Você está na página 1de 202

UNIVERSIDADE TECNICA

DE LISBOA

INSTITUTO SUPERIOR TECNICO

MODELING AND NONLINEAR CONTROL


FOR AIRSHIP AUTONOMOUS FLIGHT

Alexandra Bento Moutinho


(Mestre em Engenharia Mecanica)

Dissertacao para a obtencao do


Grau de Doutor em Engenharia Mecanica

Orientador:

Doutor Jose Raul Carreira Azinheira

J
uri:
Presidente:
Vogais:

Reitor da Universidade Tecnica de Lisboa


Doutor Joao Manuel Lage de Miranda Lemos
Doutor Jose Manuel Gutierrez Sa da Costa
Doutor Felix Mora-Camino
Doutor Jorge Manuel Miranda Dias
Doutor Pedro Manuel Goncalves Lourtie
Doutor Jose Raul Carreira Azinheira

Dezembro de 2007

To all, and to a few in particular.

Resumo
O trabalho desenvolvido e apresentado nesta tese foca o projecto, validacao
e comparacao de diferentes solucoes de controlo nao-linear que permitam um
dirigvel navegar autonomamente. De modo a atingir esta meta, e desenvolvido um modelo nao-linear de seis graus de liberdade do dirigvel baseado
nas equacoes de Lagrange, reproduzindo a resposta do dirigvel a entradas
dos actuadores e a perturbacoes de vento. A linearizacao deste modelo para
diferentes condicoes de equilbrio resulta no conhecido desacoplamento dos
movimentos longitudinal e lateral, e permite uma analise exaustiva do problema de projecto de controlo de dirigveis em todo o envelope de voo. Existem
entao condicoes para propor solucoes alternativas de controlo nao-linear de
modo a obter uma u
nica lei de controlo, valida para diferentes missoes, independente das condicoes de voo, e robusta a perturbacoes realistas de vento.
As metodologias de controlo desenvolvidas neste trabalho com vista ao voo
autonomo de dirigveis sao o Escalonamento de Ganho, a Dinamica Inversa
e o Backstepping. Alem da analise de problemas especficos inerentes ao projecto e implementacao de cada um dos controladores, sao tambem definidos
criterios desejados de desempenho, permitindo a comparacao das diferentes
solucoes. Esta analise, baseada em resultados de simulacao para missoes de
voo completas definidas desde a descolagem ate `a aterragem, e considerando
perturbacoes de vento realistas, e importante de modo a estabelecer a viabilidade da implementacao dos controladores a bordo da plataforma experimental
do dirigvel. Esta tese e parte da investigacao feita na area de controlo de
voo nao-linear de dirigveis nos projectos AURORA e DIVA do Instituto de
Engenharia Mecanica (IDMEC) do Instituto Superior Tecnico, Universidade
Tecnica de Lisboa.
Keywords: Controlo nao-linear, Escalonamento de ganho, Dinamica inversa,
Backstepping, Dirigvel, Modelacao, Controlo de voo.

ii

Abstract
The work developed and presented in this thesis focuses on the design, validation and comparison of different nonlinear control solutions allowing an
airship to navigate autonomously. To accomplish this task, a six-degrees-offreedom nonlinear model of the airship is developed based on the Lagrangian
equations, reproducing the airship response to actuator and wind disturbances
inputs. The linearization of this model for trim conditions over the flight envelope results in the known decoupling of the longitudinal and lateral motions,
and allows a thorough analysis of the airship control design problem over the
entire aerodynamic range. The conditions are then set to propose alternative
nonlinear control solutions so as to have a single control law valid for different
missions, independent of the flight region, and robust to realistic wind disturbances. The control methodologies developed in this work for the airship
autonomous flight are Gain Scheduling, Dynamic Inversion and Backstepping.
Besides the analysis of specific problems inherent to the design and implementation of each controller, desired performance criteria are also defined, allowing
the comparison of the different solutions. This assessment, based on simulation results for complete flight missions defined from take-off to landing, and
considering realistic wind disturbances, is important in order to establish the
viability of the controllers implementation onboard the experimental airship
platform. This thesis is part of the research made in the area of nonlinear
flight control of airships for the AURORA and DIVA projects of the Institute
of Mechanical Engineering (IDMEC) in Instituto Superior Tecnico, Technical
University of Lisbon.
Keywords: Nonlinear control, Gain Scheduling, Dynamic inversion, Backstepping, Airship, Modeling, Flight Control.

iii

iv

Acknowledgments
My first words of appreciation are undoubtedly to my supervisor, Professor
Jose Raul Azinheira. In his words, A PhD thesis is not intended to close
doors, but to open some more. His broad knowledge on subjects like modeling,
control, aerodynamics and instrumentation certainly opened a lot of doors for
me, being at the basis of clarifying and motivating discussions for this work.
I also want to thank Doctor Ely Carneiro de Paiva and Doctor Samuel Siqueira
Bueno of the AURORA project in CenPRA Brazil, one of the leading projects
in autonomous airships research, for the joint work developed even from such
a distance. A more recent project in this area is the Portuguese DIVA project,
whose team I also want to thank for the insight provided in the different aspects
of building an airship.
My appreciation goes also to my colleagues at GCAR, namely Miguel Pedro
Silva for always being available to discuss some of my doubts, and Mario
Mendes for all the technical support provided.
Last but definitely not least, I want to thank Carlos, my family and my friends
for being there. Bem hajam!

vi

Contents
Resumo

Abstract

iii

Acknowledgments

List of Figures

xi

List of Tables

xv

Notation

xvii

Acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii
Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviii
Variables description . . . . . . . . . . . . . . . . . . . . . . . . . . . xix
1 Introduction

1.1

Airships and their history . . . . . . . . . . . . . . . . . . . . .

1.2

The motivation - airship applications . . . . . . . . . . . . . . .

1.3

The control of airships flight . . . . . . . . . . . . . . . . . . . .

1.4

Structure of this work . . . . . . . . . . . . . . . . . . . . . . . 11

1.5

Contributions of this thesis . . . . . . . . . . . . . . . . . . . . . 13

2 The Airship Model

17

2.1

Airship platform . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.2

Airship equations of motion . . . . . . . . . . . . . . . . . . . . 19


2.2.1

Airship dynamics . . . . . . . . . . . . . . . . . . . . . . 20

2.2.2

Airship kinematics . . . . . . . . . . . . . . . . . . . . . 32

2.2.3

Airship simulator . . . . . . . . . . . . . . . . . . . . . . 34
vii

viii

CONTENTS

2.3

2.4

Airship linearized models . . . . . . . . . . . . . . . . . . . . . . 34


2.3.1

Trim or equilibrium conditions . . . . . . . . . . . . . . . 35

2.3.2

Model linearization . . . . . . . . . . . . . . . . . . . . . 38

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3 Common Concepts and Tools


3.1

45

Position errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.1.1

Path-following . . . . . . . . . . . . . . . . . . . . . . . . 46

3.1.2

Path-tracking . . . . . . . . . . . . . . . . . . . . . . . . 46

3.2

Attitude reference and wind estimation . . . . . . . . . . . . . . 47

3.3

Controllers performance evaluation . . . . . . . . . . . . . . . . 49


3.3.1

Case-study mission . . . . . . . . . . . . . . . . . . . . . 50

3.3.2

Sensitivity and robustness . . . . . . . . . . . . . . . . . 51

4 Classical Approach: Linear Control

53

4.1

Airspeed and altitude regulation model . . . . . . . . . . . . . . 55

4.2

Lateral models . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.2.1

No-roll approximation . . . . . . . . . . . . . . . . . . . 56

4.2.2

Space domain approximation

. . . . . . . . . . . . . . . 57

4.3

Linear Quadratic Regulator . . . . . . . . . . . . . . . . . . . . 58

4.4

Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . 60

4.5

4.4.1

Airspeed and altitude regulation model . . . . . . . . . . 60

4.4.2

No-roll vs. space domain . . . . . . . . . . . . . . . . . . 62

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

5 Gain Scheduling
5.1

67

More linear models . . . . . . . . . . . . . . . . . . . . . . . . . 68


5.1.1

Groundspeed and altitude regulation . . . . . . . . . . . 68

5.1.2

Complete 12-states linear model . . . . . . . . . . . . . . 69

5.2

Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

5.3

Robustness analysis . . . . . . . . . . . . . . . . . . . . . . . . . 75

5.4

5.3.1

Performance robustness

. . . . . . . . . . . . . . . . . . 75

5.3.2

Stability robustness . . . . . . . . . . . . . . . . . . . . . 80

Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.4.1

Groundspeed and altitude regulation . . . . . . . . . . . 87

CONTENTS

ix

5.4.2

Case-study mission . . . . . . . . . . . . . . . . . . . . . 89

5.4.3

Sensitivity and robustness . . . . . . . . . . . . . . . . . 93

5.5

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

6 Dynamic Inversion
6.1

99

General theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 101


6.1.1

Local coordinates transformation . . . . . . . . . . . . . 102

6.1.2

Exact linearization via feedback . . . . . . . . . . . . . . 105

6.1.3

Asymptotic output tracking . . . . . . . . . . . . . . . . 107

6.2

New formulation for cascaded systems

6.3

Application to airship path-tracking problem . . . . . . . . . . . 111

6.4

Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . 114

6.5

. . . . . . . . . . . . . . 109

6.4.1

Case-study mission . . . . . . . . . . . . . . . . . . . . . 114

6.4.2

Sensitivity and robustness . . . . . . . . . . . . . . . . . 118

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

7 Backstepping

125

7.1

Wind estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

7.2

Backstepping design approach . . . . . . . . . . . . . . . . . . . 128

7.3

Application to the path-tracking problem . . . . . . . . . . . . . 128

7.4

Control design with saturation constraints . . . . . . . . . . . . 132

7.5

Control implementation . . . . . . . . . . . . . . . . . . . . . . 136


7.5.1

7.6

7.7

Adapted control law to deal with underactuation . . . . 136

Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . 138


7.6.1

Case-study mission . . . . . . . . . . . . . . . . . . . . . 138

7.6.2

Sensitivity and robustness . . . . . . . . . . . . . . . . . 142

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

8 Comparison of controllers performance

149

8.1

Performance for case-study mission . . . . . . . . . . . . . . . . 149

8.2

Sensitivity test results . . . . . . . . . . . . . . . . . . . . . . . 152

8.3

Computational effort . . . . . . . . . . . . . . . . . . . . . . . . 154

8.4

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

9 Conclusions and Future Work

157

CONTENTS

A Referentials

161

A.1 Frames definition . . . . . . . . . . . . . . . . . . . . . . . . . . 161


A.1.1 Earth-Centered Inertial (ECI) frame . . . . . . . . . . . 161
A.1.2 North-East-Down (NED) or {i} frame . . . . . . . . . . 162
A.1.3 Aircraft-Body Centered (ABC) or {l} frame . . . . . . . 162

A.1.4 Aerodynamic or {a} frame . . . . . . . . . . . . . . . . . 163

A.2 Changing frame . . . . . . . . . . . . . . . . . . . . . . . . . . . 163


B Dryden Model For Continuous Gust

165

C Differential geometry and topology

167

C.1 Lie derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168


C.2 Diffeomorphisms and state transformations . . . . . . . . . . . . 169
Bibliography

171

List of Figures
1.1

The history of airships. . . . . . . . . . . . . . . . . . . . . . . .

1.2

Manned airships. . . . . . . . . . . . . . . . . . . . . . . . . . .

1.3

Unmanned airships. . . . . . . . . . . . . . . . . . . . . . . . . .

2.1

The AURORA airship. . . . . . . . . . . . . . . . . . . . . . . . 18

2.2

AURORA airship sensors and actuators. . . . . . . . . . . . . . 19

2.3

Simulator block diagram for airship open-loop model. . . . . . . 34

2.4

Trim values of state and control input. . . . . . . . . . . . . . . 37

2.5

Poles of linearized longitudinal dynamics vs. airspeed. . . . . . . 40

2.6

Poles of linearized lateral dynamics vs. airspeed . . . . . . . . . 42

3.1

Path-following errors definition. . . . . . . . . . . . . . . . . . . 46

3.2

Path-tracking errors definition. . . . . . . . . . . . . . . . . . . 47

3.3

Wind and yaw reference estimation. . . . . . . . . . . . . . . . . 48

3.4

Case-study mission reference. . . . . . . . . . . . . . . . . . . . 50

4.1

Linear control block diagrams. . . . . . . . . . . . . . . . . . . . 60

4.2

Trajectory and altitude for airspeed and altitude regulation. . . . 61

4.3

Longitudinal groundspeed and airspeed for airspeed and altitude


regulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

4.4

Control action for airspeed and altitude regulation. . . . . . . . . 62

4.5

Trajectory, lateral error and yaw angle for lateral control comparison. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4.6

Sideslip angle, rudder deflection and roll angle for lateral control
comparison. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5.1

Poles of linearized dynamics vs. airspeed. . . . . . . . . . . . . . 71

5.2

Evolution of B matrix coefficients with airspeed. . . . . . . . . . 73


xi

xii

LIST OF FIGURES

5.3

Evolution of B matrix Tx and Tz coefficients with airspeed. . . . 74

5.4

Gain scheduling diagram block. . . . . . . . . . . . . . . . . . . . 74

5.5

Closed-loop nominal system. . . . . . . . . . . . . . . . . . . . . 75

5.6

Disturbed feedback control system. . . . . . . . . . . . . . . . . . 76

5.7

Singular values relations. . . . . . . . . . . . . . . . . . . . . . . 79

5.8

Frequency analysis of the MIMO nominal system. . . . . . . . . 79

5.9

Frequency-domain performance specifications - disturbance rejection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

5.10 Robust stability analysis of the uncertain systems. . . . . . . . . 83


5.11 Stability robustness to plant parameter variation. . . . . . . . . . 86
5.12 Airship position coordinates and errors. . . . . . . . . . . . . . . 88
5.13 Airship ground velocity components and aerodynamic variables. . 88
5.14 Airship actuators input. . . . . . . . . . . . . . . . . . . . . . . 89
5.15 Airship position coordinates and errors. . . . . . . . . . . . . . . 90
5.16 Airship north-east position and attitude. . . . . . . . . . . . . . 91
5.17 Airship ground velocity components and aerodynamic variables. . 92
5.18 Airship actuators input. . . . . . . . . . . . . . . . . . . . . . . 92
5.19 Airship north-east trajectory and attitude, and wind attitude. . . 93
5.20 Airship position errors and aerodynamic variables. . . . . . . . . 94
5.21 Airship actuators input. . . . . . . . . . . . . . . . . . . . . . . 95
6.1

Normal form representation, with no internal dynamics. . . . . . 105

6.2

Closed-loop system, with new reference input . . . . . . . . . . 107

6.3

Closed-loop system, with model reference input . . . . . . . . . 109

6.4

Airship position coordinates and errors. . . . . . . . . . . . . . . 115

6.5

Airship north-east position and attitude. . . . . . . . . . . . . . 116

6.6

Airship ground velocity components and aerodynamic variables. . 116

6.7

Airship actuators input. . . . . . . . . . . . . . . . . . . . . . . 117

6.8

Airship north-east trajectory and attitude, and wind attitude. . . 119

6.9

Airship position errors and aerodynamic variables. . . . . . . . . 120

6.10 Airship actuators input. . . . . . . . . . . . . . . . . . . . . . . 120


7.1

Air velocity reference estimation (2D). . . . . . . . . . . . . . . 129

7.2

Airship position coordinates and errors. . . . . . . . . . . . . . . 139

7.3

Airship north-east position and attitude. . . . . . . . . . . . . . 140

LIST OF FIGURES

xiii

7.4

Airship ground velocity components and aerodynamic variables. . 140

7.5

Airship actuators input. . . . . . . . . . . . . . . . . . . . . . . 141

7.6

Airship north-east trajectory and attitude, and wind attitude. . . 142

7.7

Airship position errors and aerodynamic variables. . . . . . . . . 143

7.8

Airship actuators input. . . . . . . . . . . . . . . . . . . . . . . 144

8.1

Comparison of airship 3D trajectories. . . . . . . . . . . . . . . 150

8.2

Comparison of position errors. . . . . . . . . . . . . . . . . . . . 151

8.3

Comparison of north-east trajectories with airship heading. . . . 151

8.4

Comparison of actuators request . . . . . . . . . . . . . . . . . . 153

A.1 Relationship between the different coordinate systems. . . . . . . 162


A.2 ABC and wind frames. . . . . . . . . . . . . . . . . . . . . . . . 163
B.1 Block diagram for gust generator. . . . . . . . . . . . . . . . . . 165

xiv

LIST OF FIGURES

List of Tables
5.1

Robustness tests on model parameters. . . . . . . . . . . . . . . 96

6.1

Robustness tests on model parameters. . . . . . . . . . . . . . . 122

7.1

Robustness tests on model parameters. . . . . . . . . . . . . . . 145

8.1

Comparison of baseline results.

8.2

Computational effort comparison. . . . . . . . . . . . . . . . . . 155

8.3

Overall controllers comparison. . . . . . . . . . . . . . . . . . . 155

xv

. . . . . . . . . . . . . . . . . . 154

xvi

LIST OF TABLES

Notation
Acronyms
ABC
AF
AURORA
CB
CG
CV
DIVA
ECI
GPS
HF
IMU
LQR
LTA
LTI
MIMO
NED
PP
RMS
UAV

Airship Body Centered


Aerodynamic Flight
Autonomous Unmanned Remote mOnitoring Robotic Airship
Center of Buoyancy
Center of Gravity
Center of Volume
Dirigvel Instrumentado para Vigilancia Aerea
Earth-Centered Inertial
Global Positioning System
Hover Flight
Inertial Measurement Unit
Linear Quadratic Regulator
Lighter-Than-Air
Linear Time-Invariant
Multi-Input / Multi-Output
North-East-Down
Pole Placement
Root Mean Square
Unmanned Aerial Vehicle

xvii

xviii

NOTATION

Nomenclature
Typeface
italic
bold

scalar variables
vector or matrix variables

Subscripts
a
B
e
h
v
w

apparent for masses; airspeed for velocities


buoyancy
equilibrium condition
horizontal plane (lateral mode)
vertical plane (longitudinal mode); virtual for masses
wind

Superscripts
c
o

regarding the CG
regarding the CV

Operations

cross-product

Others
. variation relative to equilibrium condition or reference

xix

NOTATION

Variables description
Symbol

Domain

Unit

Definition

0i

a
e
r
v

Rii
R
R
R
R
R
R
R
R
R
R
R31

none
rad
rad
m
m
rad
rad
rad
rad
m
rad
rad

R
R
R33
R31

rad
rad
rad/s
rad/s

6
ag
A
B
C
C3
d
h
Ii
J
J
JB

R66
R31
Rnn
Rnm
R31
R33
R61
R
Rii
R33
R66
R33

rad/s
m/s2

m
m

m
none
kg.m2
none
kg.m2

Jv
m
Ma

R33
R
R66

kg.m2
kg

mB
MB

R
R66

kg

MBa

R66

zero matrix
angle of attack
sideslip angle
vertical position error
lateral position error
aileron deflection
elevator deflection
rudder deflection
main propellers vectoring angle
longitudinal position error
pitch angle
angular position vector relative to {i} frame
with Euler angles, [, , ]T
roll angle
yaw angle
cross-product matrix equivalent to
angular velocity and components in {l} frame,
[p, q, r]T
see equation (2.45)
inertial gravity acceleration vector, [0, 0, g]T
n-state dynamic matrix
m-input matrix
CG, center of mass of the airship
cross-product matrix equivalent to OC
disturbance vector
altitude, pD
identity matrix
inertia matrix of the airship
see equation (2.77)
inertia matrix of the buoyancy air, diagonal
matrix
virtual inertia matrix, diagonal matrix
airship mass
generalized apparent mass matrix of the airship
with masses and inertias, Mo + Mv
buoyancy mass
generalized inertial mass matrix of the buoyancy
air with masses and inertias, diag(mB I3 , JB )
generalized apparent mass matrix of the buoyancy
air with masses and inertias, MB + Mv

xx

NOTATION

Symbol

Domain

Unit

Mo

R66

Mv
Mv

R33
R66

mw
O
OC
P
p

R
R31
R31
R61
R31

p
pN or N
pE or E
pD or D
p

R
R
R
R
R31

q
r
R
S
TD
TL
TR
Tx
Ty
Tz
u
u

u
ue
V
v
v
V3
V6
Vt
w
W
x

x
xe
XT

R
R
R33
R33
R
R
R
R
R
R
R
R71
R71
R71
R61
R31
R
R33
R66
R
R
R
R121
R121
R121
R

Definition

generalized mass matrix of the airship with masses


and inertias, see equation (2.11)
kg
virtual mass matrix

generalized virtual mass matrix with masses and


inertias, diag(Mv , Jv )
kg
weighting mass, m mB
m
center of buoyancy CB = CV, origin of {l} frame
m
vector from CV to CG, C O = [ax , 0, az ]T

position vector, [pT , T ]T


m
cartesian position vector and components in {i}
frame, [pN , pE , pD ]T
rad/s roll rate in {l} frame
m
position north
m
position east
m
position down
m/s
linear velocity and components in {i} frame,
[pN , pE , pD ]T
rad/s pitch rate in {l} frame
rad/s yaw rate in {l} frame
none see equation (2.18)
none see equation (2.17)
N
engines differential thrust, TL TR
N
left engine thrust
N
right engine thrust
N
engines longitudinal thrust
N
tail motor thrust
N
engines vertical thrust
m/s
forward speed in {l} frame

input vector, [e , TL , TR , v , a , r , Ty ]T

perturbation input vector, u ue

equilibrium input vector, [ee , TLe , TRe , ve , ae , re , Tye ]T

velocity vector, [vT , T ]T


m/s
linear velocity and components in {l} frame, [u, v, w]T
m/s
lateral speed in {l} frame
m/s
cross-product matrix equivalent to v
m/s
see equation (2.45)
p
m/s
true airspeed, u2a + va2 + wa2
m/s
vertical speed in {l} frame
J
kinetic energy

state vector, [vT , T , pT , T ]T

perturbation state vector, x xe

equilibrium state vector, [veT , Te , pTe , Te ]T


N
engines total thrust, TL + TR

Chapter 1
Introduction
Contents
1.1

Airships and their history . . . . . . . . . . . . . .

1.2

The motivation - airship applications . . . . . . .

1.3

The control of airships flight . . . . . . . . . . . . .

1.4

Structure of this work . . . . . . . . . . . . . . . . .

11

1.5

Contributions of this thesis . . . . . . . . . . . . .

13

H
a um ditado que ensina o genio e uma grande paciencia;
sem pretender ser genio, teimei em ser um grande paciente. As
invencoes s
ao, sobretudo, o resultado de um trabalho teimoso, em
que n
ao deve haver lugar para o esmorecimento.
Alberto Santos-Dumont, 19181

1.1

Airships and their history

An airship or dirigible is a buoyant lighter-than-air aircraft that can be steered


and propelled through the air. Airships were the first aircraft to do controlled,
powered flight, being widely used prior to the 1940s.
Airships were developed from the free balloon. Early balloons were not truly
navigable. Attempts to improve maneuverability included elongating the balloon shape and using a powered screw to push it through the air. Credit for the
1

There is a saying that teaches the genius is a great patience; not intending to be a
genius, I insisted on being very patient. The inventions are, most of all, the result of a
persistent work, where no wilting shall take place. Alberto Santos-Dumont, 1918

CHAPTER 1. INTRODUCTION

construction of the first navigable full-sized airship goes to French engineer,


Henri Giffard, who, in 1852, attached a small, steam-powered engine to a huge
propeller and chugged through the air for seventeen miles at a top speed of
five miles per hour (fig. 1.1(a)).
The first airship to demonstrate its ability to return to its starting place in
a light wind was La France (fig. 1.1(b)), developed in 1884 by the French
inventors Charles Renard and Arthur Krebs. It was propelled by an electrically
driven propeller. However, it was not until the invention of the gasolinepowered engine in 1896 that practical airships could be built.
Count Ferdinand von Zeppelin, the German inventor, completed his first airship in 1900; this airship had a rigid frame and served as the prototype of
many subsequent models. The first Zeppelin airship consisted of a row of 17
gas cells individually covered in rubberized cloth; the whole was confined in a
cylindrical framework covered with smooth-surfaced cotton cloth. It was about
128m long and 12m in diameter; the hydrogen-gas capacity totalled 11.3 million liters. The ship was steered by rudders fore and aft and was driven by two
11kW Daimler internal-combustion engines, each turning two propellers. Passengers, crew, and engine were carried in two aluminium gondolas suspended
forward and aft. At its first trial, on July 2, 1900, the airship carried five
people (fig. 1.1(c)); it attained an altitude of 396m and flew a distance of 6km
in 17 minutes.
In 1898, the Brazilian Alberto Santos-Dumont was the first to construct and fly
a gasoline-powered airship, developing a series of 14 airships in France. In his
No. 6, he circled the Eiffel Tower in 1901 (fig. 1.1(d)). The American inventor
Thomas S. Baldwin built a 53-foot airship, the California Arrow, winning a
one-mile race in October 1904. Walter Wellman failed in an effort to cross the
Atlantic Ocean in an airship in 1910, after five unsuccessful attempts to reach
the North Pole. Although many successful flights were made before 1910, the
best engine available for use in the early airship was too heavy in proportion
to its power.
At the beginning of World War I, Germany had ten zeppelins. By 1918, 67
zeppelins had been constructed, and 16 survived the war. Those not captured
were surrendered to the Allies by the terms of the Treaty of Versailles in
1919. Airships were operated in a number of nations between the two world
wars. The major operators of rigid airships were Britain, the United States
and Germany, with Italy and France also operating a few. Italy, the Soviet

1.1. AIRSHIPS AND THEIR HISTORY

Union, the United States and Japan operated semi-rigid airships, while blimps
(nonrigid airships) were operated in many nations.
The war, however, disclosed the vulnerability of airships to aeroplane attack,
and caused the abandonment of the dirigible for offensive military purposes.
The United States was the only power to use airships during World War II,
and the airships played a small but important role. The Navy used them
for minesweeping, search and rescue, photographic reconnaissance, scouting,
escorting convoys, and antisubmarine patrols. Airships accompanied many
oceangoing ships, both military and civilian. Of the 89.000 ships escorted by
airships during the war, not one was lost to enemy action. The Akron and
Macon were two rigid airships built in the United States for the U.S. Navy.
They were the only airships that could launch and retrieve planes in midair
(fig. 1.1(e)).
When the various restrictions imposed by the Treaty of Versailles on Germany
were lifted, Germany was again allowed to construct airships. It built three
giant rigid airships: the LZ-127 Graf Zeppelin, LZ-l29 Hindenburg, and LZ-l30
Graf Zeppelin II. The Graf Zeppelin is considered the finest airship ever built.
It flew more miles than any airship had done to that time or would in the future.
Its first flight was on September 18, 1928. In August 1929, it circled the globe.
Its flight began with a trip from Friedrichshaften, Germany, to Lakehurst, New
Jersey, stopping only at Tokyo, Japan, Los Angeles, California, and Lakehurst.
The trip took twelve days, less time than the ocean trip from Tokyo to San
Francisco. During the ten years the Graf Zeppelin flew, it made 590 flights
including 144 ocean crossings. It flew more than 1.609.344km, visited the
United States, the Arctic, the Middle East, and South America, and carried
13.110 passengers.
When the Hindenburg was built in 1936, Zeppelins had been accepted as a
quicker and less expensive way to travel long distances than ocean liners provided. After making ten transatlantic crossings in regular commercial service
in 1936, when it was preparing to land at Lakehurst, New Jersey, its hydrogen
ignited and the airship exploded and burned (fig. 1.1(f)).
Since the destruction of the Hindenburg, airship activity has been confined
to the nonrigid type of craft. Although airships are no longer used for passenger transportation, they continued to be used for other purposes such as
advertising and sightseeing.

CHAPTER 1. INTRODUCTION

(a) The first flight of an airship, Henri Giffards steam


airship, 1852.

(b) La France made the


first
deliberate
circle
through the air in 1884.

(c) The first ascent of Ferdinand von Zeppelins LZ-1


in 1900.

(d) Alberto Santos-Dumont


wins the Deutsch prize in
1901 for circling the Eiffel
Tower in his No.6 airship.

(e) Akron, the worlds largest dirigible,


pays its first visit to Washington, D.C.
in 1931.

(f) Explosion of the Hindenburg, at Lakehurst, New Jersey, in 1937.

Figure 1.1: The history of airships.

1.2. THE MOTIVATION - AIRSHIP APPLICATIONS

1.2

The motivation - airship applications

After half century of hibernation, the interest in using airships for several
different applications is increasing worldwide nowadays [1, 2].
The lift of airships is mainly aerostatic, as opposed to aerodynamic as in
airplanes and helicopters. Consequently, and comparing to other aerial vehicles, airships spend most energy moving and compensating wind disturbances,
rather than trying to keep themselves on air. For this reason, they need less
powerful engines, leading to a lower energy consumption, as well as less noise
or vibrations. They possess a big load capacity and long endurance, and they
can fly at low speeds or even hover. Airships also present a slow degradation
in case of failure and are intrinsically more stable than other platforms.
Considering these characteristics, airships have a wide spectrum of applications as observation and data acquisition platforms. They can be used in
several fields related to biodiversity, ecological and climate research and monitoring. Inspection oriented applications cover different areas such as mineral
and archaeological prospecting, agricultural and livestock studies, crop yield
prediction, land use surveys in rural and urban regions, fire detection and also
inspection of man-made structures such as pipelines, power transmission lines,
dams and roads.
Besides their use in advertising2 and leisure flights3 , manned airships are being used in some of the above mentioned applications, among other. The
US/LTA conducts remote sensing experiments with airships since 1992. In
2000, SkyKitten maiden flight takes place in Cardiff. It is capable of landing
virtually anywhere on land or water without need of ground infrastructure and
carrying heavy payloads. In 2001, the first test of a new airship from CargoLifter, also designed to carry up heavy loads, happens in Berlin. The Russian
company RosAeroSystem commercializes the Au-30 Patrol Airship series. The
Total Pole Airship Project (fig. 1.2(a)), for instance, aims to measure the thickness of the pack ice layer covering the Arctic Ocean, using one of the series.
Two others are used for surveillance of power lines in Russia and one other is
scheduled to monitor traffic conditions in Moscow. With a more humanitarian
purpose, Mineseeker (fig. 1.2(b)) is an airship-based mine detection system
with optical, electro-optical and ground penetrating radar sensors, tested in
2

http://www.globalskyships.com/, http://www.airshipman.com/,
http://www.lightships.com/
3
http://www.zeppelinflug.de/, http://www.nac-airship.com/

CHAPTER 1. INTRODUCTION

Kosovo by the United Nations.

(a) Total Pole airship 4 .

(b) The Mineseeker airship at the


KFOR base in Podujevo, Kosovo 5 .

Figure 1.2: Manned airships.

High altitude flight provides a unique vantage point for scientific exploration as
well as for observation and surveillance. An airship, with its heavy lifting capacity, provides the potential for carrying certain types of payloads that would
not be practical for other types of high altitude long endurance vehicles. The
main interest in high altitude airships [3] has been for communications or wide
area surveillance. For civilian applications, high altitude airships represent a
low cost alternative to a geostationary satellite. For the military it represents
a versatile platform that can be positioned over key areas of interest quickly
and provide continuous wide area coverage for extended periods of time.
With these goals in mind, in the USA, Lockheed Martin finished a detailed
design of a high-altitude airship prototype airship in support of the Department of Defense, in order to demonstrate, among others, launch and recovery,
station-keeping and flight control capabilities. The Japan Aerospace eXploration Agency (JAXA) [4] has developed and flight tested a 60m-class unmanned airship successfully to the altitude of 4km in October 2004. In Korea [5], the 50m unmanned airship system KARI (fig. 1.3(a)) is developed and
4
5

http://www.jeanlouisetienne.com/.
http://www.airforce-technology.com/projects/mineseeker/.

1.3. THE CONTROL OF AIRSHIPS FLIGHT

flight tested to acquire basic technologies required to develop a station keeping


electrical powered airship.
A regional navigation system using geosynchronous satellites and stratospheric
airship is proposed by Won [6], and stratospheric communications platforms
for rural applications is developed by Ilcev and Singh [7]. Hurd et al. [8]
consider airships for deep space optical communications. Rao et al. [9] argue
that unmanned airships present a unique potential in emergency management
cycle and discuss their applications for surveillance, search and rescue, and
communication.
Looking further beyond, airships are also being considered for the exploration
of planetary bodies with an atmosphere [10]. NASA [11] is already designing
and testing a robotic lighter-than-air vehicle (fig. 1.3(b)) for the exploration
of planets and moons such as Venus, Mars, Titan and the gas giants.
With such a wide spectrum of applications, and considering the quest for autonomy, airships present characteristics and competitive costs when compared
to other aircrafts, certainly constituting an important option for research, development and experimental validation in autonomous aerial robotics. Moreover, most of the solutions established for this kind of air vehicle may be
transferred or adapted for airplanes or helicopters, where the risks and costs
involved in testing new methodologies are obviously higher.

1.3

The control of airships flight

During the last decades Unmanned Aerial Vehicle (UAV) systems have evolved
into highly capable machines, used mostly for surveillance and data acquisition
purposes. For a rapid unmanned capability advancement, and from a military
perspective, the US Dept. of Defense presented a UAV roadmap in 2005 [12],
containing a survey of platforms and UAV technologies. From the civilian
side, a capabilities assessment of UAVs use in Earth observations is presented
by NASA [13], addressing the technologies and capabilities required for viable
UAV missions.
Many of the UAV applications require the capacity for autonomous flight,
involving the development of a flight control and navigation system. Several
advances made in this field have been published, applying different control
solutions to a variety of UAVs [14, 15, 16, 17, 18, 19, 20, 21, 22].

CHAPTER 1. INTRODUCTION

To provide airships with autonomous operation capacity is a recent focus of


investigation worldwide. Project AURORA - Autonomous Unmanned Remote
mOnitoring Robotic Airship (fig. 1.3(c)) results from a partnership between
DRVC/CenPRA in Brazil, IDMEC/IST in Portugal and ICARE/INRIA in
France [23]. It focuses on the establishment of the technologies required to
substantiate autonomous operation of unmanned robotic airships for environmental monitoring and aerial inspection missions. This includes sensing and
processing infrastructures, control and guidance capabilities, and the ability to
perform mission, navigation, and sensor deployment planning and execution.
Other important researches related to outdoor autonomous airships in the
world at this moment are the Lotte Project at Germany (fig. 1.3(d)) [24], the

French projects at LAAS-CNRS (fig. 1.3(e)) [25], and LSC in Universite dEvry
(fig. 1.3(f)) [26]. In the USA there is a partnership between the projects of
STWing-SEAS6 [27] of University of Pennsylvania and the EnviroBLIMP at
CMU7 (fig. 1.3(g)). Recently, Project DIVA - Dirigvel Instrumentado para
Vigilancia Aerea (fig. 1.3(h)) started in Portugal8 , sharing a partnership with
the AURORA Project.
Aiming at the autonomous airship goal, aerial platform positioning and pathtracking should be assured by a control and navigation system. Such a system
needs to cope with the highly nonlinear and underactuated airship dynamics,
ranging from hover flight (defined here as a flight in low airspeed condition) to
cruise or aerodynamic flight. In addition, the abrupt and continuous dynamics
transition between the two regions, and the different use of actuators necessary
within each region, makes that a very difficult issue to be dealt with by the
control scheme.
The most common solution to the highly nonlinear airship dynamics lies in its
linearization. One important result of the linearization approach is the separation of two independent motions: the motion in the vertical plane, named
longitudinal, and the motion in the horizontal plane, named lateral. This
decoupling allows the design of independent controllers for the two motions.
Following this approach, experimental results were obtained for the AURORA
airship for path following through a set of pre-defined points in latitude/longitude,
along with an automatic altitude control [23]. Also based on a linearized airship model, Wimmer et al. [24] introduced a robust controller design method
6

http://www.stwing.org/blimp/
http://www.frc.ri.cmu.edu/projects/enviroblimp/
8
http://paloma.isr.uc.pt/diva/
7

1.3. THE CONTROL OF AIRSHIPS FLIGHT

(a) Korean airship KARI.

(b) NASA JPL aerobot.

(c) AURORA airship.

(d) Lotte airship.

(e) LAAS-CNRS airship Karma.

