Aircraft Stability
Aircraft Stability
Aircraft Stability
Kalmanje Krishnakumar
John Kaneshige
a
(b)
0 0.1 0.2 0.3 0.4 0.5
0.2
0.15
0.1
0.05
0
(a)
Fraction of Left Wing Loss
C
L
,
a
0 0.1 0.2 0.3 0.4 0.5
0
1
2
3
4
5
6
7
8
x 10
3
(d)
Fraction of Left Wing Loss
C
n
,
e
0 0.1 0.2 0.3 0.4 0.5
8
6
4
2
0
x 10
3
(c)
Fraction of Left Wing Loss
C
Y
,
e
Fig. 3 - Control Derivatives Due to Wing Loss at = 12
o
and = 0
o
For an ideal symmetric aircraft, the aileron deection effects the roll control with insignicant contribution to the
lift and pitching moment coefcients. The effect of wing loss causes the aileron deection to induce a change in the lift
coefcient as well as a change in the pitching moment coefcients as seen in Fig. 3(a)-(b). The abrupt changes in the
lift control and pitch control derivatives are due to the complete loss of one of the ailerons for a wing loss that extends
beyond 25% span. The consequence of this is that the damaged aircraft would exhibit a pitch-roll coupling when the
ailerons are deected asymmetrically. To maintain a trim state, the ight control must compensate for the unwanted
pitch motion with the elevators. The situation is similar for the elevator control as the effect of wing loss introduces
a change in the side force coefcient and a change in the yawing moment coefcient as seen in Fig. 3(c)-(d). Thus,
a deection of the elevators would result in a pitch-yaw coupling that must be compensated within the ight control
system by adjusting the rudder control accordingly. Because of the asymmetry, the general motion of a damaged
aircraft is coupled in all the three axes. As a result, any adaptive ight control strategy must be able to effectively
handle this cross-coupled effect.
III. Flight Dynamics of Asymmetric Aircraft
The longitudinal motion of a symmetric aircraft is typically symmetric with respect to the aircraft fuselage ref-
erence line. The lateral motion is uncoupled from the longitudinal motion owing to the aircraft symmetry. For a
damaged aircraft, the symmetry may no longer be preserved depending on the nature of the damage such as wing
damage. The asymmetry of the damaged aircraft thus causes the longitudinal motion and lateral motion to couple
together. Furthermore, the C.G. is shifted away from the x z plane. The motion of an asymmetric damaged aircraft,
therefore, must be understood in order to evaluate any ight control design. To this end, we consider an asymmetric
aircraft with a C.G. offset from some reference location as shown in Fig. 4. The reference location is a xed point
located at the coordinate (x
0
, y
0
, z
0
) on the aircraft which may be taken as the original C.G. of the undamaged aircraft
4 of 24
American Institute of Aeronautics and Astronautics
in order to maintain the same coordinate reference frame. The C.G. of the damaged aircraft can move relative to this
xed reference point.
CG
O
y
x
x
z
z
y
Fig. 4 - C.G. Shift Relative to Reference Point O
The damage effect resulting from a wing loss creates a larger C.G. shift in the pitch axis y than in the other two
axes as shown in Fig. 5. This results in an additional rolling moment that the ight control must be able to compensate
for using the available control surfaces in order to maintain the damaged aircraft in a trim state.
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0.01
0.005
0
0.005
0.01
0.015
0.02
0.025
Fraction of Left Wing Loss
C
G
S
h
i
f
t
a
s
F
r
a
c
t
i
o
n
o
f
H
a
l
f
W
i
n
g
S
p
a
n
x
y
z
Fig. 5 - C.G. Shift due to Wing Loss
A. Linear Acceleration
To understand the effect of the C.G. shift, the standard equations of motion for ight dynamics of a symmetric aircraft
must be modied to allow for the asymmetry. Assuming a at-earth model for a rigid body aircraft, the force vector
in the body-xed reference frame of the aircraft is
F
B
= m
dv
dt
+
d
dt
_
_
rdm
_
W (1)
5 of 24
American Institute of Aeronautics and Astronautics
where W = mg
_
sin cossin coscos
_
T
is the gravitational force vector, =
_
p q r
_
T
is the
aircraft angular rate vector, r = r+r is the position vector of the reference location such that r is the position vector
of the C.G. and r =
_
x y z
_
T
is the displacement vector from the C.G. to the reference location.
The aircraft mass is assumed to undergo a change so that
m = m
+ m (2)
where m
is the original mass of the aircraft and m < 0 is the mass change due to damages.
Assuming that the change in the mass of the aircraft is instantaneous, the force vector then becomes
F
B
= m
dv
dt
+ m
d
dt
r + m
dr
dt
W (3)
where
dr
dt
is the speed of the C.G. relative to the reference location which is assumed to be small relative to v and
therefore may be neglected.
Transforming from the body-xed reference frame to the inertial reference frame yields
F = F
B
+ m v (4)
Expanding Eq. (4) gives
X = m( u + qz ry rv + qw + gsin) (5)
Y = m( v pz + rx + ru pw gcossin) (6)
Z = m( w + py qx qu + pv gcoscos) (7)
The angular acceleration terms appearing in Eqs. (5)-(7) are a result of the C.G. shift. Thus, the linear acceleration
of an asymmetric aircraft is coupled with its angular acceleration.
B. Angular Acceleration
We consider the angular momentum vector in the body-xed reference frame
H
B
=
_
[r ( r)] dm +
_
(r v) dm (8)
Expanding this expression yields
H
B
= I + mr v (9)
where I is the mass moment of inertia matrix with respect to the reference frame at the reference location.
The time rate of change in the angular momentumgives rise to the moment equation in the inertial reference frame
M=
dH
B
dt
+ H
B
= I
d
dt
+ mr
dv
dt
+ I + m (r v) (10)
Expanding Eq. (10) results in the following moment equations
L = I
xx
p I
xy
q I
xz
r + I
xy
pr I
xz
pq + (I
zz
I
yy
) qr + I
yz
_
r
2
q
2
_
+ m(qv + rw) x + m( w qu) y m( v + ru) z (11)
M = I
xy
p + I
yy
q I
yz
r + I
yz
pq I
xy
qr + (I
xx
I
zz
) pr + I
xz
_
p
2
r
2
_
m( w + pv) x + m(pu + rw) y + m( u rv) z (12)
N = I
xz
p I
yz
q + I
zz
r + I
xz
qr I
yz
pr + (I
yy
I
xx
) pq + I
xy
_
q
2
p
2
_
+ m( v pw) x m( u + qw) y + m(pu + qv) z (13)
Equations (11)-(13) indicate that the C.G. offset effectively creates additional moments on the aircraft. Cross
coupling in both the linear and angular accelerations are present. Thus, the longitudinal and lateral motions of the
aircraft are generally coupled and the aileron or elevator commanded input therefore will affect the aircraft motion in
both stability axes.
