Dynamic Positioning of Ships
A nonlinear control design study
Proefschrift
ter verkrijging van de graad van doctor
aan de Technische Universiteit Delft,
op gezag van de Rector Magniﬁcus prof. ir. K.C.A.M Luyben,
voorzitter van het College voor Promoties,
in het openbaar te verdedigen op xxxxxxxxxxxxxxxx
door
Shah MUHAMMAD,
Master of Science (M.Sc.) in Applied Mathematics
Islamia University Bahawalpur, Pakistan
geboren te Dhurnal, Talagang, Pakistan.
Dit proefschrift is goedgekeurd door de promotor:
Prof. dr. ir. A.W. Heemink
Copromotor:
Dr. J.W. van der Woude
Samenstelling promotiecommissie:
Rector Magniﬁcus voorzitter
Prof. dr. ir. A.W. Heemink Technische Universiteit Delft, promotor
Dr. J.W. van der Woude Technische Universiteit Delft, copromotor
Prof.dr.ir. Technische Universiteit Delft, NL
Prof.dr.ir. , NL
Prof.dr.ir. , NL
Prof. Dr. , NL
Prof. Dr. University, , , SP
This thesis has been completed in partial fulﬁllment of the requirements of Delft Uni
versity of Technology, Delft, The Netherlands, for the award of the PhD degree. The
research presented in this thesis was supported in parts by two institutions: Delft Uni
versity of Technology and HEC Pakistan. I thank them sincerely for their support.
Published and distributed by: Shah Muhammad
Email: abu.ash.hal@gmail.com
ISBN # 978909xxxxxx6
Keywords: Dynamic positioning, Nonlinear control design, portHamiltonian sys
tems.
Copyright © 2012 by Shah Muhammad
All rights reserved. No part of the material protected by this copyright notice may
be reproduced or utilized in any form or by any means, electronic or mechanical,
including photocopying, recording or by any information storage and retrieval system,
without written permission of the author.
Printed in The Netherlands
Contents
1 Introduction 1
1.1 What is a Dynamic Positioning (DP) System? . . . . . . . . . . . . . 1
1.1.1 Applications of DP Systems . . . . . . . . . . . . . . . . . . 2
1.1.2 Focus of this Research . . . . . . . . . . . . . . . . . . . . . 3
1.2 An Overview of this Thesis . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Contributions of this Thesis . . . . . . . . . . . . . . . . . . . . . . . 5
2 Mathematical Model of a Sea Vessel 7
2.1 Motion of an Aﬂoating Vessel . . . . . . . . . . . . . . . . . . . . . 7
2.2 Mathematical Model Describing the Dynamics of an Aﬂoating Vessel 9
2.2.1 The Dynamical Equations of Motion of the Vessel . . . . . . 10
2.2.2 The Disturbances Model . . . . . . . . . . . . . . . . . . . . 13
2.2.3 The Measurement Model . . . . . . . . . . . . . . . . . . . . 16
2.2.4 Wave Filtering . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Summary of the Mathematical Model . . . . . . . . . . . . . . . . . 16
2.4 Properties of the Model . . . . . . . . . . . . . . . . . . . . . . . . . 18
3 SDC Parametrization and Stability Analysis of Autonomous Nonlinear
Systems 21
3.1 State Dependent Coeﬃcient Parametrization . . . . . . . . . . . . . . 21
3.2 Local Asymptotic Stability Analysis . . . . . . . . . . . . . . . . . . 23
3.3 Global Asymptotic Stability Analysis . . . . . . . . . . . . . . . . . 23
3.4 Exponential Boundedness and Global Asymptotic Stability . . . . . . 24
3.4.1 Counter Example Showing Exponentially Boundedness does
not Guarantee Global Asymptotic Stability . . . . . . . . . . 25
3.5 Periodicity and Global Asymptotic Stability . . . . . . . . . . . . . . 28
3.5.1 Counter Example Showing that the Periodicity of the System
Matrix does not Guarantee Global Asymptotic Stability . . . . 29
3.5.2 An LMI Based Approach for Global Asymptotic Stability . . 34
3.5.3 Infeasibility of the LMI Feasibility Problem . . . . . . . . . . 39
3.5.4 Further Analysis for Global Asymptotic Stability . . . . . . . 42
v
vi Contents
3.6 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . 43
4 State Dependent Riccati Equation based Control Design 45
4.1 Optimal Control Problem and the SDARE . . . . . . . . . . . . . . . 46
4.2 Control System Design for the DP Vessel . . . . . . . . . . . . . . . 47
4.2.1 Nonlinear Regulation Problem for the LF Model . . . . . . . 47
4.2.2 Nonlinear Regulation Problem for the Complete Model . . . . 52
4.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5 The Fourier Series Interpolation Method 75
5.1 The Fourier Series Interpolation (FSI) Method . . . . . . . . . . . . . 76
5.2 Performance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6 PortHamiltonian Formulation and Passivity Based Control Design 83
6.1 HamiltonianBased Control . . . . . . . . . . . . . . . . . . . . . . . 84
6.1.1 PortHamiltonian Modeling . . . . . . . . . . . . . . . . . . 84
6.1.2 The IDAPBC Technique . . . . . . . . . . . . . . . . . . . . 84
6.2 Dynamic Positioning Problem . . . . . . . . . . . . . . . . . . . . . 86
6.3 Ship Model in PortHamiltonian Framework . . . . . . . . . . . . . . 87
6.3.1 Ship Model in Cartesian Coordinates . . . . . . . . . . . . . 87
6.3.2 Ship Model in PortHamiltonian Coordinates . . . . . . . . . 87
6.4 Classical IDAPBC Design . . . . . . . . . . . . . . . . . . . . . . . 88
6.4.1 A Quadratic Energy Shaping . . . . . . . . . . . . . . . . . . 90
6.4.2 A Trigonometric Energy Shaping . . . . . . . . . . . . . . . 90
6.5 Extended IDAPBC Design . . . . . . . . . . . . . . . . . . . . . . . 92
6.5.1 Motivating Problem . . . . . . . . . . . . . . . . . . . . . . 92
6.5.2 Target Extended System . . . . . . . . . . . . . . . . . . . . 95
6.5.3 A Quadratic Energy Shaping . . . . . . . . . . . . . . . . . . 97
6.5.4 A Trigonometric Energy Shaping . . . . . . . . . . . . . . . 97
6.5.5 Analysis in Presence of Disturbances . . . . . . . . . . . . . 97
6.6 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.6.1 Simulation Results for the classical IDAPBC design . . . . . 99
6.6.2 Simulation Results in Presence of Disturbances . . . . . . . . 101
6.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7 Conclusions and Recommendations for Future Work 103
A Proof of asymptotic stability of (6.32) at (6.33) 107
B Glossary 109
Bibliography 111
List of Abbreviations 117
Summary 119
Contents vii
Samenvatting 123
List of publications 125
Curriculum Vitae 126
viii
Chapter 1
Introduction
T
nis chapter explains the fundamental objective of this thesis and gives an overview
of the related details. This study is about the control system design of a dynamic
positioning (DP) vessel. A DP vessel is a vessel whose motion is controlled by a
dynamic positioning system rather than by the conventional motion control techniques
like mooring or anchoring.
1.1 What is a Dynamic Positioning (DP) System?
A DP system is a computer controlled system. The objective of this system is to
keep the vessel within speciﬁed position and heading limits exclusively by using the
propulsion system consisting of thrusters and propellers. Diﬀerent types of thrusters,
for instance, tunnel thrusters which produce thrust in sideway directions and azimuth
thrusters which are ﬁtted under the hull of vessel, are used to produce the desired ef
fects. Azimuth thrusters can be rotated through 360 degrees and thus produce thrust
in all directions in the horizontal plane. This is particularly useful because the envi
ronmental forces and moments change over time both in magnitude and direction.
A vessel in sea is subjected to various forces and moments due to waves, wind,
sea currents, propulsion system, and unmodeled disturbances due to the environmental
eﬀects and the propulsion system. In practice, aﬂoating vessel cannot maintain a com
pletely static position at sea. Therefore, for practical reasons, position keeping means
maintaining the desired position and heading within limits that reﬂect the environmen
tal eﬀects and the system capability. This limit may vary from centimeters to meters
depending upon the nature of the operation. For instance, centimeter accuracy is de
sired for the operations like automatic berthing of ships and maneuvering in shallow
and conﬁned waters. An eﬃcient DP system would be the one which achieves these
goals with minimum fuel consumption and also tolerates transient errors or failures in
the propulsion and measurement systems.
A complete DP system consists of three major parts: the vessel’s power system,
the thrusters system, and the DP control system. Figure 1.1 shows an overview of a
1
2 Chapter 1. Introduction
DP system, [72].
Positioning Control
System
(DP Computer)
Communication
Network
Vessel and
Environmental
Sensor Systems
Positioning Reference
System
DP Control System
Dynamic Positioning System
Main Properties
and Rudders
Thrusters, Drive
Units and
Auxiliaries
Cabling
and
Routing
Thruster Control
Units
Thruster System
Distribution System
(cables, routing,
etc,)
Uninterruptable
Power Supplies
(UPS)
Engine Auxiliary
Systems
(fuel, lub oil,
cooling water,
vetillation etc,)
Power System
Power
Management
System
Electrical
Switchboards
Electrical
Generators
Prime Movers
(engines)
Figure 1.1: Major components of a DP system
1.1.1 Applications of DP Systems
DP vessels are used to achieve a variety of objectives in the oﬀshore industry. The
main vessel types used for various oﬀshore operations include diving support vessels,
drilling vessels (drill ships and semisubmersible drill ships), ﬂoating production stor
age and oﬄoading (FPSO) units, landing platform docks, maritime research vessel,
mine sweepers, pipe laying ships, platform supply vessels and anchor handling ves
sels, shuttle tankers, and survey ships. Figure 1.2 shows an FPSO unit, a naval vessel
and a supply vessel.
Figure 1.2: A supply vessel, a naval ship and an FPSO vessel in sea (from left to right).
Dynamic positioning is vital in various oﬀshore operations. From the operational
aspects (for example in case of heavy lifts) it is important to maintain the vessel within
precise navigational limits. For this, a very stable and accurate position keeping is
required. There are situations in which a failure in position keeping, i.e., failure in
maintaining the position and heading angle, may result in serious safety and ﬁnancial
hazards. For instance, in case of diving vessels a failure in position keeping may
1.1. What is a Dynamic Positioning (DP) System? 3
result in the death or injury of the divers. In situations where the operation is being
carried out very close to a ﬁxed structure, then a position keeping failure may result
in a collision. Consequently, damage to the structure or vessel, equipment, or a delay
in the operation may occur. For instance, if a drilling vessel working in deep waters
makes widely twitchy movement then it will cause damage to riser pipe or drilling
pipe and subsequently the drilling operation will be abrupted.
The position keeping failure may occur because of multiple reasons; technical
failure of the DP equipment, operator’s error, extreme weather conditions not executed
in the control design strategy, etc. In many ships and various operations, over actuation
feature is included to enhance the operational continuity by reducing the chances of
failure of the propulsion and measurement systems. This feature gives rise to the
problem of optimal allocation because in the presence of this feature there can be
many possible combinations of actuators to yield a speciﬁc control action.
With the growing demand of the oﬀshore industry, the development in the DP
technologies is proliferating to meet the stringent safety, production and exploration
demands. This has made the users and the manufacturers of the DP systems strive
hard towards more reﬁnements in the DP related equipments and expertise. Conse
quently, there have been developments in all the faculties of the DP technologies like
navigation, control, propulsion and power units, and other subsidiary components.
DP systems have emerged as a popular replacement for the conventional position
keeping techniques: anchoring and jackup barge. While the conventional tools have
no or limited maneuverability, DP systems have excellent maneuverability and can
be easily moved from one place to another. No additional external equipment like
the anchoring tugs are required for DP systems. The anchoring may take several
hours but DP has very quick setup. The conventional techniques are limited by the
sea obstructions and sea depth but DP systems do not have such limitations. For
more information on the design, principles, and applications of dynamic positioning
systems interested readers are referred to [21].
1.1.2 Focus of this Research
It is clear from the foregoing discussion that a DP system consists of several compo
nents. The focus of this thesis is the design and analysis of the positioning control
system of the vessel, a subcomponent of the DP control system. This component
may well be considered as the heart of the DP system as it interacts with the rest
of the components of DP system. There can be diﬀerent control design objectives
depending upon the nature and demands of a DP operation. Some of these control ob
jectives include position and heading regulation, path following, trajectory tracking,
and waveinduced motion reduction. We focus on the position and heading regulation.
The basic element of a positioning control system is a mathematical model of the
vessel which is an approximation of the reality. We consider a nonlinear vessel model
from [22] and this model serves as a prototype for this study. The model has been
introduced in Chapter 2. We study the design and analysis of the control laws to sta
bilize (or regulate) the model to a desired equilibrium point. From a physical point
of view, we desire to maintain the position and heading of the vessel within desired
limits. In control design, it is important to take into account the size and the dynamic
4 Chapter 1. Introduction
response of the thrust devices which must be adequate to cope with various environ
mental conditions in diﬀerent oﬀshore operations. In practice, the maximum thrust
forces and moments to maintain the position and heading in diﬀerent environmen
tal conditions are estimated and then the capability of the thrust devices to meet the
demands is analyzed. This study is called capability study.
The prototype vessel model is nonlinear due to the heading angle of the vessel.
In [7], the state dependent coeﬃcient (SDC) parametrization is introduced which is a
strategy to transform the nonlinear system into a pseudolinear form. The advantage
of this approach is that it provides an opportunity to use concepts from linear system
theory to study the nonlinear vessel model. We use the SDC framework throughout
this thesis to study the control system design for the DP vessel.
The stability analysis is the most important feature of any control system design.
An unstable system is useless and potentially dangerous. Qualitatively, a dynamic
system is called stable if starting from a position somewhere near its equilibrium or
operating point implies that it will stay around the point ever after. Due to complex
and exotic behavior of the nonlinear systems, more reﬁned concepts of stability such
as (local and global) asymptotic stability are required to describe the behavior of non
linear systems. The asymptotic behavior implies that beside being stable, the system
will converge to its equilibrium or operating position as time goes on.
Lyapunov stability theory is the most commonly used tool to study the stability
properties of nonlinear systems. The prototype vessel model has a typical nonlinearity,
when described in pseudolinear form by using the SDC parametrization. We begin
our study with the stability analysis of pseudolinear systems similar to the prototype
DP vessel system. The special form of the vessel model motivated us to combine the
Lyapunov stability theory with linear matrix inequalities (LMIs) to come up with a
new method to analyze the global asymptotic stability of the pseudolinear systems of
the form similar to the prototype vessel model.
PID controllers are commonly used in practice. We use the SDC framework to
come up with the nonlinear version of the PID controller by using the state dependent
algebraic Riccati equation (from now on we call it the SDARE) technique for the
design of a stabilizing control law for the DP vessel. The computation of the controller
and the observer gains require online computation of the solution of the SDARE. It
can require large computation time, especially, for large systems. There are various
oﬀtheshelf methods for the solution of the SDARE. We come up with a new method,
the Fourier series interpolation (FSI) method, to solve the SDARE corresponding to
the DP vessel model. The FSI method reduces the computation time for the SDARE
in comparison with the Schur decomposition method.
The portHamiltonian formulation has also become a popular technique to study
the physical systems since a decade. We transform the DP vessel model into port
Hamiltonian form and then use the IDAPBC design approach to come up with a
family of control laws. These control laws may also be seen as the nonlinear version
of the wellknown PID controllers, in the portHamiltonian framework.
1.2. An Overview of this Thesis 5
1.2 An Overview of this Thesis
Chapters 1 and 2 contain the basic introductory material about the main theme of this
thesis. Our focus in this thesis is to address the control system design problem for
dynamic positioning of a sea vessel. The ﬁrst chapter of this thesis introduces the
dynamic positioning problem. The DP system is illustrated and the importance of
dynamic positioning is highlighted by describing its applications in various oﬀshore
and onshore operations. The second chapter introduces the details of the mathematical
model we use in this thesis to describe the vessel motion. It highlights the necessary
details of the vessel motion in mathematical form.
The main subject of the third chapter is the study of global asymptotic stability of
a special type of nonlinear systems which are similar to the prototype vessel model.
The SDC framework is used to express the nonlinear system in a pseudolinear form
and then the stability is analyzed based on the properties of the state dependent system
matrix. Two counter examples are presented in this chapter. The ﬁrst counter example
shows that the conditions, that the system matrix in pseudolinear form is continuous,
Hurwitz, and exponentially bounded, as reported in the literature on this subject, are
not suﬃcient for global asymptotic stability of the pseudolinear system. In the second
counter example, in addition to the set of conditions mentioned in the ﬁrst counter
example, additionally, we also assume that the system matrix is periodic. It is shown
that the extended set too does not constitute the set of suﬃcient conditions for global
asymptotic stability of the pseudolinear system. Apart from this, we also propose in
this chapter, a method for proving global asymptotic stability of the special pseudo
linear systems by combining the Lyapunov stability theory and the LMIs.
The fourth chapter addresses the control system design problem. The SDARE
based control design and estimation technique is used to design an SDARE controller
and an SDARE observer for dynamic positioning of the vessel. The ﬁfth chapter is
about the FSI method for the approximation of the solution of the SDARE. The FSI
method reduces the online computations of the solution of the SDARE by performing
the computationally expensive tasks oﬄine. The sixth chapter is also about the control
system design problem of the DP vessel. The main idea is to transform the vessel
model in the portHamiltonian structure and then use the IDAPBC design approach to
address the control design problem. The thesis is concluded with the seventh chapter
which brieﬂy summarizes the thesis and provides some concluding remarks. The
hindsight ideas for future research are also presented in this chapter.
1.3 Contributions of this Thesis
We study the stability analysis of nonlinear systems in the SDC framework. There
had been some existing results on this subject. Our main contribution on this topic
are two counter examples. It is claimed in the literature that it is suﬃcient for global
asymptotic stability of a pseudolinear system that the system matrix in its SDC form
is continuous, Hurwitz, and exponentially bounded. In a ﬁrst counter example, we
show that this claim is not valid. Motivated by the special type of state dependence of
the system matrix in DP vessel model, we assume additionally that the system matrix
6 Chapter 1. Introduction
is periodic and show by means of another counter example that an additional condition
also does not guarantee the stability of the nonlinear system. Each of these counter
examples have separately been published, see [53] and [54].
The special form of the nonlinearity in the vessel model and the Lyapunov stability
theory has lead us to propose a new approach to prove global asymptotic stability of
the special type of pseudolinear systems which resembles the prototype DP vessel
model. This approach makes use of the LMIs to achieve global asymptotic stability.
The approach is useful in particular for the vessel model and in general for the systems
having similar structure as the vessel model.
Another contribution is the SDARE controller design for the DP vessel. The model
based SDARE controller is a state feedback controller. The complete state of the DP
vessel model is not available in practice. Therefore, a state observer is also required.
We also used the SDARE observer to ﬁnd the state estimate. It has been shown that the
SDARE controller in combination with the SDARE observer gives the desired stability
and performance of the DP vessel. Alongside the SDARE controller and observer, a
numerical method for the approximation of the solution of the SDARE is proposed.
We call this the Fourier series interpolation (FSI) method. This method is proved to
be very handy in reducing the online computation time of the SDARE for controller
and observer gains computations. The concept of the FSI method has been presented
in a conference paper, see [52].
The ﬁnal contribution of this thesis is the use of the portHamiltonian structure
and the passivity theory for the ﬁrst time for DP vessel control design. We propose a
family of passivity based controllers for the DP vessel. Passivity idea is very attractive
in a sense that it helps in assigning the physical meaning to various variables and
quantities. The stability and performance of the family of the IDAPBC designs are
discussed. This idea was presented at a conference (see [50]) and it has recently been
accepted in a journal (see [51]).
Chapter 2
Mathematical Model of a Sea
Vessel
T
ne details of a vessel model for DP considerations are presented in this chapter.
The prototype vessel model described in this chapter will be used in the subse
quent chapters for studying the control system design of the DP vessel.
2.1 Motion of an Aﬂoating Vessel
In this section, we explain various terms associated with the motion of the vessel in
the sea. Motion of an aﬂoating vessel can be described by six degrees of freedom
(DOF), i.e., a vessel can move in six diﬀerent directions. We can categorize the six
DOF in two categories:
1. The translational motion in the following three directions,
• Surge: motion in backward (aft/stern) and forward (bow/fore) directions
• Sway: motion along sideways (transversial directions): starboard (right side of
the ship) and port (left side of the ship) directions
• Heave: motion in upward and downward directions
2. The rotational motion in the following three directions,
• Roll: rotation about the surge axis
• Pitch: rotation about the sway axis
• Yaw: rotation about the heave axis
Various modes of motion and the forces acting on the vessel are shown in Figure 2.1
and summarized in Table 2.1. A DP system is concerned primarily with control of
7
8 Chapter 2. Mathematical Model of a Sea Vessel
HEAVE
YAW
SWAY
PITCH
SURGE
ROLL
Figure 2.1: Six DOF of motion and forces acting on a ﬂoating vessel [Figure courtesy
of www.km.kongsberg.com].
Table 2.1: Nomenclature of the vessel motion
axis x y z
Translation surge sway heave
position x y z
velocity u v w
force X Y Z
+ direction forward starboard downward
Translation roll pitch yaw
angle φ θ ψ
rate p q r
torque K M N
+ direction starboard aft down right turn
the vessel in the horizontal plane, i.e., only the motions along surge, sway and yaw
directions are considered for DP purposes.
In Table 2.1, X and Y represent the forces in the surge and sway directions and N
denotes the turning eﬀect because of the thrusters and environmental eﬀects.
Dynamic positioning literature is very rich in terminology. For the convenience
of the reader, we brieﬂy explain some important terms which we use in this thesis. A
superstructure is an upward extension projected above the main deck of the vessel.
Maximum legal load amount on a vessel is characterized by water line. The distance
from the water line to the upper deck level is called freeboard. Water line and free
board are the characteristic of buoyancy level. Buoyancy level is a characteristics of
waters: it is low for fresh and warm waters and it is high for cold and saline waters
which have more density. Height and weight of the superstructure aﬀects the free
board. More the height and weight of the superstructure less will be the freeboard
area.
The rear or aft part of the vessel is termed as stern. Usually, during the night time,
2.2. Mathematical Model Describing the Dynamics of an Aﬂoating Vessel 9
the stern of the vessel is indicated with a white navigation light on it. The foremost
part of the vessel, opposite to the stern part, when the vessel is underway, is referred
as the bow. The right hand side of the vessel as perceived by a person on board facing
the bow is termed as starboard. The opposite part of the vessel on the left hand side
will then be called as port. All these terms are linked with the main deck of the vessel
and has nothing to do with the location of the superstructure on the deck. Figure 2.2
illustrates all these terms.
Super Structure
Water line
Free board
Aft or Stern
Fore or Bow
Starboard side
Port side
Figure 2.2: Commonly used terms in literature on dynamic positioning
2.2 Mathematical Model Describing the Dynamics of
an Aﬂoating Vessel
Modern DP control systems for ships use controllers based on a mathematical model
of the ship. This mathematical model describes the hydrodynamic, damping, environ
mental and control forces and moments acting on the vessel. For rigorous details on
the mathematical modeling of the dynamics of the ﬂoating vessel, interested readers
are referred to [22] and for an uptodate study we refer to [17] and [63]. The informa
tion from the measurement systems is transmitted to the controller and signals from
the controller are sent to the propulsion systems consisting of the thrusters and pro
pellers (planted at at least one of the aft, starboard, port and fore sides), propellers, and
rudders to generate the desired activity to maintain the required position and heading
10 Chapter 2. Mathematical Model of a Sea Vessel
of the vessel. Figure 2.3 shows an overview of the working of the DP system.
Optimum
Controller
Thruster
Allocation
Mathematical Model
SDP SYSTEM
Data from Reference
Systems and Sensors
Feed
Forward
Wind
Figure 2.3: Flow chart of the DP control system.
The testbed vessel model used in this thesis is introduced in this chapter and the
details describing the motion of an aﬂoating vessel are also explained. In the subse
quent chapters, this model will be used for various design and analysis purposes. This
section is divided into three major subsections. In the ﬁrst one, dynamical equations
of motion are explained. In the second subsection, we discuss how the perturbations
or the environmental disturbances are incorporated into the mathematical model of the
vessel dynamics. In the last subsection, the measurement models are explained.
2.2.1 The Dynamical Equations of Motion of the Vessel
The study of the dynamical equations of motion of a mechanical systemcan be divided
into two parts: the kinematic equations of motion which deal with the geometrical
aspects of the equations of motion, and the kinetic equations of motion which deal
with the analysis of the forces causing the motion.
The Kinematic Equations of Motion
An aﬂoating vessel has six degrees of freedom. Two frames of reference are used
to describe the motion: an Earthﬁxed inertial frame of reference and a bodyﬁxed
relative frame of reference. Figure 2.4 explains the description of both frames of
reference. For DP purposes, only the motion in the horizontal plane is considered. Let
2.2. Mathematical Model Describing the Dynamics of an Aﬂoating Vessel 11
y
X
O
x
r
y
r
O
r
Figure 2.4: The Earthﬁxed and the vesselﬁxed frames of reference.
η = [x y ψ]
T
describe the position (x, y) and heading ψ of the vessel in the inertial
frame of reference and ν = [u v r]
T
describe the velocities of the vessel in the relative
frame of reference. Then, the kinematic equations of motion in vectorial form are
given by
˙ η = J(ψ)ν. (2.1)
where the transformation matrix is given by
J(ψ) =
_
¸
¸
¸
¸
¸
¸
¸
¸
_
cos ψ −sin ψ 0
sin ψ cos ψ 0
0 0 1
_
¸
¸
¸
¸
¸
¸
¸
¸
_
. (2.2)
The kinematic transformation (2.2) relates the bodyﬁxed velocities to the derivative
of the positions in the inertial frame of reference. Note that the transformation matrix
is nonsingular and orthogonal, i.e., J(ψ)J
T
(ψ) = J
T
(ψ)J(ψ) = I
3
, ∀ψ ∈ R. This
property is very important from the design and stability analysis perspective as we
will see in the subsequent chapters. Also, we see that there is no uncertainty associated
with (2.1) as it describes the wellknown geometrical aspects of the model.
For conventional ships, it is an appropriate assumption that both the pitching and
rolling motions are oscillatory with zero mean and limited amplitude. Also due to
metacentric stability, there exist restoring moments in roll and pitch directions. There
fore, only the rotation matrix in yaw will be used to describe the kinematic equations
of motion.
12 Chapter 2. Mathematical Model of a Sea Vessel
The Kinetic Equations of Motion
The nonlinear kinetic equations of motion can be formulated by using Newtonian or
Lagrangian mechanics, for instance see [22] for a detailed study. In this work, the
equations in the Newtonian formulation based on Newton’s second law of motion are
considered. The nonlinear kinetic equations of motion in vectorial form can be written
as
˙ ν = −M
−1
Dν + M
−1
τ + M
−1
J
T
(ψ)b, (2.3)
τ = B
u
u. (2.4)
In (2.3) and (2.4), the vector τ = [X, Y, N]
T
∈ R
3×1
represents the control forces
and moment acting on the vessel in the bodyﬁxed frame of reference, provided by
the propulsion system of the ship consisting of propellers and thrusters. The vector
u ∈ R
r×1
(r ≥ 1) describes the control inputs and the matrix B
u
∈ R
3×r
is a constant
matrix describing the actuator conﬁgurations. The vector u is the command to the
actuators, which are assumed to have much faster dynamic response than the vessel;
thus the coeﬃcient B
u
represents the mapping from the actuator command to the force
generated by the actuators. In the following chapters, we assume a fully actuated
vessel model and we will take B
u
= I
3
. In the forthcoming chapters, we use the
vectors τ and u interchangeably, unless it is speciﬁed. The matrices M and D are 3×3
inertia and damping matrices, respectively. The vector b ∈ R
3×1
represents the slowly
varying bias forces and moments in the Earthﬁxed inertial frame of reference, due to
the waves, wind, sea currents, and other environmental factors surrounding the vessel.
For DP consideration, the inertia matrix has the following form
M =
_
¸
¸
¸
¸
¸
¸
¸
¸
_
m − X
˙ u
0 0
0 m − Y
˙ v
mx
G
− Y
˙ r
0 mx
G
− N
˙ v
I
z
− N
˙ r
_
¸
¸
¸
¸
¸
¸
¸
¸
_
R
3×3
, (2.5)
where m is the vessel mass, I
z
is the moment of inertia about the vesselﬁxed zaxis,
and x
G
denotes the longitudinal position of the center of gravity of the vessel with
respect to the relative frame of reference. The added masses due to acceleration in the
surge, sway, and yaw directions are deﬁned as
X
˙ u
=
∂X
∂˙ u
, Y
˙ v
=
∂Y
∂˙ v
, N
˙ r
=
∂N
∂˙ r
, Y
˙ r
=
∂Y
∂˙ r
, N
˙ v
=
∂N
∂˙ v
. (2.6)
Note that the inertia along the surge direction is decoupled from the inertia eﬀects
along the sway and yaw directions. Due to small velocities and starboardport sym
metries of the vessel, the added mass in sway due to the angular acceleration in yaw
is equal to the added mass in yaw due to sway acceleration, i.e., Y
˙ r
= N
˙ v
. Hence,
in DP applications, it is assumed that the matrix M is symmetric and strictly positive
deﬁnite, i.e., M = M
T
> 0. This assumption is very useful for the purpose of analysis.
The vessel motion generates waves. This means energy is transferred from vessel
to the ﬂuid and this energy is modeled by the linear damping term. The linear damping
matrix D for DP is taken as
D =
_
¸
¸
¸
¸
¸
¸
¸
¸
_
−X
u
0 0
0 −Y
v
−Y
r
0 −N
v
−N
r
_
¸
¸
¸
¸
¸
¸
¸
¸
_
R
3×3
. (2.7)
2.2. Mathematical Model Describing the Dynamics of an Aﬂoating Vessel 13
In most DP applications, the damping matrix is assumed to be real, nonsymmetrical,
and positive deﬁnite. However, for low speed applications where the damping matrix
is reduced to 2.7, it can be assumed that N
v
= Y
r
. In such a case, we assume the damp
ing matrix D to be real, symmetric, and positive deﬁnite. The damping components in
surge, sway, and yaw directions are deﬁned by
X
u
=
∂X
∂u
, Y
v
=
∂Y
∂v
, N
r
=
∂N
∂r
, Y
r
=
∂Y
∂r
, N
v
=
∂N
∂v
. (2.8)
Decoupling of the surge mode from the sway and yaw modes is beneﬁcial for the
convergence of parameter estimation algorithms, see [26]. An a priori estimate of the
mass and damping parameters of the vessel can be obtained by using semiempirical
methods and hydrodynamic computations. See [20] for details about the identiﬁcation
and estimation of vessel model parameters. Often the mass and damping parameters
estimates are updated based on the data obtained from the practical experiments in
calm waters.
2.2.2 The Disturbances Model
The forces acting on a sea vessel can be categorized in two main categories [35]: the
internal and the external forces and moments. The internal forces and moments are
formulated as functions of acceleration, velocities, propeller propulsions, and rudder
excitations. These have partially been discussed in the previous subsection. Here
we explain the external forces acting on the vessel. These forces are also termed as
external disturbances. The external disturbances can be distinguished into 3 major
categories [78]:
• Additive disturbances  These are the disturbances due to wind, waves, sea cur
rents, etc. These forces act additively on the vessel. To model and analyze these
forces, the model of the ship is extended by adding additional states.
• Multiplicative disturbances  A vessel in sea is also subject to the time varying
parameters such as load conditions, water depth, trim, speed changes, etc. These
disturbances are called multiplicative disturbances.
• Measurement disturbances  These are the disturbances due to the wrong func
tioning or noise in the measurement devices like DGPS and gyro compass.
The external disturbances due to unmodeled dynamics, waves, wind, and sea cur
rent acting on the vessel, are distinguished into two categories: second order low
frequency (LF) disturbances and ﬁrst order wave induced wave frequency (WF) dis
turbances. See for instance, [24], [27], and [47]. Along this thesis, we call the motion
of the vessel corresponding to these disturbances the LF motion and the WF motion,
respectively. The total vessel motion is then deﬁned to be the sum of the LF and WF
motions. The eﬀect of the WF disturbances is incorporated in the measurement model,
described in the next section. Figure 2.5 illustrates the concept of the slowly varying
LF and the oscillatory WF motions.
In what follows, both the LF and WF disturbances are characterized by respective
dynamical models. In the following, some explanation of these dynamic models is
given.
14 Chapter 2. Mathematical Model of a Sea Vessel
0 10 20 30 40 50 60 70 80 90 100
−2
0
2
4
6
8
10
Time [sec]
A Description of LF and WF Motion
WF Motion
LF Motion
LF and WF Motion
Figure 2.5: The LF and the WF motions
Second Order LF Disturbances
The LF disturbances are also sometimes termed as slowly varying bias forces and
moment in surge, sway, and yaw directions. The low frequency motions are caused
by the forces generated by the thrusters and propellers, wind forces, waveinduced
forces, and hydrodynamic forces. In marine control applications, these forces and
moment can be described, [22], by the ﬁrst order Markov process given by
˙
b = −T
−1
b + Ψw
b
, (2.9)
where b ∈ R
3×1
is a vector of bias forces and moment, the vector w
b
∈ R
3×1
represents
the zeromean Gaussian white noise process, i.e., w
b
∼ (0, Q
c,b
), T ∈ R
3×3
is a diag
onal matrix of positive bias time constants and Ψ ∈ R
3×3
is a diagonal matrix scaling
the amplitude of the noise vector w
b
. The matrix T is known as the time constant. In
this context it will have relatively large values as sea states change very slowly. We
can also interpret (2.9) as a lowpass ﬁlter.
In many applications, see for instance [25, 75], it is considered more appropriate
from a physical point of view to use
˙
b = Ψw
b
to describe the bias model. This
may be described as integration of the noise signal which in fact is a random walk
phenomenon. Thus bias forces and moments are sometimes modeled as a random
walk process. Another case could be that the bias forces and moment are constant.
Then the bias model will be
˙
b = 0.
2.2. Mathematical Model Describing the Dynamics of an Aﬂoating Vessel 15
First Order Wave Induced WF Disturbances
The fundamental assumption for the development of the WF motion model is that the
sea state is known and can be described by a spectral density function. The ﬁrst order
wave induced WF disturbances in surge, sway, and yaw directions are modeled as
second order harmonic oscillations which are driven by Gaussian white noise process.
It was Balchen who ﬁrst modeled the WF motion in this way, [5]. For each of the
three directions, the WF disturbances model in frequency domain is given by
ξ
i
(s) =
σ
i
s
s
2
+ 2ζ
i
ω
0i
s + ω
2
0i
w
ξi
(s), i = 1, 2, 3 (2.10)
where ω
0i
is the dominating (sometimes also termed as undamped) wave frequency,
ζ
i
is the relative damping ratio, and σ
i
is the wave intensity parameter. The input w
ξi
represents the Gaussian white noise process, i.e., w
ξi
∼ (0, Q
c,ξi
). The damping ratio
ζ
i
is a measure to describe how the oscillations in the system (2.10) decay when a
disturbance is introduced. Normally, the damping ratio deﬁnes the level of damping
(underdamped, overdamped, criticallydamped, and undamped) of the system. The
dominating wave frequency ω
0i
is obtained by spectral analysis.
In state space representation, the WF disturbances model for each direction can be
written as
˙
ξ
(i)
1
= ξ
(i)
2
˙
ξ
(i)
2
= −ω
2
0i
ξ
(i)
1
− 2ζ
i
ω
0i
ξ
(i)
2
+ σ
i
w
i
, i = 1, 2, 3 (2.11)
A compact state space realization of the WF model is given by
_
˙
ξ
1
˙
ξ
2
_
=
_
O
3
I
3
−Ω
2
−2ZΩ
_ _
ξ
1
ξ
2
_
+
_
O
3
Σ
_
w
ξ
, (2.12)
where ξ
1
= [ξ
(1)
1
, ξ
(2)
1
, ξ
(3)
1
]
T
, ξ
2
= [ξ
(1)
2
, ξ
(2)
2
, ξ
(3)
2
]
T
, Ω = diag¦ω
01
, ω
02
, ω
03
¦, Z =
diag¦ζ
1
, ζ
2
, ζ
3
¦, and Σ = diag¦σ
1
, σ
2
, σ
3
¦. The matrix O
3
∈ R
3×3
is a zero matrix.
The WF motion parameters ω
0
, ζ, and σ depend on the sea states, structure of the
vessel and the direction of the incident waves. The vector w
ξ
describes the Gaus
sian white noise process, i.e., w
ξ
∼ (0, Q
c,ξ
) The state vector may or may not have a
physical interpretation depending on the particular statespace realization used.
For DP operations only the LF motion is required to be controlled. Therefore,
the oscillatory WF motion is required to be ﬁltered or separated from the LF motion.
This is important to avoid unnecessary power consumption and possible wear and tear
of the actuators. The WF response is required to be controlled in certain operations
like ride control of a passenger vessel, where reducing pitch and role motion helps
avoiding motion sickness. The stochastic nature of environmentally induced forces
and moments has made the Kalman ﬁlter an essential part of the modern sea vessel
motion control systems.
Filtering of the WF motion can be done either by using appropriate classical ﬁl
tering techniques or it can be done by state estimation. Using ﬁltering techniques, it
is important to know the threshold frequency. Another problem with the ﬁltering ap
proach is possible phase lag due to delay of the signals. For an estimation approach,
16 Chapter 2. Mathematical Model of a Sea Vessel
we need to know the parameters of the system. Even in the linearized case, parameters
are required. It is also important to keep in mind that the linearized model may not be
a good approximation of the actual model or system.
2.2.3 The Measurement Model
The position and heading of the vessel in the inertial frame of reference can be mea
sured by using a diﬀerential global positioning system (DGPS) and a gyrocompass.
For reliability, some vessels have multiple sensors. The measurement model can be
described, using the superposition principle, by the following vector equation
y = y
b
+ y
ξ
+ v, (2.13)
where y
b
= η and y
ξ
= ξ
1
are, respectively, the position and heading measurements of
the vessel corresponding to the LF and the WF motions and the vector v ∈ R
3×1
is the
Gaussian white noise process, i.e., v ∼ (0, R
c
). The vector v describes the measurement
noise.
2.2.4 Wave Filtering
In (2.13), the measured output is assumed to essentially contain the LF and WF mo
tion components. The separation of the WF component from the LF component is
termed as wave ﬁltering. This action is also important to avoid thruster modulation, a
phenomenon which gives rise to high frequency ﬂuctuations in the thrust demand in
the control loop. Knowledge of the sea states is required to determine the WF motion
of the vessel. Sea states can be distinguished in 9 diﬀerent forms depending on the
signiﬁcant wave height [22].
Lowpass, notch, and deadband ﬁlters were the most commonly used wave ﬁlter
ing techniques in earlier DP systems. The main drawback of these techniques was
due to a signiﬁcant phase lag to meet the high gain control requirements. In earlier
DP systems, wave ﬁltering was accomplished by using a proportional controller with
a deadband nonlinearity. This deadband produced a null control action until the con
trol signal was inside the deadband. The length of this deadband could be increased
by the operator with changing weather conditions. This change in length was termed
as ‘weather’ as it was subject to the weather conditions [77].
2.3 Summary of the Mathematical Model
For more insight and a clear picture of the model of a dynamic positioning vessel, we
summarize all the modeling details from the previous section of this chapter. The LF
motion model is described by combining the equations (2.1), (2.3), and (2.9), and it is
given by
2.3. Summary of the Mathematical Model 17
˙ η = J(ψ)ν, (2.14)
˙ ν = −M
−1
Dν + M
−1
τ + J
T
(ψ)b, (2.15)
˙
b = −T
−1
b + Ψw
b
. (2.16)
In matrix form, we can write the LF motion model as
_
¸
¸
¸
¸
¸
¸
¸
¸
_
˙ η
˙ ν
˙
b
_
¸
¸
¸
¸
¸
¸
¸
¸
_
=
_
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
J(ψ) O
3
O
3
−M
−1
D M
−1
J
T
(ψ)
O
3
O
3
−T
−1
_
¸
¸
¸
¸
¸
¸
¸
¸
_
.,,.
A
b
(ψ)
_
¸
¸
¸
¸
¸
¸
¸
¸
_
η
ν
b
_
¸
¸
¸
¸
¸
¸
¸
¸
_
.,,.
x
b
+
_
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
M
−1
O
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
.,,.
B
b
τ +
_
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
O
3
Ψ
_
¸
¸
¸
¸
¸
¸
¸
¸
_
.,,.
E
b
w
b
, (2.17)
The output equation of the LF model is
y
b
=
_
I
3
O
3
O
3
_
.,,.
C
b
_
¸
¸
¸
¸
¸
¸
¸
¸
_
η
ν
b
_
¸
¸
¸
¸
¸
¸
¸
¸
_
+ v. (2.18)
The dynamics of the WF motion model and its output are given by
_
˙
ξ
1
˙
ξ
2
_
=
_
O
3
I
3
−Ω
2
−2ZΩ
_
.,,.
A
ξ
_
ξ
1
ξ
2
_
.,,.
x
ξ
+
_
O
3
Σ
_
.,,.
E
ξ
w
ξ
, (2.19)
y
ξ
=
_
I
3
O
3
_
.,,.
C
ξ
_
ξ
1
ξ
2
_
. (2.20)
The complete model of motion of the vessel can be described by augmenting both
the LF and WF models and it can be written in the following form.
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
˙ η
˙ ν
˙
b
˙
ξ
1
˙
ξ
2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
J(ψ) O
3
O
3
O
3
O
3
−M
−1
D M
−1
J
T
(ψ) O
3
O
3
O
3
O
3
−T
−1
O
3
O
3
O
3
O
3
O
3
O
3
I
3
O
3
O
3
O
3
−Ω
2
−2ZΩ
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
.,,.
A(ψ)
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
η
ν
b
ξ
1
ξ
2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
.,,.
x
+
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
M
−1
O
3
O
3
O
3
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
.,,.
B
τ +
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
O
3
O
3
O
3
Ψ O
3
O
3
O
3
O
3
Σ
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
.,,.
E
_
w
b
w
ξ
_
.,,.
w
.
(2.21)
The output of (2.21) can be obtained by using the superposition principle, see (2.13),
and is written in matrix form as
y =
_
I
3
O
3
O
3
I
3
O
3
_
.,,.
C
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
η
ν
b
ξ
1
ξ
2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
+ v. (2.22)
18 Chapter 2. Mathematical Model of a Sea Vessel
The complete model of the vessel in compact form is written as
˙ x =A(ψ)x + Bτ + Ew, (2.23)
y =Cx + v. (2.24)
The vectors w and v represent the Gaussian white noise processes, i.e., w ∼ (0, Q
c
) and
v ∼ (0, R
c
), where Q
c
= diag¦Q
c,b
, Q
c,ξ
¦. The system (2.23)(2.24) is a pseudolinear
system because of dependency of the system matrix A(ψ) on the heading angle. We
call this pseudolinear form the state dependent coeﬃcient (SDC) parametrization of
nonlinear system. Chapter 3 explains the concept of the SDC parametrization in more
detail.
2.4 Properties of the Model
In this thesis, we deal with the control design and estimation problems of the DP vessel
discussed in the previous sections. For this purpose, it is important to know certain
properties of the vessel model which play a fundamental part in the control design and
estimation techniques which we are going to use in the subsequent chapters. These
properties include controllability, observability, stabilizability, and detectability. In
the following, we recall some important results about these properties in the context
of the pseudolinear systems, presented in [6] .
Deﬁnition 2.4.1. (Controllability in terms of rank condition) The pseudolinear
system of the form (2.23)(2.24) with an ndimensional state vector is pointwise con
trollable iﬀ the controllability matrix
( =
_
B A(ψ)B A
2
(ψ)B A
n−1
(ψ)B
_
, (2.25)
has rank equal to the dimension of the state vector x, for each ψ ∈ R. This is also
termed as the pair (A(ψ), B) is pointwise controllable.
Deﬁnition 2.4.2. (Observability in terms of rank condition) The pseudolinear sys
tem of the form (2.23)(2.24) with an ndimensional state vector is pointwise observ
able iﬀ the observability matrix
O =
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
C
CA(ψ)
CA
2
(ψ)
.
.
.
CA
n−1
(ψ)
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
, (2.26)
has rank equal to the dimension of the length of the state vector x, for each ψ ∈ R.
This is also termed as the pair (C, A(ψ)) is pointwise observable.
It can easily be checked that the controllability matrices corresponding to the sys
tems (2.17) and (2.21) have rank 6, i.e., only the position and the velocities can be
controlled. This is not restrictive as the LF bias forces and the WF motions cannot
2.4. Properties of the Model 19
be controlled. The observability matrices corresponding to both the systems have full
column ranks for all ψ ∈ R. So the systems are pointwise observable, i.e., we can
build the states of the system from the knowledge of the input and the output.
The stabilizability and detectability are weaker conditions than the controllability
and observability, respectively. These properties are important from the point of view
of the existence of the solution of the SDARE. In the following, we deﬁne a necessary
and suﬃcient condition for the pointwise stabilizability and detectability of a pseudo
linear system, see [6] for more details.
Deﬁnition 2.4.3. (Pointwise Stabilizability) The pseudolinear system of the form of
(2.23)(2.24) with an ndimensional state vector is pointwise stabilizable iﬀ
rank
_
λI − A(ψ) B
_
= n, (2.27)
for each eigenvalue λ of A(ψ) which has a nonnegative real part (Re(λ ≥ 0)) and for
all ψ ∈ R. This is also termed as the pair (A(ψ), B) is stabilizable.
Deﬁnition 2.4.4. (Pointwise Detectability) The pseudolinear system of the form
(2.23)(2.24) with an ndimensional state vector is pointwise detectable iﬀ
rank
_
λI − A(ψ)
C
_
= n, (2.28)
for each eigenvalue λ of A(ψ) which has a nonnegative real part (Re(λ ≥ 0)) and for
all ψ ∈ R. This is also termed as the pair (C, A(ψ)) is detectable.
Due to the special structure of the system matrices A
b
(ψ) in (2.17) and A(ψ) in
(2.21), the eigenvalues of the system matrices do not change with the variable ψ. The
only nonnegative eigenvalue of both A
b
(ψ) and A(ψ) is 0 with algebraic multiplicity
3. This makes it an easy task to compute the rank conditions (2.27) and (2.28). It
can be checked that the rank is 9 for both stabilizability and detectability conditions
corresponding to A
b
(ψ) and it is 15 for A(ψ). Thus both the systems (2.17) and (2.21)
are pointwise stabilizable and detectable.
20
Chapter 3
SDC Parametrization and
Stability Analysis of Autonomous
Nonlinear Systems
1
T
ne stability analysis of nonlinear systems has always been a challenging task. This
is mainly because of phenomena like ﬁnite escape time and limit cycles, see for
instance [30], [54], [74], [76], and [82]. Numerous techniques for stability analysis of
nonlinear systems have been proposed over time, for further details see [40], [42], and
[81]. One such approach is to ﬁrst write the nonlinear system dynamics in linearlike
form using a state dependent coeﬃcient (SDC) parametrization and then analyze the
possible extension of the results of linear systems theory for the stability analysis of
nonlinear systems. The SDC representation provides a systematic way to analyze the
extension of the results of linear systems theory for the stability analysis of nonlinear
systems.
3.1 State Dependent Coeﬃcient Parametrization
Let Ω ⊆ R
n
and f (x) be a vector function from Ω to R
n
. Consider the following
nonlinear system
˙ x = f (x), x
0
= x(t
0
), (3.1)
where x ∈ Ω is the state of the system. If the vector function f : Ω −→ R
n
, is continu
ously diﬀerentiable and f (0) = 0, then it is always possible [43] to write f (x) = A(x)x.
Let us call the matrix A(x) as the state dependent coeﬃcient (SDC) parametrization of
f (x). It is important to mention that the SDC parametrization is not unique unless f (x)
1
Section 3.4.1 and Section 3.5.1 of this chapter have been published in the form of two separate articles
in the IMA Journal of Mathematical Control and Information, see [53] and [54].
21
22 Chapter 3. SDC Parametrization and Stability Analysis
is a scalar function. For example, if A
1
(x) and A
2
(x) are two distinct parametrizations
of f (x) then for 0 ≤ α ≤ 1,
αA
1
(x)x + (1 − α)A
2
(x)x = αf (x) + (1 − α) f (x) = f (x),
i.e.
αA
1
(x) + (1 − α)A
2
(x)
is also a parametrization of f (x). In fact inﬁnitely many parametrizations are possible
but one has to chose only those which are appropriate for the desired objectives. For
more details on the SDC parametrization, interested readers are referred to [34] and to
the references therein.
An important property of the SDCparametrization is that it preserves the lineariza
tion of nonlinear systems. If A(x) is any parametrization of f (x) then A(0) = Vf ¦
x=0
.
The following Lemma from [6] establishes this fact.
Lemma 3.1.1. For any SDC parametrization A(x) of f (x), A(0)x is the linearization
of f (x) at the zero equilibrium.
Proof. See [6].
From here onward, we will use the notions of the coeﬃcient matrix A(x) and
the system matrix interchangeably. Consider the following pseudolinear autonomous
system
˙ x = A(x)x, x
0
= x(t
0
). (3.2)
In the remainder of this chapter, we analyze the stability properties of the pseudo
linear system (3.2). We conﬁne our analysis to the following conditions on the system
matrix of (3.2).
C.1 The matrix function, A : Ω −→ R
n×n
is a C
1
function.
C.2 A(x) is pointwise asymptotically stable (Hurwitz) matrix, i.e., all eigenvalues of
A(x) lie in the open left half plane for all x ∈ Ω. Consequently, we see that the
origin ¯ x = 0 is the only equilibrium point of the system (3.2).
C.3 The system matrix, A(x), is exponentially bounded i.e., ¦¦e
A(x)t
¦¦ ≤ M for some
real M > 0 and ∀x ∈ Ω, ∀t ∈ [0, ∞).
C.4 A(x) is a periodic function with a period θ, i.e., A(x + θ) = A(x), ∀ x ∈ Ω.
The ﬁrst two conditions imply that there is only one isolated equilibrium point, ¯ x = 0,
of (3.2). Therefore, the stability analysis of (3.2) will be with reference to this equi
librium point. The conditions C.1 and C.2 are suﬃcient to prove local asymptotic
stability. The conditions C.3 and C.4 are imposed to analyze global asymptotic stabil
ity.
3.2. Local Asymptotic Stability Analysis 23
3.2 Local Asymptotic Stability Analysis
We start with the local asymptotic stability considerations. It can be deﬁned as follows
[68]:
Deﬁnition 3.2.1. An equilibrium point ¯ x of the nonlinear system (3.2) is (locally)
asymptotically stable if it is stable, and if in addition there exists some r > 0 such that
¦¦x(0)¦¦ < r implies that x(t) → ¯ x as t → ∞.
Since A(x) is continuously diﬀerentiable, therefore, col¦A(x)¦ ∈ C
1
. By col¦A(x)¦,
we mean the set of columns of A(x). Applying the Mean Value Theorem [49] to
col¦A(x)¦, we can write
col
j
¦A(x)¦ = col
j
¦A(0)¦ +
∂col
j
¦A(z
j
)¦
∂x
x, j = 1, 2, ..., n (3.3)
where the vector z
j
is a point, on the line connecting the origin and the point x, which
yields equality in the jth equation of (3.3). By col
j
¦A(x)¦, we mean the jth column of
A(x). Using (3.3) in (3.2), we can write
˙ x = A(0)x +
_
∂col
1
¦A(z
1
)¦
∂x
x
∂col
2
¦A(z
2
)¦
∂x
x ...
∂col
n
¦A(z
n
)¦
∂x
x
_
x,
= A(0)x +
n
j=1
n
i=1
x
i
x
j
∂col
j
¦A(z
j
)¦
∂x
i
.
Multiplying and dividing the second term by ¦¦x¦¦ and deﬁning
ψ(x, z
1
, z
2
, ..., z
n
)
=
n
j=1
n
i=1
x
i
x
j
¦¦x¦¦
∂col
j
¦A(z
j
)¦
∂x
i
,
we get,
˙ x = A(0)x + ψ(x, z
1
, z
2
, ..., z
n
)¦¦x¦¦. (3.4)
Since
lim
¦¦x¦¦→0
ψ(x, z
1
, z
2
, ..., z
n
) = 0, (3.5)
and A(0) is Hurwitz, therefore ¯ x is a locally asymptotically stable equilibrium point of
(3.4). This means, that the conditions C.1 and C.2 ensure that ¯ x is a locally asymptot
ically stable equilibrium point of (3.2).
3.3 Global Asymptotic Stability Analysis
Now we proceed to the global asymptotic stability considerations. The requirements
for global asymptotic stability of a nonlinear system of the form (3.2) are the follow
ing.
i. There is only one equilibrium point, ¯ x ∈ Ω, of the system.
24 Chapter 3. SDC Parametrization and Stability Analysis
ii. The equilibrium point is locally asymptotically stable.
iii. lim
t→∞
x(t, x
0
) = ¯ x, ∀ x
0
∈ Ω, i.e., starting from any point x
0
∈ Ω, the state of the
system converges to the equilibrium point ¯ x as time goes to inﬁnity.
The conditions C.1 and C.2 are not suﬃcient to guarantee global asymptotic sta
bility of (3.2). Global asymptotic stability of nonlinear systems in this form was ﬁrst
studied by Banks and Mhana [7]. They came up with the following remark:
Remark 3.3.1. If
¯
A(x) is a continuous matrixvalued function which is asymptotically
stable for each x, then the equation
˙ x =
¯
A(x)x, x(0) = x
0
, (3.6)
is asymptotically stable for all x
0
.
This statement of Banks and Mhana is an optimistic extension of the eigenvalue
based stability test for linear systems, to analyze the stability of the pseudolinear
systems (3.2). Ultimately, it was proved wrong, independently, in [44] and [76] by a
simple counter example. The counter example is the following nonlinear system
˙ x =
_
−1 x
2
1
0 −1
_
x, x
0
= x(t
0
). (3.7)
This system satisﬁes Banks and Mhana’s hypothesis: A(x
1
) is continuous and asymp
totically stable. But if the initial condition is taken as x
0
(0) = [2 2]
T
, then simple
calculations yield the following solution of (3.7):
