Professional Documents
Culture Documents
JOHAN KNS
Examensarbete
Stockholm, 2011
Abstract
This report examines different control design methods, linear as well as nonlinear,
for a suborbital reusable launch vehicle. An investigation of the natural vehicle
behavior is made, after which a baseline linear controller is constructed to fulfill
certain handling quality criteria. Thereafter the nonlinear cascade control methods
block backstepping and nonlinear dynamic inversion via time scale separation are
examined and used to construct two nonlinear controllers for the vehicle. Optimal
controllers, in terms of three different criteria, are found using the genetic optimization algorithm differential evolution. The optimal controllers are compared, and it
is found that nonlinear dynamic inversion via time scale separation performs better
than block backstepping with respect to the cases investigated. The results suggest control design by global optimization is a viable and promising complement to
classical methods.
Keywords: Flight control systems, nonlinear control, global optimization
Acknowledgments
I am indebted to my supervisors Fredrik Berefelt and John W.C. Robinson at the
Swedish Defence Research Agency, FOI, for their continual guidance and support.
I would also like to thank Professor Xiaoming Hu at the Division of Optimization
and Systems Theory at the Department of Mathematics at KTH Royal Institute of
Technology for his advice and support.
Contents
Aerodynamic Nomenclature
1. Background
8
1.1. Scope and Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2. Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2. Vehicle Model
11
2.1. Actuator Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3. Vehicle Characteristics
3.1. Control Effectors . . . . . .
3.2. Open Loop Behavior . . . .
3.2.1. Pull Up . . . . . . .
3.2.2. Velocity Vector Roll
3.3. Results . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
14
14
15
16
16
16
4. Baseline Controller
21
4.1. Longitudinal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.2. Lateral Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.3. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5. Nonlinear Control
5.1. Motivating Idea . . . . . . . . .
5.2. Block Backstepping . . . . . . .
5.2.1. Extension of the Model
5.3. Nonlinear Dynamic Inversion .
5.4. Application to Aircraft . . . . .
5.4.1. BBS . . . . . . . . . . .
5.4.2. NDI-TSS . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
34
34
35
37
39
40
42
43
6. Optimization
44
6.1. Differential Evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.2. Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Contents
6.3. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.4. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
A. Derivations and Proofs
56
A.1. Dynamics of , , and . . . . . . . . . . . . . . . . . . . . . . . . . . 56
A.2. Invertibility of the Virtual Control Gain Matrix
. . . . . . . . . 58
Aerodynamic Nomenclature
a
e
r
J
n
F aero
Mb
u
v
g
l
m
m
n
p
q
r
u
V
angle of attack
sideslip angle
aileron angle
elevator angle
rudder angle
moment of inertia tensor
undamped actuator frequency
heading
actuator time constant
elevation
bank
control surface angles
rotational velocity
aerodynamic force
moment
control surface inputs
velocity
actuator damping coefficient
acceleration due to gravity
rolling moment
airframe mass
pitching moment
yawing moment
roll rate
pitch rate
yaw rate
velocity along body-fixed x-axis
airspeed
Contents
v
w
xe
ye
ze
CHAPTER
Background
Since the advent of the Space Age, the need for more inexpensive and reliable access to
space has been ever-increasing. The technological and economical barriers involved in
launching vehicles into space have made space access prohibitive to many prospects,
commercial and otherwise. In recent years, significant efforts have been made in
the research and development of fully reusable launch vehicles (RLVs) in order to
provide lower-cost and more reliable access to space (see [4]), not only in scientific
contexts such as transportation to and from the International Space Station, but
also for novel ideas such as space tourism. These initiatives have been undertaken
by several organizations in various parts of the world, including NASA in the United
States and the European Space Agency (ESA) and Virgin Galactic in Europe.
lay the groundwork for future high-altitude high-speed transportation. The initiative consists of two sub-projects, one concerning the implementation of the suborbital
aircraft ALPHA (Airplane Launched PHoenix Aircraft) of Fig. 1.1, whose primary
mission is essentially to provide the means for space tourism, and the other with
a hypersonic passenger aircraft, SpaceLiner. Both projects have many technological
obstacles, including exceedingly high temperatures upon atmospheric reentry and excessive accelerations (possibly unsuitable for humans). To tackle these complications,
the projects have been divided into smaller tasks and distributed to 17 specialist companies and institutions in various countries across Europe. This study was conducted
at one of these institutions: the Swedish Defence Research Agency, FOI.
10
Chapter 1. Background
using time scale separation. The performance of these controllers is investigated using
global optimization tools (Chapter 6).
120
Ascent and Initial Descent
Final Descent
Altitude [km]
100
80
60
40
20
200
400
600
Time [s]
800
1000
1200
1.2. Notation
Vectors and matrices are written in bold font. Vectors are by default column vectors.
The Euclidean norm of a vector v is written as kvk. A variable preceded by a
signifies a deviation from the nominal value. A primed variable signifies a new
variable related to the unprimed variable of the same name. Rotational matrices are
expressed as T ab , where a and b are the source and destination reference frames,
respectively. The derivative of a function f : Rm Rn with respect to a vector
x Rm is denoted by the Jacobian matrix f /x, which at row i and column j
has the derivative fi /xj . The transpose of a matrix A is written AT . The open
interval between a and b is denoted by (a, b) and the closed, left-open, and right-open
by [a, b], (a, b], and [a, b), respectively. The set R+ denotes the strictly positive real
axis. A diagonal square matrix A Rnn with elements a1 , a2 , . . . , an along the
diagonal is written diag(a1 , a2 , . . . , an ). No notational distinction is made between a
function and its Laplace transform; the distinction is evident by the context in which
the function appears.
CHAPTER
Vehicle Model
The model of the airframe forms the foundation of all flight simulations and control
designs made in this study. To construct the model, two right-handed Cartesian
coordinate systems are utilized: the aircraft-fixed and the Earth-fixed coordinate
systems. The aircraft-fixed coordinate system has its origin at the aircraft center of
mass and its x-, y-, and z-axes pointing through the nose, through the starboard
wing, and down, respectively. The Earth-fixed coordinate system has its x-, y-, and
z-axes pointing north, east, and down, respectively. The origin is usually set at the
starting point of a flight, at ground level.
Using these coordinate systems, the following standard six-degree-of-freedom (henceforth 6-DOF) model describing the translation and rotation of the airframe in its
own coordinates is used as a foundation for the model (the derivations of which can
be found in several standard references, e.g. [10, 15]):
0
1
v = F aero + T eb 0 v
(2.1)
m
g
= J 1 (M b J ),
(2.2)
T
where v = u v w is the vehicles velocity, m its mass, F aero R3 the aero
T
dynamic force, g the magnitude of the acceleration due to gravity, = p q r
the angular velocity, J R33 the moment of inertia tensor (here assumed positive
T
definite), and M b = l m n the external moment1 .