6 of 24
American Institute of Aeronautics and Astronautics
C. Aerodynamic and Propulsive Forces and Moments
Assuming that the engine thrust vector is aligned with the x-axis of the aircraft, then the forces and moments due to
aerodynamics and the propulsion are
X =
T
T
max
+ (C
L
+ C
L
) QS sin (C
D
+ C
D
) QS cos cos (14)
Y = (C
Y
+ C
Y
) QS (C
D
+ C
D
) QS sin (15)
Z = (C
L
+ C
L
) QS cos (C
D
+ C
D
) QS sin cos (16)
L = (C
l
+ C
l
) QS c (17)
M = (C
m
+ C
m
) QS c +
T
T
max
(z
e
z
0
) (18)
N = (C
n
+ C
n
) QS c +
T
T
max
y
e
+
T
T
max
y
0
(19)
where (x
e
, y
e
, z
e
) are the centers of thrust and the subscript * denotes the force and moment coefcients for the
undamaged aircraft evaluated at the reference location.
We assume that the left and right engines produce the same amount of thrust with a combined maximum thrust
equal to T
max
and are symmetrically positioned with respect to the aircraft fuselage reference line. Then
T
where
0
T
1 is the throttle position corresponding to a desired total engine thrust, and
T
where
1
2
T
1
2
is the throttle differential position difference that results in a desired engine differential thrust equal to the left engine
thrust minus the right engine thrust. The incremental changes in these coefcients due to damages are dened as
C = C
0
+ C
+ C
+ C
(20)
where C =
_
C
L
C
L
C
D
C
D
C
Y
C
Y
C
l
C
l
C
m
C
m
C
n
C
n
_
T
, =
_
a
e
r
_
T
is the ight control surface deection vector, the subscripts , , and denote the derivatives, and the subscript 0
denotes the coefcients at = 0 and = 0.
IV. Trim Analysis
The aerodynamic forces on asymmetric aircraft include a non-zero side force component that is generally not
experienced on symmetric aircraft. For a steady ight, the side force equation becomes
mgcossin + (C
Y
+ C
Y
) QS (C
D
+ C
D
) QS sin = 0 (21)
The side force trim for the asymmetric aircraft can therefore be accomplished by trimming the aircraft at a non-zero
bank angle with zero sideslip angle . However, this would result in a limitation in the bank angle in coordinated
turn maneuvers. Another side force trim approach is to trim the aircraft level with zero bank angle but at a non-zero
sideslip angle . In either case, the aircraft would have to be trimmed in both the longitudinal and lateral directions
simultaneously by searching for the steady state solution of Eqs. (14) to (16) with = 0 or = 0. The trim analysis
thus computes the trim values for the angle of attack , bank angle or sideslip angle , and engine throttle position
T
as functions of the aileron deection
a
, elevator deection
e
, and rudder deection
r
for a given aircraft Mach
number and altitude. We assume that the engine thrusts will be symmetric at all times so that
T
= 0.
If an undamaged symmetric aircraft has a mass m
, sideslip angle
T
corresponding to a
lift coefcient C
L
, drag coefcient C
D
, and side force coefcient C
Y
= 0. Then for small changes in the aircraft mass
and aerodynamic coefcients, we can determine the incremental trim angle of attack, bank angle, and throttle position
to maintain approximately the same trim airspeed V and ight path angle
T
T
max
+(C
L
+ C
L,
+ C
L,
+ C
L,
) QS sin
+(C
L
+ C
L,
+ C
L,
+ C
L,
) QS cos
(C
D
+ C
D,
+ C
D,
+ C
D,
) QS cos
+ (C
D
+ C
D,
+ C
D,
+ C
D,
) QS sin
mg cos (
) mg sin (
) = 0 (22)
7 of 24
American Institute of Aeronautics and Astronautics
(C
Y
+ C
Y,
+ C
Y,
+ C
Y,
) QS (C
D
+ C
D,
+ C
D,
+ C
Y,
) QS
+ mg [cos (
) sin (
) ] = 0 (23)
(C
L
+ C
L,
+ C
L,
+ C
L,
) QS cos
+ (C
L
+ C
L,
+ C
L,
+ C
L,
) QS sin
(C
D
+ C
D,
+ C
D,
+ C
D,
) QS sin
(C
D
+ C
D,
+ C
D,
+ C
D,
) QS cos
mg sin (
) + mgcos (
) = 0 (24)
To nd the trim bank angle at zero sideslip angle, we set = 0 in the equations above. Equation (24) then is a
quadratic equation in terms of whose solution can easily be computed as
=
b
2a
+
_
b
2a
_
2
c
a
(25)
with
a = (C
L,
sin
C
D,
cos
) QS (26)
b = [(C
L
+ C
L,
C
D,
) sin
(C
D
+ C
D,
+ C
L,
) cos
] QS mg sin (
) (27)
c = [(C
L
+ C
L,
) cos
+ (C
D
+ C
D,
) sin
] QS + mgcos (
) (28)
From Eq. (23), we now nd the trim bank angle
=
(C
Y
+ C
Y,
+ C
Y,
) QS
mg [cos (
) sin (
) ]
(29)
Finally, the incremental trim throttle position can be solved directly from Eq. (22).
Trimming the damaged aircraft with bank angle will result in a reduced bank angle limitation. This would poten-
tially affect the aircrafts turn capability. Moreover, the aircraft will not y wing-level which would not be acceptable
for a landing approach. Therefore, the damaged aircraft can be trimmed alternatively with the sideslip angle. This will
enable the aircraft to maintain a level ight but the control authority of the rudder control surface will be reduced since
it has to compensate for the non-zero sideslip angle. To obtain the trim sideslip angle, we set = 0 in the Eq. (23)
and solve Eqs. (22) to (24) simultaneously.
In examining Eq. (23) with = 0, it is noted that if the undamaged aircraft is in a cruise phase at a minimum
drag, then the trim sideslip angle for the damaged aircraft could be large if the damage develops a signicant side
force. Typically, it is not advisable to y the aircraft at a high sideslip angle because of the stability issue. Depending
on the extent of damages, an effective trim approach may be one that uses a combination of the trim bank angle and
sideslip angle.
In cases where the rudder control power is insufcient due to damages, then the engine differential thrust throttle
position
T
could be used to provide an additional control effector to trim the aircraft in yaw. Using engine differen-
tial thrust for yaw control requires examining the issue associated with a slow engine response relative to the responses
of typical ight control surfaces. While in theory the engine thrust can be used to trim the aircraft in yaw, often by
the time the engine thrust is adjusted differentially to the correct trim value, the aircraft may have reached a different
dynamic state due to the loss in airspeed and or altitude if the damage condition is severe enough to cause the aircraft
performance to rapidly deteriorate in its ight envelope. Because of the time scale difference between traditional ight
control surfaces and engines, engine actuator dynamics must be accounted for in the overall ight control strategy.
In addition to using the engine differential thrust as a control effector, other ight control surfaces can be used in
an overall control redundancy design strategy. This investigation would examine the control effectiveness of a various
combinations of ight control surfaces. For example, wing spoilers can be used for roll control and the wing ap
extension or deection can be used for pitch control. Some of these control surfaces may have different time latency
characteristics such as wing aps versus spoilers. In the control and stability analysis, actuator dynamic model of slow
systems should be included in the overall ight dynamic modeling.