x
1
(t) =
2x
2
(t)
x
2
2
(t) − 2
and x
2
(t) = 2e
−t
for all t ∈ [0, T
c
) with T
c
= ln
√
2. It is obvious that as t tends to T
c
then x
2
tends to
√
2 and consequently x
1
tends to ∞. Therefore, the system (3.7) is not asymptotically
stable for all x
0
. Such a departure of a state variable to inﬁnity at a ﬁnite time is called
the ﬁnite escape time phenomenon. From this counter example it is obvious that the
pointwise asymptotic stability of the system matrix A(x) does not help us to draw any
conclusion about the stability of the nonlinear system (3.2).
In the following two sections, we continue with global asymptotic stability anal
ysis. In Section 3.4, global asymptotic stability is analyzed with respect to the expo
nential boundedness of the system matrix. In Section 3.5, it is analyzed with respect
to the periodicity assumption on the system matrix.
3.4 Exponential Boundedness and Global Asymptotic
Stability
In this section, we continue with the ﬁndings of Langson and Alleyne and ultimately
give a counter example to show that global asymptotic stability is not guaranteed when
the system matrix is exponentially bounded. Langson and Alleyne [44] studied this
topic further and came up with the following two hypothesis:
3.4. Exponential Boundedness and Global Asymptotic Stability 25
Proposition 3.4.1. Consider the system ˙ x = A(x)x, where A : R
n
−→ R
n×n
is uni
formly continuous in x and A(x) is a stable matrix ∀x ∈ R
n
. The origin of the given
system is an asymptotically stable equilibrium point.
Corollary 3.4.1. If the hypothesis of Proposition 3.4.1 is satisﬁed with ¦¦e
A(x)t
¦¦ ≤
M for some real M > 0 and ∀x ∈ R
n
, ∀t ∈ [0, ∞), then the system ˙ x = A(x)x is
asymptotically stable for any arbitrary ﬁnite initial condition.
In the following subsection, a counter example [53] is presented in which the
hypothesis of Langson and Alleyne is satisﬁed but where the system is not globally
asymptotically stable.
3.4.1 Counter Example Showing Exponentially Boundedness does
not Guarantee Global Asymptotic Stability
Example 3.4.1. We start with the following SDC formulation of a nonlinear system
in a general setting
˙ x =
_
a b − c(x)
−b − c(x) a
_
x, x
0
= x(t
0
), (3.8)
where a, b ∈ R and c(x) is a smooth function: c : R
2
→ R. We show that the coeﬃcient
matrix in (3.8) satisﬁes the hypothesis of Langson and Alleyne, for certain choices of
the parameters a and b, and the scalar function c(x): a < 0 and b > ¦c(x)¦ for all
x ∈ R
2
.
1. Continuity: From the description of the coeﬃcient matrix in (3.8), it is obvi
ous that the coeﬃcient matrix is a continuous function: A : R
2
→ R
2×2
is a
continuous function.
2. Asymptotic Stability: The general expression for the eigenvalues of the system
matrix in (3.8) has the following form
λ
1,2
= a ±
_
c
2
(x) − b
2
. (3.9)
Clearly, if a < 0 and b
2
> c
2
(x) for all x ∈ R
2
, then A(x) is Hurwitz (asymptot
ically stable).
3. Exponential Boundedness: Under this subject, we derive a general expression
for the upper bound of the matrix exponential of the coeﬃcient matrix in (3.8).
For the sake of convenience, in the sequel we write c instead of c(x). We proceed
as follows
26 Chapter 3. SDC Parametrization and Stability Analysis
e
_
¸
¸
¸
¸
¸
¸
¸
_
a b − c
−b − c a
_
¸
¸
¸
¸
¸
¸
¸
_
t
= e
_
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
_
a 0
0 a
_
¸
¸
¸
¸
¸
¸
¸
_
+
_
¸
¸
¸
¸
¸
¸
¸
_
0 b − c
−b − c 0
_
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
_
t
= e
_
¸
¸
¸
¸
¸
¸
¸
_
a 0
0 a
_
¸
¸
¸
¸
¸
¸
¸
_
t
e
_
¸
¸
¸
¸
¸
¸
¸
_
0 b − c
−b − c 0
_
¸
¸
¸
¸
¸
¸
¸
_
t
= e
at
e
_
¸
¸
¸
¸
¸
¸
¸
_
0 b − c
−b − c 0
_
¸
¸
¸
¸
¸
¸
¸
_
t
. (3.10)
We use here the fact that if A
1
and A
2
commute then e
A
1
+A
2
= e
A
1
e
A
2
. Now
consider the following transformation to make the antidiagonal entries of the
matrix in the second exponent of (3.10) the additive inverse of each other.
_
1 0
0 γ
−1
_ _
0 b − c
−b − c 0
_ _
1 0
0 γ
_
=
_
0 −k
k 0
_
,
_
0 (b − c)γ
(−b − c)γ
−1
0
_
=
_
0 −k
k 0
_
.
Solving the pair of equations
(b − c)γ = −k and γ
−1
(−b − c) = k,
for γ and k, we get
γ = ±
_
b + c
b − c
and k = ±
√
b
2
− c
2
.
We know that
e
At
= Te
(T
−1
AT)t
T
−1
. (3.11)
Therefore, by taking the positive value of γ, we have
e
_
¸
¸
¸
¸
¸
¸
¸
_
0 b − c
−b − c 0
_
¸
¸
¸
¸
¸
¸
¸
_
t
=
_
1 0
0 γ
_
e
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
0
√
b
2
− c
2
−
√
b
2
− c
2
0
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
t
_
1 0
0 γ
−1
_
.
Hence, we can write (3.10) as
e
_
¸
¸
¸
¸
¸
¸
¸
_
a b − c
−b − c a
_
¸
¸
¸
¸
¸
¸
¸
_
t
= e
at
_
1 0
0 γ
_
e
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
0
√
b
2
− c
2
−
√
b
2
− c
2
0
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
t
_
1 0
0 γ
−1
_
. (3.12)
3.4. Exponential Boundedness and Global Asymptotic Stability 27
We know that
e
_
¸
¸
¸
¸
¸
¸
¸
_
0 x
−x 0
_
¸
¸
¸
¸
¸
¸
¸
_
=
_
cos x sin x
−sin x cos x
_
. (3.13)
Taking norm (we use the spectral norm) on both sides of (3.12) and using (3.13),
we get
¦¦e
_
¸
¸
¸
¸
¸
¸
¸
_
a b − c
−b − c a
_
¸
¸
¸
¸
¸
¸
¸
_
t
¦¦
≤ e
at
¦¦
_
1 0
0 γ
_
¦¦ ¦¦
_
¸
¸
¸
¸
_
cos
√
b
2
− c
2
t sin
√
b
2
− c
2
t
−sin
√
b
2
− c
2
t cos
√
b
2
− c
2
t
_
¸
¸
¸
¸
_
¦¦ ¦¦
_
1 0
0 γ
−1
_
¦¦.
(3.14)
We have derived the general expressions for the eigenvalues, (3.9), and the upper
bound of the matrix exponential, (3.14), for the system matrix in (3.8). Now, we show
that the hypothesis of Langson and Alleyne, is satisﬁed if we take a suitable combi
nation of the parameters a and b with scalar function c(x). For example, we take
a = −0.1, b = 3, and c(x) = −
8
π
2
tan
−1
x
1
tan
−1
x
2
. Then e
−0.1t
≤ 1,
1
√
5
< γ <
√
5, and
¦¦
_
¸
¸
¸
¸
_
cos
√
b
2
− c
2
t sin
√
b
2
− c
2
t
−sin
√
b
2
− c
2
t cos
√
b
2
− c
2
t
_
¸
¸
¸
¸
_
¦¦ = 1,
¦¦
_
1 0
0 γ
−1
_
¦¦ ≤
√
5 and ¦¦
_
1 0
0 γ
_
¦¦ ≤
√
5.
Therefore from (3.9) and (3.14), we have
Re(λ
max
) = −0.1 and ¦¦e
_
¸
¸
¸
¸
¸
¸
¸
_
−0.1 3 − c(x)
−3 − c(x) −0.1
_
¸
¸
¸
¸
¸
¸
¸
_
t
¦¦ ≤ 5,
for all x ∈ R
2
and t ∈ [0, ∞).
The system matrix has the following form
A(x) =
_
−0.1 3 +
8
π
2
tan
−1
x
1
tan
−1
x
2
−3 +
8
π
2
tan
−1
x
1
tan
−1
x
2
−0.1
_
. (3.15)
It is clear from the foregoing discussion that the system matrix in (3.15) is continu
ous, asymptotically stable (Hurwitz), and exponentially bounded. Thus the hypothesis
of Langson and Alleyne is satisﬁed.
Fig. 3.1 shows the phaseportrait of the system dynamics in (3.8) using the system
matrix (3.15) with an initial condition, x
0
= [1.2 0]
T
. It indicates that the states of
the system move away from the origin as time goes to inﬁnity although the coeﬃcient
matrix satisﬁes the suﬃcient conditions (as claimed in [44]) for global asymptotic
stability.
28 Chapter 3. SDC Parametrization and Stability Analysis
−2 −1 0 1 2 3
−4
−3
−2
−1
0
1
2
3
xy−trajectories
State x
1
S
t
a
t
e
x
2
Figure 3.1: Phaseportrait of the system dynamics (3.8)
3.5 Periodicity and Global Asymptotic Stability
From the counter example in the previous section, it is clear that the hypothesis of
Langson and Alleyne is not suﬃcient to endorse global asymptotic stability of non
linear systems of the form (3.2). At this point, a natural question is, what additional
conditions would be required to establish global asymptotic stability of nonlinear sys
tems of the form (3.2)? In addition to the smoothness conditions and exponential
boundedness, we study the case that A(x) is also a θ−periodic matrix, i.e.,
A(x) = A(x + θ) for all x ∈ Ω and some θ ∈ R. (3.16)
Condition C.4, that the system matrix A(x) is periodic, ensures that the ﬁnite escape
time phenomenon will not occur. If a function is continuous and periodic on Ω then it
must be bounded. In the following proposition, we prove howthe periodicity condition
rules out the possibility of the ﬁnite escape time phenomenon.
Proposition 3.5.1. Suppose that the system(3.2) satisﬁes the usual regularity (smooth
ness) conditions. If the system matrix in (3.2) is continuous, Hurwitz and periodic then
the system has bounded solutions for all times t < ∞.
Proof. The solution of (3.2) exists, because regularity conditions are assumed to be
satisﬁed, and can be written as
x(t) = x
0
+
_
t
0
A(x(s))x(s)ds. (3.17)
3.5. Periodicity and Global Asymptotic Stability 29
Since the system matrix A(x) is continuous, Hurwitz, and periodic for all x ∈ Ω, it
must be bounded. Therefore, ∃M > 0, s.t. ¦¦A(x)¦¦ < M, ∀x ∈ R
n
. Taking norm on
both sides of (3.17) and using the fact that A(x) is bounded, it follows that
¦¦x(t)¦¦ ≤ ¦¦x
0
¦¦ +
_
t
0
M¦¦x(s)¦¦ds.
Using Gronwall’s Lemma [80, Chapter 1, page:5], we get
¦¦x(t)¦¦ ≤ ¦¦x
0
¦¦e
Mt
.
This expression shows that the state of the system is bounded for all t < ∞, i.e., the
state does not blow up in ﬁnite time. This completes the proof.
Further research on this topic results in yet another counter example which leads
us to the conclusion that adding the periodicity condition alone does not suﬃce to
guarantee global asymptotic stability. The details are explained in the following sub
section.
3.5.1 Counter Example Showing that the Periodicity of the System
Matrix does not Guarantee Global Asymptotic Stability
Example 3.5.1. This example has two parts: ﬁrst we analyze the system matrix and
then analyze the corresponding system.
Analysis of Coeﬃcient Matrix
In order to present the counter example, we ﬁrst consider the matrix
A(x
1
) =
_
2α −2
2 + 4α
2
−4α
_
, (3.18)
where α = sin
2
x
1
+ε, with ε ∈ R such that 0 < ε < 1. Hence, α is a real number such
that 0 < α ≤ 1 + ε < 2 for any x
1
∈ R.
Note that the matrix A(x
1
) is periodic with period π, i.e., A(x
1
) = A(x
1
+ π) for all
x
1
∈ R. Further, note that A(x
1
) depends on x
1
in a continuous way.
With ε as above, the eigenvalues of A(x
1
) are given by
λ
1,2
= −α ± i
√
4 − α
2
,
where “i” denotes the imaginary unit. Clearly, the eigenvalues of A(x
1
) are complex,
since 4 − α
2
> 0. Furthermore, the eigenvalues of A(x
1
) are located in the open left
half of the complex plane, since α > 0. Hence, for all x
1
∈ R, the matrix A(x
1
) is a
socalled Hurwitz matrix, i.e., all eigenvalues of A(x
1
) have a negative real part.
An eigenvector for the eigenvalue λ
1
= −α + i
√
4 − α
2
is given, for instance, by
the nonzero complexvalued vector
v
1
=
_
2
3α
_
+ i
_
0
−
√
4 − α
2
_
.
30 Chapter 3. SDC Parametrization and Stability Analysis
It can indeed be veriﬁed that A(x
1
)v
1
= λ
1
v
1
. Note that A(x
1
) is a realvalued
matrix, but that A(x
1
)v
1
and λ
1
v
1
are complexvalued expressions. Then, equating the
real and imaginary parts of A(x
1
)v
1
and λ
1
v
1
, it follows that
A(x
1
)
_
2
3α
_
= −α
_
2
3α
_
−
√
4 − α
2
_
0
−
√
4 − α
2
_
and
A(x
1
)
_
0
−
√
4 − α
2
_
=
√
4 − α
2
_
2
3α
_
− α
_
0
−
√
4 − α
2
_
.
Combining the expressions, it follows that
A(x
1
)
_
2 0
3α −
√
4 − α
2
_
=
_
2 0
3α −
√
4 − α
2
_ _
¸
¸
¸
¸
_
−α
√
4 − α
2
−
√
4 − α
2
−α
_
¸
¸
¸
¸
_
,
or
A(x
1
)T = TD,
with
T =
_
2 0
3α −
√
4 − α
2
_
, D =
_
¸
¸
¸
¸
_
−α
√
4 − α
2
−
√
4 − α
2
−α
_
¸
¸
¸
¸
_
.
By the restrictions on ε and α, it is clear that matrix T is invertible. Hence, it
follows that
A(x
1
) = TDT
−1
,
and, consequently, that
e
A(x
1
)t
= Te
Dt
T
−1
. (3.19)
Because of the special form of the matrix D, it follows (see any book on diﬀerential
equations such as [9, page. 332]) that for all t ≥ 0,
e
Dt
= e
−αt
_
¸
¸
¸
¸
_
cos
√
4 − α
2
t sin
√
4 − α
2
t
−sin
√
4 − α
2
t cos
√
4 − α
2
t
_
¸
¸
¸
¸
_
.
The matrix on the righthand side in the above equation has a ﬁnite norm. Indeed,
considering the Frobenius norm, i.e., the square root of the sum of squared moduli of
all matrix elements, see [31], it follows that the Frobenius norm of the real matrix
_
¸
¸
¸
¸
_
cos
√
4 − α
2
t sin
√
4 − α
2
t
−sin
√
4 − α
2
t cos
√
4 − α
2
t
_
¸
¸
¸
¸
_
,
is equal to
√
2. The Frobenius norm of e
Dt
is then given by
√
2e
−αt
. The Frobenius
norms of the matrices T and T
−1
are given by
√
8 + 8α
2
and
_
2+2α
2
4−α
2
, respectively.
Let ¦¦ ¦¦
F
denote the Frobenius norm. It is wellknown (or easy to prove) that the
Frobenius norm is submultiplicative, i.e., ¦¦UV¦¦
F
≤ ¦¦U¦¦
F
¦¦V¦¦
F
for any two square
matrices U and V of the same size. Therefore, it follows from (3.19) that for all t ≥ 0,
¦¦e
A(x
1
)t
¦¦
F
≤ ¦¦T¦¦
F
¦¦e
Dt
¦¦
F
¦¦T
−1
¦¦
F
≤
_
8 + 8(1 + ε)
2
√
2e
−αt
_
2 + 2(1 + ε)
2
4 − (1 + ε)
2
.
3.5. Periodicity and Global Asymptotic Stability 31
Since α > 0, it follows that 0 ≤ e
−αt
≤ 1, for all t ≥ 0. Hence, for all t ≥ 0,
¦¦e
A(x
1
)t
¦¦
F
≤ 2
√
2
(2 + 2(1 + ε)
2
)
_
4 − (1 + ε)
2
.
So, there exists a number M ∈ R, depending on ε, such that ¦¦e
A(x
1
)t
¦¦
F
≤ M, for all
t ≥ 0 and for all x
1
∈ R.
Recapitulating, the matrix A(x
1
) deﬁned in (3.18), with α = sin
2
x
1
+ ε and with ε
a ﬁxed number such that 0 < ε < 1, satisﬁes the following conditions:
• A(x
1
) is Hurwitz for all x
1
∈ R, i.e., all eigenvalues of A(x
1
) have negative real
part for all x
1
∈ R,
• A(x
1
) depends continuously on x
1
,
• e
A(x
1
)t
is “uniformly” bounded in t ≥ 0 and x
1
∈ R, i.e., there exists an M ∈ R
such that ¦¦e
A(x
1
)t
¦¦
F
≤ M, for all t ≥ 0 and x
1
∈ R.
In [44], it is claimed that the above conditions are suﬃcient for global asymptotic
stability of the associated diﬀerential equation. In [53], we showed that this claim is
not correct.
Here, in addition to the above properties, we also have that
• A(x
1
) is periodic with period π, i.e., A(x
1
) = A(x
1
+ π) for all x
1
∈ R.
However, this additional property in general, does not oﬀer anything extra as far as
global asymptotic stability is concerned. This is shown in the following analysis of the
diﬀerential equation.
Analysis of the Diﬀerential Equation
To show that the above conditions are generally not suﬃcient to guarantee global
asymptotic stability, we consider the diﬀerential equation
˙ x = A(x
1
)x, with x =
_
x
1
x
2
_
, (3.20)
where A(x
1
) is as deﬁned in (3.18), with α = sin
2
x
1
+ ε, and with ε a ﬁxed number
such that 0 < ε < 1.
Because A(x
1
) is Hurwitz for all x
1
∈ R, see the previous subsection, A(x
1
) is
invertible for all x
1
∈ R. Therefore, diﬀerential equation (3.20) has only one equi
librium point, namely the origin. Moreover, since A(x
1
) is Hurwitz for x
1
= 0, it fol
lows that the linearization of (3.20) around the origin, given by ˙ x = A(0)x, is locally
asymptotically stable. It is then well known, see for instance [80], that there exists a
neighborhood of the origin such that solutions of (3.20), starting in this neighborhood,
converge to the origin for time going to inﬁnity. However due to the nonlinearity of
(3.20), and often also in general, there are initial conditions for which this conver
gence does not take place. Indeed, as numerical simulations show, there are initial
conditions such that the resulting solution does not converge.
32 Chapter 3. SDC Parametrization and Stability Analysis
More speciﬁcally, for = 0.05, clearly satisfying 0 < < 1, starting from x
1
(0) =
1, x
2
(0) = 0, the solution of the diﬀerential equation spirals outwards. It further
more follows from simulations that for x
1
(0) = 1.5, x
0
= 0, the resulting solution
spirals inwards, but does not converge to the origin. As a consequence, by a Poincar´ e
Bendixson kind of reasoning, see [80], both the solutions converge to a limit cycle that
intersects the x
1
−axis between 1 and 1.5. See also Figure 3.2, where the solutions are
displayed starting from x
1
(0) = 1, x
2
(0) = 0 and x
1
(0) = 1.5, x
0
= 0, respectively.
In Figure 3.2, the limit cycle to which the two solutions converge is depicted by a
dashdotted line. This limit cycle intersects the positive x
1
−axis around x
1
= 1.265.
Figure 3.2 also contains a second smaller limit cycle, also depicted by a dashdotted
line. This second limit cycle intersects the positive x
1
−axis around x
1
= 0.460. It
bounds the neighborhood around the origin in which solutions converge to the origin.
Outside this neighborhood, the solutions converge towards the larger limit cycle.
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
xy−trajectories
State x
1
S
t
a
t
e
x
2
Figure 3.2: Phase portrait of the system dynamics (3.20).
Figure 3.2 indicates that the solution starting in certain initial states never reaches
the only equilibrium state, as time goes to inﬁnity, although the coeﬃcient matrix
satisﬁes the conditions in [44]. Hence, the origin, being the only equilibrium state, is
not globally asymptotically stable. Also it is clear that the additional requirement of
periodic state dependency of the coeﬃcient matrix does not help in attaining global
asymptotic stability.
3.5. Periodicity and Global Asymptotic Stability 33
Concluding Remarks
From the foregoing discussion, it is clear that the conditions C.1, C.2, and C.3 are not
suﬃcient to guarantee global asymptotic stability of a nonlinear system written as a
pseudolinear system in state dependent coeﬃcient form (3.2), even with the periodic
ity condition as well, see [53].
In general, periodicity of the system matrix does not imply global asymptotic sta
bility of (3.2). However, this assumption helps in analyzing its global asymptotic
stability as we have seen in Example (3.5.1). For instance, we assume that the system
matrix A(x) is a function of a single state component and it is θ−periodic as well. Then
(3.16) becomes
A(x
1
+ θ) = A(x
1
), for all x
1
∈ and θ ∈ R, (3.21)
where, without loss of generality, x
1
is the ﬁrst component of the state vector.
The eigenvalue criterion is a wellknown approach to analyze asymptotic stability
of linear systems. It is stated as follows [13, Chapter 7, page:203].
Theorem 3.5.1. Eigenvalue Criterion for Stability. The system
˙ x = Ax, x(0) = x
0
, (3.22)
is asymptotically stable if and only if all the eigenvalues of the matrix A have negative
real parts.
Before the advent of fast computing machines like modern age computers, compu
tation of eigenvalues was a cumbersome task. In 1892, a Russian mathematician, A.
M. Lyapunov (18571918) proposed a method to analyze asymptotic stability of linear
systems. This method does not require eigenvalue computation. Lyapunov narrated
his method as follows [13, Chapter 7, page:205].
Theorem 3.5.2. Lyapunov Stability Theorem. The system (3.22) is asymptotically
stable if and only if, for any symmetric positive deﬁnite matrix Q, there exists a unique
symmetric positive deﬁnite matrix P satisfying the equation:
PA + A
T
P = −Q. (3.23)
In (3.23), we can easily transform the matrix equation into a matrix inequality to
get
PA + A
T
P ≺ 0, ∵ −Q ≺ 0. (3.24)
For a given N × N matrix P, the constraint P(¸) > 0 would denote that the matrix P is
positive (semi)deﬁnite and the constraint P(¸) ≺ 0 would denote that the matrix P is
negative (semi)deﬁnite.
34 Chapter 3. SDC Parametrization and Stability Analysis
3.5.2 An LMI Based Approach for Global Asymptotic Stability
In this subsection, an LMI based approach is proposed and discussed to establish
global asymptotic stability of the pseudolinear systems of the form (3.2) when the
system matrix has the form (3.21). The motivation for this approach comes from
[66, Chapter 4] where asymptotic stability of linear systems is analyzed subject to
parameter uncertainties. We see a similar situation in (3.21) where the system matrix
depends on a single state component which varies periodically with period θ.
This approach has two steps. The ﬁrst step is to formulate an LMI feasibility
problem using the concept introduced in the Theorem 3.5.2. The solution of this LMI
feasibility problem gives a constant symmetric positive deﬁnite matrix. In the second
step, we use the solution matrix from the LMI feasibility problem (Step I) to prove
the global asymptotic stability of the system under consideration by using a quadratic
Lyapunov function. The details of this approach are as follows.
Step I LMI Feasibility Problem Formulation
Inspired by the Lyapunov Stability Theorem, we propose an equivalent result for the
pseudolinear system (3.2). Before proceeding further, it is important to mention that
the LMI solver SeDuMi which we use in the sequel, does not distinguish between >
and ¸ (likewise between ≺ and ¸). Therefore, we formulate the LMI feasibility prob
lem with nonstrict equality constraints. The system (3.2) is pointwise asymptotically
stable if and only if for each x
1
∈ R, ∃ P(x
1
) = P
T
(x
1
) ¸ 0 s.t.
A
T
(x
1
)P(x
1
) + P(x
1
)A(x
1
) + I
n
¸ 0. (3.25)
It is important to mention that, without loss of generality, we take Q = I
n
as a constant
identity matrix. The choice Q = I
n
helps us in stability analysis by ensuring strict
inequalities, as we will see in the sequel. Intuitively, we are interested to investigate
whether there exists a constant matrix P = P
T
¸ 0 for which the following holds
A
T
(x
1
)P + PA(x
1
) + I
n
¸ 0, ∀ x
1
∈ R. (3.26)
Therefore, if the LMI in (3.26) holds true for all x
1
in the principal interval [−
θ
2
,
θ
2
]
then it will hold for all x
1
∈ R.
To verify this we proceed as follows. We mark k (suﬃciently large) number of
points on [−
θ
2
,
θ
2
) and formulate an LMI feasibility problem as follows:
Does there exist a P = P
T
> 0 which satisﬁes the LMI in (3.26) at k points, x
1
k
∈
[−
θ
2
,
θ
2
)?
This gives rise to k +1 LMIs. The system matrix A(x
1
) is θperiodic which means that
it will be same at x
1
= −
θ
2
and x
1
=
θ
2
. Hence, to ensure that all LMIs are distinct,
we take x
1
k
∈ [−
θ
2
,
θ
2
). Mathematically, we formulate the LMI feasibility problem as
follows:
3.5. Periodicity and Global Asymptotic Stability 35
Does there exist a symmetric matrix P which satisﬁes the following set of LMIs
P − I
n
¸ 0
A
T
(x
11
)P + PA(x
11
) + I
n
¸ 0
A
T
(x
12
)P + PA(x
12
) + I
n
¸ 0
.
.
.
A
T
(x
1k
)P + PA(x
1k
) + I
n
¸ 0
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
? (3.27)
We solve this LMI feasibility problem with YALMIP package (see [46] for more de
tails) in MATLAB
T M
using the solver SeDuMi (see [73] for more details).
Step II Global Asymptotic Stability Analysis
Here we explain how we use the solution of the LMI feasibility problem to estab
lish global asymptotic stability of the system under consideration. Before proceeding
further, we recall the important BarbashinKrasovskii theorem from [40, Chapter 4,
page:124] which we use in the stability analysis.
Theorem 3.5.3. Let ¯ x = 0 be an equilibrium point for (3.2). Let V : Ω −→ R be a
continuously diﬀerentiable function such that
V(0) = 0 and V(x) > 0, ∀x 0, (3.28)
¦¦x¦¦ −→ ∞ ⇒ V(x) −→ ∞, (3.29)
˙
V(x) < 0, ∀x 0, (3.30)
then ¯ x = 0 is globally asymptotically stable.
We deﬁne a quadratic Lyapunov function as:
V(x) = x
T
Px. (3.31)
Here P = P
T
> 0 is the solution of the LMI feasibility problem (3.27). Diﬀerentiating
(3.31) w.r.t ‘t’
˙
V(x) = x
T
[A
T
(x
1
)P + PA(x
1
)]x. (3.32)
The function in (3.31) satisﬁes all requirements stated in Theorem 3.5.3 provided
A
T
(x
1
)P + PA(x
1
) ¸ −I
n
≺ 0, ∀x
1
∈ R.
Therefore, global asymptotic stability of the system (3.2) can immediately be estab
lished if the LMI feasibility problem, (3.27), has a feasible solution. We illustrate our
approach by the following example.
36 Chapter 3. SDC Parametrization and Stability Analysis
Example 3.5.2. We consider a fourth order nonlinear system of the form (3.2). We
begin by deﬁning a system matrix of the following form.
¯
A(x
1
) = A
0
+ A
1
cos x
1
+ B
1
sin x
1
. (3.33)
We generate the constant matrices A
0
, A
1
, and B
1
at random. In particular, we ran
domly generate A
0
as a diagonal matrix and A
1
and B
1
as full matrices. Thus we form
a system matrix which is continuous and periodic. To obtain a continuous, periodic,
and Hurwitz matrix, we deﬁne,
β
= (1 + δ)λ
max
(
¯
A(x
1
)), ∀x
1
∈ [−π, π], (3.34)
where δ > 0 is a small number and λ
max
denotes the maximum real part of the eigen
values of
¯
A(x
1
). Then the matrix
A(x
1
) =
¯
A(x
1
) − βI, (3.35)
where I is the identity matrix of the size of
¯
A(x
1
), will be Hurwitz. Figure 3.3 shows
the behavior of the state dynamics of a system of the form (3.2), having system matrix
(3.35) and A
0
, A
1
, and B
1
are as given later in this example, with four diﬀerent arbi
trarily chosen initial conditions. We see that the states converge asymptotically to the
origin. This indicates that the states of the system may converge asymptotically to the
origin for all initial conditions x
0
= (x
1
, x
2
, x
3
, x
4
) ∈ R
4
.
In retrospect of the theory discussed earlier, the ﬁrst step is to formulate an LMI
feasibility problem. We present three diﬀerent results each with a diﬀerent choice of the
number of LMIs in the feasibility problem. The system matrix (3.35) is 2π− periodic.
We mark (k =)120, 300, and 440 points on the interval [−π, π). In this way, we form a
system of 121, 301, and 441 LMIs analogous to (3.27). As has been mentioned earlier,
we generate the set of matrices A
0
, A
1
, and B
1
at random. We see in our experiments
that for all such sets of matrices, the state dynamics of the nonlinear system, with
arbitrarily chosen initial states, converge asymptotically to the origin. In a particular
set, we have the following matrices
A
0
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
0.3146 0 0 0
0 0.8596 0 0
0 0 0.1287 0
0 0 0 0.0166
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
,
A
1
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
−0.0728 0.9884 0.6450 0.2224
−0.9943 −0.5990 −1.3099 1.8713
−0.7474 1.4766 −0.8674 0.1100
−0.0308 −0.8138 −0.4742 −0.4113
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
,
B
1
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
0.5112 −0.2958 1.6777 0.3630
−1.1991 −0.1680 1.9969 −0.5670
−0.0964 0.1795 0.6970 −1.0442
0.4458 0.4211 −1.3664 0.6971
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
.
3.5. Periodicity and Global Asymptotic Stability 37
0 2 4 6
−4
−2
0
2
4
S
t
a
t
e
s
(a)
State x
1
State x
2
State x
3
State x
4
0 5 10
−4
−2
0
2
4
6
S
t
a
t
e
s
(b)
State x
1
State x
2
State x
3
State x
4
0 1 2 3 4 5
−5
0
5
10
Time [sec]
S
t
a
t
e
s
(c)
State x
1
State x
2
State x
3
State x
4
0 5 10
−8
−6
−4
−2
0
2
4
6
8
Time [sec]
S
t
a
t
e
s
(d)
State x
1
State x
2
State x
3
State x
4
Figure 3.3: Behavior of the state dynamics of the system (3.2)(3.35) with four
arbitrary initial conditions (a).(
π
5
, 3, −4, 5), (b).(
2π
5
, −2, 6, −4), (c).(−
π
2
, −3, −7, 9),
(d).(−
3π
4
, −6, 6, 7).
We solve the three LMI feasibility problems consisting of 121, 301, and 441 LMIs
and we get the following feasible solutions, respectively. For k = 120, in 9 iterations,
SeDuMi yields
P
1
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
3.4727 −0.0050 −0.5492 0.0835
−0.0050 2.8798 −0.4821 −0.0967
−0.5492 −0.4821 4.1219 0.1140
0.0835 −0.0967 0.1140 3.7498
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
.
The eigenvalues of the matrix P
1
are 2.6626, 3.2274, 3.7827, and 4.5515. Since all
eigenvalues are positive, P
1
is a SPD matrix. For k = 300, in 7 iterations, SeDuMi
yields
P
2
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
4.3305 −0.0830 −0.6174 0.1847
−0.0830 3.5683 −0.4611 −0.0278
−0.6174 −0.4611 5.0721 −0.0345
0.01847 −0.0278 −0.0345 4.6450
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
.
The eigenvalues of the matrix P
2
are 3.3606, 4.0759, 4.6743, and 5.5051. Since all
eigenvalues are positive, P
2
is a SPD matrix. For k = 440, in 7 iterations, SeDuMi
38 Chapter 3. SDC Parametrization and Stability Analysis
yields
P
3
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
8.8841 −0.2495 −1.1634 0.4852
−0.2495 7.3402 −0.8016 0.0303
−1.1634 −0.8016 10.3775 −0.2747
0.4852 0.0303 −0.2747 9.5379
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
.
The eigenvalues of the matrix P
3
are 6.9676, 8.4366, 9.5000, and 11.2355. Since all
eigenvalues are positive, P
3
is a SPD matrix.
Now, a quadratic Lyapunov function V(x) = x
T
P
i
· x, i
·
= 1, 2, 3 satisﬁes all the
conditions for global asymptotic stability stated in Theorem 3.5.3. Therefore, this
fourth order nonlinear autonomous system is globally asymptotically stable.
Figure 3.4 shows the maximum eigenvalue proﬁles of A
T
(x
1
i
)P
i
· + P
i
· A(x
1
i
) for
all x
1
i
∈ [−π, π), i = 1, 2, ..., k vs x
1
corresponding to LMI systems consisting of
121, 301 and 441 constraints. We see that the curves lie below the horizontal axis.
This implies that the symmetric matrices A
T
(x
1
i
)P
i
· + P
i
· A(x
1
i
), ∀ x
1
i
∈ [−π, π], are
negative deﬁnite and the pattern of the eigenvalues proﬁles shows that they become
more negativedeﬁnite as the size of the system of LMIs grows.
−4 −3 −2 −1 0 1 2 3 4
−35
−30
−25
−20
−15
−10
−5
0
x1
λ
m
a
x
Profiles of maximum eigenvalues
k=120
k=300
k=440
Figure 3.4: Proﬁles of the maximum eigenvalues of the symmetric matrices
A
T
(x
1i
)P
i
· + P
i
· A(x
1i
), ∀ x
1
i
∈ [−π, π], i = 1, 2, ..., k and i
·
= 1, 2, 3 vs state com
ponent x
1
.
3.5. Periodicity and Global Asymptotic Stability 39
The above described LMI based Lyapunov method to prove global asymptotic
stability of the systems of the form(3.2), depends on the solution of the LMI feasibility
problem. If the LMI feasibility problem does not have a feasible solution then we
can not directly deduce global asymptotic stability with this method. Although, this
method may still be useful with a naive modiﬁcation in the LMI constraints. The
following subsection explains this situation.
3.5.3 Infeasibility of the LMI Feasibility Problem
Example 3.5.3. Consider the following autonomous nonlinear system
_
˙ η
˙ ν
_
=
_
O
3
J(ψ)
O
3
−M
−1
D
_ _
η
ν
_
+
_
O
3
M
−1
_
τ. (3.36)
This model has been introduced in Chapter 2 where more details about the model can
be found.
We see that the system matrix in (3.36) is not Hurwitz. But it is a stabilizable
system. We use the following stabilizing state feedback control law
τ =
_
−MJ
T
(ψ) D − 2M
_
_
η
ν
_
. (3.37)
This results in a continuous, Hurwitz, and 2π−periodic feedback system matrix given
by
A
c
(ψ) =
_
O
3
J(ψ)
−J
T
(ψ) −2I
3
_
. (3.38)
Figure 3.5 shows the state dynamics of the closed loop system (3.36)(3.37) with four
diﬀerent arbitrarily chosen initial conditions. Clearly, the states converge asymptoti
cally to the origin.
We formulate an LMI feasibility problem as described in (3.27) by using the closed
loop system matrix (3.38). The outcome from the LMI solver SeDuMi is: the LMI fea
sibility problem is infeasible. This means, the LMI based Lyapunov stability approach
is not working in this case.
Remark 3.5.1. Consider a naive modiﬁcation in the inequality constraints (3.27)
which results the following form
A
T
c
(ψ)P + PA
c
(ψ) +
_
O
3
O
3
O
3
I
3
_
¸ 0. (3.39)
We get a feasible solution for this LMI feasibility problem from SeDuMi. But we
cannot use Theorem 3.5.3 because
˙
V(x) is only positive semideﬁnite in this case. We
require LaSalle’s Principle for stability analysis in such situations.
Remark 3.5.2. A more practical and useful form of (3.37) can be
τ =
_
−MJ
T
(ψ) F − 2M
_
_
η
ν
_
, (3.40)
where F denotes the additional damping eﬀect introduced via controller action.
40 Chapter 3. SDC Parametrization and Stability Analysis
0 5 10
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
S
t
a
t
e
s
(a)
0 5 10
−2
−1
0
1
2
S
t
a
t
e
s
(b)
0 5 10
−3
−2
−1
0
1
Time [sec]
S
t
a
t
e
s
(c)
0 5 10
−10
−5
0
5
Time [sec]
S
t
a
t
e
s
(d)
x
y
ψ
u
v
r
x
y
ψ
u
v
r
x
y
ψ
u
v
r
x
y
ψ
u
v
r
Figure 3.5: Behavior of the state dynamics of the system in Example
3.5.3 with four arbitrary initial conditions (a).(2, 2,
π
6
, 0, 0, 0), (b).(2, −2,
π
3
, 0, 0, 0),
(c).(−1, −1, −
8π
9
, 0, 0, 0), (d).(−10, 4, −
2π
3
, 4, −3, 4).
To investigate the infeasibility issue in more detail, we approach the LMI feasibil
ity problem associated with the system in Example 3.5.3, analytically, in the following
subsection.
Analytic approach to the LMI problem
For the closed loop system matrix (3.38), we would like to ﬁnd a SPD matrix P s.t.
A
T
c
(ψ)P + PA
c
(ψ) ≺ 0, ∀ψ ∈ [−π, π].
Suppose that P has the following form
P =
_
P
a
P
T
b
P
b
P
c
_
.
For the sake of convenience, from here onward in this subsection, we omit the argu
ment ψ from J(ψ). Now consider,
A
T
c
P + PA
c
≺ 0,
3.5. Periodicity and Global Asymptotic Stability 41
⇐⇒
_
O
3
−J
J
T
−2I
3
_ _
P
a
P
T
b
P
b
P
c
_
+
_
P
a
P
T
b
P
b
P
c
_ _
O
3
J
−J
T
−2I
3
_
≺ 0,
⇐⇒
_
−JP
b
− (JP
b
)
T
P
a
J − 2P
T
b
− JP
c
(P
a
J − 2P
T
b
− JP
c
)
T
(P
b
J)
T
+ P
b
J − 4P
c
_
≺ 0. (3.41)
Or equivalently,
⇐⇒ −
_
−JP
b
− (JP
b
)
T
P
a
J − 2P
T
b
− JP
c
(P
a
J − 2P
T
b
− JP
c
)
T
(P
b
J)
T
+ P
b
J − 4P
c
_
> 0. (3.42)
To be brief, we focus on the leading principal block diagonal entry
−JP
b
− (JP
b
)
T
.
The inequality 3.41 holds true only if
JP
b
+ (JP
b
)
T
> 0, ∀ψ ∈ [−π, π]. (3.43)
In the sequel, we show that the necessary condition (3.43) for (3.42) to hold true, does
not hold.
Consider P
b
in the following form
P
b
=
_
P
b
11
∗
∗ ∗
_
,
where P
b
11
is a 2 × 2 block entry while the entries of P
b
which are not relevant for the
analysis are marked by aestericks. Since (3.43) must be true for all ψ ∈ [−π, π], we
take the following two cases:
Case 1. ψ = 0,
JP
b
+ (JP
b
)
T
=
_
P
b
11
+ P
T
b
11
∗
∗ ∗
_
.
Case 2. ψ = π,
JP
b
+ (JP
b
)
T
=
_
−P
b
11
− P
T
b
11
∗
∗ ∗
_
.
Now the inequality (3.43) holds only if the following holds
P
b
11
+ P
T
b
11
> 0 and − P
b
11
− P
T
b
11
> 0.
This is an improbable scenario. Therefore, about the LMI feasibility problem of Ex
ample 3.5.3, we conclude the following
“ P = P
T
> 0 s.t. A
T
c
(ψ)P + PA
c
(ψ) ≺ 0, holds true for all ψ ∈ [−π, π].”
In the foregoing discussion, we have proved analytically that the LMI problem
corresponding to the system (3.36) has no solution. In the following, we continue
with global stability analysis of the system (3.36) by using Lyapunov stability theory
and LaSalle’s invariance theorem.
42 Chapter 3. SDC Parametrization and Stability Analysis
3.5.4 Further Analysis for Global Asymptotic Stability
Before proceeding further, we recall an important result from stability theory which
was ﬁrst formulated by a famous French priest and education reformer, J. B. de LaSalle
(16511719). It is stated as follows [40, Chapter 4, page:128]:
Theorem 3.5.4. The LaSalle’s Invariance Theorem (Principle). Let Ω
0
⊂ Ω be a
compact set that is positively invariant with respect to (3.2). Let V : Ω −→ R be a
continuously diﬀerentiable function such that
˙
V(x) ≤ 0 in Ω
0
. Let E be the set of all
points in Ω
0
where
˙
V(x) = 0. Let M be the largest invariant set in E. Then every
solution starting in Ω
0
approaches M as t → ∞.
An invariant set is deﬁned as:
Deﬁnition 3.5.1. (Invariant Set) A set M is an invariant set for a dynamic system
if every system trajectory which starts from a point in M remains in M for all future
time.
The system matrix (3.38) has a special structure with a skew symmetric part. We
try to use this structure to analyze global asymptotic stability of the system. Let us
consider, a Lyapunov function of the form
V(x) = x
T
x. (3.44)
Diﬀerentiating the Lyapunov function, we get
˙
V(x) = x
T
_
O
3
J(ψ)
−J
T
(ψ) −2I
3
_
x,
⇒
˙
V(x) = −4ν
T
ν,
⇒
˙
V(x) ≤ 0, ∀ x 0.
Now,
˙
V(x) = 0, ⇒ ν = 0, and η is arbitrary.
Therefore,
E = ¦x :
˙
V(x) = 0¦
= ¦ν = 0, η is free¦.
The next step is to ﬁnd the largest invariant subset of E
l
. Suppose η(t) and ν(t) is a
solution pair of the closed loop system (3.36) and (3.37) that belongs identically to E
l
.
Then
ν(t) ≡ 0, ⇒ ˙ ν(t) ≡ 0,
⇒ −J
T
(ψ)η(t) ≡ 0,
⇒ η(t) ≡ 0. (3.45)
Therefore, from (3.45), we conclude that the only solution of the closed loop sys
tem (3.36) and (3.37) which stays identically in E
l
, is the trivial solution. Hence
(η, ν) = (0, 0) is a globally asymptotically stable equilibrium point of the the closed
loop system (3.36) and (3.37) .
3.6. Summary and Conclusions 43
3.6 Summary and Conclusions
This chapter focuses on the stability analysis of a special type of nonlinear systems.
These nonlinear systems when transformed to pseudolinear form by using the SDC
framework, the system matrix in new representation depends on a single state variable.
Local asymptotic stability is easy to analyze but the major topic of interest in this
chapter is global asymptotic stability analysis. We start with a brief review of the
past results on this topic. The main contributions are two counter examples. Example
3.4.1 shows that the exponential boundedness of the system matrix together with the
continuity and Hurwitzness conditions does not imply global asymptotic stability of
the nonlinear dynamical system (3.2). Example 3.5.1 shows that, in addition, even if
the periodicity of the system matrix is assumed, it does not ensure global asymptotic
stability of (3.2).
However, as a consequence of the periodicity of the system matrix, we introduce
another method to prove the global asymptotic stability of the pseudolinear system
(3.2). This method combines the concepts of the Lyapunov stability theory and the
LMIs. This approach is motivated from the idea of LPV approach introduced in [66]
to the pseudolinear systems of the form (3.2). We use the periodicity of the system
matrix together with the Lyapunov stability theory and the LMIs to formulate an LMI
feasibility problem. This LMI feasibility problem is solved by using the YALMIP
interface under MATLAB. YALMIP uses SeDuMi (an LMI solver) which transforms
the LMI feasibility problem into a convex optimization problem. The positive (semi
)deﬁnite solution of the LMI feasibility problem is then used to get a quadratic Lya
punov function. Global asymptotic stability is then followed by Lyapunov stability
theory.
We ﬁrst explained this method by using an academic example of a fourth order
pseudolinear system. Afterwards, we proceed to prove global asymptotic stability of
the nominal vessel model (3.36) subject to a naive state feedback controller (3.37). It
turns out that the LMI feasibility problem associated with the closed loop system ma
trix (3.38) does not have a feasible solution. This outcome is then proved analytically.
More investigation to this bottleneck revealed that a small modiﬁcation in the LMI
feasibility problem, see (3.39), resolves this problem. The consequence of this mod
iﬁcation is that the derivative of the Lyapunov function will be positive semideﬁnite
in this situation. In this case, the LaSalle’s invariance principal is used to establish
global asymptotic stability of the pseudolinear system.
44
Chapter 4
State Dependent Riccati Equation
based Control Design
1
T
ne use of DP systems for the positioning control of ships started in the early sixties.
The ﬁrst DP control system used the conventional PID type control law together
with a notch ﬁlter, to control each of the surge, sway, and yaw motions. The purpose of
the notch ﬁlter was to ﬁlter out the eﬀects of the WF motions from the feedback loop.
There were two main disadvantages of this approach: the integral action of the PID
controller is very low due to the coupling between the surge, sway, and yaw motions
and there is a phase lag in the control loop due to the notch ﬁlter.
The disadvantages of the PIDcontrol lawmotivated a Norwegian company, Kongs
berg, to initiate eﬀorts to invent new DP control systems using the concepts of optimal
control theory and Kalman ﬁltering from the modern control theory. These eﬀorts
made way for the modern concepts for wave ﬁltering like state estimators or observers.
Observers are more useful than the classical ﬁltering techniques because they also give
estimates of the unknown states of the system.
Observer based wave ﬁltering techniques have been reported by many researchers,
for instance see [1], [27], and [47]. One such approach is reported in [69] and it
is based on the state dependent algebraic Riccati equation, also known as SDARE.
An SDARE observer is proposed to estimate the states of a DP vessel model. This
observer was then separately used in combination with a backstepping controller [69]
and a PID controller [70].
In this chapter, we extend the ideas of LQR theory to address the nonlinear control
design and estimation problems for the DP vessel. The nonlinear model of the vessel
motion described in Chapter 2 can be assumed as a pseudolinear system by consid
ering the SDC parametrization which is discussed in Chapter 3. Our aim is to study
the SDARE controller and observer design techniques in connection with the control
design problem of a DP vessel. The state dependent algebraic Riccati equation plays
1
Part of this work has been published in the proceedings of the UKACC International Conference on
Control (CONTROL 2010), see [52].
45
46 Chapter 4. SDARE based control Design
a fundamental role in this theory. We will see that the SDARE controller is a state
feedback or an output feedback controller.
4.1 Optimal Control Problem and the SDARE
We mimick the LQR problem for the linear systems to address the control design
problem of the pseudolinear vessel system. Let us start by deﬁning the optimal con
trol (OC) or linear quadratic regulation (LQR) problem for the pseudolinear system
describing the sea vessel motion. We adapt the deﬁnition of the LQR problem from
[13] to customize it according to the pseudolinear model under consideration.
Deﬁnition 4.1.1. (LQR/OC problem for pseudolinear system of vessel model)
Given matrices Q and R, ﬁnd a control signal u(t) such that the quadratic cost function
J(x
0
, u(t)) =
1
2
_
∞
t
0
(x
T
(t)Qx(t) + u
T
(t)Ru(t))dt, (4.1)
is minimized, subject to
˙ x = A(ψ)x(t) + Bu(t), x(t
0
) = x
0
. (4.2)
In (4.1) and (4.2), the ndimensional vector x ∈ R
n
describes the state of the
system, Q = H
T
H ∈ R
n×n
(H
T
∈ R
n×p
) is SPSD matrix, u ∈ R
m
is the control input
vector, R ∈ R
m×m
is SPD matrix, A(ψ) ∈ R
n×n
is the system matrix, and the matrix
B ∈ R
n×m
describes the controller conﬁguration.
The matrices Q and R in (4.1) are termed as weighting or penalty matrices. These
are also called the tuning parameters of the SDARE controller as the performance of
the SDARE controller can be manipulated by playing around with these matrices. Al
though, we assume Q and R to be constant matrices but they can be state dependent as
well. Physically, the expression x
T
(t)Qx(t) describes the deviation of the state vector
from the origin (equilibrium state) and the expression u
T
Ru describes the control cost.
In [6], it has been proved that the problem deﬁned in Deﬁnition 4.1.1 has a subop
timal solution. The following theorem states this result.
Theorem 4.1.1. (The SDARE controller) Consider a nonlinear system in the SDC
form
˙ x = A(x)x + B(x)u. (4.3)
The SDC form implies that the matrices A(x) and B(x) are continuous in x for all
¦¦x¦¦ ≤ ˆ r, ˆ r > 0, see Section 3.1. Assume further that the matrices A(x), B(x), and Q
are such that the pair (H, A(x)) is detectable and the pair (A(x), B(x)) is stabilizable
for all x ∈ Ω ⊆ J
ˆ r
(0), where Ω is a nonempty neighborhood of the origin.
Then the system (4.3) with control
u(x) = −K
c
(x)x, (4.4)
where the controller gain is
K
c
(x) = R
−1
B
T
(x)Π
c
(x), (4.5)
4.2. Control System Design for the DP Vessel 47
is locally asymptotically stable. In (4.5), Π
c
(x) is the solution of the SDARE associated
with (4.1) and (4.3), which is given by
Π
c
(x)A(x) + A
T
(x)Π
c
(x) + Q − Π
c
(x)B(x)R
−1
B
T
(x)Π
c
(x) = 0. (4.6)
Proof. See [6].
4.2 Control System Design for the DP Vessel
In this section, we illustrate the SDARE based controller and observer design tech
nique for the control system design of the DP vessel. A DP vessel model is given in
Chapter 2. The SDARE (4.6) in context of the vessel model becomes
Π
c
(ψ)A(ψ) + A
T
(ψ)Π
c
(ψ) + Q − Π
c
(ψ)BR
−1
B
T
Π
c
(ψ) = 0. (4.7)
This is a special SDARE in the sense that only the system matrix depends on state
and this state dependence too is only on a single state component in a periodic way.
Many solutions of (4.7) could be possible, but the one which is positive (semi)deﬁnite
is desired for most of the applications. The suﬃcient conditions for the pointwise
existence of the solution of the SDARE (4.7) as given in [49], are: the pairs (A(ψ), B)
and (H, A(ψ)) are, respectively, pointwise stabilizable and detectable for all ψ ∈ R.
We divide our study into two steps. First, we assume that there is no WF motion
in the measured output and only LF motion is to be regulated. Later on, we assume
that the WF motion component is also present in the measured output.
4.2.1 Nonlinear Regulation Problem for the LF Model
The LF motion as described in Chapter 2 is given by
˙ η =J(ψ)η, (4.8)
˙ ν = − M
−1
Dν + M
−1
J
T
(ψ)b + M
−1
τ, (4.9)
˙
b = − T
−1
b + Ψw
b
, (4.10)
y
b
=C
b
η + v, (4.11)
where η = [x, y, ψ]
T
is position coordinate vector, ν = [u, v, r]
T
is velocity coordinate
vector, b ∈ R
3
describes the environmental disturbances, and τ represents the control
forces and torque. The vectors w
b
and v represent the process and measurement noise,
respectively. They are described by the zeromean Gaussian white noise processes,
i.e., w
b
∼ (0, Q
c,b
) and v ∼ (0, R
c
). We study the LF motion regulation problem by
dividing it in two parts. In the ﬁrst part, we assume that the bias forces are completely
given or their estimate is available. In the second part, we assume that slowly varying
bias forces and moments are unknown and hence should be estimated.
I. Disturbance vector b is known
We assume that the disturbance vector b is known or its estimate
ˆ
b is available and
there is no measurement noise. The LF model in the SDC form may then be written
48 Chapter 4. SDARE based control Design
as
_
˙ η
˙ ν
_
=
_
O
3
J(ψ)
O
3
−M
−1
D
_ _
η
ν
_
+
_
O
3
M
−1
_
u +
_
O
3
M
−1
J
T
(ψ)
_
b, (4.12)
y
b
=
_
I
3
O
3
_
_
η
ν
_
. (4.13)
The corresponding controllability and observability matrices are given by
C
b
=
_
O
3
J(ψ)M
−1
J(ψ)(−M
−1
D)
4
M
−1
M
−1
−M
−1
DM
−1
(−M
−1
D)
5
M
−1
_
, (4.14)
and
O
b
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
I
3
O
3
O
3
J(ψ)
.
.
.
.
.
.
O
3
J(ψ)(−M
−1
D)
4
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
. (4.15)
The matrices deﬁned in (4.14) and (4.15) have full rowand column ranks, respectively.
Therefore, the system (4.12) is both pointwise controllable and observable. Therefore,
by using Theorem 4.1.1, we deﬁne a locally asymptotically stable SDARE controller
for (4.12) as follows
u = −R
−1
M
−1
_
Π
21
(ψ) Π
22
(ψ)
_
_
η
ν
_
− J
T
(ψ)b. (4.16)
The control law (4.16) can be described as a state feedback law with a feedforward
term ‘−J
T
(ψ)b’. Moreover, Π
21
(ψ) and Π
22
(ψ) comes from
Π
c
(ψ) =
_
Π
11
(ψ) Π
12
(ψ)
Π
21
(ψ) Π
22
(ψ)
_
. (4.17)
which is the symmetric solution of the SDARE associated with (4.1) and (4.12). From
(4.13), it is clear that we can only measure the position of the vessel and its velocities
are not available for feedback. But we can also write (4.16) as
u = −R
−1
M
−1
_
Π
21
(ψ) Π
22
(ψ)J
T
(ψ)
_
_
η
˙ η
_
− J
T
(ψ)b. (4.18)
This control law can be seen as PDtype control law with feedforward term ‘−J
T
(ψ)b’.
Stability analysis of the controller (4.16)
The local asymptotic stability of the state feedback SDARE controller follows from
the Theorem 4.1.1. We prove by using the LMI based technique discussed in Chapter
3 that the SDARE feedback controller stabilizes the vessel globally asymptotically
around the origin. We can write the SDARE state feedback controller, (4.16), as
u = −
_
K
1
(ψ) K
2
(ψ)
_
_
η
ν
_
− J
T
(ψ)b. (4.19)
4.2. Control System Design for the DP Vessel 49
The closed loop system (4.12) and (4.19) is given by
_
˙ η
˙ ν
_
=
_
O
3
J(ψ)
−M
−1
K
1
(ψ) ∗
_
.,,.
A
cl
(ψ)
_
η
ν
_
. (4.20)
It is wellknown fact that the SDARE controller results in an asymptotically stable
closed loop matrix. Therefore, the closed loop system matrix (4.20) is pointwise Hur
witz. Using the LMI approach (see Chapter 3 for more details), it turns out that there
exists an SPD matrix P such that
A
T
cl
(ψ)P + PA
cl
(ψ) +
_
0 0
0 I
3
_
≤ 0, ∀ψ ∈ [0, 2π]. (4.21)
Now we consider the quadratic Lyapunov function: V(x) = x
T
Px. Diﬀerentiating
w.r.t. ‘t’, we get
˙
V(x) = x
T
(A
T
cl
(ψ)P + PA
cl
(ψ))x,
≤ x
T
_
0 0
0 −I
3
_
x, using (4.21),
≤ −ν
T
ν,
≤ 0.
Since
˙
V(x) ≤ 0, therefore, we use LaSalle’s invariance principle to further analyze
the stability of the closed loop system. For this we ﬁrst ﬁnd the set of points where
˙
V(x) = 0.
E
l
= ¦x :
˙
V(x) = 0¦,
= ¦ν = 0, η is free¦.
Now we need the invariant sets of the closed loop system. To show that the origin is
the only equilibrium point of the closed loop system, we need to show that no solution
other than the trivial solution stays identically in E
l
, i.e., η(t) ≡ 0 and ν(t) ≡ 0 for all
‘t’.
Let η(t) and ν(t) be a solution pair of the closed loop system (4.20) that belongs
identically to E
l
. Then
ν(t) ≡ 0, ⇒ ˙ ν(t) ≡ 0,
⇒ −M
−1
K
1
(ψ)η(t) ≡ 0. (4.22)
To show that η(t) ≡ 0, we need to prove that K
1
(ψ) is a nonsingular matrix. We prove
this by contradiction. Suppose that it is singular then ∃ v 0 such that
M
−1
K
1
(ψ)v = 0,
⇒
_
O
3
J(ψ)
−M
−1
K
1
(ψ) ∗
_ _
v
0
_
= 0
_
v
0
_
.
50 Chapter 4. SDARE based control Design
This implies that λ = 0 is an eigenvalue of the closed loop system matrix in (4.20).
But it is a wellknown fact that the closed system (4.20) is pointwise asymptotically
stable (all eigenvalues lie in the left half plane). Therefore, it is a contradiction. Hence,
K
1
(ψ) is nonsingular.
Therefore, from (4.22), we conclude that the only solution of the closed loop sys
tem which stays identically in E
l
is the trivial solution. Hence (η
∗
, ν
∗
) = (0, 0) is a
globally asymptotically stable equilibrium point of the the closed loop system (4.20)
which means that the SDARE controller (4.19) globally asymptotically stabilizes the
vessel.
II. Disturbance vector b is unknown
We now assume that the LF disturbances are slowly varying (as described by the ﬁrst
order Markov process
˙
b = −T
−1
b + Ψw
b
) but unknown and should be estimated. The
LF motion in this case, in the SDC form, is described as
_
¸
¸
¸
¸
¸
¸
¸
¸
_
˙ η
˙ ν
˙
b
_
¸
¸
¸
¸
¸
¸
¸
¸
_
=
_
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
J(ψ) O
3
O
3
−M
−1
D M
−1
J
T
(ψ)
O
3
O
3
−T
−1
_
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
_
η
ν
b
_
¸
¸
¸
¸
¸
¸
¸
¸
_
+
_
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
M
−1
O
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
u +
_
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
O
3
Ψ
_
¸
¸
¸
¸
¸
¸
¸
¸
_
w
b
, (4.23)
y
b
=
_
I
3
O
3
O
3
_
_
¸
¸
¸
¸
¸
¸
¸
¸
_
η
ν
b
_
¸
¸
¸
¸
¸
¸
¸
¸
_
+ v. (4.24)
The controllability and observability matrices for (4.23)(4.24) are given by
C
b
=
_
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
J(ψ)M
−1
J(ψ)(−M
−1
D)
7
M
−1
M
−1
−M
−1
DM
−1
(−M
−1
D)
8
M
−1
O
3
O
3
O
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
, (4.25)
and
O
b
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
I
3
O
3
O
3
O
3
J(ψ) O
3
O
3
−J(ψ)M
−1
D J(ψ)M
−1
J
T
(ψ)
.
.
.
.
.
.
.
.
.
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
. (4.26)
Since the matrices M, D, and J
T
(ψ) are nonsingular, therefore, controllability matrix
(4.25) has rank 6. This means that only the states η and ν are controllable. This is
not restrictive as the bias forces are always uncontrollable. The observability matrix
(4.26) has full rank for all ψ ∈ R. An SDARE associated with (4.23) is given by
Π
c
(ψ)A
b
(ψ) + A
T
b
(ψ)Π
c
(ψ) + Q − Π
c
(ψ)B
b
R
−1
B
T
b
Π
c
(ψ) = 0. (4.27)
Using Theorem 4.1.1, a locally asymptotically stable SDARE based suboptimal
control law is given by
u = −R
−1
M
−1
_
Π
21
(ψ) Π
22
(ψ) Π
23
(ψ)
_
_
¸
¸
¸
¸
¸
¸
¸
¸
_
η
ν
b
_
¸
¸
¸
¸
¸
¸
¸
¸
_
. (4.28)
4.2. Control System Design for the DP Vessel 51
The gain of the SDARE controller is deﬁned by
K
c
(ψ) = R
−1
M
−1
_
Π
21
(ψ) Π
22
(ψ) Π
23
(ψ)
_
, (4.29)
where Π
21
(ψ), Π
22
(ψ), and Π
23
(ψ) are obtained fromthe solution of the SDARE (4.27).
The control law (4.28) can be described as a state feedback controller. From (4.24), we
notice that only the LF position and heading can be measured by the sensors. There
fore, a state estimator is required to estimate the complete state. A suboptimal locally
asymptotically stable SDARE controller for the system (4.23)(4.24) will be
u = −R
−1
M
−1
_
Π
21
(
ˆ
ψ) Π
22
(
ˆ
ψ) Π
23
(
ˆ
ψ)
_
_
¸
¸
¸
¸
¸
¸
¸
¸
_
ˆ η
ˆ ν
ˆ
b
_
¸
¸
¸
¸
¸
¸
¸
¸
_
. (4.30)
Since the system (4.23)(4.24) is pointwise observable, the corresponding dual system
will be pointwise controllable. Therefore, we can design a suboptimal locally asymp
totically stable SDARE observer, see [6]. The following theorem states this result.
Theorem 4.2.1. (The SDARE state estimator) Assume that a system in SDC form
˙ x = A(x)x + B(x)u, (4.31)
y = C(x)x, (4.32)
is such that f (x) = A(x)x and
∂f (x)
∂x
j
, j = 1, 2, ..., n are continuous in x for all ¦¦x¦¦ ≤
ˆ r, ˆ r > 0 and x = 0 is a stable equilibrium point of (4.31). Assume further that the
pair (A(x), C(x)) is pointwise detectable and the matrices A(x) and C(x) are locally
Lipschitz for all x ∈ Ω ⊆ J
ˆ r
(0), where Ω is a nonempty neighborhood of the origin.
Then the estimated state given by
˙
ˆ x = A( ˆ x) ˆ x + B( ˆ x)u + K
o
( ˆ x)(y − ˆ y), (4.33)
will converge locally asymptotically to the state. In (4.33), K
o
( ˆ x) is the observer gain
and is given by
K
o
( ˆ x) = Π
o
( ˆ x)C
T
V
−1
. (4.34)
In (4.34), Π
o
( ˆ x) is the solution of the dual SDARE
Π
o
( ˆ x)A
T
( ˆ x) + A( ˆ x)Π
o
( ˆ x) + U − Π
o
( ˆ x)C
T
V
−1
CΠ
o
( ˆ x) = 0, (4.35)
associated with the system (4.31)(4.32).
Proof. See [6].
An SDARE observer for the system (4.23)(4.24) is given by
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
˙
ˆ η
˙
ˆ ν
˙
ˆ
b
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
=
_
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
J(
ˆ
ψ) O
3
O
3
−M
−1
D M
−1
J
T
(
ˆ
ψ)
O
3
O
3
−T
−1
_
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
_
ˆ η
ˆ ν
ˆ
b
_
¸
¸
¸
¸
¸
¸
¸
¸
_
+
_
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
M
−1
O
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
u + K
o
(
ˆ
ψ)(y
η
− ˆ y
η
), (4.36)
ˆ y
η
=
_
I
3
O
3
O
3
_
_
¸
¸
¸
¸
¸
¸
¸
¸
_
ˆ η
ˆ ν
ˆ
b
_
¸
¸
¸
¸
¸
¸
¸
¸
_
. (4.37)
52 Chapter 4. SDARE based control Design
The third term in (4.36) corrects the observer model by a nonlinear feedback of the
diﬀerence between the measured output y
η
and the estimated output ˆ y
η
. The observer
gain K
o
(
ˆ
ψ) in (4.36) is deﬁned by
K
o
(
ˆ
ψ) =
_
¸
¸
¸
¸
¸
¸
¸
¸
_
Π
11
(
ˆ
ψ)
Π
21
(
ˆ
ψ)
Π
31
(
ˆ
ψ)
_
¸
¸
¸
¸
¸
¸
¸
¸
_
V
−1
, (4.38)
where Π
i1
, i = 1, 2, 3 are obtained fromthe solution of the corresponding dual SDARE
of the form (4.35).
So far we have deﬁned a suboptimal locally asymptotically stable nonlinear con
trol law(4.30) and a suboptimal locally asymptotically stable nonlinear observer (4.36).
A combination of both is called a nonlinear compensator. The separation properties
for combination of linear controllers and observers are well deﬁned, see for instance
[55]. The separation principles for the nonlinear compensators are not well established
yet. The following theorem from [6] states one particular result about stability of an
SDARE compensator.
Theorem 4.2.2. (The SDARE state compensator) Assume that the system in SDC
form (4.31)(4.32) is such that f (x) = A(x)x and
∂f (x)
∂x
j
, j = 1, 2, ..., n are continuous
in x for all ¦¦x¦¦ ≤ ˆ r, ˆ r > 0. Assume further that A(x) and B(x) are continuous. Let the
matrices A(x), B(x), and C(x) are chosen such that the pair (A(x), C(x)) is detectable
and the pair (A(x), B(x)) is stabilizable for all x ∈ Ω ⊆ J
ˆ r
(0), where Ω is a nonempty
neighborhood of the origin.
Then (x, e) = (0, 0) is a locally asymptotically stable point. Here, e = x − ˆ x is the
error between the state x from (4.31) and the state estimate ˆ x from (4.33).
Proof. See [6].
A combination of the SDARE controller (4.30) and observer (4.36) is an SDARE
pseudolinear compensator illustrated in a block diagram shown in Figure 4.1 and is
given by
_
˙ x
˙
ˆ x
_
=
_
A
b
(ψ) −B
b
K
c
(
ˆ
ψ)
K
o
(
ˆ
ψ)C
b
A
b
(
ˆ
ψ) − B
b
K
c
(
ˆ
ψ) − K
o
(
ˆ
ψ)C
b
_ _
x
ˆ x
_
+
_
E
b
O
O K
o
(
ˆ
ψ)
_ _
w
b
v
_
w
b
.
(4.39)
4.2.2 Nonlinear Regulation Problem for the Complete Model
Now, we take into account the eﬀect of the wave frequency (WF) motion in the mea
surement. The WF motion in each of the surge, sway, and yaw directions is modeled
as second order harmonic oscillation, see Chapter 2 for more details. The WF motion
model is given by
_
˙
ξ
1
˙
ξ
2
_
=
_
O
3
I
3
−Ω
2
−2ZΩ
_ _
ξ
1
ξ
2
_
+
_
O
3
Σ
_
w
ξ
, (4.40)
y
ξ
=
_
I
3
O
3
_
_
ξ
1
ξ
2
_
, (4.41)
4.2. Control System Design for the DP Vessel 53
b b b b b b
x A ( ) x B u E w
b b b
y C x
b b b b
ˆ
ˆ ˆ x A ( ) x B u
o b b
ˆ
ˆ
ˆ K ( ) (y ‐ y )
b b b
ˆ ˆ y C x ˆ
K ( )
c
ˆ
‐K ( )
b b
ˆ y ‐ y
b
ˆ x
u
b
x
b
ˆ y