1
The variable m can be assumed to refer to the aircraft mass unless explicitly referred to as pitching
moment.
12
To describe the vehicles orientation and position in the Earth-fixed coordinate system, the following six kinematic equations are needed, which can be found in [10]:
x e
y e = T be v,
ze
(2.3)
(2.4)
where xe , ye , and ze are the positions in the Earth coordinate system. The angles ,
, and are the Euler angles representing heading, elevation, and bank, respectively,
which describe three successive transformations used to rotate the Earth frame to
the aircraft frame, namely rotations about (1) the Earth z-axis, (2) the y-axis of the
rotated coordinate system, and finally (3) the x-axis of the second rotated coordinate
system. The order in which the rotations are performed is important since one triple
of angles does not give a unique orientation for all rotational orders. The order given
above (heading, elevation, bank) is a convention used in the aircraft community.
The force F aero and the moment M b in Eqs. (2.1) and (2.2) are functions of the
aircraft state as well as its position and orientation. In this study, they are approximated using aerodatadata collected from CFD2 calculations and wind tunnel
measurements. These data are parameterized by, among other variables, Mach number. The lowest Mach number measured is 0.2, which is used as a lower Mach number
limit for designs and simulations in this study as well.
Additional variables that are used throughout this text are the airspeed V , angle of
attack , and sideslip angle , which in terms of the defined states can be expressed
as
V = kvk
w
= arctan
u
v
= arcsin .
V
(2.5)
(2.6)
(2.7)
cos cos
v = V sin .
sin cos
(2.8)
The angles of attack and sideslip are shown visually in Figs. 2.1 and 2.2, respectively.
Their domains are set to (2 , 24 ) and (30 , 30 ), respectively, in accordance with
aerodata limits.
2
13
x
v
(2.9)
where is the actuators time constant and and u the vectors of actuator positions
and control inputs, respectively. The time constant for the ALPHAs actuators is
unknown and is therefore approximated by taking that of the actuators of the similar
NASA X-38 vehicle, which according to [7] have an undamped natural frequency
n = 26 s1 and damping ratio = 0.707, i.e. = 1n 54 ms.
CHAPTER
Vehicle Characteristics
For an appropriate control design to be made, the basic characteristics and behavior
of the aircraft need to be investigated. For example, an airframe may be naturally
stable and thus not require a stability augmentation system, or it may be naturally
unstable and have insufficient control authority. This analysis is done by observing
how the airframe responds to control surface deflections as well as determining the
conditions necessary for the vehicle to be in an equilibrium state.
15
Name
Primary control
Rudder
Body flap
In- and outboard elevators
In- and outboard ailerons
Speed brakes
RU
BF
EI & EO
AI & AO
SB
Yaw
Pitch (primarily for trimming)
Pitch
Roll
Velocity
Control surface
Positive deflection
Rudder
Elevators and body flap
Ailerons
and therefore can exert more moment using the same deflection. Unless otherwise
specified, elevator henceforth refers to EI and aileron to AO.
Speed brake
Rudder
Body flap
Inboard elevator/aileron
Outboard elevator/aileron
Figure 3.1.: The ALPHA and its control surfaces.
Using this knowledge of the control surfaces, one can begin to investigate the aircraft
open loop behavior.
16
10
0
20
10
15
10
20
10
40
20
10
20
10
20
RU
SB
AO
AI
BF
EO
EI
vestigated in the context of equilibrium states, i.e. states in which the aircraft is
trimmed (which are also used as linearization points in the linear control design in
Chapter 4). It is first useful to make the observation that the most general steady
state motion (in terms of the body equations) an aircraft can make is helical. This
motion can be decomposed into rotations about the lateral axis (pull ups) and about
the velocity vector (velocity vector rolls).
3.2.1. Pull Up
In flight with constant angle of attack and constant pitching rate the aircraft is said to
be performing a pull up. Contained in this class of maneuvers are not only upward and
downward rotations but also straight and level flight. The states are found by solving
T
T
q(x)
(x)
(x)
where x is as defined above.
3.3. Results
Figures 3.2 and 3.3 show the longitudinal control surface deflections (body flap and
elevator) required for steady state pull ups of different normal accelerations at different Mach numbers. First, the body flap is used to attempt a steady state. If it
saturates, the elevator is engaged. The essential trend is that with higher dynamic
17
3.3. Results
pressure, less deflection is needed. In particular, the body flap is capable of keeping
the vehicle in steady state at high dynamic pressures. The more the height and/or
speed increases, however, the more the elevator needs to be engaged.
Elevator [deg]
3
2.5
2.5
2
nz
nz
1.5
1.5
0.5
0.5
0
0.2
-15
0
0.2
0.4
0.6
Mach
-10
-5
-15
0.4
0.6
Mach
-10
-5
Figure 3.2.: Body flap and elevator deflections needed for a steady state pull up at different
normal acceleration factors nz = qV /g and Mach numbers. The Machnz region that is solid
white in both graphs represents untrimmable states. Altitude: 0 km.
Figures 3.4 and 3.5 show the control surface deflections and angles of attack and
sideslip for different roll rates at 0 km and 12.5 km, respectively. It is apparent in
Fig. 3.5 that the airframe cannot sustain a roll at low dynamic pressures (see the
region between Mach 0.2 and 0.4, for example). The limiting factor appears to be
predominantly the angle of attack boundaries (2 < < 24 ) at lower roll rates
and the elevator at higher roll rates. In other words, there is not enough lift to
keep the airframe in the roll. At high roll rates, one might expect the ailerons to
saturate. However, it is apparent from both figures that the limiting factor is the
rudder limitations.
18
Elevator [deg]
3
2.5
2.5
2
nz
nz
1.5
1.5
0.5
0.5
0
0.2
-15
0
0.2
0.4
0.6
Mach
-10
-5
-10
0.4
0.6
Mach
-5
Figure 3.3.: Body flap and elevator deflections needed for a steady state pull up at different
normal acceleration factors nz = qV /g and Mach numbers. The Machnz region that is solid
white in both graphs represents untrimmable states. Altitude: 12.5 km.