The trim analysis shows that upon damage, the damaged aircraft would have to be retrimmed with a new trim
angle of attack =
+ , new bank angle = or new sideslip angle = , and new throttle position
T
=
T
+
T
. The trim , or , and
T
are all functions of the ight control surface deection as well as the
aircraft damage conguration. In general, the stability and control derivatives needed to retrim the damaged aircraft
8 of 24
American Institute of Aeronautics and Astronautics
are not known. Thus, it is necessary that these parameters be identied in ight by a parameter identication process.
Assuming that the effect of damage on the aerodynamics can be estimated, then a trim strategy is to initially retrim the
damaged aircraft with zero control surface deection by setting = 0 in Eqs. (22) to (24). Then, using the inner-loop
rate-command-attitude-hold (RCAH) control, the control surface deection for the damaged aircraft can be obtained.
This allows the trim values to be rened. Depending on the nature of damage, the trim renement may be repeated
until the damaged aircraft becomes completely trimmed.
V. Damage Adaptive Flight Control
Most conventional ight control systems utilize extensive gain-scheduling in order to achieve desired handling
qualities. While this approach has proved to be very successful, the development process can be expensive and often
results in aircraft specic implementations. Over the past several years, various adaptive control techniques have been
investigated.
7
Damaged aircraft presents a challenge to the conventional ight control systems because the aircraft
dynamics may deviate fromits known dynamics substantially due to a signicant degradation in the ight performance
of the damaged aircraft. This makes it difcult for the conventional ight control systems to cope with changes in
the stability and control of the damaged aircraft. Adaptive ight control provides a possibility for maintaining the
stability of a damage aircraft by means of being able to quickly adapt to uncertain system dynamics. Research in
adaptive control has spanned several decades, but challenges in obtaining robustness in the presence of unmodeled
dynamics, parameter uncertainties, or disturbances as well as the issues with certication, verication and validation
of adaptive ight control software prevent it from being implemented in ight control systems.
8
Adaptive control laws
may be divided into direct and indirect approaches. Indirect adaptive control methods provide the ability to compute
control parameters from on-line neural networks that estimate plant parameters.
9
Parameter identication techniques
such as recursive least squares and neural networks have been used in indirect adaptive control methods.
10
In recent
years, model-reference direct adaptive control using neural networks has been a topic of great research interests.
1113
Lyapunov stability theory has been used to establish robustness of neural network adaptive control to ensure that
adaption laws for neural network weight updates are asymptotically stable.
In the current research, we adopt the work by Rysdyk and Calise
11
to develop a neural network adaptive con-
trol with dynamic inversion for damaged aircraft. The adaptive ight control is able to provide consistent handling
qualities without requiring extensive gain-scheduling or explicit system identication for a damaged aircraft. This
particular architecture uses both pre-trained and on-line learning neural networks, and reference models to specify
desired handling qualities. Pre-trained neural networks are used to provide estimates of aerodynamic stability and
control characteristics required for model inversion. On-line learning neural networks are used to compensate for
errors and adapt to changes in aircraft dynamics. As a result, consistent handling qualities may be achieved across
ight conditions and for different damage congurations. An architecture of the neural network adaptive ight control
is shown in Fig. 6. Furthermore, we will extend this architecture to include an indirect adaptive control element that
provides an on-line estimation of the true plant dynamics. The estimation approach is provided by an adaptive law
based on the Lyapunov stability analysis. In addition, we also consider a recursive least square method for the on-line
estimation.
Fig. 6 - Direct Neural Network Adaptive Flight Control Architecture
9 of 24
American Institute of Aeronautics and Astronautics
A. Linearized Plant Dynamics
First, we need to arrive at a linear dynamics of the damaged aircraft for the feedback linearization control. To maintain
airspeed and altitude, the damaged aircraft has to be retrimmed using the trim method above. The damaged aircraft
stability must be recovered by the RCAH controller. This results in control surface deections necessary to maintain a
desired angular rate command. To design a linear RCAH controller, we want to eliminate the linear acceleration terms
in Eqs. (11) to (13) corresponding to the uncompensated damaged aircraft dynamics of linear motion resulting from
damages. Combining Eqs. (5) to (7) with Eqs. (11) to (13) yields
I
xx
p
I
xy
q
I
xz
r + I
xy
pr I
xz
pq + (I
zz
I
yy
) qr + I
yz
_
r
2
q
2
_
+ m(qv + rw) x mpvy mpwz =
_
C
l
+
C
l
_
QS c (30)
I
xy
p +
I
yy
q
I
yz
r + I
yz
pq I
xy
qr + (I
xx
I
zz
) pr + I
xz
_
p
2
r
2
_
mqux + m(pu + rw) y mqwz =
_
C
m
+
C
m
_
QS c +
T
T
max
(z
e
z
0
) (31)
I
xz
p
I
yz
q +
I
zz
r + I
xz
qr I
yz
pr + (I
yy
I
xx
) pq + I
xy
_
q
2
p
2
_
mrux mrvy + m(pu + qv) z =
_
C
n
+
C
n
_
QS c +
T
T
max
y
0
(32)
where
C
l
= C
l
+ C
y
z
c
C
z
y
c
mg
QS
_
cos cos
y
c
cos sin
z
c
_
(33)
C
m
= C
m
C
x
z
c
+ C
z
x
c
+
mg
QS
_
cos cos
x
c
+ sin
z
c
_
(34)
C
n
= C
n
+ C
x
y
c
C
y
x
c
mg
QS
_
cos sin
x
c
+ sin
y
c
_
(35)
where C
x
, C
y
, and C
z
are X, Y , and Z force coefcients normalized to the dynamic pressure force QS.
The linear dynamics of the damaged aircraft is computed by linearizing Eqs. (30) to (32)
_
I
_
d
dt
= (f
1
+ f
1
) + (f
2
+ f
2
) + (g
+ g) (36)
where =
_
p q r
_
T
is the angular rate vector, =
_
T
_
T
is the trim parameter
vector, and
f
1
= QS c
_
_
C
l,p
C
l,q
C
l,r
C
m,p
C
m,q
C
m,r
C
n,p
C
n,q
C
n,r
_
_
f
1
= QS c
_
C
l,p
C
l,q
C
l,r
C
m,p
C
m,q
C
m,r
C
n,p
C
n,q
C
n,r
_
_ + m
_
_
vy + wz vx wx
uy ux + wz wy
uz vz ux + vy
_
_
f
2
= QS c
_
_
C
l,
C
l,
0 0
C
m,
+
T Tmax,
QS
zez0
c
C
m,
+
T T
max,
QS
zez0
c
0
Tmax
QS
zez0
c
C
n,
+
T Tmax,
QS
y0
c
C
n,
+
T T
max,
QS
y0
c
0
Tmax
QS
y0
c
_
_
f
2
= QS c
_
C
l,
C
l,
C
l,
C
l,T
C
m,
C
m,
C
l,
C
l,T
C
n,
C
m,
C
l,
C
l,T
_
_
10 of 24
American Institute of Aeronautics and Astronautics
g
= QS c
_
_
C
l,a
C
l,e
C
l,r
C
m,a
C
m,e
C
m,r
C
n,a
C
n,e
C
n,r
_
_
g = QS c
_
C
l,a
C
l,e
C
l,r
C
m,a
C
m,e
C
m,r
C
n,a
C
n,e
C
n,r
_
_
Equation (36) is the angular acceleration equation of the asymmetric aircraft which can be written in a state-space
form as
= (F
1
+ F
1
) + (F
2
+ F
2
) + (G+ G) (37)
where F
1
=
I
1
f
1
, F
2
=
I
1
f
2
, G =
I
1
g
, F
1
=
I
1
(f
1
+ f
1
) F
1
, F
2
=
I
1
(f
2
+ f
2
) F
2
, and
G =
I
1
(g
+ g) G.