b
C
b
y v
b
C
b
C
b
y
v
Figure 4.1: A nonlinear compensator for LF model
where the diagonal matrix Ω = diag¦ω
01
, ω
02
, ω
03
¦ is such that each of the ω
0i
’s rep
resents the dominating wave frequency and the diagonal matrix Z = diag¦ζ
1
, ζ
2
, ζ
3
¦
is such that each of the ζ
i
’s represents the relative damping ratio in surge, sway, and
yaw direction,respectively. The diagonal matrix Σ = diag¦σ
1
, σ
2
, σ
3
¦ scales the noise
vector w
b
and its elements represents the wave intensity in each degree of freedom.
The response of the WF disturbances is incorporated in the model by assuming that
the total ship motion is the sum of the LF motion and the WF motion components.
The precept of the DP problem is to attenuate only the LF motion. Therefore, the
WF motion component must be removed from the measured output when designing
an output feedback controller. As mentioned earlier, there are various methods for
wave ﬁltering. We use the state estimation technique for wave ﬁltering. For this, we
augment the LF and WF models and assume that the superposition principle holds
true, so that we can add the outputs of both the LF an WF models. The complete
54 Chapter 4. SDARE based control Design
(augmented) model can be written as
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
˙ η
˙ ν
˙
b
˙
ξ
1
˙
ξ
2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
J(ψ) O
3
O
3
O
3
O
3
−M
−1
D M
−1
J
T
(ψ) O
3
O
3
O
3
O
3
−T
−1
O
3
O
3
O
3
O
3
O
3
O
3
I
3
O
3
O
3
O
3
−Ω
2
−2ZΩ
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
η
ν
b
ξ
1
ξ
2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
+
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
M
−1
O
3
O
3
O
3
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
u
+
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
O
3
O
3
O
3
Ψ O
3
O
3
O
3
O
3
Σ
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
_
w
b
w
ξ
_
, (4.42)
y =y
b
+ y
ξ
+ v
=
_
I
3
O
3
O
3
I
3
O
3
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
η
ν
b
ξ
1
ξ
2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
+ v. (4.43)
The controllability matrix of (4.42) is given by
C =
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
J(ψ)M
−1
−J(ψ)M
−1
DM
−1
−J(ψ)(M
−1
D)
13
M
−1
M
−1
−M
−1
DM
−1
(M
−1
D)
2
M
−1
(M
−1
D)
14
M
−1
O
3
O
3
O
3
O
3
O
3
O
3
O
3
O
3
O
3
O
3
O
3
O
3
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
,
(4.44)
Since the matrices J(ψ), M, and D have full row rank for all ψ ∈ R. Therefore, the
controllability matrix has rank 6. This means that only η and ν are controllable states
while the slowly varying disturbances and wave frequency motion are not controllable.
This is not restrictive as the objective is to control only the low frequency motion.
Using the Theorem 4.1.1, a locally asymptotically stable SDARE controller for (4.42)
is deﬁned by
u = −R
−1
B
T
Π
c
(ψ)x
= −R
−1
M
−1
_
Π
21
(ψ) Π
22
(ψ) Π
23
(ψ) Π
24
(ψ) Π
25
(ψ)
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
η
ν
b
ξ
1
ξ
2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
, (4.45)
where Π
2i
(ψ), i = 1, 2, 3, 4, 5 are obtained from the SDARE associated with (4.1) and
(4.42). Again as we know that only the position measurement is available. Therefore,
to realize the controller (4.45), we need estimates of the states. Then this control law
4.2. Control System Design for the DP Vessel 55
will be
u = −R
−1
M
−1
_
Π
21
(
ˆ
ψ) Π
22
(
ˆ
ψ) Π
23
(
ˆ
ψ) Π
24
(
ˆ
ψ) Π
25
(
ˆ
ψ)
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
ˆ η
ˆ ν
ˆ
b
ˆ
ξ
1
ˆ
ξ
2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
. (4.46)
Before addressing the observability of (4.42), we recall the following result from
[69].
Theorem 4.2.3. Consider the system
˙ x
1
= A
1
x
1
, (4.47)
˙ x
2
= A
1
x
2
, (4.48)
y = C
1
x
1
+ C
2
x
2
, (4.49)
where x
1
∈ R
n
1
and x
2
∈ R
n
2
are states and y ∈ R
p
is the output of the system. If A
1
and A
2
have no common eigenvalues, i.e.,
σ(A
1
) ∩ σ(A
2
) = φ, (4.50)
then the system (4.47)(4.49) is observable iﬀ (C
1
, A
1
) and (C
2
, A
2
) are both observ
able.
Proof. See [69].
The systems (4.23) and (4.40), respectively, describe the slowly varying low fre
quency and wave frequency motions with diﬀerent eigenfrequencies. Therefore, the
system matrices A
b
(ψ) and A
ξ
in (4.23) and (4.40) do not have any common eigenval
ues. Moreover, the observability matrices (4.26) and
O
ξ
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
I
3
O
3
O
3
I
3
.
.
.
.
.
.
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
(4.51)
have full row ranks. Therefore, the subsystems (A
b
(ψ), C
b
) and (A
ξ
, C
ξ
) of the system
(4.42) are observable. This means that by the Theorem 4.2.3, we can conclude that
the system (4.42)(4.43) is pointwise observable for all ψ ∈ R.
56 Chapter 4. SDARE based control Design
An SDARE observer for the system (4.42)(4.43) is deﬁned by
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
˙
ˆ η
˙
ˆ ν
˙
ˆ
b
˙
ˆ
ξ
1
˙
ˆ
ξ
2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
J(
ˆ
ψ) O
3
O
3
O
3
O
3
−M
−1
D M
−1
J
T
(
ˆ
ψ) O
3
O
3
O
3
O
3
−T
−1
O
3
O
3
O
3
O
3
O
3
O
3
I
3
O
3
O
3
O
3
−Ω
2
−2ZΩ
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
ˆ η
ˆ ν
ˆ
b
ˆ
ξ
1
ˆ
ξ
2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
+
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
M
−1
O
3
O
3
O
3
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
u
+ K
o
(
ˆ
ψ)(y − ˆ y), (4.52)
ˆ y =
_
I
3
O
3
O
3
I
3
O
3
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
ˆ η
ˆ ν
ˆ
b
ˆ
ξ
1
ˆ
ξ
2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
. (4.53)
The observer gain K
o
(
ˆ
ψ) in (4.52) is given by
K
o
(
ˆ
ψ) = Π
o
(
ˆ
ψ)C
T
V
−1
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
Π
11
(
ˆ
ψ) + Π
14
(
ˆ
ψ)
Π
21
(
ˆ
ψ) + Π
24
(
ˆ
ψ)
Π
31
(
ˆ
ψ) + Π
34
(
ˆ
ψ)
Π
41
(
ˆ
ψ) + Π
44
(
ˆ
ψ)
Π
51
(
ˆ
ψ) + Π
54
(
ˆ
ψ)
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
V
−1
(4.54)
where Π
i1
(
ˆ
ψ) and Π
i4
(
ˆ
ψ), i = 1, 2, 3, 4, 5 are obtained from the solution of the dual
SDARE associated with the (4.42)(4.43).
A combination of the nonlinear controller (4.46) and the nonlinear observer (4.52)
is called a nonlinear compensator and is shown in Figure 4.2. The nonlinear SDARE
compensator is given by
_
˙ x
˙
ˆ x
_
=
_
A(ψ) −BK
c
(
ˆ
ψ)
K
o
(
ˆ
ψ)C A(
ˆ
ψ) − BK
c
(
ˆ
ψ) − K
o
(
ˆ
ψ)C
_ _
x
ˆ x
_
+
_
E O
O K
o
(
ˆ
ψ)
_ _
w
v
_
.
(4.55)
Before concluding this section, it is important to mention that in reality, the WF
model is considered as a disturbance to the model. For designing a control law, it is
desired that the eﬀects of the WF disturbances do not enter into the feedback loop.
In this particular example, it is required that no part of the disturbance vectors ξ
1
and
ξ
2
enters in the control action. Here, the controller gain matrix turns out to have the
form, K
c
(
ˆ
ψ) = [K
c1
(
ˆ
ψ), K
c2
(
ˆ
ψ), K
c3
(
ˆ
ψ), O
3
, O
3
] which ensures the attenuation of the
disturbance; in the feedback the product term, K
c
(ψ) ˆ x, eliminates
ˆ
ξ
1
and
ˆ
ξ
2
from the
feedback loop.
4.3 Simulation Results
Suppose the objective is to design a controller to stabilize the vessel around the origin.
By the stabilized mode we mean that the vessel is positioned at rest at (x, y) = (0, 0)
and its heading is along the positive x−axis.
4.3. Simulation Results 57
+ + + =
b b b b b b
x A ( ) x B u E w
=
b b b
y C x
( )( ) = + + + ¢ ÷
0
ˆ
ˆ ˆ ˆ x A( ) x Bu K y y
= ˆ ˆ y Cx
b
C
+
0
ˆ
K ( )
C
c
ˆ
‐K ( ) +
ç
C
ç ç
ç = Oç +
E w
ç ç
= ç y C
ç
= + +
b
y y y v
ˆ y ‐ y
ˆ x
u
b
x b
y
y
ç
ˆ y
+