19
3.3. Results
Elevator [deg]
6
4
2
0
-2
100
0
0.2
0.3
0.4
0.5
0.6
Mach
Rudder [deg]
0.7
0.8
8
6
4
2
0
100
0
0.2
0.3
0.4
0.5
0.6
Mach
Aileron [deg]
0.7
0.8
0
-5
-10
-15
100
0
0.2
0.3
0.5
0.6
Mach
Angle of Attack [deg]
0.4
0.7
0.8
12
10
8
6
4
2
100
0
0.2
0.3
0.4
0.5
0.6
Mach
Sideslip Angle [deg]
0.7
0.8
0.5
100
0
-0.5
0
0.2
0.3
0.4
0.5
Mach
0.6
0.7
0.8
Figure 3.4.: Control surface deflections as well as angles of attack and sideslip needed for
velocity vector rolls of varying roll rates and Mach numbers. Altitude: 0 km.
20
Elevator [deg]
15
10
5
0
-5
100
0
0.2
0.3
0.4
0.5
0.6
Mach
Rudder [deg]
0.7
0.8
8
6
4
2
0
100
0
0.2
0.3
0.4
0.5
0.6
Mach
Aileron [deg]
0.7
0.8
0
-5
100
-10
-15
0
0.2
0.3
0.4
0.5
0.6
Mach
Angle of Attack [deg]
0.7
0.8
20
100
0
0.2
10
0.3
0.4
0.5
0.6
Mach
Sideslip Angle [deg]
0.7
0.8
0.5
0
-0.5
-1
100
0
0.2
0.3
0.4
0.5
Mach
0.6
0.7
0.8
Figure 3.5.: Control surface deflections as well as angles of attack and sideslip needed for
velocity vector rolls of varying roll rates and Mach numbers. Altitude: 12.5 km.
CHAPTER
Baseline Controller
In order to have a means for evaluating the performance and response of the nonlinear
controller, a baseline linear controller is designed for the family of general linearized
systems
f (x, u)
f (x, u)
x =
x +
u,
(4.1)
x x=x0
u x=x0
u=u0
u=u0
(4.2)
T
uT
at several equilibrium points xT
in the flight envelope. The variables and
0
0
dynamics are defined in more detail below.
The airframes longitudinal and lateral modes are in the linear approximation assumed to be decoupled. The linear controller design can therefore be divided into
two design tasks: pitch channel control (longitudinal) and roll channel control (lateral). Since the vehicle in question is not intended to make any extreme maneuvers,
this approximation is valid for most of the vehicles flight and the controllers relying
on it can be expected to behave well.
The design procedure for each of the control loops of the two channels is as follows:
1. A loop structure is posited with gains left undefined.
2. Desired handling qualities (HQs) are constructed as closed loop pole boundaries
in the complex plane. Because of its ubiquity, a subset of the HQ requirements
outlined by the US Department of Defense in [2] is used (shown in Table 4.1).
22
(4.4)
if and only if
0
d0 () d1 () . . . dn ()
a(q ) =
0
0
d0 () . . . dn1 ()
(4.5)
(4.6)
d1 () = 2()
(4.7)
di () = 2()di1 2 () + 2 () di2 , i = 2, 3, . . . , n.
(4.8)
23
HQ
Minimum
Maximum
Damping
CAP
Im s 6= 0
Damping
Natural frequency n
n
Im s 6= 0
Time constant
Im s = 0
0.3
0.085
2
3.6
0.08
0.4
0.1
0
1.4
24
Im
n,min
n,max
min
max
0
Re
nz g
V
(4.10)
holds. To find a suitable K, the two components of C are set equal to each other:
nz = Kq = K
nz g
,
VCO
(4.11)
where VCO is the velocity at which the two quantities should have equal influence.
The authors of [18] choose VCO = 122 m s1 , resulting in K = 12.4.
Similar to the longitudinal control loops in [12] and [15], a control loop for C is
constructed, shown in Fig. 4.2. Using the variables mentioned above, the transfer
function from elevator to nz can be expressed as Gnz (s) = Nnz (s)/D(s) and from
elevator to q as Gq (s) = Nq (s)/D(s) using the reduced linearized system
x = Alng x + B lng ue ,
(4.12)
T
T
where x = q nz e , x = q nz e , and the state and control matrices Alng R44 and B lng R4 , respectively, are Jacobian matrices deter
mined by numerical linearization. The transfer function from Ccmd
to C can thus
25
Ccmd
KI /s
EI
ALPHA
nz
KP
C
(4.13)
q
12.4
A coordinated turn is, as defined in [15], a turn where there is no aerodynamic force along the
lateral axis. Furthermore, the airspeed and angular velocity need to be constant for the aircraft
to be in steady state.
26
(n )min
n,min
n,min
Real roll root
Roll time constant
Im
1/1.4
Re
KI /s
cmd Parameter cmd
Lookup
RUcmd
K1
K2
AO
ALPHA
Figure 4.4.: Lateral control loop.
The bank angle and rudder position RU needed for turn coordination
turn rate .
for the specified are computed offline for a range of different turn rates at different
parts of the flight envelope. This computation also includes the necessary values for
the longitudinal states, which are passed to the integrated C controller (not shown
in Fig. 4.3).
Through this loop, the transfer function from cmd to is
G(s) =
K1 N (s)(s + KI )
,
sD(s) sK2 Np (s) + K1 N (s)(s + KI )
(4.14)
27
4.3. Results
depends on the sideslip angle , roll rate p, yaw rate r, roll angle , and aileron
actuator position:
x = Alat x + B lat ua ,
(4.15)
T
T
where x = p r a , x = p r a , and the state
and control matrices Alat R55 and B lat R5 , respectively, are Jacobian matrices
determined by numerical linearization.
As in the longitudinal case, Theorem 1 is used to map the boundaries (the shapes
of which are shown in Fig. 4.3) from the complex plane to parameter space to find
satisfactory K1 and K2 . Before applying Theorem 1, however, KI is set to 1/50 (chosen by trial-and-error) to keep the parameter space two-dimensional for simplicity.
Contrary to the longitudinal controller, the lateral controller needs to fulfill the HQs
for multiple modes. Therefore, the boundaries mapped into parameter space need to
be satisfied simultaneously. In other words, the region of intersection of all allowable
regions needs to be found.
4.3. Results
Figure 4.5 shows the mapping of the denominator of the longitudinal controllers
transfer function in Eq. (4.13) to parameter space for three different Mach numbers at
one altitude. Apparently, the two active conditions are maximum natural frequency
and minimum damping, the region S between which is that of acceptable values. This
region shrinks with increasing Mach number as it becomes more difficult to damp
oscillations and lower the natural frequency. For several altitudes and Mach numbers,
a point inside the region is chosen for usage in interpolation in a gain scheduler, i.e. a
function parameterized by the flight envelope position (altitude and Mach number)
T
which gives a vector KP KI as output.