(f) LSC airship CEMIF.

(g) STWing-SEAS/EnviroBLIMP airship.

(h) DIVA airship.

Figure 1.3: Unmanned airships.

10

CHAPTER 1. INTRODUCTION

to compensate for the lack of knowledge about the Lotte airship dynamic behavior and model parameters. The decoupled longitudinal and lateral control
systems both consist of an inner H -controller for the dynamics and an outer
SISO P- or PI-controller for the remaining states. Experimental results are
shown therein for the pitch and velocity control. We remark that, as far as
we are aware, both experimental results (from Lotte and AURORA Projects)
on automatic control for outdoor airships are the only ones reported in the
literature at this moment.
Also based on linearized decoupled models of the airship, and for aerodynamic
flight, solutions for the lateral control include H [28], H2 /H approach for
the design of a lateral PD-PI controller [29] and state feedback with integral
control [25]. Considering the control of both lateral and longitudinal motions
different solutions are also proposed, namely one-loop-at-a-time PID [30] and
PI control [31], sliding modes techniques [32, 33], vision-based [34, 35, 36, 37],
fuzzy logic [9] and fuzzy logic improved with genetic algorithms [38]. The
LAAS/CNRS autonomous blimp project [39, 40] proposes a global control
strategy including hover and aerodynamic flight. It is achieved by switching
between four sub-controllers based on linear and backstepping solutions, one
for each of the independent flight phases considered, take-off and landing, and
longitudinal and lateral navigation.
The use of linearized model dynamics restricts the validity of the controller to
trim points, or implies the scheduling between controllers. Using a nonlinear
model avoids these limitations, allowing to design an automatic control system
covering all the aerodynamic range, such that the different flight regions, from
hover to aerodynamic flight, are considered inside a sole formulation. For
security reasons, as well as simplicity and flexibility, a global nonlinear control
is more interesting than a linearized and decoupled one.
Considering a nonlinear model, but assuming a simplified case where the dirigible motion is limited to the horizontal plane, Bestaoui and Hima [41] propose
an input-output linearization control. Using a six-degrees-of freedom nonlinear
model of the airship, Beji et al. [26] introduce a backstepping tracking feedback control for ascent and descent flight maneuvers, where the objective is to
stabilize the airship engine around trimmed flight trajectories. Park et al. [42]
also propose an input-output linearization with a neural network applied to
compensate the underlying model errors, to control velocity, pitch and yaw.
Image-based solutions are also used in the control of indoor airships [43, 44, 45].

1.4. STRUCTURE OF THIS WORK

11

Recently, Guzman [46] compared different control laws that assured reference
tracking in velocity, altitude and heading during aerodynamic flight. The
compared controllers are the classic PID strategy, a generalized predictive
controller and nonlinear first order techniques.
None of these works, with the exception of the LAAS-CNRS group solution
with decoupled controllers [40, 47], presented results for complete missions
including take-off and landing, path-tracking and stabilization. Moreover, seldom are the ones that consider such an important issue as robustness to wind
disturbances. This work, inserted in the AURORA and DIVA projects, aims to
accomplish this task: to develop and compare airship control solutions, valid
for the entire flight envelope and capable of executing realistic missions, while
being robust to wind input.

1.4

Structure of this work

Part of the AURORA and DIVA projects objectives lies on the development of
control solutions to provide an airship with autonomous flight capacities. The
progress in this area comprises different stages: (i) control design theoretical
development and analysis; (ii) implementation of the controller in the simulator, and consequent validation; and (iii) experimental validation in autonomous
flight. This work focuses on the first two steps. We propose to design, validate
on simulation and compare different nonlinear control solutions that will allow
an airship to navigate autonomously.
The first milestone is obviously the modeling of the airship, which is a complex dynamic system with six-degrees-of-freedom. Chapter 2 is dedicated to
this objective. In order to provide a general idea of this kind of aerial vehicle,
like usual sensor and actuators available, their configuration and limitations,
the AURORA prototype is described. A model description allows us to better understand the airship behavior using a simulator, and is at the basis of
model-based control laws. The airship equations of motion are comprised of
both dynamics and kinematics, and include the wind input. The dynamics are
obtained using the Lagrangian approach, instead of the Newtonian method
used in [48]. The airship equations of motion provide a good system description, and therefore allow to reproduce the airship response to actuator and
wind disturbances inputs. For demonstration purposes, a simulator based on
the AURORA airship characteristics is available [29]. The complexity of the

12

CHAPTER 1. INTRODUCTION

equations, however, justifies the search for a linear version, a usual practice
in aeronautics. The linearization of the airship model around an equilibrium
condition results in the decoupling of the longitudinal and lateral modes. The
analysis of the different models obtained for trim points over the flight envelope provides a good knowledge of the airship behavior at different airspeeds,
important in order to develop a controller valid for any type of mission.
Before advancing to the control design, some common concepts and tools are
described in Chapter 3. These include references and errors definition, wind
estimation, and the criteria used in the comparison of the different nonlinear
controllers presented.
Chapter 4 is justified as an introduction for the control design part. Considering only linear control, and therefore a single model and controller, the
applicability of the solutions presented is restricted to the regulation of the
state errors, assuring the validity of the linearization at the chosen trim point.
This limitation vanishes if the linear systems and controllers are not fixed but
change with the measured airspeed (which defines an equilibrium condition).
This is the idea of the Gain Scheduling technique presented in Chapter 5,
which extends the validity of the linearization approach to a range of operating
points, instead of a single one. For each linear model obtained, a control law
is designed, and the overall control synthesis is achieved by switching between
models and respective controllers as function of scheduling variables. This is
the first of the three nonlinear control solutions we propose.
This approach, however, depends mostly on the engineer knowledge of the
system for a good choice of the scheduling parameters, resulting in an iterative
and time consuming process. Moreover, its guarantees of success depend on
extensive simulations covering different possible scenarios. This leads to the
search of a single control law, more related to the airship system itself than to
the control designer experience. The Dynamic Inversion approach described in
Chapter 6 is such a methodology. By inverting the system model, a control law
is obtained that cancels existing deficient or undesirable dynamics by replacing
them with a set of desired ones.
The third and last solution is Backstepping, a Lyapunov-based control design
approach presented in Chapter 7. By formulating a scalar positive function
of the system states and then choosing a control law to make this function
decrease, it is guaranteed that the nonlinear control system thus designed will
be stable. Moreover, it will be robust to some unmatched uncertainties.

1.5. CONTRIBUTIONS OF THIS THESIS

13

Any of the three control solutions presents its advantages and disadvantages,
many of which are discussed in the respective chapters. Yet, a common comparison between them is important as to provide a better overview of the
different control options. In Chapter 8 this assessment is made considering
parameters such as path-tracking trajectory errors and actuators request for
a case-study complete mission, controller performance robustness in face of
model parameter uncertainty and computational effort. These factors, together with some implementation issues, are relevant to evolve to the next
phase, the experimental validation in autonomous flight.
In Chapter 9, the conclusions of this work summarize the knowledge gained,
and point the directions of our forthcoming investigation.

1.5

Contributions of this thesis

The main contributions of this thesis are the following:


the derivation of the airship dynamic equations of motion using the Lagrangian approach;
a thorough analysis of the airship control design problem over the entire
flight envelope;
the proposal of alternative control solutions so as to have a single control law valid for different missions, independent of the flight region and
robust to realistic wind disturbances, their evaluation and comparison.
This covers the following control methodologies:
Gain Scheduling, providing a linearized reference for comparison;
Dynamic Inversion, with a new formulation for cascaded systems
described in terms of velocity and position, and where the output
of interest is the position;
Backstepping, including input saturations.
analysis of specific problems, as well as desired performance criteria.
Parts of the work related with this thesis have been previously published.
Related with Chapter 5, the application of gain scheduling to airship pathtracking is presented in:

14

CHAPTER 1. INTRODUCTION

Alexandra Moutinho and Jose Raul Azinheira. A Gain-Scheduling


Approach for Airship Path-Tracking. In Proceedings of the 4th International Conference on Informatics in Control, Automation and
Robotics, Setubal, Portugal, August 2006.
The application to hover control is described in:
Alexandra Moutinho and Jose Raul Azinheira. A Gain-Scheduling
Approach for Airship Stabilization. In Proceedings of the 7th Portuguese Conference on Automatic Control, Lisbon, Portugal, September 2006.
Related with Chapter 6, the application of dynamic inversion to the lateral
control synthesis of the Aurora airship and its comparison with linear control
can be found in:
Alexandra Moutinho and Jose Raul Azinheira. Path control of
an autonomous airship using dynamic inversion. In Proceedings
of the 5th IFAC/EURON Symposium on Intelligent Autonomous
Vehicles, Lisbon, Portugal, July 2004.
A sensitivity and robustness analysis of the dynamic inversion controller was
first presented in:
Alexandra Moutinho and Jose Raul Azinheira. Stability and robustness analysis of the AURORA airship control system using
dynamic inversion. In Proceedings of the IEEE International Conference on Robotics and Automation, Barcelona, Spain, April 2005.
The hover control based in the dynamic inversion approach is described in:
Alexandra Moutinho and Jose Raul Azinheira. Hover Stabilization
of an Airship using Dynamic Inversion. In Proceedings of the 8th
International IFAC Symposium on Robot Control, Bologna, Italy,
September 2006.
Preliminary results on the dynamic inversion and backstepping solutions are
included in this AURORA project report:

1.5. CONTRIBUTIONS OF THIS THESIS

15

Ely Carneiro de Paiva, Jose Raul Azinheira, Josue G. Ramos Jr.,


Alexandra Moutinho and Samuel Siqueira Bueno, Project AURORA: Infrastructure and Flight Control Experiments for a Robotic
Airship. In Journal of Field Robotics, Vol.23 (3-4), pp. 201-222,
March/April 2006.
Related with Chapter 7, a backstepping solution using quaternions for airship
stabilization is proposed in:
Jose Raul Azinheira, Alexandra Moutinho and Ely Carneiro de
Paiva. Airship Hover Stabilization using a Backstepping Approach.
In Journal of Guidance, Control and Dynamics, Vol.29 (4), pp.
903-914, July/August 2006.
while a backstepping solution for the stabilization of a generic UAV, also using
quaternions, is described in:
Jose Raul Azinheira and Alexandra Moutinho. Hover Control of
a UAV with Backstepping Design Including Input Saturations. In
press, IEEE Transactions on Control and Systems Technology.

16

CHAPTER 1. INTRODUCTION

Chapter 2
The Airship Model
Contents
2.1

Airship platform . . . . . . . . . . . . . . . . . . . .

18

2.2

Airship equations of motion . . . . . . . . . . . . .

19

2.3

2.4

2.2.1

Airship dynamics . . . . . . . . . . . . . . . . . . . . 20

2.2.2

Airship kinematics . . . . . . . . . . . . . . . . . . . 32

2.2.3

Airship simulator . . . . . . . . . . . . . . . . . . . . 34

Airship linearized models . . . . . . . . . . . . . . .

34

2.3.1

Trim or equilibrium conditions . . . . . . . . . . . . 35

2.3.2

Model linearization . . . . . . . . . . . . . . . . . . . 38

Conclusions . . . . . . . . . . . . . . . . . . . . . . .

43

A good knowledge of the airship model and behavior is essential for a successful control design. For this reason, this chapter presents the airship modeling.
In order to provide a general idea of this kind of aerial vehicle, like usual sensor and actuators available, their configuration and limitations, the AURORA
prototype is described in Section 2.1. The airship nonlinear dynamic model
is introduced in Section 2.2, formed from both dynamic and cinematic equations. The complexity of the nonlinear model justifies the search for a linear
simplified version. Section 2.3 describes the linearization procedure that leads
to the decoupled longitudinal and lateral state-space models. For the insight
it provides on the airship behavior, the longitudinal motion of the airship is
analyzed as function of the airspeed.
Whenever necessary for demonstration purposes, the AURORA airship platform characteristics and configuration are used.
17

18

2.1

CHAPTER 2. THE AIRSHIP MODEL

Airship platform

The Lighter-Than-Air (LTA) robotic prototype AURORA has been built as


an evolution of the Airspeed Airships AS800. It is a nonrigid airship 10.5m
long, with 3.0m diameter and 34m3 of volume. The payload capacity is approximately 10kg and maximum speed is around 50km/h (see fig. 2.1).

Figure 2.1: The AURORA airship.

The main control and navigation sensors currently used on the airship are (see
fig. 2.2(a)): a GPS with differential correction that provides the inertial position coordinates and velocity; an Inertial Measurement Unit (IMU), which
provides the roll, pitch, and yaw attitude, the angular rates and body axes
linear acceleration, serving as an inclinometer and compass as well; a Wind
sensor that measures the relative airship air speed in all three axes, the aerodynamic incidence angles, as well as the barometric altitude; and a Camera
that provides aerial images for vision processing algorithms.
The airship actuators are its deflection surfaces and two main propellers disposed on each side of the gondola (see fig. 2.2(b)). The four deflection surfaces
at the stern, arranged in a shape with allowable deflections in the range
[25o , +25o ], generate the equivalent rudder r and elevator e commands of
the classical + tail. The aileron command a is obtained with the rotation of the four deflection surfaces in the same direction. The two engines,
with a vectoring angle v within the interval [30o , +120o ], are driven by twostroke engines providing total XT (within [0, 80]N ) and differential TD (within
[0, 40]N ) thrusts. A small lateral stern thruster Ty may also be available, perpendicular to the airship longitudinal axis, to introduce one extra horizontal
input actuation during hovering tasks.

19

2.2. AIRSHIP EQUATIONS OF MOTION

GPS antenna
Wind sensor

e
v

XT , TD

_
+

+
_

Camera
GPS, IMU

TY

+ _

(a) Sensors.

(b) Actuators.

Figure 2.2: AURORA airship sensors and actuators. (The aileron command
a is obtained with the rotation of the four deflection surfaces in the same
direction.).

2.2

Airship equations of motion

The airship model is a mathematical description of the airship motion. It


is given by a set of differential equations called equations of motion, which
represent the relations between the control inputs and the state variables. A
description of the referentials mentioned herein may be found in Appendix A.
The airship nonlinear model results in the dynamic equation expressed in a
state-space form:
x = f (x, u, d)

(2.1)

where:
the state x = [vT , T , pT , T ]T includes the linear v = [u, v, w]T and
angular = [p, q, r]T inertial velocities of the airship expressed in the
{l} frame, the cartesian position p = [pN , pE , pD ]T of its center of volume
in the {i} frame, and the attitude of the airship = [, , ]T given by
the Euler angles;
the input vector u = [e , TL , TR , v , a , r , TY ]T includes the elevator deflection e , the left and right engines thrust TL , TR , the engines vectoring
angle v , the aileron deflection a , the rudder deflection r and the lateral thrust TY (since it is not yet implemented in the AURORA airship,
although mentioned, it will not be used for control);

20

CHAPTER 2. THE AIRSHIP MODEL

the disturbance vector d includes the wind input (wind velocity) expressed in the {i} frame with a constant term and a six components
vector modeling the atmospheric turbulence (nonconstant wind). It is
represented by linear velocity p w = [pNw , pEw , pDw ]T (a horizontal representation in polar components is [psw , phw ]T , with wind strength psw and
heading phw ) and angular velocity w .
In the attempt to establish a workable mathematical model of the airship flight,
a number of considerations have to be taken into account [48]:
1. the airship displaces a very large volume of air and its virtual (added)
mass and inertia properties become significant, i.e., the LTA vehicle behaves as if it had a mass and moments of inertia substantially higher
than those indicated by conventional physical methods;
2. three kinds of masses and inertia matrices must be considered: the mass
and inertia (m, J) of the vehicle itself; the mass and inertia (mB , JB ) of
the buoyancy air, corresponding to the air displaced by the total volume
of the airship; and the virtual mass and inertia (Mv , Jv ), which may be
regarded as the mass of air around the airship and displaced with the
relative motion of the airship in the air;
3. the airship mass changes in flight due to ballonet deflation or inflation.
However, fuel changes are ignored;
4. the airship is assumed to be a rigid body, and the aeroelastic effects are
neglected.
The airship model (2.1) can be described by two equations. The first one
characterizes the system dynamics with respect to the {l} frame, while the
second one represents the cinematic relation between the {l} and {i} frames.
The next two sections present these two equations, that together describe the
airship nonlinear dynamic model.

2.2.1

Airship dynamics

When the displaced fluid mass is not negligible, as is the case for airships, the
equations of motion are usually derived using the Lagrangian approach [49, 50].

2.2. AIRSHIP EQUATIONS OF MOTION

21

Let the motion of the airship be described by its inertial velocity V = [vT , T ]T ,
a 6D vector including the inertial linear (v) and angular () velocities. Let
the surrounding air be described by an inertial wind velocity Vw = [vwT , Tw ]T .
The airship has thus a relative air velocity Va equal to the difference of the
previous two:
Va = V Vw

(2.2)

The total kinetic energy W is defined as a sum [48]:


W = W c + WBo + Wvo

(2.3)

accounting for:
the vehicle motion, expressed in the center of gravity C, with M c =
diag(mI3 , J) is the generalized mass matrix:
1
W c = VcT M c Vc
2

(2.4)

the kinetic energy added to the buoyancy air (displaced by the airship
volume), expressed in the center of buoyancy O where the {l} frame is
fixed (we will drop the O to lighten the notation), and where MB =
diag(mB I3 , JB ) is the generalized mass matrix of the buoyancy air:
1
1
WBo = VT MB V + VaT MB Va
2
2

(2.5)

the energy due to an extra virtual mass, also expressed in O, and where
Mv = diag(Mv , Jv ) is the generalized virtual mass matrix:
1
Wvo = VaT Mv Va
2

(2.6)

Therefore we may write W as:


1
1
1
1
W = VcT M c Vc VT MB V + VaT MB Va + VaT Mv Va
2
2
2
2

(2.7)

We can represent all terms of the kinetic energy in the {l} frame considering
that the linear speed of the CG (vc ) is related to the linear speed of the CV

22

CHAPTER 2. THE AIRSHIP MODEL

(vo = v) through the angular speed:


vc = vo + OC = v OC

(2.8)

From (2.8) and knowing that c = o = , we have:


Vc =

"

I3 C3
03 I3

(2.9)

where C3 denotes the antisymmetric cross-product matrix corresponding to


the operation OC. 1
Substituting (2.9) into (2.7) gives:
1
1
1
1
W = VT Mo V VT MB V + VaT MB Va + VaT Mv Va
2
2
2
2

(2.10)

with:
Mo =

"

I3 03
C3 I3

Mc

"

I3 C3
03 I3

"

mI3
mC3
mC3 J mC23

(2.11)

Introducing (2.2) into (2.10) we obtain:


1
1
W = VT (Mo MB )V + (V Vw )T (MB + Mv )(V Vw )
2
2
1 T
1 T
= V (Mo + Mv )V + Vw (MB + Mv )Vw VT (MB + Mv )Vw (2.12)
2
2
Finally, we have the kinetic energy expressed as function of the airship and
wind inertial velocities:
1
1
W = VT Ma V + VwT MBa Vw VT MBa Vw
2
2

(2.13)

with Ma = Mo + Mv and MBa = MB + Mv .

The antisymmetric cross-product matrix is defined for a generic vector a = [ax , ay , az ]T

as:

0
a = az
ay

az
0
ax

ay
ax
0

23

2.2. AIRSHIP EQUATIONS OF MOTION

The Lagrangian or Euler-Lagrange equations of motion may be given by [51]:


d
q) =
F(q,
dt

W
q

W
q

(2.14)

is the system kinetic energy expressed as function of the genwhere W (q, q)


q) is the
and F(q,
eralized coordinates q vector and its time derivative q,
generalized forces vector.
We will apply equation (2.14) to each of the three terms of equation (2.13)
separately. We start with the first term, repeating after the procedure for the
remaining two terms.
Let us define the generalized coordinates vector as:

T
q = [pN , pE , pD , , , ]T = pT , T

(2.15)

whose time derivative is related with the velocity vector V by:


"

"

ST 03
03 R

#"

q = J V

(2.16)

with the orthogonal transformation matrix S defined by

cos cos
sin cos
sin

S = cos sin sin sin cos sin sin sin + cos cos cos sin
cos sin cos + sin sin sin sin cos cos sin cos cos

(2.17)

and the coefficient matrix R given by:

1 sin tan cos tan

R= 0
cos
sin
0 sin / cos cos / cos

(2.18)

The first term of the kinetic energy corresponding to the case with no wind is
then:
1
W1 = VT Ma V
2

(2.19)

24

CHAPTER 2. THE AIRSHIP MODEL

or, using equation (2.16):


1
1

W1 = q T JT
M a J q
2

(2.20)

The partial derivative of the kinetic energy W1 relative to q is:



1  T
W1
T
1
T

(J Ma J1
=
) + (J Ma J ) q
q
2
1

= JT
M a J q

(2.21)

since Ma = Ma . Computing now its time derivative leads to:


d
dt

W1
q

1
1
+ JT M J1 q
+ JT
= (JT
a

Ma (J )q
) Ma J q

= JT
Ma V + (J ) Ma V

(2.22)

where we used (2.16) and the relation:


1

(J1
)J + J J = 0

(2.23)

The partial derivative of the kinetic energy W1 now relative to q is:


W1
1 T T

=
(q J Ma J1
q)
q
2 q
= KMa V

(2.24)

with:



J1
K=
q
q

"

03 03
K1 K2

"

 #

q
J1

J1
q

(2.25)

The generalized force relative to the kinetic energy with no wind is then obtained from the difference between (2.22) and (2.24) according to (2.14):

= JT
F1 (q, q)
Ma V + (J ) Ma V KMa V

(2.26)

We will now proceed applying (2.14) to the second term of equation (2.13).
We start defining the wind coordinates vector as:

T
qw = [pNw , pEw , pDw , w , w , w ]T = pTw , Tw

(2.27)

25

2.2. AIRSHIP EQUATIONS OF MOTION

whose time derivative is related with the wind velocity vector Vw by:
"

p w
w

"

ST 03
03 R

#"

vw
w

q w = J Vw

(2.28)

The second term of the kinetic energy (2.13) corresponds to:


1
W2 = VwT MBa Vw
2
1 T T
w
= q w J MBa J1
q
2

(2.29)

Applying (2.14) to (2.29), with q defined in (2.15), we obviously have:


d
dt

W2
q

=0

(2.30)

The partial derivative of the kinetic energy W2 relative to q is:


W2
1 T T
w)
=
(q J MBa J1
q
q
2 q w
= Kw MBa Vw

(2.31)

with:


w
Kw =
J1
q
q

"

03
03
Kw1 Kw2

"

w
J1
q

 #

w
J1
q

(2.32)

The generalized force relative to W2 is then obtained from the difference between (2.30) and (2.31) according to (2.14):
= Kw MBa Vw
F2 (q, q)

(2.33)

Finally, applying the same procedure to the last term of (2.13):


W3 = VT MBa Vw

(2.34)

we obtain:
=
F3 (q, q)

JT

h
iT
1

MBa Vw (J ) MBa Vw + Kw MBa V + KMBa Vw

(2.35)

26

CHAPTER 2. THE AIRSHIP MODEL

Summing up all the generalized forces we have:


=F1 (q, q)
+ F2 (q, q)
+ F3 (q, q)

F(q, q)

=JT
Ma V + (J ) Ma V KMa V Kw MBa Vw
h
i
) T M V + K M V + KM V
w (J1
JT
V
M
Ba
Ba w
w
Ba
Ba w

(2.36)

Considering that:

F(V) = JT F(q, q)

(2.37)

we have:

T
T
+ (JT (JT
F(V) =Ma V

) J K)Ma V J Kw MBa Vw

 h
iT
1
T

MBa Vw J (J ) J K MBa Vw + JT Kw MBa V (2.38)


Moreover, considering the following equalities:
V3 = RT K1
= RT ((RT ) K2 )

(2.39)
(2.40)

Vw3 = RT Kw1

(2.41)

0 = RT Kw2

(2.42)

where V3 , and Vw3 denote the antisymmetric cross-product matrices corresponding, respectively, to the operations v, and vw , leads to:

T
JT (JT
) J K = 6 + V6

(2.43)

JT Kw = Vw6

(2.44)

whereas:
V6 =

"

03 03
V3 03

6 =

"

03
03

Vw6 =

"

03 03
Vw3 03

(2.45)

The dynamics equation of the airship in the inertial frame is then given by:
+ (6 + V6 )Ma V MBa V
w (6 + V6 )MBa Vw Vw6 MBa (V Vw )
F = Ma V
(2.46)

27

2.2. AIRSHIP EQUATIONS OF MOTION

in accordance with the equations derived by Thomasson in [50], using quasicoordinates, without the gradient terms.
Let us now deduce the dynamics equation in the air frame which we will
limit to the case of constant translation wind velocity in the inertial frame.
Considering that (see Appendix A.2):
w =
p

dvw
= v w + vw
dt

(2.47)

we have:
constant wind v w = vw

(2.48)

translation wind w = 0

(2.49)

which leads to:


w = 6 Vw
V

(2.50)

Substituting (2.2), (2.50) and the matricial relation:


V6 = Va6 + Vw6

(2.51)

(Va6 is defined similarly to V6 in (2.45) but considering the airspeed velocity


va ) into (2.46) results in:
a Ma 6 Vw + 6 Ma Va + 6 Ma Vw + (Va6 + Vw6 )Ma (Va + Vw )
F = Ma V
+MBa 6 Vw 6 MBa Vw (Va6 + Vw6 )MBa Vw Vw6 MBa Va


a + 6 Ma Va + 6 (Ma MBa ) (Ma MBa )6 Vw +
= Ma V
+(Va6 + Vw6 )(Ma MBa )(Va + Vw ) + Va6 MBa Va

(2.52)

Considering that:
Ma MBa =

"

(m mB )I3
mC3
mC3
J mC23 JB

(2.53)

28

CHAPTER 2. THE AIRSHIP MODEL

we have the following equalities:




6 (Ma MBa ) (Ma MBa )6 Vw =

"
"

0
m(C3 C3 )vw

0
(Va6 + Vw6 )(Ma MBa )(Va + Vw ) =
(Va3 + Vw3 )mC3
#
"
0
Va6 MBa Va =
Va3 Mv va

(2.54)
#

(2.55)
(2.56)

Substituting equalities (2.54)-(2.56) into equation (2.52) and knowing that the
cross-product satisfies the Jacobi Identity2 , leads to:
a + 6 Ma Va +
F = Ma V

"

0
Va3 (Mv va mC3 )

(2.57)

Finally, since:
"

0
Va3 (Mv va mC3 )

= Va6 Ma Va

(2.58)

we obtain the dynamics equation of the airship in the air frame for a constant
translation wind:
a + (6 + Va6 )Ma Va
F = Ma V

(2.59)

which has the same form as equation (2.46) for the no-wind case, i.e., equation (2.46) is invariant under a steady translation, as pointed out in [52].
Let us now go back to equation (2.46). We can write it as:
= Fkw + F
Ma V

(2.60)

with the kinematics and wind forces given by:


w + (6 + V6 )MBa Vw + Vw6 MBa (V Vw )
Fkw = (6 + V6 )Ma V + MBa V
(2.61)
and where F = Fg + Fa + Fp contains gravitational, aerodynamic and propulsion forces. The aerodynamic force contains all terms proportional to the dy2

a (b c) + b (c a) + c (a b) = 0

29

2.2. AIRSHIP EQUATIONS OF MOTION

namic pressure ( 12 Vt2 , with the air density), identified in wind tunnel tests.
This means all terms proportional to the square of the airspeed in Fkw are
already included in Fa and are at the moment accounted twice in the dynamic
model.
A closer look at (2.61) shows that:
FV V6 Ma V + V6 MBa Vw + Vw6 MBa Va
= V6 (Ma MBa )V Va6 MBa Va

(2.62)

According to (2.56), the term Va6 MBa Va in (2.62) is proportional to the


square of the airspeed and is already accounted for in Fa . Therefore it should
be removed from Fkw , which can now be written as:
w + 6 MBa Vw
Fkw = 6 Ma V V6 (Ma MBa )V + MBa V

(2.63)

According to [48], the gravitational force Fg , which adds the weight force
applied at the GC and the buoyancy force applied at the CV, is a function of
the transformation matrix S:
#
# "
"
m w I3
S(m mB )ag
Sag = Eg Sag
(2.64)
=
Fg =
mC3
OC Smag
with the gravity acceleration ag = [0, 0, g]T given in the {i} frame and the
airship weighting mass defined as the difference between its weight and its
buoyancy, mw = m mB .
Finally, the dynamic equation of the airship in the inertial frame is given by:
= 6 Ma V V6 (Ma MBa )V + MBa V
w + 6 MBa Vw + Eg Sag +Fa + Fp
Ma V
{z
} | {z }
|
Fkw

Fg

(2.65)

while in the air frame it is given by (considering translation constant wind):


a = 6 Ma Va Va6 (Ma MBa )Va + Eg Sag +Fa + Fp
Ma V
{z
} | {z }
|
Fkw

(2.66)

Fg

where the previous search of terms


in Fa was applied to the
" already included
#
0
= Va6 (Ma MBa )Va .
term (2.58) of (2.59) and where
mVa3 C3

30

CHAPTER 2. THE AIRSHIP MODEL

2.2.1.1

Forces to actuators input

As previously mentioned, the aerodynamic force vector Fa contains all dynamic


terms proportional to the dynamic pressure, identified in wind tunnel tests.
These include any control surface deflection effects corresponding to aileron
a , rudder r and elevator e , which are usually presented as a separate entity.
Therefore, if we divide the aerodynamic force such that:
Fa = Fa (Va ) + Fa ()

(2.67)

with the state only depending part Fa (Va ) and the control surfaces force input
Fa (), we may rewrite equations (2.65)-(2.66), respectively, as:
= 6 Ma V V6 (Ma MBa )V + MBa V
w + 6 MBa Vw
Ma V
+ Eg Sag + Fa (Va ) + uf

(2.68)

a = 6 Ma Va Va6 (Ma MBa )Va + Eg Sag + Fa (Va ) + uf (2.69)


Ma V
where uf = Fa ()+Fp contains both the control surfaces and propulsion forces
input.
The relation between the actuators represented in fig. 2.2(b) and the control
force inputs uf depends on the flight region, as explained in more detail at the
end of Section 2.3.1:
In the low airspeed region, the tail surfaces have reduced authority since
the action from the surface deflections is a function of the dynamic pressure and varies as the square of the airspeed Vt , according to the aerodynamic characteristics of the airship [53]. This leaves the airship to be
mainly controlled by the propulsion force inputs. The two main propellers correspond to 3 inputs (TL , TR , v ) - left and right thrust and
vectoring angle - providing longitudinal and vertical forces, pitching and
rolling torques. If available, the tail lateral thruster adds one input (TY ),
providing a side force and a yawing torque. These force actuators are
slightly influenced by the airspeed, but may be considered as independent
on a first approach.
In aerodynamic flight, the vectoring angle is no longer necessary, leaving
the airship with a reduced vertical force. The maneuvering is mostly
accomplished by the tail fins. The surface deflections correspond to
the three standard inputs of aileron, elevator and rudder deflections

31

2.2. AIRSHIP EQUATIONS OF MOTION

(a , e , r ), which mostly correspond to torque inputs for the control of


the pitch and yaw motions, keeping the airship with reduced lateral force
input.

With the six actuators inputs (e , TL , TR , v , a , r ) to control six forces (three


forces and three torques), the airship does not seem underactuated, but numerous limitations severely reduce its controllability:

no actuator is really available to oppose the aerodynamic side force;


the main engines provide four coupled force components with only three
inputs;

the tail surfaces depend on the airspeed and their authority vanishes in
the no-wind case, leaving the airship to be controlled by the force inputs
only;

all the actuators have level and rate saturation limits, that cannot be
avoided;

the force actuators, in particular, have their own dynamics, with limited
response times, that must be taken into account.

The relation between actuators and force inputs may then be established for
design purposes, neglecting the actuators dynamics, using the airspeed measurement and resolving the possible redundancies according to the usual operation of the airship [23] (the airship aerodynamic angles also have their effect,
but they may be neglected on a first approach, assuming small angles):
uf = fu (u, Vt )

(2.70)

where u = [e , TL , TR , v , a , r ]T is the real actuators input (the lateral thrust


TY is not considered since it is not yet implemented in the AURORA airship),
Vt is the true airspeed, and uf = [fu , fv , fw , fp , fq , fr ]T is the force vector,
solution of the system composed by the six equations below, in agreement

32

CHAPTER 2. THE AIRSHIP MODEL

with the AURORA airship configuration:


fu = XT cos v + k1 e

(2.71a)

fv = k2 r

(2.71b)

fw = XT sin v + k3 e

(2.71c)

fp = k2 l4 a + b4 sin v TD

(2.71d)

fq = XT b3 cos v + k5 e

(2.71e)

fr = k2 l6 r + b4 cos v TD

(2.71f)

where XT = TL +TR is the total thrust, TD = TL TR is the differential thrust,


(bj , lj ) are geometrical constants of the airship, and kj (Vt ) are second order
polynomials expressing the airspeed depending authority of the tail deflections.

2.2.2

Airship kinematics

For control and navigation purposes, the velocity vector V, expressed in the
airship {l} frame, must be transformed to the {i} frame. This leads to the
cinematic relations.
Consider the airship position is given by its coordinates in the {i} frame and the
attitude is described in terms of the Euler angles (, , ). Then, the airship
position may simply be regarded as the integration of the inertial velocity in
the {i} frame:

N
u

T
E =S v
D
w

(2.72)

where S is the orthogonal transformation matrix (2.17) that satisfies the equation:
S = S

(2.73)

Similarly, the time derivatives of the Euler angles may be related to the local

33

2.2. AIRSHIP EQUATIONS OF MOTION

angular rates (p, q, r):

p


= R q

(2.74)

with the angular transformation matrix R given by (2.18) and satisfying:


= R(R1 )R
R

(2.75)

If P = [pN , pE , pD , , , ]T defines the 6D position and V = [u, v, w, p, q, r]T


the local velocity, the position cinematic equation is expressed by:
= J V
P

(2.76)

with:
J =

"

ST 03
03 R

(2.77)

and:
J = J CJ , with CJ =

"

03
03 (R1 )R

(2.78)

The kinematics equation may also be given considering the air relative velocity
by:
= J (Va + Vw )
P

(2.79)

or, considering again translation constant wind:


= J Va + BI p w
P
with BI =

"

I3
03

(2.80)

34

CHAPTER 2. THE AIRSHIP MODEL

2.2.3

Airship simulator

Based on the 6 degrees of freedom nonlinear model compounded by equations (2.68), (2.71) and (2.76), a MATLABr /Simulinkr -based simulator was
built, allowing the design and validation of flight control and guidance strategies [29]. The simulator block diagram of the airship open-loop model is represented in fig. 2.3.
wind / turbulence
input

uc

actuators
model

airship
nonlinear model
x

Figure 2.3: Simulator block diagram for airship open-loop model.

The airship nonlinear model, as described by equation (2.1), is a function of


the state variables x, the actuators input u and wind disturbances d. These
last include both constant wind and atmospheric turbulence, modeled here by
the Dryden model (see Appendix B). The sensors described in Section 2.1
are considered ideal while the actuators model includes the propellers and
control surfaces dynamics, like delays and saturations, applied to the actuators
command input uc . The output vector y consists of interest variables to be
monitored.

2.3

Airship linearized models

The complexity of the airship nonlinear dynamic equations presented before


justifies the search for a linear version, also important in order to analyze
and evaluate the characteristics of the airship dynamics, and usual practice in
aeronautical systems.
In the following sections we will describe the procedure applied on the linearization of the airship model given by equations (2.68), (2.71) and (2.76),
and analyze the airship characteristics at the different equilibrium points over
the flight envelope.

