Content uploaded by Ahmet Duyar
Author content
All content in this area was uploaded by Ahmet Duyar on Jan 05, 2014
Content may be subject to copyright.
JOURNAL
OF
GUIDANCE, CONTROL,
AND
DYNAMICS
Vol.
15, No. 2,
March-April
1992
Fault
Diagnosis
for the
Space Shuttle
Main
Engine
Ahmet Duyar*
Florida
Atlantic
University,
Boca
Raton,
Florida 33431
and
Walter
Merrillt
NASA
Lewis
Research
Center,
Cleveland,
Ohio
44135
A
conceptual design
of a
model-based fault detection
and
diagnosis system
is
developed
for the
Space Shuttle
main
engine.
The
design approach consists
of
process modeling, residual generation,
and
fault
detection
and
diagnosis.
The
engine
is
modeled using
a
discrete
time,
quasilinear state-space representation. Model parameters
are
determined
by
identification. Residuals generated from
the
model
are
used
by a
neural network
to
detect
and
diagnose engine component faults. Fault diagnosis
is
accomplished
by
training
the
neural network
to
recognize
the
pattern
of the
respective
fault
signatures. Preliminary results
for a
failed valve, generated using
a
full,
nonlinear
simulation
of the
engine,
are
presented. These results indicate
that
the
developed approach
can be
used
for
fault detection
and
diagnosis.
The
results also show that
the
developed model
is an
accurate
and
reliable
predictor
of the
highly nonlinear
and
very complex engine.
Introduction
T
HIS
paper describes
a
model-based fault diagnosis system
based
on a
neural network classifier
for the
Space
Shuttle
main engine (SSME).
The
system
may be
used
to
monitor
the
life
cycle
of
engine components
and for the
early detection,
isolation,
and the
diagnosis
of
engine faults.
As
such,
the
proposed system
will
be one
part
of a
larger, engine health
monitoring
system.1
The
health monitoring system
will
allow
for
accommodation
of
faults, reduce maintenance
cost,
in-
crease engine availability,
and be one
part
of an
integrated,
intelligent
control
system2
for the
SSME.
A
description
of
SSME dynamics
and its
modeling
is
given
in a
study
by
Duyar
et
al.3
A
summary
of the
major failures
of the
SSME
that
have
occurred
are
outlined
by
Cikanek.4 Several
authors5'8
survey
the
available methods
and
approaches
for
fault detection
and
diagnosis.
In
particular,
the
survey
by
Isermann6 gives several
examples
of the use of
identification techniques
for
process
fault
detection.
A
fault
is the
abnormal behavior
of a
component
due to
physical
change
in the
component.
A
fault event often impairs
or
deteriorates
the
system's
ability
to
perform
its
specified
tasks
or
mission.
The
detection task
is
defined
as the act of
identifying
the
presence
of an
unspecified fault. After
a
fault
is
detected, then
the
fault must
be
isolated
to the
component
that
has
failed. During
the
process
of
isolation,
the
magnitude
of
the
fault
may be
estimated. Fault diagnosis
is the
isolation
and
estimation
of a
fault
mode.
Once
a
fault
is
detected
and
diagnosed,
the
fault
can be
accommodated through reconfigu-
ration
of the
system. Reconfiguration includes
both
hardware
actions (e.g., activating backup systems)
and
software tasks
(e.g., adjusting
the
feedback control gains).
The
detection
and
diagnostic tasks
may be
accomplished
by an
onboard
proces-
sor,
on
line
and in
real time
for
fault accommodation,
as
well
as by an
off-line
processor that analyzes recorded
data
for
life-cycle
analysis
and
preventive maintenance.
Received July
7,
1990; revision received Oct.
26,
1990;
accepted
for
publication Feb.
21,
1991.
Copyright
©
1991
by the
American
Insti-
tute
of
Aeronautics
and
Astronautics, Inc.
No
copyright
is
asserted
in
the
United States under
Title
17,
U.S.
Code.
The
U.S. Government
has a
royalty-free license
to
exercise
all
rights under
the
copyright
claimed herein
for
Governmental purposes.
All
other
rights
are re-
served
by the
copyright owner.
*Professor, Mechanical Engineering Department.
fDeputy
Branch Chief, Advanced
Controls
Technology Branch,
Mail
Stop
77-1,
21000
Brookpark
Road.
Senior Member
AIAA.
384
Initially,
a
brief description
of the
conceptual design
of the
model-based fault detection
and
diagnostic system (FDDS)
is
given.
This
is
followed
by a
description
of the
process model-
ing,
the
residual generation method,
and the
detection
and
diagnostic system design. Finally, results
of the
application
of
the
FDDS
to the
detection
of a
stuck valve fault using simu-
lated
data
are
presented.
Conceptual Design
Model-based
fault
detection methods rely
on the
determina-
tion
of
changes appearing
in the
system
due to the
existence
of
a
fault,
in
comparison with
the
normal status
of the
system.
For
example,
in
aerospace applications control actuator faults
may
be
represented
as
shifts
in the
parameters
of the
control
gain matrix. Sensor faults
may be
represented
as
abrupt
changes
in the
parameters
of the
output matrix
or
increases
in
measurement
noise. These changes
are
determined
by
compar-
ing
the
parameters
of the
observed process with
the
parame-
ters
obtained
from
the
model
of the
normal process.
The
differences
between these parameters
are
called
residuals.
The
residuals
and
their patterns
are
analyzed
for
fault detection
and
diagnosis
by
comparing them with
the
known fault signa-
tures
of the
process.
Fault signatures, which show
the
effect
of a
fault
on the
parameters,
are
generated
by
inducing faults
in the
nonlinear
dynamic
simulation
of the
process. Fault diagnosis
is
accom-
plished
by
training
a
neural network classifier
to
recognize
the
pattern
of the
respective fault signatures.
The
design
of the
FDDS
is
accomplished
in
three stages:
process modeling, residual generation
and
fault
detection,
and
diagnostic
classifier design.
In the
following sections
the
meth-
ods
used
in
these stages
are
briefly
explained.
Process Modeling
A
complete nonlinear dynamic simulation
of
SSME perfor-
mance
was
developed
by
Rocketdyne Division
of
Rockwell
International
Corporation.9
In
this study, this nonlinear
model
is
considered
as the
unknown process.
It is
used
for the
generation
of
fault
signatures
by
modifying
the
actuator mod-
els
to
include
a
fault
model.
The
input-output
data
generated
from
this simulation
are
also used
to
identify
the
parameters
of
the
engine.
Due to its
size
and
complexity
(40 min of CPU
time
for 20 s of
real-time operation with
a VAX
8800), this
nonlinear
simulation cannot
be
used
to
generate data
in
real
time
to
describe
the
normal mode
of
operation.