Figures 4.6 and 4.7 show the responses to a unit step in commanded C for two
altitudes at different Mach numbers using the gain scheduled longitudinal C control
loop. The vehicle exhibits a fast and well-damped response to the input under the
regulation of the longitudinal control loop.
Proceeding in a similar manner as in the the longitudinal control loop implementation, the denominator of the lateral controllers transfer function in Eq. (4.14) is
mapped to parameter space for three different Mach numbers at one altitude, shown
in Figure 4.8. As in the longitudinal case, the region S shrinks as the Mach number
increases. The active boundaries in this case, however, are the ones that ensure real
roll subsidence and imaginary Dutch roll roots as well as a satisfactory time constant
T
of the roll subsidence mode. A gain scheduler with the vector of gains K1 K2 as
output is made by choosing a point inside S for several altitudes and Mach numbers
and then interpolating between them.
28
KI
0
-5
-10
-3
-2
-1
KP
-5
KI
KI
-10
-3
-5
-10
-2
-1
KP
-3
-2
-1
KP
Figure 4.5.: Boundaries for the longitudinal parameters mapped by Theorem 1 at an altitude
of 3 km for, from left to right and top to bottom, Mach numbers 0.4, 0.6, and 0.8. The map
for Mach 0.2 is missing since the aircraft is untrimmable with only the body flap at that point.
Origins of boundaries as defined in Fig. 4.1.
Governed by this gain scheduled controller, two coordinated turns with different turn
rates at different parts of the flight envelope are made, the results of which are
shown in Figures 4.9 and 4.10. In both cases the turn rates reach the commanded
rates, and so do the bank angles (with a small steady state error, however). The
overshoots in are quite large, but the damping ratios are within the acceptable
range. With the exception of some excursions in the transient region when the
vehicle is entering the turn, the aerodynamic side force Faero,y = mgny tends to zero,
which, along with the condition of constant angular velocity and constant airspeed,
makes the turn coordinated.
29
4.3. Results
C*
1
Response
Command
0.5
0
10
15
10
15
10
15
Time [s]
C*
1
0.5
0
5
Time [s]
C*
1
0.5
0
5
Time [s]
Figure 4.6.: C unit step responses at an altitude of 3 km at, in order from top to bottom,
Mach 0.4, 0.6, and 0.8.
30
C*
1
Response
Command
0.5
0
10
15
10
15
10
15
Time [s]
C*
1
0.5
0
5
Time [s]
C*
1
0.5
0
5
Time [s]
Figure 4.7.: C unit step responses at an altitude of 9 km at, in order from top to bottom,
Mach 0.6, 0.7, and 0.8.
31
4.3. Results
15
K2
10
5
0
-60
-20
-40
K1
10
10
S
5
0
-60
K2
15
K2
15
-40
-20
K1
5
0
0
-60
-40
-20
K1
Figure 4.8.: Boundaries for the lateral parameters mapped by Theorem 1 at an altitude of
3 km for, from left to right and top to bottom, Mach numbers 0.4, 0.6, and 0.8. The map for
Mach 0.2 is missing since the aircraft is untrimmable with only the body flap at that point.
Origins of boundaries as defined in Fig. 4.3.
32
[deg/s]
20
Response
Command
10
0
0
10
Time [s]
15
20
10
Time [s]
15
20
10
Time [s]
15
20
[deg]
100
50
0
ny
0.5
0
-0.5
33
4.3. Results
[deg/s]
20
Response
Command
10
0
0
10
Time [s]
15
20
10
Time [s]
15
20
10
Time [s]
15
20
[deg]
100
50
0
ny
0.5
0
-0.5
CHAPTER
Nonlinear Control
The advantages of using nonlinear control methods over linear ones are manifold,
in particular the class of controllers of systems in cascade form on which this study
is focused. This class of controls includes, for example, backstepping and nonlinear
dynamic inversion. In contrast to linear control methods, these cascade methods
are not based on linear approximations of the systems dynamics and are thus built
on a more solid foundation. Since the nonlinear system description is valid globally
(in terms of the flight envelope), gain scheduling is not required. Furthermore, the
implementation of these controllers is straightforward.
This is not to say, however, that these controllers are devoid of difficulties. For
example, the dynamics must be in strict-feedback form. Additionally, there is no
clear-cut manner in which to control output variables as in the linear case.
Despite these difficulties, the benefits of this methodology outweigh its drawbacks in
terms of performance and effort required in implementation. This is true particularly
in the case of aircraft, whose motion is essentially governed by a chain of integrators:
commanded control surface deflection rates are integrated into moments, moments
are integrated into rotational velocities, which finally are integrated into orientational
angles. This inherent property makes aircraft well-suited for cascade designs.
i = 1, 2, . . . , n
(5.1)
35
of order n, with input xn+1 . The idea behind the recursive backstepping and nonlinear dynamic inversion via time scale separation methods is to control xi by using
xi+1 as a virtual control, which is chosen in such a way as to steer xi to the desired
point. The first virtual control x2 is chosen in such a way as to steer x1 to the origin
while at the same time giving it certain desired dynamics f1des (x1 ). Next, x3 is chosen
to steer x2 to its desired value. This process continues recursively for all remaining
states xi , i = 3, 4, . . . , n, and can be done using varying degrees of accuracy, as is
evident below.
(5.2)
(5.3)
(5.4)
1
2
36
) = h(x, y
+y des (x)) and k(x,
) =
chasing the desired state y des (x). Define h(x,
y
y
des
3
+ y (x)). Equations (5.2) and (5.3) can then be written as
k(x, y
x = f des (x) + g(x)
y
) + k(x,
)u y des (x, y
).