2.3. AIRSHIP LINEARIZED MODELS

2.3.1

35

Trim or equilibrium conditions

An equilibrium or trim point corresponds to a condition at which a dynamical


system is steady or at rest.
We consider here no disturbance, being the deterministic nonlinear model of
the airship given by:
x = f (x, u)

(2.81)

A trim point of the system (2.81) is then a point (x, u) = (xe , ue ) such that
the airship is in equilibrium, with a subset y e of its derivatives null:
x e = f (xe , ue )

(2.82)

y e = Cx e = 0

(2.83)

This implies there is a balance between forces acting on the airship and the
airship will remain in that particular flight condition until some disturbance
or some control input occurs.
The first step, prior to the system linearization, is to find the solutions (xe , ue )
of (2.82)-(2.83) over the flight envelope, i.e., for varying airspeed. Due to the
complex functional dependence of the aerodynamic data, this cannot be done
analytically. A numerical way to do so is to specify a convex optimization problem, with the following constraints for the restricted case problem of a straight
level flight at a given constant altitude he = De and constant airspeed Vte :
for steady flight, the derivatives of the linear and angular velocities are
= 0;
zero: v,
for steady straight flight, the derivative of the vertical position and the
= 0;
angular velocity are zero: D,
for symmetric pure longitudinal flight, the sideslip and roll angles are
zero: , = 0;
from the cinematic relation (2.74), the derivatives of the Euler angles are
= 0;
zero:
for still straight flight, the left and right engines thrust is equal: TL = TR ;
no need for lateral actuation: a , r , Ty = 0;

36

CHAPTER 2. THE AIRSHIP MODEL

for constant altitude: D = De ;


for constant airspeed:

u2 + v 2 + w2 = Vte ;

for level flight: = .


The position variables N, E and are not necessary to find a steady-state
condition, allowing the airship to take any straight level direction.
The minimization of the convex cost function, that fulfils the above constraints,
provides trim values of state and control input:
xe = [ue , 0, we , 0, 0, 0, , , De , 0, e , ]T

(2.84)

ue = [e , TLe , TRe , ve , 0, 0, 0]T

(2.85)

where indicates the variables for which no direct constraint is set.


The results obtained repeating the procedure for a range of reference airspeeds
allow us to analyze the changes in the longitudinal motion behavior of the
AURORA airship over the flight envelope.
Figure 2.4 represents the equilibrium values of pitch e , elevator e , total thrust
XTe = TLe + TRe and vector angle ve necessary for the AURORA airship, with
different weighting masses mw = 1, 3, 5kg, to maintain a steady straight level
flight at 50m altitude, and at different airspeeds Vt varying from 0 to 15m/s
in steps of 0.1m/s.
At low airspeeds, the propellers vectoring angle is necessary in order to compensate for the loss of lift force from aerodynamics, as may be seen in fig. 2.4(a).
It can also be observed that the heavier the airship, the later (or at higher airspeeds) it will need the propellers to be vectored.
For very low airspeeds, the vectoring angle is almost 90o and the model is
essentially that of aerostatic forces, with the weight excess being compensated
by the propellers vectoring. For very high airspeeds, the propellers vectoring
is not necessary, and the model is essentially that of aerodynamic forces. The
fast transition from low to high vectoring angles separates two flight regions,
namely cruise or aerodynamic flight (AF) and low airspeed or hover flight (HF).
In this transition there exists compensation from both aerodynamic forces and
propellers vectoring. For these reasons, this should be the most difficult region
to be corroborated in the model validation process and also for the control
design.

37

2.3. AIRSHIP LINEARIZED MODELS

100

12

90

mw = 1kg
mw = 3kg
mw = 5kg

80

70
60

6
e (deg)

ve (deg)

mw = 1kg
mw = 3kg
mw = 5kg

10

50
40

4
2

30
20

10
-2

0
-10

7
8
Vt (m/s)

10

11

12

13

14

-4

15

(a) Thrusters vector angle ve vs. airspeed


Vt .

8
7
Vt (m/s)

10

11

12

14

13

15

(b) Pitch angle e vs. airspeed Vt .

60

mw = 1kg
mw = 3kg
mw = 5kg

50

2
0

40
XTe (N )

ee (deg)

-2
-4

30

-6
20

-8
mw = 1kg
mw = 3kg
mw = 5kg

-10
-12
-14

8
7
Vt (m/s)

10

11

12

13

14

(c) Elevator deflection ee vs. airspeed Vt .

10

15

7
8
Vt (m/s)

10

11

12

13

14

15

(d) Thrust input XTe vs. airspeed Vt .

Figure 2.4: Trim values of state and control input for different weighting
masses mw over the flight envelope.

It can be seen in fig. 2.4(b) that the need for a higher pitch angle occurs in
the mid-range airspeeds, when the airship is asking for more aerodynamic lift.
At low airspeeds, as this extra up force is supplied by the propellers vectoring
and the aerodynamic efficiency decreases, this is no longer necessary.
Also observed is, the lower the weighing mass, the lower the pitch angle necessary to provide the sufficient aerodynamic force to maintain the equilibrium
condition.
The necessary lift is obtained through the pitch angle, which in turn produces
a pitch rate, that must be compensated by the elevator control surface action
represented in fig. 2.4(c), which justifies the similarity of both curves. At low
airspeeds, the elevator, as all the control surfaces, has reduced authority since
its action is a function of the dynamic pressure and varies with the squared

38

CHAPTER 2. THE AIRSHIP MODEL

airspeed [53].
The elevator deflection is almost the same for the three kinds of weighting
conditions, except in the transition area.
At low airspeeds, the total thrust, represented in fig. 2.4(d), is demanded basically to generate the upward force for weight compensation. At high airspeeds,
however, it is demanded mainly to compensate for the drag forces, which increase with the square of the airspeed [1] and do not depend on the weighting
condition.

2.3.2

Model linearization

The linearization of the system (2.81) corresponds to the first-order term of


its Taylor expansion around the point of interest, in this case, (xe , ue ):


f
f
(x xe ) +
(u ue )
x f (xe , ue ) +
x x=xe ,u=ue
u x=xe ,u=ue

(2.86)

Substituting the jacobian matrices:


f
A
x x=xe ,u=ue

f
B
u

(2.88)

= x xe
x

(2.89)

= u ue
u

(2.90)

(2.87)

x=xe ,u=ue

and the variations around the equilibrium values:

into (2.86) we obtain the airship linear model:


= A
x
x + B
u

(2.91)

in the absence of disturbances (deterministic case).


This procedure, however, is only possible if all terms in the dynamic model
are analytical. Generally, aerodynamic data is included in an aircraft model
in the form of lookup tables. Hence, the linearization must be done numerically, performing a numerical differentiation by finite difference of the nonlinear
equations. Numerical linearization is done perturbing each state or input sig-

2.3. AIRSHIP LINEARIZED MODELS

39

nal at a time, and then computing the resulting accelerations. The elements
of the A and B matrices are approximated as:
fi (xj ) fie
fi

xj
xj
fi
fi (uj ) fie
Bij =

uj
uj
Aij =

(2.92)
(2.93)

where fi (xj ), fi (uj ) are the accelerations at the disturbed state and input,
xj , uj are the perturbation value of the jth state and input and fie is the
value of fi at the equilibrium condition.
The linearized model (2.91), i.e., the dynamic matrix A and the input matrix
B, depend on the trim point chosen for the linearization, and in particular of
the chosen airspeed Vte and altitude he .
Note that, as seen in Section 2.2.1, the airship dynamics represented in the
inertial frame when no wind disturbance is present corresponds to the airship
dynamics represented in the air frame for constant translation wind. This
means the state vectors (2.89) may contain either v or va , as long as there is
no wind or it is a steady translation one.
As a result of the system linearization, and as usual in aeronautics, two independent (decoupled) motions may be considered: the motion in the vertical
plane, named longitudinal, and the motion in the horizontal plane and rolling,
named lateral. The corresponding linearized models will be presented next.
The analysis of the eigenvalues of the dynamic matrices A will allow us to
make an approximate description of the airship stability modes [1].

2.3.2.1

Longitudinal model

T and the input


v = [
In the longitudinal case, the state vector is x
u, w,
q, ]
T , v ]T , where all the variables represent the
v = [e , X
vector is given by u
variations around the trim value. Therefore, the longitudinal dynamic equation
is given by:

u
u

e
w

= Av w + Bv
T
(2.94)

X
q
q

40

CHAPTER 2. THE AIRSHIP MODEL

T = TL + TR is the change in
where e is the change in elevator deflection, X
thrust demand and v is the change in the vectoring angle.
When necessary, the altitude may be introduced as an additional integrating
state of the longitudinal motion, since:
Vte w
h

(2.95)

Sometimes, it is interesting to consider the thrust input expressed in its carte T , v ) by


sian coordinates instead of the polar ones, replacing the pair (X
(Tx , Tz ), which are, respectively, the changes of thrust demand in the forward
and down axes.
Figure 2.5 shows the evolution of the poles of the linearized longitudinal model
with the airspeed Vt .
1.5

Imaginary axis (rad/s)

3

0.5

-0.5

1 surge
:

2 heave / pitch subsidence


:
-1

-1.5
-12

3 longitudinal pendulum
:

-10

-8

-4
-6
Real axis (rad/s)

-2

Figure 2.5: Poles of linearized longitudinal dynamics vs. airspeed Vt (: 0m/s,


: 15m/s).

The surge mode is described by a real pole with a long time constant and is
associated to the forward speed u. The influence of the airspeed increase in this
mode is hardly noticeable. The other real pole, associated with the vertical

41

2.3. AIRSHIP LINEARIZED MODELS

speed w
(or, equivalently, the angle of attack ), describes, at hover (poles
indicated by ), the heave mode. As the airspeed increases (poles indicated
by ), the mode becomes faster and develops into a pitch subsidence. This
mode is described by the faster of the two real poles.
The longitudinal pendulum mode corresponds to the complex pair of poles that
is associated to the pitch angle and the pitch rate q. In the hover condition the
damping is zero and the pendulum oscillation property in this mode becomes
evident. As the airspeed increases the frequency decreases. The damping, on
the other hand, augments, reaching its maximum value in the transition region
with a near coupling of the four modes.
All modes are stable (marginally for air-hover) over the flight envelope.

2.3.2.2

Lateral model

In the lateral case, the state vector considered for the dynamic characteristics
T and the input vector is given by u
h = [
h = [a , r , TD ]T . All the
is x
v , p, r, ]
variables represent the variations around the trim value, which for the lateral
case corresponds to xeh = 0 and ueh = 0 (see Section 2.3.1).
Therefore, the lateral dynamic equation is given by:

v
p
r

= Ah

v
p
r

+ Bh

r
TD

(2.96)

where a is the aileron deflection, r is the rudder deflection and TD is the


differential thrust TD = TL TR between left and right propellers.
When necessary, the yaw angle may be introduced as a supplementary integrating state of the lateral motion, since:
= r/ cos(e )

(2.97)

Figure 2.6 shows the change in the poles of the linearized lateral model with
the airspeed Vt .
In the hover condition (poles indicated by ), the zero damping of the complex
pair of poles characterizes the oscillatory roll mode related to the roll rate p

42

CHAPTER 2. THE AIRSHIP MODEL

3
3

Imaginary axis (rad/s)

-1

-2

-3
-6

1 sideslip subsidence
:
2
: yaw subsidence
3 roll oscillation
:

-5

-4

-2
-3
Real axis (rad/s)

-1

Figure 2.6: Poles of linearized lateral dynamics vs. airspeed Vt (: 0m/s, :


15m/s).
This oscillatory rolling movement is the lateral equivalent
and the roll angle .
of the longitudinal pendulum oscillation and arises for similar reasons, i.e., the
fact that the center of gravity is located below the center of buoyancy of the
airship.
With the increase in airspeed (poles indicated by ), the general stability
improves, since the damping ratio of the oscillatory mode increases and the
eigenvalues of the real modes become more negative. However, in contrast to
what happens in the longitudinal case, the frequency of the oscillatory mode
stays virtually unchanged throughout the speed range and only the damping
ratio undergoes an increase, which in turn appears to be directly proportional
to the speed variation.
Referring to the two real poles, the slow mode, usually named sideslip subsidence mode, is associated with the lateral speed v (or equivalently, the sideslip
and it does not appear to be much affected by the airspeed increase.
angle ),
The fast mode is the yaw subsidence mode related to the yaw rate r and
presents a time constant that decreases with the airspeed.
As in the longitudinal case, all modes are stable (marginally for air-hover) over

2.4. CONCLUSIONS

43

the flight envelope.

2.4

Conclusions

This chapter presented the airship nonlinear model, based on both dynamics
and kinematics analysis. The UAV model introduced considers all forces that
act upon it, namely, aerodynamics, gravity, propulsion, kinematics and wind.
The nonlinear model is, however, too complex to allow a system analysis.
Therefore, a linearization procedure was followed to obtain the linearized models of the decoupled airship motions, lateral and longitudinal. The observation
of the poles location of each individual system as function of airspeed, provides
a good knowledge of the airship behavior over the flight envelope. The fact, for
example, that the damping of the oscillatory roll mode approximates zero near
hover, indicates that this motion should not be overlooked at low airspeeds.
The airship is controlled by the action of two vectored propellers and control
surfaces. It was seen, however, that these actuators authority or influence is
not constant. In fact, it varies as function of the airspeed. This indicates that
the action of the different actuators shall depend not only on the goal mission
which may include, for instance, groundspeed tracking, but also on the wind
disturbances present since they have influence in the resulting airspeed.
All this information and knowledge is undoubtedly essential to the next chapters, where the control of the airship will be addressed.

44

CHAPTER 2. THE AIRSHIP MODEL

Chapter 3
Common Concepts and Tools
Contents
3.1

Position errors . . . . . . . . . . . . . . . . . . . . .

45

3.1.1

Path-following . . . . . . . . . . . . . . . . . . . . . 46

3.1.2

Path-tracking . . . . . . . . . . . . . . . . . . . . . . 46

3.2

Attitude reference and wind estimation . . . . . .

47

3.3

Controllers performance evaluation

49

. . . . . . . .

3.3.1

Case-study mission . . . . . . . . . . . . . . . . . . . 50

3.3.2

Sensitivity and robustness . . . . . . . . . . . . . . . 51

The design, implementation on a simulation environment and evaluation of


control laws that solve the airship path-tracking problem requires the prior
definition of some concepts and tools.
This chapter gathers all the common elements used hereafter. Section 3.1 defines the position errors. Section 3.2 presents the reference attitude considered
in this work and a for wind estimation method. Section 3.3 describes the criteria used as common baseline for the performance evaluation of the different
controllers.

3.1

Position errors

The linearization procedure described in Section 2.3 considers straight level


flights, independent of the N E-direction chosen as reference, as equilibrium
of the linear system (2.91) included the
condition. If the dynamic state x
cartesian position error in the {i} frame, it could only be used with north45

46

CHAPTER 3. COMMON CONCEPTS AND TOOLS

aligned references. By considering the errors in the reference trajectory frame,


the linear model is valid for any horizontal direction chosen. Therefore, it is
necessary to define a coordinate transformation to the reference local frame.
This coordinate transformation depends, however, on the mission objective,
whether it is a path-following or path-tracking case.

3.1.1

Path-following

The path-following problem considers a reference trajectory defined by waypoints, not being therefore time-dependent. This means the lateral and
vertical position errors at a given time are defined as the closest distances
(respectively, in the horizontal and vertical planes) between the airship position
p and the reference trajectory (see fig. 3.1). This reference, given any AB
segment, defines the {r} frame.
vh

vv

B
xr

dh

dv
yr

p
zr

Figure 3.1: Path-following errors definition.


Therefore, the position errors are given by the projections of the vectors d
onto the lines v, perpendicular to the reference line AB:
= |projvh dh | = vhT dh

= |projvv dv | = vvT dv

3.1.2

(3.1)
(3.2)

Path-tracking

The path-tracking problem differs from the path-following one by the fact that
the reference path is in this case time-dependent. The reference coordinates
may be obtained from given way-points if a desired velocity is also set. This is
the case of the missions considered in this work, as will be seen ahead in this
chapter.

47

3.2. ATTITUDE REFERENCE AND WIND ESTIMATION

Consider the airship is at position p(t) when it is supposed to be at the reference position pr (t). This reference point lies in the desired trajectory defined
by the given way-points A and B, which define the reference frame {r} (see
fig. 3.2).
xi

xr
B
r

xl

p(t)

Nr

pr(t)

yl

A
Er

E
yr

yi

Figure 3.2: Path-tracking errors definition (2D).


Therefore, the position errors given in the reference frame are obtained from:

N Nr

= = Sr E Er
D Dr

(3.3)

with Sr = S(r ) defined in (2.17) and r = [0, r , r ]T corresponding to the


angles that define the transformation from {i} to {r} frames, i.e., the angles
the reference inertial velocity p r does with the {i} frame.
Note that when the objective is ground-hover, p r = 0 and r , r are therefore
undetermined. In this case, we arbitrarily define r = r = 0 for simplicity.
This leads to Sr = I and = p pr .

3.2

Attitude reference and wind estimation

The first idea is that the attitude reference shall be coincident with the reference trajectory attitude r . However, there are two situations when this is
not desirable: in the presence of wind disturbances (a certainty when flying
outdoors) and if the objective is ground-hover (since r is arbitrarily defined).
An aircraft of conventional shape must fly against the apparent wind in order
to have low drag [53]. This is also true for airships, moreover because of the

48

CHAPTER 3. COMMON CONCEPTS AND TOOLS

lateral underactuation. So, whenever there is wind, the airship will try to align
itself with the relative air, reducing the sideslip angle (see Appendix A). For
this reason, the relative air attitude ar is chosen as attitude reference.
We compute ar following these four steps (see fig. 3.3):
1. with the airship attitude and the aerodynamic variables Vt , and
(all measured variables), compute de inertial air velocity p a :

Vt

p a = ST va = ST STa 0
0

(3.4)

with Sa defined by (A.1) and S by (2.17);


estimate the wind inertial velocity
2. using the measured inertial velocity p,
vector p w by:
p w = p p a

(3.5)

3. compute the relative air velocity reference as:


p ar = p r p w

(3.6)

4. finally, compute ar = [0, ar , ar ]T , which corresponds to the angles p ar


does with the the {i} frame.
xi

B
.
pw

.
pr

.
p

r
pr

.
pa

.
pw

xl
.
pa

A
yl

yi

Figure 3.3: Wind and yaw reference estimation.


Note that, according to the methodology described:

3.3. CONTROLLERS PERFORMANCE EVALUATION

49

in the no-wind case, p w = 0 and ar is coincident with the reference


trajectory attitude r ;
for the ground-hover objective, p r = 0 implies the reference trajectory
attitude is undefined and ar = w , reducing the lateral effort by
minimizing the drag force.
Remark that ar computed in this way does not consider the equilibrium pitch.
Therefore, to reduce the airship lift, a corrected value shall be used:
a r = ar + e

(3.7)

The corrected attitude reference is then given by:


ar = [0, a r , ar ]T

(3.8)

The wind attitude w = [w , w , w ]T may also be computed, corresponding


to the angles p w does with the {i} frame (with w = 0).

3.3

Controllers performance evaluation

One of the objectives of this work is to compare the performance of the different
control methodologies used. The performance will be evaluated according to
the following three criteria:
Airship behavior for a selected case-study mission. This mission is defined to be representative of a realistic case. The controller performance
will be mainly evaluated by the airship path-tracking errors (see Section 3.1.2) and the actuators request.
Sensitivity and robustness to parameter uncertainty. The controller
should guarantee the stability of the closed-loop system even in the
present of wind disturbances and model parameter uncertainty.
Computational effort. For a real-time implementation to be possible, the
computational time taken by the controller is an important measure of
its performance.

50

3.3.1

CHAPTER 3. COMMON CONCEPTS AND TOOLS

Case-study mission

The simulation examples presented throughout this work consider a 3kg weighting mass. The controllers are implemented at 10Hz. This frequency is high
enough when compared with the frequency of the airship dynamic system,
that the continuous control design is still applicable. The fastest frequency
obtained in open-loop is approximately 2rad/s 0.32Hz for the roll oscillation pendulum.
Although other missions might occasionally be used to demonstrate the controllers performance, the following airship mission, realistic and in agreement
with the airship characteristics, will be used for comparison between controllers. It starts with a vertical take-off, a path-tracking with two semicircles, airship stabilization for ground-hover, and finally a vertical landing
(see fig. 3.4).

60
50

h (m)

40

1
3

30
20
10
0
200

100

0

300
200

100

-100
-200
N (m)

0
-100
E (m)

Figure 3.4: Case-study mission reference. North N , east E and altitude reference (bold) and projections (normal).
The same initial conditions are considered, namely at the position (Ni , Ei , hi ) =
(30, 20, 1)m and with attitude (i , i , i ) = (10, 1, 20)o . The airship has
then 15s to be stabilized at the initial reference point pr0 = (Nr0 , Er0 , hr0 ) =
(30, 20, 5)m so as to be stable and ready to start the mission. From this
point, the vertical take-off at 1m/s climbing rate begins, finishing at the first
point of the horizontal path-tracking pr1 = (Nr1 , Er1 , hr1 ) = (30, 20, 50)m.
At 7m/s groundspeed, the airship is to track a reference path provided, comprised of straight lines and two semi-circles of 200m diameter. Although in
this mission we do not always use a straight line reference, with a groundspeed

3.3. CONTROLLERS PERFORMANCE EVALUATION

51

of 7m/s and a 200m circle radius, the approximation is quite acceptable since
the yaw rate is fairly small. Obviously, the angular reference must be adapted
to the case. When reaching the point pr2 = (Nr2 , Er2 , hr2 ) = (100, 0, 50)m,
the path-tracking gives place to the airship stabilization at the coordinates
pr3 = (Nr3 , Er3 , hr3 ) = (30, 20, 50)m during 40s, preparing it for vertical
landing at pr4 = (Nr4 , Er4 , hr4 ) = (30, 20, 1)m at 0.5m/s descent rate.
In order to test the controllers robustness to wind disturbances, the airship is
submitted to a 4m/s constant wind blowing from northwest at 20o , added to
a 3D 3m/s continuous turbulence.
This mission, defined to be representative and illustrative of a realistic case,
clearly represents a challenge for the automatic control system, as (i) the dynamics varies from air-hover to aerodynamic flight during the path-tracking,
(ii) the wind input has different incidence angles (as the trajectory is circular) and also stochastic components, and (iii) the mission includes vertical
maneuvers.

3.3.2

Sensitivity and robustness to parameter uncertainty

Some of the parameters that describe the airship system are likely to be uncertain. These parameters are mostly the aerodynamic model parameters,
obtained in wind tunnel experiments. The weighting mass or heaviness, which
represents the difference between the weight and buoyancy forces, is also considered, since the equilibrium flight is mostly affected by its value. The parameters for which some uncertainty is assumed are then:
mw - weighting mass;
Clp , CMq , CNr - roll, pitch and yaw damping aerodynamic coefficients;
CD0 , CL0 - drag and lift coefficients;
CM - pitching moment coefficient;
CDi , CY , CL - aerodynamic force coefficient derivatives;
Cl , CM , CM , CM , CM , CN - aerodynamic torque coefficient derivatives;
CLe , CYr , CMe , CNr - aerodynamic input coefficient derivatives.

52

CHAPTER 3. COMMON CONCEPTS AND TOOLS

The mission considered for the evaluation of the controllers sensitivity and
robustness to parameter uncertainty corresponds to a straight line at 50m
altitude aligned with the north axis, which the airship is to follow at 8m/s
groundspeed.
The control laws are designed considering a deterministic model of the airship,
named nominal. However, the real airship system has a wind disturbance
input, since in a real flight wind disturbances are always present. The following
wind perturbation, with two components, is considered:
constant wind blowing from west at 4m/s;
turbulent gust, with an intensity of 3m/s, which is an intermediate value,
in a scale from 0m/s for clear air with no turbulence, to 7m/s for a
hurricane [54].
For the baseline simulation, we consider no error in the model parameters,
only wind disturbance input for the aerodynamic flight described. For selected
variables, we then compare the Root Mean Square (RMS) values obtained in
this baseline simulation with the RMS results obtained repeating the simulation varying each of the above parameters. The parameters vary one at a time
in order to allow the evaluation of the influence of each one.

Chapter 4
Classical Approach: Linear
Control
Contents
4.1

Airspeed and altitude regulation model . . . . . .

55

4.2

Lateral models . . . . . . . . . . . . . . . . . . . . .

56

4.2.1

No-roll approximation . . . . . . . . . . . . . . . . . 56

4.2.2

Space domain approximation . . . . . . . . . . . . . 57

4.3

Linear Quadratic Regulator . . . . . . . . . . . . .

58

4.4

Simulation results . . . . . . . . . . . . . . . . . . .

60

4.5

4.4.1

Airspeed and altitude regulation model . . . . . . . 60

4.4.2

No-roll vs. space domain

. . . . . . . . . . . . . . . 62

Conclusions . . . . . . . . . . . . . . . . . . . . . . .

65

The airship equations of motion, derived following the Lagrangian approach,


were described in Chapter 2. These equations represent the airship dynamics
and were useful in the construction of a simulator [29], necessary to better
understand the behavior of the prototype airship and also to evaluate the
performance of potential controllers.
On the other hand, this 12-state nonlinear model is too complex to allow an
analysis of the airship dynamic characteristics, since most analysis and design
tools require a linear representation of the system. For this reason, Section 2.3
was dedicated to the linearization of the airship model, resulting in the usual
decoupling of the longitudinal and lateral motions [1, 53].
In view of control implementation, and bearing in mind that the linear models
describe the dynamics of the perturbations about a given equilibrium condi53

54

CHAPTER 4. CLASSICAL APPROACH: LINEAR CONTROL

tion, some variations of these models are considered in Section 4.1.


For the design of the linear controllers, several control design methodologies
are available within Classical and Modern Control theories. The latter, however, allows to base the control design directly on the state-variable model, an
important resource for Multi-Input / Multi-Output (MIMO) systems.
Pole Placement (PP) and Linear Quadratic Regulator (LQR) are among the
most popular modern controller design techniques for MIMO systems. Pole
Placement or Eigenvector Assignment [55] allows to allocate the poles of the
MIMO system to desired locations in one step by solving equations for the feedback gains. The airship system requirements, however, are not easily specified
in terms of eigenvalues/vectors. For aircraft design, the desirable pole locations
are usually found in flying qualities specifications, which consider for instance
the size, weight and maneuverability of the aircraft and the pilot workload.
These specifications, however, may not be suitable for an autonomous airship
flight. Moreover, the PP strategy does not confer any stability robustness to
the closed-loop system. This is an important factor since any (linearized or
even nonlinear) model of the system is an approximation of the real nonlinear
airship dynamics. Furthermore, these models are usually deterministic, not
taking into account disturbances such as wind gusts or sensor measurement
noise. Robustness to model parameters errors and to disturbances is, in fact,
a key issue in the choice of the controller.
LQR is the solution to an optimization problem that has some very attractive properties. Namely, the optimal controller automatically ensures a stable
closed-loop system, achieves guaranteed levels of stability robustness, and is
simple to compute. LQR is the control that minimizes a quadratic cost function subject to constraints imposed by the system dynamics. Typically, LQR
controllers design is carried out by choosing values for the design weights,
synthesizing the control law, evaluating how well the control law achieves the
desired robustness and performance, and iterating through the process until a
satisfactory controller is found. The design weights (state and control weighting matrices) are the designers tools to balance the state errors against the
control effort. In the airship control case, the control weighting matrix is a
specially important tool in the sense that it allows the designer to change the
control effort of the different actuators over the flight envelope.
A restrictive aspect of LQR and PP controllers is that they are full state
feedback controllers. This means that every state that appears in the model

55

4.1. AIRSPEED AND ALTITUDE REGULATION MODEL

of the physical system must be measured by a sensor. This is not a problem


for the AURORA airship, since all the state variables may be easily measured.
Due to the clear advantages that the LQR brings for the airship control design,
this will be the technique applied as linear control methodology. Section 4.3
briefly reviews its theory. For a more in depth survey see [56, 55].

4.1

Airspeed and altitude regulation model

The longitudinal model presented in Section 2.3.2.1 considered the state x


v =
T
T

[
u, w,
q, ] and the input u
v = [e , XT , v ] .
With the purpose of maintaining the trim conditions chosen for linearization,
namely a straight level flight at a chosen airspeed Vte and altitude he , the
regulation of these two variables is an important issue.
Assuming either no wind is present or a steady translation one is, we can
substitute the groundspeeds u and w respectively by the airspeeds ua and wa
in the state vector x
v .
Also, as referred in Section 2.3.2.1, the altitude may be added to the model
as an additional integrating state of the longitudinal motion. Using equation (2.95) and for D = h:
w Vte
D

(4.1)

Finally, the airspeed and altitude regulation model is:

u a
w a
q

av11
av21
av31
0
av41

where avij and bvij

av12
av22
av32
1
av42

av13
av23
av33
0
av43

0 av14
0 av24
0 av34
0 Vte
0 av44

ua
wa
q

bv11 bv12 bv13


bv21 bv22 bv23
bv31 bv32 bv33
0
0
0
0
0
0

e
T
X

(4.2)
are the coefficients of the constant matrices Av and Bv

56

CHAPTER 4. CLASSICAL APPROACH: LINEAR CONTROL

defined in (2.94). The state variables are given by:


ua = ua uae = Vt cos cos Vte cos e

(4.3)

wa = wa wae = Vt sin cos Vte sin e

(4.4)

q = q
= e

(4.5)
(4.6)

and is the vertical position error measured in the trajectory reference frame
(see Section 3.1).

4.2

Lateral models

The lateral model used in Section 2.3.2.2 for the dynamic characteristics anal T and the input u
h = [
h = [a , r , TD ]T .
ysis considered the state x
v , p, r, ]
However, two different models will be used for control purposes.

4.2.1

No-roll approximation

The three degrees of freedom approximation that describes the coupling between the yawing and rolling oscillations is called Dutch-roll motion in the
flight control literature [57, 53]. For the airship case we may neglect the
rolling motion, p, 0, and we designate the remaining side slipping and
yawing motions as no-roll mode. In this case, the aileron a , whose function
is to regulate the roll movement, is not used for control. For the same reason,
the differential thrust TD is not used in hover flight (when v 90o ). When
in aerodynamic flight (v 0o ), the differential thrust TD and the rudder r ,
both control the yaw angle. Since at these airspeeds the control surfaces show
a high authority, the rudder will be used over the differential thrust.
Assuming the reference path aligned with north, and for position control or
guidance, this no-roll approximation can be complemented with the lateral
whose dynamic equations are given by:
position E and the yaw angle ,
E Vte sin Vte
= r/ cos
e

(4.7)
(4.8)

For a generic
where e is the trim value of the pitch angle, and sin .

57

4.2. LATERAL MODELS

reference heading, the lateral error is given by , measured in the trajectory


reference frame (see Section 3.1).
The lateral model here named no-roll approximation is then:


v a

r

ah11
ah13
ah31
ah33
0
0
0 1/ cos e

0 0
0 0
0 Vte
0 0

va
r

bh12
bh32
0
0

(4.9)

where ahij and bhij are the coefficients of the Ah and Bh constant matrices
defined in (2.96). The state variables are given by:

4.2.2

va = va vae = Vt sin

(4.10)

r = r
= ar

(4.11)
(4.12)

Space domain approximation

A simpler approach may be obtained for aerodynamic flight if we assume an


additional simplification in the lateral dynamics, considering it as a first order
system relating the yaw rate r and the rudder deflection r :
r kVte r

(4.13)

where the positive constant k is obtained by observing the yaw rate originated
by different values of airspeed and rudder deflection. The negative sign is
due to the convention that a positive rudder deflection leads to a negative
yaw rate. This equation results from simulation and flight observations, which
show that the yaw rate obeys an almost proportional relation with the product
of airspeed and rudder deflection, when in aerodynamic flight.
Substituting this simplified dynamics into (4.8), yields:
(kVte / cos e )r

(4.14)

=
The time derivative of a variable z may be written as the product z = dz
dt
z x
x
where t = u is the longitudinal groundspeed. Assuming u Vte , valid
x t
in the no-wind case, it is possible to rewrite equations (4.7) and (4.14) now in

58

CHAPTER 4. CLASSICAL APPROACH: LINEAR CONTROL

the space domain:


E =
= kv r
where kv = k/ cos e and z =

dz
dx

(4.15)
(4.16)

is the space derivative.

This leads to the following space domain lateral model [58]:


"

"

0 1
0 0

#"

"

0
kv

(4.17)

which, for the assumptions made earlier, is only valid when in aerodynamic
flight.

4.3

Linear Quadratic Regulator

This section briefly describes the LQR theory. The reader is referred to [56, 55]
for a more in depth survey.
Consider the linear time-invariant (LTI) dynamical system:
= A
x
x + B
u,

(t0 ) = x
0
x

(4.18)

is the state, the m1 vector u


is the control input and
where the n1 vector x
is the measured output. If the pair (A, B) is stabilizable1 ,
the p 1 vector y
then there exists a solution to an infinite time LQR problem. The controls will
be state feedbacks of the form (see fig. 4.1(a)):
= K
u
x

(4.19)

where K is the matrix of constant feedback coefficients to be determined by


the design procedure.
The objective of state regulation of the airship is to drive any initial condition
error to zero, thus guaranteeing stability. This may be achieved by selecting
1

The pair (A, B) is stabilizable if there exists a real matrix K such that A BK is
stable.

4.3. LINEAR QUADRATIC REGULATOR

59

that minimizes the quadratic cost function:


the control input u
J=

T R
(
xT Q
x+u
u)dt

(4.20)

t0

is weighted relative to the effort


In this performance index, the size of state x
through the weighting matrices Q and R, respectively
of the control action u
nonnegative and positive definite matrices. The minimization of J is a generalized minimum energy problem. The objective is to minimize the energy in
the states without using too much control energy. A larger control-weighting
, while a larger state-weighting
matrix R leads to a smaller control action u
go to zero more quickly with time.
matrix Q makes x
The LQR problem with state feedback is the following: given the linear system (4.18), find the Kalman gain matrix K in the control input (4.19) that
minimizes the value of the quadratic cost functional (4.20). This is achieved
by solving the algebraic Riccati equation for the symmetric positive definite
matrix P:
PA + AT P PBR1 BT P + Q = 0
(4.21)
Finally, the Kalman gain K is computed by:
K = R1 BT P

(4.22)

Under the assumptions made above a unique solution exists, and the closedloop dynamics, obtained by substituting (4.19) into (4.18):
= (A BK) x

(4.23)

are guaranteed to be stable. The dynamic matrix of the closed-loop system is


given by:
Ac = A BK

(4.24)

While the design of the linear controller only involves the state and input
), its implementation produces and requires the complete
variations (
x and u
variables (x and u). For instance, the actuation request u has a feedback
and a feedforward component ue , as illustrated in fig. 4.1(b).
component u

60

CHAPTER 4. CLASSICAL APPROACH: LINEAR CONTROL

xe

= A
x
x + B
u

ue

u
- l
-

++

K 

(a) Design.

airship

?
- l

K 

(b) Implementation.

Figure 4.1: Linear control block diagrams.

4.4

Simulation results

This section presents illustrative simulation results of the airship linear control
using the longitudinal and lateral models presented, and implemented according to the block diagram in fig. 4.1(b).

4.4.1

Airspeed and altitude regulation model

This section focuses on the longitudinal control using model (4.2). The purpose
is airspeed and altitude regulation, so the desired operating conditions are
maintained.
The chosen equilibrium values are 10m/s for airspeed and 50m for altitude.
This corresponds to an aerodynamic flight. In order to observe only the longitudinal behavior, the trajectory coincides with a straight line. The simulation
starts with no wind, and at t = 20s a 3m/s wind starts blowing from north.
Figure 4.2 shows the cartesian position variables. The motion is only made
along the longitudinal plane, as may be noticed by the trajectory and altitude
graphics. The altitude regulation is well achieved, with an error inferior to
0.6m. However, it presents a small static error, which (if considered significant)
may be canceled including an integrator state in the model.
The state Q and input R weighting matrices are set as (in SI units):
R
) = diag(1, 1, 1, 1, 1, 0.1)
Q(
ua , wa , q, , ,
T , v ) = diag(500, 0.1, 1000)
R(e , X

(4.25)
(4.26)

The longitudinal groundspeed u and airspeed Vt are represented in fig. 4.3.


Even with the wind perturbation at t = 20s, the airspeed regulation at 10m/s
is well accomplished (not presenting static error, and therefore not needing

61

4.4. SIMULATION RESULTS

50.6
50.4
h (m)

N (m)

400
200

50.2
50.0

0
-200

0
E (m)

49.8

200

40
20
Time (s)

60

Figure 4.2: N E trajectory and altitude h (. equilibrium value, real value


without integrator, real value with D integrator) for airspeed and altitude
regulation.

11

14

10

13
Vt (m/s)

u (m/s)

integrator), to the obvious expense of the groundspeed reduction.

