates the type and severity of the faults on the setup.
As an example, a feature space plot is shown in Figure 15.
The figure shows the basic faults along the axes of the plot.
Since each feature is uniquely linked to a particular fault,
these fault regions fall on the primary axes of the plot. For
example, the developed leakage fault parameter is a good
indicator of leakage faults, and none of the other faults
simulated on the setup cause changes in this parameter.
Similarly, the valve fault parameter indicates valve
malfunction. Thus, combinations of these faults cause the
corresponding features to fall on the primary planes of the
plot, as indicated by the fault regions on the X-Y, Y-Z, and
X-Z planes.
Figure 15 also shows a set of features derived from a
simulated fault run at 100
?
F, 10 psi and 6 GPM of flow
(represented with a teal circle), wherein the faults simulated
consisted of flow loss and valve malfunction. As seen, the
Euclidean distance between this set of features and the
various fault regions is the least for the fault region
corresponding to a combination of a mild valve fault and
mild flow loss due to leakage. Thus, the classifier
successfully indicates the presence of both valve and
leakage faults on the system.
The authors have also applied the developed Evolutionary
Prognostics to model fault-to-failure progression and
determine the time remaining before the current health state
progresses to functional failure. As indicated in an earlier
section, a Double Exponential Smoothing (DXS) algorithm
was implemented to track and project the features across
feature space.
Figure 15 Feature Space Plot and Uncertainty Estimation for Fault Classification
10
Figure 16 Evolutionary Prognostics
As Figure 16 shows, the DXS algorithm begins tracking the
path of the features through feature space as soon as
observations of the feature values become available. At any
point of time, the algorithm may be used to predict the
subsequent path of the features. This is indicated in Figure
16 by the blue cone, marked DXS Forecast. This cone may
approach the functional failure regions for one or more
faults on the system. The time interval over which the
current feature set evolves (as predicted by the DXS
algorithm) into a feature set indicative of failure, then
indicates the time to failure, as shown in Figure 16.
PHM Validation
The PHM routines were validated with additional
experimental data from the FSTB. Faults were seeded into
the setup, and the results of the PHM routines, in the form
of the severities of the various faults, were compared to the
actual seeded faults. Figure 17 displays some of these
results at various temperature, pressure and flow setpoints.
As the figure shows, the seeded faults were successfully
detected by the PHM algorithms.
As seen from the figure, the severity of the leakage fault
reduces with increasing flow. This is because the valve was
opened by a fixed amount to seed the leakage fault, thus
causing roughly the same amount of leakage flow each time.
Thus, the leakage flow as a fraction of the flow through the
test section is lower as the flow setpoint is increased.
Leakage Gear Cavitation Valve
0
0.25
0.5
0.75
1
Fault
F
a
u
l
t

S
e
v
e
r
i
t
y
/
C
o
n
f
i
d
e
n
c
e
Seeded Faults: None


Setpoints:
100 F, 10 psi, 6 GPM
Leakage Gear Cavitation Valve
0
0.25
0.5
0.75
1
Fault
F
a
u
l
t

S
e
v
e
r
i
t
y
/
C
o
n
f
i
d
e
n
c
e
Seeded Faults: Valve Fault with Leakage


Setpoints:
100 F, 25 psi, 12 GPM
Leakage Gear Cavitation Valve
0
0.25
0.5
0.75
1
Fault
F
a
u
l
t

S
e
v
e
r
i
t
y
/
C
o
n
f
i
d
e
n
c
e
Seeded Faults: Leakage


Setpoints:
120 F, 10 psi, 6 GPM
Leakage Gear Cavitation Valve
0
0.25
0.5
0.75
1
Fault
F
a
u
l
t

S
e
v
e
r
i
t
y
/
C
o
n
f
i
d
e
n
c
e
Seeded Faults: Gear Fault with Leakage


Setpoints:
120 F, 40 psi, 15 GPM
Severity
Confidence
Severity
Confidence
Severity
Confidence
Severity
Confidence
Figure 17 PHM validation
11
Real-time Implementation and Demonstration
The authors have developed a software Graphical User
Interface (GUI) to enable real-time implementation of the
PHM analysis techniques described in this paper. A
screenshot of this GUI is shown in Figure 18.