Under ideal situations, the plant dynamics of an undamaged aircraft is assumed to be known. However, for a
damaged aircraft, the plant dynamics become uncertain as the stability and control derivative matrices F
1
, F
2
, and
G are usually unknown. Consequently, the ight control needs to be able to adapt to the uncertain plant dynamics
of the damaged aircraft. The angular acceleration vector
of the damaged aircraft may be written as the sum of an
ideal angular acceleration vector
i
of the undamaged aircraft and a differential angular acceleration vector due
to damage as
=
i
+ (38)
The ideal, undamaged aircraft plant dynamics can be written as
i
= F
1
+F
2
+G (39)
where the stability and control matrices for the undamaged aircraft F
1
, F
2
, and Gare assumed to be known.
B. Direct Neural Network Adaptive Control
The goal of the adaptive ight control is to be able to y the damaged aircraft whose handling characteristics is
specied by a reference model. The control adaptation must be able to accommodate damages using the available
ight control surfaces. A reference model is used to lter a rate command vector
c
into a reference angular rate
vector
m
and a reference angular acceleration vector
m
via a rst-order model
m
+
n
m
=
n
c
(40)
where
n
= diag (
p
,
q
,
r
) is the frequency matrix.
The reference frequency parameters must be chosen appropriately in order to obtain a good transient response that
satises position and rate limits on the control surface deection. For transport aircraft, typical values of the reference
model frequencies
p
,
q
, and
r
are 3.5, 2.5, and 2.0, respectively.
3
In cases when the reference model is over- or
under-specied, the parameters of the reference model must be adjusted. The tuning of the reference model parameters
can be performed using an adaptive-critic approach to ensure that the ight control can track the reference model in
order to achieve desired handling qualities.
16
The reference model angular rate vector
m
are compared with the actual angular rate output to form a tracking
error signal
e
=
m
. A pseudo-feed back control vector u
e
is constructed using a proportional-integral (PI)
feedback scheme to better handle errors detected from the roll rate, pitch rate, and yaw rate feedback. The error
dynamics, dened by proportional and integral gains, must be fast enough to track the reference model, yet slow
enough to not interfere with actuator dynamics. The issue with the integrator windup during a control saturation is
addressed by a windup protection which limits the integrator at its current value when a control surface is commanded
beyond its limit. The pseudo-control vector u
e
is computed as
u
e
= K
P
e
+K
I
_
t
0
e
d (41)
11 of 24
American Institute of Aeronautics and Astronautics
In order to ensure low-gain error handling performance, the error dynamics is designed with natural frequencies
that match the reference model frequencies in the roll, pitch, and yaw axes. A damping ratio is chosen with
p
=
q
=
r
= 1/
2. These frequencies and damping ratio are incorporated into the proportional and integral gains as
K
P
= diag (2
p
p
, 2
q
q
, 2
r
r
) (42)
K
I
= diag
_
2
p
,
2
q
,
2
r
_
(43)
Adynamic inversion is performed to obtain an estimated control surface deection command
to achieve a desired
angular acceleration vector
d
using the known plant dynamics of the undamaged aircraft from Eq. (39) as
= G
1
(
d
F
1
F
2
) (44)
assuming that Bis invertible.
In order for the dynamic inversion control to track the reference model angular acceleration rate vector
m
, the
desired angular acceleration vector
d
is set to be equal to
d
=
m
+u
e
u
ad
(45)
where u
ad
is an adaptive control augmentation designed to cancel out the dynamic inversion error, so that in an ideal
setting, the desired angular acceleration rate
d
is equal to the reference model angular acceleration rate
m
as the
tracking error goes to zero asymptotically.
Because the true plant dynamics of the damaged aircraft is unknown and is different from the undamaged aircraft
plant dynamics as can be seen from Eq. (37), a dynamic inversion will result from the control surface deection
.
This error is equal to
=
d
=
F
1
F
2
G
(46)
Comparing with Eq. (37), we see that the dynamic inversion error can also be expressed in terms of the unknown
plant dynamics due to the damage effects
= = F
1
+ F
2
+ G
(47)
Substituting Eq. (45) into Eq. (46) results in
=
e
u
e
+u
ad
(48)
Combining Eq. (41) with Eq. (48) yields
e = Ae +B(u
ad
) (49)
where e =
_
_
t
0
e
d
e
_
and
A =
_
0 I
K
I
K
P
_
, B =
_
0
I
_
The adaptive control augmentation vector u
ad
is based on a neural network adaptation law by Rysdyk and Calise
11
that guarantees boundedness of the tracking error and of the network weights using a single-hidden-layer sigma-pi
neural network
u
ad
= W
T
(C
1
, C
2
, C
3
) (50)
where is a vector of basis functions computed using a nested Kronecker product with C
1
, C
2
, C
3
as inputs into the
neural network consisting of control commands, sensor feedback, and bias terms.
The network weights Ware computed by an adaption law, which incorporates an adaptation gain > 0 and an
e-modication term > 0
14
according to the update law
W=
_
e
T
PB+
_
_
e
T
PB
_
_
W
_
(51)
where the matrix P solves the Lyapunov equation A
T
P + P
T
A = Q for some positive-denite matrix Q and the
norm is a Frobenius norm.
12 of 24
American Institute of Aeronautics and Astronautics
The e-modication term provides a robustness in the adaptation law.
14
The update law in Eq. (23) guarantees the
stability of the network weights and the tracking error. The proof of this update law using the Lyapunov method is
provided by Rysdyk and Calise.
11
Solving for the matrix P with Q = I, the update law can be rewritten as
W= (V+ V W) (52)
where
V =
1
2
T
e
K
1
p
_
I +K
1
I
_
+
1
2
_
t
0
T
e
dK
1
I
(53)
VI. Hybrid Direct-Indirect Adaptive Control Concept
While the direct neural network adaptive law has been extensively research and has been used with good successes
in a number of applications, the possibility of high gain control due to aggressive learning can be an issue. Aggressive
learning is characterized by setting the learning rate high enough so as to reduce the dynamic inversion error rapidly.
This can potentially lead to a control augmentation command that may saturate the control authority. Moreover, high
gain control may also excite unmodeled dynamics of the plant that can adversely affect the stability of the adaptive
law. To address this issue, we are considering a modication to the present direct adaptive law to include an indirect
adaptive law that provides an opportunity to perform an on-line estimation of the plant dynamics of the damaged
aircraft explicitly. We call this approach as a hybrid direct-indirect adaptive control concept. The indirect adaptive
law will provide an estimated plant dynamics that will be used in the dynamic inversion. If successful, the control
command will result in a smaller dynamic inversion error so that the learning of the direct adaptation neural network
can be reduced. An architecture of the proposed hybrid adaptive control concept is shown in Fig. 7. In the current
study, we are developing some initial indirect adaptive laws for the on-line estimation of plant dynamics based on the
Lyapunov stability theory and also the well-known recursive least-square method. Future research still remains ahead
to rigorously investigate this proposed concept followed by high-delity simulations.