9
8

11
10

7
6

12

20
40
Time (s)

60

40
20
Time (s)

60

Figure 4.3: Longitudinal groundspeed u and airspeed Vt (. equilibrium value,


real value without integrator, real value with D integrator) for airspeed
and altitude regulation.

Finally, in fig. 4.4 it is possible to see the control action. As expected, only
the engines thrust and the elevator deflection are necessary to accomplish the
airspeed and altitude regulation objectives (the vectoring angle variation is
negligible). In face of the wind incidence, the thrust XT is reduced until
the equilibrium airspeed value is attained, which is possible by reducing the
groundspeed. The engines vectoring v is close to zero as expected for an
aerodynamic flight. The rudder deflection is zero, since no lateral control is
used.

62

32

0.76

31

0.74
v (deg)

XT (N )

CHAPTER 4. CLASSICAL APPROACH: LINEAR CONTROL

30
29

20
40
Time (s)

0.66

60

40
20
Time (s)

60

40
20
Time (s)

60

0.5
r (deg)

e (deg)

1.0

3
2

0.0
-0.5

1
0

0.70
0.68

28
27

0.72

20
40
Time (s)

60

-1.0

Figure 4.4: Control action for airspeed and altitude regulation: total thrust XT ,
vectoring angle v , elevator e and rudder r (. equilibrium value, real
value without integrator, real value with D integrator).

4.4.2

No-roll vs. space domain

The purpose of this section is to compare the performance of the lateral approximation models, no-roll approximation and space domain. The equilibrium conditions, necessary to the validity of the models, are guaranteed by the
longitudinal controller, regulating both airspeed and altitude.
The simulation considers an aerodynamic flight at Vte = 10m/s and De =
50m subject to three tests:
initial alignment on a straight line segment, with no wind incidence.
The airship starts deviated from the reference trajectory at (Ni , Ei ) =
(0, 10)m and with an orientation i = 10o ;
reference trajectory following. The airship has to track a two-segment
trajectory in the shape of a 50o elbow, corresponding to the crossing of
a route way-point, again with no wind incidence;
robustness to disturbances. At t = 60s wind starts blowing from northwest at 3m/s.

63

4.4. SIMULATION RESULTS

For the no-roll approximation, the state Q and input R weighting matrices
are set as (in SI units):
= diag(1, 1, 1, 1)
Q(
va , r, , )
R(r ) = diag(1000)

(4.27)
(4.28)

while for the space domain approximation they are set as:
= diag(1, 1)
Q(, )
R(r ) = diag(1000)

(4.29)
(4.30)

The N E trajectory, lateral position error and orientation may be seen in


fig. 4.5. Both approximations allow a good lateral control, with the airship
10
800
(m)

700
600

N (m)

500

0
-10
-20

50
Time (s)

100

50
Time (s)

100

400
80

300

60
(deg)

200
100

40
20
0

0
0

200
400
E (m)

600

-20

Figure 4.5: North-east trajectory, lateral error and yaw angle (. reference trajectory, no-roll approximation, space domain approximation).
always being able to track the reference path, after the initial deviation and
orientation are corrected. The space domain approximation presents slightly
higher errors and slower corrections than the no-roll model, namely when reacting to the wind disturbance step.
Note that when the wind starts blowing at t = 60s, the reference yaw angle is
no longer the trajectory reference angle r but something in between it and the

64

CHAPTER 4. CLASSICAL APPROACH: LINEAR CONTROL

blowing wind (w = 45o ). This yaw reference corresponds to the yaw angle
the relative air inertial velocity p ar does with the {i} frame, ar , as described
in Section 3.2. Allowing the airship to align itself with the relative air reduces
the drag force and minimizes the sideslip angle , as seen in fig. 4.6. In the
no-wind case, ar r .
The lateral control actions (only the rudder r is used in both cases) are very
similar for both cases (see fig. 4.6). Although it is the no-roll approximation
which requests higher rudder deflections, both approximations reach the lower
saturation limit in the elbow curve.
5

(deg)

0
-5
-10
-15
-20

100

50
Time (s)
5

30

10

(deg)

r (deg)

20

0
-10

-20
-30

50
Time (s)

100

-5

50
Time (s)

100

Figure 4.6: Sideslip angle , rudder deflection r and roll angle ( no-roll
approximation, space domain approximation).

In fig. 4.6 it is also possible to see the roll angle . For both models, the
maximum values of the roll angle confirm the validity of the approximation
made, neglecting the rolling motion.
The small differences between the results obtained with both approximations
and the fact that the space domain model only requires the measurement of
two variables, leads to a major advantage of this approximation over the no-roll
one. On the other hand, the latter one is valid over the entire flight envelope,
while the space domain model is only applicable when in aerodynamic flight.

4.5. CONCLUSIONS

4.5

65

Conclusions

This chapter analyzes the linear control of the AURORA airship using the
Linear Quadratic Regulator.
The use of linear control is obviously limited to the existence of linear models
of the airship. For this reason, the linearization of the nonlinear system around
a given equilibrium condition is performed, as described in Section 2.3. The
linearization, as usual in aerial systems, leads to simplified decoupled longitudinal and lateral linear models of the airship motion.
In order for the linear models to be valid, the equilibrium condition for which
the linearization is performed must be guaranteed. Usually, the trim coincides
with a straight level flight with no wind incidence. Therefore, the linear control
so that
presented in this chapter aims at the regulation of the state variables x
x xe . The longitudinal controller is responsible to regulate both airspeed
and altitude, while the lateral approximations are in charge of correcting the
lateral error and the yaw angle.
Simulation results demonstrate the good performance of the LQ regulator applied to the three models, even in the presence of wind disturbances. Position
errors are corrected, and so is the airspeed, to compensate for the variation of
wind.
However, linear control is limited, since the controller designed for a given
system is only valid in the vicinity of the equilibrium condition considered.
For instance, mission objectives like ground-hover that involve groundspeed
regulation are not possible, since the equilibrium airspeed would be unknown
a priori in the presence of wind.
This problem is solved if the linear systems and controllers are not fixed but
change with the measured airspeed and altitude (which define an equilibrium
condition). This is the idea of the gain scheduling technique, which extends the
validity of the linearization approach to a range of operating points, instead of
a single one. This control technique, already considered a nonlinear one, shall
be presented in the second part of this work.

66

CHAPTER 4. CLASSICAL APPROACH: LINEAR CONTROL

Chapter 5
Gain Scheduling
Contents
5.1

More linear models . . . . . . . . . . . . . . . . . .

68

5.1.1

Groundspeed and altitude regulation . . . . . . . . . 68

5.1.2

Complete 12-states linear model . . . . . . . . . . . 69

5.2

Scheduling . . . . . . . . . . . . . . . . . . . . . . . .

72

5.3

Robustness analysis . . . . . . . . . . . . . . . . . .

75

5.4

5.5

5.3.1

Performance robustness . . . . . . . . . . . . . . . . 75

5.3.2

Stability robustness . . . . . . . . . . . . . . . . . . 80

Simulation results . . . . . . . . . . . . . . . . . . .

87

5.4.1

Groundspeed and altitude regulation . . . . . . . . . 87

5.4.2

Case-study mission . . . . . . . . . . . . . . . . . . . 89

5.4.3

Sensitivity and robustness . . . . . . . . . . . . . . . 93

Conclusions . . . . . . . . . . . . . . . . . . . . . . .

97

Today, Gain Scheduling is the prevailing flight control design methodology [59].
While the linear control presented in the previous chapter is only valid around
a single equilibrium condition, this conventional solution performs point designs for a large set of trim conditions and then constructs a gain schedule by
considering gains with respect to flight conditions.
Using several linear models to describe the aircraft dynamics over the flight
envelope, allows the control designer to make use of all the classical design and
analysis tools. For this reason, gain scheduling is the obvious next step in the
AURORA airship control. It will hopefully provide a better knowledge of the
control design issues, as well of possible solutions.
Depending on the process considered for linearization, it is sometimes possible
67

68

CHAPTER 5. GAIN SCHEDULING

to find auxiliary variables that correlate well with the changes in the process
dynamics. In the airship case, these variables mostly correspond to the airspeed and altitude. Still, the altitude influence may be disregarded for low
altitude flights where the envelope pressure is kept practically constant. For
this reason, from now on we will only consider the airspeed as trim parameter.
This chapter presents the gain scheduling approach applied to the airship pathtracking control. In Section 5.1 two new linear models are introduced, one for
the longitudinal motion only, the other considering all the 12-states, combining the longitudinal and lateral motions in the same model. Section 5.2 describes the scheduling procedure and in Section 5.3 the closed-loop robustness
is analyzed, considering both performance and stability issues. Finally, some
simulation results are presented in Section 5.4, followed by some conclusions
in Section 5.5.

5.1

More linear models

This section introduces alternative linear models which are a natural evolution
of the previously presented ones when considering gain scheduling.

5.1.1

Groundspeed and altitude regulation

Although airspeed regulation is important to assure the linearized model is


valid, inspection oriented applications usually demand control of the groundspeed. Also, for tracking purposes, the north or longitudinal position N should
also be considered. Assuming the reference path aligned with north, its dynamics may approximately be given by:
cos e u + sin e w u + e w
N

(5.1)

For a generic reference heading, the longitudinal error is given by , measured


in the trajectory reference frame (see Section 3.1).
The complete groundspeed regulation model, which in fact includes all 6 longi-

69

5.1. MORE LINEAR MODELS

tudinal variables, is then:

u a
w a
q

av11
av21
av31
1
0
av41

av12
av22
av32
e
1
av42

av13
av23
av33
0
0
av43

0
0
0
0
0
0

0 av14
0 av24
0 av34
0
0
0 Vte
0 av44

bv11 bv12 bv13


bv21 bv22 bv23

b
v31 bv32 bv33
XT
+
0
0
0


v
0
0
0


0
0
0
(5.2)
constant matrices Av and Bv

ua
wa
q

where avij and bvij are the coefficients of the


defined in (2.94). The state variables are given by:
ua = ua uar

(5.3)

wa = wa war

(5.4)

q = q
=

ar

(5.5)
(5.6)

and is the vertical position error, measured in the trajectory reference frame
(see Section 3.1).
Given the desired groundspeed reference p r and having estimated previously
the wind velocity p w , the air velocity reference var = [uar , var , war ]T is obtained
from (see Section 3.2):
var = Sar (p r p w )

(5.7)

with Sar = S(ar ) and ar given by (3.8). The equilibrium values xe and
ue , as well as the linear matrices Av and Bv , are obtained in real-time from a
lookup table, function of the measured airspeed Vt .

5.1.2

Complete 12-states linear model

The next and final linear model evolution is obviously the complete 12-states
model. Although the longitudinal and lateral motions are still decoupled (a
consequence of the linearization), the advantage of joining both models in a
single one is having also a single controller instead of two. This requires that
all state variables are measured so the full state feedback is possible. This is
the AURORA case (see Section 2.1), otherwise a state estimator may be used.

70

CHAPTER 5. GAIN SCHEDULING

The complete 12-states linear model is then given by:

a
v

a
v

+ B
= A

e
T
X
v
a
r
TD

(5.8)

with:
a = va var
v
= r

= [, , ]T
=

ar

with var given by (5.7)

(5.9)

with r = [0, 0, rr ]T

(5.10)

given by (3.3)

(5.11)

with ar given by (3.8)

(5.12)

Note that, although we have assumed a rectilinear reference path for the linearization, the approach may also be extended to the cases where the reference path varies slowly, with negligible derivatives when compared to the state
derivative. In this case, the angular velocity reference must be adapted to the
case, reason for which we considered r 6= 0.
Again, the equilibrium values xe and ue , as well as the linear matrices A and
B, are obtained in real-time from a lookup table, function of the measured
airspeed Vt .
The evaluation of the longitudinal and lateral decoupling due to the lin =
earization is easily noticeable if we rearrange the A matrix considering x
Tlat ]T . In fact, the A matrix can be partitioned into four distinct sub[
xTlong , x
matrices as:
#
"
A
0
long
(5.13)
A =
0
Alat
where all elements in the lower left submatrix are zero, while some few in the
upper right submatrix are not exactly zero but of smaller magnitude when
compared with the elements in Along and Alat .
state. The resulting
We can also rearrange the B matrix lines considering the x

71

5.1. MORE LINEAR MODELS

B matrix can also be partitioned into four distinct submatrices as:


B =

"

Blong 0
0
Blat

(5.14)

The elements relative to either longitudinal and lateral position variables are
obviously null since the kinematics does not directly depend on the input. The
lower left and upper right submatrices are null. This shows the decoupling be t , v ]T
long = [e , X
tween what we already called longitudinal actuators inputs u
]
T , as well
long = [e , Tx , Tz ]T ) and the lateral states x
lat = [
(or u
va , p, r, , ,
lat = [a , r , TD ]T and the longitudinal states
as of the lateral actuators input u
T.
long = [
x
ua , wa , q, , , ]
Figure 5.1 describes the evolution of the 12-states linear model poles with the
airspeed. As expected, 8 of the 12 modes match the longitudinal and lateral
ones described in sections 2.3.2.1 and 2.3.2.2, respectively.
3
3 lat

2
3 long

Imaginary axis

1 long

1 lat

2 long

-1

-2

-3
-12

2 lat

1 long : surge

2 long : heave/pitch subsidence

3 long : longitudinal pendulum

1 lat : sideslip subsidence

2 lat : yaw subsidence

3 lat : roll oscillation

-10

-8

-4
-6
Real axis

-2

Figure 5.1: Poles of linearized dynamics vs. airspeed Vt (: 0m/s, : 15m/s).

The remaining 4 modes rest in the origin over the entire flight envelope. This
is expected since they correspond to the 4 natural position integrators N, E, D

72

CHAPTER 5. GAIN SCHEDULING

and . Although unstable, the system is controllable since:


rank(B AB . . . An1 B) = 12

5.2

(5.15)

Scheduling

The linear models presented in Section 5.1 vary with the airspeed (we are
assuming the altitude influence may be disregarded for low altitude flights
where the envelope pressure is kept practically constant). The airspeed is thus
the considered scheduling variable. In applications of gain scheduling, the
practice has been that one can schedule on time-varying variables, as is the
case of the airspeed, as long as they are slow enough relative to the dynamics
of the system [60].
The gain scheduling methodology may be defined as a routine of four steps,
executed for different airspeeds over the flight envelope:
1. definition of the equilibrium condition;
2. linearization of the nonlinear system equations around the trim;
3. computation of the control gain;
4. computation of the control input obtained from full state feedback.
At each sampling time t, the true airspeed Vt is measured. The first step is
then to obtain the corresponding equilibrium variables, xe and ue , solving the
optimization problem described in Section 2.3.1. We may then proceed to the
second step, where the nonlinear system is numerically linearized about the
trim, obtaining the A and B system matrices (see Section 2.3.2). We now have
almost all variables necessary to complete the third step, where we compute
the control gain K, solution of the LQR problem described in Section 4.3.
First we have to define the state Q and input R weighting matrices.
In order to have an idea of the actuators influence on the system dynamics, the evolution of the B matrix coefficients with the airspeed are shown in
fig. 5.2. The more obvious conclusion is that the longitudinal/lateral actuators
only influence the longitudinal/lateral states. In fact, observing fig. 5.2(a), we
note that the coefficients relative to the lateral states v, p, r are null, while in
q coefficients null.
fig. 5.2(b) we have the longitudinal states u, w,

73

5.2. SCHEDULING

10
B(a )

B(e )

0
0
-10

10

-20
-40

15

Vt (m/s)

0.00
0

10

-10

15

15

10

15

Vt (m/s)

0.0

0.05
B(TD )

B(v )

10

Vt (m/s)

-0.5
-1.0

15

10

0.05

B(r )

T )
B(X

0.10

-0.05

10
Vt (m/s)

10

15

Vt (m/s)

(a) Longitudinal actuators ( v, p, r,


u
, : w,
. q).

0.00

-0.05

5
Vt (m/s)

, w,
q, v, :
(b) Lateral actuators ( u
p, . r).

Figure 5.2: Evolution of B matrix coefficients with airspeed.

We also observe that the influence of the actuators depends on the available
airspeed. For instance, the reduced authority of the control surfaces e , a
and r is noticeable at low airspeeds, where the respective coefficients are null.
T ) coefficients evolution, we note the total
From the observation of the B(X
T mostly influences the longitudinal speed u in aerodynamic flight,
thrust X
while at low airspeeds the influence is on the vertical speed w.
This is justified
o
o
for the change of the vectoring angle v from 0 to 90 . The differential thrust
TD makes its more significant contribution at low airspeeds, being the sole
responsible for the rolling motion control.
For ease of implementation, we have chosen the cartesian (Tx , Tz ) version of
the thrust longitudinal input over the polar (XT , v ) one. This way we consider
the forces produced by the propellers and the forces produced by the control
surfaces deflection. The evolution of the respective B coefficients is represented
in fig. 5.3.
We can then see that the B matrix coefficients have the necessary information
on the actuators behavior over the flight envelope. Therefore, the R matrix
may be constant with Vt , only having to be adjusted to the different types
(units) of actuators presents, namely engines vs. control surfaces. Although
being constant, the R matrix coefficients are chosen by the designer (with
an iterative process) such that the performance of the closed-loop system can
satisfy the desired requirements.

74

CHAPTER 5. GAIN SCHEDULING

B(Tx )

0.1
0.0
-0.1

10

15

10

15

Vt (m/s)

B(Tz )

0.05
0.00

-0.05

5
Vt (m/s)

Figure 5.3: Evolution of B matrix Tx and Tz coefficients with airspeed (


v, p, r, u, : w,
. q).
The weighting matrix R is then defined as:
R(e , Tx , Tz , a , r , TD ) = diag(1000, 0.1, 0.1, 5000, 1000, 0.5)

(5.16)

where the different weights between a and e , r means there is a preference


of these two over the former. The same happens with TD and Tx , Tz .
The weighting matrix Q is defined as:
,
)
= diag(1, 1, 1, 1, 1, 1, 10, 10, 10, 1, 1, 1) (5.17)
Q(
u, v, w,
p, q, r, , , , ,
where the higher weights of , and indicate the request of a faster correction
of these three variables over the remaining ones.
The Kalman gain K may now be computed from (4.22). Considering a realtime implementation, the three first steps may be condensed, taking xe , ue and
K from a lookup-table that gathers all these variables for different airspeeds
over the flight envelope. Lastly, the fourth and final step, the control input u is
obtained from (4.19)-(2.90). The gain scheduling diagram block is represented
is fig. 5.4.
regulator
operating
parameters scheduling  condition
xe , ue , K
Vt

? control
- regulator signal- airship
u

output
y

state
x

Figure 5.4: Gain scheduling diagram block.

We have seen in Section 4.3 that the optimal gains at each gain scheduling point

75

5.3. ROBUSTNESS ANALYSIS

guarantee the stability of the closed-loop system. However, they should also
guarantee robust stability and performance. This means, they should guarantee stability and good performance at points near the design equilibrium point.
Such robustness can be verified after the LQR design by using multivariable
frequency-domain techniques [53]. This shall be done in the sequence.

5.3

Robustness analysis

In this section, we analyze the robustness of the closed-loop 12-states system.


Although what implementation concerns the result is the same as the full
state feedback described in Section 4.3, for the robustness analysis we will
consider the closed-loop nominal system represented in fig. 5.5. It comprises
T ]T and an external feedback
int = [
T,
an internal feedback look of y
vaT ,
p = .
loop of the cartesian position errors y
Gp
p - ju
pu
0 - je

Kp
x = A
x + B
u
+
+
6
6

x
Cp

p y

int
y
Kint 
Cint 

Figure 5.5: Closed-loop nominal system.

In the following, two important robustness issues will be analyzed:


performance robustness, which is the ability to guarantee acceptable performance even though the system may be subject to disturbances;
stability robustness, which is the capacity to provide stability in spite of
modeling errors due to incorrect dynamics coefficients identification and
plant parameter variations.

5.3.1

Performance robustness

We begin with the analysis of the closed-loop system performance robustness


according to [53].
Consider the general closed-loop system represented in fig. 5.6. The plant
is G(s), and K(s) is the compensator. The plant output is z(t), the plant

76

CHAPTER 5. GAIN SCHEDULING

control input is u(t), and the reference input, null for the regulation problem,
is r(t). The uncertainties are characterized by a disturbance d(t) acting on the
d(s)

d
-
+ 
6

r(s)

e (s)

K(s)

u(s)

G(s)

+?
+-


+

?
+




z(s)

n(s)

Figure 5.6: Disturbed feedback control system.

system (wind gusts, for instance), and sensors measurement noise n(t). The
disturbances occur typically at low frequencies, below some value d , while the
measurement noise has its predominant effect at high frequencies, above some
value n .
The regulation error is:
e(t) = z(t)

(5.18)

Due to the presence of noise n(t), e(t) may not be represented in fig. 5.6. The
signal ed (t) is given by:
ed (t) = z(t) n(t) = e(t) n(t)

(5.19)

In terms of Laplace transforms, we may take the following relations regarding


the closed-loop system:
z(s) = G(s)K(s)ed (s) + d(s)
ed (s) = e(s) n(s)
e(s) = z(s)

(5.20)
(5.21)
(5.22)

We can rewrite z(s) and e(s) as:


z(s) = T(s)n(s) + S(s)d(s)

(5.23)

e(s) = S(s)d(s) + T(s)n(s)

(5.24)

defining the system sensitivity as:


S(s) = (I + GK)1

(5.25)

77

5.3. ROBUSTNESS ANALYSIS

and:
T(s) = GK(I + GK)1 = (I + GK)1 GK

(5.26)

S(s) + T(s) = I

(5.27)

Since:

T(s) is called the complementary sensitivity, or cosensitivity. The loop gain is


G(s)K(s).
To ensure small regulation errors, we must have S(j) small at those frequencies where the disturbance d(t) is large. This will yield good disturbance
rejection. On the other hand, for satisfactory sensor noise rejection, we should
have T(j) small at those frequencies where n(t) is large.
We will now make use of the singular values to obtain performance specifications in the frequency domain. In fact, for any input, the magnitude of a
transfer function H(j) at any given frequency , may be bounded above by
its maximum singular value, denoted (H(j)), and below by its minimum
singular value, denoted (H(j)). Therefore, our results need only take into
account these two constraining values of magnitude.
Some facts we shall use in this discussion are:
(GK) 1 (I + GK) (GK) + 1

(5.28)

(M) = 1/(M1 )

(5.29)

(AB) (A)(B)

(5.30)

for any matrices A, B, GK and M, with M nonsingular. Let us also define


the L2 operator gain, denoted ||H||2 , as:
||H||2 = max[(H(j))]

(5.31)

Next, we consider the low and high-frequency specifications on the singular


value plot.

78

CHAPTER 5. GAIN SCHEDULING

Low-frequency specifications For low-frequencies, let us suppose that the


sensor noise n(t) is zero so that (5.24) becomes:
e(s) = S(s)d(s)

(5.32)

Thus, to keep the regulation norm ||e(t)||2 small, it is only necessary to ensure
that the L2 operator norm ||S||2 is small at all frequencies where d(j) is
appreciable. This may be achieved by ensuring that, at such frequencies,
(S(j)) is small. So, since at low frequencies (see fig. 5.7(a)):
(S) = [(I + GK)1 ] =

1
1

(I + GK)
(GK)

(5.33)

this may be guaranteed if we select:


(GK(j)) 1, for d

(5.34)

where d(s) is significant for d .

High-frequency specifications We now turn to the high-frequency performance specifications. The sensor noise is generally appreciable at frequencies
above some known value n . Thus, according to (5.24), to keep the regulation
norm ||e(t)||2 small in face of measurement noise, we should ensure the operator norm ||T||2 is small at high frequencies above this value. Since at high
frequencies (see fig. 5.7(b)):
(T) = [GK(I + GK)1 ] (GK)

(5.35)

this is guaranteed if:


(GK(j)) 1, for n

(5.36)

Figure 5.8 represents the singular values of the closed-loop system gain of
the nominal system in fig. 5.5, together with a graphical representation of
conditions (5.34) and (5.36). The minimum (GK) and maximum (GK)
singular values, are represented respectively in figs. 5.8(a) and 5.8(b). Referring to our nominal system represented in fig. 5.5, the loop gain corresponds
to GK = Gp Kp , where Gp is the transfer-function matrix of the inner-loop
delimited by a dashed line. Since different equilibrium conditions, i.e., different

79

150

150

100

100

50

50
Magnitude (dB)

Magnitude (dB)

5.3. ROBUSTNESS ANALYSIS

-50

-50

-100

-100

(S(j))

(T(j))

1/(GK(j)) --150
103

10

10

(GK(j)) -0

10
10
Frequency (rad/s)

10

10

-150
103

(a) (S(j)) and 1/(GK(j)).

10

100
101
101
Frequency (rad/s)

102

103

(b) (T(j)) and (GK(j)).

Figure 5.7: Singular values relations for Vt = 15m/s.


airspeeds Vt , lead to different linear models and respective controllers, fig. 5.8
contains the singular values of the different systems over the flight envelope
range 0 Vt 15 m/s (steps of 0.5m/s).
60

60

40

40

20

20

0
Low-frequency
condition

-40

 Vt

-60
-80

-60
-80
-100

-120

-120

-140

-140

-160

-160
102

High-frequency
condition

-40

-100

-180
103

Vt

-20
(GK) (dB)

(GK) (dB)

-20

101
101
100
Frequency (rad/s)

102

(a) Minimum singular values, (GK).

103

-180
103

102

101
101
100
Frequency (rad/s)

102

103

(b) Maximum singular values, (GK).

Figure 5.8: Frequency analysis (singular values) of the MIMO nominal system
over the flight envelope 0 Vt 15m/s (steps of 0.5m/s).
Observing fig. 5.8 we see that the cutting frequency of (GK) varies between
0.05 and 0.4rad/s, while for (GK) it varies between 0.3 and 0.5rad/s. According to conditions (5.34) and (5.36), we need to know what is the frequency
range where wind disturbances and measurement noise are significant (d and
n , respectively).
We start by determining d . In order to represent the perturbations introduced
in the airship system by the nonhomogeneous properties of the surrounding air,

80

CHAPTER 5. GAIN SCHEDULING

a turbulence model is used. The model chosen to represent the atmospheric


turbulence is the spectral model of Dryden [54]. The perturbation introduced
is represented as gust velocity components constant in time but spatially distributed. Appendix B describes how to obtain the turbulence velocity from
filtered white noise. Observing the filters transfer-functions (B.2)-(B.4), we
notice that the higher cut-off frequency is the one from Gw , which leads to:
d cGw =

Vt
h

(5.37)

Notice that this is a conservative value 1 that depends on the airspeed and
altitude of the airship. So one solution is to evaluate the value of (GK(j))
at = d . The magnitude of (GK(jd )) corresponds to the factor by which
wind disturbances will be attenuated (if positive) or amplified (if negative).
Figure 5.9(a) describes the variation of (GK(jd )) with airspeed Vt and altitude h, which we see is always positive. We also see that, the higher the altitude, the higher the wind disturbances attenuation. The magnitudes of (GK)
and of the gust transfer functions (B.2)-(B.4) are represented in fig. 5.9(b) for
Vt = 0.5m/s and in fig. 5.9(c) for Vt = 15m/s, both for h = 50m. These
figures show us that, in fact, (GK)(j) 1 for < d , which we may
consider here as : 20 log10 |Gi (j)| > 0. In face of this analysis, we conclude
the closed-loop system is robust to wind disturbances, i.e., condition (5.34) is
verified.
Measurement noise is well attenuated, since usually n 1 rad/s. At frequencies > n , (GK()) 0 dB (see fig. 5.8), which validates condition (5.36).
So far we analyzed the performance robustness of the closed-loop system. We
now proceed to the evaluation of its stability robustness.

5.3.2

Stability robustness

It is unusual for the plant model to be exactly known. Two basic sorts of
modeling errors are incorrect dynamics coefficients identification and plant
parameter variation. It is therefore important to determine if the closed-loop
system remains stable in the case these errors occur, i.e., if the system is
robustly stable.
1

Considering (5.37) valid beyond the ground boundary layer, we reasonably assume h 0
so that d 9 .

81

5.3. ROBUSTNESS ANALYSIS

35

30

(GK(jd )) (dB)

25

20



15













h = 50 500m

10

10

15

Vt (m/s)

60

60

40

40

20

20

-20

-20

Magnitude (dB)

Magnitude (dB)

(a) Wind disturbance attenuation.

-40
-60
-80
-100
-120

20 log10 |Gu | :
20 log10 |Gv |
20 log10 |Gw | .
20 log10 (GK)

-60
-80
-100
-120

-140
-160
103

-40

20 log10 |Gu | :
20 log10 |Gv |
20 log10 |Gw | .
20 log10 (GK)

-140
102

101
101
100
Frequency (rad/s)

102

(b) Vt = 0.5m/s and h = 50m.

103

-160
103

102

100
101
101
Frequency (rad/s)

102

103

(c) Vt = 15m/s and h = 50m.

Figure 5.9: Frequency-domain performance specifications - disturbance rejection.


Stability robustness to incorrect dynamics coefficients identification
The airship motion is described by nonlinear dynamic and cinematic equations.
While the kinematics is quite exact, in the dynamics model case this is not so
straightforward. Some model parameters are not so easy to measure, namely
the weighting mass (difference between airship weight and buoyancy) and the
aerodynamic coefficients, usually identified in wind tunnel tests. Obviously,
an incorrect parameter identification will lead to an incorrect dynamic model.
While the closed-loop system is surely stable for the nominal plant, what
happens if the actual plant is different?
We analyze here the closed-loop system stability robustness to incorrect dynamics coefficients identification. We shall do so using the Matlabr Robust
Control Toolbox [61].

82

CHAPTER 5. GAIN SCHEDULING

We consider here that the airship nominal linear model (5.8) has a parameter uncertainty of 20%, i.e., each nonzero coefficient of the A and B matrices
has 20% uncertainty over its nominal value. Although having the same maximum value, the uncertainty of each coefficient is defined independently of the
uncertainty of the remaining coefficients.
The performance of a nominally stable uncertain system will generally degrade
for specific values of its uncertain elements. Moreover, the maximum possible
degradation increases as the uncertain elements are allowed to further and
further deviate from their nominal values.
The robust stability margin is the size of the smallest deviation from nominal of the uncertain elements that leads to system instability, and allows us
to evaluate the stability robustness of uncertain systems. A nominally stable
uncertain system is generally unstable for specific values of its uncertain elements. If the uncertain system is stable for all values of uncertain elements
within their allowable ranges, the uncertain system is robustly stable. Conversely, if there is a combination of element values that cause instability, and
all lie within their allowable ranges, then the uncertain system is not robustly
stable. A robust stability margin greater than one means that the uncertain
system is stable for all values of its modeled uncertainty. A robust stability
margin less than one implies that certain allowable values of the uncertain
elements, within their specified ranges, lead to instability.
As with other uncertain-system analysis tools, only bounds on the exact stability margin are computed. The precise value is guaranteed to lie between these
upper and lower bounds. Figure 5.10 expresses the evolution of the lower and
upper bounds of the stability margin with the airspeed Vt in steps of 0.25m/s.
In the analysis we consider the worst scenario described by the lower bound.
We can see that for airspeeds between 2 and 4.5m/s and above 8.25m/s the
20% uncertain system is robustly stable since both bounds are higher than 1.
A margin of 1.3, for example, implies that the uncertain system remains stable
for all values of uncertain elements up to 30% outside their modeled uncertain
ranges (i.e., in this case the system is robustly stable for an uncertainty of 26%
over all matrices coefficients).
For the remaining airspeeds the stability margin is less than one, which means
the system is not robustly stable for some values of uncertainty. A margin
of 0.5, for instance, implies the uncertain system remains stable for all values

83

5.3. ROBUSTNESS ANALYSIS

3.0

2.5
Upper bound
Lower bound

Stability Margin

2.0

1.5

1.0

0.5

0.0

10

15

Vt (m/s)

Figure 5.10: Robust stability analysis of the uncertain systems.

of uncertain elements that are less than 0.5 normalized units away from their
nominal values (which corresponds to an uncertainty of less than 10% in our
case) and, there is a collection of uncertain elements that are more than or
equal to 0.5 normalized units away from their nominal values that results in
instability (which corresponds to an uncertainty between 10% and 20%).
We can see that the stability margin reduces for low airspeeds and during the
transition to higher ones. These are the more problematic regions, one for the
lack of controllability from the control surfaces, the other due to the change of
the actuators control action. Considering that usually Vt > 2m/s, we observe
that a 10% error in all coefficients assures a robustly stable closed-loop system
whatever the airspeed. This, however, is still a low error margin.

Stability robustness to plant parameter variations


Finally, we consider the closed-loop system stability robustness to model parameter variations due to changes in the linearization equilibrium point of the
nonlinear model. This is a low-frequency phenomenon, which we will analyze
according to [53].
It is important for the control gains K in (4.19) to stabilize the system at
all points near the design operating point for gain scheduling to be effective.
However, in passing from operating point to operating point, the parameters
of the state variable model vary.
Consider again the nominal model (4.18) used for design, which has the transfer

84

CHAPTER 5. GAIN SCHEDULING

function:
G(s) = C(sI A)1 B

(5.38)

However, due to operating point changes, the actual system is described by:
= (A + A)
x
x + (B + B)
u

(5.39)

= (C + C)
y
x

(5.40)

where the plant parameter variation matrices are A, B and C. This


results in the transfer function:
G (s) = G(s) + G(s)

(5.41)

with:
G(s) = C(sI A)1 B + C(sI A)1 B C(sI A)1 A(sI A)1 B
(5.42)
where second-order effects have been neglected. A state-space realization of
G is given by:
#
#
"
B
A A
u
x +
x =
B
0
A
h
i
y = C C x
"

(5.43)
(5.44)

Since we may write the additive uncertainty equation (5.41) in the multiplicative form:
G (j) = [I + G(j)G1 (j)]G(j) [I + M(j)]G(j)

(5.45)

(where we substituted s = j), we shall proceed considering multiplicative


uncertainties in the form (5.45), where the unknown discrepancy satisfies the
bound:
(M(j)) < m()

(5.46)

with m() known for all .


So suppose we have designed a controller K so that the nominal closed-loop

85

5.3. ROBUSTNESS ANALYSIS

system (4.23) is stable. A frequency-domain condition that guarantees the


stability of the actual closed-loop system, which contains not G(s) but G (s)
satisfying (5.45)-(5.46), will now be derived using the multivariable Nyquist
condition:
Theorem 5.1 (Generalized (MIMO) Nyquist Theorem). Let Pol denote the
number of open-loop unstable poles in GK. The closed-loop system with loop
transfer function GK and negative feedback is stable if and only if the image
of det(I + GK) as s goes clockwise around the Nyquist D-contour (right-half
of the s-plane including the j axis)
1. makes Pol anti-clockwise encirclements of the origin, and
2. does not pass through the origin.
Proof. See [62], pp. 146.
So it is required that the encirclement count of the map det(I+G K) be equal
to the negative number of unstable open-loop poles of G K. By assumption,
this number is the same as that of GK. Thus, the number of encirclements of
det(I+G K) must remain unchanged for all G allowed by (5.46). This is assured if and only if det(I+G K) remains nonzero as G is warped continuously
toward G , or equivalently:
0 < [I + [I + M(s)]G(s)K(s)]

(5.47)

for all 0 1, all M(s) satisfying (5.46), and all s on the standard Nyquist
contour.
Since G vanishes on the infinite radius segment of the Nyquist contour, and
assuming for simplicity that no indentations are required along the j-axis
portion, this reduces to the following condition:
(GK(I + GK)1 ) = (T(j)) <

1
m()

(5.48)

for all 0 < . Thus, stability robustness in face of parameter variations


A, B and C translates into a requirement that the cosensitivity T(j)
be bounded above by the reciprocal of the multiplicative modeling discrepancy
bound m().

86

CHAPTER 5. GAIN SCHEDULING

The m() bound for robustness to gain scheduling model parameter variation
is obtained the following way. Define Gij (s) as the transfer function matrix
of the state-space system (5.43)-(5.44) with A = Ai Aj , B = Bi Bj
and C = Ci Cj , and i and j representing two consecutive linear system
models. From (5.45) we have that:
M(j) = G(j)G1 (j)

(5.49)

The bound m() is obtained from:


(M(j)) (G)(G1 ) = (G)

1
= m()
(G)

(5.50)

where we used relations (5.29)-(5.30).