Figure 18 Screenshot of real-time i mplementation
interface
The interface allows the user to control the flow,
temperature and pressure setpoints to the setup. The user
may choose to view the feedback signals from the setup, or,
as shown in Figure 18, to observe the results of the PHM
analysis. These PHM results are computed on-the-fly using
real-time sensor data from the setup.
The interface allows the seeding of software faults such as
gear and valve faults into the setup. The PHM results shown
in the figure include health indices and mean times to failure
(and confidences thereof) for the pump and valve, as well as
the actual and model-predicted flow, pressure and RPM
signals. The parameters obtained from the model-based
analysis are also displayed, along with the baseline values of
these parameters. In addition, the interface has indicators to
alert the user to certain critical faults in the setup, such as
gear damage and cavitation, which might require immediate
shutdown and maintenance on the setup.
5. CONCLUSIONS
Gas turbine engine accessory components such as pumps,
valves, and actuators create a majority of the maintenance
issues in aircraft gas turbine systems. A combination of
model-based and data-driven approaches was implemented
to address the problem of predicting and forecasting the
health of these accessory components. Experimental data
were collected on a test setup representative of aircraft fuel
and lubrication systems, and these data were used to train a
developed model of the system, as well as data-driven
routines. The model-based routines yielded features that
uniquely correlated to pump and valve faults. The data-
driven routines yielded a feature which picked out pump
cavitation. There was a one-to-one correspondence between
the developed features and the faults indicated by these
features. The routines were validated with additional
experimental data from the FSTB setup. The authors have
also developed evolutionary prognostics routines based on
the DXS algorithm for the purpose of trending and tracking
the progress of fault features in feature space. In addition, a
demonstration interface was developed to enable real-time
implementation of the developed techniques. These
techniques constitute a significant initial step towards
addressing the issue of on-board diagnostics and prognostics
for gas turbine accessory components.
6. FUTURE WORK
Work will continue to streamline the developed PHM
routines, and to develop data fusion routines to merge the
results of model-based and data-driven algorithms. The
applicability of the developed PHM algorithms may be
extended to a variety of gas turbine engine fluid systems,
such as engine compressor and nozzle systems, starting
systems and Auxiliary Power Units (APUs), and fuel and
lube systems. Engine fluid system reliability data may be
included to improve the fidelity of the PHM algorithms in
predicting impending failures and Remaining Useful Life
(RUL), given current health status and envisioned use. In
addition, dedicated PHM system prototypes may be
developed for fuel and lubrication system health assessment,
and these prototypes may be tested on OEM test cells.
7. ACKNOWLEDGMENTS
This work has significantly benefited from the invaluable
support and technical input from the Wright Patterson Air
Force Base. The authors would also like to acknowledge
financial support for this work provided through the Air
Force Small Business Innovative Research (SBIR) program.
In addition, we would like to recognize the contributions of
our colleague Matthew Smith.
8. REFERENCES
[1] DUSD(LMR) CBM+Memorandum, November 2002,
Available at Defense Acquisition Guidebook,
http://akss.dau.mil/dag/TOC_GuideBook.asp?sNode=
R5-2-1-2&Exp=Y
[2] Department of Defense Web site, DoD Instruction 5.2
http://akss.dau.mil/dag/DoD5002/DoD5002-
3.9.2.asp#3.9.2
[3] Pratt and Whitney Web site http://www.pratt-
whitney.com/unique/html/releases/102601.html
[4] J. Rhoades, JSF System Development and
Demonstration, CodeOne Magazine, 17, 2002.
Available on-line at :
http://www.codeonemagazine.com/archives/2002/artic
les/arp_02/jsf/index.html
12
[5] U.S. Air Force Web site http://www.airforce-
technology.com/projects/jsf/
[6] Jelali, M. and A. Kroll, Hydraulic Servo-systems :
Modeling, Identification and Control, London:
Springer-Verlag, 2003.
[7] F.M. White, Fluid Mechanics, 4th ed., New York:
McGraw Hill, 2000.
[8] Roemer, M, Ghiocel, D. A Probabilistic Approach to
the Diagnosis of Gas Turbine Engine Faults
9. BIOGRAPHY
Carl S. Byington is a Professional
Engineer and the Director of
Systems Engineering at Impact
Technologies and possesses over
17 years in the design, analysis
and testing experience with
mechanical, thermal, and fluid
power systems. He currently
performs as the Principal
Investigator on the development of
innovative prognostics and
health management technologies for military and
commercial customers, and he has successfully lead
numerous programs for the Army, Navy, Air Force, DARPA,
and military OEMs. Prior to joining Impact, Carl worked at
NASA Langley Research Center performing air breathing
propulsion research, and more recently, the Penn Stat e
Applied Research Laboratory (PSU ARL) as the
Department Head in Condition-Based Maintenance. Carl is
a lead and co-author on 2 patents related to PHM
technology, and he has published over 65 papers, book
chapters, magazine and journal articles related to signal
processing, data fusion, statistical analysis, model-based
prognostics, and predictive health management.
Matthew J . Watson is a Manager
of Dynamic Systems at Impact
Technologies with 8 years
experience in the design,
development, and testing of
diagnostic and prognostic systems.
He has participated in the design
of advanced feature development,
fault classification, and dynamic
systems modeling techniques for a
variety of applications, including
gas turbine, flight control, power transmission, drive
train,electrochemical, fluid, and hydraulic systems. Prior to
joining Impact, Matt worked in the Condition-Based
Maintenance department of PSU-ARL, where he focused on
model-based PHM development of electrochemical and fuel
systems. He has co-authored 23 papers related to advanced
sensing techniques, signal processing, diagnostics and
control, model -based prognostics, data fusion, and
machinery health management and is co-author on 2
patents.
Dr. Sudarshan P Bharadwaj
received his undergraduate
degree in Mechanical
Engineering from the Indian
Institute of Technology, Madras
(India), as well as Masters and
Ph.D. degrees in M.E. from the
Pennsylvania State University. He
has published several papers in
his field. He has been involved
with a number of projects at
Impact related to the health
management of gas turbine engines, aircraft, and ship
systems, and has considerable experience in developing
diagnostics and prognostics routines for these systems.