Fig. 7 - Hybrid Direct-Indirect Neural Network Adaptive Flight Control Architecture
A. Indirect Neural Network Adaptive Control
We would like to estimate the unknown plant matrices using a linear-in-parameter neural network approach as
F
1
= W
T
(54)
F
2
= W
T
(55)
G = W
T
(56)
where the hat symbol denotes the estimated plant matrices and
BW
T
BW
T
B (57)
13 of 24
American Institute of Aeronautics and Astronautics
where < is the residual error from the estimation of the plant matrices.
We now propose the following adaptive laws for the estimation of W
, W
, and W
e
T
PB (58)
e
T
PB (59)
e
T
PB (60)
where
. We then let W = W
+
W, W
= W
+
W
, W
= W
+
W
, and
W
= W
+
W
where the asterisk symbol denotes the ideal weight matrices to cancel out the residual error
and the tilde symbol denotes the weight deviations.
The ideal weight matrices are unknown but they may be assumed constant and bounded to stay within a -
neighborhood of the residual error so that
= sup
,,
_
_
_W
T
W
T
W
T
W
T
_
_
_ (61)
We dene the following Lyapunov function
V = e
T
Pe + tr
_
W
T
W
W
T
W
T
W
T
_
(62)
where P 0 and tr (A) denotes the trace of a matrix A.
The time derivative of the Lyapunov function is computed as
V = e
T
Pe +e
T
P e + 2tr
_
W
T
W
T
W
T
W
T
_
(63)
Substituting Eqs. (57) and (51) into the above equation yields
V = e
T
Qe + 2e
T
PB
_
W
T
+
W
T
_
2e
T
PB
_
W
T
+
W
T
2e
T
PB
_
W
T
+
W
T
2e
T
PB
_
W
T
+
W
T
2e
T
PB
+ 2tr
_
W
T
e
T
PB
W
T
_
_
e
T
PB
_
_
_
W
+
W
_
+
W
T
W
T
W
T
_
(64)
We note that tr (AB) = tr (BA) , so that
e
T
PB
W
T
= tr
_
e
T
PB
W
T
_
= tr
_
W
T
e
T
PB
_
(65)
e
T
PB
W
T
= tr
_
e
T
PB
W
T
_
= tr
_
W
T
e
T
PB
_
(66)
e
T
PB
W
T
= tr
_
e
T
PB
W
T
_
= tr
_
W
T
e
T
PB
_
(67)
e
T
PB
W
T
= tr
_
e
T
PB
W
T
_
= tr
_
W
T
e
T
PB
_
(68)
Also, by completing the square, we have
2tr
_
W
T
_
_
e
T
PB
_
_
_
W
+
W
__
= 2
_
_
e
T
PB
_
_
_
_
_
_
_
W
2
+
W
_
_
_
_
2
_
_
_
_
W
2
_
_
_
_
2
_
(69)
14 of 24
American Institute of Aeronautics and Astronautics
Since
=
W
=
W
, and
=
W
V e
T
Qe + 2e
T
PB2
_
_
e
T
PB
_
_
_
_
_
_
_
W
2
+
W
_
_
_
_
2
_
_
_
_
W
2
_
_
_
_
2
_
2tr
_
W
T
_
W
e
T
PB
_
+
W
T
_
W
e
T
PB
_
+
W
T
_
W
e
T
PB
__
(70)
Since B = 1, we establish that
e
T
Qe (Q) e
2
(71)
e
T
PB (P) e (72)
_
_
e
T
PB
_
_
_
_
_
_
W
2
_
_
_
_
2
(P) e
_
_
_
_
W
2
_
_
_
_
2
(73)
where (Q) and (P) are the spectral radii of Qand P.
In order to guarantee that
V 0, we require that the trace operator be equal to zero, thus resulting in the adaptive
laws in Eqs. (58) to (60). In addition, we also require that
e >
(P)
2 (Q)
_
4 + W
2
_
(74)
The time rate of change of the Lyapunov function is then strictly negative and therefore it would guarantee that the
signals are bounded. We note that e, , ,
but e L
2
since
_
0
e
T
Qedt (Q)
_
0
e
2
dt V (0) V (t ) + 2 (P)
_
0
e dt
2 (P)
_
0
e
_
_
_
_
_
W
2
+
W
_
_
_
_
2
_
_
_
_
W
2
_
_
_
_
2
_
dt < (75)
Utilizing Eq. (74), we have
V (t ) V (0) 2 (P)
_
0
e
_
_
_
_
W
2
+
W
_
_
_
_
2
dt < (76)
Thus, the value of V as t is bounded. Therefore, we establish that
_
_
_
W
_
_
_ 0,
_
_
_
_
_
_ 0,
_
_
_
_
_
_ 0,
and
_
_
_
_
_
_ 0 imply e 0 as t . This means that the adaptive laws will result in a convergence of the
estimated F
1
, F
2
, and Gto their steady state values. In practice, the inputs =
_
T
T
_
T
must be
sufciently rich that contain enough frequencies to capture all the plant dynamics. In order for the on-line estimation
to converge the their correct values, the inputs need to be a persistent excitation (PE) class of signals such that if there
exist
0
,
1
, T
0
> 0 then
9
0
I
_
t+T0
t
()
T
() d
1
I (77)
We can also robustify the adaptive laws similar to Eq. (52) to better handle unmodeled dynamics and distur-
bances by adding an e-modication term
14
to Eqs. (58) to (60) as
e
T
PB
_
_
e
T
PB
_
_
W
_
(78)
e
T
PB
_
_
e
T
PB
_
_
W
_
(79)
e
T
PB
_
_
e
T
PB
_
_
W
_
(80)
15 of 24
American Institute of Aeronautics and Astronautics
in which case the time rate of change of the Lyapunov function becomes
V 2
_
_
e
T
PB
_
_
_
_
_
_
W
2
+
W
_
_
_
_
2
2
_
_
e
T
PB
_
_
_
_
_
_
_
2
+
W
_
_
_
_
_
2
2
_
_
e
T
PB
_
_
_
_
_
_
_
2
+
W
_
_
_
_
_
2
2
_
_
e
T
PB
_
_
_
_
_
_
_
2
+
W
_
_
_
_
_
2
(81)
The effect of the e-modication is to increase the negative time rate of change of the Lyapunov function so that
as long as the effects of unmodeled dynamics and or disturbances do not exceed the value of
V , the adaptive signals
should remain bounded. The e-modication thus makes the adaptive law robust to unmodeled dynamics so that the PE
condition may not be needed.