Figure 5.11 compares the reciprocal of the multiplicative modeling discrepancy bound, 1/m(), with the maximum singular value of the cosensitivity,
(T(j)), for the entire flight envelope. Figure 5.11(a) considers a different
system every 0.01m/s, while for fig. 5.11(b) the scheduling is made for Vt
steps of 0.1m/s. While condition (5.48) is always fulfilled in the first case,
60

40
1/m()

1/m()

40

20

20

0
-20
Magnitude (dB)

Magnitude (dB)

0
-20
-40
-60

-60
-80

-80

(T(j))

(T(j))
-100

-100

-120

-120
-140
103

-40

102

101
101
100
Frequency (rad/s)

102

103

(a) For different systems every 0.01m/s


steps in Vt .

-140
103

102

100
101
101
Frequency (rad/s)

102

103

(b) For different systems every 0.1m/s


steps in Vt .

Figure 5.11: Stability robustness to plant parameter variation (darker:


(T(j)), lighter: 1/m()).
thus demonstrating the stability robustness of the scheduled system, for the
second case it is not always the case. In fact, if a new system is considered
every 0.1m/s, the robustness to plant parameter variations is only assured for
the systems in the intervals [2.90, 5.70]m/s and [7.40, 15]m/s. This means a
thinner schedule should be used in the remaining flight envelope, correspond-

5.4. SIMULATION RESULTS

87

ing to the lower airspeed and transition regions, clearly where the dynamics
more rapidly change.
If there is no problem in keeping and accessing a big dimension lookup table in real-time, the 0.01m/s grid should be used for system scheduling. In
practice, this corresponds to consider a new dynamic system at each airspeed
measurement (assuming at least 0.01m/s resolution of the wind sensor).

5.4

Simulation results

This section presents illustrative simulation results of the airship gain scheduling control.

5.4.1

Groundspeed and altitude regulation

This section focuses on the longitudinal control using the model (5.2). The
purpose is to follow groundspeed and altitude profiles. The groundspeed profile
is given in terms of velocity along the reference path, i.e., vr = [ur , 0, 0]T m/s.
In order to observe only the longitudinal behavior, the trajectory coincides
with a straight line. The groundspeed profile starts at 10m/s and at tu1 = 10s
accelerates to 12m/s, with a limit of 1m/s2 . At tu2 = 65s goes back to 10m/s,
with a rate limit of 0.5m/s2 . Concerning the altitude profile, it starts at 50m
and at th1 = 40s it goes up to 60m. At th2 = 100s goes down again to 50m.
The ascent and descent rates are 1m/s. The simulation considers constant
wind incidence from north at 3m/s.
The N ED trajectory, longitudinal () and vertical () position errors, and the
altitude profile are shown in fig. 5.12. As expected, there is only motion in the
vertical plane, as may be noticed by the perfect following of the straight reference segment. The position errors oscillate around zero, with and inferiors
to 2m. The and errors correspond respectively to the transient response
of the speed and altitude profiles following. Although with an overshoot and
a small delay, the altitude profile is well followed.
The groundspeed profile and output are represented in fig. 5.13, together with
the aerodynamic variables. The groundspeed components v = [u, v, w]T are
described in fig. 5.13(a) and the airspeed Vt , the sideslip angle and the
angle of attack may be seen in fig. 5.13. Notice that the difference between

88

(m)

CHAPTER 5. GAIN SCHEDULING

65

2
1
0
-1
-2

55
(m)

h (m)

60

50
45

200

1000

100

h (m)

40
1500

500
-100
0

N (m)

-200

2
1
0
-1
-2

65
60
55
50
45

20

40

60
80
Time (s)

100

120

140

20

40

60
80
Time (s)

100

120

140

20

40

60
80
Time (s)

100

120

140

E (m)

(a) Comparison of airship north N , east E


and altitude h position () and projections
(.) with reference ().

(b) Longitudinal () and vertical () errors


and altitude profile ( output, reference).

Figure 5.12: Airship position coordinates and errors.

11
10
9

20

40

60
80
Time (s)

100

120

140

Vt (m/s)

u (m/s)

13
12

16
15
14
13
12

20

40

80
60
Time (s)

100

120

140

20

40

80
60
Time (s)

100

120

140

20

40

80
60
Time (s)

100

120

140

1
(deg)

v (m/s)

1
0
-1

20

40

60
80
Time (s)

100

120

140

-1

(deg)

w (m/s)

2
1
0
-1
-2

20

40

60
80
Time (s)

100

120

140

(a) Groundspeed: longitudinal u, lateral v


and vertical w ( reference, output).

10
5
0
-5
-10

(b) Airspeed Vt , sideslip angle and angle


of attack .

Figure 5.13: Airship ground velocity components and aerodynamic variables.


the u and Vt (approximately equal to ua ) lies on the 3m/s north wind. The
longitudinal speed presents also an overshoot and a small delay, but in the
overall a good regulation is achieved. The lateral speed and sideslip angle
are null, as expected in this case. The vertical groundspeed reference is null,
indicating the airship is requested to follow the reference path at ur = 10m/s
at all times, even during the altitude profile. This request does not take
into account the angle of attack (pitch) necessary to provide lift at a given
airspeed, reason for which the output w presents an offset (probably close to
the equilibrium value we not used as reference).

89

5.4. SIMULATION RESULTS

The actuators input is represented in fig.5.14. The groundspeed and altitude


regulation is possible through the action of the longitudinal actuators, which
control effort is represented in fig. 5.14(a). The engines thrust XT is mainly
responsible for the airspeed control, while the altitude is responsibility of the elevator deflection e . The variation of the two control variables is not, however,
independent. A change in the groundspeed (altitude) reference, and hence
output, provokes a variation in the altitude (groundspeed) output, corrected
by the elevator (engines thrust). This may be observed comparing the respective figures. The engines vectoring v is small during the entire simulation,
as expected for an aerodynamic flight. The lateral actuation, represented in
1

10

a (deg)

e (deg)

20
0
-10
-20

20

40

60
80
Time (s)

100

120

-1

140

r (deg)

XT (N )

60
40
20
0

20

40

60
80
Time (s)

100

120

40

80
60
Time (s)

100

120

140

20

40

80
60
Time (s)

100

120

140

20

40

80
60
Time (s)

100

120

140

TD (N )

v (deg)

20

0
-1

140

6
0
-3
-6

80

20

40

60
80
Time (s)

100

120

140

(a) Longitudinal actuators: elevator e , total thrust XT and vectoring v .

0
-1

(b) Lateral actuators: aileron a , rudder r


and differential thrust TD .

Figure 5.14: Airship actuators input.


fig. 5.14(b) for completeness, is obviously null for all three inputs, aileron a ,
rudder r and differential thrust TD .

5.4.2

Case-study mission

This section demonstrates the performance of the gain scheduling approach


using the complete 12-states model (5.8) executing the case-study mission
described in Section 3.3.1. The scheduling is made between airspeed steps
of 0.1m/s. The mission covers a wide range of the flight envelope, with the
airspeed Vt varying between 3 and 12m/s.
The airship position coordinates and errors are represented in fig. 5.15. The
vertical take-off and landing are well perceived in fig. 5.15(a), as well as the
path-tracking performance. Figure 5.15(b) displays the longitudinal , lateral

90

CHAPTER 5. GAIN SCHEDULING

and vertical errors, allowing to identify the more problematic mission points,
namely when the wind is aft the airship, at the end of the first half-circle
and the transition from the vertical ascent and the horizontal path-tracking
(see fig. 5.16(a)). The remaining noticeable errors correspond to instantaneous
references changes before the second stabilization, which the airship smoothly
corrects. In order to avoid saturation of the propellers, probable when the
controller tries to rapidly correct the longitudinal position, the error is limited. This approaches the idea of Teel [63] which will be better explored in
Section 7.4. The limitation of can be noticed by the constant rate at which
the north position is corrected in fig. 5.15(b). Note that the existence of instant position errors might be caused by a transition in the mission objectives,
namely from path-tracking to stabilization, but also from a discontinuity in
the position provided by the GPS when the available satellites change.

(m)

60

-40
-60
-80

50

100

150

200
Time (s)

250

300

350

400

10

20
10
0
-10
-20
-30
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

50
40
30

(m)

h (m)

-20

20

100

300
200

100

-100
-200
N (m)

(m)

200

-5

-100
E (m)

(a) Airship north N , east E and altitude h


position (bold) and projections (normal).

(b) Longitudinal (), lateral () and vertical () errors.

Figure 5.15: Airship position coordinates and errors.


Figure 5.16 represents the airship horizontal trajectory and its attitude during
the mission. The airship north-east coordinates and heading during the mission
are described in fig. 5.16(a). The preferential alignment with the wind during
take-off and landing is well recognized, while in maneuvers at low airspeeds the
airship appears as slightly crabbing. The Euler angles evolution is displayed
in fig. 5.16(b). The roll angle (), with a null reference, presents a higher
oscillation in the two above mentioned problematic parts. The pitch () and
yaw () angles try to follow the respective references described in Section 3.2.

91

5.4. SIMULATION RESULTS

The pitch angle shows higher errors during take-off and landing, as well as
during stabilization.
10

200

(deg)

250
wind
heading

0
-10

150

50

100

150

200
Time (s)

250

300

350

400

-20
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

100

20
(deg)

N (m)

50

-50

-100

(deg)

-150

-200

-250
-100

100
E (m)

200

300

(a) North-east position with airship


heading ( reference, output).

450
360
270
180
90
0
-90

(b) Roll , pitch and yaw ( reference, output).

Figure 5.16: Airship north-east position and attitude.


The airship ground velocity and the aerodynamic variables are depicted in
fig. 5.17. The ground velocity components are represented in fig. 5.17(a).
The longitudinal groundspeed u mostly follows the reference that varies between 0m/s for stabilization, take-off and landing, and 7m/s during the pathtracking. Along the circular segments, the errors are more noticeable due to
the change of the wind incidence angle while the airship is turning. The error is higher when the wind is aft the airship. The lateral velocity v is also
mostly influenced by the circular segments and during the tail wind segment.
The vertical velocity w follows its reference, with the 1 and 0.5m/s steps
corresponding to the take-off and landing vertical motion. The airspeed and
aerodynamic angles can be seen in fig. 5.17(b). During the whole mission, the
airspeed Vt varies significantly, from values around 3m/s up to 12m/s. The
airship covers a wide flight envelope, from hover to the aerodynamic flight,
crossing the troublesome transition region between the two. The sideslip angle and the angle of attack vary between 20o . Although more directly
related with wa , the behavior of is also clearly correlated with w.
The actuators input is described in fig. 5.18, with the longitudinal actuators
elevator e , total thrust XT and vectoring angle v in fig. 5.18(a) and the
lateral ones, aileron a , rudder r and differential thrust TD , in fig. 5.18(b).

92

CHAPTER 5. GAIN SCHEDULING

Vt (m/s)

u (m/s)

8
6
4
2
0
50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

-20
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

20
10

0
-2
-4
-6
0

50

100

150

200
Time (s)

250

300

350

400

0
-10

20

10
(deg)

w (m/s)

2
(deg)

v (m/s)

14
12
10
8
6
4
2

0
-1
-2

0
-10
-20

50

100

150

200
Time (s)

250

300

350

-30

400

(a) Groundspeed: longitudinal u, lateral v


and vertical w ( reference, output).

(b) Airspeed Vt , sideslip angle and angle


of attack .

Figure 5.17: Airship ground velocity components and aerodynamic variables.

30
15

0
-15
-30

XT (N )

a (deg)

30
15

100

150

200
Time (s)

250

300

350

0
-15
-30

400

50

100

150

200
Time (s)

250

300

350

400

-30
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

80

30

60

15

40
20
0
0

50

100

150

200
Time (s)

250

300

350

400

0
-30

0
-15

20

120
90
60
30

TD (N )

v (deg)

50

r (deg)

e (deg)

The elevator e action corresponds to the correction of the vertical error ,


while the rudder r has a higher command with tail wind, due to the reduced
authority at lower airspeeds. The aileron a is responsible for the control of
the roll angle. When the control surface loses authority at low airspeeds, this
function is assumed by the differential thrust TD . The vectoring angle v is

50

100

150

200
Time (s)

250

300

350

400

(a) Longitudinal actuators: elevator e , total thrust XT and vectoring v .

10
0
-10
-20

(b) Lateral actuators: aileron a , rudder r


and differential thrust TD .

Figure 5.18: Airship actuators input.


responsible for the airship lift when the airspeed Vt is too low to provide the

93

5.4. SIMULATION RESULTS

necessary aerodynamic lift. The correlation between these two variables is


obvious comparing the graphics of v and Vt .

5.4.3

Sensitivity and robustness to parameter uncertainty

Although the closed-loop system robustness has been analyzed in Section 5.3,
we present here the results of the sensitivity and robustness test described in
Section 3.3.2 as a tool of comparison between controllers performance.
For the baseline simulation, we consider no variation of the model parameters,
only wind disturbance input for the aerodynamic flight at 8m/s groundspeed
and 50m altitude. Figure 5.19(a) shows the airship north-east position and
heading when following the straight line reference aligned with north, while
subject to the 4m/s constant wind blowing from west, plus 3m/s turbulent
gust. We notice the airship is able to follow the reference, although with an
(deg)

10
1200

-10
1000

50

100

150

50

100

150

50

100

150

50

100

150

50

100

150

(deg)

10

800

5
0
-5
-20

(deg)

N (m)

wind
heading
600

-25
-30
-35

400

w (deg)

200

0
-5
-10

w (deg)

100
0

-100

-50

50

E (m)

(a) Airship north-east position with


airship heading ( reference,
output).2

90

80

Time (s)

(b) Airship attitude : roll , pitch and


yaw ( reference, output), and estimated wind attitude, w , w .

Figure 5.19: Airship north-east trajectory and attitude, and wind attitude.
2

The different scales might give a wrong idea of the airship heading ( 27o ).

94

CHAPTER 5. GAIN SCHEDULING

orientation that helps it minimize the drag force produced by the lateral wind.
Figure 5.19(b) represents the airship attitude references (see Section 3.2) and
output, as well as the wind estimated attitude. Notice that neither w nor w
are constant, since they echo both constant and gust wind components. The
excitation of the signals is due to the wind turbulence.
As may be expected, the airship position errors oscillate around zero instead
of converging, due to the wind turbulence input, as may be seen in fig. 5.20(a).
The aerodynamic variables are represented in fig. 5.20(b). The around 9m/s
airspeed corresponds to the relative air speed between the 8m/s groundspeed
heading north and the 4m/s wind speed from west. The sideslip angle is
close to zero, showing the airship is aligned with the relative airspeed.
1

Vt (m/s)

(m)

10
0

-1

100

50

150

Time (s)

50

100

150

100

150

100

150

time (s)
1

(deg)

(m)

5
0

-1

50

100

-5

150

10

0
-1

100

50

50
time (s)

(deg)

(m)

Time (s)

150

Time (s)

(a) Airship longitudinal (), lateral () and


vertical () errors.

0
-5

50
time (s)

(b) Airspeed Vt , sideslip angle and angle


of attack .

Figure 5.20: Airship position errors and aerodynamic variables.


The actuators input applied to the AURORA airship is represented in fig. 5.21,
with the longitudinal actuators action given in fig. 5.21(a), and the lateral
actuation in fig. 5.21(b). The vectoring angle v is near zero in aerodynamic
flight, as is the differential thrust TD .
Table 5.1 shows the RMS values of selected variables. They are the airship
positions errors, namely longitudinal , lateral and vertical errors, and the
true airspeed Vt , the angle of attack and the sideslip angle , together with
the groundspeed error eu relative to the 8m/s reference. The first row has
the RMS values obtained for the baseline case, and is to serve as reference

95

5.4. SIMULATION RESULTS

a (deg)

e (deg)

10

0
-5

50

100

-2

150

50

Time (s)

150

100

150

100

150

10
r (deg)

XT (N )

80
60
40
20
0

100
Time (s)

50

100

-10

150

50

Time (s)

Time (s)

120

1
TD (N )

v (deg)

90
60
30

0
-30

50

100

150

-1

50

Time (s)

(a) Longitudinal actuators input: elevator


e , total thrust XT and vectoring v .

Time (s)

(b) Lateral actuators input: aileron a ,


rudder r and differential thrust TD .

Figure 5.21: Airship actuators input.


for the remaining lines where each of the listed coefficients is varied one at a
time. For a parameter uncertainty of 70%, the controlled airship performed
qualitatively like the baseline case. We then increased the parameter uncertainty, and present here the results obtained for 90% uncertainty around the
nominal value.
Only for some parameters the uncertainty leads to significant deviations from
the baseline RMS values, either because the control action is insufficient due
to the actuators saturation (the gain scheduling demands a too high control
input) and/or just because the tested coefficient appears to be a more sensitive
model parameter. These cases are represented in bold in table 5.1.
With noticeable deviations from the baseline RMS values, for CL and CD0
(both for 90% uncertainty) the gain scheduling still controls the AURORA
airship within acceptable bounds. CL is the lift coefficient derivative due to
angle of attack; and CD0 is maybe the most important aerodynamic coefficient,
expressing the drag suffered by the airship envelope at zero lift.
For the remaining coefficients, CMe , CNr , CMq and CNr (all for 90% uncertainty), either saturation of the actuators occurred or the uncertainty of
the model parameter is too important for the gain scheduling controller to
overcome it. These parameters correspond respectively to the authority of the
elevator and rudder as pitching and yawing control inputs, and the pitch and
yaw damping derivatives. While the gain scheduling demonstrates to be ro-

96

CHAPTER 5. GAIN SCHEDULING

Table 5.1: Robustness tests on model parameters (RMS values of selected variables).
Baseline

(m)

(m)

(m)

Vt (m/s)

(deg)

(deg)

eu (m/s)

0.24

0.33

0.45

9.01

3.41

1.17

0.16

Cl

90%
+90%

0.20
0.34

0.28
0.45

0.45
0.45

9.01
9.02

3.47
3.22

0.93
1.72

0.12
0.25

CM0

90%
+90%

0.24
0.24

0.33
0.33

0.46
0.45

9.01
9.01

3.47
3.35

1.17
1.17

0.16
0.16

CM

90%
+90%

0.33
0.25

0.35
0.33

1.07
0.69

9.01
9.01

4.09
3.12

1.20
1.15

0.16
0.15

CM

90%
+90%

0.24
0.24

0.33
0.33

0.45
0.45

9.01
9.01

3.40
3.42

1.17
1.17

0.16
0.16

CM

90%
+90%

0.24
0.24

0.33
0.34

0.45
0.46

9.01
9.01

3.35
3.48

1.17
1.17

0.16
0.16

CM

90%
+90%

0.24
0.24

0.33
0.33

0.45
0.45

9.01
9.01

3.40
3.42

1.17
1.17

0.16
0.16

CMe

90%
+90%

166.29
0.23

91.32
0.34

100.08
0.48

13.64
9.01

2.87
3.61

1.97
1.18

7.30
0.15

CN

90%
+90%

0.21
0.25

0.29
0.35

0.45
0.45

9.01
9.01

3.44
3.40

1.15
1.22

0.12
0.19

CNr

90%
+90%

23.75
0.16

9.07
0.12

2.57
0.44

9.28
9.01

4.10
3.43

7.71
1.05

1.37
0.11

CY

90%
+90%

0.38
0.21

0.56
0.30

0.51
0.45

9.02
9.01

3.48
3.42

2.56
0.85

0.22
0.14

CYr

90%
+90%

0.21
0.26

0.28
0.37

0.44
0.45

9.01
9.01

3.44
3.39

1.22
1.53

0.13
0.19

CD0

90%
+90%

1.78
0.78

0.51
0.32

0.66
0.65

9.02
9.00

4.26
2.75

1.24
1.12

0.27
0.14

CDi

90%
+90%

0.31
0.24

0.33
0.34

0.47
0.44

9.01
9.01

3.49
3.35

1.17
1.17

0.17
0.15

CL0

90%
+90%

0.24
0.24

0.33
0.33

0.45
0.45

9.01
9.01

3.33
3.50

1.17
1.17

0.16
0.16

CL

90%
+90%

6.49
0.24

0.41
0.34

7.18
0.59

9.05
9.01

19.78
2.00

1.57
1.12

0.36
0.15

CLe

90%
+90%

0.23
0.26

0.33
0.33

0.43
0.68

9.01
9.01

4.08
3.16

1.20
1.15

0.16
0.17

Clp

90%
+90%

0.26
0.24

0.33
0.34

0.51
0.45

9.01
9.01

3.39
3.41

1.29
1.15

0.18
0.16

CMq

90%
+90%

401.88
0.43

102.32
0.37

11.96
1.37

8.55
9.03

20.31
4.44

13.76
1.19

2.81
0.21

CNr

90%
+90%

626.99
0.43

281.01
0.79

7.55
0.48

3.76
9.01

39.40
3.43

75.92
1.43

3.96
0.22

mw

90%
+90%

0.28
0.23

0.32
0.34

1.33
1.45

9.01
9.01

1.67
6.54

1.03
1.31

0.16
0.16

5.5. CONCLUSIONS

97

bust to a 90% uncertainty in the remaining parameters, for these four cases,
the mismatch between the airship system and the model considered in the gain
scheduling controller design is too significant for the control action to overcome
it.
In any case, the gain scheduling controller may be considered robust to wind
disturbances and plant uncertainties. Among the list selected, these six parameters CL , CD0 , CMe , CNr , CMq and CNr (and specially the last four) are
in fact the model parameters for which a more careful identification or determination should take place, though the required precision could merely remain
inside a 70% margin.

5.5

Conclusions

This chapter covers the analysis and results obtained applying a gain scheduled
state-feedback optimal controller to the airship path-tracking control problem
over the entire flight envelope. Considering the 12-states model, and for each
equilibrium condition considered, the control law provides in a single action
actuator commands to regulate both lateral and longitudinal motions.
Although at each equilibrium point the closed-loop system is guaranteed to
be stable by the optimal controller, it is important to analyze the robustness
to input disturbances as well as model uncertainties and parameter variation.
Doing so we have come to the following conclusions about the closed-loop
system:
its performance is robust to wind disturbances;
its performance is robust to measurement noise;
it is robustly stable to model parameters uncertainties up to 10% for
airspeeds over 2m/s. A 20% uncertainty in the parameters still leads to
stable systems, except for very low airspeeds and in the transition region.
This indicates that a better identification of the model should be made
for these airspeeds, namely of the aerodynamic coefficients;
it is robustly stable to parameter variations for a 0.01m/s scheduling.
This means the closed-loop system remains stable even if the actual
system does not correspond to the equilibrium condition considered. For
a 0.1m/s grid, the parameter variation still leads to stable systems, again

98

CHAPTER 5. GAIN SCHEDULING

except for very low airspeeds and in the transition region. If such a
scheduling is necessary, one should try to avoid missions that induce the
airship to fly at such airspeeds for significant periods of time.
These robustness properties, together with (many) satisfactory simulation results and implementation simplicity, indicate the gain scheduling control is a
possible solution to the AURORA airship path-tracking problem.
However, is it the best? The control solution is optimal for given weighting
matrices, which are still obtained empirically. Moreover, we are only considering a linearized version of the airship, eventually discarding important features
of the system. These issues do not guarantee an optimal overall result, reason
for which other nonlinear solutions will be considered.

Chapter 6
Dynamic Inversion
Contents
6.1

General theory . . . . . . . . . . . . . . . . . . . . .

101

6.1.1

Local coordinates transformation . . . . . . . . . . . 102

6.1.2

Exact linearization via feedback . . . . . . . . . . . . 105

6.1.3

Asymptotic output tracking . . . . . . . . . . . . . . 107

6.2

New formulation for cascaded systems . . . . . . .

109

6.3

Application to airship path-tracking problem

. .

111

6.4

Simulation results . . . . . . . . . . . . . . . . . . .

114

6.5

6.4.1

Case-study mission . . . . . . . . . . . . . . . . . . . 114

6.4.2

Sensitivity and robustness . . . . . . . . . . . . . . . 118

Conclusions . . . . . . . . . . . . . . . . . . . . . . .

121

As in most complex engineering problems we are facing nowadays, flight control design is generally based in a divide-and-conquer approach. First, the
nonlinear equations of motion of the air vehicle are linearized about selected
operating points over the flight envelope. The tools of linear control theory can
then be used to design individual compensators to satisfy closed-loop specifications. Finally, a gain scheduled control is obtained by switching between the
individual compensators according to predefined scheduling variables. This
approach, for its relevance in flight control, was outlined in Chapter 5.
However, the gain scheduling procedure is time consuming, costly to iterate,
and still relies substantially on the engineers knowhow. An alternate methodology to flight control design, that avoids this iterative tuning process and
directly considers the nonlinear nature of the problem, is Dynamic Inversion.
Dynamic Inversion is a methodology to design closed-loop control laws for
99

100

CHAPTER 6. DYNAMIC INVERSION

nonlinear systems [64, 65]. As opposed to gain scheduling, dynamic inversion


searches for a global controller from a single global nonlinear model of the
plant. Its application in flight control [66] is justified since it can explicitly
address the nonlinearities in the aircraft or airship dynamics and provides a
control law which is valid for the whole flight envelope.
With dynamic inversion, the set of existing deficient or undesirable dynamics
are canceled out and replaced by a designer selected set of favored dynamics.
This is accomplished by careful algebraic selection of a feedback function,
reason for which the dynamic inversion methodology is also called Feedback
Linearization.
The feedback linearization is viewed as a generalization of pole placement for
linear systems. Its basic idea is to first transform a nonlinear system into
a (fully or partially) linear one, and then use the well-known and powerful
linear design techniques to complete the control design. However, it does not
guarantee robustness in face of parameter uncertainty or disturbances.
Two main assumptions are made in the dynamic inversion methodology, (i) the
plant dynamics is perfectly modeled, and (ii) the system states are measured
or estimated accurately.
In practice, neither of these assumptions is realistic, and the robustness of
the closed-loop system must be secured in order to suppress any undesirable
behavior. The use of an outer-loop controller to improve a dynamic inversion
inner-loop controller robustness has been reported in some works. A Linear
Quadratic Gaussian outer-loop is addressed by Ito et. al [67], and -synthesis
is used by Reiner et. al [68] and Bennani and Looye [69]. Yedavalli et. al [70]
present a stability robustness analysis of dynamic inversion based control laws
used for flight control with uncertainties in model data.
Before advancing to a complete nonlinear control design, we took an intermediate step between linear and nonlinear control. It is an hybrid approach,
that considers the linearized dynamic models presented in Chapter 4, and the
nonlinear kinematics of the airship. The lateral and longitudinal motions are
decoupled, reason for which the kinematics, though nonlinear, allows to keep
the independence of the controllers. This was applied to the lateral control of
the airship and compared with LQR proportional feedback in [71].
A fully, yet simplified, nonlinear airship model was used in the hover stabilization of the AURORA airship using dynamic inversion [72]. At hover conditions

6.1. GENERAL THEORY

101

the relative wind dynamics may be insignificant, downsizing the aerodynamic


forces when compared with the kinematic, gravity and propulsion ones. The
control law is therefore obtained by inversion of a simplified dynamic model,
discarding the aerodynamic forces at low airspeeds.
In this chapter, the dynamic inversion methodology will be applied to the
path-tracking problem of the AURORA airship, by inversion of the 12-state
complete nonlinear dynamic system deduced in Chapter 2. Although the classical approach is reviewed in Section 6.1, a new formulation, applicable to
cascaded systems described in terms of velocity and position, and where the
output of interest is the position, is developed in Section 6.2.
The performance of the dynamic inversion closed-loop system applied to the
airship path-tracking problem is analyzed in Section 6.4. We present the simulation results obtained for the case-study mission, and examine the controller
sensitivity and robustness to parameter uncertainty. Finally, Section 6.5 closes
the chapter with some final remarks.

6.1

General theory

This section presents the classical dynamic inversion theory, and mostly follows
reference [64].
The objective of the dynamic inversion approach is to change a nonlinear
system into a linear and controllable one by means of feedback and coordinates
transformation. With this in mind, and for MIMO (multi-input multi-output)
square systems (i.e., n-states systems with the same number m of inputs and
outputs) in a neighborhood of a point xo , we first describe a suitable change of
coordinates in state space that allows us to represent the system in a normal
form of special interest. It is then based on this normal form that we obtain
a state feedback control law which applied yields a linear and controllable
closed-loop system. Finally, we consider the problem of asymptotic output
tracking.

102

CHAPTER 6. DYNAMIC INVERSION

6.1.1

Local coordinates transformation

Consider the affine or linear in control system:


x = f (x) +

m
X

gi (x)ui

(6.1a)

i=1

y1 = h1 (x)

(6.1b)

ym = hm (x)
where x is the state n-vector, u is the control input m-vector (of components
ui ) and y is the output m-vector (of components yi ); f (x) and h(x) are smooth1
n- and m-vector fields respectively, and g(x) is an nm matrix whose columns
are smooth vector fields gi . A more condensed form of (6.1) is:
x = f (x) + g(x)u

(6.2a)

y = h(x)

(6.2b)

We will start our development with the multivariable version of relative degree
(see Appendix C for the Lie derivative Lkf h(x) used notation). A multivariable
nonlinear system of the form (6.1) has a (vector) relative degree {r1 , . . . , rm }
at a point xo if:
(i)
Lgj Lkf hi (x) = 0

(6.3)

for all 1 j m, for all 1 i m, for all k < ri 1, and for all x in a
neighborhood of xo ;
(ii) the m m matrix:

A(x) =

Lg1 Lrf 1 1 h1 (x)


Lg1 Lrf 2 1 h2 (x)

rm 1
Lg1 Lf
hm (x)

Lgm Lrf 1 1 h1 (x)


Lgm Lrf 2 1 h2 (x)

rm 1
hm (x)
Lgm Lf

(6.4)

is nonsingular at x = xo .
1

A vector field f (x) is considered smooth if it has continuous partial derivatives of any
required order.

103

6.1. GENERAL THEORY

Note that ri is exactly the number of times one has to differentiate the i-th
output yi (t) at t = to in order to have at least one component of the input
vector u(to ) explicitly appearing.

Proposition 6.1. Suppose a system has a (vector) relative degree {r1 , , rm }


at xo . Then:
r1 + . . . + rm n

(6.5)

Set, for 1 i m:
i1 (x) = hi (x)
i2 (x) = Lf hi (x)

(6.6)

iri (x) = Lrf i 1 hi (x)


If r = r1 + . . . + rm is strictly less than n, it is always possible to find n r
more functions r+1 (x), . . . , n (x) such that the mapping:
m
T
(x) = [11 (x), . . . , 1r1 (x), . . . , m
(6.7)
1 (x), . . . , rm (x), r+1 (x), . . . , n (x)]

has a jacobian matrix which is nonsingular at xo and therefore qualifies as a


local coordinates transformation in a neighborhood of xo .

Proof. See [64], pp. 237.

In the remaining of this section, we will restrict our description to the systems
where the sum r = r1 + r2 + . . . + rm is exactly equal to the dimension n of
the state space. In this case, the set of functions:
ik (x) = Lk1
hi (x)
f

for 1 k ri , 1 i m

(6.8)

defines completely a local coordinates transformation at xo . Differentiating

104

CHAPTER 6. DYNAMIC INVERSION

with respect to time, we obtain:


di1
= i2 (t)
dt

diri 1
= iri (t)
dt
m
X
diri
ri
Lgj Lfri 1 hi (x(t))uj (t)
= Lf hi (x(t)) +
dt
j=1

(6.9)

for all 1 i m. Note that the coefficient that multiplies uj (t) in the latter
equation is exactly equal to the (i, j) entry of the matrix A(x) in (6.4).
Set now:

1i
2i

ri i

i1 (x)
i2 (x)

i
ri (x)

= [ 1 , . . . , m ]T

for 1 i m

(6.10)

(6.11)

Then, the equations in question can be rewritten as:


1i = 2i

ri i 1 = ri i
ri i = bi () +

m
X

(6.12)
aij ()uj (t)

j=1

yi =

1i

for 1 i m and with:


aij () = Lgj Lrf i 1 hi (1 ())
bi () = Lrf i hi (1 ())

for 1 i, j m

(6.13)

for 1 i m

(6.14)

The structure of equations (6.12) characterizes the normal form of the equations (see fig. 6.1) describing (locally around a point xo ) a nonlinear system,
with m inputs and m outputs, having a (vector) relative degree {r1 , . . . , rm }
at xo , with r = r1 + . . . + rm exactly equal to the dimension n of the state

105

6.1. GENERAL THEORY

space. In this case, when r = n, the system has no internal (zero) dynamics,
and so it is minimum phase by default. Note that in (6.12) the coefficients
uj

- bi () + Pm aij ()uj
j=1
6 6

ri i

- R

ri i

2i

- R

1i = yi

Figure 6.1: Normal form representation, with no internal dynamics.

aij () are exactly the entries of matrix (6.4), with x replaced by 1 (), and
the coefficients bi () are the entries of a vector:

b(x) =

Lrf 1 h1 (x)
Lrf 2 h2 (x)

rm
Lf hm (x)

(6.15)

again with x replaced by 1 ().

6.1.2

Exact linearization via feedback

The main problem dealt with in this section is that of using feedback and
coordinates transformation to the purpose of changing a nonlinear system into
a linear and controllable one. Formally, the problem in question can be stated
the following way:

State-space exact linearization problem Given a set of vector fields f (x)


and g1 (x), . . . , gm (x) and an initial state xo , find (if possible), a neighborhood
U of xo , a pair of feedback functions (x) and (x) defined on U , a coordinates
transformation z = (x) also defined in U , a matrix A Rnn and a matrix
B Rnm , such that:


(f (x) + g(x)(x))
= Az
x
x=1 (z)



(g(x)(x))
=B
x
x=1 (z)

(6.16)
(6.17)

106

CHAPTER 6. DYNAMIC INVERSION

and:
rank(B AB An1 B) = n

(6.18)

Our point of departure will be the normal form developed in the previous
section. Recall that in a neighborhood of the point o = 1 (xo ) the matrix
A() is nonsingular and therefore the equations:
= b() + A()u

(6.19)

with = [1 , 2 , , m ]T the new reference input, can be solved for u. The


input u solving these equations has the form of a state feedback:
u = A1 ()(b() + )

(6.20)

Imposing this feedback yields a closed-loop system characterized by the m sets


of equations:
1i = 2i

ri i 1 = ri i
ri i = i

(6.21)

for 1 i m, which is clearly linear and controllable.


In terms of the original description of the system, the linearizing feedback has
the form:
u = (x) + (x)

(6.22)

with (x) and (x) given by:


(x) = A1 (x)b(x)
(x) = A1 (x)

(6.23)
(6.24)

and A(x) and b(x) as in (6.4) and (6.15). The linearizing coordinates are
defined as:
hi (x)
ki (x) = Lk1
f

for 1 k ri , 1 i m

(6.25)

107

6.1. GENERAL THEORY

The closed-loop system obtained applying the control law (6.22) to the system (6.2) is (see fig. 6.2):

x = f (x) + g(x)(x) + g(x)(x)

(6.26a)

y = h(x)

(6.26b)

- (x) + (x)
6

- x = f (x) + g(x)u

y = h(x)
x

Figure 6.2: Closed-loop system, with new reference input .

The conditions that the system, for some choice of output functions h1 (x), . . . ,
hm (x), has a (vector) relative degree {r1 , . . . , rm } at xo , and that r1 +. . .+rm =
n, imply the existence of a coordinates transformation and a state feedback,
defined locally around xo , which solve the state space exact linearization problem. The following Lemma shows that these conditions are also necessary.
Lemma 6.1. Suppose the matrix g(xo ) has rank m. Then, the state space
exact linearization problem is solvable if and only if there exists a neighborhood
U of xo and m real-valued functions h1 (x), . . . , hm (x) defined on U , such that
the system (6.2) has some (vector) relative degree {r1 , . . . , rm } at xo and r1 +
. . . + rm = n.
Proof. See [64], pp. 247.