ç
v
Figure 4.2: Nonlinear compensator
For numerical simulations to illustrate the eﬀectiveness of the SDARE controllers
and observers designed in previous subsection, we use the data of a supply vessel from
[47]. For simplicity of numerics, Bisscaled [23] normalized parameters of the vessel
model are used. The normalized matrices M and D are:
M =
_
¸
¸
¸
¸
¸
¸
¸
¸
_
1.1274 0 0
0 1.8902 −0.0744
0 −0.0744 0.1278
_
¸
¸
¸
¸
¸
¸
¸
¸
_
, D =
_
¸
¸
¸
¸
¸
¸
¸
¸
_
0.0358 0 0
0 0.1183 −0.0124
0 −0.0124 0.0308
_
¸
¸
¸
¸
¸
¸
¸
¸
_
.
(4.56)
We assume that the ship is positioned at a point (x, y) = (−10, −10) making an angle
ψ(= 4 rad) with the positive xaxis. The control objective is to steer the system to the
stable equilibrium position, (x, y) = (0, 0) and heading along the positive xaxis.
First we consider the case when the disturbance vector b is known. Suppose b =
[0.05, 0.05, 0.01]
T
and we use the control law (4.18) in (4.12). We present the simu
lation results corresponding to the weighting matrices Q = 10
3
diag¦0.1, 1, 0.1, 0, 0, 0¦
and R = 10
2
I
3
. Figure 4.3 shows the position and heading proﬁles as they evolve with
time. It is clear that the regulation to the desired point is steady and smooth. Figure
4.4 shows: (a) the trajectory proﬁle of the vessel, (b) the cost functional proﬁle, and
(c) the norm of the state vector. Figure 4.5 shows change in the heading of the ves
sel along the trajectory. Figure 4.6 shows the controller recommended force in each
degree of freedom of the vessel. Once more, it is clear that the actuator input to the
vessel is steady and smooth with a smooth overshoot.
As we have already mentioned that the weighting matrices can be tuned to im
prove the performance of the controller. We elaborate this feature in the following.
For instance, we penalize the velocities and take Q = 10
3
diag¦0.1, 1, 0.1, 1, 1, 1¦. Fig
ures 4.7 and 4.8 show the control input signals and the heading orintation of the vessel
corresponding to the tuned weighting matrix Q. It is clear that new proﬁles for the
heading orientation and the control input signal are smoother than the previous pro
ﬁles. To emphasize more on tuning factor, we include another heading orientation
58 Chapter 4. SDARE based control Design
proﬁle corresponding to R = 10
0
I
3
, see Figure 4.9.
0 2 4 6 8 10 12 14 16 18 20
−10
−5
0
x
[
m
]
(a)
0 2 4 6 8 10 12 14 16 18 20
−10
−5
0
y
[
m
]
(b)
0 2 4 6 8 10 12 14 16 18 20
−1
0
1
2
3
4
5
Time [sec]
ψ
[
r
a
d
]
(c)
Figure 4.3: The position and heading proﬁles from the closed loop system (4.12) and
(4.18)
.
4.3. Simulation Results 59
−12 −10 −8 −6 −4 −2 0 2
−10
−5
0
x [m]
y
[
m
]
(a)
0 5 10 15
0
1
2
3
x 10
5
C
o
s
t
:
x
T
Q
x
+
u
T
R
u
(b)
0 5 10 15
0
5
10
15
Time [sec]
N
o
r
m
:


x


(c)
Figure 4.4: The evolution of the phase, cost, and norm (of the state vector) proﬁles
with time.
−12 −10 −8 −6 −4 −2 0 2 4
−14
−12
−10
−8
−6
−4
−2
0
2
x
1
[m]
Heading angle variation
x
2
[
m
]
Figure 4.5: The heading orientation over the trajectory of the vessel.
60 Chapter 4. SDARE based control Design
0 1 2 3 4 5 6 7 8 9 10
−40
−20
0
20
40
u
1
[
k
N
]
(a)
0 1 2 3 4 5 6 7 8 9 10
−20
−10
0
10
20
u
2
[
k
N
]
(b)
0 1 2 3 4 5 6 7 8 9 10
−6
−4
−2
0
2
4
Time [sec]
u
3
[
k
N
m
]
(c)
Figure 4.6: The proﬁles of the individual components of the control input vector.
−14 −12 −10 −8 −6 −4 −2 0 2 4
−14
−12
−10
−8
−6
−4
−2
0
2
x
1
[m]
Heading angle variation
x
2
[
m
]
Figure 4.7: The heading orientation over the trajectory of the vessel.
4.3. Simulation Results 61
0 1 2 3 4 5 6 7 8 9 10
−30
−20
−10
0
10
u
1
[
k
N
]
(a)
0 1 2 3 4 5 6 7 8 9 10
−10
−5
0
5
u
2
[
k
N
]
(b)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−5
−4
−3
−2
−1
0
1
Time [sec]
u
3
[
k
N
m
]
(c)
Figure 4.8: The proﬁles of the individual components of the control input vector.
−12 −10 −8 −6 −4 −2 0 2 4
−14
−12
−10
−8
−6
−4
−2
0
2
4
x
1
[m]
Heading angle variation
x
2
[
m
]
Figure 4.9: The heading orientation over the trajectory of the vessel.
62 Chapter 4. SDARE based control Design
Next, we consider the case when the bias vector b is unknown. We consider ex
actly the same situation regarding the control objective and the initial position of the
vessel, as in the nominal case, i.e., when b is available. The simulation results show
the performance of the SDARE based control law and the SDARE observer to achieve
the desired objective.
For the simulation results shown in the following ﬁgures, we use the bias time
constant as T = diag¦100, 100, 100¦ while the controller weighting matrices are
Q = 10
−1
diag¦100, 100, 10, 1, 1, 1, 0, 0, 0¦ and R = 10
−1
diag¦1, 1, 0.01¦,
and the observer weighting matrices are
U = 10
0
diag¦1, 1, 1, 1, 1, 1, 1, 1, 10¦ and V = 10
−2
diag¦1, 1, 1¦.
Figure 4.10 shows the measured position and heading proﬁles and their estimates
obtained from the SDARE observer. As desired, the proﬁles show a steady and smooth
estimation of the noisy measurements. Figure 4.11 shows the proﬁles of linear velocity
in surge and sway directions and the angular yaw velocity and their estimates. Figure
4.12 shows the proﬁles of the bias estimates in surge, sway, and yaw directions. Figure
4.13 shows: (a) the trajectory proﬁle, (b) the cost functional proﬁle, and (c) the norm
of the state vector and its estimate. Figure 4.14 shows the heading orientation with the
trajectory. The proﬁle shows a smooth and steady path and orientation of the vessel.
Figure 4.15 shows the control input signals from the controller in each degree of free
dom. The proﬁles show that the controller adjusts in the start and then signals become
steady and smooth. From all these ﬁgures, we see that the SDARE compensator states
show nice asymptotic convergence to the desired point.
To emphasize the role of the weighting matrices, we include some more simula
tions. For instance, we take
Q = 10
−2
diag¦100, 100, 10, 1, 1, 1, 0, 0, 0¦ and R = 10
−3
diag¦1, 1, 0.01¦.
This means that we reduce the penalty on both the vectors x and u. Figures 4.16 and
4.17 show the corresponding proﬁles of the trajectory and the control input signals.
The vessel takes a relatively loger route but its smoother than the one we obtained
with the previous weighting matrices. The control input signals are also ampliﬁed as
expected.
4.3. Simulation Results 63
0 5 10 15 20 25 30
−15
−10
−5
0
5
10
x
a
n
d
ˆx
[
m
]
(a)
x
ˆ x
0 5 10 15 20 25 30
−20
−10
0
10
y
a
n
d
ˆy
[
m
]
(b)
y
ˆ y
0 5 10 15 20 25 30
−2
0
2
4
6
ψ
a
n
d
ˆ
ψ
[
r
a
d
]
Time [sec]
(c)
ψ
ˆ
ψ
Figure 4.10: The proﬁles of the actual and the estimated position and orientation as
obtained from (4.39).
0 1 2 3 4 5 6 7 8 9 10
−10
0
10
20
30
u
a
n
d
ˆu
[
m

s
e
c
−
1
]
(a)
u
ˆ u
0 1 2 3 4 5 6 7 8 9 10
−10
−5
0
5
10
v
a
n
d
ˆv
[
m

s
e
c
−
1
]
(b)
v
ˆ v
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−40
−20
0
20
r
a
n
d
ˆr
[
r
a
d

s
e
c
−
1
]
(c)
Time [sec]
r
ˆ r
Figure 4.11: The actual and the estimated surge, sway, and yaw velocities as obtained
from (4.39).
64 Chapter 4. SDARE based control Design
0 5 10 15
−10
−8
−6
−4
−2
0
2
ˆ b
1
[
k
N
]
(a)
0 5 10 15
−10
−5
0
ˆ b
2
[
k
N
]
(b)
0 5 10 15
−2
0
2
4
6
ˆ b
3
[
k
N

m
]
(c)
Time [Sec]
Figure 4.12: The bias estimates in surge, sway, and yaw directions as obtained from
(4.39).
−12 −10 −8 −6 −4 −2 0 2 4 6
−15
−10
−5
0
5
x [m]
y
[
m
]
xy−trajectory
0 1 2 3 4 5 6 7 8 9 10
0
1000
2000
3000
4000
C
o
s
t
:
x
T
Q
x
+
u
T
R
u
(b)
0 5 10 15 20 25 30
0
10
20
30
40
Time [sec]


x


a
n
d


ˆx


(c)
x
ˆ x
Figure 4.13: The trajectory and the proﬁles of the cost functional and the norms of the
measured state and its estimate as obtained from (4.39).
4.3. Simulation Results 65
−14 −12 −10 −8 −6 −4 −2 0 2 4 6
−16
−14
−12
−10
−8
−6
−4
−2
0
2
4
x
1
[m]
Heading angle variation
x
2
[
m
]
Figure 4.14: The heading orientation along the trajectory.
0 1 2 3 4 5 6 7 8 9 10
−50
0
50
100
150
u
1
[
k
N
]
(a)
0 1 2 3 4 5 6 7 8 9 10
−100
0
100
u
2
[
k
N
]
(b)
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
−50
0
50
Time [sec]
u
3
[
k
N
−
m
]
Figure 4.15: The control input signals for each degree of freedom computed from
(4.30).
66 Chapter 4. SDARE based control Design
−12 −10 −8 −6 −4 −2 0 2 4 6
−20
−15
−10
−5
0
5
x
1
[m]
Heading angle variation
x
2
[
m
]
Figure 4.16: The heading orientation along the trajectory.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−100
0
100
200
u
1
[
k
N
]
(a)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−200
0
200
u
2
[
k
N
]
(b)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−80
−60
−40
−20
0
20
40
Time [sec]
u
3
[
k
N
−
m
]
(c)
Figure 4.17: The control input signals for each degree of freedom computed from
(4.30).
4.3. Simulation Results 67
Finally, we consider the case when the WF motion is also playing a role. For the
simulation results shown in the following ﬁgures, we use the bias time constant as
T = diag¦100, 100, 100¦, the dominating wave frequencies as Ω = diag¦0.9, 0.9, 0.9¦,
the relative damping ratios as Z = diag¦0.2, 0.2, 0.2¦, the controller weighting matrices
as
Q = 10
−1
diag¦100, 100, 10, 1, 1, 1, O
1×6
¦, and R = 10
−1
diag¦1, 1, 0.01¦
and the observer weighting matrices as
U = 10
0
diag¦1, 1, 1, 1, 1, 1, 1, 1, 10, 1, 1, 1, 1, 1, 1¦ and V = 10
−1
diag¦1, 1, 1¦.
Figure 4.18 shows the proﬁles of the measured position and heading, and their es
timates obtained from the SDARE observer. The proﬁles depict a smooth and steady
regulation of the vessel to the desired position. Figure 4.19 shows the velocity pro
ﬁles of the vessel motion and their estimates along surge, sway, and yaw directions.
Figure 4.20 shows the bias estimates obtained from the SDARE observer. Figure 4.21
shows: (a) the trajectory of the vessel motion, (b) the cost functional proﬁle, and (c)
the norm of the measured state vector and its estimate. In Figure 4.22, we show the
heading orientation along the trajectory. The traectory and orientation are nice and
smooth. Figure 4.23 shows the control input signal in each degree of freedom com
puted from (4.46). As in the previous case, the control signals need some adjustment
in the beginning and then become smooth and steady.
To further elaborate the importance of the weighting matrices, we add some more
simulation results. Now, Lets take the matrices
Q = 10
−2
diag¦100, 100, 10, 1, 1, 1, O
1×6
¦ and R = 10
−3
diag¦1, 1, 0.01¦.
Figures 4.24 and 4.25 show the corresponding trajectory and control input signals.
It can be seen that the vessel follows a relatively longer path but its slightly better
than the previous path. The control input signals are ampliﬁed with this choice of the
tuning matrices.
Further lets modify the observer weighting matrix U: we now take
U = 10
0
diag¦10, 10, 10, 1, 1, 1, 1, 1, 10, 1, 1, 1, 1, 1, 1¦.
Figures 4.26 and 4.27 show the corresponding proﬁles of the trajectory and the control
input signals. The trajectory proﬁle has further improved and control input proﬁles
show an increased control input.
68 Chapter 4. SDARE based control Design
0 5 10 15 20 25 30
−20
−10
0
10
x
a
n
d
ˆx
[
m
]
(a)
x
ˆ x
0 5 10 15 20 25 30
−20
−10
0
10
y
a
n
d
ˆy
[
m
]
(b)
y
ˆ y
0 5 10 15 20 25 30
−5
0
5
ψ
a
n
d
ˆ
ψ
[
r
a
d
]
Time [sec]
(c)
ψ
ˆ
ψ
Figure 4.18: The position and heading proﬁles of the vessel motion as obtained from
the compensator (4.55).
0 1 2 3 4 5 6 7 8 9 10
−10
0
10
20
u
a
n
d
ˆu
[
m

s
e
c
−
1
]
(a)
u
ˆ u
0 1 2 3 4 5 6 7 8 9 10
−10
−5
0
5
10
v
a
n
d
ˆv
[
m

s
e
c
−
1
]
(b)
v
ˆ v
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−40
−20
0
20
r
a
n
d
ˆr
[
r
a
d

s
e
c
−
1
]
Time [sec]
(c)
r
ˆ r
Figure 4.19: The velocities and their estimates along surge, sway, and yaw directions
as obtained from (4.55).
4.3. Simulation Results 69
0 5 10 15
−10
−8
−6
−4
−2
0
2
ˆ b
1
[
k
N
]
(a)
0 5 10 15
−10
−5
0
ˆ b
2
[
k
N
]
(b)
0 5 10 15
−2
0
2
4
6
ˆ b
3
[
k
N

m
]
(c)
Time [Sec]
Figure 4.20: The bias estimates in surge, sway, and yaw directions as obtained from
(4.55).
−12 −10 −8 −6 −4 −2 0 2 4
−15
−10
−5
0
5
x [m]
y
[
m
]
xy−trajectory
0 1 2 3 4 5 6 7 8 9 10
0
1000
2000
3000
4000
C
o
s
t
:
x
T
Q
x
+
u
T
R
u
(b)
0 5 10 15 20 25 30
0
10
20
30
40
Time [sec]