15
It should be noted that we have so far assumed that , ,
. Suppose that , ,
/ L
since
lim
1 +
T
= 0 (82)
Therefore, the normalized adaptive laws for unbounded signals should be
_
1 +
T
_
1
_
e
T
PB
_
_
e
T
PB
_
_
W
_
(83)
_
1 +
T
_
1
_
e
T
PB
_
_
e
T
PB
_
_
W
_
(84)
_
1 +
_
1
_
e
T
PB
_
_
e
T
PB
_
_
W
_
(85)
B. Recursive Least-Square Parameter Identication
While the indirect adaptive laws above provide a computational method for on-line estimation of the plant dynamics,
it would be incomplete to not consider the well-known least-square method which is equally robust in parameter
identication process. If the dynamic inversion error is somehow can be estimated, then we should be able to apply a
recursive least-square method to determine the weight matrices W
, W
, and W
W
T
W
T
_
, =
_
_
T
, and is the computational error in the
estimated dynamic inversion error , which may contain noise resulting from the on-line derivative computation of
since
=
F
1
F
2
G
(87)
where
is the estimated angular acceleration which may be subject to computational errors.
One method of computing
is to use a backward nite-difference method
i
=
i
i1
t
(88)
to estimate
at the i-th time step, but this method can result in a signicant error if t is either too small or too large.
Another approach is to collect n number of data points which will be used to generate an at least C
1
smooth curve
in time using a cubic or B-spline method. This curve is then differentiated at their knots to nd the estimated derivative
values. In either case, the derivative computation will introduce an error source . If the error is unbiased, i.e., it can
be characterized as a white noise about the mean value, then the least-square method can be applied to estimate the
plant dynamics.
16 of 24
American Institute of Aeronautics and Astronautics
We consider a minimization of the following cost function
J (W
, W
, W
) =
_
t
0
_
_
_
T
_
_
_
2
d (89)
Our objective is to nd recursive least-square adaptive laws for W
, W
, and W
T
=
_
t
0
_
T
_
T
d = 0 (90)
The recursive least-square formula using the gradient method is
= R
_
T
_
(91)
where
R = R
T
R (92)
To show this, we see that from Eq. (90)
_
t
0
T
d =
_
t
0
T
d (93)
Let
R
1
=
_
t
0
T
d > 0 (94)
Then, differentiating Eqs. (93) and (94) results in
R
1
+
R
1
=
T
(95)
R
1
=
T
(96)
Also, we have
RR
1
= I
RR
1
+R
R
1
= 0 (97)
Substituting Eq. (96) into Eqs. (95) and (96) and solving for
and
R yield the recursive least-square adaptive
law. The matrix Ris called the covariance matrix and the recursive least-square formula has a very similar form to the
Kalman lter where Eq. (92) is a differential Riccati equation for a zero-order plant dynamics. We will show that the
recursive weight update law is stable and results in bounded signals as follows:
We let =
+
with the hat and tilde symbols denoting ideal weights and weight deviations, respectively.
Then, the error dynamics can be written as
e Ae +B
W
T
B
T
+B (98)
We choose the following Lyapunov function
L = V + tr
_
T
R
1
_
(99)
where V is the Lyapunov function for the direct neural network adaptive control and we have established that
V 0.
The time rate of change of the Lyapunov function is computed as
L =
V + tr
_
2
T
R
1
R
1
_
(100)
The weights can be shown to converge to the ideal weights
9
so that
= R
T
(101)
17 of 24
American Institute of Aeronautics and Astronautics
Substituting Eq. (101) into Eq. (100) results in
L =
V
0 (102)
Thus, the recursive least square weight update law is stable.
In practice, the recursive least-square method can be used to estimate the plant dynamics either continuously or
discretely at every n data samples. Continuous time estimation requires solving the differential equations (91) and
(92) at each time step. On the other hands, the discrete-time sampling estimation provides more exibility in that the
estimation can be executed after a specied number of data points have been collected. This would ensure that the
signals contained in the sampled data are sufciently rich to enable an accurate convergence. Another advantage of the
recursive least-square method is that it provides an optimal noise ltering to minimize noise effects in the estimation
of the plant dynamics. The discrete-time recursive least square formula is
k
=
k1
+R
k1
k
_
T
k
T
k
k1
_
(103)
R
k
= R
k1
1
R
k1
k
_
I +
1
T
k
R
k1
k
_
1
T
k
R
k1
(104)
where k denotes the update cycle that repeats every n data samples,
T
k
=
_
W
T
,k
W
T
,k
W
,k
T
_
is the weight
matrix at k-th update cycle, is a forgetting factor that can be used to discount past data, and
T
k
=
_
T
knn+1
T
knn+1
F
T
1
T
knn+1
F
T
2
T
knn+1
G
T
T
knn+2
T
knn+2
F
T
1
T
knn+2
F
T
2
T
knn+2
G
T
.
.
.
T
kn
T
kn
F
T
1
T
kn
F
T
2
T
kn
G
T
_
_
(105)
T
k
=
_
_
T
knn+1
T
knn+1
T
knn+1
T
knn+2
T
knn+2
T
knn+2
.
.
.
.
.
.
.
.
.
T
kn
T
kn
T
kn
_
(106)
VII. Control Simulations
The neural network adaptive ight control for damaged aircraft is evaluated in a medium-delity simulation test
environment. The ight simulator is a xed-motion simulator equipped with a pilot station, programmable displays,
and a 120
o
eld-of-view visual system as shown in Fig. 8. Pilot command inputs are received through a control stick,
a rudder pedal, and a throttle quadrant. Flight control software includes a ight dynamics model of damaged aircraft
as developed herein. Simulations are performed at a 30 Hz frequency.
The damaged GTM is evaluated with various wing loss congurations. Fig. 9 shows the angular rates of the
damaged GTM with and without the neural network adaptive ight control. The neural network control augmentation
can be seen to quickly adapt to the changing dynamics of the damaged GTM. The roll, pitch, and yaw rates are quickly
brought to zero to stabilize the damaged aircraft. In contrast, without a neural network control augmentation, the
aircraft rates are changing rapidly, particularly in the roll axis. A rapid increase in the pitch attitude can result in the
damaged GTM reaching its stall angle of attack that would render the aircraft in a dangerous situation.
Fig. 10 shows the control surface deections corresponding to the neural network control augmentation. The right
aileron is commanded to move substantially to correct for a left turning rolling moment resulting from a left wing
damage. A maximum aileron limit of 35
o
is nearly reached. Thus, it is possible that for certain damage scenarios, the
control augmentation will not be able to stabilize the damaged aircraft due to the control power limitation. In such
situations, other types of control surfaces must be considered to provide additional control authorities for stabilization.
Optimal control allocation approach must be incorporated into the neural network adaptive ight control to maximize
the control effectiveness of all the available control authorities.