y
= h(x,
y
y
(5.5)
(5.6)
) = x T Qx x + y
V (x, y
2
2
(5.7)
with positive definite weighting matrices Qx Rmm and Qy Rnn , and choosing
a control law u such that the clfs time derivative
) = xT Qx x + y
T Qy y
V (x, y
= xT Qx f des (x) + g(x)
y
T Qy h(x,
) + k(x,
)u y des (x, y
)
+y
y
y
(5.8)
is negative definite along the solution trajectory. The control law needs to cancel
potentially indefinite terms in V and is additionally exploited to synthesize new
) for y. One possible control law is
error dynamics hdes
(x, y
y
1
T
des
)1 hdes
(x,
y
)
h(x,
y
)
Q
g(x)
Q
x
+
y
(x,
y
)
,
u = k(x,
y
x
y
(5.9)
where it is assumed that the desired dynamics f des (x) for the x-subsystem and
) for the y
-subsystem are smooth and chosen such that xT Qx f des (x) and
hdes
(x, y
y
T Qy hdes
), respectively, are negative definite (e.g. f des (x) = Ax, where A
y
(x, y
y
nn
R
is negative definite). Combining Eqs. (5.5), (5.6), and (5.9), one arrives at the
system
x = f des (x) + g(x)
y
1
des
) Q g(x)T Qx x,
y
= hy (x, y
(5.10)
(5.11)
37
(5.12)
, u
) = f D (x, y
+ y (x), u
+ u (x, y
)) and g
D (x, y
, u
) =
the functions f D (x, y
+ y des (x), u
+ udes (x, y
)) defined, where udes (x, y
) is the virtual control law
g D (x, y
for the y-subsystem chosen to satisfy Eq. (5.9) in place of u. This error variable
, where Qu Rpp , which is
T Qu u
is then used in a positive definite function 21 u
appended to the clf given in Eq. (5.7) to include the new dynamics:
1 T
, u
) = V (x, y
) + u
.
Qu u
VD (x, y
2
(5.13)
) + k(x,
)u y des (x, y
)
y
= h(x,
y
y
des
) + k(x,
)(u u (x, y
) + udes (x, y
)) y des (x, y
)
= h(x,
y
y
(5.14)
) + k(x,
)
)udes (x, y
) y des (x, y
),
= h(x,
y
y
u + k(x,
y
))4
leading to the clf time derivative (with terms already canceled by udes (x, y
, u
) = xT Qx x + y
T Qy y
T Qu u
V D (x, y
+ u
= xT Qx f des (x)
T Qy hdes
+y
(x,
y
)
+
k(x,
y
)
u
y
, u
) + g
D (x, y
, u
)v u des (x, y
, u
) .
T Qu f D (x, y
+u
(5.15)
, u
) is the derivative of udes (x, y
) along the solution trajectory. The u
The function u des (x, y
dependence appears because of the dependence thereon by y
.
38
Q1
k(x,
y
)
Q
y
+
u
(x,
y
,
u
)
,
y
u
, u
) is the desired error dynamics for u, chosen such that the function
where f des
(x, y
D,u
des
T
) + k(x,
)
y
= hy (x, y
y
u Q1 g(x)T Qx x
(5.17)
, u
) Q1
)T Qy y
.
u
= f des
(x, y
D,u
k(x, y
u
(5.19)
(5.18)
(5.21)
= T Q (f I () + g I ()x)
+ xT Qx (f des (x) + g(x)
y)
T Qy h(x,
) + k(x,
)u y des (x, y
) .
y
y
+y
5
Note that it is allowed that all lower states also be functions of , e.g. h = h(, x, y).
(5.22)
39
This function is made negative definite along the solution trajectory by using the
augmented function
T
f des0 (, x) = f des (x) Q1
x g I () Q
(5.23)
) h(x,
)
) = k(x,
)1 hdes
y
(5.24)
udes (x, y
y
(x, y
y
, u
)
, u
) f D (x, y
, u
)1 f des
D (x, y
(5.25)
v=g
(x, y
D,u
in place of the BBS control laws in Eqs. (5.9) and (5.16), respectively. The virtual
control law y des (x) is as defined in Eq. (5.4) and the desired dynamics for the xsubsystem as in Eq. (5.23). With the exception of certain terms being absent, the
control laws look very similar to those of the BBS controller. They are in fact a
simplified case of the BBS control laws. For each subsystem, the time derivative
of the desired signal as well as the inter-subsystem term that guarantees a negative
definite clf are omitted. Comparing the closed loop system
x = f des (x) + g(x)
y
des
) + k(x,
)
)
y
= hy (x, y
y
u y des (x, y
, u
) u des (x, y
, u
)
u
= f des
(x, y
D,u
(5.26)
(5.27)
(5.28)
with that resulting from the BBS method given in Eqs. (5.17) to (5.19), one sees
that the omission of the terms in the control laws leads to the time derivatives of the
virtual controls remaining. These terms are small in comparison to the other terms
40
when there is time scale separation. As noted in [17], in the limit of complete time
scale separation as well as when Qy = k1 I, Qu = k2 I, k1 , k1 /k2 0, the BBS
and NDI-TSS methods formally converge.
In terms of stability, BBS is clearly superior since it is based on fewer assumptions.
The cost for this guaranteed stability is two extra terms, one of which involves a
potentially complicated differentiation.
des
(, )
b = M b M des (, ,
),
M
b
(5.29)
(5.30)
(5.31)
(5.32)
They are also completely decoupled from the linear controller presented in Chapter 4.
(5.33)
41
sec 0
1
sin
0
cos
0
1
f (, ) =
F tot
cos sin cos sin sin
mV
0
0
cos tan
1
sin tan
sin
0
cos
g() =
1
sin tan cos tan
F tot = F tot (, ).
(5.34)
(5.35)
(5.36)
For derivations of these equations, see Section A.1. The dependence of the force
F tot and the function f on the vector of actuator positions is due to deflections of
the control surfaces not purely exerting moments on the vehicle, but also (relatively
small) forces. This dependence makes a standard BBS design impossible and is in
this study neglected so that F tot = F tot () and f = f ().
= f (
+ cmd )
cmd 7 and g
= g(
+ cmd ). Then, differentiation
()
Let f ()
of Eq. (5.30) yields
+g
= f ()
()
+g
des (, )
+g
.
()
()
= f ()
(5.37)
Using Eq. (2.2), the dynamics for the rotational velocity can be written as
= h() + kM b ,
(5.38)
= h(
des
)
)
)
b.
(, ,
+ kM des (, ,
+ kM
= h(
(5.39)
In order to account for model errors and synthesize approximate second order dynamics, a (virtual) integrator of the orientational error states is included in the controller
as a state R3 :
= K + ,
(5.40)
where K = diag(K, , K, , K, ) has strictly positive diagonal elements. Driving
this state to zero forces to cmd as t . The term K makes the integrator leak and is necessary for the positive definiteness of the clf, or, indirectly,
to ensure stability of the origin of .
7
Here, the formulation is that of a tracking problem. As such, the system is nonautonomous (due
to the time dependence of cmd ). However, the Lyapunov stability arguments presented above,
which are formulated for a setpoint problem, can be shown to be valid in a tracking setting as
well by a generalization of the autonomous case as shown in Theorems 4.8 and 4.9 in [8].