x


a
n
d


ˆx


(c)
x
 ˆ x
Figure 4.21: The trajectory and proﬁle of the cost functional and the norms of the
measured state vector and its estimate corresponding to the closed loop system (4.55).
70 Chapter 4. SDARE based control Design
−14 −12 −10 −8 −6 −4 −2 0 2 4 6
−14
−12
−10
−8
−6
−4
−2
0
2
4
x
1
[m]
Heading angle variation
x
2
[
m
]
Figure 4.22: The heading orientation along the trajectory.
0 1 2 3 4 5 6 7 8 9 10
−50
0
50
100
150
u
1
[
k
N
]
(a)
0 1 2 3 4 5 6 7 8 9 10
−100
0
100
u
2
[
k
N
]
(b)
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
−100
−50
0
50
Time [sec]
u
3
[
k
N
−
m
]
Figure 4.23: The control input signals for each degree of freedom as obtained from
(4.55).
4.3. Simulation Results 71
−12 −10 −8 −6 −4 −2 0 2 4 6
−14
−12
−10
−8
−6
−4
−2
0
2
4
x
1
[m]
Heading angle variation
x
2
[
m
]
Figure 4.24: The heading orientation along the trajectory.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−100
0
100
200
300
u
1
[
k
N
]
(a)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−200
0
200
u
2
[
k
N
]
(b)
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
−100
−50
0
Time [sec]
u
3
[
k
N
−
m
]
Figure 4.25: The control input signals for each degree of freedom as obtained from
(4.55).
72 Chapter 4. SDARE based control Design
−12 −10 −8 −6 −4 −2 0 2 4 6
−14
−12
−10
−8
−6
−4
−2
0
2
x
1
[m]
Heading angle variation
x
2
[
m
]
Figure 4.26: The heading orientation along the trajectory.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−100
0
100
200
300
u
1
[
k
N
]
(a)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−400
−200
0
200
u
2
[
k
N
]
(b)
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
−150
−100
−50
0
50
Time [sec]
u
3
[
k
N
−
m
]
Figure 4.27: The control input signals for each degree of freedom as obtained from
(4.55).
4.4. Conclusions 73
4.4 Conclusions
We study the regulation problem of a DP vessel by using the SDARE technique.
We extend the concept of the LQR problem to nonlinear system by using the SDC
parametrization. This gives rise to an SDARE which must be solved to obtain a sub
optimal solution of the nonlinear regulation problem.
The DP vessel model shows that only the position and heading measurements
are available from the sensors. This is an inadequate information for the SDARE
control design problem which requires complete state knowledge. Using the stabi
lizability and detectability of the vessel model together with the duality concept, we
propose a suboptimal locally asymptotically stable SDARE observer. The combina
tion of the locally asymptotically stable SDARE regulator and SDARE observer is a
locally asymptotically stable SDARE compensator.
We divide the study of the DP vessel regulation problem in two steps. First, we
study the LF motion regulation without taking into account the WF motions. We
begin by considering a nominal case (the bias disturbance vector is known or its esti
mate is available) and come up with a PDtype control law with a feedforward term.
The controller is proved to be globally asymptotically stable by using the LMIs and
the Lyapunov stability theory. Afterwards, we assume that the bias vector is unknown
and the measured signals from the sensors are contaminated by noise. We combine the
concept of the SDARE regulation and estimation (to estimate the position, velocity,
and bias vectors) to come up with a locally asymptotically stable SDARE compen
sator. Second, we take into account both the LF and WF motions and use again the
SDARE method to address the regulation and estimation problems. A locally asymp
totically stable compensator is obtained by combining locally asymptotically stable
controller and estimator. We use the data of a supply vessel for simulation experi
ments. Simulation results presented in this chapter conﬁrm the asymptotic stability of
the SDARE regulator and the SDARE estimator.
74
Chapter 5
The Fourier Series Interpolation
Method
1
In Chapter 4, the controller and observer gains are state dependent, so they are re
quired to be obtained by online solving the concerned SDAREs. In the process, us
ing the particular state dependence of the system matrix, we propose a Fourier series
based interpolation method. This method reduces the online computation time of the
solution of the SDARE by performing the major computational task oﬄine.
The earliest traces of the use of the algebraic Riccati equation in control related
problems can be found in the works on the solution of the LQ optimal control and
ﬁltering problem [38] and on approximation methods in optimal control [62]. To date
its scope has been extended to address problems of nonlinear state regulation and
estimation, estimator based feedback control synthesis, H
∞
−control, H
2
−control, and
nonlinear ﬁlter design, for instance see [6], [12], [29], [48], and [61].
The SDARE has emerged in many practical applications. Especially, the SDARE
has become an important research topic in ongoing developments in the areas of con
trol and estimation theory. It has been regarded as an eﬃcient tool to tackle problems
related with the nonlinear control design, ﬁltering, synthesis, and analysis. Conse
quently, emphasis has been put on developing computationally eﬃcient solution meth
ods for the SDARE.
In a broader perspective, we divide the solution methods in two categories: itera
tive and direct. Both type of methods have positive and negative features. Mainly, the
direct methods are computationally faster than the iterative methods, especially when
the problem is poorly conditioned or when a goodquality initial guess is not available.
On the other hand, the direct methods require more storage capacity than the iterative
methods. Some of the wellknown direct methods are the Taylor series method [6],
the generalized eigenvector approach [13], the matrix sign function method [15], the
information ﬁlter algorithm [36], and the Schur decomposition method [45]. Some
1
This work has been published in the proceedings of the UKACC International Conference on Control
(CONTROL 2010), see [52].
75
76 Chapter 5. The Fourier Series Interpolation Method
of the wellknown iterative methods are the Kleinman algorithm [2], the interpola
tion methods [6], the Newtonmethod and its modiﬁed forms [8, 16, 37], solution via
spectral factorization, doubling algorithm [11, 41], and Chandrasekhar algorithm [67].
In (4.5), we observe that the controller gain is not ﬁxed but it varies with the state
of the system. The computation of the controller gain requires the solution of the
SDARE (4.6). Therefore, in the feedback loop, controller gain is computed online.
As pointed out in [10] the reduction in the online computation time of the solution of
the SDARE is a major issue because of its high computational cost. In this chapter,
we give the Fourier series based numerical method for the solution of the SDARE of
type (4.7) which reduces the online computations. Henceforth, we call this method the
Fourier series interpolation (FSI) method. This method of solution is computationally
very simple and hence, it is valuable in the situations where online computation of the
solution of the SDARE is required.
5.1 The Fourier Series Interpolation (FSI) Method
First we explain why it is reasonable to use the Fourier series for the solution of the
algebraic Riccati equations of the type (4.7) and then we will illustrate the details of
the FSI method. For a general formulation of the FSI method, we make a slight change
of notation in the SDARE (4.7). We consider the following form
Π(x
1
)A(x
1
) + A
T
(x
1
)Π(x
1
) + Q − Π(x
1
)BR
−1
B
T
Π(x
1
) = 0, (5.1)
where x
1
∈ R, without loss of generality, is the ﬁrst component of the state vector
x ∈ R
n
, A(x
1
) ∈ R
n×n
is C
∞
and θ−periodic, Π(x
1
) ∈ R
n×n
, and B ∈ R
n×m
.
The matrix A(x
1
) is smooth and θ−periodic, i.e., A(x
1
) = A(x
1
+ θ). The solution
matrix Π(x
1
), provided it exists, is also smooth and θ−periodic, see [14]. At this point,
we recall two important results from [33] which endorse the idea of using the Fourier
series for ﬁnding an approximate solution of the SDARE (5.1). First, we recall the
Fourier convergence theorem due to Dirichlet.
Theorem 5.1.1. (The Fourier Convergence Theorem) Let f (x) be a function which
is piecewise smooth on the interval −L ≤ x ≤ L, then the Fourier series of f (x)
converges
1. to f (x), if f (x) is continuous at x.
2. to the average of the two limits,
1
2
[ f (x+) + f (x−)],
if f (x) has a jump discontinuity at x, where f (x−) and f (x+) means, respectively, the
left and the right hand limits of the function f at point x.
The periodic extension of a 2L−periodic function is always admissible. The fol
lowing corollary gives the necessary and suﬃcient condition for the continuity of the
periodic extension of a piecewise smooth function.
Corollary 5.1.1. For a 2L−periodic piecewise smooth function f (x), the periodic
extension of f (x) is continuous for all x ∈ R if and only if f (x) is continuous and
f (−L) = f (L).
5.1. The Fourier Series Interpolation (FSI) Method 77
In the context of the vessel model, the solution Π
c
(ψ) of the SDARE (4.7) is a
smooth matrix function with period θ. Therefore, by the above results, it is possible
to represent the matrix function Π(x
1
) in the form of a Fourier series. This means that
the Fourier series of Π(x
1
) exists and converges to Π(x
1
). Thus we write
Π(x
1
) =
a
0
2
+
∞
k=1
(a
k
cos
2kπx
1
θ
+ b
k
sin
2kπx
1
θ
). (5.2)
The Fourier coeﬃcients a
0
, a
k
∈ R
n×n
and b
k
∈ R
n×n
for k = 1, 2, 3, ..., are deﬁned by
a
k
=
2
θ
_ θ
2
−
θ
2
Π(x
1
) cos
2kπx
1
θ
dx
1
, k = 0, 1, 2, 3, .... (5.3)
and
b
k
=
2
θ
_ θ
2
−
θ
2
Π(x
1
) sin
2kπx
1
θ
dx
1
, k = 1, 2, 3, .... (5.4)
Let us now recall a wellknown result by Riemann and Lebesgue which is stated
in the following lemma.
Lemma 5.1.1. (The RiemannLebesgue Lemma) If f (x) is integrable on [−
θ
2
,
θ
2
]
then
lim
k→∞
_ θ
2
−
θ
2
f (x) sin kx dx = 0,
and
lim
k→∞
_ θ
2
−
θ
2
f (x) cos kx dx = 0,
From the fact that the series on the right hand side of (5.2) is convergent and
from the statement of the Lemma 5.1.1, it follows that the Fourier coeﬃcients with
suﬃciently high indices are negligibly small (their norms are of order zero) and can
be left out. Therefore, for any small positive number , there exists an integer N
·
and
such that ¦¦a
n
¦¦
2
< and ¦¦b
n
¦¦
2
< , ∀n > N
·
. We use the spectral matrix norm in this
chapter.
We explain the procedure to compute the Fourier coeﬃcients as follows. First, we
decide a stopping criterion for the computation of the Fourier coeﬃcients. We deﬁne
a tolerance level say
·
> 0, such that the process of the computation of the Fourier
series coeﬃcients stops when
¦¦Π(x
1
) − (
a
0
2
+
N
k=1
(a
k
cos
2kπx
1
θ
+ b
k
sin
2kπx
1
θ
))¦¦
2
<
·
, (5.5)
is satisﬁed for all x
1
∈ [−
θ
2
,
θ
2
).
Then, we start with N = 1 such that there are 2N+1(= 3) Fourier coeﬃcients to be
determined. Equation (5.2) holds identically true for all x
1
∈ [−
θ
2
,
θ
2
). Take m > 2N+1
78 Chapter 5. The Fourier Series Interpolation Method
points x
1
t
∈ [−
θ
2
,
θ
2
), t = 1, 2, ..., m. To each point x
1
t
∈ [−
θ
2
,
θ
2
), there corresponds an
equation of the following form
a
0
2
+
N
k=1
(a
k
cos
2kπx
1
t
θ
+ b
k
sin
2kπx
1
t
θ
) = Π(x
1
t
). (5.6)
In this way, we get the following nonhomogeneous system of m linear matrix equa
tions
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
1
2
cos
2πx
1
1
θ
cos
2Nπx
1
1
θ
sin
2πx
1
1
θ
sin
2Nπx
1
1
θ
1
2
cos
2πx
1
2
θ
cos
2Nπx
1
2
θ
sin
2πx
1
2
θ
sin
2Nπx
1
2
θ
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
2
cos
2πx
1m
θ
cos
2Nπx
1m
θ
sin
2πx
1m
θ
sin
2Nπx
1m
θ
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
a
0
a
1
.
.
.
a
N
b
1
.
.
.
b
N
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
Π
1
Π
2
.
.
.
Π
m
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
,
where Π
i
= Π(x
1
i
). We can write in short form as
7X = J (5.7)
However, note that the dimensions of the matrices in (5.7) do not conform with
the usual matrix multiplication rules. We use the Kronecker matrix product to remove
this discrepancy. For this we multiply the matrix 7from right by I
n
in the Kronecker
sense. This results in
(7⊗ I
n
)X = J. (5.8)
The equation (5.8) can be solved by using the MATLAB to get
X = (7⊗ I
n
)\J. (5.9)
This solution yields 2N + 1 Fourier coeﬃcients. Therefore, the Fourier series approx
imation of the solution of the SDARE (5.1) is
Π(x
1
) ≈
a
0
2
+
N
k=1
(a
k
cos
2kπx
1
θ
+ b
k
sin
2kπx
1
θ
). (5.10)
Now the stopping criterion (5.5) is tested. This can be done by plotting
·
− g(x
1
)
vs x
1
on the interval [−
θ
2
,
θ
2
), where g(x
1
) is the LHS of (5.5). If the stopping criterion
is satisﬁed (plot lies above x
1
axis) then we stop. Otherwise, we repeat the above pro
cedure for the next integer value of N (=2) and so on, and continue until the stopping
criterion is satisﬁed. The algorithm for ﬁnding the Fourier coeﬃcients is summarized
5.2. Performance Analysis 79
below
2
.
Algorithm 5.1.1: Comrtrxrrox or rnr Fotarra Corrrrcrrxrs (FCs)
input: The tolerance parameter
·
and an integer N(= 1).
assumption: The m number of interpolation points must satisfy m > 2N + 1.
output: The Fourier Coeﬃcients a
0
, a
1
, a
2
, ..., a
N
, b
1
, b
2
, ..., b
N
.
for N ← 1 to some f inite large natural number
step 1: Take m > 2N + 1.
step 2: Take x
t
i
∈ [−
θ
2
,
θ
2
), i = 1, 2, 3, ...m.
step 3: Solve (5.1) at m interpolation points to get Π
1
, Π
2
, ..., Π
m
. (i)
step 4: Use statement (i) of this algorithm and (5.6) to get (5.7).
step 5:
_
Use Kronecker matrix product to get a linear system of
equations (5.8).
step 6:
_
Solve the linear system (5.8) to get the desired Fourier
coeﬃcients.
if the tolerance condition is satis f ied
then
_
Fourier Coe f f icients a[1], a[2], ..., a[N], b[1], b[2], ...., b[N]
are obtained.
else
continue the f or loop until tolerance condition is satis f ied.
The main step of the algorithm is step 6 of the for loop. The repetition of the loop
continues until the stopping criteria is satisﬁed. We conclude this section with the
following remark about the above algorithm.
Remark 5.1.1. We take the number m of distict points in [−
θ
2
,
θ
2
), greater than the
number of Fourier coeﬃcients 2N+1 to be determined to generate an overdetermined
system of linear equations. The system (5.8) has a solution because rank(7 ⊗ I
n
) =
rank(7⊗ I
n
, J).
5.2 Performance Analysis
In this section we explain why the FSI method is eﬀectively a better method for online
computation of the solution of the SDARE than “care” in terms of computation time.
We do so on the basis of ﬂop count. Equations (4.20), (4.39) and (4.55) describe the
feedback dynamics of the SDARE nonlinear compensator. The presence of K
c
(
ˆ
ψ) and
K
o
(
ˆ
ψ) therein indicates that two SDAREs have to be solved at each iteration when
the ode systems (4.20), (4.39) and (4.55) are solved numerically. This demands a
computationally eﬃcient method for ﬁnding the solution of the SDARE.
2
The author is thankful to Professor Donald Kreher for his help on the pseudocode environment.
80 Chapter 5. The Fourier Series Interpolation Method
We compare the FSI method with the Schur decomposition method [13]. The
MATLAB routine “care” is based on this method. For each ψ, the SDARE (4.7)
reduces to a continuous time algebraic Riccati equation (CARE)
ΠA + A
T
Π + Q − ΠBR
−1
B
T
Π = 0, (5.11)
We solve the closed loop system equations by using the MATLAB solver ode45. Dur
ing the process of computation of the numerical solution of the closed loop system,
MATLAB routine “care” is called multiple times to solve (5.11). In the following we
explain some details of the Schur decomposition method for a better understanding of
the computational complexity of this method.
The Schur decomposition method requires two major computations for solving
the SDARE. First step is the reduction of the Hamiltonian matrix associated with the
SDARE to an ordered real Schur form (RSF). The Hamiltonian matrix associated with
(5.11) is given by
1 =
_
A −BR
−1
B
T
−Q −A
T
_
. (5.12)
The following result ensures that the Hamiltonian matrix (5.12) does not have any
eigenvalue on the imaginary axis.
Theorem 5.2.1. (Relationship between the Hamiltonian Matrix and the Riccati
Equations) Let (A,B) be stabilizable and (A,Q) be detectable. Then the Hamiltonian
matrix 1 has n eigenvalues with negative real parts, no eigenvalues on the imaginary
axis, and n eigenvalues with positive real parts. In this case the CARE (5.11) has a
unique stabilizing solution Π.
Now using the QRfactorization, 1 can be transformed into RSF form according
to the following result.
Theorem 5.2.2. (The Real Schur Triangularization Theorem) Let 1 be a 2n × 2n
Hamiltonian matrix. Then there exists a 2n × 2n matrix Q
rs
such that
Q
T
rs
1Q
rs
= R
rs
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
R
11
R
12
R
1k
0 R
22
R
2k
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 R
kk
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
, (5.13)
where each R
ii
is either a scalar or a 2 × 2 matrix. The scalar diagonal entries corre
spond to real eigenvalues, and each 2×2 matrix on the diagonal has a pair of complex
conjugate eigenvalues. The matrix R
rs
is known as the RSF of 1.
The eigenvalues in (5.13) are not ordered so an additional step is made to get an
ordered RSF which is important for the next step which uses the invariant subspace of
the stable eigenvalues of 1 to solve (5.11). The following deﬁnition and subsequent
theorem and corollary summarize the concluding step of the Schur decomposition
method.
5.2. Performance Analysis 81
Deﬁnition 5.2.1. (Basis of an Invariant Subspace from RSF) Let
Q
T
rs
1Q
rs
= R
rs
=
_
R
11
R
12
0 R
22
_
, (5.14)
and let us assume that eigenvalues of R
11
and R
22
respectively are the negative and
positive eigenvalues of the Hamiltonian matrix 1. Then the ﬁrst p columns of Q
rs
,
where p is the order of R
11
, form an orthonormal basis for the invariant subspace
associated with the stable eigenvalues of 1.
Theorem 5.2.3. A matrix Πis a solution of the CARE (5.11) if and only if the columns
of
_
I
Π
_
span an ndimensional invariant subspace of the Hamiltonian matrix 1.
Corollary 5.2.1. If the columns of
_
Π
1
Π
2
_
span an ndimensional invariant subspace
of the Hamiltonian matrix 1 associated with the CARE (5.11) and Π
1
is invertible,
then Π = Π
2
Π
−1
1
is a solution of the CARE (5.11).
The Schur method is based on reduction of the Hamiltonian matrix 1 to RSF
which is done by QR factorization. The exact ﬂop count is not possible as QR
factorization is an iterative procedure. However, by the empirical observations it is
assumed that on average each eigenvalue requires 2 QRfactorizations, [13]. The QR
factorization of a matrix requires O(n
3
) ﬂops which means that the RSF obtained in
the ﬁrst step of the Schur decomposition method also requires O(n
3
) ﬂops. The order
ing of the eigenvalues in RSF requires O(n) ﬂops. In the second step, the computation
of Π = Π
2
Π
−1
1
is done by using the leastsquares algorithm and this also requires
O(n
3
) ﬂops. Therefore the Schur decomposition requires O(n
3
) ﬂops each time CARE
(5.11) is solved online. On the other hands, the FSI method requires simple matrix
addition (O(n
2
) ﬂops) and scalarmatrix multiplication (O(n
2
) ﬂops). Therefore, the
FSI method requires O(n
2
) ﬂops when the online solution of (5.11) is computed by
this method.
The computationally expensive task in FSI method is the computation of the Fouri
er coeﬃcients which is done oﬄine, only once. From the algorithm given in Section
5.1, we see that the computation of the Fourier coeﬃcients involve two major steps.
First, the computation of the solution of the SDARE at a known (m) number of points.
Second, the solution of a linear system of equations. Each requires O(n
3
) ﬂops.
Based on the foregoing discussion, we conclude that the FSI method reduces the
online computations in a feedback dynamics based on an SDARE control law. In
Table 5.1, by using the simulation results discussed in Section 4.3, we quantitatively
demonstrate that the FSI method reduces the online computation time for the solution
of the SDARE. Each of the closed loop systems (4.20), (4.39) and (4.55) are solved
with the MATLAB routine ode45 on a time interval [0, 30]. Table 5.1 shows a com
parison between the computer time required to solve the closed loop systems with the
FSI method and by the MATLAB routine “care”. We notice that with the increase in
the size of the system, the performance of the FSI method is still comparable to its
performance with a lower order system. The simulations have been performed on an
Intel Centrino duo running at the frequency of 2 GHz having 2 GB RAM.
82 Chapter 5. The Fourier Series Interpolation Method
Table 5.1: Performance Analysis
Online comput. Online comput. Percentage (%) reduc.
time (sec) time (sec) in online
with “care” with FSI method computation time
For (4.20) 11 6 45
11 6.8 37
118 48 59
For (4.39) 96 67 30
274 192 32
For (4.55) 191 90 53
467 252 46
495 276 44
Looking at the proﬁle summary obtained fromthe MATLAB, we notice that “care”
is called 2164, 2479, and 17948 times when (4.20) is solved on interval [0, 30]. It
is called 12352 and 34406 times when (4.39) is solved on interval [0, 30] and it is
called 12406, 34547, and 34405 times when (4.55) is solved on interval [0, 30] for
each K
c
(
ˆ
ψ) and K
o
(
ˆ
ψ). The times for the oﬄine computations of the Fourier series
coeﬃcients are 0.1, 0.13, and 0.15 seconds, respectively. This clearly means that
the FSI method which requires only simple arithmetic for online computation of the
solution of the CARE is computationally much faster than “care” which uses the Schur
decomposition method to solve the Riccati equation.
Chapter 6
PortHamiltonian Formulation
and Passivity Based Control
Design
1
T
nis chapter focuses on the design of the control law for dynamic positioning as
suming that the ﬁltering and estimation problems have been solved, i.e., the
measured output contains only the noise free LF position and heading measurements
and the disturbances vector is either known or its estimate is available. The Port
Hamiltonian framework is a very useful concept which oﬀers more capacity to im
prove the performance of the controllers and to analyze stability properties. Passivity
based techniques were ﬁrst introduced in [58] and have been used since then in many
applications, for instance in underwater vehicle [3], the control of electrical machines
[19], the design of power system stabilizers [56], underactuated mechanical systems
[59], power converters [64], and magnetic levitation systems [65].
A nice feature of the passivitybased control design is the physical meaning of the
resulting control laws and concepts, such as storage energy or dissipation, which play
a fundamental role in analyzing stability and the performance of the controller. Sta
bility properties, based on the Lyapunov theory, can easily be studied for the closed
loop systems obtained. The passivity based approach has two versions. The classical
Euler Lagrange (ELPBC) allows the change in the potential energy function and the
generalized inertia matrix, see [60]. In the last decade, the Interconnection and Damp
ing Assignment Passivity Based Control (IDAPBC) methodology has emerged as an
easy and a (quasi) stepbystep methodology to obtain passivity based controllers, see
for instance [60]. IDAPBC also allows changes in the interconnection matrix.
This chapter introduces the portHamiltonian model of a vessel with 3 degrees
of freedom in a horizontal plane. The obtained portHamiltonian equations are the
1
This work has been published in the proceedings of the Dynamic Positioning Conference DP2010 in
Houston, see [50]. It is also accepted for publication at the IET Journal of Control Theory and Applications,
see [51].
83
84 Chapter 6. pH Formulation and PBC Design
starting point for the design of controllers using the IDAPBC technique. The main
contribution of this chapter is a family of passivitybased controllers, in the port
Hamiltonian framework, which uses the energy shaping of the closed loop system
to ensure (local/global) asymptotic stability. The resulting control laws are of output
feedback form and are robust in presence of eventual disturbances.
6.1 HamiltonianBased Control
6.1.1 PortHamiltonian Modeling
A large class of physical systems of interest in control applications can be modeled
in the general form of portHamiltonian systems (PHS), see [79]. PHS generalize
the Hamiltonian formalism of classical mechanics to physical systems connected in
a powerpreserving way and encodes the detailed energy transfer and storage in the
systems, and is thus suitable for control schemes based on IDAPBC.
A PHS can be written, in an implicit form, as
˙ x = (_ − F)∂H(x) + g(x)u, (6.1)
where x ∈ R
n
is the state (or Hamiltonian variables) vector, _(x) ∈ R
n×n
is the skew
symmetric interconnection matrix (_ = −_
T
), F(x) ∈ R
n×n
is the dissipation (or
damping) matrix (symmetric positive semideﬁnite, F = F
T
≥ 0), g(x) ∈ R
n×m
is
the external connection matrix, u ∈ R
m
is the control input vector, and H(x) is the
Hamiltonian (energy or storage) function
2
. In this formulation, the matrix F describes
the energy losses of the system, the interconnection matrix _ describes the ﬂow of
energy inside the system, and the port matrix g describes the ﬂow of energy from
outside (for instance from the controller) to the system. In a passivity based control
design approach, the controller can be considered as a source injecting energy into the
system.
The socalled passive output, y ∈ R
m
, is given by
y = g
T
(x)∂H(x), (6.2)
and the product u
T
y usually has unity of power. In the literature on portHamiltonian
systems, the socalled passive output is a relative degree one output.
6.1.2 The IDAPBC Technique
The Interconnection and Damping Assignment Passivity Based Control (IDAPBC),
[60], is a technique for designing controllers based on the portHamiltonian frame
work. It uses the passive stability properties to ensure the convergence of the system
to the desired ﬁxed point.
The main idea behind the IDAPBC is to deﬁne a new closed loop (or target)
system with a Hamiltonian structure. The design problem summarizes into ﬁnding a
2
The ∂
x
(or ∂, if no confusion arises) operator deﬁnes the gradient of a function of x, and in what follows
we will take it as a column vector.
6.1. HamiltonianBased Control 85
control law such that the system behaves as
˙ x = (_
d
− F
d
)∂H
d
, (6.3)
where _
d
(x) = −_
T
d
(x), F
d
(x) = F
T
d
(x) ≥ 0 and H
d
(x) has a minimum at the desired
regulation point x
d
, x
d
= arg min(H
d
(x)). The stability of this system can easily be
proved by using H
d
as a Lyapunov function (
˙
H
d
(x) = −(∂H
d
)
T
F
d
∂H
d
≤ 0, see for
instance, [57] and [60] for a detailed discussion).
The design procedure reduces to ﬁnding matrices _
d
(x) and F
d
(x) and a desired
closed loop energy function H
d
(x), which solve the socalled matching equation
(_ − F)∂H + gu = (_
d
− F
d
)∂H
d
. (6.4)
Then, the control law becomes
u = (g
T
g)
−1
g
T
((_
d
− F
d
)∂H
d
− (_ − F)∂H). (6.5)
A drawback of the IDAPBC controllers is that they are, in general, not able to
reject disturbances. To remove this discrepancy of the control design, usually a dy
namic extension of the system is introduced to obtain an integral action on the output
error. Extension of the closed loop dynamics in the IDAPBC framework can be done,
in a natural way, only for passive outputs, see [57]. A completely diﬀerent approach
addresses nonpassive outputs (or higher relative degree one outputs). In this case, a
Hamiltonian based controller with an integral action can be obtained via a change of
variables, see [18].
Let us assume that the x
o
∈ R
r
are the higher relative degree one (or nonpassive)
outputs. The main idea is to introduce a new variable z
e
∈ R
r
, which is used to enforce
the equilibrium point of the closed loop system to the desired one, and a change of
variables z = f (x, z
e
) ∈ R
n−r
to cast the target system in a Hamiltonian structure as
follows
_
¸
¸
¸
¸
¸
¸
¸
¸
_
˙ x
o
˙ z
˙ z
e
_
¸
¸
¸
¸
¸
¸
¸
¸
_
=
_
¸
¸
¸
¸
¸
¸
¸
¸
_
J
o
− R
o
J
zo
− R
zo
J
e
−J
T
zo
− R
T
zo
J
z
− R
z
O
3
−J
T
e
O
3
O
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
∂H
de
. (6.6)
The powerpreserving interconnection structure of the proposed target system is de
ﬁned by J
o
= −J
T
o
∈ R
r×r
, J
z
= −J
T
z
∈ R
(n−r)×(n−r)
, J
zo
∈ R
r×(n−r)
and J
e
∈ R
r×r
.
Dissipation is given by R
o
= R
T
o
∈ R
r×r
, R
z
= R
T
z
∈ R
(n−r)×(n−r)
, and R
zo
∈ R
r×(n−r)
.
Finally, the Hamiltonian function, H
de
takes the form
H
de
= H
d
(x
o
, z) +
1
2
z
T
e
K
e
z
e
, (6.7)
where H
d
has a minimum at the desired regulation point (x
d
o
, f (x
d
o
, 0), 0).
Following the idea of the IDAPBC technique, the stability is guaranteed if the
dissipative matrix is positive semideﬁnite, i.e., R
o
≥ 0, R
z
> 0 and R
o
−R
zo
R
−1
z
R
T
zo
≥ 0.
The key point of the PHS structure in (6.6) is that, the existence of a minimum of
H
de
in x
d
o
, implies ∂
x
o
H
d
¦
x
d
o
= 0 which, evaluated in the z
e
dynamics,
˙ z
e
= −J
T
e
∂
x
o
H
d
(6.8)
ensures that x
d
o
is an equilibrium point.
86 Chapter 6. pH Formulation and PBC Design
6.2 Dynamic Positioning Problem
A dynamic positioning (DP) system is a computer controlled system which automati
cally maintains a vessel’s position and heading by using propellers and thrusters. The
computer program contains a mathematical model of the vessel which includes infor
mation pertaining to the wind and current drag of the vessel and the location of the
thrusters. This knowledge, combined with the sensor information, helps the computer
calculate the required steering angle and thruster output for each thruster.
Generally, in DP problems, only position and heading measurements are available.
This leads to the use of observers to estimate the state (mainly the velocities and
the bias term) which are required for feedback into the control law. This problem
is studied in many papers. Some examples include a nonlinear observer designed
in [24], a passivitybased scheme considered in [27] and [47], and the Luenberger
observer used in [71].
Furthermore, the measured position and heading signals are noisy and, also, with
two diﬀerent frequency components (see Figure 2.5, Chapter 2). The total ship motion
can be seen as a superposition of a low frequency component (due to the wind, sea
currents and thruster forces and moments) and an oscillatory term (the socalled wave
induced wave frequency motion), which represents the eﬀect of the waves. See [22]
for more details.
However, DP only considers the slow variations and, consequently, the motion due
to the waves should be removed before it enters in the controller algorithm. Kalman
ﬁltering techniques were proposed in [28] and [32]. See [25] for a recent overview. As
pointed out in [27], Kalman ﬁlters require the use of a linear model, and the nonlinear
motion should be linearized at various operation points. To overcome this drawback, a
wavefrequency observer is added to compensate the wave disturbances, see [27] and
[47].
Due to the important role of the estimation and ﬁltering process, the motion control
system in the DP problem can be grouped in two basic subsystems: the observer
system (or wave ﬁlter), and the controller, see Figure 6.1.
Figure 6.1: Basic scheme of components of a ship motioncontrol system.
Various controllers have been proposed to stabilize the ship to the desired position.
PI controllers are often used [22], however more advanced techniques are applied to
6.3. Ship Model in PortHamiltonian Framework 87
this problem resulting in interesting control algorithms. Backstepping design, which
also includes the observer stage, are presented in [24] and [69]. In [47] the stability
of a PDtype controller with a passive observer is proved using a separation principle
argument. Recently, sampleddata control theory has also been applied to the DP
problem for designing control law [39]. As mentioned at the start of this chapter, we
assume that the estimation and ﬁltering problem has been solved and our focus is the
control design for dynamic positioning using the portHamiltonian framework.
6.3 Ship Model in PortHamiltonian Framework
6.3.1 Ship Model in Cartesian Coordinates
We recall here the vessel model from Chapter 2. In this chapter, we consider a model
in the following form.
_
˙ η
˙ ν
_
=
_
O
3
J(ψ)
O
3
−M
−1
D
_ _
η
ν
_
+
_
O
3
M
−1
_
τ +
_
O
3
M
−1
J
T
(ψ)
_
b. (6.9)
The details of the vectors and the matrices in (6.9) are given in Chapter 2. The en
vironmental disturbances due to the sea currents, waves, and wind are represented by
b = [b
1
b
2
b
3
]
T
∈ R
3
in the Earthﬁxed reference frame. This bias term is constant
in the Earthﬁxed reference frame, under assumption of constant or slowly varying
currents.
In this chapter, it is assumed that these natural eﬀects (sometimes called bias forces
and moments), which can also be modeled as a ﬁrstorder Markov process, [22, Chap
ter 3, page:89], are either known or an estimate of the bias vector is available. We
assume that the measurement system gives us noise free position and orientation mea
surements and that the wave frequency (WF) components from the measured output
are ﬁltered or estimated. Hence, in this chapter, we skip the dynamics of the bias and
the WF components.
The main goal in the dynamic positioning problem is to stabilize the ship in a given
ηcoordinate. Without loss of generality, our objective is to design an appropriate con
trol law τ which stabilizes the system to the origin (x, y, ψ) = (0, 0, 0). Additionally, as
the measurement of the relative velocity vector is not available, the control law should
be independent of ν, and must be able to reject unknown disturbances or uncertainties.
6.3.2 Ship Model in PortHamiltonian Coordinates
We can write the system described in (6.9) in a PHS form (6.1) by using as a state
x
T
= [q
T
, p
T
] ∈ R
6
, where q = [q
1
q
2
q
3
]
T
∈ R
3
represents the Earthﬁxed position
and heading, and the momentum p = [p
1
p
2
p
3
]
T
∈ R
3
, is deﬁned as p = Mν.
Substituting η = q and ν = M
−1
p in (6.9), we get the following system
˙ x = (_(q
3
) − F)∂H + g
τ
τ + g
b
(q
3
)b (6.10)
with the following interconnection and damping matrices
_(q
3
) =
_
O
3
J(q
3
)
−J
T
(q
3
) O
3
_
, F =
_
O
3
O
3
O
3
D
_
, (6.11)
88 Chapter 6. pH Formulation and PBC Design
the external connection matrices
g
τ
=
_
O
3
I
3
_
, g
b
(q
3
) =
_
O
3
J
T
(q
3
)
_
, (6.12)
and the Hamiltonian function given by
H =
1
2
p
T
M
−1
p. (6.13)
Note that the Hamiltonian function contains only a kinetic energy term, associated
with the momentum variable. A potential energy, artiﬁcially added by the controller
will play a key role to stabilize the ship in the desired position. From (6.2), we observe
that the passive output for the system (6.10) is the velocity vector which does not
correspond to the actual output of the system, the position and the orientation. This
is an important consideration for the control design, especially for the dynamic IDA
PBC control design in Section 6.5.
Two diﬀerent energy functions are proposed in this chapter. We start by illustrat
ing the methodology using a quadratic and a trigonometric Hamiltonian function and
recover a simple classical IDAPBC controller which guarantees asymptotic stability.
The energy shaping, based on a trigonometric function, improves the heading control.
These controllers do not produce the desired regulation properties in the presence of
unknown disturbances. Consequently, in order to achieve the desired performance, a
dynamic extension is proposed, and it results in a control law that can be interpreted
as a nonlinear version of the conventional PID controller. A salient feature of the pro
posed controllers is that they do not require the relative velocity measures and, thanks
to a dynamic extension, they also ensure a good regulation behavior even in presence
of disturbances or unknown (or nonestimated) terms.
6.4 Classical IDAPBC Design
A family of output feedback controllers can be obtained via the IDAPBC methodol
ogy. In the design process, a nominal case is considered (i.e., where the disturbances
are assumed to be completely known), and then the stability against an unknown dis
turbance vector is also analyzed.
As presented in Section 6.1, the control laws are obtained from matching the dy
namical system (6.10) with the target dynamics (6.3). To solve this, the desired inter
connection matrix is ﬁxed as in (6.10), i.e.,
_
d
=
_
O
3
J(q
3
)
−J
T
(q
3
) O
3
_
, (6.14)
the dissipation matrix set as
F
d
=
_
O
3
O
3
O
3
R
p
_
, (6.15)
where R
p
∈ R
3×3
is a symmetric positive deﬁnite matrix (R
p
= R
T
p
> 0), and the closed
loop energy function, H
d
(x), is shaped as
H
d
(q, p) = Ψ(q) +
1
2
p
T
M
−1
p, (6.16)
6.4. Classical IDAPBC Design 89
where Ψ(q) has a minimum at the origin, i.e., ∂
q
Ψ¦
q
∗
=0
= 0 and ∂
2
q
Ψ¦
q
∗ = 0 > 0. This
implies that the desired energy function has the minimum at the desired stabilizing
point (q
∗
, p
∗
) = (0, 0). From a physical point of view, the controller adds some po
tential energy, in the q coordinates, with respect to the original Hamiltonian function
(6.13).
From the resulting matching equation (6.4), the equality corresponding to the ﬁrst
row is automatically satisﬁed while, from the second equality, we obtain the following
control law
τ = −J
T
(q
3
)(∂
q
Ψ + b) − (R
p
− D)M
−1
p. (6.17)
From the q dynamics in (6.10), we get p = MJ
T
(q
3
) ˙ q, and deﬁning K
D
:= R
p
−D,
the state feedback algorithm (6.17) takes the form
τ = − J
T
(q
3
)∂
q
Ψ − K
D
J
T
(q
3
) ˙ q − J
T
(q
3
)b, (6.18)
which can be seen as a nonlinear output feedback PD (ProportionalDerivative) con
troller with a feedforward term, −J
T
(q
3
)b. Notice that with the choice R
p
= D, the
controller simpliﬁes because K
D
= 0. But, we keep the general result because, as
we point out in the simulations that by increasing the dissipation of the system, the
performance of the closed loop system improves considerably .
Proposition 6.4.1. Consider the dynamical system (6.10) in a closed loop with the
control law (6.18), where the origin q = 0 is a local minimum of Ψ(q), and that
the bias vector b and the matrix D are known. Then, the desired regulation point
(q
∗
, p
∗
) = (0, 0) is locally asymptotically stable. Furthermore, if q = 0 is the global
minimum of Ψ(q), then (q
∗
, p
∗
) = (0, 0) is globally asymptotically stable.
Proof. Substituting (6.18) in (6.10), we get a PHS in the form (6.3) with (6.14) and
(6.15) as follows.
˙ x =
_
O
3
J(q
3
)
−J
T
(q
3
) −D
_ _
O
3
M
−1
p
_
+
_
O
3
I
3
_
_
−J
T
(q
3
) −K
D
J
T
(q
3
)
_
∂H
d
+
_
O
3
−J
T
(q
3
)
_
b +
_
O
3
J
T
(q
3
)
_
b
=
_
J(q
3
)M
−1
p
−J
T
(q
3
)∂
q
H
d
− R
p
∂
p
H
d
_
=
_
O
3
J(q
3
)
−J
T
(q
3
) −R
p
_ _
∂
q
H
d
∂
p
H
d
_
=(_
d
− F
d
)∂H
d
. (6.19)
Using the Hamiltonian function (6.16) as a Lyapunov function, we get
˙
H
d
=(∂H
d
)
T
˙ x
=(∂H
d
)
T
_
O
3
J(q
3
)
−J
T
(q
3
) −R
p
_
∂H
d
=
_
(∂
q
Ψ)
T
(M
−1
p)
T
_
T
_
O
3
O
3
O
3
−R
p
_ _
∂
q
Ψ
M
−1
p
_
= − p
T
(M
−1
)
T
R
p
M
−1
p ≤ 0. (6.20)
90 Chapter 6. pH Formulation and PBC Design
Since
˙
H
d
is negative semideﬁnite, equilibrium point theorems do not establish asymp
totic stability of the closed loop system (6.19). However, invoking LaSalle’s invari
ance principle, asymptotic stability can be proved.
The set of points at which
˙
H
d
vanishes is
E
l
=¦(q, p) : p = 0 and q is free¦. (6.21)
The invariant points of the closed loop system (6.19) are given by
M
l
=¦(q
∗
, p
∗
) : p
∗
= 0 and q
∗
is such that ∂
q
Ψ(q)¦
q
∗ = 0¦. (6.22)
In case, q
∗
= 0 is the global minimum of the function Ψ(q) then the set M
l
consists of
only the origin (q
∗
, p
∗
) = (0, 0). This means that the largest invariant subset M
l
of E
l
is
a singleton set. Hence the closed loop system(6.19) converges globally asymptotically
to the desired point. In case, q
∗
= 0 is only a local minima of Ψ(q) then the set
M
l
may consists of multiple points. The closed loop system then converges locally
asymptotically to the set M
l
.
The role of the energy function in this study is similar to that of a Lyapunov func
tion. From the Lyapunov stability theory, we know that the stability properties of a
dynamical system and the minima of the Lyapunov function have a close connection.
In the sequel, we explain how this connection can be exploited to improve the perfor
mance of the controllers in this study. In (6.18), we have a rather general expression
for the control law depending on Ψ(q). What follows are two special cases, a quadratic
and a trigonometric, of the control law depending upon two diﬀerent energy shapings.
6.4.1 A Quadratic Energy Shaping
The simplest function with a global minimum at the origin has a quadratic form:
Ψ(q) =
1
2
q
T
Kq, where K = K
T
> 0 is a gain matrix. It is easy to see that q = 0
is the only minimum of Ψ(q) and the desired energy function (6.16) becomes
H
d1
(q, p) =
1
2
q
T
Kq +
1
2
p
T
M
−1
p, (6.23)
which implies, from (6.18), the following control law
τ = −J
T
(q
3
)(Kq + b) − K
D
J
T
(q
3
) ˙ q. (6.24)
Using Proposition 6.4.1, we can conclude that the closed loop system (6.10) with
(6.24), is globally asymptotically stable.
6.4.2 A Trigonometric Energy Shaping
Inspired by the energy function of a pendulum, we propose to shape the desired Hamil
tonian containing a trigonometric function with the form
Ψ(q) =
1
2
q
T
12
C
12
q
12
+ c
3
(1 − cos q
3
), (6.25)
6.4. Classical IDAPBC Design 91
where q
12
= [q
1
q
2
]
T
, C
12
= diag¦c
1
, c
2
¦ and c
1
, c
2
, c
3
> 0. This function, has multiple
local minimums at q = (0, 0, a2π), where a ∈ Z. With this choice the desired energy
function (6.16) becomes
H
d2
(q, p) =
1
2
q
T
12
C
12
q
12
+ c
3
(1 − cos q
3
) +
1
2
p
T
M
−1
p. (6.26)
Figure 6.2 illustrates the diﬀerence between the two proposed energy shapings (6.23)
and (6.26) in terms of the level surfaces.
H
d1
H
d2
q
1
H
d
( q
1
, q
3
)
q
3
q
1
Figure 6.2: Comparison between the quadratic and the trigonometric functions, H
d1
and H
d2
, respectively, in the coordinates q
1
and q
3
.
The main motivation for this kind of energy shaping is that, for certain applications
where there are no constraints (for instance, links with external objects), stabilization
in q
3
= 0 or q
3
= 2π is exactly the same. Figure 6.3, shows a possible scenario, where
the path for stabilizing in q
3
= 2π is shorter than stabilizing in q
3
= 0.
The Hamiltonian (6.26) with the target closed loop system deﬁned by (6.14) and
(6.15), implies that the control law (6.18) takes the ﬁnal form as
τ = −J
T
(q
3
)C
_
¸
¸
¸
¸
¸
¸
¸
¸
_
q
1
q
2
sin q
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
− K
D
J
T
(q
3
) ˙ q − J
T
(q
3
)b, (6.27)
where C = diag¦c
1
, c
2
, c
3
¦ is a positive deﬁnite gain matrix.
As in the previous subsection, using Proposition 6.4.1, we can conclude that the
closed loop system (6.10) with (6.27), is (locally) asymptotically stable.
92 Chapter 6. pH Formulation and PBC Design
Trigonometic path Quadratic path
y
y
r
x
r
Figure 6.3: A possible scenario when it is advantageous to use the trigonometric en
ergy function instead of the quadratic energy function.
6.5 Extended IDAPBC Design
6.5.1 Motivating Problem
The control law (6.18) assumes that b is known and that (6.10) perfectly models the
ship motion. In practical situations, the ship model only represents some simple dy
namics of the actual systemand furthermore the bias vector, b, has to be estimated. Let
us study the inﬂuence of unmodeled behaviors, wrong estimations or, in the general
case, the presence of disturbances.
In order to analyze the performance of the proposed passivitybased controller, we
study the following perturbed system
_
˙ q
˙ p
_
=
_
O
3
J(q
3
)
−J
T
(q
3
) −D
_
∂H +
_
O
3
I
3
_
τ +
_
O
3
J
T
(q
3
)
_
b +
_
O
3
I
3
_
∆, (6.28)
where ∆ ∈ R
3
is a vector which represents the disturbances. As a ﬁrst approximation,
and only to motivate the use of the extended dynamics, we consider this vector as a
constant.
The system (6.28) in a closed loop with (6.18) results in the following system
_
˙ q
˙ p
_
=
_
O
3
J(q
3
)
−J
T
(q
3
) −R
p
_
∂H
d
+
_
O
3
I
3
_
∆, (6.29)
which has an equilibrium point which satisﬁes ∂Ψ¦
q=q
∗ = J(q
∗
3
)∆ and p
∗
= 0. It
implies that in presence of disturbances, the closed loop system (6.28) with control
law (6.18) has a diﬀerent equilibria than (0, 0). However, stability should also be
further analyzed.
6.5. Extended IDAPBC Design 93
Moreover, in a special case where the disturbances are also aﬀected by J
T
(q
3
)
given by
_
˙ q
˙ p
_
=
_
O
3
J(q
3
)
−J
T
(q
3
) −D
_
∂H +
_
O
3
I
3
_
τ +
_
O
3
J
T
(q
3
)
_
b +
_
O
3
J
T
(q
3
)
_
∆, (6.30)
it can be seen that some stability properties still remain. With the control law (6.24),
global asymptotic stability can be proved by shifting the Hamiltonian function to the
new equilibrium point, i.e., with
˜
H
d1
=
1
2
(q − q
∗
)
T
K(q − q
∗
) +
1
2
p
T
M
−1
p (6.31)
as a Lyapunov function, where q
∗
= K
−1
∆. See Appendix A for more details.
Similarly, the control law (6.27) derived by using the trigonometric energy shaping
in (6.30), gives the following closed loop system
_
˙ q
˙ p
_
=
_
O
3
J(q
3
)
−J
T
(q
3
) −R
p
_ _
∂
q
H
d2
∂
p
H
d2
_
+
_
O
3
J
T
(q
3
)
_
∆, (6.32)
where ∂
q
H
d2
= C
_
¸
¸
¸
¸
¸
¸
¸
¸
_
q
1
q
2
sin q
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
and ∂
p
H
d2
= M
−1
p. The new set of equilibrium points of
(6.32) is
q
∗
=
_
∆
1
c
1
,
∆
2
c
2
, arcsin
_
∆
3
c
3
_
− 2kπ
_
T
, and p
∗
= 0, k ∈ Z. (6.33)
The new set of equilibrium points (6.33) is not a set of minima of (6.26). Unlike the
quadratic case, shifting of (6.26) to the new equilibrium (6.33) does not make them a
set of minima of the shifted Hamiltonian function. This is because of the presence of
transcendental term in (6.26). Therefore, we require another Hamiltonian function for
the stability analysis of (6.32).
We seek another Hamiltonian function (energy shaping)
˜
H
d2
(q, p) which has a
minimum value at (q
∗
, 0), i.e., we require
∂
q
˜
H
d2
(q, p)¦
q
∗ = 0, ∂
p
˜
H
d2
(q, p)¦
p
∗ = 0, and ∂
2
˜
H
d2
(q, p)¦
(q
∗
, p
∗
)
> 0. (6.34)
Let us deﬁne
˜
H
d2
(q, p) =
1
2
c
1
(q
1
−
∆
1
c
1
)
2
+
1
2
c
2
(q
2
−
∆
2
c
2
)
2
+ c
3
˜
ψ
3
(q
3
) +
1
2
p
T
M
−1
p, (6.35)
where
˜
ψ
3
(q
3
) is such that
∂
q
3
˜
ψ
3
¦
q
∗
3
= 0,
⇒ ∂
q
3
˜
ψ
3
= sin q
3
−
∆
3
c
3
.
Integrating both sides,
˜
ψ
3
(q
3
) = −cos q
3
−
∆
3
c
3
q
3
+ e
1
,
94 Chapter 6. pH Formulation and PBC Design
where e
1
is a constant of integration. Without loss of generality, we take e
1
= 1, then
˜
ψ
3
(q
3
) = 1 − cos q
3
−
∆
3
c
3
q
3
. (6.36)
Therefore, the new desired energy function is
˜
H
d2
(q, p) =
1
2
c
1
_
q
1
−
∆
1
c
1
_
2
+
1
2
c
2
_
q
2
−
∆
2
c
2
_
2
+ c
3
_
1 − cos q
3
−
∆
3
c
3
q
3
_
+
1
2
p
T
M
−1
p.
(6.37)
The Hessian of
˜
H
d2
(q, p) is given by
∂
2
˜
H
d2
=
_
∂
2
q
˜
H
d2
∂
2
qp
˜
H
d2
∂
2
pq
˜
H
d2
∂
2
p
˜
H
d2
_
=
_
∂
2
q
˜
H
d2
O
3
O
3
M
−1
_
,
where
∂
2
q
˜
H
d2
=
_
¸
¸
¸
¸
¸
¸
¸
¸
_
c
1
0 0
0 c
2
0
0 0 c
3
cos q
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
.
Using the Schur complement argument, the Hessian, ∂
2
˜
H
d2
¦
(q
∗
, p
∗
)
is positive deﬁnite
iﬀ ∂
2
q
˜
H
d2
¦
(q
∗
, p
∗
)
> 0. This requires
c
3
cos q
∗
3
> 0,
⇒ c
3
cos(sin
−1
∆
3
c
3
) > 0,
⇒
_
1 − (
∆
3
c
3
)
2
> 0,
⇒
∆
3
c
3
< 1.
Using the LaSalle’s principle, local asymptotic stability of this new set of equilibria
can be proved with the Hamiltonian function
˜
H
d2
(q, p) =
1
2
(q
12
−q
∗
12
)
T
C
12
(q
12
−q
∗
12
)+c
3
_
1 − cos q
3
− q
3
∆
3
c
3
_
+
1
2
p
T
M
−1
p, (6.38)
which has local minima if
∆
3
c
3
< 1. More details of this result have been explained in
Appendix A. Figure 6.4 explains this condition for existence of the set of minima for
the energy function (6.38).
Summarizing, the presence of unknown disturbances results in a bad positioning
of the ship. Although a high gain in the K and C matrices (in controllers (6.24) and
(6.27), respectively) implies stabilization close to the desired equilibrium point. This
may not be a rational strategy due to practical limitations of the propulsion units.
This discrepancy in the desired performance of the control law motivates the use of a
dynamic extension in order to achieve the stabilization at the desired point.
6.5. Extended IDAPBC Design 95
Figure 6.4: Visualization of the condition for existence of the set of minima of (6.38).
6.5.2 Target Extended System
In [4], it has been proved that if a system cannot be stabilized by using the classical
IDAPBC approach then it is impossible to stabilize the system by dynamic extension.
This result is not restrictive in our case; our systemis stabilizable but the discrepancy is
that it does not stabilize to the desired equilibrium point. To overcome this drawback,
we introduce an integral action by dynamic extension.
A dynamic extension for a non passive output, maintaining the portHamiltonian
structure, is possible by means of a change of coordinates. The controller is designed
for a nominal case, without disturbances, and then the presence of unknown terms
is analyzed. Following the idea in [18], we introduce a new state variable, z
e
∈ R
3
,
which is used to enforce the equilibrium point of the closed loop system to the desired
one, and a change of variables z = f (q, p, z
e
). We deﬁne the target system, with the
form of (6.6), as
_
¸
¸
¸
¸
¸
¸
¸
¸
_
˙ q
˙ z
˙ z
e
_
¸
¸
¸
¸
¸
¸
¸
¸
_
=
_
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
J(q
3
) J(q
3
)
−J
T
(q
3
) −R
z
O
3
−J
T
(q
3
) O
3
O
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
_
∂
q
H
de
∂
z
H
de
∂
z
e
H
de
_
¸
¸
¸
¸
¸
¸
¸
¸
_
, (6.39)
where R
z
= R
T
z
≥ 0, is a 3 × 3 matrix to be deﬁned. The desired Hamiltonian function
is deﬁned as
H
de
(q, z, z
e
) = Ψ(q) +
1
2
z
T
M
−1
z +
1
2
z
T
e
K
e
z
e
(6.40)
96 Chapter 6. pH Formulation and PBC Design
where K
e
= diag¦k
e1
, k
e2
, k
e3
¦ > 0. As earlier in the classical IDAPBC design, Ψ(q)
must be designed with a minimum in the desired regulation point. Then, the Hamilto
nian function (6.40) has a minimum at (q
∗
, z
∗
, z
∗
e
) = (q
∗
, 0, 0).
Matching the q dynamics, from (6.10) and (6.39), the change of variables z is
deﬁned as
z = p − MK
e
z
e
. (6.41)
The state feedback control law is obtained from the second row of (6.39) and the time
derivative of (6.41),
τ = −K
P
J
T
(q
3
)∂
q
Ψ − K
D
M
−1
p + K
I
z
e
− J
T
(q
3
)b (6.42)
˙ z
e
= −J
T
(q
3
)∂
q
Ψ (6.43)
where we deﬁne
K
P
:= (MK
e
+ I
3
), (6.44)
K
D
:= R
z
− D, (6.45)
K
I
:= K
e
. (6.46)
Similarly to the classical IDAPBC controller in the previous section, using p =
MJ
T
(q
3
) ˙ q, the control law (6.42)(6.43) takes the following form
τ = −K
P
J
T
(q
3
)∂
q
Ψ − K
I
_
J
T
(q
3
)∂
q
Ψdt − K
D
J
T
(q
3
) ˙ q − J
T
(q
3
)b (6.47)
which has the same structure as a nonlinear PID controller with a feedforward term,
−J
T
(q
3
)b.
Proposition 6.5.1. Assume that q is measurable, and that the disturbances vector b
and the matrices M and D are known. If K
e
= diag¦k
e1
, k
e2
, k
e3
¦ > 0, and Ψ(q) has
a (local) minimum at the origin, q = 0, then the system (6.10) in a closed loop with
(6.47), is (locally) asymptotically stable at the point (q, z, z
e
) = (0, 0, 0).
Furthermore, if q = 0 is a global minimum of Ψ(q), then the origin of (6.47) is
globally asymptotically stable.
Proof. The closed loop system (6.10) with (6.47) takes the form of (6.39). Then, the
Hamiltonian function (6.40) is a Lyapunovcandidate function and its time derivative
is
˙
H
de
= −z
T
(M
−1
)
T
R
z
M
−1
z ≤ 0. (6.48)
Then, the stability can be proved invoking LaSalle’s invariance principle as we already
show in proof of Proposition 6.4.1.
Equation (6.47) is a rather general formulation of the control law. What follows
are two special cases depending on two diﬀerent formulations of the energy shaping
(6.40).
6.5. Extended IDAPBC Design 97
6.5.3 A Quadratic Energy Shaping
Let us ﬁrst take the same quadratic energy shaping as in Section 6.4, Ψ(q) =
1
2
q
T
Kq
with K = diag¦k
1
, k
2
, k
3
¦ > 0, the Hamiltonian function (6.40) becomes
H
de1
(q, z, z
e
) =
1
2
q
T
Kq +
1
2
z
T
M
−1
z +
1
2
z
T
e
K
e
z
e
. (6.49)
with a global minimum at the origin, (0, 0, 0). The control law (6.42)(6.43) becomes
τ = −K
P
J
T
(q
3
)Kq + K
I
z
e
− K
D
J
T
(q
3
) ˙ q − J
T
(q
3
)b (6.50)
˙ z
e
= −J
T
(q
3
)Kq. (6.51)
From Proposition 6.5.1 the controller (6.50)(6.51) ensures the global asymptotic sta
bility.
6.5.4 A Trigonometric Energy Shaping
The trigonometric energy shaping proposed in the previous section can also be con
sidered for the extended controller. Taking again
Ψ(q) =
1
2
q
T
12
C
12
q
12
+ c
3
(1 − cos q
3
), (6.52)
where q
12
= [q
1
q
2
]
T
and C
12
= diag¦c
1
, c
2
¦, and c
1
, c
2
, c
3
> 0 the desired Hamiltonian
function (6.40) becomes
H
de2
(q, z, z
e
) =
1
2
q
T
12
C
12
q
12
+ c
3
(1 − cos q
3
) +
1
2
z
T
M
−1
z +
1
2
z
T
e
K
e
z
e
. (6.53)
As previously mentioned, this function has multiple local minimums at q
∗
= (0, 0, a2π),
a ∈ Z. Using Proposition 6.5.1, this case ensures local asymptotic stability, and the
control law (6.42)(6.43) becomes
τ = −K
P
J
T
(q
3
)C
_
¸
¸
¸
¸
¸
¸
¸
¸
_
q
1
q
2
sin q
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
+ K
I
z
e
− K
D
J
T
(q
3
) ˙ q − J
T
(q
3
)b (6.54)
˙ z
e
= −J
T
(q
3
)C
_
¸
¸
¸
¸
¸
¸
¸
¸
_
q
1
q
2
sin q
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
(6.55)
where C = diag¦c
1
, c
2
, c
3
¦.
6.5.5 Analysis in Presence of Disturbances
The main contribution of the extended dynamics excels in the presence of unknown
disturbances. As before, let us analyze the closed loop system in the presence of
disturbances. First, we consider constant (or, at least, very slow) disturbances. Taking
the perturbed system (6.28) with the designed control law (6.47), we obtain
_
¸
¸
¸
¸
¸
¸
¸
¸
_
˙ q
˙ z
˙ z
e
_
¸
¸
¸
¸
¸
¸
¸
¸
_
=
_
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
J(q
3
) J(q
3
)
−J
T
(q
3
) −R
z
O
3
−J
T
(q
3
) O
3
O
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
_
∂
q
Ψ
M
−1
z
K
e
z
e
_
¸
¸
¸
¸
¸
¸
¸
¸
_
+
_
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
I
3
O
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
∆. (6.56)
98 Chapter 6. pH Formulation and PBC Design
From the z
e
dynamics, and due to ∂
q
Ψ¦
q
∗ = 0, this system has the following equilib
rium point
3
q
∗
= 0 (6.57)
z
∗
= MR
−1
z
∆ (6.58)
z
∗
e
= −K
e
R
−1
z
∆. (6.59)
Notice that, contrary to the classical controller, in the presence of unknown distur
bances, the origin is still the equilibrium in the position coordinates.
The question that now arises is about the stability of (6.56). A simple stability
analysis is possible using the error coordinates ˜ z = z − z
∗
and ˜ z
e
= z
e
− z
∗
e
. Shifting
the energy function (6.40) to the new equilibria, we get
˜
H
de
(q, ˜ z, ˜ z
e
) = Ψ(q) +
1
2
˜ z
T
M
−1
˜ z +
1
2
˜ z
T
e
K
e
˜ z
e
, (6.60)
and (6.56) can be written as
_
¸
¸
¸
¸
¸
¸
¸
¸
_
˙ q
˙
˜ z
˙
˜ z
e
_
¸
¸
¸
¸
¸
¸
¸
¸
_
=
_
¸
¸
¸
¸
¸
¸
¸
¸
_
O
3
J(q
3
) J(q
3
)
−J
T
(q
3
) −R
z
O
3
−J
T
(q
3
) O
3
O
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
_
∂
q
Ψ
M
−1
˜ z
K
e
˜ z
e
_
¸
¸
¸
¸
¸
¸
¸
¸
_
. (6.61)
This closedloop system still has the desired structure (6.6), with a energy function
which has a minimum at (6.57)(6.59). The asymptotic stability is automatically guar
anteed.
In a more realistic case, the disturbance could be considered as a vector which
depends on time, i.e., ∆ = ∆(t). Then, we can no longer study the stability of an
equilibrium point, and we only can expect that the solution of the system becomes
bounded due to the special structure of the closed loop system, that can be seen as a
forced oscillator with damping in the ˜ z coordinates.
6.6 Simulations
In order to test the performance of the designed controllers we performed some nu
merical simulations. For this validation we used the data of a supply ship from [24].
The (Bisscaled nondimensional [22]) matrices M and D are given by
M =
_
¸
¸
¸
¸
¸
¸
¸
¸
_
1.1274 0 0
0 1.8902 −0.0744
0 −0.0744 0.1278
_
¸
¸
¸
¸
¸
¸
¸
¸
_
and
D =
_
¸
¸
¸
¸
¸
¸
¸
¸
_
0.0358 0 0
0 0.1183 −0.0124
0 −0.0124 0.0308
_
¸
¸
¸
¸
¸
¸
¸
¸
_
.
3
Notice that for the case of the trigonometric controller, the equilibria become q
∗
= (0, 0, a2π).
6.6. Simulations 99
The bias vector has been set to b = [0.05, 0.05, 0.01]
T
. For all simulations, we consid
ered that the initial conditions of the ship are (q
1
, q
2
) = (−10, −10), and the heading
angle q
3
= 4 rad, and the desired stabilization position is the origin. The desired head
ing direction is along the positive xaxis. More generally, the starting heading angle
is set greater than π to show the ability of the socalled trigonometric controller to
stabilize to the closer minimum, in this case 2π.
6.6.1 Simulation Results for the classical IDAPBC design
In this subsection, we present the simulation results for the classical (quadratic and
trigonometric) control laws (6.24) and (6.27), respectively. The gain matrices we used
are K = C = diag¦0.05, 0.05, 0.01¦ and R
p
= diag¦0.75, 0.75, 0.1¦. Precisely, for this
system, we enlarged the damping (about one order of magnitude in the ﬁrst and third
components), to improve the performance. In Figure 6.5 we show the trajectories
of the same classical controller (the quadratic case) with extra dissipation, setting
R
p
to the values proposed before, and keeping the original damping, R
p
= D. This
comparison justiﬁes the use of the extra damping to obtain more suitable paths.
−12 −10 −8 −6 −4 −2 0 2 4 6 8
−14
−12
−10
−8
−6
−4
−2
0
2
4
6
q
1
[m]
q
2
[
m
]
Figure 6.5: Simulation results: ship position trajectories in the q
1
, q
2
plane, for the
classical IDAPBC (quadratic controller) design with extra damping (solid line) and
the original damping coeﬃcient (dashed line).
Figure 6.6 shows the trajectories of the position coordinates q
1
and q
2
, and the
heading angle, q
3
, of the quadratic and the trigonometric versions. Both controllers
stabilize the ship at the desired position and angle. The notable point is the diﬀerence
in the orientation proﬁles. While the quadratic controller stabilizes the heading angle
at q
3
= 0, the trigonometric controller does so at q
3
= 2π.
Figure 6.7 compares the trajectories in the q
1
q
2
plane for two versions of the clas
sical IDAPBC design. In both cases, the performance is similar but, even though the
controllers for the q
1
and q
2
coordinates are the same (with the same gain values),
the trajectories take diﬀerent paths. This fact is associated with the diﬀerent heading
angle trajectories.
100 Chapter 6. pH Formulation and PBC Design
0 20 40 60 80 100 120 140 160 180 200
−10
−5
0
5
q
1
[
m
]
0 20 40 60 80 100 120 140 160 180 200
−10
−5
0
5
q
2
[
m
]
0 20 40 60 80 100 120 140 160 180 200
0
2
4
6
8
q
3
[
r
a
d
]
time[s]
Quadratic
Trigonometric
Quadratic
Trigonometric
Trigonometric
Quadratic
Figure 6.6: Simulation results: ship position coordinates, q
1
and q
2
, and heading
angle, q
3
, for the quadratic (solid line) and trigonometric (dashed line) controllers of
the classical IDAPBC design.
−12 −10 −8 −6 −4 −2 0 2 4
−14
−12
−10
−8
−6
−4
−2
0
2
x [m]
y
[
m
]
Heading angle orientation
Figure 6.7: Simulation results: ship position trajectories in the q
1
q
2
plane, for the
quadratic (solid line) and trigonometric (dashed line) controllers of the classical IDA
PBC design together with the respective heading orientation.
6.7. Conclusions 101
6.6.2 Simulation Results in Presence of Disturbances
In this subsection, we present the simulation results in the presence of disturbances.
The key point is to show that the controllers from the extended IDAPBC design ap
proach are able to reject unknown terms. For this scenario we assume that the dis
turbance due to the bias term, b, is not available. Consequently, the feedforward
term, J
T
(q
3
)b, is removed in all the tested controllers. The gain matrices for the clas
sical controllers, (6.24) and (6.27), are the same as in the previous subsection. The
corresponding gain matrices used for the controllers from extended IDAPBC design,
(6.51) and (6.55), are R
z
= diag¦0.75, 0.75, 0.4¦, K = C = diag¦0.05, 0.05, 0.025¦ and
K
e
= diag¦0.01, 0.01, 0.015¦.
0 20 40 60 80 100 120 140 160 180 200
−10
−5
0
q
1
[
m
]
0 20 40 60 80 100 120 140 160 180 200
−10
−5
0
q
2
[
m
]
0 20 40 60 80 100 120 140 160 180 200
0
2
4
6
8
q
3
[
r
a
d
]
time[s]
Extended controllers
Classical controllers
Extended controllers
Classical controllers
Extended Trig.
Classical Quad.
Classical Trig .
Extended Quad.
Figure 6.8: Simulation results: ship position coordinates, q
1
and q
2
, and heading
angle, q
3
, for the both the quadratic and trigonometric versions of the controllers from
the classical and extended IDAPBC designs.
In Figure 6.8, q trajectories for the four controllers are plotted. Clearly the con
trollers from IDAPBC design steer the ship to the desired equilibrium position while
the controllers from the classical IDAPBC design fails to do so and have a steady
state error. This diﬀerence between the performance of both the controllers can also
be seen from the respective trajectory proﬁles. See Figure 6.9.
6.7 Conclusions
A passivitybased approach called IDAPBC is used to obtain a set of output feedback
controllers for the dynamic positioning of a ship. This methodology is based on the
102 Chapter 6. pH Formulation and PBC Design
−12 −10 −8 −6 −4 −2 0 2 4
−12
−10
−8
−6
−4
−2
0
2
4
q
1
[m]
q
2
[
m
]
Extended Quad.
Classical
Quad.
Extended Trig.
Classical Trig.
Figure 6.9: Simulation results: ship position trajectories in the q
1
, q
2
plane, for the
both the quadratic and trigonometric versions of the controllers from the classical and
extended IDAPBC designs.
portHamiltonian description which gives a physical interpretation of the dynamical
systems. Under this point of view, the controller design problem is addressed as to
shape the energy function of the closed loop system. After a general formulation we
propose two diﬀerent controllers: ﬁrst with a quadratic energy function and second,
inspired by the physics of a pendulum, with a trigonometric energy function. Also, the
presence of disturbances is studied and it turns out that the control laws obtained by
using the classical IDAPBC methodology, do not stabilize the system at the desired
position. This discrepancy is the starting point for a second set of controllers which
consist of a dynamic extension of the system which provides stability at the desired
regulation point, also in presence of disturbances. Simulations are done to validate
and compare the performance of the controllers designed.
It is worth mentioning that the obtained control laws, with a general form of state
feedback, can be easily converted to output feedback algorithms that only requires
the position measurement. Furthermore, they exhibit a simple structure that can be
interpreted as nonlinear versions of PID controllers. An important observation is that
the trigonometric energy shaping improves the heading angle control. We also observe
that by keeping the diﬀerential term, which physically increases the damping eﬀect,
in the controller formulation helps in getting a smoother trajectory proﬁle.
Future work can be aimed at determining other energy functions, Ψ(q), to improve
the performance, as well as to consider the optimization of the resulting path. Further
analysis depending on the nature of the disturbance vector (including the wave fre
quency and wind models) is required. Also, this work could be a starting point for a
new design, using the portHamiltonian perspective, of the complete motioncontrol
system (controller and observer) for a DP problem.
Chapter 7
Conclusions and
Recommendations for Future
Work
In this chapter, we summarize the main conclusions of this work and present some
recommendations for future work based on our knowledge and experiences acquired
during the process of reaching to this point. The ﬁrst major topic studied in this thesis
is the stability analysis of a special type of nonlinear system in which nonlinearity is
because of a single state variable. The study is based on the SDC framework which
transforms the nonlinear system in a pseudolinear form with a state dependent system
matrix. The stability analysis of such systems is based on the properties of the system
matrix. From the perspective of the DP system we are interested in the asymptotic be
havior of the system. The local asymptotic stability of the system is easy to establish,
the major concern always remains the analysis of the global asymptotic stability.
Fromthe ﬁrst counter example discussed in Section 3.4.1, we can conclude that the
conditions that the system matrix is continuous, pointwise Hurwitz, and exponentially
bounded are not suﬃcient to ascertain the global asymptotic stability of the pseudo
linear system. In the prototype vessel model, the system matrix is a periodic function
of the heading angle of the vessel. The periodicity assumption also ensures that the
states of the system are bounded, i.e., they will not blow up at ﬁnite time. Further
research has proved that the extended set of conditions that the system matrix is con
tinuous, pointwise Hurwitz, exponentially bounded, and periodic does not constitute
the set of suﬃcient conditions for the global asymptotic stability of the pseudolinear
system. This fact is explained in the second counter example presented in Section
3.5.1. This study concludes that more research is required to explore the set of suf
ﬁcient conditions which establish the global asymptotic stability of the pseudolinear
systems with a system matrix depending on a single state variable.
The periodicity property of the vessel model has lured us to combine the Lyapunov
stability theory with the LMIs to come up with a new approach to prove the global
103
104 Chapter 7. Conclusions and Recommendations for Future Work
asymptotic stability of the pseudolinear system which has a periodic system matrix.
This approach has proved successful to establish the global asymptotic stability of the
vessel model, in particular. Further research is required to investigate the additional
features like for instance, the best solutions of the LMI feasibility problem, the impact
of the number of LMIs on the solutions of the LMI feasibility problem, conditions for
existence and nonexistence of the solutions of the LMI feasibility problem, etc.
We present the SDARE technique for control design of the vessel model. The
design objective is to stabilize the vessel at the desired equilibrium point. The SDARE
technique gives a state feedback controller which means that the complete state of
the system should be known. It is not the case for the vessel model in this work
as only the noisy position and heading measurements are available through sensors
(GPS and gyro). So it is required to estimate the remaining states of the vessel model.
We have also used the SDARE technique to address the estimation problem. The
SDARE technique has many characteristic features. It addresses the performance of
the controller and the observer by specifying a performance index. The state and
the control weights can be adjusted to inﬂuence the performance. For instance, an
increase in the state weighting matrix Q results in a faster regulation of the vessel
at the expense of a greater control eﬀort. Generally, the SDC parametrization is not
unique and this feature oﬀers an extra degree of freedom for design and performance
improvements. While the technique oﬀers ways to use the results from linear system
theory, it maintains the nonlinear character of the system which is not the case in
linearization methods like the extended Kalman ﬁlter (EKF) etc.
The simulation results show promising performance of the SDARE technique
when applied to address the regulation problem of a DP vessel. The controller and
the observer gains can be tuned by using the weighting matrices, to obtain the desired
performance. This study is limited to the regulation problem of the DP vessel. An ex
tension of the SDARE technique to address the path following and trajectory tracking
problems of the DP vessel remains still an open research topic.
We present the FSI method which is a fast solution method for the speciﬁc type
of the SDARE associated with a pseudolinear system whose system matrix depends
on a single state variable in a periodic way. This method is based on the Fourier
series and involves the concept of interpolation. The emphasis of the FSI method is
on oﬄine computation of the Fourier series coeﬃcients. It has been shown that it
makes the online computation of the solution of the SDARE faster by reducing the
online computations. The method is particularly useful for the systems having small
time constant or fast dynamic response. For such systems, the actuators are required
to have faster dynamic response than the system they are controlling. The method was
motivated by the periodicity of the pseudolinear vessel model. The vessel system
in this work is a slow system (with large time constant), the FSI method has shown
good results in comparison with the MATLAB routine “care” which uses the Schur
decomposition to solve the SDARE. This work is limited to just one application of the
FSI method on the vessel model. The search for more applications of the FSI method
on physical systems is still open for future research.
In this work, the use of the portHamiltonian system formulation to study the con
trol design problem of the DP vessel has revealed some important and interesting
features. The Hamiltonian (storage) function is a powerful component in this formu
105
lation. We have seen that it not only simpliﬁes the stability analysis but also oﬀers
more options to improve the performance of the controller. This can be done by us
ing diﬀerent formulations of the Hamiltonian function to design diﬀerent controllers.
The performance of the controllers can then be analyzed to see which one suits the
desired performance objective. This fact is clear from the trigonometric and quadratic
formulations of of the Hamiltonian function. A better heading control of the vessel is