18 of 24
American Institute of Aeronautics and Astronautics
Fig. 8 - Flight Simulation Test Environment
0 1 2 3 4 5 6 7 8 9 10
1
0
1
t, sec
p
,
r
a
d
/
s
e
c
(a)
0 1 2 3 4 5 6 7 8 9 10
1
0
1
t, sec
q
,
r
a
d
/
s
e
c
(b)
0 1 2 3 4 5 6 7 8 9 10
1
0
1
t, sec
r
,
r
a
d
/
s
e
c
(c)
with NN control without NN control
with NN control without NN control
with NN control without NN control
Fig. 9 - Rate Control with and without Neural Network Adaptation
19 of 24
American Institute of Aeronautics and Astronautics
0 1 2 3 4 5 6 7 8 9 10
20
30
40
t, sec
a
,
d
e
g
(a)
0 1 2 3 4 5 6 7 8 9 10
2
0
2
(b)
t, sec
e
,
d
e
g
0 1 2 3 4 5 6 7 8 9 10
0
0.5
1
t, sec
r
,
d
e
g
(c)
Fig. 10 - Control Surface Deection
To evaluate the hybrid adaptive ight control with the indirect adaptive law and the recursive least square method,
a simulation was performed in MATLAB environment. A damage conguration corresponding to a 30% loss of the
left wing is selected. A step input pitch doublet is simulated. The tracking performance of the three control laws is
compared in Fig. 11.
0 5 10 15 20
0.05
0
0.05
t, sec
q
,
r
a
d
/
s
e
c
(a)
0 5 10 15 20
0.05
0
0.05
t, sec
q
,
r
a
d
/
s
e
c
(b)
0 5 10 15 20
0.05
0
0.05
t, sec
q
,
r
a
d
/
s
e
c
(c)
Ref. Model Direct NN
Ref. Model RLS + Direct NN
Ref. Model Hybrid NN
Fig. 11 - Pitch Doublet Tracking Performance
20 of 24
American Institute of Aeronautics and Astronautics
0 5 10 15 20
10
7
10
6
10
5
10
4
10
3
10
2
10
1
10
0
t, sec
|
|
|
|
Direct NN
Hybrid NN
RLS + Direct NN
Fig. 12 - Tracking Error Norm
As can be seen, the hybrid adaptive control with the indirect adaptive law is able to improve the tracking perfor-
mance of the direct neural network adaptive control. The combined direct adaptive control with the recursive least
square parameter identication actually outperforms both the direct and hybrid adaptive control approaches as the
tracking error is signicantly reduced as seen in Fig. 12. The control surface deections to achieve this pitch maneu-
ver are shown in Fig. 13. The elevator deection for this pitch maneuver is nearly saturated. The direct neural network
adaptive control produces more overshoot than the hybrid adaptive control and the recursive least square approach.
The left rolling moment is compensated by the right aileron input and the adverse yaw is compensated by a small
rudder input. For this simulation, the actuator dynamics is not included in the study.
0 1 2 3 4 5
0
20
40
t, sec
a
,
d
e
g
(a)
0 1 2 3 4 5
40
20
0
t, sec
e
,
d
e
g
(b)
0 1 2 3 4 5
1
0.5
0
t, sec
r
,
d
e
g
(c)
Direct NN Hybrid NN RLS + Direct NN
Direct NN Hybrid NN RLS + Direct NN
Direct NN Hybrid NN RLS + Direct NN
Fig. 13 - Control Surface Deection
21 of 24
American Institute of Aeronautics and Astronautics
VIII. Discussions
The current research in the damage effect aerodynamic modeling has focused on single damage sites. Multiple
damage sites can also exist in a damage event. The next step in damage effect aerodynamic modeling is to generate
a predictive capability for multiple damage sites. An approach would be to conduct CFD modeling for representative
multiple damage patterns. These modeling results are compared to single damage site models and a learning system
can be developed to establish a surface response mapping between multiple and single damage sites. Using this
damage surface response mapping, damage effects for any arbitrary damage pattern can be rapidly estimated.
In the current adaptive ight control research, only traditional control effectors that include ailerons, elevators, and
rudder are used. In severe damage situations, these control effectors may not be sufcient to stabilize and maintain
good handling qualities of the damaged aircraft. Therefore, any adaptive ight control method must include a control
allocation strategy that utilizes other potential control effectors that are otherwise not used in a conventional ight
control system. These control effectors can include engine differential thrust for yaw control, wing spoilers for roll
control, and wing ap extension or deection for pitch control. Research in the areas of control redundancy design and
recongurable control will be conducted to investigate optimal control allocation strategies for these control effectors.
The issues of time-scale separation due to actuator dynamics are important for systems with different time latency
such as engines and aps and thus will be an area of adaptive ight control research.
17
Damage effects can present a serious challenge to conventional ight control systems because the aircraft ight
dynamics may deviate from its nominal ight dynamics substantially as result of the degradation in the ight per-
formance of the damaged aircraft. This makes it difcult for the conventional ight control systems to cope with
changes in the stability and control of the damaged aircraft. While neural network adaptive control offers a promise
of being able to adapt to changes in ight dynamics of damaged aircraft, rigorous validation by simulations and ight
testing will be pursued to explore areas of concern in the neural network adaptive ight control. One of the unresolved
concerns is the learning characteristics of a neural network. If the dynamic inversion error is large due to a large dis-
crepancy between the true and nominal plant dynamics used in the dynamic inversion control, the learning rate must
be set sufciently high in order for the neural network to reduce the error rapidly. As a consequence of the aggressive
learning, the neural network tends to generate high gain control signals that may not be dynamically achievable. This
potentially can cause a number of problems including control saturation, load constraints during ight being exceeded,
excitation of unmodeled dynamics, and others. One potential solution is to introduce the proposed hybrid adaptive
control that incorporates an explicit parameter identication based on an adaptive law derived from the Lyapunov
stability method or a recursive least-square method to estimate the true plant dynamics, which would be used for the
dynamic inversion control rather than the nominal plant dynamics. This approach potentially offers a way to reduce
the dynamic inversion error that the neural network has to compensate for.
Integrated ight dynamics modeling is another area research that addresses interactions among many types of
physics problems during ight. An integrated ight dynamics model will be developed to include an aeroservoelas-
ticity interaction model of a exible-body vehicle dynamics with a propulsion model and its actuator dynamics. This
integrated model will capture the combined effects of the 6-dof rigid body dynamics, structural dynamics of airframe,
and propulsion model. Post-stall aerodynamics can be included in the aerodynamic coefcients and derivatives which
can inuence the utter margin and the aerodynamic damping of the airframe.
Structural interaction with a ight control system is critical to any ight control development.
1820
Elastic de-
ection and mode shapes can adversely contribute to the vehicle stability and control, resulting in problems such as
utter, control reversal, structural frequency interaction within the ight control bandwidth, and others. Research
in the area of aeroservoelasticity is very important for advancing the knowledge of damage adaptive ight control.
Recent advances in uid-structure interaction modeling using coupled computational uid dynamics-nite element
method provide a predictive capability for aeroservoelastic effects on the stability and control of damaged aircraft.
21
New adaptive ight control methods will need to observe and obey structural load constraints imposed on a damaged
airframe. The resulting adaptive ight control methods therefore would be more dynamically achievable. Aeroser-
voelastic frequency interaction with a safety-critical ight control system will be investigated in order to develop an
integrated approach for dealing with potential issues with high frequency signals from elastic modes injecting into the
frequency bandwidth of the rigid-body aircraft dynamics. Research in aeroservoelastic ltering and structural iden-
tication for ight control will provide methods for assessing the elastic contribution of the airframe and developing
adaptive ight control methods that can effectively lter out unwanted structural resonant modes within the ight
control bandwidth.