42
Since the actuators are modeled as first order systems, the moment exerted on the
vehicle cannot be controlled directly. However, by controlling the actuator derivatives, one is essentially controlling the moments derivative. Therefore, the model is
extended with the dynamics8
b = .
M
(5.41)
Combining Eqs. (5.37) and (5.39) to (5.41) and differentiating Eq. (5.32), the complete system can be written as
= K +
+g
des (, )
+g
= f ()
()
()
(5.42)
)
)
+ kM
b
des (, ,
+ kM des
= h(
b (, , )
(5.44)
b =
M
des (, ,
,
b ).
M
M
b
(5.43)
(5.45)
This system is in strict-feedback form and can therefore be controlled using BBS or
,
b
and M
NDI-TSS. To choose a control law, desired dynamics for the states ,
are chosen as the linear functions
des
= K
Q
f
(, )
des ()
= K
h
des
b ) = K M
b,
jM
b (M
Mb
(5.46)
(5.47)
(5.48)
respectively, where K
= diag(Kp , Kq , Kr ), and K M
b =
= diag(K , K , K ), K
diag(Kl , Km , Kn ) have strictly positive diagonal elements and Q , Q
R+ . Then,
is chosen in accordance with Eq. (5.4) as
des (, )
= g()
1 f des
des (, )
(,
)
f
(
)
,
(5.49)
5.4.1. BBS
can be chosen:
From Eq. (5.49), the virtual control M des
b (, , )
Q
des
1 des
des
T
=k
h()
+ (, , )
() ,
M b (, , )
h ()
g
Q
8
(5.50)
Determining the signal to send to the actuators can be decoupled from the design and solved as a
b = M b ( u) for , where M b = M b ()
separate problem, namely solving the equation M
43
M
= jM
b (M
b
QM
b
(5.51)
where QM
b R+ .
Combining the system given by Eqs. (5.42) to (5.45) with the desired dynamics given
by Eqs. (5.46) to (5.48) and the control laws given by Eqs. (5.49) to (5.51), the closed
loop system becomes
= K +
Q + g
= K
()
Q T
b
g
() + kM
= K
Q
b Q kT .
b = K M
M
Mb
QM
b
(5.52)
(5.53)
(5.54)
(5.55)
5.4.2. NDI-TSS
The NDI-TSS controller construction is completed in a similar manner, the differ )
and
ence being that the time derivatives of the desired dynamics, des (, ,
des
(5.56)
M des
b (, , )
des
b ),
= jM
b (M
(5.57)
respectively, in place of Eqs. (5.50) and (5.51). This results in the closed loop system
= K +
+g
= K
()
(5.58)
)
b
des (, ,
+ kM
= K
(5.60)
b
b = K M
M
Mb
des (, ,
,
b ).
M
M
b
(5.59)
(5.61)
CHAPTER
Optimization
In nonlinear control, analytical tools such as Lyapunov theory are used extensively.
The usage of these tools leads to control laws which guarantee stability, but which
often are conservative. This, in turn, implies potentially excessive control loop gains
and unnecessary control action.
The field of mathematical optimization provides a means of analyzing and removing
the shortcomings of relying solely on analytical methods as well as making their
outcomes more applicable to the real world. Many control problems can be posed as
or augmented with the constrained nonlinear optimization problem
min
J()
(6.1)
subject to h() = 0,
where J : X R is the objective function (for engineering applications generally not
convex), X = { Rn : h() = 0} the feasible set, and h : Rn Rm the m equality
constraints (see [5] for an in-depth discussion of the subject).
In general, one is interested in the global minimum. Since the function J is potentially
non-convex, global optimization methods need to be employed. In practice, these are
much more complex than local optimizers. This, coupled with the fact that designs
in aircraft control often depend on many parameters, implies programs which in
the past would have been essentially unsolvable due to their numerical complexity.
However, with the exponential increase in computing power over the last few years,
solving them is no longer unfeasible, even when they depend on many parameters.
This leads to a new dimension in control design which can be used to solve myriad
problems.
45
(6.2)
The agents in the population are not required to be unique. Therefore the population is not a set.
46
Chapter 6. Optimization
T
with
c) Crossover: generate a recombinant vector y = y1 y2 . . . yn
elements
(
x0j
for j {l {1, 2, . . . , n} : rl kCO } {r}
yj =
(6.3)
xi,j otherwise,
j {1, 2, . . . , n}, where r1 , r2 , . . . , rn are chosen randomly from [0, 1], kCO
is the crossover constant chosen from [0, 1] to facilitate diversity, and
r is an integer chosen randomly from {1, 2, . . . , n} to ensure that not all
elements are taken from xi .
d) If J(y) < J(x), then set z i := y. Otherwise set z i := xi .
3. Set xi := z i
i I.
6.2. Application
Other studies, e.g. [14], have shown the efficacy of global optimization methods in
shaping time domain responses. Here, the feasibility of the methodology presented
above is illustrated with three design objectives concerned with inherent system properties, formulated as separate minimizations of (a) the systems L2 gain, (b) actuator
flutter, and (c) the maximum actuator rate. Each minimization is done for both
BBS and NDI-TSS at several points i in the flight envelope using the control laws
given in Section 5.4. The properties are optimized using the DE algorithm with
crossover constant kCO = 0.8, amplification factor F = 0.5, and population size
N = 10. For the remainder of this section, a dot in a subscript implies the expression
is valid for both the BBS and NDI-TSS cases, e.g. f = g implies fBBS = gBBS and
fNDITSS = gNDITSS (but not fBBS = gNDITSS ).
In order for the simulations to be realistic, constraints need to be added. The normalized time domain response of a cmd = 2 pull up (i.e. a pitching down motion
of 2 ) at point i in the flight envelope
yi, (t) =
i, (t) 0
,
cmd
(6.4)
where 0 is the initial (trimmed) value of the angle of attack, is required to lie within
the bounds presented in Fig. 6.1. Furthermore, constraints are added to keep the
elevator from saturating (see Table 3.3 for its limits) and its rate below rmax (see
Section 3.1). In order for the time derivative of the commanded value (which is used
by the controllers) to be finite, the step is replaced with a ramp of duration 0.5 s, so
that
(
t
cmd 0.5
for 0 t < 0.5
cmd (t) = 0 +
(6.5)
cmd
for 0.5 t,
47
6.2. Application
which, normalized, is
u(t) =
cmd (t) 0
.
cmd
(6.6)
The remaining commanded values cmd and cmd in the vector cmd are set to their
trimmed values. The constraints are formulated as equality constraints and collected
in the non-negative function
(t)
e,min
e,i,
XZ T
(6.8)
(6.10)
48
Chapter 6. Optimization
1.4
1.2
1
0.8
0.6
0.4
0.2
0
3
Time [s]
Here, the output y(t) is replaced with the pitching moment m(t) caused by the input
u(t), which in turn is replaced with the normalized commanded angle of attack u(t),
so that the gain gives a measure of how violently the controller performs maneuvers.