obtained from the trigonometric controller than the quadratic controller.
The dissipation (damping) matrix in the desired dynamics in the IDAPBC design
technique is very important from the perspective of the tuning of the controller gain
matrices. It makes it possible to manipulate the damping of the system. Through the
controller action, we can increase the damping of the system and it eventually helps
us to steer the vessel along a stable and smooth trajectory.
This study addresses only the regulation problem with passivity based approach
using the portHamiltonian formulation. With the encouraging results obtained, the
possible extension of this concept to address more involved problems like path fol
lowing, trajectory tracking, and state estimation is a challenging and interesting topic
for future research.
106
Appendix A
Proof of asymptotic stability of
(6.32) at (6.33)
The objective is to show that (6.33) is an asymptotically stable equilibrium point of
the perturbed closed loop system (6.32). For convenience of the reader, we recall both
equations mentioned above.
_
˙ q
˙ p
_
=
_
O
3
J(q
3
)
−J
T
(q
3
) −R
p
_ _
∂
q
H
d2
∂
p
H
d2
_
+
_
O
3
J
T
(q
3
)
_
∆, (A.1)
where
H
d2
(q, p) =
1
2
q
T
12
C
12
q
12
+ c
3
(1 − cos q
3
) +
1
2
p
T
M
−1
p. (A.2)
The equilibrium point of (A.1) is given by
q
∗
=
_
∆
1
c
1
∆
2
c
2
arcsin
_
∆
3
c
3
_
− 2kπ
_
T
, and p
∗
= 0, k ∈ Z. (A.3)
To prove the above mentioned asymptotic stability, we proceed as follows. Let us
consider the following closed loop system.
_
˙
˜ q
˙
˜ p
_
=
_
O
3
J(q
3
)
−J
T
(q
3
) −R
p
_ _
∂
˜ q
˜
H
d2
∂
˜ p
˜
H
d2
_
, (A.4)
where ˜ q =
_
¸
¸
¸
¸
¸
¸
¸
¸
_
q
1
− q
∗
1
q
2
− q
∗
2
q
3
_
¸
¸
¸
¸
¸
¸
¸
¸
_
, ˜ p = p, and
˜
H
d2
( ˜ q, ˜ p) =
1
2
˜ q
T
12
C
12
˜ q
12
+ c
3
(1 − cos q
3
−
∆
3
c
3
q
3
) +
1
2
˜ p
T
M
−1
˜ p. (A.5)
As a ﬁrst step, we use the LaSalle’s invariance principle to prove asymptotic sta
bility of (A.4). The energy function (A.5) is a scalar valued function and its ﬁrst partial
107
108 Appendix A
derivatives exist. Diﬀerentiating it w.r.t ‘t’, we get
˙
˜
H
d2
=(∂
˜
H
d2
)
T
˙
˜ x
=(∂
˜
H
d2
)
T
_
O
3
J(q
3
)
−J
T
(q
3
) −R
p
_
∂
˜
H
d2
= − ˜ p
T
(M
−1
)
T
R
p
M
−1
˜ p ≤ 0, (A.6)
where ˜ x =
_
˜ q
˜ p
_
. The set of points where
˙
˜
H
d2
= 0 is given by
E
l
=¦ ˜ x ∈ R
6
:
˙
˜
H
d2
= 0¦
=¦ ˜ p = 0 and ˜ q is free¦. (A.7)
The equilibrium (invariant) points of (A.4) are
˜ q
∗
=
_
˜ q
1
, ˜ q
2
, ˜ q
3
_
T
=
_
0, 0, arcsin(
∆
3
c
3
) − 2kπ
_
T
, and p
∗
= 0, k ∈ Z. (A.8)
The largest invariant subset M
l
of E
l
consists of the equilibrium points (A.8). Hence,
LaSalle’s invariance principle implies that the closed loop system (A.4) converges
locally asymptotically to either one of the equilibrium points.
As second step, we show that both the closed loop systems (A.1) and (A.4) are
equivalent. It is a straightforward observation that ˜ q and q have the same dynamics.
We show the equivalence of the dynamics of ˜ p and p.
˙
˜ p = − J
T
(q
3
)∂
˜ q
˜
H
d2
− R
p
∂
˜ p
˜
H
d2
= − J
T
(q
3
)∂
q
H
d2
− R
p
∂
p
H
d2
+ J
T
(q
3
)∆
=˙ p. (A.9)
Hence, (A.1) and (A.4) are equivalent. This implies that if (A.8) is an asymptotically
stable equilibrium point of (A.4) then (A.3) is an asymptotically stable equilibrium
point of (A.1).
Similarly, asymptotic stability of
_
˙ q
˙ p
_
=
_
O
3
J(q
3
)
−J
T
(q
3
) −R
p
_ _
∂
q
H
d1
∂
p
H
d1
_
+
_
O
3
J
T
(q
3
)
_
∆, (A.10)
where
H
d1
(q, p) =
1
2
q
T
Cq +
1
2
p
T
M
−1
p. (A.11)
w.r.t. the equilibrium point (q
∗
, p
∗
) = (−K
−1
∆, 0) follows. The asymptotic stability in
this case will be global because the point (q
∗
, p
∗
) = (−K
−1
∆, 0) is global minima of
the shifted Hamiltonian function associated with (A.11).
Appendix B
Glossary
This appendix contains deﬁnitions and additional details of some notions used in this
thesis.
BISScaling
Scaling is done to normalize the vessel steering equations of motion. This helps to
study the scaled models of the vessel and then compare and analyze the performance
of the full scale vessel models. This idea is more economical both in terms of time
and money, than doing the full scale experiments in the sea. There are two main meth
ods of scaling: the Primesystem of SNAME introduced in 1950 and the Bissystem
introduced in 1970 by Norrbin. The Primesystem uses the instantaneous velocity to
normalize the time unit. Therefore, it is not used for the DP vessels or other low speed
applications.
The Bissystem uses the length unit L
pp
(length between fore and aft perpendic
ulars), body mass density ratio µ =
m
ρ∇
, and the time unit
_
L
pp
g
as the basic scaling
parameters. The parameters µ, ρ, ∇, and g represent the level of buouncy, mass density
of the ﬂuid, hull contour displacement, and acceleration due to gravity, respectively.
The scaling parameters for other variables like velocity, acceleration, etc. can be found
in terms of these basic scaling parameters.
First order waveinduced forces
First order waveinduced wave forces (loads or disturbances) are forces whose mag
nitude is proportional to the wave height and their frequency is the same as that of the
incident waves. These forces are also called WF forces (disturbances or loads).
Gaussian white noise process
In Chapter 2, we have mentioned that the process and measurement uncertainties are
stationary Gaussian random white noise process. For the convenience of the reader,
109
110 Appendix B
some details of this concept are presented here. The vector w(t) =
_
¸
¸
¸
¸
¸
¸
¸
¸
_
w
1
(t)
w
2
(t)
w
3
(t)
_
¸
¸
¸
¸
¸
¸
¸
¸
_
consists
of three uncorrelated independent Gaussian white noise processes: w
1
(t), w
2
(t), and
w
3
(t). This implies the mean
E(w(t)) = 0, (B.1)
and covariance
E(w(t)w
T
(s)) = Wδ(t − s). (B.2)
The δ in (B.2) represents the Dirac delta (or continuous time impulse response) func
tion and is given by
δ(t − s) =
_
∞ if t = s
0 if t s
(B.3)
Kronecker Product
Let A ∈ R
m×n
and B ∈ R
p×q
. Then the Kronecker (direct or tensor) product of A and B
is deﬁned as the matrix
A ⊗ B =
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
a
11
B a
1m
B
.
.
.
.
.
.
.
.
.
a
m1
B a
mn
B
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
∈ R
mp×nq
. (B.4)
LaSalle’s Invariance Theorem (Principle)
An important result in stability theory which was ﬁrst formulated by a famous French
priest and education reformer, J. B. de LaSalle (16511719), is stated as follows
“Let Ω
0
⊂ Ω be a compact set that is positively invariant with respect to (3.2). Let
V : Ω −→ R be a continuously diﬀerentiable function such that
˙
V(x) ≤ 0 in Ω
0
. Let
E be the set of all points in Ω
0
where
˙
V(x) = 0. Let M be the largest invariant set in
E. Then every solution starting in Ω
0
approaches M as t → ∞.”
Second order waveinduced forces
Second order waveinduced wave forces (loads or disturbances) are forces whose mag
nitude is proportional to the square of the wave height. These forces are also called
LF forces (disturbances or loads).
Bibliography
[1] M. F. Aarset, J. P. Strand, and T. I. Fossen. Nonlinear vectorial observer back
stepping with integral action and wave ﬁltering for ships. Proceedings of IFAC
Conference on Control Applications in Marine Systems CAMS’98, October 20
27, 1998, Fukuoka, Japan.
[2] B. D. O. Anderson. Secondorder convergent algorithms for the steadystate
Riccati equation. International Journal of Control, 28(2):295–306, 1978.
[3] A. Astolﬁ, D. Chhabra, and R. Ortega. Asymptotic stabilization of some equi
libria of an underactuated underwater vehicle. Systems & Control Letters,
45(3):193–206, 2002.
[4] A. Astolﬁ and R. Ortega. Dynamic extension is unnecessary for stabilization
via interconnection and damping assignment passivitybased control. Systems &
Control Letters, 58(2):133–135, 2009.
[5] J. G. Balchen, N. A. Jenssen, and S. Sælid. Dynamic positioning using kalman
ﬁltering and optimal control theory. pages 183–186. Proceedings of IFAC/IFIP
Symposium on Automation in Oﬀshore Oil Field Operation,ly 611, 2008, Am
sterdam, The Netherlands.
[6] H. T. Banks, B. M. Lewis, and H. T. Tran. Nonlinear feedback controllers and
compensators: a statedependent Riccati equation approach. Computational Op
timization and Applications, 37(2):177–218, 2007.
[7] S. P. Banks and K. J. Mhana. Optimal control and stabilization of nonlinear
systems. IMA Journal of Mathematical Control and Information, 9(3):179–196,
1992.
[8] P. Benner and R. Byers. An exact line search method for solving generalized
contineoustime algebraic Riccati equation. IEEE Transactions on Automatc
Control, 43(1):101–107, 1998.
[9] W. E. Boyce and R. C. DiPrima. Elementary Diﬀerential Equations and Bound
ary Value Problems. John Wiley & Sons, Inc., Eighth edition, 2004.
111
112 BIBLIOGRAPHY
[10] T. C¸ imen. State dependent Riccati equation control: A survey. pages 3761–3775.
Proceedings of the International Federation of Automatic Control (IFAC) World
Congress, July 611, 2008, Seoul, South Korea.
[11] E. K. W. Chu, H. Y. Fan, and W. W. Lin. A structure preserving doubling al
gorithm for continuoustime algebraic Riccati equations. Linear Algebra and its
Applications, 396:55–80, 2005.
[12] J. R. Cloutier. State dependent Riccati equation technique: an overview. pages
932–936. Proceedings of American Control Conference, June 46, 1997, New
Mexico, USA.
[13] B. N. Datta. Numerical Methods for Linear Control Systems. Elsevier Academic
Press, 2003.
[14] D. F. Delchamps. Analytic feedback control and the algebraic Riccati equation.
IEEE Transactions on Automatic Control, 29(11):1031–1033, 1984.
[15] D. E. Denman and A. N. Beavers Jr. The matrix sign function and computations
in systems. Applied Mathematics and Computations, 2(1):63–94, 1976.
[16] L. Dieci. Some numerical considertions and Newton’s method revisited for solv
ing algebraic Riccati equations. 36(5):608–616, 1991.
[17] K. D. Do and J. Pan. Control of Ships and Underactuated Vehicles. Springer,
2009.
[18] A. Donaire and S. Junco. On the addition of the integral action to portcontrolled
Hamiltonian systems. Automatica, 45(8):1910–1916, 2009.
[19] A. D` oriaCerezo. Modeling, Simulation and Control of a DoublyFed Induc
tion Machine Controlled by a BacktoBack Converter. PhD thesis, Univesitat
Polit` ecnica de Catalunya, Barcelona, Spain, 2006.
[20] O. M. Faltinsen. Hydrodynamics of High Speed Marine Vehicles. Cambridge
University Press, 2005.
[21] H. Fa¨ y. Dynamic Positioning Systems: Principles, Design and Applications.
¯
Editions Technip, 1990.
[22] T. I. Fossen. Guidance and Control of Ocean Vehicles. John Wiley & Sons, Inc.,
1994.
[23] T. I. Fossen. Marine Control Systems: Guidance, Navigation and Control of
Ships, Rigs and Underwater Vehicles. John Wiley & Sons, Inc., 2002.
[24] T. I. Fossen and Å. Grøvlen. Nonlinear output feedback control of dynamically
positioned ships using vectorial observer backstepping. IEEE Transactions on
Control Systems Technology, 6(1):121–128, 1998.
[25] T. I. Fossen and T. Perez. Kalman ﬁltering for positioning and heading control
of ships and oﬀshore rigs. IEEE Control Systems Magazine, 29(6):32–46, 2009.
BIBLIOGRAPHY 113
[26] T. I. Fossen, S. I. Sagatun, and A. J. Sørensen. Identiﬁcation of dynamically
positioned ships. Modeling, Identiﬁcation, and Control, 17(2):153–165, 1996.
[27] T. I. Fossen and J. P. Strand. Passive nonlinear observer design for ship using
Lyapunov methods: Fullscale experiments with a supply vessel. Automatica,
35(1):3–16, 1999.
[28] TK. P. Fung and M. J. Grimble. Dynamic ship positioning using a selftuning
Kalman ﬁlter. IEEE Transactions on Automatic Control, 28(3):339–350, 1983.
[29] W. L. Garrard and J. M. Jordan. Design of nonlinear automatic ﬂight control
systems. Automatica, 13(5):497–505, 1977.
[30] H. Giacomini, J. Llibr, and M. Viano. On the nonexistence, existence, and
uniqueness of limit cycles. Nonlinearity, 9(2):501–516, 1996.
[31] G. Golub and C. F. Van Loan. Matrix Computations. The Johns Hopkins Uni
versity Press, Baltimore, USA, Third edition, 1996.
[32] M. J. Grimble, R. J. Patton, and D. S. Wise. Use of Kalman ﬁltering techniques
in dynamic shippositioning systems. IEE Proceedings, 127(3):93–102, 1980.
[33] R. Haberman. Elementary Applied Partial Diﬀerenrial Equations With Fourier
Series and Boundary Value Problems. Prentice Hall, 1998.
[34] K. D. Hammett, C. D. Hall, and D. B. Ridgely. Controllability issues in non
linear state dependent Riccati equation control. Journal of Guidance, Control,
and Dynamics, 21(5):767–773, 1998.
[35] A. D. Hansen. Predictive Controler Design and Identiﬁcation in Connection with
Steering Dynamics. PhD thesis, Technical University of Denmark, Denmark,
1996.
[36] K. Hiramoto, H. Doki, and G. Obinata. Optimal sensor/actuator placement for
active vibration control using explicit solution of albegraic Riccati equation.
Journal of Sound and Vibration, 229(5):1057–1075, 2000.
[37] J. Imae, H. Sagami, T. Kobayashi, and G. zhai. Nonlinear control design method
based on statedependent Riccati equation SDRE via quasinewton method. Pro
ceedings of the 43rd IEEE Conference on Decision and Control, December 14
17, 2004, Atlantis, The Bahamas.
[38] R. E. Kalman. Contributions to the theory of optimal control. Boletin Sociedad
Matem´ atica Mexicana, 5:102–119, 1960.
[39] H. Katayama. Nonlinear sampleddata stabilization of dynamically positioned
ships. IEEE Transactions on Control Systems Technology, 18(2):463–468, 2010.
[40] H. K. Khalil. Nonlinear Systems. Prentice Hall, Third edition, 2002.
[41] M. Kimura. Doubling algorithm for continuoustime algebraic Riccati equation.
International Journal of Systems Science, 20(2):191–202, 1989.
114 BIBLIOGRAPHY
[42] V. Lakshmikantham, V. M. Matrosov, and S. Sivasundaram. Vector Lyapunov
Functions and Stability Analysis of Nonlinear Systems. Kluwer, 1991.
[43] W. Langson. Optimal and suboptimal control of a class of nonlinear sys
tems. Master’s thesis, Graduate College of the University of Illinois Urbana
Champaign, USA, November 1997.
[44] W. Langson and A. Alleyne. A stability result with application to nonlinear reg
ulation. Journal of Dynamic Systems, Measurement, and Control, 124(3):452–
456, 2002.
[45] A. Laub. A Schur method for solving algebraic Riccati equations. IEEE Trans
actions on Automatic Control, 24(6):913–921, 1979.
[46] J. L¨ ofberg. YALMIP: A toolbox for modeling and optimization in MATLAB.
Proceedings of the CACSD Conference, September 24, 2004, Taipei, Taiwan.
[47] A. Loria, T. I. Fossen, and E. Panteley. A separation principle for dynamic po
sitioning of ships: theoretical and experimental results. IEEE Trans. on Control
Systems Technology, 8(2):332–344, 2000.
[48] P. K. Menon, T. M. Lam, L. S. Crawford, and V. H. L. Cheng. Real time compu
tational methods for sdre nonlinear control of missiles. Proceedings of the Amer
ican Control Conference Workshop, May 810, 2002, Anchorage, AK, USA.
[49] C. P. Mracek and J. R. Cloutier. Control designs for the nonlinear benchmark
problem via the state dependent Riccati equation method. International Journal
of Robust Nonlinear Control, 8(45):401–433, 1998.
[50] S. Muhammad and A. D` oriaCerezo. Output feedback passivity based controllers
for dynamic positioning of ships. Proceedings of Dynamic Positioning Confer
ence, October 1213, 2010, Houston, USA.
[51] S. Muhammad and A. D` oriaCerezo. Passivitybased control applied to the dy
namic positioning of ships. IET Control Theory and Applications, Accepted for
Publication.
[52] S. Muhammad and J. W. van der Woude. The Fourier series interpolation method
for the solution of the SDARE. Proceedings of the UKACC International Con
ference on Control (CONTROL 2010), September 710, 2010, Coventry, UK.
[53] S. Muhammad and J. W. van der Woude. A counter example to a recent result
on the stability of nonlinear systems. IMA Journal of Mathematical Control and
Information, 26(3):319–323, 2009.
[54] S. Muhammad and J. W. van der Woude. On stability conditions for systems
with periodic state dependent coeﬃcients: a counter example. IMA Journal of
Mathematical Control and Information, 28(1):97–102, 2011.
[55] G. J. Olsder and J. W. van der Woude. Mathematical Systems Theory. VSSD,
Third edition, 2005.
BIBLIOGRAPHY 115
[56] R. Ortega, M. Galaz, A. Bazanella, and A. Stankovic. An energyshaping ap
proach to excitation control of synchronous generators. Proceedings of the
American Control Conference, June 2527, 2001, Arligton, VA, USA.
[57] R. Ortega and E. Garc´ıCanseco. Interconnection and damping assignment
passivitybased control: A survey. European Journal of Control, 10(5):432–450,
2004.
[58] R. Ortega and M. Spong. Adaptive motion control of rigid robots: A tuttorial.
Automatica, 25(6):877–888, 1989.
[59] R. Ortega, M. W. Spong, F. G´ omezEstern, and F. Blankenstein. Interconnection
and damping assignment passivitybased control of portcontrolled Hamiltonian
systems. IEEE Transactions on Automatic Control, 47(8):1218–1233, 2002.
[60] R. Ortega, A. van der Schaft, B. Maschke, and G. Escobar. Interconnection
and damping assignment passivitybased control of portcontrolled Hamiltonian
systems. Automatica, 38(4):585–596, 2002.
[61] D. K. Parrish and D. B. Ridgely. Altitude control of a satellite using the SDRE
method. Proceedings of the American Control Conference, June 46, 1997, New
Mexico, USA.
[62] J. D. Pearson. Approximation methods in optimal control. Journal of Electronics
and Control, 13(5):453–465, 1962.
[63] T. Perez. Ship Motion Control. Springer, 2005.
[64] H. Rodriguez, R. Ortega, and N. Barabanov. A robustly stable output feedback
saturated controller for the boost DCtoDC converter. Systems and Control Let
ters, 40(1):1–8, 2000.
[65] H. Rodriguez, R. Ortega, and I. Mareels. Nonlinear control of magnetic levita
tion systems via energybalancing. Proceedings of the American Control Con
ference, June 2830, 2000, Chicago, IL, USA.
[66] C. Scherer and S. Weiland. Lecture Notes: DISC Course on Linear Matrix In
equalities in Control. 2006.
[67] D. Simon. Optimal State Estimation. John Wiley & Sons, Inc., 2006.
[68] J. J. E. Slotine and W. Li. Applied Nonlinear Control. Prentice Hall International
Inc., 1991.
[69] J. G. Snijders. Wave ﬁltering and thruster allocation for dynamic positioned
ships. Master’s thesis, Mathematical Systems Theory Group, DIAM, EWI Fac
ulty, TU Delft, The Netherlands, November 2005.
[70] J. G. Snijders and J. W. van der Woude. Nonlinear observer design for dynamic
positioning. Proceedings of Dynamic Positioning Conference, September 2728,
2005, Houston, USA.
116 BIBLIOGRAPHY
[71] A. J. Sørensen, S. I. Sagatun, and T. I. Fossen. Design of a dynamic positioning
system using modelbased control. Control Engineering Practice, 4(3):359–368,
1996.
[72] J. Spouge. Review of methods for demonstrating redundancy in dynamic posi
tioning systems for the oﬀshore industry. Research Report 15, DNV, Consulting,
London, 2004.
[73] J. F. Sturm. Using SeDuMi 1.01, a MATLAB toolbox for optimization over
symmetric cones. Optimization Methods and Software, 11(1):625–653, 1999.
[74] V. Torre. Existence of limit cycles and control in complete Keynesian system by
theory of bifurcations. Econometrica, 45(6):1457–1466, 1977.
[75] G. Torsetnes, J. Jouﬀroy, and T. I. Fossen. Nonlinear dynamic positioning of
ships with gainscheduled wave ﬁltering. Proceedings of the 43rd IEEE Confer
ence on Decision and Control, December 1417, 2004, Atlantis, The Bahamas.
[76] P. Tsiotras, M. Corless, and M. Rotea. Counter example to a recent result on
the stability of nonlinear systems. IMA Journal of Mathematical Control and
Information, 13(2):129–130, 1996.
[77] J. van Amerongen. Adaptive Steering of Ships: a model reference approach to
improved manoevring and economical course keeping. PhD thesis, Delft Uni
versity of Technology, Delft, The Netherlands, 1982.
[78] P. van der Klugt. Rudder Roll Stabilization. PhD thesis, Delft University of
Technology, Delft, The Netherlands, 1987.
[79] A. J. van der Schaft. L
2
−Gain and Passivity Techniques in Nonlinear Control.
Springer, 2000.
[80] F. Verhulst. Nonlinear Diﬀerential Equations and Dynamical Systems. Springer,
1996.
[81] V. Vidyasagar. Nonlinear Systems Analysis. SIAM, Second edition, 2002.
[82] A. A. Zaher. Nonlinear control of systems with multiple equilibria and unknown
sinusoidal disturbance. Communications in Nonlinear Science and Numerical
Simulation, 12(8):1518–1533, 2007.
List of Abbreviations
DP DynamicPositioning
EKF ExtendedKalmanFilter
FPSO Floating Production, Storage and Oﬄoading
IDA InterconnectionDampingAssignment
KF KalmanFilter
LMI LinearMatrixInequality
LF LowFrequency
LQ LinearQuadratic
LQG LinearQuadraticGaussian
LQR LinearQuadraticRegulation
MATLAB MATrixLABoratory
OCP OptimalControlProblem
PBC PassivityBasedControl
PHS PortHamiltonianSystems
RSF RealSchurForm
SDARE StateDependentAlgebraicRiccatiEquation
SDC StateDependentCoeﬃcient
SeDuMi SelfDualMinimization
WF WaveFrequency
YALMIP YetAnotherLinearMatrixInequalityParser
117
118
Summary
Dynamic positioning (DP) refers to maintaining the position and heading of a sea
vessel under speciﬁed limits exclusively by using thrusters and propellers. This was
usually achieved by using conventional techniques like anchoring and mooring, until
late sixties. Since then more sophisticated DP systems are being used for dynamic
positioning.
A DP system consists of three main units: power system, DP control system, and
thruster system. A vessel whose motion is controlled by DP system is called a DP
vessel. With the fast growing oﬀshore industry, the demand for more accurate and
stable DP systems has increased to ensure safety of the DP operations.
The focus of this research is the control system design for DP vessels. Starting
point of the study of the DP control system design is a mathematical model of the
sea vessel. The vessel motion in sea is divided into two motion components: a low
frequency (LF) motion component and a wave frequency (WF) motion component.
The LF motion is due to the slowly varying forces and moments due to the control
actions, waves, wind, and sea currents. The WF motion is caused by the ﬁrst order
wave induced wave loads. The measurement signal from the sensors (GPS and gyro)
essentially contains both motion components. An important consideration in the con
trol system design for many dynamic positioning problems is to counter only the LF
motion of the vessel. This is to avoid unnecessary wear and tear of the propulsion
units. As the measurement signal plays an important role in the control design proce
dures, ﬁltering of the WF component from the measurement signal is essential so that
only the LF component is transmitted to the controller. This problem of separating LF
motion component from WF motion component is called wave ﬁltering.
An important aspect of the control system design is its stability. In this thesis, we
ﬁrst study the local and global aspects of the asymptotic stability of the DP vessel.
The vessel model is nonlinear and this nonlinearity is due to the heading angle of the
vessel. The socalled state dependent coeﬃcient (SDC) parametrization will trans
form the nonlinear vessel model into a pseudolinear form with the system matrix
being function of the heading angle. Some results on the global asymptotic stabil
ity studies of similar systems are recapitulated. We then present a counter example
showing that the conditions that the system matrix is continuous, Hurwitz, and expo
nentially bounded, as reported in the literature on global asymptotic stability analysis
of such systems are not suﬃcient to establish the global asymptotic stability of the
119
120 Summary
pseudolinear systems whose system matrix depends on a single state component. In
the second counter example, we show that in addition to the above mentioned set of
conditions with the ﬁrst counter example, if we assume the periodicity of the sys
tem matrix, this still does not make it the set of suﬃcient conditions for the global
asymptotic stability of the pseudolinear system.
For the case of a periodic dependence of the system matrix on a single state com
ponent, we propose a technique to prove the global asymptotic stability of the pseudo
linear system. This technique is a combination of the linear matrix inequalities (LMIs)
and the Lyapunov stability theory. We used this technique to prove the global asymp
totic stability of the nominal vessel model with a naive state feedback controller.
In this thesis, the socalled state dependent algebraic Riccati equation (SDARE)
technique is used to address the regulation problem of the DP vessel. We start with
a nominal case when there is no WF motion involved and measurements are noise
free. The SDARE control design is presented to regulate the vessel to the desired
equilibrium position. We then proceed with the socalled nonnominal case in which
disturbance vector is unknown and measurements also contain noise. The disturbance
vector is modeled as a ﬁrst order Markov process. Therefore, the disturbance vector
is concatenated with the state of the vessel model. The SDARE control design is
a state feedback method. Therefore, knowledge of the entire state of the system is
required. The output equation in the model of DP vessel indicates that only the noisy
position and heading measurements are available for feedback. Therefore, an observer
is required to estimate the complete state of the system. The pseudolinear SDARE
observer is used to get the state estimates. We conclude the SDARE control design
concept by taking into account the WF component in the measurement model. In this
case, the noisy measurements are contaminated with the WF component as well which
requires wave ﬁltering. We use the SDARE technique to address the estimation and
ﬁltering problems. Simulation experiments performed by using the data of a supply
vessel have shown good results as far as the performance of the SDARE controller and
observer is concerned.
We present the Fourier series interpolation (FSI) method which reduces the com
putation time of the algebraic Riccati equation. The eﬀectiveness of the FSI method is
shown in the SDARE approach for DP control system design. This is especially useful
in the situations where the solution of the SDARE is required online. The fast com
putation of the SDARE may be required in the cases when this technique is used with
the systems whose dynamic response is fast/high. The idea of the method discussed
in this thesis is limited to the special case of the SDARE whose coeﬃcient matrices
depend on a single state component in a periodic way. The underlying idea behind
the fast computation of the solution of the SDARE is the oﬄine computation of the
Fourier coeﬃcients which eﬀectively reduces the online computations.
The regulation problem of the DP vessel is also addressed by using the port Hamil
tonian formulation. In this context, we study only the control design problem. The
estimation and wave ﬁltering are assumed to be addressed separately and we move on
from this point forward to address the control design problem. The interconnection
damping assignment passivity based control (IDAPBC) design technique is used to
introduce two control laws: the classical IDAPBC design and the extended IDAPBC
design. The former controller provides the desired performance in the nominal (all the
external disturbances are known or there are no disturbances) case but it fails to stabi
Summary 121
lize the system at the desired location in presence of unknown external disturbances.
To address this bottleneck in the design, the extended IDAPBC design is introduced
to yield the desired performance.
There are two notable features in the port Hamiltonian based study of the DP prob
lem. The ﬁrst one is that there is an additional degree of freedom in the form of the
Hamiltonian (also sometimes called as energy of the system) function which helps in
improving the performance and analysis of the controller. In the study of the DP prob
lem, we consider two diﬀerent formulations of the Hamiltonian function: a quadratic
and a trigonometric. Simulation results have revealed that each formulation gives a
diﬀerent heading angle proﬁle of the vessel in cases when the initial heading angle is
above π radians. The proﬁle corresponding to the trigonometric formulation gives the
optimal heading angle proﬁle. The second feature of the port Hamiltonian formulation
is that it oﬀers the freedom to add additional damping through the controller action
which helps in improving the performance of the controller and simpliﬁes the tuning
of the gains.
In conclusion, the search for the suﬃcient conditions for the global asymptotic
stability of the pseudolinear systems with state dependency of the system matrix on a
single state component, is still open for further research. The DP problem has a wide
range of issues associated with it. This study is mainly limited to the regulation of
the DP vessel to a desired equilibrium point. The ideas and techniques introduced in
this dissertation can be employed also to study various other DP problems like path
following, trajectory tracking, etc. The study of the port Hamiltonian framework for
ﬁltering and estimation problems of the DP vessel is also an active research problem.
122
Samenvatting
add dutch summary here
123
124
Publications
SHAH MUHAMMAD AND JACOB VAN DER WOUDE
A counter example to a recent result on the stability of nonlinear systems, IMA Jour
nal of Mathematical Control and Information. Vol. 26, No. 3, August 2009, pp 319323.
SHAH MUHAMMAD AND JACOB VAN DER WOUDE
The Fourier Series Interpolation (FSI) Method for the Solution of the SDARE. In
proceedings of the UKACC International Conference on Control (CONTROL 2010),
September 710, 2010, Coventry, UK.
SHAH MUHAMMAD AND ARNAU D
`
ORIACEREZO
Output Feedback Passivitybased Controllers for Dynamic Positioning of Ships. In
Proceedings of the 14th Edition of the Dynamic Positioning Conference (DP CON
FERENCE 2010), October 1213, 2010, Houston, USA.
SHAH MUHAMMAD AND JACOB VAN DER WOUDE
On stability conditions for systems with periodic state dependent coeﬃcients, IMA
Journal of Mathematical Control and Information. Vol. 28, No. 1, March 2011, pp 97102.
SHAH MUHAMMAD AND ARNAU D
`
ORIACEREZO
Passivity Based Control Applied to the Dynamical Positioning of Ships. To appear in
IET Journal of Control Theory and Applications.
125
126
Curriculum Vitae
Shah Muhammad was born on the 6
th
of January, 1977 in Talagang, Pakistan. He got
his primary and secondary schooling mainly from Government Schools in his native
town. He did his B.Sc. in 1997 from the university of the Punjab and subsequently
he got M.Sc. in Applied Mathematics from Islamia University Bahawalpur in 1999.
Later on he worked as a Scientiﬁc Oﬃcer for four years in Ghulam Ishaq Khan (GIK)
Institute of Engineering Sciences and Technology, Topi, Pakistan. This stint ended on
June 30, 2005. Since September 2006, he has been in the Netherlands, pursuing his
PhD in the Mathematical System Theory and Optimization group at the Delft Institute
of Applied Mathematics, Delft University of Technology.
He has presented several papers during his PhD in various seminars and confer
ences related to dynamical systems and control theory. Prominent activities include
his participation and presentation in the 1st Elgersburg School on Mathematical Sys
tems Theory, organized by Technische Universit¨ at Ilmenau on March 30April 3, 2009
Elgersburg, Germany, Benelux meetings on systems and control (2009 and 2010 edi
tions), UKACC Conference on Control held in Coventry, UK in 2010, and the DP
Conference (2010 edition) held in Houston, USA. He has reviewed a number of arti
cles in his ﬁeld and has two published and one accepted journal article.
Sports and reading formed Shah’s extraacademic activities during his student life.
At Islamia University Bahawalpur, he won the ﬁrst prize in 5 Km long distance race
and ended up third in 1500 m race. He is a good player of cricket. He is a ﬂuent top
order batsman who likes to play slog sweep and reverse sweep with profound eager
ness. He is also a very agile ﬁelder and has a very safe pair of hands any where in the
out ﬁeld and in the slip cordon. Shah is a strong believer that sports are indispensable
part of an individual’s life and he looks forward to promote sporting culture among
the masses and society as a whole.
Shah’s post PhDplans are to join some dynamic and growing organization/institute
where he can use his knowledge and expertise acquired during his academic and re
search career. He likes to share and transfer his experience to the prospective col
leagues/students. He believes that research culture is a key to solve the problems of
daily life and to overcome the challenges of modern era. Shah intends to work for
the promotion of research culture especially in the developing and under developed
societies.
127
Dit proefschrift is goedgekeurd door de promotor: Prof. dr. ir. A.W. Heemink Copromotor: Dr. J.W. van der Woude
Samenstelling promotiecommissie: Rector Magniﬁcus Prof. dr. ir. A.W. Heemink Dr. J.W. van der Woude Prof.dr.ir. Prof.dr.ir. Prof.dr.ir. Prof. Dr. Prof. Dr. voorzitter Technische Universiteit Delft, promotor Technische Universiteit Delft, copromotor Technische Universiteit Delft, NL , NL , NL , NL University, , , SP
This thesis has been completed in partial fulﬁllment of the requirements of Delft University of Technology, Delft, The Netherlands, for the award of the PhD degree. The research presented in this thesis was supported in parts by two institutions: Delft University of Technology and HEC Pakistan. I thank them sincerely for their support. Published and distributed by: Shah Muhammad Email: abu.ash.hal@gmail.com
ISBN # 978909xxxxxx6 Keywords: tems. Dynamic positioning, Nonlinear control design, portHamiltonian sys
Copyright © 2012 by Shah Muhammad All rights reserved. No part of the material protected by this copyright notice may be reproduced or utilized in any form or by any means, electronic or mechanical, including photocopying, recording or by any information storage and retrieval system, without written permission of the author. Printed in The Netherlands
Contents
1
Introduction 1.1 What is a Dynamic Positioning (DP) System? 1.1.1 Applications of DP Systems . . . . . 1.1.2 Focus of this Research . . . . . . . . 1.2 An Overview of this Thesis . . . . . . . . . . 1.3 Contributions of this Thesis . . . . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
1 1 2 3 5 5 7 7 9 10 13 16 16 16 18
2
Mathematical Model of a Sea Vessel 2.1 Motion of an Aﬂoating Vessel . . . . . . . . . . . . . . . . . . . . . 2.2 Mathematical Model Describing the Dynamics of an Aﬂoating Vessel 2.2.1 The Dynamical Equations of Motion of the Vessel . . . . . . 2.2.2 The Disturbances Model . . . . . . . . . . . . . . . . . . . . 2.2.3 The Measurement Model . . . . . . . . . . . . . . . . . . . . 2.2.4 Wave Filtering . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Summary of the Mathematical Model . . . . . . . . . . . . . . . . . 2.4 Properties of the Model . . . . . . . . . . . . . . . . . . . . . . . . . SDC Parametrization and Stability Analysis of Autonomous Nonlinear Systems 3.1 State Dependent Coeﬃcient Parametrization . . . . . . . . . . . . . . 3.2 Local Asymptotic Stability Analysis . . . . . . . . . . . . . . . . . . 3.3 Global Asymptotic Stability Analysis . . . . . . . . . . . . . . . . . 3.4 Exponential Boundedness and Global Asymptotic Stability . . . . . . 3.4.1 Counter Example Showing Exponentially Boundedness does not Guarantee Global Asymptotic Stability . . . . . . . . . . 3.5 Periodicity and Global Asymptotic Stability . . . . . . . . . . . . . . 3.5.1 Counter Example Showing that the Periodicity of the System Matrix does not Guarantee Global Asymptotic Stability . . . . 3.5.2 An LMI Based Approach for Global Asymptotic Stability . . 3.5.3 Infeasibility of the LMI Feasibility Problem . . . . . . . . . . 3.5.4 Further Analysis for Global Asymptotic Stability . . . . . . . v
3
21 21 23 23 24 25 28 29 34 39 42
. Conclusions and Recommendations for Future Work
6
. . . 4. . . . 6. . . . . .2 Simulation Results in Presence of Disturbances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Motivating Problem . . . . . . .5 Extended IDAPBC Design . .5. .6. . . . . . . . . . .2 Target Extended System . . . . . . . . .
. . . .5 Analysis in Presence of Disturbances . . . . . . . . . . . . . . . .1 Optimal Control Problem and the SDARE . . . . .5. 43 45 46 47 47 52 56 73 75 76 79 83 84 84 84 86 87 87 87 88 90 90 92 92 95 97 97 97 98 99 101 101 103 107 109 111 117 119
State Dependent Riccati Equation based Control Design 4. . .3 Simulation Results . .
. .
7
A Proof of asymptotic stability of (6.4 Classical IDAPBC Design . . . . . . . . . . . . . . . . .2 Performance Analysis .4 A Trigonometric Energy Shaping . . . . . . . . . . . . . . .1 Simulation Results for the classical IDAPBC design . . . 6. . . . . . . . . . . . . . . . . . . . . . . . 6. .3. . . . .2 Dynamic Positioning Problem . .3. . . . . . . . . . . . . . . 4. . . . . . . . . . . . . . . . . . . . 6. . . . . . . . 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6. . . . .5. . 6. . .1. . . . . .1 Ship Model in Cartesian Coordinates . . .4 Conclusions . . . . . . . . .5. . . 4. . . . . . . . . 6. . . . . . . . . . . . 4.2 The IDAPBC Technique . . . .1 A Quadratic Energy Shaping . . . . .7 Conclusions . . . . . .1 The Fourier Series Interpolation (FSI) Method . . . .5. .2.1 Nonlinear Regulation Problem for the LF Model . . . . . . .6 4
Contents Summary and Conclusions . 6. . . . . 5. . .3 A Quadratic Energy Shaping . . 6. . . . . . 6. . . .2 Nonlinear Regulation Problem for the Complete Model 4. . . . . . . . . . . . . . . .2. . . . . . . . . . .2 A Trigonometric Energy Shaping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PortHamiltonian Formulation and Passivity Based Control Design 6. . . . . . 6. 6.1 PortHamiltonian Modeling . . . . . . 6. .
. .32) at (6. . . .2 Control System Design for the DP Vessel . . . . . . . . . . . . . . . . .1. . . . . . . . . . . 6. . 6. . .33) B Glossary Bibliography List of Abbreviations Summary
. . . . . . . .6. . . . . . .4. . . . . . 6. . . . . . . .6 Simulations . . . . .
. . .2 Ship Model in PortHamiltonian Coordinates . .
. . . . . .4. . . . . . . 6. .1 HamiltonianBased Control . . . . . . . . . . . .vi 3. . . 6. . . . .3 Ship Model in PortHamiltonian Framework . . . . . . . . . . . . . .
5
The Fourier Series Interpolation Method 5. . . .
Contents Samenvatting List of publications Curriculum Vitae
vii 123 125 126
.
viii
.
Diﬀerent types of thrusters. Figure 1. position keeping means maintaining the desired position and heading within limits that reﬂect the environmental eﬀects and the system capability. and unmodeled disturbances due to the environmental eﬀects and the propulsion system. For instance. Therefore. In practice. propulsion system. for practical reasons. are used to produce the desired effects. sea currents. Azimuth thrusters can be rotated through 360 degrees and thus produce thrust in all directions in the horizontal plane.1 shows an overview of a 1
.Chapter
1
Introduction
T
is chapter explains the fundamental objective of this thesis and gives an overview of the related details. A DP vessel is a vessel whose motion is controlled by a dynamic positioning system rather than by the conventional motion control techniques like mooring or anchoring. wind. aﬂoating vessel cannot maintain a completely static position at sea. An eﬃcient DP system would be the one which achieves these goals with minimum fuel consumption and also tolerates transient errors or failures in the propulsion and measurement systems. The objective of this system is to keep the vessel within speciﬁed position and heading limits exclusively by using the propulsion system consisting of thrusters and propellers. tunnel thrusters which produce thrust in sideway directions and azimuth thrusters which are ﬁtted under the hull of vessel. A complete DP system consists of three major parts: the vessel’s power system. for instance.
1. This is particularly useful because the environmental forces and moments change over time both in magnitude and direction. the thrusters system. centimeter accuracy is desired for the operations like automatic berthing of ships and maneuvering in shallow and conﬁned waters. A vessel in sea is subjected to various forces and moments due to waves. and the DP control system. This limit may vary from centimeters to meters depending upon the nature of the operation. This study is about the control system design of a dynamic positioning (DP) vessel.1 What is a Dynamic Positioning (DP) System?
A DP system is a computer controlled system.
drilling vessels (drill ships and semisubmersible drill ships). failure in maintaining the position and heading angle.
Figure 1. For this.
Chapter 1..2: A supply vessel. Introduction
Dynamic Positioning System
Power System
Power Management System
DP Control System
Thruster System
Distribution System
(cables. in case of diving vessels a failure in position keeping may
. From the operational aspects (for example in case of heavy lifts) it is important to maintain the vessel within precise navigational limits. mine sweepers. and survey ships. lub oil. Dynamic positioning is vital in various oﬀshore operations. a naval vessel and a supply vessel. etc. platform supply vessels and anchor handling vessels. routing. a very stable and accurate position keeping is required. For instance. vetillation etc.2 DP system. There are situations in which a failure in position keeping. ﬂoating production storage and oﬄoading (FPSO) units. may result in serious safety and ﬁnancial hazards. maritime research vessel.1
Applications of DP Systems
DP vessels are used to achieve a variety of objectives in the oﬀshore industry.e. pipe laying ships. The main vessel types used for various oﬀshore operations include diving support vessels. Figure 1.1: Major components of a DP system
1. shuttle tankers.)
Positioning Control System (DP Computer)
Thruster Control Units
Electrical
Switchboards
Uninterruptable Power Supplies (UPS)
Communication Network
Cabling and Routing
Electrical Generators
Prime Movers (engines)
Engine Auxiliary Systems (fuel. Drive Units and Auxiliaries
Positioning Reference System
Main Properties and Rudders
Figure 1. [72]. cooling water. i.)
Vessel and Environmental Sensor Systems
Thrusters. landing platform docks.2 shows an FPSO unit.1. a naval ship and an FPSO vessel in sea (from left to right).
equipment. Consequently. we desire to maintain the position and heading of the vessel within desired limits. Consequently. With the growing demand of the oﬀshore industry. propulsion and power units. damage to the structure or vessel. and applications of dynamic positioning systems interested readers are referred to [21]. What is a Dynamic Positioning (DP) System?
3
result in the death or injury of the divers. DP systems have emerged as a popular replacement for the conventional position keeping techniques: anchoring and jackup barge. Some of these control objectives include position and heading regulation. control. path following. principles. a subcomponent of the DP control system. This component may well be considered as the heart of the DP system as it interacts with the rest of the components of DP system. production and exploration demands.
1. We consider a nonlinear vessel model from [22] and this model serves as a prototype for this study. operator’s error. over actuation feature is included to enhance the operational continuity by reducing the chances of failure of the propulsion and measurement systems. In control design. technical failure of the DP equipment. DP systems have excellent maneuverability and can be easily moved from one place to another. if a drilling vessel working in deep waters makes widely twitchy movement then it will cause damage to riser pipe or drilling pipe and subsequently the drilling operation will be abrupted. The focus of this thesis is the design and analysis of the positioning control system of the vessel. In many ships and various operations. In situations where the operation is being carried out very close to a ﬁxed structure. No additional external equipment like the anchoring tugs are required for DP systems. The conventional techniques are limited by the sea obstructions and sea depth but DP systems do not have such limitations. it is important to take into account the size and the dynamic
. then a position keeping failure may result in a collision. While the conventional tools have no or limited maneuverability. For more information on the design. The position keeping failure may occur because of multiple reasons. extreme weather conditions not executed in the control design strategy. etc. the development in the DP technologies is proliferating to meet the stringent safety. From a physical point of view. The model has been introduced in Chapter 2.2
Focus of this Research
It is clear from the foregoing discussion that a DP system consists of several components. The basic element of a positioning control system is a mathematical model of the vessel which is an approximation of the reality. and waveinduced motion reduction. There can be diﬀerent control design objectives depending upon the nature and demands of a DP operation. For instance. there have been developments in all the faculties of the DP technologies like navigation. trajectory tracking.1. We study the design and analysis of the control laws to stabilize (or regulate) the model to a desired equilibrium point. This feature gives rise to the problem of optimal allocation because in the presence of this feature there can be many possible combinations of actuators to yield a speciﬁc control action. and other subsidiary components. or a delay in the operation may occur. The anchoring may take several hours but DP has very quick setup. This has made the users and the manufacturers of the DP systems strive hard towards more reﬁnements in the DP related equipments and expertise.1. We focus on the position and heading regulation.1.
to solve the SDARE corresponding to the DP vessel model. The special form of the vessel model motivated us to combine the Lyapunov stability theory with linear matrix inequalities (LMIs) to come up with a new method to analyze the global asymptotic stability of the pseudolinear systems of the form similar to the prototype vessel model.4
Chapter 1. In practice. the state dependent coeﬃcient (SDC) parametrization is introduced which is a strategy to transform the nonlinear system into a pseudolinear form. This study is called capability study. The FSI method reduces the computation time for the SDARE in comparison with the Schur decomposition method. An unstable system is useless and potentially dangerous. The asymptotic behavior implies that beside being stable. Due to complex and exotic behavior of the nonlinear systems. when described in pseudolinear form by using the SDC parametrization. Qualitatively. The prototype vessel model has a typical nonlinearity. We begin our study with the stability analysis of pseudolinear systems similar to the prototype DP vessel system. The advantage of this approach is that it provides an opportunity to use concepts from linear system theory to study the nonlinear vessel model.
. Introduction
response of the thrust devices which must be adequate to cope with various environmental conditions in diﬀerent oﬀshore operations. Lyapunov stability theory is the most commonly used tool to study the stability properties of nonlinear systems. for large systems. These control laws may also be seen as the nonlinear version of the wellknown PID controllers. The portHamiltonian formulation has also become a popular technique to study the physical systems since a decade. The prototype vessel model is nonlinear due to the heading angle of the vessel. The stability analysis is the most important feature of any control system design. more reﬁned concepts of stability such as (local and global) asymptotic stability are required to describe the behavior of nonlinear systems. in the portHamiltonian framework. We transform the DP vessel model into portHamiltonian form and then use the IDAPBC design approach to come up with a family of control laws. the maximum thrust forces and moments to maintain the position and heading in diﬀerent environmental conditions are estimated and then the capability of the thrust devices to meet the demands is analyzed. We use the SDC framework to come up with the nonlinear version of the PID controller by using the state dependent algebraic Riccati equation (from now on we call it the SDARE) technique for the design of a stabilizing control law for the DP vessel. PID controllers are commonly used in practice. There are various oﬀtheshelf methods for the solution of the SDARE. The computation of the controller and the observer gains require online computation of the solution of the SDARE. In [7]. It can require large computation time. We come up with a new method. a dynamic system is called stable if starting from a position somewhere near its equilibrium or operating point implies that it will stay around the point ever after. We use the SDC framework throughout this thesis to study the control system design for the DP vessel. the system will converge to its equilibrium or operating position as time goes on. the Fourier series interpolation (FSI) method. especially.
Our focus in this thesis is to address the control system design problem for dynamic positioning of a sea vessel. Two counter examples are presented in this chapter. The FSI method reduces the online computations of the solution of the SDARE by performing the computationally expensive tasks oﬄine. in addition to the set of conditions mentioned in the ﬁrst counter example. The hindsight ideas for future research are also presented in this chapter. The SDARE based control design and estimation technique is used to design an SDARE controller and an SDARE observer for dynamic positioning of the vessel. It is shown that the extended set too does not constitute the set of suﬃcient conditions for global asymptotic stability of the pseudolinear system. we also propose in this chapter. It is claimed in the literature that it is suﬃcient for global asymptotic stability of a pseudolinear system that the system matrix in its SDC form is continuous.1. additionally. There had been some existing results on this subject. The DP system is illustrated and the importance of dynamic positioning is highlighted by describing its applications in various oﬀshore and onshore operations. The fourth chapter addresses the control system design problem. The ﬁrst chapter of this thesis introduces the dynamic positioning problem. we show that this claim is not valid. An Overview of this Thesis
5
1. The ﬁfth chapter is about the FSI method for the approximation of the solution of the SDARE. The sixth chapter is also about the control system design problem of the DP vessel. a method for proving global asymptotic stability of the special pseudolinear systems by combining the Lyapunov stability theory and the LMIs. The ﬁrst counter example shows that the conditions. The SDC framework is used to express the nonlinear system in a pseudolinear form and then the stability is analyzed based on the properties of the state dependent system matrix. as reported in the literature on this subject. and exponentially bounded. The main subject of the third chapter is the study of global asymptotic stability of a special type of nonlinear systems which are similar to the prototype vessel model. The main idea is to transform the vessel model in the portHamiltonian structure and then use the IDAPBC design approach to address the control design problem. that the system matrix in pseudolinear form is continuous. we assume additionally that the system matrix
.2 An Overview of this Thesis
Chapters 1 and 2 contain the basic introductory material about the main theme of this thesis.3 Contributions of this Thesis
We study the stability analysis of nonlinear systems in the SDC framework. Apart from this. In a ﬁrst counter example. and exponentially bounded. are not suﬃcient for global asymptotic stability of the pseudolinear system. Motivated by the special type of state dependence of the system matrix in DP vessel model. The second chapter introduces the details of the mathematical model we use in this thesis to describe the vessel motion.
1. Hurwitz. It highlights the necessary details of the vessel motion in mathematical form. Hurwitz. In the second counter example. The thesis is concluded with the seventh chapter which brieﬂy summarizes the thesis and provides some concluding remarks.2. Our main contribution on this topic are two counter examples. we also assume that the system matrix is periodic.
6
Chapter 1. Introduction
is periodic and show by means of another counter example that an additional condition also does not guarantee the stability of the nonlinear system. Each of these counter examples have separately been published, see [53] and [54]. The special form of the nonlinearity in the vessel model and the Lyapunov stability theory has lead us to propose a new approach to prove global asymptotic stability of the special type of pseudolinear systems which resembles the prototype DP vessel model. This approach makes use of the LMIs to achieve global asymptotic stability. The approach is useful in particular for the vessel model and in general for the systems having similar structure as the vessel model. Another contribution is the SDARE controller design for the DP vessel. The model based SDARE controller is a state feedback controller. The complete state of the DP vessel model is not available in practice. Therefore, a state observer is also required. We also used the SDARE observer to ﬁnd the state estimate. It has been shown that the SDARE controller in combination with the SDARE observer gives the desired stability and performance of the DP vessel. Alongside the SDARE controller and observer, a numerical method for the approximation of the solution of the SDARE is proposed. We call this the Fourier series interpolation (FSI) method. This method is proved to be very handy in reducing the online computation time of the SDARE for controller and observer gains computations. The concept of the FSI method has been presented in a conference paper, see [52]. The ﬁnal contribution of this thesis is the use of the portHamiltonian structure and the passivity theory for the ﬁrst time for DP vessel control design. We propose a family of passivity based controllers for the DP vessel. Passivity idea is very attractive in a sense that it helps in assigning the physical meaning to various variables and quantities. The stability and performance of the family of the IDAPBC designs are discussed. This idea was presented at a conference (see [50]) and it has recently been accepted in a journal (see [51]).
Chapter
2
Mathematical Model of a Sea Vessel
T
e details of a vessel model for DP considerations are presented in this chapter. The prototype vessel model described in this chapter will be used in the subsequent chapters for studying the control system design of the DP vessel.
2.1 Motion of an Aﬂoating Vessel
In this section, we explain various terms associated with the motion of the vessel in the sea. Motion of an aﬂoating vessel can be described by six degrees of freedom (DOF), i.e., a vessel can move in six diﬀerent directions. We can categorize the six DOF in two categories: 1. The translational motion in the following three directions, • Surge: motion in backward (aft/stern) and forward (bow/fore) directions • Sway: motion along sideways (transversial directions): starboard (right side of the ship) and port (left side of the ship) directions • Heave: motion in upward and downward directions 2. The rotational motion in the following three directions, • Roll: rotation about the surge axis • Pitch: rotation about the sway axis • Yaw: rotation about the heave axis Various modes of motion and the forces acting on the vessel are shown in Figure 2.1 and summarized in Table 2.1. A DP system is concerned primarily with control of 7
8
Chapter 2. Mathematical Model of a Sea Vessel
HEAVE
YAW ROLL SURGE
PITCH
SWAY
Figure 2.1: Six DOF of motion and forces acting on a ﬂoating vessel [Figure courtesy of www.km.kongsberg.com]. Table 2.1: Nomenclature of the vessel motion axis x y z Translation position velocity force + direction Translation angle rate torque + direction surge x u X forward roll φ p K starboard sway y v Y starboard pitch θ q M aft down heave z w Z downward yaw ψ r N right turn
the vessel in the horizontal plane, i.e., only the motions along surge, sway and yaw directions are considered for DP purposes. In Table 2.1, X and Y represent the forces in the surge and sway directions and N denotes the turning eﬀect because of the thrusters and environmental eﬀects. Dynamic positioning literature is very rich in terminology. For the convenience of the reader, we brieﬂy explain some important terms which we use in this thesis. A superstructure is an upward extension projected above the main deck of the vessel. Maximum legal load amount on a vessel is characterized by water line. The distance from the water line to the upper deck level is called freeboard. Water line and freeboard are the characteristic of buoyancy level. Buoyancy level is a characteristics of waters: it is low for fresh and warm waters and it is high for cold and saline waters which have more density. Height and weight of the superstructure aﬀects the freeboard. More the height and weight of the superstructure less will be the freeboard area. The rear or aft part of the vessel is termed as stern. Usually, during the night time,
starboard.2: Commonly used terms in literature on dynamic positioning
2. propellers.2. Figure 2.2 Mathematical Model Describing the Dynamics of an Aﬂoating Vessel
Modern DP control systems for ships use controllers based on a mathematical model of the ship. opposite to the stern part.
Super Structure Starboard side
Aft or Stern
Free board
Fore or Bow
Port side
Water line
Figure 2. All these terms are linked with the main deck of the vessel and has nothing to do with the location of the superstructure on the deck. This mathematical model describes the hydrodynamic. The information from the measurement systems is transmitted to the controller and signals from the controller are sent to the propulsion systems consisting of the thrusters and propellers (planted at at least one of the aft. The opposite part of the vessel on the left hand side will then be called as port. The foremost part of the vessel. The right hand side of the vessel as perceived by a person on board facing the bow is termed as starboard. when the vessel is underway.2. interested readers are referred to [22] and for an uptodate study we refer to [17] and [63]. damping. and rudders to generate the desired activity to maintain the required position and heading
. environmental and control forces and moments acting on the vessel.2 illustrates all these terms. Mathematical Model Describing the Dynamics of an Aﬂoating Vessel
9
the stern of the vessel is indicated with a white navigation light on it. For rigorous details on the mathematical modeling of the dynamics of the ﬂoating vessel. port and fore sides). is referred as the bow.
Let
. the measurement models are explained. This section is divided into three major subsections. Mathematical Model of a Sea Vessel
of the vessel.2. In the subsequent chapters. The Kinematic Equations of Motion An aﬂoating vessel has six degrees of freedom.10
Chapter 2. this model will be used for various design and analysis purposes. In the last subsection. only the motion in the horizontal plane is considered.3: Flow chart of the DP control system.3 shows an overview of the working of the DP system. dynamical equations of motion are explained. we discuss how the perturbations or the environmental disturbances are incorporated into the mathematical model of the vessel dynamics. In the ﬁrst one. The testbed vessel model used in this thesis is introduced in this chapter and the details describing the motion of an aﬂoating vessel are also explained.4 explains the description of both frames of reference.
Data from Reference Systems and Sensors
Wind
Thruster Allocation
Feed Forward
Optimum Controller
Mathematical Model
SDP SYSTEM
Figure 2. In the second subsection. Figure 2. and the kinetic equations of motion which deal with the analysis of the forces causing the motion. Figure 2.1
The Dynamical Equations of Motion of the Vessel
The study of the dynamical equations of motion of a mechanical system can be divided into two parts: the kinematic equations of motion which deal with the geometrical aspects of the equations of motion.
2. Two frames of reference are used to describe the motion: an Earthﬁxed inertial frame of reference and a bodyﬁxed relative frame of reference. For DP purposes.
η = [x y ψ]T describe the position (x. only the rotation matrix in yaw will be used to describe the kinematic equations of motion. ˙ where the transformation matrix is given by cos ψ J(ψ) = sin ψ 0 − sin ψ cos ψ 0 0 0 1 . This property is very important from the design and stability analysis perspective as we will see in the subsequent chapters. Therefore. it is an appropriate assumption that both the pitching and rolling motions are oscillatory with zero mean and limited amplitude. Note that the transformation matrix is nonsingular and orthogonal.. (2.2. Also. we see that there is no uncertainty associated with (2. the kinematic equations of motion in vectorial form are given by η = J(ψ)ν. ∀ψ ∈ R.4: The Earthﬁxed and the vesselﬁxed frames of reference.
. y) and heading ψ of the vessel in the inertial frame of reference and ν = [u v r]T describe the velocities of the vessel in the relative frame of reference.1)
(2.e. Mathematical Model Describing the Dynamics of an Aﬂoating Vessel
11
y
yr
xr Or
O
X
Figure 2. i.1) as it describes the wellknown geometrical aspects of the model. there exist restoring moments in roll and pitch directions.2)
The kinematic transformation (2. Then. J(ψ)J T (ψ) = J T (ψ)J(ψ) = I3 . For conventional ships.2. Also due to metacentric stability.2) relates the bodyﬁxed velocities to the derivative of the positions in the inertial frame of reference.
and yaw directions are deﬁned as ∂X ∂Y ∂N ∂Y ∂N . Nv = .3) and (2.e. (2. Nr = . For DP consideration.e. i. The vessel motion generates waves. we use the vectors τ and u interchangeably.12
Chapter 2. for instance see [22] for a detailed study. (2. The vector b ∈ R3×1 represents the slowly varying bias forces and moments in the Earthﬁxed inertial frame of reference. Hence. The vector u is the command to the actuators. wind. and xG denotes the longitudinal position of the center of gravity of the vessel with respect to the relative frame of reference. Yr = Nv . The added masses due to acceleration in the surge. sea currents.4). N]T ∈ R3×1 represents the control forces and moment acting on the vessel in the bodyﬁxed frame of reference. Due to small velocities and starboardport symmetries of the vessel.. i. This assumption is very useful for the purpose of analysis. which are assumed to have much faster dynamic response than the vessel. sway.3) (2. unless it is speciﬁed. In the forthcoming chapters. τ = Bu u. ˙ ˙ in DP applications. In the following chapters. the vector τ = [X. The vector u ∈ Rr×1 (r ≥ 1) describes the control inputs and the matrix Bu ∈ R3×r is a constant matrix describing the actuator conﬁgurations.7) D= 0 0 −Nv −Nr Xu = ˙
. This means energy is transferred from vessel to the ﬂuid and this energy is modeled by the linear damping term. we assume a fully actuated vessel model and we will take Bu = I3 . the inertia matrix has the following form 0 0 m − Xu ˙ 0 m − Yv mxG − Yr R3×3 . The matrices M and D are 3 × 3 inertia and damping matrices. (2. the equations in the Newtonian formulation based on Newton’s second law of motion are considered. Y. The linear damping matrix D for DP is taken as 0 0 −Xu −Yv −Yr R3×3 . it is assumed that the matrix M is symmetric and strictly positive deﬁnite. and other environmental factors surrounding the vessel. the added mass in sway due to the angular acceleration in yaw is equal to the added mass in yaw due to sway acceleration. provided by the propulsion system of the ship consisting of propellers and thrusters. (2.5) M= ˙ ˙ 0 mxG − Nv Iz − Nr ˙ ˙ where m is the vessel mass. In this work. respectively. Yr = .4)
In (2. Mathematical Model of a Sea Vessel
The Kinetic Equations of Motion The nonlinear kinetic equations of motion can be formulated by using Newtonian or Lagrangian mechanics.6) ˙ ˙ ˙ ˙ ∂˙ u ∂˙ v ∂˙ r ∂˙ r ∂˙ v Note that the inertia along the surge direction is decoupled from the inertia eﬀects along the sway and yaw directions. The nonlinear kinetic equations of motion in vectorial form can be written as ˙ ν = −M −1 Dν + M −1 τ + M −1 J T (ψ)b. Yv = . thus the coeﬃcient Bu represents the mapping from the actuator command to the force generated by the actuators.. Iz is the moment of inertia about the vesselﬁxed zaxis. due to the waves. M = M T > 0.
wind. An a priori estimate of the mass and damping parameters of the vessel can be obtained by using semiempirical methods and hydrodynamic computations. sea currents. Nr = . velocities. However. (2. and [47]. The total vessel motion is then deﬁned to be the sum of the LF and WF motions. the damping matrix is assumed to be real. waves. it can be assumed that Nv = Yr . Yv = . sway.2. In such a case. In what follows. propeller propulsions. waves. The eﬀect of the WF disturbances is incorporated in the measurement model.8) ∂u ∂v ∂r ∂r ∂v Decoupling of the surge mode from the sway and yaw modes is beneﬁcial for the convergence of parameter estimation algorithms. and yaw directions are deﬁned by ∂X ∂Y ∂N ∂Y ∂N .These are the disturbances due to the wrong functioning or noise in the measurement devices like DGPS and gyro compass. and rudder excitations. See for instance. [27]. Figure 2. trim. are distinguished into two categories: second order low frequency (LF) disturbances and ﬁrst order wave induced wave frequency (WF) disturbances.
. and positive deﬁnite.7. and positive deﬁnite. Often the mass and damping parameters estimates are updated based on the data obtained from the practical experiments in calm waters. In the following. Yr = . respectively.2
The Disturbances Model
The forces acting on a sea vessel can be categorized in two main categories [35]: the internal and the external forces and moments. both the LF and WF disturbances are characterized by respective dynamical models. symmetric. These have partially been discussed in the previous subsection. The external disturbances due to unmodeled dynamics. water depth. some explanation of these dynamic models is given. Xu =
2. nonsymmetrical. See [20] for details about the identiﬁcation and estimation of vessel model parameters.2.2. Here we explain the external forces acting on the vessel. see [26]. • Multiplicative disturbances . • Measurement disturbances . The internal forces and moments are formulated as functions of acceleration. the model of the ship is extended by adding additional states. [24]. Mathematical Model Describing the Dynamics of an Aﬂoating Vessel
13
In most DP applications. and sea current acting on the vessel. speed changes. These disturbances are called multiplicative disturbances.These are the disturbances due to wind. described in the next section.5 illustrates the concept of the slowly varying LF and the oscillatory WF motions. The damping components in surge. Along this thesis.A vessel in sea is also subject to the time varying parameters such as load conditions. etc. we call the motion of the vessel corresponding to these disturbances the LF motion and the WF motion. To model and analyze these forces. we assume the damping matrix D to be real. for low speed applications where the damping matrix is reduced to 2. These forces are also termed as external disturbances. etc. Nv = . These forces act additively on the vessel. The external disturbances can be distinguished into 3 major categories [78]: • Additive disturbances .
it is considered more appropriate ˙ from a physical point of view to use b = Ψwb to describe the bias model. The low frequency motions are caused by the forces generated by the thrusters and propellers. The matrix T is known as the time constant.9) as a lowpass ﬁlter. [22]. by the ﬁrst order Markov process given by ˙ b = −T −1 b + Ψwb . and yaw directions. Another case could be that the bias forces and moment are constant. In this context it will have relatively large values as sea states change very slowly. (2. Thus bias forces and moments are sometimes modeled as a random walk process.b ).e.9)
where b ∈ R3×1 is a vector of bias forces and moment. wind forces. ˙ Then the bias model will be b = 0. i. We can also interpret (2.14
Chapter 2. 75]. and hydrodynamic forces.5: The LF and the WF motions
Second Order LF Disturbances The LF disturbances are also sometimes termed as slowly varying bias forces and moment in surge. see for instance [25. Qc. wb ∼ (0. In marine control applications. these forces and moment can be described. waveinduced forces.
. Mathematical Model of a Sea Vessel
A Description of LF and WF Motion
10
LF Motion
8
LF and WF Motion
6
4
WF Motion
2
0
−2
0
10
20
30
40
50 60 Time [sec]
70
80
90
100
Figure 2.. the vector wb ∈ R3×1 represents the zeromean Gaussian white noise process. T ∈ R3×3 is a diagonal matrix of positive bias time constants and Ψ ∈ R3×3 is a diagonal matrix scaling the amplitude of the noise vector wb . sway. In many applications. This may be described as integration of the noise signal which in fact is a random walk phenomenon.
0i
i = 1. The WF motion parameters ω0 . it is important to know the threshold frequency. ζ2 . ζ3 }. i. Z = diag{ζ1 . Mathematical Model Describing the Dynamics of an Aﬂoating Vessel First Order Wave Induced WF Disturbances
15