6.1.3

Asymptotic output tracking

In this section we consider the problem of tracking the output of a reference


model, which in turn is subject to some input (t). Consider the linear model,
described by:
= A + B
yr = C

(6.27a)
(6.27b)

The asymptotic model matching problem is solved by finding a feedback control


which causes, irrespectively of what the initial states of the system and of the

108

CHAPTER 6. DYNAMIC INVERSION

model are and for every input (t) to the model, an output y(t) asymptotically
converging to the corresponding output yr (t) produced by the model under the
effect of (t).
Suppose the model has a (vector) relative degree equal to the (vector) relative
degree {r1 , . . . , rm } of the system and r1 = . . . = rm = r. In this case, since:
CB = CAB = . . . = CAr2 B = 0

(6.28)

we have that:
yr(i) (t) = CAi (t)

for all 0 i r 1

yr(r) (t) = CAr (t) + CAr1 B(t)

(6.29a)
(6.29b)

Consider again the normal form (6.12) and choose the control input as:

u = A1 (x) b(x) + yr(r) q

(6.30)

with A(x) and b(x) as in (6.4) and (6.15) and q a column vector with elements
P
(i1)
qj = ri=1 ci1 (ij yrj ) for 1 j m where c0 , . . . , cr1 are real numbers.

Define an error e(t) as the difference between the real output y(t) and the
model output yr (t):
e(t) = y(t) yr (t)

(6.31)

(i1)

Since, by construction, ij = yj
= Li1
f hj (x) for 1 i r, we may write
Pr
(i1)
qj = i=1 ci1 ej . Substituting (6.29) into (6.30) leads to:

Pr

i1
i=1 ci1 (Lf h1

u = A1 b + CAr + CAr1 B

C 1 Ai1 )

Pr
i1
i1
)
i=1 ci1 (Lf hm C m A

(6.32)

Note that imposing the input (6.32) implies:


(r1)
(r)
. . . c1 e j c0 ej
cr1 ej
rj = yj = yr(r)
j

(6.33)

i.e.:
(r)

(r1)

ej + cr1 ej

+ . . . + c1 e j + c0 ej = 0

(6.34)

109

6.2. NEW FORMULATION FOR CASCADED SYSTEMS

for 1 j m. The error functions ej (t) satisfy a linear differential equation


of order r whose coefficients can be arbitrarily preset.
Thus, by construction, the system (6.2), subject to an input of the form (6.32)
with coefficients c0 , . . . , cr1 appropriately chosen, will produce an output
asymptotically converging to the output yr (t) of the model.
Note that the input (6.32) depends explicitly on the state x(t) of the system,
on the input (t) of the model, and on the state (t) of the model, which in
turn obeys the differential equation (6.27) (see fig. 6.3). This represents a more
-

= A + B

- (, x) + (, x)

x = f (x) + g(x)u
- y
= h(x)

6
x

Figure 6.3: Closed-loop system, with model reference input .

general form of state feedback, in that includes also an internal dynamics. A


feedback of this form is called a dynamic state feedback.

6.2

New formulation for cascaded systems

Section 6.1 presents the general theory of the dynamic inversion approach for
MIMO systems with the same number m of inputs and outputs, and for which
the sum of the relative degrees ri , for 1 i m, equals the dimension n of
the state space. A new formulation for dynamic systems which respect these
assumptions will now be presented here.
Consider an affine in control system whose description is given by the following
dynamics, kinematics and output equations:
= fv (V, P) + gv (V, P)u
V
= fp (V, P)
P

(6.35b)

y = h(V, P) = P

(6.35c)

(6.35a)

110

CHAPTER 6. DYNAMIC INVERSION

where the n-state vector x = [VT , PT ]T contains the velocity V and position
P components and the m-vectors u and y are the input and output vectors,
respectively; fv (V, P) and fp (V, P) are smooth m-vector fields and gv (V, P)
is an m m matrix with rank m for x = xo and whose columns are smooth
vector fields gvi (V, P).
This new dynamic inversion formulation simplifies the dynamic inversion implementation for systems with the cascaded form (6.35a)-(6.35b), and for which
the output variable of interest is the position vector, usual in path-tracking
problems.
Let us differentiate the output (6.35c) in order to have the input u explicitly
appearing. We then have:
= fp (V, P)
y = P
= fp fv + fp fp + fp gv u
=P
y
V
P
V
= fp (V)fv + fp (P)fp + fp (V)gv u

(6.36)

(6.37)

It is easy to verify that every output i of this system has relative degree ri = 2
at a point x = xo where the m m matrix fp (V)gv is nonsingular, and that
Pm
i=1 ri = n. Therefore, according to the theory reviewed in Section 6.1, there
is a local coordinates transformation:
T ]T
T , . . . , PT , P
(x) = [PT1 , P
1
m
m

(6.38)

such that we may represent (locally around xo ) the nonlinear system (6.35) in
the normal form:
= fp (V, P)
P
= fp (V)fv + fp (P)fp + fp (V)gv u
P
y=P

(6.39)
(6.40)
(6.41)

Moreover, applying the state feedback:


u = (fp (V)gv )1 ( fp (V)fv fp (P)fp )

(6.42)

where = [1 , . . . , m ]T is a new reference input, we obtain the linear and

6.3. APPLICATION TO AIRSHIP PATH-TRACKING PROBLEM

111

controllable closed-loop system:


= fp (V, P)
P
=
P

(6.43)
(6.44)

In order to guarantee an asymptotic output tracking, the reference input


may have the form:
r C1 (P
P
r ) C0 (P Pr )
=P

(6.45)

where Pr may be either a time position reference Pr (t) or a dynamic model


position output Pm (t). The diagonal matrices C0 and C1 define the roots of
the characteristic equation:
+ C1 e + C0 e = 0
e

(6.46)

where e = P Pr is the position error.


In the next section we apply this dynamic inversion formulation to the AURORA airship path-tracking problem.

6.3

Application to the airship path-tracking


problem

This section applies the dynamic inversion approach to the airship path-tracking
problem, using the formulation described in the previous section.
Consider the deterministic no-wind case where the inertial and air velocities
are equal (V = Va ). Moreover, consider the position is given, not relative to
the inertial frame, but to the reference trajectory. In this scenario, we may
represent the airship equations of motion as:
a = M1 (6 Ma Va + Va6 (Ma MBa )Va Eg Sag Fa ) + M1 uf
V
a
a
(6.47a)
= J V V
P
(6.47b)

= [T ,
T ]T corresponds to the position error relative to the reference
where P
trajectory. The transformation matrix between the local and the reference

112

CHAPTER 6. DYNAMIC INVERSION

with
= r the attitude
trajectory frames is given by J = J(),
between frames. Vr corresponds to the reference groundspeed given in the
reference trajectory frame, as is assumed constant.
We may represent (6.47a)-(6.47b) in the general form (6.35a)-(6.35b) as 2 :
a = fv (Va , P)
+ gv uf
V
= f (V , P)

P
p

(6.48a)
(6.48b)

with:
= M1 (6 Ma Va + Va6 (Ma MBa )Va Eg Sag Fa ) (6.49)
fv (Va , P)
a

fp (Va , P) = J Va Vr
(6.50)
1

gv = Ma

(6.51)

Take the position relative to the reference trajectory as the output variables
of interest:

y=P

(6.52)

According to (6.42), applying the force control input:




p
uf = (fp (Va )gv )1 fp (Va )fv fp (P)f

(6.53)

with:

c c c s s s c c s c + s s 0

s s
c

s
c

s c s s s + c c s s c c s

s
c s
c c

fp (Va ) =
0
0
0

0
0
0

0
0
0

s c
c

(6.54)

c
c

+ r )
We may write the transformation matrix in the gravity force as S = S() = S(
and assume r to be constant.
2

6.3. APPLICATION TO AIRSHIP PATH-TRACKING PROBLEM

=
fp (P)

0
0
0
0
0
0

0
0
0
0
0
0

113

0 (c s c + s s ) v + (c s s + s c ) w
0 (s s c c s ) v + (s s s c c ) w
0
c c v c s w

s c q
s s r
0
c c
0
s q c r
c q
s r
0
c c

(6.55)

c s u + c c s v + c c c w s c u + (s s s c c ) v + (s s c + c s ) w

s s u + s c s v + s c c w

c c u + (c s s s c ) v + (c s c + s s ) w

c u s s v s c w

s q +

(s )2 s q
(c )2

+ c r +

(s )2 c r
(c )2

0
s qs
(c )2

0
c rs
(c )2

= c and sin()
= s , will lead to an
using the condensed notation cos()
asymptotic output tracking if the new reference input has the form (6.45).
In the airship case we verified a time reference excites unmodelled dynamics.
To solve this problem, and in order to provide some robustness to the dynamic
r P
m.
inversion solution, we consider tracking a model dynamics, i.e., P
The obvious choice is the linear 12-states system with the LQR state feedback
control described in Chapter 5, which, besides robust, takes into account the
real limitations of the airship system. The model dynamics is described by:
m
m = (A BK)
x
xm = Ac x
m
ym = C
xm = P

(6.56)
(6.57)

with the position as output. We then have:


m
y m = CAc x

(6.58)

m
m = CA2c x
y

(6.59)

and the new input in (6.53) is:


C
m C1 (J Va CAc x
m ) C0 (P
xm )
= CA2c x

(6.60)

The control input (6.53) exists if fp (V)gv is nonsingular, i.e, as long as


6= /2, a reasonable assumption in the case of the stable airship

114

CHAPTER 6. DYNAMIC INVERSION

platform.
The control input, however, is a force input, which cannot be directly fed to
the airship. The control law (6.53) considers an input uf that includes both
forces and torques. However, the real inputs of an airship are its actuators.
For this reason, a conversion from forces to actuators inputs is necessary for
the proper implementation of the attained controller.
The actuators input u = [e , TL , TR , v , a , r ]T is obtained solving the equations system (2.71). As referred in Section 2.2.1.1, although we have six actuators inputs to control the six forces, several limitations lead to the underactuation of the airship. Moreover, the system of equations (2.71) is not directly invertible, which implies an empiric allocation in some situations.

6.4

Simulation results

This last section demonstrates the performance of the dynamic inversion approach applied to the airship path-tracking problem. We present the results
obtained for the case-study mission and to the sensitivity and robustness to
parameter uncertainty test.

6.4.1

Case-study mission

The first results we present concern the case-study mission described in Section 3.3.1. The mission covers a wide range of the flight envelope, with the
airspeed varying from 3 to 14m/s.
The airship position coordinates and errors are represented in fig. 6.4. The
vertical take-off and landing are well perceived in fig. 6.4(a), as well as the
path-tracking performance. Figure 6.4(b) displays the longitudinal , lateral
and vertical errors. The end of the first half-circle, when the wind is aft
the airship, and the transition between the ascent and the horizontal tracking,
remain problematic mission points, with the last one showing higher position
errors than in the gain scheduling case. But other regions also rise difficulties:
the beginning of the second curve and the stabilization prior to the descent.
In those occasions we notice an increase of the position errors. In order to
avoid saturation of the propellers, probable when the controller tries to rapidly
correct the longitudinal position, the error is limited to 2m, as was in the

115

6.4. SIMULATION RESULTS

(m)

gain scheduling approach.

60
50

20
0
-20
-40
-60
-80

50

100

150

200
Time (s)

250

300

350

400

-30
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

40

(m)

h (m)

30
30
20

15
0
-15

10
0

10
100

300
200

0
100

-100
-200
N (m)

(m)

200

5
0
-5
-10

-100
E (m)

(a) Airship north N , east E and altitude h


position (bold) and projections (normal).

(b) Longitudinal (), lateral () and vertical () errors.

Figure 6.4: Airship position coordinates and errors.


Comparing the airship trajectory executed with the dynamic inversion controller (fig. 6.4(a)) with the one controlled by the gain scheduling approach
(fig. 5.15(a)), we observe that both control laws lead to the accomplishment of
the mission within acceptable deviations from the trajectory reference. However, the dynamic inversion resulting trajectory is more erroneous.
Figure 6.5 represents the airship horizontal trajectory and its attitude during
the mission. The airship north-east coordinates and heading during the mission
are described in fig. 6.5(a). The preferential alignment with the wind during
take-off and landing is again well recognized. Just as in the gain scheduling
case, the airship appears as slightly crabbing during the maneuvers at low
airspeeds. The Euler angles evolution is displayed in fig. 6.5(b). The roll angle
() presents higher oscillations in the transition from vertical to horizontal
tracking and when the wind is aft, same as in the gain scheduling case but
with higher amplitude. The pitch () and yaw () angles try to follow the
respective references described in Section 3.2.
The airship ground velocity and the aerodynamic variables are depicted in
fig. 6.6 and are somewhat similar to the ones obtained in the gain scheduling
case (see fig. 5.17). The ground velocity components are described in fig.
6.6(a). The longitudinal groundspeed u mostly follows the reference that varies
between 0m/s for stabilization, take-off and landing, and 7m/s during the

116

CHAPTER 6. DYNAMIC INVERSION

250

30
wind
heading

(deg)

200

150

15
0
-15
-30

50

100

150

200
Time (s)

250

300

350

400

30
20
10
0
-10
-20
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

100

(deg)

N (m)

50

-50

-100

(deg)

-150

-200

-250
-100

100
E (m)

200

360
270
180
90
0
0

300

(a) North-east position with airship


heading ( reference, output).

(b) Roll , pitch and yaw ( reference, output).

Figure 6.5: Airship north-east position and attitude.

Vt (m/s)

10
8
6
4
2
0

4
2
0
-2
-4
-6
0
2
1
0
-1
-2
-3

50

100

150

200
Time (s)

250

300

350

400

14
12
10
8
6
4
2

50

100

150

200
Time (s)

250

300

350

400

-20
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

20

50

100

150

200
Time (s)

250

300

350

400

(deg)

w (m/s)

v (m/s)

(deg)

u (m/s)

path-tracking. Along the circular segments, the errors are more noticeable due
to the change of the wind incidence angle while the airship is turning. The
error is higher when the wind is aft the airship. The lateral velocity v is also
mostly influenced by the circular segments and during the tail wind segment.
The vertical velocity w follows its reference, with the 1 and 0.5m/s steps
corresponding to the take-off and landing vertical motion. The airspeed and

50

100

150

200
Time (s)

250

300

350

400

(a) Groundspeed: longitudinal u, lateral v


and vertical w ( reference, output).

10
0
-10

20
10
0
-10
-20
-30

(b) Airspeed Vt , sideslip angle and angle


of attack .

Figure 6.6: Airship ground velocity components and aerodynamic variables.


aerodynamic angles can be seen in fig. 6.6(b). During the whole mission, the
airspeed Vt varies significantly, from values around 3m/s up to 14m/s. The

117

6.4. SIMULATION RESULTS

airship covers a wide flight envelope, from hover to the aerodynamic flight,
crossing the troublesome transition region between the two. Here, the behavior
of shows also correlation with w.

30
15

0
-15
-30

XT (N )

a (deg)

30
15

100

150

200
Time (s)

250

300

350

0
-15
-30

400

50

100

150

200
Time (s)

250

300

350

400

-30
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

80

30

60

15

40
20
0
0

50

100

150

200
Time (s)

250

300

350

400

120
90
60
30
0
-30

0
-15

-10
TD (N )

v (deg)

50

r (deg)

e (deg)

The actuators input is described in fig. 6.7, with the longitudinal actuators
elevator e , total thrust XT and vectoring angle v in fig. 6.7(a) and the lateral
input, aileron a , rudder r and differential thrust TD , in fig. 6.7(b). The
elevator e , while responsible for the altitude and pitch control, shows a more
constant demand during the vertical displacements. The rudder r , mainly
responsible for the lateral position and airship yaw, has a higher command
with tail wind, due to the reduced authority at lower airspeeds. The airship
roll is controlled by the aileron a at higher airspeeds, and by the differential
thrust TD when the control surface loses authority (which usually corresponds
to a vectoring angle close to 90o , allowing TD to effectively control the roll and
not the yaw). This actuator has a negligible action in aerodynamic flight (the
control surfaces authority is sufficient for the rudder r to control the airship
yaw ), an option taken when converting the forces input to an actuators
request.

10
0

50

100

150

200
Time (s)

250

300

350

400

(a) Longitudinal actuators: elevator e , total thrust XT and vectoring v .

(b) Lateral actuators: aileron a , rudder r


and differential thrust TD .

Figure 6.7: Airship actuators input.


The vectoring angle v is responsible for the airship lift when the airspeed Vt
is too low to provide the necessary aerodynamic lift. The correlation between
these two variables, although not as obvious as in the gain scheduling case,
is visible when comparing the graphics of v and Vt . At higher airspeeds, the
dynamic inversion control law chooses to use this actuator to help the elevator

118

CHAPTER 6. DYNAMIC INVERSION

on the altitude and pitch control, unlike the gain scheduling controller.
As a remark, note that the engines inputs are coupled, as are the tail surfaces
due to the -shape. Note, for instance, that when the rudder r is saturated,
the other two control surfaces inputs, elevator e and aileron a , are zero.
In the overall, the airship under dynamic inversion control executed the mission
satisfactorily, although with a more erroneous behavior than under the gain
scheduling control.

6.4.2

Sensitivity and robustness to parameter uncertainty

A fundamental assumption in the dynamic inversion methodology is that the


plant dynamics can be perfectly modeled and may be canceled exactly. In
practice, this assumption is obviously not realistic, and the robustness of the
closed-loop dynamics must be secured, in order to suppress any undesired
behavior due to plant uncertainties and wind disturbances.
This section evaluates the stability and performance robustness of the dynamic
inversion control methodology when solving the path-tracking control problem
of the AURORA airship. Due to the nonlinearity of both system and dynamic
inversion control law, we cannot make use of the analysis tools used in the gain
scheduling robustness analysis. Therefore, we limit our analysis of the dynamic
inversion closed-loop system robustness to the test described in Section 3.3.2.
For the baseline simulation, we consider no variation of the model parameters,
only wind disturbance input for the aerodynamic flight at 8m/s groundspeed
and 50m altitude. Figure 6.8(a) shows the airship north-east position and
heading when following the straight line reference aligned with north, while
subject to the 4m/s constant wind blowing from west, plus 3m/s turbulent
gust. We notice the airship is able to follow the reference, although with an
orientation that helps it minimize the drag force produced by the lateral wind.
Figure 6.8(b) represents the airship attitude references (see Section 3.2) and
output, as well as the wind estimated attitude. Notice that neither w nor w
are constant, since they echo both constant and gust wind components. The
excitation of the signals is due to the wind turbulence.
As may be expected, the airship position errors oscillate around zero instead
of converging, due to the wind turbulence input, as may be seen in fig. 6.9(a).

119

6.4. SIMULATION RESULTS

10
(deg)

1200

-10
1000

50

100

150

50

100

150

50

100

150

50

100

150

50

100

150

(deg)

10

800

5
0
-5
-20

(deg)

N (m)

wind
heading
600

-25
-30
-35

400

w (deg)

5
0
-5

200

-10

w (deg)

100
0

-100

-50

50

E (m)

(a) Airship north-east position with


airship heading ( reference,
output).3

90

80

Time (s)

(b) Airship attitude : roll , pitch and


yaw ( reference, output), and estimated wind attitude, w , w .

Figure 6.8: Airship north-east trajectory and attitude, and wind attitude.
The aerodynamic variables are represented in fig. 6.9(b). The around 9m/s
airspeed corresponds to the relative airspeed between the 8m/s groundspeed
heading north and the 4m/s wind speed from west. The sideslip angle is
close to zero, showing the airship is aligned with the relative airspeed.
The actuators input applied to the AURORA airship is represented in fig. 6.10,
with the longitudinal actuators action given in fig. 6.10(a), while the lateral
actuation is in fig. 6.10(b). The vectoring angle v , which is expected to have
a negligible action in aerodynamic flight, is clearly helping the elevator to
cancel the altitude error . The differential thrust TD , represented here for
completeness, has a negligible control action in aerodynamic flight, an option
taken when converting from forces to actuators.
Table 6.1 shows the RMS values of selected variables. They are the airship
positions errors, namely longitudinal , lateral and vertical errors, and the
true airspeed Vt , the angle of attack and the sideslip angle , together with
3

The different scales might give a wrong idea of the airship heading ( 27o ).

120

CHAPTER 6. DYNAMIC INVERSION

10
Vt (m/s)

(m)

1
0
-1

50

100

9
8

150

50

(deg)

(m)

1
0
-1
0

100

50

-5

150

50

150

100

150

time (s)
10
(deg)

(m)

100

Time (s)
3
2
1
0
-1
-2
-3

150

-2

100
time (s)

Time (s)

50

100

5
0
-5

150

50

Time (s)

time (s)

(a) Airship longitudinal (), lateral () and


vertical () errors.

(b) Airspeed Vt , sideslip angle and angle


of attack .

Figure 6.9: Airship position errors and aerodynamic variables.


5
a (deg)

e (deg)

20
10
0
-10

50

100

0
-5

150

50

20
10

r (deg)

XT (N )

80
60
40
20
0

100

50

100

150

100

150

0
-20

150

50
Time (s)

1
TD (N )

v (deg)

150

-10

Time (s)
120
90
60
30
0
-30

100
Time (s)

Time (s)

50

100

150

0
-1

50

Time (s)

(a) Longitudinal actuators input: elevator


e , total thrust XT and vectoring v .

Time (s)

(b) Lateral actuators input: aileron a ,


rudder r and differential thrust TD .

Figure 6.10: Airship actuators input.

the groundspeed error eu relative to the 8m/s reference. The first row has the
RMS values obtained for the baseline case, and is to serve as reference for the
remaining lines where each of the listed coefficients is varied one at a time.
For a parameter uncertainty of 70%, the controlled airship still performed
qualitatively like the baseline case, except for a few cases. For the coefficients
CMe and CMq , a +30 and 30% uncertainty respectively already influenced
the controller performance, while for CL this influenced appeared for 50%.
The dynamic inversion controller appeared to be affected only by the increase

6.5. CONCLUSIONS

121

of parameter uncertainty of these three longitudinal coefficients, respectively


the authority of the elevator as pitching control input, the pitch damping
derivative and the lift coefficient derivative due to angle of attack.
To see how the dynamic inversion controller behaved for higher levels of uncertainty, we then increased the parameter uncertainty to 90% around the
nominal value. We present the results obtained in table 6.1, with the cases
which lead to significant deviations from the reference RMS values represented
in bold.
For the previous coefficients, CMe (90%), CMq (90%) and CL (90%),
the uncertainty still leads to an inefficient control action, either because it is
insufficient due to actuators saturation (the dynamic inversion demands a too
high control input) and/or just because the tested coefficient appears to be
a more sensitive model parameter. Besides these three parameters, we now
notice the influence of the uncertainty in lateral coefficients, CNr and CNr
(both for 90%), respectively the authority of the rudder as yawing control
input, and the yaw damping derivative.
While the dynamic inversion demonstrates to be robust to a 90% uncertainty
in the remaining parameters, for these five parameters, the mismatch between
the airship system and the model considered in the dynamic inversion controller
design is too significant for the control action to overcome it.
In any case, the dynamic inversion controller may be considered robust to
wind disturbances and plant uncertainties. Among the list selected, these
five parameters are in fact the model parameters for which a more careful
identification or determination should take place.

6.5

Conclusions

This chapter describes the general theory of the dynamic inversion approach.
Bearing in mind systems like the airship whose dynamic equations are given
by a cascaded description, we formulate a more straightforward procedure to
obtain the dynamic inversion control law for this type of systems.
The next step is obviously its application to the AURORA airship pathtracking problem, where the controlled system shows a satisfactory performance in the execution of realistic missions.
The dynamic inversion approach is based on the cancelation of the system

122

CHAPTER 6. DYNAMIC INVERSION

Table 6.1: Robustness tests on model parameters (RMS values of selected variables).
Baseline

(m)

(m)

(m)

Vt (m/s)

(deg)

(deg)

eu (m/s)

0.60

0.45

0.78

9.01

3.26

1.21

0.14

Cl

90%
+90%

0.59
0.62

0.41
0.51

0.78
0.74

9.01
9.01

3.31
3.03

0.95
1.80

0.12
0.19

CM0

90%
+90%

0.58
0.61

0.45
0.45

0.79
0.76

9.01
9.01

3.32
3.20

1.22
1.21

0.14
0.14

CM

90%
+90%

0.48
0.65

0.44
0.46

1.31
0.73

9.01
9.01

3.96
2.80

1.24
1.20

0.14
0.14

CM

90%
+90%

0.60
0.60

0.45
0.45

0.77
0.78

9.01
9.01

3.25
3.28

1.21
1.21

0.14
0.14

CM

90%
+90%

0.62
0.58

0.45
0.45

0.76
0.80

9.01
9.01

3.19
3.34

1.21
1.22

0.14
0.14

CM

90%
+90%

0.60
0.60

0.45
0.45

0.77
0.78

9.01
9.01

3.25
3.27

1.21
1.21

0.14
0.14

CMe

90%
+90%

19.40
62.17

2.66
1.84

6.79
2.05

8.81
9.91

2.83
4.11

1.47
1.47

0.44
1.44

CN

90%
+90%

0.59
0.60

0.43
0.45

0.76
0.78

9.01
9.01

3.28
3.25

1.33
1.13

0.12
0.15

CNr

90%
+90%

19.98
0.49

16.05
0.34

6.61
0.79

9.44
9.01

4.47
3.31

5.77
1.35

1.98
0.19

CY

90%
+90%

0.55
0.61

0.72
0.41

0.80
0.78

9.02
9.01

3.30
3.27

2.93
0.88

0.17
0.13

CYr

90%
+90%

0.58
0.61

0.44
0.45

0.77
0.77

9.01
9.01

3.33
3.19

1.25
1.50

0.12
0.16

CD0

90%
+90%

0.53
1.07

0.49
0.46

1.02
0.98

9.01
9.01

4.13
2.60

1.26
1.17

0.18
0.14

CDi

90%
+90%

0.55
0.64

0.45
0.45

0.81
0.75

9.01
9.01

3.36
3.18

1.21
1.21

0.15
0.13

CL0

90%
+90%

0.60
0.59

0.45
0.45

0.77
0.78

9.01
9.01

3.18
3.34

1.21
1.22

0.14
0.14

CL

90%
+90%

140.23
0.69

42.03
0.46

87.44
0.60

9.38
9.01

27.25
1.87

8.11
1.17

3.68
0.13

CLe

90%
+90%

0.56
0.63

0.43
0.46

0.72
1.04

9.01
9.01

3.82
3.10

1.24
1.20

0.13
0.15

Clp

90%
+90%

0.61
0.60

0.45
0.46

0.79
0.77

9.01
9.01

3.18
3.28

1.40
1.18

0.15
0.14

CMq

90%
+90%

265.78
0.78

3.12
0.52

1.52
2.09

12.79
9.04

3.41
4.72

1.33
1.21

4.65
0.24

CNr

90%
+90%

597.96
0.64

189.82
0.77

6.01
0.80

4.19
9.01

36.24
3.23

73.23
1.40

4.81
0.16

mw

90%
+90%

0.67
0.33

0.48
0.58

1.29
1.45

9.02
9.02

1.52
3.92

1.52
1.37

0.21
0.16

6.5. CONCLUSIONS

123

nonlinearities by inverting the dynamic model, resulting in a stable and linear


closed-loop system. However, this procedure assumes a complete knowledge of
the system. Therefore, the analysis of the controlled system performance and
stability robustness in the presence of wind disturbances and model parameter
uncertainties is very important. The dynamic inversion controller, robust to
wind disturbances, shows to be tolerant to uncertainties in most of the model
parameters tested. However, for some aerodynamic coefficients, namely CMe ,
CMq , CL , CNr and CNr , a more careful identification or determination should
take place.
The dynamic inversion controller requires a time reference or a model dynamics
is given as reference. In the airship case we realized a time reference excites
unmodeled dynamics. In fact, the dynamic inversion of the airship nonlinear
model results in a forces input, which is assumed by the controller to be fully
available. This is however not the case. The airship has serious actuation
constraints regarding the lateral force and, although not so severe, with the
downward force as well. An obvious choice for reference is then the closed-loop
gain scheduling system, which we have seen to be robustly stable within certain
limits, and that intrinsically provides information on the actuation constraints.
However, giving this model as reference, can we expect the dynamic inversion
to outperform it? If it is not the case, what is then the advantage of the
dynamic inversion solution over the gain scheduling one?
With the purpose of avoiding this type of problems, an in the absence at the
time of a better reference for the dynamic inversion controller, we considered
a different nature of nonlinear control, which we will describe in the sequence.

124

CHAPTER 6. DYNAMIC INVERSION

Chapter 7
Backstepping
Contents
7.1

Wind estimator

. . . . . . . . . . . . . . . . . . . .

126

7.2

Backstepping design approach . . . . . . . . . . . .

128

7.3

Application to the path-tracking problem . . . . .

128

7.4

Control design with saturation constraints . . . .

132

7.5

Control implementation

136

7.5.1
7.6

7.7

. . . . . . . . . . . . . . .

Adapted control law to deal with underactuation . . 136

Simulation results . . . . . . . . . . . . . . . . . . .

138

7.6.1

Case-study mission . . . . . . . . . . . . . . . . . . . 138

7.6.2

Sensitivity and robustness . . . . . . . . . . . . . . . 142

Conclusions . . . . . . . . . . . . . . . . . . . . . . .

146

Although Lyapunovs direct method is originally a method of stability analysis, an important application is the design of nonlinear controllers. Backstepping [60] is a recursive procedure that interlaces the choice of a Lyapunov
function with the design of feedback control. In the backstepping approach,
by formulating a scalar positive function of the system states and then considering a control law that makes this function decrease, we have the guarantee
that the nonlinear control system thus designed will be asymptotically stable,
and still robust to some unmatched uncertainties.
Several successful applications of the backstepping approach for UAV control
have been reported [73, 15, 74, 75, 76].
Backstepping as also been used in the specific case of airships. A backstepping technique has been proposed by the LAAS/CNRS autonomous blimp
project [39, 40]. The global control strategy studied is obtained by switching
125

126

CHAPTER 7. BACKSTEPPING

between four sub-controllers, one for each of the flight phases considered. Each
controller is however still based on linearized models of the airship, what leads
to the separate control of the longitudinal and lateral motions. An imagebased control solution for airship tracking is based by Fukao et al. [43] on
backstepping techniques for underactuated vehicles. The airship model considered is built from the kinematics between the camera used as only sensor
and the target. Robustness issues like wind disturbance rejection are still to
be improved.
In this chapter we propose a backstepping based control solution to the airship
path-tracking problem. Based on the six-degrees-of-freedom nonlinear model
of the airship, it is valid for missions over the entire flight envelope.
Before going into the backstepping airship path-tracking control design, we
first describe in Section 7.1 a wind estimator. The interest of using this estimator instead of the wind estimation method described in Section 3.2 lies, not
in a better estimation, but in the fact that it provides useful bounds, as we
will later see. We then present a general backstepping control approach in Section 7.2, applying it to the path-tracking problem in Section 7.3. Saturation
limits are included in the control design in Section 7.4 and important implementation issues are discussed in Section 7.5.1. The controller performance is
evaluated in Section 7.6, with simulation results concerning the case-study mission and the sensitivity and robustness to parameter uncertainty test. Finally,
the conclusions are drawn in Section 7.7.

7.1

Wind estimator

The dynamics equations (2.66) or (2.69) and the kinematics equation (2.80),
expressed in the air frame, assume a constant translation wind. However, the
wind disturbance is unknown, being necessary to build an estimator based
on (2.48)-(2.49) and (2.80). Since the wind input is not affecting the angular
position part in (2.80), only the cartesian position p of the airship should be
considered:
p = ST va + p w

(7.1)

127

7.1. WIND ESTIMATOR

w ), and its dynamics may be chosen as:


The estimator states may then be (
p, p
"

p
w
p

"

ST va
0

"

Lp I 3
Lpw 03

#"

pp

p w

(7.2)

leading to an estimation error vector dynamics obtained from (7.1)-(7.2) and


given by:
=

"

Lp I3
Lpw 03

#"

pp
w
p w p

= A

(7.3)

where the two constant matrices (Lp , Lpw ) are chosen so that A be Hurwitz.1
Therefore, the origin of (7.3) is asymptotically stable and there exists a Lyapunov function:
We = T P

(7.4)

with a symmetric positive definite matrix P > 0 whose time derivative is


given by:
e = T P + T P = T (P A + AT P ) = T Q
W

(7.5)

Choosing the symmetric positive definite matrix Q with a block diagonal


form, the P matrix will then be the solution of the Riccati equation:
P A + AT P = Q =

"

Qp 03
03 Qpw

(7.6)

with Qp and Qpw diagonal matrices.


Defining the estimation errors as:
=pp

w
w = p w p
p

(7.7)
(7.8)

the estimator Lyapunov function derivative (7.5) may be rewritten as:


e =

Tw Qpw p
p
W
pT Qp p
2

(7.9)

For instance, taking Lp = aI3 and Lpw = ab I3 leads to three pairs of poles at a2 (1i)
q

2
1
with = 4b
b , i.e., with a damping factor = 2 b.
1

128

7.2

CHAPTER 7. BACKSTEPPING

Backstepping design approach

Let us consider a generic control problem with output y. We first define two

auxiliary outputs involving the output y and its derivative y:


(

y1 = ay + y

y2 = y

y 1 = ay + y

y 2 = y

(7.10)

where a is a positive scalar to be used as design parameter. It is easily seen


that when both auxiliary outputs are taken to the origin, the regulation of the
main output y is then achieved.
A candidate Lyapunov function may be:
1
1
W0 = y1T y1 + y2T y2
2
2

(7.11)

Its derivative is:


0 = y1T y 1 +y2T y 2 = (ay + y)
= (ay + 2y)
T (ay + y
)ay T y
T (ay + y
)+ y T y
W
(7.12)
If the control is chosen in order to give:
= (ay + 2y)

ay + y

(7.13)

where = T is a positive definite matrix, then the derivative:


0 = (ay + 2y)
T (ay + 2y)
ay T y
W

(7.14)

will clearly be negative definite and the system will be globally asymptotically
stable.

7.3

Application to the path-tracking problem

We shall now proceed applying the control design described in the previous
section to the path-tracking problem.
Let us assume a point pr with a constant ground velocity vr is to be tracked
with constant attitude along a rectilinear path AB (see figure 7.1):
p r = STr vr

(7.15)

129

7.3. APPLICATION TO THE PATH-TRACKING PROBLEM

where Sr = S(r ) given by (2.17) is the constant transformation matrix from


the inertial frame to the reference path.
B
Sa r

vw

vr
va r
p

Figure 7.1: Air velocity reference estimation (2D).

As the wind velocity vw is considered, the desired air velocity var may be
deduced. Moreover, since the airship is being aligned with this air velocity
we have a reference for the attitude given by the transformation Sar from the
inertial frame to the air velocity var , which is described by the desired attitude
vector ar . This leads to the reference position:
Pr =

"

pr
ar

(7.16)

The derivative of this reference position is:


r =
P

"

STr vr
0

= Jr Vr

where the reference velocity state is Vr =

"

vr
0

(7.17)

and Jr =

"

STr 03
03 Rr

with Rr = R(r ) given by (2.18).


Note that, although we have assumed a rectilinear reference path, the approach
may also be extended to the cases where the reference path varies slowly, with
negligible derivatives when compared to the state derivative.
Let us now consider a Lyapunov function candidate similar to (7.11):
1
1
Wt = y1T y1 + y2T y2
2
2

(7.18)

where the output auxiliary variables y1 and y2 are again derived from the

130

CHAPTER 7. BACKSTEPPING

but where y = P Pr is the position tracking


output y and its derivative y,
error:
(
P
r
y1 = ay + y = a(P Pr ) + P
(7.19)
P
r
y2 = y = P
Using equations (2.80) and (7.17) we have:
(

y1 = a(P Pr ) + J Va + BI p w Jr Vr
y2 = J Va + BI p w Jr Vr

(7.20)

The derivative of the Lyapunov function candidate (7.18) is:


t = y1T y 1 + y2T y 2
W
T (ay + y
) + y T y

= (ay + y)
T (ay + y
) ay T y
= (ay + 2y)

(7.21)

If the control is chosen as:


= (ay + 2y)

ay + y

(7.22)

or:
a=
a(J Va + BI p w Jr Vr ) + J CJ Va + J V
(a(P Pr ) + 2(J Va + BI p w Jr Vr )) (7.23)
where we used:
a + J CJ Va
= J V
y

(7.24)

this leads to the control law:


a = a(P Pr ) J CJ Va (aI6 + 2)(J Va + BI p w Jr Vr ) (7.25)
J V
As wind is estimated, the suggested control law is:
a = a(P Pr ) J CJ Va (aI6 + 2)(J Va + BI p
w Jr Vr ) (7.26)
J V
and (7.22) should be rewritten as:
w
= (ay + 2y)
+ (aI6 + 2)BI p
ay + y

(7.27)

7.3. APPLICATION TO THE PATH-TRACKING PROBLEM

131

and defining G = ( a2 1 + I6 )BI , the tentative


Introducing y0 = ay + 2y,
Lyapunov derivative appears as:
t = y0T (y0 2Gp
w ) ay2T y2
W

(7.28)

or, completing the squares2 :


t = (y0 Gp
w ay2T y2
Tw GT Gp
w ) + p
w )T (y0 Gp
W

(7.29)

If we now consider a corrected tentative Lyapunov function with the wind


estimator from Section 7.1:
W = Wt + We

(7.30)

may be written using (7.29) and (7.9) as:


the derivative W
= (y0 Gp
w
Tw (Qpw GT G)p
w ) ay2T y2 p
w )T (y0 Gp
p
T Qp p
W
(7.31)
which is negative definite if:
Qpw GT G > 0

(7.32)

The control law may be deduced from equations (2.69) and (7.26), leading to:
a KVa ) Eg Sag Fa
uf = Ma (V
a = aJ1 (P Pr ) CJ Va J1 2 (J Va + BI p
w Jr Vr )
V
2

(7.33)
(7.34)

where K = Ma (6 Ma + Va6 (Ma MBa )) and 22 = (aI6 + 2).