Two simplifications are made for a rough numerical approximation of the gain:
1. Only a single input is considered.
2. T is set to 6 s, the maximum time for which the time domain response envelope
is defined (see Fig. 6.1). This limit is sufficient to capture the essential behavior
of the response.
The objective is to find the maximum of the gains with respect to the examined
points i of the flight envelope, i.e.
!1/2
RT
2
0 mi, (t) dt
f, ( ) = max
(6.12)
RT
2
i
0 u(t) dt
2
n
For X R, Ln
2 (X) = {f : X R :
R
X
kf (t)k2 dt < }.
49
6.2. Application
is used as an unconstrained objective function, where mi, (t) is the pitching moment
(implicitly dependent on ) starting at point i. Essentially, one is minimizing the
pitching moment the controller requires for the vehicle to reach the commanded angle
of attack.
Another useful criterion to consider when choosing control loop gains is the movement
of the actuators. For energy saving purposes as well as to minimize wear and tear
on the actuators, one would like to minimize their movement. Here, the sum of the
elevator actuator time derivatives L2 norms
fL2 , ( ) =
X Z
i
1/2
2
|e,i, (t)| dt
(6.13)
(6.14)
(6.15)
2. Actuator flutter for comparing energy usage of the actuators using the objective
function
1
JL2 , ( ) = fL2 , ( ) + h ( )T h ( ).
2
(6.16)
3. Actuator rates for comparing maximum power of the actuators using the objective function
1
J, ( ) = f, ( ) + h ( )T h ( ).
2
(6.17)
50
Chapter 6. Optimization
6.3. Results
The results of the three optimizations described above are shown in Tables 6.1 to 6.3.
For each controller, several generations are listed with the function values given by the
optimal agents in each respective population, as well as the dispersion of the agents
normalized by the bounds of the parameters. The dispersion is a useful measure to
consider when determining when to terminate the optimization. The less dispersed
the agents become, the more certain one can be about having found a local minimum
(that may also is a global one).
All of the BBS optimizations are done with completely random initial populations (as
is evident by thein comparison to the final valuesvery large function values in the
early generations, caused by penalty terms), and the optimizer manages to find the
feasible set in each minimization. In the case of the NDI-TSS optimizations, however,
the optimizer has difficulty locating the feasible set. Therefore, the optimizations are
restarted with one agent in the initial population in the feasible set. The remaining
agents are, as in the case of the BBS optimizations, chosen randomly.
Generation
Dispersion
BBS
1
50
100
500
1000
5000
5.422 1010
3.386 102
3.358 102
3.349 102
3.349 102
3.349 102
6 101
1 101
3 102
2 103
8 104
9 1017
NDI-TSS
1
50
100
500
1000
5000
4.799 102
7.537 101
7.462 101
7.459 101
7.459 101
7.459 101
6 101
2 101
3 102
2 106
5 107
6 1017
The results from the L2 gain optimizations are summarized in Table 6.1. The NDITSS controller reaches a smaller value than the BBS controller, suggesting it in
general uses less pitching moment to reach the values it is given as commands. The
differences in behavior of the two optimized controllers can be seen visually in Fig. 6.2.
The optimal BBS controller is much faster than the optimal sluggish but economical
NDI-TSS controller. The trade-off for this speed is more elevator deflection, hence
more pitching moment exertion.
As in the L2 gain optimizations, the NDI-TSS controller is capable of a smaller L2
norm, in this case measuring the actuator energy usage, than the BBS controller. The
51
6.4. Conclusion
Generation
Dispersion
BBS
1
50
100
500
1000
5000
1.238 107
2.804 101
2.674 101
2.533 101
2.513 101
2.512 101
7 101
3 101
3 101
3 104
1 106
6 109
NDI-TSS
1
50
100
500
1000
5000
3.586 101
5.013
4.848
4.820
4.820
4.820
6 101
2 101
7 102
1 102
1 103
2 1017
results of the optimizations are shown in Table 6.2. The superiority of the optimal
NDI-TSS controller is visually apparent in Fig. 6.3 by the movement of the elevator.
It keeps the elevator fairly constant throughout the maneuver, whereas the optimal
BBS controller outputs some relatively large deflections in the transient region. Just
like the controllers optimal in L2 gain, the BBS controller is fast and the NDI-TSS
controller sluggish.
Finally, the uniform norm optimizations, i.e. power minimizations, are summarized
in Table 6.3. As in the previous two cases, NDI-TSS is more conservative. Unlike
the other two optimizations, one can see from the response in Fig. 6.4 that the
BBS controller is slower and has a larger overshoot. Also, the elevator movement is
smoother, which is what one would expect when minimizing elevator power.
In all three cases, the respective optimal NDI-TSS controllers are better in the
sense that the their optimizations reach lower objective function values than their
optimal BBS counterparts.
6.4. Conclusion
As can be seen in the responses of the optimal controllers, the NDI-TSS controllers
are very sluggish and seem to not even reach steady state. A fairer comparison
between BBS and NDI-TSS could be made if the time domain boundaries were made
stricter and if the objective function also were a function of the deviation of the
response from a desired one, such as a typical second order response. Furthermore,
more extreme maneuvers could be investigated. These investigations may reveal
BBS to be the superior control method of the two. In addition to these changes, the
52
Chapter 6. Optimization
Generation
Dispersion
BBS
1
50
100
500
1000
5000
5.423 107
3.509 101
3.355 101
1.160 101
1.157 101
1.156 101
6 101
1 101
2 101
2 105
1 106
1 1016
NDI-TSS
1
50
100
500
1000
5000
4.217 101
4.330
4.305
4.290
4.290
4.290
6 101
1 101
2 101
7 104
8 1010
8 1010
objective functions could be made more accurate. For example, the L2 gain can be
better approximated by testing more inputs, and the norms examined by testing more
initial conditions, i.e. performing maneuvers in more points of the flight envelope.