The fundamental assumption for the development of the WF motion model is that the sea state is known and can be described by a spectral density function. It was Balchen who ﬁrst modeled the WF motion in this way. The ﬁrst order wave induced WF disturbances in surge. [5].ξi ). 3 s2 + 2ζi ω0i s + ω2 0i (2. i = 1. the damping ratio deﬁnes the level of damping (underdamped. (2.. Therefore. criticallydamped. the oscillatory WF motion is required to be ﬁltered or separated from the LF motion. wξi ∼ (0.2. ζ. This is important to avoid unnecessary power consumption and possible wear and tear of the actuators.2. σ2 . Normally. Qc. and σi is the wave intensity parameter. and undamped) of the system. and σ depend on the sea states. and yaw directions are modeled as second order harmonic oscillations which are driven by Gaussian white noise process. ξ1 . overdamped. ξ1 ]T . 2.
. Filtering of the WF motion can be done either by using appropriate classical ﬁltering techniques or it can be done by state estimation. For an estimation approach. wξ ∼ (0. The input wξi represents the Gaussian white noise process. ω02 . For DP operations only the LF motion is required to be controlled. The matrix O3 ∈ R3×3 is a zero matrix. Using ﬁltering techniques. structure of the vessel and the direction of the incident waves. The vector wξ describes the Gaussian white noise process. ζi is the relative damping ratio.e.12)
(1) (2) (3) (1) (2) (3) where ξ1 = [ξ1 .e. 2. Qc.ξ ) The state vector may or may not have a physical interpretation depending on the particular statespace realization used. the WF disturbances model for each direction can be written as
(i) ˙(i) ξ1 = ξ2 (i) (i) ˙(i) ξ2 = −ω2 ξ1 − 2ζi ω0i ξ2 + σi wi . Another problem with the ﬁltering approach is possible phase lag due to delay of the signals.10)
where ω0i is the dominating (sometimes also termed as undamped) wave frequency. i. The damping ratio ζi is a measure to describe how the oscillations in the system (2. The stochastic nature of environmentally induced forces and moments has made the Kalman ﬁlter an essential part of the modern sea vessel motion control systems. Ω = diag{ω01 . In state space representation. ξ2 ]T .10) decay when a disturbance is introduced. For each of the three directions. ξ2 = [ξ2 . 3
(2. ω03 }. sway. where reducing pitch and role motion helps avoiding motion sickness. The dominating wave frequency ω0i is obtained by spectral analysis.. the WF disturbances model in frequency domain is given by ξi (s) = σi s wξi (s). σ3 }. ξ2 . The WF response is required to be controlled in certain operations like ride control of a passenger vessel. and Σ = diag{σ1 .11)
A compact state space realization of the WF model is given by ˙ ξ1 ˙ ξ2 = O3 −Ω2 I3 −2ZΩ ξ1 ξ2 + O3 Σ wξ .
Mathematical Model of a Sea Vessel
we need to know the parameters of the system. The LF motion model is described by combining the equations (2. wave ﬁltering was accomplished by using a proportional controller with a deadband nonlinearity. the position and heading measurements of the vessel corresponding to the LF and the WF motions and the vector v ∈ R3×1 is the Gaussian white noise process.13)
where yb = η and yξ = ξ1 are. notch. and deadband ﬁlters were the most commonly used wave ﬁltering techniques in earlier DP systems.3). a phenomenon which gives rise to high frequency ﬂuctuations in the thrust demand in the control loop. (2. Sea states can be distinguished in 9 diﬀerent forms depending on the signiﬁcant wave height [22]. some vessels have multiple sensors.e.1). by the following vector equation y = yb + yξ + v. This change in length was termed as ‘weather’ as it was subject to the weather conditions [77]. we summarize all the modeling details from the previous section of this chapter. The main drawback of these techniques was due to a signiﬁcant phase lag to meet the high gain control requirements. and it is given by
..13). Lowpass. It is also important to keep in mind that the linearized model may not be a good approximation of the actual model or system. (2.3 Summary of the Mathematical Model
For more insight and a clear picture of the model of a dynamic positioning vessel. respectively. The separation of the WF component from the LF component is termed as wave ﬁltering. parameters are required.
2. using the superposition principle.2. v ∼ (0.9). For reliability. and (2.
2. This deadband produced a null control action until the control signal was inside the deadband. Knowledge of the sea states is required to determine the WF motion of the vessel. In earlier DP systems. The measurement model can be described. The vector v describes the measurement noise. the measured output is assumed to essentially contain the LF and WF motion components.3
The Measurement Model
The position and heading of the vessel in the inertial frame of reference can be measured by using a diﬀerential global positioning system (DGPS) and a gyrocompass.16
Chapter 2.4
Wave Filtering
In (2.2. This action is also important to avoid thruster modulation. Rc ). The length of this deadband could be increased by the operator with changing weather conditions.
2. i. Even in the linearized case.
˙ ˙ ν = −M Dν + M τ + J (ψ)b. see (2.
−1 −1 T
(2.
˙ η ˙ ν ˙ b ˙ ξ1 ˙ ξ2 = O3 O3 O3 O3 O3 J(ψ) −M −1 D O3 O3 O3 O3 M −1 J T (ψ) −T −1 O3 O3
A(ψ)
O3 O3 O3 O3 −Ω2
O3 O3 O3 I3 −2ZΩ
η ν b ξ1 ξ2
x
+
O3 M −1 O3 O3 O3
B
τ +
O3 O3 Ψ O3 O3
E
O3 O3 O3 O3 Σ
wb wξ
w
. we can write the LF motion model as η O3 ˙ ˙ ν = O3 ˙ O3 b J(ψ) −M −1 D O3
Ab (ψ)
O3 M −1 J T (ψ) −T −1
η O3 ν + M −1 O3 b
xb Bb
O3 3 τ + O wb .17)
The output equation of the LF model is yb = η ν + v.20)
The complete model of motion of the vessel can be described by augmenting both the LF and WF models and it can be written in the following form.3.
(2. Summary of the Mathematical Model
17
η = J(ψ)ν. Ψ
Eb
(2.19)
yξ =
I3
Cξ
O3
ξ1 ξ2
. b
I3
O3
Cb
O3
(2.22) y = I3 O3 O3 I3 O3 b + v.
(2.2.21) can be obtained by using the superposition principle. ˙ b = −T −1 b + Ψwb .18)
The dynamics of the WF motion model and its output are given by ˙ ξ1 ˙ ξ2 = O3 −Ω2 I3 −2ZΩ
Aξ
ξ1 ξ2
xξ
+
O3 Σ
Eξ
wξ .13).14) (2.15) (2. ξ 1 C ξ2
. and is written in matrix form as η ν (2.16)
In matrix form.
(2.21)
The output of (2.
(Observability in terms of rank condition) The pseudolinear system of the form (2. Chapter 3 explains the concept of the SDC parametrization in more detail.4. i. (2. presented in [6] . Deﬁnition 2. ˙ y =Cx + v. Mathematical Model of a Sea Vessel
The complete model of the vessel in compact form is written as x =A(ψ)x + Bτ + Ew. A(ψ)) is pointwise observable. These properties include controllability. The system (2.25)
has rank equal to the dimension of the state vector x. where Qc = diag{Qc.b .4 Properties of the Model
In this thesis.
2. This is not restrictive as the LF bias forces and the WF motions cannot
. stabilizability. This is also termed as the pair (C. . (2. n−1 CA (ψ) .1.23)(2. In the following. This is also termed as the pair (A(ψ).18
Chapter 2. We call this pseudolinear form the state dependent coeﬃcient (SDC) parametrization of nonlinear system. . it is important to know certain properties of the vessel model which play a fundamental part in the control design and estimation techniques which we are going to use in the subsequent chapters. Rc ).24) with an ndimensional state vector is pointwise controllable iﬀ the controllability matrix C= B A(ψ)B A2 (ψ)B · · · An−1 (ψ)B .26)
has rank equal to the dimension of the length of the state vector x.23)(2.4.24) with an ndimensional state vector is pointwise observable iﬀ the observability matrix C CA(ψ) 2 O = CA (ψ) .e. we recall some important results about these properties in the context of the pseudolinear systems. Qc. we deal with the control design and estimation problems of the DP vessel discussed in the previous sections. for each ψ ∈ R. and detectability. Deﬁnition 2. It can easily be checked that the controllability matrices corresponding to the systems (2.21) have rank 6.17) and (2..e. B) is pointwise controllable.
(2.23) (2. observability..24) is a pseudolinear system because of dependency of the system matrix A(ψ) on the heading angle. i. For this purpose. (Controllability in terms of rank condition) The pseudolinear system of the form (2.24)
The vectors w and v represent the Gaussian white noise processes. Qc ) and v ∼ (0.2. for each ψ ∈ R.23)(2. w ∼ (0.ξ }. only the position and the velocities can be controlled.
3.4.21).4.21) are pointwise stabilizable and detectable. These properties are important from the point of view of the existence of the solution of the SDARE. The stabilizability and detectability are weaker conditions than the controllability and observability.23)(2. Properties of the Model
19
be controlled. This is also termed as the pair (A(ψ). (Pointwise Detectability) The pseudolinear system of the form (2. (Pointwise Stabilizability) The pseudolinear system of the form of (2. Due to the special structure of the system matrices Ab (ψ) in (2. So the systems are pointwise observable. The only nonnegative eigenvalue of both Ab (ψ) and A(ψ) is 0 with algebraic multiplicity 3.28). we can build the states of the system from the knowledge of the input and the output.e.17) and (2.23)(2. (2. i.27) and (2. see [6] for more details.4. In the following. It can be checked that the rank is 9 for both stabilizability and detectability conditions corresponding to Ab (ψ) and it is 15 for A(ψ). the eigenvalues of the system matrices do not change with the variable ψ.17) and A(ψ) in (2. Deﬁnition 2.2. This makes it an easy task to compute the rank conditions (2. The observability matrices corresponding to both the systems have full column ranks for all ψ ∈ R..
.24) with an ndimensional state vector is pointwise detectable iﬀ rank λI − A(ψ) C = n. A(ψ)) is detectable.28)
for each eigenvalue λ of A(ψ) which has a nonnegative real part (Re(λ ≥ 0)) and for all ψ ∈ R. (2. B) is stabilizable. Deﬁnition 2. This is also termed as the pair (C.4. Thus both the systems (2. we deﬁne a necessary and suﬃcient condition for the pointwise stabilizability and detectability of a pseudolinear system.27)
for each eigenvalue λ of A(ψ) which has a nonnegative real part (Re(λ ≥ 0)) and for all ψ ∈ R. respectively.24) with an ndimensional state vector is pointwise stabilizable iﬀ rank λI − A(ψ) B = n.
20
.
˙ x0 = x(t0 ). This is mainly because of phenomena like ﬁnite escape time and limit cycles.4.Chapter
3
1
SDC Parametrization and Stability Analysis of Autonomous Nonlinear Systems
e stability analysis of nonlinear systems has always been a challenging task. Consider the following nonlinear system x = f (x).1 of this chapter have been published in the form of two separate articles in the IMA Journal of Mathematical Control and Information. [76]. see [53] and [54]. and [82].1)
where x ∈ Ω is the state of the system.
T
3. One such approach is to ﬁrst write the nonlinear system dynamics in linearlike form using a state dependent coeﬃcient (SDC) parametrization and then analyze the possible extension of the results of linear systems theory for the stability analysis of nonlinear systems. The SDC representation provides a systematic way to analyze the extension of the results of linear systems theory for the stability analysis of nonlinear systems. and [81]. [74]. [42]. see for instance [30]. Numerous techniques for stability analysis of nonlinear systems have been proposed over time.1 and Section 3.5. [54]. for further details see [40]. then it is always possible [43] to write f (x) = A(x)x.1 State Dependent Coeﬃcient Parametrization
Let Ω ⊆ Rn and f (x) be a vector function from Ω to Rn . (3. Let us call the matrix A(x) as the state dependent coeﬃcient (SDC) parametrization of f (x). It is important to mention that the SDC parametrization is not unique unless f (x)
1 Section 3. is continuously diﬀerentiable and f (0) = 0. If the vector function f : Ω −→ Rn .
21
.
αA1 (x)x + (1 − α)A2 (x)x = α f (x) + (1 − α) f (x) = f (x). if A1 (x) and A2 (x) are two distinct parametrizations of f (x) then for 0 ≤ α ≤ 1. For any SDC parametrization A(x) of f (x).1. Consequently.3 The system matrix.. A(0)x is the linearization of f (x) at the zero equilibrium.2). In fact inﬁnitely many parametrizations are possible but one has to chose only those which are appropriate for the desired objectives. A : Ω −→ Rn×n is a C1 function. C. The conditions C. See [6].1 and C.4 A(x) is a periodic function with a period θ.1. i. An important property of the SDC parametrization is that it preserves the linearization of nonlinear systems. the stability analysis of (3.
.
The ﬁrst two conditions imply that there is only one isolated equilibrium point. is exponentially bounded i. From here onward.e. Therefore..2) will be with reference to this equilibrium point.22
Chapter 3.4 are imposed to analyze global asymptotic stability.2 are suﬃcient to prove local asymptotic stability.1 The matrix function. SDC Parametrization and Stability Analysis
is a scalar function. For example. ˙ x0 = x(t0 ). x = 0. ¯ of (3.2). The conditions C.e. (3. If A(x) is any parametrization of f (x) then A(0) = f  x=0 .e. ∀ x ∈ Ω. Consider the following pseudolinear autonomous system x = A(x)x. The following Lemma from [6] establishes this fact. i. eA(x)t  ≤ M for some real M > 0 and ∀x ∈ Ω. C. C. interested readers are referred to [34] and to the references therein. αA1 (x) + (1 − α)A2 (x) is also a parametrization of f (x). ∀t ∈ [0. ∞).2)
In the remainder of this chapter.. we analyze the stability properties of the pseudolinear system (3. i. we see that the origin x = 0 is the only equilibrium point of the system (3. Lemma 3. A(x + θ) = A(x). all eigenvalues of A(x) lie in the open left half plane for all x ∈ Ω. we will use the notions of the coeﬃcient matrix A(x) and the system matrix interchangeably. A(x). For more details on the SDC parametrization. We conﬁne our analysis to the following conditions on the system matrix of (3.2). ¯ C.2).3 and C. Proof.2 A(x) is pointwise asymptotically stable (Hurwitz) matrix.e.
1 and C. . . By col{A(x)}. x ∈ Ω. It can be deﬁned as follows [68]: Deﬁnition 3. ∂x j = 1. Local Asymptotic Stability Analysis
23
3. n (3.2.3) in (3. that the conditions C. The requirements for global asymptotic stability of a nonlinear system of the form (3.2).3). 2. therefore. .. x ∂xi
we get. .4)
lim ψ(x.. z1 .. By col j {A(x)}. we mean the set of columns of A(x). we mean the jth column of A(x).
(3. z1 . z2 . Using (3. z n) =
j=1 i=1
xi x j ∂col j {A(z j )} . therefore x is a locally asymptotically stable equilibrium point of ¯ (3.. An equilibrium point x of the nonlinear system (3. There is only one equilibrium point. z1 ..
∂coln {A(z n )} x ∂x
x.
3.2) are the following.2 Local Asymptotic Stability Analysis
We start with the local asymptotic stability considerations.2).. z n) = 0.
xi x j
∂xi
. of the system. z2 ... z2 . Applying the Mean Value Theorem [49] to col{A(x)}.4).. on the line connecting the origin and the point x.3)
where the vector z j is a point. ¯ Since A(x) is continuously diﬀerentiable.... x = A(0)x + ψ(x.2 ensure that x is a locally asymptot¯ ically stable equilibrium point of (3.2) is (locally) ¯ asymptotically stable if it is stable. ˙ Since
x→0
(3. we can write x = A(0)x + ˙ = A(0)x +
j=1 i=1 ∂col1 {A(z1 )} x ∂x n n ∂col2 {A(z2 )} x ∂x j ∂col {A(z j )}
.
Multiplying and dividing the second term by x and deﬁning
n n
ψ(x.
. we can write col j {A(x)} = col j {A(0)} + ∂col j {A(z j )} x. which yields equality in the jth equation of (3.3. This means.. and if in addition there exists some r > 0 such that x(0) < r implies that x(t) → x as t → ∞. z n)x.2.3 Global Asymptotic Stability Analysis
Now we proceed to the global asymptotic stability considerations.5)
and A(0) is Hurwitz.. col{A(x)} ∈ C 1 . ¯ i.1.
x1 (t) =
3. lim x(t. Ultimately. it is analyzed with respect to the periodicity assumption on the system matrix.e. the state of the t→∞ ¯ system converges to the equilibrium point x as time goes to inﬁnity. then the equation ¯ x = A(x)x. But if the initial condition is taken as x0 (0) = [2 2]T .1.2 are not suﬃcient to guarantee global asymptotic stability of (3. The equilibrium point is locally asymptotically stable. x0 ) = x. Langson and Alleyne [44] studied this topic further and came up with the following two hypothesis:
.4 Exponential Boundedness and Global Asymptotic Stability
In this section. The conditions C. The counter example is the following nonlinear system x= ˙ −1 0
2 x1 −1
x(0) = x0 . the system (3. Therefore. T c ) with T c = ln 2. From this counter example it is obvious that the pointwise asymptotic stability of the system matrix A(x) does not help us to draw any conclusion about the stability of the nonlinear system (3.7)
This system satisﬁes Banks and Mhana’s hypothesis: A(x1 ) is continuous and asymptotically stable. we continue with the ﬁndings of Langson and Alleyne and ultimately give a counter example to show that global asymptotic stability is not guaranteed when the system matrix is exponentially bounded.4.
(3. it was proved wrong. It is obvious that as t tends to T c then x2 tends to 2 and consequently x1 tends to ∞.5. Global asymptotic stability of nonlinear systems in this form was ﬁrst studied by Banks and Mhana [7]. In Section 3. Such a departure of a state variable to inﬁnity at a ﬁnite time is called the ﬁnite escape time phenomenon. ∀ x0 ∈ Ω. If A(x) is a continuous matrixvalued function which is asymptotically stable for each x.. independently.7) is not asymptotically stable for all x0 .6)
x.
x0 = x(t0 ).2).3. global asymptotic stability is analyzed with respect to the exponential boundedness of the system matrix.
(3.2). then simple calculations yield the following solution of (3. They came up with the following remark: ¯ Remark 3. we continue with global asymptotic stability analysis. SDC Parametrization and Stability Analysis ii. ˙ is asymptotically stable for all x0 .24
Chapter 3. In the following two sections.1 and C. i.
¯ iii.7): 2x2 (t) and x2 (t) = 2e−t 2 x2 (t) − 2 √ for √ all t ∈ [0.2). This statement of Banks and Mhana is an optimistic extension of the eigenvalue based stability test for linear systems. in [44] and [76] by a simple counter example. to analyze the stability of the pseudolinear systems (3. starting from any point x0 ∈ Ω. In Section 3.
8). then A(x) is Hurwitz (asymptotically stable). x0 = x(t0 ). Consider the system x = A(x)x. we derive a general expression for the upper bound of the matrix exponential of the coeﬃcient matrix in (3. and the scalar function c(x): a < 0 and b > c(x) for all x ∈ R2 . ∀t ∈ [0.1 is satisﬁed with eA(x)t  ≤ M for some real M > 0 and ∀x ∈ Rn . then the system x = A(x)x is ˙ asymptotically stable for any arbitrary ﬁnite initial condition. Exponential Boundedness: Under this subject. We start with the following SDC formulation of a nonlinear system in a general setting x= ˙ a b − c(x) −b − c(x) a x. If the hypothesis of Proposition 3. Continuity: From the description of the coeﬃcient matrix in (3.1. b ∈ R and c(x) is a smooth function: c : R2 → R. in the sequel we write c instead of c(x).4. For the sake of convenience. The origin of the given system is an asymptotically stable equilibrium point. (3. a counter example [53] is presented in which the hypothesis of Langson and Alleyne is satisﬁed but where the system is not globally asymptotically stable.9)
Clearly. In the following subsection.8) satisﬁes the hypothesis of Langson and Alleyne.8). We proceed as follows
.
3.1
Counter Example Showing Exponentially Boundedness does not Guarantee Global Asymptotic Stability
Example 3.4. ∞). for certain choices of the parameters a and b. 3. 2. where A : Rn −→ Rn×n is uni˙ formly continuous in x and A(x) is a stable matrix ∀x ∈ Rn . Exponential Boundedness and Global Asymptotic Stability
25
Proposition 3. if a < 0 and b2 > c2 (x) for all x ∈ R2 .4. (3. We show that the coeﬃcient matrix in (3.4.8) has the following form λ1.3.4. 1.1.1. it is obvious that the coeﬃcient matrix is a continuous function: A : R2 → R2×2 is a continuous function.8)
where a.2 = a ± c2 (x) − b2 . Asymptotic Stability: The general expression for the eigenvalues of the system matrix in (3.4. Corollary 3.
1 0 0 γ−1 0 −b − c b−c 0 1 0 0 γ 0 k 0 k −k 0 −k 0
= =
.10)
We use here the fact that if A1 and A2 commute then eA1 +A2 = eA1 eA2 .
Hence. we get γ=± We know that eAt = T e(T b+c b−c and
(b − c)γ 0
γ−1 (−b − c) = k.
(3. Now consider the following transformation to make the antidiagonal entries of the matrix in the second exponent of (3.26
Chapter 3.11)
Therefore. .
(3.10) the additive inverse of each other. we can write (3. by taking the positive value of γ.12)
.
0 (−b − c)γ−1 Solving the pair of equations (b − c)γ = −k for γ and k.
(3.
−1
AT )t
T −1 .
and
√ k = ± b2 − c2 . SDC Parametrization and Stability Analysis
a b−c −b − c a e
t
=e
a 0 0 a 0 a
+ t
0 −b − c
b−c 0 b−c 0
t
t t
a =e 0
0 −b − c e b−c 0
0 −b − c at =e e
.10) as a b−c −b − c a e = eat 1 0 0 γ
t
√
√ 0 2 2 e − b −c
b 2 − c2 0
t
1 0
0 γ−1
. we have 0 −b − c e = 1 0 0 γ
b−c 0
t
√ 0 − b2 − c2 e
√ b 2 − c2 0
t
1 0
0 γ−1
.
Exponential Boundedness and Global Asymptotic Stability We know that 0 −x e
27
x 0
=
cos x − sin x
sin x cos x
.
. and c(x) = − π2 tan−1 x1 tan−1 x2 . Then e−0. and √ cos b2 − c2 t √  − sin b2 − c2 t  1 0 0 γ−1 √  ≤ 5 and  √ sin √b2 − c2 t cos b2 − c2 t 1 0 0 γ  = 1. asymptotically stable (Hurwitz).
for all x ∈ R2 and t ∈ [0.1.
(3.14).1t ≤ 1.4. For example. x0 = [1.1
and
 ≤ 5. √5 < γ < 5. Now. b = 3.
Therefore from (3. Fig.14)
We have derived the general expressions for the eigenvalues. we have −0.13). and exponentially bounded. (3. √ take we 8 1 a = −0.15)
It is clear from the foregoing discussion that the system matrix in (3.1 tan−1 x1 tan−1 x2
3+
8 π2
tan−1 x1 tan−1 x2 −0.15) with an initial condition.9) and (3.3.13)
Taking norm (we use the spectral norm) on both sides of (3. we get
a b−c −b − c a e ≤ eat  1 0 0 γ
 √ cos b2 − c2 t √  ·  − sin b2 − c2 t
t
√ sin √b2 − c2 t cos b2 − c2 t
 · 
1 0
0 γ−1
. we show that the hypothesis of Langson and Alleyne. The system matrix has the following form A(x) = −3 +
8 π2
−0.1 shows the phaseportrait of the system dynamics in (3.
(3. ∞). 3. It indicates that the states of the system move away from the origin as time goes to inﬁnity although the coeﬃcient matrix satisﬁes the suﬃcient conditions (as claimed in [44]) for global asymptotic stability.8).14). is satisﬁed if we take a suitable combination of the parameters a and b with scalar function c(x). Thus the hypothesis of Langson and Alleyne is satisﬁed.1
. (3. for the system matrix in (3.9).  ≤ √ 5.1 e
t
Re(λmax ) = −0.2 0]T . (3.8) using the system matrix (3.12) and using (3. and the upper bound of the matrix exponential.1 3 − c(x) −3 − c(x) −0.15) is continuous.
Suppose that the system (3. because regularity conditions are assumed to be satisﬁed. A(x) = A(x + θ) for all x ∈ Ω and some θ ∈ R. If a function is continuous and periodic on Ω then it must be bounded.8)
3. we prove how the periodicity condition rules out the possibility of the ﬁnite escape time phenomenon.2) satisﬁes the usual regularity (smoothness) conditions. what additional conditions would be required to establish global asymptotic stability of nonlinear systems of the form (3.17)
. If the system matrix in (3. In the following proposition.28
Chapter 3. ensures that the ﬁnite escape time phenomenon will not occur. Proposition 3. a natural question is. and can be written as x(t) = x0 +
0 t
A(x(s))x(s)ds.5 Periodicity and Global Asymptotic Stability
From the counter example in the previous section. Hurwitz and periodic then the system has bounded solutions for all times t < ∞. SDC Parametrization and Stability Analysis
xy−trajectories 3
2
1
0 State x2
−1
−2
−3
−4 −2 −1 0 State x1 1 2 3
Figure 3.5..e.1: Phaseportrait of the system dynamics (3. i. At this point. Proof. that the system matrix A(x) is periodic.2)? In addition to the smoothness conditions and exponential boundedness. The solution of (3. it is clear that the hypothesis of Langson and Alleyne is not suﬃcient to endorse global asymptotic stability of nonlinear systems of the form (3.4.
(3. we study the case that A(x) is also a θ−periodic matrix.2).2) exists.16)
Condition C.1.2) is continuous. (3.
Chapter 1..5. for all x1 ∈ R. ∀x ∈ Rn . Analysis of Coeﬃcient Matrix In order to present the counter example.5. the matrix A(x1 ) is a socalled Hurwitz matrix. i. i. Furthermore.2 = −α ± i 4 − α2 . Clearly.17) and using the fact that A(x) is bounded. With ε as above. page:5]. This completes the proof. it must be bounded. Periodicity and Global Asymptotic Stability
29
Since the system matrix A(x) is continuous. Further research on this topic results in yet another counter example which leads us to the conclusion that adding the periodicity condition alone does not suﬃce to guarantee global asymptotic stability.1..t. (3. we ﬁrst consider the matrix A(x1 ) = 2α 2 + 4α2 −2 −4α . Hence. the eigenvalues of A(x1 ) are given by √ λ1. Note that the matrix A(x1 ) is periodic with period π. since 4 − α2 > 0. the state does not blow up in ﬁnite time. s.1
Counter Example Showing that the Periodicity of the System Matrix does not Guarantee Global Asymptotic Stability
Example 3.. A(x1 ) = A(x1 + π) for all x1 ∈ R. we get x(t) ≤ x0 e Mt .18)
where α = sin2 x1 + ε. ∃M > 0. and periodic for all x ∈ Ω.
.5. Taking norm on both sides of (3. Therefore. This example has two parts: ﬁrst we analyze the system matrix and then analyze the corresponding system. Hence. the eigenvalues of A(x1 ) are located in the open left half of the complex plane. Hurwitz. The details are explained in the following subsection. A(x) < M. note that A(x1 ) depends on x1 in a continuous way. since α > 0. α is a real number such that 0 < α ≤ 1 + ε < 2 for any x1 ∈ R. where “i” denotes the imaginary unit.
Using Gronwall’s Lemma [80.e. by the nonzero complexvalued vector v1 = 2 3α +i √0 − 4 − α2 .e. √ An eigenvector for the eigenvalue λ1 = −α + i 4 − α2 is given. Further. i. all eigenvalues of A(x1 ) have a negative real part.
3.e. the eigenvalues of A(x1 ) are complex. This expression shows that the state of the system is bounded for all t < ∞. for instance. with ε ∈ R such that 0 < ε < 1. it follows that x(t) ≤ x0  +
0 t
Mx(s)ds.3.
it is clear that matrix T is invertible. Indeed. consequently. see [31]. but that A(x1 )v1 and λ1 v1 are complexvalued expressions. (3.e. It is wellknown (or easy to prove) that the Frobenius norm is submultiplicative. e =e − sin 4 − α2 t cos 4 − α2 t The matrix on the righthand side in the above equation has a ﬁnite norm. it follows that A(x1 ) and A(x1 ) √0 − 4 − α2 = 2 3α = −α 2 3α − √ 4 − α2 √0 − 4 − α2 √0 − 4 − α2
√ 4 − α2
2 3α
−α
.19) Because of the special form of the matrix D. −α 2 3α √0 − 4 − α2 = 2 3α √0 − 4 − α2 √−α − 4 − α2 √ 4 − α2 −α . 4 − (1 + ε)2
. D= √−α − 4 − α2 √ 4 − α2 .
Combining the expressions.19) that for all t ≥ 0. equating the real and imaginary parts of A(x1 )v1 and λ1 v1 . it follows that A(x1 ) = T DT −1 .. Then. it follows that the Frobenius norm of the real matrix √ √ cos 4 − α2 t sin 4 − α2 t . UVF ≤ UF VF for any two square matrices U and V of the same size.. √ √ 2 4 sin √4 − α2 t Dt −αt cos √ − α t . i. SDC Parametrization and Stability Analysis
It can indeed be veriﬁed that A(x1 )v1 = λ1 v1 . considering the Frobenius norm.30
Chapter 3.
By the restrictions on ε and α. The Frobenius √ 2 norms of the matrices T and T −1 are given by 8 + 8α2 and 2+2α2 . it follows from (3. Therefore. respectively. The Frobenius norm of eDt is then given by 2e−αt . page. it follows that A(x1 ) or A(x1 )T = T D. and. 332]) that for all t ≥ 0. that eA(x1 )t = T eDt T −1 . i. 4−α Let  · F denote the Frobenius norm. Hence. √ √ − sin 4 − α2 t cos 4 − α2 t √ √ is equal to 2. Note that A(x1 ) is a realvalued matrix.e. with T= 2 3α √0 − 4 − α2 . the square root of the sum of squared moduli of all matrix elements. it follows (see any book on diﬀerential equations such as [9. eA(x1 )t F ≤ T F eDt F T −1 F ≤ √ 8 + 8(1 + ε)2 2e−αt 2 + 2(1 + ε)2 .
3.5. Periodicity and Global Asymptotic Stability Since α > 0, it follows that 0 ≤ e−αt ≤ 1, for all t ≥ 0. Hence, for all t ≥ 0, √ (2 + 2(1 + ε)2 ) . eA(x1 )t F ≤ 2 2 4 − (1 + ε)2
31
So, there exists a number M ∈ R, depending on ε, such that eA(x1 )t F ≤ M, for all t ≥ 0 and for all x1 ∈ R. Recapitulating, the matrix A(x1 ) deﬁned in (3.18), with α = sin2 x1 + ε and with ε a ﬁxed number such that 0 < ε < 1, satisﬁes the following conditions: • A(x1 ) is Hurwitz for all x1 ∈ R, i.e., all eigenvalues of A(x1 ) have negative real part for all x1 ∈ R, • A(x1 ) depends continuously on x1 , • eA(x1 )t is “uniformly” bounded in t ≥ 0 and x1 ∈ R, i.e., there exists an M ∈ R such that eA(x1 )t F ≤ M, for all t ≥ 0 and x1 ∈ R. In [44], it is claimed that the above conditions are suﬃcient for global asymptotic stability of the associated diﬀerential equation. In [53], we showed that this claim is not correct. Here, in addition to the above properties, we also have that • A(x1 ) is periodic with period π, i.e., A(x1 ) = A(x1 + π) for all x1 ∈ R. However, this additional property in general, does not oﬀer anything extra as far as global asymptotic stability is concerned. This is shown in the following analysis of the diﬀerential equation. Analysis of the Diﬀerential Equation To show that the above conditions are generally not suﬃcient to guarantee global asymptotic stability, we consider the diﬀerential equation x = A(x1 )x, ˙ with x= x1 x2 , (3.20)
where A(x1 ) is as deﬁned in (3.18), with α = sin2 x1 + ε, and with ε a ﬁxed number such that 0 < ε < 1. Because A(x1 ) is Hurwitz for all x1 ∈ R, see the previous subsection, A(x1 ) is invertible for all x1 ∈ R. Therefore, diﬀerential equation (3.20) has only one equilibrium point, namely the origin. Moreover, since A(x1 ) is Hurwitz for x1 = 0, it follows that the linearization of (3.20) around the origin, given by x = A(0)x, is locally ˙ asymptotically stable. It is then well known, see for instance [80], that there exists a neighborhood of the origin such that solutions of (3.20), starting in this neighborhood, converge to the origin for time going to inﬁnity. However due to the nonlinearity of (3.20), and often also in general, there are initial conditions for which this convergence does not take place. Indeed, as numerical simulations show, there are initial conditions such that the resulting solution does not converge.
32
Chapter 3. SDC Parametrization and Stability Analysis
More speciﬁcally, for = 0.05, clearly satisfying 0 < < 1, starting from x1 (0) = 1, x2 (0) = 0, the solution of the diﬀerential equation spirals outwards. It further more follows from simulations that for x1 (0) = 1.5, x0 = 0, the resulting solution spirals inwards, but does not converge to the origin. As a consequence, by a Poincar´ e Bendixson kind of reasoning, see [80], both the solutions converge to a limit cycle that intersects the x1 −axis between 1 and 1.5. See also Figure 3.2, where the solutions are displayed starting from x1 (0) = 1, x2 (0) = 0 and x1 (0) = 1.5, x0 = 0, respectively. In Figure 3.2, the limit cycle to which the two solutions converge is depicted by a dashdotted line. This limit cycle intersects the positive x1 −axis around x1 = 1.265. Figure 3.2 also contains a second smaller limit cycle, also depicted by a dashdotted line. This second limit cycle intersects the positive x1 −axis around x1 = 0.460. It bounds the neighborhood around the origin in which solutions converge to the origin. Outside this neighborhood, the solutions converge towards the larger limit cycle.
xy−trajectories 2.5 2 1.5 1 0.5 State x2 0 −0.5 −1 −1.5 −2 −2.5 −2
−1.5
−1
−0.5
0 State x1
0.5
1
1.5
2
Figure 3.2: Phase portrait of the system dynamics (3.20).
Figure 3.2 indicates that the solution starting in certain initial states never reaches the only equilibrium state, as time goes to inﬁnity, although the coeﬃcient matrix satisﬁes the conditions in [44]. Hence, the origin, being the only equilibrium state, is not globally asymptotically stable. Also it is clear that the additional requirement of periodic state dependency of the coeﬃcient matrix does not help in attaining global asymptotic stability.
3.5. Periodicity and Global Asymptotic Stability Concluding Remarks
33
From the foregoing discussion, it is clear that the conditions C.1, C.2, and C.3 are not suﬃcient to guarantee global asymptotic stability of a nonlinear system written as a pseudolinear system in state dependent coeﬃcient form (3.2), even with the periodicity condition as well, see [53]. In general, periodicity of the system matrix does not imply global asymptotic stability of (3.2). However, this assumption helps in analyzing its global asymptotic stability as we have seen in Example (3.5.1). For instance, we assume that the system matrix A(x) is a function of a single state component and it is θ−periodic as well. Then (3.16) becomes A(x1 + θ) = A(x1 ), for all x1 ∈ and θ ∈ R, (3.21)
where, without loss of generality, x1 is the ﬁrst component of the state vector. The eigenvalue criterion is a wellknown approach to analyze asymptotic stability of linear systems. It is stated as follows [13, Chapter 7, page:203]. Theorem 3.5.1. Eigenvalue Criterion for Stability. The system x = Ax, ˙ x(0) = x0 , (3.22)
is asymptotically stable if and only if all the eigenvalues of the matrix A have negative real parts. Before the advent of fast computing machines like modern age computers, computation of eigenvalues was a cumbersome task. In 1892, a Russian mathematician, A. M. Lyapunov (18571918) proposed a method to analyze asymptotic stability of linear systems. This method does not require eigenvalue computation. Lyapunov narrated his method as follows [13, Chapter 7, page:205]. Theorem 3.5.2. Lyapunov Stability Theorem. The system (3.22) is asymptotically stable if and only if, for any symmetric positive deﬁnite matrix Q, there exists a unique symmetric positive deﬁnite matrix P satisfying the equation: PA + AT P = −Q. (3.23)
In (3.23), we can easily transform the matrix equation into a matrix inequality to get PA + AT P 0, −Q 0. (3.24)
For a given N × N matrix P, the constraint P( ) positive (semi)deﬁnite and the constraint P( ) negative (semi)deﬁnite.
0 would denote that the matrix P is 0 would denote that the matrix P is
Before proceeding further.21) where the system matrix depends on a single state component which varies periodically with period θ.34
Chapter 3.2) when the system matrix has the form (3. (3.25)
It is important to mention that. We see a similar situation in (3. Chapter 4] where asymptotic stability of linear systems is analyzed subject to parameter uncertainties. The ﬁrst step is to formulate an LMI feasibility problem using the concept introduced in the Theorem 3. x1k ∈ θ θ [− 2 . Intuitively. 2 ). does not distinguish between and (likewise between and ).
To verify this we proceed as follows. an LMI based approach is proposed and discussed to establish global asymptotic stability of the pseudolinear systems of the form (3. The choice Q = In helps us in stability analysis by ensuring strict inequalities. θ θ we take x1k ∈ [− 2 .t. The motivation for this approach comes from [66. ∀ x1 ∈ R. The system matrix A(x1 ) is θperiodic which means that θ θ it will be same at x1 = − 2 and x1 = 2 . AT (x1 )P(x1 ) + P(x1 )A(x1 ) + In 0.26) at k points. we formulate the LMI feasibility problem with nonstrict equality constraints. The solution of this LMI feasibility problem gives a constant symmetric positive deﬁnite matrix.21). SDC Parametrization and Stability Analysis
3. as we will see in the sequel. 2 )? This gives rise to k + 1 LMIs.5. The details of this approach are as follows. we are interested to investigate whether there exists a constant matrix P = PT 0 for which the following holds AT (x1 )P + PA(x1 ) + In 0.LMI Feasibility Problem Formulation Inspired by the Lyapunov Stability Theorem. if the LMI in (3.
Step I.2). ∃ P(x1 ) = PT (x1 ) 0 s. In the second step. Therefore. (3. 2 ) and formulate an LMI feasibility problem as follows: Does there exist a P = PT 0 which satisﬁes the LMI in (3. without loss of generality. we use the solution matrix from the LMI feasibility problem (Step I) to prove the global asymptotic stability of the system under consideration by using a quadratic Lyapunov function.5.26)
θ θ Therefore.2. 2 ] then it will hold for all x1 ∈ R.2
An LMI Based Approach for Global Asymptotic Stability
In this subsection. The system (3. we take Q = In as a constant identity matrix. We mark k (suﬃciently large) number of θ θ points on [− 2 . to ensure that all LMIs are distinct. it is important to mention that the LMI solver SeDuMi which we use in the sequel. we formulate the LMI feasibility problem as follows:
. we propose an equivalent result for the pseudolinear system (3. Hence. Mathematically.2) is pointwise asymptotically stable if and only if for each x1 ∈ R.26) holds true for all x1 in the principal interval [− 2 . This approach has two steps.
t ‘t’ ˙ V(x) = xT [AT (x1 )P + PA(x1 )]x. has a feasible solution. Periodicity and Global Asymptotic Stability Does there exist a symmetric matrix P which satisﬁes the following set of LMIs P − In AT (x11 )P + PA(x11 ) + In AT (x12 )P + PA(x12 ) + In AT (x1k )P + PA(x1k ) + In ? 0 0 0 0 . ¯
∀x
0. (3.2) can immediately be established if the LMI feasibility problem. (3.
(3. We illustrate our approach by the following example. .29)
˙ V(x) < 0.2). Let x = 0 be an equilibrium point for (3.31) w.
35
(3. global asymptotic stability of the system (3.27).30)
We deﬁne a quadratic Lyapunov function as: V(x) = xT Px. we recall the important BarbashinKrasovskii theorem from [40.31)
Here P = PT 0 is the solution of the LMI feasibility problem (3. Step II.32)
The function in (3.27)
We solve this LMI feasibility problem with YALMIP package (see [46] for more details) in MATLABT M using the solver SeDuMi (see [73] for more details). ∀x 0.Global Asymptotic Stability Analysis Here we explain how we use the solution of the LMI feasibility problem to establish global asymptotic stability of the system under consideration.3.27). (3. Chapter 4. Before proceeding further.28)
x −→ ∞
⇒
V(x) −→ ∞.5. .
Therefore.5.3 provided AT (x1 )P + PA(x1 ) −In 0.
(3. (3.r. ∀x1 ∈ R.3. Let V : Ω −→ R be a ¯ continuously diﬀerentiable function such that V(0) = 0 and V(x) > 0.31) satisﬁes all requirements stated in Theorem 3.
. Theorem 3.5. then x = 0 is globally asymptotically stable. Diﬀerentiating (3. page:124] which we use in the stability analysis.
8138 0. Then the matrix ¯ A(x1 ) = A(x1 ) − βI.9969 0.1795 0. We begin by deﬁning a system matrix of the following form.4766 −0. and B1 at random. . We present three diﬀerent results each with a diﬀerent choice of the number of LMIs in the feasibility problem. We consider a fourth order nonlinear system of the form (3.1991 −0. In a particular set. SDC Parametrization and Stability Analysis
Example 3.6450 −1.4458 0. We mark (k =)120.3630 −0. A1 . converge asymptotically to the origin.0964 0. having system matrix (3. x3 .35) is 2π− periodic. The system matrix (3.4113 0. Thus we form a system matrix which is continuous and periodic.35)
¯ where I is the identity matrix of the size of A(x1 ). 301. In particular. We see that the states converge asymptotically to the origin. π).9943 −0. We see in our experiments that for all such sets of matrices.8596 0 0 0 0.2). and 441 LMIs analogous to (3. x4 ) ∈ R4 .6970 −1. ∀x1 ∈ [−π.0308 −0. ¯ β = (1 + δ)λmax (A(x1 )).3 shows the behavior of the state dynamics of a system of the form (3. we generate the set of matrices A0 .6777 1. 300.33)
We generate the constant matrices A0 . we randomly generate A0 as a diagonal matrix and A1 and B1 as full matrices.5112 −0.0166 0. with arbitrarily chosen initial states. In retrospect of the theory discussed earlier. the state dynamics of the nonlinear system.5670 −1. A1 .1100 −0.2224 1.36
Chapter 3.35) and A0 .2). In this way. x2 .7474 1. π].5. This indicates that the states of the system may converge asymptotically to the origin for all initial conditions x0 = (x1 .8713 0. and B1 are as given later in this example.9884 −0.3664 . and 440 points on the interval [−π. periodic. will be Hurwitz.3099 −0. and Hurwitz matrix. As has been mentioned earlier.4211
0.8674 −0.1287 0 0 0 0.3146 0 A0 = 0 0 0 0 0 0. A1 . Figure 3. To obtain a continuous.27). and B1 at random. (3.0442 0. we have the following matrices 0.34)
where δ > 0 is a small number and λmax denotes the maximum real part of the eigen¯ values of A(x1 ). with four diﬀerent arbitrarily chosen initial conditions.4742 1. we deﬁne. (3.0728 0. the ﬁrst step is to formulate an LMI feasibility problem.2.
−0. we form a system of 121.2958 −1.1680 B1 = −0. ¯ A(x1 ) = A0 + A1 cos x1 + B1 sin x1 . (3.5990 A1 = −0.6971
. .
6174 −0.0830 3.3: Behavior of the state dynamics of the system (3. SeDuMi yields 4.3.0830 −0.5051.4611 5.0835 −0.0967 . P2 is a SPD matrix. For k = 440. and 4. −4.7498 The eigenvalues of the matrix P1 are 2.0278 −0.7827.1847 −0. Since all eigenvalues are positive. (b). 3. in 7 iterations. respectively.0721 −0.6174 0.6450 The eigenvalues of the matrix P2 are 3.6626. SeDuMi
.4611 −0.0278 P2 = −0. 4.5492 0. For k = 120.5683 −0.0050 2. and 5. 6. −6.5515. and 441 LMIs and we get the following feasible solutions.0835 −0.0050 −0.5492 −0. −7.4727 −0.( π .2)(3. in 7 iterations. 6.1140 3. −2. 4. 5 5 2 (d).0759.4821 4. P1 = −0.5.(− 3π .0345 . 3.8798 −0. −4). in 9 iterations.1219 0.3606. 5).0967 0. Periodicity and Global Asymptotic Stability
(a) State x1 State x2 State x3 States State x4 6 4 2 0 −2 −4 0 2 4 6 0 5 10 (b) State x1 State x2 State x3 State x4
37
4 2 States 0 −2 −4
(c) 10 State x1 State x 5 States
2
(d) 8 6 4 2 States 0 −2 −4 State x1 State x2 State x3 State x4 0 5 Time [sec] 10
State x3 State x4
0
−5
−6 −8 0 1 2 3 Time [sec] 4 5
Figure 3. (c).(− π .4821 −0. P1 is a SPD matrix. 7).01847 −0. 4 We solve the three LMI feasibility problems consisting of 121.35) with four arbitrary initial conditions (a). SeDuMi yields 3. −3. 9). For k = 300.2274.3305 −0.1140 0. 301. 3. Since all eigenvalues are positive. 0.0345 4.6743.( 2π .
2. k vs x1 corresponding to LMI systems consisting of 121.0303 −0. i = 1. Therefore.2747 9. 3 vs state component x1 . 2.8841 −0. π].2495 −0.1634 −0.3775 −0.8016 0. a quadratic Lyapunov function V(x) = xT Pi x.
Profiles of maximum eigenvalues 0
k=120 −5
−10
−15 λmax k=300
−20
−25
−30 k=440 −35 −4
−3
−2
−1
0 x1
1
2
3
4
Figure 3.2355.9676. We see that the curves lie below the horizontal axis. Since all eigenvalues are positive. π]..
. and 11. This implies that the symmetric matrices AT (x1i )Pi + Pi A(x1i ). i = 1. 8. P3 is a SPD matrix.0303
−1. .1634 −0. 3 satisﬁes all the conditions for global asymptotic stability stated in Theorem 3.. 2. ∀ x1i ∈ [−π.5.3..4366..4 shows the maximum eigenvalue proﬁles of AT (x1i )Pi + Pi A(x1i ) for all x1i ∈ [−π. are negative deﬁnite and the pattern of the eigenvalues proﬁles shows that they become more negativedeﬁnite as the size of the system of LMIs grows. this fourth order nonlinear autonomous system is globally asymptotically stable. 2.. 301 and 441 constraints.4: Proﬁles of the maximum eigenvalues of the symmetric matrices AT (x1i )Pi + Pi A(x1i ). π). Figure 3.2495 7.38 yields
Chapter 3.8016 10.
The eigenvalues of the matrix P3 are 6.4852 0.4852 0. .2747
0. 9. ∀ x1i ∈ [−π.5000. Now.3402 P3 = −1. k and i = 1. SDC Parametrization and Stability Analysis
8. i = 1..5379
.
But it is a stabilizable system.36)(3.3 because V(x) is only positive semideﬁnite in this case.
. (3.36)
This model has been introduced in Chapter 2 where more details about the model can be found.40)
where R denotes the additional damping eﬀect introduced via controller action.
3. Hurwitz.5. If the LMI feasibility problem does not have a feasible solution then we can not directly deduce global asymptotic stability with this method. We require LaSalle’s Principle for stability analysis in such situations. We formulate an LMI feasibility problem as described in (3.5 shows the state dynamics of the closed loop system (3. depends on the solution of the LMI feasibility problem.2. the states converge asymptotically to the origin.38)
Figure 3. This means. We use the following stabilizing state feedback control law τ= −MJ T (ψ) D − 2M η ν . (3. Clearly. We see that the system matrix in (3. (3.36) is not Hurwitz. Remark 3.39)
We get a feasible solution for this LMI feasibility problem from SeDuMi.37) can be τ= −MJ T (ψ) R − 2M η ν .3.5. The outcome from the LMI solver SeDuMi is: the LMI feasibility problem is infeasible. (3. Consider a naive modiﬁcation in the inequality constraints (3.5.3
Infeasibility of the LMI Feasibility Problem
Example 3. Periodicity and Global Asymptotic Stability
39
The above described LMI based Lyapunov method to prove global asymptotic stability of the systems of the form (3. Consider the following autonomous nonlinear system η ˙ ˙ ν = O3 O3 J(ψ) −M −1 D η ν + O3 M −1 τ.3. Although.27) by using the closed loop system matrix (3.5.1.37)
This results in a continuous. Remark 3.2).
(3.5. this method may still be useful with a naive modiﬁcation in the LMI constraints. But we ˙ cannot use Theorem 3. The following subsection explains this situation. the LMI based Lyapunov stability approach is not working in this case.27) which results the following form AT (ψ)P + PAc (ψ) + c O3 O3 O3 I3 0.5.38).37) with four diﬀerent arbitrarily chosen initial conditions. and 2π−periodic feedback system matrix given by Ac (ψ) = O3 −J T (ψ) J(ψ) −2I3 . A more practical and useful form of (3.
− 8π .38). 0.5. SDC Parametrization and Stability Analysis
(a) 2.5 0 −0. 0. −2.(2. − 2π .(2.
For the sake of convenience. 2.
.5 0 5 10 −1 −2 0 5 10 x y ψ u v r 2 1 States 0 x y ψ u v r (b)
(c) 5 1 0 x y ψ u v r 0 5 Time [sec] 10 States
(d)
0 States
−1
−5
−2
−10 0 5 Time [sec]
x y ψ u v r 10
−3
Figure 3. 0. we approach the LMI feasibility problem associated with the system in Example 3.5 1 States 0. 6 3 (c).5: Behavior of the state dynamics of the system in Example 3.t.(−1. π . we would like to ﬁnd a SPD matrix P s. Analytic approach to the LMI problem For the closed loop system matrix (3. from here onward in this subsection. analytically. 0). 0. π . 4).5 −1 −1. 0. AT (ψ)P + PAc (ψ) c Suppose that P has the following form P= Pa Pb PT b Pc . 0). 9 3 To investigate the infeasibility issue in more detail. Now consider.3 with four arbitrary initial conditions (a). AT P + PAc c 0. −3. (b). 4. 0. in the following subsection.5. π].3. −1. we omit the argument ψ from J(ψ). (d).5 2 1. 0).40
Chapter 3.(−10. 0. ∀ ψ ∈ [−π. 4.
41 holds true only if JPb + (JPb )T 0. Therefore. The inequality 3. In the following. ∀ ψ ∈ [−π. (3. about the LMI feasibility problem of Example 3. π].3. does not hold.3.36) has no solution. Consider Pb in the following form Pb = Pb11 ∗ ∗ ∗ .
Now the inequality (3.36) by using Lyapunov stability theory and LaSalle’s invariance theorem. we continue with global stability analysis of the system (3. JPb + (JPb )T = −Pb11 − PT11 b ∗ ∗ ∗ .
(3.42) to hold true.
This is an improbable scenario.41)
−JPb − (JPb )T (Pa J − 2PT − JPc )T b
Pa J − 2PT − JPc b (Pb J)T + Pb J − 4Pc
0.5.”
In the foregoing discussion. we have proved analytically that the LMI problem corresponding to the system (3. ψ = π. JPb + (JPb )T = Case 2.43)
In the sequel. we show that the necessary condition (3. we conclude the following “ P = PT 0 s.42)
To be brief. π]. ⇐⇒ −
−JPb − (JPb )T (Pa J − 2PT − JPc )T b
Pa J − 2PT − JPc b (Pb J)T + Pb J − 4Pc
0.t. Since (3.5.
.43) for (3. Periodicity and Global Asymptotic Stability O3 JT −J −2I3 Pa Pb PT b Pc Pa Pb PT b Pc O3 −J T J −2I3
41
⇐⇒
+
0. holds true for all ψ ∈ [−π.
where Pb11 is a 2 × 2 block entry while the entries of Pb which are not relevant for the analysis are marked by aestericks. we focus on the leading principal block diagonal entry −JPb − (JPb )T .43) must be true for all ψ ∈ [−π. AT (ψ)P + PAc (ψ) c 0.
⇐⇒ Or equivalently. we take the following two cases: Case 1.
(3. Pb11 + PT11 b ∗ ∗ ∗ . π]. ψ = 0.43) holds only if the following holds Pb11 + PT11 b 0 and − Pb11 − PT11 b 0.
Hence (η. The system matrix (3.5. is the trivial solution.36) and (3. we recall an important result from stability theory which was ﬁrst formulated by a famous French priest and education reformer. Let V : Ω −→ R be a ˙ continuously diﬀerentiable function such that V(x) ≤ 0 in Ω0 . we conclude that the only solution of the closed loop system (3. J. Suppose η(t) and ν(t) is a solution pair of the closed loop system (3.45). SDC Parametrization and Stability Analysis
3. from (3.5.36) and (3. Therefore. ⇒ −J T (ψ)η(t) ≡ 0. Let M be the largest invariant set in E. An invariant set is deﬁned as: Deﬁnition 3. ˙ E = {x : V(x) = 0} = {ν = 0. The LaSalle’s Invariance Theorem (Principle). (Invariant Set) A set M is an invariant set for a dynamic system if every system trajectory which starts from a point in M remains in M for all future time.37) which stays identically in El . (3.4.5. ∀ x 0.38) has a special structure with a skew symmetric part. page:128]: Theorem 3. Diﬀerentiating the Lyapunov function.2). Then ˙ ν(t) ≡ 0. a Lyapunov function of the form V(x) = xT x.42
Chapter 3.44)
˙ V(x) = −4νT ν.37) that belongs identically to El .36) and (3. It is stated as follows [40. we get ˙ V(x) = xT ⇒ ⇒ Now.37) . ⇒ η(t) ≡ 0. and η is arbitrary. Let us consider.
.4
Further Analysis for Global Asymptotic Stability
Before proceeding further. 0) is a globally asymptotically stable equilibrium point of the the closed loop system (3.45) ⇒ ν = 0. ˙ V(x) = 0. B.1. η is free}. Let E be the set of all ˙ points in Ω0 where V(x) = 0. Let Ω0 ⊂ Ω be a compact set that is positively invariant with respect to (3. (3. The next step is to ﬁnd the largest invariant subset of El . de LaSalle (16511719). We try to use this structure to analyze global asymptotic stability of the system. ⇒ ν(t) ≡ 0. Chapter 4. Then every solution starting in Ω0 approaches M as t → ∞.
Therefore. O3 −J T (ψ) J(ψ) −2I3 x. ˙ V(x) ≤ 0. ν) = (0.
This method combines the concepts of the Lyapunov stability theory and the LMIs. More investigation to this bottleneck revealed that a small modiﬁcation in the LMI feasibility problem.36) subject to a naive state feedback controller (3. It turns out that the LMI feasibility problem associated with the closed loop system matrix (3. This outcome is then proved analytically. Local asymptotic stability is easy to analyze but the major topic of interest in this chapter is global asymptotic stability analysis. This approach is motivated from the idea of LPV approach introduced in [66] to the pseudolinear systems of the form (3. Example 3. We ﬁrst explained this method by using an academic example of a fourth order pseudolinear system. the system matrix in new representation depends on a single state variable.1 shows that.2). Global asymptotic stability is then followed by Lyapunov stability theory.4. However.2).2). Afterwards. The positive (semi)deﬁnite solution of the LMI feasibility problem is then used to get a quadratic Lyapunov function. These nonlinear systems when transformed to pseudolinear form by using the SDC framework. YALMIP uses SeDuMi (an LMI solver) which transforms the LMI feasibility problem into a convex optimization problem. Summary and Conclusions
43
3.39). we introduce another method to prove the global asymptotic stability of the pseudolinear system (3.1 shows that the exponential boundedness of the system matrix together with the continuity and Hurwitzness conditions does not imply global asymptotic stability of the nonlinear dynamical system (3.37). resolves this problem.6. In this case. in addition.38) does not have a feasible solution. The consequence of this modiﬁcation is that the derivative of the Lyapunov function will be positive semideﬁnite in this situation. The main contributions are two counter examples. it does not ensure global asymptotic stability of (3.5. as a consequence of the periodicity of the system matrix. Example 3.6 Summary and Conclusions
This chapter focuses on the stability analysis of a special type of nonlinear systems.3. even if the periodicity of the system matrix is assumed. we proceed to prove global asymptotic stability of the nominal vessel model (3. We use the periodicity of the system matrix together with the Lyapunov stability theory and the LMIs to formulate an LMI feasibility problem.
.2). the LaSalle’s invariance principal is used to establish global asymptotic stability of the pseudolinear system. see (3. This LMI feasibility problem is solved by using the YALMIP interface under MATLAB. We start with a brief review of the past results on this topic.
44
.
and yaw motions. The purpose of the notch ﬁlter was to ﬁlter out the eﬀects of the WF motions from the feedback loop. for instance see [1]. This observer was then separately used in combination with a backstepping controller [69] and a PID controller [70]. An SDARE observer is proposed to estimate the states of a DP vessel model. Our aim is to study the SDARE controller and observer design techniques in connection with the control design problem of a DP vessel. These eﬀorts made way for the modern concepts for wave ﬁltering like state estimators or observers. Kongsberg. The state dependent algebraic Riccati equation plays
T
1 Part of this work has been published in the proceedings of the UKACC International Conference on Control (CONTROL 2010). One such approach is reported in [69] and it is based on the state dependent algebraic Riccati equation. and yaw motions and there is a phase lag in the control loop due to the notch ﬁlter. to control each of the surge. also known as SDARE. The nonlinear model of the vessel motion described in Chapter 2 can be assumed as a pseudolinear system by considering the SDC parametrization which is discussed in Chapter 3.
45
. There were two main disadvantages of this approach: the integral action of the PID controller is very low due to the coupling between the surge. In this chapter. sway. to initiate eﬀorts to invent new DP control systems using the concepts of optimal control theory and Kalman ﬁltering from the modern control theory. and [47]. sway.Chapter
4
1
State Dependent Riccati Equation based Control Design
e use of DP systems for the positioning control of ships started in the early sixties. [27]. we extend the ideas of LQR theory to address the nonlinear control design and estimation problems for the DP vessel. Observers are more useful than the classical ﬁltering techniques because they also give estimates of the unknown states of the system. Observer based wave ﬁltering techniques have been reported by many researchers. see [52]. The disadvantages of the PID control law motivated a Norwegian company. The ﬁrst DP control system used the conventional PID type control law together with a notch ﬁlter.
Deﬁnition 4. see Section 3.
4.1. u ∈ Rm is the control input vector. In [6]. ˆ Then the system (4.3) with control u(x) = −Kc (x)x. Let us start by deﬁning the optimal control (OC) or linear quadratic regulation (LQR) problem for the pseudolinear system describing the sea vessel motion. (LQR/OC problem for pseudolinear system of vessel model) Given matrices Q and R. Theorem 4. (4.1. R ∈ Rm×m is SPD matrix.2). The following theorem states this result. ˙ (4.1. x(t0 ) = x0 .46
Chapter 4.5) (4.1) are termed as weighting or penalty matrices.1 Optimal Control Problem and the SDARE
We mimick the LQR problem for the linear systems to address the control design problem of the pseudolinear vessel system. subject to x = A(ψ)x(t) + Bu(t). the ndimensional vector x ∈ Rn describes the state of the system. the expression xT (t)Qx(t) describes the deviation of the state vector from the origin (equilibrium state) and the expression uT Ru describes the control cost. We will see that the SDARE controller is a state feedback or an output feedback controller.
t0
(4. The matrices Q and R in (4. where the controller gain is Kc (x) = R−1 BT (x)Πc (x). B(x)) is stabilizable for all x ∈ Ω ⊆ Br (0). u(t)) = is minimized. and the matrix B ∈ Rn×m describes the controller conﬁguration. ﬁnd a control signal u(t) such that the quadratic cost function J(x0 .1 has a suboptimal solution. Q = H T H ∈ Rn×n (H T ∈ Rn×p ) is SPSD matrix. it has been proved that the problem deﬁned in Deﬁnition 4. and Q ˆ ˆ are such that the pair (H.4)
. where Ω is a nonempty neighborhood of the origin. B(x). r > 0. We adapt the deﬁnition of the LQR problem from [13] to customize it according to the pseudolinear model under consideration. A(ψ) ∈ Rn×n is the system matrix. ˙ (4. A(x)) is detectable and the pair (A(x).2) 1 2
∞
(xT (t)Qx(t) + uT (t)Ru(t))dt. These are also called the tuning parameters of the SDARE controller as the performance of the SDARE controller can be manipulated by playing around with these matrices. Assume further that the matrices A(x). SDARE based control Design
a fundamental role in this theory.1) and (4.1)
In (4. Physically. we assume Q and R to be constant matrices but they can be state dependent as well.1. (The SDARE controller) Consider a nonlinear system in the SDC form x = A(x)x + B(x)u. Although.1.1.3)
The SDC form implies that the matrices A(x) and B(x) are continuous in x for all x ≤ r.
The SDARE (4. In (4. respectively.6)
4. The suﬃcient conditions for the pointwise existence of the solution of the SDARE (4. i. (4. Control System Design for the DP Vessel
47
is locally asymptotically stable. Rc ). which is given by Πc (x)A(x) + AT (x)Πc (x) + Q − Πc (x)B(x)R−1 BT (x)Πc (x) = 0. They are described by the zeromean Gaussian white noise processes.2 Control System Design for the DP Vessel
In this section.
4. Qc. Πc (x) is the solution of the SDARE associated with (4.7) as given in [49]. In the ﬁrst part.6) in context of the vessel model becomes Πc (ψ)A(ψ) + AT (ψ)Πc (ψ) + Q − Πc (ψ)BR−1 BT Πc (ψ) = 0. The vectors wb and v represent the process and measurement noise. Proof. wb ∼ (0. are: the pairs (A(ψ).1) and (4. b ∈ R3 describes the environmental disturbances.
−1 −1 T −1
The LF motion as described in Chapter 2 is given by (4. A(ψ)) are. pointwise stabilizable and detectable for all ψ ∈ R.2. A DP vessel model is given in Chapter 2. First. r]T is velocity coordinate vector. (4. ν = [u. we assume that there is no WF motion in the measured output and only LF motion is to be regulated. and τ represents the control forces and torque. we illustrate the SDARE based controller and observer design technique for the control system design of the DP vessel.5). We study the LF motion regulation problem by dividing it in two parts. The LF model in the SDC form may then be written
. Many solutions of (4.b ) and v ∼ (0. We divide our study into two steps. B) and (H. we assume that slowly varying bias forces and moments are unknown and hence should be estimated.7)
This is a special SDARE in the sense that only the system matrix depends on state and this state dependence too is only on a single state component in a periodic way. Disturbance vector b is known ˆ We assume that the disturbance vector b is known or its estimate b is available and there is no measurement noise. respectively.9) (4.10) (4.3).7) could be possible. ψ]T is position coordinate vector.2.11)
yb =Cb η + v. we assume that the bias forces are completely given or their estimate is available. In the second part.. but the one which is positive (semi)deﬁnite is desired for most of the applications.
where η = [x. we assume that the WF motion component is also present in the measured output. ˙ ˙ ν = − M Dν + M J (ψ)b + M τ.1
Nonlinear Regulation Problem for the LF Model
η =J(ψ)η.8) (4. See [6]. y. ˙ b = − T −1 b + Ψwb .e.4. v. I. Later on.
But we can also write (4.18)
This control law can be seen as PDtype control law with feedforward term ‘−J T (ψ)b’.12) is both pointwise controllable and observable.12) as follows u = −R−1 M −1 Π21 (ψ) Π22 (ψ) η ν − J T (ψ)b.1) and (4. O3 M −1 J(ψ)M −1 −M −1 DM −1 ··· ··· J(ψ)(−M −1 D)4 M −1 (−M −1 D)5 M −1 . We can write the SDARE state feedback controller. (4. the system (4. Therefore. it is clear that we can only measure the position of the vessel and its velocities are not available for feedback. . η ν +
Chapter 4. (4. .19)
. From (4. we deﬁne a locally asymptotically stable SDARE controller for (4.14)
(4. We prove by using the LMI based technique discussed in Chapter 3 that the SDARE feedback controller stabilizes the vessel globally asymptotically around the origin. . (4. SDARE based control Design
O3 M −1
u+
O3 M −1 J T (ψ)
b.12) (4.1.48 as ˙ η ˙ ν = yb = O3 O3 I3 J(ψ) −M −1 D O3 η ν .16).1. O3 O3 J(ψ) . (4. J(ψ)(−M −1 D)4 . Π21 (ψ) and Π22 (ψ) comes from Πc (ψ) = Π11 (ψ) Π21 (ψ) Π12 (ψ) Π22 (ψ) .16) The local asymptotic stability of the state feedback SDARE controller follows from the Theorem 4.16)
The control law (4.12).15) have full row and column ranks.14) and (4. Therefore. by using Theorem 4. Stability analysis of the controller (4.1.
(4.1.16) as u = −R−1 M −1 Π21 (ψ) Π22 (ψ)J T (ψ) η η ˙ − J T (ψ)b. . (4.13)
The corresponding controllability and observability matrices are given by Cb = and Ob = I3 O3 .13). respectively. (4. as u=− K1 (ψ) K2 (ψ) η ν − J T (ψ)b. Moreover.16) can be described as a state feedback law with a feedforward term ‘−J T (ψ)b’.17)
which is the symmetric solution of the SDARE associated with (4.15)
The matrices deﬁned in (4.
⇒ O3 −M −1 K1 (ψ) J(ψ) ∗ v 0 =0 v 0 .2. 2π]. therefore..
≤ −νT ν. Suppose that it is singular then ∃ v 0 such that M −1 K1 (ψ)v = 0.
(4. it turns out that there exists an SPD matrix P such that AT (ψ)P + PAcl (ψ) + cl 0 0 0 I3 ≤ 0. Now we need the invariant sets of the closed loop system.12) and (4.22)
To show that η(t) ≡ 0. we use LaSalle’s invariance principle to further analyze the stability of the closed loop system.e. ‘t’.20)
It is wellknown fact that the SDARE controller results in an asymptotically stable closed loop matrix. ˙ Since V(x) ≤ 0. ≤ 0.19) is given by ˙ η ˙ ν = O3 −M −1 K1 (ψ)
Acl (ψ)
49
J(ψ) ∗
η ν
. We prove this by contradiction. η is free}. i. Diﬀerentiating w. Control System Design for the DP Vessel The closed loop system (4. (4. Using the LMI approach (see Chapter 3 for more details).
.r. Let η(t) and ν(t) be a solution pair of the closed loop system (4. η(t) ≡ 0 and ν(t) ≡ 0 for all ‘t’.20) is pointwise Hurwitz. the closed loop system matrix (4. Then ˙ ν(t) ≡ 0. To show that the origin is the only equilibrium point of the closed loop system. (4. we get ˙ V(x) = xT (AT (ψ)P + PAcl (ψ))x. ˙ El = {x : V(x) = 0}. ⇒ ν(t) ≡ 0. ∀ ψ ∈ [0. ⇒ −M −1 K1 (ψ)η(t) ≡ 0. using (4.20) that belongs identically to El . Therefore.t. = {ν = 0. For this we ﬁrst ﬁnd the set of points where ˙ V(x) = 0. we need to prove that K1 (ψ) is a nonsingular matrix.21)
Now we consider the quadratic Lyapunov function: V(x) = xT Px.21). cl ≤ xT 0 0 0 −I3 x. we need to show that no solution other than the trivial solution stays identically in El .4.
26)
Since the matrices M. . The LF motion in this case.23)(4.23) is given by Πc (ψ)Ab (ψ) + AT (ψ)Πc (ψ) + Q − Πc (ψ)Bb R−1 BT Πc (ψ) = 0. O3 O3 J(ψ)M −1 J T (ψ) . u = −R M (4. Cb = M O3 O3 ··· O3 and Ob = I3 O3 O3 . .1. ˙ (4.20) is pointwise asymptotically stable (all eigenvalues lie in the left half plane). b b (4.
(4. The observability matrix (4. SDARE based control Design
This implies that λ = 0 is an eigenvalue of the closed loop system matrix in (4. II.23) ˙ −1 O3 Ψ O3 O3 −T b b η (4.27)
Using Theorem 4. b The controllability and observability matrices for (4.25)
(4.28) b
. D. O3 J(ψ) −J(ψ)M −1 D . and J T (ψ) are nonsingular. . K1 (ψ) is nonsingular. Therefore. . from (4. Hence (η∗ .25) has rank 6. controllability matrix (4. . This means that only the states η and ν are controllable. we conclude that the only solution of the closed loop system which stays identically in El is the trivial solution. An SDARE associated with (4. is described as J(ψ) O3 O3 η O3 η O3 ˙ −1 ν = O3 −M −1 D M −1 J T (ψ) ν + M u + O3 wb . Disturbance vector b is unknown We now assume that the LF disturbances are slowly varying (as described by the ﬁrst ˙ order Markov process b = −T −1 b + Ψwb ) but unknown and should be estimated. This is not restrictive as the bias forces are always uncontrollable.1. .24) yb = I3 O3 O3 ν + v.24) are given by J(ψ)M −1 · · · J(ψ)(−M −1 D)7 M −1 O3 −1 −M −1 DM −1 · · · (−M −1 D)8 M −1 . it is a contradiction. Hence. a locally asymptotically stable SDARE based suboptimal control law is given by η −1 −1 Π21 (ψ) Π22 (ψ) Π23 (ψ) ν . . 0) is a globally asymptotically stable equilibrium point of the the closed loop system (4. in the SDC form.50
Chapter 4. therefore.22).19) globally asymptotically stabilizes the vessel.26) has full rank for all ψ ∈ R.20) which means that the SDARE controller (4.20). Therefore. ν∗ ) = (0. But it is a wellknown fact that the closed system (4.
. the corresponding dual system will be pointwise controllable. n are continuous in x for all x ≤ r. we can design a suboptimal locally asymptotically stable SDARE observer. ˆ ˆ ˆ ˆ ˆ ˆ associated with the system (4. An SDARE observer for the system (4.23)(4. ˆ ˆ ˆ ˆ ˆ ˆ
(4.34).29)
where Π21 (ψ). a state estimator is required to estimate the complete state. we notice that only the LF position and heading can be measured by the sensors.. Π22 (ψ). where Ω is a nonempty neighborhood of the origin.30) ˆ b Since the system (4. and Π23 (ψ) are obtained from the solution of the SDARE (4.
51
(4.33).33)
will converge locally asymptotically to the state. Therefore. ˆ (4.37) ˆ b (4.4. Theorem 4. Assume further that the ˆ ˆ pair (A(x). (4.31) (4.24). 2. A suboptimal locally asymptotically stable SDARE controller for the system (4.36) ˆ ˆ −1 ˙ ˆ O3 O3 O3 −T b b η ˆ ν y η = I 3 O3 O3 ˆ .2.32)
f (x) is such that f (x) = A(x)x and ∂∂x j . From (4.23)(4.32). ˙ y = C(x)x. (4.31)(4. Πo ( x) is the solution of the dual SDARE ˆ Πo ( x)AT ( x) + A( x)Πo ( x) + U − Πo ( x)C T V −1CΠo ( x) = 0..34)
. C(x)) is pointwise detectable and the matrices A(x) and C(x) are locally Lipschitz for all x ∈ Ω ⊆ Br (0). In (4.27). Proof. ˆ ˆ In (4. (The SDARE state estimator) Assume that a system in SDC form x = A(x)x + B(x)u.2. Control System Design for the DP Vessel The gain of the SDARE controller is deﬁned by Kc (ψ) = R−1 M −1 Π21 (ψ) Π22 (ψ) Π23 (ψ) . r > 0 and x = 0 is a stable equilibrium point of (4. ˆ Then the estimated state given by
˙ x = A( x) x + B( x)u + Ko ( x)(y − y). j = 1.24) is given by ˙ ˆ η O3 ˆ J(ψ) O3 η O3 ˆ ν −1 ˙ ˆ ˆ ˆ ν = O3 −M −1 D M −1 J T (ψ) ˆ + M u + Ko (ψ)(yη − yη ).24) will be η ˆ ˆ ˆ ˆ ˆ u = −R−1 M −1 Π21 (ψ) Π22 (ψ) Π23 (ψ) ν .1. (4. The control law (4. Therefore. See [6]. The following theorem states this result.31). Ko ( x) is the observer gain ˆ and is given by Ko ( x) = Πo ( x)C T V −1 .35) (4. .23)(4. see [6].28) can be described as a state feedback controller.24) is pointwise observable.
The following theorem from [6] states one particular result about stability of an SDARE compensator.2. SDARE based control Design
The third term in (4. C(x)) is detectable and the pair (A(x).36). The separation properties for combination of linear controllers and observers are well deﬁned. A combination of both is called a nonlinear compensator. 0) is a locally asymptotically stable point. 2. Assume further that A(x) and B(x) are continuous. and yaw directions is modeled as second order harmonic oscillation. e) = (0. e = x − x is the ˆ error between the state x from (4.39)
4.2. where Ω is a nonempty ˆ neighborhood of the origin. see Chapter 2 for more details..1 and is given by ˙ x ˙ x ˆ = Ab (ψ) ˆ Ko (ψ)Cb ˆ −Bb Kc (ψ) ˆ ˆ ˆ Ab (ψ) − Bb Kc (ψ) − Ko (ψ)Cb x x ˆ + Eb O O ˆ Ko (ψ) wb wb . So far we have deﬁned a suboptimal locally asymptotically stable nonlinear control law (4. v (4. (4..30) and a suboptimal locally asymptotically stable nonlinear observer (4. 3 are obtained from the solution of the corresponding dual SDARE of the form (4.35). B(x)) is stabilizable for all x ∈ Ω ⊆ Br (0).36) is an SDARE pseudolinear compensator illustrated in a block diagram shown in Figure 4. A combination of the SDARE controller (4.52
Chapter 4. ˆ Proof.41)
. ˆ Π31 (ψ) where Πi1 .31) and the state estimate x from (4. j = 1.31)(4.2
Nonlinear Regulation Problem for the Complete Model
Now. see for instance [55]. The WF motion model is given by ˙ ξ1 ˙ ξ2 = yξ = O3 −Ω2 I3 I3 −2ZΩ O3 ξ1 ξ2 . The WF motion in each of the surge. i = 1. Then (x. Let the ˆ ˆ matrices A(x). The separation principles for the nonlinear compensators are not well established yet. 2. Theorem 4. r > 0. . n are continuous in x for all x ≤ r. Here.32) is such that f (x) = A(x)x and ∂∂x j . sway. and C(x) are chosen such that the pair (A(x). See [6].36) is deﬁned by ˆ Π11 (ψ) ˆ ˆ (4. B(x).2.30) and observer (4.36) corrects the observer model by a nonlinear feedback of the ˆ diﬀerence between the measured output yη and the estimated output yη ..33). (The SDARE state compensator) Assume that the system in SDC f (x) form (4.40) (4. we take into account the eﬀect of the wave frequency (WF) motion in the measurement. The observer ˆ gain Ko (ψ) in (4.38) Ko (ψ) = Π21 (ψ) · V −1 . ξ1 ξ2 + O3 Σ wξ .
and yaw direction. the WF motion component must be removed from the measured output when designing an output feedback controller. Therefore. As mentioned earlier. sway.respectively. ω02 . We use the state estimation technique for wave ﬁltering. The diagonal matrix Σ = diag{σ1 . σ3 } scales the noise vector wb and its elements represents the wave intensity in each degree of freedom. σ2 . The precept of the DP problem is to attenuate only the LF motion.
The response of the WF disturbances is incorporated in the model by assuming that the total ship motion is the sum of the LF motion and the WF motion components.1: A nonlinear compensator for LF model
where the diagonal matrix Ω = diag{ω01 . The complete
. ζ2 .2. Control System Design for the DP Vessel
53
u
xb Ab ( ) xb Bbu Eb wb
yb Cb xb
xb
CbCb
yb
v
ˆ ‐Kc ()
yb v
ˆ xb
ˆ ˆ ˆ xb Ab () xb Bbu ˆ ˆ ˆ Ko () (yb ‐ yb )
ˆ ˆ y b Cb x b
ˆ K ()
ˆ yb ‐ yb
ˆ yb
Cb