The force control input is then given by:


uf = Ma A1 (J Va + BI p w Jr Vr ) + B1 (P Pr ) + C1 Va Eg Sag Fa
(7.35)
1
2
with A1 = J1
2 , B1 = aJ and C1 = CJ + K, resulting in an asymptotically stable closed-loop system.


However, the force control input, as is it, may result in excessively high de2

From the expansion of a square:

w )T (y0 Gp
w ) = y0T (y0 Gp
w )(Gp
w )T (y0 Gp
w ) = y0T (y0 2Gp
w )+(Gp
w )T Gp
w
(y0 Gp
w ) = (y0 Gp
w )T (y0 Gp
w ) p
Tw GT Gp
w
it is easily deduced that: y0T (y0 2Gp

132

CHAPTER 7. BACKSTEPPING

mands for a real system subject to input constraints. In the next section the
control solution (7.35) will be adapted to deal with this matter.

7.4

Control design with saturation constraints

In order to include saturation limits into the control design, let us rewrite
equation (7.22), corresponding to a second derivative demand:
= ay (ay + 2y)
= (aI6 + 2)y ay = 22 y 1 y
y

(7.36)

with 1 = a and 22 as defined in the previous section.


Defining the second Lyapunov function as:
1
W2 = y2T y2
2

(7.37)

its derivative may be expressed as:


with, as before, y2 = y,
2 = y T y
= y T (22 y + 1 y) = zT2 (z2 + z1 )
W

(7.38)

Writing (7.36) as function of z1 and z2


where z1 = 1
2 1 y and z2 = 2 y.
yields:
= 2 (z2 + z1 )
y

(7.39)

Before proceeding, we will now define linear saturation as well as its properties,
and provide an important theorem used in the proof of stability of the saturated
control.

Definition 7.1. As a particular case and extension of the linear saturation


definition proposed by Teel [63], let us introduce the elementwise nondecreasing
saturation function : Rn Rn , defined by a vector m of n positive values
mi , with mi > r > 0, and such that:
z Rn , [z] = z

(7.40)

7.4. CONTROL DESIGN WITH SATURATION CONSTRAINTS

133

where the diagonal matrix is defined by:


|zi | < mi i = 1
mi
|zi | mi i =
|zi |

(7.41)

Properties 7.1. It may easily be verified that the definition yields the following
properties [63]:

z Rn ; zT [z] > 0

(7.42)
z Rn ; |[z]| R

|z| < r [z] = z

P
where |z| = zT z is the norm of vector z as defined in Rn and R2 = ni=1 m2i .

Theorem 7.1. If two saturations 1 and 2 are defined, such that R1 < 21 r2 ,
then:
1
(7.43)
(z1 , z2 ) Rn , |z2 | > r2 zT2 2 [z2 + 1 [z1 ]] > 0
2
Proof. Since |z2 | > 12 r2 and | 1 [z1 ]| R1 < 21 r2 , one can write the orthogonal
projection of the saturated vector 1 [z1 ] on z2 as:
1 [z1 ] = 1 z2 + v1

(7.44)

where |1 | < 1, zT2 v1 = 0, and |1 z2 + v1 | < 12 r2 .


Then:
zT2 2 [z2 + 1 [z1 ]] = zT2 2 [(1 + 1 ) z2 + v1 ]
= zT2 2 ((1 + 1 ) z2 + v1 )

(7.45)

= (1 + 1 ) zT2 2 z2 > 0

We can now proceed and introduce the second derivative (7.39) saturated
demand:
s = 2 2 [z2 + 1 [z1 ]]
y

(7.46)

2 = zT2 2 [z2 + 1 [z1 ]] will be


From Theorem 7.1, if |z2 | > 12 r2 , then W
negative definite for saturations 1 such that | 1 [z1 ]| R1 < 21 r2 .
Since the saturated system is asymptotically stable, after a time T2 the variable
z2 will enter the linear zone of its saturation and remain inside of it, and namely

134

CHAPTER 7. BACKSTEPPING

with |z2 | < 12 r2 .


After time T2 the saturated demand will be equal to:
s = 2 (z2 + 1 [z1 ])
y

(7.47)

Introducing (7.47) into (7.21) yields:


ts = (ay + 2y)
T (ay + y
s ) ay T y
W

T (ay 2 (z2 + 1 [z1 ])) ay T y


= (ay + 2y)

T (ay 2 z2 2 1 [z1 ]) ay T y
= (ay + 2y)

T (ay (aI6 + 2)y 2 1 [z1 ]) ay T y


= (ay + 2y)
T (2y + 1 2 1 [z1 ]) ay T y
= (ay + 2y)

(7.48)

Using the definition of the saturation 1 [z1 ] = z1 = a1


2 y, from (7.48)
we get:

ts = (2y + ay)T 2y + 1 2 1 1
W
ay
ay T y
2

(7.49)

Two scenarios are now possible: (i) z1 is not saturated, in which case 1 = I,
ts < 0; or (ii) z1 is saturated and 1 = m1i 1. Let us further
resulting in W
i
|z1i |
analyze this case.
s = a1/2 1 y, and Z = 1
Taking z0 = 21/2 y,
1 , we have:
ts = (z0 + Zs)T (z0 + s) ay T y
W

(7.50)

If we consider the decomposition of the vectors in their components z0 = [zi ]


|z |
and s = [si ] and also the diagonal matrix Z = [i ] with elements i = m11i 1,
i
then:
(z0 + s)T (z0 + Zs) =

(zi + si )(zi + i si )

(7.51)

Noting that s and z0 have behaviors similar to, respectively, z1 and z2 , and
that z2 is in its linear zone and converging, we have that after some time

7.4. CONTROL DESIGN WITH SATURATION CONSTRAINTS

135

|zi | < |si |, i and then zi = i si with |i | < 1, so that:


(z0 + s)T (z0 + Zs) =

(i si + si )(i si + i si )

(7.52)

(si )2 (i + 1)(i + i )

(7.53)

X
i

which shows that each term is positive, making the result of the sum also
ts is negative definite.
positive. Therefore, W
To include the input forces limitations into the control law design, let us con=y
s . From (7.24) and (7.47)
sider the desired demand is a saturated one, y
we obtain:
a + J CJ Va =
J V


(7.54)
2 2 2 (J Va + BI p w Jr Vr ) + 1 [1
2 1 (P Pr )]

a:
or, solving for V



a = J1 2 2 2 (J Va + BI p w Jr Vr ) + 1 [1 1 (P Pr )] CJ Va
V
2

(7.55)
Substituting now (7.55) into (7.33) leads to the control law:


1

V
)
+

(P

P
)]
p

(J
V
+
B

uf s = Ma J1
r
1
1
r
w

a
I
2
2
r
2

Ma C1 Va Eg Sag Fa

(7.56)

Again, as the wind is estimated, the control law that considers the force input
saturations is finally given by:
i
h
w Jr Vr ) + 1 [1

(P

P
)]
p

(J
V
+
B

uf s = Ma J1
1
r
2

a
I
2
2
2

Ma C1 Va Eg Sag Fa

(7.57)

where 1 and 2 are the velocity saturation matrices obtained from (7.57)
with uf s corresponding to the input force maximum values related to the
actuators limits (see sections 2.1 and 2.2.1.1), and that satisfy the condition
R1 < 12 r2 < |z2 |. This control law will lead to an asymptotically stable closedloop system as long as the estimation error is bounded according to (7.31).

136

7.5

CHAPTER 7. BACKSTEPPING

Control implementation

The control law (7.57) solves the airship path-tracking problem in the presence
of constant translational wind while taking into account the limitations of the
demanded forces input. However, this control law cannot be directly fed into
the system, and needs to be adapted.
Although the control law assumes 3 forces and 3 torques are fully available, this
isnt really the case, since the airship is an underactuated vehicle, as described
in Section 2.2.1.1. The available actuation results in weak lateral and vertical
forces responses, and therefore, the position and velocity references used in
the backstepping control law should be shaped to deal with this scenario.
This matter is analyzed in the next section.
As in the dynamic inversion controller case, the backstepping control input is a
force input, which cannot be directly fed to the airship. The control law (7.57)
considers an input uf s that includes both forces and torques. However, the
real inputs of an airship are its actuators. For this reason, a conversion from
forces to actuators inputs is necessary for the proper implementation of the
attained controller.
The actuators input u = [e , TL , TR , v , a , r ]T is obtained solving the equations system (2.71). As referred in Section 2.2.1.1, although we have six actuators inputs to control the six forces, several limitations lead to the underactuation of the airship. Moreover, the system of equations (2.71) is not directly invertible, which implies an empiric allocation in some situations.

7.5.1

Adapted control law to deal with underactuation

As referred, the present configuration of the AURORA airship actuators results


in an underactuated system. At very low airspeeds we reach the worst-case
scenario, with the airship being uncontrollable due to the lack of authority from
the control surfaces. The implementation of the proposed control law assumes
this situation is not reached, therefore requiring that the true airspeed does
not drop below a minimum, Vt > Vtmin = 2m/s (note that it is quite realistic
in outdoor conditions to assume a wind intensity above this level).
Even if this limit is respected, the airship may still be underactuated as the
transversal forces available are too weak. This means the controllers force
request for a straightforward correction of eventual lateral and vertical position

7.5. CONTROL IMPLEMENTATION

137

errors might find insufficient response on the actuators side. In the following,
we adapt the control law (7.57) to deal with this scenario, obtaining a faster
error correction with smoother input requests.
Consider the approximated kinematic relations:
E Vt
D Vt

(7.58)
(7.59)

where and are the pitch and yaw Euler angles [53] that describe the airship
orientation. Equations (7.58)-(7.59) allow us to relate the airship orientation
with its lateral and vertical positions.
Consider now the airship is to track a rectilinear path with orientation (r , r )
and has lateral and vertical errors respectively y and z. The angular errors are
defined as:
= r

(7.60)

= r

(7.61)

Due to the airship underactuation, if we try to independently correct the position and angular errors, depending on their magnitude, we will probably have
input saturation. However, if we consider the relation between position and
attitude, we may consider instead the following expressions:
= r ky y
= r kz z

(7.62)
(7.63)

where the constants ky and kz are dependent of the airspeed Vt . This means
we will postpone the angular corrections and use them to annulate the position
errors first. The angular references used in Pr in the control law (7.57) will
then be:
r = r + ky y

(7.64)

r = r + kz z

(7.65)

138

7.6

CHAPTER 7. BACKSTEPPING

Simulation results

This last section demonstrates the performance of the backstepping approach


applied to the airship path-tracking problem. We present the results obtained
for the case-study mission and to the sensitivity and robustness to parameter
uncertainty test.

7.6.1

Case-study mission

The first results we present concern the case-study mission described in Section 3.3.1.
The airship position coordinates and errors are represented in fig. 7.2. The
vertical take-off and landing are well perceived in fig. 7.2(a), as well as the
path-tracking performance. Figure 7.2(b) displays the longitudinal , lateral
and vertical errors. Like in the gain scheduling and dynamic inversion cases,
the airship deviates from the reference trajectory when the wind is at the rear,
at the end of the first half-circle, and in the transition from ascent to horizontal
tracking (see fig. 7.3(a)). Other problematic mission point, shared with the
dynamic inversion but showing higher errors, is the stabilization prior to the
descent. The correction of the longitudinal and lateral position errors due
to the instantaneous reference change induces a significant vertical error. With
the backstepping solution, the stabilization is more slowly achieved. Remember
that in this solution the controller has no prior information of the actuation
available (the gain scheduling considers the actuators input and the dynamic
inversion indirectly knows the actuation limitation through the gain scheduling
model used as reference), and therefore the errors correction is not optimized.
Comparing the airship trajectory executed with the backstepping controller
(see fig. 7.2(a)) with the ones controlled by the gain scheduling (see fig. 5.15(a))
and dynamic inversion (see fig. 6.4(a)) approaches, we observe that all control
laws lead to the accomplishment of the mission within acceptable deviations
from the trajectory reference. The backstepping resulting trajectory is less
erroneous than the dynamic inversion one, but shows higher vertical errors
than the gain scheduling.
Figure 7.3 represents the airship horizontal trajectory and its attitude during
the mission. The airship north-east coordinates and heading during the mission

139

7.6. SIMULATION RESULTS

(m)

60

-40
-60
-80

50
40
30

(m)

h (m)

-20

20
10
0

50

100

150

200
Time (s)

250

300

350

400

20
10
0
-10
-20
-30
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

4
100

300
200

0
100

-100
-200
N (m)

(m)

200

0
-4
-8

-100
E (m)

(a) Airship north N , east E and altitude h


position (bold) and projections (normal).

(b) Longitudinal (), lateral () and vertical () errors.

Figure 7.2: Airship position coordinates and errors.

are described in fig. 7.3(a). The preferential alignment with the wind during
take-off and landing is again well recognized. During the maneuvers at low
airspeeds the airship motion is smoother than in the previous two cases, due to
the references shaping described in Section 7.5.1. The Euler angles evolution
is displayed in fig. 7.3(b). The roll angle (), with a null reference, has a
significant amplitude in the transition from vertical to horizontal tracking,
which is then corrected. During the descent, the roll is well controlled. The
pitch () and yaw () angles approximately follow the respective references
described in Section 7.5.1.
The airship ground velocity and the aerodynamic variables are depicted in
fig. 7.4 and are somewhat similar to the ones obtained in the previous cases.
The ground velocity components are described in fig. 7.4(a). The longitudinal groundspeed u mostly follows the reference that varies between 0m/s for
stabilization, take-off and landing, and 7m/s during the path-tracking. Along
the circular segments, the errors are more noticeable due to the change of the
wind incidence angle while the airship is turning. The error is higher when
the wind is aft the airship. The lateral velocity v is also mostly influenced by
the circular segments and during the tail wind segment. The vertical velocity
w follows its reference, with the 1 and 0.5m/s steps corresponding to the
take-off and landing vertical motion. The airspeed and aerodynamic angles
can be seen in fig. 7.4(b). During the whole mission, the airspeed Vt varies
significantly, from values around 2.5m/s up to 12.5m/s. The airship covers a

140

CHAPTER 7. BACKSTEPPING

250

(deg)

200

wind
heading

150

60
45
30
15
0
-15
-30

50

100

150

200
Time (s)

250

300

350

400

15
10
5
0
-5
-10
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

100

(deg)

N (m)

50

-50

-100

(deg)

-150

-200

-250
-100

100
E (m)

300

200

(a) North-east position with airship


heading ( reference, output).

450
360
270
180
90
0
-90

(b) Roll , pitch and yaw ( reference, output).

Vt (m/s)

10
8
6
4
2
0
-2

4
2
0
-2
-4
-6
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

(deg)

w (m/s)

2
0
-1
-2

14
12
10
8
6
4
2

50

100

150

200
Time (s)

250

300

350

400

-20
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

20
(deg)

v (m/s)

u (m/s)

Figure 7.3: Airship north-east position and attitude.

50

100

150

200
Time (s)

250

300

350

400

(a) Groundspeed: longitudinal u, lateral v


and vertical w ( reference, output).

10
0
-10

20
10
0
-10
-20
-30

(b) Airspeed Vt , sideslip angle and angle


of attack .

Figure 7.4: Airship ground velocity components and aerodynamic variables.


wide flight envelope, from hover to the aerodynamic flight, crossing the troublesome transition region between the two. Here, the behavior of shows also
correlation with w.
The actuators input is described in fig. 7.5, with the longitudinal actuators
elevator e , total thrust XT and vectoring angle v in fig. 7.5(a) and the lateral input, aileron a , rudder r and differential thrust TD , in fig. 7.5(b). The
elevator e , while responsible for the altitude and pitch control, shows a more
constant demand during the ascent and stabilization phases. The rudder r ,

141

7.6. SIMULATION RESULTS

30
15

0
-15
-30

XT (N )

a (deg)

30
15

100

150

200
Time (s)

250

300

350

0
-15
-30

400

50

100

150

200
Time (s)

250

300

350

400

-30
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

80

30

60

15

40
20
0
0

50

100

150

200
Time (s)

250

300

350

400

0
-15

30

120
90
60
30
0
-30

TD (N )

v (deg)

50

r (deg)

e (deg)

mainly responsible for the lateral position and airship yaw, has a higher command with tail wind, due to the reduced authority at lower airspeeds. The
airship roll is controlled by the aileron a at higher airspeeds, and by the
differential thrust TD when the control surface loses authority (which usually
corresponds to a vectoring angle close to 90o , allowing TD to effectively control
the roll and not the yaw). This actuator has a negligible action in aerodynamic
flight (the control surfaces authority is sufficient for the rudder r to control
the airship yaw ), an option taken when converting the forces input to an
actuators request.

50

100

150

200
Time (s)

250

300

350

400

(a) Longitudinal actuators: elevator e , total thrust XT and vectoring v .

15
0
-15
-30

(b) Lateral actuators: aileron a , rudder r


and differential thrust TD .

Figure 7.5: Airship actuators input.

The vectoring angle v is responsible for the airship lift when the airspeed Vt
is too low to provide the necessary aerodynamic lift. The correlation between
these two variables, although not as obvious as in the gain scheduling case,
is visible when comparing the graphics of v and Vt . At higher airspeeds, the
backstepping control law chooses to use this actuator to help the elevator on
the altitude and pitch control, unlike the gain scheduling controller.
In the overall, the airship under backstepping control executed the mission
satisfactorily.

142

CHAPTER 7. BACKSTEPPING

7.6.2

Sensitivity and robustness to parameter uncertainty

This section evaluates the stability and performance robustness of the backstepping control methodology when solving the path-tracking control problem
of the AURORA airship. Due to the nonlinearity of both system and backstepping control law, we cannot make use of the analysis tools used in the gain
scheduling robustness analysis. Therefore, we limit our analysis of the backstepping closed-loop system robustness to the test described in Section 3.3.2.
For the baseline simulation, we consider no variation of the model parameters,
only wind disturbance input for the aerodynamic flight at 8m/s groundspeed
and 50m altitude. Figure 7.6(a) shows the airship north-east position and

1200

(deg)

10

1000

-10

50

100

150

50

100

150

50

100

150

50

100

150

50

100

150

(deg)

10

800

5
0
-5

N (m)

wind
heading

-20
(deg)

600

-25
-30
-35

400

w (deg)

200

0
-5
-10

w (deg)

100
0

-100

-50

50

E (m)

(a) Airship north-east position with


airship heading ( reference,
output).3

90

80

Time (s)

(b) Airship attitude : roll , pitch and


yaw ( reference, output), and wind
attitude, w , w .

Figure 7.6: Airship north-east trajectory and attitude, and wind attitude.
heading when following the straight line reference aligned with north, while
subject to the 4m/s constant wind blowing from west, plus 3m/s turbulent
3

The different scales might give a wrong idea of the airship heading ( 27o ).

143

7.6. SIMULATION RESULTS

gust. We notice the airship is able to follow the reference, although with an
orientation that helps it minimize the drag force produced by the lateral wind.
Figure 7.6(b) represents the airship attitude references (see Section 7.5.1) and
output, as well as the wind estimated attitude. The excitation of the signals
is due to the wind turbulence.
As may be expected, the airship position errors oscillate around zero instead
of converging, due to the wind turbulence input, as may be seen in fig. 7.7(a).
The aerodynamic variables are represented in fig. 7.7(b). The around 9m/s
airspeed corresponds to the relative air speed between the 8m/s groundspeed
heading north and the 4m/s wind speed from west. The sideslip angle is
close to zero, showing the airship is aligned with the relative airspeed.
10
Vt (m/s)

(m)

1
0
-1

50

100

9
8

150

50

100

50

-5

150

50

150

100

150

time (s)

10
(deg)

(m)

100

Time (s)

0
-1

150

5
(deg)

(m)

-1

100
time (s)

Time (s)

50

100

150

Time (s)

(a) Airship longitudinal (), lateral () and


vertical () errors.

5
0
-5

50
time (s)

(b) Airspeed Vt , sideslip angle and angle


of attack .

Figure 7.7: Airship position errors and aerodynamic variables.


The actuators input applied to the AURORA airship is represented in fig. 7.8,
with the longitudinal actuators action given in fig. 7.8(a), while the lateral
actuation is in fig. 7.8(b). The vectoring angle v , which is expected to have
a negligible action in aerodynamic flight, is clearly involved in the altitude
error and pitch control. The differential thrust TD , represented here for
completeness, has a negligible control action in aerodynamic flight, an option
taken when converting from forces to actuators.
Table 7.1 shows the RMS values of selected variables. They are the airship
positions errors, namely longitudinal , lateral and vertical errors, and the
true airspeed Vt , the angle of attack and the sideslip angle , together with
the groundspeed error eu relative to the 8m/s reference. The first row has

CHAPTER 7. BACKSTEPPING

20

20

10

10

a (deg)

e (deg)

144

0
-10
-20

100

50

0
-10
-20

150

50

80

20

60

10

40
20
0
0

50

150

100

100

150

100

150

0
-10
-20

150

50
Time (s)

Time (s)
120
90
60
30
0
-30

1
TD (N )

v (deg)

100
Time (s)

r (deg)

XT (N )

Time (s)

100

50

150

0
-1

50

Time (s)

(a) Longitudinal actuators input: elevator


e , total thrust XT and vectoring v .

Time (s)

(b) Lateral actuators input: aileron a ,


rudder r and differential thrust TD .

Figure 7.8: Airship actuators input.


the RMS values obtained for the baseline case, and is to serve as reference
for the remaining lines where each of the listed coefficients is varied one at
a time. For a parameter uncertainty of 70%, the controlled airship still
performed qualitatively like the baseline case. We then increased the parameter
uncertainty, and present here the results obtained for 90% uncertainty around
the nominal value.
Only for four parameters the uncertainty leads to an less efficient or inefficient
control action, either because it is insufficient due to actuators saturation (the
backstepping demands a too high control input) and/or just because the tested
coefficient appears to be a more sensitive model parameter. The cases which
lead to significant deviations from the reference RMS values are represented
in bold in table 7.1.
With noticeable deviations from the baseline RMS values, for CNr (90%
uncertainty) the backstepping still controls the AURORA airship within acceptable bounds. CNr corresponds to the authority of the rudder control
surface as yawing control input.
For the remaining coefficients, CMe , CL and CMq (all for 90% uncertainty),
either saturation of the actuators occurred or the uncertainty of the model
parameter is too important for the backstepping controller to overcome it.
These parameters correspond respectively to the authority of the elevator as
pitching control input, the lift coefficient derivative due to angle of attack
and the pitch damping derivative. While the backstepping demonstrates to

145

7.6. SIMULATION RESULTS

Table 7.1: Robustness tests on model parameters (RMS values of selected variables).
Baseline

(m)

(m)

(m)

Vt (m/s)

(deg)

(deg)

eu (m/s)

0.40

0.23

0.29

9.00

2.87

1.13

0.12

Cl

90%
+90%

0.37
0.45

0.22
0.24

0.29
0.28

9.00
9.00

2.91
2.78

0.99
1.33

0.11
0.13

CM0

90%
+90%

0.40
0.39

0.23
0.24

0.27
0.31

9.00
9.00

2.90
2.82

1.13
1.15

0.12
0.12

CM

90%
+90%

0.34
0.36

0.20
0.25

0.12
0.48

9.00
9.00

3.19
2.53

1.08
1.18

0.11
0.17

CM

90%
+90%

0.39
0.40

0.23
0.22

0.29
0.29

9.00
9.00

2.85
2.87

1.13
1.12

0.12
0.12

CM

90%
+90%

0.39
0.40

0.24
0.22

0.31
0.27

9.00
9.00

2.82
2.90

1.14
1.12

0.12
0.12

CM

90%
+90%

0.39
0.40

0.23
0.23

0.29
0.29

9.00
9.00

2.86
2.87

1.13
1.13

0.12
0.12

CMe

90%
+90%

1.51
0.38

6.16
0.20

27.06
0.15

8.99
9.00

2.18
3.14

1.56
1.08

0.29
0.10

CN

90%
+90%

0.37
0.40

0.23
0.27

0.28
0.30

9.00
9.00

2.89
2.81

1.17
1.19

0.11
0.13

CNr

90%
+90%

3.56
0.38

6.78
0.08

3.16
0.27

9.16
9.00

4.16
2.91

3.88
1.01

0.77
0.09

CY

90%
+90%

0.39
0.39

0.60
0.21

0.34
0.29

9.01
9.00

2.83
2.87

3.13
0.87

0.20
0.11

CYr

90%
+90%

0.38
0.59

0.09
1.04

0.27
0.67

9.00
9.03

2.91
2.63

1.13
3.76

0.09
0.39

CD0

90%
+90%

1.36
0.86

0.43
0.22

0.52
0.34

9.02
9.00

3.18
2.68

1.54
1.13

0.33
0.13

CDi

90%
+90%

0.36
0.43

0.23
0.23

0.28
0.30

9.00
9.00

2.90
2.83

1.13
1.12

0.12
0.12

CL0

90%
+90%

0.41
0.38

0.23
0.23

0.30
0.28

9.00
9.00

2.76
2.97

1.13
1.13

0.12
0.12

CL

90%
+90%

49.86
0.47

0.74
0.27

10.84
0.44

9.72
9.00

19.32
1.52

1.52
1.18

1.06
0.13

CLe

90%
+90%

0.39
0.42

0.19
0.29

0.10
0.54

9.00
9.00

4.18
2.10

1.08
1.22

0.11
0.15

Clp

90%
+90%

0.42
0.38

0.23
0.24

0.29
0.29

9.00
9.00

2.76
2.87

1.33
1.10

0.13
0.12

CMq

90%
+90%

265.78
0.41

3.12
0.30

1.52
0.75

12.79
9.01

3.41
2.74

1.33
1.21

4.65
0.20

CNr

90%
+90%

0.38
0.79

0.07
1.52

0.28
0.90

9.00
9.05

2.88
2.91

0.98
2.84

0.10
0.38

mw

90%
+90%

0.46
0.23

0.37
0.30

0.95
0.32

9.02
9.01

1.99
3.01

1.16
1.31

0.25
0.13

146

CHAPTER 7. BACKSTEPPING

be robust to a 90% uncertainty in the remaining parameters, for these three


cases, the mismatch between the airship system and the model considered in
the backstepping controller design is too significant for the control action to
overcome it.
In any case, the backstepping controller may be considered robust to wind
disturbances and plant uncertainties. Among the list selected, these four parameters CNr , CMe , CL and CMq (and specially the last three) are in fact
the model parameters for which a more careful identification or determination
should take place, though the required precision could merely remain inside a
70% margin.

7.7

Conclusions

This chapter introduces a backstepping approach for the airship path-tracking


problem. The asymptotically stable backstepping controller is designed formulating a scalar positive function of the system states and then choosing
a control law based on the airship six-degrees-of-freedom nonlinear model to
make this function decrease.
Some practical issues have to be addressed, and the control law is improved
to take into account input saturations and wind disturbances, maintaining its
asymptotic stability for a bounded wind estimation error. Prior to implementation, further issues are considered, namely control allocation and reference
shaping to deal with the airship underactuation. Reference shaping is vital for
the control law implementation on a underactuated airship. Remember the
control law considers six forces inputs are fully available, being blind to the
control allocation problem (changing from forces to actuators request), and
therefore to the available actuation. This is a step that, for now, is executed
after the controller design.
The application of the proposed backstepping solution to the AURORA airship
path-tracking problem resulted in a satisfactory performance in the execution
of missions including different phases as take-off and landing, path-tracking
and stabilization, even in the presence of realistic wind disturbances.
The backstepping approach is based on the six-degrees-of-freedom nonlinear
airship model with constant translation wind input. However, a real wind disturbance is stochastic and the real airship parameters might differ from the

7.7. CONCLUSIONS

147

ones of the model. Therefore, the analysis of the controlled system performance and stability robustness in the presence of realistic wind disturbances
and model parameter uncertainties is very important. The backstepping controller, robust to wind disturbances, shows to be tolerant to uncertainties in
most of the model parameters tested. However, for some aerodynamic coefficients, namely CNr , CMe , CL and CMq , a more careful identification or
determination should take place.

148

CHAPTER 7. BACKSTEPPING

Chapter 8
Comparison of controllers
performance
Any of the three control solutions described in the previous chapters presents
its advantages and disadvantages, many of which are discussed in the respective
chapters. Yet, an overall comparison between them is important as to provide a
better overview of the different control options. In this chapter this assessment
is made considering parameters such as path-tracking performance for a casestudy complete mission (Section 8.1), robustness in face of model parameter
uncertainty (Section 8.2) and computational effort (Section 8.3). These factors,
together with some implementation issues, are relevant to evolve to the next
phase, the experimental validation in autonomous flight.

8.1

Performance for case-study mission

The case-study mission described in Section 3.3.1 was used to test the pathtracking performance of each of the controllers. It considers important phases
of a generic mission, like take-off and landing, path-tracking and stabilization.
The reference trajectory considered, as well as the resulting trajectories for
each of the three controllers are gathered in fig. 8.1. Although all three controllers are able to accomplish the entire mission, the different performances
are noticeable. The gain scheduling and backstepping solutions both show
smaller deviations from the reference trajectory, with the backstepping controller presenting more difficulties in the vertical positioning control during the
transition from vertical ascent to horizontal tracking and during the stabiliza149

150 CHAPTER 8. COMPARISON OF CONTROLLERS PERFORMANCE

60
60

50

50

40

h (m)

h (m)

1
3

2

40
30

30
20

20
10
10
0

200

0

4

200
100

100

300
100

-100
-200
N (m)

100

-100

-200

-100

N (m)

E (m)

0
-100
E (m)

(b) Gain scheduling.

(a) Reference.

60

60

50

50

40

40
h (m)

h (m)

300
200

200

30

30

20

20

10

10

200

200
100

300
200

300

-200

200

100

-100
N (m)

100
100

-100

-200

-100
E (m)

(c) Dynamic inversion.

N (m)

0
-100
E (m)

(d) Backstepping.

Figure 8.1: Comparison of airship 3D trajectories.

tion. Notice that these two points represent nonsmooth tracking references.
The dynamic inversion visibly results in a more erroneous tracking.
In order to better quantify the position errors obtained with each of the controllers, the longitudinal , lateral and vertical local errors are represented
in fig. 8.2, together with an indication of the mission phase being executed.
Observing with this detail, it is obvious that none of the controllers is better
(or worse) during the entire mission. Dynamic inversion, for instance, although
usually having the higher deviations presents a smoother longitudinal stabilization.
Figure 8.3 allows us to compare the horizontal path-tracking results. Both
gain scheduling and dynamic inversion solutions lead to a higher crabbing of

151

(m)

(m)

8.1. PERFORMANCE FOR CASE-STUDY MISSION

20
0
-20
-40
-60
-80

ascent

horizontal path-tracking
descent
stabilization

50

100

150

200
Time (s)

250

300

350

400

20
10
0
-10
-20
-30
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

(m)

5
0
-5
-10

Figure 8.2: Comparison of position errors ( gain scheduling, dynamic


inversion, . backstepping).

the airship during the tail wind periods. The backstepping control results in
a smoother overall trajectory, consequence of the references shaping to deal
with the airship underactuation.

250

250
wind
heading

200

250
wind
heading

200

150

150

100

100

100

50

50

50

N (m)

150

N (m)

N (m)

200

-50

-50

-50

-100

-100

-100

-150

-150

-150

-200

-200

-200

-250
-100

-250
-100

100
E (m)

200

(a) Gain scheduling.

300

100
E (m)

200

(b) Dynamic inversion.

300

wind
heading

-250
-100

100
E (m)

200

(c) Backstepping.

Figure 8.3: Comparison of north-east trajectories with airship heading.

300

152 CHAPTER 8. COMPARISON OF CONTROLLERS PERFORMANCE

The UAV capabilities report [13] analyzes 53 proposed missions in the Earth
observation scope, and 16 capabilities required. One of the requirements refers
to the precision of trajectories. They define four levels of accuracy, namely
level 5, where the trajectory is to be based on a position accuracy better than
5m; level 3, that requires a position accuracy between 5m and 50m; level
1, where the mission requires some sensitivity to vehicle trajectory, absolute
or relative, but position accuracy can be less than 50m; and level 0 for
missions that do not involve a precision trajectory. Observing these limits, and
considering only the continuous path-tracking part of the mission (neglecting
the second stabilization, at the end of the second curve), we observe that the
three controllers, with position errors below 30m, respect the level 3 limits.
The control solutions presented here are therefore appropriate, what trajectory
precision concerns, for missions such as topographic mapping and topographic
change, river discharge and urban management.
For long endurance applications, energy management is an important autonomy issue. A lower fuel and batteries consumption requires a reduced control
effort. Figure 8.4 allows us to compare the actuators request made by each
of the three controllers during the execution of the case-study mission. The
dynamic inversion and backstepping controllers show higher and more oscillatory requests, denoting the nonlinearity of the control laws and of the control
allocation procedure. The smooth gain scheduling control effort is a direct
consequence of the linearization procedure, resulting in a linear airship model
with actuators input instead of forces.

8.2

Sensitivity test results comparison

The sensitivity and robustness to parameter uncertainty test described in Section 3.3.2 investigates the stability of the closed-loop systems even in the
present of wind disturbances and model parameter uncertainty.
The baseline simulation considered the nominal model of the airship subject to
wind disturbances. The baseline results obtained for the particular mission of
a north-aligned straight line tracking with lateral wind from the three control
solutions are gathered in table 8.1. The dynamic inversion controller leads to
the higher position errors, while the backstepping solution results in smaller
ground velocity and lateral and vertical position errors.
Regarding the robustness to model parameter uncertainty, both gain schedul-

153

30
15

0
-15
-30

XT (N )

a (deg)

30
15

100

150

200
Time (s)

250

300

350

0
-15
-30

400

50

100

150

200
Time (s)

250

300

350

400

-30
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

80

30

60

15

40
20
0
0

50

100

150

200
Time (s)

250

300

350

400

0
-30

0
-15

20

120
90
60
30

TD (N )

v (deg)

50

r (deg)

e (deg)

8.2. SENSITIVITY TEST RESULTS

50

100

150

200
Time (s)

250

300

350

10
0
-10
-20

400

30
15

0
-15
-30

XT (N )

a (deg)

30
15

100

150

200
Time (s)

250

300

350

0
-15
-30

400

50

100

150

200
Time (s)

250

300

350

400

-30
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

80

30

60

15

40
20
0
0

50

100

150

200
Time (s)

250

300

350

400

120
90
60
30
0
-30

0
-15

-10
TD (N )

v (deg)

50

r (deg)

e (deg)

(a) Gain scheduling actuators request.

10
0

50

100

150

200
Time (s)

250

300

350

400

30
15

0
-15
-30

XT (N )

a (deg)

30
15

100

150

200
Time (s)

250

300

350

0
-15
-30

400

50

100

150

200
Time (s)

250

300

350

400

-30
0

50

100

150

200
Time (s)

250

300

350

400

50

100

150

200
Time (s)

250

300

350

400

80

30

60

15

40
20
0
0

50

100

150

200
Time (s)

250

300

350

400

0
-15

30

120
90
60
30
0
-30

TD (N )

v (deg)

50

r (deg)

e (deg)

(b) Dynamic inversion actuators request.