Global optimization using genetic algorithms proved a useful tool not only in choosing
parameters for BBS and NDI-TSS controllers, but also in comparing them. This
method of design allows for the efficient choice of system parameters for a host of
objectives, e.g. minimization of energy loss and actuator movement. These results
suggest that global optimization in nonlinear control is a viable and promising design
methodology.
replacemen
6.4. Conclusion
53
1.2
1
0.8
0.6
BBS
NDI
Command
Boundaries
0.4
0.2
0
3
Time [s]
3
Time [s]
EI
0
-5
-10
Figure 6.2.: Time domain responses and elevator deflections using BBS and NDI-TSS controllers with parameters optimized for minimum moment exertion.
replacemen
54
Chapter 6. Optimization
1.2
1
0.8
0.6
BBS
NDI
Command
Boundaries
0.4
0.2
0
3
Time [s]
3
Time [s]
EI
0
-5
-10
Figure 6.3.: Time domain responses and elevator deflections using BBS and NDI-TSS controllers with parameters optimized for actuator energy conservation.
replacemen
6.4. Conclusion
55
1.2
1
0.8
0.6
BBS
NDI
Command
Boundaries
0.4
0.2
0
3
Time [s]
3
Time [s]
EI
0
-5
-10
Figure 6.4.: Time domain responses and elevator deflections using BBS and NDI-TSS controllers with parameters optimized for actuator power conservation.
APPENDIX
F tot
0
= F aero + T eb 0 ,
mg
(A.1)
(A.2)
57
Similarly, using Eqs. (2.1), (2.7), and (2.8) and the above definition of F tot , the
sideslip dynamics is:
v
d
arcsin
=
dt
V
1
V v v V
=q
2
V2
1 v
V2
1
v
v V
V
V 2 v2
1
v T
=p
v
v v
V2
V 2 (1 sin2 )
sec
v
0 1 0 2 v T v
=
V
V
T
sec
0 1 0 sin cos cos sin sin cos
=
v
V
1
cos sin sec tan sin sin sin v
=
V
F tot
1
cos sin cos sin sin
=
v
V
m
1
cos sin cos sin sin F tot
=
mV
(A.3)
The combination of Eqs. (A.2) and (A.3) along with Eq. (2.3) leads to the relation
= f (, , , ) + g(, , ),
(A.4)
where
sec 0
1
sin
0
cos
0
1
f (, , , ) =
F tot
cos sin cos sin sin
mV
0
0
cos tan
1
sin tan
sin
0
cos .
g(, , ) =
1
sin tan cos tan
(A.5)
(A.6)
58
Proposition 1. The virtual control gain matrix g() given in Eq. (5.35) is invertible
for all satisfying || 45 and all , [30 , 30 ].
Proof. Assume g() is singular for some satisfying || 45 . Then its determinant
det g() = cos tan sin tan cos
sin (cos tan + sin tan sin tan ) cos
= cos2 tan sin tan sin2 sin tan tan
(A.7)
=
+
3 3
3
= 7/3,
i.e.
(A.8)
(A.9)
3 7
|| arctan
> 45 .
(A.10)
7
This is an obvious contradiction of the initial assumption and thus completes the
proof.
Bibliography
[1] J
urgen Ackermann. Robust Control. Springer Verlag, 1997.
[2] Anonymous. Military specification flying qualities of piloted airplanes. Specification, Department of Defense, United State of America, November 1980.
MIL-F-8785C.
[3] Anonymous. Department of defense handbook flying qualities of piloted aircraft. Handbook, Department of Defense, United States of America, December
1997. MIL-HDBK-1797.
[4] Anonymous. The space launch initiative: Technology to pioneer the space frontier. Technical Report FS-2002-04-87-MSFC, NASA Marshall Space Flight Center, April 2002.
[5] Dimitri P. Bertsekas. Nonlinear Programming. Athena Scientific, New Hampshire, second edition, 1999.
[6] Ola H
arkeg
ard. Flight Control Design Using Backstepping. Licentiate thesis
no. 875, Department of Electrical Engineering, Linkoping University, SE-581 83
Linkoping, Sweden, March 2001.
[7] Daigoro Ito, Jennifer Georgie, John Valasek, and Donald T. Ward. Reentry
vehicle flight controls design guidelines: Dynamic inversion. Technical report,
NASA, March 2002.
[8] Hassan K. Khalil. Nonlinear Systems. Prentice Hall, third edition, 2002.
[9] Petar V. Kokotovic. The joy of feedback: Nonlinear and adaptive. Control
Systems, IEEE, 12(3):717, June 1992.
[10] Warren F. Phillips. Mechanics of Flight. John Wiley & Sons, Inc., second edition,
2010.
[11] John W.C. Robinson. Block backstepping for nonlinear flight control law design.
In D.G. Bates and M. Hagstr
om, editors, Nonlinear Analysis and Synthesis
Techniques for Aircraft Control, pages 231257. Springer, Berlin, 2007. Lecture
Notes in Control and Information Sciences Series, Vol. 365.
60
Bibliography
[12] D. Saussie, L. Saydy, and O. Akhrif. Longitudinal flight control design with handling quality requirements. Aeronautical Journal, 110(1111):627637, September
2006. Paper No. 2989.
[13] Corey J. Schumacher, Pramod P. Khargonekar, and N. Harris McClamroch.
Stability analysis of dynamic inversion controllers using time-scale separation.
In Proceedings of the AIAA Guidance, Navigation, and Control Conference and
Exhibit, Boston, Massachusetts, August 1998. AIAA paper 1998-4322.
[14] Marc L. Steinberg and Anthony B. Page. Nonlinear adaptive flight control
with genetic algorithm design optimization. International Journal of Robust
and Nonlinear Control, 9:10971115, December 1999.
[15] Brian L. Stevens and Frank L. Lewis. Aircraft Control and Simulation. John
Wiley & Sons, Inc., second edition, 2003.
[16] Rainer Storn and Kenneth Price. Differential evolution a simple and efficient
heuristic for global optimization over continuous spaces. Journal of Global Optimization, 11:341359, December 1997.
[17] Johan Thunberg and John W.C. Robinson. Block backstepping, NDI and related
cascade designs for efficient development of nonlinear flight control laws. In
Proceedings of the AIAA Guidance, Navigation and Control Conference and
Exhibit, Honolulu, Hawaii, August 2008. AIAA paper AIAA 2008-6960.
[18] Harold N. Tobie, Elden M. Elliott, and Lawrence G. Malcom. A new longitudinal
handling qualities criterion. In Proceedings of the National Aerospace Electronics
Conference, pages 9399, Dayton, Ohio, May 1966.
[19] A.J. van der Schaft. L2 -gain analysis of nonlinear systems and nonlinear state
feedback H control. IEEE Transactions on Automatic Control, 37(6):770784,
1992.
Examensarbete E357 i
Optimeringslra och systemteori
September 2011
www.math.kth.se/optsyst/