22 of 24
American Institute of Aeronautics and Astronautics
IX. Conclusions
This paper has presented recent results on the modeling, control, and simulation of damaged aircraft as part of
the aviation safety research at NASA. The damage effect aerodynamic modeling has been performed to provide an
understanding of the control and stability of an asymmetric damaged aircraft. The effects of aerodynamic and control
coupling in all the three stability axes are revealed from the modeling results. A 6-dof ight dynamics of asymmetric
aircraft is derived in order to account for the effect of the center of gravity shift resulting from the damage. An
approach for trimming the damaged aircraft for the translational motion is presented. The trim procedure provides
initial estimates of the trim values for the angle of attack, angle of sideslip, and engine thrust. The inuence of the
control surfaces on these trimmed values is then accounted for by adjusting the initial trim values with the control
surface deections obtained from the ight control. A hybrid direct-indirect neural network adaptive ight control
concept has been proposed to provide an opportunity to estimate plant dynamics in conjunction with the current direct
adaptive control augmentation strategy. The on-line estimation of the plant dynamics is provided by an adaptive
law derived from the Lyapunov stability theory and the recursive least-square method. The adaptive ight control is
designed to track a reference model that species desired handling characteristics for a class of transport aircraft. The
feedback control augmentation uses a proportional and integral scheme to handle errors in the roll, pitch, and yawrates.
A control simulation of the direct adaptive control law is performed in a ight simulator to assess the stability recovery
of a damaged generic transport model using the neural network adaptive ight control. The results of the simulation
show that the direct neural network control augmentation scheme is able to stabilize a damaged aircraft. In the near
future, a control simulation of the hybrid adaptive control law will be conducted to investigate the potential benets
offered by this proposed scheme in reducing the possibility of high gain control in the present direct adaptive control
strategy. Moreover, adaptive ight control research will advance the knowledge in the area of integrated ight control
with propulsion and airframe effects in order to address interactions between vehicle dynamics, propulsion dynamics,
and structural dynamics that may be present. The assumption of rigid-body aircraft ight dynamics no longer holds
true as the aircraft will have to be treated as an elastic body. This will give rise to challenges in developing adaptive
ight control that can handle aeroservoelastic effects of damaged aircraft.
References
1
National Transportation Safety Board, C., In-Flight Separation of Vertical Stabilizer American Airlines Flight 587 Airbus Industrie A300-
605R, N14053 Belle Harbor, New York, November 12, 2001, NTSB/AAR-04/04, 2004.
2
Hughes, D. and Dornheim, M.A., DHL/EAT Crew Lands A300 With No Hydraulics After Being Hit by Missile, Aviation Week Space &
Technology, pp. 42, 12/08/2003.
3
Kaneshige, J., Bull, J., and Totah, J., Generic Neural Flight Control and Autopilot System, AIAA Guidance, Navigation, and Control
Conference, AIAA-2000-4281, 2000.
4
Atkins, E., Dynamic Waypoint Generation Given Reduced Flight Performance, 42nd AIAA Aerospace Sciences Meeting and Exhibit,
AIAA-2004-779, 2004.
5
Jacklin, S.A., Schumann, J.M., Gupta, P.P., Richard, R., Guenther, K., and Soares, F., "Development of Advanced Verication and Validation
Procedures and Tools for the Certication of Learning Systems in Aerospace Applications", Proceedings of Infotech@aerospace Conference,
Arlington, VA, Sept. 26-29, 2005.
6
Totah, J., Kinney, D., Kaneshige, J., and Agabon, S., An Integrated Vehicle Modeling Environment, AIAA Atmospheric Flight Mechanics
Conference and Exhibit, AIAA-1999-4106, 1999.
7
Steinberg, M.L., A Comparison of Intelligent, Adaptive, and Nonlinear Flight Control Laws, AIAA Guidance, Navigation, and Control
Conference, AIAA-1999-4044, 1999.
8
Rohrs, C.E., Valavani, L., Athans, M., and Stein, G., Robustness of Continuous-Time Adaptive Control Algorithms in the Presence of
Unmodeled Dynamics, IEEE Transactions on Automatic Control, Vol AC-30, No. 9, pp. 881-889, 1985.
9
Ioannu, P.A. and Sun, J. Robust Adaptive Control, Prentice-Hall, 1996.
10
Eberhart, R.L. and Ward, D.G., Indirect Adaptive Flight Control System Interactions, International Journal of Robust and Nonlinear
Control, Vol. 9, pp. 1013-1031, 1999.
11
Rysdyk, R.T. and Calise, A.J., Fault Tolerant Flight Control via Adaptive Neural Network Augmentation, AIAA Guidance, Navigation,
and Control Conference, AIAA-1998-4483, 1998.
12
Kim, B.S. and Calise, A.J., Nonlinear Flight Control Using Neural Networks, Journal of Guidance, Control, and Dynamics, Vol. 20, No.
1, pp. 26-33, 1997.
13
Johnson, E.N., Calise, A.J., El-Shirbiny, H.A., and Rysdyk, R.T., Feeback Linearization with Neural Network Augmentation Applied to
X-33 Attitude Control, AIAA Guidance, Navigation, and Control Conference, AIAA-2000-4157, 2000.
14
Narendra, K.S. and Annaswamy, A.M., A New Adaptive Law for Robust Adaptation Without Persistent Excitation, IEEE Transactions on
Automatic Control, Vol. AC-32, No. 2, pp. 134-145, 1987.
15
Lewis, F.W., Jagannathan, S., and Yesildirak, A., Neural Network Control of Robot Manipulators and Non-Linear Systems, CRC, 1998
23 of 24
American Institute of Aeronautics and Astronautics
16
Krishnakumar, K., Limes, G., Gundy-Burlet, K., and Bryant, D., An Adaptive Critic Approach to Reference Model Adaptation, AIAA
Guidance, Navigation, and Control Conference, AIAA-2003-5790, 2003.
17
Naidu, D.S. and Calise, A.J., Singular Perturbations and Time Scales in Guidance and Control of Aerospace Systems: A Survey, Journal
of Guidance, Control, and Dynamics, Vol. 24, No. 6, 0731-5090, pp. 1057-1078, 2001.
18
Brenner, M.J. and Prazenica, R.J., Aeroservoelastic Model Validation and Test Data Analysis of the F/A-18 Active Aeroelastic Wing,
NASA/TM-2003-212021, 2003.
19
Meirovitch, L. and Tuzcu, I, Integrated Approach to Flight Dynamics and Aeroservoelasticity of Whole Flexible Aircraft - Part I: System
Modeling, AIAA Guidance, Navigation, and Control Conference, AIAA-2002-4747, 2002.
20
Meirovitch, L. and Tuzcu, I, Integrated Approach to Flight Dynamics and Aeroservoelasticity of Whole Flexible Aircraft - Part II: Control
Design, AIAA Guidance, Navigation, and Control Conference, AIAA-2002-5055, 2002.
21
Livne, E., Future of Airplane Aeroelasticity, Journal of Aircraft, Vol. 40, No. 6, 2003.
24 of 24
American Institute of Aeronautics and Astronautics