Figure 4. For this. ζ3 } is such that each of the ζi ’s represents the relative damping ratio in surge. so that we can add the outputs of both the LF an WF models. ω03 } is such that each of the ω0i ’s represents the dominating wave frequency and the diagonal matrix Z = diag{ζ1 .4. there are various methods for wave ﬁltering. we augment the LF and WF models and assume that the superposition principle holds true.
54 (augmented) model can be written as ˙ η ˙ ν ˙ b ˙ ξ1 ˙ ξ2 =
Chapter 4. SDARE based control Design
O3 J(ψ) O3 O3 −M −1 D M −1 J T (ψ) O3 O3 −T −1 O3 O3 O3 O3 O3 O3 O 3 O3 O O 3 3 w b + Ψ O3 , O O wξ 3 3 O3 Σ
O3 O3 O3 O3 −Ω2
O3 O3 O3 I3 −2ZΩ
η ν b ξ1 ξ2
+
O3 M −1 O3 O3 O3
u
(4.42)
y =yb + yξ + v η ν b ξ1 ξ2 + v.
=
I3
O3
O3
I3
O3
(4.43)
The controllability matrix of (4.42) is given by C= O3 M −1 O3 O3 O3 J(ψ)M −1 −M −1 DM −1 O3 O3 O3 −J(ψ)M −1 DM −1 (M −1 D)2 M −1 O3 O3 O3 · · · −J(ψ)(M −1 D)13 M −1 ··· (M −1 D)14 M −1 ··· O3 ··· O3 ··· O3 , (4.44) Since the matrices J(ψ), M, and D have full row rank for all ψ ∈ R. Therefore, the controllability matrix has rank 6. This means that only η and ν are controllable states while the slowly varying disturbances and wave frequency motion are not controllable. This is not restrictive as the objective is to control only the low frequency motion. Using the Theorem 4.1.1, a locally asymptotically stable SDARE controller for (4.42) is deﬁned by u = −R−1 BT Πc (ψ)x η ν b ξ1 ξ2 ,
= −R−1 M −1
Π21 (ψ)
Π22 (ψ)
Π23 (ψ)
Π24 (ψ)
Π25 (ψ)
(4.45)
where Π2i (ψ), i = 1, 2, 3, 4, 5 are obtained from the SDARE associated with (4.1) and (4.42). Again as we know that only the position measurement is available. Therefore, to realize the controller (4.45), we need estimates of the states. Then this control law
4.2. Control System Design for the DP Vessel will be η ˆ ˆ ν ˆ b ˆ ξ1 ˆ ξ2 .
55
u = −R−1 M −1
ˆ Π21 (ψ)
ˆ Π22 (ψ)
ˆ Π23 (ψ)
ˆ Π24 (ψ)
ˆ Π25 (ψ)
(4.46)
Before addressing the observability of (4.42), we recall the following result from [69]. Theorem 4.2.3. Consider the system x1 = A1 x1 , ˙ x2 = A1 x2 , ˙ y = C1 x1 + C2 x2 , (4.47) (4.48) (4.49)
where x1 ∈ Rn1 and x2 ∈ Rn2 are states and y ∈ R p is the output of the system. If A1 and A2 have no common eigenvalues, i.e., σ(A1 ) ∩ σ(A2 ) = φ, (4.50)
then the system (4.47)(4.49) is observable iﬀ (C1 , A1 ) and (C2 , A2 ) are both observable. Proof. See [69].
The systems (4.23) and (4.40), respectively, describe the slowly varying low frequency and wave frequency motions with diﬀerent eigenfrequencies. Therefore, the system matrices Ab (ψ) and Aξ in (4.23) and (4.40) do not have any common eigenvalues. Moreover, the observability matrices (4.26) and Oξ = I3 O3 . . . O3 I3 . . .
(4.51)
have full row ranks. Therefore, the subsystems (Ab (ψ), Cb ) and (Aξ , Cξ ) of the system (4.42) are observable. This means that by the Theorem 4.2.3, we can conclude that the system (4.42)(4.43) is pointwise observable for all ψ ∈ R.
56
Chapter 4. SDARE based control Design An SDARE observer for the system (4.42)(4.43) is deﬁned by ˙ η O3 ˆ ˆ J(ψ) O3 O3 O3 η ˆ ˙ ν ˆ ˆ O3 −M −1 D M −1 J T (ψ) O3 ˆ O3 ν ˙ ˆ ˆ b −1 b + = O3 O3 −T O3 O3 ˆ ˙ ˆ ξ1 ξ O3 O3 O3 O3 I3 1 2 ˙ ˆ O3 O3 O3 −Ω −2ZΩ ξ2 ˆ2 ξ ˆ + Ko (ψ)(y − y), ˆ η ˆ ˆ ν ˆ b . y = I3 O3 O3 I3 O3 ˆ ˆ ξ1 ˆ ξ2 ˆ The observer gain Ko (ψ) in (4.52) is given by ˆ ˆ Ko (ψ) = Πo (ψ)C T V −1 ˆ ˆ Π11 (ψ) + Π14 (ψ) Π (ψ) + Π (ψ) 21 ˆ 24 ˆ ˆ ˆ = Π31 (ψ) + Π34 (ψ) Π (ψ) + Π (ψ) 41 ˆ 44 ˆ ˆ ˆ Π51 (ψ) + Π54 (ψ)
O3 M −1 O3 O3 O3
u (4.52)
(4.53)
−1 ·V
(4.54)
ˆ ˆ where Πi1 (ψ) and Πi4 (ψ), i = 1, 2, 3, 4, 5 are obtained from the solution of the dual SDARE associated with the (4.42)(4.43). A combination of the nonlinear controller (4.46) and the nonlinear observer (4.52) is called a nonlinear compensator and is shown in Figure 4.2. The nonlinear SDARE compensator is given by ˙ x ˙ x ˆ = A(ψ) ˆ Ko (ψ)C ˆ −BKc (ψ) ˆ ˆ ˆ A(ψ) − BKc (ψ) − Ko (ψ)C x x ˆ + E O O ˆ Ko (ψ) w v . (4.55) Before concluding this section, it is important to mention that in reality, the WF model is considered as a disturbance to the model. For designing a control law, it is desired that the eﬀects of the WF disturbances do not enter into the feedback loop. In this particular example, it is required that no part of the disturbance vectors ξ1 and ξ2 enters in the control action. Here, the controller gain matrix turns out to have the ˆ ˆ ˆ ˆ form, Kc (ψ) = [Kc1 (ψ), Kc2 (ψ), Kc3 (ψ), O3 , O3 ] which ensures the attenuation of the ˆ ˆ disturbance; in the feedback the product term, Kc (ψ) x, eliminates ξ1 and ξ2 from the ˆ feedback loop.
4.3 Simulation Results
Suppose the objective is to design a controller to stabilize the vessel around the origin. By the stabilized mode we mean that the vessel is positioned at rest at (x, y) = (0, 0) and its heading is along the positive x−axis.
4.0124 0.7 and 4. we penalize the velocities and take Q = 103 diag{0.05. Figure 4.18) in (4. it is clear that the actuator input to the vessel is steady and smooth with a smooth overshoot. 1.1. Once more.0358 0 0. 1. 1}. 0. y) = (−10. For instance. We elaborate this feature in the following. 0. 1. As we have already mentioned that the weighting matrices can be tuned to improve the performance of the controller. we use the data of a supply vessel from [47]. −10) making an angle ψ(= 4 rad) with the positive xaxis. Bisscaled [23] normalized parameters of the vessel model are used. Figure 4. 1. 0. D = M= 0 −0.4 shows: (a) the trajectory proﬁle of the vessel. 0 1. It is clear that new proﬁles for the heading orientation and the control input signal are smoother than the previous proﬁles. Figure 4. 0. 0) and heading along the positive xaxis.0744 0. (x.1278 (4. y) = (0. 0} and R = 102 I3 . To emphasize more on tuning factor. and (c) the norm of the state vector. Suppose b = [0. 0. The normalized matrices M and D are: 0 0 0 0 1.3.1274 0. First we consider the case when the disturbance vector b is known. 0. We present the simulation results corresponding to the weighting matrices Q = 103 diag{0.6 shows the controller recommended force in each degree of freedom of the vessel.8902 −0. Simulation Results
57
E w
y C
C
y
u
ˆ ‐Kc ()
xb= Ab ( ) xb Bbu Eb wb
yb Cb xb
xb
yb Cb
y yb y v
v
ˆ x
ˆ ˆ ˆ ˆ x A() x Bu K 0 y y
ˆ y‐y
ˆ K0 ()
+
ˆ ˆ y Cx
ˆ y
C
Figure 4.01]T and we use the control law (4.3 shows the position and heading proﬁles as they evolve with time.0744 .12).1.1.5 shows change in the heading of the vessel along the trajectory.05.0124 .8 show the control input signals and the heading orintation of the vessel corresponding to the tuned weighting matrix Q.56) We assume that the ship is positioned at a point (x. The control objective is to steer the system to the stable equilibrium position. Figures 4. (b) the cost functional proﬁle.0308 0 −0. For simplicity of numerics. we include another heading orientation
. It is clear that the regulation to the desired point is steady and smooth.2: Nonlinear compensator For numerical simulations to illustrate the eﬀectiveness of the SDARE controllers and observers designed in previous subsection. Figure 4.1.1183 −0.
see Figure 4.
(a) 0 x [m]
−5
−10 0 2 4 6 8 10 (b) 0 y [m] 12 14 16 18 20
−5
−10 0 2 4 6 8 10 (c) 5 4 ψ [rad] 3 2 1 0 −1 0 2 4 6 8 10 12 Time [sec] 14 16 18 20 12 14 16 18 20
Figure 4.
.3: The position and heading proﬁles from the closed loop system (4.18) .58
Chapter 4.9. SDARE based control Design
proﬁle corresponding to R = 100 I3 .12) and (4.
cost. Simulation Results
59
(a) 0 y [m]
−5
−10 −12 x 10
5
−10
−8
−6 x [m] (b)
−4
−2
0
2
3 Cost: x Qx+u Ru
T T
2 1 0 0 5 (c) 10 15
15 Norm: x 10 5 0 0 5 Time [sec] 10 15
Figure 4.4: The evolution of the phase.
Heading angle variation 2
0
−2
−4 x [m]
2
−6
−8
−10
−12
−14 −12
−10
−8
−6
−4 x1 [m]
−2
0
2
4
Figure 4.5: The heading orientation over the trajectory of the vessel.4. and norm (of the state vector) proﬁles with time.3.
.
60
Chapter 4.
. SDARE based control Design
(a) 40 20 u1 [kN] 0 −20 −40
0
1
2
3
4
5 (b)
6
7
8
9
10
20 10 u2 [kN] 0 −10 −20 0 1 2 3 4 5 (c) 4 2 u [kNm] 0 −2 −4 −6 0 1 2 3 4 5 Time [sec] 6 7 8 9 10 6 7 8 9 10
Figure 4.7: The heading orientation over the trajectory of the vessel.
3
Heading angle variation 2
0
−2
−4 x2 [m]
−6
−8
−10
−12
−14 −14
−12
−10
−8
−6
−4 x1 [m]
−2
0
2
4
Figure 4.6: The proﬁles of the individual components of the control input vector.
9: The heading orientation over the trajectory of the vessel.
3
Heading angle variation 4
2
0
−2
x2 [m]
−4
−6
−8
−10
−12
−14 −12
−10
−8
−6
−4 x1 [m]
−2
0
2
4
Figure 4.8: The proﬁles of the individual components of the control input vector.5 2 2.5 1 1.5 Time [sec] 3 3. Simulation Results
61
(a) 10 0 u [kN] −10 −20 −30 0 1 2 3 4 5 (b) 5 0 u [kN] −5 −10 0 1 2 3 4 5 (c) 1 0 u [kNm] −1 −2 −3 −4 −5 0 0.5 5 6 7 8 9 10
2
1
6
7
8
9
10
Figure 4.5 4 4.3.4.
.
0. The control input signals are also ampliﬁed as expected.
. 1. 100} while the controller weighting matrices are Q = 10−1 diag{100. 1.15 shows the control input signals from the controller in each degree of freedom. Figure 4. we see that the SDARE compensator states show nice asymptotic convergence to the desired point. and (c) the norm of the state vector and its estimate. 0. We consider exactly the same situation regarding the control objective and the initial position of the vessel. 0} and R = 10−3 diag{1. 10} and V = 10−2 diag{1. sway.16 and 4. 0. Figure 4.11 shows the proﬁles of linear velocity in surge and sway directions and the angular yaw velocity and their estimates. 1. 0. 1.12 shows the proﬁles of the bias estimates in surge. SDARE based control Design
Next. as in the nominal case. For instance. i. 0. 1. and the observer weighting matrices are U = 100 diag{1. (b) the cost functional proﬁle. As desired. 1. The simulation results show the performance of the SDARE based control law and the SDARE observer to achieve the desired objective. Figure 4.. 0} and R = 10−1 diag{1.17 show the corresponding proﬁles of the trajectory and the control input signals. This means that we reduce the penalty on both the vectors x and u. Figure 4. 10. when b is available.14 shows the heading orientation with the trajectory.01}. To emphasize the role of the weighting matrices. 1. the proﬁles show a steady and smooth estimation of the noisy measurements. 1. 1. The proﬁles show that the controller adjusts in the start and then signals become steady and smooth. 100. we take Q = 10−2 diag{100.e. 1. we consider the case when the bias vector b is unknown. 1}. Figure 4. and yaw directions. From all these ﬁgures.10 shows the measured position and heading proﬁles and their estimates obtained from the SDARE observer. 1. we use the bias time constant as T = diag{100. For the simulation results shown in the following ﬁgures. The vessel takes a relatively loger route but its smoother than the one we obtained with the previous weighting matrices.01}. 1. Figures 4. 10. 1. The proﬁle shows a smooth and steady path and orientation of the vessel. 1. we include some more simulations. 0.13 shows: (a) the trajectory proﬁle. 100. 100.62
Chapter 4. Figure 4. 1. 1.
.11: The actual and the estimated surge.3.5 3 Time [sec]
3. Simulation Results
(a) 10 5 x and x [m] ˆ 0 −5 −10 −15 0 5 10 15 (b) 10 y and y [m] ˆ 20 25 x x ˆ 30
63
0
−10 y y ˆ −20 0 5 10 15 (c) 6 ψ ˆ ψ 20 25 30
ˆ ψ and ψ [rad]
4 2 0 −2 0 5 10 15 Time [sec] 20 25
30
Figure 4.5
2
2.10: The proﬁles of the actual and the estimated position and orientation as obtained from (4.
(a) 30 u and u [msec−1 ] ˆ 20 10 0 −10 0 1 2 3 4 5 (b) v and v [msec−1 ] ˆ 10 5 0 −5 −10 0 1 2 3 4 5 (c) 20 r and r [radsec−1 ] ˆ r r ˆ 0 6 7 8 9 10 v v ˆ 6 7 8 9 10 u u ˆ
−20
−40
0
0.5
4
4.5
5
Figure 4. sway.4.39). and yaw velocities as obtained from (4.39).5
1
1.
39).12: The bias estimates in surge. sway. SDARE based control Design
(a) 2 0 ˆ1 [kN] b −2 −4 −6 −8 −10 0 5 (b) 0 ˆ2 [kN] b 10 15
−5
−10 0 5 (c) 6 4 ˆ3 [kNm] b 2 0 −2 0 5 Time [Sec] 10 15 10 15
Figure 4. and yaw directions as obtained from (4.
xy−trajectory 5 0 y [m] −5 −10 −15 −12 −10 −8 −6 −4 −2 x [m] (b) 0 2 4 6
Cost: xT Qx + uT Ru
4000 3000 2000 1000 0 0 1 2 3 4 5 (c) 40 x and ˆ x 30 20 10 0 0 5 10 15 Time [sec] 20 25 30 x ˆ x 6 7 8 9 10
Figure 4.13: The trajectory and the proﬁles of the cost functional and the norms of the measured state and its estimate as obtained from (4.64
Chapter 4.
.39).
.15: The control input signals for each degree of freedom computed from (4.3.6 0.14: The heading orientation along the trajectory.8 2
Figure 4.4 0.8 1 1.4.4 1.6 1. Simulation Results
65
Heading angle variation 4 2 0 −2 −4 x [m] −6 −8 −10 −12 −14 −16 −14
2
−12
−10
−8
−6
−4 x1 [m]
−2
0
2
4
6
Figure 4.
(a) 150 100 u1 [kN] 50 0 −50 0 1 2 3 4 5 (b) 100 u2 [kN] 0 −100 0 1 2 3 4 5 6 7 8 9 10 6 7 8 9 10
50 u3 [kN−m]
0
−50 0 0.2 0.30).2 Time [sec] 1.
5 5 3 3.
(a) 200 u1 [kN] 100 0 −100 0 0.5 1 1.30). SDARE based control Design
Heading angle variation 5
0
−5 x2 [m] −10 −15 −20 −12
−10
−8
−6
−4
−2 x1 [m]
0
2
4
6
Figure 4.5 (b) 200 u2 [kN] 0 −200 0 0.5 5 3 3.5 2 2.5 (c) 40 20 u3 [kN−m] 0 −20 −40 −60 −80 0 0.5 4 4.5 2 2.5 1 1.5 1 1.5 Time [sec] 3 3.5 4 4.5 4 4.5 5
Figure 4.17: The control input signals for each degree of freedom computed from (4.16: The heading orientation along the trajectory.66
Chapter 4.
.5 2 2.
1. 0. The traectory and orientation are nice and smooth.26 and 4.25 show the corresponding trajectory and control input signals. 10. Figures 4. 1.2}. Lets take the matrices Q = 10−2 diag{100. 1} and V = 10−1 diag{1.24 and 4.9}. Further lets modify the observer weighting matrix U: we now take U = 100 diag{10.18 shows the proﬁles of the measured position and heading. 1. 1. The proﬁles depict a smooth and steady regulation of the vessel to the desired position. and (c) the norm of the measured state vector and its estimate. Figures 4. 0. 10. (b) the cost functional proﬁle. In Figure 4. 1.2. For the simulation results shown in the following ﬁgures. 1. Figure 4. 1. 0. the controller weighting matrices as Q = 10−1 diag{100. the dominating wave frequencies as Ω = diag{0. Simulation Results
67
Finally.01} and the observer weighting matrices as U = 100 diag{1. 100}. Figure 4. 1. 1. 1. the control signals need some adjustment in the beginning and then become smooth and steady. and yaw directions. we use the bias time constant as T = diag{100. 10.9.2. 1. 1. 1. sway.01}. 10. 1}. As in the previous case. we show the heading orientation along the trajectory. we consider the case when the WF motion is also playing a role. 10.46). 1.19 shows the velocity proﬁles of the vessel motion and their estimates along surge. Figure 4. It can be seen that the vessel follows a relatively longer path but its slightly better than the previous path. 1. 0. 100. and R = 10−1 diag{1. 1. The control input signals are ampliﬁed with this choice of the tuning matrices.4. 1. O1×6 } and R = 10−3 diag{1. 1. 1. O1×6 }. 1. 1. 1.27 show the corresponding proﬁles of the trajectory and the control input signals.3. 1. we add some more simulation results. The trajectory proﬁle has further improved and control input proﬁles show an increased control input. 10. 0.9.20 shows the bias estimates obtained from the SDARE observer. the relative damping ratios as Z = diag{0. 100. Figure 4. 1. 1.23 shows the control input signal in each degree of freedom computed from (4. and their estimates obtained from the SDARE observer. 1. To further elaborate the importance of the weighting matrices. 1. 1}. 1. 1.
. 0. 1.22. Now. 1. 100. Figure 4.21 shows: (a) the trajectory of the vessel motion.
18: The position and heading proﬁles of the vessel motion as obtained from the compensator (4. sway.5
2
2.5 3 Time [sec]
3.55).55).68
Chapter 4.
(a) u and u [msec−1 ] ˆ 20 10 0 −10 u u ˆ
0
1
2
3
4
5 (b)
6
7
8
9
10
v and v [msec−1 ] ˆ
10 5 0 −5 −10 0 1 2 3 4 5 (c) 20 6 7 8 9
v v ˆ
10
r and r [radsec−1 ] ˆ
r r ˆ 0
−20
−40
0
0.5
1
1. and yaw directions as obtained from (4.5
4
4. SDARE based control Design
(a) 10 x and x [m] ˆ
0
−10 x x ˆ −20 0 5 10 15 (b) 10 20 25 30
y and y [m] ˆ
0
−10 y y ˆ −20 0 5 10 15 (c) 5 20 25 30
ˆ ψ and ψ [rad]
0 ψ ˆ ψ −5 0 5 10 15 Time [sec] 20 25 30
Figure 4.5
5
Figure 4.19: The velocities and their estimates along surge.
.
3.55). and yaw directions as obtained from (4.21: The trajectory and proﬁle of the cost functional and the norms of the measured state vector and its estimate corresponding to the closed loop system (4. sway. Simulation Results
(a) 2 0 ˆ1 [kN] b −2 −4 −6 −8 −10 0 5 (b) 0 ˆ2 [kN] b 10 15
69
−5
−10 0 5 (c) 6 4 ˆ3 [kNm] b 2 0 −2 0 5 Time [Sec] 10 15 10 15
Figure 4.
.20: The bias estimates in surge.55).
xy−trajectory 5 0 y [m] −5 −10 −15 −12
−10
−8
−6
−4 x [m] (b)
−2
0
2
4
Cost: xT Qx + uT Ru
4000 3000 2000 1000 0 0 1 2 3 4 5 (c) 40 x and ˆ x 30 20 10 0 0 5 10 15 Time [sec] 20 25 30 x ˆ x 6 7 8 9 10
Figure 4.4.
70
Chapter 4.8
1 1.55).23: The control input signals for each degree of freedom as obtained from (4.
(a) 150 100 u1 [kN] 50 0 −50 0 1 2 3 4 5 (b) 100 u2 [kN] 0 −100 0 1 2 3 4 5 6 7 8 9 10 6 7 8 9 10
50 u3 [kN−m]
0
−50
−100
0
0.6
1.4
0.4
1. SDARE based control Design
Heading angle variation 4
2
0
−2
−4 x2 [m] −6 −8 −10 −12 −14 −14
−12
−10
−8
−6
−4 x1 [m]
−2
0
2
4
6
Figure 4.8
2
Figure 4.6
0.
.22: The heading orientation along the trajectory.2
0.2 Time [sec]
1.
5 2 2.25: The control input signals for each degree of freedom as obtained from (4.2 Time [sec]
1.4
1.
.5 2 2.3.2
0.6
0.
(a) 300 200 u1 [kN] 100 0 −100 0 0.8
2
Figure 4.24: The heading orientation along the trajectory.6
1.4
0.55).5 1 1.5 3 3.5 4 4.5 4 4.5 5
200 u2 [kN] 0 −200 0 0. Simulation Results
71
Heading angle variation 4
2
0
−2
−4 x [m]
2
−6
−8
−10
−12
−14 −12
−10
−8
−6
−4
−2 x [m]
1
0
2
4
6
Figure 4.8
1 1.5 (b) 3 3.4.5 1 1.5 5
u3 [kN−m]
0
−50
−100
0
0.
2
0.
(a) 300 u1 [kN] 200 100 0 −100 0 0.8
1 1.5 5
200 u2 [kN] 0 −200 −400
0
0.2 Time [sec]
1.4
0.8
2
Figure 4. SDARE based control Design
Heading angle variation 2
0
−2
−4 x2 [m]
−6
−8
−10
−12
−14 −12
−10
−8
−6
−4
−2 x1 [m]
0
2
4
6
Figure 4.5
2
2.5
1
1.5 2 2.5
5
50 u3 [kN−m] 0 −50 −100 −150
0
0.6
1.6
0.5 (b) 3 3.5 4 4.72
Chapter 4.4
1.5
3
3.55).
.5
4
4.27: The control input signals for each degree of freedom as obtained from (4.5 1 1.26: The heading orientation along the trajectory.
First. we take into account both the LF and WF motions and use again the SDARE method to address the regulation and estimation problems.4. We use the data of a supply vessel for simulation experiments. we assume that the bias vector is unknown and the measured signals from the sensors are contaminated by noise. The DP vessel model shows that only the position and heading measurements are available from the sensors. Simulation results presented in this chapter conﬁrm the asymptotic stability of the SDARE regulator and the SDARE estimator.4 Conclusions
We study the regulation problem of a DP vessel by using the SDARE technique. A locally asymptotically stable compensator is obtained by combining locally asymptotically stable controller and estimator. The combination of the locally asymptotically stable SDARE regulator and SDARE observer is a locally asymptotically stable SDARE compensator. Second. we propose a suboptimal locally asymptotically stable SDARE observer. Afterwards.4. We begin by considering a nominal case (the bias disturbance vector is known or its estimate is available) and come up with a PDtype control law with a feedforward term. velocity. we study the LF motion regulation without taking into account the WF motions. The controller is proved to be globally asymptotically stable by using the LMIs and the Lyapunov stability theory. Using the stabilizability and detectability of the vessel model together with the duality concept. We extend the concept of the LQR problem to nonlinear system by using the SDC parametrization. We divide the study of the DP vessel regulation problem in two steps. and bias vectors) to come up with a locally asymptotically stable SDARE compensator. This gives rise to an SDARE which must be solved to obtain a suboptimal solution of the nonlinear regulation problem. We combine the concept of the SDARE regulation and estimation (to estimate the position. This is an inadequate information for the SDARE control design problem which requires complete state knowledge. Conclusions
73
4.
.
74
.
In a broader perspective. H∞ −control. Both type of methods have positive and negative features. the controller and observer gains are state dependent. the generalized eigenvector approach [13]. and the Schur decomposition method [45]. Some
1 This work has been published in the proceedings of the UKACC International Conference on Control (CONTROL 2010). The earliest traces of the use of the algebraic Riccati equation in control related problems can be found in the works on the solution of the LQ optimal control and ﬁltering problem [38] and on approximation methods in optimal control [62]. emphasis has been put on developing computationally eﬃcient solution methods for the SDARE. Consequently. To date its scope has been extended to address problems of nonlinear state regulation and estimation. [12]. and nonlinear ﬁlter design.
75
. and analysis. estimator based feedback control synthesis. Some of the wellknown direct methods are the Taylor series method [6].Chapter
5
1
The Fourier Series Interpolation Method
In Chapter 4. we divide the solution methods in two categories: iterative and direct. Mainly. using the particular state dependence of the system matrix. the direct methods require more storage capacity than the iterative methods. H2 −control. [48]. for instance see [6]. The SDARE has emerged in many practical applications. the direct methods are computationally faster than the iterative methods. ﬁltering. we propose a Fourier series based interpolation method. the SDARE has become an important research topic in ongoing developments in the areas of control and estimation theory. It has been regarded as an eﬃcient tool to tackle problems related with the nonlinear control design. and [61]. the matrix sign function method [15]. so they are required to be obtained by online solving the concerned SDAREs. In the process. On the other hand. [29]. see [52]. synthesis. Especially. This method reduces the online computation time of the solution of the SDARE by performing the major computational task oﬄine. especially when the problem is poorly conditioned or when a goodquality initial guess is not available. the information ﬁlter algorithm [36].
A(x1 ) ∈ Rn×n is C ∞ and θ−periodic. 37]. where f (x−) and f (x+) means. controller gain is computed online. The matrix A(x1 ) is smooth and θ−periodic. Corollary 5. The following corollary gives the necessary and suﬃcient condition for the continuity of the periodic extension of a piecewise smooth function. At this point. in the feedback loop.1 The Fourier Series Interpolation (FSI) Method
First we explain why it is reasonable to use the Fourier series for the solution of the algebraic Riccati equations of the type (4.1. This method of solution is computationally very simple and hence. Theorem 5. (5. if f (x) is continuous at x. As pointed out in [10] the reduction in the online computation time of the solution of the SDARE is a major issue because of its high computational cost. The solution matrix Π(x1 ). we give the Fourier series based numerical method for the solution of the SDARE of type (4.7) and then we will illustrate the details of the FSI method. the left and the right hand limits of the function f at point x. is the ﬁrst component of the state vector x ∈ Rn . solution via spectral factorization.1). Therefore. For a 2L−periodic piecewise smooth function f (x).. 1 [ f (x+) + f (x−)]. For a general formulation of the FSI method. First.e. we call this method the Fourier series interpolation (FSI) method. to the average of the two limits. the periodic extension of f (x) is continuous for all x ∈ R if and only if f (x) is continuous and f (−L) = f (L).7) which reduces the online computations. Π(x1 ) ∈ Rn×n .1. 2. We consider the following form Π(x1 )A(x1 ) + AT (x1 )Π(x1 ) + Q − Π(x1 )BR−1 BT Π(x1 ) = 0. the Newtonmethod and its modiﬁed forms [8. we recall two important results from [33] which endorse the idea of using the Fourier series for ﬁnding an approximate solution of the SDARE (5. respectively. to f (x). doubling algorithm [11. (The Fourier Convergence Theorem) Let f (x) be a function which is piecewise smooth on the interval −L ≤ x ≤ L.7). 2 if f (x) has a jump discontinuity at x. and Chandrasekhar algorithm [67]. it is valuable in the situations where online computation of the solution of the SDARE is required.1. provided it exists. we recall the Fourier convergence theorem due to Dirichlet. The computation of the controller gain requires the solution of the SDARE (4.1. the interpolation methods [6]. and B ∈ Rn×m . In (4. without loss of generality. 16.
. The periodic extension of a 2L−periodic function is always admissible. is also smooth and θ−periodic. A(x1 ) = A(x1 + θ).5). The Fourier Series Interpolation Method
of the wellknown iterative methods are the Kleinman algorithm [2].
5. 41]. see [14]. Henceforth. then the Fourier series of f (x) converges 1.1)
where x1 ∈ R. In this chapter. we make a slight change of notation in the SDARE (4.76
Chapter 5. i.6). we observe that the controller gain is not ﬁxed but it varies with the state of the system.
Then. such that the process of the computation of the Fourier series coeﬃcients stops when Π(x1 ) − ( a0 + 2
N
(ak cos
k=1
2kπx1 2kπx1 + bk sin ))2 < θ θ
. it follows that the Fourier coeﬃcients with suﬃciently high indices are negligibly small (their norms are of order zero) and can be left out. θ θ
(5. We deﬁne a tolerance level say > 0. 3. there exists an integer N and such that an 2 < and bn 2 < . ak ∈ Rn×n and bk ∈ Rn×n for k = 1..
(5. Thus we write Π(x1 ) = a0 + 2
∞
(ak cos
k=1
2kπx1 2kπx1 + bk sin ). for any small positive number .
and
k→∞
θ 2
lim
θ −2
f (x) cos kx dx = 0.
θ θ Lemma 5.. are deﬁned by ak = and bk = 2 θ
θ 2 θ −2
2 θ
θ 2 θ −2
Π(x1 ) cos
2kπx1 dx1 . 2 ). 2. 2.1.. 2.1. 2 ] then
θ 2
k→∞
lim
θ −2
f (x) sin kx dx = 0. it is possible to represent the matrix function Π(x1 ) in the form of a Fourier series. 3. 2 ). we decide a stopping criterion for the computation of the Fourier coeﬃcients. ∀n > N . 3. This means that the Fourier series of Π(x1 ) exists and converges to Π(x1 ).5)
θ θ is satisﬁed for all x1 ∈ [− 2 .1.
From the fact that the series on the right hand side of (5. the solution Πc (ψ) of the SDARE (4. (The RiemannLebesgue Lemma) If f (x) is integrable on [− 2 .7) is a smooth matrix function with period θ. 1.5. k = 1. Take m > 2N +1
. Therefore.2)
The Fourier coeﬃcients a0 .3)
Π(x1 ) sin
2kπx1 dx1 . First. We explain the procedure to compute the Fourier coeﬃcients as follows.. k = 0.1.. by the above results.. . we start with N = 1 such that there are 2N +1(= 3) Fourier coeﬃcients to be θ θ determined. We use the spectral matrix norm in this chapter. θ
(5. Therefore... .4)
Let us now recall a wellknown result by Riemann and Lebesgue which is stated in the following lemma. θ
(5..2) is convergent and from the statement of the Lemma 5. Equation (5. The Fourier Series Interpolation (FSI) Method
77
In the context of the vessel model.1. .2) holds identically true for all x1 ∈ [− 2 .
···
cos cos
. Otherwise. .
2Nπx11 θ 2Nπx12 θ
sin sin
.5) is tested. (5.5). This can be done by plotting − g(x1 ) θ θ vs x1 on the interval [− 2 .
where Πi = Π(x1i ).9) (5. the Fourier series approximation of the solution of the SDARE (5. . .
1 2
cos cos
2πx11 θ 2πx12 θ
··· ··· .6)
In this way. . If the stopping criterion is satisﬁed (plot lies above x1 axis) then we stop. . The algorithm for ﬁnding the Fourier coeﬃcients is summarized
.. .7)
However.1) is Π(x1 ) ≈ a0 + 2
N
(ak cos
k=1
2kπx1 2kπx1 + bk sin ). . where g(x1 ) is the LHS of (5.8)
This solution yields 2N + 1 Fourier coeﬃcients. We can write in short form as AX = B (5. This results in (A ⊗ In )X = B.. note that the dimensions of the matrices in (5. . we get the following nonhomogeneous system of m linear matrix equations
a0 a1 . 2 ). . there corresponds an equation of the following form
a0 + 2
N
(ak cos
k=1
2kπx1t 2kπx1t + bk sin ) = Π(x1t ). aN b1 . ···
sin sin
. we repeat the above procedure for the next integer value of N (=2) and so on. t = 1. .7) do not conform with the usual matrix multiplication rules. .
cos
2πx1m θ
cos
2Nπx1m θ
sin
2πx1m θ
sin
2Nπx1m θ
. . Πm
1 2 1 2 . . bN Π1 Π 2 = . θ θ
(5. . . The equation (5. For this we multiply the matrix A from right by In in the Kronecker sense.78
Chapter 5. The Fourier Series Interpolation Method
θ θ θ θ points x1t ∈ [− 2 . m. . 2.10)
Now the stopping criterion (5. To each point x1t ∈ [− 2 .
2Nπx11 θ 2Nπx12 θ
. . We use the Kronecker matrix product to remove this discrepancy. . 2 ). Therefore.8) can be solved by using the MATLAB to get X = (A ⊗ In )\B. . 2 ).
2πx11 θ 2πx12 θ
··· ··· . and continue until the stopping criterion is satisﬁed. .. . θ θ
(5.
39) and (4.8) has a solution because rank(A ⊗ In ) = rank(A ⊗ In . ..55) are solved numerically.
. . .
θ θ step 2: Take xti ∈ [− 2 . (4. b1 . Πm . a1 . else continue the f or loop until tolerance condition is satis f ied. The system (5. .39) and (4. Π2 .8).. a[N]. b2 .55) describe the ˆ feedback dynamics of the SDARE nonlinear compensator. ... i = 1.
θ θ Remark 5. for N ← 1 to some f inite large natural number step 1: Take m > 2N + 1. aN . This demands a computationally eﬃcient method for ﬁnding the solution of the SDARE. a2 . a[2]. The main step of the algorithm is step 6 of the for loop.2. We conclude this section with the following remark about the above algorithm.5.. 2 ).1.
5.20). We take the number m of distict points in [− 2 . b[N] then are obtained..2 Performance Analysis
In this section we explain why the FSI method is eﬀectively a better method for online computation of the solution of the SDARE than “care” in terms of computation time.6) to get (5. b[2].
2 The
author is thankful to Professor Donald Kreher for his help on the pseudocode environment. Solve the linear system (5..20). (4.. 2 )..1... 2.. b[1]...1: C F C (FCs) input: The tolerance parameter and an integer N(= 1). bN . greater than the number of Fourier coeﬃcients 2N +1 to be determined to generate an overdetermined system of linear equations. Algorithm 5.7). 3.1) at m interpolation points to get Π1 .8) to get the desired Fourier coeﬃcients. assumption: The m number of interpolation points must satisfy m > 2N + 1.m. We do so on the basis of ﬂop count...
79
step 3: Solve (5. step 4: Use statement (i) of this algorithm and (5.
(i)
if the tolerance condition is satis f ied Fourier Coe f f icients a[1]. step 5: step 6: Use Kronecker matrix product to get a linear system of equations (5. Performance Analysis below2 . .1. Equations (4. The presence of Kc (ψ) and ˆ Ko (ψ) therein indicates that two SDAREs have to be solved at each iteration when the ode systems (4. B).. output: The Fourier Coeﬃcients a0 . The repetition of the loop continues until the stopping criteria is satisﬁed..
and n eigenvalues with positive real parts. In this case the CARE (5. Theorem 5. The scalar diagonal entries correspond to real eigenvalues.2. First step is the reduction of the Hamiltonian matrix associated with the SDARE to an ordered real Schur form (RSF).11) is given by H= A −Q −BR−1 BT −AT . . .
(5.13) are not ordered so an additional step is made to get an ordered RSF which is important for the next step which uses the invariant subspace of the stable eigenvalues of H to solve (5. (The Real Schur Triangularization Theorem) Let H be a 2n × 2n Hamiltonian matrix. Then the Hamiltonian matrix H has n eigenvalues with negative real parts. (5. MATLAB routine “care” is called multiple times to solve (5.1. H can be transformed into RSF form according to the following result.2.12) does not have any eigenvalue on the imaginary axis. The eigenvalues in (5. For each ψ. The matrix Rrs is known as the RSF of H.80
Chapter 5. The MATLAB routine “care” is based on this method.13)
where each Rii is either a scalar or a 2 × 2 matrix. .11). Theorem 5. . During the process of computation of the numerical solution of the closed loop system. no eigenvalues on the imaginary axis. (5.12)
The following result ensures that the Hamiltonian matrix (5.7) reduces to a continuous time algebraic Riccati equation (CARE) ΠA + AT Π + Q − ΠBR−1 BT Π = 0. (Relationship between the Hamiltonian Matrix and the Riccati Equations) Let (A.B) be stabilizable and (A. Now using the QRfactorization. 0 R12 R22 . 0 R1k R2k .11)
We solve the closed loop system equations by using the MATLAB solver ode45. 0 ··· ··· .
. . The Schur decomposition method requires two major computations for solving the SDARE.Q) be detectable.11) has a unique stabilizing solution Π. In the following we explain some details of the Schur decomposition method for a better understanding of the computational complexity of this method. Rkk .2. . Then there exists a 2n × 2n matrix Qrs such that R11 0 T Qrs H Qrs = Rrs = . The following deﬁnition and subsequent theorem and corollary summarize the concluding step of the Schur decomposition method. and each 2×2 matrix on the diagonal has a pair of complex conjugate eigenvalues. . .11). the SDARE (4. The Fourier Series Interpolation Method
We compare the FSI method with the Schur decomposition method [13]. The Hamiltonian matrix associated with (5.
Performance Analysis Deﬁnition 5. we quantitatively demonstrate that the FSI method reduces the online computation time for the solution of the SDARE.
.11) and Π1 is invertible.2.2. 30]. (Basis of an Invariant Subspace from RSF) Let QT H Qrs = Rrs = rs R11 0 R12 R22 . by using the simulation results discussed in Section 4. the FSI method requires simple matrix addition (O(n2 ) ﬂops) and scalarmatrix multiplication (O(n2 ) ﬂops). then Π = Π2 Π−1 is a solution of the CARE (5. Then the ﬁrst p columns of Qrs .1. 1 Corollary 5. Theorem 5. First. On the other hands.1. The simulations have been performed on an Intel Centrino duo running at the frequency of 2 GHz having 2 GB RAM. we conclude that the FSI method reduces the online computations in a feedback dynamics based on an SDARE control law. where p is the order of R11 .3. the computation of Π = Π2 Π−1 is done by using the leastsquares algorithm and this also requires 1 O(n3 ) ﬂops.1. Second. form an orthonormal basis for the invariant subspace associated with the stable eigenvalues of H. The ordering of the eigenvalues in RSF requires O(n) ﬂops.1. Table 5. Each of the closed loop systems (4.14)
and let us assume that eigenvalues of R11 and R22 respectively are the negative and positive eigenvalues of the Hamiltonian matrix H. However.11) is computed by this method. we see that the computation of the Fourier coeﬃcients involve two major steps. In the second step.5. the computation of the solution of the SDARE at a known (m) number of points. (4.11) if and only if the columns I of span an ndimensional invariant subspace of the Hamiltonian matrix H. by the empirical observations it is assumed that on average each eigenvalue requires 2 QRfactorizations.3. A matrix Π is a solution of the CARE (5. If the columns of The Schur method is based on reduction of the Hamiltonian matrix H to RSF which is done by QR factorization. Based on the foregoing discussion. Each requires O(n3 ) ﬂops. the solution of a linear system of equations.1 shows a comparison between the computer time required to solve the closed loop systems with the FSI method and by the MATLAB routine “care”.20). From the algorithm given in Section 5. In Table 5. the FSI method requires O(n2 ) ﬂops when the online solution of (5.2. [13].
81
(5.11) is solved online.39) and (4.2. The QRfactorization of a matrix requires O(n3 ) ﬂops which means that the RSF obtained in the ﬁrst step of the Schur decomposition method also requires O(n3 ) ﬂops. the performance of the FSI method is still comparable to its performance with a lower order system. Therefore. Therefore the Schur decomposition requires O(n3 ) ﬂops each time CARE (5.55) are solved with the MATLAB routine ode45 on a time interval [0. The computationally expensive task in FSI method is the computation of the Fourier coeﬃcients which is done oﬄine.11). only once. Π Π1 span an ndimensional invariant subspace Π2 of the Hamiltonian matrix H associated with the CARE (5. We notice that with the increase in the size of the system. The exact ﬂop count is not possible as QRfactorization is an iterative procedure.
and 17948 times when (4.15 seconds.1: Performance Analysis Online comput. The Fourier Series Interpolation Method
For (4. 30]. we notice that “care” is called 2164.
. and 0. Online comput.55) is solved on interval [0.8 37 118 48 59 96 67 30 274 192 32 191 90 53 467 252 46 495 276 44
Looking at the proﬁle summary obtained from the MATLAB.55)
Table 5.1. 34547. The times for the oﬄine computations of the Fourier series coeﬃcients are 0.39) is solved on interval [0. 0. 30] and it is called 12406. respectively. and 34405 times when (4. 2479.20)
For (4.39) For (4. Percentage (%) reduc.13. This clearly means that the FSI method which requires only simple arithmetic for online computation of the solution of the CARE is computationally much faster than “care” which uses the Schur decomposition method to solve the Riccati equation. It is called 12352 and 34406 times when (4. 30] for ˆ ˆ each Kc (ψ) and Ko (ψ).82
Chapter 5. time (sec) time (sec) in online with “care” with FSI method computation time 11 6 45 11 6.20) is solved on interval [0.
and magnetic levitation systems [65]. power converters [64].e. the Interconnection and Damping Assignment Passivity Based Control (IDAPBC) methodology has emerged as an easy and a (quasi) stepbystep methodology to obtain passivity based controllers. for instance in underwater vehicle [3]. underactuated mechanical systems [59]. can easily be studied for the closed loop systems obtained. This chapter introduces the portHamiltonian model of a vessel with 3 degrees of freedom in a horizontal plane. The PortHamiltonian framework is a very useful concept which oﬀers more capacity to improve the performance of the controllers and to analyze stability properties.
83
. the design of power system stabilizers [56].. which play a fundamental role in analyzing stability and the performance of the controller. The obtained portHamiltonian equations are the
1 This work has been published in the proceedings of the Dynamic Positioning Conference DP2010 in Houston. The classical Euler Lagrange (ELPBC) allows the change in the potential energy function and the generalized inertia matrix. It is also accepted for publication at the IET Journal of Control Theory and Applications. see [60]. Stability properties. Passivitybased techniques were ﬁrst introduced in [58] and have been used since then in many applications.Chapter
6
1
PortHamiltonian Formulation and Passivity Based Control Design
T
is chapter focuses on the design of the control law for dynamic positioning assuming that the ﬁltering and estimation problems have been solved. The passivity based approach has two versions. IDAPBC also allows changes in the interconnection matrix. i. see for instance [60]. the control of electrical machines [19]. such as storage energy or dissipation. based on the Lyapunov theory. see [51]. see [50]. In the last decade. A nice feature of the passivitybased control design is the physical meaning of the resulting control laws and concepts. the measured output contains only the noise free LF position and heading measurements and the disturbances vector is either known or its estimate is available.
The socalled passive output. the socalled passive output is a relative degree one output. see [79]. J(x) ∈ Rn×n is the skewsymmetric interconnection matrix (J = −J T ). The design problem summarizes into ﬁnding a
2 The ∂ (or ∂.1. and the port matrix g describes the ﬂow of energy from outside (for instance from the controller) to the system. in the portHamiltonian framework. The resulting control laws are of outputfeedback form and are robust in presence of eventual disturbances. The main contribution of this chapter is a family of passivitybased controllers. ˙ (6.1. R(x) ∈ Rn×n is the dissipation (or damping) matrix (symmetric positive semideﬁnite. A PHS can be written. and in what follows x we will take it as a column vector. and is thus suitable for control schemes based on IDAPBC. (6. [60]. in an implicit form. as x = (J − R)∂H(x) + g(x)u. y ∈ Rm .2
The IDAPBC Technique
The Interconnection and Damping Assignment Passivity Based Control (IDAPBC).84
Chapter 6. u ∈ Rm is the control input vector. PHS generalize the Hamiltonian formalism of classical mechanics to physical systems connected in a powerpreserving way and encodes the detailed energy transfer and storage in the systems.
.
6. the controller can be considered as a source injecting energy into the system. pH Formulation and PBC Design
starting point for the design of controllers using the IDAPBC technique. the matrix R describes the energy losses of the system. which uses the energy shaping of the closed loop system to ensure (local/global) asymptotic stability. is given by y = gT (x)∂H(x).2)
and the product uT y usually has unity of power. and H(x) is the Hamiltonian (energy or storage) function2 . In the literature on portHamiltonian systems. It uses the passive stability properties to ensure the convergence of the system to the desired ﬁxed point. R = RT ≥ 0).
6.1 HamiltonianBased Control
6. g(x) ∈ Rn×m is the external connection matrix.1)
where x ∈ Rn is the state (or Hamiltonian variables) vector. The main idea behind the IDAPBC is to deﬁne a new closed loop (or target) system with a Hamiltonian structure. is a technique for designing controllers based on the portHamiltonian framework.1 PortHamiltonian Modeling
A large class of physical systems of interest in control applications can be modeled in the general form of portHamiltonian systems (PHS). In this formulation. the interconnection matrix J describes the ﬂow of energy inside the system. In a passivity based control design approach. if no confusion arises) operator deﬁnes the gradient of a function of x.
and a change of variables z = f (x. see [18]. evaluated in the ze dynamics. 0). [57] and [60] for a detailed discussion). Rz > 0 and Ro − Rzo R−1 RT ≥ 0.
85
(6. Jzo ∈ Rr×(n−r) and Je ∈ Rr×r . Extension of the closed loop dynamics in the IDAPBC framework can be done. x = arg min(Hd (x)). The design procedure reduces to ﬁnding matrices Jd (x) and Rd (x) and a desired closed loop energy function Hd (x). Ro ≥ 0.8)
ensures that xd is an equilibrium point. Then. the Hamiltonian function.7)
where Hd has a minimum at the desired regulation point (xd .4)
(6. z zo The key point of the PHS structure in (6..5)
A drawback of the IDAPBC controllers is that they are. z Finally. the existence of a minimum of Hde in xd . see for instance. Rd (x) = Rd (x) ≥ 0 and Hd (x) has a minimum at the desired d d regulation point x . see [57]. i. in a natural way. the stability is guaranteed if the dissipative matrix is positive semideﬁnite. o o
T ˙ ze = −Je ∂ xo Hd
(6. 0). o
.
(6. A completely diﬀerent approach addresses nonpassive outputs (or higher relative degree one outputs). only for passive outputs. which solve the socalled matching equation
(J − R)∂H + gu = (Jd − Rd )∂Hd . f (xd . usually a dynamic extension of the system is introduced to obtain an integral action on the output error.e. z) + 1 T z Ke ze . Hde takes the form Hde = Hd (xo . Rz = RT ∈ R(n−r)×(n−r) . 2 e (6. not able to reject disturbances. o o Following the idea of the IDAPBC technique. a Hamiltonian based controller with an integral action can be obtained via a change of variables.6) is that. T Dissipation is given by Ro = Ro ∈ Rr×r .6. implies ∂ xo Hd  xd = 0 which.6) zo zo T −Je O3 O3 ze ˙ The powerpreserving interconnection structure of the proposed target system is deT T ﬁned by Jo = −Jo ∈ Rr×r .3)
T T where Jd (x) = −Jd (x). To remove this discrepancy of the control design. The stability of this system can easily be ˙ proved by using Hd as a Lyapunov function (Hd (x) = −(∂Hd )T Rd ∂Hd ≤ 0.1. which is used to enforce the equilibrium point of the closed loop system to the desired one. Let us assume that the xo ∈ Rr are the higher relative degree one (or nonpassive) outputs. in general. HamiltonianBased Control control law such that the system behaves as ˙ x = (Jd − Rd )∂Hd . Jz = −Jz ∈ R(n−r)×(n−r) . ze ) ∈ Rn−r to cast the target system in a Hamiltonian structure as follows Jzo − Rzo Je x o Jo − R o ˙ z = −J T − RT ˙ Jz − Rz O3 ∂Hde . In this case. the control law becomes u = (gT g)−1 gT ((Jd − Rd )∂Hd − (J − R)∂H). (6. and Rzo ∈ Rr×(n−r) . The main idea is to introduce a new variable ze ∈ Rr .
Furthermore.2 Dynamic Positioning Problem
A dynamic positioning (DP) system is a computer controlled system which automatically maintains a vessel’s position and heading by using propellers and thrusters. also. This problem is studied in many papers. However. with two diﬀerent frequency components (see Figure 2. pH Formulation and PBC Design
6. only position and heading measurements are available. which represents the eﬀect of the waves. a passivitybased scheme considered in [27] and [47]. combined with the sensor information. helps the computer calculate the required steering angle and thruster output for each thruster. Kalman ﬁlters require the use of a linear model. and the Luenberger observer used in [71].
Figure 6. see Figure 6. the motion due to the waves should be removed before it enters in the controller algorithm. This leads to the use of observers to estimate the state (mainly the velocities and the bias term) which are required for feedback into the control law. sea currents and thruster forces and moments) and an oscillatory term (the socalled wave induced wave frequency motion). As pointed out in [27]. DP only considers the slow variations and. a wavefrequency observer is added to compensate the wave disturbances. consequently. To overcome this drawback. The computer program contains a mathematical model of the vessel which includes information pertaining to the wind and current drag of the vessel and the location of the thrusters. Some examples include a nonlinear observer designed in [24]. PI controllers are often used [22]. Various controllers have been proposed to stabilize the ship to the desired position.1: Basic scheme of components of a ship motioncontrol system. the measured position and heading signals are noisy and.5. and the controller. This knowledge. Kalman ﬁltering techniques were proposed in [28] and [32]. Due to the important role of the estimation and ﬁltering process. the motion control system in the DP problem can be grouped in two basic subsystems: the observer system (or wave ﬁlter). in DP problems. and the nonlinear motion should be linearized at various operation points. See [25] for a recent overview. Chapter 2).86
Chapter 6. See [22] for more details.1. however more advanced techniques are applied to
. Generally. see [27] and [47]. The total ship motion can be seen as a superposition of a low frequency component (due to the wind.
Ship Model in PortHamiltonian Framework
87
this problem resulting in interesting control algorithms. are either known or an estimate of the bias vector is available. and must be able to reject unknown disturbances or uncertainties.
6.2
Ship Model in PortHamiltonian Coordinates
We can write the system described in (6. waves. under assumption of constant or slowly varying currents. Hence. The main goal in the dynamic positioning problem is to stabilize the ship in a given ηcoordinate. page:89]. our objective is to design an appropriate control law τ which stabilizes the system to the origin (x. Recently. Additionally. 0. In this chapter.9) in a PHS form (6. Without loss of generality. 0).9). (6.3. is deﬁned as p = Mν. The environmental disturbances due to the sea currents. pT ] ∈ R6 .1 Ship Model in Cartesian Coordinates
We recall here the vessel model from Chapter 2. Chapter 3. In this chapter. ˙ η ˙ ν = O3 O3 J(ψ) −M −1 D η ν + O3 M −1 τ+ O3 M −1 J T (ψ) b. In [47] the stability of a PDtype controller with a passive observer is proved using a separation principle argument. we assume that the estimation and ﬁltering problem has been solved and our focus is the control design for dynamic positioning using the portHamiltonian framework. in this chapter.1) by using as a state xT = [qT . This bias term is constant in the Earthﬁxed reference frame. which also includes the observer stage. are presented in [24] and [69].9) are given in Chapter 2.10)
.6. [22.
6. where q = [q1 q2 q3 ]T ∈ R3 represents the Earthﬁxed position and heading.3 Ship Model in PortHamiltonian Framework
6. ψ) = (0. as the measurement of the relative velocity vector is not available. which can also be modeled as a ﬁrstorder Markov process. the control law should be independent of ν. y. it is assumed that these natural eﬀects (sometimes called bias forces and moments). we skip the dynamics of the bias and the WF components.3. we consider a model in the following form. and the momentum p = [p1 p2 p3 ]T ∈ R3 . (6. Backstepping design. and wind are represented by b = [b1 b2 b3 ]T ∈ R3 in the Earthﬁxed reference frame. we get the following system ˙ x = (J(q3 ) − R)∂H + gτ τ + g b (q3 )b with the following interconnection and damping matrices J(q3 ) = O3 −J T (q3 ) J(q3 ) O3 . R= O3 O3 O3 D .11) (6. Substituting η = q and ν = M −1 p in (6.9)
The details of the vectors and the matrices in (6. As mentioned at the start of this chapter. We assume that the measurement system gives us noise free position and orientation measurements and that the wave frequency (WF) components from the measured output are ﬁltered or estimated.3. sampleddata control theory has also been applied to the DP problem for designing control law [39].
13) 2 Note that the Hamiltonian function contains only a kinetic energy term.10) with the target dynamics (6. (6. and it results in a control law that can be interpreted as a nonlinear version of the conventional PID controller.10). the control laws are obtained from matching the dynamical system (6. a nominal case is considered (i. a dynamic extension is proposed. Hd (x). i. improves the heading control.e. In the design process. From (6. A potential energy. H=
6.e. These controllers do not produce the desired regulation properties in the presence of unknown disturbances. p) = Ψ(q) + 1 T −1 p M p.3). Two diﬀerent energy functions are proposed in this chapter.4 Classical IDAPBC Design
A family of output feedback controllers can be obtained via the IDAPBC methodology. associated with the momentum variable.. pH Formulation and PBC Design
g b (q3 ) =
O3 J T (q3 )
. (6. we observe that the passive output for the system (6. To solve this. We start by illustrating the methodology using a quadratic and a trigonometric Hamiltonian function and recover a simple classical IDAPBC controller which guarantees asymptotic stability. Jd = the dissipation matrix set as Rd = O3 O3 O3 Rp .5. As presented in Section 6. 2 (6. the desired interconnection matrix is ﬁxed as in (6. they also ensure a good regulation behavior even in presence of disturbances or unknown (or nonestimated) terms. in order to achieve the desired performance. especially for the dynamic IDAPBC control design in Section 6.16)
.10) is the velocity vector which does not correspond to the actual output of the system. The energy shaping. where the disturbances are assumed to be completely known).15) O3 −J T (q3 ) J(q3 ) O3 . A salient feature of the proposed controllers is that they do not require the relative velocity measures and.. artiﬁcially added by the controller will play a key role to stabilize the ship in the desired position. thanks to a dynamic extension. Consequently.1.14)
where R p ∈ R3×3 is a symmetric positive deﬁnite matrix (R p = RT > 0). the position and the orientation. This is an important consideration for the control design. (6.
(6. is shaped as Hd (q.2).12)
and the Hamiltonian function given by 1 T −1 p M p. based on a trigonometric function.
Chapter 6. and the closed p loop energy function.88 the external connection matrices gτ = O3 I3 . and then the stability against an unknown disturbance vector is also analyzed.
0) is locally asymptotically stable.17) takes the form ˙ τ = − J T (q3 )∂ q Ψ − KD J T (q3 ) q − J T (q3 )b. and deﬁning KD := R p − D. Using the Hamiltonian function (6. Furthermore. in the q coordinates. i.15) as follows. Consider the dynamical system (6.20)
(M −1 p)T
= − pT (M −1 )T R p M −1 p ≤ 0. we get a PHS in the form (6. and that the bias vector b and the matrix D are known.18)
which can be seen as a nonlinear output feedback PD (ProportionalDerivative) controller with a feedforward term. Proof.18).10). (6. as we point out in the simulations that by increasing the dissipation of the system. with respect to the original Hamiltonian function (6.10). −J T (q3 )b.14) and (6.4. we keep the general result because.4. the controller adds some potential energy. Proposition 6. p∗ ) = (0. (6.
. ∂ q Ψ q∗ =0 = 0 and ∂2 Ψ q∗ = 0 > 0. we get p = MJ T (q3 ) q. 0).e. Notice that with the choice R p = D. from the second equality. Then. we obtain the following control law τ = −J T (q3 )(∂ q Ψ + b) − (R p − D)M −1 p. the equality corresponding to the ﬁrst row is automatically satisﬁed while. where the origin q = 0 is a local minimum of Ψ(q). then (q∗ . the performance of the closed loop system improves considerably . x= ˙ O3 −J T (q3 ) O3 −J T (q3 ) = = J(q3 ) −D b+ O3 M −1 p O3 J T (q3 ) b + O3 I3 −J T (q3 ) −KD J T (q3 ) ∂Hd +
J(q3 )M −1 p −J T (q3 )∂ q Hd − R p∂ p Hd O3 −J T (q3 ) J(q3 ) −R p ∂ q Hd ∂ p Hd (6.17) ˙ From the q dynamics in (6. Substituting (6. if q = 0 is the global minimum of Ψ(q).3) with (6.4).19)
=(Jd − Rd )∂Hd . From the resulting matching equation (6.18) in (6.16) as a Lyapunov function..13). the controller simpliﬁes because KD = 0. Classical IDAPBC Design
89
where Ψ(q) has a minimum at the origin. p∗ ) = (0. This q implies that the desired energy function has the minimum at the desired stabilizing point (q∗ . p∗ ) = (0.10) in a closed loop with the control law (6. From a physical point of view. the desired regulation point (q∗ . 0) is globally asymptotically stable.1. we get ˙ Hd =(∂Hd )T x ˙ =(∂Hd )T = (∂ q Ψ)T O3 −J T (q3 ) J(q3 ) −R p
T
∂Hd O3 O3 O3 −R p ∂q Ψ M −1 p (6. the state feedback algorithm (6. But.6.
from (6.10) with (6. ˙ The set of points at which Hd vanishes is El ={(q.18). we explain how this connection can be exploited to improve the performance of the controllers in this study. 2 2 (6. It is easy to see that q = 0 2 is the only minimum of Ψ(q) and the desired energy function (6. we can conclude that the closed loop system (6. p∗ ) = (0.
6.24). we propose to shape the desired Hamiltonian containing a trigonometric function with the form Ψ(q) = 1 T q C12 q12 + c3 (1 − cos q3 ). is globally asymptotically stable.19). p) = 1 1 T q K q + pT M −1 p. of the control law depending upon two diﬀerent energy shapings. (6. we have a rather general expression for the control law depending on Ψ(q).18).4. In case. q∗ = 0 is only a local minima of Ψ(q) then the set Ml may consists of multiple points. pH Formulation and PBC Design
˙ Since Hd is negative semideﬁnite.4. However. the following control law ˙ τ = −J T (q3 )(K q + b) − KD J T (q3 ) q. equilibrium point theorems do not establish asymptotic stability of the closed loop system (6.25)
. 2 12 (6.90
Chapter 6. invoking LaSalle’s invariance principle. (6.22) (6. 0). The role of the energy function in this study is similar to that of a Lyapunov function. a quadratic and a trigonometric. Hence the closed loop system (6.21)
In case.1. This means that the largest invariant subset Ml of El is a singleton set.19) are given by Ml ={(q∗ . p) : p = 0 and q is free}. In (6. we know that the stability properties of a dynamical system and the minima of the Lyapunov function have a close connection.24)
Using Proposition 6. What follows are two special cases.16) becomes Hd1 (q. The closed loop system then converges locally asymptotically to the set Ml . In the sequel. where K = K T > 0 is a gain matrix.1
A Quadratic Energy Shaping
The simplest function with a global minimum at the origin has a quadratic form: Ψ(q) = 1 qT K q.2
A Trigonometric Energy Shaping
Inspired by the energy function of a pendulum. From the Lyapunov stability theory. q∗ = 0 is the global minimum of the function Ψ(q) then the set Ml consists of only the origin (q∗ . p∗ ) : p∗ = 0 and q∗ is such that ∂ q Ψ(q) q∗ = 0}.4. The invariant points of the closed loop system (6.
6.19) converges globally asymptotically to the desired point.23)
which implies. asymptotic stability can be proved.
p) = 1 T 1 q12C12 q12 + c3 (1 − cos q3 ) + pT M −1 p. where a ∈ Z. c3 } is a positive deﬁnite gain matrix. has multiple local minimums at q = (0. c2 .27). The main motivation for this kind of energy shaping is that.14) and (6.4.1. Classical IDAPBC Design
91
where q12 = [q1 q2 ]T .26)
Figure 6. is (locally) asymptotically stable. c2 } and c1 .18) takes the ﬁnal form as τ = −J T (q3 )C q1 q2 sin q3 − KD J T (q3 ) q − J T (q3 )b.15). implies that the control law (6.q3)
Hd1
Hd2
q3
q1
Figure 6. C12 = diag{c1 .4. for certain applications where there are no constraints (for instance. As in the previous subsection. a2π). respectively. c3 > 0. we can conclude that the closed loop system (6. using Proposition 6.23) and (6. 2 2 (6. ˙
(6. The Hamiltonian (6.26) with the target closed loop system deﬁned by (6.16) becomes Hd2 (q. c2 .2: Comparison between the quadratic and the trigonometric functions. links with external objects). stabilization in q3 = 0 or q3 = 2π is exactly the same. Hd1 and Hd2 .10) with (6. in the coordinates q1 and q3 . 0. This function.
Hd(q1. shows a possible scenario. where the path for stabilizing in q3 = 2π is shorter than stabilizing in q3 = 0.6.27)
where C = diag{c1 .2 illustrates the diﬀerence between the two proposed energy shapings (6.
.3.26) in terms of the level surfaces. Figure 6. With this choice the desired energy function (6.
92
Chapter 6.
.5 Extended IDAPBC Design
6. In order to analyze the performance of the proposed passivitybased controller.3: A possible scenario when it is advantageous to use the trigonometric energy function instead of the quadratic energy function. As a ﬁrst approximation. The system (6. b.29)
which has an equilibrium point which satisﬁes ∂Ψ q=q∗ = J(q∗ )∆ and p∗ = 0. It 3 implies that in presence of disturbances.18) results in the following system ˙ q ˙ p = O3 −J T (q3 ) J(q3 ) −R p ∂Hd + O3 I3 ∆. In practical situations. 0). However. we study the following perturbed system ˙ q ˙ p = O3 −J T (q3 ) J(q3 ) −D ∂H + O3 I3 τ+ O3 J T (q3 ) b+ O3 I3 ∆. wrong estimations or.1 Motivating Problem
The control law (6. has to be estimated. the closed loop system (6.28) with control law (6. in the general case. (6. and only to motivate the use of the extended dynamics. (6. Let us study the inﬂuence of unmodeled behaviors. the ship model only represents some simple dynamics of the actual system and furthermore the bias vector.
6.18) has a diﬀerent equilibria than (0.10) perfectly models the ship motion. the presence of disturbances.5. we consider this vector as a constant.28)
where ∆ ∈ R3 is a vector which represents the disturbances. stability should also be further analyzed. pH Formulation and PBC Design
y
yr
Quadratic path
Trigonometic path
xr
Figure 6.18) assumes that b is known and that (6.28) in a closed loop with (6.
24).32). ˜ We seek another Hamiltonian function (energy shaping) Hd2 (q. Let us deﬁne 1 ∆1 2 1 ∆2 2 1 ˜ ˜ Hd2 (q. global asymptotic stability can be proved by shifting the Hamiltonian function to the new equilibrium point.5. The new set of equilibrium points of sin q3 q∗ = ∆1 ∆2 ∆3 .35) ˜ ∂ p Hd2 (q. in a special case where the disturbances are also aﬀected by J T (q3 ) given by ˙ q ˙ p = O3 −J T (q3 ) J(q3 ) −D ∂H + O3 I3 τ+ O3 J T (q3 ) b+ O3 J T (q3 ) ∆. 2 c1 2 c2 2 ˜ where ψ3 (q3 ) is such that ˜ ∂q3 ψ3 q∗ = 0.e.31)
as a Lyapunov function. (6. c3 (6. we require ˜ ∂ q Hd2 (q.26) to the new equilibrium (6. .. (6. p) p∗ = 0. Unlike the quadratic case.26). 3 ⇒ Integrating both sides. See Appendix A for more details. i. and p∗ = 0. arcsin − 2kπ c1 c2 c3
T
where ∂ q Hd2 (6.33) is not a set of minima of (6. Similarly.33)
The new set of equilibrium points (6. ˜ ψ3 (q3 ) = − cos q3 − ∆3 q3 + e1 . 0).e. gives the following closed loop system ˙ q O3 O3 J(q3 ) ∂ q Hd2 + = ∆.32) ˙ p ∂ p Hd2 J T (q3 ) −J T (q3 ) −R p q1 = C q2 and ∂ p Hd2 = M −1 p. p)(q∗ . p) q∗ = 0. where q∗ = K −1 ∆.
(6.27) derived by using the trigonometric energy shaping in (6.6.32) is
. Therefore. k ∈ Z. c3 ˜ ∂q3 ψ3 = sin q3 − ∆3 . and ˜ ∂2 Hd2 (q.30).34)
. i.. the control law (6. p) = c1 (q1 − ) + c2 (q2 − ) + c3 ψ3 (q3 ) + pT M −1 p. With the control law (6. (6. shifting of (6.26). Extended IDAPBC Design
93
Moreover.33) does not make them a set of minima of the shifted Hamiltonian function. p) which has a minimum value at (q∗ . This is because of the presence of transcendental term in (6. p∗ ) > 0. we require another Hamiltonian function for the stability analysis of (6.30)
it can be seen that some stability properties still remain. with 1 1 ˜ Hd1 = (q − q∗ )T K(q − q∗ ) + pT M −1 p 2 2 (6.
∂2 Hd2 (q∗ . local asymptotic stability of this new set of equilibria can be proved with the Hamiltonian function ∆3 1 T −1 1 ˜ Hd2 (q. c3 1−( ∆3 2 ) > 0. p) = (q12 − q∗ )T C12 (q12 − q∗ )+c3 1 − cos q3 − q3 + p M p. p∗ ) is positive deﬁnite ˜ iﬀ ∂2 Hd2 (q∗ . p∗ ) > 0. O3 M −1
˜ Using the Schur complement argument. p) = c1 q1 − 2 c1
2
(6. the new desired energy function is 1 ∆1 ˜ Hd2 (q. ˜ ˜ ∂2 Hd2 ∂2 p Hd2 q q 2 ˜ 2 ˜ ∂ pq Hd2 ∂ p Hd2 ˜ ∂2 Hd2 O3 q .38) 12 12 2 c3 2 which has local minima if ∆33 < 1.24) and (6.4 explains this condition for existence of the set of minima for the energy function (6. Although a high gain in the K and C matrices (in controllers (6. c3 ∆3 < 1.37)
˜ The Hessian of Hd2 (q. p) is given by ˜ ∂2 Hd2 = = where ˜ ∂2 Hd2 q c1 = 0 0 0 c2 0 0 0 c3 cos q3 . This discrepancy in the desired performance of the control law motivates the use of a dynamic extension in order to achieve the stabilization at the desired point.
. c3 2 (6. pH Formulation and PBC Design
where e1 is a constant of integration. This may not be a rational strategy due to practical limitations of the propulsion units. the Hessian.94
Chapter 6. then ∆3 ˜ ψ3 (q3 ) = 1 − cos q3 − q3 . More details of this result have been explained in c Appendix A. Without loss of generality. 3 ∆3 c3 cos(sin−1 ) > 0. Figure 6. Summarizing. (6.27). respectively) implies stabilization close to the desired equilibrium point. c3 Therefore. c3
⇒ ⇒ ⇒
Using the LaSalle’s principle. This requires q c3 cos q∗ > 0. the presence of unknown disturbances results in a bad positioning of the ship. we take e1 = 1.38).36)
1 ∆2 + c2 q2 − 2 c2
2
+ c3 1 − cos q3 −
∆3 1 q3 + pT M −1 p.
and then the presence of unknown terms is analyzed. is possible by means of a change of coordinates. A dynamic extension for a non passive output. as O3 J(q3 ) J(q3 ) ∂ q Hde q ˙ z = −J T (q3 ) −Rz O3 ∂ z Hde . ze ) = Ψ(q) + 1 T −1 1 z M z + zT Ke ze 2 2 e (6. ze ∈ R3 . which is used to enforce the equilibrium point of the closed loop system to the desired one.38). ze ). it has been proved that if a system cannot be stabilized by using the classical IDAPBC approach then it is impossible to stabilize the system by dynamic extension.6. our system is stabilizable but the discrepancy is that it does not stabilize to the desired equilibrium point. and a change of variables z = f (q. with the form of (6. Following the idea in [18].5.5. z. is a 3 × 3 matrix to be deﬁned. We deﬁne the target system. Extended IDAPBC Design
95
Figure 6. without disturbances.4: Visualization of the condition for existence of the set of minima of (6.
6. p. (6. The desired Hamiltonian function z is deﬁned as Hde (q. we introduce a new state variable. This result is not restrictive in our case. The controller is designed for a nominal case. we introduce an integral action by dynamic extension.40)
.6).2
Target Extended System
In [4].39) ˙ T ∂ ze Hde −J (q3 ) O3 O3 ze ˙ where Rz = RT ≥ 0. maintaining the portHamiltonian structure. To overcome this drawback.
47) is globally asymptotically stable.48) Then. z∗ ) = (q∗ .41). e Matching the q dynamics. What follows are two special cases depending on two diﬀerent formulations of the energy shaping (6.1.
. ke2 . the stability can be proved invoking LaSalle’s invariance principle as we already show in proof of Proposition 6. (6. the Hamiltonian function (6.45) (6. q = 0. Furthermore. is (locally) asymptotically stable at the point (q. the change of variables z is deﬁned as z = p − MKe ze . Ψ(q) must be designed with a minimum in the desired regulation point.46) = = −KP J T (q3 )∂ q Ψ − KD M −1 p + KI ze − J T (q3 )b −J (q3 )∂ q Ψ
T
(6.96
Chapter 6.43)
Similarly to the classical IDAPBC controller in the previous section. the control law (6. from (6.39).47). Proposition 6.39). then the origin of (6. Proof. Equation (6.47) takes the form of (6.5. Then. 0).41)
The state feedback control law is obtained from the second row of (6. Rz − D. (6. Then. then the system (6.47)
which has the same structure as a nonlinear PID controller with a feedforward term. −J T (q3 )b. ke3 } > 0. (6. Ke . 0. 0. τ ˙ ze where we deﬁne KP KD KI := := := (MKe + I3 ). the Hamiltonian function (6.40) is a Lyapunovcandidate function and its time derivative is ˙ Hde = −zT (M −1 )T Rz M −1 z ≤ 0.1. pH Formulation and PBC Design
where Ke = diag{ke1 . ke2 .44) (6. and Ψ(q) has a (local) minimum at the origin. z∗ . and that the disturbances vector b and the matrices M and D are known.42)(6.39) and the time derivative of (6.40) has a minimum at (q∗ .47) is a rather general formulation of the control law. 0). z. ke3 } > 0.42) (6. As earlier in the classical IDAPBC design.43) takes the following form τ = −KP J T (q3 )∂ q Ψ − KI ˙ J T (q3 )∂ q Ψdt − KD J T (q3 ) q − J T (q3 )b (6.4.10) and (6. if q = 0 is a global minimum of Ψ(q). If Ke = diag{ke1 .10) in a closed loop with (6. The closed loop system (6. Assume that q is measurable. using p = ˙ MJ T (q3 ) q.10) with (6. ze ) = (0.40).
51)
From Proposition 6. the Hamiltonian function (6.40) becomes 1 T 1 1 q12C12 q12 + c3 (1 − cos q3 ) + zT M −1 z + zT Ke ze .3
A Quadratic Energy Shaping
Let us ﬁrst take the same quadratic energy shaping as in Section 6. Extended IDAPBC Design
97
6. Taking the perturbed system (6.50)(6. this function has multiple local minimums at q∗ = (0.5.28) with the designed control law (6.53) 2 2 2 e As previously mentioned.54) sin q3 q1 ˙ ze = −J T (q3 )C q2 (6.1 the controller (6.55) sin q3 Hde2 (q. we consider constant (or.4
A Trigonometric Energy Shaping
The trigonometric energy shaping proposed in the previous section can also be considered for the extended controller.43) becomes Hde1 (q. As before.40) becomes 1 T 1 1 q K q + zT M −1 z + zT Ke ze . Taking again Ψ(q) = 1 T q C12 q12 + c3 (1 − cos q3 ). 0.49) 2 2 2 e with a global minimum at the origin. Using Proposition 6. k2 .50) (6. a ∈ Z. (6. z. very slow) disturbances. c2 .42)(6.6.5.5. c3 > 0 the desired Hamiltonian function (6. Ψ(q) = 1 qT K q 2 with K = diag{k1 . 0. 2 12 (6. ze ) = τ ˙ ze = = ˙ −KP J T (q3 )K q + KI ze − KD J T (q3 ) q − J T (q3 )b −J T (q3 )K q.52)
where q12 = [q1 q2 ]T and C12 = diag{c1 .5
Analysis in Presence of Disturbances
The main contribution of the extended dynamics excels in the presence of unknown disturbances.
6. c3 }. at least. and c1 . (6. (6. a2π). The control law (6. we obtain O3 J(q3 ) J(q3 ) ∂ q Ψ O3 q ˙ −1 T ˙ z = −J (q3 ) −Rz O3 M z + I3 ∆.43) becomes q1 ˙ τ = −KP J T (q3 )C q2 + KI ze − KD J T (q3 ) q − J T (q3 )b (6.
6.42)(6. (6.1. let us analyze the closed loop system in the presence of disturbances.56) O3 Ke ze −J T (q3 ) O3 O3 ze ˙
.51) ensures the global asymptotic stability. k3 } > 0. c2 }.5. (0. c2 .4. 0).5. z. First. and the control law (6.5.47). this case ensures local asymptotic stability. ze ) = where C = diag{c1 .
. contrary to the classical controller.59)
z∗ e
Notice that.57)(6. z
(6.98
Chapter 6. z.. we get 1 1 ˜ ˜ ˜ ˜ ˜ ˜ ˜ Hde (q.1278
0 0. The asymptotic stability is automatically guaranteed. The (Bisscaled nondimensional [22]) matrices M and D are given by 1.56) can be written as O3 q ˙ ˙ z = −J T (q3 ) ˜ ˙ −J T (q3 ) ze ˜ (6.0744
0 −0.
(6.
that for the case of the trigonometric controller.40) to the new equilibria.6 Simulations
In order to test the performance of the designed controllers we performed some numerical simulations.e. i.1274 0 M= 0 and 0. in the presence of unknown disturbances.0124 0.0358 0 D= 0
3 Notice
0 1.6). ∆ = ∆(t). a2π). we can no longer study the stability of an equilibrium point.0308
. Shifting e the energy function (6.60)
J(q3 ) −Rz O3
J(q3 ) O3 O3
∂q Ψ −1 M z ˜ ˜ Ke ze
. 0.57) (6. For this validation we used the data of a supply ship from [24]. the equilibria become q∗ = (0. pH Formulation and PBC Design
From the ze dynamics.0744 0. In a more realistic case.
6.58) (6.0124
0 −0. that can be seen as a ˜ forced oscillator with damping in the z coordinates. A simple stability ˜ ˜ analysis is possible using the error coordinates z = z − z∗ and ze = ze − z∗ .59).1183 −0.8902 −0. with a energy function which has a minimum at (6.61)
This closedloop system still has the desired structure (6. 2 2 e and (6. ze ) = Ψ(q) + zT M −1 z + zT Ke ze .56). Then. and we only can expect that the solution of the system becomes bounded due to the special structure of the closed loop system. this system has the following equilibrium point3 q∗ z
∗
= 0 = = MR−1 ∆ z −Ke R−1 ∆. the origin is still the equilibrium in the position coordinates. and due to ∂ q Ψ q∗ = 0. the disturbance could be considered as a vector which depends on time. The question that now arises is about the stability of (6.
01} and R p = diag{0.05. for the classical IDAPBC (quadratic controller) design with extra damping (solid line) and the original damping coeﬃcient (dashed line).6. of the quadratic and the trigonometric versions. 0.1
Simulation Results for the classical IDAPBC design
In this subsection. even though the controllers for the q1 and q2 coordinates are the same (with the same gain values).75. Both controllers stabilize the ship at the desired position and angle. respectively. 0.75. For all simulations. 0. the starting heading angle is set greater than π to show the ability of the socalled trigonometric controller to stabilize to the closer minimum. The gain matrices we used are K = C = diag{0. the performance is similar but. the trajectories take diﬀerent paths.24) and (6. The desired heading direction is along the positive xaxis. we enlarged the damping (about one order of magnitude in the ﬁrst and third components). setting R p to the values proposed before. q2 plane. In both cases. Figure 6. we considered that the initial conditions of the ship are (q1 . and keeping the original damping.05.6. 0.05.
6. More generally. This fact is associated with the diﬀerent heading angle trajectories. to improve the performance. in this case 2π.5: Simulation results: ship position trajectories in the q1 . the trigonometric controller does so at q3 = 2π. Figure 6. and the desired stabilization position is the origin.5 we show the trajectories of the same classical controller (the quadratic case) with extra dissipation. This comparison justiﬁes the use of the extra damping to obtain more suitable paths. 0. we present the simulation results for the classical (quadratic and trigonometric) control laws (6. and the heading angle.6. Precisely.
6 4 2 0 −2 q2[m] −4 −6 −8 −10 −12 −14 −12
−10
−8
−6
−4
−2 q1[m]
0
2
4
6
8
Figure 6. q3 .6 shows the trajectories of the position coordinates q1 and q2 .
. R p = D.27). In Figure 6. The notable point is the diﬀerence in the orientation proﬁles.7 compares the trajectories in the q1 q2 plane for two versions of the classical IDAPBC design. While the quadratic controller stabilizes the heading angle at q3 = 0. and the heading angle q3 = 4 rad.1}. q2 ) = (−10.05.01]T . 0. −10). Simulations
99
The bias vector has been set to b = [0. for this system.
.
Heading angle orientation 2
0
−2
−4 y [m]
−6
−8
−10
−12
−14
−12
−10
−8
−6
−4 x [m]
−2
0
2
4
Figure 6. pH Formulation and PBC Design
Quadratic
0 q1[m] −5 Trigonometric −10 0 20 40 60 80 100 120 140 160 180 200 5 Trigonometric 0 q2[m] −5 Quadratic −10 0 20 40 60 80 100 120 140 160 180 200 8 6 q3[rad] 4 Quadratic 2 0 0 20 40 60 80 100 time[s] 120 140 160 180 200 Trigonometric
Figure 6. and heading angle. q1 and q2 .100
5
Chapter 6.7: Simulation results: ship position trajectories in the q1 q2 plane. for the quadratic (solid line) and trigonometric (dashed line) controllers of the classical IDAPBC design together with the respective heading orientation. for the quadratic (solid line) and trigonometric (dashed line) controllers of the classical IDAPBC design. q3 .6: Simulation results: ship position coordinates.
is not available. is removed in all the tested controllers.05. 0. Extended Quad.01. we present the simulation results in the presence of disturbances. The corresponding gain matrices used for the controllers from extended IDAPBC design.27).6. 0.05.8: Simulation results: ship position coordinates. Clearly the controllers from IDAPBC design steer the ship to the desired equilibrium position while the controllers from the classical IDAPBC design fails to do so and have a steady state error.4}. are Rz = diag{0.015}.55). are the same as in the previous subsection. In Figure 6. See Figure 6. for the both the quadratic and trigonometric versions of the controllers from the classical and extended IDAPBC designs.
Classical controllers 0 q1 [m] Extended controllers −5
−10
0
20
40
60
80
100
120
140
160
180
200
Classical controllers 0 q2 [m] Extended controllers −5
−10
0
20
40
60
80
100
120
140
160
180
200
8 Classical Trig .6. 140 160 180 200
Figure 6. This methodology is based on the
. K = C = diag{0. and heading angle. Consequently. Classical Quad.75.01. 0. The key point is to show that the controllers from the extended IDAPBC design approach are able to reject unknown terms.7. (6.
6.8.9. q3 .7 Conclusions
A passivitybased approach called IDAPBC is used to obtain a set of output feedback controllers for the dynamic positioning of a ship. 0. For this scenario we assume that the disturbance due to the bias term. This diﬀerence between the performance of both the controllers can also be seen from the respective trajectory proﬁles. q1 and q2 .75. 6 q3 [rad] 4 2 0 0 20 40 60 80 100 time[s] 120 Extended Trig . J T (q3 )b. (6. 0. Conclusions
101
6.025} and Ke = diag{0.51) and (6. q trajectories for the four controllers are plotted. 0.2
Simulation Results in Presence of Disturbances
In this subsection.24) and (6. the feedforward term. The gain matrices for the classical controllers. b.
portHamiltonian description which gives a physical interpretation of the dynamical systems. q2 plane. the controller design problem is addressed as to shape the energy function of the closed loop system. Under this point of view.
−10
−8
−6
−4 q1[m]
−2
0
2
4
Figure 6. this work could be a starting point for a new design. to improve the performance. It is worth mentioning that the obtained control laws. also in presence of disturbances. inspired by the physics of a pendulum. An important observation is that the trigonometric energy shaping improves the heading angle control.102
Chapter 6. which physically increases the damping eﬀect.
. in the controller formulation helps in getting a smoother trajectory proﬁle. Simulations are done to validate and compare the performance of the controllers designed. do not stabilize the system at the desired position. −8 −10 −12 −12 Classical Quad. can be easily converted to output feedback algorithms that only requires the position measurement. Also. using the portHamiltonian perspective. 2 0 Classical Trig. After a general formulation we propose two diﬀerent controllers: ﬁrst with a quadratic energy function and second. −2 q2[m] −4 −6 Extended Quad. Also. Future work can be aimed at determining other energy functions. with a general form of state feedback. for the both the quadratic and trigonometric versions of the controllers from the classical and extended IDAPBC designs. pH Formulation and PBC Design
4 Extended Trig. with a trigonometric energy function. as well as to consider the optimization of the resulting path. Ψ(q). the presence of disturbances is studied and it turns out that the control laws obtained by using the classical IDAPBC methodology. We also observe that by keeping the diﬀerential term.9: Simulation results: ship position trajectories in the q1 . Furthermore. of the complete motioncontrol system (controller and observer) for a DP problem. they exhibit a simple structure that can be interpreted as nonlinear versions of PID controllers. This discrepancy is the starting point for a second set of controllers which consist of a dynamic extension of the system which provides stability at the desired regulation point. Further analysis depending on the nature of the disturbance vector (including the wave frequency and wind models) is required.
pointwise Hurwitz.e. Further research has proved that the extended set of conditions that the system matrix is continuous. From the perspective of the DP system we are interested in the asymptotic behavior of the system. From the ﬁrst counter example discussed in Section 3. The study is based on the SDC framework which transforms the nonlinear system in a pseudolinear form with a state dependent system matrix. The periodicity assumption also ensures that the states of the system are bounded. the system matrix is a periodic function of the heading angle of the vessel.1. The local asymptotic stability of the system is easy to establish.4. they will not blow up at ﬁnite time. we can conclude that the conditions that the system matrix is continuous. pointwise Hurwitz.Chapter
7
Conclusions and Recommendations for Future Work
In this chapter. exponentially bounded. and exponentially bounded are not suﬃcient to ascertain the global asymptotic stability of the pseudolinear system. the major concern always remains the analysis of the global asymptotic stability. The periodicity property of the vessel model has lured us to combine the Lyapunov stability theory with the LMIs to come up with a new approach to prove the global 103
. and periodic does not constitute the set of suﬃcient conditions for the global asymptotic stability of the pseudolinear system. This study concludes that more research is required to explore the set of sufﬁcient conditions which establish the global asymptotic stability of the pseudolinear systems with a system matrix depending on a single state variable. This fact is explained in the second counter example presented in Section 3. In the prototype vessel model. i. The ﬁrst major topic studied in this thesis is the stability analysis of a special type of nonlinear system in which nonlinearity is because of a single state variable. The stability analysis of such systems is based on the properties of the system matrix.5.1.. we summarize the main conclusions of this work and present some recommendations for future work based on our knowledge and experiences acquired during the process of reaching to this point.
to obtain the desired performance. So it is required to estimate the remaining states of the vessel model.104
Chapter 7. The method was motivated by the periodicity of the pseudolinear vessel model. It has been shown that it makes the online computation of the solution of the SDARE faster by reducing the online computations. We present the FSI method which is a fast solution method for the speciﬁc type of the SDARE associated with a pseudolinear system whose system matrix depends on a single state variable in a periodic way. It addresses the performance of the controller and the observer by specifying a performance index. the FSI method has shown good results in comparison with the MATLAB routine “care” which uses the Schur decomposition to solve the SDARE. Further research is required to investigate the additional features like for instance. etc. in particular. An extension of the SDARE technique to address the path following and trajectory tracking problems of the DP vessel remains still an open research topic. The emphasis of the FSI method is on oﬄine computation of the Fourier series coeﬃcients. The vessel system in this work is a slow system (with large time constant). The simulation results show promising performance of the SDARE technique when applied to address the regulation problem of a DP vessel. The design objective is to stabilize the vessel at the desired equilibrium point. The SDARE technique has many characteristic features. This approach has proved successful to establish the global asymptotic stability of the vessel model. For such systems. This study is limited to the regulation problem of the DP vessel. The Hamiltonian (storage) function is a powerful component in this formu
. For instance. the impact of the number of LMIs on the solutions of the LMI feasibility problem. the actuators are required to have faster dynamic response than the system they are controlling. It is not the case for the vessel model in this work as only the noisy position and heading measurements are available through sensors (GPS and gyro). Generally. conditions for existence and nonexistence of the solutions of the LMI feasibility problem. The SDARE technique gives a state feedback controller which means that the complete state of the system should be known. The method is particularly useful for the systems having small time constant or fast dynamic response. The search for more applications of the FSI method on physical systems is still open for future research. In this work. Conclusions and Recommendations for Future Work
asymptotic stability of the pseudolinear system which has a periodic system matrix. an increase in the state weighting matrix Q results in a faster regulation of the vessel at the expense of a greater control eﬀort. the use of the portHamiltonian system formulation to study the control design problem of the DP vessel has revealed some important and interesting features. This work is limited to just one application of the FSI method on the vessel model. the SDC parametrization is not unique and this feature oﬀers an extra degree of freedom for design and performance improvements. We present the SDARE technique for control design of the vessel model. The controller and the observer gains can be tuned by using the weighting matrices. We have also used the SDARE technique to address the estimation problem. This method is based on the Fourier series and involves the concept of interpolation. the best solutions of the LMI feasibility problem. it maintains the nonlinear character of the system which is not the case in linearization methods like the extended Kalman ﬁlter (EKF) etc. The state and the control weights can be adjusted to inﬂuence the performance. While the technique oﬀers ways to use the results from linear system theory.
and state estimation is a challenging and interesting topic for future research. trajectory tracking. With the encouraging results obtained. The dissipation (damping) matrix in the desired dynamics in the IDAPBC design technique is very important from the perspective of the tuning of the controller gain matrices. This can be done by using diﬀerent formulations of the Hamiltonian function to design diﬀerent controllers. It makes it possible to manipulate the damping of the system.
. A better heading control of the vessel is obtained from the trigonometric controller than the quadratic controller.105 lation. Through the controller action. We have seen that it not only simpliﬁes the stability analysis but also oﬀers more options to improve the performance of the controller. This fact is clear from the trigonometric and quadratic formulations of of the Hamiltonian function. This study addresses only the regulation problem with passivity based approach using the portHamiltonian formulation. the possible extension of this concept to address more involved problems like path following. we can increase the damping of the system and it eventually helps us to steer the vessel along a stable and smooth trajectory. The performance of the controllers can then be analyzed to see which one suits the desired performance objective.
106
.
32).
(A. 2 2 (A.2) = O3 −J T (q3 ) J(q3 ) −R p ∂ q Hd2 ∂ p Hd2 + O3 J T (q3 ) ∆. Let us consider the following closed loop system. (A. and p∗ = 0.1) is given by q∗ =
∆1 c1 ∆2 c2
arcsin
∆3 c3
− 2kπ
T
.3)
To prove the above mentioned asymptotic stability. p) = qT C12 q12 + c3 (1 − cos q3 − ˜ 12 ˜ q3 ) + pT M −1 p.1)
The equilibrium point of (A. For convenience of the reader. k ∈ Z. ˙ q ˙ p where Hd2 (q. and ˜ (A. The energy function (A.5) is a scalar valued function and its ﬁrst partial 107
. (A. we proceed as follows. we use the LaSalle’s invariance principle to prove asymptotic stability of (A. we recall both equations mentioned above. ˙ q ˜ ˙ p ˜ q1 − q∗ 1 where q = q2 − q∗ ˜ 2 q3 = O3 −J T (q3 ) J(q3 ) −R p ˜ ∂ q Hd2 ˜ ˜ ∂ p Hd2 ˜ .4)
.33) is an asymptotically stable equilibrium point of the perturbed closed loop system (6.4). ˜ ˜ 2 c3 2
As a ﬁrst step.32) at (6. p = p.Appendix
A
Proof of asymptotic stability of (6.33)
The objective is to show that (6.5)
1 ∆3 1 ˜ ˜ ˜ Hd2 ( q. p) = 1 1 T q12C12 q12 + c3 (1 − cos q3 ) + pT M −1 p.
˜ ˙ ˜ ˜ p = − J T (q3 )∂ q Hd2 − R p ∂ p Hd2 ˜ ˜ ˜ = − J T (q3 )∂ q Hd2 − R p ∂ p Hd2 + J T (q3 )∆ = p.10)
w.8)
The largest invariant subset Ml of El consists of the equilibrium points (A. arcsin(
∆3 ) − 2kπ c3
T
. (A. The asymptotic stability in this case will be global because the point (q∗ . q2 .4) then (A.108 derivatives exist.3) is an asymptotically stable equilibrium point of (A.r.8) is an asymptotically stable equilibrium point of (A. Diﬀerentiating it w. 0.8). and p∗ = 0.r.
(A. k ∈ Z.11). ˜ ˜ where x = ˜ q ˜ ˙ ˜ . the equilibrium point (q∗ . Hence. 2 2 (A. p) = 1 1 T q Cq + pT M −1 p. p∗ ) = (−K −1 ∆.4) are q∗ = q1 .6)
(A. 0) follows. As second step.4) are equivalent. It is a straightforward observation that q and q have the same dynamics. we show that both the closed loop systems (A. ˜ ˜ The equilibrium (invariant) points of (A. Similarly.11) = O3 −J T (q3 ) J(q3 ) −R p ∂ q Hd1 ∂ p Hd1 + O3 J T (q3 ) ∆.
.1).7)
= 0.t ‘t’.4) converges locally asymptotically to either one of the equilibrium points.9)
Hence. The set of points where Hd2 = 0 is given by p ˜ ˙ ˜ El ={ x ∈ R6 : Hd2 = 0} ˜ ={ p = 0 and q is free}. q3 ˜ ˜ ˜ ˜
T
(A. p∗ ) = (−K −1 ∆. ˙ (A. ˜ We show the equivalence of the dynamics of p and p. asymptotic stability of ˙ q ˙ p where Hd1 (q. LaSalle’s invariance principle implies that the closed loop system (A.4) are equivalent. This implies that if (A.1) and (A. (A. we get ˙ ˙ ˜ ˜ Hd2 =(∂Hd2 )T x ˜ ˜ =(∂Hd2 )T O3 −J T (q3 ) J(q3 ) −R p ˜ ∂Hd2
Appendix A
= − pT (M −1 )T R p M −1 p ≤ 0.1) and (A. 0) is global minima of the shifted Hamiltonian function associated with (A.t.
The parameters µ. respectively. hull contour displacement. The Primesystem uses the instantaneous velocity to normalize the time unit. we have mentioned that the process and measurement uncertainties are stationary Gaussian random white noise process. These forces are also called WF forces (disturbances or loads). The scaling parameters for other variables like velocity. mass density of the ﬂuid. than doing the full scale experiments in the sea. it is not used for the DP vessels or other low speed applications. . Gaussian white noise process In Chapter 2. and g represent the level of buouncy. acceleration. body mass density ratio µ = ρm . This helps to study the scaled models of the vessel and then compare and analyze the performance of the full scale vessel models. ρ. BISScaling Scaling is done to normalize the vessel steering equations of motion. and acceleration due to gravity. There are two main methods of scaling: the Primesystem of SNAME introduced in 1950 and the Bissystem introduced in 1970 by Norrbin. 109
. etc.
L
First order waveinduced forces First order waveinduced wave forces (loads or disturbances) are forces whose magnitude is proportional to the wave height and their frequency is the same as that of the incident waves. and the time unit g as the basic scaling parameters. Therefore.Appendix
B
Glossary
This appendix contains deﬁnitions and additional details of some notions used in this thesis. can be found in terms of these basic scaling parameters. For the convenience of the reader. This idea is more economical both in terms of time and money. The Bissystem uses the length unit L pp (length between fore and aft perpendicpp ulars).
” Second order waveinduced forces Second order waveinduced wave forces (loads or disturbances) are forces whose magnitude is proportional to the square of the wave height. (B. A⊗B= . Then the Kronecker (direct or tensor) product of A and B is deﬁned as the matrix a11 B · · · a1m B . Let ˙ E be the set of all points in Ω0 where V(x) = 0.2). These forces are also called LF forces (disturbances or loads). .1)
The δ in (B. This implies the mean E(w(t)) = 0. ∞ 0 if t = s if t s (B. Let ˙ V : Ω −→ R be a continuously diﬀerentiable function such that V(x) ≤ 0 in Ω0 . and covariance E(w(t)wT (s)) = Wδ(t − s). de LaSalle (16511719).2) (B. w2 (t). J. Let M be the largest invariant set in E.3)
. .4) ..2) represents the Dirac delta (or continuous time impulse response) function and is given by δ(t − s) = Kronecker Product Let A ∈ Rm×n and B ∈ R p×q . Then every solution starting in Ω0 approaches M as t → ∞. and w3 (t). . . . The vector w(t) = w2 (t) consists w3 (t) of three uncorrelated independent Gaussian white noise processes: w1 (t). ∈ Rmp×nq .110
Appendix B
w1 (t) some details of this concept are presented here. (B. am1 B · · · amn B LaSalle’s Invariance Theorem (Principle) An important result in stability theory which was ﬁrst formulated by a famous French priest and education reformer. is stated as follows “Let Ω0 ⊂ Ω be a compact set that is positively invariant with respect to (3. B.
Banks. Dynamic extension is unnecessary for stabilization via interconnection and damping assignment passivitybased control. Strand. 111
. IEEE Transactions on Automatc Control.Bibliography
[1] M. [5] J. 1998. C. Jenssen.ly 611. 43(1):101–107. and T. 2008. Mhana. 1998. Amsterdam. and R. E. 45(3):193–206. [8] P. Ortega. O. 28(2):295–306. An exact line search method for solving generalized contineoustime algebraic Riccati equation. pages 183–186. D. Proceedings of IFAC/IFIP Symposium on Automation in Oﬀshore Oil Field Operation. and S. Astolﬁ and R. 58(2):133–135. 1992. Anderson. G. 37(2):177–218. DiPrima. [6] H. Nonlinear vectorial observer backstepping with integral action and wave ﬁltering for ships. [2] B. [3] A. F. Astolﬁ. Benner and R. Elementary Diﬀerential Equations and Boundary Value Problems. P. Inc. Optimal control and stabilization of nonlinear systems. B. J. D. 2009. [7] S. October 2027. The Netherlands. 2002. Dynamic positioning using kalman ﬁltering and optimal control theory. Byers. 9(3):179–196. 2007. John Wiley & Sons. Systems & Control Letters. T. [4] A. A. Aarset. Systems & Control Letters. Computational Optimization and Applications. I. Japan. Tran. Chhabra. P. T. Sælid. Fossen. International Journal of Control. 1978. M. Fukuoka. [9] W. Secondorder convergent algorithms for the steadystate Riccati equation. Lewis. N. Balchen. Eighth edition. and H. J. Nonlinear feedback controllers and compensators: a statedependent Riccati equation approach. Banks and K. 2004. IMA Journal of Mathematical Control and Information. Ortega. Proceedings of IFAC Conference on Control Applications in Marine Systems CAMS’98. Asymptotic stabilization of some equilibria of an underactuated underwater vehicle. Boyce and R..
R. Proceedings of American Control Conference. Cambridge University Press. Springer. 45(8):1910–1916. I. Univesitat Polit` cnica de Catalunya. I. 6(1):121–128. 1984. [23] T. 2006. ¸ Proceedings of the International Federation of Automatic Control (IFAC) World Congress. Linear Algebra and its Applications. Analytic feedback control and the algebraic Riccati equation. W. IEEE Transactions on Control Systems Technology. I. Guidance and Control of Ocean Vehicles. Perez. Faltinsen. 1998. H. Hydrodynamics of High Speed Marine Vehicles. Beavers Jr. Some numerical considertions and Newton’s method revisited for solving algebraic Riccati equations. [19] A. Datta. 36(5):608–616. [25] T. [24] T. D` riaCerezo. Barcelona. 2003. 396:55–80. K. Fossen. [13] B. Grøvlen. Elsevier Academic Press. Fan. M.112
BIBLIOGRAPHY
[10] T. IEEE Transactions on Automatic Control. Y. State dependent Riccati equation technique: an overview. The matrix sign function and computations in systems. Cloutier. 2009. Pan. Seoul. W. Fossen and Å. On the addition of the integral action to portcontrolled Hamiltonian systems. E. Kalman ﬁltering for positioning and heading control of ships and oﬀshore rigs. [18] A. Dieci. Junco. D. pages 932–936. 2002. I. Fossen and T. Nonlinear output feedback control of dynamically positioned ships using vectorial observer backstepping. N. Chu. IEEE Control Systems Magazine. Dynamic Positioning Systems: Principles. Simulation and Control of a DoublyFed Induco tion Machine Controlled by a BacktoBack Converter. Cimen. [15] D. 1990. Design and Applications. June 46. [16] L. N. 1976. Fossen.
. [12] J. South Korea. 29(6):32–46. y ¯ Editions Technip. Fa¨ . PhD thesis. Navigation and Control of Ships. Applied Mathematics and Computations. State dependent Riccati equation control: A survey. 2005. Rigs and Underwater Vehicles. and W. 2008. Delchamps. Denman and A. 2009. 2009. Control of Ships and Underactuated Vehicles. Inc. Automatica.. A structure preserving doubling algorithm for continuoustime algebraic Riccati equations. 1991. John Wiley & Sons. 2(1):63–94. Donaire and S. Do and J. Inc.. [21] H. 29(11):1031–1033. [22] T. New Mexico. F. 1994. pages 3761–3775. Spain. Marine Control Systems: Guidance. Modeling. July 611. 2005. John Wiley & Sons. [14] D. Numerical Methods for Linear Control Systems. USA. [17] K. [11] E. e [20] O. Lin. 1997.
[33] R. On the nonexistence. Automatica. [36] K. existence. Prentice Hall. and G. L. Dynamic ship positioning using a selftuning Kalman ﬁlter. Giacomini. K. J. Predictive Controler Design and Identiﬁcation in Connection with Steering Dynamics. Control. and D. Technical University of Denmark. Katayama. C. Patton. 2010. 1996. Obinata. [29] W. I. D. Llibr. 1996. 1998. 127(3):93–102. Sagami.BIBLIOGRAPHY
113
[26] T. Nonlinear sampleddata stabilization of dynamically positioned ships. Golub and C. and G. Garrard and J. 2004. T. Third edition. 1977. The Johns Hopkins University Press. IEEE Transactions on Control Systems Technology. [35] A. Kobayashi. PhD thesis. Modeling. J. 21(5):767–773. Passive nonlinear observer design for ship using Lyapunov methods: Fullscale experiments with a supply vessel. S. 9(2):501–516. 20(2):191–202. Grimble. Use of Kalman ﬁltering techniques in dynamic shippositioning systems. [41] M. Atlantis. M. Imae. Wise. R. 17(2):153–165. IEE Proceedings. 1996. and M. Sørensen. Kimura. USA. Grimble.
. Strand. December 1417. Denmark. Baltimore. and Control. Haberman. 13(5):497–505. 229(5):1057–1075. Boletin Sociedad Matem´ tica Mexicana. Hammett. zhai. Van Loan. Viano. [34] K. [32] M. Identiﬁcation of dynamically positioned ships. 1983. [40] H. Identiﬁcation. Doki. E. Fung and M. The Bahamas. a [39] H. Hiramoto. [28] TK. 28(3):339–350. Hall. Sagatun. 35(1):3–16. Nonlinear control design method based on statedependent Riccati equation SDRE via quasinewton method. 2002. I. Elementary Applied Partial Diﬀerenrial Equations With Fourier Series and Boundary Value Problems. 1999. IEEE Transactions on Automatic Control. J. P. S. and A. [38] R. Fossen. Proceedings of the 43rd IEEE Conference on Decision and Control. 1998. H. P. Hansen. I. Nonlinear Systems. D. Contributions to the theory of optimal control. Fossen and J. Prentice Hall. Kalman. 1989. and D. 1960. F. Journal of Guidance. Controllability issues in nonlinear state dependent Riccati equation control. 1980. B. [31] G. Nonlinearity. Khalil. [37] J. Third edition. Jordan. Journal of Sound and Vibration. D. [27] T. and uniqueness of limit cycles. Doubling algorithm for continuoustime algebraic Riccati equation. Ridgely. International Journal of Systems Science. 18(2):463–468. [30] H. Matrix Computations. J. Optimal sensor/actuator placement for active vibration control using explicit solution of albegraic Riccati equation. 1996. H. Automatica. 5:102–119. 2000. Design of nonlinear automatic ﬂight control systems. J. and Dynamics.
[55] G. W. W. [48] P. Laub. Panteley. 24(6):913–921. [52] S. 26(3):319–323. 1998. Third edition. 8(45):401–433. AK. 2000. [45] A. 1991.
. A counter example to a recent result on the stability of nonlinear systems. A Schur method for solving algebraic Riccati equations. [54] S. van der Woude. van der Woude. 2009. [53] S. Mracek and J. Matrosov. [43] W. and S. Muhammad and J. Accepted for Publication. Optimal and suboptimal control of a class of nonlinear systems. Lakshmikantham. H. R. Lam. D` riaCerezo. K. 28(1):97–102. M. Passivitybased control applied to the dyo namic positioning of ships. October 1213. A stability result with application to nonlinear regulation. Sivasundaram. Master’s thesis. Taiwan. November 1997. Journal of Dynamic Systems. 2011. Proceedings of the UKACC International Conference on Control (CONTROL 2010). D` riaCerezo. W. Cloutier. Mathematical Systems Theory. Muhammad and J. 124(3):452– 456. and Control. International Journal of Robust Nonlinear Control. [51] S. Kluwer. [50] S. [46] J. Real time computational methods for sdre nonlinear control of missiles. and V. L¨ fberg. S. 2010. 1979. IMA Journal of Mathematical Control and Information. M. van der Woude. IMA Journal of Mathematical Control and Information. Olsder and J. and E. UK. Houston. Menon. Taipei. [47] A. L. IEEE Trans. VSSD. 8(2):332–344. Proceedings of Dynamic Positioning Conference. May 810. Alleyne. 2004. Control designs for the nonlinear benchmark problem via the state dependent Riccati equation method. 2005. V. Muhammad and J. Coventry. L. 2010. On stability conditions for systems with periodic state dependent coeﬃcients: a counter example. Crawford. van der Woude. USA. Vector Lyapunov Functions and Stability Analysis of Nonlinear Systems. Output feedback passivity based controllers o for dynamic positioning of ships. IET Control Theory and Applications. I. T. September 24. A separation principle for dynamic positioning of ships: theoretical and experimental results. USA. o Proceedings of the CACSD Conference. USA. T.114
BIBLIOGRAPHY
[42] V. 2002. September 710. Measurement. 2002. Loria. The Fourier series interpolation method for the solution of the SDARE. Cheng. Langson. [44] W. W. Proceedings of the American Control Conference Workshop. P. [49] C. YALMIP: A toolbox for modeling and optimization in MATLAB. Muhammad and A. IEEE Transactions on Automatic Control. Graduate College of the University of Illinois UrbanaChampaign. Langson and A. Anchorage. on Control Systems Technology. Fossen. J. Muhammad and A.
Chicago. June 46. VA. E. W. Snijders. B. Maschke. Altitude control of a satellite using the SDRE method. June 2830. Stankovic. Snijders and J. A robustly stable output feedback saturated controller for the boost DCtoDC converter. Proceedings of the American Control Conference. J. Simon. Spong. G. R. Systems and Control Letters. Ortega. Master’s thesis. 2000. Interconnection o and damping assignment passivitybased control of portcontrolled Hamiltonian systems. Nonlinear control of magnetic levitation systems via energybalancing. and G. Ortega. Rodriguez. [65] H. Weiland. 2004. W. Perez. and A. Garc´Canseco. G. Ridgely. 1989. A. [57] R. [58] R. Wave ﬁltering and thruster allocation for dynamic positioned ships. 2000. EWI Faculty. Escobar. Nonlinear observer design for dynamic positioning. The Netherlands. An energyshaping approach to excitation control of synchronous generators. Barabanov. [63] T. John Wiley & Sons. and I. June 2527. Springer. Scherer and S. IL. Ortega and M.. [60] R. and N. van der Woude. Journal of Electronics and Control. 47(8):1218–1233. 1997. G´ mezEstern. M. Parrish and D. 2002. D. Proceedings of the American Control Conference. 2005. F.BIBLIOGRAPHY
115
[56] R. 1991. Houston. Prentice Hall International Inc. 2002. Optimal State Estimation. DIAM. Spong. [66] C. Pearson. M. 38(4):585–596. 13(5):453–465. Arligton. Ship Motion Control. 10(5):432–450. Applied Nonlinear Control. [67] D. 2006.
. Slotine and W. Interconnection and damping assignment ı passivitybased control: A survey. Interconnection and damping assignment passivitybased control of portcontrolled Hamiltonian systems. 25(6):877–888. K. van der Schaft. 2005. Galaz. USA. USA. R. New Mexico. [61] D. Mareels. 2006. September 2728. Inc. European Journal of Control. [62] J. Ortega. USA. TU Delft. Automatica. [70] J. 40(1):1–8. Blankenstein. Proceedings of Dynamic Positioning Conference. [69] J. Automatica. Ortega and E. [64] H. Rodriguez. IEEE Transactions on Automatic Control. Li.. Proceedings of the American Control Conference. Approximation methods in optimal control. and F. Mathematical Systems Theory Group. Ortega. USA. 2001. [68] J. November 2005. 1962. Adaptive motion control of rigid robots: A tuttorial. Lecture Notes: DISC Course on Linear Matrix Inequalities in Control. B. A. Bazanella. Ortega. [59] R.
Sturm. [75] G. Zaher. 1982. and M. 2004. 45(6):1457–1466. J. Research Report 15. 1996. [80] F. Design of a dynamic positioning system using modelbased control. [78] P. Fossen.116
BIBLIOGRAPHY
[71] A. Rotea. Sørensen. L2 −Gain and Passivity Techniques in Nonlinear Control. [82] A. Corless. Adaptive Steering of Ships: a model reference approach to improved manoevring and economical course keeping. [81] V. Spouge. S. SIAM. [73] J. 12(8):1518–1533. Tsiotras. Nonlinear dynamic positioning of ships with gainscheduled wave ﬁltering. The Bahamas. 11(1):625–653. Atlantis. Consulting. Delft. and T. 2004. Springer. Torsetnes. Review of methods for demonstrating redundancy in dynamic positioning systems for the oﬀshore industry. Using SeDuMi 1. London. Econometrica. [72] J. Nonlinear Diﬀerential Equations and Dynamical Systems. 1987. van der Schaft. Second edition. [79] A.
. 1999. I. [77] J. J. M. December 1417. The Netherlands. Vidyasagar. Nonlinear control of systems with multiple equilibria and unknown sinusoidal disturbance. van Amerongen. Nonlinear Systems Analysis. [76] P. F. 13(2):129–130. PhD thesis. Delft. a MATLAB toolbox for optimization over symmetric cones. [74] V. Communications in Nonlinear Science and Numerical Simulation. I. Springer. Torre. A. and T. Fossen. 1996. Rudder Roll Stabilization. I. 4(3):359–368.01. DNV. 1977. 2007. 2000. IMA Journal of Mathematical Control and Information. Delft University of Technology. J. Jouﬀroy. 2002. Verhulst. 1996. Optimization Methods and Software. Sagatun. Proceedings of the 43rd IEEE Conference on Decision and Control. The Netherlands. PhD thesis. Existence of limit cycles and control in complete Keynesian system by theory of bifurcations. Delft University of Technology. van der Klugt. Control Engineering Practice. Counter example to a recent result on the stability of nonlinear systems.
Storage and Oﬄoading InterconnectionDampingAssignment KalmanFilter LinearMatrixInequality LowFrequency LinearQuadratic LinearQuadraticGaussian LinearQuadraticRegulation MATrixLABoratory OptimalControlProblem PassivityBasedControl PortHamiltonianSystems RealSchurForm StateDependentAlgebraicRiccatiEquation StateDependentCoeﬃcient SelfDualMinimization WaveFrequency YetAnotherLinearMatrixInequalityParser
117
.List of Abbreviations
DP EKF FPSO IDA KF LMI LF LQ LQG LQR MATLAB OCP PBC PHS RSF SDARE SDC SeDuMi WF YALMIP
DynamicPositioning ExtendedKalmanFilter Floating Production.
118
.
DP control system. The focus of this research is the control system design for DP vessels. As the measurement signal plays an important role in the control design procedures. The vessel motion in sea is divided into two motion components: a low frequency (LF) motion component and a wave frequency (WF) motion component. This was usually achieved by using conventional techniques like anchoring and mooring. The LF motion is due to the slowly varying forces and moments due to the control actions.Summary
Dynamic positioning (DP) refers to maintaining the position and heading of a sea vessel under speciﬁed limits exclusively by using thrusters and propellers. Since then more sophisticated DP systems are being used for dynamic positioning. A vessel whose motion is controlled by DP system is called a DP vessel. This is to avoid unnecessary wear and tear of the propulsion units. wind. In this thesis. This problem of separating LF motion component from WF motion component is called wave ﬁltering. waves. we ﬁrst study the local and global aspects of the asymptotic stability of the DP vessel. Some results on the global asymptotic stability studies of similar systems are recapitulated. the demand for more accurate and stable DP systems has increased to ensure safety of the DP operations. and thruster system. and sea currents. Starting point of the study of the DP control system design is a mathematical model of the sea vessel. With the fast growing oﬀshore industry. and exponentially bounded. An important aspect of the control system design is its stability. An important consideration in the control system design for many dynamic positioning problems is to counter only the LF motion of the vessel. Hurwitz. The measurement signal from the sensors (GPS and gyro) essentially contains both motion components. until late sixties. We then present a counter example showing that the conditions that the system matrix is continuous. The WF motion is caused by the ﬁrst order wave induced wave loads. The socalled state dependent coeﬃcient (SDC) parametrization will transform the nonlinear vessel model into a pseudolinear form with the system matrix being function of the heading angle. as reported in the literature on global asymptotic stability analysis of such systems are not suﬃcient to establish the global asymptotic stability of the 119
. ﬁltering of the WF component from the measurement signal is essential so that only the LF component is transmitted to the controller. A DP system consists of three main units: power system. The vessel model is nonlinear and this nonlinearity is due to the heading angle of the vessel.
The former controller provides the desired performance in the nominal (all the external disturbances are known or there are no disturbances) case but it fails to stabi
. Therefore. we propose a technique to prove the global asymptotic stability of the pseudolinear system. We start with a nominal case when there is no WF motion involved and measurements are noise free. Therefore. The SDARE control design is presented to regulate the vessel to the desired equilibrium position. We present the Fourier series interpolation (FSI) method which reduces the computation time of the algebraic Riccati equation. the socalled state dependent algebraic Riccati equation (SDARE) technique is used to address the regulation problem of the DP vessel.120
Summary
pseudolinear systems whose system matrix depends on a single state component. the disturbance vector is concatenated with the state of the vessel model. In this context. The idea of the method discussed in this thesis is limited to the special case of the SDARE whose coeﬃcient matrices depend on a single state component in a periodic way. This technique is a combination of the linear matrix inequalities (LMIs) and the Lyapunov stability theory. we study only the control design problem. this still does not make it the set of suﬃcient conditions for the global asymptotic stability of the pseudolinear system. This is especially useful in the situations where the solution of the SDARE is required online. We used this technique to prove the global asymptotic stability of the nominal vessel model with a naive state feedback controller. The output equation in the model of DP vessel indicates that only the noisy position and heading measurements are available for feedback. The underlying idea behind the fast computation of the solution of the SDARE is the oﬄine computation of the Fourier coeﬃcients which eﬀectively reduces the online computations. we show that in addition to the above mentioned set of conditions with the ﬁrst counter example. In the second counter example. Therefore. the noisy measurements are contaminated with the WF component as well which requires wave ﬁltering. In this case. We conclude the SDARE control design concept by taking into account the WF component in the measurement model. The interconnection damping assignment passivity based control (IDAPBC) design technique is used to introduce two control laws: the classical IDAPBC design and the extended IDAPBC design. if we assume the periodicity of the system matrix. The SDARE control design is a state feedback method. The estimation and wave ﬁltering are assumed to be addressed separately and we move on from this point forward to address the control design problem. The regulation problem of the DP vessel is also addressed by using the port Hamiltonian formulation. knowledge of the entire state of the system is required. We use the SDARE technique to address the estimation and ﬁltering problems. an observer is required to estimate the complete state of the system. The pseudolinear SDARE observer is used to get the state estimates. We then proceed with the socalled nonnominal case in which disturbance vector is unknown and measurements also contain noise. Simulation experiments performed by using the data of a supply vessel have shown good results as far as the performance of the SDARE controller and observer is concerned. In this thesis. For the case of a periodic dependence of the system matrix on a single state component. The disturbance vector is modeled as a ﬁrst order Markov process. The fast computation of the SDARE may be required in the cases when this technique is used with the systems whose dynamic response is fast/high. The eﬀectiveness of the FSI method is shown in the SDARE approach for DP control system design.
The study of the port Hamiltonian framework for ﬁltering and estimation problems of the DP vessel is also an active research problem. In the study of the DP problem. the search for the suﬃcient conditions for the global asymptotic stability of the pseudolinear systems with state dependency of the system matrix on a single state component.Summary
121
lize the system at the desired location in presence of unknown external disturbances. In conclusion. The second feature of the port Hamiltonian formulation is that it oﬀers the freedom to add additional damping through the controller action which helps in improving the performance of the controller and simpliﬁes the tuning of the gains. is still open for further research. This study is mainly limited to the regulation of the DP vessel to a desired equilibrium point. Simulation results have revealed that each formulation gives a diﬀerent heading angle proﬁle of the vessel in cases when the initial heading angle is above π radians. the extended IDAPBC design is introduced to yield the desired performance. The proﬁle corresponding to the trigonometric formulation gives the optimal heading angle proﬁle. The ideas and techniques introduced in this dissertation can be employed also to study various other DP problems like path following. trajectory tracking.
. we consider two diﬀerent formulations of the Hamiltonian function: a quadratic and a trigonometric. The ﬁrst one is that there is an additional degree of freedom in the form of the Hamiltonian (also sometimes called as energy of the system) function which helps in improving the performance and analysis of the controller. To address this bottleneck in the design. etc. There are two notable features in the port Hamiltonian based study of the DP problem. The DP problem has a wide range of issues associated with it.
122
.
Samenvatting
add dutch summary here
123
.
124
.
IMA Journal of Mathematical Control and Information. pp 319323. 26. To appear in IET Journal of Control Theory and Applications. ` SHAH MUHAMMAD AND ARNAU DORIACEREZO Output Feedback Passivitybased Controllers for Dynamic Positioning of Ships.
125
. September 710. 2010. IMA Journal of Mathematical Control and Information. 2010. In proceedings of the UKACC International Conference on Control (CONTROL 2010). pp 97102. March 2011.Publications
SHAH MUHAMMAD AND JACOB VAN DER WOUDE A counter example to a recent result on the stability of nonlinear systems. October 1213. In Proceedings of the 14th Edition of the Dynamic Positioning Conference (DP CONFERENCE 2010). 3. 1. No. Coventry. ` SHAH MUHAMMAD AND ARNAU DORIACEREZO Passivity Based Control Applied to the Dynamical Positioning of Ships. SHAH MUHAMMAD AND JACOB VAN DER WOUDE On stability conditions for systems with periodic state dependent coeﬃcients. USA. Vol. August 2009. No. UK. Houston. SHAH MUHAMMAD AND JACOB VAN DER WOUDE The Fourier Series Interpolation (FSI) Method for the Solution of the SDARE. Vol. 28.
126
.
Sc. he won the ﬁrst prize in 5 Km long distance race and ended up third in 1500 m race. Benelux meetings on systems and control (2009 and 2010 editions). Shah intends to work for the promotion of research culture especially in the developing and under developed societies. in 1997 from the university of the Punjab and subsequently he got M. Prominent activities include his participation and presentation in the 1st Elgersburg School on Mathematical Systems Theory.
127
. This stint ended on June 30. Shah’s post PhD plans are to join some dynamic and growing organization/institute where he can use his knowledge and expertise acquired during his academic and research career. Sports and reading formed Shah’s extraacademic activities during his student life. and the DP Conference (2010 edition) held in Houston. Since September 2006. At Islamia University Bahawalpur. Germany. in Applied Mathematics from Islamia University Bahawalpur in 1999. he has been in the Netherlands. 2005. He likes to share and transfer his experience to the prospective colleagues/students.Sc. He has presented several papers during his PhD in various seminars and conferences related to dynamical systems and control theory.Curriculum Vitae
Shah Muhammad was born on the 6th of January. organized by Technische Universit¨ t Ilmenau on March 30April 3. Later on he worked as a Scientiﬁc Oﬃcer for four years in Ghulam Ishaq Khan (GIK) Institute of Engineering Sciences and Technology. He believes that research culture is a key to solve the problems of daily life and to overcome the challenges of modern era. 1977 in Talagang. He got his primary and secondary schooling mainly from Government Schools in his native town. He is also a very agile ﬁelder and has a very safe pair of hands any where in the out ﬁeld and in the slip cordon. Topi. Pakistan. Shah is a strong believer that sports are indispensable part of an individual’s life and he looks forward to promote sporting culture among the masses and society as a whole. He is a good player of cricket. 2009 a Elgersburg. USA. He has reviewed a number of articles in his ﬁeld and has two published and one accepted journal article. Pakistan. UKACC Conference on Control held in Coventry. Delft University of Technology. He did his B. UK in 2010. pursuing his PhD in the Mathematical System Theory and Optimization group at the Delft Institute of Applied Mathematics. He is a ﬂuent top order batsman who likes to play slog sweep and reverse sweep with profound eagerness.