50

100

150

200
Time (s)

250

300

350

400

15
0
-15
-30

(c) Backstepping actuators request.

Figure 8.4: Comparison of actuators request (elevator e , total thrust XT ,


vectoring v , aileron a , rudder r and differential thrust TD ).

154 CHAPTER 8. COMPARISON OF CONTROLLERS PERFORMANCE

Table 8.1: Comparison of baseline results (RMS values).


(m)

(m)

(m)

eu (m/s)

Gain Scheduling

0.24

0.33

0.45

0.16

Dynamic Inversion

0.60

0.45

0.78

0.14

Backstepping

0.40

0.23

0.29

0.12

ing and backstepping controllers demonstrated robustness to a 70% uncertainty in all analyzed parameters (changing one at a time), while the dynamic
inversion presented control problems for lower values of uncertainties for some
aerodynamic coefficients. For uncertainties up to 90%, the aerodynamic coefficients that resulted in an inefficient control, namely CL , CD0 , CMe , CNr ,
CMq and CNr , are the ones for which a more careful identification or determination should take place.

8.3

Computational effort

For a real-time implementation to be possible, the computational time taken


by the controller is an important measure of its performance. The controller
is to be implemented onboard the airship platform (as was in the simulator)
at 10Hz.
Table 8.2 represents the computational time taken by each of the three controllers tested. The computational time obviously depends on the characteristics of the machine. Therefore, not only absolute time is presented but also
the relative time between controllers, based on the higher computational time.
This measure provides a better comparison of the computational effort.
The approximate values obtained for the gain scheduling and dynamic inversion controllers is justified since the dynamic inversion execution code also runs
the gain scheduling in order to obtain the model reference used. Although with
a more complex control law, the time taken to execute the backstepping controller code is almost 50% less than the time taken by the gain scheduling
controller. This may be justified by the fact that the gain matrix K used in
the gain scheduling is being computed online rather than being obtained from
a lookup table. The computational effort is obviously also a function of the
code optimization.

155

8.4. CONCLUSIONS

Table 8.2: Computational effort comparison.

8.4

Total time1 (s)

Relative time (%)

Gain Scheduling

0.0475

95.4

Dynamic Inversion

0.0498

100.0

Backstepping

0.0271

54.4

Conclusions

This chapter provides an overview of the advantages and disadvantages of


the three control solutions considered in this work, namely gain scheduling,
dynamic inversion and backstepping, when applied to the path-tracking airship
problem.
In the previous sections we compared the results obtained for each controller
regarding a case-study mission, a sensitivity and robustness test for model parameter uncertainty, and the computational effort. In order to better visualize
the relative results, table 8.3 presents a qualitative overall comparison between
controllers.
Table 8.3: Overall controllers comparison (+ good,  average, poor).
Gain Scheduling

Dynamic Inversion

Backstepping

Path-tracking performance:
Path-tracking errors
Tracking smoothness
Requested control effort

+

+

Robustness to:
Wind disturbances
Parameters uncertainty

+
+

+
+

+


+
+

Implementation issues:
Computational effort
Code simplicity


+




+



+

Others:
Design parameters tuning
Possible evolution

The table is divided into four parts, three of which, path-tracking performance,
controllers robustness and implementation issues, were already analyzed in the
previous sections. The last one contains more subjective, designer experience
1

Computational time measured in a Pentium IV with 512M B at 2.8GHz, running


MATLABr R2006a.

156 CHAPTER 8. COMPARISON OF CONTROLLERS PERFORMANCE

related, but also relevant issues, namely design parameters tuning and possible
evolution.
The evaluation of the design parameters tuning provides a comparative idea of
the necessary effort of the designer to correctly tune the controllers parameters.
While the dynamic inversion, using a model reference, has the decision of which
model to follow, the gain scheduling requires the adjustment of the state and
input control matrices parameters. The backstepping performance depends of
a proper choice of the reference shaping parameters.
The last item refers to the possible evolution of each solution. This evaluation
is merely based on the knowledge acquired throughout this work, and serves
as an indication of the future work yet to be developed for each of the control
solutions (see next chapter).

Chapter 9
Conclusions and Future Work
Considering their particular features, airships have a wide spectrum of applications as observation and data acquisition platforms. If we also consider
the quest for autonomy, airships present characteristics and competitive costs
when compared to other aircrafts, certainly constituting an important option
for research, development and experimental validation in autonomous aerial
robotics. Moreover, most of the solutions established for this kind of air vehicle may be transferred or adapted for airplanes or helicopters, where the risks
and costs involved in testing new methodologies are obviously higher.
The role of airships as UAVs depends, however, of their autonomous flight
capacity. This implies the development of control solutions for the airship
autonomous flight, that allow the execution of different missions even in the
presence of wind disturbances.
So far, a global control solution as not yet been presented for airships, with the
exception of the LAAS-CNRS group solution with decoupled controllers [40,
47], for complete missions including take-off and landing, path-tracking and
stabilization. Moreover, seldom are the ones that consider such an important
issue as robustness to wind disturbances. This work, inserted in the AURORA
and DIVA projects, made a breakthrough in this topic, developing and comparing airship control solutions, valid for the entire flight envelope, and capable
of executing realistic missions, while being robust to wind input.
An airship is an highly nonlinear system. The dynamics when in hover or aerodynamic flight varies greatly, with different combinations of actuators available,
which leads to a problematic transition region between the two. The successful
development of an overall control solution depends therefore on a good knowl157

158

CHAPTER 9. CONCLUSIONS AND FUTURE WORK

edge of the system behavior over the flight envelope, and on a good model of
the airship. With this in mind, a six-degrees-of-freedom nonlinear model was
developed based on the Lagrangian approach. The linearization of this model
over the aerodynamic range lead to the usual decoupling of the longitudinal
and lateral motions. A detailed analysis of these linear models provided the
necessary insight of the airship behavior characteristics, allowing the design of
the first control solution.
The Gain Scheduling approach is based on the linear description of the airship. In order for the solution to be valid over the entire flight envelope, for
each linear model obtained (one for each equilibrium condition defined), an
optimal state feedback control law is designed. The overall control synthesis
is achieved by switching between models and respective controllers as function
of the scheduling variable airspeed. The main advantages of this method are
the simplicity of both linear model and controller, allowing to use the classic
control tools, and the fact that the model inputs are the airship actuators, a
result of the linearization procedure. A disadvantage is the time consuming
tuning of the control design parameters, namely the state and input control
matrices.
The Dynamic Inversion solution results of the inversion of the six-degrees-offreedom nonlinear airship model, obtaining a control law that cancels existing
deficient or undesirable dynamics by replacing them with a set of desired ones.
For systems, like the airship, described in a cascaded form based on dynamics
and kinematics, a new dynamic inversion formulation is presented, allowing an
easier implementation. However, a control solution based on the nonlinear description of the system presents a disadvantage if the system is underactuated,
as is the airship. The nonlinear model considers forces as input. Therefore,
the controller, obtained by inversion of this model, computes a forces request
considering all six forces are fully available. This is however not the case.
The airship has serious actuation constraints regarding the lateral force and,
although not so severe, with the downward force as well. If this information is
not provided a priori to the controller, the resulting forces request will demonstrate to be inadequate. A first solution to this problem was found by providing
the gain scheduling closed-loop dynamics as reference to the dynamic inversion
controller, instead of the reference trajectory dynamics, since the linear model
provides indirect information on the airship actuation limits. This solution,
however, limits the performance of the dynamic inversion controller to that of
the gain scheduling provided as model.

159

The Backstepping controller is designed formulating a scalar positive function


of the system states and then choosing a control law based on the airship sixdegrees-of-freedom nonlinear model to make this function decrease, therefore
guarantying the asymptotic stability of the controller. Some practical issues
are addressed, and the control law is improved to take into account input
saturations and wind disturbances, maintaining its asymptotic stability for a
bounded wind estimation error. The control law implementation again raises
the problem of the airship underactuation. This time, the solution found in
the dynamic inversion case is not applicable. The answer to the problem was
then to provide the controller with shaped attitude references. The idea is to
delay the attitude rectification in benefit of the correction of the transversal
lateral and vertical errors.
All three control solution, Gain Scheduling, Dynamic Inversion and Backstepping, proved to be capable of executing complete missions considering take-off
and landing, path-tracking and stabilization, in the presence of realistic wind
disturbances. An assessment of the advantages and disadvantages of each controller, as well as a comparison between them, was also made, providing an
overall insight of the autonomous airship control problem and of the solutions
proposed.
These solutions, having already demonstrated their value, have still issues to
evolve. Variations of the proposed solutions, namely the ones based on the nonlinear airship model, are already under development. One is the evaluation of
the reference shaping solution used in the Backstepping as an alternative to using the Gain Scheduling as dynamic model in the Dynamic Inversion solution.
Other evolution pertains the control allocation from forces to actuators, which
is not a straightforward procedure since the relation between forces and actuators is not invertible, requiring sometimes an empiric solution. Moreover, the
allocation process is, till now, executed after the control request computation,
leaving the controller blind to the actuators limitations, reason for which an
attitude references shaping is necessary. The inclusion of an improved control
allocation, even keeping the references shaping to minimize the airship underactuation effects, into the controller design is a key factor in obtaining a more
satisfying overall nonlinear control solution.

160

CHAPTER 9. CONCLUSIONS AND FUTURE WORK

Appendix A
Referentials
Contents
A.1 Frames definition . . . . . . . . . . . . . . . . . . . .

161

A.1.1 Earth-Centered Inertial (ECI) frame . . . . . . . . . 161


A.1.2 North-East-Down (NED) or {i} frame . . . . . . . . 162
A.1.3 Aircraft-Body Centered (ABC) or {l} frame . . . . . 162

A.1.4 Aerodynamic or {a} frame . . . . . . . . . . . . . . 163

A.2 Changing frame . . . . . . . . . . . . . . . . . . . .

163

In this chapter we define the referentials used in the description of the airship
dynamics and kinematics, and how they are related.

A.1

Frames definition

To describe the dynamics of an airship, one needs to set up a coordinate frame.


Different coordinate frames may be used to describe the airship motion. The
following summarizes some of the coordinate frames [1, 53] used while modeling
the airship motion. Figure A.1 shows the spatial relationship between the
coordinate systems.

A.1.1

Earth-Centered Inertial (ECI) frame

The ECI frame is centered at the origin of the Earth. The z-axis coincides
with the Earths spin axis, pointing to the north Pole and the x-axis points in
the direction of the vernal equinox (the vernal equinox is an imaginary point
161

162

APPENDIX A. REFERENTIALS

y
E
x

ABC frame
z

xl

NED frame
z

ECI frame
Equator

Figure A.1: Relationship between the different coordinate systems.


in space which lies along the line representing the intersection of the Earths
equatorial plane and the plane of the Earths orbit around the Sun or the
ecliptic). Finally, to complete an orthogonal right handed system, the y-axis
is perpendicular to the xz-plane.

A.1.2

North-East-Down (NED) or {i} frame

The NED frame is centered on the Earths surface at the point vertically below
the airships Center of Gravity (CG), at its initial location, where it is fixed.
The xy-plane is tangent to the Earths surface. The x-axis points in the north
direction, the y-axis to the east and the z-axis is normal to the Earths surface,
pointing inward.
In this work the Earth will be assumed as flat and taken as an inertial frame.
It will be referenced as {i} frame.

A.1.3

Aircraft-Body Centered (ABC) or {l} frame

The ABC or local frame (referenced as {l} frame) is a right handed orthogonal
axis system fixed to the air vehicle. In order to accommodate the constantly
changing CG, the ABC frame is centered at the airships Center of Volume
(CV), assumed to be also the Center of Buoyancy (CB), and constrained to
move with it. The x-axis is coincident with the axis of symmetry of the envelope and the xz-plane coincides with the longitudinal plane of symmetry of
the airship (see fig. A.2). It is reasonable to assume both the CV and the CG

163

A.2. CHANGING FRAME

lie on the axis of symmetry of the envelope.

q,

yl

p,

r,
zl

xl

relative
air (Vt)
xa

Figure A.2: ABC and wind frames.

A.1.4

Aerodynamic or {a} frame

The Aerodynamic frame considers the relative aerodynamic incidence angles.


It is obtained from the {l} frame with two rotations (see fig. A.2):
1. rotation about the yl -axis for the angle of attack ,
2. rotation about the resulting z-axis for the side-slip angle .
This makes the x-axis coincide with the direction of the total relative air
velocity Vt . The angles and are known as the aerodynamic angles and are
needed to specify the aerodynamic forces and moments.
The complete transformation from the body {l} frame to the aerodynamic {a}
frame is then given by the Sa matrix, expressed as function of the aerodynamic
angles and , and given by:

cos cos sin sin cos

Sa = cos sin cos sin sin


sin
0
cos

A.2

(A.1)

Changing frame

The time derivative is defined in the inertial frame. The time derivative from
inertial {i} to local {l} frame introduces the Coriolis acceleration:
dv
dv
=
+ v = v + v
dt {i}
dt {l}

(A.2)

164

APPENDIX A. REFERENTIALS

Following the assumption of a rigid body, the linear velocity of the CG (point
C) is related to the linear velocity of the CV (O) through the angular velocity:
vc = v0 + OC = v OC

(A.3)

The transformation from the inertial reference to the local frame is achieved
by the following sequence of rotations:
1. rotation about the zi -axis (positive yaw angle);
2. rotation about the resulting y-axis (positive pitch angle);
3. rotation about the resulting x-axis (positive roll angle).
where the roll , pitch and yaw angles are commonly referred to as Euler
angles (see fig. A.2).
The complete transformation from the inertial {i} to the local {l} frame is
then given by the S matrix (often called Direction Cosine Matrix), expressed
as function of the Euler angles = [, , ]T and given by (2.17).

Appendix B
Dryden Model For Continuous
Gust
This Appendix mostly follows reference [54].
To generate the gust signals with the required intensity, scale lengths and
power spectral density (PSD) functions for some given velocity and height,
a white-noise source with a PSD function N () = 1 is used to provide the
input signal to a linear filter, chosen such that it has an appropriate frequency
response so that the output signal from the filter will have a PSD function
i (). The scheme is represented in the block diagram shown in fig. B.1.

linear filter
white noise N()
generator

Gi (s)

i ()

Figure B.1: Block diagram for gust generator.

The relation of the PSD function of the output signal to the PSD function of
the input signal is given by:
i () = |Gi (s)|2s=j N ()

(B.1)

The filters needed to generate the appropriate spectral densities for the trans165

166

APPENDIX B. DRYDEN MODEL FOR CONTINUOUS GUST

lational gust velocities are:

Ku
s + u
p
s + v
Gv (s) = Kv
(s + v )2
p
s + w
Gw (s) = Kw
(s + w )2
Gu (s) =

(B.2)
(B.3)
(B.4)

where
3Vt v2
3Vt w2
2Vt u2
, Kv =
, Kw =
Lu
Lv
Lw
Vt
Vt
v =
, w =
3Lv
3Lw

Ku =

u = Vt /Lu , v = Vt /Lv , w = Vt /Lw

(B.5)
(B.6)
(B.7)

The turbulence intensity reaches its maximum value of 7 m/s in a thunderstorm scenario. The turbulence scale length varies with height. The dependence of scale length on height is defined in this manner:
h > ho Lu = Lv = Lw = ho
p
h ho Lu = Lv = ho h, Lw = h

(B.8)
(B.9)

where ho = 533 m and h is the height of the airship encountering the turbulence.

Appendix C
Differential geometry and
topology
Contents
C.1 Lie derivatives . . . . . . . . . . . . . . . . . . . . .

168

C.2 Diffeomorphisms and state transformations . . . .

169

The purpose of this appendix is to introduce some mathematical tools from


differential geometry and topology, in the context of nonlinear dynamical systems [65, 64].
A vector function f : Rn Rn is called a vector field in Rn . Only smooth
vector fields shall be considered, which means the function f (x) has continuous
partial derivatives of any required order.
Given a smooth scalar function h(x) : Rn R of the state x, the gradient of
h(x) is defined as:
h
(C.1)
h(x) =
x
The gradient is represented by a row-vector of elements (h)j = h/xj .
Similarly, given a vector field f (x), the Jacobian of f is defined as:
f (x) =

f
x

and is represented by a n n matrix of elements (f )ij = fi /xj .


167

(C.2)

168

C.1

APPENDIX C. DIFFERENTIAL GEOMETRY AND TOPOLOGY

Lie derivatives

Given a scalar function h(x) and a vector field f (x), a new scalar function
Lf h(x) is defined, called the Lie derivative of h with respect to f .

Definition C.1 (Lie Derivative). Let h : Rn R be a smooth scalar function,


and f : Rn Rn be a smooth vector field on Rn . Then the Lie derivative of h
with respect to f is a scalar function defined by Lf h = h f .

Thus, the Lie derivative Lf h(x) is simply the directional derivative of h(x)
along the direction of the vector f (x).
If h is being differentiated k times along f , the notation Lkf h is used; in other
words, the function Lkf h satisfies the recursion
L0f h(x) = h(x)

(C.3)

Lkf h(x) = Lf (Lk1


h) = (Lk1
h) f
f
f

for k = 1, 2, ...

(C.4)

Similarly, if g(x) is another vector field, then the scalar function Lg Lf h(x) is
Lg Lf h(x) = (Lf h) g(x)

(C.5)

This simple example will show the relevance of Lie derivatives to dynamic
systems. Consider the following single-output system:
x = f (x)

(C.6)

y = h(x)

(C.7)

The derivatives of the output are


h
x = Lf h(x)
x
(Lf h)
x = L2f h(x)
y =
x
y =

and so on.

(C.8)
(C.9)

C.2. DIFFEOMORPHISMS AND STATE TRANSFORMATIONS

C.2

169

Diffeomorphisms and state transformations

The concept of diffeomorphism can be viewed as a generalization of the familiar


concept of coordinate transformation. It is formally defined as follows:
Definition C.2 (Diffeomorphism). A function : Rn Rn , defined in a
region , is called a diffeomorphism if it is smooth, and if its inverse 1
exists and is smooth.
If the region is the whole space Rn , then (x) is called a global diffeomorphism. Global diffeomorphisms are rare, and therefore one often looks for local
diffeomorphisms, i.e., for transformations defined only in a finite neighborhood
of a given point. Given a nonlinear function (x), it is easy to check whether
it is a local diffeomorphism by using the following lemma:
Lemma C.1. Let (x) be a smooth function defined in a region in Rn . If
the Jacobian matrix is nonsingular at a point x = x0 of , then (x)
defines a local diffeomorphism in a subregion of .
A diffeomorphism can be used to transform a nonlinear system into another
nonlinear system in terms of a new set of states, similarly to what is commonly
done in the analysis of linear systems. Consider the dynamic system described
by
x = f (x) + g(x)u

(C.10)

y = h(x)

(C.11)

and let a new set of states be defined by


z = (x) .

(C.12)

Differentiation of z yields
z =

(f (x) + g(x)) .
x =
x
x

(C.13)

One can easily write the new state-space representation as


z = f (z) + g (z)u

(C.14)

y = h (z)

(C.15)

170

APPENDIX C. DIFFERENTIAL GEOMETRY AND TOPOLOGY

where x = 1 (z) has been used, and the functions f , g and h are defined
obviously.

Bibliography
[1] Gabriel Alexander Khoury and John David Gillett. Airship Technology,
volume 10 of Cambridge Aerospace Series. Cambridge University Press,
1999.
[2] O. J. Netherclift. Airships today and tomorrow. Airship Association
Publication, (4), 1993. The Airship Association, Ltd.
[3] Anthony Colozza. Initial feasibility assessment of a high altitude long
endurance airship. Technical report, NASA - Dryden Flight Research
Center, December 2003.
[4] K. Eguchi, Y. Yokomaku, and M. Mori. Overview of stratospheric
platform airship R&D program in japan. In Proceedings of the AIAA
14th Lighter-Than-Air Technical Committee Convention and Exhibition,
Akron, USA, July 2001.
[5] Yung-Gyo Lee, Dong-Min Kim, and Chan-Hong Yeom. Development of
korean high altitude platform systems. International Journal of Wireless
Information Networks, 13(1):3142, January 2006.
[6] Chang-Hee Won. Regional navigation system using geosynchronous satellites and stratospheric airships. IEEE Transactions on Aerospace and
Electronic Systems, 38(1):271278, January 2002.
[7] St.D. Ilcev and A. Singh. Development of stratospheric communications
platforms (SCP) for rural applications. In Proceedings of the IEEE 7th
AFRICON Conference in Africa, volume 1, pages 233238, Gaborone,
Botswana, September 2004.
[8] W.J. Hurd, B.E. MacNeal, G.G. Ortiz, R.V. Moe, J.Z. Walker, M.L.
Dennis, E.S. Cheng, D.A. Fairbrother, B. Eegholm, and K.J. Kasunic.
171

172

BIBLIOGRAPHY

Exo-atmospheric telescopes for deep space optical communications. In


Proceedings of the IEEE Aerospace Conference, March 2006.
[9] Jinjun Rao, Zhenbang Gong, Jun Luo, and Shaorong Xie. Unmanned
airships for emergency managment. In Proceedings of the IEEE International Workshop on Safety, Security and Rescue Robotics, pages 125130,
Kobe, Japan, June 2005. IEEE Press.
[10] Alberto Elfes, Samuel S. Bueno, Marcel Bergerman, Ely C. de Paiva,
Josue G. Ramos JR., and Jose R. Azinheira. Robotic airships for exploration of planetary bodies with an atmosphere: Autonomy challenges.
Autonomous Robots, (14):147164, 2003.
[11] J.L Hall, V.V Kerzhanovich, A.H. Yavrouian, J.A. Jones, C.V. White,
and B.A.Dudik. An aerobot for global in situ exploration of Titan. Advances in Space Research, 37(11):21082119, 2006. The Next Generation
of Scientific Balloon Missions.
[12] Stephen A. Cambone,
Kenneth J. Krieg,
Peter Pace,
and
Linton
Wells
II.
Unmanned
aerial
vehicles
roadmap.
Technical
report,
US
Dept.
of
Defense,
http://www.fas.org/irp/program/collect/uav_roadmap2005.pdf,
2005.
[13] Timothy H. Cox et al. Earth observations and the role of UAVs - a
capabilities assessment. Technical report, NASA - Dryden Flight Research
Center, August 2006.
[14] Anibal Ollero and Luis Merino. Control and perception techniques for
aerial robotics. Annual Reviews in Control, 28(2):167178, 2004.
[15] Emilio Frazzoli, Munther A. Dahleh, and Eric Feron. Trajectory tracking control design for autonomous helicopters using a backstepping algorithm. In Proceedings of the American Control Conference, pages 4102
4107, Chicago, Illinois, USA, June 2000.
[16] Xinyan Deng, Luca Schenato, and Shankar Sastry. Hovering flight control of a micromechanical flying insect. In Proceedings of the 40th IEEE
Conference on Decision and Control, Orlando, Florida, USA, December
2001.

BIBLIOGRAPHY

173

[17] Lingzhong Guo, Chris Melhuish, and Quanmin Zhu. Towards neural adaptive hovering control of helicopters. In Proceedings of the IEEE International Conference on Control Applications, pages 5458, Glasgow, Scotland, UK, September 2002.
[18] Sahjendra N. Singh, Marc L. Steinberg, and Anthony B. Page. Nonlinear adaptive and sliding mode flight path control of F/A-18 model. IEEE
TRansactions on Aerospace and Electronic Systems, 39(4):12501262, October 2003.
[19] Ciann-Dong Yang and Wen-Hsiung Liu. Nonlinear h decoupling hover
control of helicopter with parameter uncertainties. In Procedings of the
American Control Conference, pages 34543459, Denver, Colorado, USA,
June 2003.
[20] Xin Chen and Changchun Pan. Application of h control and inverse
dynamic system in direct side force control of UAV. Journal of Nanjing
University of Aeronautics & Astronautics, 38(1):3336, February 2006.
[21] C. Patel and I. Kroo. Control law design for improving UAV performance
using wind turbulence. In Proceedings of the 44th AIAA Aerospace Sciences Meeting and Exhibit, Nevada, USA, January 2006.
[22] Dan Necsulescu, Yi-Wu Jiang, and Bumsoo Kim. Neural network based
feedback linearization control of an unmanned aerial vehicle. International
Journal of Automation and Computing, 4(1):7179, January 2007.
[23] Ely Carneiro de Paiva, Jose Raul Azinheira, Jr. Josue G. Ramos, Alexandra Moutinho, and Samuel Siqueira Bueno. Project AURORA: Infrastructure and flight control experiments for a robotic airship. Journal of
Field Robotics, 23(3/4):201222, March/April 2006.
[24] Dirk-A. Wimmer, Michael Bildstein, Klaus H. Well, Markus Schlenker,
Peter Kungl, and Bernd-H. Kroplin. Research airship Lotte: Development and operation controllers for autonomous flight phases. In Workshop
on Aerial Robotics, IEEE International Conference on Intelligent Robots
and Systems, pages 5568, Lausanne, Switzerland, October 2002.
[25] Emmanuel Hygounenc and Philippe Sou`eres. Lateral path-following GPSbased control of a small-size unmanned blimp. In Proceedings of the IEEE
International Conference on Robotics and Automation, volume 1, pages
540545, Taipei, Taiwan, September 2003. IEEE Press.

174

BIBLIOGRAPHY

[26] L. Beji and A. Abichou. Tracking control of trim trajectories of s blimp for
ascent and descent flight manoeuvres. International Journal of Control,
78(10):706 719, July 2005.
[27] G.A. Kantor, D. Wettergreen, J.P. Ostrowski, and S. Singh. Collection
of environmental data from and airship platform. In Proceedings of the
SPIE Conference on Sensor Fusion and Decentralized Control in Robotic
Systems IV, volume 4571, October 2001.
[28] Swee B. Tan and Bellur L. Nagabhushan. Robust heading-hold autopilot
for an advanced airship. In Proceedings of the 12th AIAA Lighter-ThanAir Technology Conference, July 1997.
[29] Ely C. de Paiva, Samuel S. Bueno, S. B. V. Gomes, J. J. G. Ramos,
and M. Bergerman. A control system development environment for AURORAs semi-autonomous robotic airship. In Proceedings of the IEEE
International Conference on Robotics and Automation, volume 3, pages
23282335, Detroit, USA, May 1999.
[30] J. Mueller and M. Paluszek. Development of an aerodynamic model and
control law design for a high altitude airship. In Proceedings of the AIAA
3rd Unmanned Unlimited Technical Conference, Workshop and Exhibit,
Chicago, USA, September 2004.
[31] A. Elfes, J. Montgomery, J. Hall, S. Joshi, J. Hall, J. Payne, and C. Bergh.
Autonomous Flight Control for a Planetary Exploration Aerobot. In
Proceedings of the 8th International Symposium on Artificial Intelligence,
Robotics and Automation in Space, volume 603 of ESA Special Publication, 2005.
[32] Guoqing Xia and Dan R. Corbett. Cooperative control systems of searching targets using unmanned blimps. In Proceedings of the 5th Worth
Congress on Intelligent Control and Automation, volume 2, pages 1179
1183, Hangzhou, P.R. China, June 2004. IEEE Press.
[33] Ely de Paiva, Fabio Benjovengo, and Samuel Bueno. Sliding mode control
for the path following of an unmanned airship.
[34] Sjoerd van der Zwaan, Matteo Perrone, Alexandre Bernardino, and Jose
Santos-Victor. Control of an aerial blimp based on visual input. In Proceedings of the 8th International Symposium on Intelligent Robotic Systems, Reading, UK, July 2000.

BIBLIOGRAPHY

175

[35] Jose R. Azinheira, Patrick Rives, Jose R. H. Carvalho, Geraldo F. Silveira, Ely C. de Paiva, and Samuel S. Bueno. Visual servo control for
the hovering of an outdoor robotic airship. In Proceedings of the IEEE
International Conference on Robotics & Automation, pages 27872792,
Washington, DC, USA, May 2002.
[36] Geraldo F. Silveira et al. Optimal visual servoed guidance of outdoor
autonomous robotic airships. In Proceedings of the American Control
Conference, pages 779784, Anchorage, May 2002.
[37] Geraldo F. Silveira, Jose R. Azinheira, Patrick Rives, and Samuel S.
Bueno. Line following visual servoing for aerial robots combined with
complementary sensors. In Proceedings of the 11th International Conference on Advanced Robotics, Coimbra, Portugal, June 2003.
[38] Jinjun Rao, Zhenbang Gong, Jun Luo, and Shaorong Xie. A flight control
and navigation system of a small size unmanned airship. In Proceedings
of the IEEE International Conference on Mechatronics & Automation,
pages 14911496, Niagara Falls, Canada, July 2005. IEEE Press.
[39] Emmanuel Hygounenc and Philippe Soueres. Automatic airship control involving backstepping techniques. In Proceedings of the IEEE International Conference on Systems, Man and Cybernetics, Hammamet,
Tunisia, October 2002.
[40] Emmanuel Hygounenc, Il-Kyun Jung, Philippe Soures, and Simon
Lacroix. The autonomous blimp project of LAAS-CNRS: Achievements in
flight control and terrain mapping. The International Journal of Robotics
Research, 23(4-5):473511, AprilMay 2004.
[41] Y. Bestaoui and S. Hima. Trajectory tracking of a dirigible in a high
constant altitude flight. In Proceedings of the 5th IFAC symposium on
Nonlinear control systems, Saint Petersburg, Russia, July 2001.
[42] Chang-Su Park, Hyunjae Lee, Min-Jea Tahk, and Hyochoong Bang. Airship control using neural network augmented model inversion. In Proceedings of the IEEE Conference on Control Applications, pages 558563,
June 2003.
[43] Takanori Fukao, Kazushi Fujitaniy, and Takeo Kanade. Image-based
tracking control of a blimp. In Proceedings of the 42nd IEEE Confer-

176

BIBLIOGRAPHY

ence on Decision and Control, pages 34623467, Hawaii, USA, December


2003.
[44] J. Hacker and B.-H. Kroplin. An experimental study of visual flight trajectory tracking and pose prediction for the automatic computer control
of a miniature airship. In Proceedings of the SPIE International Society
for Optical Engineering, pages 2536, 2003.
[45] Yasunori Kawai, Satoshi Kitagawa, Shintaro Izoe, and Masayuki Fujita.
An unmanned planar blimp on visualfeedback control: Experimental results. In Proceedings of the 42nd SICE Annual Conference, pages 680685,
2003.
[46] Leonardo Guzman. Modelado, control y navegacion para el vuelo
autonomo de dirigibles. PhD thesis, Universidad de los Andes and Institut National des Sciences Appliquees de Toulouse, 2007.
[47] Emmanuel Hygounenc. Mod`elisation et commande dun dirigeable pour le
vol autonome. PhD thesis, Laboratoire dAnalyse et dArchitecture des
Syst`emes du CNRS, Universite Paul Sabatier de Toulouse, 2003.
[48] Jose Raul Azinheira, Ely Carneiro de Paiva, and Samuel Siqueira Bueno.
Influence of wind speed on airship dynamics. Journal of Guidance, Control, and Dynamics, 25(6):11161124, November-December 2002.
[49] H. Lamb. The inertia coefficients of an ellipsoid moving in fluid. Repts.
and Memoranda 623, Aeronautic Research Committee, October 1918.
[50] P. G. Thomasson. Equations of motion of a vehicle in a moving fluid.
Journal of Aircraft, 37(4):630639, July 2000.
[51] Mark W. Spong and M. Vidyasagar. Robot Dynamics and Control. John
Wiley & Sons, 1989.
[52] P. G. Thomasson. On calculating the motion of a vehicle in a moving
fluid. In Proceedings of the 3rd International Conference on Nonlinear
Problems in Aviation and Aerospace, Daytona Beach, Florida, USA, May
2000.
[53] Brian L. Stevens and Frank L. Lewis. Aircraft Control and Simulation.
John Wiley and Sons, Inc., USA, 1992.
[54] Donald McLean. Automatic Flight Control Systems. Prentice Hall, 1990.

177

BIBLIOGRAPHY

[55] Frank L. Lewis and Vassilis L. Syrmos.


Interscience, second edition, 1995.

Optimal Control.

Wiley-

[56] Brian D. O. Anderson and John B. Moore. Optimal Control: Linear


Quadratic Methods. Prentice-Hall, 1989.
[57] Robert C. Nelson. Flight Stability and Automatic Control. Aerospace
Science & Technology Series. McGraw-Hill, second edition, 1997.
[58] Jose Raul Azinheira, Ely C. de Paiva, Josue J. G. Ramos, Alexandra
Moutinho, and Samuel S. Bueno. Estrategias de controle lateral para um
dirigvel robotico autonomo. In Proceedings of the Congresso Brasileiro
de Autom
atica, Gramado, RS, Brazil, September 2004.
[59] Wilson J. Rugh and Jeff S. Shamma. Research on gain scheduling - survey
paper. Automatica, 36(10):14011425, October 2000.
[60] Hassan K. Khalil. Nonlinear Systems. Prentice-Hall, third edition, 2000.
[61] Gary Balas, Richard Chiang, Andy Packard, and Michael Safonov. Robust
Control Toolbox - For Use With Matlabr . MathWorks, third edition, 2005.
[62] Sigurd Skogestad and Ian Postlethwaite. Multivariable Feedback Control
- Analysis and Design. John Wiley & Sons, 1996.
[63] Andrew R. Teel. Global stabilization and restricted tracking for multiple
integrators with bounded controls. Systems & Control Letters, 18(3):165
171, March 1992.
[64] Alberto Isidori. Nonlinear Control Systems. Springer-Verlag, 2nd edition,
1989.
[65] Jean-Jacques E. Slotine and Weiping Li.
Prentice-Hall, 1991.

Applied Nonlinear Control.

[66] Dale Enns, Dan Bugajski, Russ Hendrick, and Gunter Stein. Dynamic
inversion: an evolving methodology for flight control design. International
Journal of Control, 59(1):7191, January 1994.
[67] Daigoro Ito, Donald T. Ward, and John Valasek. Robust dynamic inversion controller design and analysis for the X-38. In Proceedings of the 2001
AIAA Guidance, Navigation, and Control Conference, Montreal, Canada,
August 2001. AIAA-2001-4380.

178

BIBLIOGRAPHY

[68] Jacob Reiner, Gary J. Balas, and William L. Garrard. Robust dynamic
inversion for control of highly maneuverable aircaft. Journal of Guidance,
Control and Dynamics, 18(1):1824, January-February 1995.
[69] Samir Bennani and Gertjan Looye. Flight control law design for a civil
aircraft using robust dynamic inversion. In Proceedings of the 2nd IMACSIEEE/SMC International Multiconference on Computational Engineering
in Systems Applications, Nabeul-Hammamet,Tunisia, April 1998.
[70] Rama K. Yedavalli, Praveen Shankar, and David B. Doman. Robustness
study of a dynamic inversion based indirect adaptive control system for
flight vehicles under uncertain model data. In Proceedings of American
Control Conference, pages 10051010, Denver, USA, June 2003.
[71] Alexandra Moutinho and Jose Raul Azinheira. Path control of an autonomous airship using dynamic inversion. In Proceedings of the 5th
IFAC/EURON Symposium on Intelligent Autonomous Vehicles, Lisbon,
Portugal, July 2004.
[72] Alexandra Moutinho and Jose Raul Azinheira. Hover stabilization of an
airship using dynamic inversion. In Proceedings of the 8th International
IFAC Symposium on Robot Control, Bologna, Italy, September 2006.
[73] R. Mahony, T. Hamel, and A. Dzul. Hover control via Lyapunov control
for an autonomous model helicopter. In Proceedings of the 38th Conference
on Decision & Control, Phoenix, Arizona, USA, December 1999.
[74] Ki-Seok Kim and Youdan Kim. Robust backstepping control for slew maneuver using nonlinear tracking function. IEEE Transactions on Control
Systems Technology, 11(6):822829, November 2003.
[75] Najib Metni, Tarek Hamel, and Franois Derkx. A UAV for bridges inspection: Visual servoing control law with orientation limits. In Proceedings
of the 5th IFAC/EURON Symposium on Intelligent Autonomous Vehicles,
Lisbon, Portugal, July 2004.
[76] Jay Farrell, Manu Sharma, and Marios Polycarpou. Backstepping-based
flight control with adaptive function approximation. Journal of Guidance,
Control and Dynamics, 28(6):10891102, NovemberDecember 2005.

Você também pode gostar